2021 Book AdvancesInConditionMonitoringA

Lecture Notes in Mechanical Engineering
Len Gelman
Nadine Martin
Andrew A. Malcolm
Chin Kian (Edmund) Liew Editors
Advances in
Condition
Monitoring and
Structural Health
Monitoring
WCCM 2019
Series Editors
Francisco Cavas-Martínez, Departamento de Estructuras, Universidad Politécnica
de Cartagena, Cartagena, Murcia, Spain
Fakher Chaari, National School of Engineers, University of Sfax, Sfax, Tunisia
Francesco Gherardini, Dipartimento di Ingegneria, Università di Modena e Reggio
Emilia, Modena, Italy
Mohamed Haddar, National School of Engineers of Sfax (ENIS), Sfax, Tunisia
Vitalii Ivanov, Department of Manufacturing Engineering Machine and Tools,
Sumy State University, Sumy, Ukraine
Young W. Kwon, Department of Manufacturing Engineering and Aerospace
Engineering, Graduate School of Engineering and Applied Science, Monterey,
CA, USA
Justyna Trojanowska, Poznan University of Technology, Poznan, Poland
Lecture Notes in Mechanical Engineering (LNME) publishes the latest develop-
ments in Mechanical Engineering—quickly, informally and with high quality.
Original research reported in proceedings and post-proceedings represents the core of
LNME. Volumes published in LNME embrace all aspects, subfields and new
challenges of mechanical engineering. Topics in the series include:
• Engineering Design
• Machinery and Machine Elements
• Mechanical Structures and Stress Analysis
• Automotive Engineering
• Engine Technology
• Aerospace Technology and Astronautics
• Nanotechnology and Microengineering
• Control, Robotics, Mechatronics
• MEMS
• Theoretical and Applied Mechanics
• Dynamical Systems, Control
• Fluid Mechanics
• Engineering Thermodynamics, Heat and Mass Transfer
• Manufacturing
• Precision Engineering, Instrumentation, Measurement
• Materials Engineering
• Tribology and Surface Technology
To submit a proposal or request further information, please contact the Springer
Editor of your location:
China: Dr. Mengchu Huang at mengchu.huang@springer.com
India: Priya Vyas at priya.vyas@springer.com
Rest of Asia, Australia, New Zealand: Swati Meherishi at
swati.meherishi@springer.com
All other countries: Dr. Leontina Di Cecco at Leontina.dicecco@springer.com
To submit a proposal for a monograph, please check our Springer Tracts in
Mechanical Engineering at http://www.springer.com/series/11693 or contact
Leontina.dicecco@springer.com
Indexed by SCOPUS. All books published in the series are submitted for
consideration in Web of Science.
More information about this series at http://www.springer.com/series/11236

Len Gelman Nadine Martin
• •
Andrew A. Malcolm Chin Kian (Edmund) Liew

•
Editors
Advances in Condition
Monitoring and Structural
Health Monitoring
WCCM 2019
123
Editors
Len Gelman Nadine Martin
University of Huddersfield CNRS Grenoble INP, GIPSA-Lab
Huddersfield, UK University Grenoble Alpes
Grenoble, Rhône, France
Andrew A. Malcolm
Advanced Remanufacturing Chin Kian (Edmund) Liew
and Technology Centre Singapore Institute of Technology
Singapore, Singapore Singapore, Singapore
ISSN 2195-4356 ISSN 2195-4364 (electronic)

ISBN 978-981-15-9198-3 ISBN 978-981-15-9199-0 (eBook)
https://doi.org/10.1007/978-981-15-9199-0
© Springer Nature Singapore Pte Ltd. 2021
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Contents
Signal Processing: Fault Diagnosis for Complex System Monitoring

Detection and Localization of a Gear Fault Using Automatic
Continuous Monitoring of the Modulation Functions . . . . . . . . . . . . . . . 3
X. Laval, N. Martin, P. Bellemain, and C. Mailhes
Multi-harmonic Demodulation for Instantaneous Speed Estimation
Using Maximum Likelihood Harmonic Weighting . . . . . . . . . . . . . . . . . 15
C. Peeters, J. Antoni, and J. Helsen
A Variational Bayesian Approach for Order Tracking in Rotating
Machines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Y. Marnissi, D. Abboud, and M. Elbadaoui
Reconstructing Speed from Vibration Response to Detect Wind
Turbine Faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
K. Marhadi and D. Saputra
Non-regular Sampling and Compressive Sensing for Gearbox
Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
C. Parellier, T. Bovkun, T. Denimal, J. Gehin, D. Abboud, and Y. Marnissi
Change Detection in Smart Grids Using Dynamic Mixtures
of t-Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
A. Samé, M. L. Abadi, and L. Oukhellou
Unbalance Detection and Prediction in Induction Machine Using
Recursive PCA and Wiener Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
A. Picot, J. Régnier, and P. Maussion
Automatic Fault Diagnosis in Wind Turbine Applications . . . . . . . . . . . 77
D. Saputra and K. Marhadi
Blade Monitoring in Turbomachines Using Strain Measurements . . . . . 87
D. Abboud, M. Elbadaoui, and N. Tableau
v
vi Contents
Measuring the Effectiveness of Descriptors to Indicate Faults

in Wind Turbines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Gearbox Condition Monitoring Using Sparse Filtering and
Parameterized Time–Frequency Analysis . . . . . . . . . . . . . . . . . . . . . . . . 105
Shiyang Wang, Zhen Liu, and Qingbo He
Rotating Machinery Fault Diagnosis with Weighted Variational
Manifold Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Quanchang Li, Xiaoxi Ding, Wenbin Huang, and Yimin Shao
Recent Advances on Anomaly Detection for Industrial Applications

A Data-Driven Approach to Anomaly Detection and Health
Monitoring for Artificial Satellites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Takehisa Yairi, Yusuke Fukushima, Chun Fui Liew, Yuki Sakai,
and Yukihito Yamaguchi
Robust Linear Regression and Anomaly Detection in the Presence
of Poisson Noise Using Expectation-Propagation . . . . . . . . . . . . . . . . . . 143
Yoann Altmann, Dan Yao, Stephen McLaughlin, and Mike E. Davies
Spacecraft Health Monitoring Using a Weighted Sparse
Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Barbara Pilastre, Jean-Yves Tourneret, Loïc Boussouf,
and Stéphane D’escrivan
When Anomalies Meet Aero-elasticity: An Aerospace Industry
Illustration of Fault Detection Challenges . . . . . . . . . . . . . . . . . . . . . . . 169
P. Goupil, J.-Y. Tourneret, S. Urbano, and E. Chaumette
Anomaly Detection with Discrete Minimax Classifier for Imbalanced
Dataset or Uncertain Class Proportions . . . . . . . . . . . . . . . . . . . . . . . . . 179
Cyprien Gilet and Lionel Fillatre
A Comparative Study of Statistical-Based Change Detection Methods
for Multidimensional and Multitemporal SAR Images . . . . . . . . . . . . . . 189
Ammar Mian and Frédéric Pascal
Bearing Anomaly Detection Based on Generative Adversarial
Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Jun Dai, Jun Wang, Weiguo Huang, Changqing Shen, Juanjuan Shi,
Xingxing Jiang, and Zhongkui Zhu
Classification/Analysis for Fault or Wear Diagnosis

Specific Small Digital-Twin for Turbofan Engines . . . . . . . . . . . . . . . . . 219
J. Lacaille
Contents vii
Fault Diagnosis of Single-Stage Bevel Gearbox by Energy Operator

and J48 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
V. Kumar, A. K. Verma, and S. Sarangi
An Efficient Multi-scale-Based Multi-fractal Analysis Method
to Extract Weak Signals for Gearbox Fault Diagnosis . . . . . . . . . . . . . . 241
Ruqiang Yan, Fei Shen, and Hongxing Tao
A Novel Method of PEM Fuel Cell Fault Diagnosis Based
on Signal-to-Image Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Zhongyong Liu, Weitao Pan, Yousif Yahia Ahmed Abuker, and Lei Mao
Fault Classification of Polymer Electrolyte Membrane Fuel Cell
System Based on Empirical Mode Decomposition . . . . . . . . . . . . . . . . . 261
Yousif Yahia Ahmed Abuker, Zhongyoug Liu, Lisa Jackson, and Lei Mao
Fleet-Oriented Pattern Mining Combined with Time Series Signature
Extraction for Understanding of Wind Farm Response to Storm
Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
P.-J. Daems, L. Feremans, T. Verstraeten, B. Cule, B. Goethals,
and J. Helsen
Efficiency and Durability of Spur Gears with Surface Coatings
and Surface Texturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Giovanni Iarriccio, Antonio Zippo, Marco Barbieri,
and Francesco Pellicano
System Model for Condition Monitoring and Prognostic

Wheel-Rail Contact Forces Monitoring Based on Space Fixed-Point
Strain of Wheel Web and DIC Technology . . . . . . . . . . . . . . . . . . . . . . 295
Xiaochao Wang, Zhenggang Lu, Juyao Wei, and Yang He
Experimental Validation of the Blade Excitation in a Shaft
Vibration Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
V. Vasicek, J. Liska, J. Strnad, and J. Jakl
Seismic Risk Evaluation by Fragility Curves Using Metamodel
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
H. Z. Yang and C. G. Koh
Reliability Centred Maintenance (RCM) Assessment of Rail Wear
Degradation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
H. J. Hoh, J. L. Wang, A. S. Nellian, and J. H. L. Pang
Prognostic of Rolling Element Bearings Based on Early-Stage
Developing Faults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
M. Hosseini Yazdi, M. Behzad, and S. Khodaygan
viii Contents
Recent Technologies for Smart and Safe Systems

Fault-Tolerant Actuation Architectures for Unmanned Aerial
Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
M. A. A. Ismail, C. Bosch, S. Wiedemann, and A. Bierig
Towards Certifiable Fault-Tolerant Actuation Architectures
for UAVs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
C. Bosch, M. A. A. Ismail, S. Wiedemann, and M. Hajek
Development of a Patrol System Using Augmented Reality
and 360-Degree Camera . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Kenji Osaki, Tomomi Oshima, Naoya Sakamoto, Tetsuro Aikawa,
Yuya Nishi, Tomokazu Kaneko, Makoto Hatakeyama,
and Mitsuaki Yoshida
Research on Evaluation Method of Safety Integrity for Safety-Related
Electric Control System of Amusement Ride . . . . . . . . . . . . . . . . . . . . . 375
Weike Song, Peng Cheng, and Ran Liu
Study on Warning System of Transportable Pressure Equipment
from Regulatory Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
Z. R. Yang, Z. Z. Chen, Z. W. Wu, and P. Yu
Industry 4.0: Why Machine Learning Matters? . . . . . . . . . . . . . . . . . . . 397
T. H. Gan, J. Kanfoud, H. Nedunuri, A. Amini, and G. Feng
Condition Monitoring of Electromechanical Systems

A Method for Online Monitoring of Rotor–Stator Rub in Steam
Turbines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
J. Liska and J. Jakl
Implementation of Instantaneous Power Spectrum Analysis
for Diagnosis of Three-Phase Induction Motor Faults Under
Mechanical Load Oscillations; Case Study of Mobarakeh
Steel Company, Iran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
S. Mani, M. Kafil, and H. Ahmadi
Analysis of Unbalanced Magnetic Pull in a Multi-physic Model
of Induction Machine with an Eccentric Rotor . . . . . . . . . . . . . . . . . . . 429
Xiaowen Li, Adeline Bourdon, Didier Rémond, Samuel Koechlin,
and Dany Prieto
Bearings Fault Diagnosis in Electromechanical System Using
Transient Motor Current Signature Analysis . . . . . . . . . . . . . . . . . . . . . 439
Alok Verma, Haja Kuthubudeen, Viswanathan Vaiyapuri,
Sivakumar Nadarajan, Narasimalu Srikanth, and Amit Gupta
Contents ix
SHM/NDT Technologies and Applications

Assessment of RC Bridge Decks with Active/Passive NDT . . . . . . . . . . . 451
T. Shiotani, H. Asaue, K. Hashimoto, and K. Watabe
Estimation of In-Plane Strain Field Using Computer Vision to Capture
Local Damages in Bridge Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
T. J. Saravanan and M. Nishio
Wireless Strain Sensors for Dynamic Stress Analysis of Rail Steel
Structural Integrity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
X. P. Shi, Z. F. Liu, K. S. Tsang, H. J. Hoh, Y. Liu, Kelvin Tan Kaijun,
Phua Song Yong, Kelvin, and J. H. L. Pang
Purpose and Practical Capabilities of the Metal Magnetic
Memory Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 481
A. A. Dubov, S. M. Kolokolnikov, M. N. Baharin, M. F. M. Yunoh,
and N. M. Roslin
Development and Application of Inner Detector in Drilling Riser
Auxiliary Lines Based on Magnetic Memory Testing . . . . . . . . . . . . . . . 489
Xiangyuan Liu, Jianchun Fan, Wei Zhou, and Shujie Liu
Defect Recognition and Classification Techniques for Multi-layer
Tubular Structures in Oil and Gas Wells by Using Pulsed Eddy
Current Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Hang Jin, Hua Huang, Zhaohe Yang, Tianyu Ding, Pingjie Huang,
Banteng Liu, and Dibo Hou
Corrected Phase Reconstruction in Diffraction Phase Microscopy
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
V. P. Bui, S. Gorelik, Z. E. Ooi, Q. Wang, and C. E. Png
Study on Fatigue Damage of Automotive Aluminum Alloy Sheet
Based on CT Scanning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
Ya li Yang, Hao Chen, Jie Shen, and Yongfang Li
IoT-Web-Based Integrated Wireless Sensory Framework
for Non-destructive Monitoring and Evaluation of On-Site
Concrete Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527
A. O. Olaniyi, T. Watanabe, K. Umeda, and C. Hashimoto
A Study of Atmospheric Corrosive Environment in the North East
Line Tunnel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
L. T. Tan, S. M. Goh, H. Y. Gan, and Y. L. Foo
x Contents
Acoustic Emission
Fatigue Crack Growth Monitoring Using Acoustic Emission:
A Case Study on Railway Axles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
R. Marks, A. Stancu, and S. Soua
Acoustic Emission Monitoring for Bending Failure of Laminated
Glass Used on Glass Water Slide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
Ran Liu, Gongtian Shen, Yong Zhang, Wei Zhou, and Pengfei Zhang
Source Location of Acoustic Emission for Anisotropic Material
Utilizing Artificial Intelligence (WCCM2019) . . . . . . . . . . . . . . . . . . . . . 573
Y. Mizutani, N. Inagaki, K. R. Kholish, Q. Zhu, and A. Todoroki
Standard Development and Application on Acoustic Emission
Condition Monitoring of Amusement Device . . . . . . . . . . . . . . . . . . . . . 579
J. J. Zhang, G. T. Shen, Z. W. Wu, and Y. L.Yuan
Fault Diagnosis of Pitch Bearings on Wind Turbine Based on Acoustic
Emission Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
Guang-hai Li, Yang ming An, Tao Guo, and Guang-quan Hou
Ultrasonic Testing
In Situ Condition Monitoring of High-Speed Rail Tracks Using Diffuse
Ultrasonic Waves: From Theory to Applications . . . . . . . . . . . . . . . . . . 601
K. Wang, W. Cao, and Z. Su
Theoretical Analysis of Scattering Amplitude Calculation of Torsional
Wave on a Pipe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611
Jaesun Lee, Jan D. Achenbach, and Younho Cho
Detection of a Micro-crack on a Rotating Steel Shaft Using Noncontact
Ultrasonic Modulation Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . 617
I. Jeon, P. Liu, and H. Sohn
Measurement of Axial Force of Bolted Structures Based on Ultrasonic
Testing and Metal Magnetic Memory Testing . . . . . . . . . . . . . . . . . . . . 625
Siqi Yang, Laibin Zhang, and Jianchun Fan
Ultrasonic Testing of Anchor Bolts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
Y. M. Tong, A. R. K. Rajkumar, H. G. Chew, and C. K. Liew
Weld Inspection of Piping Elbow Using Flexible PAUT Probe . . . . . . . . 647
S. J. Lim, T. S. Park, M. Y. Choi, and I. K. Park
Laser Ultrasonic Imaging of Wavefield Spatial Gradients
for Damage Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655
S. Y. Chong, Z. Wu, and M. D. Todd
Contents xi
Guided Waves Testing

Guided Ultrasonic Waves in Glass Laminate Aluminium
Reinforced Epoxy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 667
M. Rennoch, M. Koerdt, A. S. Herrmann, N. Rauter, and R. Lammering
Ultrasonic Guided Waves in Unidirectional Fiber-Reinforced
Composite Plates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
Ductho Le, Jaesun Lee, Younho Cho, Duy Kien Dao,
Truong Giang Nguyen, and Haidang Phan
Mode Selectivity and Frequency Dependence of Guided Waves
Generated by Piezoelectric Wafer Transducers in Rebars
Embedded in Concrete . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687
Rajeshwara Chary Sriramadasu, Ye Lu, and Sauvik Banerjee
How Temperature Variations Affect the Propagation Behaviour
of Guided Ultrasonic Waves in Different Materials . . . . . . . . . . . . . . . . 695
M. Rennoch, J. Moll, M. Koerdt, A. S. Herrmann, T. Wandowski,
and W. M. Ostachowicz
Nonlinearity Study of Damaged Spring Materials Using
Guided Wave . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
Jeongnam Kim, Yonghee Lee, Sungun Lee, and Younho Cho
Transducer Arrangement for Baseline-Free Damage Detection
Methods With Guided Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
U. Lieske, T. Gaul, and L. Schubert
Mode Switchable Guided Elastic Wave Transducer Based
on Piezoelectric Fibre Patches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723
Y. Kim, K. Tschöke, L. Schubert, and B. Köhler
Lamb Wave Excitation Using a Flexible Laser Ultrasound Transducer
for Structural Health Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
Wei Li, Jitao Xiong, and Wenbin Huang
Lamb Wave Propagation of S0 and A0 Modes in Unidirectional CFRP
Plate for Condition Monitoring: Comparison of FEM Simulation
with Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 741
Subin Kim, Joonseo Song, Seungmin Kim, Younho Cho,
and Young H. Kim
Thermographic Testing
Three-Dimensional Reconstruction of Leaked Gas Cloud Image
Based on Computed Tomography Processing of Multiple Optical
Paths Infrared Measurement Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 751
M. Uchida, D. Shiozawa, T. Sakagami, and S. Kubo
xii Contents
Experimental Analysis on Fault Prediction of AC-Contactor

by Infrared Thermography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Yue Yu, Wei Li, Chao Ye, and Bin Hu
Research on the Detecting Method of Motor Running State
of Amusement Device Based on Infrared Thermography
Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 769
C. Ye, Y. Yu, J. J. Zhang, R. Liu, and G. T. Shen
Application of Ultrasound Thermometry to Condition Monitoring
of Heated Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 781
I. Ihara, R. Sawada, Y. Ogawa, and Y. Kawano
Gas Leak Source Identification by Inverse Problem Analysis
of Infrared Measurement Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 791
H. Nishimura, D. Shiozawa, T. Sakagami, and S. Kubo
About the Editors
Len Gelman, Ph.D., Dr. of Sciences (Habilitation) joined Huddersfield University

as Professor, Chair in Signal Processing/Condition Monitoring and Director of
Centre for Efficiency and Performance Engineering, in 2017 from Cranfield
University, where he worked as Professor and Chair in Vibro-Acoustical
Monitoring since 2002. Len developed novel condition monitoring technologies
for aircraft engines, gearboxes, bearings, turbines and centrifugal compressors. Len
published more than 250 publications, 17 patents and is Co-editor of 12 Springer
books. He is Fellow of: BINDT, International Association of Engineers and
Institution of Diagnostic Engineers, Executive Director, International Society for
Condition Monitoring, Honorary Technical Editor, International Journal of
Condition Monitoring, Editor-in-Chief, International Journal of Engineering
Sciences (SCMR), Chair, annual International Condition Monitoring Conferences,
Honorary Co-chair, annual World Congresses of Engineering, Co-chair,
International Congress COMADEM 2019 and Chair, International Scientific
Committee of Third World Congress, Condition Monitoring. Len is Editorial Board
member of journals Applied Sciences, Electronics, Sensors, International Journal
Insight, International Journal of Acoustics and Vibration, International Journal of
Prognostics and Health Management, International Journal of Basic and Applied
Sciences and International Journal of Pure and Applied Mathematics. He was
Chair, First World Congress, Condition Monitoring, Chair, Second World
Congress, Engineering Asset Management and Chair, International Committee of
Second World Congress, Condition Monitoring. Len is Chair of International CM
Groups of ICNDT and EFNDT and Member of ISO Technical Committee,
Condition Monitoring. Len made 42 plenary keynotes at major international con-
ferences. He was Visiting Professor at ten Universities abroad. Len received two
Rolls-Royce (UK) Awards for Innovation, COMADIT Prize (by British Institute of
NDT) for significant contribution through research/development in condition
monitoring, Oxford Academic Health Science Network Award, William Smith
Prize by UK Institution of Mechanical Engineers and USA Navy Award.
xiii
xiv About the Editors
Nadine Martin is a senior researcher at the CNRS, Nat. Centre of Scientific

Research, within GIPSA-lab, Grenoble Image Speech Signal and Automatic,
University of Grenoble Alpes, France. In the signal-processing domain, her
research interests are the analysis, detection and modeling of non-stationary signals
in time-frequency domain with a particular interest in conditional preventive
monitoring of complex systems such as wind turbines or paper machines. His
current research focuses on expert-level automatic signal processing for continuous
fault detection and localization with severity tracking. N. Martin is the author of
around 150 papers, 2 patents, 7 book chapters, and co-editor of the book
Time-Frequency Decision. Currently, she is an expert for MINALOGIC, the French
Global Innovation Cluster for Digital Technologies, Chair of ISCM, the Int. Society
for Condition Monitoring, Vice-Editor-in-Chief of IJCM, the International Journal
of Condition Monitoring and a member of a working group of the International
Committee of Nondestructive testing, ICNDT. Within ISCM, she is active in the
scientific program of Condition Monitoring conferences.
Andrew A. Malcolm is the Deputy Group Manager for the Intelligent Process
Verification Group, and Senior Technical Lead for the Non-Destructive Evaluation
Team at the Advanced Remanufacturing and Technology Centre (ARTC) under the
Agency for Science, Technology and Research (A*STAR) in Singapore. He
received his M.Sc. in Computer-Aided Engineering and his Ph.D. for a thesis on
precision measurement through fringe pattern analysis from Liverpool John Moores
University in 1987 and 1995 respectively. Prior to joining A*STAR in early 1996,
he worked for various UK and European companies developing vision systems for
non-destructive testing, process control and product inspection, and for Liverpool
John Moores University as an Industrial Research Fellow and Sessional Lecturer.
He has a background in non-destructive testing and inspection, machine and
computer vision, image processing, visual guidance systems for autonomous
vehicles, stereo vision and teleoperation control. He is currently involved with a
number of research and industry projects at ARTC developing and implementing
novel techniques in the areas of ultrasonic testing, eddy current testing, thermog-
raphy and X-ray imaging and computed tomography focusing on technology
transfer for industrial implementation. Key application domains include aerospace,
heavy machinery and fast-moving consumer goods (FMCG). His current research
interests included X-ray computed tomography for visualisation and dimensional
measurement, X-ray inspection and advanced image processing, non-destructive
testing and evaluation, application of artificial intelligence within NDE, and the role
of NDE in Industry 4.0. He is a Fellow of the Non-destructive Testing Society of
Singapore and currently chairman of the certification committee responsible for the
internationally-recognised SGNDT certification scheme accredited by the
Singapore Accreditation Council in accordance with ISO 9712:2012 and EN4179
schemes.
About the Editors xv
Chin Kian (Edmund) Liew has more than 15 years of teaching and research
experience in the field of non-destructive testing and structural health monitoring.
His early research interests were in guided wave ultrasound where he developed
pattern recognition techniques to identify defects in metal beams. Subsequently, he
researched on several emerging NDT methods to inspect helicopter composite
airframe structures for Australian Aerospace, including phased array ultrasound,
radiography, and thermography. He joined the Singapore Institute of Technology
(SIT) in 2013 where he was instrumental in establishing an Authorised Training
Organisation (ATO) at the university for the Singapore NDT (SGNDT) professional
certification. He is currently an Associate Professor at SIT where he continues to be
active in the field of NDT, delivering education and training, as well as conducting
applied research for a diverse range of industry projects including for the rail,
aerospace, civil, and marine engineering sectors.
Signal Processing: Fault Diagnosis
for Complex System Monitoring
Detection and Localization of a Gear
Fault Using Automatic Continuous
Monitoring of the Modulation Functions
X. Laval, N. Martin, P. Bellemain, and C. Mailhes
Abstract In the context of automatic and preventive condition monitoring of rotating

machines, this paper presents a case study of a naturally-worn parallel straight gear
by monitoring the evolution of the modulation functions. The Hilbert demodulation
is automatically performed considering only the frequency content of the signals
detected by the AStrion software. The gear has been worn over 3000 h with a constant
axial load. A particular focus is set on the amplitude modulation function in order to
assess its efficiency to characterize both the severity of the wear and the most worn
part of the gear. The results are confronted with on-site observation of the teeth. For
this purpose, the evolution of both amplitude and phase modulations over several
meshing harmonics are compared, as well as demodulation on both original and
residual signals. Indicators to automatically classify the wear are discussed.
Keywords Condition monitoring · Signal processing · Hilbert demodulation ·

Gear fault detection · Continuous surveillance
1 Introduction
Gearboxes are very important parts of mechanical systems. An unforeseen failure

in an industrial context can lead to important shutdowns and replacement costs.
Thus, the need of a functional and efficient condition monitoring applied to these
parts is crucial. However, the day-to-day preventive maintenance consists mainly in
monitoring standard features such as peak-to-peak, root mean square or Kurtosis.
However, these methods are limited when the goal is an early fault detection of
low vibrations embedded in strong other vibrations. A lot of technics, based on
advanced signal processing are reviewed in [1] and [2] such as wavelet [3], empirical
X. Laval (B) · N. Martin · P. Bellemain

University Grenoble Alpes, CNRS, Grenoble INP, GIPSA-Lab, 11 rue des Mathématiques,
Grenoble Campus BP46, F-38402 Saint Martin D’heres, France
e-mail: xavier.laval@gipsa-lab.fr
C. Mailhes
University of Toulouse, IRIT/INPT-ENSEEIHT/TéSA, Toulouse, France
© Springer Nature Singapore Pte Ltd. 2021 3

L. Gelman et al. (eds.), Advances in Condition Monitoring and Structural
Health Monitoring, Lecture Notes in Mechanical Engineering,
https://doi.org/10.1007/978-981-15-9199-0_1
4 X. Laval et al.
mode decomposition [4], Teager-Kaiser energy operator [5]. A common method is

the Hilbert demodulation, which efficiency has been proven by McFadden [6] and
Randall [7].
This paper presents a case study of a gearbox wear through automatic methods
with the AStrion software using the Hilbert demodulation method. AStrion is a
preventive maintenance software, developed coinjointly between the GIPSA-lab in
Grenoble, France and the IRIT in Toulouse, France. This software, which princi-
ples have been presented in [8–10], operates by automatically detecting peaks and
classifying them in harmonics and sideband families. The sideband families can
be automatically demodulated to estimate the temporal modulation functions. This
paper aims at characterizing the wear by monitoring the evolution of these functions
with simple operators requiring no a priori knowledge of the system. The aim is to
propose automatic methods to characterize spallings without mechanical modelling,
using sensors on different positions to give credit to the diagnosis. The demodula-
tion process will be shortly presented first. The GOTIX bench, from where the data
come from, will be introduced in a second part. The monitoring of the modulation
functions on a set of sensors will be presented in a third part. Specific indicators are
defined from these modulation functions. The fourth part shows the ability of the
indicator trends along timestamp signals to characterize the wear. Discussion of the
method and its limitations in operational context will be undertaken in the last part.
2 Demodulation Process in AStrion and Physical Meaning

of Modulation Functions
From the identified harmonics and sideband families, AStrion is able to demodulate
a specific family in signals where this family has been detected. The demodulation
module first applies a multi-rate filter on a band corresponding either to the number
of detected sidebands or a specific user-defined number of sidebands. This narrow
band spectrum is then shifted to zero, before applying a synchronous averaging at the
modulation frequency. Finally, the Hilbert transform is applied without removing the
deterministic part. The method has been presented in details in [11]. After demod-
ulation, 3 temporal signal functions representing the amplitude modulation (AM),
the phase modulation (PM) and the frequency modulation (FM) are given over one
cycle of the modulation frequency.
If the demodulated harmonic corresponds to a meshing frequency, the AM is the
envelope of the gearbox vibration. Its mean corresponds to the energy of the demod-
ulated harmonic. It can thus be described as the instantaneous acceleration, which is
an image of the force generated at the point of meshing. If the modulating frequency
corresponds to a shaft connected to the gearbox, once applied the synchronous aver-
aging, the AM represents the variations of the force generated by the meshing over
one rotation of the gear. This force is influenced in particular by shaft misalign-
ments, load asymetry and teeth meshing. Furthermore, a tooth with a defect such as
Detection and Localization of a Gear Fault … 5
spalling has a different stiffness compared to a healthy one [12]. Since the acceler-
ation measure is directly influenced by stiffness variation, it is expected that for a
damaged gear, the AM variations will be representative of the fault.
The FM is the derivative of the PM and represents the instantaneous frequency, its
mean being equal to the demodulated harmonic frequency. In the case of a gearbox
modulated by a shaft, after the synchronous averaging at the modulation frequency,
the FM will represent the instantaneous speed of meshing. It can be influenced for
example by fissured teeth. However, in particular for low energy harmonics, or for
strong spalling, the AM can drop close to zero and be folded, causing a 180° phase
shift and an artificial surge in the FM. Thus the FM is more problematic to use, since
it can lead to false interpretation.
3 Data from the GOTIX Experimental Bench
The GOTIX bench, hosted in the GIPSA-lab, is an experimental setup consisting in a

gearbox driven by an asynchronous motor and loaded by a DC-generator. A synchro-
nized set of data comprising vibrations, currents, voltages and angular positions, are
recorded through an OROS38 acquisition system. More information can be found
on the GOTIX website http://www.gipsa-lab.grenoble-inp.fr/projet/gotix/presentat
ion.html from which data can be downloaded. The main experiment is a long-term
natural wear experiment on a parallel straight tooth multiplier gearbox composed of
a Z gear = 57 teeth gear and a Z pinion = 15 teeth pinion. Thus, the vibration signals
are all dominated by the meshing harmonics. The different harmonics are modulated
by both the input and output shafts. AStrion is able to detect all harmonics and modu-
lations automatically, up to 57 between the strongest harmonics. AStrion results in
a frequency zoom are presented in Fig. 1. It is noticeable that the sidebands ±19 of
the gear shaft rotation are predominant. This is due to the assembly phase frequency,
generated by the fact that the number of gear teeth and pinion teeth are not prime
Z gear Z
numbers: 19 = pinion
5
= 3. In order to avoid interferences from this frequency and
its sidebands, the maximum number of sidebands has been set to 9 on each side of
the harmonic.
The bench has been rotating for over 6000 h with the same pair of gears, slowly
wearing them in a completely natural way. Recordings were performed every 5 h.
Currently, the gears are worn, with spalling marks on several teeth, but not yet out of
use. The last change in the operating conditions happened 3300 h ago for technical
resaons. Since then, the speed and load have been near constant at respectively
475 rpm and 760 Nm on the driving gear shaft. Thus, the study has been performed
on these last 3300 h, with 650 signals investigated, each signal corresponding to 80 s
of recording. Six accelerometers are located at 3 different positions, measuring 2 axes
at each position. The considered sensors in the following parts, denominated sensor 1
and sensor 2, are both radially orientated. Sensor 1 is on the gear while sensor 2 is on
the pinion. AStrion results in Fig. 1 show a predominance of the third harmonic. This
6 X. Laval et al.
Fig. 1 AStrion results in a frequency zoom of a GOTIX signal estimated. The meshing harmonics
are identified in (a), and an additional zoom shows the identified sidebands around the third
harmonics in (b)
is true for the 4 radial accelerometers on GOTIX and probably indicates a system
resonance. However, the choice here is to not take into account any resonance but to
demodulate all harmonics and to compare their similarities in order to identify the
potential faulty parts of the gear.
4 Monitoring of the Modulation Functions
In order to obtain comparable modulation functions, the signals are all synchronized
using a top-tour signal from the modulating gear shaft. Thus, resulting modulation
functions start from the same angular position of the gear. Harmonics with a sufficient
number of sidebands are all investigated, whatever their amplitude. Furthermore, the
AM and FM functions are normalized by their mean value in order to compare
only the shape of these functions and be independent from their harmonic amplitude
and frequency variations. These two features are also monitored independently by
AStrion. As can be seen on Fig. 2, for sensor 1, the different functions present a
continuous evolution as the gear is slowly wearing itself.
One can already notice trends in these plots, in particular around 200° and 280°:
at these angular positions, AM seems to present a significant evolution, either as an
Fig. 2 Evolution of AM and FM functions for 3 different harmonics (H1, H3 and H6) over the
hours of gear rotation for sensor 1. Each function length corresponds to one gear rotation, from 0°
to 360°
increase (H1) or a decrease (H3 and H6). This is also true when studying other
harmonics. However, it can be very long and fastidious to study each of these
harmonics for every gear present in a system. The evolution is also present on FM,
but can be masked by important peaks, visible on H6 in particular. These peaks are
issued from a 180° phase shift in the PM, caused by the folding of the AM when its
value comes close to zero. This effect is illustrated in Fig. 3. They can be problematic
in the context of automatic detection. They can however be used to automatically
remove AM presenting this folding, their shape being distorted.
In order to confirm the diagnostic, it can be interesting to compare accelerometers
at different positions. Figure 4 shows AM for 2 different accelerometers, among the
8 available in the system, both oriented radially: accelerometer 1 is located on the
pinion shaft while accelerometer 2 is located on the gear shaft. While differences are
visible, the trends are similar. By automatically demodulating both, it is possible to
confirm or refute a diagnostic.
Even if a confirmed operator could establish a diagnosis, the process of scan-
ning the demodulation functions for every harmonic generated by every mechanical
elements of a complex system can be very time-consuming and inefficient. Hence
8 X. Laval et al.
Fig. 3 AM and FM from the 6th harmonic for 2 folded cases. The link between the FM peaks and
the AM folding is clearly visible on datatips
Fig. 4 AM for H1 and H3 for sensors 1 and 2 from 2 different accelerometers
the necessity to have features representing the evolutions that are monitorable with
automatic classifications technics.
5 Demodulation Function Features
The aim of this part is to extract as much information as possible from the AM
and FM functions using simple features. Thus, only classical features such as peak-
to-peak (PP), root mean square (RMS) and Kurtosis (K) values have been used.
First, these features are computed for each measure on the global AM functions.
But in order to improve the method as well as the number of features available for
the automatic classification, a gliding feature is computed on each function with a

rectangular angular window of length 38° which amounts to dividing the functions
in 19 parts, with a 50% overlap between each part to avoid side effects. As shown
in [11], the filtering due to the choice of the demodulation band, ± 9 sidebands in
the present case, strongly alters the shape of the modulation functions. The choice
of 19 parts is empirical, and corresponds to Nparts = n sideband left + n sideband right + 1.
It can thus be determined automatically. When linking it to the mechanics, each
part contains 3 teeth, plus an overlapping of around 1 tooth and a half on each side.
The first six harmonics are examined. In the following, the x-axis is labelled with
zone numbers, each zone corresponding to 3 teeth of the gear as above-mentioned.
Figure 5 shows the PP feature across the 19 angular zones over the 3300 h of rotation.
The AM functions of H1, H3 and H6 highlight angular zones where PP increases, in
particular in zones 9 and 10, reflecting the slow wearing of the gear.
However, this phenomenon is not so clear on the FM evolution. As seen in Fig. 5,
H1 FM does not present any evolution, while H3 shows some maximum in zones 10
Fig. 5 Gliding Peak-to-peak features for 3 harmonics on amplitude and frequency modulations
functions and for all the time-stamps. The blank lines correspond to the signals for which the
modulation sidebands have not been detected. The x-axis corresponds to 19 angular zones of 3
teeth each
10 X. Laval et al.
Fig. 6 For harmonics H1 to H6, PP, RMS and Kurtosis on the whole amplitude modulation function
(a), (b), (c). PP, RMS and Kurtosis for zone 10 (d), (e), (f). PP for zones 3, 13 and 19 (g), (h), (i)
and 11, but it does not evolve continuously. Only H6 shows an evolution. Focusing on
the AM, Fig. 6 confirms the interest of the PP feature, both on the whole AM and on
different zones. For clarity and manual comparison purpose, the features have been
normalized (divided by their mean value) for every figure. In an automatic context,
this step may not be necessary. However, a median filter has also been applied to
smooth the results. Since its purpose is also to improve the classification results, this
step is more important.
The PP value on the whole AM (a) steadily increases for 4 harmonics. Further-
more, it increases for 5 harmonics in zone 10 (d), all of them starting their increase
around the same moment, between 2000 and 2400 h of rotation. Only H4 in fact
starts to decrease at this moment, after a steadily increase. This shows that this zone
in particular is more worn than the rest of the gear. On the other zones, only zone
13 (h) is showing signs of wear with 3 on 6 harmonics increasing. On the contrary,
no trend can be discerned on zones 3 (g) and 19 (i). The results from Kurtosis and
RMS, whether on the whole AM (b and c) or in the worn zone 10 (e and f) are also
inconclusive. The fact that the PP increases in a particular zone for several harmonics
is a strong validation of the PP increasing on the whole AM. It increases the likeli-
hood of an advanced wear on this zone, and can reinforce an automatic diagnostic.
When inspecting the state of the gear teeth in Fig. 7, it is confirmed that the zone 10
is the most worn. Indeed, large spalling marks on teeth 30, 31 and 46 can be seen,
the largest being on teeth 31. Teeth 30 and 31 belongs to zone 10. Smaller spalling
marks have been observed on 18 teeth, such as tooth 21.
Fig. 7 Teeth 21, 30, 31 and 46 from the gear of the GOTIX bench. Circled in red are spalling marks
6 Limitations of the Method
The strongest limitation of the presented method is probably the fact that this method
needs to have quite stable operating conditions, in particular speed and load, for each
measure. The shapes of modulation functions are strongly dependent on the excitation
of the system. This article shows that with constant operating conditions, such as on
an experimental bench like GOTIX, the AM and FM present a continuous evolution,
pertinent for the detection of defects. However, in the context of industrial exploita-
tion, it is rarely the case. A solution could be to monitor the state parameters (speed
and torque in particular) to trigger a measurement only when the system is operating
at a predefinite condition. It can work for systems remotely and continuously moni-
tored, such as wind turbines. When the measures are performed by technicians, they
can previously set these conditions before performing the measurements.
Another limitation is the necessity to have angularly synchronized measures,
starting all from the same gear angle. On the results presented here, the measures
have all been synchronized using a top-tour on the gear shaft. However, these are not
always present and can not be mounted easily in industrial context.
The last limitation is the changes in the system due to maintenance: a greasing
for example is susceptible to change the profile of the AM function brutally and
12 X. Laval et al.
needs to be taken into account in the context of automatic monitoring, by a feed of

information to the monitoring software
7 Conclusions and Perspectives
This paper presented a case study of a gearbox natural wearing and how to auto-
matically assess it by monitoring the AM and FM function evolution for several
harmonics. It has been shown that the wear is highlighted in the very continuous
evolution of the AM, more than the FM, and that a clear indicator of it is the use
of the peak-to-peak feature, both on the whole AM and on specific parts associated
to the wornest part of the gear. The confirmation that the wear is concentrated on
a specific part of the gear, as well as the fact that other sensors presents the same
characteristics, is used as a double check to ensure the fiability. The double advantage
of dividing the AM from several harmonics in several zones is both to provide more
features for an automatic classification, increasing the fiability of the alarm raised,
and to provide technicians a zone to check on the gear and decide if the gear should
be replaced.
The GOTIX experiments on the worn gear will resume after a reparation of the
driving system, and will go on until the gearbox breaks naturally. It will be interesting
then to check if the zones identified as worn now will be the one to break, and to
see the future evolution of the AM functions. Furthermore, the same study should
be conducted on other type of mechanical pieces, bearings or planetary gearbox for
example, as well as in industrial context, in order to confirm its interest.
References
1. Fanning P, Carden EP (2004) Vibration based condition monitoring: a review. SHM 3(4):355–
377
2. Márquez FPG et al (2012) Condition monitoring of wind turbines: techniques and methods.
Renew Energy 46:169–178. https://doi.org/10.1016/j.renene.2012.03.003
3. Wang WJ, McFadden PD (1995) Application of wavelets to gearbox vibration signals for fault
detection. JSV 192(5):927–939
4. Liang M et al (2012) Fault diagnosis for wind turbine planetary gearboxes via demodulation
analysis based on ensemble empirical mode decomposition and energy separation. Renew
Energy 47:112–126. https://doi.org/10.1016/j.renene.2012.04.019
5. Antoniadou I et al (2015) A time–frequency analysis approach for condition monitoring of a
wind turbine gearbox under varying load conditions. MSSP 64–65:188-216. http://dx.doi.org/
10.1016/j.ymssp.2015.03.003
6. McFadden PD, Smith JD (1985) A signal processing technique for detecting local defects in a
gear from the signal average of the vibration. In: Proceedings of the institution of mechanical
engineers, vol 199, no 43, pp 287–292. https://dx.doi.org/10.1243/PIME_PROC_1985_199_
125_02
7. Randall RB (1982) A new method of modelling gear faults. ASME, J Mech Des 104:259–267.
https://dx.doi.org/10.1115/1.3256334
8. Firla M et al (2016) Automatic characteristic frequency association and all-sidebands demod-

ulation for detection of a bearing fault of a wind turbine test rig. MSSP 80:335–348. https://
doi.org/10.1016/j.ymssp.2016.04.036
9. Martin N, Mailhes C (2018) Automatic data-driven spectral analysis based on a multi-estimator
approach. Signal Process 146:112–125. https://doi.org/10.1016/j.sigpro.2017.12.024
10. Gerber T, Martin N, Mailhes C (2015) Time-frequency tracking of spectral structures estimated
by a data-driven method. IEEE Trans Ind Electron Spec Session 52(10):6616–6626. https://dx.
doi.org/10.1109/TIE.2015.2458781
11. Laval X et al (2018) Vibration response demodulation, shock model and time tracking. In: CM
2018 and MFPT 2018, Nottingham, UK, 10–12 September 2018
12. Chaari F et al (2008) Effect of spalling or tooth breakage on gearmesh stiffness and dynamic
response of a one-stage spur gear transmission. Eur J Mech A/Solids 27:691–705. https://doi.
org/10.1016/j.euromechsol.2007.11.005
Multi-harmonic Demodulation
for Instantaneous Speed Estimation
Using Maximum Likelihood Harmonic
Weighting
C. Peeters, J. Antoni, and J. Helsen
Abstract This paper details a novel way to estimate the instantaneous angular speed
of a rotating shaft directly from its vibration signal measurement. Speed estimation
through single-harmonic phase demodulation has proven itself already multiple times
in the past to be a fairly reliable way to obtain accurate instantaneous angular speed
estimations. However, if the chosen harmonic experiences any disturbances such
as crossing orders or noise increases, the end result can be severely compromised.
The technique in this paper utilizes the phase of multiple harmonics in the signal
simultaneously to obtain an estimate of the instantaneous speed. The novelty lies
in the simultaneous phase demodulation formulation and in the way the phases of
the harmonics are combined using a maximum likelihood estimation approach. To
examine the robustness of this technique, this paper investigates the potential benefits
and hindrances of using a multi-harmonic demodulation approach as compared to a
single-harmonic approach. Simulation and experimental results show that a signifi-
cant increase in estimation accuracy can be gained by employing more information
contained within the phases of the harmonics in the vibration signal.
Keywords Vibrations · Instantaneous speed estimation · Tacholess · Harmonics ·

Phase demodulation
1 Introduction
Vibration analysis has become one of the main tools in condition monitoring of
rotating machines. A reoccurring problem in the analysis of rotating machinery vibra-
tions is the fluctuation of the rotation speed causing non-stationary behavior of the
vibration signal content. This issue has received much attention in recent years since
C. Peeters (B) · J. Helsen

Applied Mechanics, Vrije Universiteit Brussel, Pleinlaan 2, Brussels, Belgium
e-mail: cedric.peeters@vub.be
J. Antoni
Laboratoire Vibrations Acoustique, INSA-Lyon, Villeurbanne, France

https://doi.org/10.1007/978-981-15-9199-0_2
16 C. Peeters et al.
knowledge of the rotation speed is an integral part of almost every vibration-based

condition monitoring scheme. In general, the overall goal of condition monitoring is
to obtain an as accurate as possible understanding of the behavior of the machine and
its components. Therefore, rotation speed estimation has become one of the mainstays
in vibration analysis to assess machine health. Accurate knowledge of the rotation
speed allows to compensate for the speed variation. Typically, this compensation is
done by transforming the signal from the time to the angle domain.
While it is possible to install an encoder or tachometer to measure the IAS directly
on the machine, this imposes an additional cost for the machine operator since a
technician needs to dedicate time and resources to the instrumentation [1, 2]. Because
of this increased cost, there has been a significant amount of research done on the
estimation of the instantaneous angular speed from the vibration signal itself. A
recent review of several commonly used signal processing techniques to estimate the
speed from the vibration is provided in [3]. It turns out that the majority of techniques
for speed estimation from the vibration signal can be divided into two groups. One
group focuses on tracking harmonics in a time-frequency representation (TFR) of the
signal. Methodologies that fall into this group typically try to either generate a TFR
that is as accurate as possible, such as in [4–6], or they try to improve the tracking
itself in the TFR as much as possible, such as in [7–9]. The second group utilizes
phase demodulation of a harmonic. Typically, a single harmonic with a high signal-
to-noise ratio is chosen and then demodulated using its analytic signal representation;
examples are given in [10–13]. Methods that do not fall into these two groups employ
various different ways to estimate the speed and include techniques such as Kalman
filters, scale transforms, or wavelets.
This paper discusses the derivation of a multi-harmonic phase demodulation
methodology and can thus be categorized as a member of the second group of tech-
niques. The methodology section describes how the phase of multiple harmonics can
be combined and how the technique can be made more robust to low signal-to-noise
ratio harmonics. The latter improvement significantly increases the applicability of
this technique to industrial cases. The proposed methodology is illustrated on a
simulated signal and on an experimental case study using vibrations from a plane
jet engine gearbox. The obtained results of the multi-harmonic demodulation are
compared to the conventional single-harmonic demodulation method. This compar-
ison clearly indicates the significant gain in accuracy and robustness that can be
made by incorporating more harmonics and thus more signal content in the speed
estimation process.
2 Methodology
The idea of the method is based on extending the conventional single-harmonic phase
demodulation to multiple harmonic phase demodulation. Additionally, a maximum
likelihood estimation methodology is added to estimate weights for every harmonic.
Both steps are described in Sects. 2.1 and 2.2.
Multi-harmonic Demodulation for Instantaneous Speed Estimation … 17
2.1 Multi-harmonic Demodulation (MHD)
The vibration signal x(n) (of length N) is considered to be a sum of a set of K complex
harmonic signals x k (n) defined as a complex exponential:

K
x(n) = Ak (n)e j(αk θ(n)+φk ) (1)
k=1
with n being the sample number, k the harmonic number, K the total number of
harmonics, A the real amplitude, α the harmonic order, φ the constant phase, θ the
instantaneous angle of rotation.
For just one harmonic k this sum can be written as:
xk (n) = Ak (n)e j(αk θ(n)+φk ) + νk (n) (2)
In practice, this harmonic is filtered out using band-pass filtering. Calculating the
derivatives ẋk (n), gives:

ẋk (n) = Ȧk (n)e j(αk θ(n)+φk ) + Ak (n) e j(αk θ(n)+φk ) jαk θ̇ (n) (3)
The linear relationship between x k (n) and ẋk (n) can be written as follows:
ẋk (n) = βk (n)xk (n) (4)
where βk (n) is equal to:
Ȧk (n)
βk (n) = + jαk θ̇ (n) (5)
Ak (n)
Now, we introduce process noise νk (n) (e.g., due to imperfect sampling and
numerical differentiation):
ẋk (n) = βk (n)xk (n) + νk (n) (6)
The least-square error estimator of θ̇ (n) using all the filtered narrow-band
harmonic signals is given by:

K

θ̇ˆ (n) = arg min ẋk (n) − jαk θ̇ (n)xk (n) (7)
k
Calculating the derivative and setting it to zero gives after some algebra in the end
the following:
K
k=1 ẋk (n)xk∗ (n)
θ̇ˆ (n) = K (8)
k=1 αk |xk (n)|2
where {.} denotes the imaginary part.
2.2 Maximum Likelihood Estimation of Harmonic Weighting
Equation 8 gives a uniform weighting to every harmonic. In an attempt to make

the weights more dependent on the signal-to-noise ratio, a maximum likelihood
estimation is derived.
Three assumptions are made:
1. The process noise is a univariate
complex
Gaussian distribution with zero bias
and unknown variance ∼ N 0, σk2 (n)
2. The statistics of the process noise are independent of time and the number of the
harmonic k
3. The process noises of the K different harmonics are assumed to be uncorrelated.
The second assumption is wrong since νk (n) does depend on time and the
harmonic number, but it allows us to use a simple Gaussian distribution. By incor-
porating the complex Gaussian probability density function of the process noise, we
arrive at the likelihood function:
K −|ẋk (n)−βk (n)xk (n)|2

e
k=1 σk2 (n)

P x1 , x2 , . . . , xk |θ̇ˆ (n) = K = L θ̇ (n) (9)
πKN k σk (n)
2
To find θ̇ˆ that maximizes Eq. 11, we minimize the negative log-likelihood:

δ − ln L θ̇ˆ (n)
=0 (10)
δ θ̇ˆ
Solving the derivative for θ̇ˆ gives, after some algebra:

K ẋk (n)xk∗ (n)αk

θ̇ˆ (n)ML =
k=1 σk2 (n)
K (11)
|xk (n)|2 αk2
k=1 σk2 (n)
The optimal weights for each harmonic are thus given by:
αk
wk (n) = (12)
σk (n)
2
The noise variance is estimated using again a maximum likelihood estimation and
it is assumed to be constant over a short time interval [n − I, n + I ]. The PDF of the
noise variance is given by:

n+I 2
1 −|ẋk (i)−βσk2(i)xk (i)|
P xk |σk2 = e k (13)
i=n−I
π σk2
Deriving the maximum log-likelihood gives:
1
n+I

ẋk (i) − jαk θ̇ (i)xk (i)2

σk (n)ML ≈ (14)
2I + 1 i=n−I
2.3 Practical Considerations
The main benefit of the formulation in Eq. 11 is that the denominator has a small
probability of getting close to zero since it uses the amplitudes of multiple harmonics.
In practice, the complex harmonic signals xk (n) are obtained using complex demodu-
lation with a roughly estimated speed. The rough speed can be determined in various
ways that need little input (e.g., through maximum tracking in the STFT). For the
differentiation involved, a filter can be chosen that does not amplify high-frequency
noise. Differentiating filters typically increase high-frequency noise which can be
observed by looking at their transfer function. Four examples of transfer function
amplitude responses are shown in Fig. 1. This paper employs a first-order difference
scheme since it has the lowest computation time and the produced results did not
Fig. 1 Four examples of the

magnitude responses of
various differentiation filters
vary significantly with the differentiation filter type. The MHD method is run iter-
atively by estimating in an alternating fashion first θ̇ˆ (n)ML and then σk (n)ML until

convergence or the maximum number of iterations has been reached.
3 Simulation Results
In order to highlight the influence of the maximum likelihood weighting, a simple

example is chosen where a signal is simulated that consists of the following:
• 5 sinusoidal harmonics with a base rotation speed of 31 Hz, but the harmonic
numbers are chosen to be odd. The 5th harmonic however is chosen to be asyn-
chronous with the speed. This means the speed synchronous signal content only
consists of the 1st, 3rd, 7th, and 9th harmonic.
• 1 lightly damped resonance in the middle of the 7th harmonic, which is detrimental
for the phase demodulation of that harmonic.
• Time-varying and unitary amplitudes for all 5 harmonics (i.e., mono-component
sinusoidal modulation).
• Additive white Gaussian noise at an SNR of −5 dB.
The spectrogram of the signal is shown in Fig. 2. The red rectangles indicate the
lightly damped mode coinciding with the 7th harmonic and the asynchronous 5th
harmonic. The simulated signal has a sample rate of 700 Hz and a duration of 20 s.
In total, the MHD is run for 20 iterations.
Fig. 2 Spectrogram of the simulated signal

Fig. 3 Left: Speed estimation results comparison for MHD with and without MLE weighting.
Right: Normalized average weights per harmonic and per iteration
Figure 3 shows the speed estimation results with and without maximum likelihood
estimation of the weights. It is clear that without proper weighting, the estimation
fails. The figure also displays the evolution of the weights, which clearly shows the
proper selection of the optimal harmonics for the speed estimation.
4 Experimental Results
The experimental data set comes from the Safran contest at the Surveillance 8 confer-
ence, held in Roanne, France [14]. A diagram of the aircraft engine with the faulty
bearings and the locations of the accelerometers is shown in Fig. 4. A spectrogram
of the processed data of sensor 2 is displayed in Fig. 5. It is obtained using a Hanning
window with a length of 211 samples with an overlap of 95%.
For comparison, the single-harmonic phase demodulation method is used on the
38th harmonic of the high-pressure shaft. The multi-harmonic demodulation method
uses the first 60 harmonics as input. The estimated speed profiles for both methods
are shown in Fig. 6 on the left. Visually, it is difficult to tell which method performs
best. Therefore, the estimated speeds are compared to the measured encoder speed.
Fig. 4 Left: Engine and accessory gearbox. Shafts are identified by labels in amber color. Right:
Diagram of the kinematics of the gearbox
Fig. 5 Spectrogram of the aircraft engine data
Fig. 6 Left: Estimated speed profiles. Right: Mean and median absolute errors
Figure 6 shows on the right the mean and median absolute deviations to the encoder
speed for both methods. The multi-harmonic method clearly outperforms the single-
harmonic demodulation method.
5 Conclusions
This paper discusses the potential gain in accuracy and robustness when using
multiple harmonics for phase demodulation as compared to just one single harmonic.
The proposed methodology presents a framework for incorporating multiple demod-
ulated phase signals into a singular demodulation step. It also provides a means to
optimize the available set of harmonics such that the actual speed estimation is more
accurate and less influenced by asynchronous signal components or poorly excited
harmonics.
Acknowledgements The authors of VUB would like to thank the agency for Innovation by Science
and Technology in Belgium for supporting the SBO HYMOP project and also SIM for the support
with the MASIWEC project. They would also like to thank FWO for providing funding for a long
stay abroad of Cédric Peeters at the University of INSA Lyon.
References
1. Potter R (1990) A new order tracking method for rotating machinery. Sound Vibr 24(9):30–34
2. Fyfe K, Munck E (1997) Analysis of computed order tracking. Mech Syst Signal Process
11(2):187–205
3. Peeters C, Leclère Q, Antoni J, Lindahl P, Donnal J, Leeb S, Helsen J (2019) Review and
comparison of tacholess instantaneous speed estimation methods on experimental vibration
data. Mech Syst Signal Process 129:407–436
4. Peng Z, Meng G, Chu F, Lang Z, Zhang W, Yang Y (2011) Polynomial chirplet transform with
application to instantaneous frequency estimation. IEEE Trans Instrum Meas 60(9):3222–3229
5. Kwok HK, Jones DL (2000) Improved instantaneous frequency estimation using an adaptive
short-time fourier transform. IEEE Trans Signal Process 48(10):2964–2972
6. Cheung S, Lim JS (1992) Combined multiresolution (wide-band/narrow-band) spectrogram.
IEEE Trans Signal Process 40(4):975–977
7. Barrett RF, Holdsworth DA (1993) Frequency tracking using hidden markov models with
amplitude and phase information. IEEE Trans Signal Process 41(10):2965–2976
8. Wang S, Chen X, Wang Y, Cai G, Ding B, Zhang X (2015) Nonlinear squeezing time-frequency
transform for weak signal detection. Sig Process 113(195):210
9. Schmidt S, Heyns PS, De Villiers JP (2018) A tacholess order tracking methodology based
on a probabilistic approach to incorporate angular acceleration information into the maxima
tracking process. Mech Syst Signal Process 100:630–646
10. Urbanek J, Barszcz T, Sawalhi N, Randall R (2011) Comparison of amplitude-based and
phase-based methods for speed tracking in application to wind turbines. Metrol Meas Syst
18(2):295–304
11. Urbanek J, Barszcz T, Antoni J (2013) A two-step procedure for estimation of instantaneous
rotational speed with large fluctuations. Mech Syst Signal Process 38(1):96–102
12. Combet F, Gelman L (2007) An automated methodology for performing time synchronous
averaging of a gearbox signal without speed sensor. Mech Syst Signal Process 21(6):2590–2606
13. Bonnardot F, El Badaoui M, Randall R, Daniere J, Guillet F (2005) Use of the acceleration
signal of a gearbox in order to perform angular resampling (with limited speed fluctuation).
Mech Syst Signal Process 19(4):766–785
14. Antoni J, Griaton J, Andre H, Avendano-Valencia LD, Bonnardot F, Cardona-Morales O,
Castellanos-Dominguez G, Daga AP, Leclere Q, Vicuna CM et al (2017) Feedback on the
surveillance 8 challenge: vibration-based diagnosis of a Safran aircraft engine. Mech Syst
Signal Process 97:112–144
A Variational Bayesian Approach
for Order Tracking in Rotating Machines
Y. Marnissi, D. Abboud, and M. Elbadaoui
Abstract Order tracking is important in rotating machine signal processing. It

consists in tracking the complex envelope of phase varying sinusoids associated
with distinct rotating mechanical elements. The resulting magnitudes and phases
are usually used for vibration level control and system resonance identification. A
common approach to tackle this issue is the Vold-Kalman filter. However, the perfor-
mance of this filter is highly conditioned by the choice of some tuning parameters
that control its selectivity. In this paper, a new order tracking method is proposed to
jointly estimate the targeted envelopes and the optimal values of these parameters in
an automatic manner. To do so, the order tracking problem is formulated within the
Bayesian framework. In details, prior distributions are affected to the target variables
(including the tuning parameters and the noise variance); posterior distributions are
then derived and estimated using a bariational Bayesian approximation algorithm.
Results on simulated and real helicopter vibration data demonstrate the effectiveness
of the proposed approach.
Keywords Order tracking · Vold-Kalman filter · Bayesian framework ·

Vibrations · Nonstationary · Rotating machines
1 Introduction
In rotating machines, each mechanical source generates a unique pattern as the

machine operates. For many components, the vibratory response can be modelled
accurately through a sum of phase and amplitude modulated sinusoids whose instan-
taneous frequencies are linearly related to the rotating frequency of the rotating
component. The estimation of the amplitude and the phase modulations, also known
Y. Marnissi · D. Abboud (B) · M. Elbadaoui

Information & Signal Technology center, SafranTech, Rue des Jeunes Bois, Châteaufort, 78772
Magny-les-Hameaux, France
e-mail: dany.abboud@safrangroup.com
M. Elbadaoui
LASPI, EA3059, Université de Lyon, UJM-St-Etienne, 42023 Saint-Etienne, France

https://doi.org/10.1007/978-981-15-9199-0_3
26 Y. Marnissi et al.
as order tracking, is of high importance in rotating machine signal processing. On

the one hand, it provides the instantaneous vibratory level through the magnitude as
well as potential “phase jumps” which are associated with resonance modes. On the
other hand, it can be used to estimate the deterministic components and separate them
from other random interfering contributions (e.g. environmental noise, electromag-
netic interference, measurement errors, etc.). When the instantaneous frequencies of
the targeted components are known, a common approach to tackle this issue is the
Vold-Kalman filter [1–5]. Several variants of this algorithm have been proposed in
Refs. [6–8]. This filter models the order tracking problem through two equations.
The first one is the data equation which establishes the link between the observation
(raw signal) and the component(s) of interest. The second is the structural equation
that adds a constraint on the smoothness of the estimated envelopes. This constraint
is dictated by some tuning parameters which control the degree of smoothness or,
equivalently, the selectivity of the filtering [9]. Nowadays, these parameters are very
often empirically set by the operator, though their values can significantly affect
the estimation. This paper addresses the shortcomings of the Vold-Kalman filter
by adopting a new approach based on the Bayesian framework. One of the main
advantages of the proposed formulation is that it allows joint estimation of the target
envelopes and the tuning parameters from the measured signal, while providing a
good approximation of the minimum mean square error estimators.
This paper is organized as follows: Sect. 2 opens with a brief review on the
original Vold-Kalman filter as well as a discussion about its shortcomings. In Sect. 3,
the theoretical fundamentals of the newly proposed Bayesian method are presented
together with an efficient iterative algorithm. In Sect. 4, the efficiency of the new
method is demonstrated and compared with the Vold-Kalman method on simulations
and helicopter vibrations data. Finally, some conclusions are drawn in Sect. 5.
2 A Brief Review of the Vold-Kalman Filter
2.1 Method Description
Let K > 0 be the number of target components (i.e. sinusoids) to be extracted

from a measured signal of length N > 0. The data equation states that each target
component k ∈ {1, . . . , K } is equal to the observed signal up to an additive error
δk (n). This can be formulated as follows:
∀k ∈ {1, . . . , K } ∀n ∈ {1, . . . , N } y(n) − xk (n) exp( jφk (n)) = δk (n) (1)
The latter formulation is called a mono-component estimation problem since each

of the K components is estimated separately and independently from the others.
Another approach to tackle this issue is to jointly consider the K components. In this
case, Eq. (1) turns to a multi-component estimation problem with an additive error
A Variational Bayesian Approach for Order Tracking in Rotating … 27
δ(n) as follows:

K
∀n ∈ {1, . . . , N } y(n) − xk (n) exp( jφk (n)) = δ(n) (2)
k=1
The Vold-Kalman filter also requires prior information on the target components
that will be integrated into the structural equation. More specifically, for every k ∈
{1, . . . , K }, the complex envelope ak (n) is assumed to have a smooth and regular
variation. This is ensured using the slow-varying discrete difference of the complex
envelope to act as a low pass filter:
∀k ∈ {1, . . . , K } ∀n ∈ {1, . . . , N } ∇ p+1 xk (n) = k (n) (3)
where ∇ p+1 is the difference operator of order p and k (n) is an additive error term.
Given the data and the structural Eqs. (1) and (3), respectively, the least square esti-
mator of each complex envelope xk = [xk (1), . . . , xk (N )]T in the mono-component
estimation problem reduces to the minimizer of the 22 norm of the residual errors
δk (n) and k (n). The solution reads
−1
∀k ∈ {1, . . . , K } x̂k = I + rk A Tp A p CkH y (4)
where rk is a positive weighting factor, A p is the matrix of discrete difference of order

p, Ck is the diagonal matrix formed by the terms s exp( jφk (1)), . . . , exp( jφk (N )) and
(.) H denotes the conjugate transpose. Similarly, in the case of the multi-component
estimation problem, the least square estimator of the vector of complex envelopes
T
x = x1T , . . . , x KT can be expressed in matrix form:
−1
x̂ = C H C + T R C H y (5)
where C = [C1 , . . . , C K ], R is the diagonal matrix formed with r1 , . . . , r K and is

the block diagonal matrix formed by K matrices A p .
2.2 Shortcomings of the Method
The quality of the estimator is very sensitive to the choice of weighting factor r k . On
the one hand, the larger r k is, the more the solution is influenced by the structural
equation resulting into a very smooth envelope. This also amounts to say that the
larger r k is, the narrower the bandwidth of the filter is and the more important the
selectivity is. However, using large values of r k may cause numerical issues due
the bad conditioning of the matrix C H C + T R that increases linearly with r k
resulting to in nonsensical results. On the other hand, decreasing the value of r k may
lead to an amplification of the noise level in the estimated amplitude. The setting of
the weighting factor turns to be more complicated in the case of a multi-component
estimation problem. In fact, considering the same value for all r k is not generally a
reliable solution since the different components may have different energy ranges.
3 Proposed Method
3.1 Bayesian Formulation
The considered observation model is similar to that of (2). The coefficients of the
noise δ(n) are assumed to be i.i.d according to a zero mean complex Gaussian
noise of unknown variance 2τ −1 which means that their real and imaginary parts
are real Gaussian random variables of variance τ −1 (i.e. τ denotes the precision of
the real and imaginary variables). Given the Gaussian assumption, the observation
distribution given x and τ reduces to a complex Gaussian distribution with mean C x
and variance 2τ −1 . Just like the Vold-Kalman method, a constraint on the smoothness
of the envelopes is needed. Equation (3) can be relaxed to the probabilistic framework
by considering instead the probability density function of x given the precision τ and
the regularization parameters r1 , . . . , rk :

K
p(x|r1 , . . . , rk , τ ) = p(xk |rk , τ ) (6)
k=1
with

1 2
∀k ∈ {1, . . . , K } p(xk |rk , τ ) = Cte.rkN τ N
exp − τ.rk A p xk (7)
2
where . denotes the 2 -norm and C t e is a constant independent of r k , τ and

x k . The noise precision τ and the regularization parameters r 1 , . . . , r k will be
estimated jointly with the complex envelopes. For simplification reasons, conju-
gate priors are adopted, namely for each r k , a Gamma prior with real positive
parameters a k > 0, bk > 0, i.e. p(r k |a k , bk ) ∝ r akk −1 exp(−bk .r k ) and for
the noise precision parameter a gamma prior with positive real hyperparameters
a > 0, b > 0, i.e. p(τ|a, b) ∝ τa−1 exp(−b.τ). Note that the hyperparameters
a k , bk , a, b are fixed according to our prior knowledge about the parameters r k and
τ . In the case where a prior information is unavailable, one can choose small values
for these parameters to obtain diffuse prior distributions. The Bayes rule allows
to derive the posterior distribution from the observation and the considered prior
distribution of the set of unknown parameters = {x, r 1 , . . . , r k , τ } as follows:
K
p(y|x, τ ) p(x|r 1 , . . . , r k , τ) p(τ|a, b) p(r k |a k , bk )
p(|y) = k=1
(8)
p( y)
The negative logarithm of this conditional distribution reads up to an additive

constant independent of :
τ
K
τ
f (, y) = C x − y2 + r k A p x 2k + bτ − (a + N(K + 1))log(τ )
2 2
k=1

K
K
+ bk r k − (N + a k )log(r k ) (9)
k=1 k=1
The aim now is to derive estimations of the variables in from their joint poste-
rior distribution (8). As the posterior distribution has an intricate form, a varia-
tional Bayesian approximation (VBA) approach is adopted to approximate it. This
is addressed in the next subsection.
3.2 Algorithmic Solution
3.2.1 Variational Bayes Approximation (VBA) Principle
The VBA approach consists in approximating p(|y) with a distribution q() which
is as close as possible to p(|y) in the sense of the Kullback–Leibler divergence
criterion [10]. In order to have a tractable solution, the approximate probability
density is assumed to have the following factorized form:

K
K
q() = qxk (xk ) qrk (rk )qτ (τ ). (10)
k=1 k=1
The advantage of this factorization form is to enable a disjoint estimation of the

different components while ensuring a multi-component modelling of the problem.
In this case, an expression of the optimal density approximation for each variable
is expressed as the exponential of the expectation of f (, y) with respect to the
densities of the remaining variables, i.e.

∀r ∈ qr (r ) ∝ exp E qi (i )

{ f (, y)} (11)
i =r
3.2.2 Densities Update
Using the expression (11), one can easily show that the optimal density of the envelope
xk reduces to a circularly symmetric complex normal distribution, that is qxk (xk ) ≡
CCN(m k , 2. k ) where
⎛ ⎞

−1
β αk T α H⎝
= Id + A Ap and m k = C y− Ck m k ⎠ (12)
k
α βk p β k k
k =k
which means that the real and imaginary parts of xk follow Gaussian distributions
with means equal to the real and the imaginary parts of m k respectively, and with
the same real covariance matrix k . It can be noted that the computation of m k
requires inverting a matrix that has the same structure as the one that intervenes
in the mono-component Vold-Kalman estimation. However, contrary to the mono-
component estimation, an additional term is added to the approximate posterior mean
m k namely k =k Ck m k to account for the bias introduced by the other components.
Similarly, the optimal densities reduce to Gamma distributions for rk and τ , i.e.
qrk (rk ) ≡ G(αk , βk ) and qτ (τ ) ≡ G(α, β) where

αk = a k + N
α (13)
βk = bk + 2β A p m 2k + 2.trace ATp A p k
⎧
⎨ α = a + N (K + 1)
⎪
K 2 2
1 K
αk + 2.trace AT A p k
⎪
⎩β = 2 Ck m k − y + trace[ k ] + A p m k
k=1 k=1
2βk p
(14)
3.2.3 Iterative Algorithm
Due to the implicit relations existing between the variables, these approximate densi-
ties are computed iteratively by updating the separable density of one variable while
fixing the others which results in the following algorithm
1. Set initial values: αk(0) ,βk(0) , m (0) (0) (0)
k , k , k ∈ {1, . . . , K }, α , β
(0)
2. For t = 0, 1, . . .
a. For k = 1, . . . , K
i. Update k(t+1) and m (t+1)
k using (12)
(t+1) (t+1)
ii. Update αk , βk using (13)
b. Update α (t+1) , β (t+1) using (14).
The approximate posterior means at convergence are then used as estimators of

opt
α opt

αk
opt
rk , τ, xk that is τ̂ = β opt
, rk = opt
βk
and xk = m k .
3.2.4 Implementation Issues
Since the main objective is the computation of the posterior mean, the calculation and
the storage of k is not required. However, the update of the approximate densities of
the other parameters requires the knowledge of second-order statistics of x k , being
dependent on k . Several solutions can be adopted, for instance, using circulant
border conditions for the discrete difference matrix A p . This results in a circulant
matrix k that is diagonalizable in the Fourier domain so that all second-order
statistics can be computed easily in the Fourier domain. Additional solutions using
diagonal approximation or Monte Carlo approach can be also found in [10].
4.1 Numerical Evaluation
The synthetic signal is a three-second length signal with a sampling frequency F s =

5000 Hz, comprising two sinusoidal components whose instantaneous frequencies
are given in Fig. 1a. The instantaneous frequency of the first component, plotted with
a black continuous line, is a hanning window over the signal length; while that of the
second component, plotted with a red continuous line, consists of a monotonically
increasing function over 1 s, cascaded with a constant frequency (220 Hz) over the
remaining two seconds. In order to evaluate the methods in a challenging setting, the
second profile (red line) is purposely chosen close in the increasing part and intersects
in the remaining part (precisely at 2 s) with the other profile. A complex Gaussian
noise with a precision equal to 21.38 × 10−4 is also added which corresponds to a
signal-to-noise ratio equal to 1; this gives K = 2 and N = 15, 000. Amplitudes of
the estimated envelopes using the proposed approach are given in Fig. 1b with dotted
Fig. 1 Results on simulated data. a Speed profile. The actual magnitude (dotted lines) and the
estimated magnitudes using the b proposed approach, c mono-component Vold-Kalman and d multi-
component Vold-Kalman
Fig. 2 Results on helicopter vibration signal. a Initial spectrogram. b Spectrogram of the estimated
component. c Estimated envelope
lines, together with the corresponding actual magnitude (dotted lines). The estimated

values for the other parameters are respectively r 1 = 1.58 × 108 , r 2 = 1.09 × 1010
and τ̂ = 21.02 × 10−4 . r 1 and r 2 are then used as inputs in the standard Vold-

Kalman filter in order to provide estimators using the mono-component and multi-
component models. The obtained estimators are exposed in Fig. 1c, d. One can
see that the multi-component estimation removes the artifacts of close and crossing
frequencies. Despite the factorization form (10) assuming that the densities of the
different components are separable, the Bayesian approach gives comparable results
as for the multi-component Vold-Kalman filter, yet not requiring a manual setting of
the filter bandwidths.
4.2 Application to Helicopter Vibration Data
The proposed method is now applied on a helicopter vibration signal acquired with
a constant sampling frequency in a run up to automatically extract the high pressure
shaft component. The spectrograms of the initial signal and the extracted component
as well as the amplitude of the estimated envelope are displayed in Fig. 2. Note that,
for confidentiality issues, the sampling frequency is set to one and the acceleration
signal is normalized by its standard deviation. The estimated envelope shows an
obvious and smooth increase of the amplitude caused by the passage through a
structural resonance.
5 Conclusion
This paper proposed a new approach for vibro-acoustic component extraction for
rotating machines using a Bayesian framework. The data and the structural equation
proposed by the standard Vold-Kalman filter have been expressed in probabilities
forms. The regularization parameters controlling the selectivity of the filter are also
modelled as random variables with Gamma distributions. Estimators for the enve-
lope, the noise variance and the regularization parameters are then derived from
their resulting posterior distribution using a Variational Bayesian approach. Exper-
iments performed on simulated data and a real case of a helicopter vibration signal
demonstrate the good performance of the method.
References
1. Vold H, Mains M, Blough J (1997) Theoretical foundations for high performance order tracking
with the Vold-Kalman tracking filter. In: SAE transactions, pp 3046–3050
2. Vold H, Herlufsen H, Mains M, Corwin-Renner D (1997) Multi axle order tracking with the
Vold-Kalman tracking filter. Sound Vibr 5:30–35
3. Herlufsen H, Gade S, Konstantin-Hansen H (1997) Characteristics of the Vold-Kalman order
tracking filter. Sound Vibr 33:34–44
4. Pan MC, Lin YF (2006) Further exploration of Vold–Kalman-filtering order tracking with
shaft-speed information—I: Theoretical part, numerical implementation and parameter inves-
tigations. MSSP 5:1134–1154
5. Pan MC, Lin YF (2006) Further exploration of Vold–Kalman-filtering order tracking with
shaft-speed information—II: Engineering applications. MSSP 6:1410–1428
6. Feldbauer C, Holdrich R (2000) Realization of a Vold-Kalman tracking filter—a least squares
problem. In: COST G-6, Limerick, pp 1–4
7. Pan MC, Lin YF (2010) Extended angular-velocity Vold–Kalman order tracking. JDSMC
3:031001
8. Pan MC, Lin YF (2010) High efficient crossing-order decoupling in Vold-Kalman filtering
order tracking based on independent component analysis. MSSP 6:1756–1766
9. Tuma J (2005) Setting the passband width in the Vold-Kalman order tracking filter. In: ICSV,
Lisbon, Portugal, pp 11–14
10. Marnissi Y, Zheng Y, Chouzenoux E, Pesquet JC (2017) A variational Bayesian approach for
image restoration—application to image deblurring with Poisson-Gaussian noise. IEEE Trans
Comput Imaging 4:722–737
Reconstructing Speed from Vibration
Response to Detect Wind Turbine Faults
K. Marhadi and D. Saputra
Abstract Speed information is often necessary to detect certain failure modes in

wind turbines, such as unbalance, looseness, misalignment, and gear related prob-
lems. This is normally obtained from tacho sensors targeting certain shafts, e.g.
generator shaft or main shaft. Tacho pulses can be used to track rotational speed
of certain components as well as to perform angular resampling. Most importantly,
knowing machine speed is necessary for alarming speed dependent failure modes.
Correct speed information may be unavailable due to broken tacho sensor and incor-
rect setting of the sensor. Consequently, many failure modes cannot be detected
when the information is not available. Unable to detect these failure modes early
may cause unexpected secondary damages that leads to unexpected downtime of
the wind turbines and loss of production. Maintenance work on correcting speed
sensor is costly, especially for offshore wind turbines, and may result in downtime.
Furthermore, condition monitoring system (CMS) of wind turbines is significantly
cheaper without speed sensors; considering a wind farm normally consists of many
wind turbines. Therefore, it is important to acquire speed information without the
need for having a tacho sensor. This paper presents an industrial implementation
of speed reconstruction based on vibration response to detect speed-related faults,
such as unbalance and gear faults. Speed is determined by frequency demodula-
tion of response vibration signals. Hilbert transform is used to extract the instanta-
neous frequency, where the derivative of the analytic signal is done in the frequency
domain. Using data from wind turbines monitored by Brüel & Kjær Vibro, we show
this method can detect speed-related faults as early as when the speed informa-
tion is available through real speed sensor. The method was able to detect unbalance,
misalignment and looseness on wind turbine generators, as well as gear-related prob-
lems on high speed stage, intermediate stage, and planetary stage gearboxes. This
offers the possibility of reducing the cost of wind turbines CMS significantly.
K. Marhadi (B) · D. Saputra

Bruel & Kjær Vibro, Skodsborgvej 307B, 2850 Nærum, Denmark
e-mail: kun.marhadi@bkvibro.com
D. Saputra
e-mail: dhany.saputra@bkvibro.com

https://doi.org/10.1007/978-981-15-9199-0_4
36 K. Marhadi and D. Saputra
Keywords Wind turbine CMS · Tacho pulses · Hilbert transform · Machine

speed · Order cepstrum
1 Introduction
Condition monitoring of wind turbine requires monitoring magnitudes of certain

frequencies, also known as descriptors. These descriptors describe the state of a
component and are trended for alarming purpose. Some of them require speed infor-
mation of certain shafts to monitor, for example, unbalance in a generator and gear-
related faults in a gearbox [1] and [2]. This information is normally obtained from a
speed sensor.
In a wind turbine, the speed sensor is normally attached to the generator shaft.
Shaft speeds and tooth mesh frequencies at various stages can be determined from the
generator shaft speed using ratios of various shafts in the gearbox. A non-contacting
eddy current displacement sensor is normally used for speed reference sensor. This
type of sensor needs to be adjusted correctly to obtain correct speed measurement.
The generator shaft rotation often drifts over time, which could cause the initial speed
sensor setting to be invalid. Incorrect speed sensor setting could cause speed reference
measurement to be invalid, which results in unreliable speed-dependent descriptors.
As a result, failure modes that require speed information for their detection may be
missed.
Several works to mitigate the problem of invalid speed sensor reading have been
presented [3–8]. They mostly involve detection of speed information from vibra-
tion; thus speed sensor is not required. In [8], speed information is estimated for
wind turbine applications in case speed sensor becomes unavailable. In that paper,
prior information from the sensor is required before it becomes unavailable for the
algorithm to work.
Those works mainly focused on estimating the speed information itself rather than
using it for faults detection. This paper focuses on using estimated speed to detect
faults in wind turbines. The speed is estimated using frequency demodulation of
response vibration signals using Hilbert transformation as presented in [9] and [10].
The work in [9] shows that it is the most accurate method to determine instantaneous
speed of a shaft by frequency demodulation.
2 Hilbert Transform to Estimate Speed
As described in [9] and [10], instantaneous phase and amplitude of a vibration signal
can be determined by Hilbert transform techniques, which result in an analytic signal
that corresponds to the real signal. This signal can be represented as:
xa (t) = A(t)exp( jφ(t)) (1)
where φ̇(t) = ω(t) = 2π f (t).

Reconstructing Speed from Vibration Response … 37
The derivative of the instantaneous phase of an analytic signal can be represented

as:

ẋa (t)
ω(t) = φ̇(t) = I m (2)
xa (t)
where ẋa (t) is the derivative of xa (t). As explained in [9], the exact derivative can
be obtained by multiplying the spectral values by jω over the frequency ranges to
be demodulated before performing inverse transform of the spectrum.
With Eq. 2, the instantaneous speed of a shaft can be determined from vibration
signal at all times. However, the initial estimated speed needs to be smoothed out
because the differentiation introduces noise in real vibration signal as described in
[10]. In the current paper, initial estimated speed is smoothed out using a FIR filter
in the time domain as presented in [10].
3 Fault Detection using Speed Estimation
In Bruel and Kjær Vibro wind turbine condition monitoring applications, time wave-
forms data are obtained approximately every two days. The time waveforms contain
vibration signals from all components monitored. The length of each time waveform
is 10.24 s with sampling rate of 25,600 Hz. The type of wind turbine analyzed in this
work has a gearbox configuration of one planetary and two helical stages. For brevity,
this paper only presents detection of generator unbalance, gear faults at intermediate
or 2nd stage and planetary or 1st stage based on estimated speed information.
A critical part in estimating speed without a speed sensor is to select a frequency
band to be demodulated; from which the speed is to be determined. The carrier
frequency to be demodulated in this band must be as clean as possible from contam-
inations from other components/frequencies. Furthermore, it must be able to take
into account variation of the speed to be estimated, e.g. not more than 15% speed
variations. Based on work in [8], a band is selected to demodule the first tooth mesh
frequency of the high speed or 3rd stage of the wind turbine gearbox.
Speed is estimated from high speed stage vibration signal for the duration of time
waveform obtained, i.e. 10.24 s. The estimated speed is then converted in terms of
rotation angle to perform order analysis, i.e. angular resampling vibration signals, and
to track the orders of interest. This is done by integrating the estimated instantaneous
speed over time to get the cumulative rotations as a function of time. All failure
modes detected here are based on tracking orders that indicate faults.
To detect generator unbalance, the first order magnitude (1X) of the generator
shaft is tracked over a period of four months as shown in Fig. 1 for a 2 MW turbine.
An increasing trend of 1X can be seen clearly starting mid April, indicating devel-
opment of unbalance. The tacho information from the generator shaft, which is also
contained in the 10.24 s time waveform data, can be used to verify the estimated
Fig. 1 Trend of first order magnitude (1X) of generator shaft
speed. Throughout the paper, it is the generator shaft speed that is estimated as refer-
ence. For comparison, the first order magnitude based on real tacho is also presented.
The maximum difference of 1X based on estimated and real tacho is 9.75%, and the
mean difference is 4.56%.
Figure 2 shows comparison of the actual averaged speed and the average estimated
speed in 10.24 s over the period of four months. The magnitude of 1X is proportional
to rotational speed. As Fig. 2 shows, there are small differences between actual
and estimated speeds. The maximum difference between average actual and average
estimated speed is 0.34%. Despite the differences, unbalance can still be detected,
and its development is tracked.
Detecting gear related problems involves identification of sidebands around gear
tooth mesh frequencies. Spacing of the sidebands will indicate which gear has
Fig. 2 Trend of average actual and estimated speeds of turbine with unbalance generator
the problem. Spacing of sidebands indicates modulation of shaft where the gear
is attached. As described in [11], it is very convenient to identify family of sidebands
using cepstrum. Thus, order cepstrum, which is defined as the inverse Fourier trans-
form of the log-order spectrum, is used in current work to identify sidebands related
to gear problems.
In detecting gear problem at the second stage (intermediate stage), vibration signal
from intermediate stage sensor is order-tracked according to the intermediate stage
shaft rotations. Thus, gear related problems at this shaft manifest as sidebands with
spacing of one order of the intermediate shaft in order spectrum. It is convenient
to track one cycle in the order cepstrum to indicate development of gear related
problem at the intermediate shaft. Figure 3 shows the development of gear problem
at the intermediate shaft from a 2 MW turbine over a period of three months. As the
figure shows, gear problem occurred after the 21st of January. The average difference
between one cycle tracked based on real and one cycle tracked based on estimated
speed is 48.07%, and the maximum difference is 286%. Figure 4 shows comparison
between the actual and estimated generator speeds. Maximum difference between
actual and estimated generator speed is 16.86%. Despite a maximum speed difference
of approximately 17%, gear problem can be detected.
Gear problem at the planetary stage, which is a very difficult failure mode to be
detected in wind turbines, can also be detected by estimating speed. A wind turbine
had a planetary ring defect. The ring is stationary and the input from main rotor of
the turbine drives the planetary stage carrier. The ring problem resulted in strong
modulation of three times carrier speed, known as ring defect frequency, which is
tracked in order cepstrum. Figure 5 shows the trend of ring defect of the planet carrier
from order cepstrum over a period of six months. Significant development of the fault
can be seen in the end of March as magnitude of the ring defect increases. The average
difference between ring defect based on estimated speed and the one based on real
Fig. 3 Trend of first cycle of intermediate shaft from an order cepstrum

Fig. 4 Trend of average actual and estimated generator shaft speeds of turbine with gear problem
at the intermediate shaft
Fig. 5 Trend of ring defect from an order cepstrum
tacho is 8.29%, and the maximum difference is 60.76%. Figure 6 shows comparison
between the actual and estimated generator speeds of the turbine. The maximum
difference between actual and estimated generator speed is 0.49%. Although there
are few clear differences at certain points in time, problem at the planetary stage can
be detected.
The descriptors calculated based on estimated speed in the present work can
have big differences with the ones calculated based on real speed at certain times.
However, the failure modes can still be detected over a long period. In fact, faults
were detected in approximately the same time as the ones trended using descriptors
calculated based on real speed. Descriptors based on estimated and real speeds follow
the same trend over time. Table 1 presents summary of the result.
Fig. 6 Trend of average actual and estimated generator shaft speeds of turbine with planetary stage
problem
Table 1 Summary of differences between estimated and real speeds

Turbines Average absolute Maximum average Average absolute Maximum
generator speed generator speed descriptor absolute
difference (%) difference (%) difference (%) descriptor
difference (%)
Turbine with 0.04 0.34 4.56 9.75
unbalance
Turbine with 2nd 0.48 16.86 48.09 286
stage gear fault
Turbine with ring 0.04 0.49 8.29 60.76
defect
4 Conclusion
Wind turbine failure modes that require speed information to detect can be identified
with estimated speed information. The faults can be detected even when the estimated
speed is not 100% accurate. They are identified at the same time as when descriptors
based on real speed show development of faults. This gives the possibility of condition
monitoring of wind turbines without speed sensor, which could significantly reduce
the overall cost of wind turbine condition monitoring.
References
1. Bartelmus W, Zimroz R (2009) Vibration condition monitoring of planetary gearbox under

varying external load. Mech Syst Sig Process 23(1):246–257
2. Taylor JI (2000) The gear analysis handbook. The Vibration Institute
3. Coats MD, Sawalhi N, Randall R (2009) Extraction of tach information from a vibration signal
for improved synchronous averaging. Acoustics, Adelaide
4. Urbanek J, Zimroz R, Barszcz T, Antoni J (2012) Reconstruction of rotational speed from
vibration signal– comparison of methods. In: The 9th conference on condition monitoring and
machine failure prevention technologies, London, pp 866–878
5. Zhao M, et al (2013) A tacho-less order tracking technique for large speed variations. Mech
Syst Signal Process 40(1):76–90
6. Zimroz R, Milioz F, Martin N (2010) A procedure of vibration analysis from planetary gearbox
under non-stationary cyclic operations by instantaneous frequency estimation in time-frequency
domain. In: The 7th international conference on condition monitoring and machinery failure
prevention technologies, Stafford
7. Zimroz R et al (2011) Measurement of instantaneous shaft speed by advanced vibration signal
processing—application to wind turbine gearbox. Metrol Meas Syst 18(4):701–712
8. Skrimpas GA et al (2015) Speed estimation in geared wind turbines using the maximum
correlation coefficient. In: Annual conference of the prognostic and health management society,
San Diego
9. Randall RB, Smith WA (2018) Accuracy of speed determination of a machine from frequency
demodulation of response vibration signals. In: The 15th international conference on condition
monitoring and machinery failure prevention technologies, Nottingham
10. Randall R, Smith W (2016) Use of the teager kaiser energy operator to estimate machine speed.
In: European conference of the prognostic and health management society, Bilbao
11. Marhadi K, Skrimpas GA (2017) Using cepstrum information to detect gearbox faults in wind
turbine planetary stage gearbox. In: The 1st word congress on condition monitoring, London
Non-regular Sampling and Compressive
Sensing for Gearbox Monitoring
C. Parellier, T. Bovkun, T. Denimal, J. Gehin, D. Abboud, and Y. Marnissi
Abstract Vibration-based condition monitoring is a powerful field to achieve

preventive maintenance and control the efficiency of mechanical systems. The
common monitoring scheme consists of capturing vibration signals at a sufficiently
high sampling frequency to consider the dynamic and kinematic properties of the
machine, and then condition indicators are constructed and saved or sent to a ground-
based surveillance station. A major drawback of such an approach is that these
indicators may not be sufficient to establish an accurate diagnosis, and the signal
itself is needed in various situations. As the frequency rate of vibration signals and,
consequently, their size highly depends on the highest frequency to be monitored, it is
generally set high in order to respect the Nyquist theorem. As a consequence, sending
these data can be compromised because of the limited bandwidth of the transmis-
sion channel. This paper proposes a non-regular sampling strategy via compressive
sensing to significantly reduce the sampling frequency (below the Nyquist limit)
and the data size, while preserving enough information for diagnosis. It is based on
the hypothesis that the measured signal consists of a sum of sinusoids corrupted by
some random Gaussian noises. The signal recovery problem from random samples
is then formulated through an 2 − 1 optimization problem, which is solved via the
iterative shrinking-thresholding algorithm. At last, the potentiality of the suggested
approach is demonstrated on real vibration signals measured from a spur gearbox
with a spalling fault that progresses over 12 days.
Keywords Health monitoring · Compressive sensing · Non-regular sampling ·

1 -minimization · Iterative shrink thresholding · Gearbox
C. Parellier · T. Bovkun · T. Denimal · J. Gehin

Télécom Paristech, 46 Rue Barrault, 75013 Paris, France
D. Abboud (B) · Y. Marnissi
Safran Tech, Rue de Jeunes Bois, 78772 Châteaufort, Magny-les-Hameaux, France

https://doi.org/10.1007/978-981-15-9199-0_5
44 C. Parellier et al.
1 Introduction
Gearboxes are mechanical equipment used across a large range of industries. They
consist in at least a pair of gears engaged with each other to ensure power transmis-
sion. Their geometry determines the gear ratio and, therefore, the speed and torque
ratio. Consequently, the design and chipping state of the tooth have a major impact on
the gearbox efficiency. A broken tooth may lead to severe accidents in some critical
applications. In this respect, their in-service monitoring is crucial.
Vibration-based condition monitoring comprises a class of well-established
methods used in a predictive maintenance strategy. Sensors (e.g., accelerometers)
are mounted in the vicinity of the monitored components. Then, traditional methods
consist in acquiring signals on a regular basis, i.e., with a constant time step. The
sampling rate must be chosen high enough to consider the dynamic and kinematic
properties of the machine. Afterwards, the signal is processed to compute condi-
tion indicators, which provide health information about the rotating machine. While
this approach strongly condenses the data, other useful health information may be
missed. In fact, it is more suitable to have a reasonably close signal to the real one. In
many applications (e.g., aircraft engine monitoring), vibration signals are acquired at
relatively high frequency rates due to the very high rotating frequency of the engine
components. In detail, the sampling rate must respect the Nyquist/Shannon (NS)
theorem, which states that the sampling frequency must be greater than twice the
maximum frequency in the spectrum. This results in a huge volume of data, which
is constraining in terms of transmission bandwidth and/or memory storage.
Aiming to address the issues above, compressive sensing (CS) [1] is a continuously
growing research area in signal processing since 2006. It is an efficient data acquisi-
tion theory based on sparsity property of input signal. CS allows acquiring a sampled
signal and faithfully recovering the original signal while the (average) sampling rate
is much lower than the NS rate. Because of these advantages, CS is widely used in
numerous engineering areas such as electrical, electronic, and telecommunications
[2–4]. CS has proven its efficiency in rotating machinery faults detection [5] where
the sparse representation classification algorithm is applied to solve problems based
on 1 -norm minimization. Höglund et al. [6] applied CS to bridge damage detec-
tion, using a real-world vibration dataset, and proved its ability to recover a signal
sampled at a rate of 90%. Their solution used for sensing a random matrix with either
i.i.d. Gaussian values or random values from {1, −1} i.i.d according to a Bernoulli
distribution.
This paper intends an application of compressive sensing to vibration signals
produced by a spur gearbox. In order to detect gear damages, the implemented method
should guarantee a reliable recovered signal from the under-sampled observed signal.
The rest of the paper is organized as follows: Sect. 2 succinctly reviews the model
of vibrations generated by gears, in addition to some common gear condition indi-
cators. Then, the proposed monitoring approach is detailed. Section 3 focuses on
the demonstration of the recovering quality on real signals acquired on a test bench,
Non-regular Sampling and Compressive Sensing for Gearbox … 45
including a local gear fault in a spur gearbox that evolves across the experiment. The
paper is sealed with a conclusion in Sect. 4.
2 Gear Diagnosis by Compressive Sensing
This section starts with a brief review of the characteristic frequencies of a gearbox
subjected to a fault and related gear condition indicators. Next, the non-regular
sampling method is detailed. Last, the problem is formulated and an efficient
algorithm for signal recovery is proposed.
2.1 Gear Vibrations
The studied system consists of a simple spur gearbox with two wheels. The driver
gear transmits a rotational motion by meshing with the driven gear. These meshing
forces are the main source of vibration. In ideal situation, this force is spread evenly
among the tooth, thus its temporal profile is periodic (in the stationary case) with a
frequency called the gear meshing frequency, subsequently denoted by FGM ∈ R.
Each time a local fault in a tooth makes a contact under a certain load with another
tooth, an abrupt change in the contact force occurs leading to an impulse in the
vibration signal. The generated impulses are likely to modulate the meshing signal
and generate multiple sidebands in the spectrum carried by the FGM and its harmonics,
spaced by the faulty gear frequency. The number and energy of these sidebands are
likely to increase with the progression of the local fault.
FGM = Ni f i , where Ni ∈ N is the number of teeth of gear i ∈ N and f i ∈ R is
the rotating frequency of gear i. Since frequencies FGM and f 1,2 are known, a gear
condition indicator (GCI) Ii can be constructed:
L 2
k=−L S( f GM + k ∗ f i )
Ii = , (1)
S( f GM )2
where S( f ) denotes the signal spectrum, f is the frequency (in [Hz]) and L ∈ N is
defined such that 2L + 1 is the considered number of sidebands.
The crest factor (CF) is a condition indicator used in vibration-based condition
monitoring, known to be sensitive during early phases of the damage:
Speaks (s)
CF(s) = , (2)
RMS(s)
where Speaks (s) is the maximum value of the signal s and RMS(s) the root mean
square of signal s.
Indicators (1) and (2) are used afterwards as a part of the validating process of the
proposed approach Sect. 3.
2.2 Non-regular Sampling Method
One considers a continuous-time signal c : T → R, where T ⊂ R is the interval

observation and its highest frequency is denoted f max ∈ R. NS theorem imposes a
regular sampling at a frequency f NS = 1/TNS ≥ 2 f max to capture all the information
of c. This condition leads to a discrete time signal s ∈ R N , where N ≥ T /TNS =
2T f max .
A non-regular sampling can be performed to reduce the overall size of data needed
to describe periodic signals and then to obtain an observed signal y ∈ R M such that
M N . To achieve the non-regular sampling and control the sampling rate, the
sampled instants τm ∈ R are selected following the jittered random sampling:
T
τm = m + δm , (3)
M
where δm is an i.i.d. random variable with the uniform probability distribution U
bounded between 0 and T /M, which means δm ∼ U (0, T /M).
Numerically, the sampling process is computed with:
y = s, (4)
where ∈ R M×N is the measurement matrix.

To illustrate the previous method, matrix 1,2 ∈ R N /2×N are 2 realizations of
a simplified case with sampling rate M = N /2. One instant is sampled on each
interval highlights with a red square (5).
(5)
2.3 Problem Formulation
The non-regular sampling produces an observed signal y contaminated with inter-

fering and sampling noises. The compressed sensing framework gives an efficient
approach to recover a signal reasonably close to s [7]. Applying this technique, two
constraints need to be fulfilled: sparsity and incoherence.
Quasi-stationary mechanical signals are often quasi and/or multi-periodic in the
time domain and, thus, can be faithfully represented in the Fourier domain through
a small number of non-zero coefficients (which defines the Fourier coefficients of
a periodic signal). In other words, mechanical signals are likely to exhibit a sparse
spectrum characterized by numerous harmonics whose frequency depends on the
system kinematics. For this reason, the Fourier basis ∈ N N ×N is perhaps the
most convenient representation of the signal s. Incidentally, the incoherence property
between representation basis and measurement matrix is guaranteed by the random
measurement matrix previously detailed in Sect. 2.2. Equation (4) can be then
reformulated as:
y = −1 x + , (6)
where x ∈ R N is the signal s in frequency basis and a Gaussian noise independent

of signal.
2.4 Proposed Solution and Algorithm
To fulfill the sparsity condition on x while reducing the Gaussian noise , a common
approach to solve (6) is the least squares method with 1 -norm regularization, also
referred to as LASSO [8]:
2
x ∈ argmin y − −1 x 2 + λ x 1 ,

(7)
where . i is the i -norm and the λ ∈ R+ is the regularization term. The larger λ
is, the more the solution is influenced. A large λ results in a very sparse signal and
may lead to a poor estimation of peaks with low amplitudes. Conversely, decreasing
the value of λ may result in poorly sparse signal and high level of residual noise.
Therefore, the choice of λ can significantly affect the estimation.
To solve (7) numerically, one of the most popular method is the class of “iter-
ative shrinking-thresholding” algorithms called ISTA [9]. In order to improve the
algorithmic efficiency, it has been coded from scratch for the specific problem (7)
(Fig. 1).
In fact, the computation of Fourier Transform matrix is avoided due to its high
computational cost. Instead, the Fast Fourier Transform implemented in Scipy [10]
is used.
Fig. 1 Flowchart of the implementation of ISTA for the specific problem (7)
3 Application
Raw signals are measured at high definition on a test bench with an accelerometer
mounted in the vicinity of gearbox (Sect. 2.1). One signal is acquired per day, for
12 days, at a sampling frequency of 20 kHz. The gear 1 and gear 2 have, respectively,
N1 = 20 teeth and N2 = 21 teeth. FGM is around 330 Hz. The target phenomenon
to identify is a progressive local fault in gear 1.
Figure 2 illustrates the raw temporal signal and the relative spectrum of the first
day (healthy state) and of the last day (faulty state). In the latter, the presence of
a high number of sidebands with significant magnitude compared to the meshing
frequency magnitude is an obvious sign of the presence of a damage. The increase
of this ratio indicates a progression of the damage.
Fig. 2 Temporal and frequency signal at day 1, healthy state, and day 12, faulty state
Figures 3, 4 and 5 provides an overview of the recovered spectrum superposed on

the raw one. Since gear 1 fault begins to increase significantly from day 9, figures
are focused on days 1, 9, and 12. One observes that the nine harmonics of FGM are
recovered with a low information loss, both in terms of frequency and amplitude,
even for the frequencies higher than NS criterion.
Figure 6 compares the GCI (1) of raw signals and recovered ones of both gears. It
is computed on the four harmonics of sidebands around FGM . The high peak of gear
1 at day 11 indicates a strong spalling on this gear. The GCI values of gear 2 remain
low, which denotes a good gear condition. Finally, the recovering is efficient up to a
sampling rate of 0.2, since spalling condition of gear 1 is identified.
Figure 7 shows the CF (2) results. One observes that the trend is preserved with
any sampling rates: There is no significant fault before day 8 and a local fault appears
and strongly increases from day 9.
Fig. 3 Frequency spectrum of raw and recovered signal at day 1, for a sampling rate of 0.2
Fig. 4 Frequency spectrum of raw and recovered signal at day 9, minor fault, for a sampling rate
of 0.2
Fig. 5 Frequency spectrum of raw and recovered signal at day 12, major fault, for a sampling rate
of 0.2
Fig. 6 Evolution per day of GCI (1) computed on spur gear 1 and 2 for sampling rate ∈
{0.5, 0.4, 0.3, 0.2}
Fig. 7 Evolution of crest factor through days, computed for sampling rate ∈ {0.5, 0.4, 0.3, 0.2}
4 Conclusion
This study focused on the ability to transmit a mechanical vibration signal lightweight
but informative in condition monitoring framework. The sampling based on uniform
distribution via compressive sensing showed good performance in recovering the
signal. In particular, it was demonstrated that the proposed approach was able to
recover high frequency peaks in the spectrum, though being beyond the Nyquist
frequency. However, since the problem is reduced to a LASSO optimization problem,
its solution depends on the threshold hyper parameter λ which has a major impact on
the recovering efficiency. The search of an ideal parameter was performed through
a manual iterative method, which may be insufficient to deal with a large amount of
signals. Studying more input signals, one may be able to determine a set of value
of λ relevant to some signal classes. Moreover, this step may be automatized via a
Bayesian approach which takes part of some prior knowledge on the model variables.
Acknowledgements We thank A. Sabourin and C. Tillier of Télécom ParisTech (LTCI) who

provided a rich assistance.
References
1. Donoho DL (2006) Compressed sensing. IEEE Trans Inf Theor 52(4):1289–1306

2. Herman MA, Strohmer T (2009) High-resolution radar via compressed sensing. IEEE Trans
Signal Process 57(6):2275–2284
3. Duarte MF et al (2008) Single-pixel imaging via compressive sampling. IEEE Signal Process
Mag 25(2):83–91
4. Tian Z, Giannakis GB (2007) Compressed sensing for wideband cognitive radios. In:
International conference on acoustics, speech and signal processing, Honolulu, pp 1357–1360
5. Tang G (2015) Sparse classification of rotating machinery faults based on compressive sensing
strategy. Mechatronics 31:60–67
6. Höglund J, Wei B, Hu W, Karoumi R (2014) Compressive sensing for bridge damage detection.
In: Proceedings of 5th Nordic workshop, on system and network optimization for wireless
7. Candès EJ, Romberg JK, Tao T (2006) Stable signal recovery from incomplete and inaccurate
measurements. Commun Pure Appl Math 59(8):1207–1223
8. Tibshirani R (1996) Regression shrinkage and selection via the lasso. J Roy Stat Soc Ser B
58(1):267–288
9. Combettes PL, Pesquet JC (2011) Proximal splitting methods in signal processing, Springer
series optimization A, vol 49. Springer, New York, pp 185–212
10. Jones E et al (2019) SciPy: open source scientific tools for python, 2019-06. https://www.scipy.
org/
Change Detection in Smart Grids Using
Dynamic Mixtures of t-Distributions
A. Samé, M. L. Abadi, and L. Oukhellou
Abstract The analysis of massive amounts of data collected via smart grids can
contribute to manage more efficiently energy production and water resources. Within
this framework, a method is proposed in this article to detect changes in panel data
from smart electricity and water networks. The adopted approach is based on a repre-
sentation of the data by clusters whose occurrence probability varies in the course
of time. Because of the cyclical nature of the studied data, these probabilities are
designed to be piecewise periodic functions whose break points reflect the intended
changes. A mixture of t-distributions with a hierarchical structure forms the basis of
our proposal. This model was also chosen for its robustness properties. The weights of
the mixture, at the top of the hierarchy, constitute the change detectors. At the lowest
hierarchy level, the mixture weights are designed to model the periodic dynamics of
the data. The incremental strategy adopted for parameter estimation makes it possible
to process large amounts of data. The practical relevance of the proposed method is
illustrated through realistic urban data.
Keywords Energy and water monitoring · Smart grid · Panel data · Change
detection · Segmentation and clustering · Incremental learning
1 Introduction
Nowadays, with the advent of smart cities, high-resolution data acquired through
sensor networks allow for a more accurate monitoring of energy and water, simul-
taneously at different points of the urban area. Processing this data can help ensure
a best compromise between supply and demand, thus leading to natural resource
savings. In this context, the data usually collected are consumption measurements
collected and transmitted via smart meters. As these data describe the dynamics of a
set of entities (individual houses or collective buildings) over time, we can formalize
them as panel data. New algorithms based on flexible latent variable models are
A. Samé (B) · M. L. Abadi · L. Oukhellou

COSYS-GRETTIA, Univ Gustave Eiffel, IFSTTAR, F-77447 Marne-la-Vallée, France
e-mail: allou.same@ifsttar.fr

https://doi.org/10.1007/978-981-15-9199-0_6
54 A. Samé et al.
useful for the descriptive and predictive analysis of these data. As shown in the
following section, several researches have already been conducted to address this
complex problem.
The main purpose of this paper is the (off-line) change detection in such panel data,
a problem which can also be seen as the data segmentation along the time axis. The
proposed approach is based on the representation of panel data, which are dynamic
in nature, by a set of classes whose occurrence probability varies over time. To detect
changes while taking into account the cyclic nature of the data, the prior probability
of classes is assumed to be piecewise periodic. In this way, we assimilate the break
times of the latter probabilities to the change points. To formalize these assumptions,
we rely on a dynamic mixture of t-distributions with a hierarchical structure, which
has also been chosen for its robustness properties. To allow the processing of large
amounts of data, the model parameters are estimated using an incremental expecta-
tion–maximization algorithm derived from the variational method initiated by Neal
and Hinton [17].
The article is structured as follows. Section 2 raises the problem of change detec-
tion from panel data and provides a brief state of the art on this issue. Section 3
introduces the proposed hierarchical mixture model dedicated to change detection
from panel data. Section 4 describes the incremental strategy adopted for parameter
estimation, and Sect. 5 is dedicated to the application of this method on a realistic
dataset issued from a smart grid.
2 Problem Statement and State of the Art
In this paper, we focus on energy, water, or gas consumption data from smart
meters associated to individual houses or collective buildings in an urban area. We
assume that, for each meter, the consumption is available at hourly intervals during
several days. As the main interest of this article lies in the dynamics of a panel of
consumers on a daily time scale, the data will be denoted as (x 1 , . . . , x T ), where
x t = (x t1 , . . . , x tn ) will designate the cross-sectional sample extracted on day t ≤ T
from n meters, with x ti ∈ Rd . In this study, each individual x ti consists of d = 24
values (hourly consumption).
The problem addressed in this paper is that of change detection in these panel
data. Let us recall that this problem consists in finding time instants at which changes
occur in the data distribution. This process requires the data distributions, before and
after the change point, to be determined. Figure 1 illustrates the problem of change
detection in a panel of daily consumption curves.
Several studies have already focused on the problem of change detection in panel
data. In [11, 12], the so-called random break model has been proposed. It assumes
that each time series has its own break point. The likelihood, which is similar to that
of a mixture model, is maximized via a specific Expectation–Maximization (EM)
algorithm. In [2, 3, 10], a statistical approach is used for estimating a common break
in the mean and the variance of panel data. The parameters estimation is performed
Change Detection in Smart Grids Using Dynamic Mixtures … 55
Fig. 1 Change detection in a panel of daily consumption curves from a water grid
via a quasi-maximum likelihood method. To detect a change in the mean of panel

data, an adaptation of the cumulative sum (CUSUM) method has been developed in
[4, 9], and a CUSUM-based statistic has been proposed in [14] to detect a common
variance change in panel data. The problem of multiple change point detection in
panel data has been addressed in [5] through the use of a double-CUSUM statistic
and a bootstrap procedure.
3 Simultaneous Change Detection and Clustering
The method proposed for change detection in panel data is based on a mixture of
multivariate t-distributions whose mixing weights, themselves, are modeled as a
temporal mixture model which is piecewise periodic. A two-way clustering of panel
data is derived from this approach: a clustering of observations x ti to one of K clusters
and a clustering of the time instants t to one of L contiguous segments. This latter
clustering constitutes the detection of change points. The model and its parameter
estimation strategy are successively described in the following subsections.
3.1 Robust Mixture Modeling
According to the proposed model, each observation x ti is assumed to be generated

from the following hierarchical mixture distribution:
L

K
p(x ti ; θ ) = πl (t; α)πlk (ut ; β l ) f (x ti ; μk , k , ϑk ), (1)
k=1 l=1
where θ represents the set of parameters of the model, L and K are the numbers of
mixture components at the two levels of the hierarchy, and f is the density function of
a multivariate t-distribution with mean μ ∈ Rd , scale matrix ∈ Rd×d , and degrees
of freedom ϑ, defined by
ϑ+d
− 1 (ϑ+d)
(x − μ) −1 (x − μ) 2
f (x; μ, , ϑ) = 1+
2
, (2)
(π ϑ)1/2 ||1/2 ϑ2 ϑ
56 A. Samé et al.
where is the Gamma function. Here, we use the multivariate t-distribution rather
than the classic Gaussian distribution for its more generic character (a Gaussian
density is obtained when ϑ tends to infinity) and for its robustness to outliers [16].
L
πl (t; α)πlk (ut ; β l ) = pk (t) allow for a
def
For our model, the global weights l=1
segmentation of the time axis into contiguous parts, while remaining periodic in each
segment. To this end, the mixing weights πl are defined by
exp(αl1 t + αl0 )
πl (t; α) = , (3)
h exp(αh1 t + αh0 )
with α = (αl0 , αl1 )l=1,...,L ∈ R2L . Using linear functions of time inside the logistic
transformation leads to the desired segmentation [18]. For instance, for L = 2
segments, the single change point (t ∗ = −αl0 /αl1 ) corresponds to the intersection
of two logistic functions [18].
In order to allow the proportions πlk vary periodically in the course of time while
taking into account special calendar events (e.g., public holidays); they are defined
as

exp β lk ut
πlk ut ; β l = , (4)
h exp β lh ut

where β l = β l1 , . . . , β l K is the set of parameters to be estimated, with β lk =
(βlk0 , . . . , βl,k,2q+1 ) ∈ R2q+2 , and ut = (1, C1 (t), S1 (t), . . . , Cq (t), Sq (t), 1sd (t))
is the corresponding covariate vector, with

2π jt 2π jt 1 if day t is special
C j (t) = cos , S j (t) = sin , 1sd (t) = (5)
m m 0 otherwise.
It should be noticed that m = 7 (weekly seasonality), and q(≤ m/2)

is the required number of trigonometric terms, where a is the integer part
of a. If there is no special day in the studied period, then we use ut =
(1, C1 (t), S1 (t), . . . , Cq (t), Sq (t)) as covariate vector with the new set of parame-
ters β lk = (βlk0 , . . . , βl,k,2q ) . For identifiability purposes, the parameters (α L0 , α L1 )
and β l K are set to be the null vector.
3.2 Formulation in the Form of a Generative Model
From a generative point of view, the proposed model can be defined as follows:
1. Generation of the cluster labels z ti ∈ {1, . . . , K } of observations x ti as follows:
(a) drawing of segment labels wti ∈ {1, . . . , L} according to the multinomial
distribution M(1; π1 (t; α), . . . , π L (t; α));
(b) drawing of cluster label z ti ∈ {1, .. . , K }, given wti = l, according
to the
multinomial distribution M 1; πl1 ut ; β l , . . . , πl K ut ; β l ;
2. generation of x ti , given z ti = k, from the t-distribution with parameters
μk , k , ϑk as follows:

(a) generation of yti according to the Gamma distribution ϑ2k , ϑ2k ;

(b) generation of x ti , given yti , according to the distribution N μk , ytik .
Therefore, the proposed model involves the discrete latent variables w =

(w1 , . . . , w T ) and z = (z 1 , . . . , z T ), and the continuous latent variables y =
( y1 , . . . , y T ), with wt = (wti )i=1,...,n , z t = (z ti )i=1,...,n , and yt = (yti )i=1,...,n .
The graphical representation corresponding to the proposed model is given in Fig. 2.
Figure 3 displays examples of weights π l and π lk , and cluster labels z ti generated
using (L = 2, K = 3). As it can be seen, the mixture proportions of the highest level
of the hierarchy allow for an effective segmentation of the time axis (change point
at t = 36), while the mixture proportions of the bottom of the hierarchy take into
account the periodic dynamics of the data. The cluster labels generated from the
combination pk (t) of these proportions show two successive periodic behaviors.
4 Incremental Learning
Given the temporal data x = (x 1 , . . . , x T ) and their covariates (u1 , . . . , u T ), the

parameter vector θ is estimated by maximizing the log-likelihood

L(θ ) = log πl (t; α)πlk ut ; β l f (x ti ; μk , k , ϑk ) (6)
t,i k,l
Fig. 2 Graphical probabilistic representation associated to the proposed model

58 A. Samé et al.
Fig. 3 Examples of mixture proportions obtained with L = 2, K = 3, and q = 3, and cluster

labels zti generated from these probabilities: (top) the mixture proportions at the highest level of
the hierarchy allow the detection of a change at t = 36; (middle) the mixture proportions at the
bottom of the hierarchy take into account the periodic nature of the data; (bottom) the cluster labels
zti generated from the combination pk (t) of these proportions show two different periodic behaviors
through the Expectation–Maximization (EM) algorithm [7, 15]. For our specific
mixture model, each step of this algorithm requires the computation of two categories
of posterior probabilities, in addition to the use of multivariate t-distributions. In the
perspective of processing large amounts of data, an incremental version of the EM
algorithm is proposed. Before describing this version, we start by developing the
off-line version of the algorithm.
4.1 EM Algorithm
Using the generative formulation of the model, the complete log-likelihood can be
written as

ϑk ϑk k
CL(θ ) = wtil z tik log πl (t; α)πlk ut ; β l yti ; , N x ti ; μk , ,
t,i,k,l
2 2 yti
(7)
where (.; a, b) is the Gamma probability density function with parameters a and b;
N (.; μ, ) is the Gaussian density with mean μ and covariance matrix . Variables
wtil and z tik are defined as follows: wtil = 1 if wti = l and 0 otherwise, and z tik = 1
if z ti = k and 0 otherwise. The EM algorithm starts from an initial parameter θ (0)
and iterates the following two steps until convergence.
Expectation Step
Computation of the expected complete log-likelihood conditionally on the observed
data and the current parameter θ (c) , which is given by the following auxiliary function
Q:

Q θ , θ (c) = E CL(θ )|x 1 , . . . , x T ; θ (c)
(c) (c) (c)
= Q (c)
1 (α) + Q 2l β l + Q 3k μk , k + Q 4k (ϑk ), (8)
l k k
with

Q (c)
1 (α) =
(c)
τtilk logπl (t; α) (9)
t,i,l,k
(c)
Q (c)
2l β l = τtilk logπlk ut ; β l (10)
t,i,k

k
Q (c)
3k μk , k = (c)
τtilk logN x ti ; μk , (11)
t,i,l
yti

ϑk ϑk
Q (c)
4k (ϑk ) = (c)
τtilk yti ; ,
t,i,l
2 2
(c)
(c) ϑk ϑk + d ϑk(c) + d
+ τtilk −1 ψ − log . (12)
t,i,l
2 2 2
where ψ refers to the Digamma function defined by ψ(a) = ∂ log (a)/∂a =

(c)
(a)/ (a), τtilk is the posterior probability that x ti originates from the kth cluster
and the lth segment given the current parameter, and λ(c) tik is the conditional expectation
60 A. Samé et al.
of yti given the current parameter. These expectations can be written as
(c)
τtilk = E wtil z tik |x ti ; θ (c)

πl t; α (c) πlk ut ; β l(c) f x ti ; μ(c) k , (c)
k , ϑk
(c)
=
(c)

, (13)
l,k πl t; α
(c) π
lk ut ; β l f x ti ; μk , k , ϑk(c)
(c) (c)

λ(c)
tik = E yti |z ti = k, x ti ; θ
(c)
ϑk(c) + d
=

−1
. (14)
ϑk(c) + x ti − μ(c)
k (c)
k x ti − μ (c)
k
Maximization Step
Computation of the new parameter θ (c+1) by maximizing the Q function with respect
to θ . The maximizations of Q (c) (c)
1 and Q 2l are logistic regression problems weighted
(c)
by the probabilities τtilk , which cannot be solved in a closed form. The Iteratively
Reweighted Least Squares (IRLS) algorithm [8, 13] is exploited to this end. In a
similar way as for the standard
mixture
of Gaussian distributions [15], we perform
the maximization of Q (c)
3k μ k , k .
4.2 Incremental Algorithm
As is usually the case with smart meters data, the EM algorithm can become slow
for massive panel data (T × n relatively large) since each of its iterations, and more
particularly the E-step, requires visiting the whole dataset. A solution would consist
in distributing the E-step among several process units [6, 19], which are made possible
by the fact that the posterior probabilities τtilk that can be computed independently.
In this article, we opted for another approach, which consists in adapting the incre-
mental strategy initiated by Neal and Hinton [17] to our context of periodic mixture
modeling. Before describing our incremental algorithm, it should first be emphasized
that the EM algorithm previously described can also be written by using only the
following sufficient statistics:
(c)
(c) (c)
(c)
S1lk = s1tlk where s1tlk = τtilk (15)
t i
(c)
(c) (c)

S2lk = s2tlk where s2tlk = τ (c) λ(c) (16)
t i tilk tik
(c)
(c) (c)

S3lk = s3tlk where s3tlk = τ (c) logλ(c) (17)
t i tilk tik
(c)
(c) (c)

S4lk = s4tlk where s4tlk = τ (c) λ(c) x ti (18)
t i tilk tik
(c)
(c) (c)

S5lk = s5tlk where s5tlk = τ (c) λ(c) x ti x ti . (19)
t i tilk tik
To accelerate the EM algorithm without visiting the whole dataset, we use partial
E-steps and incremental parameter updating, which are based on the computation
of sufficient statistics [17]. This strategy also has the advantage of accelerating the
convergence of the algorithm while keeping its monotony property. We opted in
this article for specific partial E-steps, where the posterior probabilities τtilk and
λtik are computed only for a day t. The additive nature of the expected sufficient
statistics defined by (15)–(19) enables their incremental updating. The main steps
of this procedure is detailed by the following pseudo-code. It will be called IEM-
SPMIX, where I stands for “incremental,” EM for “Expectation–Maximization,” S
for “segmental,” P for “periodic,” and MIX for “mixture.”

Once the estimate θ is provided by the algorithm, a contiguous segmentation of

the time instants can be obtained by setting El (l = 1, . . . , L) as follows:

El = t ∈ [1; T ]; πl t; α = max1≤h≤L πh t; α . (20)
The estimated change points are therefore the starting time stamp of segments E l .
A partition of the data is deduced by setting

z ti = k ⇔ k = argmax1≤h≤L P z ti = h|x ti ; θ = argmax1≤h≤L τtilk . (21)
l
where τtilk is given by (13).

62 A. Samé et al.
5 Experiments
In this section, realistic consumption data similar to those collected from real world
water networks are used to evaluate the proposed change detection algorithm. They
are obtained through simulations from a specific dynamic model [1]. With these data,
the ground truth segmentation is available, thereby allowing the method evaluation.
Let us recall that each observation x ti ∈ R24 corresponds to the behavior adopted
by a user i during a day t (there is one numeric value per hour). The dataset has L
= 3 segments (change at t = 36 and t = 71), and K = 8 clusters. The numbers of
time instants is T = 105 days (which corresponds to 10 weeks,) and the number
of individuals is n = 100. These data can also be formalized by a tensor of 100
× 105 × 24 numeric values. A part of these data is displayed in Fig. 4, where
each subplot represents the set of observations (x ti )i=1,...,n of a day t, x ti being
displayed longitudinally as a time series of 24 values. Since the main concern of
this study is change detection in consumption behavior (time series shape) rather
than consumption volume, an additional transformation has been applied to the data:
each series x ti = (x tis )s=1,...,24 was standardized by setting xtis to (x tis − x ti )/σ (x ti ),
where x ti and σ (x ti ) are, respectively, the mean and standard deviation of x ti .
Fig. 4 Extract of a realistic panel dataset (20 smart meters observed during 16 days)
64 A. Samé et al.
Table 1 BIC criterion in relation to the number of segments and the number of clusters
K =6 K =7 K =8 K =9 K = 10
L=1 371,954.51 371,014.72 369,707.24 370,405.08 370,614.00
L=2 369,621.13 368,902.14 367,214.11 367,838.20 368,290.31
L=3 368,327.83 367,193.04 365,951.70 366,382.93 366,992.62
L=4 368,536.86 367,563.87 366,413.30 366,883.04 367,651.21
L= 5 368,833.13 367,990.21 366,940.57 367,544.82 368,241.98
Table 2 BIC criterion based on Gaussian mixture model, in relation to the numbers of segments
and clusters
K =6 K =7 K= 8 K =9 K = 10
L=1 393,606.63 392,564.63 391,145.70 391,627.69 392,105.33
L= 2 391,304.51 390,223.47 388,635.08 389,140.04 389,729.94
L=3 389,826.11 388,684.20 387,099.98 387,729.74 388,307.09
L=4 390,081.98 389,019.08 387,494.16 388,189.17 388,861.90
L= 5 390,413.55 389,427.82 387,923.91 388,692.74 389,448.64
5.1 Model Selection
To select the best couple (L, K), we have relied on the minimization of the Bayesian

Information Criterion (BIC), which is defined as B I C(L , K ) = −2L θ +

m L ,K log(nT ), where m L ,K is the number of free parameters of the model, and θ

is the maximum likelihood estimates provided by the IEM-SPMIX algorithm. This
estimation algorithm has thus been launched for different pairs (L, K), with L varying
in the set {1, …, 5} and K varying in the set {6, …, 10}. The results achieved are
given in Table 1. The minimum value of BIC is obtained for the pair (L = 3, K = 8),
which corresponds to the reality of the data.
We also compared these BIC values with those achieved by using a Gaussian
distribution instead of the t-distribution. The results obtained are given in Table 2.
As it can be observed, the resulting values of BIC in this case are higher than those
obtained with the proposed model, which shows that our model is better adapted to
the considered data.
5.2 Results
Using the number of segments (L = 3) and the number of clusters (K = 8) which

have been selected in the previous section, the algorithm IEM-SPMIX has been
launched on the panel data (x ti )t=1,...,T,i=1,...,n . Figure 5 shows the resulting daily
Fig. 5 Daily cluster labels and change points (vertical lines) obtained with IEM-SPMIX (left);
consumption patterns corresponding to the clusters (right)
cluster labels together with the estimated change points (vertical lines at day 36 and

day 71) and the eight estimated means μk of the t-distributions. The estimated change
points coincide with the true change points. The displayed means constitute typical
consumption behavior patterns. For instance, pattern μ2 which is a typical behavior
of active residents (two well-separated consumption peaks) and is frequent during
the working days of segment 2 while pattern μ7 is frequent during the weekends of
segment 3.
Thanks to our robust mixture modeling approach, Fig. 6 allows us to see, within
the panel data, the observations that can be considered as noise and displays
some
of these observations.
−1 An observation
x ti has been considered as noise if
τ
l,k tilk x ti − μk k x ti − μ k > χ d;0.95 where χd;α is the α-percentile of the
2
, 2
Chi-square distribution with d degrees of freedom [16]. In practice, we observed

that these noisy data could generally be attributed to atypical behaviors specific to
the studied panels and, depending of the period, to exogenous factors (temperature,
precipitation).
Fig. 6 Noisy observations within the panel (left); corresponding consumption curves (right)
66 A. Samé et al.
The estimated mixture proportions are displayed in Fig. 7. As shown in Fig. 7a,
the change points are given by t = 36 and t = 71. The segment–specific periodic
proportions are displayed in Fig. 7b–d. They show the local dynamics of the panel data
distribution. The global dynamic behavior
of the data is
given by Fig. 7e, which gives

L

the time evolving probability pk (t) = P z ti = k; θ = l=1 πl (t; α )πlk (ut ; β l ).

Using this plot, we can distinguish the differences between segments, which can help
give insight to the data.

Fig. 7 Estimated mixture proportions: a πl t; α , b–d πlk ut ; β l , and e pk (t)
6 Conclusion
A change detection method, specially dedicated to periodic panel data emanating

from smart electricity and water networks, has been proposed in this article. A hier-
archical mixture of t-distributions whose weights evolve over time that forms the
basis of this method. The weights of the mixture, at the top of the hierarchy, consti-
tute the change detectors. At the lowest hierarchy level, the mixture weights are
designed to model the periodic dynamics of the data. An incremental strategy, which
is based on the expected sufficient statistics of the model, has been adopted for
parameter estimation. It allows keeping the monotonic convergence properties of the
log-likelihood. The proposed method proved to be effective on realistic simulated
data. As short-term perspectives of this work, comparisons of our method with a
two-phase approach involving clustering and then segmentation will be conducted.
Although BIC can help selecting the number of clusters and segments, we observed
in practice that an underestimation of the number of clusters could lead to poor detec-
tion of changes, which is not the case when the number of clusters is over-estimated.
The consideration of this fact in the model selection phase therefore deserves to be
studied.
References
1. Abadi ML, Samé A, Oukhellou L, Cheifetz N, Mandel P, Féliers C, Chesneau O (2017) Predic-
tive classification of water consumption time series using nonhomogeneous markov models.
In: IEEE DSAA, pp 323–331
2. Bai J (2010) Common breaks in means and variances for panel data. J Econometrics 157(1):78–
92
3. Bai J, Carrion-I-Silvestre JL (2009) Structural changes, common stochastic trends and unit
roots in panel data. Rev Econ Stud 76(2):471–501
4. Chan J, Horváth L, Hušková M (2013) Darling–erdős limit results for change-point detection
in panel data. J Stat Plann Infer 143(5):955–970
5. Cho H et al (2016) Change-point detection in panel data via double cusum statistic. Electron J
Stat 10(2):2000–2038
6. Dean J, Ghemawat S (2008) MapReduce: simplified data processing on large clusters. ACM
51(1):107–113
7. Dempster AP, Laird NM, Rubin DB (1977) Maximum likelihood from incomplete data via the
EM algorithm. J Roy Stat Soc B 39:1–38
8. Green P (1984) Iteratively reweighted least squares for maximum likelihood estimation, and
some robust and resistant alternatives. J Roy Stat Soc B 46(2):149–192
9. Horváth L, Hušková M (2012) Change-point detection in panel data. J Time Ser Anal 33(4):631–
648
10. Im KS, Lee J, Tieslau M (2005) Panel lm unit-root tests with level shifts. Oxford Bull Econ
Stat 67(3):393–419
11. Joseph L, Wolfson DB (1992) Estimation in multi-path change-point problems. Commun Stat
Theor Methods 21(4):897–913
12. Joseph L, Wolfson DB (1993) Maximum likelihood estimation in the multi-path change-point
problem. Ann Inst Stat Math 45(3):511–530
68 A. Samé et al.
13. Krishnapuram B, Carin L, Figueiredo MAT, Hartemink AJ (2005) Sparse multinomial logistic
regression: fast algorithms and generalization bounds. IEEE Trans Pattern Anal Mach Intell
27(6):957–968
14. Li F, Tian Z, Xiao Y, Chen Z (2015) Variance change-point detection in panel data models.
Econ Lett 126:140–143
15. McLachlan GJ, Krishnan T (2008) The EM algorithm and extensions. Wiley, New York
16. McLachlan GJ, Ng SK, Bean R (2016) Robust cluster analysis via mixture models. Aust J Stat
35(2):157–174
17. Neal R, Hinton G (1998) A view of the EM algorithm that justifies incremental, sparse, and
other variants. In: Jordan MI (ed) Learning in graphical models. Springer, Dordrecht (1998)
18. Samé A, Chamroukhi F, Govaert G, Aknin P (2011) Model-based clustering and segmentation
of time series with changes in regime. Adv Data Anal Classif 5(4):301–321
19. Wolfe J, Haghighi A, Klein D (2008) Fully distributed EM for very large datasets. In: ICML’08,
pp 1184–1191
Unbalance Detection and Prediction
in Induction Machine Using Recursive
PCA and Wiener Process
A. Picot, J. Régnier, and P. Maussion
Abstract This paper focuses on the detection of unbalance in induction machines

and the prediction of its dynamic evolution based on the analysis of the current
signals. The proposed method is based on the combination of recursive Principal
Component Analysis (PCA) and a Wiener process. The recursive PCA is processed
on the different features computed from the current signals in order to choose the
most relevant ones. This way the fault detection is processed with the only knowledge
of the healthy state of the machine. A linear Wiener process is then used to model
the behavior of PCA components and predict their evolution over time in order to
estimate the dynamic of the tracked fault along with the remaining useful life (RUL).
The proposed method is applied on real data from a 5.5 kW induction machine with
three different levels of unbalance and obtains very promising results.
Keywords Principal component analysis · Wiener process · Induction machine ·

Unbalance diagnosis · Current-based diagnosis
1 Introduction
Condition monitoring of electrical machines has become an important industrial

concern for safety and reliability. Electrical and mechanical faults have been well
documented over the years in the case of single faults scenarios. Unfortunately, the
industrial reality is often more complex, especially when there is no a priori on the
faults to happen. Mechanical faults have different origins: unbalance, eccentricity,
or bearings damages. The most popular approach to diagnose these different faults
A. Picot (B) · J. Régnier · P. Maussion

LAPLACE, Université de Toulouse, CNRS, Toulouse, France
e-mail: antoine.picot@laplace.univ-tlse.fr
J. Régnier
e-mail: jeremi.regnier@laplace.univ-tlse.fr
P. Maussion
e-mail: pascal.maussionr@laplace.univ-tlse.fr

https://doi.org/10.1007/978-981-15-9199-0_7
70 A. Picot et al.
is called the Motor Current Signal Analysis (MCSA) and has been widely docu-
mented [1]. This method is based on the spectral analysis of motor currents, gener-
ally computed using the Fast Fourier Transform (FFT ). Mechanical faults modify
the spectral contents of the currents through specific harmonic components. These
harmonic components depend on the fault and the system characteristics [2–4].
Unfortunately, the electro-mechanical complexity is generally increased in indus-
trial applications due to the use of gearboxes, screws, external bearings and couplings
among others. It results on a high risk of fault effects overlapping, even more consid-
ering multiple operating conditions. Artificial intelligence techniques have proven to
be very effective to tackle the issue of multi-faults scenarios [5–7]. In the particular
case where only the healthy functioning of the system is known, novelty detections
techniques such as Self-Organising Maps (SOM) [8] or recursive Principal Compo-
nent Analysis (PCA) [9] have shown promising results. For health management,
predicting the useful life left at a particular time of operation is more interesting
than detecting a fault. This way, it is possible to correct the system before the failure
happens and so to decrease the length of production shutdowns. This is called the
Remaining Useful Life (RUL). A review on the different techniques for the estima-
tion of RUL is given in [10]. Among these different techniques, the Wiener process is
an interesting way to model the problem. The RUL is estimated as a time dependant
drift added with Brownian noise [11, 12].
This paper proposes to use recursive PCA along with a Wiener process in order to
detect unbalance in induction machines. The recursive PCA is used in order to select
the most relevant features for the detection knowing only the healthy behavior of the
machine. A 2D Wiener process is then applied in the PCA plan in order to predict the
evolution of the PCA features across time. Both techniques are presented in Sect. 2.
The proposed method is evaluated in Sect. 3 on experimental data measured on a
5.5 kW induction machine with different unbalance levels.
2 Proposed Method
2.1 Recursive PCA
PCA is a multivariate statistic method that transforms possibly correlated variables

into uncorrelated ones by maximizing the variance between them. The new variables
are called Principal Components and are linear combinations of the original ones.
This well known technique is often used to reduce the size of the data set by extracting
the most important information from the data set [13]. The idea of recursive PCA as
presented in [9] is to process PCA on the first observations that constitute a healthy
database. The data are then projected in this new space. If several new observations
in a row are far away from the healthy group, it is considered as a new fault and a new
PCA is processed on the healthy data and the faulty data in order to obtain the best
subspace to separate them. If not, the PCA is not updated. This process is repeated
Unbalance Detection and Prediction in Induction Machine … 71
Fig. 1 Algorithm of the recursive PCA
every time the observations do not fit existing groups. The algorithm of recursive
PCA is depicted in Fig. 1.
2.2 Wiener Process
As described in [11, 12], a Wiener process based degradation X(t) is generally

described as
X (t) = X (0) + λt + σ B(t) (1)
where X(0) is the initial degradation value (usually equal to 0), λ is the drift parameter,
σ > 0 is the diffusion coefficient, and B(t) is the standard Brownian motion. Such a
model assumes that the drift rate is constant. The parameters λ and σ are recursively
estimated from the data using a Maximum Likelyhood Estimator (MLE). Then,
the degradation estimation at time t i with current observation x i can be estimated
according to (2) for t > t i .
72 A. Picot et al.
X (t) = xi + λ(t − ti ) + σ B(t − ti ) (2)
In this work, the Wiener process is applied on PCA features in order to follow and
predict the system evolution in the PCA space. However, the PCA subspace is likely
to change when a new fault is detected because of the recursive PCA algorithm. Let
be x and y the coordinates in the first PCA subspace (before fault detection) and x
and y the coordinates in the new one (after fault detection). Variable (x,y) can be
expressed as Wiener processes according to (3).

x(t) = λx t + σx B(t)
(3)
y(t) = λ y t + σ y B(t)
It is supposed that new coordinates (x ,y ) can be expressed as a linear combination
of (x,y) as

x (t) = ax(t) + by(t)
(4)
y (t) = cx(t) + dy(t)
with a, b, c, d some constants that can be computed using already known observations.
The new coordinates can then be expressed as a Wiener process where the new Wiener
parameters can be expressed as a linear combination of old ones according to (5).
λx = aλx + bλ y
σx = aσx + bσ y
(5)
λy = cλx + dλ y
σy = cσx + dσ y
3 Material and Results
3.1 Motor Unbalance Experiments
The present method is tested on current signals recorded on a 5.5 kW induction

machine (IM) with two poles pairs. The load level of the IM is regulated by a direct
current generator. An iron disk placed on the shaft is used to create different unbalance
levels. The experimental system is depicted in Fig. 2.
The IM is fed by an open-loop inverter at 40 Hz and operates at a low load level.
The phase currents of the IM are acquired with a sample frequency of f e = 10 kHz.
A total of 250 5 s-recordings of the IM currents have been measured on the machine.
These measurements include 100 recordings with no unbalance corresponding to the
healthy state of the machine, 50 recordings with a low level of unbalance, noted B1 ,
created by a small weight (77.5 g) fixed to the iron disk, 50 recordings with a medium
Fig. 2 Experimental test bench
level of unbalance, noted B2 , created by a medium weight (133 g), and 50 recordings
with a high level of unbalance, noted B3 , created by a heavy weight (274 g). These
unbalances produce load torque oscillations at f r with an amplitude of respectively
0.15%, 0.27%, and 0.55% of the IM rated torque.
Several temporal and frequential features are computed from the different
recordings. The list of the different features can be found in [9].
3.2 Results and Discussion
The distance between the current observation and the healthy state is displayed in
Fig. 3. The recursive PCA detected a new fault at recording 100 and a new PCA has
been performed at this time. It is enlighted in Fig. 3 as huge step on the distance
criteria that is due to the apparition of B1. This step is indeed amplified by the PCA
as explained in [9]. Then, the distance criteria remains high but no significative
difference is found between the 3 unbalance levels. The algorithm considers it as
the same fault and no new PCA is processed. So, the default is well detected by the
method but no difference is found between the 3 faults severities. It is interesting to
note that the fault was detected knowing only the healthy state and without any a
priori on the features to monitor.
The 2 main components only are kept with the PCA. It is then possible to follow
the evolution of the system in modelling these 2 components as a Wiener processes.
Figure 4 displays 3 examples of prediction at 50 recordings. Figure 4a is just before
the PCA is processed and Fig. 4b right after (recordings 100 and 101). Figure 4c
displays the prediction at recording 220. In this figure, the observation at t + 50
recordings is displayed as a magenta square and its prediction as a red circle. The
74 A. Picot et al.
Fig. 3 Distance between the current observation and the healthy state
Fig. 4 Examples of prediction at time t = 100 (a), t = 101 (b) and t = 220 (c)
dashed circle represents the standard deviation of the prediction. The blue crosses
are the past observations at current time t.
At time t = 100 recordings, the prediction is not good because the standard
deviation is large and the observation is far from the predicted point. This is due to
the fact that the subspace is not optimized for the fault detection. Once the fault is
detected and the PCA is processed, the new subspace is better suited for the fault
detection and the prediction is more accurate: the observation is closer to the predicted
one and is inside the standard deviation circle (t = 101 recordings). At time t = 220
recordings, the prediction gets better and better. It is very close to the observation
and the error circle is reduced. Figure 5 shows the error between the prediction and
the observation for the proposed method (blue curve). As an indication, the error
obtained when using the Wiener process only is depicted as the dashed red curve.
The prediction error is decreasing a lot at the beginning because the Wiener process
parameters are identified more and more precisely. At recording 100, the PCA is
processed and the prediction error becomes smaller for the proposed method. This
is due to the new subspace that is optimized for unbalance monitoring. At the end,
the prediction error decreases again. Overall, the prediction error is smaller with the
proposed method.
Fig. 5 Error between predictions and observations
Nevertheless, the prediction error remains quite constant after the fault detection
while a decrease could have been expected because the Wiener parameters should be
better estimated with time. This might be due to the fact that the degradation here is
not really linear. The analysis of the distance between the observation and the healthy
state in Fig. 4 shows that the method considers the unbalance as 1 fault and does
not make the difference between the 3 severities. Other models of drift parameters
should be experimented in order to reach better results.
4 Conclusions
This paper presents an original method for the automated detection of mechanical
unbalance in induction machine using current siganls. The proposed method is based
on the use of recursive PCA, i.e. the PCA is updated every time a new fault is detected,
in order to choose the best subspace and features for the fault detection. This way,
the method only needs to know the healthy behavior of the machine and can be seen
as a novelty detection algorithm. The fault is tracked using a Wiener process applied
on the PCA components in order to predict its evolution in the PCA subspace.
The proposed method has been tested on real experimental data from a 5.5 kW
induction machine for 3 different levels of unbalance. The unbalance was correctly
detected but no specific difference between the 3 unbalance levels is found. The
features chosen by the PCA should be analysed to better understand this point.
The prediction of the fault evolution seems correct and gets more and more precise
with time. However, a better precision could be expected and different form of drift
76 A. Picot et al.
parameters should be tested in order to increase the robustness of the proposed

method.
References
1. Thomson WT, Culbert I (2017) Motor current signature analysis for induction motors.
Wiley/IEEE Press
2. Henao H, Razik H, Capolino G-A (2005) Analytical approach of the stator current frequency
harmonics computation for detection of induction machine rotor faults. IEEE Trans Ind Appl
41(3):801–807
3. Trajin B, Chabert M, Régnier J, Faucher J (2009) Hilbert versus con- cordia transform for three-
phase machine stator current time-frequency monitoring. Mech Syst Sig Process 23(8):2648–
2657
4. Picot A, Obeid Z, Régnier J, Poignant S, Darnis O, Maussion P (2014) Statistic-based spectral
indicator for bearing fault detection in permanent-magnet synchronous machines using the
stator current. Mech Syst Signal Process 46(2):424–441
5. Delgado M, Cirrincione G, Espinosa AG, Ortega JA, Henao H (2013) Dedicated hierarchy
of neural networks applied to bearings degradation assessment. In: 9th IEEE international
symposium on diagnostics for electric machines, power electronics and drives (SDEMPED),
pp 544–551
6. D’Angelo MF, Palhares RM, Cosme LB, Aguiar LA, Fonseca FS, Caminhas WM (2014) Fault
detection in dynamic systems by a fuzzy/Bayesian network formulation. Appl Soft Comput
21:647–653
7. Rauber TW, de Assis Boldt F, Varejão FM (2015) Heterogeneous feature models and feature
selection applied to bearing fault diagnosis. IEEE Trans Ind Electron 62(1):637–646
8. Carino JA, Delgado-Prieto M, Zurita D, Millan, M, Redondo JAO, Romero Troncoso R (2016)
Enhanced industrial machinery condition monitoring methodology based on novelty detection
and multi-modal analysis. IEEE Access 4:7594–7604
9. Picot A, Régnier J, Maussion P (2019) Mechanical faults detection in induction machine
using recursive PCA with weighted distance. In: IEEE international conference on industrial
technology (ICIT), pp 1–6
10. Si X-S, Wang W, Hu C-H, Zhou D-H (2011) Remaining useful life estimation—a review on
the statistical data driven approaches. Eur J Oper Res 213:1–14
11. Si X-S, Wang W, Hu C-H, Chen M-Y, Zhou D-H (2013) A Wiener-process-based degradation
model with a recursive filter algorithm for remaining useful life estimation. Mech Syst Sig
Process 35:219–237
12. Zhai Q, Ye Z-S (2017) RUL prediction of deteriorating products using an adaptive wiener
process model. IEEE Trans Ind Inform 13(6):2911–2921
13. Abdi H, Williams LJ (2010) Principal component analysis. Wiley Interdisc Rev: Comput Stat
2(4):433–459
Automatic Fault Diagnosis in Wind
Turbine Applications
Abstract A wind turbine may have any—out of hundreds—types of bearing on its

generator or other components. The bearing type is often unknown until a possible
defect occurs or until a diagnostic engineer performs a diagnosis. We identify faults
on different wind turbine components by matching their fault frequencies with
frequency harmonics and sidebands found in the spectrum analysis. It is crucial
to know the fault frequencies, such as ball pass inner and outer race defect frequen-
cies, before the diagnosis. Lack of expertise and experience in identifying bearing
type and frequency may result in an incorrect diagnosis. Moreover, not knowing the
bearing frequency leads to the inability to track bearing fault development contin-
uously. In this paper, we propose automatic diagnosis, a novel method to identify
prominent peaks above carpet level in power spectrum analysis, and then label them
for potential faults. Prominent peaks are noticeable spectral peaks above the noise
floor, and we calculate the noise floor using median filtering. Once we extract the
peaks, we implement priority labeling. The fittest bearing frequency is intelligently
estimated using approximate peak matching and envelope of bandpass segment. We
use the estimated frequency to identify, quantify, and track early and late bearing
faults without knowing the bearing type. Using data from wind turbines monitored
by Brüel & Kjær Vibro, we show that automatic diagnosis accurately detects faults,
such as ball pass outer race fault, ball spin frequency fault, generator imbalance, and
gear fault. We expect automatic diagnosis to reduce manual identification of fault
frequencies. Implementing automatic diagnosis could significantly reduce the cost
of simultaneously monitoring the condition of hundreds, or even thousands, of wind
turbines.
Keywords Power spectrum analysis · Wind turbine · Fault frequency ·

Sidebands · Harmonics · Automatic diagnosis
D. Saputra (B) · K. Marhadi

Brüel & Kjær Vibro, Skodsborgvej 307B, 2850 Nærum, Denmark
K. Marhadi

https://doi.org/10.1007/978-981-15-9199-0_8
78 D. Saputra and K. Marhadi
1 Introduction
A condition monitoring system (CMS) is vital for effective maintenance and ensuring
wind turbines’ maximum uptime. It is even crucial for offshore wind turbines. The
implementation of CMS in the wind turbine industry has been presented in [1]. Most
CMS for wind turbines is based on vibration analysis [2–4] presented a case study of
using vibration data to monitor fault progression in the generator bearing of a wind
turbine in real industrial application.
As described in [4], the primary components monitored in wind turbine CMS are
generator, gearbox, main bearing, and tower. Accelerometers are installed on these
components, and there could be up to ten accelerometers installed in a wind turbine.
The data acquisition unit in a wind turbine collects vibration data continuously from
each sensor. Different vibration measurements, also known as descriptors, monitor
any fault development in a wind turbine. To monitor generator bearings, for example,
several descriptors are used at different frequency ranges. The data acquisition unit
continuously calculates the overall vibration RMS level (ISO RMS [10-1000 Hz]),
the high-frequency bandpass (HFBP [1 k–10 k Hz]), the high-frequency crest factor
(HFCF), and several harmonics or orders of the running speed of the generator (e.g.,
1X, 2X) from each sensor. Different failure modes can be monitored from various
sensors using a different set of descriptors.
[4] showed that each descriptor is trended over time to detect specific failure
modes in standard industrial practice. Alarm or warning of particular failure mode
is triggered when the trend from specific descriptors crosses over predefined thresh-
olds. Diagnostic engineers, subsequently, perform vibration analysis to confirm and
describe the alarms.
Broadband descriptors are prone to generate false alarms. For example, line
frequency and crosstalk among different turbine components can falsely increase
ISO RMS, crest factor, and peak-to-peak. An alarm from a broadband descriptor
such as ISO RMS and crest factor can mean a) one very severe fault, b) multiple
faults but not too severe, or c) no critical faults, for example, stand still marks on
generator bearings. Also, broadband descriptors cannot detect some severe faults.
For example, a generator bearing can have severe outer race defect without any
alarm due to relatively low ISO RMS. Moreover, any broadband descriptors cannot
detect roller defect and cage defects because their peaks are in the same frequency
region as the peaks from the generator running speed harmonics. Furthermore, the
bearing type is often unknown because replacement of generator bearing may be
cheap and frequent, requiring continuous update of the bearing information on the
CMS. Incorrect bearing fault frequencies could lead to misdiagnosis of the turbine
component.
When monitoring hundreds or thousands of turbines, human involvement in
confirming the alarms, identifying faults without alarms, and finding fault frequencies
before performing diagnosis is potentially cumbersome. Thus, it is vital to automate
the whole process as much as possible. This paper proposes a novel method, namely
Automatic Fault Diagnosis in Wind … 79
automatic diagnosis, that calculates and trends narrow band descriptors to detect
faults and tracks unknown bearing frequencies.
2 Automatic Diagnosis
There are four essential strategies in automatic diagnosis: angular resampling, iden-
tification of peaks, priority labeling, and frequency tracking. Angular resampling is
essential in performing accurate machine diagnosis. It overcomes the speed variation
issue and eliminates smearing in the power spectrum, making fault peaks clear and
easy to track. Spectral peaks are identified based on the following criteria:
1. Signals larger than 3 dB relative to the spectrum carpet level are considered as
peaks.
2. The carpet level is determined by performing median filtering.
3. Peaks are always local maxima.
4. For pruning purpose, chosen peaks should be within the first and third quantiles
plus 1.5 interquartile range of all identified peaks.
The chosen peaks are then prioritized and labeled according to the failure modes
to be detected. For instance, on detecting a bearing fault, the shaft’s running speed
harmonics are removed, then the bearing fault peaks such as inner race ball pass
frequency (BPFI) and outer race ball pass frequency (BPFO) are identified. However,
line frequency should always be the first to be removed as it is mainly related to
electrical interference, not related to mechanical faults. Some fault frequencies with
the same priority, such as BPFO and sidebands of tooth meshing frequency (TMF),
cannot be removed as they can coexist at the same frequency and accumulate in
amplitude. Finally, peaks related to a known failure mode are then summed and
trended over time as a descriptor. In other words, one failure mode is described by
one descriptor.
A unique feature in automatic diagnosis is its ability to track unknown bearing
frequencies. When monitoring thousands of wind turbines from different turbine
owners, it is considered fortunate for diagnostic engineers to know which bearing is
installed, especially at the generator side. Frequency tracking is beneficial in these
three diagnosis cases:
1. The diagnosed shaft has only one bearing, and we have a list of possible
bearing types. In that case, the bearing frequency consistently converges to a
number when a bearing fault develops. However, this case is relatively infrequent
as sensors usually crosstalk among each other.
2. The diagnosed shaft has many bearings, and we have a list of possible bearing
types for each bearing position. In that case, the bearing fault peaks are grouped
by bearing position, and the best bearing type is presented for each position.
Bearing types will still be unknown until a fault develops. However, when several
bearings are failing, automatic diagnosis can identify which bearing is more
severe than the others. This advantage solves the crosstalk issue.
3. There is no list of possible bearing types. In that case, the correct bearing
frequencies can be discovered, given an adequately more expansive frequency
search space.
3 Automatic Diagnosis in Real Wind Turbines
Over the years, Brüel & Kjær Vibro has monitored more than 8,000 turbines of
various types from different manufacturers. We collect the descriptor scalars every
half hour and the time waveform data from various sensors at different turbine compo-
nents every two days, or once every a few hours in some turbines. The time wave-
form sampling rate from each sensor is 25.6 kHz with a length of 10.24 seconds.
In this paper, we use these time waveform data to perform automatic diagnosis over
specific periods. We focus on detecting a planetary defect, 3rd stage gear wheel fault,
and generator bearing fault on turbines with one-planetary-and-two-helical (1P2H)
gearbox configurations. We use the terms “frequency” and “order” interchangeably.
We indicate a fault by the increasing trend in automatic diagnosis descriptor value,
and we determine the criteria for automatic detection beforehand.
3.1 Planetary Stage Fault
A planetary stage in a wind turbine gearbox consists of a sun gear, a ring gear, a
planet carrier, and typically three planet gears. The planet gears revolve around a sun
gear while rotating on planet shafts at the same time. Once the planet gear finishes
one revolution, the number of meshing teeth of the sun gear becomes equal to the
number of meshing teeth of the ring gear. Consequently, the ring and the sun defect
frequencies equal the number of planets, times the meshing frequency, divided by
the number of gear teeth, while the planet defect frequency equals to the number of
planets that are failing, times meshing frequency over the number of planet gear teeth.
Sidebands around planetary TMF with the spacing of the planetary defect frequency
indicate the presence of a planetary defect. Considering that the planetary gears move
at the same speed with the carrier shaft, sidebands with the spacing of the carrier
shaft speed around planetary defect sidebands could also present, as described in [5].
The sum of all these sidebands is used as a planetary defect indicator.
In automatic diagnosis, these sidebands are tracked in the frequency order spec-
trum. The time waveforms obtained from a sensor at the intermediate stage are
angular resampled based on the low-speed shaft that connects the planetary stage
gear and the 2nd stage helical gear. Figure 1 shows the order spectrum of a turbine
having a planetary defect. Figure 2 shows all the ring defect peaks. Figure 3 shows
the sum of all these peaks and the sums of all peaks related to sun and planet defects
as descriptors trended over a year. According to the figure, the fault was developing
Fig. 1 The black line is the order spectrum of a turbine with a planetary defect seen from
an intermediate-speed stage sensor with a resolution of 0.078. The small red bar is the third TMF
of the planetary stage. The other planetary TMFs are too small to be peaks
Fig. 2 The blue bars are the ring defect peaks around its TMF. The small red bar is the third TMF
of the planetary stage. The other planetary TMFs are too small to be considered as peaks
since the April 30, 2012. The turbine owner performed maintenance on August 8,
2012, and the trends returned to normal.
3.2 Third Stage Wheel Fault
To detect wheel fault at the 3rd stage gearbox, we trend the sum of sidebands of its
TMF with spacing of the running speed of intermediate-speed shaft. Figure 4 displays
Fig. 3 The trend of the planetary defect descriptors over a year
Fig. 4 The trend of the 3rd stage wheel and pinion faults descriptors over a year
the trend of the descriptor for a turbine having 3rd stage wheel fault. The fault started
to develop in December 2015. The turbine owner performed maintenance on April
28, 2017, and since then, the descriptor trend was back to a low level. Figures 5 and
6 depict the turbine’s order spectrums when the 3rd stage wheel and pinion faults
present, respectively. Figure 6 indicates that a pinion fault could be present. However,
as indicated in Fig. 5, the wheel fault is more severe than the pinion fault because
there are more wheel fault sidebands with high amplitudes. The pinion fault could
be a secondary effect of the wheel fault.
Fig. 5 The red bars are the 3rd stage wheel fault sidebands around its TMF, angular resampled with
the high-speed shaft, seen from high-speed stage sensor with a resolution of 0.004
Fig. 6 The pink bars are the 3rd stage pinion fault sidebands around its TMF
3.3 Generator Bearing Inner and Outer Race Defects
When the bearing fault frequencies, for example, BPFI and BPFO, are unknown,
the most probable frequencies must be determined first. This step requires having
an adequately wide frequency band to search. The predicted bearing fault frequency
can be determined when the frequency trend is relatively consistent for some time.
Descriptor value that indicates bearing type fault in automatic diagnosis is the sum
of harmonics of the predicted frequency and their sidebands.
Fig. 7 The trend of the outer race and inner race defects on the generator bearing over 4.5 years.
Misalignment, outer race defect, and inner race defect are in m/s2 . BPFO and BPFI are in order of
generator running speed
Figure 7 displays the trends of bearing outer and inner race defects in a turbine.
Outer race defect is the sum of amplitudes of BPFO harmonics. Meanwhile, inner
race defect is the sum of amplitudes of BPFI harmonics and their running speed
sidebands. Misalignment is the amplitude of twice of the generator running speed.
BPFO and BPFI are the tracked frequencies of outer and inner race defects, respec-
tively. Generally, the outer race bearing fault is always high before July 2017, which
is confirmed by the consistent trend of BPFO since January 1, 2015. Meanwhile, the
inner race bearing defect starts to develop on May 17, 2016, confirmed by the consis-
tent trend of BPFI from May 17, 2016, to July 2017. We can infer the shift of BPFI
and BPFO in August 2017 as the generator bearing type change. The predicted
BPFO and BPFI frequencies fluctuate after August 2017. However, we can see the
consistent frequency, which indicates no bearing fault or mild bearing fault. After
bearing replacement, the bearing defect trend is still relatively high, which could
indicate the presence of a real bearing defect. Around spring 2018, the inner race
bearing defect trend slightly increases, yet there is no indication of bearing change.
This pattern could be due to a misalignment issue since January 2018, indicated as
the purple line. Once the generator was realigned back by early summer 2018, the
inner race defect trend returned to its previous level.
3.4 Generator Bearing Rolling Element Fault
When the ball spin frequency (BSF) and the fundamental train frequency (FTF)
are unknown, the most probable frequencies must be determined to monitor the
rolling element defect. The predicted BSF and FTF can be determined when both
are relatively consistent for some time.
Figure 8 displays the trends of outer and inner race defects in a turbine. Rolling
element fault (with sidebands) is the sum of amplitudes of BSF harmonics and their
FTF sidebands. BSF is the tracked ball spin frequency. In general, the rolling element
fault descriptor suddenly increased on February 22, 2015, confirmed by the consistent
BSF trend. Figure 9 showed the turbine’s order spectrum when the rolling element
fault was at the maximum level on February 22, 2015.
Fig. 8 The trend of the rolling element fault on the generator bearing over three months. Rolling
element fault is in m/s2 , and BSF is in Hz
Fig. 9 The order spectrum of a turbine with BSF fault on generator bearing, resampled with
generator shaft, seen from generator drive-end sensor
4 Conclusion
Automatic diagnosis represents a failure mode as a single descriptor, which signifi-

cantly simplifies fault detection. Automatic diagnosis can also eliminate manual iden-
tification of fault frequencies when such information is not available. We can fully
automate the whole process using predetermined criteria for fault detection. Auto-
matic diagnosis will help monitor a thousand turbines with minimum human effort,
especially with IoT infrastructure availability. In the end, an automatic diagnosis
makes wind turbine condition monitoring very cost-effective.
References
1. Andersson C, Gutt S, Hastings M (2007) Cost-effective monitoring solution using external

surveillance centre. In: The 2nd world congress on engineering asset management and the 4th
international conference on condition monitoring, Harrogate, UK
2. Tavner P (2012) Offshore wind turbines. The Institute of Engineering and Technology
3. Crabtree C (2011) Condition monitoring techniques for wind turbines, unpublished doctoral
dissertation, Durham University
4. Marhadi K, Hilmisson R (2013) Simple and effective technique for early detection of rolling
element bearing fault: a case study in wind turbine application. In: COMADEM, Helsinki
5. Miao Q, Zhou Q (2015) Planetary gearbox vibration signal characteristics analysis and fault
diagnosis. Shock and Vibration
Blade Monitoring in Turbomachines
Using Strain Measurements
D. Abboud, M. Elbadaoui, and N. Tableau
Abstract Blade rubbing is an undesirable but unavoidable phenomenon in turbo-

machinaries. It decreases the engine performance and can lead to blades or/and
stator damages. Thus, blade monitoring is of high importance not only to ensure
security, but also to optimize the engine efficiency. In particular, the detection and
the identification of the instants of contacts between the blades and the stator are
a valuable information that can be used through a preventive maintenance scheme.
This paper proposes a new detection approach based on strain gauge measurements.
Since the rubbing phenomenon creates a wave shock that deform the structure, these
latter can be efficiently measured by a strain gauge mounted on the stator. However,
the strain signal is generally corrupted with other mechanical and electromagnetic
interferences. To tackle this issue, the authors propose a signal separation technique
based on modeling the blades’ signal as cyclo-non-stationary. Also, the use of the
synchronous fitting is proposed to separate blades’ component from environmental
noise. The proposed technique is validated on real signals captured from a strain
gauge mounted on the stator of a turbomachine.
Keywords Health monitoring · Blades · Turbomachines · Cyclo-non-stationary ·

Source separation · Strain gauges
D. Abboud (B) · M. Elbadaoui

Signal and Information Technology Center, Safran Tech, Rue des Jeunes Bois, 78772 Châteaufort,
Magny-les-Hameaux, France
M. Elbadaoui
LASPI, EA3059, Université de Lyon, UJM-St-Etienne, 42023 Saint-Etienne, France
N. Tableau
R&T HP Turbine, Safran Aircraft Engines, Rond Point René Ravaud, 77550 Réau,
Moissy-Cramayel, France

https://doi.org/10.1007/978-981-15-9199-0_9
88 D. Abboud et al.
1 Introduction
Turbomachines aims at transforming the kinetic energy of a fluid into mechanical

energy (and vice versa) via a rotary assembly called a rotor. They have different
designs according to their functions: turbines, pumps, compressors, turbochargers,
etc. Nevertheless, most of them have a common architecture consisting of a rotor (a
wheel mounted on a shaft) which is connected to a fixed housing (the stator) via a
bearing. To establish a relative rotational movement between the stator and the rotor,
the rotor is necessarily spaced from the stator and this spacing is commonly called
the blade clearance of the turbomachine. Blade clearance is an important param-
eter of design and operation of a turbomachine. Its minimization prevents exces-
sive amounts of fluid to bypass the rotating row of blades and, therefore, improves
the general efficiency of the turbomachine. However, this clearance is subjected to
fluctuations due to thermal expansions and mechanical phenomena dependent on
the operating regime of the turbomachine as well as other environmental conditions
(e.g., temperature) [1]. Therefore, minimal clearance often induces contacts between
the blade tip and the stator that leads to blade’s rubbing. Though being inevitable,
the rubbing phenomenon decreases the engine performance and can lead to blades
or/and stator damages [2]. Thus, detecting and identifying the instants of contacts
between the blades and the stator are precious and offers at least two advantages.
The first one is to get information about the existing wear of the turbomachine which
can help to make a preventive maintenance scheme. The second one is to provide
the clearance control system with the contacts information. In practice, contact can
be detected using stator vibrations, acoustic signals, and/or dynamic pressure [3, 4].
In this paper, it is proposed to measure the deformation (or the displacement) using
strain gauges mounted on the stator. As for any sensor, it was found that the signal
is highly corrupted by other electromagnetic interference and background noise. A
proper signal processing techniques are thus needed to reduce environmental noise
so that the signals associated with the contact can be accurately distinguished. In
this paper, a proper signal processing method is proposed to extract the mechanical
component produced by the blades’ rub from noisy perturbations. In this regard,
a model for the strain signal produced by the blade rubbing phenomenon within
the “cyclo-non-stationary” framework is proposed. This framework was originally
proposed to model vibration signals captured under nonstationary regimes [5]. In this
paper, it will be shown that this framework is large enough to enfold strain signals
resulting from the blade rubbing phenomenon. This framework offers a straight-
forward way to extract blades’s contribution from environmental noise through the
synchronous fitting [6, 7]. The rest of the paper is organized as follow: Sect. 2
describes the benchmark used to demonstrate the proposed approach and formulates
the problem. In Sect. 3, it is proved how the strain signal resulting from blades’
rubbing is cyclo-non-stationary and the synchronous fitting previously proposed in
Ref. [6]. Reference [7] is reviewed and proposed to separate the blades contributions
from the noise. Section 4 exposes and discusses the results. The paper is sealed with
a conclusion in Sect. 5.
Blade Monitoring in Turbomachines Using Strain Measurements 89
rotor
blades
Stator section blade
Strain gauges
Stator
integration
tool
Sliding table
(c)
(b)
(a)
Fig. 1 Illustration of the rubbing test bench (a), focus on the integrated stator section (b), and the
strain gauges bonded on the stator section (c)
2 Problem Statement and Formulation
This section describes the experiment setup used to validate the blade rub detection
approach proposed in this paper. A mathematical model of the strain gauge signal is
then proposed.
2.1 Experiment Setup
To test blade rubbing phenomenon and its consequences on the stator, a specific
rubbing test bed has been used (Fig. 1a). The experimental setup consists in a
rotor combining three identical metallic blades, and a fully integrated stator section
mounted on a sliding table (Fig. 1b). Several strain gauges have been bonded on the
stator parts (Fig. 1c) to be able to measure strains values near the contact location
during tests. Test has been conducted using a rotor speed of 15,000 rpm. The pene-
trating speed (relative approach between stator and rotor) has been fixed between 0.01
and 0.15 mm/s. Maximum penetrating depth of 0.4 mm has been reached during tests.
To be able to measure properly dynamic phenomena, the strain sampling frequency
has been fixed to 240 kHz during tests.
2.2 Problem Modeling
When a blade tip rubs with the stator, a sudden variation in the stress occurs at the
interface owing to the transfer of the kinetic energy into elastic energy. Often, the
90 D. Abboud et al.
shock waves propagate into and excites the stator structure which responds according
to its normal modes. While the occurrence of these forces are cyclic due to the rotating
nature of the rotor, its magnitude is likely to change from cycle-to-cycle as the latter
can depend on the depth of the rubbing and other operational conditions.
Based on this, considering a rotor with R blades, it is proposed to approximate
the contribution of the rth blade in the measured strain signal as follow:
⎛ ⎞
Q
∀n ∈ {1, . . . , L} dr [n] = ⎝ aqr δ[n − q.N − τr ]⎠ ⊗ h r [n]
q=1

Q
= aqr h r [n − q.N − τr ] (1)
q=1
where
n stands for the discrete angle variable,
N is the number of points for which the rotor turns one cycles,
L is the whole signal length,
Q = L/N denotes the number of cycles,
⊗ is the operator of numerical convolution,
δ[n] is the Kronecker function,
h r [n] is the impulse response of the stator with the r th blade,
aqr is the magnitude of the contact force (the excitation) generated by the r th
blade at the qth cycle (note that aqr = 0 in the absence of rubbing) and
τr = τ0 + (r − 0.5)/R is the angular shift expressed in number of samples and
τ0 the angular shift of the first blade.
It is then assumed that the correlation length of the impulse response of the
structure is smaller than the cycle, Eq. (1) become:
∀n ∈ {1, . . . , L} dr [n] ≈ ar [n]cr [n] (2)

Q
where c p [n] = q=1 h r [n − q.N − τr ] is a periodic function of period N and
ar [n] is a smooth function such as ar [n − q.N − τr ] = aqr . The measured signal
is obtained by summing up the contributions of all the blades and an environmental
noise denoted by W [n]:
∀n ∈ {1, . . . , L} X [n] = D[n] + W [n] (3)
where

R
∀n ∈ {1, . . . , L} D[n] = dr [n] (4)
r =1
3 Proposed Solution
This section shows the cyclo-non-stationary nature of the blades’ signal and reviews
the synchronous fitting technique.
3.1 Cyclo-Non-stationary Modeling of the Bades’ Signal
Let’s consider the smoothness of the modulation functions, a1 [n] . . . a R [n], then by
applying the Weirstrass theorem, one may rewrite dr [n] as a Pth polynomial with
periodic coefficients:

P
∀r ∈ {1, . . . , R}, ∀n ∈ {1, . . . , L} dr [n] = c̃r p [n]n p (5)
p=1
where c̃r [n] = e pr cr [n] and e pr are the Pth polynomial coefficients such as: ar [n] =
P p
p=1 e pr n . Then, the signal of interest, D[n], can be also expressed through a Pth
polynomial with periodic coefficients:

P
∀n ∈ {1, . . . , L} D[n] = C p [n]n p (6)
p=1

where C p [n] = rR=1 c̃r p [n] is a periodic function of period N (because it is a sum of
periodic function with the same period). This is typically a deterministic cyclo-non-
stationary signal. The next subsection addresses the extraction of this component
from asynchronous noise.
3.2 The Synchronous Fitting
The synchronous fitting is a cyclo-non-stationary method that estimates deterministic

cyclo-non-stationary components with a single fundamental period when this latter
is corrupted by other asynchronous components. It can be seen as a generalization
of the synchronous average, a common cyclostationary tool [8], to the cyclo-non-
stationary case. Hereafter, this method is applied to estimate the signal of interest
D[n]. Let’s first denote by n = (n − 1)/N + 1 the angular position in the period N
(a/b stands for the remainder of the division of a by b). Because C p [n] is periodic
with a period N , one has C p [n] = C p [n + (q − 1)N ] for any integer q = 1, . . . , Q
(Q is the number of cycles). One can rewrite Eq. (6) as:
92 D. Abboud et al.

P
∀q ∈ {1, . . . , Q}∀n ∈ {1, . . . , N } D[n + (q − 1)N ] = C p [n](n + (q − 1)N ) p
p=0
(7)
p p
Applying the binomial theorem (n + (q − 1).N ) p = i=0 li N i (n − N ) p−i q i ,
p
where li are the binomial coefficients, one can deduce that the sample associated
with the position n, sq [n] = D[n + (q − 1).N ] for all integer q ∈ {1, . . . , Q},
defines also a Pth order polynomial with constant coefficients, viz

P
∀q ∈ {1, . . . , Q}∀n ∈ {1, . . . , N } sq [n] = b p [n]q p (8)
p=0
j
where b p [n] = N p Pj= p C p (n − N ) j− p c j [n]. It is important to note that b p [n] is
parametrized by n. An estimation of the predictible part is to find the curve s[n] =
T
s1 [n], . . . , s Q [n] for all n ∈ {1, . . . , N } that reduces the noise to find an estimation
of b[n] = [b1 [n], . . . , b P [n]]T for each n ∈ {1, . . . , N } to a given order P of the
polynomial order. This can be made by minimizing the mean square error, i.e.
⎛ ⎞

Q
∀n ∈ {1, . . . , N } b[n] = argmin⎝ W [n + (q − 1)N ]2 ⎠
q=1
⎛ ⎞

Q
2
= argmin⎝ D[n + (q − 1)N ] − xq [n] ⎠
q=1
⎛⎛ ⎞2 ⎞

Q
P
= argmin⎝ ⎝ b p [n]q p − xq [n]⎠ ⎠ (9)
q=1 p=0
where xq [n] = X [n + (q − 1)N ]. The matrix with dimension Q × (P + 1) is

defined such as q, p = q p−1 (with q ∈ {1, . . . , Q} and p ∈ {1, . . . , P + 1}, and
T
x[n] = x1 [n], . . . , x Q [n] . The above equation can be rewritten as:

∀n ∈ {1, . . . , N } b[n] = argminb[n] − x[n]2 (10)
and its solution writes:

−1
∀n ∈ {1, . . . , N } b[n] = T T x[n] (11)
Now the coefficients are calculated, the signal of interest located at n can be found
as:

−1
∀n ∈ {1, . . . , N } s[n] = b[n] = T T x[n] (12)
Last, the deterministic component can be estimated as follows:

∀n ∈ {1, . . . , L} D [n] = sq [n] avec n = (n − 1)/N + 1 et q = 1 + (n − n)/N

(13)
4 Results
This section is dedicated to validating the proposed approach on the signal obtained
from the experiment described in Sect. 2.1. The time signal is displayed in Fig. 2
(top plot) together with its spectrogram. The spectrogram indicates the presence of
strong electromagnatic interferences that mask the mechanical component generated
by the blades; these components appear as horizontal lines in the spectrogram (being
typically harmonics with fixed frequencies asynchronous to the shaft rotation). The
rubbing occurs after 18 s and stay almost two seconds during which the depth of
the rubbing changes. The occurrence of this phenomenon is seen in the spectrogram
and manifests through a wideband waveform that extends over a large frequency
band. As expected, the excitation is angle-periodic with respect to the blade rotation
Fig. 2 Raw strain gauge signal and its spectrogram

94 D. Abboud et al.
and blade pass frequencies. Nevertheless, it is important to estimate the strain signal
associated with the rubbing itself to clearly identify and quantify the phenomenon.
It is thus proposed to apply the synchronous fitting with respect to the blade rotation
frequency. Though being acquired under a stationary condition, the time signal is
first resampled using a blade position sensor that measures the passage of the blades
to compensate small speed fluctuations. The obtained signal is now synchronized
with the blade positions and its plot is exposed in Fig. 2 (top plot).
The synchronous fitting is then applied using an 8th order polynomial and the
obtained results are shown in Fig. 3. Very interestingly, the estimated signal returns
small values before and after the rubbing phenomenon, but high values during. This
clearly indicates the presence and the strength of the rubbing. The residual signal,
which basically consists on electromagnetic interferences, seems more regular and
has similar time-statistics. This is expected as this latter is independent from the
mechanical condition of the system. It is important to note that the strain signal
contains time-varying DC components which turns out to be useless in the detection
process. This component is separated by simply applying a narrow-band low-pass
filter and then disregarded. Eventually, a close-up of the waveform is displayed
in Fig. 4 during the rubbing, showing three events in each cycle associated with
the response of each blade in the rotor (the rotor consists of three blades). Results
highlight the potentiality of the proposed approach to detect and quantify rubbing in
turbomachines.
Fig. 3 Raw strain gauge signal in the angular domain, the blades’ component estimated by the
synchronous fitting, and the corresponding residue that basically consists of the electrical part
Fig. 4 Close-up in the rubbing zone of the raw strain gauge signal, the blades’ component estimated
using the synchronous fitting, and the corresponding residue that basically consists of the electrical
part
5 Conclusion
In this paper, a blade rub detection technique based on strain gauge measurements
is proposed. A mathematical model for the strain signal is proposed to model the
rub phenomenon. It is shown that the related response is deterministic cyclo-non-
stationary. While the periodicity is clearly justified by the rotating nature of the
blades in the angular domain, the cycle-to-cycle change of the statistics is physically
caused by the presence or the absence of the rub on a hand, and its depth on the
other hand. This finding has offered a straightforward way to extract this component
from other environmental noise. Accordingly, the synchronous fitting is used to
perform the separation (or denoising) step. An experiment has been conducted on
real turbomachines to validate the technique. Furthermore, filtered signals have been
used to calibrate a dedicated numerical model [9]. The results highlight the efficiency
of the synchronous fitting in extracting the blades’ component and in indicating the
occurrence of blade rub.
References
1. Jia B, Zhang X (2011) An optical fiber blade tip clearance sensor for active clearance control
applications. Procedia Eng 15:984–988
2. Barschdorff D, Korthauer R (1986) Aspects of failure diagnosis on rotating parts of turbomachin-
erys using computer simulation and pattern recognition methods. In: International conference
on condition monitoring, Brighton, United Kingdom, May 21–23 (1986)
96 D. Abboud et al.
3. Abdelrhman AM, Hee LM, Leong MS, Al-Obaidi S (2014) Condition monitoring of blade in
turbomachinery: a review. Adv Mech Eng 2014:10
4. Abdelrhman AM et al (2015) Time frequency analysis for blade rub detection in multistage rotor
system. Appl Mech Mater 773–774:95–99
5. Antoni J, Ducleaux N, Nghiem G, Wang S (2013) Separation of combustion noise in IC engines
under cyclo-non-stationary regime. Mech Syst Signal Process 38(1, SI):223–236
6. Daher Z, Sekko E, Antoni J, Capdessus C, Allam L (2010) Estimation of the synchronous average
under varying rotating speed condition for vibration monitoring. In: Proceedings of ISMA 2010
including USD 2010
7. Abboud D, Marnissi Y, Assoumane A, Elbadaoui M (2019) Synchronous fitting for deterministic
signal extraction in non-stationary regimes: application to helicopter vibrations. In: Proceeding
of the international conference of surveillance, VISHNO and EVA (SURVISHNO 2019), 8–10
July 2019, Lyon, France
8. Antoni J, Bonnardot F, Raad A, El Badaoui M (2004) Cyclostationary modelling of rotating
machine vibration signals. Mech Syst Signal Process 18(6):1285–1314
9. Nyssen F, Tableau N, Lavazec D, Batailly A (2019) Experimental and numerical characterization
of a ceramic matrix composite shroud segment under impact loading. Int J Mech Sci (submitted)
Measuring the Effectiveness
of Descriptors to Indicate Faults in Wind
Turbines
Abstract Over the years, numerous descriptors have been proposed to indicate the
health and various failure modes of wind turbine components. Description and math-
ematical interpretation of descriptors, such as kurtosis, crest factor, peak to peak,
mesh, and residual RMS, have been shown in several works to detect bearing and gear
faults. However, there are only a few works that evaluate the effectiveness of various
descriptors in detecting faults. This paper proposes Descriptor Ranking, a novel
machine learning application that measures descriptors’ effectiveness to indicate
different failure modes accurately. The proposed application uses different feature
selection algorithms to rank the importance of the descriptors. Results show that
Descriptor Ranking can correctly rank the importance of any calculable descriptors.
Also, we evaluate the performance of different feature selection algorithms, such as
random forest regressor and logistic regression. Given a balanced set of vibration data
with apparent faults and no faults, Descriptor Ranking visualizes different descrip-
tors’ effectifeness ranks, including newly proposed descriptors, in detecting different
failure modes. Knowing the effectiveness of a descriptor in indicating specific failure
modes is essential to optimize the condition monitoring strategy of wind turbines or
any other machinery. Having multiple descriptors to identify the same failure mode
can be avoided. Hence, we can reduce false alarms, and faults can be detected more
accurately. With fewer false alarms, monitoring the condition of many wind turbines
is going to be more efficient.
Keywords Wind turbine faults · Descriptor · Machine learning · Feature selection
D. Saputra (B) · K. Marhadi

Brüel & Kjær Vibro, Skodsborgvej 307B, 2850 Nærum, Denmark
K. Marhadi

https://doi.org/10.1007/978-981-15-9199-0_10
1 Introduction
The strategy for monitoring the condition of a wind turbine has evolved over the
last decade. Routine analysis of the vibration signal from sensors near the monitored
components is one popular strategy. We perform vibration analysis by calculating
statistical features, called descriptors, that indicate various failure modes. Different
descriptors have been proposed and mathematically interpreted to detect bearing and
gear faults, such as residual RMS, kurtosis, and peak to peak [1]. These descriptors
are subsequently trended over time to monitor any developing faults.
There have not been many evaluations on the effectiveness of various descrip-
tors known to identify different failure modes. We should avoid having ineffec-
tive descriptors as they may lead to false alarms. Also, an excessive number of
descriptors to be calculated requires unnecessary computational resources. Hence, we
should not implement newlyproposed descriptors without thorough evaluation. This
paper proposes Descriptor Ranking, a novel machine learning application that selects
descriptors that accurately indicate different failure modes. We compare some well-
known descriptors and descriptors calculated based on a method called Automatic
Diagnosis [2]. In a nutshell, Automatic Diagnosis descriptors calculates the sum of
the amplitudes of peaks related to particular failure modes in the order spectrum.
2 Machine Learning Methods
Machine learning offers a brilliant way to select the best descriptor that explains a
failure mode through feature selection. A useful feature selection algorithm intelli-
gently recognizes redundant and irrelevant descriptors. We can thus remove these
descriptors without incurring much loss of supervised learning prediction power.
The initial, most important step in feature selection is data gathering. There should
be plenty and a balanced amount of real vibration data containing mechanical faults
and those with no faults. The biased result may occur if there are not enough data or
more data in one group than the other.
The subsequent crucial step is data cleaning. A short length of time waveform
data can cause erratic behavior in fault detection, primarily when evaluating many
of these time waveform data over a certain period. To illustrate, suppose there are
20 files in the time waveform format, each of which has a duration of 10 seconds.
A specific fault signature appears swiftly in the waveform every 20 seconds. Hence,
10 seconds is not enough window to ensure the presence of the fault. Consequently,
ten waveform files might indicate the presence of a fault and the other ten files indicate
no faults. The machine learning training set should be free from such false-negative
cases.
Another common mistake in data cleaning is misdiagnosis, i.e., when a time
waveform file is labelled as faulty while there is no fault or vice versa. Multiple
faults considered as just one fault is also a case of misdiagnosis. Also, we should not
Measuring the Effectiveness of Descriptors to Indicate … 99
use data without a confirmed diagnosis. In wind turbine operation and maintenance,
it is possible to miss actual fault, even with a physical inspection. The inability to
physically identify actual fault could give incorrect feedback to the given diagnosis.
In such cases, we cannot use the data either.
The next step is to create a training set by calculating the descriptor values and
assigning classes to all vibration data into faulty (1) and no-fault (0). We can then feed
a well-generated training set into a machine-learning algorithm to obtain a predictive
model that describes the relationship between descriptors and the fault status, i.e.,
faulty or no-fault. The model is subsequently used to rank the effectiveness of a
descriptor in assigning fault or no-fault condition.
There are several feature selection algorithms that we can use, as listed in Table 1.
In this paper, we use the Random Forest Regressor algorithm [3] to rank descrip-
tors’ importance. The algorithm performs very well compared to the other methods
in our examples because Random Forest uses bagging and feature randomness
when building individual decision trees to create a de-correlated forest of trees.
The combined prediction is more accurate than any of those individual trees. In real
life, the wisdom of the crowds is always better than lone decision making. This is
the philosophy behind the Random Forest algorithm.
Random Forest algorithm takes vote from many decision trees, each built over a
random subset of the training set and a random subset of features. This guarantees
that the individual trees are de-correlated, thus minimizing overfitting. Each tree
divides the training set into two clusters, each of which contains sub-datasets similar
among themselves and different from the other cluster. Therefore, each descriptor’s
importance is calculated as the purity of each cluster, i.e., a measure of how often
a randomly chosen descriptor will predict incorrectly. Impurity is measured using
the Gini index in Random Forest Classifier and mean squared error (MSE) in Random
Forest Regressor. The idea behind the Random Forest algorithm is to calculate how
much the existence of each descriptor decreases the impurity. The more a descriptor
decreases the impurity, the more critical the descriptor is. The impurity decrease
from each descriptor is subsequently averaged across all trees to determine the
descriptor’s final importance, represented as the fraction of all descriptors’ total
importance.
3 Data
Over the years, Brüel & Kjær Vibro has monitored the condition of more than 8,000
turbines of various types from different manufacturers. Every half hour, we collect
various descriptors from these turbines. We also collect time waveform data from
various sensors located at different turbine components every two days or every
few hours in some turbines. The time waveform sampling rate from each sensor
is 25.6 kHz with a length of 10.24 seconds. We used these time waveform data to
generate descriptors and to evaluate their importance in detecting faults. This paper
focuses on finding the best descriptors that detect a planetary defect and a second
Table 1 List of feature selection algorithms that are compared in this paper. The evaluation is
omitted for brevity
Algorithm Detail
Pearson correlation coefficient (PCC) Linear correlation between a descriptor and the
fault status, regardless of correlation among
descriptors.
Distance correlation This algorithm addresses PCC’s problem of not
measuring variable independence.
F-regression Univariate linear regression that uses F-test to
compare the significance between a descriptor
and the fault status, regardless of correlation
among descriptors.
Random forest classifier Ensemble learning algorithm that constructs
decision trees on various subsamples and uses
modes to improve predictive accuracy and to
control overfitting. Out-of-bag samples are used
to estimate the R2 on unseen data.
Random forest regression Similar to above, but it uses averaging instead of
modes.
Extra trees classifier Similar to Random Forest, but instead of
looking for the most discriminative thresholds,
thresholds are randomly drawn for each
candidate feature, and the best threshold is
picked as the splitting rule. This reduces the
variance of the model but slightly increase bias.
Extra trees regressor Similar to above, but it uses averaging instead of
modes.
Random forest regressor with Drop columns We run a random forest regressor, then remove
method one descriptor, retrain the reduced training set,
and do that for other descriptors. This algorithm
calculates whether the other descriptors can
replace that descriptor.
Random forest classifier with Drop columns Similar to above but using random forest
method classifier.
Logistic regression A statistical model that uses a logistic function to
model a binary class training set. SAGA solver
is used as it handles L1 regularization very well
Maximal Information Correlation Linear and non-linear association between a
descriptor and the fault status, regardless of
correlation among descriptors
Ridge regression Regression algorithm using linear least squares
and L2 regularization
Lasso Regression Regression algorithm using linear least squares
and L1 regularization
Randomized lasso regression Computing Lasso on each resampling of the
training set
stage wheel fault on turbines with one-planetary-and-two-helical (1P2H) gearbox

configurations. The dataset used in this work is obtained from customers of Brüel
& Kjær Vibro and is subject to a confidentiality agreement between Brüel & Kjær
Vibro and its customers.
4 Results and Discussion
4.1 Planetary Stage Fault
The turbine owner suspected a planetary defect around April 13, 2016, see Fig. 1.
The owner finally replaced the gearbox on July 31, 2017, and there were no faults
detected since then. There are 173 data points without faults and 170 data points
containing faults during these approximately two years. This provides an excellent,
balanced training set for Descriptor.
The broadband descriptors used in this comparison are Envelope Condition
Unit (ECU), High-Frequency Band Pass (HFBP), ISOA, MESH, Residual Value
(RV), Tooth Mesh Frequencie (TMF), High-Frequency Crest Factor (HFCF), High-
Frequency Peak (HFPK), Order Magnitude (MA), and other bandpass measurements
in selected ranges. Some of these descriptors were explained in [1], and others are
Brüel & Kjær Vibro proprietary descriptors. Most of the descriptors are measured
in acceleration, but some are measured in velocity. High frequency in this work
is typically defined as above 1000 Hz. ECU is the vibration’s effective value after
Fig. 1 The trend of the planetary defect descriptors over four years using Automatic Diag-
nosis. RingDF_AutomaticDiagnosis is an Automatic Diagnosis descriptor, while RV_Planetary
and RV_IntermediateStage are the residual value after the spectrum is resampled using a low-speed
shaft and an intermediate-speed shaft, respectively
Fig. 2 The importance of various descriptors to detect a planetary defect, calculated using Random
Forest regressor algorithm
enveloping the high-frequency range into frequency ranges of the bearing defects.
The effective-value of the acceleration measurement according to the vibration stan-
dard ISO 10816 is defined as ISOA. MESH is the effective-value of the vibration
between the first and the third gear mesh frequencies. The effective-value between
the gear tooth mesh frequencies is defined as RV. TMF is the tooth mesh frequency.
In this work, TMFs are measured up to the fifth TMF (5TMF). The first- and second-
order magnitudes (1MA and 2MA) of shafts in the gearbox and generator are also
measured.
Meanwhile, the Automatic Diagnosis descriptors for comparison are the sum of
peaks of the sun defect (SunDF), planet defect (PlaDF), and ring defect (RingDF)
sidebands, resampled with carrier shaft (Carr) and low speed shaft (Lss).
Result from Descriptor Ranking, as shown in Fig. 2, shows that Automatic Diag-
nosis descriptors outperform the broadband descriptors. Moreover, AutomaticDiag-
nosis descriptors predict that the planetary fault is a ring defect, neither a sun defect
nor a planet defect.
4.2 Second Stage Wheel Fault
A second stage wheel fault was suspected on October 1, 2017, as seen in Fig. 3.
The turbine owner finally replaced the gearbox on September 24, 2018. However, the
pinion fault started to develop after the maintenance. As there are 50 more data points
during the fault period than those during the no-fault period, we need to add around
50 no-fault data points to our training set to balance our training set. We should
Fig. 3 The trend of the 2nd stage wheel fault descriptors over 4.5 years calculated using Automatic
Diagnosis. RV is multiplied by five for visualization purpose as the original RV is too small
not take data points during the period when the wheel might have started failing.
Therefore, the first 50 points in 2015, i.e., when the fault is not seen, were taken.
This provides a balanced dataset for Descriptor Ranking by having 156 no-fault data
points and 154 fault data points.
Figure 4 presents the Descriptor Ranking results. Seventeen descriptors were
ranked based on calculated importance. Some broadband descriptors from Fig. 2 were
used here, along with additional ones, such as low-frequency kurtosis (LFKurt), low-
frequency peak-to-peak (LFPK), the effective-value of certain low-frequency band-
pass according to the vibration standard ISO 10816-21 (ISOA21), high-frequency
kurtosis (HFKurt), and low-frequency crest factor (LFCF). In general, Automatic
Diagnosis descriptor 2WheelFault (2nd Stage Wheel Fault) is much more effective
in detecting the fault than the other 16 descriptors. This is because the Automatic
Diagnosis descriptor is a narrow band descriptor, as explained in [1], and the fault
in the training set is not a 2PinionFault (2nd stage pinion fault). There is a devel-
oping second stage pinion fault after the maintenance, as seen in Fig. 3. Therefore
the residual value (RV) increased after maintenance. RV is normally considered to
be the best broadband descriptors to detect this fault [1, 4]. However, RV cannot
differentiate between wheel fault and pinion fault because it just represents energy
between tooth mesh frequencies.
5 Conclusion
The effectiveness of descriptors can be determined using Descriptor Ranking. Auto-

matic Diagnosis descriptors in this study outperform other descriptors. We can
evaluate any new proposed descriptors using Descriptor Ranking. Having effective
Fig. 4 The importance of various descriptors to 2nd stage wheel fault, calculated using Random
Forest regressor algorithm
descriptors in condition monitoring could help reduce false alarms and increase the
accuracy of faults detection.
References
1. Zhu J et al (2014) Survey of condition indicators for condition monitoring systems. In: Annual
conference on prognostics and health management society, vol 5
2. Saputra D, Marhadi K (2020) On Automatic Fault Diagnosis in Wind Turbine Condition
Monitoring. In: PHM Society European Conference. Vol. 5. No. 1. 2020
3. Breiman L (2001) Random forests. Mach Learn 45(1):5–32
4. Skrimpas GA et al (2017) Residual signal feature extraction for gearbox planetary stage fault
detection. Wind Energy 28(8):1389–1404
Gearbox Condition Monitoring Using
Sparse Filtering and Parameterized
Time–Frequency Analysis
Shiyang Wang, Zhen Liu, and Qingbo He
Abstract It is challenging to monitor the condition of a gearbox being operating

under variable conditions. Considering that numerous variable conditions have the
non-stationarity nature, a time–frequency analysis (TFA) approach is very suited
to analyze the frequency components of interest. As an emerging TFA method,
parameterized TFA utilizes further parameters to parameterize kernel functions, so
a complex signal can be characterized more accurately. However, it is an issue to
estimate the instantaneous frequency with the interference of background noise.
This paper presents a method to solve this problem by using the sparse filtering
method, as the sparse filtering is capable of enhancing a particular desired feature to
achieve the desired effect. With the sparse filtering, the parameterized TFA is used
to achieve signal decomposition and clarify the signal spectrum structure. Finally,
these interferential frequencies are extracted and removed, and the processed signals
are used to monitor the gearbox. The availability of this method was verified by
simulated and experimental signal analysis in gearbox condition monitoring.
Keywords Gearbox condition monitoring · Time–frequency analysis · Sparse

filtering · Signal decomposition
1 Introduction
In modern industry, gearboxes are one of the vital components of rotating machinery
[1]. The gearbox includes components such as the shaft, gears and bearings. The
proportion of gear parts damaged in the gearbox is 60%, and the failure of the
gear itself has the greatest impact on the system [2]. Therefore, the fault must be
S. Wang
Department of Precision Machinery and Precision Instrumentation, University of Science and
Technology of China, Hefei, Anhui 230026, People’s Republic of China
Z. Liu · Q. He (B)
State Key Laboratory of Mechanical System and Vibration, Shanghai Jiao Tong University,
Shanghai 200240, People’s Republic of China
e-mail: qbhe@sjtu.edu.cn

https://doi.org/10.1007/978-981-15-9199-0_11
106 S. Wang et al.
diagnosed and eliminated before the accident occurs. If the gear in the gear term
fails, the amplitude of the meshing frequency and its harmonic components will rise,
and a modulation sideband will appear around these frequencies [3]. These shocking
signal components are too weak, and identifying these signal components becomes
a very meaningful job.
In practice, the gearbox is generally operated under variable conditions. Recently,
with adopting a matched kernel function for a pre-defined signal model, parame-
terized TFA can be used. Yang et al. [4] proposed generalized parametric time–
frequency analysis method, which extends the kernel form of parametric time–
frequency analysis from multiple to arbitrary. This method is useful for analyzing
strong time-varying signals. Kerber et al. [5] proposed chirplet atom with polynomial
local frequency delay to analyze the corresponding non-stationary signal. Therefore,
the parameterized TFA is used to achieve signal decomposition and clarify the signal
spectrum structure.
However, it is an issue to identify the specified varying frequency from the inter-
ference of background noise by parameterized TFA. This paper presents a method
to solve this problem by using the sparse filtering method. Sparse filtering is one of
the most effective algorithms among the unsupervised learning algorithms. Zennaro
[6] emphasized that sparse filtering is an optimized algorithm which contains only
one hyperparameter. Lei et al. [7] used sparse filtering to learn impulsive features
representation. Jia et al. [8] optimized the cost function, like lp/lq norm, to enhance
weak impulsive signal. The recent research proved sparse filtering can enhance a
particular desired feature, which achieves the desired effect in gearbox condition
monitoring.
In the proposed study, the sparse filtering is firstly used to enhance the particular
desired feature. After that, the parameterized TFA is used to represent the desired
signal components. This is beneficial to remove the interferential frequencies and
extract the useful components for gearbox condition monitoring. In the following,
Sect. 2 presents the theoretical background that is related to this study. Sections 3
and 4 represent the simulation and experimental results, respectively. Finally, Sect. 5
gives some concluding remarks.
2 Theory Background
2.1 Sparse Filtering
Different from usual learning algorithms that often model the training data distri-
bution, the sparse filtering is an unsupervised algorithm. Sparse filtering directly
analyzes the feature distribution of training data under the guidance of so-called
good parameters. The objective function is optimized to involve only one tunable
parameter. Strictly, sparse filtering is just a feature extractor [9]. In the analysis of
Gearbox Condition Monitoring Using Sparse Filtering and Parameterized … 107
its network structure and feature matrix, sparse filtering algorithm can be regarded
as a neural network with a two-layer structure.
For a stochastic input
quantity
x = (x1 x2 . . . xn )T , after being multiplied by the
weight matrix W = wi, j ∈ R m×n
of the network, the vector y = (y1 y2 . . . ym )T is
achieved as
y = Wx (1)
Then an activation function g is added to y, and we can obtain the activated

eigenvector f ∈ R m as
f = g(y) = g(W x) (2)

√
Soft-absolute function: g(z) = ε + z 2 ≈ |z|, where ε takes a small positive

constant. Consider a training set containing ns samples S = x (1) , x (2) , . . . , x (n s ) ,
T
recorded as: f (i) = f 1(i) , f 2(i) , . . . , f m(i) . It represents a linearly operated and
activated feature vector, where:

f (i) = g W x (i) , i = 1, 2, . . . , n s (3)
So the selection of W has a strong correlation with the quality of F. In addition,

F should satisfy the following three criteria [10]: (1) population sparsity (it requires
sparse features per example); (2) lifetime sparsity (it requires sparse features across
examples); and (3) high dispersal (it requires uniform activity distribution). The
objective function is obtained by the following three steps:
(i)
1. Row normalization: change matrix F into F̃ = ( f˜j(i) )m×n s , among f˜j(i) = || f jj ||2
f
2. Column normalization: change the matrix F̃ into F̂ = ( fˆj(i) )m×n s , among fˆj(i) =
f˜j(i)
˜
|| f (i) ||2

ns
3. The sum of the absolute values of the matrix elements: L(W ) = || fˆ(i) ||1 .
i=1
After the above three steps, L(W ) is our objective function, Using the objective
function, the sparse filtering algorithm does it to minimize it, which is:
minimize (L(W )) (4)
Get its minimum point min (W ), which is the optimal weight matrix.
108 S. Wang et al.
2.2 Parameterized TFA
Fourier transform obtains an explicit frequency spectrum in stable conditions.

However, for some special signals whose frequency varies with the time, Fourier
transform is useless [4]. Recently, with adopting a matched kernel function for a
pre-defined signal model, the parameterized TFA is more effective in characterizing
the real time–frequency pattern of non-stationary signal.
As one of the parameterized TFA methods, the chirplet transform is to use a
slanted line with an arbitrary tilt angle on the time–frequency domain to analyze a
linear frequency modulated signal. The essence of chirplet transform is to project
a signal onto a set of line-modulated wavelet functions. The definition of chirplet
transform (5) is:
+∞
Cwt (t, ω; c) = Amp (t) x(τ )gσ∗ (τ − t) exp(− jωτ )dτ (5)
−∞
⎧
⎨ R (τ ) = exp(− jcτ 2 /2); M (τ ) = exp( jctτ ); gσ (t) = √1 exp(−t 2 /4σ )
c c,t
2 πσ
⎩
x(τ ) = x(τ )c (τ )c,t (τ ); Amp (t) = exp[− j (τ c/2 − ωt)]
R M 2
(6)
The parameters in the formula (6) are explained as follows: x(τ ) is signal, cR (τ )
is frequency rotation operator of line frequency modulation parameter, cR (τ ) will
parse the signal x(τ ) in the time–frequency plane by an angle θ and tan(θ ) =
−c · c,t M
(τ ) represents a frequency shifting operator of τ period near t moment.
c,t (τ ) translates the rotated signal x(τ )cR (τ ) in the time–frequency plane with a
M
translation increment of ct. Amp is amplitude. gσ (t) is Gaussian window function.
3 The Proposed Signal Processing Method
Integrating the above theory, we propose a method by combining sparse filtering

and parameterized TFA. As shown in the scheme in Fig. 1, the gearbox signal x(t)
is first inputted into the sparse filtering constructor to find a specific atom. The
filtered signal is hoped to enhance a specified feature to achieve the desired effect.
Based on previously provided spectrum information, the parameterized TFA is used
to achieve signal decomposition and clarify the signal spectrum structure. Finally,
these interferential frequencies are extracted and removed, and the processed signals
are better used to monitor the gearbox.
Fig. 1 Scheme of the presented method
4.1 Signal Structure
For the case where gearbox is running, we used two gears to simulate the condition of
the gearbox. The first signal of gears has a meshing frequency of 770 Hz and a fault
frequency of 70 Hz. The second signal of gears has a meshing frequency of 800 Hz
and a fault frequency of 90 Hz. The two signals are superimposed, plus a white noise
with the signal-to-noise ratio as 0 dB. The spectrum and envelope spectrum of the
constructed signal are shown in Fig. 2. It can be seen that the frequency components
are mixed together.
Fig. 2 Simulated gearbox signal x(t): a x(t) waveform; b spectrum; c envelope spectrum
110 S. Wang et al.
Fig. 3 Results of sparse filtering: a signal, b spectrum, c envelope spectrum of the first signal
(meshing frequency 770 Hz); d signal, e spectrum, f envelope spectrum of the second signal
(meshing frequency 800 Hz)
4.2 Signal Processing by Sparse Filtering
The above signal contains information of multiple sets of gears, and their meshing
frequency will be more similar, and the fault frequency is modulated in the meshing
frequency. The correlation information between two similar meshing frequen-
cies becomes difficult to distinguish. The sparse filtering can separate the signals
according to the selection of the atoms and clarify the characteristics of the signals.
By selecting a specific atom, the sparse filtering can select the corresponding
features to enhance. In Fig. 3, the meshing frequencies of the two sets of signals are
770 and 800 Hz, respectively. The signal is completely separated, and the denoising
effect is well done. So sparse filtering has potential capacity to detect impulsive
components and identify different periodic components.
4.3 Signal Extraction by Parameterized TFA
Parameterized TFA has good applicability for non-stationary working conditions, so

it is used as a feasible method for signal extraction. However, it needs to clarify the
required spectral components. According to the result of two separated signals by
sparse filtering, the results extracted by the parameterized time–frequency analysis
are shown in Fig. 4.
According to the above example, it can be proved that the corresponding feature
signal is enhanced by sparse filtering and use the parameterized time–frequency
analysis to achieve signal decomposition and clarify the signal spectrum structure.
This method can be used to monitor the gearbox.
Fig. 4 Results of parameterized TFA: a signal, b spectrum, c envelope spectrum of the first signal
(meshing frequency 770 Hz); d signal, e spectrum, f envelope spectrum of the second signal
(meshing frequency 800 Hz)
5 Experiments
5.1 Data Description
To prove the validity of the previous method, this study analyzes a set of actual
gearbox signals. The measured signal duration is ten seconds, and the sampling
frequency is 25600 Hz. In this study, we take a stable signal segment of five seconds
for analysis. As shown in Fig. 5, the signal shows some large spectral components,
but the envelope spectrum cannot show valuable information, so we cannot get any
results without other information.
5.2 Signal Processing
Figure 5 shows the processed signal by sparse filtering. We can clearly find 1209 Hz
and its multiplier components in the spectrum and the envelope spectrum. After
confirming with the manufacturer, it is the resonant frequency of the gearbox, which
is the invalid interference frequency. Figure 6a shows the time–frequency diagram
where the dotted line shows the components to be removed. By directly subtracting
the interferential components, other latent frequency components are highlighted in
the time–frequency diagram as shown in Fig. 6b. Then we used the parameterized
TFA again to extract useful signal component. Figure 6c shows an extracted latent
time-varying frequency component around 6700 Hz. The time-varying frequency
provides more information related to the gearbox condition, which is beneficial for
further condition monitoring.
Combining sparse filtering and parameterized TFA, it can be seen that the compo-
nents of the main interference energy can be removed and other latent frequency
112 S. Wang et al.
Fig. 5 Raw signal and result of sparse filtering: a signal, b spectrum, c envelope spectrum of raw
signal; d signal, e spectrum, f envelope spectrum of the processed signal by sparse filtering
Fig. 6 Results of parameterized TFA: a preprocessed time–frequency diagram, b time–frequency

diagram after interference removal, c time-varying frequency after interference removal
components can be highlighted. In a signal with periodic repetitive pulses, this

method laid a good foundation for the subsequent research on related directions,
which provides the possibility for data analysis in noisy components.
6 Conclusion
This paper proposed a method by combining sparse filtering and parameterized TFA
for gearbox condition monitoring. With the specific atom, the capability of sparse
filtering was enhanced to distinguish different signal components, especially peri-
odical and impulsive signal components. But sparse filtering cannot maintain the
amplitude of the signal. In this paper, parameterized TFA is used to achieve signal
decomposition and clarify the signal spectrum structure, which preserves the original
information and can be used for time-varying signals. Results of this study verify
the effectiveness of the combination method, which can estimate the instantaneous
frequency with the interference of background noise, achieve signal decomposi-
tion and clarify the signal spectrum structure. It has laid a good foundation for the
subsequent research on gearbox condition monitoring and fault diagnosis.
Acknowledgements This work was supported by Joint Fund of Equipment Pre-research and
Ministry of Education under Grant 6141A02022141, and the National Program for Support of
Top-Notch Young Professionals.
References
1. He Q, Kong F, Yan R (2007) Subspace-based gearbox condition monitoring by kernel principal

component analysis. Mech Syst Sig Process 21(4):1755–1772
2. Randall RB (2011) Vibration-based condition monitoring: industrial, aerospace and automotive
applications. Wiley
3. Lin J, Zuo MJ (2003) Gearbox fault diagnosis using adaptive wavelet filter. Mech Syst Sig
Process 17(6):1259–1269
4. Yang Y et al (2014) General parameterized time-frequency transform. IEEE Trans Sig Process
62(11):2751–2764
5. Kerber F et al (2010) Attenuation analysis of lamb waves using the chirplet transform.
EURASIP J Adv Sig Process 2010(5):1–6
6. Zennaro FM, Chen K (2017) Towards understanding sparse filtering: a theoretical perspective.
Neural Netw Off J Int Neural Netw Soc 98:154
7. Lei Y et al (2016) An intelligent fault diagnosis method using unsupervised feature learning
towards mechanical big data. IEEE Trans Ind Electron 63:3137–3147
8. Jia X et al (2018) Sparse filtering with the generalized lp/lq norm and its applications to the
condition monitoring of rotating machinery. Mech Syst Sig Process 102:198–213
9. Xiong W, He Q, Ouyang K (2018) Feature-difference sparse filtering for bearing health moni-
toring. In: 2018 IEEE International Instrumentation and Measurement Technology Conference
(I2MTC), Houston, TX, USA, pp 1–5
10. Ngiam J et al (2011) Sparse filtering. In: Advances in neural information processing systems,
pp 1125–1133
Rotating Machinery Fault Diagnosis
with Weighted Variational Manifold
Learning
Quanchang Li, Xiaoxi Ding, Wenbin Huang, and Yimin Shao
Abstract The transient characteristics attributed to partial fault of rotating

machinery are of technical evidences for health diagnosis, which are unavoidably
overwhelmed by much noise disturbance in complex modulation bands. Usually,
only the optimal component with a selected scale is filtrated and then extracted
to be the evidence for fault diagnosis, which may ignore some important features
distributed in the other scales. Focusing on the feature distribution in a whole-band,
this paper proposed a novel weighted variational manifold learning (WVML) for tran-
sient features reinforcement and fault diagnosis in the view of multi-scale analysis.
First, a series of intrinsic mode functions, containing different modulated frequency
information, are automatically obtained via an improved variational mode decom-
position (IVMD) approach, where a signal difference average criterion is introduced
to determine the appropriate decomposition level. Time–frequency manifold (TFM)
learning is referred to then sequentially mine their intrinsic structures in the corre-
sponding time–frequency domain. Therewith, the reinforced signal for each varia-
tional mode can be achieved via phase preserve and a series of inverse transforms
in a suppression of the in-band noise. Combining the reinforced mode signals with
different weight coefficients, the reconstructed waveform can be finally synthesized.
In these manner, the desired features of all scales can be soundly reinforced with
mode manifolds mined and dynamically weighted in a self-learning way. The conse-
quence of proposed WVML scheme is verified through experimental gear defective
signal and result reveals this method yields an excellent accuracy in health diagnosis
of rotating machinery.
Keywords Variational mode decomposition · Manifold learning · Multi-scale

analysis · Rotating machinery · Fault diagnosis
Q. Li · X. Ding (B) · W. Huang · Y. Shao

State Key Laboratory of Mechanical Transmission, Chongqing University, No.174 Shapingba
Street, Shapingba District, Chongqing 400044, China
e-mail: dxxu@cqu.edu.cn

https://doi.org/10.1007/978-981-15-9199-0_12
116 Q. Li et al.
1 Introduction
Rotating machinery has been acting a crucial role in industrial manufacturing, the
partial faults of which maybe develop persistently then attract serious accidents
with economic losses [1, 2]. As a result, it is indispensable to diagnose the health
of key parts of rotating machinery such as bearings. Thanks to the periodic tran-
sient impulses in the acquired signals always give back the pivotal physical knowl-
edge concerning fault bearing dynamics [3], vibration signal is the most widespread
technique which is used in a field of rotating machinery health diagnosis.
There are varies of signal processing approaches focusing on defective vibra-
tion transient features reinforcement and noise repression. Band-pass filtering [4] is
the firstly used method to eliminate noise through removing frequency components
outside the given frequency range, whereas it cannot distinguish the fault-related
components and in-band noise in the same frequency band. As a method to pick the
selected frequency band with particular range, wavelet transform (WT) [5] shows
good characteristics of scale variations in frequency decomposition but the setting
of mother basis and redundancy of wavelet make it have certain application restric-
tions. Empirical mode decomposition (EMD) [6] achieves purposes from the time
scale of data sequence, whose effectiveness depends largely on the method of finding
extreme points, the interpolation of the extreme points to the carrier envelope and the
stopping criterion of the application. In addition, the robustness of EMD is limited
by the deficiency of system of mathematical theory. On the foundation of EMD, a
new method called variational mode decomposition (VMD) [7], is extensively intro-
duced in fault identification and diagnosis. Liu et al. [8] applied VMD into CNN to
extract planetary gears features and required 100% total recognition. Bagheri et al. [9]
proposed a VMD-based modal identification algorithm to accomplish engineering
structures mode identification and demonstrated the efficacy of it. However, VMD
itself still has some following issues. First, the number of decomposition level in
VMD needs to be given in advance while it is difficult to appropriately determine by
prior acknowledgement. Second, the VMD retains a denoising effect as the Wiener
filter to some extent, but it is powerless to remove the in-band noise of the signal.
Recently, a new data-driven approach called time–frequency manifold (TFM)
learning proposed by He et al. [10] has achieved excellent features extraction by
the way of mining the intrinsic structures contained in the nature of different faults.
Different from normal manifold learning, TFM learning can take non-stationary prop-
erty hidden in high-dimensional time–frequency distribution into consideration and
appears more effective results. Many TFM-based methods such as resonance-based
TFM feature extraction [11], TFM-based fault classification [12], and TFM sparse
reconstruction [13] have been studied and also acquired good results. However, on the
one hand, with a nonlinear manifold learning, the immense operands vastly restricts
farther applications for multi-modulation feature extraction in practical operations.
On the other hand, TFM learning cannot satisfy the request of full-band structure
mining owing to its amplitude loss.
Rotating Machinery Fault Diagnosis with Weighted Variational Manifold Learning 117
For a potential research value in realistic rotating machinery fault diagnosis,

a novel weighted variational manifold learning (WVML) method is proposed to
accomplish multi-scale transient features extraction of complex modulated signal in
a data-driven way. Thanks to the excellent performance of VMD in modulated signal
decomposition, band-limited intrinsic mode functions (IMFs) reflecting different
modulation information are firstly decomposed using an improved VMD method, so
as to ensure each band not be affected by other adjacent feature distributions. Then,
the manifold structures of decomposed IMFs are able to be mined and enhanced by
lifting each IMF to time–frequency plane for TFM learning. Among them, short-time
Fourier transform (STFT) and its inverse algorithm are applied to process signal into
time–frequency domain or return to time domain. The final reconstructed signal is
acquired via synthesizing the learned manifold signals with different weights dynam-
ically adjust the proportion of each IMF to emphasize the fault-relevant frequency
components. These merits will make it possible for both VMD and TFM tech-
nique to efficiently process in the full scale, acquire the comprehensive modulation
information, and provide a more accurate fault diagnosis result.
2 Weighted Variational Manifold Learning
2.1 Variational Mode Decomposition Based on SDA
As a novel signal decomposition approach, VMD decomposes a given signal y(t)

into K band-limited IMFs through unceasingly updating center frequencies ωk and
corresponding bandwidths to satisfy the minimum sum of estimated mode band-
widths [14]. Different from existing signal decomposition methods such as EMD or
WT, VMD redefined the physical significance of IMFs as AM-FM components with
their corresponding narrow-band attribute. With such an assumption, the different
sub-bands modulated by meshing frequency, resonant frequency, and so on in rotating
machinery can be separated. To obtain each IMF and corresponding bandwidth, the
consequent constrained variational problem needed to be solved
2
j
min ∂t δ(t) + ∗ u k (t) e k
− jω t
s.t. u k (t) = y(t) (1)
{u k (t)},{ωk } πt
2 k
where {uk (t)}: = {u1 (t), u2 (t), …, uk (t)}, {ωk }: = {ω1 , …, ωk } are, respectively,
the sets of decomposed components and corresponding center frequencies. In this
way, K different modulated signals with varies frequency ranges can be obtained.
However, the selection of VMD parameter K needed to be determined in advance,
which is important but difficult. In the framework of proposed WVML method, to
avoid the spurious components caused by inaccuracy setting of VMD decomposition
levels, a signal difference average (SDA) criterion [15] is introduced in this section
118 Q. Li et al.
to adaptively determinate the number of IMFs in VMD, that is, the selection of
parameter K. SDA is able to reflect their similarity entre the sum of IMFs and raw
one, which is calculated by
1
n
γ = (yIMFs − ys ) (2)
Nt i=1
where N t is the total number of signal sequence, yIMFs represents the reconstructed
signal by summing all the IMFs, ys is the original signal. The smaller the SDA value,
the higher similarity the IMF and signal have. The SDA-based determination can be
described as
1. For a given range k [l, u], the value of each k is input to decompose the signal
ys .
2. Acquire the reconstructed signal yIMFs by summing all the IMFs.
3. Calculate the SDA γ of each reconstructed signal yIMFs and obtain the SDA curve.
Choose the most reasonable k value to be the first-rank decompose level K opt .
4. Finally, waveforms of K opt IMFs can be obtained, which actually contain different
modulated distributions in frequency domain.
2.2 Time–Frequency Manifold Learning
The TFM signature can reflect the nonlinear manifold characteristics of non-
stationary signals in time–frequency domain. It is usually mined through operating
manifold learning at a succession of TFDs in reconstructed phase space [10]. In this
study, the TFM signature of each IMF is learned on its TFD at a clear frequency
range, as a result different modulated bands are treated due to their different distri-
bution characteristics. A TFM is generally learnt through the following procedures:
A time series analysis technique called phase space reconstruction (PSR) is firstly
applied to rebuild the manifold. For a given IMF x(t) with the number of N t points,
its ith phase point vector in m-dimensional phase space can be calculated by
X im = [xi , xi+τ , . . . , xi+(m−1)τ ] (3)
where m represents the embedding dimension, x i represents the ith point in IMF x(t),
and τ represents the time delay. Then, vectors {X im |i = 1, 2, …, n} are aligned by
chronological order to develop a time-dependent matrix in the phase space

T
P = X 1m , X 2m , . . . , X nm ∈ R m×n (4)
P( j, h) = x h+( j−1)τ (5)

where τ = 1, n = N t − m + 1, j [1, m] and h [1, n].

Then, the phase space tracking is operated in the time–frequency domain so as to
transform each row of the time-dependent matrix into a TFD by discrete STFT
∞

P j [l]w[h − l]e−i vl
2π
S j (v, h) = L (6)
l=−∞
where h represents the position on time axis, v represents the position on frequency
axis. L is the total number of discrete frequency, w(h) represents a short window
sliding on time axis. The obtained Sj (v,h) can be regarded as a series of complex
matrices with the size of L × n, where L is the total frequency points and n is the
total time points. Consequently, m TFDs which only contain the amplitude matrices
Aj , and their corresponding phase part θj are able to be obtained, respectively, through
Sj (v,h), while each of them can also be represented as a matrix. By utilizing a mani-
fold learning algorithm in the reconstructed phase space, the TFM signature can be
obtained from previous TFDs. The local tangent space alignment (LTSA) algorithm
is applied in this paper to compute therein d TFM signatures. Thus, the latent struc-
ture reflecting the desired pattern of the IMF can be automatically extracted with the
noise removed, particularly on in-band noise suppression to some extent. Further-
more, the waveform of each TFM signature in time domain is able to be transformed
via inverse STFT.
2.3 Weighted Coefficient-Based Signal Reconstruction
As Ref. [10] shows, the manifold signature will lose absolute amplitude information
during the operation of matrix normalization, as a result, the max value of ampli-
tude of each IMF will be the same, which is contradictory to the mechanism that
different IMFs contributes different weights to the signal. Therefore, a group of
weight coefficients is introduced here to reconstruct the waveform more accurately.
The Pearson correlation coefficient [13] is a statistical indicator used to measure the
degree of correlation between two variables. For the decomposed kth IMF x k (t) = [x 1 ,
x 2 ,…,x Nt ] and the raw signal y(t) = [y1 , y2 ,…,yNt ], the cross-correlation coefficient
can be calculated as follows
N t
i=1 (x i − x)(yi − y)
r k = N t (7)
Nt
i=1 (x i − x) 2·
i=1 (yi − y) 2
where x is the mean of x k (t), y is the mean of y(t). Nt is the entire number of points
in sequence x k (t). The result r k value ranges from –1 to 1. The closer the result is to
1, the more relevant the two variables are. Based on the assumption that the output
120 Q. Li et al.
IMF component containing transient information related to the fault features has a
high correlation with the raw signal, the coefficient r k of any IMF x k (t) and the raw
signal y(t) is used to give the corresponding component different proportion in the
synthesized signal. And finally, the manifold signal y (t) can be synthesized by

K opt
y(t) = rk · xk (t) (8)
k=1
2.4 Procedure of the WVML Method
Figure 1 illustrates the steps of proposed WVML method. In this manner, the new
synthesized signal can be well processed for further fault diagnosis according to
rotating machinery like bearing and gear.
3 Experimental Study
In this section, a case of experimental gear data is analyzed to validate the advantages
and effect of the proposed WVML method in manifold mining for further health
diagnosis of rotating machinery.
3.1 Experimental Facility
A gearbox setup is built for the validation of proposed WVML method. As Fig. 2
shows, it is chiefly made up of a 3-phase induction motor, two flexible couplings, a
magnetic powder brake, and a two-stage gearbox. In addition, this two-stage gearbox
contains a total transmission ratio as 3.59, and the teeth number of each gear is labeled
in Fig. 2. A thin spalling fault with length of 15 mm, width of 3 mm, and depth of
1 mm was set in the second-stage gear Z 3 .
Four acceleration meters are, respectively, installed at the bearing plates of this
gearbox. In addition, an industrial computer with a LMS acquisition system was
employed to collect the vibration defective signal for follow-up processing and anal-
ysis. Figure 3 illustrated the picture of the experimental data acquisition setup. The
sampling frequency is 5120 Hz. To initially reduce the amount of calculation, a 4×
down sampling operation was performed in the later analysis. It is explained the
vibration signal collected by acceleration meter installed at the bearing plates of gear
Z 3 is analyzed in this validation.
Fig. 1 Steps of the WVML method
3.2 Experimental Results
To practically verify the effect of the WVML method, the vibration signal collected
at 700 rpm, 60 Nm with 5120 points was analyzed as an example. Figure 4 shows the
waveform and spectra characteristics of raw spalling defective signal, from which
can see there are inconspicuous transient impulses submerged by much background
noise in both Fig. 4a, b. In addition, by Hilbert transform and Fourier transform the
frequency components are able to be represented in Fig. 4c, d, where the modulation
frequency and harmonic frequency is difficult to identify in two frequency spectra.
122 Q. Li et al.
Fig. 2 Schematic of the gearbox setup
Fig. 3 Photograph of the experimental setup
The first process of WVML is decomposing the raw signal into K opt IMFs with
respective modulated frequency information in a variational solution. The value K opt
was determined based on SDA criteria, and the curve was plotted as Fig. 5a. Within
the range of effective decomposition levels, the minimum SDA value acquired when
Fig. 4 Raw spalling defective signal: a waveform in time domain; b TFD; c amplitude spectrum;
d envelope spectrum
Fig. 5 SDA curve and modes decomposition of gear defective signal
k = 3, corresponds to the optimal decomposition levels. Figure 5b represents the

frequency bands separation of each IMF by means of VMD.
The three decomposed IMF were then transformed into time–frequency domain to
accomplish manifold learning. The comparison of IMF2 and its reconstructed signal
in Fig. 6 was taken as an example to illustrate TFM learning. From Fig. 6b, d, it can
be observed within the frequency range of [400, 640] the manifolds were success-
fully extracted from the raw TFD with full-band noise repressed. Furthermore, the
manifold signal reconstructed by a series of inverses transform shows more clear
periods of transient impulses.
All the three manifold signals were obtained to final learned signal reconstruction.
The final waveform and TFD, amplitude spectrum, and its corresponding envelope
spectrum are orderly plotted in Fig. 7. With the comparison of raw waveform in
Fig. 4, it can be observed that the periodic features have been enhanced in not
only the time domain but also time–frequency domain. From Fig. 7c, d can see
the meshing frequencies f m1 and f m2 in amplitude spectrum, as well as modulated
124 Q. Li et al.
Fig. 6 The comparison between IMF2 and its manifold signal: a raw signal of IMF2; b TFD of
IMF2; c reconstructed manifold signal of IMF2; d TFM of IMF2
Fig. 7 Reconstructed spalling defective signal: a waveform in time domain; b TFM; c amplitude
spectrum; d envelope spectrum
rotating frequency f r and its multi-harmonic frequencies in envelope spectrum were

emphasized, which is significative for gear fault diagnosis. From the application in
gear fault diagnosis, it can be concluded that based on full frequency band infor-
mation preserving, the manifold structure for the desired time–frequency impulses
is significantly enhanced via the proposed WVML method, which also provides a
more obvious spectra result for fault diagnosis.
4 Conclusion
In this paper, a novel time–frequency analysis method named weighted variational

manifold learning (WVML) is proposed for complex modulated transient features
enhancement. Different from existing methods, this WVML method focuses on
multi-scale feature reconstruction with lossless recovery of defect information. By
employing the mode self-separation theory into the framework of manifold learning,
the complex modulation information embedded in non-stationary signal can be
soundly reinforced with mode manifolds mined and weighted in a self-learning way.
Additionally, a SDA criterion is introduced for optimal decomposition level deter-
mination. Practical gear defective signal is analyzed with the result; the proposed
method presents significant effectiveness in complex multi-scale sideband informa-
tion extraction with transient features enhancement. Furthermore, it can be fore-
seen the proposed WVML method supplies a proper time–frequency architecture for
multi-scale modulation analysis in which data mining will play a greater potential.
References
1. Rai A, Upadhyay S (2016) A review on signal processing techniques utilized in the fault
diagnosis of rolling element bearings. Tribol Int 96:289–306
2. Ding X, He Q (2017) Energy-fluctuated multiscale feature learning with deep convent for
intelligent spindle bearing fault diagnosis. IEEE Trans Instrum Meas 66(8):1926–1935
3. Chen X, Wang S, Qiao B, Chen Q (2018) Basic research on machinery fault diagnosis: past,
present, and future trends. Front Mech Eng 13(2):264–291
4. Hussain S (2017) Fault diagnosis of gearbox using particle swarm optimization and second
order transient analysis. J Vibr Acoust-Trans ASME 139(2):021015
5. Pandya D, Upadhyay S, Harsha S (2015) Fault diagnosis of high-speed rolling element bearings
using wavelet packet transform. Int J Sig Imaging Syst Eng 6:390–401
6. Zvokelj M, Zupan S, Preil I (2016) EEMD-based multiscale ICA method for slewing bearing
fault detection and diagnosis. J Sound Vib 370:394–423
7. Li F, Li R, Tian L et al (2019) Data-driven time-frequency analysis method based on variational
mode decomposition and its application to gear fault diagnosis in variable working conditions.
Mech Syst Sig Process 116:462–479
8. Liu C, Cheng G, Chen X, Pang Y (2018) Planetary gears feature extraction and fault diagnosis
method based on VMD and CNN. Sensors 18(5):1523
9. Bagheri A, Ozbulut O, Harris D (2018) Structural system identification based on variational
mode decomposition. J Sound Vib 417:182–197
10. He Q, Liu Y, Long Q, Wang J (2012) Time-frequency manifold as a signature for machine
health diagnosis. IEEE Trans Instrum Meas 61(5):1218–1230
11. Yan J, Sun H, Chen H et al (2018) Resonance-based time-frequency manifold for feature
extraction of ship-radiated noise. Sensors 18(4):936
12. He M, He D, Qu Y (2016) A new signal processing and feature extraction approach for bearing
fault diagnosis using AE sensors. J Fail Anal Prev 16(5):821–827
13. Ding X, He Q (2016) Time-frequency manifold sparse reconstruction: a novel method for
bearing fault feature extraction. Mech Syst Sig Process 80:392–413
126 Q. Li et al.
14. Li Y, Cheng G, Liu C, Chen X (2018) Study on planetary gear fault diagnosis based on
variational mode decomposition and deep neural networks. Measurement 130(8):94–104
15. Isham M, Leong M, Lim M et al (2018) Variational mode decomposition: mode determination
method for rotating machinery diagnosis. J Vibroengineering 20(7):2604–2621
Recent Advances on Anomaly Detection
for Industrial Applications
A Data-Driven Approach to Anomaly
Detection and Health Monitoring
for Artificial Satellites
Takehisa Yairi, Yusuke Fukushima, Chun Fui Liew, Yuki Sakai,

and Yukihito Yamaguchi
Abstract In the operation of artificial satellites, overall health monitoring of the

system and detection of symptoms of potential troubles are very important. For
about twenty years, we have studied the data-driven health monitoring methods for
artificial satellites, in which machine learning techniques are applied to the house-
keeping data of satellites. In this paper, we review the basic of anomaly detection
based on unsupervised learning and some results in early studies and then discuss
the lessons learned and challenges to be tackled in the data-driven health monitoring
for artificial satellites.
Keywords Anomaly detection · Data-driven health monitoring · Artificial satellite
1 Introduction
One of the most important tasks in the operation of artificial satellites is health
monitoring or checking whether the system is working correctly or not. As the modern
satellite systems are getting more complicated and an increasing number of satellites
are planned to be launched in near future, it is impossible or impractical for a limited
number of experienced operators to monitor all satellites individually and carefully.
Therefore, more reliable and inexpensive ways of health monitoring for satellites are
demanded. The most simple and primitive method of spacecraft health monitoring
is the so-called limit-checking, in which upper and lower limits of normal ranges are
imposed on the variables (or sometimes called parameters) to be monitored. Then, if
any sensor measurement goes over the limit, an alarm goes off or a warning is issued.
As the limit-checking is easy to understand and implement for humans, it still remains
T. Yairi (B) · C. F. Liew

Research Center for Advanced Science and Technology, The University of Tokyo, 4-6-1 Komaba,
Meguro-ku, Tokyo, Japan
e-mail: yairi@g.ecc.u-tokyo.ac.jp
Y. Fukushima · Y. Sakai · Y. Yamaguchi
Research Center for Advanced Science and Technology, The University of Tokyo, 4-6-1 Komaba,
Meguro-ku, Tokyo, Japan

https://doi.org/10.1007/978-981-15-9199-0_13
130 T. Yairi et al.
the main monitoring method in spacecraft operation. It is not always easy, however,
to decide appropriate threshold values for tremendous number of variables. Besides,
it is very hard to detect subtle early symptoms of potentially serious troubles by
limit-checking. Rule-based [1] and model-based [2] approaches were studied and
developed as more advanced health monitoring and diagnosis methods, which are
expected to overcome the limitations of limit-checking. While those approaches have
certain advantages that they are able to detect unknown anomalies and diagnose
the causes of detected faults by the powerful inference capabilities, it is generally
expensive to construct and maintain the rule bases or model bases because a lot
of expert knowledge is required to do so. This prevents us from employing these
advanced health monitoring methods for every small satellite, as a different satellite
needs a different rule base or model. It is impractical from the viewpoint of cost.
In these circumstances, the data-driven approach to the health monitoring of arti-
ficial satellites draws much attention these days [3–5]. It does not require the expert
knowledge, unlike the rule- and model-based approaches. Instead, this approach
utilizes the big amount of past sensor data in spacecraft bus systems (i.e., house-
keeping or HK data) as its main information source. The rising popularity of machine
learning and artificial intelligence is another driving force. On the other hand, the
data-driven approach does not work at all without data. Besides, in the health moni-
toring of artificial satellites (and other critical systems), we cannot take advantage
of supervised learning techniques in most cases, as it is common that there is little
amount of past anomaly data in these applications. Authors have been studying and
developing data-driven anomaly detection methods for artificial satellites for more
than 15 years [6–9]. In the remaining of this paper, we will present the lessons learned
in the series of studies and discuss possible approaches to the challenges.
2 Data-Driven Health Monitoring for Artificial Satellites
2.1 Anomaly Detection by Machine Learning
The central part of data-driven health monitoring is the technique called anomaly
detection or novelty detection, which has been developed in the machine learning
research community. Anomaly detection is different from the supervised classifica-
tion in that training data is composed of only normal samples, whereas the supervised
classification takes both normal and anomaly samples as its training data. This differ-
ence comes from the fact that anomaly data samples are usually unavailable or rare
beforehand in actual health monitoring applications. The basic procedure of anomaly
detection is described as follows. (1) A statistical model of normal behavior of the
system is learned from training data. (2) Anomaly score for each test data sample
is computed by evaluating it based on the learned normal behavior model. To learn
the normal behavior models, a variety of unsupervised learning algorithms including
dimensionality reduction, clustering and density estimation can be employed. We
A Data-Driven Approach to Anomaly Detection and Health Monitoring … 131
can understand how the unsupervised learning works as anomaly detector in several
ways. In the following, we introduce three typical — probabilistic, geometrical and
encoder-decoder interpretations.
(a) Probabilistic interpretation
A typical use of unsupervised learning is to estimate the probability density
distribution of given data samples. Those methods are roughly classified into
two types—parametric models such as multivariate Gaussian and mixture of
Gaussians and non-parametric models such as kernel density estimation. When
we apply them to anomaly detection, we first learn the density distribution
model of normal data samples and then compute the (log-)likelihood of the test
samples. If the log-likelihood is significantly low, it is detected as anomaly.
Therefore, the minus log-likelihood can be used as the anomaly score.
(b) Geometrical interpretation
Another major topic in unsupervised learning is dimensionality reduction repre-
sented by principal components analysis (PCA). Dimensionality reduction
methods are meant to approximate the distribution of data in the originally
high-dimensional space by lower-dimensional subspaces or manifolds. In the
application to anomaly detection, we obtain a “normal” subspace representing
the distribution of normal training data and then compute the distance of each
test sample from the obtained normal subspace. If the distance is larger than
threshold, it is regarded as anomaly. Similarly, the clustering which is another
representative unsupervised learning can be also used for this purpose. In that
case, we compute the distance from a test sample to the nearest cluster center.
(c) Encoder/decoder interpretation
An autoencoder is an unsupervised neural network, which can be viewed as
a non-linear extension of PCA. It consists of an encoder which converts orig-
inal high-dimensional data into lower-dimensional data and a decoder which
recovers the original dimensionality from the compressed data. The encoder
and decoder are trained simultaneously so that they recover the original data
with minimum loss. The idea of using an autoencoder for anomaly detection
can be explained as follows. Once the autoencoder is trained using only normal
data samples, it will recover a given test data sample well if it is normal, but
will fail it if it is anomaly. Therefore, we can use the reconstruction error or the
difference between the input and output patterns as the anomaly score.
2.2 Examples of Anomaly Detection Applied to Satellite Data
In this subsection, we show a few examples of applying data-driven anomaly detec-

tion methods to the house-keeping data of satellites. The details of the satellites are
omitted here for confidentiality reasons.
(a) Detection of a solar paddle trouble
Figure 1 shows a part of electric power subsystem (EPS) data of a commu-
nication testing satellite. Although it is hard to detect by looking, a trouble
132 T. Yairi et al.
Fig. 1 HK data in EPS of a communication testing satellite
in one of solar paddles was observed at the time indicated by a red dotted
line in the figure. In this case study, we applied several unsupervised learning
algorithms to the early part of this period for training to obtain density distribu-
tion models and then evaluated each data sample in the whole period. Figure 2
shows the anomaly scores (minus log-likelihoods) evaluated by them. Since this
trouble is hard to detect even for experts, it is not surprising that simple distri-
bution models such as probabilistic PCA (PPCA) and Gaussian mixture failed
in detecting the anomaly. We found, however, that the minus log-likelihood
computed by mixture of probabilistic PCA (MPPCA) significantly increased
at the time in which the trouble occurred, which means MPPCA successfully
detected the anomaly. We consider that this is because MPPCA is more suitable
for modeling the high-dimensional and multi-modal spacecraft HK data than
others.
(b) Detection of battery trouble symptoms
Figure 3 shows a part of HK data of another satellite, consisting of 228 analog
variables. While some irregular patterns can be found in the last part of the
period, it is hard to find significant symptoms prior to them. Figure 4 shows
the anomaly score (i.e., minus log-likelihood) in this period, when MPPCA is
applied to this data. We can see non-trivial increases of the score at least three
times prior to the known troubles. Although experts examined these events
and confirmed that they were not symptoms of the subsequent battery troubles,
they admitted that detection of these events is still valuable because they were
considered to be minor issues in another battery cell. This example shows the
(a) GMM
(b) PPCA
(c) MPPCA
Fig. 2 Log-likelihood of HK data
possibility of the data-driven approach to detect subtle changes or symptoms

which are easily overlooked by operators.
(c) Detection of attitude anomaly and novel events
Figures 5 and 6 show numerical and categorical variables of HK data of yet
another satellite, respectively. In the latter figure, different categorical values
except the most frequent one are indicated by different colors, for each cate-
gorical variable. In general, spacecraft HK data consists of both numerical vari-
ables (typically corresponding to sensor measurements) and categorical vari-
ables (corresponding to discrete modes of subsystem components). In order
to incorporate the categorical variables into the statistical model, we integrated
MPPCA model with multi-dimensional categorical distribution models. Figure 7
shows the anomaly score of this period. We found that it successfully detected
“unusual” events during this period, including altitude anomaly and a rare event
which had never been encountered. As the categorical variables contain the
134 T. Yairi et al.
Known troubles
Training Test
228 variables
Time index
Fig. 3 HK data with battery troubles
information about issued commands, this model is expected to check if the state
of satellite changes as expected after a command is given.
3 Lessons Learned
3.1 Importance of Data Preprocessing
Recently, thanks to the growing interest in machine learning, we can easily use many
kinds of latest algorithms with just a few lines. However, before we apply them to HK
data of artificial satellites, it is necessary for us to preprocess the raw data appropri-
ately in some ways. First, the original HK data usually contains a high percentage of
missing values. This is partly because different sensors have different sampling rates
and partly because operators intentionally omit some portion of data to reduce the
amount of communication from the satellite to the ground. Unfortunately, since the
majority of popular machine learning algorithms are unable to deal with the missing
values, we have to complete the missing elements by some means. Based on our expe-
rience, elaborate completion methods such as high-order polynomial interpolation
are not recommended, because they have a risk of producing undesirable artifacts
as a side effect. Besides, satellite HK data occasionally contains exceptionally large
abnormal values, caused by errors in data conversion or transmission. Although those
error values can also be seen as a kind of anomaly, we are not interested in them
in most cases because they are temporary and not related to any serious troubles in
Time index
Fig. 4 Anomaly score evaluated by MPPCA
(Reconstrucon error)
Anomaly score
136 T. Yairi et al.
Fig. 5 Numerical variables
the whole system. So we call this type of value error “trivial” outlier. Trivial outliers
should be removed beforehand in the data preprocessing stage for two reasons. First,
they may hide the truly serious anomalies from the operators, because the absolute
values of the former are generally much larger than those of the latter. Second, if
outliers with extremely large values are contained in the training data, the learned
models are severely damaged.
3.2 Modeling of Satellite House-Keeping Data
A significant characteristic of satellite HK data is its heterogeneity. That is to say, the

variables consisting HK data are first divided into two types: continuous variables
that take real values and discrete variables that take categorical values. Typically,
the former corresponds to sensor measurements, whereas the latter represents some
status or modes of components. In addition, continuous variables have a diversity
in the physical quantities, units and ranges. This heterogeneity of HK data is a big
difference from other data sources such as imagery data, in which all variables (i.e.,
pixels) have the same range and quantity. Another essential property of HK data
Fig. 6 Categorical variables
Fig. 7 Anomaly score

138 T. Yairi et al.
is the temporal dependency among data samples. In other words, HK data is multi-
dimensional (or even high-dimensional) time series generated from spacecraft which
is a dynamical system. However, it is not always easy and straightforward to model
the temporal dependency, because sensor measurements may have different sampling
rates and are not synchronized. Another remarkable property of HK data is multi-
modality of data distribution or existence of cluster structure. This is because the
whole satellite system and its subsystems have multiple operational modes. Inter-
estingly, within each mode (or cluster), data samples are often distributed on almost
linear subspaces. We consider this is because artificial satellite systems are designed
to operate linearly around stable states. While deep neural networks (DNNs) have
gained great reputation in many domains such as image recognition, natural language
processing and so on, it is rather important for the start-of-the-art techniques to deal
with the properties of satellite HK data described above for successful results.
3.3 Difficulty in Unsupervised Learning
The main difficulty in anomaly detection for artificial satellites comes from the reality
that we have few data samples of anomalies in advance that we want to detect, which
makes it impossible to use supervised classification methods that are most reliable
and promising techniques. Therefore, we have no choice but to use unsupervised
learning methods which have difficulties in deciding thresholds and avoiding false
alarms when the amount of the training data is not sufficient. To overcome this
difficulty, we need domain knowledge or other information sources that compensate
the lack of anomaly data.
3.4 Expected Roles of Data-Driven Approaches
As seen in the past case studies, the data-driven approach to satellite health monitoring
has some advantages over conventional methods. For example, it has a possibility to
discover subtle changes and events in the HK data which may be symptoms of serious
troubles but are occasionally overlooked by humans. However, we consider human
operators and knowledge-based approaches are still necessary. While the data-driven
approach is good at monitoring a big amount of HK data broadly and shallowly, it is
poor in logical inference which is necessary for decision making and fault diagnosis.
4 Challenges
4.1 Explainability
Recently, the concept of “explainable AI” is spreading over many application areas.
The explainability is important also in the satellite health monitoring, as it is not
enough just to detect anomalies. It is required to answer the questions such as “Why
is it considered to be an anomaly?” or “How serious is it?”. The latter question is not
easy, because the anomaly score utilized in the data-driven anomaly detection indi-
cates how likely it is an anomaly, but not how critical the anomaly is. Obviously, an
important challenge in the data-driven health monitoring is to improve its explain-
ability. In our opinion, it is impractical to realize such a high-level explainability
only by (unsupervised) learning from data. Some logical inference to handle expert
knowledge and domain models will be necessary.
4.2 Automation of Data Preprocessing
As we pointed out in Sect. 3.1, the raw HK data sent from satellite needs appropriate
preprocessing, because it is not ready for most of machine learning algorithms. This
preprocessing is a very troublesome and time-consuming task for operators, which
could become a barrier to the wide use of data-driven health monitoring. Therefore,
the automation of the data preprocessing is considered a key.
4.3 Transfer Learning
One difficulty of data-driven approach is that we usually do not have enough amount
of training data in the early stage soon after the launch. It would be desirable to utilize
the data of other satellites instead. As mega constellation missions consisting of the
same type of satellites are being planned and there is a trend toward standardization
of bus systems, this could be realized in future. It is still challenging, however,
to perform transfer learning or to share the training data among different types of
satellites.
4.4 Interaction with Operators
As has been discussed in Sect. 3.1, the anomaly detection based on unsupervised
learning has certain limitations, because it uses only normal data samples in the past.
One possible idea to overcome them is to utilize the user feedback or reaction from
140 T. Yairi et al.
the operators who evaluate the detection results of the system as partial or incomplete
label information and to improve the performance gradually. This is closely related to
multiple topics such as active learning, relevance feedback, semi-supervised learning
and lifelong machine learning that draw attentions in machine learning community.
5 BISHAMON: A New Platform for Data-Driven Health

Monitoring
Based on the lessons and discussions in the previous sections, we are now developing
a novel platform for monitoring house-keeping data of various artificial satellites,
named BISHAMON (Big data-driven Intelligent HeAlth MONitoring). Supported
by Japan Science and Technology Agency, this project is planning to develop a
prototype by the end of FY2019 and to start an experimental service in FY2020.
6 Conclusion
In this paper, we introduced an overview of the data-driven health monitoring for

artificial satellites. This approach takes full advantage of a big amount of past house-
keeping data and machine learning techniques. At the same time, however, we still
have non-trivial challenges before this approach is widely spread and deals with
drastic increase in artificial satellites which are expected in coming years.
Acknowledgements This work was supported by JST START Grant Number JPMJST1814, Japan.
Some results in this paper are from our past studies conducted under the support of JSPS KAKENHI
Grant Numbers 26289320.
References
1. Tallo DP, Durkin J, Petrik EJ (1992) Intelligent fault isolation and diagnosis for communication
satellite systems. Telematics Inf 9(3/4):173–190
2. Williams BC, Nayak PP (1996) A model-based approach to reactive self-configuring systems.
In: Proceedings of the 13th national conference on artificial intelligence, pp 971–978
3. DeCoste D (1997) Automated learning and monitoring of limit functions. In: Proceedings
international symposium on artificial intelligence, robotics, and automation in space, pp 287–292
4. Schwabacher M, Oza N, Matthews B (2009) Unsupervised anomaly detection for liquid-fueled
rocket propulsion health monitoring. J Aerosp Comput Inf Commun 6(7):464–482
5. Iverson DL et al (2012) General purpose data-driven monitoring for space operations. J Aerosp
Comput Inf Commun 9(2):26–44
6. Fujimaki R, Yairi T, Machida K (2005) An approach to spacecraft anomaly detection problem
using kernel feature space. In: Proceeding of the 11th ACM SIGKDD international conference
on knowledge discovery in data mining, pp 401–410
7. Inui M, Kawahara Y, Goto K, Yairi T, Machida K (2009) Adaptive limit checking for spacecraft
telemetry data using kernel principal component analysis. Trans Japan Soc Aeronaut Space Sci
Space Technol Japan 7(ists26):Pf-11–Pf-16
8. Yairi T et al (2010) Spacecraft telemetry monitoring by dimensionality reduction and reconstru
ction. In: The NASA conference on intelligent data understanding
9. Yairi T et al (2017) A data-driven health monitoring method for satellite housekeeping data
based on probabilistic clustering and dimensionality reduction. IEEE Trans Aerosp Electron
Syst 53(3):1–18
Robust Linear Regression and Anomaly
Detection in the Presence of Poisson
Noise Using Expectation-Propagation
Yoann Altmann, Dan Yao, Stephen McLaughlin, and Mike E. Davies
Abstract This paper presents a family of approximate Bayesian methods for joint
anomaly detection and linear regression in the presence of non-Gaussian noise.
Robust anomaly detection using non-convex sparsity-promoting regularization terms
is generally challenging, in particular when additional uncertainty measures about
the estimation process are needed, e.g., posterior probabilities of anomaly presence.
The problem becomes even more challenging in the presence of non-Gaussian, (e.g.,
Poisson distributed), additional constraints on the regression coefficients (e.g., pos-
itivity) and when the anomalies present complex structures (e.g., structured spar-
sity). Uncertainty quantification is classically addressed using Bayesian methods.
Specifically, Monte Carlo methods are the preferred tools to handle complex mod-
els. Unfortunately, such simulation methods suffer from a significant computational
cost and are thus not scalable for fast inference in high dimensional problems. In
this paper, we thus propose fast alternatives based on Expectation-Propagation (EP)
methods, which aim at approximating complex distributions by more tractable mod-
els to simplify the inference process. The main problem addressed in this paper is
linear regression and (sparse) anomaly detection in the presence of noisy measure-
ments. The aim of this paper is to demonstrate the potential benefits and assess the
performance of such EP-based methods. The results obtained illustrate that approxi-
mate methods can provide satisfactory results with a reasonable computational cost.
It is important to note that the proposed methods are sufficiently generic to be used
in other applications involving condition monitoring.
Y. Altmann (B) · D. Yao · S. McLaughlin

School of Engineering and Physical Science, Heriot-Watt University, Edinburgh, UK
e-mail: Y.Altmann@hw.ac.uk
D. Yao
e-mail: dy14@hw.ac.uk
S. McLaughlin
e-mail: S.Mclaughlin@hw.ac.uk
M. E. Davies
School of Engineering, University of Edinburgh, Edinburgh, UK
e-mail: mike.davies@ed.ac.uk

https://doi.org/10.1007/978-981-15-9199-0_14
144 Y. Altmann et al.
Keywords Robust regression · Bayesian inference · Anomaly detection ·

Expectation-propagation · Generalized linear models
1 Introduction
Joint robust linear regression and anomaly detection is a challenging problem encoun-
tered in many industrial applications whereby sensors can be prone to random spu-
rious readings. This problem is particularly challenging when the observation noise
differs from the classical Gaussian assumption (e.g., Poisson noise) and when the
sparse anomalies present spatial or temporal structures. This is typically the case
for spectral unmixing problems [1, 2] for which the detected flux is quantified at
a photon level and is affected by Poisson noise. While Bayesian models coupled
with Monte Carlo sampling strategies stand as classical tools to handle challenging
regression problems has enable uncertainty quantification for subsequent decision
making, they still suffer from a high computational complexity which prevents their
broad deployment in industrial applications requiring fast analysis [3]. Consequently,
approximate Bayesian methods have been receiving a growing interest for their better
scalability.
A classical approach to performing approximate Bayesian inference relies on
using a Kullback-Leibler (KL) divergence as an asymmetric discrepancy criterion
between an exact but complex distribution and an approximating distribution. For
instance, minimizing the direct KL divergence between the approximate and exact
distributions [4] leads to Variational Bayes (VB) approaches. On the other hand,
it turns out that minimizing the reverse KL divergence (between the approximate
and the actual distributions) leads to a saddle point problem. Although provably
convergent double loop algorithms [5] have been proposed, they suffer from high
computational complexity. Expectation-Propagation (EP) methods form a family
of faster fixed point message passing solutions which minimize sequentially (and
locally) a set of reverse KL divergences [6]. Although their convergence is in general
not ensured (damping strategies can be adopted to mitigate this limitation), when EP
methods converge, they can provide better estimates than VB for problems where
both methods can be applied. In particular, VB approaches tend to underestimate
posterior variances more severely and EP. Gaussian approximations or messages
are classically used within the EP framework [7], leading to more tractable KL
divergence minimization problems. To reduce the number of sequential EP updates
involving one-dimensional densities, it is possible to use, locally, multivariate Gaus-
sian approximations. In particular, in high dimensional (e.g., imaging) problems,
these factors can be constrained to have diagonal covariance matrices. Recently, a
class of algorithms called Vector Approximate Message Passing (VAMP) [8] has been
derived using EP with diagonal/isotropic Gaussian messages over vectors, although
the noise model can be non-Gaussian. While Poisson likelihoods have recently been
considered within EP [9], and in particular for linear regression in [10], here we use
EP for joint robust linear regression and anomaly detection. In [10], we showed that
Robust Linear Regression and Anomaly Detection … 145
by considering extended Bayesian models, it is possible to design several EP-based

methods, while trading off computational complexity and accuracy of the approx-
imations. To handle sparse anomalies, we use a similar splitting scheme as in [11]
whereby non-log-concave spike-and-slab prior models can be considered. The pro-
posed models and associated a posteriori estimates are compared, through a series
of experiments, to results obtained via Monte Carlo sampling from the true Bayesian
model (in the absence of anomaly). Although a thorough comparison of the pro-
posed methods with MCMC methods in the presence of anomalies would be more
useful (as in [10] for the more standard linear regression problem), methods such
as those investigated in [1, 2] (see also [3]) are still too computationally demanding
and suffer from poor convergence properties in high dimensional settings and are not
considered here. Thus, here we only present the anomaly detection results obtained
by our method.
The remainder of this paper is organized as follows. The exact Bayesian model
used for linear regression is introduced in Sect. 2. Section 3 discusses the associated
factor graph, the approximation based on Expectatio-Propagation and the associated
update rules. Simulations results obtained using simulated data are reported and
presented in Sect. 4 and conclusions are finally reported in Sect. 5.
2 Exact Bayesian Model
2.1 Likelihood
We consider the robust estimation of a vector x = [x1 , . . . , xN ]T from a set of

noisy measurements y = [y1 , . . . , yM ]T , with N ≤ M . Precisely, assuming mutual
independence between the elements of y, conditioned on the value of the vector
u = [u1 , . . . , uM ]T with positive entries (and depending on x), each measurement ym
follows a Poisson distributions, i.e.,
f (ym |um ) = (um )ym exp [−um ]/ym !, ∀m ∈ 1, . . . , M .
In this work, we consider the problem of linear regression in the presence of anomalies
in the data and we assume that the vector x of interest is related to the vector u,
containing the means of the Poisson distributions above, by the following equation
u = Ax + z r (1)
where A = [a1 , . . . , aM ]T is a known M × N matrix with positive elements (the

vectors {am }m=1,...,M are the rows of A). Moreover, r = [r1 , . . . , rM ]T ∈ RM
+, z =
[z1 , . . . , zM ]T ∈ (0; 1)M is a vector of binary labels and denotes the Hadamard
(element-wise) product, such that v = z r or vm = zm rm , ∀m, i.e.,

0 if zm = 0 (absence of outlier)
vm =
rm if zm = 1 (presence of outlier).
While we could have simply used v to represent the anomalies, using the two vectors
z and r as in [1, 11] allows us to separate and assign prior distributions to the locations
(through z) and the values (through r) of the anomalies.
Using the mutual independence between the elements of y, conditioned on the
value of (x, z, r) (or u), the joint likelihood reduces to

M
M
f (y|x, z, r) = f (ym |x, z, r) = f (ym |aTm x + zm rm ), (2)
m=1 m=1
i.e., to a product of M independent Poisson distributions. Since A = [a1 , . . . , aM ]T is

known, it is omitted hereafter in all the conditional distributions to simplify notations.
2.2 Prior Distributions
To complete our Bayesian model, we need to define prior distributions for the
unknown model parameters in Eq. (2). First, we assume that the prior model f (x) is
separable in x, i.e., can be written as a product of N independent distributions

N
f (x) = fx (xn ).
n=1
Moreover, although different distributions could be assigned to the elements of x,

here we use the same distribution for all the elements of x. As discussed in [10], con-
sidering independent prior distributions based on truncated Gaussian or exponential
distributions simplifies significantly the EP-based inference but more complex prior
models, ensuring the positivity of x can be used. Without loss of generality, here we
consider a product of independent and identical gamma distributions whose shape
and scale parameters (kx , θx ) are user-defined based on prior knowledge about x (if
available). If a limited amount of information is available, weakly informative dis-
tributions (i.e., with large variances) can be used, while still enforcing the positivity
of x.
As mentioned above, using (z, r) instead of v allows us to assign independent
prior distributions to the location and amplitude of the outliers. For the anomaly
amplitudes, we use (as for x) a set of M identical and independent gamma prior
distributions, i.e.,

M
f (r) = fr (rm ).
m=1
Finally, we consider that the sparse anomalies can be grouped across some of the
M measurements. In a similar fashion to [1], in this work we cluster the data into K
groups of successive measurements and assume that if one measurement is corrupted,
its whole group of M /K measurements is corrupted. Precisely, if Ik is the set of
indices of the measurements in the kth group, we can define z̃ (k) ∈ (0; 1) such that
zm = z̃ (k) , ∀m ∈ Ik . We then assign z̃ (k) a Bernoulli distribution with mean π , i.e.,
fz (z̃ (k) = 1) = π, ∀k ∈ 1, . . . , K,
which represents the (user-defined) prior probability of a block to be corrupted by

anomalies. Note that once the user-defined groups are fixed, the prior distribution of
z is fully defined by the prior model assigned to z̃ = [z̃ (1) , . . . , z̃ (K) ]T , i.e.,

K
f (z̃) = fz (z̃ (k) ).
k=1
As illustrated in Fig. 1, if K = M , the locations of the anomalies are completely

uncorrelated. Note that here, we used a simple grouping strategy for the anomalies
and used the same prior probability π for each block, but more complex models (e.g.,
different group sizes, non-successive measurements) can be used depending on the
application.
Fig. 1 Example of sparse anomalies (first and third subplots) generated with M = 200 and π =
10%. While on average, the two anomaly sequences have the same number of non-zeros values,
the two bottom subplots have been obtained with K = M , i.e., without structured sparsity while the
two top plots have been obtained with K = 40 groups of 5 successive measurements
2.3 Joint Posterior Distribution
Using the joint likelihood f (y|x, r, z) and the prior distributions f (x), f (r) and f (z̃),
we obtain
f (y, x, r, z̃) = f (y|x, r, z̃)f (x)f (r)f (z̃) (3)

M
M
N
K
= f (ym |aTm x + zm rm ) fr (rm ) fx (xn ) fz (z̃ (k) ),
m=1 m=1 n=1 k=1
and the (marginal) posterior distribution of x, our main parameter of interest, is given
by

f (x|y) = f (y|x, r, z)f (x)f (r)f (z̃)/f (y)dr. (4)
z̃
In the absence of anomaly, estimating x via maximum a posteriori (MAP) estimation

is relatively simple if f (x) is log-concave since the MAP estimator can be obtained
by solving a convex optimization problem [12–14]. However, the MAP estimator is
only a point estimate and such optimization strategy does not provide a posteriori
confidence measures about x. Moreover, the problem becomes non-convex in the
presence of anomalies, unless a convex prior model is used for v (as in [15] for
instance). In this work, we focus on approximating the posterior mean and covariance
(or marginal variances) of f (x|y) while also estimating the posterior probabilities of
outlier presence, i.e., f (z̃ (k) = 1|y).
Unfortunately, computing the posterior mean and covariance matrix analytically
in not possible, especially when M is large as the integrations involved in Eq. (4)
become computationally intractable. Simulation methods are instead classically used
to sample from the posterior distribution and approximate the quantities of interest.
For instance, constrained Hamiltonian Monte Carlo methods [16] have been con-
sidered to solve regression problems in the presence of Poisson noise [2] (see also
[3] for comparison of samplers including a bouncy particle sampler [17]). However,
simulation-based approaches still suffer from a high computational cost and poor
scalability properties. Indeed, while satisfactory results have been obtained using
MCMC methods [2] for robust spectral unmixing, the associated computational cost
was kept reasonable low thanks to the small number of measurements per pixel
(M = 32, N ≤ 16) and the use of an Ising model promoting correlated locations
of the anomalies and improving the mixing properties of the Markov chains. Thus,
approximate Bayesian methods (including EP methods), appear as promising and
scalable alternative solutions.
3 EP for Robust Linear Regression with Poisson Noise
EP methods appear as promising alternatives to Monte Carlo simulation, in terms

of estimation performance and computational complexity. Similarly to Variational
Bayes (VB) approaches, they consist of approximating the joint posterior distribu-
tion f (x, r, z|y) by a simpler distribution (e.g., by imposing a posteriori independence
between x, r and z) whose moments are simpler to compute. However, in contrast to
VB, EP methods can be applied more generally by choosing the families of approx-
imating distributions and by imposing additional constraints to the solution. In [10]
we compared several EP-based methods for linear regression with Poisson noise,
based on Gaussian approximations with different covariance structures. In partic-
ular, we used a general data augmentation scheme and investigated the estimation
performance as additional constraints (local independence, isotropy) where imposed.
A similar approach is adopted here to handle potential outliers. The key ingredient
used in this work is the consideration of an augmented model based on Eq. (1).
Let u = [u1 , . . . , uM ]T be a vector of auxiliary variables, the model in Eq. (3) can
be extended as
f (y, u, x, r, z) = f (y|u)f (u|x, r, z)f (x)f (r)f (z̃),

with f (u|x, r, z) = M m=1 δ(um − am x − zm rm ) = δ(u − Ax − z r), where δ(·) is
T
the Dirac
delta function. Note that Eq. (3) is can be obtained by marginalizing u, i.e.,
using f (y, u, x, r, z)d u and that computing the marginal mean and covariance of
x can be achieved using either the original or the extended model. While EP-based
inference could be achieved with the original model (similarly to the work in [10],
which first considered a non-extended model), the extended model offers, as will be
discussed in this section, more flexibility in terms of parallel implementation of the
EP updates.
3.1 EP Approximation Using Auxiliary Variables
As in [10, 11], this joint density can be approximated by a product of unnormalized

multivariate Gaussian distributions (for (u, x, r)) and Bernoulli distributions (for z̃).
More precisely, we denote by q(·) approximating factors and use
⎧
⎪
⎪ f (y|u) ≈ qu,0 (u)
⎨
fr (r) ≈ qr,0 (r)
⎪
⎪ fx (x) ≈ qx,0 (x)
⎩
δ(u − Ax − z r) ≈ qu,1 (u)qx,1 (x)qr,1 (r)qz,1 (z̃)
fy (u) fx (x)
y u x
qu,0 (u) qx,0 (x)
δ(u − Ax − z r) fr (r)
r
qu,1 (u)qx,1 (x)qr,1 (r)qz,1 (z̃) qr,0 (r)
fz (z̃)
z̃ qz,0 (z̃)
Fig. 2 Factor graph used to perform EP-based robust linear regression using variable splitting. The
rectangular boxes (resp. circles) represent the factor (resp. variable) nodes and the approximating
factors are shown in blue
where the approximation are defined as

⎧
⎪
⎪ qu,i (u) ∝ N (u; mu,i , u,i ), ∀i ∈ (0; 1)
⎨ q (r) ∝ N (r; m , ), ∀i ∈ (0; 1)
r,i r,i r,i
(5)
⎪
⎪ qx,i (x) ∝ N (x; mx,i , x,i ), ∀i ∈ (0; 1)
⎩
qz,1 (z̃) ∝ Kk=1 Ber(z̃k πk,1 ),
where N (·; mu,i , u,i ) is the probability density function of the multivariate normal
distribution with mean mu,i and covariance u,i . Note that since the approximating
distributions for z̃ are products of Bernoulli distributions (as f (z̃)), the approximation
of f (z̃) is exact and we obtain qz,0 (z̃) = f (z̃). In other words, qz,0 (z̃) will not need
to be updated during the EP algorithm (see comment below). The resulting factor
graph is depicted in Fig. 2. Using this factorization, the overall approximation of
f (u, x, r, z, y), denoted by Q(u, x, r, z) (its dependence on y has been omitted for
ease of reading) is given by
Q(u, x, r, z̃) = Qu (u)Qx (x)Qr (r)Qz (z̃)

where Qu (u) ∝ 1i=0 qu,i (u), Qx (x) ∝ 1i=0 qx,i (x), Qr (r) ∝ 1i=0 qr,i (r) and Qz (z̃)

∝ 1i=0 qz,i (z̃). To simplify notations later, we also introduce
Qi (u, x, r, z̃) = qu,i (u)qx,i (x)qr,i (r)qz,i (z̃), ∀i ∈ (0; 1),
such that Q(u, x, r, z̃) = Q0 (u, x, r, z̃)Q1 (u, x, r, z̃).

3.2 EP Updates
The overall aim of EP methods is to optimize Q(u, x, r, z̃) to minimize the Kullback-
Leibler (KL) divergence1
min KL (f (u, x, r, z, y)||Q(u, x, r, z̃)) .

Q(u,x,r,z̃)
Unfortunately, this minimization problem is intractable and the EP updates consist

instead of the sequential and iterative optimization of each factor node in Fig. 2.
Precisely, one iteration of the four updates required is given by

(i) min KL f (y|u)qu,1 (u)||qu,0 (u)qu,1 (u) ,
qu,0 (u)
(ii) minKL (δ(u − Ax − z r)Q0 (u, x, r, z̃)||Q1 (u, x, r, z̃)Q0 (u, x, r, z̃)),
Q1 (u,x,r,z̃)

(iii) min KL f (x)qx,1 (x)||qx,0 (x)qx,1 (x) ,
qx,0 (x)

(iv) min KL f (r)qr,1 (x)||qr,0 (r)qr,1 (r) .
qr,0 (r)
These updates are then repeated until convergence. Note that although 5 factor nodes
are displayed in Fig. 2, since qz,0 (z̃) is in the same family of distributions as f (z̃), we
obtain qz,0 (z̃) = f (z̃) which can be set during the initialization of the method. While
several stopping criteria can be used, here we simply iterate until the variations of the
means and variances of the approximation Q(u, x, r, z̃) between successive iterations
become negligible. We also set a maximal number of iterations (e.g., 100) to stop the
algorithm if the first condition is not met. It is important to note that in general, EP
methods are not guaranteed to convergence. However, as in [10] we used a classical
damping strategy (see [11] for details) and did not experience convergence issues in
all the scenarios considered in Sect. 4.
Prior to presenting briefly how the updates (i)–(iv) are performed, it is important
to discuss the additional constraints imposed to Q(u, x, r, z̃) and in particular to the
different covariance matrices in Eq. (5). While allowing full covariance matrices
increases the flexibility and potentially the quality of the approximation, it can also
lead to numerical instabilities if too general approximations are used. For instance,
since we used products of independent prior distributions to define fx (x) and fr (r), it is
not useful to use full covariance matrices in their approximations, e.g., in qx,0 (x) and
qr,0 (x). While a full analysis of the optimal covariance matrices is interesting, it is not
the main contribution of this work and leave it for future work. In Table 1, we report
the structures of the different covariance matrices involved in Fig. 2 and will motivate
some of these choices in the next paragraph. We did not impose additional constraints
on the means of the different distributions as we did not experience numerical issues
with the covariance matrices in Table 1.
1 Notethat VB methods are also based on KL-divergence minimization, but they use the direct
KL-divergence KL (Q(u, x, r, z̃)||f (u, x, r, z, y)).
Table 1 Properties of the different covariance matrices involved in the approximation of Fig. 2.
“Isotropic” means that the covariance matrix is proportional to the identify matrix
u,0 u,1 r,0 r,1 x,0 x,1
Diagonal Isotropic Diagonal Isotropic Diagonal Full
For brevity, we only present briefly how the the four updates (i)–(iv) are com-
puted. The interesting reader is invited to consult [10, 11] for further details. Because
Gaussian and Bernoulli distributions are chosen as approximating distributions,
the KL minimization problems (i)–(iv) can be solved by computing the mean and
(co)variance of distributions on the left-hand side of the divergences (also referred
to as tilted distribution) and by matching these moments to those of the distributions
on the right-hand side of the KL divergence. If constraints are added (e.g, isotropy of
the covariance matrix), the KL minimization can be achieved using gradient-based
methods (see [11]) as it can be casted as a convex optimization problem.
3.2.1 Update of qu,0
Using the facts that u,1 is diagonal (it is even proportional to the identity matrix) and
that the likelihood is separable in u, the tilted distribution f (y|u)qu,1 (u) has a diag-
onal covariance matrix. Consequently the KL minimization in (i) can be performed
element-wise using the method proposed in [9].
3.2.2 Update of Q1
The update of Q1 relies on computing moments of the tilted distribution δ(u −

Ax − z r)Q0 (u, x, r, z̃). While it is not possible to compute analytically the full
covariance matrix associated with that tilted distribution, it is worth recalling that
Q(u, x, r, z̃) = Qu (u)Qx (x)Qr (r)Qz (z̃). Thus, we only need to compute the moments
of the marginals of the tilted distribution w.r.t. x, {um }m , {rm }m and {zm }m . These
updates are similar to those used in [10, 11] and thus not further detailed here.
However, it is important to mention here that while we used a full covariance matrix
for x,1 to capture posterior correlation between the elements of x induced by A,
it is possible, as in [10, 11] to only keep its diagonal elements. This can accelerate
the update to qx,0 below at the cost of the loss of posterior correlations between the
elements of x in Qx (·).
3.2.3 Update of qx,0
Once Q1 (·) has been updated, the update of qx,0 is performed exactly as in the “EP-
DF” method in [10]. Because x,1 is a full matrix, the marginal means and variances
of qx,0 are updated sequentially (see [10] for details). However, if x,1 is chosen to
be diagonal, the update of qx,0 is the same as that of qr,0 below.
3.2.4 Update of qr,0
Similarly to qx,0 , the update of qr,0 is performed exactly as in the "EP-II" method
in [10]. Because r,1 is diagonal, the marginal means and variances of qr,0 can be
updated in a parallel manner. The tilted distribution f (r)qr,1 (x) is proportional to
the product of M independent truncated Gaussian distributions, whose moments are
easy to compute analytically.
Since many updates can be performed in parallel, the EP scheme used is highly
scalable (when M is large). The next section presents simulation results illustrating
the benefits of the proposed inference strategy.
In this work, we illustrate the potential benefits of the proposed approach for robust
liner regression and anomaly detection through a series of experiments conducted
with simulated data, which allows us to use ground truth information for validation.
4.1 Comparison with Monte Carlo Sampling

(Without Anomalies)
As discussed at the end of Sect. 2, MCMC methods can become extremely slow for
large value of M when addressing the joint linear regression and anomaly detec-
tion problem, which makes the comparison with the reference/optimal estimators
(considering the MCMC-based posterior mean and covariances as gold standard)
difficult in practice. Thus, we first compare the performance of the proposed method
to that of an MCMC assuming no outliers and by generating data without outliers.
In all the simulations conducted in this section we used (M , N ) = (500, 20) and
the matrix A has been generated randomly such that its columns are correlated, as
can been seen in Fig. 3. First, we generated anomaly-free data using x generated
from a product of independent gamma distributions with shape and scale parameters
equal to (kx , θx ) = (1, 1). These parameters have been also used to define the prior
distribution fx (·). Moreover, we used K = M (no structured sparsity), π = 10% and
a product of independent gamma distributions with shape and scale parameters equal
to (kr , θr ) = (1500) for f (r).
Fig. 3 AT A for the matrix A used in Sect. 4. The condition number of A is κ(A) = 35.18
Figure 4 compares the posterior mean and covariance matrices estimated using the
proposed EP-method and the MCMC method used as reference in [10] (using NMC =
20, 000 Monte Carlo iterations to ensure the convergence of the chains). This figure
shows that in the absence of anomalies, the two methods provide very similar results,
illustrating the ability of the EP method to approximate the actual non-Gaussian
posterior distribution. Note that for these results all the marginal probabilities f (z k |y)
obtained via EP satisfy f (z k |y) < 10−4 , i.e., the method clearly identifies that no
anomalies are present. Moreover, although the EP method estimates more parameters
than the MCMC method (it estimates means/covariances associated with (r, z)), the
EP method converges within 50 iterations in 1.5s using Matlab 2017b implementation
running on a MacBook Pro with 16 GB of RAM and a 2.9 GHz Intel Core i7 processor,
while the MCMC method takes 10 s.
Fig. 4 Top: ground truth x (red) and posterior means of x estimated by the proposed EP-based
method (black) and reference MCMC method (dashed blue) in the absence of outliers in the data.
Bottom: associated posterior covariance matrices obtained by the two methods
4.2 Anomaly Detection
In this section, we illustrate the ability of the proposed method to perform jointly
robust linear regression and anomaly identification in the presence of Poisson noise.
First, we generated data corrupted with anomalies using (kx , θx ) = (1, 1), (kr , θr ) =
(1500), K = M and π = 10%. The main results obtained are presented in Fig. 5.
The second subplot shows the posterior mean and associated ±3 standard deviation
intervals (obtained from the estimated posterior covariance matrix) of x. The posterior
mean of r is presented in the third subplot but the confidence interval are omitted
here as they are difficult to see. The bottom subplot depicts the marginal probabilities
f (z k = 1|y) of outliers presence together with the actual values of z (k) (−z (k) is
displayed to facilitate the visual comparison). This figure shows that the proposed
method is able to identify all the significant anomalies (associated with high posterior
probabilities f (z k |y)) while the smallest anomalies are assigned lower probabilities.
Finally, we assess the performance of the proposed method in the presence of
spatially correlated anomalies. As in the previous experiment, we used (kx , θx ) =
(1, 1), (kr , θr ) = (1, 500) and π = 10% but the anomalies were generated using
K = M /5 (groups of 5 successive anomalies). Figure 6 depicts the regression and
Fig. 5 Top row: realization of y generated using Eq. (2). Second row: ground truth x (red line),
posterior mean (black line) and associated ±3 standard deviation intervals (blue dashed lines) of x
in the presence of 10% of outliers in the data. Third row: actual anomalies (red line) and posterior
mean of r (black line). Bottom: marginal probabilities f (z (k) = 1|y) of outliers presence (blue line)
together with the actual values of z (k) (red line)
anomaly detection results obtained using K = M /5 (left-hand side subplots) and

K = M (right-hand side subplots). While including additional information about the
anomaly structure does not significantly improves the regression performance in this
example (similar variances in the top row of Fig. 6), it improves the detection and
localization of the anomalies. Indeed, it further reduces the probabilities f (z (k) |y) in
the absence of anomalies, i.e., it reduces the overall probability of false alarm.
5 Conclusions
In this work, we present a new set of Expectation-Progagation methods for joint

robust linear regression and anomaly detection in the presence of Poisson noise.
Using an extended Bayesian model, it is possible to derive efficient EP updates and
handle non-convex posterior distributions induced by the use of spike-and-slab prior
models. The resulting approximate Bayesian inference scheme provides satisfactory
results, with a significantly shorter computational cost when compared to intrinsically
sequential Monte Carlo sampling strategies. Although simple experiments have been
Fig. 6 Linear regression and anomaly detection results obtained using K = M /5 (left-hand side
subplots) and K = M (right-hand side subplots) for data generated with K = M /5. Top: ground
truth x (red line), posterior mean (black line) and associated ±3 standard deviation intervals (blue
dashed lines) of x in the presence of 10% of outliers in the data. Middle: actual anomalies (red
line) and posterior mean of r (black line). Bottom: marginal probabilities f (z (k) = 1|y) of outliers
presence (blue line) together with the actual values of z (k) (red line)
conducted with 1D signals, the method can be extended to handle group-sparsity in

2D (e.g., images) or higher dimensions. Additional experiments conducted on the
robust spectral unmixing problem addressed in [2] (but not presented here) confirmed
the benefits of the proposed method for detection of structured anomalies in spectral
signatures. Future work first includes a thorough comparison with reference MCMC-
based methods (e.g., the method proposed in [1]) to confirm these preliminary results.
It would be also interesting to extend this method to other anomaly models, e.g.,
where u = (1 − z) (Ax) + z r.
Acknowledgements This work was supported by the Royal Academy of Engineering under the
Research Fellowship scheme RF201617/16/31.
References
1. Newstadt GE, III AH, Simmons J (2014) Robust spectral unmixing for anomaly detection. In:
Proceeding IEEE-SP workshop statistical and saignal processing, Gold Coast, Australia
2. Altmann Y, Maccarone A, McCarthy A, Newstadt G, Buller GS, McLaughlin S, Hero A
(2017) Robust spectral unmixing of sparse multispectral lidar waveforms using gamma Markov
random fields. IEEE Trans Comput Imaging 3(4):658–670
3. Tachella J, ltmann Y, Pereyra M, Tourneret J-Y (2018) Bayesian restoration of high-dimensional
photon-starved images. In: Proceeding European signal processing conference (EUSIPCO),
Rome, Italy
4. Blei DM, Kucukelbir A, McAuliffe JD (2017) Variational inference: a review for statisticians.
J Am Stat Assoc 112(518):859–877
5. Yuille AL, Rangarajan A (2002) The concave-convex procedure (cccp). In: Advances in neural
information processing systems. MIT Press, pp 1033–1040
6. Minka TP (2001) Expectation propagation for approximate bayesian inference. In: Proceed-
ings of the seventeenth conference on uncertainty in artificial intelligence. Morgan Kaufmann
Publishers Inc. pp 362–369
7. Seeger MW (2008) Bayesian inference and optimal design for the sparse linear model. J Mach
Learn Res 9:759–813 June
8. Schniter P, Rangan S, Fletcher AK (2016) Vector approximate message passing for the gener-
alized linear model. In: 50th asilomar conference on signals. Systems and computers. Pacific
Grove, CA, USA, Nov, , pp 1525–1529
9. Ko Y-J, Seeger MW (2016) Expectation propagation for rectified linear poisson regression. In:
Asian conference on machine learning, ser. Proceedings of machine learning research, vol 45,
Hong Kong, pp 253–268
10. Altmann Y, Perelli A, Davies ME (2019) Expectation-Propagation algorithms for linear regres-
sion with Poisson noise: application to photon-limited spectral unmixing. In: IEEE International
Conference on Acoustics, Speech, and Signal Processing (ICASSP), Brighton, UK
11. Hernández-Lobato J, Hernández-Lobato D, Suárez A (2015) Expectation propagation in linear
regression models with spike-and-slab priors. Mach Learn 99(3):437–487. [Online]. Available:
https://doi.org/10.1007/s10994-014-5475-7
12. Figueiredo M, Bioucas-Dias J (2010) Restoration of Poissonian images using alternating direc-
tion optimization. IEEE Trans Image Proc 19(12):3133–3145
13. Altmann Y, Maccarone A, Halimi A, McCarthy A, Buller GS, McLaughlin S (2016) Efficient
range estimation and material quantification from multispectral lidar waveforms. In: Proceeding
sensor signal processing for defence (SSPD) conference. Edinburgh, UK
14. Tobin R, Altmann Y, Ren X, McCarthy A, Lamb RA, McLaughlin S, Buller GS (2017) Com-
parative study of sampling strategies for sparse photon multispectral lidar imaging: towards
mosaic filter arrays. J Opt 19(9):094006
15. Févotte C, Dobigeon N (2015) Nonlinear hyperspectral unmixing with robust nonnegative
matrix factorization. IEEE Trans Image Proc 24(12):4810–4819
16. Brooks S (2011) Handbook of markov chain monte carlo, Chapman & Hall/CRC Handbooks
of Modern Statistical Methods. Taylor & Francis
17. Bouchard-Côté A, Vollmer SJ, Doucet A (2018) The bouncy particle sampler: a nonreversible
rejection-free markov chain monte carlo method. J Am Stat Assoc 113(522):855–867
Spacecraft Health Monitoring Using
a Weighted Sparse Decomposition
Barbara Pilastre, Jean-Yves Tourneret, Loïc Boussouf,

and Stéphane D’escrivan
Abstract In space operations, spacecraft health monitoring and failure prevention

are major issues. This important task can be handled by monitoring housekeep-
ing telemetry time series using anomaly detection (AD) techniques. The success of
machine learning methods makes them attractive for AD in telemetry via a semi-
supervised learning. Semi-supervised learning consists of learning a reference model
from past telemetry acquired without anomalies in the so-called learning step. In a
second step referred to as test step, most recent telemetry time-series are compared
to this reference model in order to detect potential anomalies. This paper presents
an extension of an existing AD method based on a sparse decomposition of test
signals on a dictionary of normal patterns. The proposed method has the advantage
of accounting for possible relationships between different telemetry parameters and
can integrate external information via appropriate weights that allow detection per-
formance to be improved. After recalling the main steps of an existing AD method
based on a sparse decomposition Pilastre et al (Sign Proc, 2019 [1]) for multivariate
telemetry data, we investigate a weighted version of this method referred to as W-
ADDICT that allows external information to be included in the detection step. Some
representative results obtained using an anomaly dataset composed of actual anoma-
B. Pilastre (B) · J.-Y. Tourneret

INP-ENSEEIHT, University of Toulouse, 2 Rue Camichel, 31071 Toulouse, France
e-mail: barbara.pilastre@tesa.prd.fr
J.-Y. Tourneret
e-mail: jean-yves.tourneret@enseeiht.fr
IRIT UMR5505, 118 Route de Narbonne , 31062 Toulouse, France
TéSA, 7 Boulevard de la Gare, 31500 Toulouse, France
L. Boussouf
Airbus Defence and Space, 31 rue des cosmonautes, 31400 Toulouse, France
e-mail: loic.boussouf@airbus.com
S. D’escrivan
Centre National d’Études Spatiales (CNES), 18 avenue Édouard Belin 18 av. Edouard Belin,
31400 Toulouse, France
e-mail: Stephane.d-Escrivan@cnes.fr

https://doi.org/10.1007/978-981-15-9199-0_15
160 B. Pilastre et al.
lies that occurred on several satellites show the interest of the proposed weighting
strategy using external information obtained from the correlation coefficient between
the tested data and its decomposition on the dictionary.
Keywords Machine learning · Spacecraft health monitoring · Anomaly

detection · Sparse decomposition · Dictionary learning
1 Introduction
Spacecraft health monitoring and failure prevention are major issues that can be
handled by checking housekeeping telemetry time series. Housekeeping telemetry
is composed of hundreds to thousands telemetry parameters describing the evolution
over time of physical quantities such as temperature, pressure, voltage. Detecting
anomalies in these parameters jointly can be a complicated task. Thus, a lot of state-
of-the-art telemetry AD methods consider an univariate framework which consists
of detecting anomalies in univariate time series separately via a semi-supervised
learning. The idea is to learn nominal spacecraft behaviours from past telemetry
composed of normal patterns (i.e., data without anomalies) and build a reference
model to which new data can be compared. Then, any deviation from the nominal
behaviour is considered as a fault. In this context, no assumptions are made about
anomalies whose the type is not required for the learning. This represents a real advan-
tage compared to other methods based on supervised learning in which labelled data
are needed. Examples of univariate anomalies affecting telemetry time series are dis-
played in Fig. 1 (top, red boxes). Popular machine learning (ML) methods that have
been investigated in this framework include the one-class support vector machine
[2], nearest neighbour techniques [3, 4] or neural networks [5, 6]. However, these
methods do not account for possible relationships existing between different param-
eters and thus cannot detect multivariate anomalies resulting from changes in the
relationships between several telemetry time series. An example of such multivari-
ate anomaly is displayed in Fig. 1 (bottom, red box). The detection of the abnormal
behaviour in the top time series of this figure (outlined in the red box) will be detected
more easily by using the bottom time series. This kind of situation is often referred
to as “contextual anomaly”, which has been recently investigated in [7–9].
2 Proposed Anomaly Detection Method
2.1 Preprocessing
The proposed AD method first segments the different telemetry times series into
windows of fixed size w as illustrated in Fig. 2. This segmentation step has been
Spacecraft Health Monitoring Using … 161
Fig. 1 Examples of univariate (top) and multivariate (bottom) anomalies
Fig. 2 Segmentation of telemetry into windows

used successfully in many telemetry AD methods [2, 3]. Each resulting matrix is
then transformed into a vector whose first w components correspond to the first time
serie denoted as y1 , whose components w + 1 to 2w correspond to the second time
serie denoted as y2 and so on. This preprocessing creates mixed signals composed
of telemetry time series formed by the different parameters, i.e., y = [yT1 , . . . , yTK ]T
with yk ∈ Rw , k = 1, . . . , K, where K is the number of telemetry parameters and w
is the size of the time window.
2.2 Anomaly Detection Using a Weighted Sparse

Decomposition
Sparse decompositions have received an increasing attention in many signal and

image processing applications. These decompositions were investigated for univari-
ate AD in [10]. The AD strategy introduced in [10] decomposes each multivariate
test signal as the sum of three signals: a nominal signal defined by a sparse represen-
tation into a dictionary of normal patterns, an anomaly signal and an additive noise
such that
y = x + e + b (1)
where y ∈ RN is the multivariate test signal, ∈ RN ×L is a dictionary previously

learned from data describing normal behaviours of spacecraft data, x ∈ RL is a sparse
vector of coefficients, e ∈ RN is a possible anomaly signal (with e = 0 in absence
of anomaly) and b ∈ RN is an additive noise. The proposed multivariate framework
considered in this work and the associated preprocessing divide the multivariate
signals y and e as K blocks associated with the K telemetry parameters, i.e, y =
[yT1 , . . . , yTK ]T and e = [eT1 , . . . , eTK ]T . The resulting sparse decomposition problem
can be expressed as follows [1]
1 K
min y − x − e 22 +ax1 + b ek 2 (2)
x,e 2
k=1

where x1 = Ll=1 |xl | is the 1 norm of the vector x = (x1 , ldots, xL )T and ek 2
is the euclidean norm of the vector ek . Note that (2) considers two distinct sparsity
constraints for the coefficient vector x and the anomaly signal e. This formulation
reflects the fact that a nominal continuous signal can be well approximated by a linear
combination of few atoms of the dictionary (sparsity of x using the 1 norm) and that
anomalies are rare and affect few parameters at the same time (sparsity of e using
the 21 norm). This paper generalizes the AD strategy defined by (2) referred to as
ADDICT [1] by the integration of external information via appropriate weights. The
proposed model for AD in spacecraft telemetry is similar to the well-known group-
lasso model [11] that assigns a weight to each group. In this work we consider K
groups associated with the K telemetry parameters and propose to solve the following
problem
1 K
arg min y − x − e22 + ax1 + b wk ek 2 . (3)
x,e 2
k=1
where wk > 0 is the kth weight associated with the kth parameter. The problem (3)
can be solved with the alternating direction method of multipliers (ADMM) [12] by
adding an auxiliary variable z and the constraint z = x leading to
1 K
arg min C(x, e, z) = y − x − e22 + a z1 + b wk ek 2 s.t. z = x
x,e,z 2
k=1
(4)
where “s.t” means “subject to”. Note that, contrary to Problem (3), the first and
second terms of (4) are decoupled, which allows an easier estimation of the vector
x. The ADMM algorithm associated with (4) minimizes the following augmented
Lagrangian
μ
LA (x, z, e, m, μ) = C(x, e, z) + mT (z − x) + z − x22 (5)
2
where m is a vector containing Lagrange multipliers and μ is a regularization param-
eter controlling the level of deviation between z and x. The ADMM algorithm itera-
tively updates x, z, e and m as follows
Update of x
x is classically updated as follows

1 μk k
xk+1
= arg min y − x − e 2 + m (z − x) +
k 2 k k
z − x2 .
2
(6)
x 2 2
Simple algebra leads to
xk+1 = (T + μk I )−1 [T (y − ek ) + mk + μk zk ].

(7)
Update of z
The update of z is defined as

μk
zk+1 = arg min az1 + (mk )T (z − xk+1 ) +
z − xk+1 22 . (8)
z 2
Thesolution of (8) is given by the element-wise soft thresholding operator

zk+1 =
Sγ k xk+1 − μk mk with γ k = μk , where the thresholding operator Sγ (u) is defined
1 a
by
⎧
⎪
⎨u(n) − γ if u(n) > γ
Sγ (u) = 0 if |u(n)| ≤ γ (9)
⎪
⎩
u(n) + γ if u(n) < −γ
where u(n) is the nth component of u.

Updates of m and μ
The updates of m and μ are defined as
k+1 = mk + μk (zk+1 − xk+1 ) and

m μk+1 = ρμk with ρ > 1.
Update of e
Due to the reweighted scheme investigated in this paper, the update of the anomaly
vector e changes from the non-weighted scheme studied in [1], leading to

1 K

e = arg min y − x − e2 + b
2
wk ek 2 . (10)
e 2
k=1
The problem (10) can be divided into the following K independent problems

1
ek = arg min
hk − ek 22 + bwk ek 2 . (11)
ek 2
where h = y − x and hk is the kth element of h associated with the kth telemetry
time series for k = 1, . . . , K. The solution of (11) is given by the group-shrinkage
e = Tb (h) defined by
operator
hk 2 −bwk
hk 2
hk if hk 2 > bwk
[Tb (h)]k = (12)
0 otherwise
Note that the estimation of the anomaly signal e directly depends on the residual h
resulting from the sparse decomposition of the test signal in the dictionary of normal
patterns. Moreover, a value of the weight higher than 1 tends to reduce AD whereas
a value lower than 1 promotes AD (for wk ≥ 1, the first equality is more difficult to
satisfy, which leads to
e = 0 more frequently).
2.3 Anomaly Detection Strategy

The proposed AD strategy referred to as W-ADDICT (for Weighted Anomaly Detec-
tion using a sparse decomposition on a DICTionary) is defined by comparing the
estimated anomaly signal e denoted as e to an appropriate threshold, i.e.,
e22 > SPFA

Anomaly detected if (13)
where SPFA is the threshold depending on the desired compromise between false alarm
and good detection rates. This threshold can be determined using receiver operating
characteristics (ROCs) if a ground-truth is available or using the user’s experience.

ROCs express the probability of detection PD as a function of the probability of false
alarm PFA. In this work, these probabilities were computed using ground-truth time
series and the threshold SPFA was determined from the pair (PFA, PD) located the
closest to the ideal point (0, 1).
2.4 Weight Updating Using a Correlation Coefficient

Weighting the square norms ek 2 allows external information from expert knowl-
edge to be considered. This paper investigates the usefulness of the correlation coef-
ficient between the test signal y and its sparse decomposition x, assuming that
the smaller the value of the correlation coefficient, the higher the probability that y
is affected by an anomaly. As explained before, weights interfere in the detection
process in such a way that a small value of the weight promotes the detection of
an anomaly. Conversely, a large value of the weight tends to reduce the number of
detected anomalies. In order to respect this idea, we propose to build the weights
with the following function
fα : [−1, 1] → R+
(14)
ck → wk = ((1+α)−c
1
k)
2
which is an increasing function of ck where ck is the correlation coefficient between

the test signal and its reconstruction associated with the kth parameter and α is an
appropriate hyperparameter fixed by the user. The choice of this function is moti-
vated by its simplicity since it only depends on one parameter α. Note that ck = α
corresponds to wk = 1, i.e., to the unweighted model. For ck < α, the kth weight
satisfies wk < 1, which favours AD. Conversely, when ck > α, the kth weight is such
that wk > 1, which limits AD. Our experiments have shown that α can be tuned in
the interval ] − 1, +1[ without loss of generality.
The AD method (2) and its weighted version (3) (with weights built using the corre-
lation coefficient as proposed in Sect. 2.4) have been evaluated on a representative
anomaly dataset composed of K = 7 telemetry parameters with an available ground-
truth. The hyperparameters a and b were tuned by cross validation using ground-truth
and ROC curves. In all the experiments, the dictionary was composed of L = 2000
atoms learnt using the K-SVD algorithm [13] with two months of nominal telemetry
(without anomalies), which represents approximately 5000 mixed training signals
obtained after applying the preprocessing described in Sect. 2.1 with the param-
eter w = 50 (i.e., the signal length is N = 350). The anomaly database was built
using 18 days of telemetry, i.e., was composed of 1000 signals, 82 of them being
affected by anomalies. Note that the 82 anomalies were divided into 5 anomaly peri-
ods with various durations, as illustrated in the three examples of anomalies that are
Fig. 3 Anomaly scores obtained with unweighted algorithm (2) (left) and with the weighted one
(3) (right) with ground-truth marked by red background and the detection threshold indicated by
the red horizontal lines
displayed in Fig. 1. Note also that a specific attention was devoted to the construc-
tion of a heterogeneous database (containing discrete and continuous time-series,
with anomalies of different amplitudes and durations, …). Finally, it is important to
note that the majority of anomalies are actual anomalies that have been observed in
operated satellites.
Figure 3 displays the anomaly scores e22 returned by the unweighted AD method
(2) (left) and the weighted one (3) (right) for each of the 1000 test signals of the
anomaly dataset. Note that the actual anomalies are marked by red backgrounds
whereas the red horizontal lines indicate the detection thresholds (determined using
cross validation as described in Sect. 2.3). As can be observed, the score exceeds
threshold almost exclusively in presence of anomaly, meaning that anomalies are
well detected and that few false alarms are returned. Note that the contextual corre-
sponding to the last anomaly period of the ground-truth is also well detected. These
first results show the interest of using sparse decompositions for anomaly detection
in spacecraft telemetry. Furthermore, they show that the weighted model allows the
number of false alarms to be reduced. For example, the test signals #237, #535 and
#921 indicated by blue arrows are false alarms for the non-weighted scheme whereas
they are not detected by the weighted algorithm.
Quantitative results in terms of probability of detection, probability of false alarm
and area under the curve (AUC) are reported in Table 1. These results confirm that
the weighted model reduces the number of false alarms for a fixed probability of
detection (for PD = 89%, the PFA decreases from PFA = 5.2% for the unweighted
model to PFA = 2.2% for the weighted one). Note that false alarm reduction is
very interesting for spacecraft monitoring applications because each detection of the
algorithm is examined by experts, which is potentially a very time consuming task.
Note also that the detection threshold might be optimized using other criteria, e.g.,
allowing a smaller value of PD to decrease the number of false alarms.
Table 1 Probability of detection, probability of false alarm and area under curves (AUC) for the
unweighted model (2) and the proposed weighted model (3)
Method Threshold PD (%) PFA (%) AUC
Unweighted model (2) 4.9 89 5.23 0.93
Weighted model (3) 4.3 89 2.18 0.96
4 Conclusion
This paper showed the interest of sparse decompositions for anomaly detection in
multivariate housekeeping telemetry time series. In particular, external information
provided by the user can be incorporated in the anomaly detection via appropriate
weights. Our first results indicated that weights built from the correlation coefficient
between the test signal and its sparse decomposition in a dictionary of normal patterns
allowed a significant reduction of the false alarm probability.
For future work, a scaling-up is needed to evaluate the anomaly detection method
in an operating context including hundreds to thousands telemetry time series. Con-
cerning the weighted model, it would be interesting to build weights using other
information than the correlation coefficient. Finally, we think that the model could
be updated sequentially using expert feedback in order to improve detection, e.g., in
the case of anomalies evolving with time. This opens the way for many works related
to online or sequential anomaly detection, which could be useful for spacecraft health
monitoring.
Acknowledgements The authors would like to thank Pierre-Baptiste Lambert from CNES and Clé-
mentine Barreyre from Airbus Defence and Space for fruitful discussions about anomaly detection
in spacecraft telemetry. This work was supported by CNES and Airbus Defence and Space.
References
1. Pilastre B et al (2019) Anomaly detection in mixed telemetry data using a sparse representation
and dictionary learning. Sig Proc to appear 2019
2. Fuertes S et al (2016) Improving spacecraft health monitoring with automatic anomaly detec-
tion techniques. In: Proceeding international conference space operations (SpaceOps’2016),
Daejeon, South Korea
3. Martínez-Heras J-A, Donati A, Kirksch M, Schmidt F (2012) New telemetry monitoring
paradigm with novelty detection. In: Proceeding international conference space operations
(SpaceOps’2012), Stockholm, Sweden
4. O’Meara C et al (2016) Athmos: automated telemetry health monitoring system at GSOC using
outlier detection and supervised machine learning. In: Proceeding international conferecne
space operations (SpaceOps’2016), Daejeon, South Korea
5. O’Meara C et al (2018) Applications of deep learning neural networks to satellite teleme-
try monitoring. In: Proceeding international conference space operations (SpaceOps’2018),
Marseille, France
6. Hundman K et al (2018) Detecting spacecraft anomalies using LSTMs and nonparametric

dynamic thresholding. Proceeding international conference knowledge data mining (KDD’18).
United Kingdom, London, pp 387–395
7. Takeishi N et al (2014) Anomaly detection from multivariate times-series with sparse represen-
tation. In: Proceeding IEEE international conferecne system man and cybernetics, San Siego,
CA, USA
8. Yairi T et al (2017) A data-driven health monitoring method for satellite housekeeping data
based on pobabilistic clustering and dimensionality reduction. IEEE Trans Aerosp Electron
Syst 53(3):1384–1401
9. Barreyre C (2018) Statistiques en grande dimension pour la détection d’anomalies dans les don-
nées fonctionnelles issues des satellites. Ph.D. dissertation, Université de Toulouse, Toulouse,
France
10. Adler A et al (2015) Sparse coding with anomaly detection. J Sign Process Syst 79(2):179–188
11. Yuan M et al (2006) Model selection and estimation in regression with grouped variables. J
Roy Statist Soc Series B (Methodological) 68(1):49–67
12. Boyd S et al (2010) Distributed optimization and statistical learning via alternating direction
method of multipliers. Found Trends Mach Learn 3(1):1–222
13. Aharon M et al (2006) K-SVD: an algorithm for designing overcomplete dictionaries for sparse
decomposition. IEEE Trans Sign Process 54(11):4311–4322
When Anomalies Meet Aero-elasticity:
An Aerospace Industry Illustration
of Fault Detection Challenges
P. Goupil, J.-Y. Tourneret, S. Urbano, and E. Chaumette
Abstract This paper aims at describing some industrial aerospace applications

requiring anomaly detection. After introducing the industrial context and needs, the
paper investigates two examples of oscillation/vibration detection associated with
aero-elastic phenomena: (i) the first example related to aircraft structural design
optimization is characterized by very stringent constraints on detection time; (ii)
the second example linked to equipment condition monitoring has more relaxed
constraints in terms of detection requirements.
Keywords Aerospace · Civil aircraft · Flight control system · Oscillation

detection · Vibration diagnosis
1 Introduction
The flight control system of an aircraft is an avionic system used to control its attitude,
trajectory and speed. In its electrical version, termed electrical flight control system
(EFCS, a.k.a. as fly-by-wire), it consists of all the components between the pilot
inputs (in the cockpit) and the movable surfaces called control surfaces, including
these two elements. In more details, it comprises wires, numerical busses, probes
and sensors, actuators, power sources (e.g. hydraulic) and flight control computers
managing the whole system. EFCS was first developed in the 1960s by Aerospatiale
and installed on Concorde (1969) as an analogue system. Much later, digital EFCS
was introduced first on the Airbus A310 (1982, spoilers, slats and flaps only) and
was then generalized on all control surfaces (3-axis) on the A320 (1987). This is
P. Goupil (B) · S. Urbano

AIRBUS, Aircraft Control, 316 route de Bayonne, Cedex 09, 31060 Toulouse, France
e-mail: philippe.goupil@airbus.com
J.-Y. Tourneret
INP/ENSEEIHT-IRIT-TéSA, University of Toulouse, Toulouse, France
E. Chaumette
Isae-Supaero, University of Toulouse, Toulouse, France

https://doi.org/10.1007/978-981-15-9199-0_16
170 P. Goupil et al.
now generalized to all Airbus aircraft (e.g. A350, A380). EFCS main advantages
include sophisticated control of the aircraft, flight envelope protection functions,
pilot workload alleviation and weight saving [1, 2]. EFCS is designed to meet very
stringent requirements in terms of dependability and availability. The detection of
all related anomalies (e.g. faults, failures or any abnormal behaviour) is therefore a
very important point to be considered in the aircraft design.
In this context, this paper details two particular scenarios of anomaly detection,
both related to vibrations and their specific constraints: the oscillatory failure cases
(OFC) and the limit cycle oscillation (LCO)-induced vibrations, both affecting the
aircraft during flight due to coupling with aero-elastic phenomena. However, these
scenarios have different origins and consequences as well as diverse requirements
in terms of fault detection. OFC leads to very stringent constraints on detection time
and is related to aircraft structural design optimization. Conversely, LCO has more
relaxed constraints in terms of detection requirements and is linked to equipment
condition monitoring. Both scenarios must be compliant with severe robustness rules.
Dedicated solutions must be eligible to real-time limitations.
The paper is organized as follows: Sect. 2 is devoted to a review of the main real-
time constraints with the goal to make the reader sensitive to the strong industrial
limitations that must be taken into account from the early phases of the solution
development. Section 3 details the OFC context and describes the solutions currently
used and certified on in-service aircraft as well as current research study. Section 4 is
dedicated to LCO-induced vibrations and presents related ongoing research activities.
Finally, Sect. 5 presents some conclusions and perspectives.
2 Industrial Context and Constraints
EFCS development is an arduous process because of its criticality in the sense that any
severe malfunction could potentially affect the flight or prevent its correct achieve-
ment in the optimum configuration (e.g. aircraft performance, fuel consumption,
…). Moreover, most of the design requirements come directly from the aviation
authorities (e.g. the European Aviation Safety Agency) and taking into account these
regulations leads to develop flight control computer software according to the Design
Assurance Level A, which is the most stringent requirement for a real-time on-board
software. In this context, getting the aircraft certified means demonstrating that it
complies with the regulations mainly thanks to a lot of verification and validation
activities which can be summarized as follows: (i) Verification: are we developing
the system right ?; (ii) Validation: are we developing the right system? In practice,
and especially for complex systems like EFCS, these activities involve a big number
of tests of different natures (simulators, system bench, flight tests…) with a total
duration longer than the development phase. Non-standard tests could be required
such as the algorithm long-term behaviour in steady phase (to check that there is no
divergence). Not only is the average performance required but also the performance
in extreme and unusual conditions, at the operational boundaries.
When Anomalies Meet Aero-elasticity: An Aerospace … 171
Focusing more on anomaly detection, the general industrial state of practice [2]
provides high levels of hardware redundancy (sensors, computers, wires, etc.…) in
order to perform consistency tests, cross-checks and built-in-tests of various sophis-
tications. The flight control computers, hosting the EFCS software and especially
the anomaly detection algorithms, are multirate time-triggered, which means that
they are able to manage different sampling frequencies. It also implies potentially
mixing asynchronous signals with different sampling rates. Due to the life cycle of
a civil aircraft, the flight control computers are equipped with CPU with limited
computing power (e.g. the A320 family has been designed from the 80s and is still
flying). This means that it is tricky to embed complex and non-deterministic algo-
rithms such as optimization routines. The software is generally specified thanks to a
graphical tool describing accurately all the functions to be ensured. A limited set of
graphical symbols (e.g. adder, filter, integrator, lookup tables, very much in the style
of Simulink blocks) is used to describe each part of the algorithm. The limitations
imposed by such a process guarantee on the one hand that the code has a complexity
level which allows its implementation on the actual computer but on the other hand it
makes the designer task more complex. For instance, it is not possible to implement
easily vector or matrix computation (no dedicated graphical symbols) and thus the
corresponding calculation must be broken down to a series of scalar multiplication
and addition. Note that there is no state-space representation and no direct way to
implement conditional loops. It must also be noticed that the mastering of any innova-
tive algorithm, which is more complex than the current state of practice is crucial for
applications by non-experts. Particularly, the high-level tuning of hyper-parameters
must be clearly explained: how to tune the parameters on a different aircraft model
or on a different sensor?
3 Oscillatory Failure Cases (OFC)
Oscillatory failure case (OFC) is the term given to a class of failures occurring in
the actuator servo loop that cause unwanted oscillations of the control surface [3].
OFC can result from the faulty behaviour of an electronic component or an actuator
mechanical failure. It will take the form of a sinusoidal signal propagating through the
control loop leading to the control surface vibration. The current Airbus servo loop
principle is depicted in Fig. 1. The flight control computer has a dual channel structure
where the so-called COM (command) channel is dedicated mainly to the flight control
law computation and to the control surface servo loop. The so-called MON channel
(monitoring) is primarily dedicated to the monitoring of all EFCS components. The
OFC effects can be modelled as the injection of a periodic signal at two specific
points of the control loop: the analogic output and the analogic input (see Fig. 1).
As the actuator acts as a low-pass filter, the OFC phenomenon is generally limited
to approximately [0–12] Hz. An OFC has the main effect to generate loads on the
structure when located within the actuator bandwidth. Coupled with the aero-elastic
behaviour of the aircraft, an OFC may lead to unacceptable high loads. The OFC
Fig. 1 Airbus control surface servo loop
amplitude must be contained by system design within an envelope function of the

frequency. The capability to detect these failures is very important because it has an
impact on the structural design of the aircraft. Moreover, the load envelope constraints
must be respected. More precisely, if an OFC of given amplitude cannot be detected
and passivated, this amplitude must be considered for load computations. The result
of this computation can lead to reinforce the structure in case of load exceeding.
In order to avoid reinforcing the structure and consequently to save weight, low-
amplitude OFC must be detectable early. Thus, there is a link between anomaly
detection and aircraft weight saving which is not obvious at first sight. Concretely,
for structure-related system objectives, it is necessary to detect OFC beyond a given
amplitude in a given number of periods, whatever the unknown OFC frequency. The
latest OFC detection strategy used on board Airbus aircraft is based on analytical
redundancy. It is a classical approach in the field of fault detection, where the actual
actuator rod position, measured by a dedicated sensor, is compared to an estimated
position produced by an actuator model fed by the command signal. This residual
generation is followed by a decision-making step to confirm the oscillation, typically
by counting successive and alternate crossings of a given threshold. This kind of
technique has been chosen as on the one hand actuator models were already widely
used and mastered, and on the other hand OFC signal appears “as is” on the residual
(no transformation) which eases the threshold tuning. The frequency range to monitor
is broad and as the computing power is limited, it is not possible to use classical
spectral analysis. OFC detection in the temporal domain alleviates the computational
burden and automatically adjusts to all frequencies. The estimated position p̂ is the

result of a discrete integration of the estimated velocity ṗ expressed by the following

equation:

F (t)+Fdamp (t)

P(t) − aero
ṗ (t) = V0 (t) S
(1)
Pref
where P is the hydraulic pressure delivered in input of the actuator, Faero repre-

2
sents the aerodynamic forces applied on the control surface, Fdamp = kd ṗ is the
servo control load of the adjacent actuator in damping mode (in case of a dual
active/passive scheme, i.e. one active actuator and one passive actuator in stand-by
mode in case of a reconfiguration), S is the surface area of the actuator piston, Pref
is the differential pressure corresponding to the maximum rod speed and V0 (t) is
the maximum speed of one actuator without any load. In a first simplified version,
the parameter vector θ = (P, Faero , kd )T is assumed to be constant (fixed to its
most probable value) to reduce the computational complexity. In a second version, to
achieve better performance, i.e. an earlier detection of lower OFC amplitudes, θ can
be estimated by means of an extended Kalman filter to reduce modelling errors [4].
Both versions have been extensively validated, certified and are flying on in-service
aircraft.
However, these model-based approaches have the drawback of requiring a model
for each type of actuator (e.g. three kinds of actuators in the A380) and the corre-
sponding model must be tuned on each aircraft type (e.g. A380 and A350). This is
why more generic techniques have been investigated like data-driven approaches with
the advantages of being less dependent of the system. The main idea is to monitor
the similarity between the input (command signal) and the output (actuator rod or
control surface position) of the system under study (the control surface servo loop,
Fig. 1): in a faulty situation there is a loss of similarity between the input and output
signals, compared to the fault-free case. One interesting idea is a modified version
of the so-called complex invariant distance (CID), proposed in [5]. It consists of
computing a similarity index by weighting the Euclidean distance between input and
output by a correcting factor representing the curvilinear distance of these signals
[6]:
C I D(x, y) = E D(x, y)C D(x, y) (2)
where x is the command, y the actuator/control surface position, ED is the

Euclidean distance and CD the curvilinear distance defined as:
⎧
⎪
⎪
N
⎪
⎨ E D(x, y) = (xi − yi−τ )2
i=1 (3)
⎪ N−1
⎪
⎪
⎩ C D(x, y) = C D(y) ; C D(x) =
C D(x)
(xi − xi+1 ) + TS
2 2
i=1
where TS is the sampling time and τ is the drag error, i.e. the nominal delay
between the command and the position. The index C I D(x, y) is compared to a
given threshold and OFC is detected when it stays greater than the threshold during
a given time. Figure 2 shows an example displaying the command x and the position
y (top figure) and an OFC occuring at t = 15 s. At the bottom of the figure, it is
clear that ED is not high enough to detect the fault (same value before and after the
fault occurrence), especially due to an artificial bias between x and y. However, the
Fig. 2 Example of CID

index variations for OFC
detection
correcting factor CD is more sensitive to the fault and the global CID index allows
us to discriminate the fault-free and faulty situations.
4 Limit Cycle Oscillations (LCO)
The LCO phenomenon consists of an unwanted sustained oscillation of a control

surface due to the combined effect of aero-elastic behaviour and an increased level
of mechanical free-play in the elements that connect the control surface to the aero-
dynamic surface. Under some circumstances, it can lead to slight airframe vibrations,
which can impact the passengers and the crew comfort. The current state of practice
for LCO-vibration detection relies mainly on the pilot sensitivity and on long-term
scheduled maintenance. Once the vibrations have been detected by the crew, a main-
tenance task is triggered but the troubleshooting is tricky. In the works presented in
this paper, two goals are pursued: (i) to estimate the duration and amplitude of the
vibrations in order to compute a condition monitoring index to be communicated to
the operating airline and (ii) to ease the vibration troubleshooting in order to facil-
itate the maintenance tasks (i.e. to find directly the location of the vibration). The
challenge is to build a condition monitoring index to be communicated to the airline
to trigger a specific and accurate maintenance task and avoid any unscheduled work
that could lead to flight delays or cancellation. Thanks to a more accurate vibra-
tion troubleshooting, the maintenance task will be shortened and will not impact the
airline in terms of aircraft immobilization. The control aspect is not considered here,
i.e. the possibility to cancel the vibrations by injecting a dedicated command is not
investigated.
The main idea investigated is to monitor individually each actuator using the
already existing information (Fig. 1) in order to detect and diagnose the vibrations and
to de facto isolate the faulty component. Using the same analytical redundancy tech-
nique as for OFC is not possible here as there is no simple model of the system under
study (a control surface linked to an aerodynamic surface leading to the airframe
vibrations). Using a distance-based solution is also impossible as the vibration ampli-
tudes to detect are very small, far below the OFC objectives and the performance
observed in the OFC context. Using the same notation as for the OFC, a simple model
of the actuator dynamic is:
y(n) = λx(n − τ ) (4)
where n is the discrete time (n = 0, . . . , N −1), λ is an attenuation or amplification

factor (e.g. due to the aerodynamic forces applied on the control surface) and τ is
the aforementioned drag error. As the problem to solve can be formalized as the
detection of deterministic signals with unknown parameters, assuming the simple
model (4), the following binary hypothesis testing problem can be investigated:

H0 : y(n) = λx(n − τ ) + w(n)

(5)
H1 : y(n) = λx(n − τ ) + A cos(2π f n + φ) + w(n)

with w(n) ∼ N 0, σ 2 . Under the null hypothesis H0 , the control surface posi-
tion y(n) is simply a delayed and amplified/attenuated version of the command x(n)
corrupted by some additive Gaussian noise w(n). Under the alternative hypothesis
H1 , the control surface position is polluted by the vibration modelled as a sinusoid
of a given amplitude, frequency and phase. In the binary hypothesis testing problem
(5), the optimal decision rule is based on the probability distribution of the observa-
tions under both hypotheses and the a priori probability of each hypothesis (Bayes
detector). In absence of any a priori probability for the two hypotheses, which is the
case here, the most common detector is the likelihood ratio test (LRT) resulting from
the Neyman-Pearson test [7]. Unfortunately, optimal statistical tests such as the LRT
cannot always be implemented since there are often some unknown parameters in the
observation model, leading to the so-called composite hypothesis testing problem
[8]. A very common approach in this situation is to replace the unknown parame-
ters in the LRT by their maximum likelihood estimators (MLE), following the ideas
of the generalized likelihood ratio test (GLRT) [8]. The GLRT requires to estimate
the unknown parameters under both hypothesis: {λ, τ } under H0 and {λ, τ, A, f, φ}
under H1 , using the maximum likelihood principle. All the details regarding these
estimators are available in [9], where the test and its theoretical performance are
derived and validated. Once the GLRT test has been established, the detection rule
consists of comparing the GLRT statistic to a threshold that can be derived from the
desired false alarm rate [9] (resulting from strong industrial constraints). The GLRT
requires to estimate the parameters of the model (5), i.e. the LCO amplitude Â and

duration T . Moreover, the energy generated by the vibrations can be estimated as:
Ê LC O = ∫ Â2 dt

(6)
T
For diagnostic, it is reasonable to assume that only a fraction of the real energy
E LC O is dissipated through wear process that increases the free-play level, so a simple
index I (for a given flight F) can be:
I (F) = kw ∫ Â2F dt

(7)
T F
where kw is a wear severity coefficient. This definition can be recursively extended

to n flights:
I (Fn ) = λ f I (Fn−1 ) + kw
∫ Â2Fn dt (8)
T Fn
where λ f is a forgetting factor adjusted to reduce the undesired effects of estima-

tion errors. Figure 3 shows an example of the GLRT statistics on a flight test data for
the elevator control surface where LCO signals have been injected artificially. On
the top, the command signal x in blue is called “com”, the control surface position
y in red is called “RVDT”. A first vibration of a very small amplitude A1 (not given
for confidentiality reasons) is injected at t = T1 during 30s and a second amplitude
A2 = 2*A1 is injected at t = T2 during 20 s. A1 and A2 are so small that they
are not visible. The command and corresponding control surface positions are very
close as the drag error is small. The two graphs at the bottom of Fig. 3 show the
estimated LCO vibration duration and the corresponding estimated amplitudes. It
can be observed that the estimated amplitudes are very close to the ground truth
displayed in blue. Figure 3 also shows the accumulated duration (middle curve). The
global accuracy is correct even if the first duration is slightly underestimated and the
second one is marginally overestimated.
5 Conclusions
This paper deals with two aerospace industrial challenges related to vibration
detection with the goal to exemplify concrete applications and their engineering
constraints. In the first challenge, the vibrations must be detected very quickly in real
time to trigger a hardware reconfiguration (i.e. using the adjacent stand-by actuator).
The second challenge is dedicated to detecting and estimating very small vibrations
due to mechanical free-play resulting from a wearing effect. Note that in this second
case, the problem is not to detect the vibrations very swiftly but rather to detect
incipient faults, before they are felt by the pilots. These two different challenges
illustrate the variety of engineering applications. For future works, some appealing
*T1 = 500s
A1 deg / 30sec
*T2 = 1000s
A2=2*A1 deg / 20sec
50
30
A2
A1
Fig. 3 Example of the GLRT-based algorithm on real flight data with vibrations artificially injected
avenues are worth to be mentioned. Several techniques have been considered for
OFC during several decades. Future work could be dedicated to determine adaptive
thresholds and consider different decision strategies. For LCO, it could be helpful to
complement the vibration diagnosis by a free-play detection and diagnosis strategy,
since mechanical free-play is generally the cause of the vibrations.
References
1. Brière D, Favre C, Traverse P (1995) A family of fault-tolerant systems: electrical flight controls,
from A320/330/340 to future military transport aircraft. Micoprocessors Microsyst 19(2)
2. Goupil P (2011) AIRBUS state of the art and practices on FDI and FTC in flight control system.
Control Eng Pract 19(6):524–539
3. Goupil P (2010) Oscillatory failure case detection in the A380 electrical flight control system
by analytical redundancy. Control Eng Pract 18(9):1110–1119
4. Lavigne L, Zolghadri A, Goupil P, Simon P (2011) A model-based technique for early and robust
detection of oscillatory failure case in A380 actuators. Int J Control Autom Syst 9(1):42–49
5. Batista GEAPA, Keogh EJ, Tataw OM, De Souza VMA (2014) CID: an efficient complexity-
invariant distance for time series. Data Min Knowl Disc 28(3):634–669
6. Urbano S, Chaumette E, Goupil P, Tourneret J-Y (2017) A data-driven approach for actu-
ator servo loop failure detection. In: Proceedings of the 20th IFAC World Congress, IFAC-
PapersOnLine, vol 50, no 1, pp 13544–13549
7. Kay SM (1998) Fundamentals of statistical signal processing. Detection theory, vol 2. Prentice
Hall
8. Van Trees HL (2004) Detection, estimation, and modulation theory. John Wiley & Sons
9. Urbano S, Chaumette E, Goupil P, Tourneret J-Y (2018) Aircraft vibration detection and diag-
nosis for predictive maintenance using a GLR test. In: Proceedings of 10th IFAC Symposium on
Fault Detection, Supervision and Safety for Technical Processes (Safeprocess’2018), Warsaw,
Poland, Aug. 2018, pp 1030–1036
Anomaly Detection with Discrete
Minimax Classifier for Imbalanced
Dataset or Uncertain Class Proportions
Cyprien Gilet and Lionel Fillatre
Abstract Supervised classification is becoming a useful tool in condition monitor-

ing. For example, from a set of measured features (temperature, vibration, etc.), we
might be interested in predicting machine failures. Given some costs of class misclas-
sifications and a set of labeled learning samples, this task aims to fit a decision rule
by minimizing the empirical risk of misclassifications. However, learning a classifier
when the class proportions of the training set are uncertain, and might differ from the
unknown state of nature, may increase the misclassification risk when classifying
some test samples. This drawback can also occur when dealing with imbalanced
datasets, which is common in condition monitoring. To make a decision rule robust
with respect to the class proportions, a solution is to learn a decision rule which min-
imizes the maximum class-conditional risk. Such a decision rule is called a minimax
classifier. This paper studies the minimax classifier for classifying discrete or dis-
cretized features between several classes. Our algorithm is applied to a real condition
monitoring database.
Keywords Minimax classifier · Uncertain class proportions · Imbalanced dataset
1 Introduction
Context: This paper studies a new anomaly detection method based on supervised
classification. For example, from a set of measured features (temperature, vibration,
etc.), we might be interested in predicting potential machine failures. However, very
often we have to face three difficulties when dealing with such databases. Firstly,
The authors thank the Provence-Alpes-Côte d’Azur region for its financial support.
C. Gilet (B) · L. Fillatre

University of Côte d’Azur, CNRS,I3S Laboratory,Sophia-Antipolis, France
e-mail: gilet@i3s.unice.fr
L. Fillatre
e-mail: lionel.fillatre@i3s.unice.fr

https://doi.org/10.1007/978-981-15-9199-0_17
180 C. Gilet and L. Fillatre
both numerical and categorical features are available in the database, and it might
be necessary to preprocess the features in order to optimize predictions. In such
cases, an effective approach for performing supervised classification is to discretize
numerical features [1] in order to work only with discrete features. Secondly, costs of
misclassifications are given by experts, and have to be considered fixed. These costs
aim to penalize differently the classifications errors which may bring more or less
important consequences. Thirdly, the class proportions are imbalanced. For example
the proportion of components presenting a failure is in general much lower than the
proportion of components without failure. This problem leads most of the classifiers
to neglect the classes containing fewest samples. The risk of misclassifications can
also increase when the class proportions are uncertain, or can change in time, which
is also possible when we have no idea of the state of nature. In this paper we propose
a new supervised classification algorithm which aims to address all these difficulties.
Notations and problem statement: The task of supervised classification aims to mini-
mize the empirical risk of misclassifications from a finite set of labeled learning sam-
ples D = {(Yi , X i ), i ∈ I}. Let K ≥ 2 be the number of classes, Y := {1, . . . , K }
the set of observed labels, Ŷ = Y the set of predicted labels, X the space where all
the features are defined, and m the number of samples available in the training set.
Let us denote Yi the random variable characterizing the label of the sample i, and
X i = [X i1 , . . . , X id ] ∈ X the random vector containing all the d features describ-
ing the sample i. Let us define = {δ : X → Ŷ} the set of all the classifiers and
let us consider a decision rule δ ∈ . Finally, let us consider the matrix of costs
K ×K
C ∈ R+ such that, for all (k, l) ∈ Y × Ŷ, Ck,l is the cost of assigning the label
Ŷi = l to the sample i while its true class Yi is k. The empirical risk of misclassifi-
cations r̂ (δ) = m1 i∈I CYi ,δ(X i ) of δ can be written as [2]:

r̂ (δ) = π̂k R̂k (δ) , (1)
k∈Y
where, for all k ∈ Y, π̂k is the proportion of samples from the class k, and

R̂k (δ) = Ck,l P̂(δ(X i ) = l | Yi = k), (2)
l∈Ŷ
with P̂(δ(X i ) = l | Yi = k) the probability of predicting the label l when the class
is k.
Benefits of the minimax classifier: Let us denote δπ̂ the classifier δ ∈ fitted from the
training set D satisfying the class proportions π̂ = [π̂1 , . . . , π̂ K ]. This classifier is
then used to predict the labels Yi of new test samples from their associated features
X i ∈ X . Let us consider the test set D = {(Yi , X i ), i ∈ I } composed of m test
samples satisfying the class proportions π = [π1 , . . . , π K ]. The risk associated to the
classifier
δπ̂ and to the class proportions π is then noted and defined by r̂ (π , δπ̂ ) =

k∈Y k k (δπ̂ ). Hence, the risk evolves linearly when the class proportions π differ
π R̂
Anomaly Detection with Discrete Minimax Classifier … 181
5
Training set: average risk
Validation set: average risk
4
0
040 200 400 600
Fig. 1 Scania trucks database: Left. Selection of the number of categories T using the
kmeans quantizer. Right. Expectation of the evolution of the risks r̂ (π ,δ) for each classifier
δ ∈ {δπ̂LR , δπ̂SVM , δπ̂RF , δπ̂B , δπ̄B } as a function of π1 . Since K = 2, r̂ π , δπ̂ can be rewritten as

r̂ π , δπ̂ = π1 R̂1 (δπ̂ ) + π2 R̂2 (δπ̂ ) = π1 R̂1 (δπ̂ ) − R̂2 (δπ̂ ) + R̂2 (δπ̂ ). It is then clear that

r̂ π , δπ̂ is a linear function of π1 . Finally, let us remind that V (π ) = r̂ π , δπB is the estimation
of the minimum risk we could expect when considering the class proportions π
from π̂ . This fact is illustrated in Fig. 1, right. As described in [2], a solution to make
a decision rule robust to differences of the class proportions between D and D is to
minimize maxk∈Y R̂k (δπ̂ ) with respect to all possible changes of the class proportions.
Let us denote S the K -dimensional probabilistic simplex. This minimax problem can
be written as
δπ̄B = argmin max r̂ (π, δπ ). (3)
δ∈ π∈S
As explained in [2], the minimax classifier δπ̄B tends to equalize as more as possible
the class-conditional risks, and thus to consider fairly all classes.
Related works: A common approach to deal with imbalanced datasets is to balance
the data by resampling the training set. However, this approach may increase the mis-
classification risk when classifying some test samples which are imbalanced. Another
common approach is the cost sensitive learning [3, 4] which aims to optimize the
cost of class misclassifications in order to counterbalance the number of occurrences
of each class. But, as previously mentioned, in our context we need to consider these
costs fixed since they are given by experts. The task of learning the class-proportions
which maximize the minimum empirical risk was already studied in past years. One
of the closest work related to our context is [5]. The authors proposed an interesting
fixed-point algorithm based on generalized entropy and strict sense Bayesian loss
functions. This approach alternates a resampling step of the learning set with an eval-
uation step of the class-conditional risk, and it leads to estimate the least-favorable
priors. However, the fixed-point algorithm needs the minimax rule to be an equalizer
rule. We can show that this assumption is in general not satisfied when considering
discrete features. Moreover, in practice, re-sampling the data to precisely consider
some class proportions is not always possible due to the size of the smallest classes
in the training set.
Contributions: We consider here discrete or discretized features (X ⊂ Nd ), K ≥ 2,

K ×K
and a given matrix of costs C ∈ R+ fixed. We show that under these conditions
we can calculate the discrete Bayes classifier δπ̂B which minimizes the empirical
risk over the training set. Then we propose an algorithm to compute the minimax
classifier (3). The convergence is guaranteed. We finally applied our algorithm to
the real condition monitoring database [6] in order to illustrate the robustness of our
minimax classifier.
2 Computation of the Minimax Classifier
In order to manipulate both numerical and categorical features, we discretized the

data with the k-means quantizer. In other words, each feature vector X i ∈ Rd com-
posed of all the features is quantized with the centroid index of the centroid closest
to it, i.e., X i → Q(X i ) = xt where Q : Rd → {x1 , . . . , x T } denotes the k-means
quantizer with T centroids, and t is the index of the centroid of the cluster in which
X i belongs to. Hence, the set of features becomes X = {x1 , . . . , x T }. Let us note
T := {1, . . . , T }, Ik := {i
∈ I : Yi = k} the set of labeled learning samples belong-
ing to class k, and m k := i∈I 1{Yi =k} the number samples in class k. Let
1
p̂kt := 1{X i =xt } (4)
m k i∈I
k
be the empirical probability of observing xt in class k, and let us denote 1{·} the
indicator function. The following theorem characterizes the empirical discrete Bayes
classifier δπ̂B , and its associated empirical risk of misclassifications.
Theorem 1 The decision rule

δπ̂B : X i → arg min Ck,l π̂k p̂kt 1{X i =xt } (5)
l∈Ŷ t∈T k∈Y
minimizes (1) over , and its associated empirical risk is

r̂ π̂ , δπ̂B = Ck,l π̂k p̂kt 1{k∈Y Ck,l π̂k p̂kt =minq∈Ŷ k∈Y Ck,q π̂k p̂kt } . (6)
t∈T k∈Y l∈Ŷ
Proof The proof is omitted due to space limitations. More details are given in [7].

Let V : π → r̂ π, δπB be the empirical Bayes risk (6) extended over the simplex
S. Thus, when the probabilities p̂kt are fixed, V (π ) is the minimum risk we can
expect when considering the class proportions π . It is shown in [7] that the minimax
classifier is the Bayes classifier δπ̄B associated to the class proportions π̄ satisfying
π̄ = arg max V (π ). (7)

π∈S
As described in [7], V is concave and, in general, non-differentiable over S. Hence,

to compute π̄ , we propose to use a projected subgradient algorithm based on [8].
This algorithm is fully described in Algorithm 1. Its convergence to π̄ , as a function
of the number N of iterations, is established in [7].
Algorithm 1 Computation of the minimax classifier δπ̄B

Input: (Yi , X i )i∈I , K , N .
Compute π (1) = π̂
Compute the p̂kt ’s as described
in (4).
π̄ ← π (1) , r̄ ← r̂ π̂ , δπ̂B see (6)
for n = 1 to N do
for k = 1 to K do
(n)
gk ← R̂k δπB(n) = t∈T l∈Ŷ Ck,l p̂kt 1{ (n)
Ck,l πk p̂kt =minq∈Ŷ
(n)
Ck,q πk p̂kt }
k∈Y k∈Y
end for
(n) (n)
r (n) = k=1 K
πk gk see (1)
(n)
if r > r̄ then
r̄ ← r (n) , π̄ ← π (n)
end if
γn ← 1/n, ηn ← max{1, g (n) 2 }
π (n+1) ← PS π (n) + γn g (n) /ηn see [9] for the projection PS onto the simplex S
end for
Output: r̄ , π̄, and δπ̄B : X i → arg min t∈T k∈Y Ck,l π̄k p̂kt 1{X i =xt }
l∈Ŷ
3 Experiments
We applied our algorithm to the real condition monitoring database [6]. This public
database focuses on Air Pressure System (APS) used for various functions in Scania
trucks such as braking and gear changes. Measurements of a specific APS component
were collected from heavy Scania trucks in everyday usage. The goal is to predict a
potential failure of this component. We therefore consider K = 2 classes where the
class 1 corresponds to the APS without failures, and the class 2 corresponds to the
APS presenting failures. For this database, the costs of class misclassifications were
given by experts:
0 10
C= ,
500 0
so that the cost of predicting a non-existing failure is $10, while the cost of missing a
failure is $500. After removing missing values, the training set D contains the mea-
surements of m = 54, 731 samples for which m 1 = 54, 137 do not present any failure
and m 2 = 594 present a failure. Hence, the class proportions π̂ = [0.9891, 0.0109]
of the training set are highly imbalanced, which highly complicates the task of pre-
dicting a failure. Each sample is described by d = 129 numerical and categorical
anonymized features. For this database, an independent test dataset D is given, and
contains the measurements of m = 14, 578 samples (after removing samples with
missing values). The class proportions of this test set are π = [0.9848, 0.0152].
We compared our minimax classifier δπ̄B to the discrete Bayes classifier δπ̂B (5), and
to three others well known methods: Logistic Regression [10], Random Forest [11],
and Support Vector Machine (SVM) [12]. These three last methods were applied to
the original (non-quantized) features. For the following, we will use the notations
δπ̂LR for denoting the Logistic Regression, δπ̂RF for denoting the Random Forest, and
δπ̂SVM for denoting the SVM. Tocompare these five methods, we consider the empir-
ical risk of misclassifications r̂ π , δπ̂ , the maximum of the class-conditional risks
maxk∈Y R̂k (δπ̂ ), and the Total Cost ($) given by

TC(δ) = C1,2 1
Ŷ =2 + C2,1 1
Ŷ =1 = m r̂ π , δπ̂ .
i i
i∈I1 i∈I2
3.1 Training Step
In order to apply our algorithm, we quantized the features with the k-means algorithm
with a number T = 40 centroids. The choice of T is important. It was established
from Fig. 1, left. For each value of T , we performed a tenfold cross-validation over a
training subset. Of the 10 subsamples, a single subsample is retained as the validation
set for testing the classifier, and the remaining 9 subsamples are used as training set
for learning the classifier. The blue curve shows that the classifier is better for a large
value of T but its generalization error on the validation set, plotted in red, increases
from T around 40. We therefore chose T = 40 which is a good trade-off between
minimizing the average empirical risk over the subtraining set and the validation set.
In Table 1 we can observe that the class-conditional risks R̂1 (δ) and R̂2 (δ) are
very unbalanced for the classifiers δπ̂LR , δπ̂SVM , δπ̂RF , δπ̂B . This means that for these
classifiers, the risk of missing a failure is much higher than the risk of predicting
a non existent failure. In the case where we are interesting in minimizing the risk
of missing failures, it is therefore better to consider our minimax classifier δπ̄B since

this decision rule aims to minimize max{ R̂1 δπ̂B , R̂2 δπ̂B }, and tends to equalize as
much as possible the class-conditional risks. This fact is also illustrated in Fig. 1,
right.
In Table 2 we studied the empirical probabilities P̂(δ(X i ) = l | Yi = k), for all
δ ∈ {δπ̂LR , δπ̂SVM , δπ̂RF , δπ̂B , δπ̄B }. In Table 2, and more precisely in the confusion matrix
associated to the minimax classifier δπ̄B , we can observe that this decision rule gets
the best predictions associated to the class 2. Note that concerning the training step,
the discrete Bayes classifier δπ̂B gets very good results too, even concerning the class
Table 1 Results associated to the training dataset with π̂ =[0.9891, 0.0109]. The minimax clas-
sifier is associated to π̄ = [0.6045, 0.3955] and it satisfies r̂ π̄, δπ̄B = 6.0437.
Original features Quantized features
δ δπ̂LR δπ̂SVM δπ̂RF δπ̂B δπ̄B

r̂ π̂ , δ $/sample 1.98 5.31 0.94 0.97 5.60
R̂1 (δ) $/sample 0.011 0.004 0.003 0.486 5.592
R̂2 (δ) $/sample 181.818 489.057 86.700 45.455 6.734
TC(δ) $ 108,580 290,720 51,640 53,300 306,740
Time s 501.90 991.48 213.41 10.17 10.34
Table 2 Confusion matrices associated to the training step
δπ̂LR Ŷ = 1 Ŷ = 2 δπ̂SVM Ŷ = 1 Ŷ = 2 δπ̂RF Ŷ = 1 Ŷ = 2

Y = 1 0.9989 0.0011 Y = 1 0.9996 0.0004 Y = 1 0.9997 0.0003
Y = 2 0.3636 0.6364 Y = 2 0.9781 0.0219 Y = 2 0.1734 0.8266
δπ̂B Ŷ = 1 Ŷ = 2 δπ̄B Ŷ = 1 Ŷ = 2
Y = 1 0.9514 0.0486 Y = 1 0.4408 0.5592
Y = 2 0.0909 0.9091 Y = 2 0.0135 0.9865
1, as detailed in Tables 1 and 2. However, we can observe in Fig. 1, right, that it

would be better to consider the minimax classifier when the class proportions of the
test set π differ from π̂ , since its associated risk is more stable contrary to the dicrete
Bayes classifier δπ̂B . Finally, in Table 1, we can observe an important gap between the
computation times of the classifiers δπ̂LR , δπ̂SVM , δπ̂RF and the classifiers δπ̂B , δπ̄B . Note
that processing times were obtained on a computer using an i7 processor (3.1 GHz).
3.2 Test Step
Concerning the test step, we first applied each classifier δπ̂LR , δπ̂SVM , δπ̂RF , δπ̂B and
δπ̄B to the entire test dataset provided in the database. The results are detailed in
Table 3 and Table 4. We can observe that the risks and the class-conditional risks
associated to each classifier are pretty similar to the training step (except for the
Random Forest). Note that here π is similar to π̂ . In order to illustrate the robustness
of the classifier δπ̄B when π differs from π̂ , we considered the following experiment.
First, we generated M = 10, 000 random priors π (s) , s ∈ S := {1, . . . , M}, over the
simplex using the procedure proposed by [13]. For each s ∈ S, we then generated
a test subset containing 201 samples (size of the smallest class in D ) satisfying the
class proportions π (s) . Thus, each trained classifier δ ∈ {δπ̂LR , δπ̂SVM , δπ̂RF , δπ̂B , δπ̄B } is
tested on each test subset satisfying the class proportions π (s) . The results are detailed
Table 3 Results associated to the test dataset with π = [0.9848, 0.0152]

Original features Quantized features
δ δπ̂LR δπ̂SVM δπ̂RF δπ̂B δπ̄B

r̂ π , δ $/sample 2.59 7.41 2.90 1.14 5.65
R̂1 (δ) $/sample 0.021 0.001 0.016 0.461 5.629
R̂2 (δ) $/sample 169.683 488.687 190.045 45.259 6.787
TC(δ) $ 37,800 108,010 42,230 16,620 82,320
Table 4 Confusion matrices associated to the test step
δπ̂LR Ŷ = 1 Ŷ = 2 δπ̂SVM Ŷ = 1 Ŷ = 2 δπ̂RF Ŷ = 1 Ŷ = 2

Y = 1 0.9979 0.0021 Y = 1 0.9999 0.0001 Y = 1 0.9984 0.0016
Y = 2 0.3394 0.6606 Y = 2 0.9774 0.0226 Y = 2 0.3801 0.6199
δπ̂B Ŷ = 1 Ŷ = 2 δπ̄B Ŷ = 1 Ŷ = 2
Y = 1 0.9539 0.0461 Y = 1 0.4371 0.5629
Y = 2 0.0905 0.9095 Y = 2 0.0136 0.9864
Fig. 2 Dispersion of the risks, the Total Costs, and the maximums of the class-conditional risks,
associated to the M test subsets satisfying the class proportions π (s) , s ∈ S
in Fig. 2. We can observe that the dispersion of the risks r̂ (π (s) , δπ̄B ) is considerably
lower and thinner than for the other classifiers. This observation holds for the Total
Cost, and for the maximum of the class-conditional risks. This illustrates that the
minimax classifier δπ̄B is more robust when π differs from π̂ .
4 Conclusion
This paper proposes an algorithm to compute the minimax classifier which (i) is
suitable for working on discrete/discretized features, (ii) takes into account costs
given by experts, and (iii) is robust to the imbalanced or uncertain class proportions.
In future work, we will investigate the robustness of the classifier with respect to the
probabilities p̂kt .
References
1. Yang Y, Webb GI (Jan 2009) Discretization for naive-bayes learning: managing discretization
bias and variance. Mach Learn 74(1):39–74
2. Poor HV (1994) An introduction to signal detection and estimation, 2nd edn. Springer, Berlin
3. Ávila Pires B, Szepesvari C, Ghavamzadeh M (2013) Cost-sensitive multiclass classification
risk bounds. In: Proceedings of the 30th international conference on machine learning, vol 28
4. Drummond C, Holte RC (2003) C4.5, class imbalance, and cost sensitivity: Why under-
sampling beats oversampling. In: Proceedings of the ICML’03 workshop on learning from
imbalanced datasets
5. Guerrero-Curieses A, Alaiz-Rodriguez R, Cid-Sueiro J (2004) A fixed-point algorithm to min-
imax learning with neural networks. IEEE Trans Syst Man Cybernetics Part C Appl Rev
34(4):383–392
6. AB SC (2016) APS failure at Scania trucks data set. UCI machine learning repository (Online).
Available: https://www.kaggle.com/uciml/aps-failure-at-scania-trucks-data-set
7. Gilet C, Barbosa S, Fillatre L (2019) Minimax classifier with box constraint on the priors.
hal-02296592
8. Alber YI, Iusem AN, Solodov MV (1998) On the projected subgradient method for nonsmooth
convex optimization in a hilbert space. Math Programming 81:23–35
9. Condat L (2016) Fast projection onto the simplex and the 1 ball. Math Programming
158(1):575–585
10. McCullagh P, Nelder JA (1990) Generalized linear models, 2nd edn. Chapman & Hall
11. Breiman L (2001) Random forests. Mach Learn 45:5–32
12. Christianini N, Shawe-Taylor JC (2000) An introduction to support vector machines and other
Kernel-based learning methods. Cambridge University Press, Cambridge
13. Reed WJ (1974) Random points in a simplex. Pacific J Math 54(2):183–198
A Comparative Study
of Statistical-Based Change Detection
Methods for Multidimensional
and Multitemporal SAR Images
Ammar Mian and Frédéric Pascal
Abstract This paper addresses the problem of activity monitoring through change
detection in a time series of multidimensional Synthetic Aperture Radar (SAR)
images. Thanks to SAR sensors’ all-weather and all-illumination acquisitions capa-
bilities, this technology has become widely popular in recent decades when it con-
cerns the monitoring of large areas. As a consequence, a plethora of methodologies
to process the increasing amount of data has emerged. In order to present a clear
picture of available techniques from a practical standpoint, the current paper aims
at presenting an overview of statistical-based methodologies which are adapted to
the processing of noisy and multidimensional data obtained from the latest gener-
ation of sensors. To tackle the various big data challenges, namely the problems
of missing data, outliers/corrupted data, hetergenenous data, robust alternatives are
studied in the statistics and signal processing community. In peculiar, we investigate
the use of advanced robust approaches considering non-Gaussian modeling which
appear to be better suited to handle high-resolution heterogeneous images. To illus-
trate the attractiveness of the different methodologies presented, a comparative study
on real high-resolution data has been realized. From this study, it appears that robust
methodologies enjoy better detection performance through a complexity trade-off
with regards to other non-robust alternatives.
Keywords Change detection · Synthetic aperture radar · Time series · Activity

monitoring
A. Mian (B)
SONDRA, CentraleSupélec, 3 rue Joliot Curie, 91192 Gif-sur-Yvette, France
LISTIC, Université Savoie Mont-Blanc, 74944 Annecy le Vieux, France
e-mail: ammar.mian@centralesupelec.fr
F. Pascal
Laboratoire des Signaux et Systèmes, CentraleSupélec, 3 rue Joliot Curie,
91192 Gif-sur-Yvette, France
e-mail: frederic.pascal@centralesupelec.fr

https://doi.org/10.1007/978-981-15-9199-0_18
190 A. Mian and F. Pascal
1 Introduction
1.1 Motivations
Change detection (CD) for remotely sensed images of the Earth has been a popular
subject of study for recent decades. It has indeed attracted a plethora of scholars due to
the various applications to military (activity monitoring) or civil (geophysics, disaster
assessment, etc) applications. With the increase in the number of spatial missions
with Synthethic Aperture Radar (SAR) sensors, the amount of readily available data
has led to a big data era. In order to process this quantity, automatic algorithms to
help analyzing time series of images have to be developed.
Thus, CD algorithms have been thoroughly investigated in the recent literature
[1]. In that context, statistical based methodologies have been successful to process
the images that are subjected to the well-known speckle noise. Notably, when it
concerns the processing of multidimensional (such as Polarimetric SAR images) and
multitemporal images, various methodologies have been proposed in the literature.
The present paper aims at giving a short overview of parametric methodologies based
on popular distribution models of the SAR data.
1.2 Change Detection Overview
The literature on the subject is dense, and a variety of methodologies are available.1
Broadly speaking, as illustrated in Fig. 1, a change detection algorithm relies on three
separate elements:
• A pre-processing phase in which the time series of images have to be co-registered,
which means that by applying geometric transforms, each pixel of every image
corresponds to the same physical localization. Various methodologies also consider
a denoising step in which the speckle noise is reduced thanks to filtering techniques
[3, 4].
Lastly, features can be selected from the images for more concise or descriptive
representation for CD purposes. Some possibilities include wavelet decomposition
[5], Markov fields [6] or Principal Component Analysis [7] among others.
• A comparison step in which the features at each date are compared among them-
selves. This step consists in finding a relevant distance to measure dissimilarities
of features between the different images of the series. Many techniques exist based
on various principles.
• A post-processing step which varies depending on the methodology used for the
comparison step. It can either correspond to a thresholding [8, 9] or involves
machine learning classification algorithms [10].
1 Notably, see [1] or [2] for an overview.

A Comparative Study of Statistical-Based Change Detection Methods … 191
Pre-processing Comparison
Corrections Pixel level
Co-registration Object level
Denoising Supervised
Feature selection Non-supervised
Image Time Series
Post-processing
Thresholding
Classification
Change detection
results
Fig. 1 General procedure for a change detection methodology
We will deal presently with comparison step methodologies. More specifically,

we will place ourself in the context of pixel-level methodologies as opposed to
object-level methodologies as defined in [1]. In this paradigm, the comparison is
done on a local spatial neighborhood through the principle of the sliding windows:
for each pixel a spatial neighborhood is defined and a distance function is computed
to compare the similarities of this neighboorhood between the images at different
time.
Statistical-based methodologies rely on a probability model in order to infer about
changes. Those methods can rely on a model with associated parameters in which
case they are referred to as parametric methodologies. Nonparametric methods con-
sider statistical tools without considering parameters [11] or Bayesian models [12]
with a prior on the distribution of those parameters. We will consider presently the
parametric approach which has been the most popular subject of study in the recent
decade thanks to the seminal work by Conradsen et al. [13] which showed the attrac-
tiveness of applying statistical results on the equality of covariance matrix for CD.
We will first introduce probability models used to model SAR images with asso-
ciated parameters in Sect. 2. Then in Sect. 3, we will detail several dissimilarity
functions referenced in the literature for SAR change detection. Finally, we will
compare the performance of these functions on two distinct datasets in Sect. 5 and
conclude in Sect. 6.
p p
x17 x18 x19 xT7 xT8 xT9
x14 x15 x16 ... xT4 xT5 xT6
x11 x12 x13 xT1 xT2 xT3
t=1 t=T
Fig. 2 Illustration of the sliding windows (in grey) approach ( p = 3, N = 9). The central pixel
(x5t ) corresponds to the test pixel
2 Statistical Modeling of SAR Images
2.1 The Data
Denote by W = {X1 , . . . , XT } a collection of T mutually independent groups of

p-dimensional i.i.d complex vectors: Xt = [x1t , . . . , xtN ] ∈ C p×N . With regards to
the Single Look Complex (SLC) SAR images these sets correspond to the local
observations in a spatially sliding windows as illustrated in Fig. 2. The subscript k
correspond to a spatial index while the superscript t corresponds to a time index.
When it concerns Multilook Complex Images (MLC), the complex amplitude
data xkt is not made available. The Sample Covariance Matrix (SCM) is then released
to filter the speckle noise and compress the data. The number of pixels L used for
the averaging is called the number of looks.
2.2 Gaussian Model
The centered complex Gaussian distribution is the most encountered one in the
SAR literature, notably for modelling Polarimetric images. Indeed, since each pixel
consists of the coherent sum of the contribution of many scatterers, it is expected that,
thanks to the central limit theorem, the Gaussian model is not too far from the actual
empirical distribution. A random vector x ∈ C p is said to be distributed along the
complex Gaussian distribution (which we denote x ∼ CN (0 p , )) if its probability
density function (p.d.f) is:
1
pxCN (x; ) = exp −xH −1 x . (1)
π p ||
For multilook images, the matrix obtained through data averaging is assumed to be
Wishart distributed as a consequence.
2.3 Non-Gaussian Modeling
While the Gaussian distribution is popular, it fails to accurately describe the het-
erogeneity observed in very high-resolution images as described in [14], [15] or in
[16]. Indeed, in those images, the amount of scatterers in each pixel has been greatly
reduced with regards to low-resolution images.
To better describe the observed distribution of data, other models have been con-
sidered. For example, the K-distribution has been considered in [17, 18], the Weibull
distribution in [19] or inverse Generalized Gaussian distribution in [20]. These vari-
ous models belong to the family of complex elliptical distributions which generalizes
them as discussed in [21] which is a model depending on a density generator function
g : R+ → R+ that satisfies the condition m p,g = R+ t p−1 g(t)dt < ∞ and a pos-
p
itive definite matrix ∈ SH : a random vector x ∈ C p follows a complex elliptical
distribution (denoted x ∼ CE(0 p , g, )) if its p.d.f is the following:

pxCE (x; , g) = C p,g ||−1 g xH −1 x , (2)

where C p,g is a normalization constraint ensuring that C p pxCE (x; , g)dx = 1.
Another representation of those distributions can be found through the compound-
Gaussian model, sometimes referred as product model. In this case, a random vector
x follows a complex compound-Gaussian (CCG) distribution if:
√
x= τ z, (3)
where z ∼ CN (0 p , ξ ), called the speckle and τ , called the texture, follows a distri-
bution probability on R+ . This quantity is often assumed to be deterministic in order
to generalize the Gaussian distribution without having to consider a model on the
texture. In this case, the statistical model for the pixels in a sliding window is written:

xkt ∼ τkt zkt , (4)
where τkt ∈ R+ is deterministic and zkt ∼ CN (0 p , ξ t ).

The elliptical and CCG distributions are probability models which are in the scope
of the robust statistical literature initiated by works such as [22–24]. More details can
be found in the books [25] and [26]. Following their definition, a method is said to
be robust, in the context of this paper, when its statistical properties are independent
of the density generator function in the elliptical case or independent of the set of
texture parameters for the deterministic CCG case.
Using these various models, we will describe, hereafter, various dissimilarity
measures for those models which can be used for change detection.
3 Dissimilarity Measures
3.1 Problem Formulation
Under a parametric approach, CD can be achieved by deciding between the two

following alternative hypotheses:

H0 : θ 1 = . . . = θ T = θ 0 (no change),
. (5)
H1 : ∃(t, t ), θ t = θ t (change)
where θ corresponds to the parameters of the distribution used as a model.

In order to decide, a dissimilarity measure between the data over time is needed.
It can be seen as a function, also called statistic:
p×N ×T
ˆ : C

→R
W ˆ
→ (W),
ˆ
such that (W) is high when H1 is true and low otherwhise.
An interesting property of such function is the Constant False Alarm Rate (CFAR)
property. This property is valid when the distribution of the statistic under H0 hypoth-
esis is not a function of the parameters of the problem. This allows selecting a thresh-
old value for the detection which is directly linked to the probability of false alarm.
The threshold is often obtained by deriving the distribution of the statistic or can be
obtained by Monte-Carlo simulations when it is not available.
3.2 Hypothesis Testing Statistics
The first kind of statistics comes from the statistical literature on hypothesis testing.
Several techniques have been developed to obtain a statistic of decision.
A comparative study of this framework can be found in [27]. The analysis pre-
sented allowed to show that many approaches led to equivalent results. Thus, three
statistically independent statistics remain:
• the Generalized Likelihood Ratio Test (GLRT) statistic:

ˆ SCM T N
0
ˆG =
, (6)
T

ˆ SCM N
t
t=1
where:
1 t tH 1 SCM
N T
ˆ tSCM =
∀t, ˆ 0SCM =
xk xk and ˆ . (7)
N k=1 T t=1 t
This statistic is well known in the literature and its properties have been well
studied. [28] has derived the distribution under the null hypothesis and proposed
an approximation as a function of χ 2 distributions.
• the t1 statistic:

−1 2
1
T
ˆ t1 = Tr ˆ 0SCM ˆ tSCM
. (8)
T t=1
This statistic has the CFAR property and has a χ 2 asymptotic distribution (when
N → ∞).
• the Wald statistic:

T 2
ˆ Wald = N
ˆ SCM ˆ SCM −1
Tr I p − 1 (t ) (9)
t=2
T

T
−q N ˆ SCM −† ˆ SCM −1
(t ) ⊗ (t ) , vec ϒt , (10)
t=1 t=2
where † is the transpose operator and:

ˆ tSCM )−1 − (
ϒ t = N ( ˆ 1SCM (
ˆ tSCM )−1 ˆ tSCM )−1 , (11)
−1
q(x, ) = x x.
H
(12)
This statistic has the CFAR property and has a χ 2 asymptotic distribution (when
N → ∞). It is, however, computationally more complex than the previous ones
due to the Kroenecker product.
In order to deal with multilook images and for a pair of two images (T = 2), [29]
proposed to use the Complex Hotelling-Lawley statistic given by:
ˆ HTL = Tr((
ˆ 1SCM ),
ˆ 0SCM )−1 (13)
where the estimates correspond to the multilook covariance matrices at date 0 and
1. The authors also derived the distribution under null hypothesis and showed that it
can be approximated by a Fisher-Snecdor distribution. The study showed a potential
improvement of detection rate compared to the GLRT statistic.
Concerning non-Gaussian models, [30] has considered the derivation of the GLRT
using the compound-Gaussian model with deterministic texture parameters:
T T p
MT T N MT
q ξ̂ 0 , xk
t
ξ̂ 0
N
ˆ MT =

t=1
, (14)
T
T TE p
TE N k=1 T p
ξ̂ Wt T q ξ̂ Wt , xkt
t=1 t=1
where Wt = {xkt : 1 ≤ k ≤ N },

T
H
xkt xkt
TE p t tH
xk xk MT p N
t=1
ξ̂ Wt = , and ξ̂ 0 = , (15)
N t TE
xk ∈Wt q(ξ̂ Wt , xk )
t N k=1
MT
T
q ξ̂ 0 , xkt
t=1
and q(ξ , x) is defined at in (12).

This statistic has been shown to have better robustness to very heterogeneous data
for which the Gaussian model is very far. This comes however, at a cost of complexity.
Indeed, there is a need to compute the solution of two fixed point equations which is
computationally expensive. The statistic has the CFAR property but the distribution
under null hypothesis has not yet been derived.
3.3 Information-Theoretic Measures
Information theory tools are based on another approach which consists in measuring
quantities related to the information at disposition on the data. In order to compare
two distributions, distances can be computed. When a parametric model is used, the
distance is often dependent on the parameters of the two distributions compared.
Since in many applications, the parameters are not known, estimates are plugged
into the distance function.
The Kullback-Leibler divergence is a popular measure encountered, notably in
SAR change detection problems [31]. For the Gaussian case, a distance can be
obtained through the symmetrised version of the divergence:
1
ˆ KL =
ˆ 0SCM ,
d( ˆ 1SCM ) + d(
ˆ 1SCM ,
ˆ 0SCM ) , (16)
2
where d(A, B) = Tr(B−1 A) + log |A| |B|

. [32] has shown that this distance is asymp-
totically distributed as a χ 2 distribution.
Other measures include in Renyi entropy or Bhattacharya distance which have

been shown to behave similarly in [32]. They will thus be omitted in the present
paper.
For non-Gaussian model, [33] proposed a measure in the multilook case and
for T = 2, based on the principle of mutual information. This distance, based on a
gamma model on the texture parameters, is a generalization of the one proposed in
[34]:
ˆ MLL = MLL(W1 ) + MLL(W2 ) − MLL(W12 ), (17)
where W12 = W1 ∪ W2 ,
MLL(W) = N ν+2pL (log(Lν) − log(μ)) − N L log |ξ TE

W| (18)

+ ν−2pL −1 t t H
log Tr (ξ TE
W ) xk xk − N log |(ν)|
xkt ∈W

+ log K ν− pL 2 Tr (ξ TE ) −1 xt xt H Lν/μ ,
W k k
xkt ∈W
is the Gamma function, L is the number of equivalent looks, ξ TE W is defined in

(15), K ν (z) is the modified Bessel function of the second kind and (μ, ν) are the
parameters of a Gamma distribution found by fitting the distribution of the texture
parameters set
T = {τ̂kt = q(ξ TE
W , xk )/ p : xk ∈ W}.
t t
(19)
The method used yields a high computational cost, since the test statistic relies
on a Bessel function as well as a fitting of a Gamma distribution.
3.4 Riemannian Geometry Distances
Riemannian geometry is an alterative concept which allows us to compare distribu-

tions. Indeed, when the parameters of a distribution lie in a Riemannian manifold (for
example, covariance matrices lie in Riemannian manifold of positive definite matri-
p
ces SH [35]), it is possible to define a metric which can be related to the Kullback-
Leibler divergence. Consequently, distances between two probability distributions
which take into account the geometry properties of the parameter space, have been
considered in the litterature. Notably, for a Gaussian model, we have the following
distance [36]:
ˆ RE = log (
ˆ 2SCM (
ˆ 1SCM )− 21 ˆ 1SCM )− 21 2F , (20)
p
where log is the logarithm of matrices which for SH , having the eigenvalue decom-
position = UUH , is defined as follows:
log() = U log ()UH , (21)
where the log is pointwise logarithm.

For the elliptical case, a Riemannian CES distance can be obtained as well [37]:
p 2

p

ˆ RE = α
log2 λi + β log λi (22)
i=1 i=1
ˆ 1MLE )−1
where λi is the i-th eigenvalue of ( ˆ ∈{0,1}
ˆ 2MLE , MLE
is the MLE of the scatter
matrix in CES case and α, β depends on the density generator function. For a Student-
t distribution, we have:
d+p
N
d+p xk xk H
α= ˆ MLE =
, β = α − 1, , (23)
d + p+1 N k=1 d + xk H (ˆ MLE )−1 xk
where d is the number of degrees of freedom of the Student-t distribution.

The concept of Riemannian geometry has also been considered in [38], where a
methodology has been proposed for polarimetric SAR images based on Kennaugh
matrix decomposition. This distance, however, is not based on a probability model
and will be thus omitted in the present paper.
3.5 Optimal Transport
Finally a distance can be obtained through the concept of optimal transport which
defines a distance depending on a cost to displace parts of a reference distribution into
a target distribution. It has been shown in [39], that this distance has a closed-form
for elliptical distributions:
1
1 1 2
d(1 , 2 ) = Tr(1 ) + Tr(2 ) − 2Tr 1 2 1
2 2
. (24)
Thanks to this result, we can define two statistics for Gaussian and Elliptical
case by plugging estimates of the covariance matrices. A Wasserstein statistic for
Gaussian can be computed by taking:
ˆ WG = d(
ˆ 1SCM ,
ˆ 2SCM ). (25)
Table 1 Summary of statistics with their respective properties

Statistic Reference Model Eq. T >2 SLC/MLC CFAR Asymptotic null
distribution
ˆG
[13] Gaussian (6) Both χ 2 (asymptotic)
ˆt
1
ˆ Wald
ˆ HTL
[29] Gaussian (13) × Both Fisher-Snecdor
(approximation)
ˆ MT
[30] Deterministic (14) SLC ×
CCG
ˆ KL
[32] Gaussian (16) × Both χ 2 (asymptotic)
ˆ MLL
[33] Compound- (17) × MLC × ×
Gaussian
ˆ RG
[36] Gaussian (20) × Both × ×
ˆ RE
[37] Student-t (22) × SLC × ×
ˆ WG
[40] Gaussian (25) × Both × ×
ˆ WE
[40] Elliptical (26) × SLC × ×
A Wasserstein distance in the elliptical can be obtained by taking:

TE TE
ˆ WE = d(τ̂1 ξ̂ 1 , τ̂2 ξ̂ 2 ),
(26)
TE N TE
where ξ̂ ∈{0,1} is defined in (15) and τ̂∈{0,1} = ( p/N ) k=1 q(ξ̂ , xk ).
3.6 Summary
A summary of the different statistics with their respective properties can be found in
Table 1.
4 Analysis of Complexity
To compare the attractiveness of the different statistics described in Sect. 3 with

regards to computational complexity, we consider an experimental Monte-Carlo
analysis on synthetic Gaussian data.
A set of data W is generated with parameter values p = 10, N = 25 and T =
2. The Gaussian data is generated with a covariance matrix ()i j = 0.5|i− j| and
4000 Monte-Carlo trials have been considered. For statistics involving fixed-point
equation, we fixed the number of iterations to 15 and we choose L = 1 (Single Look
Table 2 Time consumption in seconds

ˆG
ˆ t1 ˆ Wald
ˆ HTL ˆ MT ˆ KL
ˆ MLL
ˆ RG
ˆ RE
ˆ WG
ˆ WE

0.001 0.001 0.003 0.001 0.004 0.001 0.006 0.011 0.007 0.002 0.007
data). Finally, the simulation has been done on 2.50 GHz processor in Python 3.7.
The mean execution times are provided in Table 2.
Several observations can be made:
• Most of the Gaussian derived statistics have similar time consumption since the
statistic are based on the computation of the SCM which is less expensive than a
fixed-point estimation. Wald statistic is more expensive than the Gaussian GLRT
and t1 statistics due to the inverse of a Kroenecker products which is more expen-
sive than a simple inverse. ˆ RG has the highest time consumption, even compared
to non-Gaussian methodologies, due to the fact that it requires to compute the
square root of matrices which can take a lot of time depending on the implementa-
tion (the one used correspond to the Scipy implementation which can be improved
with regards to complexity).
• Concerning non-Gaussian statistics, the time consumption is generally higher than
Gaussian methodologies due to the fixed-point estimation. The best methodology
with regards to time consumption is ˆ MT since it requires fewer operations than
the others.
5 Experiments on Real Data
5.1 Methodology and Aim of the Study
We consider comparing the performance of detection for all the statistics presented
in Sect. 3 on real SLC polarimetric SAR datasets where a ground truth is available.
In order to compute the statistics, we use a sliding windows of size 5 × 5 (N = 25).
The fixed-point estimates are computed with 5 iterations to save time. For the statistic
ˆ MLL , since it consider a multilook image, we choose L = 1. For the statistic ˆ RE
based on the Student-t distribution, we choose d = 3 in order to have a distribution
far from the Gaussian one. When the data correspond to a series of length T > 2,
only the adapted statistics are computed.
The performance of detection is measured through the Receiver Operator Curve
(ROC), which quantify the amount of good detection as a function of false alarms.
5.2 Datasets Used
We consider three scenes of UAVSAR polarimetric data obtained from

https://uavsar.jpl.nasa.gov (Courtesy NASA/JPL-Caltech) presented in Figs. 3 and
4 with their associated ground truth. The first two scenes are referenced under label
SanAnd_26524_03 Segment 4 and correspond to a pair of images at dates April 23,
2009, and May 11, 2015 (T = 2). The groud truth for these two scenes has been
fetched from [41]. The third scene is referenced under label Snjoaq_14511 and cor-
respond to a series of T = 17 images and the ground truth have been realized by the
present authors.
Fig. 3 UAVSAR Scenes 1 (top) and 2 (bottom) in Pauli representation. Left: Image at first date.
Middle: Image at second date. Right: Ground Truth
Fig. 4 UAVSAR Scene 3 in Pauli representation. Left: Image at t = 1. Right: Ground Truth on the
whole series
5.3 Reproducibility
The codes (in Python 3.7) are available at https://github.com/AmmarMian/WCCM-

2019. The datasets can be obtained from UAVSAR website (https://uavsar.jpl.nasa.
gov) using the labels given and the ground truth may be obtained from the authors.
5.4 Results and Discussion
The ROC for all statistics and three scenes are presented in Fig. 5. A transcription
of the detection performance at PFA = 0.01 is also given in Table 3.
The first observation that can be made is that ˆ MLL generally have poorer perfor-
mance of detection than the other statistics. This is mostly explained by the fact that
it has been developed for a multilook scenario which was not the case here. Thus,
even by taking L = 1, the statistic yields poorer detection performance than most of
the other statistics.
The statistics obtained through the Wasserstein distance, ˆ WG and ˆ WE appear
to not work well for this CD problem since the performance of detection compared
to others distance is lower, especially for Scene 2.
For the case T = 2, considering the Gaussian-derived statistics ˆ G, ˆ t1 ,
ˆ KL ,
ˆ HTL and ˆ RG , it is difficult to discern a best statistic since depending on the scene,
their relative performance vary. When the number of images is high, ˆ Wald appears
to be lower in all cases but the performance of detection is not far. For T > 2, ˆG
seems to have better performance than its counterpart but no conclusion can be made
since only one set of data with T > 2 was available.
Finally, concerning robust statistics ˆ MT has the overall best performance for all
scenes. In scene 1, it has slightly lower performance than ˆ HTL but on the other
scenes (especially Scene 3), the performance of detection is greatly improved. The
gain of this statistic compared to Gaussian one is explained by the heterogeneity of
the images at the transition between objects. As an example, the output of both ˆG
1
Λ̂G Λ̂t1
0.8
Λ̂HTL Λ̂RG
Λ̂KL Λ̂WG 0.6
PD
Λ̂Wald Λ̂MT 0.4
Λ̂WE Λ̂MLL
0.2
Λ̂RE
0
10−3 10−2 10−1 100
PFA
1 1
0.8 0.8
0.6 0.6
PD
PD
0.4 0.4
0.2 0.2
0 0
10−3 10−2 10−1 100 10−3 10−2 10−1 100
PFA PFA
Fig. 5 ROC plots for a window size of 5 × 5 for the three scenes. Top-right: Scene 1. Bottom-left:
Scene 2. Bottom-right: Scene 3
Table 3 Probability of detection at a false alarm rate of 1%

ˆG
Statistic ˆt
ˆ Wald
ˆ HTL
ˆ MT
ˆ KL
ˆ MLL
ˆ RG
ˆ RE
ˆ WG
ˆ WE

1
Scene 1 0.41 0.41 0.35 0.43 0.41 0.40 0.18 0.41 0.39 0.12 0.36
Scene 2 0.33 0.30 0.27 0.32 0.39 0.32 0.11 0.33 0.29 0.19 0.19
Scene 3 0.90 0.75 0.62 0.95
and ˆ MT are presented in Fig. 6. For this data, the Gaussian derived statistic, which
does not take into account the heterogeneity, yields a high value of the statistic at
the transition between the fields while the CCG statistic does not, which constitutes
a big improvement concerning false alarms. The natural Student-t statistic ˆ WE
does not have good performance of detection which is explained by a mismatch in
the degrees of freedom chosen to model the distribution of the data. Moreover, the
Student-t model can be inaccurate to model the actual data. In those regards, ˆ MT is
able to obtain better performance without relying on a specific elliptical model and
has thus a better robustness to mismatch scenarios.
6000
5000
4000
3000
2000
1000
Fig. 6 Output of ˆ G (left) and

ˆ MT (right) on scene 3. The major difference between the two
statistics is higlihted by red arrows
6 Conclusions
We have described several parametric approaches for CD in multidimensional and

multitemporal SAR images obtained from various principles and detailed their prop-
erties. The main result obtained through the comparative study on real data under-
taken in Sect. 5, was to show that the statistic ˆ MT , based on a deterministic CCG
modelling, provided the best performance of detection thanks to its inherent robust-
ness properties. Indeed, by taking into account the heterogeneity of the backscattering
power on the local window size, the statistic allowed us to reduce the number of false
alarms and kept the amount of true detection coherent with Gaussian-derived detec-
tors. This improvement comes with a higher computational cost as explored in Sect. 4
which has been shown to be reasonable with regards to its detection performance.
References
1. Hussain M, Chen D, Cheng A, Wei H, Stanley D (2013) Change detection from remotely
sensed images: From pixel-based to object-based approaches. (ISPRS) J Photogrammetry
Remote Sens 80:91–106 (Online). Available: http://www.sciencedirect.com/science/article/
pii/S0924271613000804
2. Hecheltjen A, Thonfeld F, Menz G (2014) Recent advances in remote sensing change
detection—a review. Springer Netherlands, Dordrecht, pp 145–178
3. Achim A, Tsakalides P, Bezerianos A (2003) Sar image denoising via Bayesian wavelet shrink-
age based on heavy-tailed modeling. IEEE Trans Geosci Remote Sens 41(8):1773–1784
4. Foucher S, Lopez-Martinez C (2014) Analysis, evaluation, and comparison of polarimetric
SAR speckle filtering techniques. IEEE Trans Image Process 23(4):1751–1764
5. Mian A, Ovarlez J-P, Ginolhac G, Atto AM (2017) Multivariate change detection on high
resolution monovariate SAR image using linear time-frequency analysis. In: 25th European
Signal Processing Conference (EUSIPCO). Kos, Greece
6. Wang F, Wu Y, Zhang Q, Zhang P, Li M, Lu Y (2013) Unsupervised change detection on SAR
images using triplet Markov field model. IEEE Geosci Remote Sens Lett 10(4):697–701
7. Yousif O, Ban Y (2013) Improving urban change detection from multitemporal SAR images
using pca-nlm. IEEE Trans Geosci Remote Sens 51(4):2032–2041
8. Bruzzone L, Prieto DF (2000) Automatic analysis of the difference image for unsupervised
change detection. IEEE Trans Geosci Remote Sens 38(3):1171–1182
9. Kervrann C, Boulanger J (2006) Optimal spatial adaptation for patch-based image denoising.
IEEE Trans Image Process 15(10):2866–2878
10. Gong M, Zhao J, Liu J, Miao Q, Jiao L (2016) Change detection in synthetic aperture radar
images based on deep neural networks. IEEE Trans Neural Netw Learn Syst 27(1):125–138
11. Aiazzi B, Alparone L, Baronti S, Garzelli A, Zoppetti C (2013) Nonparametric change detection
in multitemporal SAR images based on mean-shift clustering. IEEE Trans Geosci Remote Sens
51(4):2022–2031
12. Prendes J, Chabert M, Pascal F, Giros A, Tourneret J (2015) Change detection for optical and
radar images using a bayesian nonparametric model coupled with a markov random field. In:
2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),
April 2015, pp 1513–1517
13. Conradsen K, Nielsen AA, Schou J, Skriver H (2001) Change detection in polarimetric SAR
data and the complex Wishart distribution. In: IGARSS 2001. Scanning the present and resolv-
ing the future. Proceedings. IEEE 2001 International Geoscience and Remote Sensing Sym-
posium (Cat. No.01CH37217), vol 6, pp 2628–2630
14. Greco MS, Gini F (2007) Statistical analysis of high-resolution SAR ground clutter data. IEEE
Trans Geosci Remote Sens 45(3):566–575
15. Gao G (2010) Statistical modeling of SAR images: a survey. Sensors 10(1):775–795
16. Ollila E, Tyler DE, Koivunen V, Poor HV (2012) Compound-Gaussian clutter modeling with
an inverse Gaussian texture distribution. IEEE Signal Process Lett 19(12):876–879
17. Yueh S, Kong J, Jao J, Shin R, Novak L (1989) K-distribution and polarimetric terrain radar
clutter. J Electromagnetic Waves Appl 3(8):747–768 (Online). Available: https://doi.org/10.
1163/156939389X00412
18. Muller H (1994) K statistics of terrain clutter in high resolution SAR images. In: Proceedings
of IGARSS ’94 - 1994 IEEE International Geoscience and Remote Sensing Symposium, vol
4, Aug 1994, pp 2146–2148
19. Bucciarelli T, Lombardo P, Oliver CJ, Perrotta M (1995) A compound Weibull model for SAR
texture analysis. In: 1995 International Geoscience and Remote Sensing Symposium, IGARSS
’95. Quantitative Remote Sensing for Science and Applications, vol 1, July 1995, pp 181–183
20. Freitas CC, Frery AC, Correia AH (2005) The polarimetric g distribution for SAR data analysis.
Environmetrics 16(1):13–31 (Online). Available: https://onlinelibrary.wiley.com/doi/abs/10.
1002/env.658
21. Ollila E, Tyler DE, Koivunen V, Poor HV (2012) Complex elliptically symmetric distributions:
survey, new results and applications. IEEE Trans Signal Process 60(11):5597–5625
22. Maronna RA (1976) Robust M-estimators of multivariate location and scatter. Ann Statistics
4(1):51–67
23. Yohai VJ (1974) Robust estimation in the linear model. Ann Statist 2(3):562–567, 05 1974
(Online). Available: https://doi.org/10.1214/aos/1176342717
24. Martin B, Pierre D. Robust estimation of the sur model. Canadian J Statistics 28(2):277–288
(Online). Available: https://onlinelibrary.wiley.com/doi/abs/10.2307/3315978
25. Maronna RA, Martin DR, Yohai VJ (2006) Robust statistics: theory and methods. 1st edn, ser.
Wiley Series in Probability and Statistics, Wiley
26. Zoubir AM, Koivunen V, Ollila E, Muma M (2018) Robust statistics for signal processing.
Cambridge University Press
27. Ciuonzo D, Carotenuto V, Maio AD (2017) On multiple covariance equality testing with appli-
cation to SAR change detection. IEEE Trans Signal Process 65(19):5078–5091
28. Anderson T (2003) An introduction to multivariate statistical analysis, Ser. Wiley Series
in Probability and Statistics, Wiley (Online). Available: https://books.google.es/books?
id=Cmm9QgAACAAJ
29. Akbari V, Anfinsen SN, Doulgeris AP, Eltoft T, Moser G, Serpico SB (2016) Polarimetric
SAR change detection with the complex Hotelling-Lawley trace statistic. IEEE Trans Geosci
Remote Sens 54(7):3953–3966
30. Mian A, Ginolhac G, Ovarlez J-P, Atto AM (2019) New robust statistics for change detection
in time series of multivariate SAR images. IEEE Trans Signal Process 67(2):520–534
31. Inglada J, Mercier G (2007) A new statistical similarity measure for change detection in mul-
titemporal SAR images and its extension to multiscale change analysis. IEEE Trans Geosci
Remote Sens 45(5):1432–1445
32. Frery AC, Nascimento ADC, Cintra RJ (2014) Analytic expressions for stochastic distances
between relaxed complex Wishart distributions. IEEE Trans Geosci Remote Sens 52:1213–
1226
33. Liu M, Zhang H, Wang C, Wu F (2014) Change detection of multilook polarimetric SAR images
using heterogeneous clutter models. IEEE Trans Geosci Remote Sens 52(12):7483–7494
34. Beaulieu J, Touzi R (2004) Segmentation of textured polarimetric SAR scenes by likelihood
approximation. IEEE Trans Geosci Remote Sens 42(10):2063–2072
35. Skovgaard LT (1984) A Riemannian geometry of the multivariate normal model. Scandinavian
J Statistics 211–223
36. Smith ST (2005) Covariance, subspace, and intrinsic crame/spl acute/r-rao bounds. IEEE Trans
Signal Process 53(5):1610–1630
37. Breloy A, Ginolhac G, Renaux A, Bouchard F (2019) Intrinsic cramérao bounds for scatter
and shape matrices estimation in ces distributions. IEEE Signal Process Lett 26(2):262–266
38. Ratha D, De S, Celik T, Bhattacharya A (2017) Change detection in polarimetric SAR images
using a geodesic distance between scattering mechanisms. IEEE Geosci Remote Sens Lett
14(7):1066–1070
39. Gelbrich M (1990) On a formula for the l2 wasserstein metric between measures on euclidean
and hilbert spaces. Mathematische Nachrichten 147(1):185–203
40. Ghaffari N, Walker S (2018) On multivariate optimal transportation. arXiv preprint
arXiv:1801.03516
41. Nascimento ADC, Frery AC, Cintra RJ (2019) Detecting changes in fully polarimetric SAR
imagery with statistical information theory. IEEE Trans Geosci Remote Sens 57(3):1380–1392
Bearing Anomaly Detection Based
on Generative Adversarial Network
Jun Dai, Jun Wang, Weiguo Huang, Changqing Shen, Juanjuan Shi,
Xingxing Jiang, and Zhongkui Zhu
Abstract It is challenging to detect incipient fault of a bearing in a rotating machine

being operated under complex conditions. Current intelligent fault diagnosis models
require massive historical data of bearing monitoring signals under different health
states and the corresponding state labels. However, in some cases, only the samples
when the bearing is under healthy condition are available. To deal with the scenario
when anomalous samples are absent, this paper proposes a bearing anomaly detec-
tion method based on generative adversarial network (GAN). Specifically, only the
normal samples are used to train the proposed GAN model, by which the signal
data distribution of normal samples is learned. Then, the newly acquired signals with
unknown health states are taken as the input data of the trained model. Residual loss of
each new sample is calculated by comparing the dissimilarity between the generated
sample and the input sample. Because the trained model is optimized by the normal
samples, the residual losses of the anomalous samples will be large due to the different
data distribution from that of the normal samples. Finally, the bearing anomaly is
detected when the residual loss of newly acquired sample increase quickly during
continuous inspection of the bearing health condition. Life-circle vibration signals
of a bearing measured from a run-to-failure test are used to validate the proposed
method. The result indicates that the bearing anomaly can be detected earlier by
the proposed GAN model than the traditional method based on signal time-domain
statistical criterion, which is of great significance for incipient weak fault detection
of mechanical systems.
Keywords Anomaly detection · Deep learning · Generative adversarial network ·

Rolling bearing · Rotating machines
J. Dai · J. Wang (B) · W. Huang · C. Shen · J. Shi · X. Jiang · Z. Zhu

Institute of Industrial Measurement, Control and Equipment Diagnostics, School of Rail
Transportation, Soochow University, Suzhou, People’s Republic of China
e-mail: junking@mail.ustc.edu.cn

https://doi.org/10.1007/978-981-15-9199-0_19
208 J. Dai et al.
1 Introduction
In modern industries, mechanical systems tend to be more and more automatic,

precise, and efficient. Minor damage occurred may affect the normal operation of
the whole system and even cause serious accidents. In order to accurately capture
the health status information of the mechanical systems, mechanical condition moni-
toring systems have been used to collect massive data from the equipment [1], and
data-driven maintenance has become an important technique for mechanical system
condition monitoring and fault diagnosis [2].
Traditional signal analysis methods, such as wavelet transform (WT) [3], indepen-
dent component analysis (ICA) [4], and empirical mode decomposition (EMD) [5],
can extract effective fault components from original signals. However, the perfor-
mance of fault diagnosis by using these methods depends on signal processing expe-
rience of the maintenance personnel. In order to extract valid features from the
measured data efficiently, deep learning has been applied to the field of machinery
fault diagnosis. Tamilselvan and Wang [6] proposed a deep belief network (DBN)
health state classification model and applied it to aircraft engine health diagnosis.
Jia et al. [1] used an auto-encoders (AE) model and back-propagation algorithm to
extract fault features from original signals. Zhang et al. [7] constructed a convo-
lutional neural network (CNN) model for bearing fault diagnosis. Zhao et al. [8]
proposed a residual neural network (ResNet) based on dynamic weighted wavelet
coefficients for fault monitoring. These researches indicate that it has good applica-
tion prospect to adopt the deep learning in mechanical equipment condition moni-
toring and fault diagnosis [9]. However, these models are trained by inputting massive
historical data and the corresponding labels. In practical diagnosis task, it is difficult
to obtain all the samples under different fault states. In the cases when safety is of
paramount importance, such as the condition monitoring of bearings installed in a
train wheelset or an aeroengine, only the samples when the bearing is under healthy
state are available. In order to detect anomaly of a machinery without anomalous
samples, a novel method that relies on only the available normal samples is urgent
to be proposed.
Generative adversarial network (GAN) is a semi-supervised feature learning algo-
rithm which is proposed by Goodfellow in 2014 [10]. It has been applied in many
fields, including image processing [11, 12], style transfer [13], and anomaly detec-
tion [14, 15]. GAN has also been applied in the field of fault diagnosis in recent
years. Wang et al. [16] solved the problem of insufficient data samples by using the
GAN, and then used stack sparse AE to extract fault-sensitive features. Yong et al.
[2] dealt with the data imbalance problem by using the GAN to oversample the data
with less samples. A deep neural network model was then used to classify different
faults. Han et al. [17] proposed a model named categorical adversarial auto-encoder
(CatAAE) for unsupervised fault diagnosis of rolling bearings. Since the GAN is
able to learn the distribution feature of a group of data during the training process,
and discover the distribution diversity between the trained data and a new group of
Bearing Anomaly Detection Based on Generative … 209
data during the testing process, it can be used to solve the problem of the absence of
anomalous samples in practical diagnosis task.
In this paper, a new GAN-based anomaly detection method is proposed for bearing
condition monitoring. First, the samples of the bearing under health state are used to
train the model of the GAN. Then, the newly acquired signals with unknown health
states are taken as the input of the trained model. Due to that the data distributions
of the anomalous samples are different from that of the normal samples; the GAN
model trained by the normal samples are not fit for the anomalous samples. By
measuring the dissimilarity between the input data and the output data of the model,
the anomaly of the bearing can be finally detected. A group of life-circle vibration
signals of a rolling bearing validates the enhanced performance of the proposed
GAN-based anomaly detection method as compared to the traditional method based
on time-domain statistical criteria.
2 Generative Adversarial Network
The GAN contains a generator and a discriminator, as is illustrated in Fig. 1. The

generator aims to map latent vectors from latent space to high dimensions, by which
fake samples are created. The discriminator is used to tell the fake and real samples
apart. During the training process, the latent vectors can be reconstructed to be similar
to the real samples, and the data distribution of the real samples is learned. The GAN
is optimized by competing between the generator and the discriminator.
In Fig. 1, the latent vector z is a sample from a prior distribution pz . G(z) repre-
sents a generated sample. We hope the distribution of the generated sample G(z) is
close to that of the real sample x. The discriminator in the GAN is used to discrim-
inate between the real and the generated samples. The probability is given by the
discriminator to reflect the input sample is real or fake. When the real sample x is
input, the D(x) is expected to be close to 1; conversely, when the generated sample
G(z) is input, the output D(G(z)) is expected to be close to 0. The objective function
can be described as follows:

min max E x∼Pdata log D(x) + E z∼ pz [log(1 − D(G(z)))] (1)
G D
Fig. 1 Structure of the GAN x

Real sample Discriminator Real/fake
G(z)
z
Latent vector Generator
210 J. Dai et al.
3 GAN-Based Bearing Anomaly Detection Method
In the condition monitoring of a specific rolling bearing, the data collected in the
early stage are normal samples. In order to detect the anomaly of the bearing the first
time when a defect occurs without the anomalous samples, a new GAN model is
established in this paper. The proposed model is trained by only the normal samples.
That is to say, the normal samples are taken as the real samples in Fig. 1. During the
testing process, the newly collected samples are taken as the real samples succes-
sively, and the generator and discriminator are fixed. The latent vector is updated by
calculating the dissimilarity between a real sample and the corresponding generated
sample. Since the latent space has smooth transitions [18], an ideal-case scenario is
that the best latent vector z* can produce a generated sample that is as similar as
possible to the real sample x. If a new sample corresponds to the normal condition,
the dissimilarity between the generated sample and real sample will be very small.
On the contrary, if the new sample corresponds to the abnormal condition, due to
the different data distributions between the normal and the anomalous samples, the
optimal generated sample G(z∗) will hardly express the real sample. Therefore, the
dissimilarity between the real sample x and the generated sample G(z∗) is a good
evaluator for anomaly degree.
The flowchart of the proposed GAN-based bearing anomaly detection method is
shown in Fig. 2, where I represents the total iteration numbers during the training
process and K represents the total update numbers during the testing process. Firstly,
the GAN model is trained with backpropagation by the bearing vibration data under
normal condition after preprocessing via the Fourier transform and normalization
Normal samples New samples
Preprocessing Preprocessing
Backpropagation
GAN Trained GAN Latent

vector
update
N N
i<I k<K
Y Y
Testing
Training Process Residual loss Process
Fig. 2 Flowchart of the proposed GAN-based anomaly detection method

operation. Secondly, a newly acquired sample after the same preprocessing is taken
as the input data of the trained GAN model for testing. The optimal generated sample
G(z ∗ ) that has the minimum loss over the input data is obtained by updating the latent
vector for K times. Thirdly, the dissimilarity between the input data and the optimal
generated sample G(z∗), which is measured by residual loss, is calculated. Fourthly,
the newly acquired samples are continuously tested by the GAN model. A curve of
the residual loss versus time is obtained. The anomaly of the monitored bearing is
finally detected when an obvious increase on the curve is observed.
In this paper, the objective functions proposed in [14] is applied to update the latent
vector. Two parts of loss, including residual loss and discrimination feature loss, are
contained in the objective functions. The residual loss measures the dissimilarity
between the generated sample G(z) and the real input sample x, as formulated by:

L r es = min (x − G(z))2 (2)
The discrimination feature loss is defined by.

L dis = min (D f (x) − D f (G(z)))2 (3)
where D f (·) represents the hidden layer output of the discriminator. The objective
function is given by combining the residual loss and the discrimination feature loss:
L dis = (1 − λ)L res + λL dis (4)
where λ represents the weight between the two losses. λ = 0.1 in this paper.
4 Experiment Description
In this section, the life-circle vibration signals of a bearing from the center for intel-
ligence maintenance systems are used to verify the effectiveness of the proposed
method. The experiment platform is shown in Fig. 3. Four bearings (Type: Rexnord
ZA-2115 double row bearing) were installed on a shaft and driven by an AC motor.
The rotating speed was kept at 2000 RPM. Bearing vibration signals were collected
every 10 min with the sampling frequency of 20 kHz. Each sample has 20,480 points.
The system had been running continuously for 163 h before stop and 980 samples
were acquired finally. A severe outer-race fault was found in Bearing 1 after the
experiment. The bearing outer-race fault characteristic frequency is 236.4 Hz.
The time-domain vibration data of Bearing 1 is shown in Fig. 4. The maximum
amplitude begins to change apparently at the 117th hour. After the 160th hour, the
vibration amplitude increases sharply until the system shutdown. Since kurtosis is a
time-domain statistical criterion that is widely used to detect the mechanical transient
faults, the kurtosis values of all the vibration samples were first calculated and then
212 J. Dai et al.
Accelerometers Radial load Thermocouple

Vertical Accelerometer
Bearing 1 Bearing 2 Bearing 3 Bearing 4

Motor
Horizontal Accelerometer
Fig. 3 Experiment platform
Fig. 4 Time-domain vibration data of Bearing 1
normalized to be among the range of [0, 1], as displayed in the blue line with a
triangle in Fig. 5. The curve of the kurtosis versus time exhibits an obvious increase
at about the 108th hour, which is earlier than 117th hour when anomaly is observed in
Fig. 4. The data were finally analyzed by the proposed GAN-based anomaly detection
method. The samples collected from beginning to the 16th hour were used to train the
GAN model. The Adam optimization algorithm with a learning rate of 0.0005 was
adopted during the training process. During the testing process, the residual losses
for all the samples were calculated and then smoothed by a moving-average filter
with window length of 5. The obtained residual loss versus time was also normalized
to be among the range of [0, 1], as displayed in the red line with a circle in Fig. 5. It
can be seen that the residual loss increases visibly at the 85th hour, which is much
earlier than the anomaly time detected by the kurtosis criterion.
In order to confirm that a fault had already occurred at the 85th hour, the corre-
sponding vibration signal was analyzed by the envelope demodulation method. The
0.9 Normalized residual loss obtained by GAN

Normalized Kurtosis
0.8
0.7
0.6
Amplitude
0.5
0.4
0.3
0.2
0.1
0
0 20 40 60 80 100 120 140 160
Time/hours
Fig. 5 Two curves for bearing anomaly detection: kurtosis versus time and residual loss versus
time
waveform and the envelope spectrum of the vibration signal are shown in Fig. 6. A
frequency peak at 230.1 Hz is explicitly identified in the envelope spectrum, which
is close to the bearing characteristic frequency of the outer-race fault, indicating
that a defect had appeared in the outer raceway of the bearing at the 85th hour. The
experimental result validates the superior performance of the proposed GAN-based
anomaly detection method over the traditional time-domain statistical criterion in
bearing incipient fault detection.
0.5
Amplitude/g
-0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time/s
0.15
X: 230.1
Amplitude
0.1
Y: 0.06839
0.05
0
0 100 200 300 400 500 600 700 800 900 1000
Frequence/Hz
Fig. 6 Waveform and envelope spectrum of the vibration signal collected at the 85th hour
214 J. Dai et al.
5 Conclusion
This paper has proposed a new bearing anomaly detection method based on the GAN.
The main innovation is that the proposed GAN model is trained with the absence of
the bearing anomalous samples. The data distribution of the bearing normal samples
can be learned by the GAN model during the training process. When the trained GAN
model is used for bearing condition monitoring, the residual loss will be small if the
bearing is healthy, whereas will be large if a fault occurs in the bearing. The curve of
the residual loss versus time can reflect the propagation of the bearing fault. A life-
circle bearing vibration data validated the effectiveness of the proposed GAN-based
bearing anomaly detection method. Comparing with the traditional time-domain
statistical criterion, the bearing anomaly can be detected earlier by the proposed
method. The proposed method has great significance in incipient weak anomaly
detection of mechanical systems.
Acknowledgements This work was supported in part by the National Natural Science Foundation
of China under Grants 51805342, 51875376, 51875375 and 51705349, in part by the Natural Science
Foundation of Jiangsu Province under Grants BK20180842 and BK20160318, in part by the China
Postdoctoral Science Foundation under Grants 2018M640514, 2017M621811, 2017M611896 and
2019T120456, and in part by the Jiangsu Planned Projects for Postdoctoral Research Funds under
Grant 2018K006B.
References
1. Jia F, Lei Y, Lin J, Zhou X, Lu N (2016) Deep neural networks: a promising tool for fault
characteristic mining and intelligent diagnosis of rotating machinery with massive data. Mech
Syst Signal Process 72–73:303–315
2. Lee YO, Jo J, Hwang J (2017) Application of deep neural network and generative adversarial
network to industrial maintenance: a case study of induction motor fault detection. In: 2017
IEEE International Conference on Big Data (Big Data), Boston, pp 3248–3253
3. Bouzida A, Touhami O, Ibtiouen R, Belouchrani A, Fadel M, Rezzoug A (2011) Fault diagnosis
in industrial induction machines through discrete wavelet transform. IEEE Trans Ind Electron
58(9):4385–4395
4. Guo Y, Na J, Li B, Fung R-F (2014) Envelope extraction based dimension reduction for
independent component analysis in fault diagnosis of rolling element bearing. J Sound Vib
333(13):2983–2994
5. Lei Y, Lin J, He Z, Zuo MJ (2013) A review on empirical mode decomposition in fault diagnosis
of rotating machinery. Mech Syst Signal Process 35(1–2):108–126
6. Tamilselvan P, Wang P (2013) Failure diagnosis using deep belief learning based health state
classification. Reliab Eng Syst Saf 115(7):124–135
7. Zhang W, Li C, Peng G, Chen Y, Zhang Z (2018) A deep convolutional neural network with new
training methods for bearing fault diagnosis under noisy environment and different working
load. Mech Syst Signal Process 100:439–453
8. Zhao M, Kang M, Tang B, Pecht M (2018) Deep residual networks with dynamically weighted
wavelet coefficients for fault diagnosis of planetary gearboxes. IEEE Trans Ind Electron
65(5):4290–4300
9. Zhao R, Yan R, Chen Z, Mao K, Wang P, Gao RX (2019) Deep learning and its applications
to machine health monitoring. Mech Syst Signal Process 115:213–237
10. Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A, Bengio
Y (2014) Generative adversarial nets. In: Advances in neural information processing systems,
Montréal, pp 2672–2680
11. Wolterink JM, Leiner T, Viergever MA, Išgum I (2017) Generative adversarial networks for
noise reduction in low-dose CT. IEEE Trans Med Imaging 36(12):2536–2545
12. Wang C, Xu C, Wang C, Tao D (2018) Perceptual adversarial networks for image-to-image
transformation. IEEE Trans Image Process 27(8):4066–4079
13. Chen X, Xu C, Yang X, Song L, Tao D (2019) Gated-gan: adversarial gated networks for
multi-collection style transfer. IEEE Trans Image Process 28(2):546–560
14. Schlegl T, Seeböck P, Waldstein SM, Schmidt-Erfurth U, Langs G (2017) Unsupervised
anomaly detection with generative adversarial networks to guide marker discovery. In:
International conference on information processing in medical imaging, Boone, pp 146–157
15. Akçay S, Atapour-Abarghouei A, Breckon TP (2019) Skip-GANomaly: skip connected and
adversarially trained encoder-decoder anomaly detection. arXiv Prepr., arXiv1901.08954
16. Wang Z, Wang J, Wang Y (2018) An intelligent diagnosis scheme based on generative adver-
sarial learning deep neural networks and its application to planetary gearbox fault pattern
recognition. Neurocomputing 310:213–222
17. Liu H, Zhou J, Xu Y, Zheng Y, Peng X, Jiang W (2018) Unsupervised fault diagnosis
of rolling bearings using a deep neural network based on generative adversarial networks.
Neurocomputing 315:412–424
18. Radford A, Metz L, Chintala S (2015) Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv Prepr., arXiv1511.06434
Classification/Analysis for Fault or Wear
Diagnosis
Specific Small Digital-Twin for Turbofan
Engines
J. Lacaille
Abstract Ten years ago, we develop a numeric state vector to allow us following
the performance of a small fleet of turbofan engines. As engine manufacturer with
“flight by the hour” new contracts going, we needed to improve our technology;
hence, building analytic digital-twin solutions. This improvement allows to monitor
the engines as well as the behavior of the airlines. It becomes possible to identify use
classes to help design systems better adapted to our clients. The main difficulty to
upgrade this technology was to manage huge amount of data. At the time, the only
observations we had were small 4 Kb snapshots broadcasted by satellite link. And
even with such small amount of observations, the mathematics were limited to an
evolving buffer of one year and a half of flights for a ten aircrafts small company.
Today, we download full flight temporal measurements reaching more than 2 Gb
per flight and we want to manage our fleet of more than 35,000 engines. Moreover,
the life duration of the engine parts increases and the first shop visit is now more
than three years away. The solution was to use new cluster technology and adapt
algorithms. A main element is to be able to reduce the observation dimension hence
allowing mathematic computation, but specifically for each class of behavior. The
compression algorithm is now running an auto-encoder neural network on GPU and
the cluster analysis still uses a self-organizing map but deployed in parallel on a
cluster of computers. We propose to present this new improvement and conclude
with our view about operational digital-twins.
Keywords Monitoring · Turbofan · Classification · Organization
J. Lacaille (B)
Safran Aircraft Engines, Rond-Point René Ravaud, Moissy-Cramayel, France
e-mail: jerome.lacaille@safrangroup.com
DataLab, Safran Aircraft Engines, Moissy-Cramayel, France

https://doi.org/10.1007/978-981-15-9199-0_20
220 J. Lacaille
1 Engine State
Our turbofan engines are complex machines of more than 16,000 parts which should
last at least 40 years producing enough thrust to ensure safe travel of more than 150
passengers, ten times a day every days of the year. Each component is tested for
its life duration and replacement is usually done a lot earlier than needed. In fact,
one may assume than an aircraft turbofan engine has almost definitely only new
components. Hence, a real business exists in the increase of the delay between shop
visits and part replacements. A way to control the state of the engine is to follow its
behavior flight after flight and build a numeric vector of indicators able to discover
specific wear. In fact, one can build many of such vectors respectively corresponding
to each physical or virtual component of the complex propulsion system. A physical
component may be a specific hub of disk blades we want to monitor for corrosion
and a virtual component is a function to optimize like the general performance of
the engine or the unbalance vibration level. Our presentation will mainly be about
performance measured by exhaust temperature (loss of energy) and consumption.
The first step was to collect data during each flight. As we were interested
about the performance which is thrust, consumption, and compressor efficiency, we
retrieved aircraft communication addressing and reporting system (ACARS) data
during takeoff phase (the most stressful for the engine) and cruise (the longest).
Those measurements acquired once for each flight were just a static summary that
we decompose in two categories:
• endogenous data: the measurement representing the health of the subsystem of
interest (here the performance represented by the fuel flow, the normalized ratio
between high pressure and low pressure shaft, different temperatures and internal
pressures),
• exogenous data: the measurements of the acquisition context like the plane
attitude, the environment, the pilot commands, the state of the bleeds, and so
on.
Figure 1 presents a small list of measurements acquired for a first implementation
of the algorithm [1].
The analysis process has three main steps [2]
1. Normalization of the health indicators according to the external context.
2. Smoothing of the residuals signals to get rid of unexplained noise, but without
destruction of interesting patterns like mean biases and trends.
3. Building of a 2D map to represent the state on a computer screen (Fig. 2).
Results are observable positions of each engine at any given time on a nice 2D
map. Hence, the name of the whole algorithm “Engine State Cartography.” Most of
the engines evolve in a normal condition, so their position is in the middle of the map.
However, when something happens, a normal wear of abradable parts, for example,
or deposition of carbon residual on the turbine blades, we clearly see a trend to an
edge of the map. An example of such behavior is given on Fig. 3.
Specific Small Digital-Twin for Turbofan Engines 221
Fig. 1 An example list of ACARS measurements collected during the early implementation of the
cartography algorithm
Fig. 2 Three main steps for the analytic process

222 J. Lacaille
Fig. 3 This drawing represents the temporal evolution of an engine state flight after flight over the
map. It goes from a big red dot to a small blue point. We observe a normal behavior followed by a
classical increase of exhaust gaz temperature (EGT) due to simple and usual wear
We looked at all those kind of events which represent the normal life of a turbofan
engine and ask our mechanical experts for interpretation. On another hand, the use
of a self-organizing map to represent classes of engine behaviors has a big advantage
with the conservation of the topological distance between original state vectors.
Hence, two nearby cells on the map continue to represent nearby engine states in the
real domain. Using this specificity of the Kohonen map [3], it is possible applying a
hierarchical clustering algorithm between cell, using underlying distances, to define
bigger regions with similar engine states. We just had to make the link between expert
interpretation and the big regions to obtain a result like the one presented on Fig. 4.
2 Normalization
The normalization step was one of the most challenging [4]. As we did not want to
loose interpretability, we prefer a classical regression algorithm. But to keep effi-
ciency and robustness, we had to reduce the dimensionality of the exogeneous input.
For dimension reduction, we used a least absolute shrinkage and selection oper-
ator (LASSO) algorithms [5, 6] adding a linear (absolute value) L1 constraint to the
mean square error of the estimation of each health indicator. Equation (1) gives the
expression of this regression where the Yi,r j are the endogenous measurements with
r a health indicators, i an engine number, and j a snapshot. The X i,l j are expert
Fig. 4 After analysis by the experts and a hierarchical clustering of the map cells in bigger regions,
we were able to interpret the edges of the map
physical combinations of input context observations numbered by l from 1 to q such

combinations. The λrl coefficients are the learnt regression parameters.
2

q

q
r
arg min Yi, j −
r
λrl X i,l j where λ < C r (1)
l
λ∈R q i, j l=1 l=1
The use of physical computation X instead of just applying the algorithm to all
context measurements helps insertion of physics in the analysis. The L1 constraint
forces the solution to find values of the parameter at the intersection of an iso-
quadratic error (ellipsoid) and an iso-L1 constraint (hypercube). Clearly as shown on
Fig. 5, the result is mostly obtained by an intersection of an edge which corresponds
to a maximum of zero coefficients.
3 More Data and Time Series
The problem with this algorithm arrises when we get continuous data from the new
engines. The LASSO algorithm is a simple regression with big iterative process and
its complexity increases with the number of observations. Hence, we built an itera-
tive algorithm that combines linear dimension reduction and adaptation to different
224 J. Lacaille
Fig. 5 A nice interpretation of the comparison between usual ridge regression with an L2 constraints
(on the left) to reduce linear dependencies and LASSO regression (on the right) that minimizes the
number of non-null coefficients
contexts of acquisition. This algorithm [7], mixture probabilistic principal compo-

nent analysis (MP-PCA), models a mixture of hyperplane gaussian distributions.
It essentially is a classification of data with each class represented by a projec-
tion of a minimal dimension hyperplane. The calibration of the different classes and
their representation is implemented by an EM algorithm [8] (Expectation–maximiza-
tion) called high dimensional robust clustering (HDRC) with a middle step inserted
between the two standard steps to manage specific outliers during the classification.
• E step is to estimate the aposteriori probability to belong to a class;
• T step trim a given proportion of point with low likelihood;
• M step maximizes the likelihood by optimization of model parameters.
Here, the model is a mixture of flat Gaussian distributions with variance Σk =

Q k Λk Q k for class k described in (2) with the coefficients (ak1 . . . akdk , bk ) organized
in decreasing modules.
⎛⎡ ⎤ ⎞
ak1 ··· 0 ⎫
⎜⎢ . . . .. ⎥ ⎟ ⎬
⎜ ⎣ .. . . ⎦ 0 ⎟
⎜ ⎟ ⎭
dk
⎜ ⎟
⎜ 0 · · · akd
Λk = ⎜ ⎡ ⎤⎟
⎟ ⎫ (2)
⎜ bk · · · 0 ⎟ ⎬
⎜ ⎢ .. . . .. ⎥⎟
⎜ 0 ⎣ . . . ⎦⎟ ( p − dk )
⎝ ⎠ ⎭
0 · · · bk
Parameters are the direction of the hyperplanes Q k and the variance Λk including
its width parameters (characteristic values) and its dimension dk . The number of
classes is controlled with a BIC criterion (Bayesian information criterion) (Fig. 6).
This solution was really efficient and was used to monitor bench tests that produce
very long time sequences of numeric measurements. The trim step (T ) helps identi-
fying outliers with a control over the probability of false alarm [9]. The identification
of outliers as point that are not adapted to this mixture model was used to detect abnor-
malities during development tests and stop the bench before any damage occurrence
[10].
Fig. 6 Illustration of a mixture of two gaussian plan distribution in a 3D space
4 Smoothing and Organizing
Once normalized, the state vector is a representation of health indicators simply

smoothed over a time period adapted to the specific pattern we are looking for.
Performance analysis uses weekly moving averages to smooth the signal. But as we
want to keep the main changes, we use a recursive least square algorithm (RLS)
which automatically detects trend changes and avoid the destruction of important
features [11] (Fig. 7).
Finally, the Kohonen algorithm is used to build the self-organizing map (SOM)
from the smoothed health indicators.
This algorithmic process separates the building of the state vector to its represen-
tation on the map. But our real interest is about map positions which correspond to
labels. Using those successive positions, a given engine trajectory (on the map) is a
label string and it becomes possible to use Levenshtein distance (edit distances) to
Fig. 7 Smoothed version of normalized health indicators by RLS algorithm. The variables shown
are the exhaust gaz temperature (EGT), the corespeed N2 normalized according to the command,
the fuel flow, a static pressure (ps3) before combustion and after compression and the temperature
(t3) at that same station
226 J. Lacaille
compare trajectories, or even part of trajectories and find similar trajectories in the
past of other engines. Such a research is like an expert system using return of exper-
iment [12]. Experiments are the trajectories stored in our database and the expert
request find similar patterns and anticipate future behavior just by looking at what
happens after old similar observations. The edit distance computes the minimal cost
of transformation of a string into another by small replacement of labels. It naturally
uses the distance between cells on the map as unitary cost for a label replacement.
5 Neural Network
Our methodology separates the computation of the state vector of health indicators
from the representation on a map. This is just a mathematic artifact and it should
be possible to adapt at the same time the compression of the input data into state
indicators and its graphic representation. Moreover, until now, the state indicators
were defined by the system specialists and this was possible because of the small
number of input variables. But with the arrival of time series downloaded after each
flight called continuous engine operational data (CEOD), the health descriptors may
be hidden in temporal patterns or multivariate correlations including delais between
different measurements.
However, the classification by Kohonen algorithm uses a similarity measurement
and as it is well-known nowadays, bigger is the space dimension and lesser is the
significance of a distance (this is the crurse of dimensionality). Our idea is to compress
the data with a neural auto-encoder; at the same time, we build the SOM [13, 14].
The auto-encoder is calibrated by minimization of a reconstruction error Lr when
the classification corresponds to the placement of prototype vectors with error Lsom .
Both cost functions are added with a ponderation γ and learnt iteratively as shown
on Fig. 8 by minimization of Lr + γ Lsom .
On this scheme, We is the encoding weight matrix of the neural network, Wd the
decoding weights, and m L the stacked position of the prototypes vectors for the SOM
algorithm.
Fig. 8 Gradient computation of the combination of both errors used simultaneously

Fig. 9 Multivariate data sequence encoding and decoding scheme linked to an autorganizing layer
for visualization
Figure 9 corresponds to the exploitation of time sequences by this neural network.

This scheme may use recurrent versions of neural-networks but we did not test such
solution yet.
Next image (Fig. 10) presents the result of this combined algorithm (named
DESOM for Deep Embedded Self-organizing Map by its author Florent Forest
[13]) on vibration levels measured on engine’s driveshafts during accelerations. The
input data are curves representing the energy levels for each rotation speeds and the
algorithm clearly isolates automatically categories of behaviors for the experts to
analyze.
Fig. 10 Map on the right shows the result of the combined algorithm on dynamic curve observations.
The input data are vibrations energies measured on a radial driveshaft (y-axis of each small curve)
according to the shaft speed during an acceleration
228 J. Lacaille
6 Conclusion
On previous versions of this algorithm, when using small ACARS messages, we rely
on embedded on-board computations to build interesting specific health indicators.
However, as we now have access to a very big amount of temporal series, it becomes
more convenient to build those indicators on the ground using extensive parallel
computing techniques; hence, working simultaneously on a whole fleet of engines.
This way, each engineer may test new ideas, eventually optimizing meta-parameters
like temporal window sizes or ponderations between compression and classification.
But aeronautic engineers will not bother programing such small computations with
a complex framework like Spark even using a new Scala language.
Hence, we proposed a generic and scalable pipeline for large-scale analytics
of operational data oriented toward health-monitoring applications [15]. Based on
Apache Hadoop and Spark, our approach enables domain experts to scale their algo-
rithms and extract features from tens of thousands of flights stored on a cluster. All
computations are performed using the Spark framework; however, custom functions
and algorithms can be integrated without knowledge of distributed programming.
Then, our unsupervised learning algorithm is integrated for clustering and dimen-
sionality reduction of the flight features, in order to allow efficient visualization and
interpretation through a dedicated Web application.
References
1. Lacaille J, Côme E (2011) Visual mining and statistics for a turbofan engine fleet. In: IEEE
aerospace conference
2. Côme E, Cottrell M, Verleysen M, Lacaille J, Aircraft engine health monitoring using self-
organizing maps. In: Industrial conference on data mining
3. Kohonen T (1995) Self-organizing maps, vol 30. Springer
4. Lacaille J (2009) Standardized failure signature for a turbofan engine. In: IEEE aerospace
conference
5. Tibshirani R The lasso : some novel algorithms and applications, pp 1–43
6. Tibshirani R (1996) Regression shrinkage and selection via the lasso. R Stat Soc 58(1):267–288
7. Bellas A, Bouveyron C, Cottrell M, Lacaille J (2012) Robust clustering of high-dimensional
data. In ESANN, pp 25–27
8. Bouveyron C, Girard S, Schmid C (2007) High-dimensional data clustering. Comput Stat Data
Anal 52(1):502–519
9. Lacaille J, Gerez V (2011) Online abnormality diagnosis for real-time implementation on
turbofan engines and test cells. in PHM Society, pp 579–587
10. Lacaille J, Gerez V (2012) A batch detection algorithm installed on a test bench. In PHM
society, pp 1–7
11. Lacaille J, Côme E, Inrets I (2011) Sudden change detection in turbofan engine behavior. In:
CM & MFPT
12. Côme E, Cottrell M, Verleysen M, Lacaille J (2011) Aircraft engine fleet monitoring using
self-organizing maps and edit distance. In: WSOM
13. Forest F, Lebbah M, Azzag H, Lacaille J (2019) Deep embedded SOM: joint representation
learning and self-organization. In: ESANN
14. Forest F, Lebbah M, Azzag H, Lacaille J (2019) Deep architectures for joint clustering and
visualization with self-organizing maps. In: Pacific-Asia conference on knowledge discovery
and data mining (PAKDD/LDRC)
15. Forest F, Lacaille J, Mustapha L, Azzag H (2018) A generic and scalable pipeline for large-
scale analytics of continuous aircraft engine data. In: IEEE International Conference on Big
Data
Fault Diagnosis of Single-Stage Bevel
Gearbox by Energy Operator and J48
Algorithm
V. Kumar, A. K. Verma, and S. Sarangi
Abstract The main objective of this paper is to detect and classify the faults in a
single-stage bevel gearbox. In this paper, energy operator (EO) is combined with the
spectrum and cepstrum analysis is used for detection of faults and for classification
purpose, J48 algorithm is used. Usually, the vibration signal is contaminated with
noises which suppress the fault frequencies in spectrum analysis. So, a filter-free,
simple, and efficient EO method is used to reduce noise and interferences of vibration
signal coming from the gearbox test rig. All the experiments are carried out by varying
the speed and load under three health conditions of gearbox such as healthy, chipped,
and missing tooth. The vibration signal is preprocessed by EO, which is further
analyzed by spectrum and cepstrum analysis in MATLAB to detect the location of
faults. J48 decision tree algorithm is used to select the dominant feature among 12
statistical features which are extracted from the processed signal and based on these
chosen features; the overall classification accuracy is approx. 89% with a precision
of 0.917. The obtained results show that the presented method can detect and classify
the faults of gearbox efficiently.
Keywords Bevel gearbox · Fault diagnosis · Energy operator · Spectrum

analysis · Cepstrum analysis · J48 algorithm
1 Introduction
The gearbox is a complex and critical rotating machine, widely used in automobile,
petrochemical, power, cement, paper & pulp, steel, and sugar industries for the trans-
mission system. Any unexpected failure in the gearbox leads to a sudden breakdown
of machines that increase the downtime and maintenance cost. So, early detection of
V. Kumar (B) · S. Sarangi

Department of Mechanical Engineering, Indian Institute of Technology Patna, Patna, India
e-mail: 1821me18@iitp.ac.in
A. K. Verma
Rolls-Royce@NTU Corporate Lab, Nanyang Technological University, Singapore 639798,
Singapore

https://doi.org/10.1007/978-981-15-9199-0_21
232 V. Kumar et al.
these failures becomes essential in industries to reduce their catastrophic damages

and economic losses. Hence, for that reason, it becomes a significant research area
in the field of condition monitoring. A lot of research is going on this field. A fault
diagnosis technique consists of the acquisition of data, extraction of feature, selection
of feature, and classification of faults. It is capable of detecting failure from machine
parts. Vibrational signal analysis is very prominent in the fault detection because it
carries the signature of faults. In the real world, these signals are contaminated with
a lot of noise coming from the different parts of the machine. Techniques like time
synchronous averaging (TSA), angular domain averaging, filter method, etc., are used
to reduce noises from the vibrational signal, but it requires prior information about
the system. E Mayo et al. described a non-parametric (do not require any information
about the system) fault detection technique based on the energy of oscillations of the
vibrational signal [1]. Kaiser proposed and derived the formula for energy operator
(EO) and described how this method is useful to reduce noise and interference level
[2]. Verma et al. described how spectrum analysis is helpful to predict misalign-
ment faults in an induction motor [3]. Randall described how cepstrum analysis is
effective to predict the defects in the gearbox [4]. Nacib et al. detected the cracks
on gear tooth using cepstral analysis on the helicopter’s gearbox and described how
spectrum analysis is not used for fault detection [5]. Arun Kumar et al. presented
the time domain statistical parameters of the vibrational signal for feature extraction
[6]. Sugumaran et al. implemented these parameters for fault classification by using
a machine learning approach [7]. Asr et al. used the J48 decision tree algorithm and
non-naive Bayesian classifier for the diagnosis of combined faults in the rotating
machine [8]. In this paper, a method is presented to depict the location and types of
faults in the single-stage bevel gearbox, as displayed in Fig. 1. The remaining portion
of the paper is given as follow: Theoretical background of EO and J48 algorithm is
given in Sect. 2. Section 3 describes the experimental setup, while Sect. 4 gives the
detail about the result and discussion. Finally, Sect. 5 outlines the conclusion of this
research.
Fig. 1 Flowchart for single-stage bevel gearbox fault diagnosis

Fault Diagnosis of Single-Stage Bevel Gearbox … 233
2 Theoretical Background
2.1 Energy Operator (EO)
It is a non-parametric, filter-free, simple, and efficient fault detection technique. It

evaluates the energy of oscillations in a vibrational signal in real time because it does
not require any prior information about the system. It is based on the equation of
natural vibration system. The main purpose of applying this method is to reduce the
noise and interference generated during the running condition of the gearbox. The
detail and derivation of a discrete form of EO are described in [1, 2]. The discrete
form of EO of the discrete vibrational signal X n is calculated as
E(X n ) = X n2 − X n−1 X n+1 (1)
2.2 J48 Decision Tree Algorithm
J48 algorithm is a Java application of the Weka tool’s C4.5 algorithm. By using the
idea of information entropy, C4.5 produces the decision tree from the training data set.
It is a structure-like flowchart in which the interior node depicts the feature; the branch
describes the decision rule, and each node of the leaf shows the outcome/class label.
The root node, also known as a decision node is the top node in the decision tree. The
branch begins from the root node to the leaf node consists of many nodes, and each
node represents only one feature. The presence of feature in the decision tree indicates
the significance of that feature. At each decision node, J48 pick out the dominant
feature among the set of features for classification by using proper estimation criterion
based on the entropy and information gain. The maximum information gain feature
is selected to take the decision [7].
2.2.1 Information Gain and Entropy
Both originate from information theory. Information gain is loss in entropy, computed
on the basis of the difference between entropy before splitting and average entropy
after splitting the dataset based on specified feature values. For feature ‘F’,
information gain (E, F) relative to a set of examples ‘E’ is calculated as follows:
|E u |
Information Gain(E, F) = Entropy(E) − Entropy(E u ) (2)
|E|
u∈valus(F)
234 V. Kumar et al.
where value (F) denote the collection of all possible values for feature ‘F’, and ‘E u ’
is the sub-set of ‘E’ corresponds to ‘u’ value of feature ‘F’.
Entropy measures the randomness of the set of examples (E) and calculated by

m
Entropy(E) = −P j × log2 P j (3)
j=1
where ‘m’ denote the number of class, Pj is the probability of class ‘j’, calculated as
the ratio of the class ‘j’ in the set.
3 Experimental Setup
All the experiments are carried out on the gearbox test rig placed in dynamics of
machine laboratory in IIT, Patna, shown in Fig. 2. It is designed to acquire the
vibration signal data from the gearbox by introducing different types of faulty gears
under varying the speed and load condition on it. The test rig consists of an induc-
tion motor (three-phase), coupling, shafts, the pulley that connect the bevel gearbox
with belt drive, magnetic clutch (brake), triaxial accelerometer, and data acquisition
system (DAQ). The variable-speed motor controller is used to control motor speed.
Specification of gearbox test rig has shown in Table 1. A triaxial accelerometer has
installed on the housing of the gearbox. The vibration data has collected through
DAQ at a sampling frequency of 2048 Hz for 50 s by varying the motor speed from
15 to 35 Hz and load conditions (0, 0.452, 0.904 Nm) under three conditions, i.e.,
healthy, chipped, and missing tooth. Total of 27 samples have collected. These data
are imported and processed in MATLAB for further analysis.
Fig. 2 a Bevel gearbox test Rig, b Gear arrangement inside the gearbox
Table 1 Specification of parts used in the gearbox test rig

Number of teeth on driver gear 18
No. of teeth on the driven gear 27
Dia. of pulley connected with motor shaft 0.05 m
(D_PM)
Dia. of pulley connected with driver gear shaft 0.125 m
(D_PG)
Input speed of driver gear (sp) in Hz (D_PM/ D_PG) × Input speed of the motor
(Hz)
Input speed of driven gear (sg) in Hz (18/27) × Input speed of driver gear (Hz)
Gear mesh frequency (GMF) in Hz Number of teeth on driver/driven gear × Input
speed of driver/driven gear
Rahmonics of driver (sp)/driven (sg) gear 1/Input speed of driver (sp)/driven (sg) gear
(Hz)
4 Results and Discussions
4.1 Fault Identification
Figure 3 shows the spectrum analysis on the original signal and EO signal at 25 Hz
motor speed, i.e., 10 Hz driver gear speed and 0.904 Nm load. It is noticed that the
level of noise and interferences is less in the EO signal as compared to the original
signal. The peaks are more precise and easily visible in FFT-EO plot than FFT-ori
sig plot, especially the driver (sp)/driven (sg) frequency that is visible in FFT-EO
plot but not in FFT-ori sig plot.
Fig. 3 a FFT of original signal and EO signal in healthy condition at 10 Hz (sp) and 0.904 Nm
load, b FFT of original signal and EO signal in chipped condition at 10 Hz (sp) and 0.904 Nm load
236 V. Kumar et al.
4.1.1 Spectrum Analysis
Spectrum analysis is instrumental in identifying the characteristic frequencies of

gearbox such as gear mesh frequency (GMF), driver gear/pinion frequency (sp),
driven gear/gear frequency (sg) and their harmonics and sidebands. Details of spec-
trum analysis are given in [3, 9]. The amplitude of GMF is not used for detection
of faults in gearbox because other operating variables such as load can affect this
[5]. But the growing of sidebands of modulated frequency (sp, sg) near GMF and
its harmonics indicate the existence of faults. But it fails to recognize the type and
location of faults as the number of defects increases in the gearbox. So it is essential
to identify the periodicity of these sidebands so that one can identify the location of
faults.
4.1.2 Cepstrum Analysis
Cepstrum analysis overcomes these issues and predicts the location of faults based
on rahmonics generated by the corresponding faulty rotating parts in the gearbox [5].
Fundamental rahmonics (time period) of sp and sg are 0.1 s and 0.15 s, respectively.
Details of cepstrum analysis are given in [4, 9]. Figure 4 shows the cepstrum analysis
at different fault conditions. It observes that the quefrency of 0.3 s is dominant in
all three fault conditions because it is the fundamental rahmonics of the mating of
the same pair of the tooth during meshing (calculated by 3rd rahmonics of sp or 2nd
rahmonics of sg). So, the growth in the amplitude of these peaks in the cepstrum
Fig. 4 Cepstrum analysis of EO signal at 25 Hz and 0.452 Nm load

shows that the location of faults is in the gear pair, which is either from the driver or
driven gear.
If we see the figure carefully, it is noticed that in single-stage bevel gearbox, the
healthy condition have 0.1 and 0.15 s family of rahmonics while in chipped and
missing conditions, only the family of rahmonics of 0.15 s and 0.1 s are observed,
respectively. It is only used to detect the location of faults but fails to classify the types
of faults coming from the same source. Therefore, we need to develop a technique
to classify these types of faults.
4.2 Fault Classification
Totally, 12 features such as peak, minimum, mean, root mean square, variance,
standard deviation, kurtosis, skewness, crest factor, shape factor, clearance factor,
and impulse factor are extracted from the pre-processed signal [8]. Details of these
features are given in [6]. On the basis of information gain and entropy of all extracted
features, dominant features that are contributing much in the fault classification are
selected by J48 algorithm. The details about the J48 algorithm are given in [7].
Figure 5 illustrates the decision tree formed by J48 algorithm in which shape factor
(SF) and minimum (Min) value features are presented in decreasing order of their
dominance. If SF value is greater than 0.006597, then it is correctly classified as
‘Missing Tooth’ fault condition. And when SF is less than or equal to 0.006597 and
Min is greater than—45.073038, then it is classified as ‘Healthy Tooth’. Otherwise,
it is classified as ‘Chipped Tooth’ fault condition of gearbox. So, based on these 2
dominant features, the fault classification of the gearbox are done. Out of the total of
Fig. 5 Decision tree for the selection of features

238 V. Kumar et al.
Table 2 Confusion matrix by

Fault condition Healthy tooth Chipped tooth Missing tooth
J48 algorithm
of the gearbox
Healthy tooth 3 0 0
Chipped tooth 1 1 0
Missing tooth 0 0 4
Table 3 Class-wise detail accuracy by J48 algorithm

Class TP rate FP rate Precision Recall F-measure ROC area
Healthy tooth 1.000 0.167 0.750 1.000 0.857 0.917
Chipped tooth 0.500 0.000 1.000 0.500 0.667 0.893
Missing tooth 1.000 0.000 1.000 1.000 1.000 1.000
Weighted avg 0.889 0.056 0.917 0.889 0.878 0.948
27 samples, 18 samples are used to train the model and rest samples are used to test
it. After testing the 9 samples on the model, the J48 algorithm creates a confusion
matrix, as illustrated in Table 2. The elements on the diagonal of the confusion matrix
indicate the correctly classified samples.
From Table 2, it can be observed that 3 out of 3 samples are correctly classified as
healthy tooth, 2 out of 1 sample is correctly classified as chipped tooth and 1 sample
is misclassified as healthy tooth, 4 out of 4 samples are correctly classified as missing
tooth fault condition of the gearbox. So, 8 samples are correctly classified out of 9
samples and 1 sample is misclassified. Therefore, the fault classification accuracy
developed by the proposed method presented in this paper is near about 89%.
Table 3 shows the class-wise detail accuracy produced by J48 algorithm. TP rate
and FP rate must be close to 1 and 0, respectively, for the effectiveness of the classifier.
Table 3 shows that TP is close to 1 and FP is close to 0, then based on this, we can
predict that our proposed classifier is very effective for fault classification problem.
Precision (Positive predictive value) tells how many selected samples are correctly
classified while recall (sensitivity) tells how many correctly classified samples are
selected during testing. F-measure is the harmonic mean of precision and recall.
ROC area is a graphical representation that shows the performance of a classification
model at all classification thresholds.
5 Conclusion
In this paper, energy operator (EO) was combined with the spectrum and cepstrum
analysis to detect the location of faults in real time, as discussed in Sect. 4.1. But
due to the limitation of the spectrum and cepstrum analysis, it failed to classify
the different types of faults coming from the same source. For that, a classification
technique, J48 decision tree algorithm in WEKA data mining tool was used for
classification of faults in the single-stage bevel gearbox. The overall classification
accuracy through this presented method was found to be 89% with a precision of
0.917, which is acceptable for any fault classifier. So based on the results obtained,
the presented method can be utilized for the practical application and online fault
diagnosis of gearbox efficiently.
References
1. Mayo E, Liang M, Baddour N (2016) Gear fault detection with the energy operator and its
variants. IDETC2016, Charlotte, North Carolina
2. Kaiser JF (1990) On a simple algorithm to calculate the ‘energy’ of a signal. ICASSP-90,
Albuquerque, New Mexico, pp 381–384
3. Verma AK, Sarangi S, Kolekar MH (2013) Misalignment fault detection in induction motor
using rotor shaft vibration and stator current signature analysis. IJMMS 6(5/6):422–436
4. Randall RB (1973) Cepstrum analysis and gearbox fault diagnosis. Brüel and Kjær, Copenhagen,
Application Note No. 13-150
5. Nacib L, Pekpe KM, Sakhara S (2013) Detecting gear tooth cracks using cepstral analysis in
gearbox of helicopters. IJAET 5(2):139–145
6. Arun Kumar M, Manjunath TC (2016) A Brief Review/Survey of Vibration Signal Analysis in
Time Domain. SSRG-IJECE 3(3):12–15
7. Sugumaran V, Muralidharan V, Ramachandran KI (2007) Feature selection using decision tree
and classification through proximal support vector machine for fault diagnostics of roller bearing.
MSSP 21:930–942
8. Asr MY et al (2017) Diagnosis of combined faults in rotary machinery by non-naive bayesian
approach. MSSP 85:56–70
9. Tuma J (2014) Measurement, signal analysis, signal processing and noise reduction measures.
In: Vehicle gearbox noise and vibration. Wiley
An Efficient Multi-scale-Based
Multi-fractal Analysis Method to Extract
Weak Signals for Gearbox Fault
Diagnosis
Ruqiang Yan, Fei Shen, and Hongxing Tao
Abstract Weak fault signals are always embedded in mass vibration noise in many
gear systems, thus making difficulty in gear fault diagnosis. In order to extract weak
fault signals, a new multi-scale-based multi-fractal analysis (MMA) method is intro-
duced in this paper, which is based on classical multi-fractal detrended fluctua-
tion analysis (MFDFA) framework. Firstly, the Hurst surface features are utilized
to describe the characteristics with multifractal of the vibration signal, which have
been proved to be sensitive to the dynamical responses of the various gear faults.
Secondly, a moving fitting window is added to the MFDFA framework to sweep
through all the range of the scales, and then obtain final multi-scale features, whose
purpose is to magnify those features in some important scales and weaken the rest
scales. In addition, other techniques, such as the distance-based feature selection and
the random forest (RF) classifier, are also introduced into the gearbox fault diag-
nosis procedure to verify the effectiveness of extracted features for differentiating
various gear states. Experiments using the Qianpeng testbed (QT) prove that the
MMA method can effectively extract weak signals, and has higher diagnostic accu-
racy than other algorithms, such as empirical mode decomposition (EMD), wavelet
transform (WT), and classical MFDFA.
Keywords Multi-scale-based multi-fractal analysis · Weak signals · Gearbox fault

diagnosis · Random forest
1 Introduction
Gear plays an important role in the mechanical drive system. However, frequent
breakdowns appear in gears when machine works in high speed and heavy load, and
R. Yan (B) · F. Shen · H. Tao

School of Mechanical Engineering, Xi’an Jiaotong University, Xi’an, Shanxi, China
e-mail: yanruqiang@xjtu.edu.cn
R. Yan
School of Instrument Science and Engineering, Southeast University, Nanjing, China

https://doi.org/10.1007/978-981-15-9199-0_22
242 R. Yan et al.
local damages may lead to failure of whole equipment, finally rise up to economic
losses. Gearbox fault diagnosis (GFD) technology aims to detect early gear faults by
analyzing related signals of gearbox and then remind the operator [1, 2], thus being
important to prevent potential losses. However, the feature generated by gear damages
may be embedded in strong ambient noise, thus making difficulty in GFD. So, it is
necessary to choose proper technique to extract the weak features. Qin [3] discussed
a new automatic sparse representation method for detecting weak transients of gears.
Dhamande [4] proposed a gear-bearing compound fault features extracted from
continuous and discrete wavelet transform of vibration signal. Wang [5] designed a
gearbox fault feature extraction strategy based on multi-point kurtosis (MKurt) and
multi-point optimal minimum entropy deconvolution adjusted (MOMEDA). Yong
[6] successfully extracted gear feature images with multi-central frequencies and
vibration signal frequency spectrum. Methods above are effective in extracting weak
gear fault features.
In this study, we will explore a multi-scale-based multifractal analysis (MMA)
method, which is based on the classical multi-fractal detrended fluctuation analysis
(MFDFA) [7], which is of then used to detect the long-range correlations of time
sequences, so it can be used to analyze the nature of the average fluctuations of the
data after removing the local trends. In [8], the MFDFA framework was applied to
diagnose the rolling bearings faults and achieved good results, but still depending
on selected scales. Therefore, the scale will be considered as a variable in our MMA
method. In this way, the obtained Hurst surface can describe different fluctuations
of magnitudes in several scales, and is conducive to extract weak features of gear
vibration signals. In addition, other techniques, such as the distance-based feature
selection and the random forest (RF) classifier [9], will be also introduced to build
an entire GFD model.
The structure of the paper is shown as follows. Section 2 introduces the proposed
GFD framework and some related works, especially the distance-based feature selec-
tion and the RF classifier. Section 3 gives the basic MFDFA and proposed MMA
algorithm. Section 4 validates the effectiveness of the technique using Qianpeng
testbed (QT), followed by a conclusion in Sect. 5.
2 The GFD Framework and Related Works
2.1 Proposed GFD Framework
Gearbox vibration signals consist of time-related sinusoidal component, fault-related

periodic component, and noise. Generally, the signal variation due to gear faults
can be reflected by some time domain indicators or frequency domain indicators.
The purpose of GFD framework is to extract weak features and detect gear faults
effectively, as shown in Fig. 1. Firstly, the vibration signals are collected using sensors
which are installed around gearbox. Secondly, the Hurst surfaces are obtained with
An Efficient Multi-scale-Based Multi-fractal … 243
Fig. 1 Proposed GFD framework
MMA method to extract fault features. Thirdly, high-quality features are selected in
the Hurst surfaces, which are considered as the analyzed datasets. Finally, the random
forest classifier is adopted to identify final fault categories. In this GFD framework,
MMA method plays an important role because weak faults signals are separated from
strong noise at this stage, thus contributing to fault classification.
2.2 Distance-Based Feature Selection
The number of MMA features is large, and most of them are similar even in different
categories. If all redundant information is included, not only will the diagnostic
accuracy be reduced but also the time cost will be too high. So, a distance-based
method is utilized to identify effective features, which is based on the consideration
that the distance between internal samples is less than external samples. At first,
the evaluation factors of all features are calculated with a feature evaluation crite-
rion and are arranged in descending order. Then, a threshold of evaluation factor
is designed to decide the number of features. We assume that C categories exist
in target feature set: f m,c, j , m = 1, 2, . . . , Mc ; c = 1, 2, . . . , C; j = 1, 2, . . . , J ,
where f m,c, j represents the jth feature of the mth sample in c category; Mc repre-
sents the samples number in c category; J represents the number of features. The
calculation process of distance-based feature selection method is expressed by:
1 Mc

dc, j = qm,c, j − ql,c, j , l, m = 1, . . . , Mc , l = m
Mc (Mc − 1) l,m=1

C

d (w)
j = dc, j /C, v(w)
j = max dc, j /min dc, j (1)
c=1
244 R. Yan et al.
where dc, j represents the average distance between samples in c category; d (w)
j repre-
(w)
sents the average distance in C total categories; v j represents the different factors
between the same category. Then, related distance metrics are further defined and
calculated by
1 cM
1
C

u c, j = × f m,c, j , d (b) u c, j − u e, j
j =
Mc m=1 C(C − 1)
c, e = 1
c = e

v(b) − u e, j , (c, e = 1, 2, . . . , C; c = e)
j = max u c, j − u e, j /min u c, j (2)
where u c, j represents the average of the jth feature in c category; d (b)

j represents the
(b)
average distance of each sample in cross categories; v j represents cross factors of
the distance between different categories. Finally, the evaluation factor is defined by:

λ j = 1/ v(w) (w)
j /max v j + v(b) (b)
j /max v j , α j = λ j d (b) (w)
j /d j , β j

= λ j α j /max α j (3)
where λ j represents the weighting factor of the jth feature; α j represents the radio
of the distance between the same category and the cross categories; β j ∈ [0, 1]
represents normalized evaluation factor, which reflects the contribution of the jth
feature.
2.3 Random Forest Classifier
The random forest (RF) is a supervised classifier where more than one decision
tree is used to create the forest, which is more robust when more trees is used in
decision-making process. The margin function of the RF can be expressed as
mr(X, Y ) = PΘ (h(X, Θ) = Y ) − max PΘ (h(X, Θ) = j) (4)

j=Y
where mr(X, Y ) represents margin function; h(X, Θ) represents the output of classi-
fier; PΘ (·) represents error function; X represents the feature vector, and Y represents
the category. RF algorithm is an efficient model that fuses multiple decision trees
and organizes them into a single tree with the highest classification accuracy. Firstly,
given original dataset D1 with m × n matrix, a new dataset D2 is created from D1
by sampling and removing 1/3 of the row data. Secondly, model is trained from the
reduced samples and the unbiased error is calculated. Thirdly, n 1 column is selected
from total n columns at each node of D2 , where n 1 ≤ n. Finally, the best clas-
sification accuracy is obtained by simultaneously growing more than one tree and
the collection of individual decisions. Overall, RF fits many classification trees to a

dataset, and is widely used for machine fault diagnosis.
3 Multi-scale-Based Multi-fractal Analysis Method
3.1 Basic MFDFA Method
MFDFA was developed from detrended fluctuation analysis (DFA) [10], which aims
to describe the fluctuations of magnitudes and to extract most information of the
signals using the Hurst exponent. Given vibration sequence xk with N points, the

i
profile of xk is first defined by Y (i) ≡ [xk − x̄], i = 1, . . . , N , where x̄ is the mean
k=1
of xk , but it is not compulsory because the linear component will also be eliminated
in following steps. Secondly, Y (i) is divided into Ns = int(N /s) non-overlapping
local segments with length s. Meanwhile, Ns non-overlapping local segments are
constructed from the end in the same way to get 2 Ns local segments. Thirdly, the
least squares fit (LSF) equation [11] is utilized to calculate the local trend of these
segments by

s
F (s, v) =
2
{Y [(v − 1)s + i] − yv (i)}2 /s, v = 1, . . . , Ns
i=1

s
F 2 (s, v) = {Y [N − (v − Ns )s + i] − yv (i)}2 /s, v = Ns + 1, . . . , 2Ns (5)
i=1
where v represents the number of segment and yv represents the fitting polynomial
in the vth segment. Fourthly, the qth order fluctuation function is calculated based
on the average of all segments by
2N 1/q
s q/2
Fq (s) ≡ F (s, v)
2
/2Ns (6)
v=1
where q ∈ R represent the index parameters of MFDFA algorithm. If q = 2, the

MFDFA algorithm equates to the classical DFA. Finally, the scaling behavior of
Fq (s) is determined by analyzing the log-log plots for each q, by
Fq (s) ∼ s H (q) (7)
where H (q) represents the Hurst exponent, which depends on the value q. H (q) ∈
(0, 0.5) means antipersistency of the time series, H (q) = 0.5 means uncorrelated
246 R. Yan et al.
noise, H (q) ∈ (0.5, 1) means the persistency of the time series, and H (q) ≥ 1.5
means the integrated white noise.
3.2 Multi-scale-Based Multi-fractal Analysis Method
In MFDFA framework, the whole spectrum of Hurst exponents is not the only result.
Therefore, the scale will be considered as a variable in our MMA algorithm to
different fluctuations of magnitudes in several scales. Here, a moving fitting window
is added to sweep through all the range of the scales s when the function Fq (s) is
ploting. We define new Hurst exponent as follows

H (q, s) = log Fq (s) Ri / log[s Ri ] (8)
where Fq (s) Ri and s Ri represent points falling into window Ri. When fluctuation
functions Fq (s) are presented in logarithmic form, the moving fitting window expands
logarithmically, thus tending to be constant width.
For clarity, the points of the MMA graph are connected to form a colored surface.
Different from the plot of Fq (s), which is plotted with a logarithmic axis, the MMA
graph is presented using the linear axis. Notice that, two key parameters, including
the width of segment (ws ) and the length of slide (ls ), decide the shape of MMA graph
[12]. In this case, the ratio between the right endpoint and left endpoint represents
the value of ws . High quality ws contributes to get more fault information accurately,
and to improve the diagnostic performance. And high quality ls is necessary to avoid
overlapping and fluctuations of MMA graph. After obtaining MMA graph, the values
in Hurst surface can be considered as effective features for follow-up classifier. So,
those weak fault signals, which are embedded in mass vibration noise, can be finally
magnified and extracted.
4 Results and Analysis
4.1 Descriptions of Datasets
As shown in Fig. 2, dataset was collected from the Qianpeng testbed (QT) provided
by Xi’an Jiaotong University, which consists of motor, gearbox, and break. Here,
a driving gear and a driven gear were installed inside gearbox and four channels
were collected. Testing gears contain five fault categories: normal condition (NC),
light crack fault (LCF), small crack fault (SCF), medium crack fault (MCF), heavy
crack fault (HCF), and missing tooth fault (MTF). The sampling frequency was set
as 8192 Hz and the rotating speed was set as 30 Hz.
Fig. 2 Qianpeng testbed and faults contained in gears
4.2 Experimental Setup
• Datasets setup, datasets are listed as: (1) Selected sensor: total four channels;
(2) Selected faults: total six categories; (3) The number of training samples: 100;
(4) The number of test samples: 100; (5) The length of signal: 8192 point (one
second).
• MMA setup, related parameters are listed as: (1) q: −5 to 5; (2) ws = 20; (3) ls
= 2; (4) s: 10 to 400.
• RF setup, related parameters are listed as: (1) n 1 = 0.5 n; (2) Channel fusion:
total four sensors.
• Comparison methods, (1) Wavelet transform (WT) [13]; (2) Empirical mode
decomposition (EMD) [14]; (3) MFDFA model.
4.3 Experimental Results
Firstly, Fig. 3 gives Hurst surfaces under six different fault categories. From these
MMA graphs, it can be found that the feature differences exist between different
fault types, especially when the scale is small (0–100). Meanwhile, Fig. 3b shows
that MMA method can effectively extract light crack fault (LCF) from signals, which
proves the superiority of the proposed method. After feature extraction, those Hurst
values are analyzed with distance-based feature selection, which also indicates that
features in low scale (0–100) have more value than high scale (100–500), so the
low-scale features are selected.
Secondly, Table 1 gives confusion matrixes of GFD with single sensor (3#) and
fused sensors. Two conclusions can be obtained: (1) By comparing the performance
of single and fused sensors, the diagnostic accuracy of the latter (98.83%) is higher
than the former (90.00%), which shows beneficial effects of RF stratrgy. (2) By
248 R. Yan et al.
Fig. 3 Hurst surfaces under six kinds of fault categories
Table 1 Confusion matrixes of GFD with single sensor and fused sensors
Fault Single sensor (3#) Fused sensors (1#, 2#, 3#, 4#)
LCF SCF MCF HCF MTF NC LCF SCF MCF HCF MTF NC
C1 99 1 99 1
C2 0 97 3 100
C3 1 93 1 5 1 99
C4 5 80 15 98 5
C5 2 22 76 3 97
C6 5 95 100
comparing between six faults categories, minor gear faults (NC and LCF, 97.00%)
have higher diagnostic accuracy than major gear faults (HCF and MTF, 78.00%),
which indicates that MMA graphs has better promotion in major gear faults.
Finally, Table 2 compares the GFD accuracy of four feature extraction models.
From this table, it can be seen clearly that MMA with RF model has the highest
Table 2 The comparion of GFD accuracies between four models

Models LCF SCF MCF MTF NC Total
WT + RF 0.98 0.94 0.79 0.88 0.95 0.908
EMD + RF 1 0.96 0.74 0.72 0.87 0.858
MFDFA + RF 0.98 0.97 0.90 0.95 0.93 0.946
MMA + RF 0.99 0.99 0.95 0.98 0.96 0.974
fault diagnostic accuracy, which improves 2.8% performance than MFDFA with RF
model, especially in MCF and MTF types. However, the time cost of MMA with
RF model is higher than that of MFDFA with RF because more than one scale
is calculated and compared in MMA algorithm, so the feature selection becomes
particular importance. Overall, the proposed method can describe different fluctua-
tions of magnitudes in several scales, and is conducive to extract weak features of
gear vibration signals.
5 Conclusions
The new multi-scale-based multi-fractal analysis method is introduced to extract

weak signals for gear fault diagnosis, where the scale is considered as a variable. By
observing Hurst surfaces, the feature differences exist for different fault categories,
especially when the scale is small. Meanwhile, GFD model based on fused sensors
with RF classifier shows better gear fault diagnostic performance than the model
based on single sensor. In addition, MMA model has the highest diagnositic accuracy
when it is compared with classical feature extraction algorithms such as WT, EMD,
and MFDFA, especially in MCF and MTF types. Therefore, the proposed method
is expected to be a useful tool to solve the practical problem of gear fault diagnosis
under strong noise interference.
Acknowledgements The project is supported by National Natural Science Foundation of China

(51575102) and Jiangsu Postgraduate Research Innovation Program (KYCX18_0075).
References
1. Sun R et al (2018) Gear fault diagnosis based on the structured sparsity time-frequency analysis.
Mech Syst Sig Process 102:346–363
2. Hartono D, Halim D, Roberts GW (2019) Gear fault diagnosis using the general linear chirplet
transform with vibration and acoustic measurements. J Low Freq Noise Vib Act Control
38(1):36–52
3. Qin Y, Zou J, Cao F (2018) Adaptively detecting the transient feature of faulty wind turbine
planetary gearboxes by the improved kurtosis and iterative thresholding algorithm. IEEE Access
6:14602–14612
4. Dhamande LS, Chaudhari MB (2018) Compound gear-bearing fault feature extraction using
statistical features based on time-frequency method. Meas J Int Meas Confed 125:63–77
5. Wang Z et al (2017) Composite fault feature extraction of gear box based on MKurt-MOMEDA.
Zhendong Ceshi Yu Zhenduan/J Vib Meas Diagnosis 37(4):830–834
6. Li Y et al (2018) Planetary gear fault diagnosis via feature image extraction based on multi
central frequencies and vibration signal frequency spectrum. Sensors 18(6)
7. Rizvi SAR et al (2014) An analysis of stock market efficiency: developed vs islamic stock
markets using MF-DFA. Phys A Stat Mech Its Appl 407:86–99
8. Liu H, Wang X, Lu C (2015) Rolling bearing fault diagnosis based on LCD-TEO and
multifractal detrended fluctuation analysis. Mech Syst Signal Process 60:273–288
250 R. Yan et al.
9. Han T et al (2018) Comparison of random forest, artificial neural networks and support vector
machine for intelligent diagnosis of rotating machinery. Trans Inst Meas Control 40(8):2681–
2693
10. Thaxton CS et al (2018) Detrended fluctuation analysis and entropy-complexity causality
analysis of temperatures in an urbanized mountain stream. Stoch Environ Res Risk Assess
32(3):843–858
11. Möglich A (2018) An open-source, cross-platform resource for nonlinear least-squares curve
fitting. J Chem Educ 95(12):2273–2278
12. Shi W et al (2014) Multiscale multifractal detrended cross-correlation analysis of financial
time series. Phys A Stat Mech Its Appl 403:35–44
13. Huang W, Kong F, Zhao X (2018) Spur bevel gearbox fault diagnosis using wavelet packet
transform and rough set theory. J Intell Manuf 29(6):1257–1271
14. Park S, Kim S, Choi JH (2018) Gear fault diagnosis using transmission error and ensemble
empirical mode decomposition. Mech Syst Sign Process 108:58–72
A Novel Method of PEM Fuel Cell Fault
Diagnosis Based on Signal-to-Image
Conversion
Zhongyong Liu, Weitao Pan, Yousif Yahia Ahmed Abuker, and Lei Mao
Abstract This paper proposes an image processing-based approach for the fault
diagnosis of polymer electrolyte membrane (PEM) fuel cells. As more abundant
information is contained in the image than that that in 1D signal, features representing
PEM fuel cell faults could be better highlighted with the image. Experimental data
from a PEM fuel cell system at different states, including flooding and dehydration
scenarios, is used to validate the proposed method. By converting the PEM fuel cell
voltage signal into a 2D grey image, several features are extracted from the image,
their performance in discriminating different PEM fuel cell states is investigated, and
two optimal features are determined for fault diagnosis. Moreover, the diagnostic
performance of optimal features from grey image is compared with features from
PEM fuel cell voltage. Results demonstrate that better diagnostic performance could
be obtained with the proposed method.
Keywords PEM fuel cell · Fault diagnosis · Signal-to-image conversion · Feature

extraction
1 Introduction
As non-renewable energy sources like fossil energy become scarcer, the development
and utilization of alternative energy sources receive much more attention in the last
few decades. Among these alternative energies, hydrogen and fuel cell technology,
especially PEM fuel cell, attract more interest due to its properties such as zero
pollution and high efficiency, which lead to the utilization of PEM fuel cell systems
in many fields, including stationary power station, automotive, and even consumer
Z. Liu · W. Pan · Y. Y. A. Abuker · L. Mao (B)

Department of Precision Machinery and Instrumentation, University of Science and Technology
of China, Hefei, Anhui 230027, China
e-mail: leimao82@ustc.edu.cn
L. Mao
Key Laboratory of Precision Scientific Instrumentation of Anhui Higher Education Institutes,
University of Science and Technology of China, 230027 Hefei, China

https://doi.org/10.1007/978-981-15-9199-0_23
252 Z. Liu et al.
devices [1]. However, limited durability and reliability of PEM fuel cell systems still
restrict the large-scale use [2].
As multi-domain knowledge like electrochemical, material, structure, fluid and
thermal engineering, is included in PEM fuel cells, various PEM fuel cell faults could
be experienced during its operation, including membrane electrode assembly (MEA)
contamination, bad water management [3], reactant starvation [4], etc. Therefore, it is
necessary to detect and isolate different PEM fuel cell faults, so that effective mitiga-
tion strategies could be adopted for its performance extension. For the above purpose,
a set of researches has been developed regarding PEM fuel cell fault diagnosis,
including model-based technologies, data-driven methods, and knowledge-based
approaches [5, 6]. In model-based methods, a numerical model should be established
to simulate PEM fuel cell and its ancillary systems, and the fault diagnosis could
then be performed by calculating the deviation between model outputs and experi-
mental measurements [7]. However, the complexity of PEM fuel cell system brings
great challenges in developing accurate models [8, 9]. Regarding the data driving
methods, signal processing techniques would be applied to experimental measure-
ments for feature extraction, and faults could be identified with pattern recognition
algorithms [10–13]. Knowledge-based approaches could integrate prior information
and expert knowledge regarding PEM fuel cell performance degradation mechanisms
based on which the inference analysis could be performed to identify various faults
[14].
It should be mentioned that various methods proposed for PEM fuel cell fault
diagnosis always use one-dimensional (1D) data, especially fuel cell voltage, which
may indicate PEM fuel cell performance decay, but may not contain distinguishable
information for various faults. As a possible solution, multiple 1D data will be assem-
bled for the fault diagnosis, causing the increased computational time for the analysis
and would not applicable for on-line fault diagnosis purpose. Compared to 1D data,
more abundant information is contained in the two-dimensional (2D) image, which
could provide more information for fault isolation. The fault diagnostic method based
on 2D images have been used in several systems [15–17], but their application in
PEM fuel cell fault diagnosis is still in the infancy. Therefore, it is urgently required
to propose image-based PEM fuel cell fault diagnostic techniques and investigates
their effectiveness.
In the study, a novel 2D image-based PEM fuel cell diagnostic techniques is
proposed, and its performance in identifying PEM fuel cell water management issues
is investigated using test data from a PEM fuel cell system. The paper is organized as
follows. Section 2 presents the methodology of transferring 1D fuel cell voltage signal
to 2D grey image, and several commonly used image features will also presented. In
Sect. 3, the tested PEM fuel cell system and corresponding test data are described.
Section 4 investigates the performance of PEM fuel cell fault diagnosis with 2D
image features, and provides the comparison to that from 1D voltage feature. Finally,
conclusions are made in Sect. 5.
A Novel Method of PEM Fuel Cell Fault … 253
2 PEM Fuel Cell Data Generation and Feature Extraction

Techniques
2.1 PEM Fuel Cell 2D Image Generation Technique
In this study, difference method is used to obtain the approximated first-order deriva-
tive of 1D discrete signal (PEM fuel cell voltage herein), which could be expressed
as follows.
yx = [ y(2) − y(1), y(3) − y(2), . . . , y(n) − y(n − 1) ] (1)
where y(i), i = 1, …, n, denotes 1D time-domain voltage signal.

With first-order difference of PEM fuel cell voltage from Eq. (1), 2D grey image
data could then be calculated, which could express more details contained in fuel
cell voltage signal.
In the analysis, N 2 consecutive signal points are selected from yx by sequnence,
which is normalized ranging from 0 to 255, and then converted into an grey image
with Eq. (2). This has been widely used in previous studies for image conversion
[15–17] and could be written as below.

yx (( j − 1) × M + k) − min(yx )
pixel( j, k) = round × 255 (2)
max(yx ) − min(yx )
where pixel denotes the pixel strength of an image.
2.2 Image Feature Extraction Techniques
In this section, several commonly used image features are presented, with which the
performance of proposed method in identifying various PEM fuel cell faults could
be investigated.
2.2.1 Image Grey Histogram Statistical Features
As images consists of pixels of different grey values, the distribution of grey values
could be used to represent the images, which could be achieved with image grey
histogram and expressed below.
n(i)
H (i) = (3)
M×N
254 Z. Liu et al.
where i represents the grey level of the pixel, n(i) is the number of pixels with pixel
level i, and M × N term is the total number of pixels in the image. From previous
studies [18], several features could be generated and used for image classification,
including variance, skewness, kurtosis, and entropy, which are written as follows.

L−1
Variance : σ = 2
(i − μ)2 H (i) (4)
i=0
where μ is average of image pixels, L-1 is the highest pixel level of the image.
1
L−1
Skewness : μs = (i − μ)3 H (i) (5)
σ 3 i=0
1
L−1
Kurtosis : μk = (i − μ)4 H (i) − 3 (6)
σ 4 i=0

L−1
Entropy : μg = H (i) log2 [H (i)] (7)
i=0
2.2.2 Image Moment Features
The image is seen as a flat object, and the value of each pixel is considered to be the
density there. The expectation of a point is the moment at which the image is at that
point. Commonly used image moment features include gravity center and orientation
angle, which could be expressed as below [19, 20].
M10
Gravity centre : xc = (8)
M00
M01
yc = (9)
M00

where M00 = V (i, j) is zero moment, M10 = i × V (i, j) and M01 =
i j i j
j × V (i, j) are the first moments.
i j
⎛ ⎞
1 2× M M11
− x c × yc
Orientation angle : θ = arctan⎝ M20 ⎠
00
(10)
2 M
+ M02
M
− xc − yc
00 00

where M20 = i 2 × V (i, j), M02 = j 2 × V (i, j), M11 = i×j×
i j i j i j
V (i, j) are the image second moments.
3 Description of PEM Fuel Cell Test
In this study, a PEM fuel cell system rated at 80 W was used, which consists of fuel
cell stack, coolant systems, air, and hydrogen supply systems. More details about the
tested PEM fuel cell system could be found in [21].
Two tests were performed using the aforementioned test bench to obtain test data
about two PEM fuel cell faults, i.e., flooding and dehydration. Temperature of the
fuel cell was reduced to produce a flooding effect, where liquid water was condensed
and accumulated inside the fuel cell. On the other test, dehydration was achieved by
increasing the fuel cell temperature, where inlet reactant humidity was decreased,
and dry membrane was obtained. Figure 1 shows the test data from PEM fuel cell
flooding and dehydration scenarios, respectively.
It could be found that both faults could cause PEM fuel cell voltage drop, but
the decay shape shows difference, which could be highlighted with 2D image and
used for fault isolation. It can be found from above figure that in the flooding case,
the voltage drop is proportional to that of temperature change; while at dehydration
scenario, initial temperature change would not cause clearly voltage drop, while with
further rising the fuel cell temperature, the voltage drop rate becomes more sensitive
to the temperature variation.
Fig. 1 Test data from PEM fuel cell flooding (a) and dehydration (b) scenarios
256 Z. Liu et al.
4 Performance of Proposed Method in PEM Fuel Cell Fault

Diagnosis
In the study, the voltage signal from each scenario is decomposed to nine segments;
with Eqs. (1–2), a total 18 sub-images could be obtained. As the sampling frequency
of raw voltage signal is 4 Hz, the time corresponding to each image contains about
100 s voltage data, during which the fuel cell state would not change significantly.
As described in Sect. 2, several features could be extracted from the image and
used for fault diagnosis; the performance of these features in classifying two PEM
fuel cell faults is investigated, from which two optimal features could be determined.
For this purpose, the two faulty voltage signals are converted to the 2D images,
respectively, without segmentation, as shown in Fig. 2. Table 1 lists the performance
of various features in two faulty scenarios. It could be observed that skewness and
kurtosis could show the most difference at two scenarios, indicating that these two
features have the best performance in classifying two faults, thus should be used in
the following analysis.
The performance of selected features is further investigated using 18 sub-images
obtained from the raw voltage signal. Results are depicted in Fig. 3a; it can be found
clearly that the two PEM fuel cell faults could be discriminated with good quality,
and early-stage fault could also be isolated with proposed technique.
Fig. 2 Converted image from flooding (a) and dehydration (b) scenarios
Table 1 Performance of various features in fault classification

Skew Kurtosis Variance Entropy Image center Orientation angle
Flooding −4.93 39.25 217.01 0.2037 30.55, 30.50 0.4912
Dehydration −1.33 12.43 498.71 0.42 30.48, 30.48 0.36
Fig. 3 Classification results of flooding and dehydration scenarios with image-based processing
method (a) and wavelet packet energy method (b)
Furthermore, the fault diagnosis is performed using 1D voltage signals, where

the wavelet packet transform is applied to extract the features from voltage signal,
and two features with the highest energies are selected for fault diagnosis, more
details could be found in [22]. Similarly, each voltage signal is decomposed into
nine segments in the analysis, and the diagnostic result is shown in Fig. 3b. By
comparing with those from proposed method, it can be concluded that although
some faults could be distinguished using 1D voltage signal, the early-stage fault
could not be classified correctly, which hinders the use of voltage signal in practical
applications. This further confirms that the converted 2D image could contain more
distinguishable information, and could provide better diagnostic performance in PEM
fuel cell systems.
5 Conclusions
In this study, a novel method is proposed for PEM fuel cell fault diagnosis using
converted 2D grey images, and the performance of proposed method in identifying
various PEM fuel cell faults is investigated using test data from a PEM fuel cell
system. In the analysis, the diagnostic performance of various commonly used image
features is studied, from which two features of skew and kurtosis are selected as they
could provide the best classification results. The performance of selected features in
PEM fuel cell fault diagnosis is further investigated using the segmented images, and
results demonstrate the early-stage fuel cell fault could also be isolated with good
quality, which could be beneficial in practical applications, as mitigation strategies
could be taken to recover the fuel cell performance at the early fault stage. Further-
more, the diagnostic results using proposed method is compared with those using 1D
258 Z. Liu et al.
voltage data. From the results, the proposed method could provide better diagnostic
results, indicating that more distinguishable information is contained in 2D image
data.
Acknowledgements This work is supported by ‘Hundred of Talents Program of Chinese Academy

of Sciences—Young Talents (2018–2050)’, and ‘Anhui Provincial Natural Science Foundation
1908085ME161’.
References
1. Lipman TE, Edwards JL, Kammen DM (2004) Fuel cell system economics: comparing the
costs of generating power with stationary and motor vehicle PEM fuel cell systems. Energ
Policy 32:101–125
2. Wu JF et al (2008) A review of PEM fuel cell durability: degradation mechanisms and mitigation
strategies. J Power Sources 184:104–119
3. Yousfi-Steiner N et al (2008) A review on PEM voltage degradation associated with water
management: impacts, influent factors and characterization. J Power Sources 183:260–274
4. Yousfi-Steiner N et al (2009) A review on polymer electrolyte membrane fuel cell catalyst
degradation and starvation issues: causes, consequences and diagnostic for mitigation. J Power
Sources 194:130–145
5. Petrone R et al (2013) A review on model-based diagnosis methodologies for PEMFCs. Int J
Hydrogen Energy 38:7077–7091
6. Zheng Z et al (2013) A review on non-model based diagnosis methodologies for PEM fuel cell
stacks and systems. Int J Hydrogen Energy 38:8914–8926
7. Hissel D, Candusso D, Harel F (2007) Fuzzy-clustering durability diagnosis of polymer
electrolyte fuel cells dedicated to transportation applications. IEEE Trans Veh Technol
56:2414–2420
8. Mann RF et al (2000) Development and application of a generalised steady-state electrochem-
ical model for a PEM fuel cell. J Power Sources 86:173–180
9. Steiner NY et al (2011) Diagnosis of polymer electrolyte fuel cells failure modes (flooding &
drying out) by neural networks modeling. Int J Hydrogen Energy 36:3067–3075
10. Damour C et al (2015) Polymer electrolyte membrane fuel cell fault diagnosis based on
empirical mode decomposition. J Power Sources 299:596–603
11. Hua JF et al (2011) Proton exchange membrane fuel cell system diagnosis based on the signed
directed graph method. J Power Sources 196:5881–5888
12. Li Z et al (2014) Online diagnosis of pemfc by combining support vector machine and fluidic
model. Fuel Cells 14:448–456
13. Pahon E et al (2016) A signal-based method for fast PEMFC diagnosis. Appl Energy 165:748–
758
14. Riascos LAM, Simoes MG, Miyagi PE (2007) A Bayesian network fault diagnostic system for
proton exchange membrane fuel cells. J Power Sources 165:267–278
15. Do V, Chong UP (2011) Signal model-based fault detection and diagnosis for induction motors
using features of vibration signal in two-dimension domain. Strojniski Vestnik-J Mech Eng
57:655–666
16. Guo XJ, Chen L, Shen CQ (2016) Hierarchical adaptive deep convolution neural network and
its application to bearing fault diagnosis. Measurement 93:490–502
17. Wen L et al (2018) A new convolutional neural network-based data-driven fault diagnosis
method. IEEE Trans Industr Electron 65:5990–5998
18. Attallah B et al (2017) Histogram of gradient and binarized statistical image features of wavelet
subband-based palmprint features extraction. J Electron Imaging 26
19. Celikoglu A, Tirnakli U (2018) Skewness and kurtosis analysis for non-gaussian distributions.
Phys a-Statist Mech its Appl 499:325–334
20. Hmimid A, Sayyouri M, Qjidaa H (2018) Image classification using separable invariant
moments of Charlier-Meixner and support vector machine. Multimedia Tools Appl 77:23607–
23631
21. Mao L, Jackson L, Davies B (2018) Effectiveness of a novel sensor selection algorithm in pem
fuel cell on-line diagnosis. IEEE Trans Industr Electron 65:7301–7310
22. Steiner NY et al (2011) Non intrusive diagnosis of polymer electrolyte fuel cells by wavelet
packet transform. Int J Hydrogen Energy 36:740–746
Fault Classification of Polymer
Electrolyte Membrane Fuel Cell System
Based on Empirical Mode Decomposition
Yousif Yahia Ahmed Abuker, Zhongyoug Liu, Lisa Jackson, and Lei Mao
Abstract This paper investigates the empirical mode decomposition (EMD)-based

feature selection technique in identifying polymer electrolyte membrane fuel cell
water management issue. In the study, EMD is applied to fuel cell voltage signals
at various states, including fuel cell flooding and membrane drying out, from which
intrinsic mode functions (IMFs) are obtained. Sensitivity index is used to determine
the IMFs more sensitive to fuel cell fault, which are then used for the fault diagnosis.
The effectiveness of the selected IMFs in discriminating flooding and drying out
scenarios is validated using experimental data from a PEM fuel cell system. The
results show that different faults can be distinguished effectively. From the findings,
a generalized feature extraction and selection technique can be provided for fault
diagnosis in practical PEM fuel cell applications.
Keywords Polymer electrolyte membrane fuel cell · Empirical mode

decomposition · Flooding · Drying · Sensitivity · Fault diagnosis
1 Introduction
Owing to climate change and depletion of fossil resources issues, hydrogen and fuel
cell technologies have recently gained increasing research interest, thanks to their
high efficiency and zero emissions. The fuel cell system is device that transforms
chemical energy stored in fuels into electric power through redox reaction. Fuel cells
can be classified into several types based on the electrolyte used, such as solid-oxide
fuel cells (SOFCs), phosphoric acid fuel cells (PAFCs), alkaline fuel cells (AFCs),
molten carbonate fuel cells (MCFCs), and polymer electrolyte membrane fuel cells
Y. Y. A. Abuker · Z. Liu · L. Mao (B)

Department of Precision Machinery and Precision Instrumentation, University of Science and
Technology of China, Hefei, Anhui 230027, China
e-mail: leimao82@ustc.edu.cn
L. Jackson
Department of Aeronautical and Automotive Engineering, Loughborough University,
Loughborough, UK

https://doi.org/10.1007/978-981-15-9199-0_24
262 Y. Y. A. Abuker et al.
(PEMFCs). Amongst the aforementioned different types of fuel cell systems, PEMFC
is the most widespread type used in transport and stationary applications, due to char-
acteristics like the low operating temperature, the possibility to use air oxygen, swift
start-up and high power density [1–5]. However, the twin challenge of durability
and reliability of PEMFC are the major drawbacks limiting its wider commercial-
ization. A number of research studies have been dedicated to PEMFC diagnosis as
a possible solution, which can be loosely divided into two categories, model-based
and data-driven-based methods [6]. Concerning model-based methods, the presence
of malfunction in the system is determined with developed numerical model, which
can be mathematical, empirical, or knowledge-based [7]. With developed PEMFC
model, methods such as parameter identification, observer-based, and parity-space
modelling could be utilized for fault diagnosis [8]. In data-driven methods, data
mining or statistical methods are used for PEMFC faults detection and isolation [9].
In this category, several signal processing techniques, including data pre-processing
and data analyzing, would be applied to data collected from the PEMFC system
[10, 11].
Based on previous studies [12], almost all fuel cell failures can lead to a voltage
drop; thus, the PEMFC voltage is a good indicator for implementing fault diagnosis.
In numerous studies, fuel cell voltage has been used to identify faults in the PEMFC
systems. Steiner et al. [13] used wavelet transform to output voltages for PEMFC
faults identification; results showed that fuel cell flooding and non-flooding state can
be discriminated obviously. Kim et al. [14] determined the appropriate equivalent
circuit model parameters using hamming neural network and voltage output pattern
for PEMC faults identification and isolation. Legros et al. [15] used power spectral
densities (PSD) based on the on-line measurement of the electrochemical noise (EN)
to identify flooding and membrane dehydration faults. Damour et al. [16] proposed
a novel technique based on EMD to decompose a voltage signal into several IMFs,
with 60s of operation time and 1 kHz sampling frequency.
From the above studies, it can be concluded that the key factor for data-driven
PEMFC is the extraction and selection of appropriate features from raw data,
but features were selected with trial-and-error method in most previous studies,
which is time-consuming and may not achieve the best performance due to local
optimum. Therefore, a generalized feature selection technique is urgently required
for implementing reliable data-driven fault diagnosis in practical PEMFC systems.
In this study, the features extracted from PEMFC voltage using EMD are further
evaluated and studied, the optimal features are determined with sensitivity analysis,
and experimental data from a PEMFC system is used to investigate the effectiveness
of selected features in discriminating flooding and drying out faults. The rest of this
paper is organized as follows: The PEMFC system is briefly described in Sect. 2. The
theory of EMD is presented in Sect. 3, and the proposed method is also presented.
In Sect. 4, the effectiveness of selected features in fault diagnosis is presented and
compared with those from previous studies. Conclusions are made in Sect. 5.
Fault Classification of Polymer Electrolyte Membrane … 263
2 PEMFCs Typical Structure and Operating Principle
The typical PEMFCs stack is built at the centre of PEMFC from membrane electrode
assemblies (MEAs) which involve the electrolyte (proton exchange membranes), gas
diffusion layers, and catalyst layers that are inserted between two types of electrodes
(anodes and cathodes), where electrochemical redox reactions take place to generate
electricity, whenever the fuel is provided as shown in Fig. 1. [17, 18].
Usually, a functioning PEMFC provides continuously with fuel (hydrogen) on
the anode side and oxygen on the cathode side. On the anode side, as indicated by
Eq. (1), hydrogen breaks down into protons and electrons (oxidation reaction). The
protons pass through the membrane and the electrons move through the external
circuit, where they generate electrical current and achieve cathode side.
H2 → 2H+ + 2e− (1)
On the cathode side, the protons move through the membrane and the electrons
through the external circuit, the oxygen reacts by reduction reaction:
Fig. 1 Typical structure of Electric Current

PEM fuel cell
e- e-
H2 Air
e- e-
e-
O2
H2
H+
H+
H+
Excess
H2 Water
and Heat
H2O
Anode Cathode
Catalyst Layer Catalyst layer

Electrolyte
1
O2 + 2H + 2e− → H2 O (2)
2
The comprehensive reactions on the anode and cathode sides, shown in Eq. (3)
1
O2 + H2 → H2 O (3)
2
Through the transformation of chemical energy to electric power and heat, the
only byproduct of the reaction is water [19].
3 Description of EMD and Proposed Feature Selection

Method
EMD is a data-driven technique. The gist of the EMD is to decompose the orig-
inal signal step by step, creating a series of IMFs with varying scales from high to
low frequency oscillation. Compared with other decomposition techniques, EMD is
applicable to cope with nonlinear and non-stationary signal. All component of the
IMF must meet the following two requirements [20].
With EMD, in the whole dataset, the number of extrema and the number of zero-
crossings must be identical or differ at most by one, and the mean values of both
the envelope specified by the local maxima and the envelope specified by the local
minima must be zero.
The EMD decomposition process can be given in the following order:
1. For the original signal x(t), define all the local maxima and local minima.
2. Connect all the maxima and minima to generate u (t) upper envelope and l (t)
lower envelope.
u(t) + l(t)
3. m(t) = (4)
2
where m(t) represent the mean for the upper and lower envelope.
4. hi (t) = x(t)- m(t). If h1 (t) fully met the requirements of IMF, then the iteration
halted, and hi (t) is the first imf1 , otherwise go back to the first step.
5. Defines the residual r n (t)= x(t)-imfi (t), as r n (t) still have information from
low frequency; substitutes x(t) with r n (t), and continues the procedure until the
residue is below the specified threshold. The original signal can be reassembled
by using Eq. (5)

n
x(t) = (imfi (t)) + rn (t) (5)
i=1
Based on extracted IMFs from the original signal using EMD, the features could
be generated and selected for PEMFC fault diagnosis. In this study, the energy of
IMFs is calculated using Eq. (6),
+∞
Ei = |imfi (t)|2 (6)
−∞
where E i represents energy for each IMFs.

It should be mentioned that the energy features written in Eq. (6) have been
used in the previous study [16] to identify various PEMFC faults, but as multiple
energy features could be obtained with decomposed IMFs, trail-and-error analysis
was employed to select energy features for the best diagnostic performance. While in
this study, a generalized feature selection strategy is proposed and used for PEMFC
fault diagnosis, which is presented as follows.
To select the best feature that holds the most information related to flooding and
drying out faults, sensitivity value of each IMF energy is calculated, which is the
energy variation due to certain amount change in cell voltage. This could be written
in Eq. (7)

E i j − E i( j+1)
Si j = (7)
|V1 − V2 |
where S ij is the sensitivity value of jth IMF energy to the ith voltage segment, V 1
and V 2 are the maximum and minimum voltage of the corresponding decomposed
signal, E i and E i(j+1) represent energies of the jth and (j+1)th IMFs extracted from
the ith voltage segment .
With Eq. (7), the variation of IMF energy could be obtained. As the cell voltage
change is always due to the PEMFC fault, the IMF energy variation could be used
to represent the sensitivity of corresponding IMF energy to the fault. Therefore, the
IMFs having the highest sensitivity values should be selected and used in PEMFC
fault diagnosis, as these IMFs are most affected by the PEMFC faults.
4 Effectiveness of Proposed Method in PEMFC Fault

Diagnosis
4.1 The PEM Fuel Experimental Setup
80 W PEMFC system data have been used to validate the efficiency of the method,
including PEMFC stack, cooling circuit, fuel supply system, and air supply systems,
as shown in Fig. 2.
Fig. 2 PEM fuel cell test

platform
The tested PEMFC platform with 100 cm2 active area constructed by Pragma
Industries, utilizing the same technology and materials as commercial PEMFCs,
comprising Nafion polymer electrolyte membrane, carbon diffusion materials,
platinum nanoparticle catalyst, silicon seal gasket, and graphite flow field sheets.
Two types of PEMFC faults, including flooding and drying out, are tested with
the aforementioned platform by reducing and increasing PEMFC stack temperature,
respectively.
4.2 Performance of Selected Features in PEMFC Fault

Diagnosis
In the analysis, the effectiveness of the proposed method is assessed on 14 segments

with a window length of 180s, including eight segments for flooding state and six
segments for dehydration state. EMD is used to decompose the cell voltage signal,
four IMFs are then extracted from each segments. It should be noted that although
four IMFs are decomposed in the study, the proposed method could also be applied
in scenarios with more decomposed IMFs. Figure 3 illustrates the four IMFs decom-
posed by EMD of the cell voltage output of the flooding and drying out scenarios. The
energy value for each IMFs is calculated to construct feature vectors, and the fuel cell
fault classification features are selected using the sensitivity index in Eq. (7). Further-
more, to compare the performance of selected features with proposed approach with
that from previous studies, the energies values of each IMF are calculated using
Eq. (6), and the performance of IMF with those features used in previous study [16]
is also illustrated, and the results are depicted in Fig. 4.
It could be found from Fig. 4 that with selected features from the proposed method,
the two types of fault can be discriminated clearly. On the other hand, with features
obtained with are trial-and-error method, three samples were misclassified, indicating
that the proposed method could provide better diagnostic performance in PEMFC
system.
(a) (b)
0.61 0.605
Signal
signal
0.605 0.6
0.6
0.595 0.595
0 50 100 150 0 10-3 50 100 150
10-3 5
IMF1
IMF1
0
-2 0
-4 -5
0 50 100 150 0 10-3 50 100 150
10-4 5
20
IMF2
IMF2
10 0
0
-5
0 10-4 50 100 150 0 50 100 150
10-3
20 1
IMF3
IMF3
10
0 0
-1
0 50 100 150 0 10-3 50 100 150
10-3
5 5
IMF4
IMF4
0 0
-5 -5
0 50 100 150 0 50 100 150
0.6002
0.605
Res
Res
0.60018
0.6 0.60016
0 50 100 150 0 50 100 150
Time (s) Time (s)
Fig. 3 The EMD decomposition of the fuel cell voltage a flooding b drying
10 -3 10 -3
7 7
Flooding Flooding
Drying (a) Drying (b)
6 6
5 5
Feature 2
Feature 2
4 4
3 3
2 2
1 1
0 0
0 0.5 1 1.5 2 0 2 4 6
Feature 1 10 -3 Feature 1 10 -3
Fig. 4 Fault classification based a sensitivity index b highest energy

Table 1 The Euclidean

Type of feature Euclidean distance
distance of two PEMFC faults
Highest sensitivities (proposed method) 0.0021
Highest energies 0.0018
Furthermore, with each method, the centre of two different PEMFC faults are
calculated, and Euclidean distance of two faults could be obtained and used to inves-
tigate the performance of different methods in distinguishing PEMFC faults. Table 1
lists the results; it could be seen that the proposed method could improve the of
performance classification, and the slight difference in the Euclidean distance is due
to the limited IMF samples in each PEMFC fault. Therefore, the features from the
proposed method could provide even better PEMFC fault diagnostic performance
than that using trail-and-error technique, thus would be more beneficial in practical
PEMFC applications.
5 Conclusion
In this study, a EMD-based approach is used to detect and classify flooding and
dehydration fault, where features for fault classification are selected using sensitivity
analysis. The cell voltage is decomposed into several IMFs using EMD technique.
The sensitivity value is then calculated representing the relationship between IMFs
energy and voltage. Experimental data from a PEMFC system is used to verify the
proposed method. Results demonstrate that drying out and flooding faults can be
classified with high quality. Compared to diagnostic results using feature from trail-
and-error analysis for faults classification, the selected features can provide better
classification results. Further works are in progress to assess the performance of other
decomposition technique for the detection of flooding and drying out.
Acknowledgements This work is supported by National Natural Science Foundation of China

(NSFC) with no. of 51975549, Anhui Provincial Natural Science Foundation with no. of
1908085ME161, and Hundred of Talents Program of Chinese Academy of Sciences – Young Talents
with no. of 2018-050.
References
1. Wang Y et al (2011) A review of polymer electrolyte membrane fuel cells: technology,

applications, and needs on fundamental research. Appl Ener 88(4):981–1007
2. Onanena R et al (2013) Fuel cell health monitoring using self-organizing maps. Chem. Eng.
Trans. 33
3. Mao L, Jackson L, Dunnett S (2017) Fault diagnosis of practical polymer electrolyte membrane
(PEM) fuel cell system with data-driven approaches. Fuel Cells 17(2), 247–258
4. Hua J et al (2011) Proton exchange membrane fuel cell system diagnosis based on the
multivariate statistical method. Int J Hydrogen Energ 36(16):9896–9905
5. Riascos LAM, Simoes MG, Miyagi PE (2007) A Bayesian network fault diagnostic system for
proton exchange membrane fuel cells. J Power Sources 165(1):267–278
6. Dijoux E et al (2017) A review of fault tolerant control strategies applied to proton exchange
membrane fuel cell systems. J Power Sour 359, 119–133
7. Petrone R et al (2013) A review on model-based diagnosis methodologies for PEMFCs. Int J
Hydrogen Energ 38(17):7077–7091
8. Aitouche A, Yang Q, Bouamama BO (2011) Fault detection and isolation of PEM fuel cell
system based on nonlinear analytical redundancy: an application via parity space approach.
Eur Phys J-Appl Phys 54(2)
9. Li Z et al (2013) Online diagnosis of PEMFC by analyzing individual cell voltages. In: European
control conference (ECC). IEEE, 2439–2444
10. Zheng Z, et al (2013) A review on non-model based diagnosis methodologies for PEM fuel
cell stacks and systems. Int J Hyd Ener 38(21):8914–8926
11. Li Z et al (2014) Data-driven diagnosis of PEM fuel cell: a comparative study. Cont Eng Pract
28, 1–12
12. Yousfi-Steiner N et al (2008) A review on PEM voltage degradation associated with water
management: Impacts, influent factors and characterization. J Power Sour 183(1):260–274
13. Steiner NY et al (2011) Non-intrusive diagnosis of polymer electrolyte fuel cells by wavelet
packet transform. Int J Hyd Energ Commercialisation 740–746
14. Kim J et al (2012) State-of-health diagnosis based on hamming neural network using output
voltage pattern recognition for a PEM fuel cell. Int J Hydrogen Energ 37(5):4280–4289
15. Legros B et al (2011) First results on PEMFC diagnosis by electrochemical noise. Electrochem
Commun 13(12):1514–1516
16. Damour C et al (2015) Polymer electrolyte membrane fuel cell fault diagnosis based on
empirical mode decomposition. J Power Sour 299:596–603
17. Corrêa JM et al (2004) An electrochemical-based fuel-cell model suitable for electrical
engineering automation approach. IEEE Trans Ind Electron 51(5):1103–1112
18. Jensen JO et al (2010) Report Nr, Schriften des Forschungszentrums Jülich. Energy and
Environment
19. Li Z et al (2014) Data-driven diagnosis of PEM fuel cell: a comparative study. Control Eng
Pract 28(28):1–12
20. Shukla S, Mishra S, Singh B (2009) Empirical-mode decomposition with Hilbert transform
for power-quality assessment. IEEE Trans Power Delivery 24(4):2159–2165
Fleet-Oriented Pattern Mining
Combined with Time Series Signature
Extraction for Understanding of Wind
Farm Response to Storm Conditions
P.-J. Daems, L. Feremans, T. Verstraeten, B. Cule, B. Goethals, and J. Helsen
Abstract Offshore wind turbine installations are rapidly spreading around Europe
and all over the world. These turbines are typically installed in large wind farms
combining turbines of the same type. Farm owners target maximal performance of
the farm in general and particularly predictability of behaviour. The latter is getting
increasingly important since offshore wind farms are being managed more and more
as conventional power plants driven by the electricity market supply and demand
considerations. The context of zero subsidy farms exposes farm operators to fluctu-
ations in electricity market prices. As such, deep understanding of farm behaviour
is essential to come up with a good strategy to deal with these fluctuations. This
paper focusses on the automated extraction of farm-wide response to storm condi-
tions. The input data for the analysis are status logs and SCADA 1-second data. The
status logs record the important turbine controller events. Typically, they consist of a
number, a time of occurrence, and a time of de-activation. The number is linked to a
detailed description. The SCADA data consists of time series of the most important
sensors in the turbine: power produced, RPM, wind speed,… The advantage of the
1-sec data over the traditional 10-minute averages is that the dynamic event content
is much more preserved. Data of several offshore wind farms is used in the analysis
to have a solid dataset. In total, 5 years of data of more than 50 turbines is used. We
show a novel farm-wide pattern mining approach that extracts events occurring for
multiple turbines in the same time period. This allows us to identify those events
that are predominantly driven by global wind excitations (e.g., gusts) or grid events
P.-J. Daems · T. Verstraeten · J. Helsen (B)

Acoustics and Vibrations Research Group, Vrije Universiteit Brussel, Pleinlaan 2, Brussels,
Belgium
e-mail: jan.helsen@vub.be
L. Feremans · B. Goethals
Department of Mathematics and Computer Science, Universiteit Antwerpen, Antwerp, Belgium
B. Cule
Department of Accountancy and Finance, Universiteit Antwerpen, Antwerp, Belgium
T. Verstraeten
Artificial Intelligence Laboratory, Vrije Universiteit Brussel, Brussels, Belgium

https://doi.org/10.1007/978-981-15-9199-0_25
272 P.-J. Daems et al.
(e.g., low voltage ride through). From the extracted events, we lift out the storm
conditions. For these conditions, a further investigation of the time series data is
done. Using event detection algorithms, we extract the signatures of the stop events
that each turbine is performing from the time series data. We show that the extreme
change in wind speed and wind direction leads to an excessive misalignment of the
turbines in the farm, followed by a stop of those turbines. The extracted patterns
are compared to the time signatures to show their correlation and complementarity.
As such, the typical turbine response to this event is identified. This can serve as
input for identification of novel controller approaches by the farm owner and turbine
manufacturer to deal with this problem.
Keywords Pattern mining · Condition monitoring · Wind turbine · Dynamic

event · Fleet · Wind farm
1 Introduction
Wind turbines target a typical lifetime of 20 years. During those years, the machine
should produce as much energy as possible, while keeping downtime due to failure
minimal. As such predictability of future behaviour is key. On the one hand, condition
monitoring approaches are investigated extensively to achieve better failure predic-
tions and thus increased uptime [1–3]. A variety of sensor types can be used for
turbine condition monitoring ranging from accelerometers, current sensors to oil
particle count devices. An increasing attention has been given in literature to the use
of data of the supervisory control and data acquisition (SCADA) system as input
data for performing condition monitoring, since this would avoid the need to place
additional sensors on the turbine [4–6]. On the other hand, performance analysis, for
example, through power curve modelling, targets an increase of turbine output [7–9].
Today, performance analysis is typically performed on data of the SCADA system
of the turbine sampled at 10 min intervals. These values are highly interesting to
keep an overview of general trends. However, the 10-minute interval is challenging
if one intends to investigate the dynamic events the turbines are exposed to, such as
emergency stops. Another data source also provided by the SCADA system offers
potential in this regard: the status logs [10]. These logs provide a record of the most
important turbine controller actions. Typically, they are characterized by a status code
number, a descriptive message, a time of activation, a time of de-activation, and a
duration. Since the trigger times are registered directly, this data source is very limit-
edly influenced by time-shifts linked to limited sampling rates. As such the status logs
show high potential for event sequence investigations to better understand turbine
response to triggers for the controller. Gonzalez et al. define an optimized taxonomy
for categorizing the status logs and investigate the relation between component faults
being followed by failure occurrences in others [11]. Pattern mining techniques are
currently being explored to get detailed insights in these sequences. Kusiak et al.
used frequent pattern mining techniques to identify and predict sequences in the
Fleet-Oriented Pattern Mining Combined with Time Series … 273
status logs [10]. Giu et al. define two methods for alarm analysis: a time-based and
probability-based approaches and illustrate this approach on 2 years of field data
from two populations of onshore turbines [12]. Feremans et al. similarly determined
frequent item sets and used those to derive association rules to determine patterns
near alarms [13]. These investigations show that there is potential for the use of
pattern mining approaches for status code analysis to determine and better under-
stand the sequences governing the turbine controller responses to different triggers.
This paper builds further on this and introduces the fleet concept. Instead of focussing
on individual turbines, we target the detection of patterns over multiple turbines in
the farm. In other words, similar responses of multiple turbines in the farm to joint
or individual triggers.
In addition to the status logs and SCADA 10 min data, for newer turbine types,
the complete SCADA tag set or at least a subset is becoming available at sampling
rates of 1 s. The finer granularity of this data sources makes it a complementary data
source for event understanding to the status log data. Event detection algorithms
allow to extract and classify the dynamic events based on their time series signatures
[14]. This paper uses this approach to extract events from the SCADA 1 s data of
the different turbines in the farm. This data is then combined with the interesting
patterns mined from the status logs.
2 Methodology
2.1 Overview
This paper suggests a method building on the identification of fleet-wide patterns

in the status logs and in the dynamic events detected in the SCADA 1-second data,
illustrated in Fig. 1. It comprises of three parts. First, event detection is done on
the SCADA 1-second data to extract turbine start and stop events. Second, patterns
are extracted separately from status logs to detect event sequences that are present in
multiple turbines in the farm. Finally, the results of the two are brought together in an
event fusion step. The following subsections will discuss the different components
of the methodology in detail. In the next section, the methodology is illustrated by
means of an experimental example.
Step 1: Pattern mining of status log data

Prior to the discussion of the pattern mining approach targeting itemsets used in the
proposed approach, some terminology is introduced. To make it more intuitive, the
analogy is made to analysis of customer baskets in a supermarket context.
• Itemset: All status log windows in the database are considered the set of all items:
I = {i1 , i2 ,…,in }. A subset of I is defined as an itemset. In comparison, in the
Fig. 1 Farm-wide approach for event sequence extraction
field of supermarket market basket analysis, the set of items consists of all the
products that the supermarket is selling.
• Transaction: A transaction is defined as a unique tuple that is comprised of a
transaction identifier and an itemset. Thus, a transaction is noted by the tuple T i
= (tid, X). In this tuple, tid is a unique transaction identifier. X is an itemset and
thus a subset of I. Again, for the supermarket basket analysis, a transaction is a
set of items or products that one customer has in his basket.
• Transaction database: The transaction database contains all transactions. For the
supermarket basket analysis, the transaction database contains all baskets that
were sold to customers during the period of investigation.
• Support: Support is linked to an itemset. It represents the number of transactions in
the transaction database in which a certain itemset occurs. Typically, a minimum
support is defined: min_sup.
• Frequent pattern: Only frequent itemsets with a support higher than this minimum
are considered. Similarly, a maximum support can be defined. The pattern mining
approaches used in this paper pose a threshold to the minimum support neglecting
all patterns below the minimum support.
• Closed itemset: One of the main challenges for pattern mining approaches is the
fast growth of frequent patterns. Many of the identified itemsets have subsets. An
itemset of size n has 2n subsets. If the support of the itemset is high and above the
min_sup threshold, then also that of the subsets will be high, which means that
it will be above the min_sup threshold too. Therefore, we target closed itemsets.
An itemset can be defined to be closed if no superset is present with the same
support.
2.2 Frequent Pattern Mining
Frequent pattern mining algorithms target the discovery of all itemsets in the transac-
tion database with a support higher than the defined minimum support. This cannot be
done brute force since it would result in enumerating 2|I| subsets to check all possible
subsets and count occurrences. To overcome this, most frequent itemset mining algo-
rithms exploit the anti-monotonic property of itemset support in a branch-and-bound
algorithm. For any superset of an itemset, it holds that if the support of the itemset
is smaller than the minimum support, this will also be the case for the superset. This
property can be intuitively explained as follows. Assume we have three items x, y and
z and assume x to occur in 8 transactions, y in 4 transactions and z in 2 transactions.
In case, we mine for patterns with minimum support of 3, then only itemsets of size
1 {x} and {y} are frequent. For the itemsets of size 2, possibilities are {x, y}, {x,
z} and {y, z}. We can deduct that definitely the support of {x, y} will be smaller
or equal to that of {x} and {y} individually. It could be that {x, y} is present in all
4 transactions containing y which would make it a frequent itemset. However, for
itemsets of length 2 containing z, it is already known that the frequency will not be
big enough for making them frequent.
Most popular itemset mining algorithms are build further on this property. The
Apriori algorithm introduced by Agrawal et al. uses a breadth-first approach similar
to the concept introduced above [15]. Zaki defined the Eclat algorithm to exploit a
depth-first approach, which is based on re-arranging the transactional database and
computing the set of transactions for each item [16]. This allows to calculate the
support by computing a set of intersections. Han et al. developed the FP-growth
algorithm based on a divide-and-conquer strategy to keep an efficient data structure
for the database [17]. For each recursive step, the database is filtered by the current
itemset prefix, which shrinks its size, and speeds up support computation. In this
paper, the Eclat algorithm is used, more specifically the method by Borgelt [18]. We
are not sure about the fact that indeed the correct sequence in which the status codes
occur is kept when they are stored. As such we opt to use itemset-based mining
approaches and not take the sequence in the identified itemsets into account.
2.3 Data-Windowing
Status log data is time series data. The activation and de-activation time of each status
log are recorded. As such each status code will be part of an activation time series and a
de-activation time series. For simplicity, the number of the status code is used as value.
As such two time series of status values and corresponding timestamps are generated.
To make these time series compatible with the pattern mining algorithms itemsets
need to be created. This is done by windowing the data using a sliding window. In
previous work on single turbine pattern mining, we used overlapping fixed windows
spanning 2 h [13]. However, overlapping windows result in higher support values
than there are occurrences in the original time series, and need correcting. For farm-
wide event mining, we, therefore, use non-overlapping sliding windows spanning
24 h.
2.4 Farm-Wide Support
We provide two definitions for computing farm-wide support that are applicable, in
general, to a fleet’s of devices, where each device logs events over a long period of
time.
Vertical support is the number of windows where the patterns occurs in a single
turbine. The farm-wide vertical support of a pattern is the none-zero average of the
vertical support for each wind turbine in the farm.
For example, if we create windows spanning 1 day, an average vertical support
of 50 for the extreme wind direction singleton pattern, means this pattern occurs 50
times on average in a turbine.
Horizontal support measures the number of time the pattern occurs in different
wind turbines during the same window, or period.
The mean non-zero horizontal support is computed by averaging the horizontal
support over all windows.
For example, if we create windows spanning 1 day, an average horizontal support
of 5 for the extreme wind direction singleton pattern, means this pattern occurs on
average in 5 wind turbines.
We remark that certain patterns, such as pattern related to lightning, have very
low vertical support, and high horizontal support.
For mining patterns having a high non-zero mean value of vertical support, we
create a vertical transaction database for each wind turbine. We then mine closed
itemsets in each transaction database (or turbine) and set a constraint on minimal
vertical support. We then compute the non-zero mean of vertical support for each
pattern in the entire fleet by looking at the number of occurrences in all vertical
transactional databases. For example, assuming 10 turbines, a total span of 1 year
and a window of 1 day, we create 10 databases, each consisting of 365 transactions.
For mining patterns having a high none-zero mean value of horizontal support, we
create a horizontal transaction database for each day, thereby creating a transaction
for all wind turbines. We then mine closed itemsets in each horizontal transaction
database and set a constraint on minimal horizontal support. We then compute the
non-zero mean of horizontal support for each pattern in the entire fleet by looking
at the number of occurrences in all horizontal transactional databases. For example,
assuming 10 turbines, a total span of 1 year and a window of 1 day, we create 365
databases, each consisting of 10 transactions.
We remark that mining many small databases with constraints requires moderate
resources, assuming the usage of an efficient algorithm for mining closed itemsets.
The result of this step will be a set of farm-wide patterns together with their farm-
wide support. In addition the time of occurrence and turbine of occurrence of each
pattern is stored.
Step 2: Event detection in SCADA 1-second data

In this paper, we focus on start and stop behaviours of wind turbines. As such these
event types are extracted from the SCADA 1-second data. For this, we use an iden-
tification and classification approach that we previously developed and documented
[14]. In essence, this approach learns the time signature of typical annotated start and
stop events. After this training step, it is able to automatically extract these events
from time series data.
Step 3: Event fusion

In this paper, we aim to gain better insights into the co-occurrence of turbine response
throughout a wind farm. To do this, we merge the event sequences identified by the
pattern mining approaches with the events extracted from the time series. The goal
is to use the time series data as a way to go deeper in phenomenon understanding.
We use the pattern mining to scan through large datasets of several years to identify
farm-wide events. Once the events are identified, we validate them using the events
detected in the time series and perform an in-depth root-cause analysis using the raw
SCADA 1-second data.
3 Experimental Case
3.1 Overview
The experimental case follows the philosophy described in the method section. We
use multiple years of data of multiple offshore wind farms. The dataset consists of
5 years of status log data of one farm with more than 40 turbines and 6 months of
status log data and SCADA 1-second data for a farm with more than 40 turbines. The
experimental case consists of two steps. First, the 5 years of status log data of the first
farm is used to identify interesting farm-wide events. For this paper, we specifically
target the example of a heavy storm. Once the farm pattern for a storm is identified,
that specific farm pattern is searched in the status log data of the second farm. This
implies transfer learning: patterns identified on one farm are transferred to another
farm. To make it additionally challenging, the turbines types and brands of the two
farms are different. Once the storm pattern is identified, it is verified based on the
events of the SCADA 1-second data that indeed a storm took place. Then, the turbine
response is further investigated in detail using the SCADA 1-second data.
Table 1 Pattern linked to

Code Description
storm conditions identified in
status log data of farm 1 410 Yaw cable twist has been reset to a specified value°
412 Yaw cable was untwisted
460 Automatic restart after predefined time [s]
552 Start of automatic yawing action
560 Yaw control setting has been changed
568 Stop of automatic yawing action
712 Extreme direction of the wind with angle at Wind speed
× m/s
426 Nacelle position has been reset unexpectedly
3.2 Pattern Mining on Farm 1
The status log data has a long tail distribution. A very limited part of the status log data
forms 99% of the total data. These status logs typically lead to trivial patterns, as we
illustrated in [19]. Thus, we aim to filter these trivial operational codes out. All status
log data were first cleaned by removing standard operational messages depicting
instantaneous values of power produced, rpm, pitch, … Moreover, to avoid trivial
patterns only closed patterns with a minimal length of 5 were retained. Additionally,
a non-zero mean horizontal support of 10 is required. Windows of one day are
considered.
Table 1 illustrates an important pattern found from the fleet-wide pattern mining
analysis. The pattern showed a non-zero mean horizontal support of 11. This means
that the pattern was showing up at 11 turbines at the same time. Moreover, it occurred
on a non-zero mean vertical support of 9.5 times with a standard deviation of 9 over
the five-year data span. We expect the pattern might be occurring close to storm
conditions. To validate this hypothesis, we try to find similar patterns in the data of
farm 2.
3.3 Pattern Searching in Farm 2
The turbines of farm 2 are from a different manufacturer who uses a different
ontology/taxonomy for defining his status log tags. This implies that the status codes
of the first farm cannot be directly used in the pattern query for farm 2. To over-
come this, generic text components were extracted from the status log descriptions
in Table 1. These so-called key words were then used for the search in the status
log data of farm 2. In the status log data of farm 2, the pattern illustrated in Fig. 2
was found that links closely to the pattern of Table 1. Again, the subcomponents
of excessive deviation between the wind direction and nacelle direction are found
(orange dot in Fig. 2). This links to the tag “Extreme direction of the wind with angle°
Fig. 2 Status code representation for a subset of the turbines of farm 2 over the course of 1 day
at Wind speed x m/s” in Table 1. Moreover, there is extreme deviation between the
needed and current yaw angle requiring to start yawing (green dot in Fig. 2). This
links to the tag “Nacelle position has been reset unexpectedly”. These two indicate
that the turbines are completely misaligned to where the wind is coming from. As
such the turbines perform a yawing action, indicated by the tag “Yaw speed high” in
Fig. 2 and the tags “Start of automatic yawing action”, “Stop of automatic yawing
action”, and “Yaw control setting has been changed”. Finally, it seems that in many
cases, the targeted yaw angle is not achievable without prior unwinding of the power
cable. In Table 1, this is identified by the tags “Yaw cable was untwisted” and “Yaw
cable twist has been reset to a specified value°”. In Fig. 2, this is identified by the
tag “Cable auto unwind”. The comparison of the tasks shows that there is a high
correlation between the status code pattern that was found in farm 1 and in farm 2.
3.4 SCADA 1-Second Event Analysis on Farm 2
The status codes were thus used to identify the time stamps of the interesting events.
As mentioned before, we expect the event sequence that was identified using the
pattern mining approach to be linked to storm conditions. In Fig. 3, it can been
seen that although the majority of the turbines shows a certain behaviour, not all
turbines are showing exactly the same response. This means that there must be local
differences between the loading the turbines are exposed to. These local changes are
investigated further using the time series SCADA 1-second data. Figure 3 shows the
different start-up and shutdown events identified from the SCADA 1-second data.
The blue zone indicates the time period for which the farm event is expected to take
Fig. 3 Events detected from SCADA 1-second data using signature detection
place. The shutdown events are clearly present for all turbines in the farm in contrast
to the status codes where only a limited set of machines were showing events. This
is due to the fact that the status log for stop is only recorded if the blade pitch evolves
to 90°. Moreover, from the events in Fig. 3 it is clear that several other farm-wide
events occur later on. Once the turbines perform their final start-up no more events
occur. This indicates that the event sequence is located in this limited time window.
3.5 Event Fusion
To understand the different events that are taking place and why they result in the
sequence that they do, we investigate the SCADA 1-second time series deeper. A
representative turbine is taken that was showing identified pattern. Figure 4 illustrates
the characteristic signals during the event: rpm, wind speed, wind direction, and yaw
angle. Based on the wind direction and yaw angle, it can be seen that there is a rapid
change in wind direction causing an important misalignment of the turbine to the
wind. This causes the machine to perform a stop. As shown in Fig. 5, some turbines
are experiencing a large rapid wind speed increase in addition to the sudden direction
change. Many turbines are becoming strongly misaligned towards the wind at that
particular moment. This matches with the presence of the tag type “Extreme wind
direction angle °” in the status codes. The machine attempts to correct for the yaw
misalignment by performing yawing action, as was indicated in the pattern. However,
the misalignment remains substantial. After the sudden increase of the wind speed,
the wind reduces significantly in speed. For certain turbines, as shown in Fig. 4, the
turbine is not able to reach the required yaw angle due to cable winding constraints.
Thus, the turbine remains in idling condition and performs a cable unwinding proce-
dure as was identified in the pattern. Figure 6 illustrates this cable unwind. The yaw
angle changes linearly over time as the turbine unwinds itself. After this action,
Fig. 4 SCADA 1-second data of rotor speed, wind speed, wind direction, and yaw angle for period
of 2 h to provide an overview of the response to storm condition
of 1 h to illustrate extreme yaw misalignment
the machine can again meet the required yaw angle. These actions match with the
sequences identified using the pattern mining approach.
4 Conclusions
This paper showed a methodology for identification of event sequences governing

the response of wind turbines in wind farms. It comprised of farm-wide pattern
identification and SCADA 1-second-based event detection to determine the typical
event sequences following farm-wide triggers. The methodology was illustrated by
of 1 h to illustrate cable unwind
a storm event. Patterns were learned on the status logs of one farm and used to
identify the storm event in a second farm with turbines from a different manufacturer.
Moreover, SCADA 1-second data was used to show the validity of the identified
pattern and to further understand the way the turbines reacted to the storm event.
Acknowledgements The authors would like to acknowledge the IWT HYMOP project and VLAIO
SIM MaSiWEC project for their support as well as the farm owners for sharing their data in the
context of OWI-lab.
References
1. Marquez FPG, Tobias AM, Perez JMP, Papaelias M (2012) Condition monitoring of wind
turbines: techniques and methods. Renew Energy 46:169–178
2. Hameed Z, Hong YS, Cho YM, Ahn SH, Song CK (2009) Condition monitoring and fault
detection of wind turbines and related algorithms: a review. Renew Sustain Energy Rev 13(1):1–
39
3. Lu B, Li Y, Wu X, Yang Z (2009) A review of recent advances in wind turbine condition
monitoring and fault diagnosis. In: IEEE power electronics and machines in wind applications
4. Leahy K, Lily Hu R, Konstantakopoulos IC, Spanos CJ, Agogino AM (2016) Diagnosing
wind turbine faults using machine learning techniques applied to operational data. In: IEEE
conference on prognostics and health management (ICPHM)
5. Tautz-Weinert J, Watson SJ (2017) Using SCADA data for wind turbine condition monitoring—
a review. IET Renew Power Gener. 11(4)
6. Helsen J, Devriendt C, Weijtjens W, Guillaume P (2015) Condition monitoring by means of
SCADA analysis. In: European wind energy conference
7. Kusiak A, Verma A (2013) Monitoring wind farms with performance curves. IEEE Trans
Sustain Energ 4(1):192–199
8. Carrillo C, Obando Montano AF, Cidras J, Diaz-Dorado E (2013) Review of power curve
modelling for wind turbines. Renew Sustain Energ Rev 21:572–581
9. Optis M, Perr-Sauer J, Philips C, Craig AE, Lee JCY, Kemper T, Sheng S, Simley E, Willimans
L, Lunacek M, Meissner J, Fields J (2019) OpenOA: an open-source code base for operational
analysis of wind power plants, Wind Energ Sci https://doi.org/10.5194/wes-2019-12
10. Kusiak A, Verma A (2011) Prediction of status patterns of wind turbines: a data-mining
approach. J Solar Energy 133(1)
11. Gonzalez E, Reder M, Melero JJ (2016) SCADA alarms processing for wind turbine component
failure detection. J Phys Conf Ser 753:072019
12. Qiu Y, Feng Y, Tavner P, Richardson P, Erdos G, Chen B (2012) Wind turbine SCADA alarm
analysis for improving reliability. Wind Energy 15:951–966
13. Feremans L, Cule B, Devriendt C, Goethals B, Helsen J (2017) Pattern mining for learning
typical turbine response during dynamic wind turbine events. ASME IEDETC
14. Daems PJ, Gioia N, Vervaeck H, Peeters C, Verbeke J, Helsen J (2019) Automatic detection
of events critical for drivetrain health and lifetime from long-term field measurements. In:
Conference wind power drives, Aken Germany
15. Agrawal R, Mannila H, Srikant R, Toivonen H, Verkamo AI (1996) Fast discovery of association
rules. Adv Knowl Discovery Data Mining 12(1):307–328
16. Zaki MJ (2000) Scalable algorithms for association mining. IEEE Trans Knowl Data Eng
12(3):372–390
17. Han J, Pei J, Yin Y (2000) Mining frequent patterns without candidate generation. ACM Sigmod
Rec 29:1–12
18. Borgelt: Pyfim toolbox, http://www.borgelt.net/pyfim.html
19. Doro P, Scerri L, Helsen J (2016) Pattern detection in status codes as an optimization tool in
offshore wind farms. European conference if the prognostics and health management society,
Bilbao
Efficiency and Durability of Spur Gears
with Surface Coatings and Surface
Texturing
Giovanni Iarriccio, Antonio Zippo, Marco Barbieri, and Francesco Pellicano
Abstract Experimental results on mechanical efficiency and durability of spur gears

are presented. Reducing the power losses and increasing the reliability of gears are
fundamnetal aspects of the design of transmissions. In order to analyze the tribolog-
ical performance of innovative coatings and laser texturing, three different types of
spur gears were tested: standard carburized gears, WC/C-coated carburized gears,
and carburized gears with laser texture pattern. Tests have been carried out through a
power recirculating test rig equipped with a single-stage transmission; power losses
were evaluated by analyzing the torque-meters signals; several pictures of the teeth
were taken at scheduled times to monitor the wear progression. In conclusion, results
are presented and discussed.
Keywords Gears · Coatings · Texturing · Experimental · Efficiency · Pitting
1 Introduction
The efficiency and durability of gears are topics of great interest for the industry
and the scientific community: these aspects are closely related to the economy and
safety of gear transmissions. Thus, the gear life prediction [1], as well as the early
identification of damages [2, 3], are primary tasks.
It is worth mentioning those studies that aim to increase the useful life and effi-
ciency of the gears by investigating the effects of teeth surface finish and surface
coatings on gear performance [4–7].
In this experimental study, the effects of tungsten carbide/carbon (WC/C) coatings
and surface texturing on the efficiency and durability of spur are investigated: test
setup is described, and preliminary results are presented.
G. Iarriccio (B) · A. Zippo · M. Barbieri · F. Pellicano

Department of Engineering “Enzo Ferrari”, University of Modena and Reggio Emilia, Via P.
Vivarelli 10, 41121 Modena, Italy
e-mail: giovanni.iarriccio@unimore.it

https://doi.org/10.1007/978-981-15-9199-0_26
286 G. Iarriccio et al.
2 Test Setup
Experimental tests were carried out through an electric power recirculation test
rig, equipped with two 75 kW asynchronous electric motors and a single-stage
transmission, Fig. 1.
An oil-jet lubrication system allows the gear lubrication; a mineral oil Shell Omala
S2 G68 was adopted for all the tests and the lubricant temperature was measured
near the jet nozzles using a digital thermometer.
The following kind of gears were considered:
• Standard carburized gears;
• Two different kinds of WC/C-coated gears with a carburized substrate, named as
WC/C-“A” and WC/C-“B”;
• Two sets of carburized gears with laser-textured teeth flanks (Fig. 2), where one
set has a pattern density per unit area of 20% and one set has a density of 40%;
both the sets have texture craters with a diameter of 100 µm;
Gear main dimensions and properties are listed in Table 1: gear samples are
manufactured with a large crowning to concentrate the contact area between the
meshing teeth.
For the mechanical efficiency tests, each gear pair were tested at 200 N m under
a stepped speed ramp from 144 to 2736 rpm, consisting of 61 Steps, with a speed
increment of 43.2 rpm and step period of 40 s.
Regarding the durability tests, gears were tested at a constant speed of 2500 rpm
and a constant torque of 200 N m; these working conditions guaranteed reduced
dynamic effects as the tests were carried out far from the resonances of the mechanical
system. Furthermore, to verify the wear progression, accurate visual inspections of
the teeth flanks were conducted at scheduled time intervals.
Fig. 1 Gears test rig

Efficiency and Durability of Spur Gears with Surface Coatings … 287
Fig. 2 a Textured gear, b texture pattern detail
Table 1 Gear sample main

Number of teeth 62
dimensions and properties
Module 2 mm
Pressure angle 20°
Addendum factor 1.0
Crowning 0.1 mm (both sides)
Transmission ratio 1:1
Gear core material 20MnCr5 steel
Surface roughness Ra 0.2 µm
3 Experimental Results: Efficiency and Durability
In this section, the mechanical efficiency and durability results are presented. By
taking advantage of what has been developed in previous experimental researches
[8, 9], the acquired data are post-processed. Note that durability tests were performed
only on WC/C-“B”-coated gears that showed higher efficiency than standard
carburized gears, which represents the benchmark for the present study.
3.1 Efficiency
The unitary transmission ratio implies that the mechanical efficiency of the
transmission is evaluated through the signal ratio of the torque meters (1).
Tgear
η= (1)
Tpinion
WC/C-coated gears do not always have higher efficiency than standard gears: as
shown in Fig. 3, WC/C-“B” gears are the most efficient samples, with a maximum
percentage difference (2) of 0.17% at 416.6 Hz, Fig. 4; WC/C-“A” gears are less effi-
cient than the other tested gears in the whole investigated frequency range. Moreover,
WC/C-“A” gears showed delamination of the coating layer, Fig. 5.
Fig. 3 Efficiency comparison at 200 N m: standard carburized gears ID 12 (black), WC/C-“A”-

coated gears (blue), WC/C-“B”-coated gears (red)
Fig. 4 Efficiency percentage

difference between
WC/C-“B”-coated gears and
standard carburized gears
Fig. 5 Delamination of the WC/C-“A” coating layer
(ηcoated − ηuncoated )
δ= · 100 (2)
ηuncoated
Both the textured gear samples showed an efficiency lower than standard carbur-
ized gears, Fig. 6: for the considered texture densities it is possible to confirm that
the higher is the texture density the lower is the efficiency.
Fig. 6 Efficiency comparison at 200 N m: standard carburized gears (black), textured carburized
gears 40% (blue), textured carburized gears 20% (red)
3.2 Durability
Durability tests were conducted on standard carburized gears and on WC/C-“B” gear
pair, which showed the highest efficiency for all the tested gear pairs.
Pitting holes were detected on carburized gears after 223 h and a clear contact
pattern was visible on teeth flanks, Fig. 7a; After 245 h, abnormal wear was observed
on both the root and the tip of the teeth, and a crack that involves mainly the tip area
is shown in Fig. 7b.
As shown by the efficiency results, WC/C-“B”-coated gears have the best effi-
ciency performance. The durability test were stopped approximately at 160 h when
the delamination of the coating layer occurred, Fig. 8.
Fig. 7 Carburized gears: a pitting hole after 223 h of test, b abnormal wear and crack after 230 h
Fig. 8 Delamination of the

WC/C coating layer at 130 h
4 Conclusions
Preliminary results show that WC/C surface coatings can improve the efficiency
performance of gears, but these coatings have short durability due to the delamination
of the coating layer.
The gears with texturing showed poor performance: results are closely related
to the dimensional and geometric characteristics of the pattern, i.e., spacing and
depth of the texture craters. Thus, further studies are needed to find the optimum
pattern configuration and fully characterize the effects of the surface texturing on the
efficiency and durability of gears.
Acknowledgements The author(s) would like to thank the Regione Emilia Romagna for supporting
this research: Project “MetAGEAR-(PG/2015/732270)-CUP E82I15002670002.”
References
1. Townsend DP, Coy JJ, Zaretsky EV (1978) Experimental and analytical load-life relation for
AISI 9310 steel spur gears. J Mech Des Trans ASME 100:54–60
2. Gelman L, Zimroz R, Birkel J et al (2005) Adaptive vibration condition monitoring technology
for local tooth damage in gearboxes. Insight Non-Destr Test Cond Monit 47:18–21
3. Stander CJ, Heyns PS, Schoombie W (2002) Using vibration monitoring for local fault detection
on gears operating under fluctuating load conditions. Mech Syst Signal Process 16:1005–1024.
https://doi.org/10.1006/mssp.2002.1479
4. Nakatsuji T, Mori A, Shimotsuma Y (1995) Pitting durability of electrolytically polished medium
carbon steel gears. Tribol Int 38:223–232. https://doi.org/10.1080/10402009508983399
5. Krantz T, Alanou MP, Evans HP et al (2001) Surface fatigue lives of case-carburized gears with
an improved surface finish. J Tribol 123:709–716. https://doi.org/10.1115/1.1387036
6. Fujii M, Seki M, Yoshida A (2010) Surface durability of WC/C-coated case-hardened steel gear.
J Mech Sci Technol 24:103–106. https://doi.org/10.1007/s12206-009-1171-0
7. Michalczewski R, Kalbarczyk M, Tuszyn W (2013) The rolling contact fatigue of WC/C-coated
spur gears. J Eng Tribol 227:850–860. https://doi.org/10.1177/1350650113478179
8. Zippo A, Barbieri M, Iarriccio G et al (2019) Nonlinear vibrations of circular cylindrical shells
with thermal effects: an experimental study. Nonlinear Dyn 1–19. https://doi.org/10.1007/s11
071-018-04753-1
9. Pellicano F, Zippo A, Iarriccio G et al (2020) Experimental study on nonlinear random exci-
tation. In: Rizzi C, Andrisano AO, Leali F et al (eds) Design tools and methods in industrial
engineering. Springer International Publishing, Cham, pp 637–648. https://doi.org/10.1007/978-
3-030-31154-4_54
System Model for Condition Monitoring
and Prognostic
Wheel-Rail Contact Forces Monitoring
Based on Space Fixed-Point Strain
of Wheel Web and DIC Technology
Xiaochao Wang, Zhenggang Lu, Juyao Wei, and Yang He
Abstract Condition monitoring of wheel-rail contact forces has great significance

for evaluating the dynamic performance of railway vehicles. Instrumented wheelset
method has disadvantages that strain gages are complicated in arrangement, wheel-
rail forces decoupling is difficult, and the accuracy is low. Combined with the digital
image correlation (DIC) technology, this paper introduces a novel non-contact moni-
toring method based on space fixed-point strain on the surface of wheel web. Through
the finite element modeling and simulation analysis of wheelset, it is found that
the strain at some space fixed-points on the surface of wheel web is linear with
lateral/vertical forces. Based on this conclusion, the numerical equation between
single space fixed-point strain and the lateral/vertical forces is established. Two point
numerical equations can be decoupled to obtain the lateral and vertical force. The
space fixed-point strain can be obtained by non-contact using the mature DIC tech-
nology and then achieve the simple and accurate measurement method of wheel-rail
contact forces. It is verified by FEM simulation results that under the premise of
using DIC technology to get space fixed-point strain of wheel web; the force moni-
toring method introduced in this paper can obtain the wheel-rail contact forces accu-
rately and simply, which has important theoretical guiding significance for wheel-rail
contact forces monitoring.
Keywords Wheel-rail contact forces · Condition monitoring · Space fixed-point

strain · DIC
1 Introduction
Wheel-rail contact is the most basic and complicated problem in vehicles and track
systems. The accurate and simple measurement of wheel-rail forces are the basis for
in-depth study of wheel-rail contact. It is the theoretical basis for studying the train
performance and analyzing the causes of wheel-rail surface damages [1].
X. Wang · Z. Lu (B) · J. Wei · Y. He

Institute of Rail Transit, Tongji University, Shanghai 201804, China
e-mail: luzhenggang@tongji.edu.cn

https://doi.org/10.1007/978-981-15-9199-0_27
296 X. Wang et al.
At present, the measurement of wheel-rail contact forces can be carried out mainly
by special wheelsets equipped with strain gauges and slip rings or telemeters [2, 3].
Since the web strain is modulated by the rotation angle of the wheel, researchers need
to spend much effort to establish various bridge models to study how to demodulate
the wheel-rail forces more accurately [4–6]. Bridge placed on axle is simple and
convenient, and application cost is low. The disadvantage is that the axle strain is
less sensitive to the position of the wheel-rail contact point [2]. Since the instrument
wheelset method is a contact measurement, current collecting ring or telemeter for
transmitting data is required, which makes the measurement of wheel-rail force that
requires the replacement of special instrument wheelset [7]. Japanese scholars have
proposed a method for measuring the lateral force according to the lateral deformation
of the wheel. This method does not need to attach the strain gauge to the wheel pair
and install the load sensor. The vertical force can be additionally measured by the
spring deformation and strain of the bogie [8].
Digital Image Correlation (DIC) is an image measurement method using computer
vision technology. It is a non-contact method for full-field shape, deformation, and
motion measurement. It uses the randomly distributed spots on the surface of the
object or the pseudo-random distribution artificial speckle field as the deformation
information carrier (Fig. 1). It is a new experimental mechanics method for full-field
displacement and strain analysis of materials or structure surfaces under external
load or other factors. At present, DIC-based strain non-contact measurement tech-
nology has been relatively mature [9, 10]. For example, a picture of marble surface
is stretched vertically with deformation of 3.11% (Fig. 2). The DIC analysis can
calculate the surface strain of about 3.26%, which is consistent with the actual value.
Based on the DIC technology to obtain the space fixed-point strain of the wheel
web surface, the FEM analysis shows that the strain at some space fixed-points of
the wheel web is linear with lateral/vertical contact force. Based on this conclusion,
the measurement method of this paper is proposed: The numerical equation between
the space single fixed-point strain and the lateral/vertical contact force is established.
By decoupling two point equations, the lateral and vertical forces can be obtained.
Fig. 1 Photographing equipment and DIC technology analysis results

Wheel-Rail Contact Forces Monitoring Based on Space Fixed-Point … 297
Fig. 2 Image surface strain calculation based on DIC
2 Wheel Static Finite Element Model
2.1 Finite Element Modeling
According to the wheelset parameters of metro vehicle, wheelset 3D model is estab-

lished (Fig. 3). Since only the influence of lateral and vertical forces on the strain
of web is analyzed, FEM simulations are performed on one wheel. Establish wheel
finite element static model (Fig. 4), set wheel material to structural steel, material
constant: Young’s modulus E = 2.1 × 105 MPa, Poisson’s ratio ν = 0.3, wheel
center fixed. A small elliptical region is taken at the nominal rolling circle of the
wheel to approximate the contact spot, and lateral/vertical loads are applied to the
elliptical region.
Fig. 3 Wheelset 3D model

298 X. Wang et al.
Fig. 4 Finite element model

of wheel
2.2 Result Analysis
After the wheel FEM is established, the lateral and vertical loads are applied to the
contact spot, respectively, and the vertical strain of web surface is calculated.
Figures 5 and 6 show the simulation results of vertical strain on web surface after
wheel is loaded. It can be observed from the results that when wheel is only loaded by
lateral force; there is a stress concentration point below the axle of the web (Fig. 5a).
When wheel is only loaded by vertical force, there is a stress concentration point
below the junction between the rim and web (Fig. 5b). Moreover, under the inde-
pendent action of lateral force or vertical force, the strain at the stress concentration
point is linear with lateral/vertical contact force (Fig. 6). Further research found that
a. Lateral force loaded b. Vertical force loaded
Fig. 5 Vertical strain cloud of wheel web

a. Lateral force loaded b. Vertical force loaded
Fig. 6 Vertical strain at different wheel-rail contact force (stress concentration point)
vertical strain at each point on the web surface is linear with the lateral and vertical
forces.
3 Measurement Method
It can be concluded from the FEM simulation that the wheel-rail force is linear with
the vertical strain at each point on the surface of wheel web.
FL · k1 = ε1 (1)
FV · k2 = ε2 (2)
where FL /FV is the wheel-rail lateral/vertical force, and ε1 /ε2 is the vertical strain
of web surface, and k1 /k2 is the coefficient between web vertical strain and
lateral/vertical force.
Since the wheel-rail force is linear with the vertical strain of wheel web, and the
deformation of loaded wheel is very small; when the wheel is loaded by different
wheel-rail contact forces, the principle of strain superposition is satisfied.
F1 · k1 + F2 · k2 + · · · + Fn · kn = ε (3)
According to above analysis, the theory of wheel-rail force measurement based

on space fixed-point strain of web surface can be proposed. Two points are taken
on the surface of the web, which are lateral force strain sensitive points and vertical
force strain sensitive points (Fig. 7).
Vertical strains ε1 and ε2 are measured at the same time. According to strain
superposition principle, the strain at these two points is the result of simultaneous
action of lateral force and vertical force, which can be expressed as (4)
300 X. Wang et al.
Fig. 7 Space fixed-points position (two points)

FL · k11 + FV · k12 = ε1
(4)
FL · k21 + FV · k22 = ε2
where k11 , k12 , k21 , k22 are the coefficients between the vertical strain at each point
and wheel-rail force, in units of με/kN. After measuring the two vertical strains, the
wheel-rail lateral and vertical force can be obtained by (5).
−1
FL k k ε
= 11 12 · 1 (5)
FV k21 k22 ε2
4 Simulation Verification
Apply the wheel-rail force to wheel, including varying lateral and vertical force, and
measure the strain at two points on the surface of the web (Fig. 7). The wheel-rail
lateral and vertical force can be calculated according to Eq. (5). By comparing the
calculated wheel-rail force with the actual wheel-rail force, the method of wheel-rail
force measurement proposed in this paper can be verified.
Take two points on the surface of the wheel web, which are lateral force strain
sensitive point and vertical force strain sensitive point (Fig. 6), and calibrate the
coefficients k11 , k12 ,k21 ,k22 .

k11 k12 10.028 −0.678
= (6)
k21 k22 −1.839 −1.150
Applying randomly lateral forces and vertical forces, where the lateral force values
range from 40 ± 40 kN, the vertical force values range from 80 ± 80 kN, and calculate
a total of 22 time steps. According to Eq. (5) and the coefficient matrix (6), the wheel-
rail contact force can be decoupled, and the calculated force is compared with the
actual value (Fig. 8).
80 160 Actual
• • •
Calculated
• • •
140
60 120
Vertical Force /
Lateral Force /
100
40 80
60
20 40
•Actual
• •
20
•Calculated
• •
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Time step Time step

a. Lateral force b. Vertical force
Fig. 8 Actual wheel-rail contact force vs. calculated force
It can be observed from Fig. 8 that during the large-scale random variation of the
wheel-rail lateral/vertical force, the calculated force values are consistent with the
actual values, which can verify the wheel-rail force measurement method proposed
in this paper.
When wheels pass rail joint or worn track, the vertical inertia load will be gener-
ated. If the inertia force is large, then the strain and deformation of the wheel will be
affected. Therefore, it is necessary to study the effect of inertial load on wheel-rail
force to obtain a more accurate measurement result.
The inertia load can be realized by applying acceleration. It can be observed from
Fig. 9 that the inertial load causes a stress concentration point below the axle of the
web, and at the stress concentration point, the inertial load is linear with the surface
strain of the web (Fig. 10). Further research found that vertical strain at each point
on the web surface is linear with the inertial load.
Fig. 9 Vertical strain cloud

of wheel web (loaded by
inertial acceleration)
302 X. Wang et al.
Fig. 10 Vertical strain at

different inertial acceleration
(stress concentration point)
Fig. 11 Space fixed-points

position (three points)
Correct the measurement results of the wheel-rail force: add another measuring
point on the web and measure the strain at three points of the web (Fig. 11). According
to the strain superposition principle (Eq. 3), the wheel-rail force and the vertical
inertial force acting on the wheel are calculated.
⎡ ⎤ ⎡ ⎤−1 ⎡ ⎤
FL k11 k12 k13 ε1
⎣ FV ⎦ = ⎣ k21 k22 k23 ⎦ · ⎣ ε2 ⎦ (7)
Acc k31 k32 k33 ε3
where Acc is vertical inertial acceleration.

Applying randomly lateral forces, vertical forces, and vertical inertial acceleration,
where the lateral force values range from 40 ± 40 kN, the vertical force values range
from 80 ± 80 kN, the vertical inertial acceleration values range from 72.5 ± 72.5 g,
and calculate a total of 22 time steps. It can be observed from Fig. 12 that calculated
wheel-rail force, and the actual value are consistent, which verifies the wheel-rail
force measurement method when there is vertical inertial load.
80 160 Actual
• • •
Calculated
• • •
140
60 120
Vertical Force /
Lateral Force /
100
40 80
60
20 40
•Actual
• •
20
•Calculated
• •
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Time step Time step
a. Lateral force b. Vertical force
)
1500 •Actual
• •
•Calculated
• •
Vertical Acceleration / (
1000
500
0
0 5 10 15 20 25
Time step
c. Vertical inertial acceleration
Fig. 12 Actual wheel-rail contact force versus calculated force
5 Conclusions and Future Work
5.1 Conclusions
By static FEM analysis of wheel, the theory and method of wheel-rail force measure-
ment are proposed and verified by simulation. The simulation considers the influence
of inertial load. The numerical correction under inertial force conditions is performed.
Research indicates:
1. The vertical strain at each space fixed-point on the surface of the wheel web is
linear with the wheel-rail contact lateral and vertical forces.
2. The wheel-rail force measurement method based on the space fixed-point strain
of the web is feasible. When the lateral force and the vertical force change in a
large range, the force value can be accurately identified.
3. The vertical inertial load has an influence on the measurement accuracy. Through
the measurement correction method, the influence of the inertial force can be
basically eliminated, and the value of the vertical inertial acceleration can be
calculated.
304 X. Wang et al.
5.2 Future Work
1. Study the applicability of the wheel-rail force measurement method under the
conditions of wheel scratches and polygons, track flattening and wave grinding,
etc. The effect of the position change of wheel-rail contact point on the space
fixed-point strain of the web should be studied.
2. DIC technology application in the state of wheel rotation and experimental
verification.
Acknowledgements The authors thank the financial support from National Key R and D Program
of China (No. 2017YFB1201302-12A).
References
1. Gullers P, Anderssona L, Lunden R (2008) High-frequency vertical wheel-rail contact force-

field measures and influence of track irregularities. Wear 265(9):1472–1478
2. Hertber H (1981) Determination of wheel/rail forces by means of measuring wheelsets
on Deutsche Bundesbahn. In: International conference of wheel/rail load and displacement
measure techniques. Transportation Systems Center, MIT, Cambridge, pp 19–21
3. Zeng YQ et al (1998) Theory of high accuracy load measuring wheelset. J China Railway Soc
20(6):28
4. Nielsen COJ (2008) High-frequency vertical wheel-rail contact forces-validation of a prediction
model by field testing. Wear 265:1465–1471
5. Otter DE (1991) A design for next generation load measuring wheelsets. IEEE/ASME, May
1991
6. Ren Y (2011) Research on key technology of wheel/rail contact status measurement. Southwest
Jiaotong University, Chengdu
7. Magel E et al (2008) Traction, forces, wheel climb and damage in high-speed railway operations.
Wear 265:1446–1451
8. Matsumoto A (2008) A new measuring method of wheel-rail contact forces and related
considerations. Wear 265:1518–1525
9. Rommel C (2008) Strain Measurements with the digital image correlation system Vic-2D.
University of Puerto Rico, Mayag
10. Vanderesse N et al (2018) Measurement of deformation heterogeneities in additive manu-
factured lattice materials by digital image correlation: strain maps analysis and reliability
assessment. J Mech Behav Biomed Mater 86:397–408
Experimental Validation of the Blade
Excitation in a Shaft Vibration Signals
V. Vasicek, J. Liska, J. Strnad, and J. Jakl
Abstract Monitoring of blade vibrations is an important part in the maintenance of

the rotating machine and its proper operation. The signal analysis of blade vibration
plays an important role in detection of blade structure change. Early detection can
avoid unnecessary losses. This is important especially for blades in the last stage
of low-pressure turbine. The popular method used for blade vibration monitoring is
blade-tip timing, which is not installed standardly yet due to the higher installation
costs. An alternative method for blade vibration monitoring is based on relative shaft
vibration signal analysis. This signal is measured by standardly installed instrumen-
tation thus any additional sensor installation is not necessary. Recently, a few papers
were published showing that rotor oscillations contain the information about the
blade vibrations. However, the principle of the blade oscillations transition into the
shaft vibrations was not described in detail so far. Explanation of main principle of
rotor excitation by rotating blades is described in this paper. The described theory
is experimentally validated using a machinery fault simulator. The experiment was
based on the excitation of the forced blade vibrations by the piezo elements installed
on the bladed disc, which was mounted on the simulator shaft. Excitation of piezo
elements was set in an appropriate form to be able to clarify the principle and validate
theory which was described in this paper.
Keywords Blade vibrations · Shaft · Experimental validation · Relative shaft

vibrations
V. Vasicek (B) · J. Liska

NTIS—New Technologies for Information Society, Univerzitni 8, 30614 Pilsen, Czech Republic
e-mail: vasicekv@ntis.zcu.cz
J. Liska · J. Strnad · J. Jakl
Department of Cybernetics, University of West Bohemia, Pilsen, Czech Republic

https://doi.org/10.1007/978-981-15-9199-0_28
306 V. Vasicek et al.
1 Introduction
Condition monitoring is a process of measure and evaluation of a parameter, associ-

ated with machinery operation, for the purpose of determining the machine integrity
[1]. Early detection of an impending fault is important especially in a power genera-
tion industry, in which any outage is related to large financial loss. One of the most
stressed parts in a steam turbine is a moving blade. Popular method for blade vibra-
tion monitoring is a tip-timing method in nowadays [2]. Unfortunately, the measuring
system and its installation are expensive, so the blades remain often unmonitored.
Recently a few papers by the authors of this article were published in order to
introduce a use of standardly measured shaft vibration signal, which contains also the
information about the blade vibrations, and so can be used for blade diagnostics, for
example [3, 4]. Unfortunately, the principle how the blade oscillations are transmitted
into the shaft vibrations is not described in the literature yet. The decision whether
the blade oscillations cause the tension transverse wave spreading along the shaft
surface, or the shaft bending, should be presented in this paper and also validated by
the experiment.
2 Asymmetry in Blisk Vibration
The displacement of the blade causes the bending moment acting on rotor causing
the shaft bending. The consequence of acting forces is the shaft radial deflection
in a sensor plane. The blade oscillations, blade moment, and the radial vibrations
are assumed as harmonic and so the trajectory of the overall bending moment can
be described according to (1). In this complex equation, blade oscillation frequency
is represented by f , phase of blade oscillation is ϕ, i indicates the blade, N is the
number of the blades, φ represents the angle between the blade and the reference on
the disc, and Mi is the amplitude of bending moment contribution of i-th blade. It can
be shown that the trajectory of overall bending moment (1) is elliptical in general.

N
M(t) = Mi cos(2π f t + ϕi )e jφi (1)
i=1
It can be shown, that all nodal diameters except 1 ND are symmetrical and so the
overall bending moment is zero. However, in reality, the blades are mistuned and
so the deviation from the idealistic nodal diameter occurs and the overall bending
moment is nonzero. The example of the bladed disc oscillation measured by tip
timing is in Fig. 1. The amplitude of blade oscillation is expressed by the length of
the polar plot column and the color represents the phase. Overall bending moment
is represented by the blue elliptical trajectory, which was evaluated using Eq. (1).
Experimental Validation of the Blade Excitation in a Shaft … 307
Fig. 1 Illustration of blisk vibration measured by tip-timing
3 Experiments
For the validation of the theory introduced in the previous chapter, the experimental
machine fault simulator Magnum from SpectraQuest was used. The diameter of the
shaft is 0.75 in. and its length is 35 in. The shaft is connected to motor by the flexible
coupling and it is housed in 2 sleeve bearings with multiple oil inlets. Distance of
the bearings is 28.5 in. and lubrication of them is provided by oil system, which is
pressured by the external supply.
For the testing purposes the bladed disc, which is illustrated in Fig. 2, was designed
and manufactured. The bladed disc is made of the 2 mm thick sheet metal and it is
mounted on the shaft using 2 flanges slipped on the shaft from both sides and screwed
together using screws passing through each blade. By the tightening of the screws,
it is possible to change the eigenfrequency of belonging blade. The excitation of
the blades is made by the piezo disc supplied by the external voltage signals with
adjustable amplitude, frequency and phase.
There were made 3 experiments—A, B, C. The illustration of the experiments is
in Fig. 3. The parameters of the experiments are in Table 1. Excitation was applied
on blade 1 and 3 with respect to blade number 0, which is illustrated in Fig. 3 and
which is stated as the reference blade with angle equals to 0. Numbering of the blades
was made counterclockwise. The angles of piezo disc in corresponding coordinate
system are then 45° and 135°. The external excitation signal had uniform amplitude,
frequency equals to 300 Hz, which was outside of the resonance frequency of the
blades, the blisk itself and the shaft, and thus the forced vibrations were applied. The
phase of the excitation signal for each experiment is listed in Table 1.
Fig. 2 Bladed disc and piezo actuators of experimental rotor stand
Fig. 3 Experiments A, B, C and the overall bending moment forces acting on the shaft
Table 1 Excitation blade

A B C
parameters of experiments A,
B, C φ (°) 45 45, 135 45, 135
ϕ (°) 0 0, 0 180, 0
Table 2 contains the theoretical results of the shaft vibration if the principle that the
blade oscillations cause the shaft bending is true. The selection of excitation signal
phases made in order to the shaft radial movement was acted in just one direction φ.
The trajectory of overall bending moment should be a line in that case. The amplitude
of overall acting force is Fo . The results can be also read in Fig. 3, where the shaft
vibration is illustrated by the vector as the result of overall bending moment (1). The
Table 2 Theoretical shaft

A B C
radial vibrations of
experiments A, B, C φ (°) 45 90 0
√ √
Fo F 2·F 2·F
shaft radial oscillation is caused by overall force evaluated as a vector addition of

each blade force contribution. Illustration using the vectors can be made just in of
chosen parameters.
4 Results
The results of the experiments are illustrated in this chapter. To easily validate the
theory described in previous part of this paper using the experiments A, B, C, the
blisk was for each experiment turned into the position that the direction of expected
overall bending moment was the same as the direction of the shaft vibration sensor.
For measuring by the horizontal sensor, the blisk was turned by 45° clockwise in
case of experiment A, by 90° clockwise in case of experiment B, and by 0° in case of
experiment C. Then only the horizontal sensor should measure radial shaft vibration
and the vertical sensor should measure just the noise. For measuring by the vertical
sensor, the blisk was turned by 45° counterclockwise in case of experiment A, by
0° in case of experiment B, and by 90° counterclockwise in case of experiment C.
Then only the vertical sensor should measure radial shaft vibration and the horizontal
sensor should measure just the noise, because the blisk force acting on the shaft is
only in vertical direction.
Figure 5 compares the results by the experiment A and B. The left column was
measured by horizontal shaft vibration sensor, the right column was measured by
vertical shaft vibration sensor. The first row represents the case, when the direction of
acting overall bending moment was horizontal. The second row represents the case,
when the direction of acting overall bending moment was vertical. The layout of the
graphic is illustrated in Fig. 4, where the bearing pedestal, relative shaft sensors and
shaft expected movement represented by the red arrows are presented.
All the graphs in Fig. 5 are Fourier transform amplitude spectrums. It can be clearly
seen, that the excitation of the shaft can be truly measured only when direction
of overall bending moment and the sensor equals. In other case only the noise is
measured. To be able also to compare the amplitudes, whose √ values are in Table 2,
the amplitude for 300 Hz of experiment B was divided by 2. From Fig. 5 can be
also discovered, that the amplitudes are as expected, because the 300 Hz components
are almost the same.
The same results can be seen in Fig. 6, where the experiments A and C are
compared. The evaluation of the results was the same as in case of A and B.
Fig. 4 Layout of the results in Figs. 5 and 6—the direction of overall bending moment force acting
on the shaft—red arrow, and sensor used for relative shaft vibration measurement
10-3 10-3
2 2
ampl [mm]
ampl [mm]
1 1
0 0
295 300 305 295 300 305
frequency [Hz] frequency [Hz]
10-3 10-3
4 4
ampl [mm]
ampl [mm]
2 2
0 0
295 300 305 295 300 305
Fig. 5 Experiments A (blue), B (red) result comparison—left column represents horizontal sensor,
right column represents vertical sensor, the first row represents the situation when force acting on
shaft has horizontal direction, the second row represents the situation when force acting on shaft
has vertical direction
-3 -3
10 10
2 2
ampl [mm]
ampl [mm]
1 1
0 0
295 300 305 295 300 305

-3 -3
10 10
4 4
ampl [mm]
ampl [mm]
2 2
0 0
295 300 305 295 300 305
Fig. 6 Experiments A (blue), C (red) result comparison—left column represents horizontal sensor,
right column represents vertical sensor, the first row represents the situation when force acting on
shaft has horizontal direction, the second row represents the situation when force acting on shaft
has vertical direction
5 Conclusions
The results of the experiments have shown that the principle of the blade oscillations
presence in the shaft vibrations is related certainly to the shaft bending. The theoret-
ical background was described in the beginning of this paper. The experimental rotor
stand was used to experimental validation. This paper has a few important conclu-
sions. First, the shaft vibration signal contains the information about the asymmetry
of the blisk vibration, so the absolute level of blade vibrations cannot be evaluated.
The second, the asymmetry of blisk vibration, is not unique for only one config-
uration of blade oscillations. That was shown √ in A and B experiment results. If
amplitude of the excitation signal would be 2 times higher, then the results for A
and B would be the same although the different blades were excited. The third, even
the same blades are excited, the different response is measured. It is reflected by the
B and C experiment. The difference is made by the phase difference of excitation
signals. The inverse phase of excitation signal of one blade in experiment C caused
the inverse direction of shaft vibration motion. The further research should include
the conclusions of this paper and the aim should be to include also the rotation of
the disc into the measured signal and its analysis.
Acknowledgements This work was supported from ERDF under project “Research
Cooperation for Higher Efficiency and Reliability of Blade Machines (LoStr)” No.
CZ.02.1.01/0.0/0.0/16_026/0008389 and by the project from the Ministry of Education, Youth
and Sports of the Czech Republic, PUNTIS-LO1506.
References
1. Mechefske C (2005) Machine condition monitoring and fault diagnostics. In: Vibration
monitoring, testing, and instrumentation, pp 25-2
2. Heath S, Imregun M (1998) A survey of blade tip-timing measurement techniques for
turbomachinery vibration. J Eng Gas Turbines Power 120:784–791
3. Liska J, Vasicek V, Jakl J (2018) On possibilities of using relative shaft vibration signals for
rotating blades monitoring. In: ASME turbomachinery technical conference and exposition,
Oslo
4. Liska J, Vasicek V, Jakl J (2016) Turbine blades state monitoring based on advanced order
analysis. In: 13th International conference on condition monitoring and failure prevention
technologies, Paris
Seismic Risk Evaluation by Fragility
Curves Using Metamodel Methods
H. Z. Yang and C. G. Koh
Abstract This study proposed an efficient and reliable fragility estimation by meta-
model methods. To evaluate seismic risk, it is important to account for ground motions
and structural parameter uncertainties. Nevertheless, seismic fragility analysis
with uncertainties is impractical since it requires many time-consuming nonlinear
dynamic simulations. To this end, an efficient approach based on metamodel in
conjunction with Monte Carlo simulation is proposed to develop fragility curves.
Several metamodel methods such as kriging, response surface method (RSM), and
radial basis function (RBF) are investigated for this purpose. Optimum Latin hyper-
cube design is used to generate “space filling” samples for nonlinear dynamic analysis
under seismic excitations. A metamodel is constructed based on these design of exper-
iment samples which include the input parameters and output fragility results. Kriging
and RBF are better than the RSM metamodel. The fragility curves can be generated
based on metamodel by considering the randomness of earthquake ground motions
and uncertainties in material properties. Finally, seismic risk is evaluated by fragility
curves. The computation time is significantly reduced by applying the metamodel
method with acceptable accuracy. The proposed methodology also avoids nonlinear
dynamic non-convergent problems. Kriging and RBF methods complement each
other and are able to accurately evaluate fragility curves.
Keywords Fragility curves · Earthquake · Metamodel · Probabilistic seismic

assessment · Monte Carlo simulation
H. Z. Yang (B)
School of Engineering, University of Glasgow, Glasgow, UK
e-mail: Hezhen.yang@Glasgow.ac.uk
C. G. Koh
Department of Civil and Environmental Engineering, National University of Singapore,
Singapore, Singapore

https://doi.org/10.1007/978-981-15-9199-0_29
314 H. Z. Yang and C. G. Koh
1 Introduction
Consideration of uncertainty in seismic analysis is important. Primarily because

earthquake is a random natural event and no two earthquakes are alike [1]. In addition
to uncertainties in seismic ground motions, seismic risk evaluation should consider
uncertainties in structural and material parameters. Probabilistic methods provide
a quantitative means to account for these inherent uncertainties in structural safety
assessment [2]. In this context, fragility curves are very useful in presenting the proba-
bility of failure (in terms of life safety, for example) with respect to an input parameter
such as peak ground acceleration (PGA) [3]. They can be developed by employing
analytical, empirical and hybrid methods, and expert judgment in some cases. Analyt-
ical fragility methods have attracted significant attention from researchers because of
their theoretical basis over other methods. However, these methods involve solving
nonlinear dynamic equations and the computational effort is high for the multiple
simulations needed in considering seismic uncertainties.
Fragility analysis has been widely applied for seismic evaluation of bridges [4],
frame buildings [5], and masonry buildings [1]. These methods have the potential to
overcome the excessive computation time problem for seismic probabilistic evalua-
tion. However, there is a scarcity of advanced metamodel methods such as kriging and
radial basis function (RBF) that evaluate the seismic probability fragility analysis.
The main purpose of this study is to develop a methodology to obtain fragility
curves with high efficiency and accuracy. Typical random variables are chosed to
describe uncertainties affecting the structural behavior. There are two challenges
involved in this task. The first challenge is to reduce the huge computation time. The
second challenge is to overcome non-convergence problem in numerous nonlinear
dynamic analyses. A four-story building in Italy [6] is selected as an illustration case
study. Three metamodel methods are compared, such as RSM, kriging, and RBF
metamodels.
2 Theories and Methodology
2.1 Fragility Function
A fragility function is defined as:

ln(x/θ )
P(C|IM = x) = (1)
β
where P(C|IM = x) is the probability of structure to collapse, IM = x is a ground

motion with intensity measure, is the standard normal cumulative distribution, θ
is the median of the fragility function, and β is the standard deviation of ln(IM).
Seismic Risk Evaluation by Fragility Curves Using Metamodel … 315
2.2 Analytical Fragility Curves Development Methods
Vamvatsikos and Cornell [7] predicted nonlinear structural dynamic properties under
seismic ground motions by the incremental dynamic analysis (IDA). IDA utilizes
scaled ground motions to find different structural damage states till the structure
collapses. IDA aims to develop analytical fragility curves. Multiple-stripe anal-
ysis (MSA) can also establish accurate fragility curves with a minimal number of
nonlinear dynamic analyses. We assume that the observation of collapse from each
ground motion is independent records. The probability that a ground motion with
IM = x j is given by the binomial distribution as follow:

nj z n −z
p Z j collapses in n j ground motions = pjj 1 − pj j j (2)
zj
Equation (1) is substituted for pj , so that the fragility parameters are explicit as
follows:
m
nj ln x j /θ ln x j /θ
Likelihood = 1− (3)
j=1
zj β β
Estimates of the fragility function parameters are obtained by maximizing this

likelihood function as shown in the following equation.

m
nj ln x j /θ
θ̂ , β̂ = arg max ln + z j ln
θ,β
j=1
zj β

ln x j /θ
+ n j − z j ln 1 − (4)
β
2.3 Metamodel Methods
Metamodel methods are applied for predicting the relationship between the inputs
and outputs with high efficiency. As shown in Fig. 1, the basic principle of metamodel
methods can be described as follows. The original system solution space includes
input and output information. First, the input design parameters are generated with
design of experiment. Second, the output response is calculated from the original
system solution space. Third, the input and output relationships are constructed by
metamodel. If the metamodel accuracy is sufficiently high, the metamodel can be
used instead of the original system without requiring time-consuming calculations
to obtain the output responses. Convergent problem in nonlinear dynamic analysis
Fig. 1 Basic principle of metamodel methods
process is unavoidable. During DOE process, selecting sample points which are
near the convergent problem point can smooth calculation samples. In this way, the
metamodel method can overcome convergent problem or failure cases during the
seismic risk analysis for buildings.
2.4 Metamodel Accuracy Evaluation
The accuracy of the metamodel is evaluated by several statistical indicators. The

root-mean-square error (RMSE) and the R2 for the data points used to construct the
metamodel are calculated. Furthermore, the average error (AveErr), RMSE and the
maximum error (MaxErr) for all of the data points which not used in generating the
models are also checked. These statistical indicators may be defined as follows:
SS E
R2 = 1 − (5)
S yy
n 2

n
n
i=1 yi
with SS E = (yi − ȳi ) S yy =
2
yi2 − (6)
i=1 i=1
n
1 |yi − ȳi |
n
AvgErr = (7)
n i=1 yi

n
t=1 (yi − ȳi )
1 2
n
RMSE = 1 n (8)
n i=1 yi

|yi − ȳi |
MaxErr = Max (9)
i yi
3
3
3
3
3
3
3
4. 4. 4.
Fig. 2 Finite element model of the four-story building and layout for the building
where n is the number of data, yi is the measured response and ȳi is the predicted
response for each data point.
3 Numerical Cases
3.1 Four-Story Building Model
Fragility analysis is applied to a four-story building model and the layout, as shown
in Fig. 2. The model describes a typical four-story building constructed in the 1960s
in Italy and strengthened with reinforced concrete jackets. The detailed finite element
method of the four-story building models is created by Seismosoft [6]. To illustrate
the effectiveness of the proposed methodology, fragility curves are evaluated by the
metamodel. First, the optimum Latin hypercube simulation method simulated design
points for the seismic dynamic analysis. Second, RSM, RBF, and kriging methods
are employed for constructing a metamodel. Thirdly, Monte Carlo simulations based
on the metamodel are employed for the fragility analysis of the four-story building.
3.2 Uncertainties
Building performance under future earthquakes is largely unknown and cannot

be predicted with certainty. In this study, the compressive and tensile strengths
of concrete, as well as steel yield strengths, are chosen as principal random vari-
ables. PGA is assumed to have a uniform distribution with a range 0.6–1.2 g. The
distribution characteristics and the values are given in Table 1.
Twenty uncorrelated earthquake acceleration histories consistent with the design
requirements are artificially generated and used as input ground motion. Input vari-
ables include the uncertainties in construction material properties and uncertainties
Table 1 Uncertainties of random variables

Variable Mean (MPa) COV (%) Probabilistic distribution
Concrete yield strength (f c ) 24 21 Normal
Young’s modulus of concrete (E c ) 23,025 20.6 Normal
Steel yield strength (f y ) 400 10 Normal
Table 2 Limits associated

Structural performance level Permissible top drift ratio (%)
with various structural
performance levels Immediate occupancy (IO) 1
Life safety (LS) 2
Collapse prevention (CP) 4
in ground motions. Uncertainties in ground motions use a suite of artificial accelero-

grams [8, 9]. Three performance levels defined in FEMA 356 are used, as shown in
Table 2.
3.3 Comparison of Multiple Stripes Analysis

and Incremental Dynamic Analysis
The incremental dynamic analysis (IDA) can be used to obtain seismic capacity of
the structure. The different damage states could be obtained from the capacity curves.
The fragility curves described the probability of the defined limit state of a structure
as a function of ground motion intensity measures, IM.
Multiple-stripe analysis (MSA) is a collection of single-stripe analyses performed
at multiple levels of the spectral acceleration [10]. This will provide estimates of
fragility curves parameters (θ, β) for each curve. Ground motion selection needs
to be both “sufficient” and “efficient.” This work adopts the spectral representation
based simulation methodology used in stochastic engineering. For low-rise build-
ings, approximately 10–20 ground motion records are usually sufficient to provide
adequate accuracy in the estimation of seismic demand. Here, we compared 12, 20,
and 48 records. As shown in Fig. 3, the accuracy of 20 ground motions is similar
to 48 ground motions of the IDA developed fragility curves. There is a significant
error for 12 ground motions records compared with 48 records. Hence, 20 ground
motions records are selected for fragility analysis in this study.
Fig. 3 The fragility curves developed by MSA and IDA with different ground motion records
3.4 Metamodel Accuracy Evaluation
In the present study, the optimal Latin hypercube method is used as design of experi-
ment (DOE) method. Several metamodels are constructed to perform the comparison.
A good fit is achieved if the coefficient of determination R−2 is close to 1. This coef-
ficient is nevertheless often insufficient for accuracy evaluation since it increases
with the number of data points used. Hence, validating the model using additional
data points which are not used in the training set is essential. For the purpose of
these statistical tests, 45 additional combinations of input variables are generated
randomly. The resulting statistical measures are computed and displayed in Table 3.
The R2 value is also reduced when few sample data are used, as shown in data
set II and III. If there is a set of data samples with a significant difference between
the measured and predicted values, the R2 approaches zero or negative. In Table 3,
dataset II and data set III are used to conclude that there is a significant amount
of error in the metamodel. The maximum absolute error for β is larger than 50%.
The root-mean-square error for β is larger than 10%. In data set I, the 30 samples
RSM is able to predict θ with good accuracy. However, it is unable to predict β
with acceptable accuracy. In data set I, the R2 for the RSM predicting β is smaller
than 0.5, maximum absolute error for the RSM predicting β is larger than 50%. For
RBF and kriging model, the R2 is larger than 0.5 and the maximum absolute error
is smaller than 50%. Both the 30 samples for RBF and the 30 samples for kriging
are able to predict (θ, β) with acceptable accuracy. According to the above analysis,
Table 3 Metamodel accuracy evaluation indicators for the prediction of the fragility
Data Statistical criteria Fragility curves Fragility curves parameter
parameter (θ) (β)
RSM RBF Kriging RSM RBF Kriging
Set I (30 points) R2 0.94 0.95 0.95 0.34 0.773 0.67
AveErr (%) 1.66 1.49 1.53 14.68 8.577 10.39
RMSE (%) 1.40 0.83 0.79 10.59 3.769 4.68
MaxErr (%) 3.97 4.06 5.54 50.76 36.29 43.26
Set II (18 points) R2 0.92 0.91 0.91 0.16 0.252 0.22
AveErr (%) 1.83 1.95 1.95 16.53 15.573 15.92
RMSE (%) 1.39 1.26 1.24 10.75 8.791 8.74
MaxErr (%) 4.45 4.74 5.46 68.53 64.04 61.72
Set III (12 points) R2 0.71 0.90 0.87 −1.94 −0.26 −0.43
AveErr (%) 3.55 2.04 2.40 30.86 20.22 21.51
RMSE (%) 2.89 1.64 2.00 24.49 15.04 16.63
MaxErr (%) 8.63 4.34 5.92 97.28 66.59 70.79
data set I is used to construct the metamodel for the fragility analysis by considering
uncertainties.
Kriging and RBF metamodels are more appropriate than RSM metamodel for
the relationship between the input probabilistic seismic analysis and output fragility
analysis results which have higher order nonlinearity and sometime even discon-
tinuous properties. Kriging and RBF metamodels are more appropriate than RSM
metamodel for simulating such relationships. It is very important to use a suitable
metamodel to represent the most significant properties.
3.5 Development of Fragility Curves by Considering

Uncertainties
The proposed approach utilizes the metamodel method for obtaining fragility curves
considering uncertainties. The performance comparison of the fragility analysis time
with and without the use of metamodel is shown in Fig. 4. The computation time is
based on a desktop computer with a single 3.4-GHz processor. Monte Carlo simula-
tions of 10,000 samples are used for developing fragility curves considering uncer-
tainties. In the case of twenty ground motions, the computation time is 16.3 days
with applying the metamodel but would be increased by more than two orders of
magnitude without applying the metamodel. The strategy for developing fragility
curves is based on metamodel which does not require a large number of nonlinear
dynamic analyses. Hence, the metamodel method can reduce the computation time
significantly and help in avoiding nonlinear dynamic non-convergent problem.
Fig. 4 Comparison of consuming time of develop fragility curves considering uncertainties
The metamodel method is used in calculating the fragility curves for the building
by considering ground motions and material properties uncertainties. Tens of thou-
sands of fragility curves are developed based on constructed kriging and RBF, as
shown in Fig. 5. But in order to avoid underestimation of the building fragility, a
more robust means of probabilistic seismic risk evaluation is necessary. To obtain the
fragility curves, 10,000 simulations are sufficient [11]. Using the metamodel, Monte
Carlo simulations of 10,000 samples are used to estimate the mean and standard devi-
ation. Since simulations are conducted based on the metamodel, the fragility analysis
only requires several minutes (on a single 3.4 GHz processor). Fragility curves for
Fig. 5 Fragility curves generated by Monte Carlo simulation based on the metamodel
the three limit states (immediate occupancy, life safety and collapse prevention) are
determined by considering uncertainties that are scattered within some ranges.
4 Conclusions
The proposed approach utilizes the metamodel in combination with Monte Carlo
simulations (MCS) for obtaining fragility curves. This strategy makes it practical
to carry out probabilistic seismic risk evaluation by using MCS which would other-
wise be very computationally demanding. The randomness of the ground motions
and material uncertainties is considered in the fragility analysis. Among the three
metamodel methods studied, kriging and RBF are able to accurately predict the
fragility curves when compared with response surface method (RSM). Fragility
curves considering uncertainties can be predicted directly based on metamodels
for seismic risk assessment of buildings. By this way, the fragility analysis can
avoid requiring the repetition of seismic nonlinear dynamics. The advantages of the
proposed methodology are as follows.
1. Uncertainties in material properties and ground motion randomness are consid-
ered in seismic risk assessment.
2. The computational time is significantly reduced by two orders of magnitude.
3. The metamodel method could smooth calculation samples. Hence, the nonlinear
dynamic analysis convergent problem could be avoided by using the MSA
methodology.
4. The main strength of this method lies in its versatility. Uncertainties in both the
structure model and ground motion model can be considered easily.
Acknowledgements This work was financially supported by the Ministry of Education of

Singapore (Academic Research Fund R-302-000-095-112).
References
1. Rota M, Penna A, Magenes G (2010) A methodology for deriving analytical fragility curves
for masonry buildings based on stochastic nonlinear analyses. Eng Struct 32:1312–1323
2. Padgett JE, DesRoches R (2007) Sensitivity of seismic response and fragility to parameter
uncertainty. J Struct Eng ASCE 133:1710–1718
3. Billah AHMM, Alam MS (2015) Seismic fragility assessment of highway bridges: a state-of-
the-art review. Struct Infrastruct Eng 11:804–832
4. Choi ES, DesRoches R, Nielson B (2004) Seismic fragility of typical bridges in moderate
seismic zones. Eng Struct 26:187–199
5. Kircil MS, Polat Z (2006) Fragility analysis of mid-rise R/C frame buildings. Eng Struct
28:1335–1345
6. Seismosoft (2015) SeismoStruct—a computer program for static and dynamic nonlinear
analysis of framed structures. Available online from http://www.seismosoft.com
7. Vamvatsikos D, Cornell CA (2002) Incremental dynamic analysis. Earthquake Eng Struct

Dynam 31:491–514
8. Saxena V, Deodatis G, Shinozuka M, Feng MQ (2000) Development of fragility curves for
multi-span reinforced concrete bridges. In: Proceedings of international conference on Monte
Carlo simulation, pp 499–504
9. Deodatis G (1996) Non-stationary stochastic vector processes: seismic ground motion
applications. Probab Eng Mech 11:149–167
10. Baker JW (2015) Efficient analytical fragility function fitting using dynamic structural analysis.
Earthq Spectra 31:579–599
11. Saha SK, Matsagar V, Chakraborty S (2016) Uncertainty quantification and seismic fragility of
base-isolated liquid storage tanks using response surface models. Probab Eng Mech 43:20–35
Reliability Centred Maintenance (RCM)
Assessment of Rail Wear Degradation
H. J. Hoh, J. L. Wang, A. S. Nellian, and J. H. L. Pang
Abstract Rail wear degradation durability performance is an important concern that

affects railway track maintenance and strategies to mitigate and prevent costly events
of rail steel failures. An understanding of wear degradation will enable predictions
of rail wear degradation trends, providing guidance for effective rail maintenance
planning. In this work, an approach to predict rail steel wear degradation based on
rail wear records is developed. The approach is based on statistical modelling of
wear degradation data from maintenance records. For each track type and radius,
available historical wear measurements are extrapolated to wear limits (as time-to-
failure), and suitable continuous probability distributions are used to fit these failures.
The low rail head wear and high rail side wear of sharp curves deemed most critical
are emphasized, while differentiating the different track foundation types (viaduct
or tunnel). Predictions can be made for the degradation trend of rails, providing
guidance for effective rail maintenance planning. As an example, for R300 (sharp
curve radius of 300 m) high rails in the tunnel sections, rail wear failures follow an
increasing failure rate Weibull distribution, with scale parameter of about 20 years.
This indicates that a large proportion of rails will fail within that stipulated time
frame, and hence, it is mandatory for these rails to be replaced in advance prior to
failure.
H. J. Hoh (B)
SMRT-NTU Smart Urban Rail Corporate Laboratory, 50 Nanyang Avenue, Singapore 639798,
Singapore
e-mail: hjhoh@ntu.edu.sg
H. J. Hoh · J. L. Wang · A. S. Nellian · J. H. L. Pang
SMRT-NTU Smart Urban Rail Corporate Laboratory, Nanyang Technological University,
Singapore, Singapore
H. J. Hoh
School of Electrical & Electronic Engineering, Nanyang Technological University, Singapore,
Singapore
J. L. Wang · A. S. Nellian · J. H. L. Pang
School of Mechanical & Aerospace Engineering, Nanyang Technological University, Singapore,
Singapore

https://doi.org/10.1007/978-981-15-9199-0_30
326 H. J. Hoh et al.
Keywords Wear · Reliability · Degradation · Railway · Weibull distribution
1 Reliability and Rail Wear Degradation
The principle railway asset maintenance strategies are corrective maintenance,

preventive maintenance, condition-based maintenance and predictive maintenance,
with each constituting a distinct strategical approach. In this present study, a reliability
centred maintenance assessment is performed to assess rail track wear degradation
by a predictive maintenance approach. While considerable literature study has been
conducted with regards to degradation data analysis, there is limited progress on
predicting rail wear degradation trends at sharp-radius curve track sections.
Degradation refers to the deterioration of measured performance data (factors
known as covariates), with failure occurring when a specified threshold is reached
[1]. Degradation analysis allows better extrapolation than time-to-failure analysis
[2], but requires the deterioration to be measurable and directly associated to the
failure mode. Degradation data analysis can be achieved by stochastic process or
general statistical path models [3]. Foundations of the general statistical degradation
path model framework and methods to estimate a time-to-failure distribution were
established by Lu and Meeker [2].
In the railway industry, degradation applies to the deterioration of rail track in
the form of head wear, fatigue and other defects originating primarily from contact
between wheel-rail interaction [3]. While rail wear is dependent on several factors
such as train traffic, axle load, rail-wheel profile, rail metallurgy as well as envi-
ronmental conditions, it is proven to be more critical at sharp-radius curves where
maximum stresses are present due to the rail-wheel contact conditions [4]. Failure
to address rail wear degradation by adopting decisive maintenance and inspection
strategies will further result in risk of rail breaks and derailments [5]. The wear rates
are predominant on the high rail gauge side corner and the low rail head crown of a
curve track. A track degradation model was used to evaluate the impact of parameters
such as axle load, speed and vehicle types on rail replacement and maintenance costs
[6]. Likewise, an optimized Life Cycle Cost (LCC) model based on rail degradation
data was developed [7].
In general, modelling and prediction of rail track degradation via probability distri-
bution models are challenging due to constraints in acquiring the complete historical
data necessary for degradation analysis. Chattopadhyay and Kumar [8] utilized the
rail wear failure data from the Swedish National Rail Administration to analyse and
predict rail breaks for R500-600 curve sections. Time-to-failure information was
recorded in terms of Millions of Gross Tonnes (MGT), and the Weibull distribution
was used to estimate parameters for modelling and prediction of rail wear life. Freitas
et al. [9] applied linear degradation path and Weibull distribution model to rail wheel
wear analysis in the Brazilian railway network and showed that degradation provided
Reliability Centred Maintenance (RCM) Assessment of Rail … 327
better estimates than time-to-failure analysis. Additionally, Ferreira et al. [3] devel-
oped the study of rail wheel based on a more complete database that considered the
operating conditions’ impact on wheel wear.
This study will focus on studying rail wear degradation by utilizing historical wear
data measurements, based on the different track foundation type and curve radius to
evaluate time-to-failure. These failure data will then be fit to appropriate continuous
probability distribution models for prediction.
2 Rail Wear Degradation Methodology
The rail wear degradation methodology (Fig. 1) requires the data to be prepared,
followed by regression and extrapolation of degradation to determine time-to-failure.
The extrapolated time-to-failure values will then be analysed to determine Weibull
distribution parameters.
This study will focus on the sharp-radius curves where maximum stresses are
present due to rail-wheel contact conditions [4]. In the rail degradation analysis,
failure would be defined as the wear exceeding the permissible wear values. Based on
the methodology (Fig. 1), regression and extrapolation to critical wear are performed
on the historical wear data to determine ’time-to-failure’ information for the sharp-
curved tracks considered. Statistical lifetime data analysis can then be applied to the
time-to-failure set of information to be fit into the Weibull distribution model [10].
The Weibull distribution is the most widely used statistical distribution for relia-
bility, due to its flexibility to characterize different failure rate trends in the reliability
bathtub curve (decreasing, constant, increasing) (Fig. 2). The two-parameter Weibull
probability density function is given as

β t β−1 − ηt β
f (t) = e (1)
η η
2. Wear DegradaƟon
1. Data PreparaƟon 3. LifeƟme Data Analysis
Analysis
Separate covariates for radius,

Identify wear measurement
high/low rail, foundation Select distribution to be fitted
and critical values
types (Weibull)
Determine critical curve

Regression analysis
sectors
Get historical wear Extrapolate degradation to Determine parameters

measurements determine time-to-failure
Fig. 1 Rail wear degradation analysis methodology

Fig. 2 The reliability

bathtub curve Failure rate
Infant mortality Wear-out

decreasing failure rate increasing
failure rate
Useful life
constant low failure rate
Time
where time t ≥ 0, shape parameter β > 0 and scale parameter η > 0. The shape
parameter β represents the reliability bathtub curve (Fig. 2) failure rate trends that
are decreasing (β < 1), constant (β = 1) or increasing (β > 1). The scale parameter
represents the time which it takes for 63.2% of the population to fail. The cumulative
distribution function F(t) (also known as failure distribution) and the reliability
distribution R(t) are
β
− ηt
F(t) = 1 − e (2)
β
− ηt
R(t) = e (3)
while the failure rate is

β t β−1
λ(t) = (4)
η η
The historical wear measurements from year 1999 to 2017 of various metro lines
were filtered to consider only curves from the curves of sharper radius R300 (repre-
sented by data points of radius R275 to R350) and radius R500 (represented by
data points of radius R475 to 550). Other information pertaining to the rail geom-
etry (chainage to determine location, NSEW direction, rail foundation type) was
preserved. In the rail-wheel contact at the curves, head wear is observed to be more
significant on the low rail while gauge corner wear dominates on the high rail. Permis-
sible rail wear values have been specified for wear at the gauge corner (W 3 < 18 mm)
and the head (W 1 < 16 mm) (Fig. 3).
Reference
profile 45°
W1
Worn profile
W3
Fig. 3 Wear measurement of rail profile
(a) High rail gauge corner wear degradation (W2 < 18 mm) (b) Low rail head wear degradation (W1 < 16 mm)
Fig. 4 Degradation plots for radius R300 tunnel curves
3 Wear Degradation Analysis
Linear regression is used to fit [9] the historical wear measurements, which are
then extrapolated to critical values to obtain time-to-failure information. The results
are presented for sharp-radius curves of R300 in the tunnel sections (presented in
logarithmic scale in Fig. 4). Similar regression was conducted for R500 curves in
the tunnel sections (Fig. 5) and in the viaduct sections of the track (Fig. 6).
4 Lifetime Data Analysis
The failure rate of the wear failure mode in rails is expected to show an increasing
failure rate trend, corresponding to the wear out phase of the bathtub curve (Fig. 2).
This can be represented by a Weibull distribution with β > 1 and is corroborated by
the Weibull parameters estimated summarized in Table 1.
Based on the lifetime data analysis of the sharp-radius R300 tunnel section curves
(Fig. 7), the wear out behaviour is more significant (β > 2) for the high rail gauge
Fig. 5 Degradation plots for radius R500 tunnel curves
Fig. 6 Degradation plots for radius R500 viaduct curves
Table 1 Summary of Weibull parameters

Curve radius rail (high/low) Weibull parameters
β η/years
R300 (tunnel, high rail) 2.18 25.2
R300 (tunnel, low rail) 1.57 37.0
R500 (tunnel, high rail) 1.35 20.9
R500 (tunnel, low rail) 1.59 24.8
R500 (viaduct, high rail) 1.42 25.1
R500 (viaduct, low rail) 1.88 58.9
corner wear, while slight wear out behaviour is observed for the low rail head
wear (1 < β < 2). For the moderately sharp curves of radius R500, the wear
for both viaduct and tunnel foundation types were observed to exhibit slight wear
out behaviour (1 < β < 2).
Based on the Weibull parameters, the reliability of the various cases is compared
(Fig. 8). The reliability for wear of low rails in the R500 viaduct and R300 tunnel
, ,
(a) High rail gauge corner wear Weibull plot (b) Low rail head wear Weibull plot
Fig. 7 Weibull plots for R300 curves at tunnel sections
Reliability of rail based on wear failure mode

1
R300-tunnel-high-rail
0.9 R300-tunnel-low-rail
R500-tunnel-high-rail
0.8 R500-tunnel-low-rail
0.7 R500-viaduct-high-rail
Reliability
0.6
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50
Time/years
Fig. 8 Reliability of rail based on wear failure modes for various conditions
are generally higher than that of the high rail wear and the low rail R500 tunnel. In
general, the reliability of the low rail wear is higher than that of the high rail, while
wear is more prominent at the tunnel compared to the viaduct section.
5 Conclusion
An approach to predict rail track wear degradation based on historical rail wear
records is presented. Critical rail track radius of R300 for tunnels and R500 (tunnel
and viaduct) is compared, showing that wear is more prominent at the tunnel, and that
high rail gauge corner wear is more prominent than low rail head wear. Low rail head
wear for cases of R300 tunnel and R500 viaduct is shown to be less prominent than
the other cases. The Weibull parameters estimated are used to generate reliability
plots to plan timely maintenance and replacement of rails.
Acknowledgements This research work was conducted in the SMRT-NTU Smart Urban Rail
Corporate Laboratory with funding support from the National Research Foundation (NRF), SMRT
and Nanyang Technological University; under the Corp Laboratory @ University Scheme.
References
1. Gorjian N, Ma L, Mittinty M, Yarlagadda P, Sun Y (2010) A review on degradation models in

reliability analysis. In: Engineering asset lifecycle management, pp 369–384
2. Lu CJ, Meeker WQ (1993) Using degradation measures to estimate a time-to-failure
distribution. Technometrics 35(2):161–174
3. Ferreira JC, Freitas MA, Colosimo EA (2012) Degradation data analysis for samples under
unequal operating conditions: a case study on train wheels. J Appl Stat 39(12):2721–2739
4. Cannon DF, Edel KO, Grassie SL, Sawley K (2003) Rail defects: an overview. Fatigue Fract
Eng Mater Struct 26:865–886
5. Dick T, Barkan CPL, Chapman E, Stehly MP (2001) Analysis of factors affecting the location
and frequency of broken rails. In: Proceedings of the world congress on railway research,
Cologne, Germany
6. Larsson D (2004) A study of the track degradation process related to changes in railway traffic.
Licentiate thesis, monograph, Luleå Tekniska Universitet, Luleå
7. Reddy V (2007) Development of an integrated model for assessment of operational risks in
rail track, PhD
8. Chattopadhyay G, Kumar S (2009) Parameter estimation for rail degradation model. Int J
Perform Eng 5(2):119–130
9. Freitas MA, de Toledo MLG, Colosimo EA, Pires MC (2009) Using degradation data to assess
reliability: a case study on train wheel degradation. Qual Reliab Eng Int 25(5):607–629
10. Weibull W (1951) A statistical distribution function of wide applicability. J Appl Mech 13:293–
297
Prognostic of Rolling Element Bearings
Based on Early-Stage Developing Faults
M. Hosseini Yazdi, M. Behzad, and S. Khodaygan
Abstract Rolling element bearing (REB) failure is one of the general damages in
rotating machinery. In this manner, the correct prediction of remaining useful life
(RUL) of REB is a crucial challenge to move forward the unwavering quality of
the machines. One of the main difficulties in implementing data-driven methods for
RUL prediction is to choose proper features that represent real damage progression.
In this article, by using the outcomes of frequency analysis through the envelope
method, the initiated/existed defects on the ball bearings are identified. Also, new
features based on developing faults of ball bearings are recommended to estimate
RUL. Early-stage faults in ball bearings usually include inner race, outer race, ball
and cage failing. These features represent the sharing of each failure mode in failure.
By calculating the severity of any failure mode, the contribution of each mode can be
considered as the input to an artificial neural network. Also, the wavelet transform
is used to choose an appropriate frequency band for filtering the vibration signal.
The laboratory data of the ball bearing accelerated life (PROGNOSTIA) are used to
confirm the method. To random changes reduction in recorded vibration data, which
is primary in real-life experiments, a preprocessing calculation is connected to the
raw data. The results obtained by using new features show a more accurate estimation
of the bearings’ RUL and enhanced prediction capability of the proposed method.
Also, results indicate that if the contribution of each failure mode is considered as
the input of the neural network, then RUL is predicted more precisely.
Keywords Prognostic · Neural network · Rolling element bearing · Vibration ·

Remaining useful life
M. Hosseini Yazdi · M. Behzad (B) · S. Khodaygan

Mechanical Engineering, Sharif University of Technology, Azadi St., Tehran, Iran
e-mail: m_behzad@sharif.ir

https://doi.org/10.1007/978-981-15-9199-0_31
334 M. Hosseini Yazdi et al.
1 Introduction
Correct failure time prediction of mechanical elements acts as a vital role in improving
the reliability of production in different industries [1, 2]. Rolling element bearing
(REB) failure is one of the frequent damages in rotating machines, so researchers have
worked on the remaining useful life (RUL) of REB. The effectiveness of vibration
condition-monitoring analysis in this field has motivated many researchers to focus
on this subject [3]. In general, approaches used to predict failure times are divided into
three general categories: physical models, knowledge-based models and data-driven
models. In data-driven models, a feature of the recorded data is required to detect
bearing depression as well as the estimated RUL. The features must have adequate
information about the health of the machine. Artificial intelligence (AI) methods are
one of the main categories of data-driven models, which involves two stages: training
(developing an AI model), predicting the RUL by using the data records [4, 5].
Researchers have used different features to predict the RUL. Gebraeel et al. [6]
employed the neural network (NN) model with the application of the amplitude of
vibration in the failure frequency as input features to predict the RUL. Mahammad
et al. [7] used root mean square (RMS) and kurtosis as inputs to a feed-forward neural
network (FFNN) for RUL prediction. Wang et al. [8] investigated the effects of using
a recurrent neural network with neuro-fuzzy systems in failure time prediction of
bearing. Zhao et al. [9] offered a method based on time-frequency domain features to
predict the RUL. In this study, new features based on developing faults of ball bearings
are suggested to estimate RUL. The FFNN method has been used for prediction. By
calculating the severity of any failure mode, the sharing of each failure mode can be
considered as the input to an artificial neural network. Lastly, the outcomes of using
new and popular features have been compared.
2 Feature Selection
2.1 Rolling Bearing Failure Modes
Rolling bearings are one of the most commonly used elements in the industry. Randall
and Antoni [10] studied the advanced signal processing techniques of REBs diag-
nostic. Frequencies that appear in the case of various faults in the vibration signal
play an essential role in frequency analysis, and they can be computed from the
geometric properties and shaft speed. The equations for the different frequencies are
as follows: (“(1)” to “(3)”)

n fr d
Ball pass frequency, outer race : B P F O = 1 − cos ∅ (1)
2 D
Prognostic of Rolling Element Bearings … 335

n fr d
Ball pass frequency, inner race : B P F I = 1 + cos ∅ (2)
2 D
2
D d
Ball (roller)spin frequency : B S F = 1− cos ∅ (3)
2d D
Which n is the number of rolling elements, d is the ball diameter, D is the outer ring
diameter, ∅ is the angle of the load from the radial plane, and f r is the shaft speed. In
the envelope analysis, the signal in the resonance region is bandpass filtered, and then,
the fast Fourier transform (FFT) is taken. The main challenge is the determination
of the filter range, mainly when the fault factors are weak.
2.2 Selecting Appropriate Band Filter for Envelope Analysis
In this paper, the Shannon entropy of the filtered signal (wavelet coefficients) is
accepted as the objective function of the algorithm. As the defect frequency of inner
race (BPFI) is the largest bearing characteristic frequency (BCF), the bandwidth is
chosen as 5BPFI for all scales, and the value of center frequency f 0 is equal to the
sampling frequency f s [11]. There are many approaches to calculate the optimum
scale parameter S. To select the optimal scale range, some constraints need be taken
into account as bellows:
According to Nyquist sampling theorem for the upper cut-off frequency (lowest
bound of scale):

fs 5B P F I f s fs
+ → Smin (4)
Smin 2 2.5 0.4 f s − 2.5B P F I
In order to decrease the interfering of shaft harmonic effects, the lower cut-off
frequency is considered 25 times larger than the shaft speed:
fs 5B P F I

fs
− 25 f sha f t → Smax (5)
Smax 2 25 f sha f t + 2.5B P F I
In continuos wavelet transform (CWT), the admissibility condition implies that

the wavelet does not have a zero frequency. In other words, it can be said that the
wavelet function has zero mean. To achieve this condition, σf0 should be large enough.
If σf0 > 1.3, the compatibility equation is satisfied with an acceptable approximation
[12]. So, an additionally constrain for lower frequency is considered as
f0 fs fs
= > 1.3 → Smax < (6)
σ σ Smax 1.3σ
Therefore, to select the upper limit of the scale, both “(5)” and “(6)” must be
satisfied. So, upper-scale limit is selected as below

Smax = Min Smax , Smax (7)
As a conclusion, the range of the scale is selected as below:
(0.4 f s − 2.5B P F I )−1 f s < Si < Smax (8)
2.3 Degradation Features Extraction
The steps of the recommended approach of obtaining the features are described in
Fig. 1, which full description is as follows:
Step 1. Transform the time domain signals into the frequency domain by employing
fast Fourier transform (FFT): One cycle includes 2560 samples of vibration
signals, which is transformed into the frequency domain by applying FFT.
Step 2. Frequency-wise plot: The amplitude at a fixed frequency (e.g., Frq: BPFI)
varies at different cycles. Accordingly, amplitudes at various cycles are received
at a constant frequency, which is known as a frequency-wise plot. This method
is performed for three frequencies: BPFI, BPFO and BSF. The amplitude of the
envelop analysis in the defect frequencies of the inner ring, outer ring and balls
is named with IF, OF and BF, respectively.
Fig. 1 Degradation feature extraction

3 Experimental Data
In the PHM 2012 conference [13], a dataset of the ball bearing accelerated life
(PROGNOSTIA) was published. The PROGNOSTIA includes the 17 ball bearing
run-to-failure tests in three operating conditions. In this experiment, an electric motor
with a power of 250 W and 2830 rpm was used as a shaft supplier. A gearbox reduces
the electric motor’s speed to the required value in less than 2,000 rpm. In these tests,
both horizontal and vertical vibration of bearing housings have been monitored, and
the accelerometers have recorded the vibration signals with 25.6 kHz every 10 s. It
should be mentioned, the 0.1 s in every 10 s is considered as one cycle, and there
are 2560 samples in each cycle (Fig. 2). In this paper, the first operating condition is
chosen to examine the proposed approach. Because the applied load in this operating
condition is equivalent to the dynamic load of the ball bearing. Specifications of the
first operation conditions are presented in Table 1.
One of the standard critical features in the time domain is the RMS of the accelera-
tion signal [12]. In many cases, RMS is a useful tool for following the failure process.
Thus, the RMS trends of the seven datasets in operating condition one were plotted
in Fig. 3. As can be seen, the RMS feature of the acceleration signal in bearings 1
Fig. 2 Overview of PROGNOSTIA [13]
Table 1 Specifications of PROGNOSTIA test

OC Speed (RPM) Load (N) Number of Tests
1 1800 4000 7
Fig. 3 RMS trends for first working condition
and 3 has a three-stage pattern of degradation (regular operation, slow and fast degra-
dation). In bearings 5, 6 and 7, this feature does not appear to be a good indication
of failure from the beginning of the failure process. In these circumstances, RMS
cannot be a proper and reliable feature for estimating RUL and the defect growth
process.
4 Artificial Intelligence Modeling
In this paper, the FFNN is used to predict the RUL of the ball bearings. This network
comprises two layers. The number of neurons in the first layer is ten, which is
elected by try and error. A nonlinear sigmoid function is taken for the output of the
first layer, and the linear function is used to produce the final output of the network.
For eliminating noise in the trend of features, the smoothing algorithm is applied to
all the suggested features. In the FFNN model, 55% of the data points are defined to
train, 30% of the points are employed for validation, and the left 15% of the points
are used to test. In the prediction stage, the obtained features from the vibration data
should be smoothened with the identical procedure that was proposed for the training
data. Before referring to the results, it is suitable to understand the offered algorithm
as a flowchart. The main steps of the proposed method are shown in Fig. 4.
5 Results
In summary, the algorithm used in this paper can be explained as follows: The first step
is to remove the noise from the vibration signal, which is done by wavelet transform.
The next step is to calculate the features associated with the defect frequency of the
REB. Of course, determining the proper band for filtering the signal is very important.
Vibration Wavelet De-noised Shannon

Signal Transform Signal Entropy
Training
RUL Smoth Feature
Neural
Prediction Features Extraction
Network
Fig. 4 Flowchart of the proposed method
Fig. 5 a Frequency response—b Envelope Analysis (Signal No. 2750, Test 1-1)
According to Fig. 5a in test 1-1, the resonance band is between 500 and 3000 Hz.
Nevertheless, there is no sign of faults in frequency response when the filter is applied
within this range. However, the frequency band from 4000 Hz to 6000 Hz is most
suitable for faults to appear, which is done by the Shannon entropy method (Fig. 5b).
When the fault occurs and hits the surface, these blows can resonate in the inner
ring, outer rings, balls, sensors and other components of the system, so more than
one resonance happens. There are several resonant in the frequency response that
certainly all do not have the same intensity. In conventional methods, the resonance
region is selected for filtering.
In many cases, especially when defects are in the initial stages of growth, this
selected range involves other factors [14]. In the proposed method, all large and
strong resonances and weak resonances are investigated, and the best resonator is
selected, which is indicative of defect and less disturbing factors (noise). It should be
noted that each test includes a large number of signals which are in the time domain,
and they should be converted into the frequency domain. The signal amplitudes can
be selected at BCF. These values are plotted for all the signals, which represent the
share of each failure mode in failure. Table 2 shows the characteristics of the inner
Table 2 Bearing characteristic frequencies

OC BSF (Hz) BPFO (Hz) BPFI (Hz)
1 107,7 168,3 221,6
ring, outer ring and ball frequencies in the first operating condition.
In Fig. 6, the variation in the amplitude of the BCFs is shown in Test 1-1.
According to Fig. 6, it can be understood that all three features have an increasing
trend at the end of life. The more interesting result is that if these values are considered
as inputs of the neural network, the RUL is estimated more precisely. RMS has been
used for comparison with standard features because this feature presents a downward
trend in life expectancy and has been used in the literature as the main feature in RUL
prediction and performs better than other features, too. Additionally, information on
the input and output of the neural network in the two cases is presented in Table 3.
In Fig. 7, the life expectancy results for test 1-3 are given in two cases. The red
line shows that the real-life and blue line is a predicted life. Estimations of RUL are
shown as a percentage. As can be considered for real life, at the start of the test,
all life is left (100%), and no life is left at the end of the experiment (0%). RUL
reduction is also considered linear. It is understood that when the new features (IF,
OF, BF) are used, RUL predictions are achieved more accurately.
This process is also used for all bearings, and the error values in these two cases
are given in Table 4. The error is calculated according to “(9)”.
N
1 RU L predicted − RU L real
Error = (9)
N i=1 RU L real
As can be understood from Table 4, the results of the RUL will be more accurate if
the features that indicate the defects in the ball bearings are used rather than using the
RMS, which only indicates the overall bearings condition. When a general parameter
a) Inner Race Defect b) Outer Race Defect c) Ball Defect
Fig. 6 Process of changing the amplitude of the envelope in the failure frequency
Table 3 Neural network specifications in two cases

Case Number Input Output
1 RMS RUL
2 IF, OF, BF RUL
Fig. 7 Estimation of the RUL (Right: with first case input, Left: with second case inputs)
Table 4 Error values of the bearings tests in two cases

Predicting datasets Training datasets Error in case 1 Error in case 2
1 1–3, 1–4, 1–5, 1–6, 1–7 0.425 0.352
2 1–1, 1–4, 1–5, 1–6, 1–7 0.578 0.416
3 1–1, 1–4, 1–5, 1–6, 1–7 0.364 0.297
4 1–1, 1–3, 1–5, 1–6, 1–7 0.692 0.513
5 1–1, 1–3, 1-4, 1–6, 1–7 0.257 0.176
6 1–1, 1–3, 1–4, 1–5, 1–7 0.384 0.261
7 1–1, 1–3, 1–4, 1–5, 1–6 0.457 0.329
Average numerical testing error 0.451 0.334
like RMS is used, there is no distinction between the components and consider them
all the same.
6 Conclusion
In this paper, new features based on developing faults of ball bearings are suggested
to estimate RUL. These features represent the sharing of each failure mode. Also, the
wavelet transform is used to choose a proper frequency band for filtering the vibration
signal. The laboratory dataset of the bearing accelerated life in the PROGNOSTIA
tests is employed to approve the method.
These research results indicate that if the contribution of each failure mode is
considered as the input of the neural network, then life is estimated more precisely.
Applying this algorithm has two advantages: 1—More precise life is estimated. 2—
The failure mode of the ball bearing is known, which results in the maintenance being
taken to reduce the damage. In these situations, you should use condition-monitoring
methods that show the state of the equipment at any given time and can be used to
estimate the lifetime more accurately.
References
1. Albrecht PF et al (1987) Assessment of the reliability of motors in utility applications. IEEE

Trans Energy Conv 3:396–406
2. Crabtree CJ, Donatella Z, Tavner PJ (2014) Survey of commercially available condition
monitoring systems for wind turbines
3. Randall RB (2011) Vibration-based condition monitoring: industrial, aerospace and automotive
applications. John Wiley & Sons
4. Peng Y, Dong M, Zuo MJ (2010) Current status of machine prognostics in condition-based
maintenance: a review. Int J Adv Manuf Technol 50:297–313
5. Huang et al (2015) Support vector machine based estimation of remaining useful life: Current
research status and future trends. J Mech Sci Technol 29:151–163
6. Gebraeel N et al (2004) Residual life predictions from vibration-based degradation signals: a
neural network approach. IEEE Trans Ind Electron 51:694–700
7. Mahamad AK, Saon S, Hiyama T (2010) Predicting remaining useful life of rotating machinery
based artificial neural network. Comput Math Appl 60:1078–1087
8. Wang WQ, Golnaraghi MF, Ismail F (2004) Prognosis of machine health condition using
neuro-fuzzy systems. Mech Syst Sig Process 18:813–831
9. Zhao M, Tang B, Tan Q (2016) Bearing remaining useful life estimation based on time–
frequency representation and supervised dimensionality reduction. Measurement 86:41–55
10. Randall RB, Antoni J (2011) Rolling element bearing diagnostics—a tutorial. Mech Syst Sig
Process 25(2):485–520
11. Shi DF, Wang WJ, Qu LS (2004) Defect detection for bearings using envelope spectra of
wavelet transform. J Vib Acoust 126(4):567
12. Liao H, Zhao W, Guo H (2006) Predicting remaining useful life of an individual unit using
proportional hazards model and logistic regression model. In: Reliability and maintainability
symposium, pp 127–132
13. Nectoux P et al (2012) An experimental platform for bearings accelerated degradation tests.
In: IEEE international conference on prognostics and health management, PHM 2012, pp 1–8.
IEEE Catalog Number: CPF12PHM-CDR
14. Hosseini M, Yazdi M, Behzad B, Ghodrati A, Vahed T (2019) Experimental study of rolling
element bearing failure pattern based on vibration growth process. In: 29th european safety
and reliability conference (ESREL), Honover, Germany
Recent Technologies for Smart and Safe
Systems
Fault-Tolerant Actuation Architectures
for Unmanned Aerial Vehicles
M. A. A. Ismail, C. Bosch, S. Wiedemann, and A. Bierig
Abstract Rising civilian applications that make use of unmanned aerial vehicles
(UAVs) demand crucial precautions to minimize safety hazards. Future UAVs are
expected to incorporate fault-tolerant architectures for critical on-board systems to
ensure compliance with airworthiness certification. Reliability reports of in-service
UAVs showed that flight control actuators are among the highest root-causes of
UAVs mishaps. In this paper, the current state-of-the-art actuation architectures for
UAVs are reviewed to identify technical requirements for certification. This work is
part of a TEMA-UAV research project aimed at developing certifiable fault-Tolerant
Electro-Mechanical Actuators for future UAVs.
Keywords Airworthiness certification for UAVs · Electro-mechanical actuators ·

Fault-tolerant control
1 Introduction
Unmanned aerial vehicles (UAVs) have provided superior performance for many civil
and military applications due to their low acquisition and operating costs. However,
the reliability level of in-service UAVs is not equivalent to general aviation, according
to several reliability reports based on about 290,000 flight hours (FH) [1–3]. The US
Office of the Secretary of Defense has stated, “improving UA reliability is the single
most immediate and long-reaching need to ensure their success” [1]. UAVs currently
have a loss rate 10 times worse than manned aircraft category, signaling that the
M. A. A. Ismail (B) · A. Bierig

DLR (German Aerospace Center), Institute of Flight Systems, Lilienthalplatz 7, 38108
Brunswick, Germany
e-mail: Mohamed.Ismail@dlr.de
C. Bosch
Institute of Helicopter Technology, Technical University of Munich, Arcisstraße 21, 80333
Munich, Germany
S. Wiedemann
MACCON Elektroniksysteme GmbH, Aschauer Str. 21, 81549 Munich, Germany

https://doi.org/10.1007/978-981-15-9199-0_32
346 M. A. A. Ismail et al.
IAI fleet US army fleet
Fig. 1 Average source of system failures for in-service UAVs fleets [1]. IAI = Israeli Aircraft
Industries
root-causes of these losses should be identified and mitigated [3]. Here, the scope
is for UAVs of a takeoff-weight > 25 kg because of their certification potential [4].
System failures for two large fleets (the US army and Israeli Aircraft Industries) are
shown in Fig. 1, in which flight control systems are the second highest root-cause of
UAVs failures.
Flight control system failures (Fig. 1) have been investigated to the following
findings [1, 2]:
• Flight control systems failures are usually failures of the electro-mechanical
actuators (EMAs).
• The root-causes of EMA failures within US army fleets can be grouped into:
– EMAs designed for neither aerospace nor UAVs (e.g., Pioneer UAVs, with a
loss rate of 334 per 100,000 FH compared to 1 per 100,000 FH for general
aviation); and
– EMAs that partially involve components matching the manned aircraft category
(e.g., Predator UAVs with a loss rate of 32 per 100,000 FH).
The superior reliability level of the manned aircraft category is maintained by

certification requirements that are being extended to UAVs [4]. Recently, the Euro-
pean Aviation Safety Agency published safety regulations for UAVs considering three
UAV categories: open (takeoff-weight < 25 kg), specific, and certified. A detailed
discussion of the certification requirements for different UAVs is provided in a follow-
up paper [4]. In this paper, the scope is for flight control actuators that are directly
influenced by the expected certification conditions for UAVs:
C1. No single failure will result in a catastrophic-failure condition (evaluated by
an UAV level functional hazard analysis); and
C2. Each catastrophic-failure condition is extremely improbable.
The first condition implies a fault-tolerant control (FTC) design, as the flight
control function must not be interrupted after the first failure [5]. The second condition
relates to a quantitative probability of occurrence for possible catastrophic scenarios.
Two FTC approaches have been cited [6–9] for mitigating flight control actuator
failures: the aircraft and actuator FTCs shown in Fig. 2.
The aircraft FTC is implemented in the flight control computer (FCC), which acti-
vates modified flight control laws to compensate for actuator failures, as described in
Fault-Tolerant Actuation Architectures … 347
Aerodynamic Model
Reconfiguration
Monitoring
Multiple
Mission Flight Control Aircraft Sensors
Actuators
Single Failure
Aircraft based FTC for actuator failures
Health
Monitoring
Reconfiguration
Multiple
Actuators
Mission Flight Control Aircraft Sensors
Backup Actuator
Single Failure
Actuator based FTC for actuator failures
Fig. 2 FTCs for accommodating UAVs actuator failures
[6, 7]. In [6], a multi-variable FTC was developed for controlling a fixed-wing UAV
with a defective actuator, using only one operational control surface. In [7], a FTC
with a bank of observers was used to compensate for jammed flight control actua-
tors, but was limited to small jammed angles, making it insufficient for certification
requirements. The FTC actuation architectures are based on subsystem redundan-
cies and health monitoring functions [8, 9]. In this paper, the current state-of-the-art
FTC actuation technologies and architectures for UAVs are reviewed in Sects. 2 and
3, respectively. The technical challenges for airworthiness certification and a new
electrical fault-tolerant topology are discussed in Sect. 4.
2 Fault-Tolerant Technologies
2.1 Electrical Motors
Several electrical motor technologies have been evaluated briefly in the literature
for flight control actuators [9–14]. Among those technologies are machines without
magnets, such as induction machines (IMs), switched reluctance machines (SRMs)
and synchronous reluctance machines (SynRMs), and machines with permanent
magnets (PMs), such as surface-mounted permanent-magnet synchronous machines
(SPMSM or PMSM) and interior permanent-magnet synchronous machines

(IPMSM). IMs are known for their ruggedness, reliability, and inexpensive cost, but
they provide limited power-to-weight density [9]. SRMs are considered more widely
in fault-tolerant applications due to their multiple single-phase winding configu-
rations and their inherent capability to continue the operation after a phase fault.
However, they also exhibit comparatively high torque ripples (especially during a
phase loss), have comparatively higher acoustic noise, and a reduced power density
of approximately 30–50% compared to SPMSMs [10, 11]. SynRMs use distributed
three-phase windings, allowing synchronous rotor operation with the stator magnetic
field. Their torque ripples are comparative to the SPMSM and IPMSM while their
power density lies between the SPMSM and the IM [12] and is strongly dependent
on the rotor geometry.
PM-free machines have the advantages of no cogging- or drag torque (in case
of a short circuit failure), lower cost, easier manufacture, and greater robustness
against mechanical and temperature shocks. Due to their low back-electromotive
force, these machines lead to better inverter control failsafe operation, since the
inverter does not need to deal with uncontrollable rectification, even at high speed
[13]. PM machines, by contrast, are widely used as flight control actuators because of
their superior power-to-weight density and their relatively simple control techniques
[9, 14]. The IPMSM has, in comparison with the SPMSM, a higher reluctance torque
and, therefore, a higher power density. Nevertheless, PM machines have, besides
cogging effects (if the rotor or stator is not skewed), several fault-tolerant related
drawbacks, such as drag torque (caused by short circuit faults), electrical and thermal
phase coupling (triggered by distributed windings), magnet (hysteresis) losses, and
comparatively high back-electromotive force. PMs are also prone to change over time
as a function of temperature stresses (aging) and field weakening operation, as well as
speed limitation if the magnets are surface-mounted instead of buried (e.g., SPMSM).
The permanent-magnet-assisted synchronous reluctance machine (PMa-SynRM) is
a hybrid reluctance machine possessing PM elements partially inserted into the rotor
flux barriers for increased power density and, consequently, increased manufacturing
complexity [12]. Instead of rare-earth magnets, cheaper ferrites can be used, which
leads to decreased power density and actuator temperature range. That being said,
PMa-SynRMs enjoy the aforementioned PM-less advantages to a certain degree as
well as higher torque-to-weight, in comparison with the SynRM, and an extended
field weakening range compared to the SPMSM and IPMSM [13]. Efficiency and
power density depend strongly on rotor geometry and inserted magnetic material.
Therefore, it is scalable between the IPMSM and SynRM [15, 16]. Due to design
similarity, the reliability of the PMa-SynRM can be considered equivalent or better
than the IPMSM, as less magnet material is used. Less magnet material also grants a
lower drag torque compared to conventional PM machines in case of a short circuit
fault. Consequently, and in a fault-tolerant context, conventional PM machines should
be overrated from two to four times the nominal ratings [11] depending on low- or
high-speed operation. Yet, because of lower drag torque and magnet material, the
PMa-SynRM requires a smaller overrating, making it an interesting alternative to
conventional PM and reluctance machines by size and weight.
2.2 Electrical Power Drive
The power drive is responsible for controlling electric power and delivering it to
the motor through power transistors (e.g., MOSFETs or IGBTs [14]). The arrange-
ment of power transistors depends on the motor configuration, namely multiple
independent three-phase modules (MITPMs) and multiple independent single-phase
modules (MISPMs). MITPMs encompass one or more motors divided into sets of
three-phase star-connected modules, each powered by three-phase transistor bridges.
Rated power is supplied by n sets of three-phase modules connected in parallel. For
example, in the 3 + 3 system shown in Fig. 3 (left), two sets of three-phase modules
are redundant in one rotor [9, 17].
MISPMs encompass a single motor divided into electrically isolated n phases,
each powered by an H-bridge. Rated power is supplied by n sets of H-bridge modules.
For example, in the 2 + 1 system shown in Fig. 3 (right), three sets of isolated
single-phase modules are redundant in one rotor. The motor could be operated by 2
active phases (i.e., a faulty phase) using a modified pulse width modulation (PWM)
waveforms [14]. The isolated phases may comprise three to six phases [8, 9].
3 Fault-Tolerant Actuation Architectures
Fault-tolerant actuation architectures, A1–A6 in Fig. 4, comprise control units

(CONs), monitoring units (MONs), position sensors for the actuator deflection angle,
communication channels (COMs) to FCCs, clutches, and electric motors based on
technologies in Sect. 2.1. All of these architectures include fault-tolerant electrical
motors configured as dual motors (e.g., A1, A3, A4, and A5) or as single motors with
backup windings (e.g., A2 and A6). The mechanical power of the dual electric motors
is transferred to the control surface by using either a torque-summing gearbox (e.g.,
A3 and A4), an electric clutch (e.g., A5) or a directly controlled surface linkage (e.g.,
Fig. 3 Widely cited drive topologies for 3 + 3 (left) and for 2 + 1 (right) [9]
FCC FCC
C D Duplex Monitoring O O
FCC
O N N
FCC O WA N
Switching & G P
M C C
Isolation
FCC O WB
P I T PS
C D M M
CS
0
FCC O N N A2 A1
CS
CS A4 CS CS A4CS
Duplex P Triplex P Duplex P

CL CL
SG G
SG 3x P 3x P
WWW
P P
M M M M H H H
M M
N C
C C C C
C C V
O O O O O O O O O
FCC FCC FCC FCC FCC FCC FCC FCC FCC

A3 A4 A5 A6
P Position sensor M Motor G Gearbox H H-bridge of switches

V Voter unit C Control unit W Winding I Current sensor
SG Torque summing gearbox N Monitoring unit D Drive unit T Temperature sensor
CL Clutch O Communication CS Control surface PS Power supply
Fig. 4 Examples of fault-tolerant actuation architectures for UAVs: A1 [20], A2 [21], A3 [22], A4
[23], A5 [18], A6 [8]
A1). Sensor redundancies are present in position sensors because of their high failure
rates. A4 and A5 feature a typical duplex layout used for general purpose EMAs,
such as those in many in-service UAVs, but they are not fault-tolerant in-flight. A6
involves a voter to reliably tolerate incorrect commands from triple FCCs, despite
having only one control unit, making it insufficient to keep the actuator operational
after its first failure. A5 was developed for optionally piloted vehicles (OPV, a cate-
gory between UAVs and certified manned aircraft [18]). Further details on how A5
manages in-flight reconfiguration for optionally piloted vehicles are not available.
4 Discussion
4.1 Electrical Topology
An evaluation of electrical topologies, including PMSMs and PMa-SynRMs, is

shown in Fig. 5. The complexity is the number of power transistor (PT) switches
(e.g., MOSFETs, multiplied by 4 for a better illustration). The overratings for indi-
vidual power transistors in a SPMSM are calculated in [14], mainly comprising
additional torque contributions for a partial loss of actuator torque by an open circuit
lane (a single phase for MSPMs or a three-phase set for MTPMs) and a drag torque
by a short circuit lane at low speeds. The overratings for PMa-SynRMs are esti-
mated by assuming a reduction in drag torque only by a factor of 0.9, according to
the investigation in [19], while the overrating due to the open circuit faulty remains
unchanged. Considering minimum complexity as a vital selection criterion for UAVs,
topologies 2 + 1 and 3 + 3 are the best candidates within MISPMs and MITPMs,
respectively. In a 3 + 3 topology, the motor control provides ripple-free torque after
losing an active lane. In a 2 + 1 topology, the motor control must use asymmetric
switching waveforms to keep the motor running with only two phases [9, 14]. These
waveforms cause undesirably high levels of current harmonics, violating the total
harmonic distortion requirements set by DO-160 [9] that may be considered for
certification requirements. Therefore, a 3 + 3 topology of PMa-SynRMs has good
potential for certified fault-tolerant actuation for UAVs.
MSPMs MTPMs
Ref. case
Non-fault
tolerant
Complexity (4*PT) PM Over rating (%) PMa-SynRM Over rating (%)
Fig. 5 Evaluations for different electrical topologies considering their complexities and over ratings
4.2 Architecture Characterization for Airworthiness

Certification
The UAV actuation architectures cited in this study can be characterized into three
groups:
(1) Non-fault-tolerant actuation architectures
– Single-channel industrial actuators
– Actuator failures may be tolerated in FCCs by modified control laws.
Examples: Pioneer UAVs [2] and [6, 7]
(2) Partial fault-tolerant actuation architectures

– FCC communication channels are less than three, insufficient for tolerating
a drift (a challenging fault) or incorrect FCC commands in general [8, 14].
– On-ground or a partial in-flight reconfiguration after a single critical failure.
Examples: A1 [11], A3 [13], A4 [14], and A5 [15].
(3) Full fault-tolerant actuation architectures

– Triplex communications applied for reliable FCC commands by a voter.
– All catastrophic-failure conditions to which the actuator and its controller
are the main contributor or only source are mitigated by fail operational or
failsafe modes.
– In-flight actuator reconfiguration after a single critical failure.
Example: A2 [12], however, neither a prototype nor test setup has been reported.
This group complies with C1 (Sect. 1), but C2 requires a reliability analysis
discussed in a follow-up study [4].
5 Conclusion
The limited reliability of in-service UAVs has been significantly attributed to failures
caused by flight control EMAs, because they are neither fully fault-tolerant nor
qualified for UAVs. Recently, fault-tolerant actuators have been exploited by some
commercial UAVs, comprising mainly fault-tolerant capabilities for electric motors
and position sensors. These actuators are mostly based on PM motors to achieve a high
power-to-weight density. However, PM motors should be overrated two to four times
to compensate for expected electrical failures. Alternatively, the PMa-SynRM of 3 +
3 drive topology has been discussed here, using less drag torque and magnet material
to obtain a smaller overrating, making it an interesting alternative to conventional
PM-based EMAs by size and weight. The term “fault-tolerant actuator” has been used
frequently in the literature to describe EMAs that can partially maintain functionality
after some critical failures. A fault-tolerant actuator for a certified UAV demands
more rigorous requirements, namely: (a) all significant failures should be identified
within a specific probability of occurrence as defined by a certification body; (b) these
failures should be accommodated by fault detection and reconfiguration techniques;
and (c) these techniques should be operational online during flight phases.
Acknowledgements This work was supported by the TEMA-UAV project through the German
National Aerospace Research Program (Lufo V-3).
References
1. US Office of the Secretary of Defense (2005) Unmanned Aircraft Systems Roadmap 2005–
2030. US Office of the Secretary of Defense, United States. https://fas.org/irp/program/collect/
uav_roadmap2005.pdf. Accessed 2005
2. Israel K, Nesbit R (2004) Defense science board study on unmanned aerial vehicles and
uninhabited combat aerial vehicles. US Defense Science Board, United States
3. Sadeghzadeh I, Zhang Y (2011) A review on fault-tolerant control for unmanned aerial vehicles
(UAVs). Infotech@Aerospace. p 1472
4. Bosch C et al (2019, under publication) Towards certifiable fault-tolerant actuation architectures
for UAVs. In: WCCM 2019, Singapore
5. Lauffs JP, Holzapfel F (2016) Hardware-in-the-loop platform for development of redundant
smart actuators. Aircraft Eng Aerosp Technol: Int J 88(3):358–364
6. Venkataraman R, Seiler PJ (2016) Safe flight using one aerodynamic control surface. In: AIAA
guidance, navigation, and control conference, p 0634
7. Bateman F, Noura H, Ouladsine M (2011) Active fault diagnosis and major actuator failure
accommodation: application to a UAV. In: Advances in flight control systems, InTech 2011,
pp 137–158
8. Di Rito G, Galatolo R, Schettini F (2016) Self-monitoring electro-mechanical actuator for
medium altitude long endurance unmanned aerial vehicle flight controls. Adv Mech Eng
8(5):1687814016644576
9. Cao W et al (2011) Overview of electric motor technologies used for more electric aircraft
(MEA). IEEE Trans Ind Electron 59(9):3523–3531
10. Radun AV (1992) High-power density switched reluctance motor drive for aerospace
applications. IEEE Trans Ind Appl 28(1):113–119
11. Pickert V (2019) Fault Tolerant. ECPE, Leinfelden-Echterdingen, Germany
12. Donaghy-Spargo CM (2016) Synchronous reluctance motor technology: opportunities, chal-
lenges and future direction. Eng Technol Ref, pp 1–15
13. Wang B et al (2017) A fault-tolerant machine drive based on permanent magnet-assisted
synchronous reluctance machine. IEEE Trans Ind Appl 54(2):1349–1359
14. Bennett JW (2010) Fault tolerant electromechanical actuators for aircraft. Ph.D. thesis,
Newcastle University
15. Ibrahim M, Rashad E, Sergeant P (2017) Performance comparison of conventional synchronous
reluctance machines and PM-assisted types with combined star-delta winding. Energies
10(10):1500
16. Torres López Torres C (2018) Analysis and implementation of a methodology for optimal
PMa-SynRM design taking into account performances and reliability, Ph.D. thesis, Universitat
Politècnica de Catalunya
17. Zhu J, Niu X (2010) Investigation of short-circuit fault in a fault-tolerant brushless permanent
magnet ac motor drive with redundancy. In: 5th IEEE conference on industrial electronics and
applications, pp 1238–1242
18. Technical manuals Pegasus GmbH, 10 December 2018. www.pegasus.de
19. Ullah S et al (2016) A permanent magnet assisted switched reluctance machine for more electric
aircraft. In: XXII ICEM, pp 79–85
20. Flurry SJ (2010) KR Patent No. 100972516B1
21. Clarkstone JK, Morgan S, White N (2012) Australian Patent No. 2012235920B2
22. Hadeokju (2016) KR Patent No. 101658459
23. Technical manuals Volz GmbH, 10 December 2018. www.volz.de
Towards Certifiable Fault-Tolerant
Actuation Architectures for UAVs
C. Bosch, M. A. A. Ismail, S. Wiedemann, and M. Hajek
Abstract There is an increasing demand for integrating unmanned aerial vehicles

(UAVs) into civilian airspaces. Consequently, onboard systems should be evaluated
for possible threats to human life. This paper discusses future certification require-
ments for flight control actuators. A scheme is proposed for evaluating reliability
requirements for flight control electro-mechanical actuators (EMAs) considering
different flight control configurations. This work is part of the TEMA-UAV project,
which aims at developing certifiable fault-tolerant actuation for future UAVs.
Keywords Electro-mechanical actuators · Airworthy certification · Fault-tolerant

architectures
1 Introduction
Unmanned aerial vehicles (UAVs) are mainly used in leisure products and in mili-
tary applications. However, due to the continuous technological advancements of
electro-mechanical actuators (EMAs), power electronics and controllers [1], as well
as their availability at lower costs, a recent study [2] predicts increased demand
for UAVs in the next 30 years. Future UAV operation scenarios include not only
government authority missions, such as border security, maritime surveillance, and
military actions, but also their use for delivery purposes, medical supply, fire-fighting,
and agriculture. However, the increasing interest in medium-sized UAVs demands
C. Bosch (B) · M. Hajek

Institute of Helicopter Technology, Technical University of Munich, Arcisstraße 21, 80333
Munich, Germany
e-mail: Colin.Bosch@tum.de
M. A. A. Ismail
Institute of Flight Systems, DLR (German Aerospace Center), Lilienthalplatz 7, 38108
Brunswick, Germany
S. Wiedemann
MACCON Elektroniksysteme GmbH, Aschauer Str. 21, 81549 Munich, Germany

https://doi.org/10.1007/978-981-15-9199-0_33
356 C. Bosch et al.
new certification requirements, including their onboard systems and utilised actua-
tion architectures. This paper focuses on future flight control EMAs and proposes
a scheme to evaluate reliability requirements of such a system in consideration of
different flight control configurations.
2 UAV Safety Requirements Review
2.1 European Aviation Safety Agency Civil Rulemaking

and SC-RPAS.1309
Currently, aviation authorities focus on creating a regulatory basis for UAV safety
and certification. Resulting from this dynamic rulemaking, systems manufacturers
are often required to anticipate aspects of certification before applicable regulations
have been published. Following several review phases, the European Aviation Safety
Agency (EASA) recently published regulations (EU) 2019/947 [3] and 2019/945 [4]
describing top-level safety targets for manufacturers and operators, considering three
different UAV risk categories: open (takeoff weight < 25 kg), specific and certified.
The open category aims at low-risk operations and is not the focus of this discussion.
The other categories represent two different approaches to UAV certification:
1. Operations within the specific category demand an operational risk assessment in
which the operator assesses the mission risk or shows conformity with standard
scenarios. Due to geographical or temporal UAV operation, we regard this as a
mission-based certification [3, 4].
2. A certification similar to that needed for manned aviation might be required for the
certified category. UAVs being operated over assemblies of people and exceeding
a characteristic dimension of 3 m are subject to this category [4]. While EASA has
not published Certification Specifications for UAVs yet, the Joint Authorities for
Rulemaking on Unmanned Systems (JARUS) proposed CS-LURS (rotorcraft)
[5] and CS-LUAS (fixed-wing) [6], both providing regulations for UAVs lighter
than 750 kg. Furthermore, the Schiebel S-100 UAV, which has a takeoff weight
of 200 kg, was given special conditions based on CS-LURS as a certification
basis [7].
For type certification, aircraft development should follow a top-down approach,
which is extensively described in guidelines ARP4754 (system development) [8] and
ARP4761 (safety assessment) [9]. The first step consists of a structured investiga-
tion of potential functional failure conditions on aircraft level, also referred to as
functional hazard assessment (FHA). It requires information about criticalities and
their accepted quantitative probabilities. Both CS-LURS and CS-LUAS reference
AMC RPAS.1309. In 2015, EASA proposed Special Conditions SC-RPAS.1309-01
for UAVs lighter than 750 kg [10]. As depicted in Table 1, higher probabilities are
acceptable than in manned aviation [8]. In addition, SC-RPAS.1309 accepts a loss of
Towards Certifiable Fault-Tolerant Actuation … 357
Table 1 Acceptable UAV risk for SC-RPAS.1309 [10]

Failure condition classification NSE MIN MAJ HAZ CAT
Allowable quantitative probabilities [h−1 ] – 10−3 10−4 10−5 10−6
Design assurance level (DAL) – D C C B
NSE No safety effect, MIN Minor, MAJ Major, HAZ Hazardous, CAT Catastrophic
Table 2 Acceptable UAV risk according to STANAG 4671 [11]

Failure condition classification NSE MIN MAJ HAZ CAT
Allowable quantitative probabilities [h−1 ] – 10−3 10−4 10−5 10−6
vehicle as hazardous “where it can be reasonably expected that multiple fatalities will
not occur” [10]. The specified design assurance levels (DALs) also show a reduction
compared to manned aviation [8, 10].
2.2 Military Regulations
Military certification of UAVs often follows STANAG 4671. This standard provides
airworthiness requirements for UAVs with takeoff weights between 150 and
20,000 kg. Table 2 denotes accepted probabilities [11].
2.3 Consequences for TEMA-UAV Requirements
Sections 2.1 and 2.2 show that

– UAVs exceeding characteristic dimensions of 3 m are likely to be subject to a
structured type certification process following new EASA regulations.
– STANAG 4671 and SC-RPAS.1309 represent very similar safety and probability
requirements.
The Use Cases considered for TEMA-UAV (see Sect. 4) are in the same size range
as the Schiebel S-100 aircraft. Both the S-100 planned EASA certification efforts and
the Use Cases’ dimensions (>3 m) undermine the decision to plan for a regular type
certification. We regard this as a worst-case scenario in case the operator’s concept
of operations is not eligible for the specific category.
358 C. Bosch et al.
3 A Scheme for Evaluating Flight Control Actuation

Architectures
Certification requirements are usually described in terms of high-level safety

constraints. A transition to low-level reliability constraints is achievable by eval-
uating possible functional deficiencies for a given flight control architecture. We
propose a scheme for evaluating these requirements for flight control EMAs, as
shown in Fig. 1. The top-level stage is related to safety objectives that are defined
by certification regulations for UAVs flying in non-segregated airspace. This FHA
process, as already mentioned in Sect. 2, requires respective UAV criticality and
occurrence information (e.g. [10, 11]) on a design-independent level.
For a UAV use case, we define a flight controls layout comprising the number
of control surfaces for primary flight control. Applying safety requirements to flight
control levels requires a specific flight controls layout. For example, a total loss of
the pitch control has a catastrophic effect on UAV level. The accepted probability
of failure may increase if the flight controls layout for the pitch function consists of
duplex actuation channels. The outcomes of FHA involve qualitative and quantitative
conditions for a candidate fault tolerant control (FTC) architecture. The qualitative
conditions ensure that no single point of failure can conflict with safety-critical
Fig. 1 A scheme for evaluating flight control actuation architectures

actuation functions. The quantitative conditions define the required failure rates for
a candidate FTC architecture to be consistent with certification regulations.
For UAV design, we discuss a candidate FTC architecture fulfilling necessary
functional and operational requirements, such as flight control computers, actuators,
and fault-tolerant features. In order to point out potential failure modes for fault
detection and reconfiguration methods for a candidate FTC architecture, qualitative
conditions from the FHA are evaluated by failure modes, effects and criticality anal-
ysis (FMECA). Failures can be mitigated by FTC features, (e.g. redundancy) or by
maintenance inspections. Quantitative conditions from FHA are evaluated by fault
tree analysis (FTA) to ensure that all safety-critical failure modes have a probability
of occurrence lower than the threshold determined by the FHA. To be considered
certifiable, it could be necessary to update candidate FTC architectures by adding or
editing FTC features.
4 Use Case Analysis
Di Rito, Galatolo and Schettini performed an aircraft FHA for a fixed wing configu-
ration [12]. By extending their methodology for different aircraft configurations, we
can analyze design-specific actuation system criticalities. For that purpose, we regard
several UAV use cases to derive respective safety requirements for the system level.
Figure 2 illustrates the integration of different use cases in the ARP development
processes.
Fig. 2 Use case evaluation in accordance with ARP development and safety processes [8]
360 C. Bosch et al.
Table 3 Example of a simplified fixed-wing FHA (RMT: Remote)

FHA function Failure cond. Possible failure effect Class./Pr.
Control roll Loss of roll May result in uncontrollable flight state CAT/10−6 h−1
Erroneous roll Erratic roll inputs result in uncontrollable CAT/10−6 h−1
flight state
Partial err. roll Limited control might help fly to a safe HAZ/10−5 h−1
crash site
Control pitch Loss of pitch May result in uncontrollable flight state CAT/10−6 h−1
Erroneous pitch Erratic pitch inputs result in CAT/10−6 h−1
uncontrollable flight state
Partial err. pitch Limited control might help fly to a safe HAZ/10−5 h−1
crash site
Control yaw Loss of yaw In crosswinds, the workload on the RMT MAJ/10−4 h−1
crew increases
Erroneous yaw Erratic yaw inputs result in uncontrollable CAT/10−6 h−1
flight state
Partial err. yaw Limited yaw control may help fly to a safe HAZ/10−5 h−1
crash site
4.1 Aircraft FHA
Top-level failure conditions are dealt with in an aircraft FHA [9], in which we consider
different aircraft configurations while neglecting the specific flight control layout.
We regard three separate FHAs (fixed-wing, rotorcraft and gyrocopter). For every
function, respective failure conditions are assessed for their criticality. Table 3 depicts
a simplified example for a fixed-wing UAV.
4.2 Use Case Definition
According to the Unmanned Vehicles Handbook, fixed-wing UAVs had the biggest
share (≈65–70%) of production and development UAVs in the year 2008, followed
by rotorcraft configurations with a share of ≈20% [13]. To represent this, we examine
three fixed-wing flight control architectures and one rotorcraft application. Following
previous work by Bierig et al., a generic gyrocopter architecture is also included in
this analysis [14]. Figure 3 depicts Use Cases 1–5. The fixed wing Use Cases (1–3)
differ in the number of respective control surfaces for roll, pitch and yaw and were
derived from existing UAV examples. Use Case 4 features a classic main/tail rotor
configuration, while Use Case 5 describes a gyrocopter flight control architecture.
Fig. 3 Use Cases 1–5 (control surfaces and actuation legs are marked as dotted rectangles)
4.3 Derivation of Safety Requirements/Preliminary Aircraft

Safety Assessment (PASA)
Following the use case definition, we can derive quantitative and qualitative require-
ments. For this analysis, several assumptions are required. For Use Cases 1–3 we
propose:
– One actuation leg per control surface
– For pitch and roll: the FHA failure condition erroneous is applicable if at least 50%
of control surfaces of the respective axis have failed. If less than 50% of surfaces
have failed, we assume a partial erroneous state. The failure condition loss is
assessed as catastrophic on aircraft level, but this does not need to be discussed
here because surfaces would need to fail in a specific way (free floating), which
induces less strict reliability requirements for the system level.
– For yaw: the FHA failure condition erroneous is applicable, if more than 50% of
control surfaces have failed. If ≤50% of surfaces have failed, we assume a partial
erroneous state.
In the following analysis, we use Boolean logic to develop the actuation level
required probabilities (λACT ) necessary to meet the top-level target. Table 4 shows
an example for Use Case 2. For reasons of clarity, only most critical failures are
shown.
For Use Case 4 we assume:
– Any failure of the tail rotor actuation is considered erroneous and loss on aircraft
level.
362 C. Bosch et al.
Table 4 Preliminary aircraft safety assessment/actuation requirements derived from aircraft FHA
(Use Case 2)
FHA function FHA failure condition Failed surfaces Actuation requirement equation
Control roll Erroneous roll 2 of 4 ailerons 6 · λ2ACT ≤ 10−6 h−1
Partial erroneous roll 1 of 4 ailerons 4 · λACT ≤ 10−5 h−1
Control pitch Erroneous pitch 1 of 2 elevators 2 · λACT ≤ 10−6 h−1
Partial erroneous pitch
————— —————
Control yaw Erroneous yaw 2 of 2 rudders λ2ACT ≤ 10−6 h−1
Partial erroneous yaw 1 of 2 rudders 2 · λACT ≤ 10−5 h−1
– If at least one of the swashplate actuation legs fails, both top-level failure
conditions erroneous and loss are triggered.
For use case 5:
– If one of both actuation legs fails, both top-level failure conditions erroneous and
loss are triggered.
– Any failure of the rudder actuation leg is considered erroneous and loss on aircraft
level.
4.4 PASA Results
Figure 4 illustrates specific actuation requirements for previously defined use cases
for critical failure conditions. The strictest requirements can be found in the swash-
plate actuation of Use Case 4 and the rotor disk actuation of the gyrocopter appli-
cation, while average budgeted failure rates are slightly higher in fixed-wing appli-
cations. For Use Cases 1–3, a reduced number of control surfaces increases the
failure rate budget per surface if the same respective aircraft FHA failure condition
is triggered.
4.5 Consequences for TEMA-UAV
For the development of a prototype actuator, which is part of TEMA-UAV, we selected

the rotor disk actuation of Use Case 5 for the following reasons. Firstly, there is a
relatively low failure budget for the rotor disk actuation. Secondly, knowhow from
other DLR projects and the possibility for prototype testing at DLR are crucial [14].
A design-specific system FHA for this use case is depicted in Table 5.
Fig. 4 Failure rate requirement for one actuation leg λACT for Use Cases 1–5 (number of legs in
brackets)
Table 5 System FHA excerpt for one actuation leg of the rotor disk (Use Case 5)
Function Failure condition Failure effect Class./λACT,req
Command actuation leg Erroneous actuation Erratic roll/pitch inputs CAT/5•10−7 h−1
cannot be compensated
Loss of actuation Lost control of flight CAT/5•10−7 h−1
attitude
As part of the actuation requirements specification, the system FHA creates the
basis for the subsequent preliminary system safety assessment (PSSA) to investigate
different actuator architectures.
5 Conclusion
Although certification specifications have not been published yet, the probabilities
and DALs discussed in Sect. 2 represent the most important safety requirements
for aircraft level. For UAVs exceeding a size of 3 m, we recommend concentrating
on a full type certification. The use case analyses illustrated the strong dependency
between aircraft failure conditions and qualitative and quantitative safety require-
ments on actuation level, which depend on the system design. The analyses also
emphasized that the actuation legs of fixed-wing configurations might have a less
strict probability requirement than those of rotorcraft configurations. In a rotorcraft
or gyrocopter UAV, every described actuation leg is needed for safe continuation of
flight. In this work, actuation requirements are insensitive to specific reliability data.
364 C. Bosch et al.
However, at a later stage, knowledge about component failures will be required to

assess specific architectures, which can fulfill the budget determined in this paper.
Acknowledgements This work is supported by the TEMA-UAV project through the German
National Aerospace Research Program (Lufo V-3).
References
1. Ismail MAA, Bosch C, Wiedemann S, Windelberg J (2019, under publication) Fault tolerant
actuation architectures for UAVs. In: WCCM 2019, Singapore, 2–5 December 2019
2. European Drones Outlook Study—Unlocking the value for Europe (2016) SESAR, Brussels,
November
3. Commission Implementation Regulation (EU) 2019/947, EASA, Brussels, 24 May 2019
4. Commission Delegated Regulation (EU) 2019/945, EASA, Brussels, 12 March 2019
5. Certification Specification for Light Unmanned Rotorcraft Systems (CS-LURS), JARUS, 30
October 2013
6. Recommendations for Certification Specification for Light Unmanned Aeroplane Systrems
(CS-LUAS), JARUS, November 2016
7. Special Condition for EASA Type Certification Base – Camcopter S-100c RPAS, EASA, 09
July 2016
8. Guidelines for Development of Civil Aircraft and Systems, SAE ARP4754, November 1996
9. Guidelines and Methods for Conducting the Safety Assessment Process on Civil Airborne
Systems and Equipment, SAE ARP4761, December 1996
10. Special Condition—Equipment, Systems, and Installation, SC-RPAS.1309-01, 24 July 2015
11. Unmanned Aerial Vehicles Systems Airworthiness Requirements (USAR), STANAG 4671,
September 2009
12. Di Rito G, Galatolo R, Schettini F (2016) Self-monitoring electro-mechanical actuator for
medium altitude long endurance unmanned aerial vehicle flight controls. Adv Mech Eng 8(5):1–
11
13. Donaldson P, Lake D (2008) Unmanned aircraft – in production/unmanned aircraft—in
development. In: Unmanned vehicles handbook, United Kingdom, Shephard, pp 13–84
14. Bierig A et al (2018) Design considerations and test of the flight control actuators for a
demonstrator for an unmanned freight transportation aircraft. In: R3ASC 2018, Toulouse
Development of a Patrol System Using
Augmented Reality and 360-Degree
Camera
Kenji Osaki, Tomomi Oshima, Naoya Sakamoto, Tetsuro Aikawa,

Yuya Nishi, Tomokazu Kaneko, Makoto Hatakeyama, and Mitsuaki Yoshida
Abstract From the viewpoints of safety plant operation and improvement of utiliza-
tion rate of equipment, reduction of the human error in inspection work and efficient
maintenance engineering and inspection work are needed. Therefore, we developed a
maintenance support tool using augmented reality (AR) technology and 360-degree
camera. Furthermore, the self-localization technology using Visual Simultaneous
Localization and Mapping (Visual SLAM) technology is adopted. The inspection
work and data management which linked with localization data with Visual SLAM
technology are the feature of this system. These technologies help users identify
equipment, assist users’ recognition and judgment of the object conditions and
effective inspection data management.
Keywords Patrol system · Augmented reality · Visual SLAM · 360-degree camera
1 Introduction
From the viewpoints of safety plant operation and improvement of utilization rate of
equipment, reduction of the human error in inspection work and efficient maintenance
engineering and inspection work are needed. Therefore, we developed a patrol system
using augmented reality (AR) technology, 360-degree camera, and self-localization
technology. AR is the technology to merge a live view of the real world with computer-
generated images. AR makes it possible to display information required for inspection
on the screen of a portable device such as a smart phone or a tablet PC. A 360-degree
K. Osaki (B) · T. Oshima · N. Sakamoto · T. Aikawa

Energy Systems R&D Center, Toshiba Energy Systems & Solutions Corporation, 1 Toshiba-Cho,
Fuchu-shi, Tokyo, Japan
e-mail: kenji.osaki@toshiba.co.jp
Y. Nishi · T. Kaneko · M. Hatakeyama
Isogo Nuclear Engineering Center, Toshiba Energy Systems & Solutions Corporation, Tokyo,
Japan
M. Yoshida
Field Engineering Department, Toshiba Energy Systems & Solutions Corporation, Tokyo, Japan

https://doi.org/10.1007/978-981-15-9199-0_34
366 K. Osaki et al.
camera can photo the image of all the directions. The 360-degree camera can record
the spherical image of an inspection worker and enables the inspection image of
the direction to need. Early detection of change of an onsite situation was made
possible by difference processing with the past spherical image. Furthermore, the
self-localization technology using Visual Simultaneous Localization and Mapping
(Visual SLAM) technology is adopted for navigation to the inspection point. These
technologies help users identify equipment, assist users’ recognition and judgment
of the object conditions and effective inspection data management. In this paper, the
prototype under development is reported.
2 Self-localization Technology
We adopted self-localization technology using Visual Simultaneous Localization and

Mapping [1]. This technology estimates a self-position and creates a surrounding
map simultaneously using the information obtained by the camera and the depth
sensor of the portable device. The concept of this technology is shown in Fig. 1. The
position of the portable device in the three-dimensional coordinate is obtained by
computing distance to the characteristic points on surrounding the objects (Object 1
to Object 6) recognized through the pictures (Picture 1 to Picture 3) at every place. As
a result, prediction of the position of the portable device and recognition of objects
to be inspected can be achieved indoors, while global positioning system (GPS) is
unavailable. This technology helps inspection worker navigate to inspection point
and record inspection data combined with position information.
Fig. 1 Self-localization method using Visual Simultaneous Localization and Mapping

Development of a Patrol System Using Augmented Reality … 367
Fig. 2 Typical of AR content displayed on the smart phone
3 Support Function Using AR Technology [2, 3]
3.1 AR Technology
AR is the technology to merge a live view of the real world with computer-generated
images. A virtual content is superimposed on real image. Typical AR content
displayed on the smart phone is shown in Fig. 2. AR makes it possible to display
information required for inspection on the application screen of a portable devices
such as a smart phone or a tablet personal computer (PC). Furthermore, the self-
localization technology using Visual SLAM technology is adopted for navigation to
the inspection point.
3.2 Examples of AR Application
The support items, the function, and the applied technology are selected based on
the information of the sample cases such as inspection error and operation error in
maintenance tasks as shown in Fig. 3. They can help various inspection works.
3.2.1 Support of Reading Instruments
In inspection tasks, reading and recording of numerical values of analog instruments

at field sites are carried out. An application screen of this function is shown in
Fig. 4(1). In this sample, a photograph of the analog meters obtained beforehand
is registered as a photograph marker. Each meter is automatically recognized by
368 K. Osaki et al.
Fig. 3 Function and applied technology of the maintenance support tool
(1) Support of reading instruments (2) Support of sensor attachment
(3) Management of information using virtual tags
Fig. 4 Screen image of AR application
using the photograph marker, and frames and numbers are appended by AR to four
meters. Workers recognize with these frames and numbers that the tool has grasped
the meters to be inspected. It automatically reads values of analog meters from the
position of their needles. The value for each meter is displayed on the checklist and is
automatically saved when finally checked by the worker. This function helps improve
efficiency of the inspection tasks and reduces human errors as well.
3.2.2 Support of Sensor Attachment
Vibration measurement for pumps are conducted in periodical inspection. Sensors

for measurement usually have to be attached on the position specified according
to manuals. An application screen of this function is shown in Fig. 4(2). The tool
recognizes a pump to be inspected by reading the AR marker. The positions where
sensors should be attached are displayed as virtual objects, based on the positional
relationship to the AR marker.
3.2.3 Management of Information Using Virtual Tags
Workers occasionally have to record cautions and notices in the field tasks such as
inspection and maintenance. While these tasks usually have been controlled with tags
at the site, this tool manages them with electronic information. As shown in Fig. 4(3),
virtual tags that contain cautions and notices are attached on the object of interest and
displayed on the screen. Since they are recorded with position information, workers
are able to easily grasp their positions on a map at next inspection, leading to improve
the efficiency of confirmation task as well as the management of states of corrective
actions.
4 Patrol System Using 360-Degree Camera
4.1 360-Degree Video Browsing Function
The configuration of a patrol system using a 360-degree camera is shown in Fig. 5.

The 360-degree video and the movement trajectory measured by Visual SLAM tech-
nology are recorded. The time-synchronized 360-degree video and the movement
trajectory are displayed on a map such as a layout diagram. The 360-degree video at
the position selected on the trajectory can be viewed.
Fig. 5 Screen image of AR application

370 K. Osaki et al.
Fig. 6 Screen image of browsing function in corporate with 3D data
An example of the 360-degree image browsing system is shown in Fig. 6. The

position where the image is registered is indicated by the round mark on the layout
diagram of the equipment. Here, the 360-degree image on only the points to be
compared is extracted from the movement trajectory. The 360-degree image, laser
scan data, and 3D-CAD data at the same position can be compared and viewed.
This browsing function can be used to grasp changes in the site situation at design,
construction, and remodeling work.
4.2 Differential Processing Function Between Before

and After Images
We have developed a technology that automatically detects changes in two 360°

images for the purpose of quickly grasping changes in the field situation, such as
changes in equipment layout and discoloration due to leakage. Here, the result of
evaluating the applicability of differential processing with respect to changes in the
layout of equipment is shown.
The differential processing algorithm is shown in Fig. 7. Feature points in 360-
degree video are extracted and converted into 3D point cloud data. Next, the position
of the equipment is calculated from the information of the camera position and
orientation measured by the Visual SLAM technology and the information of the 3D
point cloud data. The position of the equipment at another time in the same place is
Fig. 7 Algorithm for differential processing
calculated from the video at another time by the method described above. A change
is detected by differentially processing the position of the equipment obtained from
the two videos.
The result of the difference processing (point cloud and voxel) is shown in Fig. 8.
The viewpoint in the figure is the position of the camera, and the part of voxel A
and B where the layout of the equipment has changed due to differential processing
is displayed in voxels. The part of voxel A and B show the moved equipment and
newly appeared equipment, respectively. The result of superimposing the difference
processing result on the 360-degree video is shown in Fig. 9. It was confirmed that
it was possible to detect changes in the layout of equipment about 30 cm within 3 m
from a camera.
5 Conclusions
• In order to improve safety and efficiency of maintenance activities and reduce

human error, we developed a patrol system using augmented reality (AR)
technology, 360-degree camera, and self-localization technology.
• A prototype system using AR technology was developed by assuming how it
would be used in field sites. Verification of the effects of the tool in field tests,
and improvement and addition of functions are conducted, and onsite application
is planned.
• A system that can view omnidirectional images of any location on a map indoors
has been developed using a 360° camera and a self-localization function using
Visual SLAM technology. In addition, it was confirmed that the change in
the layout of equipment can be detected from different images by differential
372 K. Osaki et al.
Fig. 8 Result of differential processing (point cloud and voxel)
Fig. 9 Result of differential processing (360-degree image)

processing. In the future, we will improve the accuracy of functional verification

and differential processing.
References
1. Fuentes-Pacheco J, Ruiz-Ascencio J, Rendón-Mancha JM (2012) Visual simultaneous localiza-

tion and mapping: a survey. Artif Intell Rev 43:55–81
2. Oshima T, Osaki K, Nishi Y (2019) Development of a maintenance support tool using augmented
reality technology. In: Proceedings of 27th international conference on nuclear engineering,
ICONE27-2029
3. EJAM Development of a Maintenance Support Tool Using Augmented Reality Technology.
http://www.jsm.or.jp/ejam/Vol.11No.2/NT/NT94/94.html
Research on Evaluation Method of Safety
Integrity for Safety-Related Electric
Control System of Amusement Ride
Weike Song, Peng Cheng, and Ran Liu
Abstract The safety-related electric control system (SRECS) of amusement ride

as an important system ensures the safe operation of the device. Its safety integrity
determines whether the SRECS can perform safety functions reliably. At present,
the professional safety integrity evaluation of the SRECS for amusement ride is still
blank all over the world. Based on the analysis of the structural type and move-
ment characteristics of the amusement rides, this paper develops a safety integrity
evaluation method suitable for the SRECS of amusement rides with reference to the
international function safety standards of safety-related system. First step is to iden-
tify the safety-related control function (SRCF) of SRECS combined the failure case
database, and then to do the risk assessment of SRCF considering the failure severity
and possibility, such as death, injury, and high altitude detention. The safety integrity
level requirements of ride (RSILr) can be obtained by risk matrix. The four reliability
indicators of the designed control loop are analyzed, and the safety integrity level
(RSIL) of the control loop is obtained by combining the safety integrity histogram.
Finally, the effectiveness of the safety integrity of the control system is confirmed
by tests.
Keywords Amusement ride · Safety-related electric control system (SRECS) ·

Safety-related control function (SRCF) · Safety integrity level of ride (RSIL) ·
Risk assessment
1 Background
With further development of amusement rides in experience diversification aspect, the

complexity of amusement rides is getting higher and higher, which requires amuse-
ment rides to be more accurate in the control system. By adopting more advanced
safety PLC control technology, taking advantage of a variety of sensors for data
W. Song (B) · P. Cheng · R. Liu

China Special Equipment Inspection and Research Institute,
Heping Xiyuan Building 2, Beijing, China
e-mail: songweike@csei.org.cn

https://doi.org/10.1007/978-981-15-9199-0_35
376 W. Song et al.
conversion and transmission, making use of advanced AC/DC driving technology

for power transmission, the precise control of amusement rides is achieved. At the
same time, the complexity of the system is increasing the requirements of reliability
of the rides. Especially, the important safety-related system which involves personal
safety, its reliability directly determines the operation safety of amusement rides [1].
Safety-related system refers to the related system which can ensure the safe oper-
ation of amusement rides, including safety-related mechanical system and safety-
related electric control system (SRECS). Safety-related systems are essential to
ensure the safe operation of amusement rides. Common safety-related systems of
amusement rides include safety interlocking system, emergency stop system, safety
deceleration braking system, anti-collision protection system, anti-roll-back protec-
tion system, anti-overspeed protection system, and so on. At present, the design of
amusement rides is more focused on safety-related mechanical systems, but pays
less attention to the safety-related control systems, and safety integrity evaluation of
safety functions is not carried out. The resulting accidents of amusement rides are
not unusual.
Based on the structure and movement characteristics of amusement ride, this paper
develops a safety integrity evaluation method suitable for the SRECS of amusement
rides with reference to the international function safety standards of safety-related
system. First step is to identify the safety-related control function (SRCF) of SRECS
combined the failure case database, and then to do the risk assessment of SRCF
considering the failure severity and possibility, such as death, injury, and high altitude
detention. The safety integrity level requirements of ride (RSILr) can be obtained by
risk matrix. The four reliability indicators of the designed control loop are analyzed,
and the safety integrity level (RSIL) of the control loop is obtained by combining
the safety integrity histogram. Finally, the effectiveness of the safety integrity of
the control system is confirmed by tests, which should be combined with the actual
operating conditions of amusement rides.
2 Process of Safety Integrity Evaluation
Safety integrity evaluation of amusement rides includes identification of SRECS,

risk assessment of SRCF, safety circuit design, RSIL calculation, safety integrity
evaluation, optimal design of safety circuit, and experiment confirmation, as shown in
Fig. 1. The identification of SRECS and risk assessment of SRCF are the prerequisites
for RSIL allocation. RSIL calculation and evaluation of safety circuit are parts of
RSIL verification. Therefore, RSIL allocation and RSIL verification are the main
body of safety integrity evaluation of amusement rides.
Research on Evaluation Method of Safety Integrity … 377
Fig. 1 Process of safety

Start
integrity evaluation
Identification of
SRECS
RSIL
Se
Allocation
Risk assessment of
SRCF
CI
(Fr+Pr+Av)
Safety circuit
design
Cat.
MTTFd
RSIL calculation
DCavg RSIL
Verification
CCF
Safety integrity
evaluation
NO
Meet the design
requirements
YES
Test confirmation
finish
3 RSIL Allocation
3.1 Identification of SRECS
The safety integrity evaluation of the SRECS of amusement rides should firstly
determine the evaluation object. SRECS which really affects the safety of amusement
rides often only accounts for a small part of the whole control system. Designers need
to accurately extract SRECS from the complex control system according to structure
and movement characteristics of amusement rides, design experience, and failure
database [2, 3].
Due to the diversity and novelty of amusement rides, the SRCFs implemented by
SRECS are also diverse, which makes it difficult for identification. Therefore, the
identification of SRECS depends more on the designer’s understanding of the safety
performance of the rides.
378 W. Song et al.
According to the current regulations and standards, the classification and grading
requirements for amusement rides, combined with the actual work experience, typical
SRECS and SRCF of common amusement rides are identified as shown in Table 1.
This table can be referred in the SRECS and SRCF identification process of actual
rides, but it must be increased or decreased in combination with specific equipment
features.
Table 1 Typical SRECS and SRCF of common amusement rides

No. Ride type Typical ride SRECS SRCF
1 Rotation ride Large pendulum Closing and locking Closing, locking, and ridet
detection system of startup interlock function
safety bar
Lifting platform closing Lift platform drop in place
detection system and ride startup interlock
function
Emergency stop system Emergency stop function
2 Roller coaster Launch coaster Closing and locking Closing, locking, and ridet
safety bar
Station safety door Station safety door closing
closing detection system and ride startup interlock
function
Vehicle position Vehicle position and start
detection system interlock, anti-collision
function
Vehicle position and brake
interlock, anti-collision
function
Station lifting friction Station friction wheel drop
wheel position detection in place and ride startup
system interlock function
3 Lifting tower Fly tower Closing and locking Closing, locking, and ridet
safety bar
Steel wire rope broken Steel wire rope broken and
detection system ride stop interlock function
Gondola position Prevent gondola bottoming
detection system and topping function
3.2 Risk Assessment of SRCF
The purpose of risk assessment of SRCF is to obtain the safety integrity level require-
ment (RSILr) of that function. Improper risk assessment method will lead to too high
or too low safety integrity level. Too high RSILr will cause waste and too low RSILr
will not meet safety requirements and lead to unacceptable risks. RSIL is a discrete
level (one of three possible levels) that specifies the safety integrity requirements
assigned to the SRCF of SRECS. Here, the security level 3 is the highest and the level
1 is the lowest. Due to the SRECS of amusement rides is different from the traditional
process control system of safety instruments, this paper uses the risk matrix method
based on the amusement ride failure database to evaluate and distribute RSILr. The
risk parameters are derived from the following elements:
Se: Severity that may result from danger
Fr: Frequency and duration of exposure
Pr: Probability of hazard time occurring
Av: Avoiding or limiting the possibility of injury.
For the parameter Se of amusement ride, personnel death, serious injury, minor
injury, personnel detention, economic losses, social influence, etc., should be consid-
ered. The three parameters of the probability of injury occurrence (Fr, Pr, and Av)
should be evaluated independently of each other. Every parameter should consider
the worst case. The classification of Se, Fr, Pr, and Av are as shown in Tables 2, 3,
Table 2 Classification of Se
Result Se
Death of the person/Unrecoverable and serious injury/High altitude detention time is 4
greater than 1 h/Direct economic loss more than 10,000 RMB
No death/Unrecoverable and serious injury, but the possibility of continuing to do the same 3
work after rehabilitation/High altitude detention time is greater than 1 h/Direct economic
loss more than 10,000 RMB
No death/Minor injury/High altitude detention time is less than 1 h/Direct economic loss 2
less than 10,000 RMB
No death/Scared/High altitude detention time is less than 1 h/Direct economic loss less than 1
10,000 RMB
Table 3 Classification of Fr
Frequency of exposure Fr(Duration > 10 min)
≤1d 5
>1d~≤1w 4
>1w~≤1 m 3
>1 m~≤1y 2
>1y 1
380 W. Song et al.
Table 4 Classification of Pr
Risk probability Pr
Extremely high, similar accidents have occurred 5
High, similar failures have occurred, and occur frequently in similar devices 4
Medium, similar failures have occurred, occur occasional in similar devices 3
Low, possible occurence, there is no such thing in in similar devices 2
Very low, impossible occurance can be ignored 1
4, and 5 considering the failure database.

The technical parameters above are selected from the performance behavior of
safety functions in specific amusement rides, referring to the knowledge database In
this method, RSIL is assigned by referring to RSIL allocation matrix, as shown in
Table 6, where the rows in table represent severity (Se) and the columns represent the
level of analysis above (CI) that is added by Fr, Pr, and Av. Row-column intersection
is the RSILr that should assign to SRCF.
Table 5 Classification of Av
Possibility Av
Almost not 5
Danger occurs quickly/Passengers cannot be evacuated from danger/The surrounding
environment is bad/Danger is not easily identified
Small 4
Danger occurs quickly/Passengers cannot be evacuated from danger/Good
surrounding/Danger is not easy to be identified
Medium 3
Danger occurs quickly/Passengers cannot be evacuated from danger/Good
surrounding/Danger is easily identified
Large 2
The danger occurs slowly (or delayed)/Passengers cannot be evacuated from
danger/Good surrounding/Danger is easily identified
Very large 1
The danger occurs slowly (or delayed)/Passengers can evacuate from danger/Good
surrounding/Danger is easily identified
Table 6 RSIL allocation matrix

Se CI
3–4 5–7 8–10 11–13 14–15
4 RSIL2 RSIL2 RSIL2 RSIL3 RSIL3
3 / / RSIL1 RSIL2 RSIL3
2 / / / RSIL1 RSIL2
1 / / / / RSIL1
4 SIL Verification
The safety control circuit is the carrier that implements the RSIL of the SRCF. After
the safety evaluation determines the RSILr value and completes the safety control
circuit design and extracts the SRCF, how to find the actual RSIL value of the SRCF
becomes the focus of the next step. In order to facilitate the actual evaluation work and
further refine RSIL of the SRCF, referring to the latest requirements for evaluating
the PL value of the SRP/CS in ISO 13849-1 [4]. The RSIL value of the SRCF should
be determined as following reliability parameters: Control category, mean time to
dangerous failure, average diagnostic coverage, and common cause fFailure. The
main methods of RSIL verification are reliability block diagram, fault tree, Markov
model, etc. However, no matter which method, it is based on the evaluation and
verification of the reliability of safety circuit components. Reliability index of safety
circuit is the premise of RSIL verification.
4.1 Reliability Parameters
4.1.1 Cat. (Control Category)
The reliability level relative to the fault and the ability of self-detection-warning under
the fault state of SRECS are the most important factor which determines RSIL. There
are five categories of safety circuits: B, 1, 2, 3, and 4. The higher the category level,
the higher the reliability. Control category is the basic characteristic of safety circuit.
4.1.2 MTTFd (Mean Time to Dangerous Failure)
Since the overall reliability of the system is determined by the superposition of the
reliability of different elements in the system, the average failure rate is also one of
the main factors determines the reliability of the system. Generally speaking, MTTFd
is the inverse function of failure rate, MTTFd = 1/λ, expressed in units of [year].
This value can belong to one component or can be a description of a system. For a
system, the average fault-free time of the system can be calculated by determine the
MTTFd of each part.
In the case of considering the risk caused by the number of mechanical operations,
the B10d (10% of the average number of component failures) shown in formula 1
should also be taken into account. In this case, the MTTFd value can be obtained
according to the following formulas by setting the annual average number of oper-
ations nop (Number of operations) where formula 3 is the calculation formula of
series circuit, and formula 4 is the calculation formula of parallel circuit [5].
382 W. Song et al.
B10d
M T T Fd = (1)
0.1 × n op
dop × h op × 3600
n op = (2)
tcycle
1 n
1
= (3)
M T T Fd i=1
M T T Fdi

2 1
M T T Fd = M T T Fd1 + M T T Fd1 − (4)
3 1
M T T Fd1
+ 1
M T T Fd2
Due to the strong seasonality and regularity of the operation of amusement rides,
reasonable parameters can be set in the initial design phase according to the operation
conditions and predicted usage of this type of equipment in the past. The MTTFd
values of safety PLC, safety relay, and so on are generally given by the manufacturer
after their verification, so the designer could choose the certified safety devices
preferentially when designing the circuit.
4.1.3 DCavg (Average Diagnostic Coverage)
The average diagnostic coverage is the average of self-diagnostic ability of each

SRECS. Reasonable self-diagnosis and self-monitoring measures can be used as a
complementary measure for adopting low reliability circuit architecture. This method
can effectively improve the RSIL level of the circuit. It is also in line with the current
situation of the design and manufacture of amusement rides in China. Which means
a higher integrity level can be achieved by selecting relatively cheap devices and
adopting reasonable means of self-detection measures.
4.1.4 CCF (Common Cause Failure)
Common cause failure refers to the common causes of each hardware structure
failure. ISO13849-1 defines that the CCF value is determined by the comprehensive
score of the eight factors specified and it shall not be less than 65 points, and only
SRECS with category 2 or more need CCF evaluation.
4.2 RSIL Verification
RSIL verification depends on the above four reliability parameters, but the most
important parameter to determine the overall safety performance of the system is the
control category. However, higher or lower safety performance can also be achieved
RSIL
MTTFd
low
RSIL
1 medium
high
RSIL
2
RSIL
3
Cat.B Cat.1 Cat.2 Cat.2 Cat.3 Cat.3 Cat.4
DCavg-N DCavg-N DCavg-L DCavg-M DCavg-L DCavg-M DCavg-H
Fig. 2 Determination of RSIL considering Cat., DCavg and MTTFd
by adjusting the other three parameters under the same control category. Therefore,
certain flexibility can be achieved at the design stage according to different design
input requirements. On the premise of ensuring safety performance, the manufac-
turing cost could be properly controlled. The final safety integrity level results can
be determined by Fig. 2.
5 Optimization and Test Confirmation
The purpose of safety integrity evaluation of amusement rides is to evaluate whether

the safety function meets the actual needs. If the actual RSIL value does not meet
RSILr through the above safety evaluation, it is necessary to re-optimize the safety
circuit or its safety components of that safety function. RSIL value of safety function
can be improved by improving the Cat. of safety circuit or by improving the reliability
index of safety components [5].
The experiment confirmation of safety function is to analyze and test the SRCF
of SRECS. The validation work includes application analysis and tests carried out
according to the validation plan if necessary. The application analysis should be
carried out at the same time with the design process in order to solve relatively easy
problems as soon as possible. During the test, various input combinations are applied
to the relevant safety components of SRECS in order to confirm the safety function,
and compare the corresponding output results with the prescribed output results. The
confirmation of safety function should be embodied in the operation and maintenance
manual of amusement rides, and the inspection cycle of the confirmation of safety
384 W. Song et al.
function should be clearly defined as well to keep the safety function continuously
functional.
6 Conclusions
The safety integrity evaluation method of amusement rides is applicable to SRECS

that directly related to personal safety, which fills in the blank of safety evalua-
tion aspect of amusement rides and has great practical significance. However, the
amusement rides are different from the traditional safety system of instruments in the
process industry, and there are great differences in RSIL allocation and verification.
At the same time, the diversity and complexity of amusement rides determine that
there are many difficulties in the evaluation of their safety function integrity, which
will also be the key point of future research. It mainly focuses on the following
aspects:
(1) RSIL allocation mainly depends on traditional experience and failure case
database. RSIL of safety circuit should be allocated more reasonably by
developing more effective and accurate quantitative evaluation methods.
(2) For RSIL verification, most of the control components used in domestic amuse-
ment rides do not have specific reliability indexes. In this case, how to determine
the reliability indexes of safety circuit components is particularly important.
(3) The RSIL of the safety circuit depends not only on the safety and reliability of
the hardware circuit, but also have more interfering factors worth to study such
as the software reliability and human factors.
Acknowledgements This work was supported by the national key research and development
program of China (Grant No. 2016YFF0203105).
References
1. Alizadeh S, Sriramula S (2018) Impact of common cause failure on reliability performance of

redundant safety related systems subject to process demand. Reliab Eng Syst Saf 172:129–150
2. Wu C, Hua J (2018) Research on methodology of functional safety research and development.
J Saf Sci Technol 14(8):23–28
3. Ma B, Sun ZQ (2016) Quantitative evaluation of safety integrity level based on MBF parameter
model. J East China Univ Sci Technol 42(1):97–103
4. Safety of machinery-Safety-related parts of control systems-Part1: General principles for design,

ISO13849-1:2015
5. Safety of machinery-Functional safety of safety-related electrical, electronic and programmable
electronic control systems, IEC62061:2015
Study on Warning System
of Transportable Pressure Equipment
from Regulatory Perspective
Z. R. Yang, Z. Z. Chen, Z. W. Wu, and P. Yu
Abstract Accidents on transportable pressure equipment happen occasionally and

some of them were due to weak supervision on its using and maintenance processes.
In order to reduce this kind of accidents, a warning system from regulatory perspective
is designed and studied in this paper. The warning system involves several significant
processes of the transportable pressure equipment including design, manufacture,
using, filling, and maintenance. First, the risks were identified and analyzed in all
processes. Second, the index system was established and the weight of each index was
calculated by using AHP and entropy methods. Third, the warning level classification
and the waning model were illuminated in this study. The warning system would
have a certain reference value to the safety management on transportable pressure
equipment.
Keywords Transportable pressure equipment · Supervision · Warning system ·

Index system
1 Introduction
With the high speed development of economy, the demands for energy resources
are increasing. Transportable pressure equipment (TPE) is a kind of container for
transporting LNG, liquid oxygen, liquid hydrogen, liquid nitrogen, and some other
liquid chemicals. So, the demands of this kind of special equipment also have a high
growth. The effective regulation and supervision on relative companies would do
benefit on protecting the safety and stable running of both equipment and industry.
P. Yu (B)
Faculty of Vehicle Engineering and Mechanics, Dalian University of Technology, No. 2 Linggong
Road, Ganjingzi District, Dalian, China
e-mail: more@mail.dlut.edu.cn
Z. R. Yang · Z. Z. Chen · Z. W. Wu
State Administration for Market Regulation (SAMR) Key Laboratory of Special Equipment
Safety and Energy-Saving, China Special Equipment Inspection and Research Institute (CSEI),
Beijing, China

https://doi.org/10.1007/978-981-15-9199-0_36
388 Z. R. Yang et al.
In China, there are several regulations and rules about the supervision on special
equipment, such as regulations on safety supervision of special equipment, rules
for on-site supervision and inspection of special equipment, regulations on accident
report, investigation and handling of special equipment, approval rules for special
equipment inspection institutions, and so on. Even though the development of regula-
tory system of special equipment is becoming thriving, there are still some accidents
happen due to weak supervision. According to the statistical data released by state
administration for Market Regulation in 2018, there are 6 accidents happened on
gas cylinders including three cases caused by illegal or improper operation, one
case caused by equipment defect and safety attachment failure, one case caused by
illegal business and one case caused by other reasons [1]. Sometimes, the accidents
of TPE may cause vast economical loss and fatalities. So, normalized regulation and
supervision should be enforced during the running of special equipment.
Presently, the researches of TPE focus on the design, manufacture, non-destructive
testing, and so on [2–8]. In order to safeguard the safety of TPE when using it,
warning system has been studied. However, most of the warning system did not take
regulation and supervision into consideration, and the main purpose of the warning
system is to keep the safety during its transportation [9–11]. So, in this paper, a
warning system from the regulatory perspective is proposed and studied, the risk
identification, warning index system, and warning model were studied.
2 Overall Design of Warning System
This regulatory warning system involves regulation and supervision in the processes
of TPE’s design, manufacture, using, filling and maintenance. The schematic diagram
is shown in Fig. 1. The supervision is conducted by the country, province, and city
regulatory administration to inspect the processes in design, manufacture, using and
filling companies. Then, according to the relevant regulations and rules, the assess-
ment of each processes could be obtained. The supervisors inspect and record every
Fig. 1 The schematic diagram of warning system

Study on Warning System of Transportable … 389
indexes, then the data are input into the warning system, the assessment results are
computed by using the warning model. Then, the corresponding warning level will
be give according to the results. Moreover, the warning information, like warning
level and the assessment results, will be sent to the regulatory administration.
The warning system can reflect the potential risks of the current equipment in time,
which make sure each equipment run safely and steadily. Meanwhile, all the data
collected by the system of each equipment, process and company could be valuable
references to evaluate the industry and it becomes easily to find the defects. It is also
helpful to determine the industry’s development direction.
3 Risk Identification and Index System Establishment
3.1 Risk Identification
The regulatory risk of TPE mainly happens in the processes of design, manufacture,
factory test, using, filling and maintenance. Most of the accidents are related to the
weak supervision. The warning systems identify and analyze the risks in the above
processes.
3.1.1 Risk in Design and Manufacture Process
Design is the primary process of the TPE, so the safety of it is significant. In this
process, the specification, design drawing, the results of numerical model test, and
the other references of the products should be scrutinized to avoid the design defects.
In the process of manufacture, the company or factory should have corresponding
qualifications of human, material, and financial resources. So, the main risks in the
process of design and manufacture are the design defect and unqualified products.
3.1.2 Risk in Type Test Process
Type test is a compulsory process for each batch of TPE; it is used to verify whether
the product is qualified in both shapes and performance. Meanwhile, it is also a
method to evaluate whether the design and manufacture companies are qualified in
this area. In this process, the main risks in this process is low qualification rate.
3.1.3 Risk in Using Process
There are multiple reasons cause hazards in using process. But, the risks are mainly
caused by disobeying regulation or operation rules when using TPE and employing
the unqualified operators. So, risks include delayed operations in the processes of
registration before or after use, routine inspection, maintenance, periodical inspec-
tion, and annual inspection. And the unqualified operators employed in the processes
mentioned before are also a kind of the risks.
3.1.4 Risk in Filling Process
Filling process is one of the common processes when using TPE, and it is also an
important and high-risk process because the substances in the TPE are hazardous
chemicals. Therefore, there should be specific equipment and qualified workers to
fulfill this process. The main risks in this process are irregular substances, unqualified
filling companies and operators, delayed inspection for filling equipments, and safety
attchment.
3.1.5 Risk in Periodical Inspection Process
Periodical inspection and maintenance are necessary parts for TPE. All the TPEs
should be inspected and maintained periodically and annually according to the spec-
ification, regulations, and rules. The risks in this process are delayed inspection and
maintenance.
3.2 Warning Index System Establishment
Based on the risk identification, the reasons of each risk were analyzed and the
warning index system was established. There are five first-class indexes and several
second-class indexes in the index system. The warning systems assess the safety
conditions and give warning information by using the data of each second-class
index.
3.2.1 First-Class Index
Firs-class index is the indicator of safety condition of the TPE which shows whether
it is safe to use the current TPE. According to the contents of risk identification,
the first-class index is divided into five indexes: design and manufacture safety, type
test safety, using safety, filling safety, and periodical inspection safety (by country,
province or city supervision administration). The warning index system consists of
several indexes related to design, manufacture, using, and filling companies. So, the
warning system is also a tool to evaluate the ability of every company.
3.2.2 Second-Class Index
The first-class indexes are macro indicators of the safety condition, so there should
be more specific indexes which can easily obtained by rating or calculating. In this
index system, there are several second-class indexes under the corresponding first-
class indexes. All the second-class indexes are chosen by professional experts in this
field and they are also indicators that are closely related to the safety conditions. The
index system including firs-class and second-class indexes is shown in Table 1.
3.3 Index Weight Calculation
Index system is a complex system, in order to get the synthetic score, the weight
of each index is calculated. Analytic Hierarchy Process (AHP) is a normal method
to calculate the index weight. It is usually used when there is less indexes in one
system. In this study, AHP is used to calculate the weight of first-class index. Firstly,
analyze the first-class indexes and establish the hierarchy model and establish the
judgement matrix. Secondly, obtain a set of overall level priorities for the hierarchy.
Thirdly, check the consistency of the judgement. Finally, the weight of each index
will be obtained.
As there are many second-class indexes, AHP is not applicable. So, entropy
method is applied to calculate the weight of second-class index. In this part, TPEs
from different design and manufacture companies and different product batches are
chosen as samples. Then calculate the weight of each index by using the steps of
entropy method as shown in Fig. 2.
4 Study on Warning Model
4.1 Classification of Warning Levels
The warning is divided into four levels, they are level I, level II, level III, and level
IV which means heavy, medium, light, and no warning, respectively. Different level
represent different safety conditions and corresponding solutions should be imple-
mented. In the case of level III and level IV warning, the company can use the
TPEs, but the company have to deal with the warning information shown in level III
warning within a given time. In the case of Level II warning, the TPEs cannot be
used at present, the supervision administration and the company should investigate
the potential risks and find the solutions. And after the problems have been solved,
the TPEs can be used again. And in the case of level I, the TPEs are forbidden to be
used. After investigation, if the risk is related to the quality problems of the TPEs,
the TPEs should be scrapped instantly, and on the other hand, if the risk is related to
Table 1 Indexes of warning index system

First-class index Second-class index Relevant company
Design and manufacture safety Design FTY Design company
Human resources Manufacture company
Site resources
Equipment resources
Factory test FTY
Type test safety Type test FTY
Using safety Timeliness rate of registration Use company
Timeliness rate of routine
inspection
Timeliness rate of
maintenance
Timeliness rate of periodical
inspection
Timeliness rate of annual
inspection
Timeliness rate of safety
attachment inspection
Implementation rate of safety
regulations
Qualification rate of safety
manager
Abnormal using condition
rate
Filling safety Qualification rate of filling Filling company
substances
Qualification rate of filling
parameters
Human resources
Site resources
Equipment resources
Timeliness rate of filling
equipment inspection
Qualification rate of filling
operators
Abnormal filling condition
rate
Periodical safety(country, Timeliness rate of periodical Use/manufacture company
province, city) inspection
Timeliness rate of annual
inspection
Fig. 2 The calculation steps of second-class indexes weight
the weak regulation of any companies, the companies should be fined or disqualified
in this field.
4.2 Warning Model
As all the analyses in the warning system are qualitative and there is less regulations
and rules determine the safety threshold of every index and the whole system. So, in
this warning system, the statistical method is used to determine the safety threshold.
For the indexes that have a concrete value, they are calculated by the definite formulas,
for example, FTY, timeliness rate, qualification rate, and so on. For the other indexes
that are difficult to get a concrete value, we adopt scoring system. And the score
would be given by the supervisor according to the regulations and rules.
The data of the second-class indexes are collected, recorded and scored by local,
city, province or country supervision administration, or come from the reports
submitted by the companies themselves. The TPEs with electronic label (such as
RFID) will record and update the relevant information in their own labels. And the
TPEs without an electronic label store the information in the computer with a certain
code. All the information will be uploaded to the warning system. Moreover, the
sum of every product of original second-class index value and its weight is set as a
synthetic result.
Several TPEs were used as a sample, but the equipment used in this sample should
be different than those used in Sect. 3.3. And the synthetic scores for every TPE are
obtained by using the calculation method mentioned before. Based on the statistics,
the value in the top 20% is set as level IV warning, the value in the arrange of
20%–50% is set as level III warning, the value in the arrange of 50%–80% is set as
level II warning, and the value in the last 20% is set as level I warning. The warning
information will be shown in the warning system and sent to all related authorities
and companies.
5 Conclusions
In this paper, a warning system from regulatory perspective is designed. This system
involves the regulation and supervision in almost every process of TPE, including
design, manufacture, using, filling, inspection, and maintenance process. The index
system was established based on the risk analysis, and the index system consists of
five first-class indexes: design and manufacture safety, type test safety, using safety,
filling safety, and periodical inspection safety. In order to have a better assessment,
several second-class indexes are set under the corresponding firs-class indexes. The
weight of every index were calculated by using AHP and entropy methods. Then, the
warning model was studied and the synthetic score of the safety conditions could be
obtained by using the warning model. In this system, there are four warning levels
and warning information of every TPE could be sent to the related authorities and
companies.
Acknowledgements We would like to thank the National Key R&D Program of China “Research
on Key Technologies of Risk Supervision and Management for Typical Transportable Pressure
Equipment” (Grant No.2017YFC0805605) for the financial support.
References
1. Notice of the state administration of market supervision on the safety of special equipment
nationwide in 2018, 04 April 2019. http://www.samr.gov.cn/tzsbj/dtzb/gzdt/201904/t20190
404_292592.html
2. Tschirschwitz R, Krentel D, Kluge M et al (2017) Mobile gas cylinders in fire: consequences
in case of failure. Fire Saf J 91:989–996
3. Ge S, Zhang YL, Xu XD (2004) Measurement and analysis of residue stress conditions in high
pressure gas cylinder. In: ICRS-7, Xi’an, pp 322–327
4. Yue Z, Yu F (2013) Research on materials corrosion failure model of liquefied natural gas
(LNG) cylinder for vehicle. In: ICMSE Dalian, pp 690–693
5. Pal U, Mukhopadhyay G, Bhattacharya S (2017) Failure analysis of an exploded CO2 gas
cylinder. Eng Fail Anal 79:455–463
6. Kingklang S, Daodon W, Uthaisangsuk V (2019) Failure investigation of liquefied petroleum
gas cylinder using FAD and XFEM. Int J Pressure Vessels Piping 171:69–78
7. Song BH, Lee JH, Jo YD et al (2016) Study on the safety management of toxic gas cylinder
distribution using RFID. Int J Saf Secur Eng 6(2):209–218
8. Kim JN, Jeong BY, Park MH (2017) Accident analysis of gas cylinder handling work based
on occupational injuries data. Hum Fact Ergon Manufact 27(6):280–288
9. Liu SJ, Chen ZZ, Huang QH (2018) Construction of monitoring and management platform for
intelligent connected special equipment-taking transportable pressure special equipment for
example. China Spec Equip Saf 34(8):1–8
10. Liu X, Rapik Saat M, Barkan CPL (2013) Integrated risk reduction framework to improve
railway hazardous materials transportation safety. J Hazard Mater 260:131–140
11. Ghaleh S, Omidvari M, Nassiri P et al (2019) Pattern of safety risk assessment in road fleet
transportation of hazardous materials (oil materials). Saf Sci 116:1–12
Industry 4.0: Why Machine Learning
Matters?
T. H. Gan, J. Kanfoud, H. Nedunuri, A. Amini, and G. Feng
Abstract Machine Learning is at the forefront of the Industry 4.0 revolution. Both
the amount and complexity of data generated by sensors make interpretation and
ingestion of data beyond human capability. It is impossible for engineers to optimise
an extremely complex production line and to understand which unit or machine is
responsible for quality or productivity degradation! It is extremely difficult for engi-
neers to process monitoring or inspection data as it requires a protocol, a trained
and certified engineer as well as experience! It is extremely difficult to guarantee
the quality of every single product particularly at high production rates! Artificial
Intelligence can help answering the above questions. Indeed, machine learning can
be used for predictive or condition-based maintenance. Even without knowing the
design of the machine (i.e., gearbox stages, bearing design, etc.), a machine learning
algorithm is capable to monitor deviation of monitoring sensors features compared
to a healthy state. Machine learning can be used to monitor the quality of produc-
tion by allowing the automation of the quality control process. Monitoring additive
manufacturing process to detect defects during printing and allow mitigation through
real-time machining and reprinting of the defective area. Ensuring the quality of very
complex and sensitive production processes such as MEMS, electronic wafers, solar
cells or OLED screens. Brunel Innovation Centre (BIC) is working on developing
algorithms combining statistical, signal/image processing for features extraction and
deep learning for automated defect recognition for quality control and for predic-
tive maintenance. Brunel Innovation Centre is also working on integrating those
technologies into the digital twin.
Keywords Industry 4.0 · Machine learning · Digital twin · In-line inspection
T. H. Gan (B) · J. Kanfoud · H. Nedunuri · A. Amini · G. Feng

Brunel University London, Kingston Lane, Uxbridge, Middlesex UB8 3PH, UK
e-mail: tat-hean.gan@brunel.ac.uk
T. H. Gan
TWI Ltd., Granta Park, Cambridge CB21 6AL, UK

https://doi.org/10.1007/978-981-15-9199-0_37
398 T. H. Gan et al.
1 Introduction
The term industrial revolution was first popularised by the English economic histo-
rian Arnold Toynbee to describe the process that led to United Kingdom economic
and industrial development from 1760 to 1840. Steam power and mechanised produc-
tion are among the most recognisable developments associated with the 1st industrial
revolution but the 1st revolution has various features technological, socioeconomic
and cultural. Among the technological changes, the use of new materials (iron), the
use of new energy sources (coal, steam), the invention of new machines (spinning
jenny), a novel organisation of work in the factory system, development of transporta-
tion and communication and the increasing application of science to the industry.
The 2nd industrial revolution (1870–1910) also referred to as the technological revo-
lution is considered the basis of modern life paving the way for widespread elec-
tricity, modern transport, water supply, sewage systems, etc. This has been achieved
through major advances such as petrol replacing coal as the main energy source and
the discovery of the Bessemer process for the mass-production of steel but also the
electrification, the invention of the internal combustion engine. The increase of the
size of industries resulted in complex production, maintenance and financial prob-
lems that led to significant developments in industrial engineering, manufacturing
engineering and business management. Such developments have been pioneered
by Frederick Winslow Taylor leading to the theory bearing his name [1] and Ford
[2] optimisation of factories for mass production of standardised products. This
was achieved thanks to precise and organised moving assembly lines using purpose-
built machines simplifying the production and allowing the employment of unskilled
workers. The 3rd industrial revolution (1950 to 1980) is the digital revolution and
built on the development of semiconductors, transistors and microprocessors. This
led to a progressive move from analogue to digital electronics in the industry. The
development of programmable logic controllers, PID Controllers, computers and
network communication resulted in control systems such as SCADA and allowed
a progressive automation of production and the centralisation of monitoring and
control of equipment, machines, processes and production.
Industry 4.0 or the 4th industrial revolution is a revolution in the making
targeting smart and flexible manufacturing with mass customization, better quality
and improved productivity. This paper is organised in 3 sections; the 1st presents
Industry 4.0, the 2nd explains why machine learning is needed for Industry 4.0 and
the final section presents examples of work developed by Brunel Innovation Centre
(BIC) in the context of Industry 4.0
2 Industry 4.0
Industry 4.0 is a strategic German initiative. The final report of the working group
[3]. It is easier to understand Industry 4.0 through its five main objectives:
Industry 4.0: Why Machine … 399
• production flexibility by optimising the steps in a manufacturing process in order

to reduce cost, optimise machine loads and reduce dead time
• convertible factories: that generate individualised products which requires a
modular factory that adapts near real time to customer specific and different
requirements while achieving mass and low-cost production
• customer-oriented solutions: by using sensor data to understand better the
customer environment and need and adapt the product or service accordingly
• optimised logistics: by linking customer orders to predictive algorithms assessing
consumables and material required and by optimising the supplier choice on the
delivery route to achieve both an optimised cost and respected deadlines
• use of data: data was already available thanks to the 3rd industrial revolution but
it is local and the SCADA standard widely used is very rigid and not suitable
for production flexibility and convertible factories. In the context of Industry
4.0, the production data can be used more intelligently for monitoring the process
efficiency, cost, health and reliability but also for monitoring the product during its
life opening up opportunities for new business model (service rather than product
or both)
• circular economy: the availability of data during a product life allows assessing
the life of a product and planning recycling as well.
The Industry 4.0 is an outcome of the current digital era but where the 3rd industrial
revolution was driven by digital electronics, the 4th revolution is driven by Internet.
Indeed, the connectivity is key to Industry 4.0 and by connectivity we mean the intra-
connectivity which was achieved in the 3rd revolution but lacked flexibility. More
importantly, by connectivity we mean the vertical connectivity with the supply chain
and the customers and the horizontal connectivity between production, maintenance,
logistics, research, marketing and finance. The near real-time flow of data between
different departments and different stakeholders is the key to achieving the above
objectives.
In practise, Industry 4.0 is achieved by mapping continuously the physical events
during production and service to the digital domain. Paolo et al. [4] lists the key
enabling technologies for Industry 4.0 as shown in Fig. 1.
3 Why Machine Learning Matters?
Industry 4.0 is built upon connectivity and data. The data is coming from the produc-
tion units, from the machines, from the supply chain, from the logistics, from the
intelligent and connected products, etc. The reason such amount of data is collected
is to support manufacturing, maintenance, industrial planning and business intelli-
gence. Wang et al. [5] states that manufacturing systems should be able to monitor
physical processes and make smart decisions through real-time communication with
humans, machines and sensors. The processing of the potential huge amount of data
and the smart decision making requires machine learning as humans are unable to
Industry 4.0
Internet of Things Cloud computing
Cybersecurity Big data
Digital Twin and

Autonoumous Robots simulation
Horizontal and Vertical System Augumented

Integration Reality
Additive Manufacturing
Fig. 1 Enabling technologies of Industry 4.0 [4]
process either complex or huge amount of data. Through a training process, machine
learning algorithms are capable of creating virtual models of a production line, a
factory or a process and use link the materials, power and consumables, the produc-
tion steps and time and the output no matter how complex it is. The trained models
allow optimising the production depending on customer specific requirements and
delivery times and optimise the production to minimise cost, respect deadlines while
guaranteeing quality. The machine learning algorithms can automatically predict
required consumables and materials and inform the supply chain as well as optimise
the logistics for both cost and guaranteed on-time delivery. The trained algorithms
could handle partially or totally all the above operations depending on the quality
and reliability of the data coming from different stakeholders. They can also provide
scenarios to human to make the final decisions.
Machine learning algorithms can be used to monitor—without extra-sensors—
the health of the process and the production by evaluating deviations in power or
time of production of the whole process as well as individual machines and alert on
deterioration and risk of failure. If the production process is equipped with relevant
sensors for condition monitoring of machinery or/and quality control of the product
at different stages, they can be used to monitor in real-time both the quality of the
products and the health of the machines allowing a high degree of certainty about
the productivity, the efficiency, the health and the quality of the production units.
4 Case Studies of Machine Learning in Industry 4.0
In the context of Industry 4.0, BIC work focuses on developing machine learning
algorithms for condition monitoring, structural health monitoring and in-line inspec-
tion. In this section, we review the impact of machine learning on various Industry
4.0 areas emphasising BIC developments
4.1 Machine Learning in Additive Manufacturing
Additive manufacturing is a key technology for Industry 4.0 as it allows production

flexibility and customer-oriented solutions. Defects that occur during the additive
manufacturing (AM) process can result in irreversible damage and structural failure
of the object after its manufacturing [6]. Machine learning can be implemented
in AM process to achieve high quality products. ML can be implemented as an
additional layer to the real-time and in situ non-destructive testing (NDT) evaluations.
The nature of some AM methods means that not all NDT techniques are effective
in characterising critical defects [6]. Thermography, X-ray computed tomography
(CT scan), or digital radiography are limited by the resolution of images and cost
and are not ideal for real-time monitoring [6]. Beyond defect identification and
quality control, ML could be applied in additive manufacturing for process parameter
optimisation, process monitoring and control. Collins et al. [7] reported a machine
learning model capable of predicting and estimating mechanical strength of a part
manufactured using electron beam direct manufacturing. The model was able to
identify distribution of yield strength using deep learning and genetic algorithms
(GA). For metal arc welding, Xiong et al. [8] have proposed a deep learning model
to optimise the bead geometry based on several process parameters such as feed rate,
speed of welding, arc voltage and nozzle to plate distance. BIC is developing in the
EM-ReSt project [6], electromagnetic acoustic transducers (EMAT) and eddy current
transducers (ECT) for real-time monitoring of AM components. Big data collection
and analysis will be performed and blended with statistical, machine learning and
big data analysis for the estimation of the likelihood of AM techniques to introduce
anomalies into the printed structures.
4.2 Machine Learning in Digital Twin
Digital twin is a digital representation of a physical system developed indepen-

dently and connected to the physical asset. The digital information could include
multi-physics-based physics model of the asset or sensory information. According
to Glaessgen and Stargel [9]: “the digital twin is an integrated multi-physics, multi-
scale, probabilistic simulation of a complex product and uses the best available
physical models, sensor updates, etc., to mirror the life of its corresponding twin”.
Multiple researchers have empirically and theoretically proven that machine learning
and big data technologies have improved the efficiency of production. Techniques
such as data mining, pattern recognition and deep learning have been demonstrated
[10–12]. However, actual and real-world applications and reported case studies of
ML in manufacturing are very limited [13]. ML and big data are being utilised in
manufacturing domains such as: value creation [14] and quality control [15]. For
tackling uncertainties in manufacturing, Monostori [16] reported a hybrid AI and
multi-strategy machine learning approach.
Machine learning can be implemented in all the above listed manufacturing
domains. Qingfei et al. [13] proposed a machine learning-based digital twin frame-
work as shown in Fig. 2. This framework aims to optimise production in petrochem-
ical industries. Similar models can be implemented in other industries to improve
the production efficiency (Fig. 3).
BIC is involved in collaborative research projects aiming at developing a digital
twin of a wind turbine and a bridge. WindTwin [17] aims to create a virtual copy
of a wind turbine. The digital twin of the asset aims to improve the understanding
Fig. 2 Machine learning-based digital twin framework [13]

Fig. 3 Project smart bridge concept [18]
of performance variations, anticipate wear and failures and implement condition-

based maintenance instead of planned scheduled maintenance. The project is devel-
oping modelica models of the machinery and application of machine learning for
enhancing condition monitoring algorithms and improving defect detection. Smart
bridge [19] aims to create a virtual model of a railway bridge located near London.
The virtual model developed allows monitoring the condition and degradation of the
asset combining data from IoT sensors and Artificial Intelligence algorithms consid-
ering environmental, operational and historic conditions. The project also aims devel-
oping risk-based inspection algorithms based on the monitoring data and loading of
the bridge.
5 Conclusion and Perspectives
Industry 4.0 and its enabling technologies are a very dynamic area of interest for
the industry and the academia. A lot of resources are being invested in developing
the enabling technologies: IIoT, clould computing, machine learning models and
libraries, etc. The Industry 4.0 still suffers from the limitations of any young concept
the lack of standards. Indeed, there is no communication protocol for Industry 4.0,
no security standards and not even a standard definition of Industry 4.0
According to Annalise Suzuki, Director of Technology and Engagement, at
Elysium Inc, there is no standard for machine to machine communication which
is a key to Industry 4.0 implementation. The development of standards for Industry
4.0 is particularly challenging given that this revolution aims to connect not only
the whole supply chain but the whole stakeholders which makes agreements more
challenging to achieve.
References
1. Taylor FW (1911) The principles of scientific management. Harper and Brothers, New York
2. Ford H, Crowther S (1922) My life and work, x edn. Dbouleday, New York
3. Platform Industrie 4.0, 26 June 2019. https://www.plattform-i40.de/PI40/Navigation/EN/
Home/home.html
4. Paolo C et al (2018) Product lifecycle management to support Industry 4.0. In: 15th IFIP WG
5.1 international conference, proceedings, Turin, Italy, vol 540. Springer
5. Wang S et al (2016) Towards smart factory for industry 4.0: a self-organized multi-agent system
with big data based feedback and coordination. Comput Netw 101:158–168
6. Project EM-ReSt. https://www.brunel.ac.uk/research/Projects/EM-ReSt
7. Collins PC et al (2014) Progress toward an integration of process–structure–property–perfor-
mance models for “three-dimensional (3-D) printing” of titanium alloys. Jom 66(7):1299–1309
8. Xiong J et al (2014) Bead geometry prediction for robotic GMAW-based rapid manufacturing
through a neural network and a second-order regression analysis. J Intell Manuf 25(1):157–163
9. Glaessgen E, Stargel D (2012) The digital twin paradigm for future NASA and US Air Force
vehicles. In: 53rd AIAA/ASME/ASCE/AHS/ASC structures, structural dynamics and materials
conference, Honululo, Hawai
10. Carbonell JG et al (1983) An overview of machine learning, pp 3–23. Morgan Kaufmann, New
York
11. Kateris D et al (2014) A machine learning approach for the condition monitoring of rotating
machinery. J Mech Sci Technol 28(1):61–71
12. Köksal G et al (2011) A review of data mining applications for quality improvement in
manufacturing industry. Expert Syst Appl 38(10):13448–13467
13. Min Q et al (2019) Machine learning based digital twin framework for production optimization
in petrochemical industry. Int J Inf Manag 49:502–519
14. ur Rahman MH et al (2016) Big data reduction framework for value creation in sustainable
enterprises. Int J Inf Manag 36(6):917–928
15. Tellaeche A, Ramón A (2016) Machine learning algorithms for quality control in plastic
molding industry. In: IEEE 18th conference on emerging technologies & factory automation,
Cagliari, Italy
16. Monostori L (2003) AI and machine learning techniques for managing complexity, changes
and uncertainties in manufacturing. Eng Appl Artif Intell 16(4):277–291
17. Project WindTwin. https://www.brunel.ac.uk/people/project/184220
18. Project SmartBridge. https://www.brunel.ac.uk/research/Institutes/Institute-of-Materials-and-
Manufacturing/Structural-Integrity/Brunel-Innovation-Centre/Project-summaries/SmartB
ridge
19. https://gtr.ukri.org/projects?ref=103883
Condition Monitoring
of Electromechanical Systems
A Method for Online Monitoring
of Rotor–Stator Rub in Steam Turbines
J. Liska and J. Jakl
Abstract In recent years, the operation of steam turbines has been associated with
the occurrence of rotor–stator rub. This is because the clearances in the turbine flow
paths decrease in an effort to increase the turbine efficiency. At present, detailed
rubbing diagnostics is implemented as offline analysis of measured data. The detec-
tion of the rotor–stator rub is based on analysis of vibration signals in time domain
when a machine operator monitors the overall level of vibration or the vector of
rotation frequency contained in the measured vibration signal. This paper deals with
design and implementation of the system for online monitoring of rotor–stator rub
implemented as the rub advanced monitoring system (RAMS), which, in addition to
detection phase, provides also localization of the place of contact occurrence, based
on additional knowledge of the rotor geometry. Except the contact on the shaft, the
contact between the blade tips and the stator seal can also be detected and localized.
The contact is in that case accompanied by periodic impacts whose frequency for
example corresponds to the product of the number of swirl brakes above blade tips
and the rotation frequency. With this information obtained from standard rotor vibra-
tion signals, the rotor operation can be diagnosed in more detail and, if necessary,
the turbine operation can be adapted to the current situation.
Keywords Rotor–stator rub · Full spectrum · Diagnostics · Remote monitoring
1 Introduction
Advanced analysis of rotor vibration signals provides additional information on rotor

condition and its operation. From the long-term point of view, a higher efficiency
of the turbine is achieved primarily by reducing the clearances between the rotor
and stator, and the seal parameters are designed to minimize steam energy losses.
The rotor–stator contact is subsequently a necessary result appearing during the
J. Liska (B) · J. Jakl

NTIS—European Centre of Excellence, Faculty of Applied Science, University of West Bohemia,
Univerzitni 8, 30614 Plzen, Czech Republic
e-mail: jinliska@ntis.zcu.cz

https://doi.org/10.1007/978-981-15-9199-0_38
408 J. Liska and J. Jakl
machine operation. It includes several physical phenomena, such as friction, impacts,

changes in system stiffness due to the physical connection of the rotor and stator,
etc. Especially after installation of the new rotor and seals, there is a high probability
of rub occurrence due to the small clearances in seals. In some cases, the seals are
abraded and the contact usually disappears. However, more intensive contact causes
the rotor to heat up at the point of contact which is accompanied by thermal bow
of the rotor. The condition may deteriorate rapidly and cause serious damage to the
machine.
Muszynska describes the rub problem in [1]. At present, diagnostics of the rotor–
stator contact is based on an offline analysis of vibration data. Within the turbine
operation, the overall rotor vibration and 1X phasor are monitored. As discussed by
Goldman in [2], the variation of 1X phase and amplitude within the constant rota-
tion speed may indicate bent shaft caused by an inhomogeneous thermal stress of
the rotor during the rub occurrence. This approach is an indirect method of rub
accompanied also by rotor bending. Several authors [3–5] have proposed more
or less advanced methods for rub monitoring and identification. These methods
utilize primarily signal analysis in time–frequency domain. Time–frequency signal
processing methods allow to analyze and describe the time dependency of signal
frequency content. The inputs for methods testing are usually obtained by numerical
simulations of the rotor–stator contact model (see, for example, [3]) or experimentally
using laboratory rotor stands and rigs (see [6]). Rotor–stator rub detection methods
are most often based on analysis of rotor vibration data—relative rotor or absolute
stator signals. Hall and Mba in [7] have also proposed acoustic emission for analysis
of rub signals.
2 Rotor–Stator Contact Diagnostics
As mentioned above, methods for offline rub diagnostics described in the literature
are based on the processing and representation of relative rotor vibration signals by
evaluating the development of 1X phasors over time (see Fig. 1).
In addition, Muszynska in [1] and [8] describes a method of calculating the full
spectrum that allows the determination of rotor precession. The change of rotor
precession from forward to backward may be additional indicator of rub presence,
with exception of rotor supported by anisotropic bearing during transient changes of
rotor speed. Periodical partial rub is a special type of rub, when rotor hits the stator
n times per m turns. As a symptom of the partial rub, an origin of subsynchronous
frequencies in the spectrum of analyzed signals could be detected. The issue of
online detection and early detection of rubbing has been solved by automation of
the full spectra processing (method of full cumulative spectrum) and by subsequent
standardization of the subharmonic frequency components of the spectrum. A novel
method for periodic and partial rub detection was described in [9] and [10] (Fig. 2).
Thresholding of the resulting characteristic values is used for automatic detection
of partial rub. The resulting algorithm (see [11] and [12]) has been implemented into
A Method for Online Monitoring of Rotor–Stator … 409
Fig. 1 Development of 1X phasors during contact—measuring planes on gearbox and generator
Fig. 2 Waveform of characteristic (standardized) values—HP (left) and LP (right) part of TG

200 MW
the rub advanced monitoring system (RAMS), which, in addition to detection phase,
provides also localization of the place of contact occurrence, based on additional
knowledge of the rotor geometry.
3 Localization of the Rub During Operation of the Turbine
The symptoms of the shaft rubbing were many times successfully detected by the
rub advanced monitoring system within the operation of the steam turbine. Except
the contact on the shaft, the contact between the blade tips and the stator seal can be
also localized.
In the following text, the case study of blade rubbing localization is presented.
The rub symptoms were significantly and simultaneously measured at three measure-
ment planes (bearing pedestals). Consequently, the localization of the contact origin
was performed. Firstly, the blade contact was recognized at the excited frequency
(276 Hz), which corresponds to the product of the rotation speed and number of
“swirl brakes” installed in the stator (see Fig. 3).
To identify the time of arrival (TOA) of rub induced vibration, a method comparing
signal characteristics in two consecutive sliding windows (length N) in predefined
frequency bands was used. The aim of the method is to identify signal changes. The
ratio of two consecutive windows shows the local properties. Let us consider a signal
s and its n-th sample. Then the quantity w describes the signal changes by dividing
signal parts between samples n and n + N − 1 and between samples n − N to n. The
rub impact TOA is identified as a maximum of the w-value.
n+N −1
|s(i) − s(i − 1)|
w(i) = i=n
n−1
(1)
i=n−N |s(i) − s(i − 1)|
The w-value is used to determine the TOA as an input for data processing in the
rub localization phase after its detection (Fig. 4).
The global maximum of w-value could be also guaranteed by appropriate choice
of window length N because of signal stationarity and low damping before the rub
Fig. 3 Time–frequency representation of the rub—absolute bearing vibration signal

Fig. 4 w-value (red) detects the beginning of impact arrival in selected vibration frequency band
(blue)
origin. The window length of 1 ms (i.e. N = 100) has been verified for signals with
the sampling frequency of 50 kHz. The advantage of this approach is its ability
to adapt itself at different frequencies to determine the rub event beginning in the
time–frequency domain.
In most cases, the transmission and propagation velocity of the rub caused vibra-
tion through the rotor is not known. Therefore, there is need to perform the simplifi-
cation of probable place detection with use of the linear localization based on knowl-
edge of the distance between the measurement planes. In this case, three measure-
ment planes are used for localization, and therefore it is not necessary to have prior
knowledge of the speed of wave propagation.
A frequency range of 270 Hz to 280 Hz was used for the analysis of the blade rub
in the specific case of the event described in this article. The frequency band of 1–
6 Hz is a band of the shaft rubbing—a contact in the labyrinth seals (see Fig. 3). The
time–frequency representation was calculated using the approximation of continuous
Gabor transformation. To estimate the TOA of the rub vibration to the measuring
sensor, the method based on (1) was implemented. Figure 5 shows the resulting
axial rub localization in the form of a histogram (width of column is 50 mm). The
picture also shows schematically the turbine HP part in the background and several
rotating blade stages. The places of sensor installation are highlighted and labeled
as SV1–SV3 (shaft vibration).
The highest occurrence of the contact was detected between the 4th and the 6th
stage of the HP rotor. The variance of the localization is partly depending on the
existing noise of the measuring chain and on the operational deflection shape of the
rotor during the rub occurrence. However, the turbine inspection has revealed that the
rub has occurred at multiple places with the greatest damage in the place determined
by the localization (see Fig. 6).
Fig. 5 Example of contact localization in the form of histogram—HP part of TG 200 MW
There is also need to identify the place of contact circumferentially, not only
axially. We use combination of both absolute bearing vibration and relative shaft
vibration signals to approximate the tangential localization. By combining the infor-
mation from both sensor types, the angle determination of the maximum contact
intensity could be achieved. The resulting tangential angle of contact is displayed in
the Fig. 7 (right-bottom part).
The shown ellipse matches the orbit of the rotor based on the relative rotor vibra-
tion measurement. The color of the ellipse is based on the evaluation of the amplitude
in the characteristic band of the absolute stator vibration signal (minimum = blue
color, maximum = red color). Thus, the maximum amplitude (red) determines the
angle of the contact between the rotor and stator. With this information obtained from
standard rotor vibration signals, the rotor operation can be diagnosed in more detail
with precise rub localization and, if necessary, the turbine operation can be adapted
to the current situation.
4 Conclusion
After the rub detection and localization, the HP turbine part was inspected and the
estimated rub places were examined in detail. The inspection outputs have confirmed
the results of the linear localization presented above (axial and tangential part). The
most significant abrasion of the swirl brakes was detected between the 4th and 6th
stage above the blade tips and the honey comb seal was also missing on several places
in the bottom HP stator part. In this case, visual inspection has helped to evaluate the
Fig. 6 Inspection of the stator part of 220 MW steam turbine after the rub occurrence
Fig. 7 Significant damage of the honeycomb seal (above blade tips) and tangential localization of
the blade rubbing
results of the rub detection and localization method and to increase the confidence
of turbine operators in its ability.
The RAMS monitoring system with the implemented method provides the rub
localization with centimeter accuracy. This fact significantly reduces duration of
forced outage since it is known in advance where exactly the rub occurred. Moreover,
planning of repair can be much more detailed because of the same reason. From the
steam turbine operation point of view, the developed technology serves as an early
alert system that prevents steam turbines from wreckage pretty easily caused by the
rub when not revealed on time.
Acknowledgements This work was supported from ERDF under project “Research
Cooperation for Higher Efficiency and Reliability of Blade Machines (LoStr)” No.
CZ.02.1.01/0.0/0.0/16_026/0008389 and by the project from the Ministry of Education, Youth
and Sports of the Czech Republic, PUNTIS-LO1506.
References
1. Muszynska A (2005) Rotordynamics. Taylor & Francis, London

2. Goldman P, Muszynska A, Bently DE (2000) Thermal bending of the rotor due to rotor-to-stator
rub. Int J Rotating Mach 6(2):91–100
3. Patel TH, Darpe AK (2009) Study of coast-up vibration response for rub detection. Mech Mach
Theory 44:1570–1579
4. Peng Z, He Y, Lu Q, Chu F (2003) Feature extraction of the rub-impact rotor system by means
of wavelet analysis. J Sound Vib 259(4):1000–1010
5. Lim MH, Leong MS (2013) Detection of early faults in rotating machinery based on wavelet
analysis. In: Advanced in mechanical engineering, vol 2013, 8 p
6. Abuzaid MA, Eleshaky ME, Zedan MG (2009) Effect of partial rotor-to-stator rub on shaft
vibration. J Mech Sci Technol 23:170–182
7. Hall LD, Mba D (2004) Diagnosis of continuous rotor-stator rubbing in large scale turbine
units using acoustic emissions. Ultrasonics 41:765–773
8. Goldman P, Muszynska A (1999) Application of full spectrum to rotating machinery
diagnostics. Orbit 20(1):17–21
9. Liska J, Jakl J, Cerny V (2011) The use of time-frequency methods in rotor/stator impact-
rubbing detection. In: Proceedings of Asme 2011 international design engineering technical
conferences & computers and information in engineering conference, IDETC/CIE 2011,
Washington, DC, pp 1–10
10. Liska J, Jakl J, Janecek E (2012) Steam turbine rotor/stator impact and rubbing detection. In:
Proceedings of 9th international conference on condition monitoring and machinery failure
prevention technologies, London, pp 1–12
11. Liska J, Cerny V A method of detecting and localizing partial rotor-stator rubbing during the
operation of a turbine. US patent US9903787B2
12. Liska J, Cerny V A method of detecting and localizing partial rotor-stator rubbing during the
operation of a turbine. European patent EP2746541
Implementation of Instantaneous Power
Spectrum Analysis for Diagnosis
of Three-Phase Induction Motor Faults
Under Mechanical Load Oscillations;
Case Study of Mobarakeh Steel
Company, Iran
S. Mani, M. Kafil, and H. Ahmadi
Abstract Induction motors are the most commonly used prime movers in various
industrial applications. Their vast amount of implementation necessitates a thor-
ough condition monitoring framework to assure their well-being, reduction of costs,
increasing their lifetimes and avoiding unwanted stoppages. Intensive research efforts
have been focused on motor current signature analysis (MCSA), the technique which
utilizes the spectral analysis of the stator current and aims to detect electrical and
mechanical defects. Although MCSA is one of the most common and efficient tech-
niques in condition monitoring of induction motors, several drawbacks hinder their
applicability and efficiency in certain conditions; one of the most important of which
is when the motor faces oscillating loads where fault-related sidebands in MCSA
spectrum are misunderstood with the ones due to load oscillation and hence it leads
to false indications . In this paper, a new fault detection technique focusing on active
and reactive instantaneous power signatures rather than stator current is proposed
for the diagnosis of faults in three-phase induction motors and prevention of false
indications due to overlapping fault sideband components with those pertained to
load oscillations. As a specific case study, a large 2 MW motor in steel produc-
tion process in Mobarakeh Steel Company, Iran, with a relatively oscillating load
is studied. A common fault in induction motors, a broken rotor bar is simulated for
verification of the results. Computer simulation is performed using MATLAB based
on the proposed case study and both approaches are compared in both healthy and
faulty situations, highlighting the advantages of the proposed method.
Keywords Motor Current Signature Analysis · Diagnosis · Instantaneous power ·

Broken rotor bar · Oscillating load
S. Mani (B) · M. Kafil · H. Ahmadi

Technical Inspection Office, Mobarakeh Steel Company, Isfahan, Iran
e-mail: manisina@gmail.com

https://doi.org/10.1007/978-981-15-9199-0_39
416 S. Mani et al.
1 Introduction
Condition-based maintenance (CbM) and condition monitoring (CM) denote the

current status of equipment and indicate type of maintenance required while
achieving optimal use of machine parts and guarantee breakdowns will not occur
unexpectedly. With the help of intelligent diagnostic condition monitoring system,
CbM would be able to present exact and valuable information about the condition
of a machine. Many methods have been developed to support the development of
condition monitoring system using specialized sensors [1–4].
The most prevalent type of motors utilized in industries are the induction motors
which make up 95% of industrial prime movers [5, 6]. So it can be posited that
induction motors are the most commonly used electromechanical machines in indus-
tries. During their operation, they are exposed to various loading and environmental
conditions which acting together with the motor natural aging process may lead to
numerous types of failures.
As a result and due to their highlighted role in industries, significant efforts have
been dedicated to induction machine fault diagnosis and many techniques have been
proposed to detect and diagnose incipient motor faults including both electrical and
mechanical faults.
In recent years, the use of analysis methods together with the aid of systems of
computerized data processing and acquisition has brought forth new areas in the
research of condition monitoring for induction motors. The computer and transducer
technologies combined with the advanced signal processing methods have resulted
in the ability to apply condition monitoring systems in a more effective manner with
lower maintenance costs [7, 8]. The key point is, it remains necessary during the
maintenance period that the data regarding the status of the machines be obtained
by on-line condition monitoring, and hence, the machinery failure can be effectively
reduced.
Motor current signature analysis (MCSA) is one of the most simple and popular
methods for detection of three-phase induction motors faults [9–12]. MCSA is based
on the analysis of the spectral components of one or more motor currents when the
motor is at the steady state [9, 10]. Its remote monitoring capability based on stator
current measurement available in the motor control center makes it an attractive and
convenient technology to use in an industrial setting. The advantage of MCSA is its
remote monitoring capability, simple setup and accurate results.
Although MCSA proves to be non-intrusive and economical for motor fault detec-
tion, the false indication in MCSA approach due to overlap of fault-related sidebands
with those pertained to load oscillations is unavoidable [13–25]. Thus, it is essential
to have a non-invasive on-line condition monitoring scheme which would have the
ability to detect the defects under various loading conditions. In this paper, in place
of the stator current and to avoid the shortcomings of this approach, the instanta-
neous power is proposed as a medium for motor signature analysis oriented toward
detection of motor abnormalities [26–28]. The procedure is known as instantaneous
power signature approach (IPSA) and uses line-to-line voltages and line currents to
Implementation of Instantaneous Power Spectrum Analysis for … 417
derive the desired signals [29–33]. Numerous literature like [34, 35] investigated this
phenomenon and made a comparison between MCSA and IPSA, narrating the advan-
tages of the latter over the former in such practical conditions. It is shown that the
amount of information carried by active and reactive instantaneous power signals
is higher than that deducible from the current signal alone. As mentioned earlier,
the advantage of IPSA over MCSA is more highlighted in cases where the load is
oscillating with time; thus, the lateral fault side bands in current signal are misun-
derstood with those pertained to load variations, hence leading to false indications.
IPSA approach can assist in such cases by distinguishing the two sidebands.
In order to investigate the effectiveness and feasibility of the proposed approach
in industrial applications, a comprehensive simulation is presented in this paper,
offering the possibility to model fault conditions and to investigate if their separa-
tion from load torque oscillations is possible, in both MCSA and IPSA approaches.
The simulation is performed in MATLAB environment, based on the industrial data
derived from a Stand Reversing Mill 2 MW induction motor, located at cold rolling
mill plant at Mobarakeh Steel Company, Iran.
2 IPSA in Healthy, Faulty and Load Oscillation Conditions
The outstanding advantages and probable drawbacks of MCSA have been elaborated
in the previous section. As mentioned, in cases accompanied with load variations,
MCSA has serious shortcomings, as the fault sidebands in the current spectrum are
misunderstood with the ones pertained to load oscillations and this may lead to false
indications. Instantaneous power signal spectrum has got more information in this
regard, thus can be assisting in such cases. In this section, the phenomenon of IPSA
with a focus on a healthy and a defective motor and having load variation scenarios
into account is investigated.
2.1 Case of a Healthy Motor
Let us take a healthy induction motor, powered from a balanced three-phase source
of sinusoidal voltages and driving a constant load into account.
Taking this motor as a reference, waveforms of the voltages and currents can be as
(1) and (2) [36, 37]. It is worth mentioning that no abnormalities in stator windings
are assumed in further considerations.
√
va (t) = 2V cos ωt (1)
√
i a (t) = 2I cos(ωt − ϕ) (2)
418 S. Mani et al.
where V and I are the stator voltage and current RMS values, respectively, and ϕ is
the phase angle. The voltage and currents of phase b and c can be reproduced using
a 2π /3 of phase difference.
Three-phase instantaneous active and reactive powers p3ph (t) and q3ph (t) are
calculated via (3) and (4).
p3ph (t) = (va × i a + vb × i b + vc × i c ) = 3V I cos ϕ (3)
√
q3ph (t) = (2/ 3)((vb − vc ) × i a + (vc − va ) × i b + (va − vb ) × i c ) = 3V I sin ϕ
(4)
2.2 Case of a Defective Motor
Let us now consider presence of a fault in the rotor which modulates the current
spectrum. As mentioned before, occurrence of a rotor cage fault or mechanical abnor-
mality is characterized by appearance of a sequence of sideband components around
the fundamental, at frequencies f s ± kf d , in the stator current spectrum (with f d =
f b , f r or f 0 for the case of broken rotor bars, air-gap eccentricity faults, or load torque
oscillations, respectively). Broken rotor bar is a relatively common fault in induction
motors and accounts for a large percentage of the total induction motor failures [9,
10]. It is taken into account as the motor fault in this study and simulation.
For simplicity, it is considered that the failure causes only an amplitude modulation
on the stream of the stator. It could also prove that phase modulations appear in current
as amplitude modulations by processed result from similar functions to the Bessel
functions, an integral function in mathematically modeling the induction motor,
which are sued in the simulation section in this study. The modulated current i a, f
can be expressed by:
∞
√
i a, f = i a (t) + 2I1,k cos ω − kω f t − ϕ1,k
k=1
√
+ 2I2,k cos ω + kω f t − ϕ2,k (5)
where i a (t) is the healthy-mode current of phase a and ωf is the angular frequency of
the failure. So the instantaneous active and reactive power signals can be calculated
by
∞

p f (t) = 3V I cos ϕ + 3V I1,k cos kω f t + ϕ1,k
k=1

−I2,k cos kω f t − ϕ2,k (6)
∞

q f (t) = 3V I sin ϕ + 3V I1,k sin kω f t + ϕ1,k
k=1

−I2,k sin kω f t − ϕ2,k (7)
Thus with some mathematical simplifications, the amplitudes of the faulty compo-
nents produced in both active and reactive spectrums are derived via (8) and
(9).

P f,k = 3V I1,k + I2,k + 2I1,k I2,k cos ϕ1,k + ϕ2,k (8)

Q f,k = 3V I1,k + I2,k − 2I1,k I2,k cos ϕ1,k + ϕ2,k (9)
Based on ϕ1,k and ϕ2,k phase angles in two rotor fault and load oscillation
scenarios, the sideband frequency is visible in one of power spectrums, as it is in the
current-related one. So the superiority of this approach is that the load oscillation can
be distinguished from the rotor faults via this method, as the oscillation frequency
component is absent in the reactive power spectrum and the rotor fault frequency
component is absent in active power frequency spectrum. So these two spectrums
can come in hand for discrimination of the two faulty frequency components.
2.3 Case of a Varying Mechanical Load
Abnormal mechanical load conditions often significantly influence the characteristic

signatures in the motor current spectrum [38, 39]. Mechanical load abnormalities
can be classified in two main types. The first type is mechanical faults, such as shaft
misalignments, that essentially create a rotor eccentricity inside the motor. So the fault
indicator frequencies directly identify their detection. The second type of mechanical
load anomalies such as the ones occurring in reciprocating compressors may also
introduce the very same sideband harmonics in the stator current spectrum while
the rotor shaft is not at all influenced. In other words, the second-type mechanical
load anomalies do not produce any rotor faults inside the motor itself [39, 40]. This
can lead to a major drawback of MCSA approach; as by analyzing the sideband
harmonics in a stator current spectrum, the rotor fault condition can be confused
with these mechanical load anomalies. This is the main issue in this study, where
a false indication is produced when a low-frequency oscillating load is confused
with a rotor fault in MCSA spectrum. The approach presented in this study aims to
compensate this specific drawback.
420 S. Mani et al.
Fig. 1 The SRM motor at MSC
3 Introducing the Case Study
Mobarakeh Steel Company (MSC), the largest steel producer in MENA region, incor-
porates numerous electric motors, most significantly electric induction motors with
various working scenarios, environmental conditions, ages and power ratings. The
importance of CbM and CM techniques in MSC, specifically on induction motors,
has urged engineers and operators to focus on novel approaches to run a vivid and
comprehensive condition monitoring (CM) framework for the motors in order to
modify the present structure mostly based on MCSA approach and tackle the prob-
able shortcomings. Several applications at MSC face oscillating loads; all of which
highlight the drawbacks of MCSA in those industrial cases. IPSA is the solution for
MSC case. The data derived from a Stand Reversing Mill (2 MW SRM motor) at
Cold Rolling Plant at MSC, shown in Fig. 1, is used in this paper to investigate the
effectiveness of the proposed approach. Table 1 presents the characteristics pertained
to the introduced case study.
4 Computer Simulation and Results
In this section, mathematical model of induction motors based on the direct-

quadrature (d-q) framework of the machine is simulated in MATLAB Simulink
environment. Figure 2 depicts a brief view of the Simulink model performed in
this study.
Table 1 The SRM motor

Parameter Value Unit
characteristics
Rated power 2 MW
Power factor 0.87 –
Rated frequency 50 Hz
Rated voltage 1100 V
Number of poles 4 –
Stator Rs 0.0061
Rotor Rr 0.0067
Stator Xls 0.1071
Rotor Xlr 0.0593
Fig. 2 The schematic MATLAB Simulink model
In order to investigate the false indications produced by presence of oscillating

torques in an induction motor, a healthy motor in oscillating torque condition is
elaborated in the first stage. The time varying oscillating torque pertained to the
SRM motor introduced in the previous section is exerted on the simulated model.
Figure 3 shows the rotor shaft speed while Figure 4 represents the shaft torque of the
motor.
All three spectrums of current, active and reactive powers are illustrated in Figs. 5,
6, 7, 8, 9 and 10, in two scenarios of healthy and faulty motor both with oscillating
loads. In the steady state working condition of the motor, stator current of the motor
driving an oscillating load is illustrated in frequency domain in Fig. 5. It is shown
422 S. Mani et al.
Fig. 3 The rotor shaft speed (m/s) versus time
Fig. 4 The rotor shaft torque (N m) versus time
0.1
0.09
0.08
0.07
0.06
Ias(db)
0.05
X: 56.53
0.04 Y: 0.03282
0.03
0.02
0.01
40 45 50 55 60
f
Fig. 5 Stator current spectrum of the healthy motor in time varying load condition
0.09
0.08
0.07
Phealthy(db) 0.06
0.05 X: 6.528
Y: 0.0405
0.04
0.03
0.02
0.01
0
2 4 6 8 10 12 14
f
Fig. 6 Frequency spectrum of instantaneous active power for the healthy motor in time varying
load condition
0.1
0.09
0.08
0.07
Qhealthy(db)
0.06
0.05
0.04
0.03
X: 2.083
0.02 Y: 0.01352
0.01
0
2 4 6 8 10 12 14
f
Fig. 7 Frequency spectrum of instantaneous reactive power for the healthy motor in time varying
load condition
that only the electrical working frequency of the motor, 50 Hz, is visible in the stator
current spectrum together with the lateral sidebands pertained to the time varying
torque. The spectrums of instantaneous active and reactive powers are also depicted
in Figs. 6 and 7. It is worth mentioning that all values in the figures in this subsection
are normalized based on the maximum values in that specific spectrum. It is shown
that a sideband at 6.28 Hz is present in both current and active power spectrums in
both healthy and faulty conditions with similar values, indicating that load oscillation
components are present in both signals. This is different in case of reactive power
424 S. Mani et al.
0.1
0.09
0.08
0.07
Iastotf(db) 0.06
0.05
X: 56.53
0.04 Y: 0.03584
0.03
0.02
0.01
40 45 50 55 60
f
Fig. 8 Stator current spectrum of the faulty motor in time varying load condition
0.1
0.09
0.08
0.07
Pfaulty(db)
0.06
0.05 X: 6.528
Y: 0.04035
0.04
0.03
0.02
0.01
0 2 4 6 8 10 12 14
f
Fig. 9 Frequency spectrum of instantaneous active power for the faulty motor in time varying load
condition
spectrum where the 2.083 Hz component is pertained to the rotor fault and not to the
load oscillation.
In the second stage, a defective motor is simulated with a broken rotor bar defect.
The fault frequency components pertained to this defect are introduced in (10).
f br b = (1 ± 2ks). f s (10)
where f s denotes the line frequency, s represents the motor slip and k is an integer.
0.1
0.09
0.08
0.07
Qfaulty(db) 0.06
0.05
0.04 X: 2.083
Y: 0.04308
0.03
0.02
0.01
2 4 6 8 10 12 14
f
Fig. 10 Frequency spectrum of instantaneous reactive power for the faulty motor in time varying
load condition
Having the defective simulated motor, Figure 8 illustrates the stator current spec-
trum of a problematic motor in oscillating load condition. The f br b frequencies
pertained to the simulated fault can be seen in this figure. Also, the instantaneous
active and reactive power in frequency domain are represented in Figs. 9 and 10.
It is shown that f f component is present in the current spectrum. In other words,
the faulty component is transmitted to the current signal; so in case of simultaneous
load oscillation and rotor faults, the current signal cannot assist in distinguishing the
two frequency sidebands. Besides, although the current spectrum sidebands in the
two mentioned conditions have different values, presence of noises of other types
of electromagnetic disturbances may alter the amplitudes of the sidebands which
indeed boosts the false indications, as the f f component is present in the current
spectrum in both cases of faulty rotor and oscillating load.
On the other hand, since the load oscillation frequency components are absent in
the instantaneous reactive power spectrum and rotor fault-related frequency compo-
nents f f are absent in active power spectrum, the fault can be distinguished in the
motor via this approach. So in case of unclear results, both instantaneous power
spectrums can be checked, and if the specific sideband appears in this spectrum, it
can be determined if it is related to the oscillating load or the rotor fault.
According to the formulation and the simulation presented in this study, it is
vividly shown that the instantaneous power spectrum method has an advantageous
features compared to MCSA. This is indeed pertained to the fact that both fault and
load oscillation-related sidebands are present in stator current which makes discrim-
ination of these two cases difficult in such simultaneous occurrences. The advantage
of instantaneous active and reactive power spectrums in this case is that the fault-
related frequency can be fully distinguished which assists in diagnosis of the defect
in (prevalent) simultaneous problems. This study incorporating a strong computer
simulation with technical data pertained to an industrial case clearly highlights the
426 S. Mani et al.
advantage of instantaneous power spectrum analysis in such specific cases, which

are in fact quite prevalent in most large industries like MSC, studied as the case study
in this research.
5 Conclusion
This study assesses motor current signature analysis (MCSA), the most prevalent
fault detection and diagnosis method for induction motors and its drawbacks in
facing with time varying loading conditions are investigated. It is shown that MCSA
has serious shortcomings due to false indication resulted from presence of rotor faults
and load oscillations at similar frequencies. In order to prevent such false indications
in MCSA spectrum, a new approach taking both current and voltage signals into
account is studied. Instantaneous power signature analysis (IPSA) is the solution to
the cases where the load is varied with time. A MATLAB Simulink simulation is
performed in this study to compare both approaches for a 2 MW Stand Reversing
Mill (SRM) motor case study at Mobarakeh Steel Company (MSC) in Iran, focusing
on its actual data, incorporating low-frequency load torque oscillations. The research
investigates advantages and disadvantages of the understudy methods and highlights
superiorities of the presented approach over the more prevalent one in industries.
Based on the derived results, it is proved that IPSA is more effective than MCSA in
discrimination of load varying conditions with rotor faults in induction motors, since
it delivers delicate result for fault detection, diagnosis and fault.
Acknowledgements The authors would like to deeply thank the Rotary Machinery and Power
Equipment Technical Inspection Office and Cold Rolling Mill Plant at Mobarakeh Steel Company,
Isfahan, Iran, for provision of information, technical infrastructure and access to the required data.
References
1. Li B, Chow MY, Tipsuwan Y, Hung JC (2000) Neural-network-based motor rolling bearing

fault diagnosis. IEEE Trans Ind Electron 47:1060–1069
2. Tavner PJ, Ran L, Pennman J, Sedding H (2008) Condition monitoring of rotating electrical
machines. Research Studies Press Ltd., Letchworth, England
3. Okoro IO (2005) Steady and transient states thermal analysis of a 7.5-kW squirrel-cage
induction machine at rated-load operation. IEEE Trans Energy Convers 20:730–736
4. Braun JM, Brown G (1990) Operational performance of generator condition monitors:
comparison of heated and unheated ion-chambers. IEEE Trans Energy Convers 5:344–349
5. Thomson WT, Fenger M (2001) Current signature analysis to detect induction motor faults.
IEEE Ind Appl Mag 7:26–34
6. Intesar A (2008) Investigation of single and multiple faults under varying load conditions using
multiple sensor types to improve condition monitoring of induction machines
7. Choi S, Akin B, Rahimian MM, Toliyat HA (2001) Implementation of fault diagnosis algorithm
for induction machines based on advanced digital signal processing techniques. IEEE Trans
Ind Electron 58:937–948
8. Siyambalapitiya DJT, McLaren PG (1990) Reliability improvement and economic benefits of

on-line monitoring system for large induction machines. IEEE Trans Ind Appl 26:1018–1025
9. Nandi S, Toliyat HA, Li X (2005) Condition monitoring and fault diagnosis of electrical
motors—a review. IEEE Trans Energy Convers 20:719–729
10. Benbouzid MEH (2000) A review of induction motors signature analysis as medium for fault
detection. IEEE Trans Ind Electron 47:984–993
11. Tandon N, Yadava GS, Ramakrishna KM (2007) A comparison of some condition monitoring
techniques for the detection of defect in induction motor ball bearing. Mech Syst Signal Process
21:244–256
12. Tavner PJ (2008) Review of condition monitoring of rotating electrical machines. IET Electr
Power Appl 2:215–247
13. Stack JR, Habetler TG, Harley RG (2004) Fault classification and fault signature production
for rolling element bearings in electric machines. IEEE Trans Ind Appl 40:735–739
14. Sun Q, Tang Y (2002) Singularity analysis using continuous wavelet transform for bearing
fault diagnosis. Mech Syst Signal Process 16:1025–1041
15. Schoen RR, Habetler TG (1997) Evaluation and implementation of a system to eliminate
arbitrary load effects in current-based monitoring of induction machines. IEEE Trans Ind Appl
33:1571–1577
16. Frosini L, Bassi E (2010) Stator current and motor efficiency as indicators for different types
of bearing faults in induction motors. IEEE Trans Ind Electron 57:244–251
17. Blodt M, Granjon P, Raison B, Rostaing G (2008) Models for bearing damage detection in
induction motors using stator current monitoring. IEEE Trans Ind Electron 1:383–388
18. Jung JH, Lee JJ, Kwon BH (2006) Online diagnosis of induction motor using MCSA. IEEE
Trans Ind Appl 53:1842–1852
19. Stack JR, Habetler TG, Harley RG (2005) Experimentally generating faults in rolling element
bearing via shaft current. IEEE Trans Ind Appl 41:25–29
20. Randall RB, Antoni J (2011) Rolling element bearing diagnostics—a tutorial. Mech Syst Signal
Process 25:485–520
21. Jardine AKS, Lin D, Banjevic D (2006) A review on machinery diagnostics and prognostics
implementing condition-based maintenance. Mech Syst Signal Process 20:1483–1510
22. Bellini A et al (2002) On-field experience with on-line diagnosis of large induction motors
cage failures using MCSA. IEEE Trans Ind Appl 38(4):1045–1053
23. Lee S et al (2013) Evaluation of the influence of rotor axial air duct design on condition
monitoring of induction motors. IEEE Trans Ind Appl 49(5):2024–2033
24. Culbert IM, Rhodes W (2007) Using current signature analysis technology to reliably detect
cage winding defects in squirrel-cage induction motors. IEEE Trans Ind Appl 43(2):422–428
25. Leith D, Rankin D (1991) Real time expert system for identifying rotor faults and mechanical
influences in induction motor phase current. In: Proceeding of fifth international conference
on electric machines and drives IET, London, UK, 11–13 Sept 1991, pp. 46–50
26. Bellini A, Filippetti F, Franceschini G, Tassoni C, Kliman GB (2001) Quantitative evaluation
of induction motor broken bars by means of electrical signature analysis. IEEE Trans Ind Appl
37(5):1248–1255
27. Liu Z, Yin X, Zhang Z, Chen D, Chen W (2004) Online rotor mixed fault diagnosis way based
on spectrum analysis of instantaneous power in squirrel cage induction motors. IEEE Trans
Energy Convers 19(3):485–490
28. Didier G, Ternisien E, Caspary O, Razik H (2006) Fault detection of broken rotor bars in
induction motor using a global fault index. IEEE Trans Ind Appl 42(1):79–88
29. Faiz J, Ojaghi M (2009) Instantaneous-power harmonics as indexes for mixed eccentricity
fault in mains-fed and open/closed-loop drive connected squirrel-cage induction motors. IEEE
Trans Ind Electron 56(11):4718–4726
30. Legowski SF, Sadrul Ula AHM, Trzynadlowski AM (1996) Instantaneous power as a medium
for the signature analysis of induction motors. IEEE Trans Ind Appl 32(4):904–909
31. Benbouzid MEH, Kliman GB (2003) What stator current processing-based technique to use
for induction motor rotor faults diagnosis? IEEE Trans Energy Convers 18(2):238–244
428 S. Mani et al.
32. Nemec M, Drobnič K, Nedeljković D, Fišer R, Ambrožič V (2010) Detection of broken bars in
induction motor through the analysis of supply voltage modulation. IEEE Trans Ind Electron
57(8):2879–2888
33. Zagirnyak MV, Mamchur DG, Kalinov AP (2010) Elimination of the influence of supply mains
low-quality parameters on the results of induction motor diagnostics. In: Proceeding of XIX
international conference on electrical machines—ICEM, Rome, Italy
34. Salles G, Filippetti F, Tassoni C, Crellet G, Franceschini G (2000) Monitoring of induction
motor load by neural network techniques. IEEE Trans Power Electron 15(4):762–768
35. Kral C, Habetler TG, Harley RG (2004) Detection of mechanical imbalances of induction
machines without spectral analysis of time domain signals. IEEE Trans Ind Appl 40(4):1101–
1106
36. Shoen RR, Lin BK, Habetler TG, Schlag JH, Farag S (1995) An unsupervised, on-line system
for induction motor fault detection using stator current monitoring. IEEE Trans Ind Appl
31:1280–1286
37. Maier R (1992) Protection of squirrel-cage induction motor utilizing instantaneous power and
phase information. IEEE Trans Ind Appl 28:376–380
38. Obaid RR, Habetler TG (2003) Effect of load on detecting mechanical faults in small induction
motors. In: Proceeding of IEEE SDEMPED, Atlanta, GA, pp 307–311
39. Long W, Habetler TG, Harley HG (2007) Review of separating mechanical load effects from
rotor faults detection in induction motors. In: Proceeding of IEEE SDEMPED, pp 221–225
40. Drif M, Benouzza N, Bendiabdellah A, Dente JA (2001) Induction motor load effect diagnostic
utilizing instantaneous power spectrum. In: Proceeding of 9th European conference on power
electronic applications
Analysis of Unbalanced Magnetic Pull
in a Multi-physic Model of Induction
Machine with an Eccentric Rotor
2nd World Congress on Condition Monitoring
(Singapore 2–5 December 2019)
Xiaowen Li, Adeline Bourdon, Didier Rémond, Samuel Koechlin,

and Dany Prieto
Abstract Based on a multi-physic model of an asynchronous electrical machine

with a strong electro-magneto-mechanical interaction, the Unbalanced Magnetic Pull
(UMP) generated from the uneven air-gap due to the rotor eccentricity is calculated
and discussed. The proposed model includes strong couplings between mechanical,
magnetical, and electrical behaviors, particularly, the relationship between angle
sampling and time sampling about the rotor position which introduces the instanta-
neous angular speed. Simulations with different input static eccentricity values are
investigated in rated operation, and the effects of UMP on the dynamic behavior of
the whole system are analyzed. Vibration analysis is performed at the supports to
extract information in vibration signals for eccentricity monitoring.
Keywords Unbalanced magnetic pull (UMP) · Asynchronous electrical machine ·

Eccentric rotor · Rotor center orbit · Multi-physic model
1 Introduction
Unbalanced Magnetic Pull (UMP) is a radial electromagnetic force generated in the

uneven air-gap of the electrical machine when a rotor eccentricity exists. Usually, it
is ignored in the vibration analysis due to the difficulty of its measurement. However,
once it is produced, it will pull the rotor toward the direction of the minimum air-gap
and continues to worse the eccentric situation. Noise and vibration are generated
inside of motors which increase the bearing wear and other NVH problems on the
external structure. In the industry, sometimes the motor needs to be installed on a
X. Li (B) · A. Bourdon · D. Rémond

University of Lyon, INSA-Lyon, CNRS, UMR5259, LaMCoS, F-69621 Villeurbanne, France
e-mail: xiaowen.li@insa-lyon.fr
X. Li · S. Koechlin · D. Prieto
NIDEC PSA Emotors, 212, Bd Pelletier, 78955 Carrieres-Sous-Poissy, France

https://doi.org/10.1007/978-981-15-9199-0_40
430 X. Li et al.
poorly stable structure like a cantilever or a long flexible shaft. Under the effect of
gravity, the buckling of the shaft increases the rotor eccentric level and potentially
leads to the rotor–stator contact during the operation that may destroy windings
or motor inside structure. Considering that an electrical machine is a multi-physic
system consisting of three different fields that operate together and interact each
other; a multi-physic analyzed model is established to study the effects of UMP on
the vibration behavior of the induction machine with an eccentric rotor. A particular
attention is paid to the rotor angle position which is considered as an output without
any assumption on its time history. This relationship offers the capability to describe
precisely the cyclic phenomena (component with discrete geometry in rotation) and
time phenomena (as the resonance in mechanical parts).
In this paper, the simulation results about 10% eccentricity of the average air-gap
distance are discussed firstly to present the effect of UMP to the whole system. After-
wards, several simulation results about different eccentricity values are compared to
highlight the linear variation with small eccentricities and nonlinear variation effect
appearing for large eccentricity values. The rotor center orbit is studied to analyze the
stability of the system. The connection between the support vibration and UMP is also
analyzed at the end of the paper to explain one of the important vibration sources
in the electrical machines and how to extract information in vibration signals for
eccentricity monitoring.
2 Method of Analysis
2.1 Multi-physic Model
Fourati et al. have proposed a multi-physic model about a 3-phases induction

machine based on angular approach in [1]. In the present paper, this electro-magneto-
mechanical interaction is reinforced by adding the coupling of the UMP (Femx ,Femy )
generated inside of the motor due to the air-gap eccentricity. The modified strong
multi-physics coupling is described in Fig. 1.
2.2 Effective Air-Gap Distance
Considering the rotor eccentricity, the air-gap distance around the periphery of the
rotor is no longer uniform. In most literatures, people used to calculate it by an
approximative expression as a function of the shaft rotation angle as in [2]. Only
the rotor whirling motion is considered so the rotor center has no variation in terms
of rotation angle. Hence in this paper, a formula of instantaneous effective air-gap
distance as a function of rotor center coordinates is defined in (1) according to Fig. 2.
Analysis of Unbalanced Magnetic Pull in a Multi-Physic Model … 431
Fig. 1 Multi-physics couplings
Fig. 2 Calculation of instantaneous air-gap associated with the stator tooth i

ei (xr , yr ) = Rs2 − 2Rs (cos θsi · xr + sin θsi · yr ) + xr2 + yr2 − Rr (1)
where Rs and Rr are the radius of stator inner ring and rotor outer ring, respectively,
θsi is the anglar position of each stator tooth, and xr and yr are the coordinate of
432 X. Li et al.
the rotor center. This is the effective air-gap distance between each pair of stator
and rotor teeth to be adapted with the Ostovic model from [9] used to calculate the
air-gap permeance.
2.3 Calculation of UMP
Oppositively, from the Maxwell stress tenseur method used in the references [2–7],
in our model, UMP is achieved by the principle of virtual work [1, 5, 8, 9] to obtain a
more realistic UMP value in a simpler manner. Therefore, each component of UMP is
calculated by deriving the magnetic co-energy along the x and y direction according
to (2) and (3).

1
ns
nr
1 ∂ pi j
Femx = ∅i j · − 2 ·
2
(2)
2 i j pi j ∂x

1
ns
nr
1 ∂ pi j
Femy = ∅i j · − 2 ·
2
(3)
2 i j pi j ∂y
where n s and n r are the stator and rotor tooth number, respectively, ∅i j is the magnetic
flux transferred in permeance network of the air-gap area, and pi j is the air-gap
permeance between each pair of stator and rotor teeth.
3 Results and Discussions
3.1 Simulation Results of the Case with 10% Ee Static

Eccentricity
10% static eccentricity of the average air-gap distance is introduced in the horizontal
axis (x axis) to perform a simulation in rated operation. In order to highlight the
strong electro-magneto-mechnical interaction of this multi-physic model called as
“Model EMM,” a similar model known as “Model EM” without mechanical coupling
is established with a fixed rotor center.
3.1.1 Variation of UMP
It can bee seen in Fig. 3 that with an input static eccentricity, the two models generate
2 UMP components that fluctuate around two constant values. It is reasonable that
120 120
Fx Fx
Fy Fy
100 100
Unbalanced magnetic pull(N)

80 80
60 60
40 40
20 20
0 0
-20 -20
195 195.5 196 196.5 197 197.5 198 198.5 199 199.5 200 195 195.5 196 196.5 197 197.5 198 198.5 199 199.5 200
Shaft position (revolution) Shaft position (revolution)
Fig. 3 Variation of UMP as a function of shaft position (Left: Model EMM; Right: Model EM)
the x component is bigger than y component because the eccentricity is set up in the
x axis direction. However, the average value of y component is not zero which means
the UMP resultant force does not point to the narrowest air-gap direction but rotates
with a small angle from it. According to [10], this can be explained as the effects of
the compensation currents induced in the rotor cage bar when facing the alternation
of the big and small air gap. If two model results are compared, even their average
values are almostly the same, those amplitudes from the model EM are bigger than
those simulated from model EMM. It illustrates that without a mechanical coupling,
UMP is overestimated.
3.2 Variation of UMP with Different Eccentricities
3.2.1 With a Relatively Small Eccentricity
The polar diagram of UMP is realized by plotting the variation of F y as a function of

F x in the polar coordinate. As shown in Fig. 4, the arrows are plotted by the average
values of (F x , F y ) with the solid line while the trajectory at the end of each arrow
represents their fluctuation in the last revolution with the dotted line. The magnitudes
of UMP increase proportionally with the increase of the input static eccentricity while
their offset angle hardly changes. The fluctuation in the case of 40% Ee is bigger
than others because of the strong modulation phenomena.
3.2.2 With a Relatively Large Eccentricity
In rated operation, once the eccentricity exceeds 40% Ee, some nonlinear effects
appear in all the simulation results. From Fig. 5, different from the variation of
UMP with 40% Ee, there is a strong pulsation emerged in the case with 50% Ee. The
434 X. Li et al.
Polar diagram of UMP in the last revolution 40%Ee

90 500
30%Ee
120 60
400
20%Ee
300
150 30 10%Ee
200
100
180 0
210 330
240 300
270
Fig. 4 Polar diagram of UMP with different eccentricities in rated operation
500 6000
Fx Fx
450 Fy 5000 Fy
400 4000
350 3000
300 2000
250 1000
200 0
150 -1000
100 -2000
50 -3000
0 -4000
980 982 984 986 988 990 992 994 996 998 1000 1380 1382 1384 1386 1388 1390 1392 1394 1396 1398 1400
Shaft position (revolution) Shaft position (revolution)
Fig. 5 Variation of UMP as a function of shaft position (Left: 40% Ee; Right: 50% Ee)
global magnitudes increase largely even the eccentricity amplifies only by 1.25 times.
Meanwhile, the variation of F y component fluctuates almost at the same level as F x
component does. That’s why in Fig. 7, the rotor center orbit of 40% Ee converges
to a small circle close to the initial rotor center position while the rotor center orbit
with 50% Ee rotates around the initial position with a relatively big radius. From the
angular spectrum of UMP magnitude given in Fig. 6, despite of the even harmonics
of the supply frequency, some modulations between the harmonics of f t and those
supply frequency harmonics will appear in the case of 50% Ee. It is not clear how to
explain the origin of f t but it will increase with the increase of rotation speed. This
Angular spectrum of UMP magnitude: 40%Ee Angular spectrum of UMP magnitude: 50%Ee
103 103
ft
102 102
ft
101 101 2fs

Amplitude (N)
Amplitude (N)
2fs
4fs
100 100
4fs
6fs
10-1 10-1
10-2 6fs 10-2
-3
10 10-3
10-4 10-4
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Angular frequency (event/revolution) Angular frequency (event/revolution)
Fig.6 Angular spectrum of UMP magnitude
Fig. 7 Rotor center orbit in xoy section
strong modulation phenomenon also appears in other simulation results with 50%
Ee, for example, the variation of electromagnetic torque will fluctuate with a larger
amplitude. This will potentially increase the risk of instability of the whole system.
3.3 Vibration Analysis from the Support Acceleration
Based on the simulation with 10% Ee input static eccentricity, the translation accel-
eration of support 1 from the mechanical part is analyzed along two directions. In
order to study their periodicity, their angular spectra are compared and presented
in Fig. 8. From three spectra, all the important characteristic frequencies can be
identified including 2 f θs = 2.032 ev/rev, Nr = 30 ev/rev and the modulation
Nr ± 2 f θs that are caused by the input eccentricity. However from the spectra of
UMP and the acceleration along x axis “Accx1,” more frequencies like the harmonics
436 X. Li et al.
Angular spectrum of UMP magnitude

2
10
X: 30
Amplitude (N)
X: 27.97
X: 2.032 Y: 1.166 Y: 1.294 X: 32.03
0
10 Y: 0.3287 Y: 0.3624
X: 12.19 X: 18.29 X: 34.06

Y: 0.01261 Y: 0.0129 X: 24.39
Y: 0.00996
-2 X: 6.096 Y: 0.006047
10 Y: 0.00213
-4
10
0 5 10 15 20 25 30 35
Angular frequency (event/revolution)
Angular spectrum of acceleration Accx1

10-4
Amplitude ( m/rad 2 )
X: 30
X: 27.97 Y: 3.698e-05
X: 2.032 Y: 3.353e-05 X: 32.03
10-6 Y: 4.601e-07 X: 6.096 X: 12.19 X: 18.29 Y: 1.023e-05

X: 24.39 X: 34.06
Y: 1.885e-07 Y: 1.905e-07 Y: 1.664e-07 Y: 1.926e-07
Y: 1.139e-07
10-8
10-10
0 5 10 15 20 25 30 35
-4
Angular spectrum of acceleration Accy1
10
Amplitude ( m/rad 2 )
X: 27.97
Y: 3.308e-05
X: 2.032 X: 32.03
-6 Y: 8.983e-06
10 Y: 4.478e-07
X: 30
Y: 2.191e-06
X: 34.06
Y: 2.094e-07
-8
10
-10
10
0 5 10 15 20 25 30 35
Fig. 8 Angular spectrum of different results (Top: UMP magnitude, middle: translation acceleration
of support 1 along x; Bottom: translation acceleration of support 1 along y)
of 6 f θs = 6.096 ev/rev allowed to be recognized although the spectrum of UMP is

richer than that of Accx1 considering that Accx1 is an indirect signal excited by the
UMP. Furthermore, by comparing the amplitude of the frequency 30 ev/rev between
the spectra of Accx1 and Accy1, its simple to distinguish the static eccentricity
appeared in which direction.
4 Conclusions
By adding the coupling of UMP with an unfixed rotor center, the multi-physic
induction machine model based on angular approach is developed in the term of
electro-magneto-mechnical interactions. The effective air-gap distance is defined as
a function of the rotor center coordinate to calculate the UMP with different rotor
eccentricities. The variation of UMP is more reasonable when compared with the
results from the model without mechanical coupling. An offset angle of UMP resul-
tant force is detected from the direction of the narrowest air-gap which does not
change with the variation of a relatively small eccentricity, and meanwhile, their
UMP magnitude increases linearly with the augementation of the eccentricity. In the
case of a relatively large eccentricity, nonlinear effects appear to all the simulation
results and the harmonics begin to show up in their angular spectrum. The whole
machine tends to be instable. The spectra of UMP and those of the translation accel-
eration from support 1 demonstrate the possibility of detecting the rotor eccentricity
from the vibration analysis of support part. It should be noted that due to the tangen-
tial component of UMP in respect of narrowest air-gap direction, the vibrations can
be detected from both the two directions.
Acknowledgements Authors gratefully acknowledge the support of Association National

Research Technology of France and Enterprise Nidec Leroy Somer.
References
1. Fourati A et al (2017) Angular-based modeling of induction motors for monitoring. J Sound

Vib 395:371–392
2. Guo D, Chu F, Chen D (2003) The unbalanced magnetic pull and its effects on vibration in a
three-phase generator with eccentric rotor. J Sound Vib 254(2):297–312
3. Smith AC, Dorrell DG (1996) Calculation and measurement of unbalanced magnetic pull in
cage induction motors with eccentric rotors. Part 1: analytical model. IEE Proc Electr Power
Appl 143(3):193–201
4. Han X, Palazzolo A (2016) Unstable force analysis for induction motor eccentricity. J Sound
Vib 370:230–258
5. Holopainen TP (2004). Electromechanical interaction in rotordynamics of cage induction
motors, Number 543. VTT Publications, p 543
6. Mahyob A et al (2007) Induction machine modelling using permeance network method for
dynamic simulation of air-gap eccentricity. In: 2007 European Conference on Power Electronics
and Applications, EPE
7. Boy F, Hetzler H (2019) A co-energy based approach to model the rotordynamics of electrical
machines. In: Mechanisms and Machine Science, vol 63
8. Xu X, Han Q, Chu F (2016) Nonlinear vibration of a generator rotor with unbalanced magnetic
pull considering both dynamic and static eccentricities. Arch Appl Mech 86(8):1521–1536
9. Ostovic V (1989) Dynamics of saturated electric machines, 1st edn. Springer, New York, USA
10. Saint-Michel J (2001) Influence of rotor eccentricity on IR motor C, 1st edn. Nidec Leroy-
Somer, France
Bearings Fault Diagnosis
in Electromechanical System Using
Transient Motor Current Signature
Analysis
Alok Verma, Haja Kuthubudeen, Viswanathan Vaiyapuri,

Sivakumar Nadarajan, Narasimalu Srikanth, and Amit Gupta
Abstract The bearing faults may lead to expensive and catastrophic failures, which
affect the reliability of power drivetrain in electromechanical system. To reduce
impact of the bearing failures, accurate analysis and timely diagnosis are required
for improved reliability of the electromechanical systems. In this work, an enhanced
transient current signature analysis has been investigated for aircraft application using
permanent magnet synchronous motor with bearing defects. The motor current under
steady-state and transient conditions is acquired from an experimental test-rig with
bearing defects at different loading and speed levels. The acquired signals are first
investigated using frequency domain analysis and then compared with the time–
frequency domain analysis such as wavelet analysis. The discrete wavelet transform
is used to analyze on the calculated residual current for the bearing fault diagnosis.
The proposed time–frequency-based technique is able to provide useful features
that characterize the condition of the bearings on transient current. Further, back-
propagation neural network is used over the calculated features which distinguishes
the defective bearing from the healthy bearing with high accuracy.
Keywords Condition monitoring · Bearing defects · Transient analysis ·

Frequency domain analysis · Wavelets · Neural network
A. Verma (B) · H. Kuthubudeen

Rolls-Royce @ NTU Corporate Lab, Nanyang Technological University, Singapore, Singapore
e-mail: akverma@ntu.edu.sg
V. Vaiyapuri · S. Nadarajan · A. Gupta
Rolls-Royce Singapore Pte Ltd., Singapore, Singapore
N. Srikanth
Energy Research Institute @ NTU, Nanyang Technological University, Singapore, Singapore

https://doi.org/10.1007/978-981-15-9199-0_41
440 A. Verma et al.
1 Introduction
Electrical machines and associated electromechanical components are becoming

integral part of aircraft engine for onboard electric power generation. But, the faults
in the electromechanical system may lead to significant damage to the machine and
system associated. The techniques like condition monitoring and faults diagnosis are
used in modern industries to overcome these problems. The condition monitoring
is a key to improve the reliability and availability of electromechanical system by
detecting and diagnosing the faults at an early stage to prevent failures.
The faults in electromechanical system can be distinguished as mechanical and
electrical faults. As per the IEEE standard 493-1997 and statistical study of EPRI,
bearing defects are the most common failures in rotating machines which constitute
more than 40–45% of all defects [1]. Large number of papers have been published
in the field of fault diagnosis of rotating machines, and most of the work has been
done on vibration monitoring of the machine to detect and classify the types of fault.
The motor current signature analysis (MCSA) is a popular technique which uses
current harmonics to detect and diagnose bearing faults because the effect of faults
is reflected to the harmonics of stator current. The MCSA is a non-invasive, cheaper,
and does not require extra sensors because current is already monitored for protec-
tion and control purposes. Schoen et al.[2] first evaluated that the bearing defect
produced vibration characteristics frequency and modulated the air gap flux density,
so stator current frequencies were produced and can be used for fault detection.
Later, comparison of vibration and current monitoring technique was also given by
Benbouzid et al.[3]. Blodt et al.[4] proposed a noble analytical model for the rolling-
element bearing faults of electrical motor which influenced the stator-based current.
Yazici and Kliman [5] presented a time–frequency domain method for the diagnosis
of broken rotor bar and bearing faults using current, torque, and speed signals from a
3/4 HP induction motor. An experimental study on permanent magnet synchronous
motor (PMSM) was also used for the diagonosis of bearing faults using short-time
frequency transform (STFT) and Gabor Spectrogram of current signals [6].
Generally, time domain, frequency domain, and time–frequency domain tech-
niques are used in the literature for bearing fault diagnosis [7, 8]. Sometimes, tech-
niques of information entropy and artificial intelligence were also used for the fault
diagnosis purpose [9, 10]. The transient current signal contains non-stationary funda-
mental components. So, time-domain and frequency-domain analyses are not useful
for the non-stationary waveform [11, 12]. Therefore, technique based on time–
frequency domain analysis such as wavelet transform can be used to analyze the
bearing defects for the non-stationary current signals [13]. Inspired by their work, this
work is focused on the bearing fault diagnosis of rotating electromechanical systems
using transient current signature analysis for aircraft application. The presented work
is first of its kind in the author’s knowledge where the diagnosis algorithm is success-
fully implemented with the experimental analysis for the diagnosis of bearing faults
using transient current signature analysis.
Bearings Fault Diagnosis in Electromechanical System … 441
2 Experimental Analysis
2.1 Experimental Setup
The experimental setup along with the major components are shown in Fig. 1. The
experimental setup has a 3- PMSM as a test machine coupled to a 3- induc-
tion machine (load machine through roller bearing, shaft, gearboxes, and flexible
coupling). The gear-ratio of the respective gearboxes was adjusted in such a way that
speed of the input shaft is almost equal to the output shaft.
The current sensor was connected to the stator winding cable in order to collect the
current signals using a 32-channel data acquisition system as shown in Fig. 2a, b. The
components and their specification used in this experimental setup are given in Table
1, and Table 2 shows the specification of bearings used in the experimental setup.
The healthy bearing, outer race defective bearing before and after the experiments
are shown in Fig. 3.
Fig. 1 Experimental setup
Fig. 2 a Current sensors attached to stator wiring, b data acquisition system

442 A. Verma et al.
Table 1 Components
Sensors/machine Sensitivity/specification
specifications for the
experimental setup Current sensor 10 mV/A
DAQ 32 channels
Speed drive 32–38 Amp, 380–500 V
PMSM 20 HP, 50 Hz, 5000 RPM
Induction motor 20 HP, 50 Hz, 5000 RPM
Table 2 Bearing
Bearing type Angular contact ball bearing
specifications
Number of balls (n) 13
Contact angle (F) 40°
Pitch diameter (PD) 38.5 mm
Ball diameter (BD) 8.9 mm
Fig. 3 a Healthy bearing, b bearing with outer race defect: before experiment and c bearing with
outer race defect: after experiment
2.2 Experimentation
The startup transient 3- current signals were collected in two sets. First set of
data were acquired from healthy system (N1 ) with the speed and load ranging from
800–1500 rpm and 0–75% of full load, respectively, at an equal interval of 50 rpm
and 25% of full load. Then, the bearing with outer race defect (O1 ) was installed to
the experimental setup to acquire the second sets of current signals, with the same
operating conditions as that of the first set. The healthy system was denoted as 0
and the system with outer race defect was designated by 1. In this work, individual
data set consisted of 60 experimental responses for transient condition so, the total
numbers of experimental responses are 120 in the two sets. The sampling rate for
each signal was set to 10KS/s. The key test cases for the experiments are given in
Table 3.
Table 3 Controlling factors

Controlling factor Level
and levels for the experiments
A Rotational speed (rpm) 800–1500
B Loading condition (%) 0–75
C Bearing conditions 0 and 1
3 Methodology
The method used in this paper consists of total five steps, i.e., experimental test-rig
arrangement, data collection, data pre-processing, feature extraction, and classifi-
cation. The acquired current signals are analyzed in time–frequency domain such
as discrete wavelet transform. The wavelet features of different bearing conditions
operating under different operating speeds and loads are fed as inputs to the back-
propagation neural network (BPNN). Thorough description of the procedure is given
in the sub Sects. 3.1 and 3.2.
The acquired signals are first investigated using frequency domain analysis and
then compared with the time–frequency domain analysis such as wavelet analysis.
Fast Fourier transform (FFT) of steady-state and transient state current signals for
healthy and outer race defective bearings conditions are shown in Fig. 4a–d. From
Fig. 4a, b, it can be observed that Fig. 4b has prominent 1X and F o is clearly visible
as compare to the Fig. 4a. F o can be expressed as

n BD
fO = fr 1− cos = 267.46 Hz (1)
2 PD
where PD is the pitch diameter, is the contact angle, n is the number of balls,
BD is the ball diameter, f r is the rotating frequency, f o is the outer race frequency
and 1X, 2X, … are frequency sidebands. The spectrum shown in Fig. 4b with F o ,
clearly demarcates that the system has bearing with outer race defect. However,
Fig. 4c, d represents the spectrum for the healthy and outer race defective bearing
using transient current signals and we can notice that F o is not at all visible for the
outer race defect fail to differentiate between healthy and outer race defect as shown
in Fig. 4d. The non-stationary fundamental components occurred for the transient
current signals. So, traditional frequency domain analysis is unable to detect faults
using transient current signals as shown in Fig. 4d. Hence, time–frequency domain
analysis needs to be use for the transient conditions.
3.1 Wavelet Transform
The wavelet transform is a well known signal processing technique for the non-
stationary signal analysis having transient behavior [13]. Continuous wavelet trans-
form (CWT) of f (t) is obtained by the sum of all signal multiplied by the scale and
444 A. Verma et al.
Fig. 4 FFT results for different bearing conditions at 1000 RPM and no load: a current spectrum
for healthy bearing at steady-state condition, b current spectrum for outer race defective bearing
at steady-state condition, c current spectrum for healthy bearing at transient state condition and
d current spectrum for outer race defective bearing at transient state condition
shifted form of the wavelet function (t).
∞
1 t−y
CWT(x, y) = √ f (t) dt (2)
|x| x
−∞
where x is a scale parameter and this is inverse of frequency, y represents as

translation parameter. Further, discrete wavelet transform (DWT) is a discretization
of CWT and expressed as
∞
1 t − 2i j
DWT(i, j) = √ f (t) dt (3)
2i 2i
−∞
where parameter i,j = 1, 2, …. High frequencies are associated wide bandwidth,

while low frequencies are associated with narrow bandwidth. Using DWT, the signal
with lower bandwidths and higher bandwidths can be separated which helps in
adjusting the resolution at high frequency.
3.2 Artificial Neural Network
In this work, BPNN is considered for the diagnosis of bearing fault using transient
current signature. The BPNN consists of input layer, hidden layer, and output layers.
Input layer is used to receive a set of inputs and forward the signals to the hidden
layer. The hidden layer processed that information and send to the output layer and
finally, output layer transfers the values to an external receiver. The nodes in a hidden
layer and number of hidden layers are variable and dependent on the convergence of
result. Further, the weight (W rs ) is used to modify the output signal of hidden layer
from rth node to sth node of the output layer. The summation of modified signal is
again adjusted here by the sigmoid function and send that signals to the output layer.
The results for output layer can be calculated as,
m

Ops = f Wrs Opr , s = 1, 2, . . . ,n (4)
r =0
Here, f is used as sigmoid transfer function. The predicted values are compared
with the desired values during training period, and the mean square error (E p ) was
calculated as
n
1 2
Ep = Dpk − Opk (5)
i=1
n
Here, Dps is the desired output value and Ops is the calculated output for the
pth pattern. The detail classification results with the training, testing, and errors are
presented in the result section.
4 Results
In this work, transient stator current signals acquired through permanent magnet
synchronous motor (PMSM) are used to calculate the features using discrete wavelet
transform by calculating the residual current signal. The acquired transient current
signals need to be further preprocessed to remove the fundamental components from
the signal. So, the average moving filter was used for this task and the residual
current signals by subtracting filtered signal from original transient signal were used
for feature extraction using DWT as shown in Fig. 5. The average moving filter can
be calculated as
1
n+(b−1)
bs (n) = xs (n) (6)
s n
446 A. Verma et al.
Fig. 5 Startup transient current analysis for different bearing conditions at 1000 RPM and different
loading conditions: a smooth current signal for system with healthy bearing at no load, b smooth
current signal for system with outer race defective bearing at no load, c wavelet decomposition at
d3 level of a healthy system at 0–75% loading and d wavelet decomposition at d3 level of a system
with outer race bearing defect at 0–75% loading
where the average value from filter is bs , n is considered as data points, smoothing
factor is s, and actual data point is x s . The smooth transient current signals are
extracted using moving average filter for healthy and faulty bearings as shown in
Fig. 5a, b. Further, wavelet decomposition was applied on the residual current signal
for both the healthy and faulty bearings at different loading conditions as shown
in Fig. 5c, d. From Fig. 5d, the significant changes in the features using DWT can
Fig. 6 a Variation of neurons in hidden layer with mean square error and b variation of experimental
value with the predicted value
Table 4 Accuracy of bearing fault diagnosis on the test set as N 1 U O1

S. No. Network No. of iterations Mean square Mean square Prediction error
architecture error training error testing (%)
1 3-10-1 893 0.014725 0.013559 12.8577
2 3-13-1 494 0.010054 0.008795 10.12103
3 3-15-1 1069 0.01287 0.009499 11.2014
be observed in between 150 and 200 samples for the defective bearing at loading
from 0 to 75%. Later, the calculated features along with back-propagation neural
network (BPNN) is employed as noble method to diagnose bearing defects. In this
presented work, the number of inputs to the BPNN is bearing condition, load, and
speed whereas output is a bearing fault diagnosis.
The BPNN was first trained and then tested for different values of the network
architecture and the BPNN architecture 3-13-1 considered best as shown in Fig. 6.
The best results for the trained and tested BPNN architecture are given in Table
4. It is noticed that the maximum error for the best neural network architecture is
approx. ±12.85% of the experimental value and minimum error for is ±10.12%. This
observation clearly demarcates that the model based on BPNN is useful to trained
and predict the bearing defects using startup transient current signal with reasonable
and good accuracy.
5 Conclusions
In this work, the combination of DWT and BPNN is suggested for the bearing fault
diagnosis in the electromechanical system using startup transient current signals.
The results claim that frequency domain analysis is not useful for the non-stationary
waveform. Therefore, a technique based on time–frequency domain analysis, such
448 A. Verma et al.
as DWT, is used on the residual current to analyze the bearing defects. The signif-
icant differences between healthy and defective bearing have been observed on the
extracted DWT features. The results and the prediction accuracy of the BPNN model
are quite close to 90%, which shows that the combination of DWT and BPNN model
using transient current signals can be used effectively.
Acknowledgements This work was performed at the Rolls-Royce@NTU Corporate Lab under the
Corp Lab@University Scheme with support from the Rolls-Royce Pte Ltd. Singapore and National
Research Foundation (NRF) Singapore
References
1. Allbrecht PF, Appiarius JC, McCoy RM, Owen EL, Sharma DK (1986) Assessment of the
reliability of motors in utility applications. IEEE Trans Energy Convers EC-1(1):39–46
2. Schoen RR, Habetler TG, Kamran F, Bartheld RG (1995) Motor bearing damage detection
using stator current monitoring. IEEE Trans Ind Appl 31:1274–1279
3. Benbouzid MEH, Vieira M, Theys C (1999) Induction motors faults detection and localization
using stator current advanced signal processing techniques. IEEE Trans Power Electron 14:14–
22
4. Blodt M, Granjon P, Raison B, Roistaing G (2008) Models for bearing damage detection in
induction motors using stator current monitoring. IEEE Trans Ind Electron 55:383–388
5. Yazıcı B, Kliman GB (1999) An adaptive statistical time-frequency method for detection of
broken bars and bearing faults in motors using stator current. IEEE Trans Ind Appl 35:442–452
6. Rosero J, Cusido J, Espinosa AG, Orteg JA, Romeral L (2007) Broken bearings fault detection
for a permanent magnet synchronous motor under non-constant working conditions by means
of a joint time frequency analysis. In: IEEE international symposium on industrial electronics
(ISIE), Spain, 4–7 June 2007, pp 3415–3419
7. Immovilli F, Bellini A, Rubini R, Tassoni C (2010) Diagnosis of bearing faults in induction
machines by vibration or current signals: a critical comparison. IEEE Trans Ind Appl 46:1350–
1359
8. Devaney MJ, Eren L (2004) Detecting motor bearing faults. IEEE Instrum Meas Mag 7:30–50
9. Verma A, Sarangi S, Kolekar M (2016) Misalignment faults detection in an induction motor
based on multi-scale entropy and artificial neural network. Electr Power Compos Syst 44:916–
927
10. Verma A, Sarangi S (2015) Fault diagnosis of broken rotor bars in induction motor using multi-
scale entropy and backpropagation neural network. In: Intelligent computing and applications,
vol 343. Springer, New Delhi, pp 393–204
11. Douglas H, Pillay P (2005) The impact of wavelet selection on transient motor current signature
analysis. In: IEEE international conference on electric machines and drives (IEMDC), San
Antonio, TX, USA, May 2005, pp 80–85
12. Widodo A, Yang B-S, Gu D-S, Choi B-K (2009) Intelligent fault diagnosis system of induction
motor based on transient current signal. Mechatronics 19:680–689
13. Douglas H, Pillay P, Ziarani AK (2004) A new algorithm for transient motor current signature
analysis using wavelets. IEEE Trans Ind Appl 40:1361–1368
SHM/NDT Technologies and Applications
Assessment of RC Bridge Decks
with Active/Passive NDT
T. Shiotani, H. Asaue, K. Hashimoto, and K. Watabe
Abstract We are facing a lot of infrastructures damaged seriously. In considera-

tion with restriction of the maintenance budget, strategy to prolong their service as
long as possible has been generally adopted. To maintain those structures reasonably,
current condition of the structures shall be properly evaluated by ideal inspections, for
example, using NDT. The authors have thus been studying NDT utilizing elastic wave
approaches to quantify the damage of infrastructures. In the paper, periodic moni-
toring and continuous monitoring with elastic wave approaches are demonstrated
respectively. Specifically elastic wave- or AE- tomography is applied to the existing
concrete bridge deck to interpret the damage distribution with their characteristic
parameters. As for the continuous monitoring, newly developed super acoustic (SA)
sensors and a four channels FPGA-based edge computing unit with event-driven
technology are demonstrated. And to introduce the NDT into in-situ maintenance
program, essential issues are summarized. As for a key technology of remote moni-
toring, several fundamental aspects of laser-induced elastic waves are experimentally
studied.
Keywords Infrastructures · NDT · Elastic wave · Tomography · SA sensor · Laser

induced AE
1 Introduction
Deterioration processes of materials/structures can be classified into several stages,

namely primary, secondary, and tertiary stages. Although ideal proactive maintenance
shall be conducted during primary or at the latest secondary stage before coming into
the seriously deteriorated level of tertiary stage, almost infrastructures constructed
T. Shiotani (B) · H. Asaue · K. Hashimoto

Department of Civil & Earth Resources Engineering, Kyoto University, C-Cluster, Nishikyo-Ku,
Kyoto, Japan
e-mail: Shiotani.tomoki.2v@kyoto-u.ac.jp
K. Watabe
Corporata Research & Development Center, Toshiba Corporation, Tokyo, Japan

https://doi.org/10.1007/978-981-15-9199-0_42
452 T. Shiotani et al.
intensively during the rapid economic era have been left without proper maintenance
resulting in reaching their original life-span of about 50 years at present. In addition
to consider the current low economic growth leading to the restricted investment for
existing infrastructures, replacement of those with new ones will not be realized,
while life extension policy for the existing structures sometimes with retrofitting is
generally adapted worldwide. When focusing even on the deteriorated tertiary stage,
there are still issues to solve, for example, how to determine the further implementa-
tion for the deteriorated structures, i.e., if it keeps leaving, repair, or replace. And for
the replacement, how to prioritize the structure for repair program is still in study.
Quantification of the repair work and estimation of life time behavior based on their
integrity are the issues in progress. The authors have thus been studying snap-shot
monitoring for periodic inspection and continuous monitoring to alert the critical
condition with NDT. In this paper, as applications of the NDT, bridge decks made
of reinforced concrete (RC) are exemplified, which is referred to as a most budget
taking member of road infrastructures in Japan [1].
2 Snap-Shot Monitoring
In Japan, periodic inspections with close visual observation in a span of five-year

have been regulated in the Road Act in 2014. Health condition of the infrastructure
is thus been rated every five years with naked eyes. As these ratings are obtained by
the surface information of infrastructures, these only suggest the condition during
the final phase of deterioration, namely ones for collective/reactive maintenance so
as to realize quantitative evaluation even during initial phase of deterioration, the
method which can evaluate inside of the infrastructures has been in great demand.
It seems difficult to directly measure the whole of infrastructures with their prop-
erties; however, there is a potential for elastic wave techniques as these provide the
health information of the structures with such parameters as elastic wave velocity
and energy attenuation. Accordingly, the authors have been studying elastic wave
techniques enabling the structural owner to obtain damage or deterioration informa-
tion. Specifically elastic wave tomography and acoustic emission tomography are
the candidates to quantify the damage. The basic analytical procedure and the appli-
cability of these techniques have already been reported as elastic wave tomography
[2–4] and AE tomography [5–7]. A typical result of AE tomography is introduced
as in the following together with the verification using core samples [7].
Two RC panels in a real 40 years old RC bridge deck, which has deterioration
such as rebar corrosion due to salt attack and fatigue failure due to mobile loads,
were selected. One referred to as Panel A is seriously deteriorated as combweb-like
cracks are readily confirmed on the bottom, and the other referred to as Panel B has
no remarkable cracks observed on the bottom. The AE measurement was carried out
with 12 AE sensors of 60 kHz resonance, set on the bottom side of the RC bridge
deck as shown in Fig. 1. Random excitations with a hammer of a φ11 mm curvature
edge were carried out for exciting elastic waves on the top surface of the panel.
Assessment of RC Bridge Decks with Active/Passive NDT 453
Fig. 1 AE sensor arrangement for AE tomography (unit: mm)
Figure 2a shows velocity tomogram obtained from AE tomography analysis and

AE sources induced by mobile load in service before implementing artificial exci-
tation with a hammer. Two facts can be clarified: small velocity areas suggesting
serious deterioration are find both cases of seriously deteriorated and minor deterio-
rated, and more AE sources are found for the case of minor deterioration (see Panel
B) than of serious case (see Panel A). These implies that internal damage cannot
always be compatible to that of the surface. As for the case that the damage devel-
oped remarkably, AE sources, namely secondary AE sources become more active
to that case of more deteriorated. Not only less numbers of secondary AE sources
and moreover large attenuation rate during AE propagation for the deteriorated case
may be attributed to this fact. In Fig. 2b, core samples’ photos corresponding to the
(a) AE sources and velocity tomogram (m/s)
(b) Core samples retrieved from the designated points of (a)
Fig. 2 Measurement results for Panel A (left column) and Panel B (right column)
Fig. 3 Damage assessment AE source density

based on AE sources and dense
elastic wave velocity [8]
Elastic wave velocity

3 2
Moderate Minor
damage damage
high
low
Severe
sound
damage
4 1
sparse
position # of Fig. 1a can be found. Indeed the areas showing low elastic wave velocity
correspond well to the condition including visible large lateral cracks. Accordingly,
damage identification would be possible when combining AE sources and velocity
distribution as shown in Fig. 3.
3 Continuous Monitoring
As a national project under the sponsorship of NEDO, super acoustic (SA) sensors
and self-powered operated wireless AE signal processing module have been devel-
oped. The SA sensor is a MEMS device expanding the responsive frequency to the
extremely wide band between Hz and MHz [9]. The SA sensor consists of four prin-
cipal parts: sensor chip, polydimethylsiloxane (PDMS), printed circuit board (PCB),
and Au/Cr electrode as shown in Fig. 4. A beam is formed on the center of a Si sensor
chip. The beam vibrates when acoustic emission signals propagate into the sensor. A
piezoresistive area is formed at the center of the beam. Au/Cr electrode is patterned
to supply current for the piezorestive area. On the one side of the sensor chip, silicone
oil is encapsulated with a parylene film. Taking advantage of the surface tension of
the silicone oil and properly limiting the size of the gap between the beam and the
surrounding chip, leakage of the oil is prevented. An appropriate amount of PDMS
is mounted on the parylene film to form the fixed area to receive the lower-frequency
vibration.
The AE signal processing module involves a wireless edge device including the
SA sensor, a concentrator, and an analysis server. Date detected by the SA sensors is
transmitted to the server via the concentrator. The module is directly installed in the
object of interest and the continuous monitoring is executed aided by event-driven
technology (detailed later on). Appearance of the module (100 × 70 × 40 mm)
is depicted as in Fig. 5. The module can be input of four channels SA sensors,
Fig. 4 Configuration of SA sensor
Fig. 5 Appearance of the AE signal processing module (left) and internal components (right)
consisting of analog frontend processor, a signal processing circuit, a self-power oper-

ation circuit generated by solar panels, and a wake-up circuit. The signal processing
circuit involves field programmable gate array (FPGA) so that parametric features
of signal waveforms can be extracted with FPGA. An Industry, Science, Medical
(ISM) frequency band is adapted to transfer the date. The module runs by solar-
generated power; however, there is a potential of power loss when enormous data of
high frequency are detected simultaneously and, therefore, event-driven technology
is introduced, where excessive accelerations by heavy vehicles wake the system
Fig. 6 AE activity with event-driven technology (left) and resultant AE sources (right)
up. Reasonable AE monitoring is thus implemented by the event-driven technology

without missing the effective date.
The SA sensors and the module are installed into the real highway bridge deck and
continuous monitoring has been performed almost one year. The number of wake-up
and those of vehicles’ passages can be found as in Fig. 6. It is found that the system
wakes when the large cars pass and corresponding AE activity is found as well.
4 Future Technologies to Introduce the NDTs

to Maintenance Program of Infrastructures
There is still space to introduce the introduced NDTs to the maintenance program
of infrastructures. Although the elastic wave velocities below 3000 m/s appear to
suggest serious damage of RC decks of bridges, exact scales of damage or deteriora-
tion and, for example, resulting elastic wave velocities have no decisive relation so far.
In the tomography spaces of sensors and obtaining resolution of tomogram (velocity
distributions) are almost clarified, i.e., a half space of sensors is the minimal reso-
lution for the analysis [10], however, several issues on frequency, namely sensors’
responsible frequency and AE source- or propagated AE wave-frequency depending
on the condition of propagation media as concrete, and moreover corresponding
detectable scale of damage are not well studied. Nevertheless, most important fact is
that: damage scale of interest for any cases of structures shall be decided in advance to
the study by owners of structures, i.e., once the damage scale of interest and degree of
damage of interest have been set, corresponding appropriate frequency and sensors’
type and space will be uniquely determined. It is noted that these discussions have
just started recently.
Other issues should be discussed are on sensors’ installation. The sensors are
currently contacted physically to the object of interest and those are connected to the
system via cables. Remote sensing, for example, air-borne sensors or laser doppler
sensors shall be developed specified for the infrastructures.
Fig. 7 Experimental overviews (left) and configuration of specimen with artificial defect (right)
Besides, excitation of elastic waves for active monitoring places an important

role to assess the structures during tomography approaches. There are so far several
ways to excite the elastic waves: occurrence of secondary AE activity from existing
defects due to stimulation of traffic loads, artificial impacts by hammers, and naturally
induced impact by such as precipitation. In the AE tomography as the information
of the excitations is unknown and identified by a conventional AE source location
algorithm, the resultant tomograms calculated based on the arrival time differences
between the excitation- and detection-time includes errors influenced by the uncertain
information of the sources. The authors are thus studying artificial stable impacts by
laser, realizing excitations at designated positions without contact to the object.
As a preliminary study, a concrete plate of 300 × 300 × 100 mm with an arti-
ficial defect of 10 mm thick Styrofoam is subjected to high-speed laser excitations
[11]. The waveforms excited by the laser propagated in the concrete are detected by
conventional AE sensors of 60 kHz resonance. The system overview and the area of
the artificial defect are shown as in Fig. 7. Two types of laser focus namely 14 J/cm2
and 82.5 J/cm2 by 250 mJ impact are studied. As found in Fig. 8, no remarkable
differences are obtained for both cases of propagation path with/without the defect.
An area of low velocity identifies with the area of artificial defect as shown in Fig. 9.
It was found that elastic wave excitations by the laser and the tomography based on
the excitations gave a pretty reasonable result implying that remote laser excitations
could be used as the sources of elastic waves for the tomography approaches.
5 Conclusions
To contribute NDT for the maintenance program of infrastructures, two approaches

shall be studied: one is for periodic monitoring, for example, a span of several year
and the other is for continuous monitoring enabling final alert to the citizens. Active
and passive elastic wave approaches are the candidates for these roles. Tomographic
approaches provide unvisible internal information of, for example, concrete slabs,
and SA sensors and the monitoring module give real-time failure evolution. To realize
these ideas in the in-situ maintenance flow, many remaining issues shall be still
studied, for example, remote excitations for tomography approaches will be the
key technology, several fundamental aspects of laser-induced elastic waves were
experimentally studied in this paper.
Fig. 8 Detected waveforms for 14 J/cm2 (left) and 82.5 J/cm2 (right) by 250 mJ impact
Fig. 9 Wave paths (left) and velocity tomogram (right)
Acknowledgements A part of the study is implemented under the sponsorship of NEDO (New
Energy Development Organization). Their financial support is greatly appreciated. The SA sensor
is concepted by Prof. Shimoyama with Toyama Prefectural University (Formerly involved in the
University of Tokyo). His notable contribution to this project is highly applauded.
References
1. W-nexco.co.jp/Expressway Renewal Project, 30 Sept 2019, https://www.w-nexco.co.jp/renewa

lproject/
2. Kobayashi Y, Shiotani T, Shiojiri H (2006) Damage identification using seismic travel time
tomography on the basis of evolutional wave velocity distribution model. Structural Faults and
Repair 2006 (CD-ROM), Edinburgh
3. Kobayashi Y et al (2007) Three-dimensional seismic tomography for existing concrete struc-

tures. In: Proceedings of second international operational analysis conference, vol 2, pp
595–600
4. Shiotani T, Aggelis DG, Momoki S (2009) Elastic wave validation of large concrete structures
repaired by means of cement grouting. Constr Build Mater 23:2647–2652
5. Kobayashi Y, Shiotani T (2012) Seismic tomography with estimation of source location for
concrete structure. Structural Faults and Repair 2012, CD-ROM, Edinburgh
6. Iwamoto K et al (2012) Application of AE tomography for fracture visualization of rock
material. Prog Acoust Emiss XV:193–196
7. Asuae H et al (2016) Applicability of AE tomography for accurate damage evaluation in actual
RC bridge deck. Structural Faults & Repair 2016, CD-ROM, Edinburgh
8. Watabe K et al (2017) Novel non-destructive technique of internal deterioration in concrete
deck with elastic wave approaches. In: Proceedings of the 12th world congress on engineering
asset management (WCEAM 2017), Paper 93, Brisbane
9. Khang-Quang N et al (2014) Multi-axis force sensor with dynamic range up to ultrasonic. In:
Proceedings of 27th IEEE MEMS conference (MEMS2014), pp 769–772
10. Hashimaoto K et al (2018) Simulation investigation of elastic wave velocity and attenuation
tomography using rain-induced excitations. Structural Faoults and Repair 2018, CD-ROM
Edingurgh
11. Shiotani T et al (2019) Application of non-contact laser impact with acoustic emission analysis
and tomography technique for damage detection in concrete. AEWG 2019, Chicago (in print)
Estimation of In-Plane Strain Field Using
Computer Vision to Capture Local
Damages in Bridge Structures
T. J. Saravanan and M. Nishio
Abstract The monitoring of variation in strain field near the ends of girder and
bearings in steel plate girder bridges to capture local damages, is of a paramount
importance for condition assessment of its service life. It is well-known that, the
thickness reduction due to corrosion and failure at bearings, possibly will decrease
the necessary functional requirements of structure. By understanding the computer
vision technique, it is implemented for a typical bending test under quasi-static
loading to aluminum plate girder bridge model specimen at laboratory level, for
estimating in-plane strain field variation near end of girder. In this research work, the
experiments are carried out using low-cost camera and two-dimensional digital image
correlation (2D-DIC) technique is employed for analysis. The experiment is repeated
under two different loading conditions and change in boundary conditions, to capture
small responses with noise in it. The in-plane strain field patterns at different time
step corresponding to the applied quasi-static load can be clearly understood from
2D-DIC results. The obtained results are in good agreement with the strain gauge
measurements. The residual strain can also be captured at the unloading condition.
From the strain captured at the Web section, it is between 15 and 35µ which is
very low strain value. Hence, the computer vision camera can capture the low strain
variations which corresponds to local damages in steel plate girder bridge structures
through 2D-DIC technique.
Keywords In-plane strain field · Computer vision · DIC · Quasi-static loading ·

Local damage · Condition monitoring
T. J. Saravanan (B)
Institute of Advanced Sciences, Yokohama National University, Kanagawa 2408501, Japan
e-mail: thiyagarajan-jothi-gh@ynu.ac.jp
M. Nishio
Department of Engineering Mechanics and Energy, University of Tsukuba, Ibaraki 3058577, Japan

https://doi.org/10.1007/978-981-15-9199-0_43
462 T. J. Saravanan and M. Nishio
1 Introduction
In Japan, many bridges constructed during the period of high economic growth
are in service for 50 years and it is necessary to maintain and manage the bridges
appropriately and efficiently [1]. Damage due to corrosion at the end of girder and
malfunction of the bearings causes serious issues in steel plate girder bridges [2].
Periodic inspection with close proximity visual inspection has become mandatory.
The existing steel plate girder bridges are subjected to various loading conditions;
however, there are a lot of uncertainties in the properties of existing structures due
to deteriorations, seismic loading histories, and the other factors during their opera-
tions [3]. As existing structures are in service, structural parameters such as material
properties and boundary conditions have undergone deterioration (aging).
Structural health monitoring (SHM) is a method in which various conventional
sensors are attached to a structure, and the overall state of the structure is quantita-
tively and objectively grasped from the acquired measurement data. Computer vision
technique based on full-field non-contact measurements at low cost using cameras
is currently used as new application in SHM field for condition monitoring of struc-
tures at local level [4]. Due to environmental effect on steel plate girder bridges, the
boundary conditions are prone to determination and malfunctioning with high uncer-
tainty. It is difficult to grasp this change and reflect it quantitatively in the numerical
model. However, if there are uncertainties at boundary conditions, it is difficult to
independently evaluate the state of other members, such as bridge girders and floor
slabs. This is because conventional monitoring considers acquisition of a single data
such as acceleration vibration data or static strain data. Under these circumstances,
digital image correlation (DIC) techniques can be used to perform a data acquisition
of a full-field strain/displacements patterns for damage localization at local level [5].
The aim of this research work is to implement computer vision-based two-
dimensional (2D) DIC technique for the measurement of in-plane strain field near
the end of steel-plate girder bridge structures. In this paper, the experimental inves-
tigations are carried out at laboratory level on aluminum plate girder bridge model.
For estimation of strain field near the end girders and bearings, the bending test is
carried out by applying quasi-static loading. The end supports are designed with steel
bearings replicating the real scenario such that it accounts for translation and rotation.
The experiment is repeated for two different loading conditions (i.e., at L/4 and L/2
locations) with different boundary conditions to capture small responses with noise
on it. The strain field measurements obtained using CMOS camera through 2D-DIC
technique are compared with conventional strain gauges for validation.
2 Experimental Study
For the experimental investigation, a simple and economical industrial camera

(Model: DFK33UP5000, The Imaging Source Europe GmbH, Germany) with
Estimation of In-Plane Strain Field Using Computer Vision … 463
complementary metal oxide semiconductor (CMOS) sensor type. The camera is

connected to a C-mount optical lens, (Model: VS-5018H1, VS Technology, USA)
with focal length 50 mm. The pixel resolution is 2592 (width) × 2048 (height), which
accounted for the pixel size of 4.8 µm × 4.8 µm and maximum frame rate of 60
frames per second. The computer vision software utilized to view is ICCapture—
Image acquisition (Version: 2.4.642.2631) and the acquired data are communicated
to the control laptop (Panasonic, CF-SV7) through USB 3.0. For in-plane strain
field estimation, a plate girder bridge specimen is considered for the bending test by
applying quasi-static loading. Figure 1a shows the fabricated beam specimen and its
boundary condition. The beam member is made of aluminum and has a total length
of 1500 mm. The angle sections are fixed with bolts to form an I-shaped section with
an outer dimension of 60 mm × 60 mm and a top and bottom flange thickness of
3 mm and Web thickness of 6 mm. The beam is installed on the pedestal via support
members at both ends say, bearing ‘A’ and ‘B’. The strain measurements are taken
near the end of bearing ‘B’ as shown in Fig. 1a. The bearing member is made with
the same mechanism as the sliding steel bearing allowing translation and rotation
with stoppers. The rotation function is ensured by placing the upper steel plate on
the curved surface of the lower plate surface, and the translation function is ensured
by allowing displacement in the direction of the bridge axis as shown in Fig. 1b. It
is difficult to evaluate the coefficient of friction at the contact surface between the
upper and lower plates, so it is hardly applied at present. However, many existing
bridges have this support system and the degradation occurs due to corrosion. It is
installed at both ends of the beam member and is used as a simple beam boundary
condition.
The conventional strain gauges are utilized along with data logger (TDS-540,
Tokyo Measuring Instruments Laboratory Co., Ltd., Japan) to measure the reference
strain values on one side of the Web section of the beam, as shown in Fig. 2a. The
measurement layout of strain gauges is illustrated in a schematic diagram in Fig. 2b.
The other side of the Web section is prepared with random speckle pattern for the
camera measurement along with the external illuminance effect for better image
clearance for DIC post processing as shown in Fig. 2c. The camera is placed at
a distance of 300 mm away from the beam specimen. The quasi-static loading of
Bearing A Bearing B
(a) (b)
Fig. 1 Beam specimen: a strain measured at bearing ‘B’; b details of boundary conditions
Fig. 2 Strain gauge measurement set up: a beam specimen; b schematic layout; c camera
measurement
76.37 kg is gradually applied and removed on the beam specimen at two different
locations, L/4 and L/2 using weight blocks as shown in Fig. 3, while only the results
corresponding to the location L/4 is provided in this paper.
The experiment is repeated for two different boundary conditions, such as fixed
support condition (stoppers are used to arrest the translation and rotation at both ends)
and simply supported condition (translation is allowed at bearing ‘B’). To analyze the
obtained data, GOM correlate 2D-DIC software is utilized to estimate the full field
strain patterns. From Figs. 4a, b, the in-plane strain field patterns obtained at different
time steps corresponding to applied quasi-static loading at the location L/4 can be
clearly seen for both fixed and simply supported boundary conditions, respectively.
At time, t = 0 (no applied load), zero strain can be measured, while at t = 150 s (no
load), the residual strain can be captured.
The strain values obtained by strain gauges are compared with the corresponding
value by 2D-DIC analysis and the comparison plots for strain values at top and bottom
of Web section near end of the bearing ‘B’ are shown in Fig. 5a, b, respectively, for
the fixed support boundary conditions.
Similarly, the comparison plots for strain values at top and bottom of Web section
near end of the bearing ‘B’ are shown in Figs. 6 (a-b), respectively, for the simply
L/4
(a) (b)
1000
800
Load (N)
600
400
200
0
0 50 100 150
time (s)
(c)
Fig. 3 Experimental setup: a loading at L/4; b loading at L/2; c applied quasi-static load
supported boundary conditions. The obtained 2D-DIC results are found to be in good
agreement with the strain gauges.
The strain variations captured at the Web section near the end of bearing ‘B’ for
quasi-static loading applied at location L/4 is in the range of 20–25µ and 30–40µ
for the fixed support and simply supported condition, respectively, which is of very
low strain value. Likewise, the experiment is repeated for the bending test on beam
specimen by applying quasi-static load at location L/2. The strain variations are
obtained and compared among 2D-DIC and strain gauges and the strain value range
from 15 to 25µ. Still, the camera with CMOS sensor can capture very small strain
variations through 2D-DIC at local level which corresponds to the damages near the
end girder caused due to corrosion and bearing failure.
3 Conclusion
In this paper, the experimental investigations are carried out to estimate the in-plane
strain field patterns using computer vision to capture local damages in bridge struc-
tures near the end of girders. The laboratory level and plate girder bridge specimen
are considered for bending test by applying quasi-static loading at two different loca-
tions with different boundary conditions to capture the small strain variation near the
(a) fixed boundary conditions
(b) simply supported boundary conditions
Fig. 4 In-plane strain field pattern obtained from 2D-DIC

(a) at top of the web section
(b) at bottom of the web section
Fig. 5 Comparison plot for strain variation obtained from 2D-DIC and strain gauges for loading
at L/4 and fixed support boundary condition
(a) at top of the web section
(b) at bottom of the web section
Fig. 6 Comparison plot for strain variation obtained from 2D-DIC and strain gauges for loading
at L/4 and simply supported boundary condition
bearings. The obtained 2D-DIC results are in good agreement with the strain gauge
values. Hence, it reveals that local in-plane strain field patterns can be obtained using
available cameras. In future, the computed deformation fields can be used to compare
with the FE model of a similar plate girder bridge specimen and the posterior distribu-
tion of the model parameters can be done by FE model updating. It helps in reducing
the actual uncertainties of model parameters.
Acknowledgements The authors are thankful to Tatsuya Suzuki, Yokohama Nationa University,
Japan for his kind support during experiments. The first author is grateful to the Japan Society for
the Promotion of Science (JSPS); whose postdoctoral fellowship program supports his stay in Japan
during this work.
References
1. Nishio M, Marin J, Fujino Y (2012) Uncertainty quantification of the finite element model of
existing bridges for dynamic analysis. J Civ Struct Health Monit 2(3–4):163–173
2. Ahn JH et al (2013) Residual shear strength of steel plate girder due to web local corrosion. J
Const Steel Res 89:198–212
3. Nishio M, Hitomi J (2015) Sequential estimation of posterior distributions of model parameter of
bridge bearing under corrosion process. J Jpn Soc Civ Eng Ser A2 (Appl Mech) 71(2):I_99–I_108
(in Japanese)
4. Marrugo W et al (2014) A vision-based system for the dynamic measurement of in-plane
displacements. In: III international congress of engineering mechatronics and automation
(CIIMA), IEEE. Cartagena, Colombia, Oct 2014, pp 1–3
5. Dizaji MS, Alipour M, Harris DK (2018) Leveraging full-field measurement from 3D digital
image correlation for structural identification. Exp Mech 58(7):1049–1066
Wireless Strain Sensors for Dynamic
Stress Analysis of Rail Steel Structural
Integrity
X. P. Shi, Z. F. Liu, K. S. Tsang, H. J. Hoh, Y. Liu, Kelvin Tan Kaijun,

Phua Song Yong, Kelvin, and J. H. L. Pang
Abstract Fatigue fractures were observed at the web-to-head and web-to-foot

interface regions of rail steel thermite weld joint on curve tracks in a mass rapid
transit tunnel. To investigate these failures, onsite wireless strain measurement was
conducted to record the train-induced stress spectrum at a thermite weld joint and the
results were presented in this paper. Dynamic finite element (FE) analysis was then
conducted using the thermite weld and rail steel structural model under train-induced
loading conditions and is validated by the strain gage sensor measurements. The train-
induced stress at critical locations of thermite weld was investigated. Combining the
strain measurements and FE analysis results, the service load stress spectrum using
the variable amplitude strain analysis spectrum at the stress concentration site was
obtained and fatigue life for a critical location of the thermite weld was predicted
using the cumulative damage summation method. This research provides a fatigue
mechanics-based approach for rail steel asset life prediction methodology.
Keywords Strain sensors · Fatigue · Finite element · Structural integrity ·

Thermite weld
X. P. Shi · Z. F. Liu · K. S. Tsang · H. J. Hoh · Y. Liu · J. H. L. Pang (B)

SMRT-NTU Smart Urban Rail Corporate Laboratory, Nanyang Technological University, 50
Nanyang Avenue, Singapore 639798, Singapore
e-mail: mhlpang@ntu.edu.sg
X. P. Shi · H. J. Hoh
School of Electrical & Electronic Engineering, Nanyang Technological University, 50 Nanyang
Avenue, Singapore 639798, Singapore
Z. F. Liu · K. S. Tsang · Y. Liu · J. H. L. Pang
School of Mechanical & Aerospace Engineering, Nanyang Technological University, 50 Nanyang
Avenue, Singapore 639798, Singapore
K. T. Kaijun · P. S. Yong · Kelvin
SMRT Trains Pte. Ltd., Permanent Way, Singapore, Singapore

https://doi.org/10.1007/978-981-15-9199-0_44
472 X. P. Shi et al.
1 Introduction
Rail steel thermite welds are widely used in railway track systems. However, due
to the geometry irregularity induced by a thermite weld, stress concentrations at the
connection between weld and parent rail can lead to fatigue failures at the thermite
weld region. Cracks may initiate from these stress concentration sites and under
repeated loading induced by wheel rolling in and out, and these cracks may cause
weld failure. Some typical weld failures are shown in Fig. 1.
Most thermite weld failures occur at curved tracks where fatigue occurs at the web-
to-head and web-to-foot transition regions due to complex in-plane and out-of-plane
loading arising from the wheel forces applied to the rail track. To investigate these
failures, strain measurement at these locations were conducted on a curved track
during train operation hours. The measurements provided important information
about train loading such as train velocity and rail track operational stresses. The
rosette strain gages are installed at a close distance from the thermite weld toe to
parent metal rail transition region. Finite element (FE) modelling of the thermite
weld was used to determine the stresses and a stress scale factor was obtained from
strain measurement at several locations to weld failure locations.
Combining the strain measurements and FE analysis results, the service load
stress spectrum using the variable amplitude strain analysis for fatigue life prediction
was conducted. A rain-flow cycle counting analysis method was performed, and the
results were used in Miner’s Law, a cumulative damage summation fatigue failure
assessment approach. The fatigue life of the critical region on the thermite weld was
then predicted.
Fig. 1 Typical weld failure for rail thermite weld, a fracture initiates at head-web region in the
field side, b fracture initiates at foot-web region in the field side, c fracture initiates at head-web
region in the gauge side
Wireless Strain Sensors for Dynamic Stress Analysis of Rail … 473
2 Strain Measurement
Onsite strain measurement was conducted, on a 300 m radius track comprising of a

concrete slab in a metro tunnel. The wireless strain measurement system is shown in
Fig. 2. Strain gages are connected, to a wireless transmitter with Wheatstone bridge
circuits powered by rechargeable batteries. The transmitter was fixed on the wall
of the tunnel and the strain signal was transmitted to a wireless receiver which was
placed beyond a safety distance from the track and was connected to a laptop for
data post processing. The instrument set-up was shown in Fig. 3.
As the centrifugal force, mainly acted on the outside rail of the curve track, the
strain measurement was conducted only on the outside rail. Four rosette strain gages
were attached near the thermite weld at the web-to-head and web-to-foot transition
locations. The rosette strain elements in 0°, 90° and 45° direction were recorded as
shown in Fig. 4. For electrical shielding protection and EMI effect elimination, strain
gage areas were covered with aluminium foil-based tape.
The instrument set-up was completed before train operating hours in the early
morning engineering hours, and the system was then turned on before the first train
Fig. 2 Wireless strain measurement system
Fig. 3 Instrument set-up for strain measurement

Fig. 4 Strain measurement locations
300
Head-web 45o strain
Head-web Longitudinal strain
Head-web Vertical strain
200
Strain (με)
100
-100
-200
0 2 4 6 8 10
Time (s)
Fig. 5 Strain measurement result for multiple trains (left) and one train (right)
starts. The strain signals were automatically recorded. A recording frequency of

1024 Hz was used in order to accurately capture the peak strains induced by passing
trains. Typical strain signals recorded was shown in Fig. 5. On the left of Fig. 5, each
peak represents a train passing by the measurement locations. The first peak was
observed by zooming in to the strain spectrum as shown on the right of the figure.
There are 24 axles in total for a 6-car train. Hence, there are 24 strain peaks on each
channel shown on the right of Fig. 5. The speed of the train can also be obtained
from analysis of Fig. 5 (right) by dividing the physical distance ‘from the first axle
of the train to the last axle’ of the train by time difference ‘between the first peak
and the last peak of the strain signals’. The train shown in this figure is running at a
speed of 52 km/h derived from the results of strain measurement.
3 Numerical Simulation
As shown in Fig. 4, in the previous section, strain gages were attached some distance
from the exact interface of parent rail and weld. A numerical model was then devel-
oped to determine a stress scale factor from strain measurement at various locations
to weld fracture location.
Fig. 6 Numerical model of train-track system
Multibody dynamics method [1, 2] and finite element (FE) method [3–7] were
usually used in simulation of the train-track system. Multibody dynamics method is
able to simulate the dynamic effects due to the structural flexibility of the vehicle
components and the track and the computation scale can be quite large [8]. However,
due to rigid body assumption used in this method, the stress distribution in the track
and thermite weld is unlikely to be solved. FE method can solve complex structural
stress problem, but the calculation cost is high when the model scale is large.
In this research, an integrated approach was developed. The multibody dynamics
analysis was firstly conducted with the commercial software Universal Mechanisms.
All the components in the railway systems are considered in this multibody dynamic
model. The model is the same as used in Bai’s work [9]. The interaction force between
wheels and tracks was exported and was used in following FE models. The FEM
simulation was conducted using commercial FEM software ABAQUS. The whole
FE model consists of a wheel, rail track with weld, base plate, rubber pad, clips and
sleepers as shown in Fig. 6. The shape of thermite weld was constructed based on
a thermite weld specimen. The bottom of the sleepers as well as the longitudinal
movement of the rail track end were both fixed. General contact in ABAQUS was
used between wheel and rail track, also between rail track and base plate, rubber pad
and clips. The friction factor used for general contact in this model was assumed as
0.3.
The track length in this research is 4.9 m which can minimize the rail end effect
meanwhile reduces the calculation time. Similar length was used in other researchers’
work [5, 10, 11]. The wheel was modelled as rigid body for simplification. A separate
study showed the stress distribution in the regions below the railhead using this
simplification is only 1% less than that of deformable wheel. Bilinear elastic–plastic
material model was used for rail steel and linear elastic model was used for other
materials. The material parameters used in this simulation were listed in Table 1.
To eliminate numerical oscillation and shock impact in explicit simulation, an
implicit–explicit sequential approach was used. The wheel was located 0.45 m away
from one end of the rail and a vertical and lateral small displacement was applied
to establish wheel–rail contact in implicit analysis. The deformation and the stress
components of the rail system were then imported into input file for explicit analysis.
Similar implicit–explicit sequential approach was also used in literature [14, 15]. In
explicit analysis, the wheel–track reaction force obtained from UM was applied on
the reference node of the wheel, and the wheel was rolling towards moving direction
Table 1 Material parameters used in FE model

Parts Material Material properties
Density Elastic Poisson’s Yield stress Ultimate
(kg/m3 ) modulus ratio (MPa) stress
(MPa) (MPa)
Rail R350 steel 7850 210,000 0.3 830 1255
[12]
Sleeper Concrete 2500 38,000 0.15 – –
Clip and Steel 7850 210,000 0.3 – –
base plate
Rubber pad Rubber 920 2.29 0.47 – –
[13]
without slipping. The model was validated by strain measurement. Figure 7 shows the
comparison result of weld web-head location, outside of the rail. The simulation result
shows the same trend exists both in longitudinal strain and vertical strain compared to
test results. In Fig. 7a, both simulation and test result show a tension peak at ~1.36 s.
The same phenomenon has also been reported in other researcher’s work [16–18] and
this tension spike was considered the main reason for crack propagation at thermite
weld. The difference in strain peak value between FEM result and test result may be
due to irregular geometry of the test weld and rail constrain condition in FE model.
To determine stress scale factor from the location of ‘strain measurement’ to the
location of ‘weld fracture’, stress profile of two elements at these two locations was
compared and the scale factor for weld fracture location at the head-web corner is 3
in this example.
140 80
Test Test
120 Simulation
Simulation 60
100
80 40
Strain (με)
Strain (με)
60
20
40
20 0
0
-20
-20
-40 -40
-60
-80 -60
-100
1.30 1.32 1.34 1.36 1.38 1.40 1.42 1.44 1.30 1.32 1.34 1.36 1.38 1.40 1.42 1.44
Time (s) Time (s)
(a) (b)
Fig. 7 FEM result comparing with test result, a longitudinal strain, b vertical strain
4 Fatigue Life Prediction
The fatigue life of thermite weld at critical location was estimated using rain-flow
counting and damage accumulation method using the strain measurement and simu-
lation result. This section shows an example of estimation for fatigue life at weld
web-head location on the outside of the rail. Rain-flow counting is conducted using
a commercial software DATS analysis. The damage on the weld due to the train-
induced stresses on the track can be calculated using the previous/past strain measure-
ment results and the stress-life curve (S–N curve) of the material. The whole fatigue
life prediction process is shown in Fig. 8. A single train signal as shown in the Fig. 5
(right) was used for this calculation. The rain-flow matrix is shown in Table 2.
Using the rain-flow matrix, the damage accumulated can be calculated using the
stress-life curve and Miner’s rule. The Palmgren-Miner rule states that failure occurs
when
I
ni
=1 (1)
i=1
Ni
where n i is the number of applied load cycles of type i, and Ni is the fatigue life
under different stress range S. The S–N curve used in this paper is from literature
[19] where the crack growth life was predicted for an initial crack size of 0.025 mm
to a final crack size of 5 mm. The S–N equation related is
Fig. 8 Flowchart for fatigue life prediction at weld joint

478
Table 2 Damage accumulation

Stress (Pa) Mean −3.64E+06 −2.35E+06 −1.06E+06 2.23E+05 1.51E+06 2.80E+06 4.08E+06 5.37E+06 6.66E+06 7.94E+06
Range After 1 2 3 4 5 6 7 8 9 10 Total Expected ni /N i
scale (ni ) life (N i )
9.12E+06 2.74E+07 1 0 0 0 0 4 2 0 0 0 1 7 3.87E+08 1.8077E−08
1.17E+07 3.50E+07 2 0 0 0 0 1 1 1 0 0 3 6 1.70E+08 3.52022E−08
1.42E+07 4.27E+07 3 0 0 0 0 0 1 1 0 0 1 3 8.83E+07 3.39813E−08
1.68E+07 5.03E+07 4 0 0 0 0 0 0 0 0 0 0 0 5.10E+07 0
1.93E+07 5.80E+07 5 0 0 1 0 0 0 0 0 0 0 1 3.18E+07 3.1422E−08
2.19E+07 6.56E+07 6 0 0 1 1 0 0 0 0 0 0 2 2.11E+07 9.4954E−08
2.44E+07 7.33E+07 7 0 2 4 0 0 0 0 0 0 0 6 1.46E+07 4.11241E−07
2.70E+07 8.09E+07 8 1 4 0 0 0 0 0 0 0 0 5 1.05E+07 4.77011E−07
2.95E+07 8.86E+07 9 2 1 0 0 0 0 0 0 0 0 3 7.76E+06 3.86643E−07
3.21E+07 9.62E+07 10 1 2 1 0 0 0 0 0 0 0 4 5.89E+06 6.79285E−07
Total 2.16782E−06
X. P. Shi et al.
−3.33
S
N= (2)
10383
The damage accumulated is then calculated as shown in Table 2. The total damage
caused per train is 2.17e−6. The expected life can be calculated as 1/total damage,
which is 461,294 trains. Assuming there are 300 trains per days, the weld fatigue
life is 1538 days (~4 years). It should be noted that as only one S–N curve was used,
effect of mean stress was not considered in this section. The paper only presents
an integrated approach to predict the fatigue life of thermite weld. More accurate
prediction can be achieved by considering more details of material properties.
5 Conclusions
In this research, onsite wireless strain measurement was conducted on a curved track
in a tunnel. The structural strain spectrum of train-induced loading at web-head
and web-foot regions of thermite weld was obtained. Numerical model was then
developed to calculate the stress scale factor. Finally, the fatigue life of thermite weld
at critical locations was predicted using rain-flow counting method under cumulative
damage summations by combining strain measurement and simulation results.
Acknowledgements This research work was conducted in the SMRT-NTU Smart Urban Rail
Corporate Laboratory with funding support from the National Research Foundation (NRF), SMRT
and Nanyang Technological University; under the Corp Lab@University Scheme.
References
1. Sharma SK (2018) Multibody analysis of longitudinal train dynamics on the passenger ride
performance due to brake application. Proc Inst Mech Eng Part K: J Multi-body Dyn 1–14
2. Ding GF, He Y, Zou YS, Li R, Ma XJ, Qin SF (2015) Multi-body dynamics-based sensitivity
analysis for a railway vehicle. Proc Inst Mech Eng Part F-J Rail Rapid Transit 229(5):518–529
3. Ranjha SA, Mutton P, Kapoor A (2014) Fatigue analysis of the rail underhead radius under
high axle load conditions. Adv Mater Res 891–892:1181–1187
4. Ranjha S, Kulkarni A, Ekambaram P, Mutton P, Kapoor A (2016) Finite element prediction of
the stress state of a rail underhead radius under high axle load conditions. Int J Cond Monit
Diagn Eng Manage 19(2):11–16
5. Salehi I, Kapoor A, Mutton P, Mutton P, Alserda J (2010) Improving the reliability of
aluminothermic rail welds under high axle load conditions. In: CORE 2010, rail—rejuvenation
and renaissance, conference on railway engineering, Wellington, New Zealand
6. Ranjha S, Ding K, Mutton P, Kapoor A (2012) Mechanical state of rail underhead region under
heavy haul operations. In: CORE 2012, global perspectives, conference on railway engineering,
Brisbane, Queensland
7. Salehi I, Kapoor A, Mutton P (2011) Multi-axial fatigue analysis of aluminothermic rail welds
under high axle load conditions. Int J Fatigue 33(9):1324–1336
8. Shabana AA, Sany JR (2001) A survey of rail vehicle track simulations and flexible multibody
dynamics. Nonlinear Dyn 26(2):179–210
9. Bai L, Li B, Shen F, Ng WY, Yusoff MFBM, Lim ST, Zhou K (2018) An integrated approach
for predicting wear and fatigue damage of wheels and rails in SMRT systems. In: International
conference on intelligent rail transportation, Singapore
10. Desimone H, Beretta S (2006) Mechanisms of mixed mode fatigue crack propagation at rail
butt-welds. Int J Fatigue 28(5):635–642
11. Ranjha SA, Ding K, Mutton PJ, Kapoor A (2012) Finite element modelling of the rail gauge
corner and underhead radius stresses under heavy axle load conditions. Proc Inst Mech Eng
Part F: J Rail Rapid Transit 226(3):318–330
12. Motameni A, Eraslan AN (2016) Fracture and fatigue crack growth characterization of
conventional and head hardened railway rail steels. J Sci Eng Res 3(3):367–376
13. Kaewunruen S, Lewandrowski T, Chamniprasart K (2018) Dynamic responses of interspersed
railway tracks to moving train loads. Int J Struct Stab Dyn 18(01):1850011
14. Yang Z, Boogaard A, Wei ZL, Liu JZ, Dollevoet R, Li ZL (2018) Numerical study of wheel-rail
impact contact solutions at an insulated rail joint. Int J Mech Sci 138:310–322
15. Zong N, Wexler D, Dhanasekar M (2013) Structural and material characterisation of insulated
rail joints. Electron J Struct Eng 13(2):75–87
16. Eisenmann J (1967) Stresses acting on the rail—recent findings. Brussels, Belgium, Interna-
tional Railway Congress Association, pp 537–550
17. Sugiyama T, Yamazani T, Umeda S (1971) Measurement of local stress on outer rail head. Q
Rep RTRI 11–13
18. Mutton PJ, Tan M, Bartle P, Kapoor A (2009) The effect of severe head wear on rolling contact
fatigue in heavy haul operations
19. Lawrence FV, Ross ET, Barkan C (2004) Reliability of improved thermite welds
Purpose and Practical Capabilities
of the Metal Magnetic Memory Method
A. A. Dubov, S. M. Kolokolnikov, M. N. Baharin, M. F. M. Yunoh,

and N. M. Roslin
Abstract This paper presents the purpose and practical capabilities of metal
magnetic memory (MMM) methods. MMM is an advanced Non-destructive Testing
(NDT) method developed through accumulation of Self-Magnetic Leakage Field
(SMLF) and its distribution analysis in Stress Concentration Zones (SCZs) area. The
SCZs is a major source of damages mechanism development and can be detected
in express-control mode using specified MMM sensor, scanning devices, and soft-
ware. It can be performed without require surface preparation and artificial magne-
tization where it utilizes residual magnetization formed naturally during product
fabrication and throughout the operation. MMM method when combined with other
NDT methods such as ultrasound dramatically improves inspection result efficiency.
Presently, MMM methods are widely employed in base metal and weld joints inspec-
tion. For inaccessible objects such as buried pipelines, Non-Contact Magnetometric
Diagnostics (NCMD) will be employed with the same fundamentals used by the
MMM method. Therefore, it is proposed that MMM is best suited for SCZs detection
in feasible diagnosis of equipment and structure and at once enables the assessment
of actual stress–strain state to be carried out for all structural elements.
Keywords Non-destructive testing · Inspection · Metal magnetic memory · Stress

concentration zones · Non-contact magnetometric diagnostic
1 Introduction
Metal magnetic memory (MMM) is an advance NDT method developed through

accumulation of SMLF and its distribution analysis in SCZs area [1]. In reference
to ISO 24497-1-2009, it is an after-effect that occurs as residual magnetization in
A. A. Dubov · S. M. Kolokolnikov
Energodiagnostica Co. Ltd., Moscow, Russia
M. N. Baharin (B) · M. F. M. Yunoh · N. M. Roslin
Technology Department, MR Technology Sdn. Bhd., 4809-1-30, Jalan Perdana 2, 63000
Cyberjaya, Selangor, Malaysia
e-mail: baharin@mrtsb.com.my

https://doi.org/10.1007/978-981-15-9199-0_45
482 A. A. Dubov et al.
the parent metal or welded joint formed during the fabrication or due to irreversible
change of the local magnetization state of inspection object in zones of stress concen-
tration or damage under loadings. SMLF that reflects the natural residual magneti-
zation during the fabrication process is principally different compared to Magnetic
Leakage Field (MLF) occurring on anomalies found in inspected object through
artificial magnetization applied in other magnetic testing method such as Magnetic
Particle Inspection (MPI) [2].
MMM method can be performed without requirement of surface preparation such
as insulation or coating removal as well as stimulating artificial magnetization into
the inspected object whereby it utilizes the natural generated residual magnetization
formed during object manufacturing or object under pressurized condition. The main
objective of conducting MMM is to locate SCZs on inspected object which indicates
early stages of damage mechanism in express-control mode by using specified sensor
scanning devices and software.
2 Stress Concentration Zones in Metal Magnetic Memory

Method
The SCZs is not only a predefined area where it creates dissimilar conditions for stress
distribution produced by external loads, but then it is also recognized as a random area
due to the initial metal heterogeneity occurrence related with the addition workload
of large structural exterior design.
The geometric characteristic of SCZs magnetic anomalies is the distance between
self-magnetic field extreme values multiple with the standard product size (thickness,
width, diameter). This distance is the minimum distance between the adjacent glide
pads or the critical size of the shell, for instance, the loss of stability in the pipe.
For quantitative evaluation of the level of stress concentration (sources of damage),
the gradient of the normal (H y ) and/or tangential (H x ) SMLF elements is as shown
below [2].

Hy dH
K in = , at x → 0 K in = (1)
x dx
where x is the distance between the adjacent points of inspection [2]. In some
cases, during the inspection of equipment Stress–Strain State (SSS) the resulting
SMLF gradient is used as below [2]:

|H | = Hx2 + Hy2 + Hz2 (2)
Numerous studies done at the manufacturing plant have shown that all metallic
object of the same category, made of the same material grade and under the same
manufacturing technology, have a similar Residual Magnetization (RM) distribution
Purpose and Practical Capabilities of the Metal Magnetic … 483
where the magnetic anomalies are only detected during inspection at residual stress
concentration areas and different structural distortions in individual metallic object
[3]. For example, during the magnetization of the thermodynamic products in their
production process, the internal pressure, rather than the weak external geomagnetic
field, plays a decisive role. During product operation, the RM is first distributed
under the impact of the workload, and magnetic anomalies, due to the geometric
displacement and standard size of the product, occur in the SCZs.
If the local SCZs does not occur in a metallic object of the same category under
the influence of the work load, the RM distribution pattern is almost the same. In
ensuring that, it is required to check thousands of units and products from the same
type. Based on established principles and extensive practice in examining numerous
units of equipment and structures, the authors propose less standardized calibration
methodologies and inspection approaches, as well as their corresponding metrology.
3 Integration of MMM Method with Other NDT Methods
It has been more than 10 decades since welding activity was introduced into indus-
tries and the most important factor in determining the reliability of weld joints is the
distribution of residual stresses. There are various NDT methods used to examining
the welding condition and MMM is best utilized to perform the comparative measure-
ment on the effect of the relationship between residual stress and the properties of
the material.
Incorporating inspection result utilizing MMM with other NDT methods will
dramatically improve the examination efficiency. Currently, MMM method is widely
employed in multipart inspection of base metal and welded joints. By using express-
control mode in MMM method, the SCZs can be detected on the Inspection Object
(IO) without any surface preparation required at where they will be classed by SMLF
gradient and the design parameter, m afterward. For verification purposes, specified
SCZs will furthermore inspected by other NDT methods such as Ultrasonic Testing
(UT).
For inspection conducted using MMM method on specific equipment unit/IO,
the development of techniques in classifying magnetic anomalies of the size on the
surface and internal section of metallic object are currently progressing. According
to some sources, in many industries, standards for unacceptable anomalies resulted
from NDT are different for same IOs. Moreover, the size of acceptable and unaccept-
able anomalies in existing guidance documents cannot be physically proven from
a mechanical point of view. On this concern, it shall be understood that a signif-
icant attribute of the MMM technique is that it evaluates the anomalies’ severity
and presume their direction and intensity by using magnetic parameters in the SCZs
which already contain macro cracks of unacceptable sizes.
In the event of increased number of damage mechanism in aging operating
metal equipment having unexpected fatigue properties, the proposed MMM method
primarily for early screening and analysis of the anomalies has obviously beneficial
asset owner over other NDT methods.
4 Development Capabilities of the MMM Method
Today, more and more experts are focusing on the development of technical analysis,
lifetime assessments and risks, and condition monitoring of equipment/IO. Nonethe-
less, the “catastrophe” of the scenario is that when evaluating the industrial conditions
of Hazardous Facilities (HIFs), experts are facing the fact that existing guidelines
do not enclose a strong definition of the concept of mechanical fracture and material
science standpoint. Some research papers provide a perspective on contemporary
knowledge of fracture mechanics as the explanation of such concepts as “limiting
metal conditions” within the local SCZs structure and “limiting the structural prop-
erties of the structure itself.” In practice, according to the above concept, the MMM
method allows the assessment of the actual condition of the equipment and structures
[4].
In addition, in the field of metal and structural element fracture mechanics and
materials science, the development of new standards or guidance documents on
technical analysis, monitoring and lifetime HIFs and risk assessment is also crucial.
In some possible circumstances, utilizing MMM method to perform 100% coverage
of inspections object and to recognize all potentially hazardous zones, suspected
damage growth, lifecycle assessments and risks will be more precise and predictable.
From the author’s point of view, MMM is currently the most suitable and practical
method for SCZs detection in structures. Simultaneously, this method enables actual
stress–strain state evaluation of total structural elements in the express-control mode
based on 100% inspection coverage to be performed.
On the other hand, the Non-Contact Magnetometric Diagnostics (NCMD) capa-
bilities of unpiggable pipeline sections are also unique. Fundamentals of NCMD are
similar with MMM method during contact inspection. NCMD recognize magnetic
anomalies in geomagnetic field distributions caused by variations in magnetization of
metal pipeline in SCZs and in corrosion-fatigue growth areas. Inspection conducted
for buried (or underwater) pipelines utilizing NCMD must be able to answer to the
question of: “When and where will the damage or accident be expected?”. In order
to solve the issues, NCMD applications will be combined with additional inspection
methods for pipelines such as UT, eddy currents, etc., since presently it is unfeasible
to detect and classify anomalies in the pipeline based solely on NCMD results.
Figure 1 shows an example of SCZs location based on NCMD data analysis.
Figure 2a shows an example of UT results based on the location of NCMD SCZs
data. Longitudinal cracks were found at the longitudinal stitch welding as shown in
Fig. 2b. The defect was found by performing Ultrasonic Test Flaw Detection (UTFD)
at high SCZs.
However, it should be taken into concern that if, according to NCMD results,
auxiliary inspections detect anomalies in only a few SCZs from many SCZs detected,
SCZ Locaon
Fig. 1 Example of SCZs location of NCMD signal
the outcome is quite efficient because it is considered that even if an accident is

prevented, for instance on a gas or oil pipeline, all expenses of the NCMD carried
out are covered.
5 Conclusion
As a conclusion, mechanical express-control device for metal anomalies and local

SCZs detection provides the advantages to industrial users as described below;
• Provides early indication of corrosion-fatigue and lifespan evaluation of equip-
ment and structures;
• Detect anomalies/defects (lamination, casting defects, etc.) in the inner metal
layer by using SMLF geometry parameters controlled by glide pads dislocations
in SCZs;
• 100% coverage of IO inspection to detect local SCZs where source of damage
mechanism is growing;
• Improve the efficiency of the IO inspection program through combination of the
MMM method and other significant NDT methods;
• Reduce cost of performing inspection as it discards the use of artificial magne-
tization through other magnetic testing method and does not requires surface
preparation such as insulation removal.
Fig. 2 a Example of UT result based on SCZs location of NCMD data, b defect on the inspection
object
Acknowledgements The authors would like to express his sincere thanks and gratitude especially
to Energodiagnostica Co. Ltd., Moscow and MR Technology Sdn. Bhd. and everyone throughout
accomplishing this paper.
References
1. Dubov A, Dobov A, Ladanyi P (2014) Tensile testing of steel specimens using the metal magnetic
memory method. In: 11th European conference on non-destructive testing (ECNDT 2014),
Prague, pp 1–6
2. Kolokolnikov S, Dobov A (2018) Physical basics, practical capabilities and purposes of the metal
magnetic memory method application for technical diagnostic of critical industrial equipment.
NDT for safety, Prague, pp 49–55

3. Dubov A, Vstovsky GV (2000) Physical base of the method of metal magnetic memory. In: 2nd
international conference on metal magnetic memory, Moscow
4. Gorkunov ES, Efimov AG, Shubochkin AE, Artemyev BV (2016) On the question of magnetic
NDT application to determine the stress-strain state of metal structures. World NDT 3:52–55
Development and Application of Inner
Detector in Drilling Riser Auxiliary Lines
Based on Magnetic Memory Testing
Xiangyuan Liu, Jianchun Fan, Wei Zhou, and Shujie Liu
Abstract Drilling riser auxiliary lines are critical components of deep-water floating
drilling platform, which mainly consist of choke line, kill line, hydraulic line, boost
line, etc. This paper discusses an inner detector in drilling riser auxiliary lines based on
metal magnetic memory testing. In contrast to conventional detect methods currently
employed, the prominent advantage of the detector is that it can scan the inner
wall directly by the pull of a winch, with no need to disassemble the buoyancy
modules. This way, the time and cost of testing would be shortened considerably.
The detector is designed with two sensor arrays assembled in staggered pattern,
and each array consists of six tunnel magnetoresistance probes radially arranged
from the canter of the shaft, much like spokes on a wagon wheel, ensuring 100%
coverage of the inner wall. The raw data acquired from the front and rear sensor arrays
can be aligned automatically by the alignment algorithms of imaging software for
further processing. For determination of the detector characteristics, the laboratory
test was conducted with 3.5 tubing that were modified with a series of prefabricated
groove and dome-shaped artificial defects. And result shows that it can present the
damage sharply. Besides, field test also confirms the design has features of reasonable
structure and stable performance.
Keywords Riser auxiliary lines · Inner detector · Imaging software · Magnetic

memory
X. Liu · J. Fan (B) · W. Zhou

College of Safety and Ocean Engineering, China University of Petroleum-Beijing, Changping,
Beijing, China
e-mail: fjc688@126.com
X. Liu
e-mail: xiangyuan.liu@qq.com
S. Liu
CNOOC Research Institute, Beijing, China

https://doi.org/10.1007/978-981-15-9199-0_46
490 X. Liu et al.
1 Introduction
In deep-water floating drilling, the drilling riser system is an indispensable equip-

ment. It refers to a section of pipe from the subsea blowout preventer to the moonpool.
Its main function is to isolate the seawater, guide the drilling tool, circulate the drilling
fluid, and trip sunsea BOP. Drilling riser auxiliary lines are critical components of
this system, which mainly consist of choke line, kill line, hydraulic line, boost line,
and so on [1]. They enable the marine riser system to keep the pressure in the well
balanced and control the subsea blowout preventer properly. These pipelines are
subjected to high pressure during the working process, and the fluid medium therein
is corrosive. The inner wall tends to form cracks, corrosion pits, and the like under
the combined action of corrosion, erosion, and stress concentration. According to
the statistics of the literature, the total downtime of the BOP of the 47 development
wells drilled by the semi-submersible platform from 1987 to 1989 was 627 h due to
the failure of the choke line and kill line, accounting for 29% of the total downtime
[2]. It can be seen that the failure of choke line and kill line on the drilling riser is
one of the main reasons for the shutdown of the floating drilling unit. Therefore,
regularly detecting the damage status of these auxiliary pipelines and ensuring their
safe operation are the key to successful drilling and completion quality. According
to industry regulations, the riser is inspected annually.
At present, conventional non-destructive testing techniques such as magnetic
powder, penetrant testing, and ultrasonic are commonly used to detect crack defects
and wall thickness of the pipe body (as shown in Fig. 1). All those methods require
the removal of the buoyancy block, which brings about problems such as long oper-
ating cycle, low efficiency, and high cost. Moreover, since the inner diameter of the
auxiliary pipeline is small and the length is up to 22.86 m, the detection process
of ultrasonic testing and permeation detection have to be adopted in the form of
sampling inspection on the oil tools base, which makes the problem of miss detec-
tion cannot be avoided. Therefore, it is of great significance to develop a detection
device being capable of performing full-line scanning detection on the riser auxiliary
pipeline.
Fig. 1 Sample inspections on the base (ultrasonic testing left-hand side, penetrant testing right-hand
side)
Development and Application of Inner Detector in Drilling Riser … 491
What is known is that magnetic flux leakage detection is a well-recognized detec-

tion method for long-distance oil and gas pipelines, but magnetic flux leakage testing
requires specialized magnetization equipment to magnetize the pipeline [3], while
the auxiliary pipeline has an inner diameter ranging from 2 to 4 in. It is very difficult
to design the magnetization module in such a small inner diameter. But the metal
magnetic memory detection technology proposed in the 1990s does not need to
consider this problem, because what it detects is the residual magnetic field induced
by the stress and the geomagnetic field with no need to magnetize the detected object,
and it has unique advantages in the early diagnosis of fatigue damage [4]. This paper
introduces magnetic memory detection into the inner detection of the drilling riser
auxiliary lines. An inner detector is designed to detect the pipeline directly without
removing the buoyancy block. The two-dimensional appearance of flaws can be
presented intuitively by the imaging software.
2 Design of the Inner Detector
2.1 Analysis of Design Requirement
The magnetic memory method is a non-destructive testing method based on the prin-
ciple of recording and analyzing the distribution of the own leakage field generated
by the stress concentration region of the ferromagnetic member. The signal picked up
by the detection is the magnetic field distribution on the surface of the ferromagnetic
member, which requires magnetic memory probe fits closely against the surface of
the component during the inspection process. Each riser typically has 4 or 5 auxil-
iary lines, and the inner diameter specifications of these auxiliary lines are varied by
different drilling tool bases, ranging from 2 to 4 in. Therefore, the detector should
have a variable diameter to meet the need of detection of different lines.
2.2 Details of the Design
The inner detector designed in this paper consists of four main modules: central shaft,
sensor boxes, screw-spring support frame, and aviation plug.
As shown in Fig. 2, the front end of the center shaft is designed with a hook for
attaching to the winch wire rope. The detector completes the scanning inside the line
by the pulling of the winch. The sensor box is packaged with a giant magnetoresistive
sensor, which can effectively pick up the tangential magnetic field signal on the
surface of the ferromagnetic component. The six sensor boxes are arranged radially
along the central axis, forming a set of spoke-like probe arrays. Screw-spring frame is
a key component of the equipment. By tightening or loosening the screw, the external
492 X. Liu et al.
Fig. 2 Design drawing of the inner detector
diameter radiated from the six sensor boxes can be adjusted within a certain range
to adapt to the internal detection of lines of different diameters.
A total of two sets of probe arrays assembled in staggered pattern are equipped with
the central axis. Figure 3 shows the front view of two probe arrays which consist of
12 channels sensors. Since the scanned portions of the two sets of arrays complement
each other, a sufficient magnetic field signal of inner wall can be obtained in one test,
thereby achieving full coverage detection in high efficiency. The tail of the central
shaft is designed with an aviation plug interface. The power and signal lines of the
12 sensors in the two probe arrays are routed through the wire slot of the central
axis to the aviation plug. The detector can be connected to the data acquisition box
by means of aviation plug to power the probe and transport data. These good wire
arrangements enable the detectors to be of good performance in maintenance.
Fig. 3 The front view of two

probe arrays consist of 12
channels sensors
Fig. 4 Etch pits (design drawing left-hand side, picture right-hand side)
Fig. 5 Crack defects (design of the groove artificial defects above, picture of tube with defects
below)
3 Laboratory Test
3.1 Artificial Flaws
For determination of the detector characteristics, two sections of 3.5 tube were
modified with artificial defects to conduct laboratory test.
Sample 1: A ball-shaped cutter with diameter of 5 mm was used in order to mill
a series of defects in analogy to etch pits to a depth of 2 and 6 mm into the 3.5 tube.
Sample 2: 3.5 tube with 13 items of grooves produced with different depths from
0.5 to 6.5 mm by wire-electrode cutting in analogy to cracks.
Figures 4 and 5 show the design drawings of the artificial defects and the detailed
pictures of tubes with artificial defects.
3.2 Data Processing and Result
Export data from the acquisition software and load it into the data imaging software.
After the data denoising, the software needs to align the data of 12 channels. In order
494 X. Liu et al.
to cover more detection areas in one test, the detectors are staggered with two front
and rear sensor arrays, so that the data detected by the previous probe array is a bit
ahead of the data of the next probe array. Therefore, the data of the two arrays must
be aligned for further gradient processing.
The signal detected by the sensor is the tangential magnetic field of the wall of
the auxiliary line, and the magnetic field of the wall is not uniform itself, so the
original signal is not able to well characterize the defect. Under the circumstances
the magnetic field needs to be gradient processed to eliminate the influence of the
unevenness of the magnetic field. The gradient becomes larger due to the magnetic
field mutation near the defect, and this feature can be used to characterize the defect.
The last step is to draw a two-dimensional cloud image. Gradient data of 12
sensors can be drawn in the form of a cloud image to realize data visualization.
Figure 6 shows that six round-hole defects are detected by the test. And the two-
dimensional contour distortion of the circular hole is within an acceptable range,
which indicates that the inner detector can be applied to the detection of etch pits.
Figure 7 shows the section result of the 3.5 tube with groove defects. 13 items
of prefabricated groove artificial defects are totally presented on the cloud image
sharply. To get better visual effect, the cloud image can be attached to the surface of
a cylinder as Fig. 8 shows.
Fig. 6 Contrast between picture of tube with artificial defects and cloud image of detection
Fig. 7 Cloud image of the groove artificial defects’ detection
Fig. 8 The surface of a cylinder attached with cloud image
Laboratory test indicates that the inner detector can present the prefabricated
groove and dome-shaped artificial defects effectively.
4 Industrial Application
To verify the reliability of the inner detector, industrial application has been carried
out in a drilling base. The auxiliary lines in this base have an inner diameter of 2.5
and 3.5 in.
The onsite inspection requires two students to cooperate with each other. As is
shown in Fig. 9: One is responsible for manipulating the winch, pulling the inner
detector forward in the auxiliary line, and the other operating the computer at the
other end to collect data and assist in conveying the cable connected to the back of
the detector by aviation plug.
After the detection, the data can be exported, and the developed data imaging
software can be used to figure two-dimensional imaging of the inner wall condition,
which is an unfolded view of the inner wall. Considering the length of the auxiliary
line is 22.86 m and its diameter is 3.5 in., the aspect ratio of the image should be
496 X. Liu et al.
Fig. 9 Industrial application of the inner detector
Fig. 10 Data imaging result
kept at 22.86: 3.5 * Pi to ensure that the contour of the defect will not be distorted.
However, because the aspect ratio is too high, in order to avoid the image being too
narrow to be viewed, the imaging software is designed to automatically cut the image
into 4 segments.
It can be seen from the data imaging result (Fig. 10) that there are several pit defects
at about 3, 13, and 22 m in this auxiliary line. In fact, as is shown in Fig. 11, the
ends of the auxiliary lines on the base are seriously corroded. The result of industrial
application shows that the inner detector has features of reasonable structure and
stable performance.
5 Conclusion and Perspective
A novel inner detector based on magnetic memory testing has been developed for the
detection of corrosion damage in drilling riser auxiliary lines. It can clearly display
the 2D contour of the defect with the aid of supporting imaging software. Due to
the high speed of data processing, it is suitable for full-line scanning of auxiliary
Fig. 11 Photographs of the auxiliary lines
lines on the drilling tools base. Its test results are instructive for the re-inspection of
lines by other means of detection, which can transform the current random sampling
detection method into targeted detection, enhancing the accuracy and efficiency of
detection.
The downside is that the advancement of the detector in the lines relies on the
pulling of the winch. It needs to pass the wire rope through the entire auxiliary line in
advance, which is a heavy workload. What is worse, the section of cable connected to
the back of the detector by aviation plug will be contaminated after passing through
the inner wall of the auxiliary line, affecting the subsequent computer operation of
the inspector. Therefore, wireless transmission should be considered in subsequent
studies. In addition, it should be considered to use the endoscope to take pictures of
damage, such as corrosion or cracks on the inner wall, and match them with the two-
dimensional contour map formed by the magnetic memory test data. After match and
confirmation, the damage map library can be established, which lays a foundation
for the future research on the intelligent identification of defects.
Acknowledgements This work was financially supported by the National Key Research and
Development Program of China (Grant No. 2017YFC0804503).
Reference
1. Zhou S, Liu Q (2016) Introduction. In: Theory and application research of deepwater drilling
riser system mechanical behavior. Beijing, China, Science Press, p 5
2. Montgomery ME (1998) Choke and kill lines—safety and downtime. In: Houston IADC
deepwater conference, Brookshire, Texas
3. Shi Y et al (2015) Theory and application of magnetic flux leakage pipeline detection. Sensors
13(9):31036–31055
4. Yao K, Wu LB, Wang YS (2019) Nondestructive evaluation of contact damage of ferromagnetic
materials based on metal magnetic memory method. Exp Tech 43(3):273–285
Defect Recognition and Classification
Techniques for Multi-layer Tubular
Structures in Oil and Gas Wells by Using
Pulsed Eddy Current Testing
Hang Jin, Hua Huang, Zhaohe Yang, Tianyu Ding, Pingjie Huang,
Banteng Liu, and Dibo Hou
Abstract Pulsed eddy current (PEC) method has attracted researchers’ attention in
the detection of multi-layer tubular structure defects in oil and gas wells because of
its advantages of non-destructive testing, not being affected by non-ferromagnetic
interference. In this paper, the detection experiment of the casing with prefabricated
defects is carried out, and the PEC signal analysis method for defect recognition and
classification of multi-layer tubular structures is studied. To deal with the problem
that it is difficult to extract time-domain features, a modified principal component
analysis method is proposed. Combining the statistical features extracted by the
method and frequency-domain features, the combined features are obtained, and the
random forest algorithm is introduced to classify the defects. This paper demonstrates
an effective method for the detection of multi-layer tubular structure defects in oil
and gas wells.
Keywords Pulsed eddy current testing · Multi-layer tubular structure · Combined

feature · Random forest
1 Introduction
In the field of oil exploitation, the well casings, which are responsible for protecting
tubing and providing a safe acquisition environment, are typical multi-layer tubular
structures [1]. However, in the actual production process, casing wall may suffer from
perforation cracks, bending deformation, multi-surface extrusion and other defects
H. Jin · Z. Yang · T. Ding · P. Huang (B) · D. Hou

State Key Laboratory of Industrial Control Technology, College of Control Science and
Engineering, Zhejiang University, Hangzhou 310027, Zhejiang, China
e-mail: huangpingjie@zju.edu.cn; hhh_box@163.com
H. Huang
Zhongyuan Oilfield Company, SINOPEC, Puyang 457001, Henan, China
B. Liu
College of Information Technology, Zhejiang Shuren University, Hangzhou 310015, Zhejiang,
China

https://doi.org/10.1007/978-981-15-9199-0_47
500 H. Jin et al.
due to the long-term influence of harsh working environment [2]. If the casing defects
are not dealt with timely and pertinently, it will affect the normal production of oil
and gas wells and cause huge economic losses.
Pulse eddy current testing is a non-destructive testing technology developed on the
basis of traditional eddy current testing technology [3]. Because of its advantages
of non-destructive testing, non-ferromagnetic interference, non-contact, and high
detection efficiency, it has been applied in the detection of defects of multi-layer
tubular structures in oil and gas wells. Zhang et al. realized the classification of
casing outer wall corrosion, inner wall corrosion, and outer wall corrosion based on
the stacked auto-encoder neural network [4]. Wang et al. conducted experimental
evaluation on the defects and ovality of coiled tubing and realized the localization,
qualitative and quantitative analysis of defects [5]. Xu et al. proposed a composite
nondestructive testing method based on PEC method and magnetic flux leakage
method and realized longitudinal crack, internal defect and transverse crack detection
of tubing [6]. At present, the research on multi-layer tubular is mainly aimed at
identifying special defects, but few studies on the classification of casing defects that
may occur in actual situations.
Due to the limitation of signal sampling in the detection process, it is difficult to
extract typical time-domain features to effectively characterize the defect types [4].
Therefore, a method of defect recognition and classification for multi-layer tubular
structures based on combined features is proposed in this paper. Firstly, based on
the statistical features extracted by the modified PCA method, the casing defects are
recognition. Then the combined features are obtained by combining the statistical
features with signal frequency-domain features. Based on the combined features,
the random forest algorithm is used to construct the data model of the relationship
between the PEC signals and defect types, and the defect classification of multi-layer
tubular structures is realized.
2 Experiment System and Specimen
2.1 Experiment System
The experimental setup is shown in Fig. 1. In the experimental system, the excitation
signal module generates the pulse wave signal and drives the excitation coil in the
compound detection probe to generate the excitation magnetic field, which is sensed
by the detection coil. The detection signal is transmitted to the signal processing
center by the signal acquisition module for storage and processing.
Figure 2 shows the structure of the PECT instrument, which mainly consists of
excitation signal module, composite detection probe, and signal acquisition module.
The instrument will place the vertical centers in the multi-layer tubular structures
for detection. The composite detection probe is composed of longitudinal probe A,
transverse probe B, and transverse probe C. The three probes are perpendicular to
Defect Recognition and Classification Techniques … 501
Fig. 1 PEC detection system
Fig. 2 Schematic diagram of PEC instrument
each other and the electromagnetic information of specimens in three-dimensional

space can be obtained. The auxiliary module includes excitation signal generating
circuit, detection signal processing circuit, coding transmission circuit, and so on,
which is used for signal transmission and processing.
2.2 Tubular Structures Defect Specimens
In the experiment, according to the actual defects in an oil field, standard defects
such as perforation crack, bending deformation, multi-surface extrusion, and so on
are defined and processed corresponding experimental specimens for low carbon steel
casing, as shown in Fig. 3. Figure 4 shows partial processed experimental specimens.
The nonlinear relationship between the mentioned casing defect types and PEC
signals and the difficulty in extracting typical time-domain features of signals to char-
acterize defects make it difficult to classify effectively [7]. Therefore, it is necessary
502 H. Jin et al.
Fig. 3 Schematic diagram

of some experimental
specimens
Fig. 4 Partial processed experimental specimens
to introduce other methods to describe defect information more comprehensively

and classify the defects [8].
3 Theory
3.1 PEC Detection Principle
PEC excites the square wave signal with constant period, amplitude, and duty ratio as
the excitation signal and induce the primary magnetic field. The change of magnetic
field will induce eddy currents in the specimen, and different defects affect the distri-
bution and intensity of eddy current, ultimately affect the total magnetic field in the
space. By analyzing the changes of detection signal, the information of the specimen
can be obtained to realize qualitative and quantitative analyse of defects [9].
The casing material used in this experiment is low carbon steel, which belongs
to the ferromagnetic material. Figure 5 shows a typical PEC response signal of
ferromagnetic materials. The secondary field of ferromagnetic material is in the same
direction as the primary field, so there is no zero-crossing time. Affected by “smoke
ring effect” [10] caused by the relatively large space size of casing, the detection is
difficult to obtain typical time-domain features such as peak time or voltage peak.
Fig. 5 PEC response signal

of ferromagnetic materials
3.2 Principal Component Analysis
Principal component analysis (PCA) is a signal dimension reduction processing tech-

nology based on the statistical analysis method. By converting multiple variables into
a few integrated variables (principal components, PC), each PC is a linear combi-
nation of the original variables and the PCs reflect most information of the original
signal. In this paper, a modified PCA method is proposed. The features are extracted
by PCA separately based on the direction of probes and combined as statistical
features.
3.3 Random Forest
Random forest algorithm (RF) is an improved decision tree method based on bagging
sampling. The basic idea is to construct a large number of classifiers to avoid over-
fitting in a single decision tree. RF will abandon the results of trees which have no
influence on prediction, adopt the results of trees which have strong influences, and
get the final results by weighted fusion. Therefore, RF has strong anti-noise ability
and prediction ability based on similar features.
In this paper, combined features are obtained by combining the frequency-domain
features of PEC signals and the statistical features extracted by the modified PCA
method. Based on combined features, a model is constructed by RF algorithm to
classify the defects of multi-layer tubular structures. The overall technology diagram
is shown in Fig. 6.
504 H. Jin et al.
Fig. 6 Diagram for defect classification
4 Results and Analysis
For non-destructive testing and evaluation of the double-layer tubular structures

commonly used in actual wells, namely the experimental scenario of tubing +
casing, casing defects are recognized first, and then classified based on defective
response signals. 1774 sets of PEC response signal were collected, including six
different categories: defect-free, perforation cracks (Defect A), single-surface extru-
sion (Defect B), double-surface extrusion (Defect C), four-surface extrusion (Defect
D), and bending deformation (Defect E). Sample information in detail is shown in
Table 1.
4.1 Defect Recognition
For sample with or without defects, defect features are obtained by the modify PCA
method, which separately extracts the PCs of probes A, B, C and integrated into
a set of features. Compared with the traditional PCA that extract the PCs of the
original signals directly, modified PCA makes use of the probes’ different detection
capabilities for defects. Table 2 shows the results of using the statistical features to
determine whether the defect exists.
Figure 9 shows the comparison of feature extraction: (a) represents features of
the first three PCs extracted by traditional PCA, (b) represents features extracted by
modified PCA. There is features’ overlapping in (a), which affects the accuracy of
defect recognition. The overlapping part disappears by the modified PCA method,
which gives a reliable explanation of the experimental results.
Table 1 PEC signal dataset

Total dataset Defect-free Defect A Defect B Defect C Defect D Defect E
Training set 1332 703 110 139 135 102 143
Testing set 442 234 36 46 45 34 47
Total 1774 937 146 185 180 136 190
Table 2 Results of defect recognition

Feature extraction Model Training set accuracy (%) Testing set accuracy (%)
Traditional PCA SVM 97.14 95.14
Modified PCA SVM 97.47 97.31
Traditional PCA RF 100 98.32
Modified PCA RF 99.93 99.55
(a) Feature Extraction Based on Traditional PCA (b) Feature Extraction Based on modified PCA
Fig. 9 Comparison of feature extraction
From the results in Table 2, RF algorithm is more accurate than SVM. This is
mainly due to the randomness of several classifiers in RF, which has a better identi-
fication of similar features compared to SVM’s single classifier. This also provides
a reliable basis for defect classification later.
4.2 Defect Classification
The classification of defect types is the main target of PECT of tubular structures.
Therefore, after the mentioned defect recognition, this paper analyzes defective
signals to classify the types of defects. This paper adopts modified PCA to extract
statistical features because of its better characterization of defects. Figure 10 shows
the result of feature extraction. There are many features’ overlapping among different
defect types, which indicates that only statistical features are not enough to classify
defects accurately.
The excitation signal of PEC is composed of sinusoidal waves at multiple
harmonic frequencies and its response contains abundant spectrum information, so
effective frequency-domain feature extraction is helpful for defect classification.
Figure 11 shows the amplitude spectrum and low-frequency amplified spectrum
of different defect signals after FFT. It is found that the FFT amplitude spectrum
peak can be used as the frequency-domain feature to characterize different defect
506 H. Jin et al.
Fig. 10 Modified PCA feature extraction
Fig. 11 FFT amplitude spectrum
types. Therefore, statistical features extracted by modified PCA and FFT ampli-
tude spectrum peak as a frequency-domain feature are proposed to form combined
features.
Table 3 Defect classification accuracy with RF model

Feature extraction Model Training set accuracy (%) Testing set accuracy (%)
Frequency-domain RF 71.70 64.42
PCA RF 92.05 90.38
Modified PCA RF 93.96 93.27
Combined features RF 99.68 99.51
Table 3 shows the classification accuracy of the RF model using different feature
extraction methods, from which the following conclusions can be drawn: Compared
with the traditional PCA method, the modified PCA method improves the classifi-
cation accuracy slightly; Compared with the single statistical feature, the combined
features only add one-dimensional feature but improves the accuracy greatly. The
classification accuracy of the training set and test set reach 99.68 and 99.51% based
on the modified PCA combined with the frequency-domain features, indicating a
high model accuracy.
To analyze the reasons for improvement of defect classification by combined
features, the confusion matrix of the testing set result based on modified PCA and
combined features is shown in Fig. 12.
From the results in Fig. 12, combined features have advantages in classify defects
such as single-surface extrusion and double-surface extrusion, which may be caused
by the spectrum peak contains the maximum mean of energy information of the
defect sample signal and the frequency-domain feature make up the shortcomings
of PCA that PCs with less contribution may contain important information of defect
types.
In order to verify the performance of the classifier of the proposed method, the
SVM model commonly applied in PEC studies is used as a reference. Figure 13 shows
the prediction results of RF and SVM model with combined features. Both models
accurately classify the defects of multi-surface extrusion and perforation cracks,
while RF model has better performance for the defects of single-surface extrusion
Fig. 12 Testing set result with different feature extraction methods

508 H. Jin et al.
Fig. 13 Confusion matrices of RF and SVM model with combined features
and bending deformation. This is attributed to the randomness of the RF model, which
improves the anti-noise ability and the classification accuracy of similar features.
5 Conclusion
In this paper, defect recognition and classification techniques for multi-layer tubular
structures using PECT is studied. Through the detection experiment of the casing
with prefabricated defects, this paper adopts modified PCA method to extract features
to realize defect recognition and constructs a model by RF algorithm based on the
combined features to realize defect classification of multi-layer tubular structures.
The following conclusions are obtained:(1) The modified PCA with RF model can
realize accurate defect recognition of multi-layer tubular structures; (2) Compared
with single statistical feature, combined features obtained from frequency-domain
features and modified PCA can effectively improve the defects classification accu-
racy of multi-layer tubular structures; (3) RF algorithm, compared with SVM, has a
better identification of similar features, which improve the defects recognition and
classification accuracy of multi-layer tubular structures.
From the foregoing, it can be seen that the defect recognition and classification
techniques of multi-layer tubular structures presented in this paper have a further
application prospect in PECT.
Acknowledgements This work is supported by the National Natural Science Foundation of China
[Grant No. 61174005] and the Technology Major Project of China [Grant No. 2016ZX0517-003
and No. 31300028-18-ZC0613-0002].
References
1. Qi C, Yu R, Cheng Q et al (2013) Research and application of the online magnetic detection

device for CT. J Nanchang Hangkong Univ Nat Sci 27(04):77–83
2. Zhang Z, Wu J, Luo Y (2002) Analysis and countermeasures of casing deformation causes in
Gangxi oil field. J Jianghan Pet Inst (04):83–84+1
3. Sophian A, Tian GY, Taylor D et al (2001) Electromagnetic and eddy current NDT: a review.
Insight-Non-Destr Test Cond Monit 43(5):302–306
4. Zhang Y, Li Y, Yan B et al (2018) Classification of subsurface corrosion in double-casing pipes
via integrated electromagnetic testing with stacked auto-encoder artificial neural network. J
Air Force Eng Univ (Nat Sci Ed) 19(01):72–78
5. Wang L, Song Z, Chang J et al (2015) Test analysis and application of coiled tubing
electromagnetic lossless detection. Oil Field Equip 44(07):60–63
6. Xu X (2005) Nondestructive testing of oil pipe defects. Daqing Petroleum Institute
7. Tang W, Shi YW (1997) Detection and classification of welding defects in friction-welded
joints using ultrasonic NDT methods. Insight 39(2):88–92
8. Huang P, Ding T, Luo Q et al (2019) Defect localisation and quantitative identification in
multi-layer conductive structures based on projection pursuit algorithm. Nondestr Test Eval
34(1):70–86
9. Sophian A, Gui YT, Taylor D et al (2003) A feature extraction technique based on principal
component analysis for pulsed Eddy current NDT. NDT E Int 36(1):37–41
10. Yang P, Qiu N (2015) Research on transient electromagnetic detection technology of Goaf
water based on smoke ring effect. Zhongzhou Coal 12:93–96
Corrected Phase Reconstruction
in Diffraction Phase Microscopy System
V. P. Bui, S. Gorelik, Z. E. Ooi, Q. Wang, and C. E. Png
Abstract This paper presents an approach for correcting phase reconstruction from
the interferogram Diffraction phase microscopy (DPM) system if there exist arte-
facts around the defect features due to noises. By averaging local region around the
artefacts of the unwrapped phase using a sliding window with iterative approach,
the correct phase information can be retrieved. The reconstructed phase images after
correction step demonstrate the ability of the proposed approach in extracting the
correct phase data from the system.
Keywords Diffraction phase microscopy · 3D reconstruction · Phase extraction ·

Phase unwrapping · Local phase approximation · Background removal
1 Introduction
Optical interferometry has the ability to reveal nanoscale information because it

can provide access to phase. Diffraction phase microscopy (DPM) is one of the
most popular optical inspection systems that uses a common-path interferometer to
obtain both the amplitude and phase information from a sample in order to achieve
3D information (shape and size) of an object. In the conventional DPM system,
one stringent requirement is that the first order of diffraction beam needs to be
produced with higher magnitude than zero order beam so that after going through
the pinhole, which also requires high accuracy alignment, these two fields interfere
at the final image plane on the CCD sensor to create the correct final interferogram
with the optimal fringe visibility [1–4]. The 3D reconstruction of interferogram
images obtained from DPM system using conventional algorithms such as Fourier
transform-, or Hilbert transform-based phase extraction, and branch cuts-, or global
V. P. Bui (B) · C. E. Png

Institute of High Performance Computing, A*STAR, 1 Fusionopolis Way, Singapore 138632,
Singapore
e-mail: buivp@ihpc.a-star.edu.sg
S. Gorelik · Z. E. Ooi · Q. Wang
Institute of Materials Research and Engineering, A*STAR, Singapore, Singapore

https://doi.org/10.1007/978-981-15-9199-0_48
512 V. P. Bui et al.
optimization-based phase unwrapping often results in very good outcome [5–16].

However, we have been observed that, in some specific cases, there are distorted
reconstructed images and lack of clarity if the defect size is below the diffraction
limit (sub-µm).
This work contributes to derive an approach for correcting phase reconstruction
from the interferogram system if there exist artefacts around the defect features.
Additionally, even though the noise level is inherently low in the DPM system, there
is still a need to quantify such noise and remove it when dealing with subdiffraction
size of defects. As we aim to develop a quasi-real-time imaging system, the first step
is to employ GPU-based Hilbert transform for extracting the phase information from
the interferograms. After that, making use of a fast algorithm based on quality map
is leveraged on for unwrapping the extracted phase, which will be gone through a
pre-corrected step by averaging region around the artefacts before dimension char-
acterization (estimation of shape and size of the defects). In this paper, we have
also attempted to reconstruct the background intensity of the resulting unwrapped
phase, from which the constructed background is then subtracted before retrieving
the correct phase information. The reconstructed phase images after corrections step
demonstrate the ability of the proposed approach in extracting the correct phase data
from the DPM system.
2 Phase Reconstruction in DPM System
2.1 Conventional Phase Reconstruction Process
The ultimate limiting factor for defect detection is not resolution, but rather noise,
which decreases image contrast and reduces sensitivity. This can be relieved by
using DPM system, which presents a low-noise optical microscope imaging system
with high sensitivity. A conventional DPM interferometer is illustrated in Fig. 1,
where a 405 nm semiconductor laser coupled into a single-mode fibre is used as
the illumination source. The light beam going through speckle reducer and optical
components is focused on the back focal plane of the objective lens for normal
incidence on the test sample. The reflected beam is collected by the same objective
lens and incident on a diffraction grating (300 lines per mm) to produce multiple
orders of beams. The first order and the zeroth order beams will then pass through a
pinhole, and 4f system which is composed of two lenses with focal lengths of 75 mm
and 400 mm, respectively. In DPM system, one stringent requirement is that the first
order of diffraction beam needs to be produced with higher magnitude than zeroth
order beam so that after going through the pinhole, which requires high accuracy
alignment, these two fields interfere at the final image plane on the CCD sensor
creating correct interferogram image with optimal fringe visibility [1–4].
The algorithm for reconstruction of phase image from the interferogram image
obtained from DPM system consists of two main steps. Firstly, phase extraction
Corrected Phase Reconstruction in Diffraction Phase Microscopy System 513
Fig. 1 A common-path diffraction phase microscope (DPM) for 3D imaging of defects
algorithm is used to extract the phase information from the raw interferogram image.
Then, phase unwrapping algorithm is employed to obtain the absolute phase image
and its 3D information. In fact, the 3D phase reconstruction is usually performed
using conventional algorithms such as Fourier transform-, or Hilbert transform-based
phase extraction [5–9], and branch cuts-, or global optimization-based phase unwrap-
ping [10–16]. In addition, there is a need to develop a background removal algorithm
in case no reference image can be provided. The background model is based on the
2D surface fitting (fourth-order polynomial) with data extracted from the unwrapped
phase. This is then used as the reference image for subtracting when retrieving the
phase information [17, 18].
The reconstruction process often results in very good outcome as illustrated in
Figs. 2, 3 and 4. The result obtained in Fig. 2 is from the interferogram image
generated by Eq. (3), where the one in Fig. 3 is the phase unwrapped from the
wrapped phase simulated by using Eq. (4). We use the measured metrics, such as peak
signal-to-noise ratio (PSNR) defined in Eq. (1) where mean-squared error (MSE) is
Fig. 2 Phase extraction from simulated interferogram image obtained by Eq. (3)
Fig. 3 Phase unwrap from simulated wrapped phase image obtained by Eq. (4)
Fig. 4 Reconstructed 2D and 3D images of SiO2 pillar with diameter of 60 µm, high of 100 nm
on Si wafer
obtained from Eq. (2), to evaluate the accuracy of the most popular extraction and
unwrapping algorithms for the results in Figs. 2 and 3, and the outcome is listed in
Table 1. It can be seen that a combination of Hilbert transform for phase extracting
and quality map for phase unwrapping can be considered as the best for subsequent
phase reconstruction. Figure 4 shows the reconstructed phase images of a 60 µm
Table 1 Accuracy level of conventional extraction and unwrapping algorithms

Extraction algorithm Fourier Hilbert Derivative Hilbert Huang
PSNR (dB) 35 57 34 46
Unwrapping algorithm Branch cuts Quality map Lp-norm Network flow
PSNR (dB) 36 36 35 36
diameter SiO2 pillar with high of 100 nm standing on the Si wafer. Note that a CCD
camera (1024-by-2048, pixel size of 5.3 µm) is used to capture the interferogram
image. The results agree very well with the test sample’s fabrication data (shape and
size).

PSNR = 10 log10 peakval2 /MSE (1)

MSE = 1/NM (X (i, j) − Y (i, j))2 (2)

I (x, y) = 128e−(x +y 2 )
+ 127e−(x +y 2 )
2 2
cos 8π x 2 + y 2 + 32π x (3)

ϕ(x, y) = Wrap To Pi 20e−(x +y ) /2/0.22 + 10(x + y) + 0.5 randn(N )
2 2
(4)
2.2 Local Corrections of Unwrapped Phase
It has been seen that DPM system is able to reduce much noise. However, because
of system calibration errors, sample-tilt, aberration, non-uniformity or time-varying
illumination intensity, there still exist noises causing artefacts during phase recon-
struction; and worst yet if the subdiffraction limit defect signal buried into noise that
will result in phase data corrupted. A calibration procedure is proposed by smoothing
based on iterative approach to approximate local phase for unwrapped noisy data as
presented in Fig. 5. The local approximation with average estimation in a sliding
window is performed to remove the artefact phase wherever there is. The procedure
Algorithm: Corrected unwrapped phase
Input: unwrapped phase, sliding window (size), threshold

Output: corrected unwrapped phase
Procedure corrected_phase(): N (number of sliding windows)
for i = 1 to N do
P = peak_detection(): Diffp > threshold, Maxp = number of peaks (artefacts)
NewPeakList.append(peak)
while count > 0 do
average_region(pixels, average parameters)
check = Diffp vs. threshold
count = number of remaining peaks (Maxp, check)
end while
Estimation = pixel value in sliding window
end for
return corrected phase
End procedure
Fig. 5 Pseudo-code of corrected unwrapped phase algorithm

Fig. 6 a Interferogram image, b Phase reconstruction, c corrected unwrapped phase, d 3D phase

image with respect to (c)
is initiated by the phase estimates obtained for the neighbouring points and the arte-
facts (containing in the sliding window), which is defined as when the magnitude
of the phase difference taking value larger than certain threshold). This is repeated
until there is no more discrete peak (artefact) within the sliding window, and the
phase becomes smoother. This process not only removes the artefact, but also helps
enhance signal-to-noise ratio (SNR) [19–21].
The experimental samples are captured by CCD camera as interferogram images and
processed using conventional phase reconstruction algorithm as shown in Fig. 6a,
b, respectively. It can be observed that strong noises disrupted the phase informa-
tion causing artefacts around the defects’ location. Figure 6c shows the corrected
unwrapped phase using the proposed procedure, and its 3D phase image is depicted
in Fig. 6d. From this result, we can see the importance of implementing the proper
correction procedure when there exist strong noises in the DPM system, and using
conventional phase reconstruction algorithm leading to undesireable outcome.
4 Conclusions and Future Works
A pre-corrected procedure of phase reconstruction in a common-path interferometer

system for quantitative phase imaging is presented. The reconstructed phase image
is corrected and able to provide more accurate 3D morphological information of
the test sample. Further investigation on many more samples will be conducted to
improve the robustness of the developed procedure.
Acknowledgements This study was supported by the Agency for Science, Technology and
Research (A*STAR) of Singapore (Grant No. 182-243-0030).
References
1. Chowdhury S, Izatt J (2014) Structured illumination diffraction phase microscopy for

broadband, subdiffraction resolution quantitative phase imaging. Opt Lett
2. Zhou R et al (2015) Semiconductor defect metrology using laser-based quantitative phase
imaging. Proc SPIE
3. Pham H et al (2011) Off-axis quantitative phase imaging processing using CUDA: toward
real-time applications. Biomed Opt Express
4. Bhaduri B et al (2014) Diffraction phase microscopy: principles and applications in materials
and life sciences. Adv Opt Photonics
5. Bhaduri B, Popescu G (2012) Derivative method for phase retrieval in off-axis quantitative
phase imaging. Opt Lett
6. Ikeda T et al (2005) Hilbert phase microscopy for investigating fast dynamics in transparent
systems. Opt Lett
7. Takeda M, Ina H, Kobayashi S (1982) Fourier-transform method of fringe-pattern analysis for
computer-based topography and interferometry. J Opt Soc Am
8. Kemao Q (2004) Windowed Fourier transform for fringe pattern analysis. Appl Opt
9. Trusiak M et al (2016) Single and two-shot quantitative phase imaging using Hilbert-Huang
Transform based fringe pattern analysis. Proc SPIE
10. Herráez M (2002) Fast two-dimensional phase-unwrapping algorithm based on sorting by
reliability following a noncontinuous path. Appl Opt
11. Goldstein R et al (1988) Satellite radar interferometry: two-dimensional phase unwrapping.
Radio Sci
12. Costantini M (1998) A novel phase unwrapping method based on network programming. IEEE
Trans Geosci Remote Sens
13. Pritt MD, Shipman JS (1994) Least-squares two-dimensional phase unwrapping using FFT’s.
IEEE Trans Geosci Remote Sens
14. Ghiglia DC, Romero LA (1996) Minimum Lp-norm two-dimensional phase unwrapping. J Opt
Soc Am A
15. Huang HYH et al (2012) Path-independent phase unwrapping using phase gradient and total-
variation (TV) denoising. Opt Express
16. Flynn TJ (1997) Two-dimensional phase unwrapping with minimum weighted discontinuity.
J Opt Soc Am A
17. Cheng C-Y, Hsieh C-L (2017) Background estimation and correction for high-precision
localization microscopy. ACS Photonics
18. Zhou R et al (2013) Detecting 20 nm wide defects in large area nanopatterns using optical
interferometric microscopy. Nano Lett
19. Katkovnik V et al (2008) Phase local approximation (PhaseLa) technique for phase unwrap
from noisy data. IEEE Trans Image Process
20. Ghiglia DC, Romero LA (1994) Robust two-dimensional weighted and unweighted phase
unwrapping that uses fast transforms and iterative methods. J Opt Soc Am A
21. Goldstein G, Creath K (2014) Quantitative phase microscopy: how to make phase data
meaningful. Proc SPIE Int Soc Opt Eng
Study on Fatigue Damage of Automotive
Aluminum Alloy Sheet Based on CT
Scanning
Ya li Yang, Hao Chen, Jie Shen, and Yongfang Li
Abstract Fatigue damage is the damage accumulation process under the repetitive
cycle load, which is accompanied by degradation of material properties. It is great
significant to study the fatigue life of aluminum alloy in the field of engineering
application. The damage characteristics of 6061 aluminum alloy are studied by three-
dimensional reconstruction and finite element method to estimate low-cycle fatigue
damage based on the X-ray computer tomography. The results show that the estimated
damage variables are accordant with those measured experimentally. Moreover, the
three-dimensional reconstruction method combined with the finite element analysis
method to calculate the material damage can meet the engineering requirements.
Keywords Damage characteristics · Fatigue test · X-ray computer tomography ·

Finite element analysis
1 Introduction
Fatigue failure is one of the most common failure modes of structures, which is caused
by gradual material degradation under cyclic loading. Material failure is originated
initially from the microvoid to grow gradually and finally forming the macro-scale
cracks [1–3]. Such damage evolution in the components steps across meso-scale and
macro-scale. Therefore, in order to track the fatigue life along with whole evolution
of damage, it is necessary to investigate the relationship between damage variables
of different scales, and establish an approach to describe damage, thereby forming a
multiscale correlation method to predict fatigue life.
The theoretical research of macroscopic damage mechanics emphasizes the influ-
ence of damage on macroscopic mechanical behavior of materials [4–7], however,
Y. Yang (B) · H. Chen · Y. Li

School of Mechanical and Automotive Engineering, Shanghai University of Engineering Science,
Long Teng Road 333, Shanghai, China
e-mail: carolyn71@163.com
J. Shen
Computer Science Department, University of Michigan, Dearborn, USA

https://doi.org/10.1007/978-981-15-9199-0_49
520 Y. Yang et al.
the process of microscopic physical changes of damage evolution is not observed.

Microscopic damage mechanics uses a mechanics average method to establish rela-
tionship between microscopic damage parameters with material damage and defor-
mation with a combination of damage primitives (microcrack, microvoid, shear zone)
[8–10].
In order to study the damage process of structure, it is impossible to solve this
problem simply from a macroscopic or microscopic view. Huang presented a model
with spring element to connect the damage evolution process on the mechanism level
to the material and component design on the continuum level [11]. Shelke extracted
lower scale feature from the macro-scale wave signal using nonlocal elasticity theory
to show progressive evolution of the damage [12]. In recent years, efforts have been
made to study effect of microstructural evolution on fatigue behavior by advanced
nondestructive testing instruments. High-resolution X-ray tomography has been
used to characterize the fatigue mechanisms [13–16]. However, few authors studied
the evolution of degradation mechanism from mesoscopic defects to macroscopic
damage. Therefore, it makes sense to establish theoretical method of multiscale
damage evolution and failure process in engineering practice.
The objective of this study is to propose a multiscale correlation method to predict
fatigue life of aluminum alloy 6061. In this method, mesoscopic defects are visualized
by X-ray computed tomography (CT), macroscopic damage is acquired through the
finite element simulation based on three-dimensional reconstruction from CT data.
An integral system of fatigue life prediction is obtained. The validity of the present
method is shown by comparing its results with experimental results.
2 Identification of Mesoscopic Defects
2.1 Material and Specimen
The composition of the material AL6061 is shown in Table 1. The specimen was
prepared in accordance with ASTM E8/E8M-15a and combined with the require-
ments of the fatigue tester. The specimen thickness is 2 mm. The specimen shape is
shown in Fig. 1.
Table 1 Property of aluminum alloy 6061

Chemical Cu Si Fe Mn Mg Zn Cr Ti Other AL
composition
Scale factor 0.25 0.6 0.7 0.15 0.85 0.25 0.16 0.15 0.15 Margin
Study on Fatigue Damage of Automotive Aluminum Alloy Sheet … 521
Fig. 1 Fatigue specimen geometry
2.2 Fatigue Test and Nondestructive Detection
In order to investigate the damage degradation of AL6061 under cyclic loading, one
material testing machine (MTS810) was selected to be carried out for the experiment
(Fig. 2). For low-cycle fatigue tests, a suitable fatigue loading was first searched
to achieve a fatigue life of slightly over 20,000 cycles. The value of 3500 N was
Fig. 2 Material testing system (MTS) and deformation measurement system

522 Y. Yang et al.
Fig. 3 X-ray computed

tomography system
obtained by using large number of specimens subject to different magnitudes of

fatigue loading. Then, three repeated tests were conducted to avoid the fortuity.
To observe the degradation of microstructure for specimen under each sub-cycle,
a high-energy and high-resolution micro-focus X-ray computed tomography system
(XCT) were used to measure all the voids and cracks in those material test specimens
(Fig. 3). Radiographic images were acquired at spatial resolution of 5 µm.
Three-dimensional reconstruction was conducted in VG studio system based on

two-dimensional initial data from XCT, and then, the FEA model was generated in
ABAQUS software, as shown in Fig. 4. For the finite element analysis, the elastic
strain energy, plastic strain energy, and residual strain energy can be output for the
damage degradation. In addition, the Young’s modulus of the specimen can also be
obtained.
Figure 5 shows that there is a linear relationship between the effective Young’s
modulus with the fatigue damage, but the residual strain energy have no obvious
relationship with the fatigue damage. This means that the effective Young’s modulus
is a more suitable for depicting the fatigue damage evolution.
Figure 6 shows the comparison between the simulated effective Young’s modulus
and the experimental test modulus. The result indicates that the degradation trend of
the effective Young’s modulus simulated by FEM is consistent with the experimented
elastic modulus, and the simulation of Young’s modulus based on three-dimensional
reconstruction from CT images could be a promising potential method for fatigue
damage evolution and estimation of residual life.
Fig. 4 The damage model

of specimen in ABAQUS
Fig. 5 The change of damage parameters in numerical computation, a the change of effective
Youngs modulus, b the change of residual strain energy
4 Conclusions
A multiscale methodology for fatigue life prediction of aluminum alloy was devel-
oped. In this technique, effective Young’s modulus as macroscopic material degra-
dation parameter is acquired through the finite element simulation based on three-
dimensional reconstruction of mesoscopic defects from X-ray computed tomography
524 Y. Yang et al.
Fig. 6 The change of

damage parameters in
numerical computation
(CT). The elastic modulus decreases linearly as the fatigue cycles increases. The
three-dimensional reconstruction method combined with the finite element analysis
method to calculate the material damage can meet the engineering requirements.
Acknowledgements This research was funded by the Natural Science Foundation of Shanghai
(18ZR1416500) and Development Fund for Shanghai Talents, and “Shuguang Program” supported
by Shanghai Education Development Foundation and Shanghai Municipal Education Commission.
References
1. Liu YB et al (2010) Advanced characterization of pores and fractures in coals by nuclear

magnetic resonance and X-ray computed tomography. Sci China (Earth Sci) 53(6):854–862
2. Jordon JB et al (2010) Microstructural inclusion influence on fatigue of a cast A356 aluminum
alloy. Metall Mater Trans A 41:356
3. Mu P et al (2014) Influence of casting defects on the fatigue behavior of cast aluminum
AS7G06-T6. Int J Fatigue 63:97–109
4. Murakami S, Ohno N (1980) A continuum theory of creep and creep damage. Creep in
structures, Leicester, pp 422–444
5. Talreja R (1985) A Continuum mechanics characterization of damage in composite materials.
R Soc 399(1817):195–210
6. Chaboche JL (1988) Continuum damage mechanics: part I-general concepts. J Appl Mech
55(1):59–64
7. Chow CL, Chen XF (1992) An anisotropic model of damage mechanics based on endochronic
theory of plasticity. Int J Fract 55(2):115–130
8. Gurson AL (1977) Continuum theory of ductile rupture by void nucleation and growth: part
I-yield criteria and flow rules for porous ductile media. J Eng Mater Technol 99(1):2–15
9. Ju JW, Lee X (1991) Micromechanical damage models for brittle solids. Part I: tensile loadings.
J Eng Mech 117(7):1495–1514
10. Beeker W, Gross D (1988) A two-dimensional micromechanical model of anisotropic elastic-
microplastic damage evolution. Ingenieur-Archiv 58(4):295–304
11. Huang Y, Gong XY, Suo Z (1997) A model of evolving damage bands in materials. Int J Solids
Struct 34(30):3941–3951
12. Shelke A, Banerjee S (2011) Multi-scale damage state estimation in composites using nonlocal
elastic kernel: an experimental validation. Int J Solids Struct 48(7–8):1219–1228
13. Marrow TJ, Buffiere JY (2004) High resolution X-ray tomography of short fatigue crack
nucleation in austempered ductile cast iron. Int J Fatigue 26(7):717–725
14. Zhang H, Toda H (2007) Three-dimensional visualization of the interaction between fatigue
crack and micropores in an aluminum alloy using synchrotron X-ray microtomography. Metall
Mater Trans 38(8):1774–1785
15. Padilla E et al (2012) Quantifying the effect of porosity on the evolution of deformation and
damage in Sn-based solder joints by X-ray microtomography and microstructure-based finite
element modeling. Acta Mater 60(9):4017–4026
16. Nicoletto G, Konecna R, Fintova S (2012) Characterization of microshrinkage casting defects
of Al-Si alloys by X-ray computed tomography and metallography. Int J Fatigue 41:39–46
IoT-Web-Based Integrated Wireless
Sensory Framework for Non-destructive
Monitoring and Evaluation of On-Site
Concrete Conditions
A. O. Olaniyi, T. Watanabe, K. Umeda, and C. Hashimoto
Abstract This research combines the use of IoT-web-based application in moni-

toring the real-time conditions of concrete, to accurately evaluate the maturity of
concrete samples from its fresh state, accurately predict the strength of concrete
among other properties, and it provides remote access to the outputs in real-time.
The predicted strength of concrete was acquired using various maturity functions
based on the temperature-time history of concrete samples. Two cement types of
fresh concrete were prepared, ordinary portland cement (OPC) and blast furnace
slag (BB) cement, at water-cement ratio 50%. Immediately, after fresh concrete
samples were cast into molds and well vibrated, the monitoring of internal average
temperature commenced, and these specimens were kept under controlled curing
condition over a 28-day period, and the temperature data were logged directly into
the database by the IoT sensors via a mobile internet hotspot connection. Graphical
outputs of temperature, concrete maturity and strength data for both concrete types
were monitored remotely; presented on the programmed web-based application and
was accessible for download on any internet-connected device.
Keywords Concrete · Temperature · Maturity · Internet-of-things · Wireless ·

Strength
1 Introduction
Valuable information of the existing conditions of concrete structural components are

enabled by continuous data-driven structural health monitoring (SHM) techniques
and these assists concrete professsionals in making timely construction management
decisions on/off-site. However, conventional approaches can be tedious and are often
time-consuming; especially when large amount of SHM data is to be processed and
presented into easy-to-understand user-friendly outputs [1]. The proliferation of the
A. O. Olaniyi · T. Watanabe (B) · K. Umeda · C. Hashimoto

Civil Engineering Department, Tokushima University, Josanjima Campus, Tokushima City, Japan
e-mail: t_watanabe@tokushima-u.ac.jp

https://doi.org/10.1007/978-981-15-9199-0_50
528 A. O. Olaniyi et al.
internet of things (IoT) micro-controllers as well as cloud-based data storage plat-

forms have greatly improved the art of SHM practices over the past few decades.
The temperature-time history of concrete otherwise known as the maturity method
has been used to reliably predict the compressive strength of concrete since around
1950, when there were several cases of building collapse due to uninformed decision
of early removal of formworks and props on-site [2]. Extensive research from that
time birthed maturity functions, and various equipment; several of which requires
proximity to site, are quite bulky to use and sometimes expensive. Remotely moni-
toring the real-time strength conditions of concrete with simple wireless techniques
is therefore desirable.
This project utilized IoT-based embedded temperature sensors and a web-based
application to evaluate the real-time conditions of concrete; the maturity of concrete
samples from fresh state, accurately predict the strength of concrete among other
properties and it provides remote access to the outputs in real-time. Graphical outputs
of temperature, concrete maturity and strength data for two concrete types were
evaluated and was made accessible for download on any internet-connected device.
Monitoring the real-time strength of concrete with the use of simple techniques
such as IoT-based sensors to monitor on-site properties of concrete is very relevant
nowadays.
2 Research Background and Objectives
2.1 Research Objectives
• To develop simple IoT senors and web-based application that allows users to
monitor concrete conditions; temperature, maturity and strength, remotely and in
real-time.
• Variation of materials and curing conditions to know how the properties of
concrete are affected.
2.2 Maturity Concept of Concrete
Maturity of concrete is a quality control mechanism that predicts the strength of

the in-situ concrete under varying temperature conditions. According to the ASTM,
maturity method can be defined as “a system for estimating concrete strength that
is based on the assumption that samples of a certain concrete mixture reaches the
same maturity index, such samples will attain equal strength” [3]. One of the earliest
researchers of the maturity technique, A.G. Saul, had defined maturity in very simple
terms; “concrete of the same mix proportion at equal maturity has almost the same
strength, at whatever combination of temperature and time to make up that maturity”
IoT-Web-Based Integrated Wireless Sensory Framework … 529
Fig. 1 Illustration of
strength-maturity
relationship [5]
[4]. This implies that there is a unique relationship between the maturity index and
strength of a concrete mixture. If two samples of a concrete mixture have the same
maturity, they will have the same strength even though they may have been exposed
to different curing times and temperatures.
At lower temperature, concrete samples take more time to reach maturity M1,
meanwhile, concrete exposed to higher temperature reaches maturity M2 faster
in less time (Fig. 1). However, if M1 = M2, these two concrete mixes will have
approximately the same strength, even though their curing times and temperatures
are different.
The effects of time and average temperature on concrete strength gain are quan-
tified through maturity equations. Evaluating the strength of concrete at its early
age is not only important for smart structural components but also in constructing
structurally sound concrete structures faster [5].
The Nurse-Saul maturity function is one of the widely used models to monitor
the temperature-time history of concrete;

t
M= (T − To )t (1)
0
where:
M = Maturity Index (°C-h or °C-day).
T = average temperature of concrete (°C).
T o = the datum temperature (often taken as 0 °C or 10 °C).
t = time interval (in days or hours).
The datum temperature employed in this equation depends on the type of cement
being used to prepare the concrete sample and the curing temperature range that
such sample will be subjected to. For example, concrete prepared with the Ordinary
Portland Cement (OPC) and exposed to a curing temperature range of 0–40 °C, a
datum temperature of 0 °C or −10 °C is applied [5].
Based on (1), the equivalent age, can be mathematically expressed as:
t
0 (T− To )
ten = t (2)
Tr − To
where:
t en = equivalent age of concrete (days or hours).
T r = the reference temperature, (°C).
Equation (2) was, however, found to represent a linear relationship, which may not
be valid for concrete samples at elevated temperatures above 60 °C. A new function
was developed in 1977 by Freiesleben and Pedersen, based on Arrhenius equation
that was used to account for the impact of the temperature on a chemical reaction
[2].

t
−E 1
− T1r
tea = e R T
t (3)
o
where:
t ea = equivalent age at ref. temp. (days/h).
E = activation energy, J/mol.
R = universal gas constant, 8.314 J/mol K.
2.3 Internet of Things (IoT) in Construction Practices
IoT involves the use of intelligently connected devices and systems to leverage data
from embedded sensors and actuators in machines [6]. It can be stated that the huge
collection and processing of data from sensors in civil structures is approaching an
advanced stage.
The application of embedded sensors within civil infrastructures is, however, no
longer at its infancy in civil engineering. Numerous instances of integrated phys-
ical/chemical sensors are being deployed in structural management, energy effi-
ciency, or transportation. Nonetheless, embedded sensors in civil infrastructures is
still far from ubiquitous. Deployment of sensors within structures are, however,
limited to specific cases of high experiences of SHM techniques [7]. The applica-
tion of temperature sensors in this project is another step further in the utilisation
of actuators for endless possibilities in the field of civil engineering, especially for
instances where remote and real-time situation of the condition of infrastructures are
desired.
3 Research Methodology
3.1 Concrete Specimen Preparation
According to the mix design provisions in the Japanese Industrial Standard; JIS
A1101, and based on two cement types; OPC and Blast Furnace Slag cement (BB
cement), two types of concrete mix were prepared in the laboratory [8]. The fine and
coarse aggregates used for the experiment were washed thoroughly to remove dirt
particles and air-dried to reduce the surface water content after washing. For each
concrete type, trial mixes were carried out to determine the required properties of
fresh concrete, i.e., slump (SLP), air content and concrete temperature (CT) as at the
time of casting, in accordance with the JIS A1101 standard. Table 1 shows the mix
proportion used in this experiment, using the two types of cement.
A total of thirty-eight cylindrical concrete specimens were cast and cured at 20 °C
(19 samples for each concrete type) [3]. To determine the compressive strengths of
each concrete cylindrical specimen, fifteen samples for each type, were subjected
to uniaxial compressive strength test at ages 1, 3, 7, 14 and 28 days. The concrete
cylinders were kept in moulds and sealed with polythene bags to prevent loss of
moisture to the atmosphere (Fig. 2).
3.2 Maturity Functions
Equation (4) is a derivative of (1) and can be used to evaluate concrete strength F c
of concrete samples with maturity M, where A is the slope and B the intercept of
equation [9].
Fc = a log M + b (4)
t
(Fc )t = f (5)
α + βt c 28
Substituting the curing age t with the equivalent age t ea , from (3), the modified
ACI model was used
as shown in (5) below, to evaluate the strength of concrete. The
28-day strength f c 28 and strength revelation factors α and β; constants 4.0 and
0.85, respectively, were applied in this study [10].
3.3 Temperature Sensors and IoT Micro-controllers
The DS18b20 waterproof temperature probes and WiFi-enabled ESP32 Arduino

micro-controllers were used in this experiment to measure the temperature of
532
Table 1 Concrete mix proportion

Type Cement type W /C (%) S/A W (kg) C (kg/m3 ) S (kg/m3 ) G1 (kg/m3 ) G2 (kg/m3 ) SLP Air CT
I OPC 50 42 7.0 14.0 28.68 23.76 15.76 12.5 7.2 20.8
II BB 7.0 14.0 28.68 23.76 15.76 12.3 4.0 20.6
A. O. Olaniyi et al.
Fig. 2 Measurement of the

temperature data of concrete
cylindrical specimen in the
laboratory
concrete samples and the temperature data was logged in real-time to the database.
The Arduino integrated development environment (IDE) was used to write Arduino-
based codes on the ESP32 board and the Application Programming Interface (API)
of the database was provided within the code, for each sensor to log its temperature
readings at intervals.
The program was written in a way that the temperature data of each concrete
specimen were logged at time intervals of thirty (30) minutes for the first forty-eight
(48) hours, and then subsequent data were logged at intervals of one hour for the
total period of the maturity test (i.e., 28 days). The SSID and password of a mobile
hotspot (provided by a dedicated portable WiFi modem) was also written within the
Arduino codes, which enables the sensors to be able to transmit their temperature
data wirelessly to the database via internet.
To confirm the accuracy of the designed sensors, the temperature of 2 extra
concrete samples each were monitored using the Keyence datalogging device
simultaneously (Fig. 3).
Fig. 3 Keyence data-logging device and the developed wireless sensor schematic diagram
Fig. 4 Temperature data of Type-I Type-II

both types of concrete 25
specimen
Temperature (⁰C)
20
15
10
5
0
Age (hour)
3.4 Web Application Development
A free and open source Python programming language-based module known as

Django was used to develop the web-based application that allows users to view
concrete maturity and strength results in real-time, irrespective of their location.
Django is a high-level Python Web framework that enables swift development and
pragmatic web-based application design. This highly sophisticated but easy-to-use
framework takes care of much of the hassle of Web development, so that the engineer
can focus on writing his conceptualised application without the need to reinvent the
wheel [11].
4.1 Concrete Temperature and Maturity Index
Both the data logger and the IoT device gave satisfactory results of the temperature-
time history of each of the concrete samples, as shown in Fig. 4. The temperature of
both concrete specimens increased significantly within the first 24 h and returned to a
constant minimum value, just a little bit below the temperature rise. The temperature
measurement procedure lasted for twenty-eight (28) days from the fresh state (Fig. 5).
4.2 Concrete Compressive Strength by Uniaxial and Maturity

Methods
The compressive strengths of both concrete mixes were evaluated via the conven-
tional uniaxial compressive strength test and the proposed maturity technique. The
28-day predicted strength of concrete type-I was approximately 100.25% of the
uniaxial (41.05 N/mm2 ) compressive strength; 41.15 N/mm2 (Fig. 6). While for
Fig. 5 Maturrity data of Type -I OPC Concrete Type II -BB Concrete
Compressive Strength (N/mm2)

concrete types I and II 45
specimens
30
15
0
0 4000 8000 12000 16000 20000
Maturity (⁰C-hr)
Fig. 6 Compressive strength type-I (OPC) type-II (BB)

of concrete types I and II 50

specimens
40
30
20
10
0
0 7 14 21 28
Age (days)
concrete type-II specimen, the strength result showed approximately 99.75% of the
uniaxial (28.69 N/mm2 ) strength; 28.63 N/mm2 (Fig. 7). The uniaxial result had
approximately 100% of the design strength for both mixes.
Fig. 7 Uniaxial compressive Uniaxial Type-I Uniaxial Type-II

and predicted strength of Pred. Type-I Pred. Type-II
type-I and type-II concrete

50
40
30
20
10
0
0 7 14 21 28
Age (days)
4.3 Conclusion and Recommendations
The aim of this research was to utilize IoT micro-controllers and temperature probes
combined with written computer programs based on the maturity method, to monitor
and evaluate the properties of concrete samples. Two types of concrete specimen
based on cement types were prepared and their temperature, maturity and compres-
sive strength were evaluated. A web-based application was developed, and it provided
a remote platform that enabled access to various data of the concrete conditions in
an interactive graphical format. This technique is applicable on-site, and several
improvements are being worked on.
The proposed technique suggests solution to some of the problems associated with
conventional methods which often require proximity to the construction site. It also
suggests an informed decision about safe early removal of formworks, where such
is required. Without having to worry about the cluster of wires and other obstruc-
tions on-site, this solution can thus be regarded as a user-friendly, and reliably safe
technique for real-time condition monitoring of concrete structures.
References
1. Boddupalli C, Sadhu A, Azar ER, Pattyson S (2019) Improved visualization of infrastructure

monitoring data using building information modeling. Struct Infrastruct Eng Maint Manage
Life-Cycle Des Perform 15(9)
2. Topçu İB, Karakurt C (2018) Strength-maturity relations of concrete for different cement,
types. In: Fırat S, Kinuthia J, Abu-Tair A (eds) Proceedings of 3rd international sustainable
buildings symposium (ISBS 2017). ISBS 2017. Lecture Notes in Civil Engineering, vol 6.
Springer, Cham
3. American Society for Testing and Materials (ASTM) (2004) Standard practice for estimating
concrete strength by the maturity method, ASTM C 1074-04. American Society for Testing
and Materials, Philadelphia, PA
4. Anderson KW, Uhlmeyer JS, Kinne C, Pierce LM, Muench S (2009) Use of the maturity method
in accelerated PCCP construction. (Report No. WA-RD 698.1). Washington State Department
of Transportation, Washington, DC
5. Kaburu A, Kaluli J, Kabubo C (2015) Application of the maturity method for concrete quality
control. J Sustain Res Eng 2(4):127–138
6. Wu W et al (2015) Improving data center energy efficiency using a cyber-physical systems
approach: integration of building information modeling and wireless sensor networks. Procedia
Eng 118:1266–1273
7. Chacón R, Posada H, Toledo Á, Gouveia M (2018) Development of IoT applications in civil
engineering classrooms using mobile devices. Comput Appl Eng Educ 26:1769–1781. https://
doi.org/10.1002/cae.21985
8. Japan Industrial Standard (JIS A1101) (2018) Standard specifications for concrete structures-
2018, ‘Design’ 2018 Edition
9. Eichi T, Asuo Y, Tetsuroo K, Isao U, Hideki O, Chikanori H (2002) Concrete engineering
handbook. ISBN 978-4-254-26476-0. C 3351, p 77
10. Dan BE, Domingo CJ (1997) Prediction of creep, shrinkage, and temperature effects in concrete
structures. ACI Committee Report 209R-92
11. ‘DjangoProject’ Documentation, 10 Oct 2018, https://www.djangoproject.com/
A Study of Atmospheric Corrosive
Environment in the North East Line
Tunnel
L. T. Tan, S. M. Goh, H. Y. Gan, and Y. L. Foo
Abstract This project was conducted to study the effect of the tunnel atmosphere
on uniform corrosion of metals. Metal coupons and environment monitoring systems
(ECMs) were deployed at buffer areas of different stations along the North East Line
to measure the rate of corrosion. Concurrently, temperature and relative humidity
were also measured through the ECM. Results from one year of monitoring are
presented in this paper.
Keywords Corrosion · Train tunnel · Atmosphere · Metal · Sensor · Exposure
1 Introduction
Corrosion refers to the degradation of a material as a result of its chemical or elec-

trochemical reaction with the environment. For corrosion of metals to take place, the
following components must be present: an anode and a cathode on the metal surface
are in contact with an electrolyte (Fig. 1) [1].
General (or uniform) corrosion refers to corrosion that occurs uniformly across
a surface area. For atmospheric corrosion, the general corrosion rate of a mate-
rial depends on the corrosivity of the environment, which may be categorized into
different levels. ISO 9223 standard (2012) [2] identifies six categories of corrosivity
for outdoor environment ranging from C1 to C5 followed by the highest corrosivity
category of CX (Table 1). The corrosivity levels can be determined directly through
mass loss measurements of standard carbon steel, copper, zinc and aluminum coupons
that have undergone corrosion for one year or indirectly through environment param-
eters consisting of relative humidity, temperature, SO2 deposition and Cl− deposi-
tion. Separately, ANSI/ISA Standard 71.04 (2013) [3] has developed four corrosivity
levels specifically for the electronics industry (Table 2). The lowest severity level is
G1 where corrosion is unlikely to be a factor in determining equipment reliability
(ANSI/ISA 2013).
L. T. Tan (B) · S. M. Goh · H. Y. Gan · Y. L. Foo

Singapore Institute of Technology, 10, Dover Dr, Singapore, Singapore
e-mail: dominic.tan@singaporetech.edu.com

https://doi.org/10.1007/978-981-15-9199-0_51
538 L. T. Tan et al.
Fig. 1 Four components for

an electrochemical corrosion
cell [1]
Table 1 Categories of
Category Corrosivity
corrosivity of outdoor
atmosphere according to ISO C1 Very low
9223 (2012) C2 Low
C3 Medium
C4 High
C5 Very high
CX Extreme
Table 2 Severity levels of corrosivity according to ANSI/ISA 17.04 (2013) standards. The overall
ISA severity level is based on the higher of the two corrosion rates of copper and silver
Severity level G1 mild G2 moderate G3 harsh GX severe
Copper reactivity level (in angstroms)* <300 <1000 <2000 ≥2000
Silver reactivity level (in angstroms)* <200 <1000 <2000 ≥2000
The most common type of corrosion monitoring system consists of physical metal
coupons which are considered to be a proven measure of corrosion damage in the
case of alloys of iron, zinc and copper (ISO 9226 2012) [4]. The coupons have flat
surface areas which are polished with grit 120 abrasive paper. They are weighed
before and after exposure to measure the rate of corrosion. The corrosion products
on the coupons may also be studied using analytical methods or microscopy such as
scanning electron microscopy (SEM).
A wide range of corrosion may occur in rail infrastructure affecting railway tracks,
rolling stock as well as signaling components. The occurrence of corrosion is due to
the exposure of the rail infrastructure to corrosive agents that may be found in the
environment, including biocorrosion (e.g., Maruthamuthu et al. 2013) [5]. Corrosion
may lead directly to failure or contribute to the initiation and propagation of cracks
such as in corrosion fatigue of railway axles (e.g., Bretta et al. 2000) [6].
Most studies on corrosion of rail infrastructure have focused either on the railway
track (e.g., Panda et al. 2008, Ritter et al. 2014) [7, 8] or rolling stock components
such as the axle and the wheel. Ritter et al. [8] performed a detailed study on the
corrosion fatigues of rail base, utilizing methods such as microstructural analysis,
A Study of Atmospheric Corrosive Environment in the North East Line Tunnel 539
salt fog testing and mechanical fatigue testing. The purpose of their tests was to
compare the corrosion resistance of different coatings on the rail base. Similarly,
Panda et al. [7] investigated the corrosion behavior of different rail steels while Liu
et al. [9] compared the performance of different coating materials on the corrosion
of overhead catenary components.
A relatively large amount of research has been conducted on atmospheric corro-
sion; however, very little has been studied on atmospheric corrosion in a tunnel
environment. In a series of papers [10, 11], metal coupons and electrical corrosion
sensors were deployed to monitor corrosion in a road tunnel. However, studies on
railway tunnels have been lacking. This study was therefore performed to obtain
some baseline data of the corrosivity level of the environment in a mass rapid transit
tunnel.
2 Methodology
2.1 Exposure Sites
Ten MRT stations along the North East Line tunnel were selected for this study. For
each station, a buffer area was selected as the exposure site. The selection of the
buffer area was based on the availability of space to deploy the test rack and power
socket to operate the sensors.
2.2 Corrosion Specimens
At each exposure site, a custom-made sensor rack was installed, as shown in Fig. 2.
Each rack is in the form of a floor standing, custom-made aluminum frame which
was secured to the wall. The overall size of the sensor rack is approximately 100 cm
length × 30 cm width × 20 cm depth. Each rack contained three different corrosion
indicators to measure the corrosion rate. The indicators consisted of one unit of
Environment Condition Monitoring System (ECM, COSASCO), six sets of copper–
silver strip coupons (CCC, Purafil) and seven sets of metal coupons (six pieces per
set, supplied by ASPI). The CCCs were later removed and not used further in the
study. The ECM was attached to the top section of the rack, and the metal coupons
were set up at the bottom section as shown in Fig. 2. The ECM sensors were installed
vertically facing the track, and the metal coupons were installed with one side facing
the direction of the moving trains.
Environment Condition
Monitoring System (ECM)
Metal Coupon
Fig. 2 Front view of the placement of the sensors and metal coupons in the sensor racks (left) and
deployment of the sensor rack to the exposure site (right)
2.3 Metal Coupons
Metal coupons of five materials, low alloy steel (C4140), low carbon steel (C1010),
stainless steel SS304, copper alloy (CDA110) and zinc were used in this study.
The coupons were in the form of rectangular plates of size 7.62 cm × 1.27 cm ×
0.15 cm. Each coupon as received was stamped with a unique serial number. Their
mass and dimensions were measured and recorded before and after they were exposed
at the sites. Photo-macrographs of the coupons before and after exposure were also
recorded.
C4140, SS304 and CDA110 coupons were exposed between June 2017 and May
2018 while C1010 and zinc coupons were exposed between June 2018 and May
2019. Two coupons of each metal were collected for analysis at predetermined expo-
sure duration. One coupon from each set was subjected to mass loss measurements,
determined according to the method specified in ASTM G1-03 (2011) while the other
coupon was subjected to scanning electron microscopy (SEM) and energy dispersive
X-ray analysis/(EDX) analysis.
3.1 Temperature and Relative Humidity
The variation of the relative humidity and temperature through the different hours of
the day is shown in Figs. 3 and 4. The relative humidity largely ranged between 40 and
Station 1 Station 2
90 90
80 80
Relative Humidity (RH%)

70 70
60 60
50 50
40 40
30 30
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time-of-a-day (Hour) Time-of-a-day (Hour)
Station 3 Station 4
90 90
80 80

70 70
60 60
50 50
40 40
30 30
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Station 5 Station 6
90 90
80 80
70 70
60 60
50 50
40 40
30 30
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Station 7 Station 8
90 90
80 80
70 70
60 60
50 50
40 40
30 30
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Station 9 Station 10
90 90
80 80
70 70
60 60
50 50
40 40
30 30
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Fig. 3 Measurements of relative humidity from ten exposure sites (May 2017 to May 2019)
Station
NE1 HBF1 Station
NE3 OTP2
40 40
35 35
Temperature (C)
Temperature (C)
30 30
25 25
20 20
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Station
NE5 CQY3 Station
NE7 LTI4
40 40
35 35
Temperature (C)
Temperature (C)
30 30
25 25
20 20
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
NE9 BNK NE10 PTP

40 40
35 35
Temperature (C)
Temperature (C)
30 30
25 25
20 20
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Station
NE12 SER7 Station
NE14 HGN8
40 40
35 35
Temperature (C)
Temperature (C)
30 30
25 25
20 20
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Station
NE16 SKG9 Station
NE17 PGL10
40 40
35 35
Temperature (C)
Temperature (C)
30 30
25 25
20 20
00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Fig. 4 Measurements of temperature from ten exposure sites (May 2017 to May 2019)
60% while the temperature was between 30 and 35 °C at all exposure sites. Relative
humidity above 70% was recorded at seven exposure sites except Station 4, Station
5 and Station 7. The large fluctuations in both relative humidity and temperature
appear to occur predominantly during maintenance hours, i.e., between 1 and 4 am.
Otherwise, the relative humidity and temperature largely remained relatively constant
within a narrow range during operating hours (between 6 am and 11 pm).
3.2 Metal Loss of Silver and Copper Sensors
Figure 5 shows the metal loss from the silver sensor of ECM from June 2017 to
June 2019. The rates of corrosion at all the exposure sites were fairly constant with
no large changes during the two years of exposure. Different degrees of corrosion
were recorded at the exposure sites, with Station 1 showing the largest degree of
corrosion. For the copper sensors of ECM, little corrosion was observed during the
same exposure duration as shown in Fig. 6. Considering both the silver and copper
corrosion, the average rates of corrosion at all sites fall into the lowest severity level
(G1) in accordance with ISA standard (Table 2) [3]. This severity level generally
indicates that corrosion due to the environment is not a factor in determining equip-
ment reliability. The corrosion of the copper sensor is not significant because copper
is highly dependent on relative humidity [12], and it was found the relative humidity
in the tunnel environment was not high.
5000
Station 1
4500
Station 2
metal loss from silver sensor (Å)
4000
Station 3
3500
Station 4
3000 Station 5
2500 Station 6
Station 7
2000
Station 8
1500
Station 9
1000 Station 10
500
0
01-06-17 00
20-06-17 08
09-07-17 16
29-07-17 00
17-08-17 08
05-09-17 16
25-09-17 00
14-10-17 08
02-11-17 16
22-11-17 00
11-12-17 08
30-12-17 16
19-01-18 00
07-02-18 08
26-02-18 16
18-03-18 00
06-04-18 08
25-04-18 16
15-05-18 00
03-06-18 08
22-06-18 16
12-07-18 00
31-07-18 08
19-08-18 16
08-09-18 00
27-09-18 08
16-10-18 16
05-11-18 00
24-11-18 08
13-12-18 16
02-01-19 00
21-01-19 08
09-02-19 16
01-03-19 01
20-03-19 09
08-04-19 17
28-04-19 01
time stamp
Fig. 5 Metal loss from silver sensor of ECM (June 2017–June 2019)
5000
Station 1
4500
Station 2
metal loss from copper sensor (Å)
4000
Station 3
3500
Station 4
3000
Station 5
2500
Station 6
2000
Station 7
1500
Station 8
1000
Station 9
500
Station 10
0
01-06-17 00
20-06-17 08
09-07-17 16
29-07-17 00
17-08-17 08
05-09-17 16
25-09-17 00
14-10-17 08
02-11-17 16
22-11-17 00
11-12-17 08
30-12-17 16
19-01-18 00
07-02-18 08
26-02-18 16
18-03-18 00
06-04-18 08
25-04-18 16
15-05-18 00
03-06-18 08
22-06-18 16
12-07-18 00
31-07-18 08
19-08-18 16
08-09-18 00
27-09-18 08
16-10-18 16
05-11-18 00
24-11-18 08
13-12-18 16
02-01-19 00
21-01-19 08
09-02-19 16
01-03-19 01
20-03-19 09
08-04-19 17
28-04-19 01
time Stamp
Fig. 6 Metal loss from copper sensor of ECM (June 2017–June 2019)
3.3 Mass Loss of Metal Coupons
The appearance of the metal coupons is shown in Fig. 7. Corrosion could be clearly
observed on the carbon steel C1010 and low alloy steel C4140. Some staining could
be observed on the stainless steel SS304, copper CDA 110 and zinc coupons, but
otherwise, little to no corrosion was measured on these metals.
Figure 8 shows mass loss of C4140 from ten exposure sites. Different exposure
sites appeared to exhibited different degrees of corrosion after one year of exposure.
C4140
CDA 110
SS 304
C1010
Zinc
Fig. 7 Appearance of metal coupons before (left) and after (right) exposed for 12 months
mass loss in terms of thickness reduction (µm)

Station 1
Station 2
Station 3
2 Station 4
Station 5
Station 6
Station 7
1 Station 8
Station 9
Station 10
0
0 4 8 12 16
exposure duration (month)
Fig. 8 Mass loss in terms of thickness reduction of C4140 (June 2017–June 2018)
The reduction in the thickness of the C4140 coupons as a result of the atmospheric
corrosion was within 2 µm after one year of exposure. Figure 9 shows the mass loss
of C1010 from ten exposure sites after one year of exposure. Similar to the corrosion
of C4140 coupons, the corrosion-induced thickness reduction was within 2 µm after
one year of exposure.
3
mass loss in terms of thickness reduction (µm)
Station 1
Station 2
Station 3
2 Station 4
Station 5
Station 6
Station 7
Station 8
1
Station 9
Station 10
0
0 4 8 12 16
exposure duration (month)
Fig. 9 Mass loss in terms of thickness reduction of C1010 (June 2018–June 2019)
Based on the thickness loss rate of low carbon steel (C1010) after one year of
exposure, it may be deduced that the corrosivity level in the train tunnel be catego-
rized as C1 (very low corrosivity) in accordance with the ISO 9223 [2]. Assuming
that the corrosion rates for the metals are linear, the reduction of thickness may be
approximated to be within 60 µm if explorated to 30 years of service life. Do note
that the carbon steel C1010 and low alloy steel C4140 were exposed at different
periods; however, as shown by the ECM data, the rates of corrosion between the
different periods are expected to be similar.
3.4 Microstructure Analysis
Representative corrosion products observed at the surface of C4140 and C1010

coupons after 12 months of exposure are shown in Fig. 10. The corrosion prod-
ucts show globular and flowery structure, which are common iron oxide products of
steel subjected to atmospheric corrosion [13, 14]. Further analysis of the corrosion
products using EDX (see Table 3) revealed trace amounts of sulfur and chlorine.
These elements are commonly detected in other studies of atmospheric corrosion
[15, 16]. The source of sulfur and chlorine may be from the air-borne pollutants,
for example, sulfur dioxide and chloride [17]. According to Sing [18], both sulfur
dioxide and chlorides change the structure and protective properties of the rust layer.
a) b)
Fig. 10 Representative morphology of corrosion products from the exposed metal coupons,
a C4140 from exposure site at Station 1, b C1010 from exposure site at Station 3
Table 3 Representative weight percentage of the elements from the exposed C4140 at Station 1
and C1010 at Station 3
Fe C N Si O Cl S
C4140 48 11.0 4.9 0.1 33 0.4 1.2
C1010 56 8.1 1.6 0 33 0.8 0.7
For salt like chloride to accelerate corrosion, the metallic surface needs to be wet
[19].
4 Conclusions and Future Work
The corrosivity of the North East Line tunnel has been monitored using metal coupons
and ECM systems. Although different sites have shown different rates of corrosion,
the data collected suggest that the corrosivity levels are generally low.
While metal coupons are typically used to measure corrosion rates, the use of the
ECM system provides a continuous measurement of corrosion. In addition, the ECM
system can potentially allow for condition-based maintenance through monitoring
of corrosion rates. In condition-based maintenance (CBM) development, wireless
sensor networks (WSNs) have been used for monitoring the railway infrastructure
such as bridges, rail tracks, track beds and track equipment along with vehicle health
monitoring such as chassis, bogies, wheels and wagons [20, 21]. The integration
of wireless networking and sensing technology to form wireless-based condition
monitoring has successfully reduced the need for humans to perform manual inspec-
tion. In addition, the safety and reliability of systems are improved through auto-
mated condition monitoring which detects faults before they escalate. In order to
enable condition-based maintenance techniques, this project has also embarked on
the development of a low power, wireless with 3G cellular module, embedded corro-
sion monitoring system (based on the ECM), cloud database and mobile app for
real-time environment monitoring based on the IEEE 1451.5 open architecture for
smart transducers.
References
1. Davis JR (2000) Corrosion: understanding the basics. ASM International, p 563

2. ISO9223 (2012) Corrosion of metals and alloys—corrosivity of atmospheres—classification,
determination and estimation
3. ISA-71-4 (2013) Environmental conditions for process measurement and control systems:
airborne contaminations
4. ISO9226 (2012) Corrosion of metals and alloys—corrosivity of atmospheres—determination
of corrosion rate of standard specimens for the evaluation of corrosivity
5. Maruthamuthu S, Dhandapani P, Kamalasekaran S, Siddarth AS, Manoharan SP, Muthuraman
K, Narayanan G (2013) Role of ureolytic bacteria on corrosion behavior of fretted grade 880
mild steel rail. Eng Fail Anal 33:315–326
6. Beretta S, Sangali F, Syeda J, Panggabean D, Rudlin J (2017) RAAI Project: Life-prediction
and prognostics for railway axles under corrosion-fatigue damage. Proc Structural Integrity
4:64–70
7. Panda B, Balasubramanian R, Dwivedi G, Mahapatra S (2008) Corrosion of Novel Rail Steels
in 3.5 % NaCl solution. Trans Indian Inst Metals
8. Ritter GW, Stuart C, Tang Y (2014) Rail base corrosion and cracking prevention: phase 2. U.S
Department of Transportation
9. Liu ZG, Song Y, Wang HR, Zhang J, Han ZW (2018) Advances of research on high-speed
railway catenary. J Modern Transp 26:1–23
10. Kreislova K, Knotkova D (2011) Corrosion behaviour of structural metals in respect to long-
term changes in the atmospheric environment. EUROCORR
11. Kreislova K, Geiplova H, Licbinsky R (2012) Corrosivity of road tunnel microclimate.
EUROCORR 2012
12. ISO11844-1 (2006) Corrosion of metals and alloys—classification of low corrosivity of indoor
atmospheres—Part 1: determination and estimation of indoor corrosivity
13. Humaira B (2015) Morphology and composition of corrosion products formed on painted mild
steel in marine atmosphere after 4, 8 and 12 months exposure. Int J Adv Res 3(4):318–325
14. Alcantra J, Chico B, Simancas J, Diaz I, Fuente D, Morcillo M (2016) An attempt to classify
the morphologies presented by different rust phases formed during exposure of carbon steel to
marine atmospheres. Mater Characterization 118:65–78
15. Corvo F, Minotas J, Delagado J, Arroyave C (2005) Changes in atmospheric corrosion rate
caused by chloride ions depending on rain regime. Corros Sci 47:883
16. Mendoza AR, Corvo F (1999) Outdoor and indoor atmospheric corrosion of carnon steel.
Corros Sci 41:75–86
17. Morcillo M, Fuente D, Diaz I (2011) Atmospheric corrosion of mild steel. Department Surface
Engineering, Corrosion and Durability. National Centre for Metallurgical Research 47(5):426–
444
18. Singh DN, Yadav S, Saha JK (2008) Role of climatic conditions on corrosion. Corros Sci
50:93–110
19. Alcántara J, Chio B, Díaz I, Fuente D, Morcillo M (2015) Airborne chloride deposit and its
effect on marine atmospheric corrosion of mild steel. Corrosion Sci 97:74–88
20. Fumeo E, Oneto L, Anguita D (2015) Condition based maintenance in railway transportation
systems based on big data streaming analysis. Proc Comput Sci 53:437–446
21. Kim KHL, Kong DWC (2012) Condition based monitoring in railway. J Int Council Electrical
Eng 2(1):99–103
Acoustic Emission
Fatigue Crack Growth Monitoring Using
Acoustic Emission: A Case Study
on Railway Axles
R. Marks, A. Stancu, and S. Soua
Abstract Railway axles are safety-critical components, which are designed to be

safe and reliable under standard operation. However, unexpected service condi-
tions, impact events and issues missed in maintenance can affect their lifetime.
Non-destructive testing (NDT) methods employed to assure structural integrity are
expensive and disruptive requiring axles to be taken out of service for inspection.
Therefore, there is need for more efficient, real-time monitoring approach. This paper
presents an overview of key challenges facing successful in-service implementation
of an acoustic emission (AE) condition monitoring system. This is followed by an
experimental study of AE data captured during a mechanical test of a railway axle.
The data is clustered using a self-organising map (SOM) to differentiate damage
signals from other sources. The results presented show consistency with the crack
growth measured using traditional NDT techniques.
Keywords Railway axle · Structural health monitoring · Acoustic emission ·

Damage detection
1 Introduction
There is increasing demand on railway assets as passenger kilometres continue to

increase [1]. With increased demand, there is requirement for train operating compa-
nies (TOCs) to reduce costs, improve service provisions and improve operations
efficiency while delivering a safe and reliable service. Rolling stock maintenance
expenditure has been identified as a significant direct operating cost, particularly in
the UK [2]. This is especially the case for unplanned maintenance which has other
implications such as service delays and passenger disruption.
R. Marks · A. Stancu (B) · S. Soua

TWI Ltd., Granta Park, CB21 6AL Cambridge, UK
e-mail: andra.stancu@twi.co.uk
R. Marks
e-mail: ryan.marks@twi.co.uk

https://doi.org/10.1007/978-981-15-9199-0_52
552 R. Marks et al.
The train axles are a large source of maintenance costs, requiring axle removal
from the bogie for manual NDT [3]; currently carried out every 30,000 km [4]. This
costly and time-consuming exercise does not fully prevent axle failure due to the
large cyclic loads and operating environments that the axles are subjected to between
inspections. Between 2006 and 2009, a total of 329 broken axles were reported in the
European Union [5]. Such failures have resulted in derailment [6], significant damage
and loss of life [7]. Crack growth models have been traditionally used to model the
expectant axle life [8]; however, these do not account for unforeseen circumstances
(such as abnormal loading, environmental conditions and impacts). Therefore, there
is a need for continuous monitoring technologies to detect the onset of damage and
its growth on axles, enabling an ‘as-required’ maintenance strategy to be adopted.
This paper presents experimental mechanical testing of a representative railway
axle with acoustic emission (AE) monitoring. An unsupervised artificial neural
network technique was then used to identify the crack growth signals.
2 Acoustic Emission Monitoring
AE is the spontaneous release of a strain wave when a structure is loaded or under-

goes a damage event such as crack growth or corrosion [9]. By mounting sensors
to the structure, AE can be detected, studied and analysed to determine structural
health. Storing all transient AE waves is computationally expensive and is inefficient.
Therefore, transient waves attributes are usually extracted such as peak amplitude
and MARSE as shown in Fig. 1.
AE has received much study for monitoring fatigue damage in metallic structures,
particularly in recent years [10, 11]. AE monitoring has therefore been proposed for
railway axle condition assessment.
Bruni et al. [12] used AE for monitoring a notched axle subjected to rotational
bending loads, leading to Carboni et al. [13] to build trends between cumulative AE
energy and vibrational harmonic amplitude. Carboni and Gianneo [14] monitored an
axle under rotational bending loads to monitor the growth of fretting fatigue, using
a SOM to filter damage signals from noise. Zhou et al. [15] presented a method of
differentiating different signals using a self-adaptive empirical mode decomposition
with and Elman neural network.
3 Self-organising Map
AE has shown to be a suitable structural monitoring method and can be reasonably

straight forward in a simple low noise, laboratory environment. However, monitoring
can be more complex in an environment with lots of background AE. To reduce the
amount of background AE captured, suitable acquisition parameters are applied.
The simplest parameter is an appropriate threshold level (achieved by assessing the
Fatigue Crack Growth Monitoring Using Acoustic Emission … 553
Fig. 1 Examples of attributes extracted from an AE wave
background noise level when no/low loading is applied to the structure). However, this
does not prevent high amplitude background AE being captured (i.e. from rubbing).
Frequency filters can remove reduce background AE acquisition although this risks
removing useful information from the AE signal acquired (e.g. a low pass filter will
effect the number of counts). Front-end AE attributes filters can remove unwanted
signal acquisition; however, this approach requires knowledge about the damage
signals in order to build a useful front-end filter.
SOMs have shown promise in differentiating damage signals from other AE. A
SOM is an artificial neural network (ANN) that does not require training. Because
SOMs are unsupervised, they are well suited to sorting AE attribute data where
information on the damage signals is unknown.
A brief SOM overview is presented. For further information, the reader is referred
to Kohonen [16]. A SOM can be thought of a layer of neurons which is fed by an
input layer as presented in Fig. 2.
A random weight is assigned to each neuron in the Kohonen layer. A piece of
input data is then randomly selected, and the Euclidean distance to each neuron is
calculated. The neuron with the smallest Euclidian distance is declared the winner,
and the weights of every neuron in Kohonen layer are updated. This process is
repeated with all input data, resulting in less updates to the Kohonen layer as similar
input data group together. Typically, there are more neurons than in the Kohonen layer
than expected clusters. The ‘U-matrix’ (or distance matrix) is a representation of the
distances between neurons. This allows clusters to be identified where each cluster
554 R. Marks et al.
Fig. 2 Schematic of a
self-organising map
has low distance values in the matrix. Computational clustering methods (such as
the K-means algorithm) can be used to cluster the U-matrix automatically.
4 Experimental Set-up
The specimen produced for this study was manufactured from stainless steel (hot
rolled round bar grade 15,230.9; a commonly used grade for tram axles) as part
of a wider investigation. The specimen represents a half-scale in-service axle with
additional material for test rig mounting. An elliptical notch 6.5 mm deep and 13 mm
was created at the mid-point by electrical discharge machining (EDM) to instigate
the fatigue crack. The axle was loaded cyclically in 3pt bending at 0.5 Hz for the first
3123 cycles, increasing to 1 Hz for subsequent cycles. A schematic of test set-up is
presented in Fig. 3.
Fig. 3 Test schematic

The test was stopped periodically to measure the growth of the crack and its
depth using ultrasonic phased array and alternating current potential drop (ACPD).
The load range was adjusted to facilitate the growth of the crack. The details of the
loading are presented in Table 1.
AE was captured using a Vallen AMSY-6 system connected to a Windows laptop
running Vallen acquisition software. Three Vallen VS-150-RIC (resonance 100–
450 kHz, 150 kHz peak, 34 dB integrated pre-amplifier) sensors were mounted to
the axle at the locations shown in Fig. 3. A 40 dB acquisition threshold was selected
to remove unwanted signals, and a 95–850 kHz band-pass filter was also applied.
Typical acquisition settings for metallic structures were used: duration discretization
time (DDT) = 400 ms, re-arm time = 3.2 ms, pre-trigger = 1000 samples and 5 MHz
sample rate. Load and displacement data was recorded simultaneously from the load
machine at 10 Hz.
5 Results
5.1 Raw Data
A front-end filter was applied to remove AE signals with less than 3 counts. Data
from Channel 1 is presented in Fig. 4 demonstrating the influence of background AE
in the raw data.
Cumulative energy is commonly used for determining the crack growth from AE.
The raw data showed spikes in cumulative energy which were not consistent with
the crack length measured. This was particularly the case at 80,000 s.
5.2 SOM Training
The SOM was built in Python based on a modified SOMPy package [17]. A 12 × 12
planar hexagonal map was used with a linear principal component analysis (PCA) for
weight initialisation. Range normalisation was applied to the data and used to train
the SOM using a Gaussian neighbourhood function and a batch training method.
Training was conducted using peak amplitude (in dB), rise time, duration, energy
(MARSE), signal strength, RMS and counts from Channel 1. The map was trained
over 75 iterations (50 rough, 25 fine) with each iteration taking approximately 1.5 s
on an Intel i5, 8th Generation Windows laptop.
Once the SOM was trained, the U-matrix was calculated (presented in Fig. 5a).
The peak distances are represented in yellow and demonstrate the cluster boundaries.
It is apparent that there are two main clusters: one in the bottom left and one in the
upper central region. The U-matrix was clustered using a K-means algorithm as
shown in Fig. 5b. It should be noted that the number of nodes in a cluster is not
556
Table 1 Loading sequence for the axle

Load, kN Number of cycles Total number of cycles Crack depth, mm Crack length, mm
Min Max Range ACPD Phased array
0 0 0 0 0 2.2 – 4 – 8
8 80 72 15,000 15,000 3.7 1.5 4 0 11.5
8 80 72 5000 20,000 4.4 0.7 6.7 2.7 14
44 80 36 10,000 30,000 4.7 0.3 7.1 0.4 17
44 80 36 25,000 55,000 5.6 0.9 8.8 1.7 18
8 80 72 5,584 60,584 6.8 1.2 10 1.2 23
8 80 72 7,416 68,000 8.2 1.4 11.1 1.1 28
8 80 72 6,824 74,824 10.1 1.9 14.3 3.2 35
R. Marks et al.
(a) (b)
Fig. 4 Raw data from Channel 1, a Amplitude versus time of hits, b Cumulative energy versus
time with measured crack length
(a) (b)
Fig. 5 Clustering of data a U-matrix, b U-matrix clustered by a K-means algorithm
representative of the amount of data assigned to that cluster, but rather the spread of
the data.
The amount of clusters was determined using three widely used performance
indices: Calinski-Harabasz, Silhouette and Davies-Bouldin. An aggregate score of
all three indices was created based on work by Crivelli et al. [18]. The results from
all of the performance indices indicated two clusters were best suited.
5.3 Clustered Data
The data from Channel 1 was fed back into the SOM and allocated a cluster. This
resulted in 259,211 AE hits in Cluster 0 and 22,126 AE hits in Cluster 1. The
558 R. Marks et al.
Fig. 6 Cumulative energy for each cluster for Channel 1 data (used to train SOM)
Fig. 7 Cumulative energy for each cluster for Channel 3 data (not used to train SOM)
cumulative energy for each cluster was then plotted against time as presented in
Fig. 6.
The data from the other channels was also clustered by the same technique. The
cumulative energy plots for the two clusters for the data from Channel 3 are presented
in Fig. 7. Cluster 0 contains 221,480 hits, and Cluster 1 contains 13,007 hits.
6 Discussion
Cluster 0 for Channel 1 shows a good consistency between the crack length growth
and the cumulative energy. It should be noted that the regions where there is no
change in cumulative energy identify the periods when the loading was suspended
for NDT. The initiation of the crack growth is identified by the increasing gradient
at the start of the test. This gradient increases as the crack continues to grow from
the notch. This is the case for each loading increment; however, the 4th loading
cycle saw a decreased cumulative energy gradient. This was most likely due to the
small amount of crack growth, which was observed from the NDT measurements.
However, a large amount of cumulative energy was observed in Cluster 1 during this
period, which may indicate increased background AE activity.
Cluster 1 however shows minimal consistency with the crack growth, particularly
with low energy hits and the outlying hit energy spikes previously mentioned. This
demonstrates that the SOM has the ability to differentiate evident outliers from signals
consistent with damage. Furthermore, using the SOM to cluster the data for Channel
3 shows consistent results with the clustered data from Channel 1.
7 Conclusion
A technique for clustering damage and non-damage AE signals from a mechanically

tested railway axle test specimen using a SOM has been presented. The technique
successfully identified apparent outliers based on seven different AE attributes. This
resulted in a data cluster that showed good consistency between the cumulative
energy and the crack length.
The results presented are from an early stage study to determine the feasibility
of the technique. More testing is planned to further refine the acquisition of the
data and to further refine the SOM parameters. A whole dataset from the test was
used for clustering; however, this runs the risk of ‘over-training’. Therefore, further
work is required to assess a suitable training length. Furthermore, influences of other
SOM variables such as map size and number of training iterations are required. This
may enable establishing more clusters and further differentiate signals from different
sources.
Acknowledgements This work is funded by Innovate UK MONAXLE project, InnovateUK Grant

104243.
References
1. Eurostat—Railway passenger transport statistics—quarterly and annually, 02/07/2019,

https://ec.europa.eu/eurostat/statistics-explained/index.php/Railway_passenger_transport_sta
tistics_-_quarterly_and_annual_data
2. Civity Management Consultants (prepared for the Office of Rail Regulation) European
Benchmarking of the costs, performance and revenues of GB TOCs, 2012
3. Hoddinott DS (2004) Railway axle failure investigations and fatigue crack growth monitoring
of an axle. Proc Inst Mech Eng Part F: J Rail Rapid Transit 218(4):283–292
560 R. Marks et al.
4. AxleInspect Consortium (2013) Axleinspect (development of novel inspection techniques for

train axles). Final Report Summary. Cordis—EU research results
5. European Railway Agency (ERA) (2011) Railway safety performance in the European Union.
https://www.era.europa.eu/Document-Register/Pages/Railway-Safety-Performance-in-the-
European-Union-2011.aspx
6. Castro I (2013) Development of novel inspection techniques for train axles. In: Internationla
Wheelset Congress, Kiev
7. Rudlin JR, Shipp R (2002) Review of techniques for rail axle inspection. Rail Safety and
Standards (RSSB)
8. Beretta S et al (2005) Fracture mechanics and scal effects in the fatigue of railway axles. Eng
Fract Mech 72(2):195–208
9. Unnþórsson R (2013) Hit detection and determination in AE bursts. In: Acoustic emission-
research and applications. IntechOpen
10. Bhuiyan MY et al (2018) Toward identifying crack-length related resonances in acoustic
emission waveforms for structural health monitoring applications. Struct Health Monit
17(3):577–585
11. Crivelli D et al (2018) Gear tooth fatigue test monitoring with continuous acoustic emission:
advanced signal processing techniques for detection of incipient failure. Struct Health Monit
17(3):423–433
12. Bruni S et al. A preliminary analysis about the application of acoustic emission and low
frequency vibration methods to the structural health monitoring of railway axles. In: Chemical
engineering
13. Carboni M et al (2013) A study on the performance of acoustic emission and low frequency
vibration methods to the real-time condition monitoring of railway axles. In: Proceedings of the
12th International conference of the Slovenian Society for non-destructive testing, pp 559–566
14. Carboni M, Gianneo A (2017) Feasibility of an Acoustic emission based structural health
monitoring for railway axle under fretting fatigue damage. In: 14th International conference
of the Slovenian society for non-destructive testing, pp 65–72
15. Zhou Y et al (2018) A new method to classify railway vehicle axle fatigue crack AE signal.
Appl Acoust 1(131):174–185
16. Kohonen T (2013) Essentials of the self-organising map. Neural Netw 1(37):52–56
17. SOMPy by V Moodavi, S Packman and I Vallés 4th March 2019
18. Crivelli D, Guagliano M, Monici A (2014) Development of an artificial neural network
processing technique for the analysis of damage evolution in pultruded composites with acoustic
emission. Compos B Eng 1(56):948–959
Acoustic Emission Monitoring
for Bending Failure of Laminated Glass
Used on Glass Water Slide
Ran Liu, Gongtian Shen, Yong Zhang, Wei Zhou, and Pengfei Zhang
Abstract In recent years, a new kind of amusement ride called glass water slide
has been rapidly increasing. As a key component of this ride, laminated glass may
be subject to all kinds of applied loads, which impose safety risks. In addition, the
test data and evaluation method for laminated glass used on glass water slide is
not enough. To address this issue, a method combining acoustic emission (AE) with
digital image correlation (DIC) is used to monitor the damage and deformation under
three-point bend loading conditions. The results show that the lower tempered glass
is broken when loaded to the maximum bending load (about 16kN), then the load
dropped rapidly. Afterward, the bending load is restored to about 8kN, at the moment,
the upper tempered glass is broken, which indicated the tempered glass completely
failure. For the laminated glass with bonding defects, more AE signals with higher
amplitude value are generated during the load process. The strain concentration and
severe displacement deformation are observed when the upper and lower parts of the
glass fail. The distribution characteristics of the displacement and strain fields can
be well correlated with the amplitude and frequency distribution of the AE signals,
revealing the whole process of the damage evolution of the laminated glass. It is
suggested that the complementary technology combining AE with DIC is effective
for damage monitoring and evaluation of the laminated glass.
Keywords Laminated glass · Three-point bend · AE · DIC
R. Liu · G. Shen (B) · Y. Zhang

China Special Equipment Inspection and Research Institute, Beijing, China
e-mail: shengongtian@csei.org.cn
W. Zhou · P. Zhang
College of Quality and Technical Supervision, Hebei University, Baoding, China

https://doi.org/10.1007/978-981-15-9199-0_53
562 R. Liu et al.
1 Introduction
A new kind of amusement ride called glass water slide has recently appeared in China.
In more than two years, there are around 1000 glass water slides and throughout many
attractions. No applicable standards can be used to guide the design, manufacture
and operation, many injuries and accidents occurred. Consequently, the security of
glass water slide should be attracted great attentions.
The main material of glass water slide is laminated glass, however, the test data
and evaluation method for laminated glass used on glass water slide was not been
reported. Laminated glass is widely used in various industries [1–3], especially in
the construction industry [4, 5].
Besides, laminated glass is a kind of typical brittle material, whose fracture and
destruction are sudden, so the safety and stability of laminated glass have always
been the focus of research. Early, the research on laminated glass mainly focused on
the material properties [6, 7] and static experimental analysis [8, 9]. In recent years,
more and more researches on laminated glass dynamic loading are reported. Van
Dam [10] verified the effect of the interlayer on mechanical behaviour and proper-
ties by impact test and finite element analysis. Zhang [11] conducted an explosion
experiment on laminated glass panel and discussed dynamic response characteris-
tics and failure modes under different explosive equivalents. Furthermore, Qiu [12]
achieved to monitor and control of brittle material damage using acoustic emission
(AE), suggesting that the AE could be used to monitor the laminated glass damage.
In this work, laminated glass of used glass water slide with defects and no
defects were prepared. The three-point bending tests of as-prepared specimens were
performed by the universal testing machine, AE with digital image correlation (DIC)
was used to monitor the damage and deformation of laminated glass during the whole
loading progress. Of particular significance was that the characteristics combining
AE with DIC of specimens were discussed.
2 Experimental Details
2.1 Specimen Preparation
The laminated glass used in the experiment consists of two tempered glass substrates
and a PVB interlayer. Among them, the samples with the qualified quality were set as
Specimen A, and samples prefabricated with bubbles and degumming defects in the
interlayer were set as Specimen B. Each type of samples with the thickness of 16 mm
were made into the size of 300 mm × 300 mm, as shown in Fig. 1. Subsequently, a
black/white paint was randomly applied to the side of the specimens for measuring
the displacement and strain fields [13].
Acoustic Emission Monitoring for Bending Failure of Laminated Glass … 563
Fig. 1 Two types of laminated glass typical specimens (Specimens A and B)
2.2 Three-Point Bending Tests and AE Monitoring
The sample was subjected to a three-point bending test on a mechanical test machine
at a speed of 2 mm/min [14]. Two types of specimens were placed on two support rolls
with a span of 260 mm, as shown in Fig. 2. During the three-point bending test, AE
signals derived from internal damage were obtained using an AE detection system
Fig. 2 Three-point bending experimental measurement system

564 R. Liu et al.
(DS-2A) while the speckle images of the sample were collected by a CMOS camera
(MER-500-14U3M). The frequency range of the two AE sensors were 100–900 kHz
and the preamplifier were 40 dB. In the cause of forming a good acoustic coupling
between the sample and the sensor, silicone grease was used as a coupling agent,
and the coupling performance was verified by a pencil break AE experiment. The
threshold was fixed at 10 mv (40 dB) by repeated testing to eliminate electrical and
mechanical noise interference. Finally, the sampling frequency was set to 3 MHz [15].
In addition, speckle image acquisition was performed on laminated glass material
samples under bending load using a CMOS camera with a resolution of 2592 × 1944
pixels.
3.1 Mechanical Performance Analysis
Three point bending tests were carried out for two kinds of laminated glass materials
Specimen A (without defects) and Specimen B (with defects), respectively, and the
load–displacement curves of specimens were shown in Fig. 3. Point A was set for
observing the damage of the specimen in the linear stage. For purpose of analysing
the damage mechanism of the two-layer glass plate, Points B and C were set at
two-inflection points, respectively. Before the failure, the load and deformation of
the two kinds of specimens basically showed a linear relationship. When the load
reaches a certain value, the specimens suddenly burst, showing obvious brittleness.
It can be interestingly found from Fig. 3 that the maximum loads of Specimen A
and B were 15.55 kN and 14.22 kN, respectively, and then the load of the laminated
Fig. 3 The load-deflection

curves for the laminated
glass specimens
Fig. 4 The damage morphology of laminated glass
glass materials decreases rapidly. However, the load of the two kinds of samples were
again raised to about 8 kN. This may be because the specimen used in this experi-
ment consists of two layers of glass plates, and the two layers of glass plates may
fail separately during the loading process, causing two load rises. By calculation, the
bending strengths of the Specimen A and B were 78.71 MPa and 72.21 MPa, respec-
tively. Therefore, it can be found that the mechanical properties of the Specimen A
were better than that of the Specimen B, indicating that the prefabricated void defect
reduces the load-carrying capacity of the laminated glass material, but the effect was
limited.
The damage morphology of the laminated glass material under three-point
bending loading was shown in Fig. 4, and it can be found that many cracks extend in
the longitudinal direction. In addition, it can observe that there was a certain differ-
ence in crack propagation between the two layers of glass. The cracks in the upper
layer glass plate were initiated by the external surface and spread downwards. This
may be due to the external surface of the upper layer glass was directly subjected
to a force, and typical pressure damage occurs. The crack of the lower layer glass
starts from the lower surface because the bottom was subjected to tensile stress, and
it was observed that the lower layer glass plate has more cracks than the upper layer
glass plate, which was attributed to the effect of tensile stress greater than compres-
sive stress during flexural loading. Moreover, Specimen B was inferior in damage
resistance to Specimen A because it had prefabricated defects such as voids, and
Specimen B exhibited more cracks and greater degree of deformation in Fig. 4.
566 R. Liu et al.
3.2 AE Characteristics Analysis
According to the AE relative energy versus time curve of Specimen A and B in

Fig. 5, it was found that there was almost no increase in the AE energy in the
initial loading of the laminated glass material, indicating that the early linear stage
specimen has no obvious damage. As shown in Fig. 5a, the change of AE relative
energy was consistent with the mechanical behaviour of the material, and higher
energy values appear at the inflection points of the mechanical curves corresponding
to the cracking of the upper layer and lower layer glass plates. In addition, it can be
observed that the high-energy AE signal was mainly distributed between the two-
inflection points B and C, indicating that the main damage of the laminated glass
material was concentrated during the final burst of the samples. For Specimen B,
the distribution characteristics of the AE relative energy were similar to those of
Specimen A, as shown in Fig. 5b. However, after the maximum load was reached,
that is, during the burst of the specimen (from the failure of the upper layer glass
plate to the failure of the lower layer glass plate), more high-energy AE signals were
generated, and the highest energy reaches 2000 mV*mS, while the maximum AE
energy of Specimen A was 1400 mV*mS, indicating that Specimen B with defect
was more seriously damaged than specimen A the qualified quality.
For purpose of studying the AE signals associated with the damage, the change
course of the AE energy and duration generated by the damage was analysed, as
shown in Fig. 6. In general, the duration of AE signal may be related to the type of
damage, and the AE energy was often associated with the severity of the damage.
According to Fig. 6a, b, the tendency of the AE energy to rise as a function of
duration increases generally indicates that the degree of damage of the specimen
was intensified. However, a large number of AE signals with low duration and high
energy also appeared, which was consistent with the phenomenon that the specimens
suddenly burst in the experiment, indicating that as a brittle material, laminated glass
has a large number of brittle fractures.
Fig. 5 The AE energy and load with increasing time curve of laminated glass specimen
Fig. 6 AE energy variation with duration of laminated glass specimen
Fig. 7 The AE cumulative count and amplitude of a Specimen A, b Specimen B with increasing
load
Figure 7 shows the variation of the force-cumulative count-amplitude over time

for the laminated glass specimen under three-point bend loading. In the linear stage
of loading, the deformation was small, only a few signals appear, and the curve of
cumulative count has almost no change, which was consistent with the results of the
above analysis. When the load was increased to point B as shown in Fig. 3, the lower
layer glass plate was broken. The bearing capacity of the specimen was lowered, and
the load decreases sharply. Meanwhile, a large number of high-amplitude AE signals
began to appear, and the number of cumulative count began to increase dramatically.
The abundant AE signal reveals the appearance of severe damage, combined with
the damage morphology of the specimens in Fig. 4, which may be a concentrated
fracture of the fibres and resin. During the fragmentation of specimens, the high-
amplitude AE signal was distributed intensively, and the cumulative count continues
to rise, which means that many damage modes such as matrix cracking, fibre pulling,
fibre breakage and overall fracture occur. In addition, comparing Fig. 7a, b, it was
observed that Specimen B has more high-amplitude signals, and the cumulative count
rises higher. The AE signals originating from the damage reveals the entire failure
568 R. Liu et al.
process, indicating that Specimen B with defects suffered more damage during the
loading process.
3.3 Displacement Field and Strain Field Distribution

of Laminated Glass
Figures 8, 9 and 10 described the displacement field changes of the laminated glass
material specimens. In order to clarify the damage evolution of specimens in the
process of bending loading, combined with the bending load curve of specimens
in Fig. 3, the full-field deformation of point A, B and C was analysed, respectively.
According to the changes of the displacement field in the x and y directions of Fig. 8a,
b, it can be seen that the displacement deformation values of Specimen A and B were
small at this time and exhibited similar distribution characteristics. In the x direction,
the deformation values of Specimen A and Specimen B were small and distributed
on the external surfaces, as well as the interlayer of the two layers of glass plates.
In the y direction, the displacement fields of Specimen A and B exhibit a ring-like
distribution, and the value of displacement deformation in the loading centre was
Fig. 8 Displacement field of a Specimen A and b Specimen B at point A
Fig. 9 Displacement field of a Specimen A and b Specimen B at point B

Fig. 10 Displacement field of a Specimen A and b Specimen B at point C
the largest, and then decreases to both sides. Due to y direction was the main bear-
loading direction in bending loading progress, the displacement in y direction was
much larger than that in x direction.
As the load increases to point B, the displacement field distribution of Specimen
A and Specimen B in both directions was basically consistent with the characteris-
tics described previously, while the displacement field value in y direction increases
significantly, corresponding to the cracking of the lower layer glass plate. In addition,
the displacement field value of Specimen B in both the x direction and the y direction
was significantly larger than that of Specimen A, indicating that the damage defor-
mation of the specimen with defects was larger than that of the specimen without
defects in the loading process, which was consistent with the damage morphology
reflected in Fig. 4.
The displacement field of Specimens A and B of laminated glass material when the
load increases to point C was shown in Fig. 10. The displacement field values of two
kinds of sample increase dramatically in both directions. However, the distribution of
displacement field in the x direction was more integrated than before. The laminated
glass material was supported by two layers of glass plates. After the lower glass
plates were broken, the two layers of glass plates were pressed tighter, and the upper
and lower layers of glass plates begin to bear the load together, which also reveals
why the load will rise again to about 8 kN after a sudden drop.
On basic of the distributions of displacement fields of two types of specimens
from Figs. 8, 9 and 10, the corresponding distributions of strain fields of laminated
glass specimens was calculated by using Eq. (1), as shown in Figs. 11, 12 and 13,
respectively.
m m
i=1 j=1 f x i , y j − f g x i , y j − g
C(u, v) =
2 (1)
m m 2 m m
i=1 j=1 f x i , y j − f i=1 j=1 g x i , y j − g
570 R. Liu et al.
where f (x, y) and g (x , y ) were the grey value of the reference image and the
deformed image, respectively. On basic of the maximum correlation coefficient C,
the deformed image was calculated to obtain the strain field.
In the linear stage, the strain values of Specimen A and Specimen B were small,
and the difference was not significant. At point A, the strain values in both directions
were relatively close, indicating that the specimen was under uniform stress at this
time. However, in the x direction, the full strain field of the two kinds of specimen
Fig. 11 Strain field of a Specimen A and b Specimen B at point A
Fig. 12 Strain field of a Specimen A and b Specimen B at point B
Fig. 13 Strain field of a Specimen A and b Specimen B at point C

was randomly distributed, while in the y direction, a higher strain value appeared at
the interlayer position of the two layers of glass plates.
When the load was continuously increased to point B, the strain field distribution
of Specimen A did not change much in the x direction, and only the value of the
strain slightly increased. For Specimen B, in addition to the increase in the value of
the strain field, a plurality of strain concentration bands extending upward from the
bottom appeared at the bottom of the lower layer glass plate, which was consistent
with the above-described mechanical analysis. In the y direction, Specimen A and
B showed significant strain concentration at the interlayer of two layers glass plates,
which indicates that this region was a weak part of the entire load-bearing process.
Therefore, the manufacturing technology of laminated glass greatly affects the safety
performance of amusement facilities such as glass water slides.
Figure 13 showed the strain field of Specimen A and B at the point C. The
maximum strain values of Specimen A and B in the x direction reached 4.701%
and 5.481%, respectively. Some strain concentration bands appeared on the upper
layer glass plate, which showed fragmentation of the upper glass plate. The strain
band shown extends from the surface to the centre, consistent with the description
of the damage feature. The maximum strain values of Specimen A and B in the
y direction were 1.754% and 2.222%, respectively, which were distributed on the
interlayer of the two layers glass plates. In addition, the strain value of Specimen B
was significantly larger than that of Specimen A, indicating that the damage of the
specimens with defects was more serious. The comprehensive displacement field and
strain field information can clearly reflect the damage deformation characteristics of
laminated glass materials.
4 Conclusions
In this study, mechanical properties, AE signal response characteristics, strain and

deformation fields of laminated glass specimens were investigated in detail. Some
important conclusions are summarised as follows:
1. The void defects result in the reduction of mechanical properties of the laminated
glass. The laminated glass specimen showed obvious brittleness, and two layers
of glass plates fail separately during the loading process, causing two load rises.
2. The abundant AE signals were mainly distributed between the inflection points
of B and C, indicating that the main damage of the laminated glass material was
concentrated in the final burst process. Some AE characteristic parameters such
as AE energy, duration, AE cumulative count and amplitude reflected the damage
evolution and failure behaviour of laminated glass.
3. The concentration of displacement and strain occurred at the interlayer of the
two-layer glass plate by DIC measurement. In addition, the displacement and
strain of specimens with defects were significantly larger than that of specimens
with the qualified quality.
572 R. Liu et al.
4. The method combining AE and DIC can effectively monitor the damage
process of the laminated glass, and provide timely safety evaluation and damage
prediction for amusement facilities such as glass water slides.
Acknowledgements The authors gratefully acknowledge the financial support of the project of
China Special Equipment Inspection and Research Institute (grant no. 2018qingnian04).
References
1. Luthy H, Filser F, Loeffel O, Schumacher M, Gauckler LJ, Hammerle CHF (2005) Dental
materials, vol 21, pp 930–937
2. Studart R, Filser F, Kocher P et al (2007) Dent Mater 23:177–180
3. Qiu Y, Bao YW, Liu YH (2006) Key Eng Mat 324–325:367–370
4. Belis J, Depauw J, Callewaert D et al (2009) Failure mechanisms and residual capacity of
annealed glass/SGP laminated beams at room temperature. Eng Fail Anal 16(6):1866–1875
5. Bennison SJ, Smith CA, Van Duser A et al (2002) Structural performance of laminated glass
made with a “Stiff” Interlayer. In: Use of glass in buildings, pp 57–65
6. Xin W, Li KF, Fan SJ et al (2013) Spectral properties of ~2 µm emission of Tm3+ doped
Bi2 O3 -SiO2 -PbO glass. J Inorganic Mater 28(2):165–170
7. Zhao XL, Liang XJ, Luo HY et al (2013) Third-order nonlinear optical properties of silver
quantum dots doped in third-order nonlinear optical properties of silver quantum dots doped
in sodium borosilicate glass. J Inorganic Mater 28(9):1003–1008
8. Xu XQ, Xu J, Chen J-J et al (2017) Investigation of dynamic multi-cracking behavior in PVB
laminated glass plates. Int J Impact Eng 100:62–74
9. Bao Y-W, Liu Z-Q (2016) Mechanism and criterion of spontaneous breakage of tempered glass.
J Inorganic Mater 31(4):401–406
10. Van Dam S (2017) Experimental analysis of the post-fracture response of laminated glass under
impact and blast loading. Ph.D. Dissertation in Ghent University
11. Zhang Q-L, Tao Z-X, Wang X et al (2013) Experimental study on dynamic response of
laminated glass curtain wall under explosive action. J Build Struct 34(4):74–80
12. Qiu Y, Bao Y-W, Liu X-G et al (2011) Monitoring and control on the damage of brittle materials
achieved by using universal testing machine and acoustic emission equipment. Eng Test 12:7–
10
13. Brynk T, Laptiev A, Tolochyn O et al (2012) The method of fracture toughness measurement
of brittle materials by means of high speed camera and DIC. Comput Mater Sci 64:221–224
14. Ajitanshu V, Pawar SJ et al (2018) Experimental and simulation studies on flexural strength
of laminated glass using ring-on-ring and three-point bending test. In: Proceedings of the
Institution of Mechanical Engineers, vol 21, pp 3930–3941
15. Grégoire D, Verdon L et al (2015) Mesoscale analysis of failure in quasi-brittle materials:
comparison between lattice model and acoustic emission data. Int J Numer Anal Meth Geomech
39:1639–1664
Source Location of Acoustic Emission
for Anisotropic Material Utilizing
Artificial Intelligence (WCCM2019)
Y. Mizutani, N. Inagaki, K. R. Kholish, Q. Zhu, and A. Todoroki
Abstract AE testing (AT) is one of the nondestructive inspection methods using

elastic waves (AE) generated by cracks or deformation of materials. The location of
sources of AE is determined by arrival time differences to the AE sensors attached on
the structures (AE source location). In this study, only one three-axis accelerometer
is used for two-dimensional source location on anisotropic thin plate as an alternative
to attaching a number of general AE sensors to the plate. As a result, arrival time
differences that required for traditional AE source location algorighm cannot be
obtained. Convolution neural network (CNN) which is one of the AI is utlilized
to realize source location in this study. Even used only one accelerometer, source
location accuracy is better than that by using a number of AE sensors.
Keywords Acoustic emission · Source location · Three-axis accelerometer ·

Convolutional neural network (CNN) · Anisotropic material · CFRP
1 Introduction
Acoustic emission (AE) refers to elastic waves generated by cracks, deformations and
transformations of materials. AE testing (AT) is one of the nondestructive inspection
methods using these AEs detected by high-sensitive sensors to evaluate material
or structural integrity. During the test, the location of sources of AE is determined
by arrival time differences to the AE sensors attached on the structures (AE source
location). When the targets of the AE testing are made of isotropic material, AE
velocity does not have orientation dependence, and AE source location can be easily
conducted. On the other hand, when the targets are made of anisotropic material,
Y. Mizutani (B) · N. Inagaki · K. R. Kholish · Q. Zhu · A. Todoroki

Department of Mechanical Engineering, Tokyo Institute of Technology, 2-12-1-I1-70, Meguro-ku
Ookayama, Tokyo, Japan
e-mail: ymizutan@mes.titech.ac.jp
K. R. Kholish
Faculty of Mechanical and Aerospace Engineering, Bandung Institute of Technology, Bandung,
Indonesia

https://doi.org/10.1007/978-981-15-9199-0_54
574 Y. Mizutani et al.
special consideration is required to realize source location. For example, the author
et al. proposed the location method for thin anisotropic plate that use of arrival time
for specific frequency components of AE which anisotropic effects can be neglected
[1].
Recent years, artiicial intelligence (AI) began to be utilized for wide range of
engineering fields. A. Ebranhimkhanolou [2] utilized AI for source location of AE
on reveted metallic panels. Asif [3] utilized AI to evaluate quality of welding by
AEs detected during the process. The author et al. [4] conducted zone location for
artificial AE source and evaluated the capability of AI for this purpose.
As mentioned above, source locations are generally determined by using arrival
time differences to the AE sensors. Even for the source location for isotropic mate-
rials, two sensors for one-dimensional source location, three for two dimensions and
four for three dimensions are minimumly required. In this study, only one three-axis
accelerometer is used for two-dimensional source location on anisotropic thin plate
as an alternative to attaching a number of general AE sensors to the plate. As a
result, arrival time differences that required for traditional AE source location algo-
rithm cannot be obtained in this study. Convolution neural network (CNN) which is
one of the AI is utlilized to realize source location in this study.
2 Experiment
2.1 Measurement of AE
Unidirectional carbon fiber reinforced plastics (CFRP) plate with aluminum liner
is prepared as a specimen. The total thickness of the plate is 2 mm. One three-
axis accelerometer (Ono Sokki, NP-3560B) with a resonance frequency of 55 kHz
was attached to the CFRP surface of the plate. As shown in Fig. 1, an equilateral
triangle with 200 mm side length is drawn on the CFRP surface so that the position
of accelerometer () becomes the center of the bottom line of the triangle. At the
48 points inside the triangle, artificial AEs are generated 10 times at each point
by impacting rubber with hemisphere shape. Then, we obtain a total of 480 AEs
by the accelerometer. Output of accelerometer sensor which is the acceleration of x,
y, z directions (z is the direction vertical to the plate) is digitized by digitizer (Pico
Technology, type: 4424) with sampling time of 34.95 µs and the sampling number
of 512 points. The data with 512 × 3 matrix is saved as a csv file.
2.2 Use of Nueural Network
A convolution neural network (CNN), which is one of the most powerful AI tools, is
used in this study. Neural network console which is opened to public by Sony Network
Source Location of Acoustic Emission for Anisotropic Material … 575
Fig. 1 Location of accelerometer () and the source AE on specimen
Communication Inc. is used for developing the structure of CNN. Figure 2 shows
the structure of the CNN constructed for this study. This structure was constructed
by modifying “LeNet.sdcproj” which is a sample project of neural network console.
This sample project is a representing CNN in the famous academic paper [5] by
Dr. Yann LeCun et al. who made the boom of the deep learning in recent years. In
this study, we put “Reshape” layer to convert accelerations in three directions (512
points wave × 3) to 3 × 512 pixel image data. The difference between CNN and
general NN is that characteristics of the images were extracted by convoluting the
filter (filtering).
In the CNN shown in Fig. 2, the filtering was conducted twice (shown as “C:
Convolution” in the figure”). After the filtering, we obtained 16 images that extracted
features of original images, but the size was reduced to 3 × 32 or 3 × 16 pixels by
pooling (shown as “M: Max pooling” in the figure). Another important layer in Fig. 2
is that of ReLU layer instead of using general sigmoid layer as an activation function.
It is known that ReLU is best function among general activation functions to complete
the learning with short time [6]. Suitable filters are also decided during the learning
phase. We set maximum epoch as 300 and batch size as 64 for the learning phase of
the NN. Batch size is the number of data extracted from the entire data used for one
learning. The number of repetitions of learning is called as the epoch number. 384
data (80%) of 480 data were used for learning, and remained 96 data (20%) were
used for the verification.
Fig. 2 Configuration of CNN
3 Result and Discussion
Figure 3 shows the result for the source location by using verifica-
tion data. The symbol shows the given locations and the shows
the estimated locations. Table 1 shows the average and maximum
errors of the source location results. For the comparison, the errors in
reference 1 are shown in the same table. Specimen used in this reference is
Source Location of Acoustic Emission for Anisotropic Material … 577
Fig. 3 Source location results using CNN with single three-axis accelerometer
Table 1 Errors for the source location

Configuration Average error (mm) Maximum error (mm)
CNN with 3-axis accelerometer 3.41 10.0
Kurokawa [1] 3.9 12.0
same as that used in this study. Although, the source for AE is different from this
study. In the reference 1, three AE sensors which attached at the corner of the
triangle were used and the specific frequency components were extracted before
estimating the source location. In this study, even we could not obtain arrival time
differences as we attached only one sensor on the specimen and no signal processing
were performed, location results are better than that in reference 1.
Figure 4 shows AE initial oscillation direction in x–y plane. Early vibration direc-
tions that shown by double arrows match the direction to the AE source from the
accelerometer. AI may use such kind of information to estimate AE source location.
In the future, we are going to clarify the basis of estimation of CNN by visualizing
the filter in convolution layers in CNN.
Fig. 4 Vibration direction in x–y plane due to the arrival of AE. (Left figure: At position 8 in Fig. 1.
Right: At 12. Single arrow shows trajectory of acceleration in time series, and double arrow shows
vibration direction with initial stage of AE)
4 Conclusion
A method of AE source location using a single three-axis accelerometer powered by

AI was proposed and compared with the conventional source location method using
a number of AE sensors. The main conclusions are as follows.
• The accuracy of AE source location by proposed method in this study is better
than that in reference 1 which is proposed by our group previously using three
AE sensors and without using AI.
• AI may use information of initial oscillation direction in x–y plane due to AE
arrival.
References
1. Kurokawa Y, Mizutani Y, Mayuzumi M (2005) Real-time executing source location system

applicable to anisotropic thin structures. J Acoustic Emission 23:224–232
2. Ebrahimkhanolou A, Dubuc B, Salamone S (2019) A generalizable deep learning framework for
localizing and characterizing acoustic emission sources in reveted metallic panels. Mech Syst
Signal Process 130:248–272
3. Asif K, Zhang L, Derrible S, Indacochea E, Ozevin D, Ziebart B (2019) Using machine learning to
predict welding quality from airbone AE sensor data. In: Presented at 9th international conference
on acoustic emission, Chicago, Illinois, USA, 17–20 June 2019
4. Tsuchida F, Mizutani Y, Todoroki A (2018) Zone location and classification of AE source by
AI. In: Proceeding of 24th international acoustic emission symposium, Sapporo, Japan, 5–9
November 2018, pp 291–296
5. LeCun Y, Bottou L, Bengio Y, Haffner P (1998) Gradient-based learning applied to document
recognition. In: Proceedings of the IEEE, November 1998.
6. LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436–444
Standard Development and Application
on Acoustic Emission Condition
Monitoring of Amusement Device
J. J. Zhang, G. T. Shen, Z. W. Wu, and Y. L.Yuan
Abstract Rotating structure, as the most important part of the amusement device, is
difficult to be disassembled after its installation. Acoustic emission (AE) technology
is an effective tool for condition monitoring and fault diagnosis of rotating machinery.
Acoustic emission tests for non-defective and defective rolling bearings were studied.
The analysis and evaluation method of AE condition monitoring on amusement
devices were introduced. Application test on a rotating amusement device was carried
out. The results indicated that the AE technique can be applied to the condition
monitoring and fault diagnosis of the amusement devices.
Keywords Acoustic emission · Condition monitoring · Rotating structure ·

Amusement device · Standard
1 Introduction
In recent years, the amusement device industry has grown rapidly with increasing
number and higher parameters. By the end of 2018, the quantity of large-scale amuse-
ment devices is come to 25,100 in China, an increase of 3.7% over 2017. There are
more than hundreds of millions times of people experienced on the amusement
equipment each year. Rotating structure is the most important part of the amuse-
ment device which bearing tremendous dynamic loads and variable speeds. While
failure occurred on the rotating part, the risk is rather high and directly threatens to
the safety of passengers’ life. Due to the rotating part is difficult to be disassembled,
condition monitoring and fault prediction are very important for safe operation of the
amusement devices. At present, inspector only observes whether there is abnormal
noise of the rotating structure during its operation. This method does not have the
ability to discover early damage. Acoustic emission (AE) technology has advantage
for online detection and has potential of early damage forecasting. The defects in the
J. J. Zhang · G. T. Shen (B) · Z. W. Wu · Y. L.Yuan

China Special Equipment Inspection and Research Institute, Building 2, Hepingjie, Xiyuan,
Chaoyang District, Beijing, China

https://doi.org/10.1007/978-981-15-9199-0_55
580 J. J. Zhang et al.
rotating structure can be found in the comparison of the AE tests during its normal
operation, thereby achieving inspection without disassemble and reducing the costs.
To investigate the AE testing methods of rotating structure of amusement device,
the laboratory tests and field tests have been carried out. The AE characteristics of
rolling bearings with or without defects were obtained and compared in lab studies.
More than 30 amusement devices were tested for acquiring the AE characteristics of
rotating structure of amusement devices. On this basis, AE technique for condition
monitoring the amusement devices was proposed and developed a national standard
[1]. There are many research reports covering the results of laboratory research and
field application testing [2–6].
In this paper, the AE characteristics of non-defective rolling bearing and rolling
bearings with outer race defect were shown, respectively. The characteristic param-
eter and evaluation method of the defective rolling bearing were obtained through
the comparative analysis. The interpretation and evaluation method of AE condition
monitoring for amusement devices in the standard were introduced. The filed appli-
cation of an amusement device Top Spin was shown to demonstrate the AE testing
effect. The AE signals were collected during the normal operation of the device,
and the testing was conducted again after an interval. The RMS curves and spec-
trum features were comparative analyzed. Also, the AE condition monitoring results
were evaluated by the standard. The results of this study provide an effective method
for the condition monitoring and fault diagnosis of rotating structure of amusement
device.
2 Acoustic Emission Testing of Rolling Bearings
2.1 Test Rig and Instrument
The test rig consists of a rolling bearing operation device, a digital multi-channel
AE system AMSY-6 (Vallen-Systeme GmbH, Germany), sensors, preamplifiers,
computer (signal processing software), etc.
The rolling bearing to be tested is deep groove ball bearing of type 6107-Z which
contains inner ring, outer ring, rolling elements and retainer. Characteristics of the
test bearings are internal diameter ψ = 35 mm; external diameter ϕ = 62 mm;
thickness d = 14 mm.
For the AE data acquisition, the VS150-M resonant sensors with peak response
frequency of 150 kHz are employed, with an operating frequency range of 100 Hz–
450 kHz. The preamplifiers have 34 dB gain with 20–1000 kHz band pass filter. The
sampling rate is set to 10 MHz, and the sample time is 819.2 µs. Set the threshold
40 dB to remove background noise in the laboratory. The probe is directly attached
to the outer race of the bearing by a magnetic clamp. Figure 1 shows the test-rig.
AE testing was carried out, respectively, on one non-defective bearing and two
bearings with different defects. A simulated wearing defect of rectangular slot was
Standard Development and Application on Acoustic Emission Condition … 581
Fig. 1 Test rig
seeded onto the outer race of the bearing artificially and the slot widths of 5.0 mm
and 2.5 mm, respectively. The rotating speed of the rolling bearing is 24 rpm (the
operating circle is 2.5 s).
2.2 AE Characteristics of Rolling Bearings
2.2.1 Acoustic Emission Parameter Characteristics
For non-defective rolling bearing, AE signals appeared occasionally with no obvious

characteristics of AE parameter courses. When the rolling bearing is defective,
acoustic emission signals with high peak amplitude were appeared. Rolling bearing
with larger defect size has higher peaks in acoustic emission signals. For the non-
defective rolling bearing and rolling bearing with different defects, changes in counts,
duration time and rise time do not reveal obvious distinguishing features. RMS
exhibits distinct peak characteristics among all acoustic emission characteristics.
Some peaks occasionally appeared in RMS course when the non-defective rolling
bearing is running. For bearings with defect on outer race, RMS clearly shows the
peak phenomenon, and as the defect size increases, the RMS peak periodicity is more
(a) non-defective rolling bearing
(b) rolling bearing with outer race defect of 2.5mm wide
(c) rolling bearing with outer race defect of 5.0mm wide
Fig. 2 AE courses of rolling bearing
obvious. In addition, the amplitude of RMS also rises with bearing defects increasing
(Fig. 2).
2.2.2 Characteristics of RMS Average Changes
According to the analysis of acoustic emission parameter characteristics, RMS is

the characteristic parameter to evaluate the rolling bearing health status. In the unit
of bearing operation cycle, the RMS of all signals in each cycle is accumulated and
divided by the sum of the hits to obtain the average value of the RMS of the cycle.
Then, the RMS average curve of the bearings with the running cycles was obtained,
as shown in Fig. 3. As can be seen in the figure, the RMS average curves of the three
bearings can be clearly distinguished.
Fig. 3 RMS average curve
3 Analysis and Evaluation Method of AE Condition

Monitoring of Amusement Devices
3.1 AE Data Analysis
According to laboratory research results, RMS and energy were discovered as the
characteristic parameters for the AE monitoring and fault diagnosis of rotating struc-
tures. The results are evaluated based on the changes of the monitoring curves of
characteristic parameters. Baseline data need to be set for comparison with future
tests to find the changes. Baseline data were obtained during the normal operating of
the amusement device which represent the initial stable conditions of the device. For
new or repaired devices, baseline data should be obtained under the stable operation
after the running-in period. For the devices that have been running for a long time,
the data acquired by the first monitoring can be used as a reference stable point for
the baseline.
3.2 Evaluation Method
The monitoring results can be divided into four levels according to the ratio K of
the average of the monitoring curve to the baseline data. See in Table 1. The value
of a, b and c is determined by experiment which reflects the damage degree of the
rotating structure. As the AE level of monitoring results increases, the health status
of the rotating structure declines with more serious damage.
Table 1 AE levels of monitoring results

AE levels of monitoring results The ratio K of the average of the monitoring curve to the
baseline data
I K<a
II a≤K<b
III b≤ K≤ c
IV K>c
Table 2 AE levels of monitoring results of rolling bearing

AE levels of monitoring results The ratio K of the average of the defective curve to the
non-defective data
I K<1.1
II 1.1≤K<1.4
III 1.4≤K≤1.7
IV K>1.7
Take rolling bearing as an example of rotating structure, Table 2 gives the recom-
mended values for the classification of the monitoring results using the ratio of the
average of RMS curves of different defect degrees to the non-defective data (baseline
data).
4 Application on the Top Spin
4.1 Testing Object and AE Instrument
Top Spin is a kind of rotating ride that is popular in amusement parks. It is composed of
rotating arms, connecting spindle, gear bearings, support pillars, ship cabin, electric
control system and other parts. The cabin can be raised and rotated around different
axes to a top speed of 30 km/h and max acceleration of 2.5 g which giving passengers
a wonderful flight experience. The height of the Top Spin is 18 m, and the cabin
capacity is 40 passengers. The rotating structure which consists of the spindle and
gear bearings was the test object of the AE monitoring. AE on-line monitoring of gear
bearing on Top Spin was carried out in 2017 and 2019, respectively. The appearance
of Top Spin was shown in Fig. 4.
The AE system consists of an AMSY-6 digital multi-channel AE instrument,
sensors, computer (signal processing software), etc. For the AE data acquisition, the
resonant sensors of model VS150-RIC with peak response frequency of 150 kHz
were employed, with an operating frequency range of 100–450 kHz. The gain of the
Fig. 4 Appearance of Top Spin
built-in amplifier in sensors is 34 dB. The AE sensors were mounted on the steel
structure close to the rotating part, see in Fig. 5.
4.2 Testing Results
Start the Top Spin running at normal condition. The AE signals were acquired contin-
uously four running cycles in each test. Figure 6 shows the RMS courses of two AE
tests. For the tests in 2017 and 2019, the RMS average was calculated of each running
cycle. The fitting curves of RMS average for the two tests were obtained in Fig. 7.
The RMS curve of 2017 can be used as the baseline data due to it was the first test
of the device. The averages of the RMS curves for 2017 and 2019 were 10.51 and
11.52, respectively. Then, the ratio K of the average of RMS curve to the baseline
data can be calculated. K was 1.09, equal to 11.52/10.51. The monitoring result was
at level I according to Table 2. According to the test results, it is recommended to
implement acoustic emission monitoring again after 1 year.
Fig. 5 Sensors arrangement
(a) 2017 (b) 2019
Fig. 6 RMS courses of the two tests

Fig. 7 RMS average curves of the two tests
5 Conclusions
This paper summarized the research results of the AE monitoring and fault diagnosis
of amusement devices. The fault characteristic parameter RMS average was extracted
by the acoustic emission test of the non-defective and defective rolling bearings in the
laboratory. An acoustic emission monitoring and evaluation method based on RMS
average was obtained, and the standard is developed. The AE tests of gear bearing
on Top Spin gave a specific case of AE condition monitoring and evaluation. Some
conclusions and revelations were obtained of this study as following.
(1) For different rotating structures of amusement devices, laboratory and field tests
are required to obtain the abc value for evaluation of AE monitoring results.
(2) Field tests should be carried out of the same device at different time inter-
vals in the future. The experience of AE monitoring data analyzing could be
accumulated by comparing the changes of RMS curves with the baseline data.
Acknowledgements This study was financially supported by the research project of National key
research and development plan project 2017YFC0805704.
References
1. Condition monitoring and fault diagnostics of amusement device —Part 2: Acoustic emission
monitoring method, GB/T 36668.2 (2018)
2. Wu ZW, Shen GT, Yuan J, Wang Y (2011) Application of acoustic emission technique in detection
of giant wheel main spindle. Nondestructive Testing 33(5):39–42
3. Wu ZW, Shen GT, Yuan J (2011) Acoustic emission signals’ characteristics of the giant wheel
main spindle. Nondestructive Testing 33(9):29–32
4. Wu ZW, Shen GT, Zhang JJ (2012) Characteristics of acoustic emission signals in the rolling
bearing on giant wheel. In: Proceedings of the 30th European conference on acoustic emission
testing & 7th international conference on acoustic emission, pp 135–143
5. He CF, Zhang JJ, Shen GT, Wu ZW (2013) Acoustic emission characteristics of rolling bearings
for giant wheel. J Beijing Univ Technol 39(5):653–659
6. Shen GT (2015) Acoustic emission technology and application. Science press, Beijing, pp 8–12
Fault Diagnosis of Pitch Bearings
on Wind Turbine Based on Acoustic
Emission Method
Guang-hai Li, Yang ming An, Tao Guo, and Guang-quan Hou
Abstract Pitch bearing is one of the critical components of wind turbines. If it broke
down, it would lead to serious phenomena, such as falling leaves and falling towers.
Therefore, the fault diagnosis of the pitch system bearing is important. Acoustic
emission technology has been widely applied to the bearings fault diagnosis. It could
capture the dynamic information of the pitch bearing when the bearing generated
microscopic deformation and crack propagation. A fault diagnosis method of pitch
bearing based on acoustic emission technology is proposed in this paper. Firstly,
the acoustic emission signal with obvious cracking characteristics is selected from
collecting signals by the kurtosis value. Then, the selected acoustic emission signal is
processed by the wavelet spectrum theory to extract the fault feature and identify the
fault pattern of pitch bearings. Lastly, an online monitoring system of pitch bearings
based on acoustic emission method is developed. The real application results showed
that the system could online monitor the work condition of the pitch bearing. The
crack fault of pitch bearing was identified by wavelet packet transform analysis and
time–frequency spectrum analysis under continuous wavelet transform.
Keywords Pitch bearing · On-line monitoring · Crack fault · Acoustic emission ·

Wavelet transform
1 Introduction
Pitch bearing is an important component of the wind turbine to realize the pitching
operation. It is installed between the hub and the blade by bolts, and its function is to
transfer the blade load to the hub. The work condition of the pitch bearing is harsh,
such as the frequent start-stop, high transmission torque, and ratio. As the angle range
G. Li (B)
e-mail: liguanghai@tsinghua.org.cn
Y. An · T. Guo · G. Hou
Bei Jing Nan Yang Si Yuan Intelligent Technology Co., Ltd., Beijing, China

https://doi.org/10.1007/978-981-15-9199-0_56
590 G. Li et al.
of pitching motion is 0–90°, the rolling elements of pitch bearing do not roll over the
total race. Then, the ring of the pitch bearing is prone to overload phenomenon. It
leads to the cracking fault. Thus, the fault diagnosis of the pitch bearing is necessary.
At present, most of the researches on pitch bearings are only based on the applica-
tion of the basic Hertz theory or the general analysis by using finite element analysis
software [1–3]. However, there is relatively little research on the online monitoring
of pitch bearings for wind turbines. In this paper, an online monitoring system is
proposed.
Acoustic emission (AE) method is often utilized to detect defect changes in the
material and it has been widely applied in the fault diagnosis of the wind turbine.
Zhang et al. [4] analyzed the fault characteristic of the blade material, which caused
by the bending effect, based on the AE technique. Zhu et al. [5] extracted the fault
features from the AE signal of the gearbox in a wind turbine by combining the eemd
with the fuzzy entropy method and obtained a good result. Shi et al. [6] proposed
an AE signal processing method based on the wavelet packet energy spectrum to
monitor and evaluate the fatigue damage degree of early material structures. Han
et al. [7] developed an online monitoring system to detect the work condition of the
blade of the wind turbine. Shen et al. [8] proposed a fault detection method for wind
turbine based on the AE signal and designed a vibration monitoring system based
on ARM embedded system to realize the function of real-time monitoring.
These research works have established a solid foundation on the fault diagnosis
of the pitch bearings in the wind turbine.In this paper, a real-time monitoring system
based on the AE technique is developed to detect the cracking fault of pitch bearing.
Firstly, WIFI communication technology is used to collect the AE signal. Then, the
impulse signal is extracted from the collected signal by kurtosis value. Finally, the
wavelet time–frequency spectrum and wavelet packet energy spectrum were used to
identify the cracking fault of the pitch bearings. The effectiveness of the proposed
method is validated by an experiment and real application.
2 Methodology
2.1 Kurtosis
As the kurtosis could effectively characterize the pulse character of a signal, it is

often used to be an important indicator for fault diagnosis of rotating machinery.
Considering the sensitivity of the kurtosis value to the impulse signal and the acoustic
emission is a transient elastic wave, the kurtosis is utilized as an indicator to extract
the AE signal of cracking fault of the pitch bearing ring from the noise environment.
The kurtosis is defined as follows:
E(u(t) − μ)4
k= (1)
σ4
Fault Diagnosis of Pitch Bearings on Wind Turbine Based … 591
where u(t) is the original signal, μ is the mean value, σ is the standard deviation, E
is the expectation operator.
2.2 Continuous Wavelet Transform (CWT)
The AE signal of a defective pitch bearing is similar to an impulse signal. The CWT is
a suitable choice to analyze such a non-stationary signal. Its principle is to decompose
the signal into a series of wavelet coefficients at different scales by wavelet basis.
The coefficient matrix of CWT is defined as:
+∞
Wc (a, b) = x(t)a,b (t)dt (2)
−∞
where x(t) is the signal, a is the scale, b is the translation, a,b (t) is the mother
wavelet:

1 t −b
a,b (t) = √ (3)
a a
It is very important to select a proper mother wavelet function. The Morlet wavelet
is selected as the mother wavelet to analyze the signal because of the shape of wavelet
is similar to the impulse component. The function of Morlet wavelet is defined as:
−1

e−iωt − e−ω0 2 e−t 2
2 2
(t) = π 4 (4)
The wavelet scalogram is defined by the wavelet coefficients:
SG(a, b; ) = |Wn (a, b; )|2 (5)
It provides a variable analysis window on a time–frequency plane with good local-

ization characteristics, can better describe the duration and frequency characteristics
of the AE signal.
2.3 Wavelet Packet Decomposition and Reconstruction
Wavelet packet method is an extension of wavelet decomposition. In the wavelet

packet decomposition, each approximation and detail coefficient is recursively
decomposed until the desired decomposition level is achieved. Assume that u(t)
is AE signal collected in the test. It is decomposed as the following equation [9]:
592 G. Li et al.
⎧ j+1,2n j,n
⎨ dk = h 2 j−k (k)dl
l
(6)
⎩ dkj+1,2n+1 = g2l−k (k)dl
j,n
j,n
where dl is the nth decomposed coefficient in the jth level, n = 0, 1, . . . , 2i − 1,
h is the low-pass coefficients, g is the high-pass coefficients, l and k are the filter
length and the number of wavelet coefficients, respectively.
The signal can be reconstructed by the following equation [10]:
j,n
j+1,2n+1 j+1,2n

dk = h k−2l dl + gk−2l dl (7)
l
In wavelet packet analysis, the details as well as the approximations are decom-
posed, and the time–frequency resolution of the signal is improved. Once the signal
was processed by the wavelet packet method, the energy of the signal is decomposed
into each frequency band. When the pitch bearing was defective, the energy distri-
bution would change obviously. The wavelet packet energy spectrum can clearly
reflect the energy distribution characteristics of the AE signal. Thus, the energy of
each frequency band is selected as an indicator to detect the fault of pitch bearings
in this paper.
3 Experiment
To study the characteristics of acoustic emission signals of crack initiation and

propagation of the material, a tensile experiment was carried out.
3.1 Tensile Experiment
The sample material is 42CrMo. It is processed according to the requirements of

GB/T 228.1–2010 “Metal material tensile test Part 1: room temperature test method".
The sample parameters are shown in Fig. 1. An AE144S sensor is used to collect
the signal with a sampling frequency of 2.5 MHz. The test sample is stretched by a
loading machine with 1 mm/min constant displacement.
3.2 Experimental Conclusions
The kurtosis trend during the experiment is shown in Fig. 2. To illustrate the effec-
tiveness of the kurtosis, two time series of different kurtosis values (one is lower than
Fig. 1 Parameters of the sample
60
50
40
Kurtosis
30
20
10
500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500
Number of Samples
Fig. 2 Kurtosis trend of the sample
15, the other is higher than 15) were analyzed by wavelet time–frequency spectrum
and wavelet packet energy spectrum. The analysis results are shown in Fig. 3.
As Fig. 3b shown, the time–frequency spectrum is stable and the main component
of the signal concentrates below 100 kHz.The energy of the first two frequency
bands account for more than 90% (0–156 kHz). The waveform of the AE signal with
kurtosis of 22.48 presented a burst characteristic. The main component of the signal
concentrated around 150–250 kHz and accounted for more than 85% of the overall
energy of the signal. For the signal with the kurtosis value lower than 15, its main
component and energy concentrated below 100 kHz. For the signal with the kurtosis
value higher than 15, its main component and energy of concentrated around 150–
250 kHz. The two kinds of the signal are corresponding to the sample stress release
and crack propagation stages. Thus, the crack propagation of the sample during the
tensile test can be effectively identified by the proposed method.
594 G. Li et al.
Kurtosis=5.33
0.15
0.1
Amplitude/V
0.05
-0.05
-0.1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time/ms Time/ms Band Sequence

(a) (b) (c)
Kurtosis=22.48
0.5 600 0.45
0.4 0.4
500
0.3 0.35
Frequency/kHz
0.2
Percentage /%
Amplitude/V
400 0.3
0.1
0.25
0 300
0.2
-0.1
200 0.15
-0.2
-0.3 0.1
100
-0.4 0.05
-0.5 0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 0 2 4 6 8 10 12 14 16
(d) (e) (f)
Fig. 3 Analysis results of the signal with different kurtosis values. a–c kurtosis value is 5.33,
d–f kurtosis is 22.48
4 System Design and Application
4.1 System Design
The system consists of hardware and software as Fig. 4 shown. The hardware mainly
consists of data acquisition analog circuit, A/D conversion circuit, FPGA logic
control module, and ARM core processing chip. The software is developed based
on Browser/Server structure. It includes three main functions: (1) Communication
with the collection terminal and the instruction issued; (2) User interaction and data
storage; (3) Remote access via the Internet.
Fig. 4 Framework of the system

Fig. 5 Diagram of system deployment
4.2 Analysis Result of Real Application Data
The AE system was installed at the wind turbine (Type 87/1500) in Pinglu, Shanxi
province. Considering the background noise of the wind turbine, the threshold value
of acoustic emission signal collection is 50 dB. The system deployment is shown in
Fig. 5.
Two groups data from the same position of the 1# pitch bearing and the 2# pitch
bearing were used to detect the work condition of the wind turbine. The kurtosis
trend of the data is shown in Fig. 6. By comparing the two datasets, it can be found
that the number of acoustic emission events of 1# pitch bearing is more than the 2#
pitch bearing, and the overall kurtosis value of the 1# pitch bearing is larger than the
2# pitch bearing. To better demonstrate the processed result, one portion signal of
each two datasets is selected to detect the cracking fault of the pitch bearing. The
analysis results were shown in Fig. 7.
For 1# pitch bearing, the original waveform presented obvious impact character-
istic. The energy of the signal is mainly concentrated on around 100 kHz. In addition,
a sideband of the energy appears near 200 kHz. The kurtosis value of the AE signal
of 2# pitch bearing is 3.36. The energy of the signal is mainly concentrated on around
80 kHz. As a comprehensive evaluation of the analysis results, 1# pitch bearing has
a risk of crack propagation, and 2# pitch bearing is ordinary stress release. Climbing
the tower and inspecting the pitch bearing, it was found that there was a cracking
fault of pitch bearing in the outer ring (as Fig. 8 shown).
1#Pitch bearing 2#Pitch bearing

80 12
10
60
Kurtosis
Kurtosis
8
40
6
20 4
0 2
0 1000 2000 3000 4000 5000 0 100 200 300 400 500 600 700
(a) Number of samples (b) Number of samples
Fig. 6 Kurtosis Trend of acoustic emission signal. a 1# pitch bearing, b 2# pitch bearing
596 G. Li et al.
Kurtosis=27.55
0.1 600 0.7
0.08
0.6
0.06 500
0.04 0.5
400
Amplitude/V
0.02 0.4
(a) 0
-0.02
300
0.3
-0.04 200
0.2
-0.06
100 0.1
-0.08
-0.1 0 0
0 0.5 1 1.5 2 2.5 3 3.5 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 2 4 6 8 10 12 14 16
Kurtosis=3.36
600 0.6
0.1 500 0.5
0.05 400 0.4

Amplitude/V
(b) 0 300 0.3
-0.05 200 0.2
-0.1 100 0.1
-0.15 0 0
0 0.5 1 1.5 2 2.5 3 3.5 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 2 4 6 8 10 12 14 16
Fig. 7 Anlysis results of signal processing. a Original waveform, time spectrum, and energy spec-
trum of 1# pitch bearing, b Original waveform, time spectrum and energy spectrum of 2# pitch
bearing
Fig. 8 Outer ring cracking of 1# pitch bearing
5 Conclusion
(1) A fault diagnosis method of pitch bearings on wind turbine is proposed in this
paper. The acoustic emission signal of bearing cracking is selected to detect the
bearing fault by the kurtosis values. The wavelet spectrum and energy spectrum
analysis were used to process the selected signal. The analysis result of the
experiment showed the cracking fault of the material was effectively identified
by the proposed method.
(2) A real-time monitoring system for the pitch bearing in the wind turbine was
developed. The real application results showed the system could detect the fault
of pitch bearing.
Acknowledgements This work was supported by a National Key R&D Program of China:
2018YFC0809004.
References
1. Amasorrain JI, Sagartzazu X, Damian J (2003) Load distribution in a four contact-point slewing
bearing. Mech Mach Theory 38:479–496
2. Zhou Z et al (2018) Stress and failure analysis of wind power pitch bearing. J Yangtze Univ
(Nat Sci Edit) 15(9):1–5
3. Tao G, Wang J, Wang Y (2015) Fracture failure analysis on pitch bearing of wind turbine. Hot
Work Technol 44(8):1–4
4. Zhang P et al (2016) Acoustic emission research of wind-power blades composites in three
point bending test. Eng Plastics Appl 44(4):1–6
5. Zhu J et al (2018) Research on fault diagnosis of planetary gearbox based on improved EEMD
and acoustic Emission technology. Noise Vib Control 38(Z1):1–4
6. Shi H, Li H, Wang Z (2019) Acoustic emission signal processing technology based on wavelet
packet energy spectrum. J Test Meas Technol 33(3):1–8
7. Han J (2010) Research of the wind turbine blade crack wireless monitoring system based on
acoustic emission technology. Beijing University of Chemical Technology, Beijing, pp 21–38
8. Shen X, Xi J (2017) Design of the monitoring system of wind turbines using rubbing acoustic
emission recognition. Chin J Electron Dev 40(4):1–6
9. Wang X et al (2015) Acoustic emission detection for mass fractions of materials based on
wavelet packet technology. Ultrasonics 60:27–32
10. Wang W, Yang S (2009) A method for choosing the Wavelet decomposition level in structural
fault analysis. Struct Environ Eng 36(4):1–8
Ultrasonic Testing
In Situ Condition Monitoring
of High-Speed Rail Tracks Using Diffuse
Ultrasonic Waves: From Theory
to Applications
K. Wang, W. Cao, and Z. Su
Abstract Real-time condition monitoring is a critical step to warrant the integrity

of rail tracks in bourgeoning high-speed railway (HSR) industry. Nevertheless,
existing damage identification, condition monitoring and structural health monitoring
(SHM) approaches, despite their proven effectiveness in laboratory demonstration,
are restricted from in-situ implementation in engineering practice. By leveraging
authors’ continued endeavours, an in situ health and condition monitoring framework,
using actively generated diffuse ultrasonic waves (DUWs) and a benchmark-free
condition-contrasting algorithm, has been developed and deployed. Fatigue cracks
in the tracks show unique contact behaviours under different conditions of external
loads and further disturb DUW propagation, and the crack growth induced by external
loads can also alternate DUW propagation. By contrasting DUW propagation traits,
fatigue cracks in rail tracks can be characterised quantitatively and the holistic condi-
tion of the tracks can be evaluated in a real-time manner. Compared with guided
wave- or acoustic emission-based methods, the DUW-driven inspection philosophy
exhibits immunity to ambient noise and measurement uncertainty, less dependence
on baseline signals, and high robustness in atrocious engineering conditions. Confor-
mance tests are performed on rail tracks, in which the evolution of fatigue damage is
monitored continuously and quantitatively, demonstrating effectiveness, reliability
and robustness of DUW-driven condition monitoring towards HSR applications.
Keywords High-speed railway · Fatigue crack · In situ structural health

monitoring · Diffuse ultrasonic wave · Condition contrasting
K. Wang · W. Cao · Z. Su (B)

Department of Mechanical Engineering, The
Hong Kong Polytechnic University, Kowloon, Hong Kong SAR
e-mail: zhongqing.su@polyu.edu.hk

https://doi.org/10.1007/978-981-15-9199-0_57
602 K. Wang et al.
1 Introduction
The integrity of rail tracks is of great importance for the railway industry, particu-
larly for the high-speed railways which support trains travelling with a speed over
200 km/h. During service lifetime, diverse factors, such as cyclic loads exerted by
moving trains, climate variation and unexpected events, can cause the generation of
fatigue damages, jeopardizing the integrity of rail tracks. To warrant the rail track
safety, non-destructive testing (NDT) methods [1, 2] have been developed and imple-
mented in railway industries. Nevertheless, inspection and maintenance using these
NDT techniques are timely consuming and labour intensive, and they cannot be
performed during the service of the railway system. Thus, these methods are inca-
pable of providing surveillance of the health status of rail tracks in an online and
in situ manner.
To monitor health status of rails continuously, a number of structural health moni-
toring (SHM) approaches [3–7] have been widely explored and investigated. Methods
based on the guided ultrasonic waves (GUWs) have been intensively attempted to
diagnose the fatigue damage. Damages in rail tracks can alter the ultrasonic wave
propagation, and by calibrating the damage-related wave scattering characteristics,
fatigue damages are assessed qualitatively and quantitatively. Nevertheless, conven-
tional GUW-based methods might lose effectiveness in practical applications. The
high complex cross section of rail tracks leads to challenges in wave signal anal-
ysis. More importantly, the detection philosophy that assesses the damage via the
comparison against baseline signals is not feasible, because in reality the interfer-
ence of noises induced by diverse factors, including the system malfunction, degra-
dation of instruments, and changes in operation environments, is such remarkable
that the damage-related features are obfuscated. This leads to the low reliability and
robustness of conventional methods, limiting their practical application.
To enhance the robustness of the health monitoring of tracks, we propose an
approach that exploits the diffuse ultrasonic waves (DUWs) and employs a condition
contrasting concept. The merits of using DUW to assess the health condition include:
(1) convenience in actuation/acquisition; (2) capability of detecting damage with a
small size; and (3) capability of evaluating the holistic health status of rail tracks in
the range covered by sensor network. With the condition-contrasting concept, the
health status can be evaluated using the signals from a current status, and thus it does
not require a benchmark process. In addition, with the condition contrasting concept,
the disturbance induced in practical applications can be excluded. Therefore, with
the proposed approach, the defects in rail tracks are detected in an in situ manner
and evaluated in a robust manner.
In Situ Condition Monitoring of High-Speed Rail Tracks Using Diffuse … 603
2 DUW-based Approach and Condition-contrasting

Concept
Considering the complexity of the geometries of rail tracks, the ultrasonic waves
emitted into the track are multiply scattered by the boundaries in the cross section.
As a consequence, analogous to the wave propagation in diffusive medium, such
as concrete, the acoustic energy is disseminated from its incident direction into
other directions, generating multiple random propagation paths [8–10] and leading
to the diffusion of incident waves. The accordingly generated wavefield is referred
to as diffused ultrasonic wave field. Despite a highly complex signal appearance, the
DUWs are reproducible and contain rich information about the medium, and they
are highly sensitive to changes in material or structural properties of rail tracks, such
as defects. Via appropriate processing, the information can be extracted for further
evaluation of health status of holistic rail track.
As stated above, in conventional methods, the health status of inspected structure
is evaluated via contrasting the measured signals against previously obtained baseline
signals. The effectiveness of this fashion can be impaired by drift of baseline signals
caused by diverse factors, and therefore it is challenging to extract the defect-related
waves features. To tackle this deficiency, a condition-contrasting concept is proposed
to minimize the adverse influence of practical factors. In this concept, DUW signals
from inspected structure, that are obtained under different conditions are contrasted.
In the process of signal acquisition, the practical factors are invariant, and thus the
variations in DUW signals are only linked to defects.
For a fatigue crack in reality, the two surfaces of crack are in contact in the vicinity
of the crack tip, owing to the crack closure phenomenon [11, 12], while an air gap
exists between the two surfaces (“open part”) for the rest part of the crack. When a
load is exerted on the track bearing fatigue crack, the opening/closing of the crack
can be altered, which will deviate DUW propagation and from that under a load-free
condition, as shown in Fig. 1a. Making use of such a characteristic, a fatigue crack
can be identified and characterized. With the acquired DUW signals, we propose a
damage index Rcc1 to measure the crack status alternation-induced discrepancies
Fig. 1 a Crack contact status alternation-induced disturbance in traversing waves; b crack growth-
induced disturbance in DUWs
604 K. Wang et al.
in DUW signals. The index Rcc1 evaluates the decrease in coherence between the
signals from load-free condition and under-load condition, and it reads.

X (t)Y (t)dt
Rcc1 = 1 − . (1)
X (t)2 dt Y (t)2 dt
In Eq. (1), X(t)is the signal acquired before the exertion of load, and Y (t) the
signal acquired when the track is subjected to a load.
In addition to the change of crack contact status, the propagation of the crack
might also be induced by the load, and this load-induced crack growth disturbs the
propagation of DUWs in the rail track, thereby causing a substantial deviation of the
DUW signals acquired after the load from those acquired before the exertion of load,
as shown in Fig. 1b. Therefore, the crack growth-induced changes in DUW can also
be used as indicators of the defects in rail tracks. To make use of the crack growth,
the second index Rcc2 is proposed to calibrate the level of decorrelation between the
signals obtained before the exertion of load and when the load is removed, and it
reads.

X (t)Z (t)dt
Rcc2 = 1 − . (2)
X (t)2 dt Z (t)2 dt
In Eq. (2), X(t) is the signal acquired before the exertion of load, and Z(t) the
signal obtained when the load is removed.
Note that the more severe the crack, the greater will be the part of crack alternating
contact status, producing a higher Rcc1. Therefore, the Rcc1 can be used to quantify
the severity of the defect. In addition, the defect grows with a probability linked
with the defect severity. That means the more severe the fatigue crack, the higher
the chance for the occurrence of crack growth when a load is exerted on the track.
Therefore, the severity of fatigue crack can also be quantitatively evaluated using the
probability of the Rcc2 reaching a certain threshold.
3 Experimental Validation
3.1 Set-Up and Test Strategy
For proof-of-concept, a rail track section measuring 1.2 m in length from China high
speed railway were examined. On this track section, the develop SHM approach
was implemented using a configured system. Two PZT wafers were installed on the
web of track with a distance of 70 mm from the rail bottom to generate and acquire
DUWs. A three-point bending test was performed on the rail track, in which the
two supporting points was apart 0.6 m that was consistent with the distance between
Fig. 2 The three-point bending test and sensor network for wave generation and acquisition
adjacent sleepers in reality. Note that in reality the weight of a typical train coach is
45,000–60,000 kg, a cyclic load (10–50 kN) with a frequency of 10 Hz was applied
at the midpoint of the upper surface of the rail track through a fatigue test machine
(ENPUDA® ), as shown in Fig. 2. At the midpoint of track bottom, a notch was
produced before the test. After ~108,000 cycles of fatigue load, a fatigue crack,
measuring 30 mm in length, was generated from the notch tip. An extensometer was
mounted on the notch to measure the crack opening.
To acquire the DUW signals at different stage of the fatigue life of rail track, the
fatigue test was suspended every 18,000 cycles, and during the suspension of cyclic
load, the DUWs were generated and acquired via the PZT wafers. In the DUW test,
ten-cycle Hanning windowed sinusoidal tone bursts with a frequency of 250 kHz
were produced by the waveform generation module. The DUWs are acquired under
different conditions: (1) when no load was applied (pre-load); then (2) when a static
load of 40 kN was exerted at the middle point of the rail head (on-load); and (3) after
a few seconds, when the static load was removed (post-unload). To acquire sufficient
information, DUW signals in a period of 10 ms were collected through the system.
To improve the signal noise ratio, the data from consecutive 128 tests were averaged.
3.2 Results
With the extensometer, the crack opening displacement which is linked with the
crack length is obtained as shown in Fig. 3a displaying the correlation between
the COD and the fatigue cycles. It is clearly demonstrated that the COD increases
monotonously with the fatigue cycle. To assess the change in opening/closing status
of the crack caused by the load, difference between DUW signals from the pre-load
and on-load conditions are calculated using Eq. (1), as detailed in Sect. 2. On this
606 K. Wang et al.
Fig. 3 a COD and Rcc1 versus fatigue cycle; b Rcc2 versus fatigue cycle
basis, the Rcc1 between DUW signals is obtained, see Fig. 3a, which apparently
shows a similar trend to the COD: a monotonic increase with the fatigue cycle.
This has evidenced that the external load can lead to crack status alternation and the
consequent distortion of DUW propagation. Therefore, the Rcc1 and its increasing
rate is capable of evaluating the development of the fatigue damage. To identify the
crack growth caused by the load, one can contrast the signals from the pre-load and
post-unload conditions, and the damage index Rcc2 is obtained using the Eq. (2), as
shown in Fig. 3b. It is clear that the Rcc2 extracted from DUWs at 36,000 fatigue
cycles is remarkable, whereas the Rcc2 obtained at other fatigue cycles is insidious,
arguing that at 36,000 fatigue cycles a crack propagation has occurred due to the
exertion of static load.
4 Deployment and Implementation of the Proposed

Approach
To examine its effectiveness and feasibility, the developed approach was implemented
on a rail track bearing fatigue defects from a marshalling station. Considering that the
load exerted on the rail by a passing train was time-varying, the load-related changes
in the DUWs were also time-varying, making it challenging to measure the changes
of DUWs related with the crack status alternation. Therefore, only the load-induced
crack growth was considered and exploited, and thus DUWs acquired before exertion
of the load and when the load is removed were employed.
To examine the effectiveness of the approach, rail tracks in which defects were
identified using an ultrasonic flaw detector are investigated. The PZT wafers were
installed on the web of the rail with a distance of 3 m and connected to the developed
system. To simulate the train passage-induced load, a hydraulic press was adopted to
exert a load of 100 kN, see Fig. 4a. With the DUW signals acquired in each DUW test,
the defined damage indices could be calculated with Eq. (2). The above experiments
Fig. 4 a The load on the rail exerted by hydraulic press; b the Rcc2 from each DUW test
demonstrated that the proposed damage indices are always lower than 1% when the
track is in a health status. Figure 4b displays the indices Rcc2 in each DUW test.
It is clear that in three tests the Rcc2 is greater than 1%, indicating the occurrence
of crack growth, and the probability of the crack growth for the track argues that a
defect with a certain severity in the switch rail is detected.
To investigate the robustness of the developed SHM approach, in situ monitoring
for the health condition of in-service rail tracks in a turnout system was performed.
Two PZT wafers was installed on the track in service in a turnout system, see Fig. 5a.
The DUW signals were acquired before and after the passing of a train and, using
Eq. (2), the proposed damage index Rcc2 was obtained, as displayed in Fig. 5b. It is
observed that no remarkable Rcc2 is obtained, arguing that the rail track in-service
is in an intact status.
It is worth noting that in condition-contrasting method, the DUW signals used
to assess the health condition of rail tracks are all obtained in a current status,
therefore the noised caused by various practical factors are substantially suppressed.
This warrants the robustness of the proposed approach in practical implementation,
benefiting its applications in condition monitoring towards HSR.
Fig. 5 a The sensor network installed on the turnout system (5 m apart); b the Rcc2 from each
DUW test
608 K. Wang et al.
5 Conclusions
Targeting in situ health monitoring of rail tracks, an approach based on diffuse ultra-
sonic waves and a benchmark-free condition-contrasting algorithm is developed and
deployed. The load applied on the rail track can alternate the crack opening/closing
status, and the DUW propagation is disturbed. The load can also induce the propaga-
tion of the crack, whereby remarkable variations are generated in the ultrasonic wave
field in rail tracks. By contrasting DUW signals acquired under different conditions,
damage indices are constructed, on which basis fatigue cracks in rail tracks can be
characterised quantitatively and the holistic condition of the tracks can be evaluated
in a real-time manner. The effectiveness of the proposed method is experimentally
examined, in which rail track subjected to cyclic load is monitored and evaluated.
Then conformance tests are performed on rail tracks in a marshalling station, demon-
strating effectiveness, reliability and robustness of DUW-driven condition monitoring
towards HSR applications.
Acknowledgements The work was supported by a Key Project (No. 51635008) and a General
Project (No. 51875492) received from the National Natural Science Foundation of China. The
authors acknowledge the support from the Hong Kong Research Grants Council via General
Research Funds (Nos.: 15201416 and 15212417).
References
1. Mair C, Fararooy S (1998) Practice and potential of computer vision for railways
2. Ph Papaelias M, Roberts C, Davis C (2008) A review on non-destructive evaluation of rails:
state-of-the-art and future development. Proc Inst Mech Eng F J Rail Rapid Transit 222:367–384
3. Cawley P, Lowe M, Alleyne D, Pavlakovic B, Wilcox P (2003) Practical long range guided
wave inspection-applications to pipes and rail. Mater Eval 61:66–74
4. Rao J, Ratassepp M, Lisevych D, Hamzah Caffoor M, Fan Z (2017) On-line corrosion moni-
toring of plate structures based on guided wave tomography using piezoelectric sensors. Sensors
17:2882
5. Wang K, Liu M, Su Z, Yuan S, Fan Z (2018) Analytical insight into “breathing” crack-
induced acoustic nonlinearity with an application to quantitative evaluation of contact cracks.
Ultrasonics 88:157–167
6. Wang K, Li Y, Su Z, Guan R, Lu Y, Yuan S (2019) Nonlinear aspects of “breathing” crack-
disturbed plate waves: 3-D analytical modeling with experimental validation. Int J Mech Sci
159:140–150
7. Zhu W, Xiang Y, Liu C-J, Deng M, Xuan F-Z (2018) A feasibility study on fatigue damage
evaluation using nonlinear Lamb waves with group-velocity mismatching. Ultrasonics 90:18–
22
8. Larose E, Zhang Y, Moreau L, Planes T, Obermann A (2017) Diffuse ultrasound monitoring of
stress and damage development on large scale concrete structures. J Acoust Soc Am 141:3831–
3831
9. Eybpoosh M, Berges M, Noh HY (2016) Sparse representation of ultrasonic guided-waves for
robust damage detection in pipelines under varying environmental and operational conditions.
Struct Control Health Monitor 23:369–391
10. Michaels JE, Michaels TE (2005) Detection of structural damage from the local temporal
coherence of diffuse ultrasonic signals. IEEE Trans Ultrason Ferroelectr Freq Control 52:1769–
1782
11. Buck O, Thompson R, Rehbein D (1988) Ultrasonic measurements of crack tip shielding by
closure. Mater Sci Eng A 103:37–42
12. Pippan R, Hohenwarter A (2017) Fatigue crack closure: a review of the physical phenomena.
Fatigue Fract Eng Mater Struct 40:471–495
Theoretical Analysis of Scattering
Amplitude Calculation of Torsional Wave
on a Pipe
Jaesun Lee, Jan D. Achenbach, and Younho Cho
Abstract The use of guided wave is widely used for inspection of long-range
pipelines due to its propagation characteristic with low attenuation. Also, recently
structural health monitoring (SHM) and condition monitoring (CM) is made a huge
paradigm shift on inspection and monitoring of large structures. However, the most
important issue of structure monitoring is quantitative defect sizing and interpreta-
tion from the signal. In this research work, the elastodynamic reciprocity is applied
to calculate scattered wave amplitude calculation. The reciprocity is a well-known
theorem to formulate the complicated scattering problem in a simpler calculation. The
torsional wave mode on a cylindrical structure is explored by the reciprocity theorem.
The scattered amplitude is presented as a function of incident wave frequency and
defect size. The defects on the surface with an inclined angle are formulated to
analyze the relation of defect angle and wavelength. The wave field is obtained by
using the reciprocity theorem and a superposition technique. The theoretical result
can provide signal interpretation for quantitative defect sizing. The closed-form solu-
tion of torsional wave scattering amplitude can be used for quantitative signal inter-
pretation. It is expected to be used in the analysis of structure condition by SHM and
CM approach as baseline data.
Keywords Scattering · Guided waves · Torsional wave · Reciprocity theorem ·

Amplitude calculation
J. Lee (B)
School of Mechanical Engineering, Changwon National University, Changwon, Korea
e-mail: jaesun@changwon.ac.kr
J. D. Achenbach
Engineering Sciences and Applied Mathematics and Mechanical Engineering, Northwestern
University, Evanston, IL, USA
Y. Cho
School of Mechanical Engineering, Pusan National University, Busan, Korea

https://doi.org/10.1007/978-981-15-9199-0_58
612 J. Lee et al.
1 Introduction
The guided waves have great potential for long-distance propagation on the waveg-
uides such as plates, rods, and pipes. Corrosion is always a concern in metallic
materials. In the engineering field, corrosion pits should be accessible for visual
inspection or general ultrasonic testing inspection. When corrosion is located on the
opposite side of the accessible transducer location, it is hard to evaluate since the
signal is scattered due to the shape and size of corrosion. For this reason, it is difficult
to inspect by a conventional method. Long-distance guided wave inspection methods
can be an alternative solution to corrosion pits. However, it is not easy to detect corro-
sion from a distance from the transducer since the signal analysis is difficult for the
scattered problem from the long distance.
Wave scattering analysis is mainly performed by a numerical approach. The finite
element method (FEM) and the boundary element method (BEM) consume lots of
computational memory and time to calculate the wave scattering signal and amplitude
from surface erosion. Mutual law is a powerful way to calculate scatter wave ampli-
tude calculations in a simple way. In general, reciprocity provides the relationship
between displacement, traction component, and body force [1]. By the reciprocity
theorem, we can provide the relationship between the incident and the scattering wave
amplitude ratio. Some applications of reciprocity theorem have been performed [2].
Recently, a new approach based on elastic dynamics theorem has been studied for
surface wave scattering amplitude by surface mechanics [3]. However, initial research
is limited due to the surface wave propagation problem in semi-space. Many engi-
neering structures are organized in a pipeline structure. Therefore, the study on the
scattering wave analysis for the hollow cylinder structure is important for practical
engineering applications.
In this paper, the calculation of the scattering wave amplitude for the cylindrical
structure by the torsional wave mode is investigated using a reciprocity theorem.
The scattered wave amplitude is calculated by the use of the reciprocity theorem in
a much simpler way.
2 Reciprocity Theorem Application
2.1 Reciprocity Theorem
The reciprocal relationship is the relationship between displacement, traction, and

physical fitness for the two elastic dynamics of a single body. The approach taken
here is described in detail by Achenbach [4]. The goal of using cleanup is to combine
information about both states to better understand one of the states of interest. For
our analysis, we will look for the amplitude of the radiation wave due to the point
load and define an imaginary wave perpendicular to the first state as the second state.
Theoretical Analysis of Scattering Amplitude Calculation … 613
The radiated wave (state A) has a physical strength in the form of a point source
acting on the surface of the structure. The virtual wave (state B) is homogeneous.
Both states will be a mechanical solution to the undamaged plate. The correlation
can be written as (1). In general, the reciprocity theorem provides the relationship
between displacements, surface tractions of the same body. The normals are directed
outward from the solid boundary through surface integration.

f i A u iB dV = τiBj u iA − τiAj u iB n j dS (1)
V S
where S is a surface around an area defined by volume V. f i is the force, u i is the

displacement, and τi j means the stresses.
2.2 Equivalent Force of Surface Traction
The reciprocity theorem is employed to calculate the amplitude of the surface wave
generated by the time-harmonic loading source. The circumferential load Pθ is
applied to the origin of the system in the θ direction given in Fig. 1.
f θ = Pθ δ(r − b)δ(z)e−iωt (2)
To apply reciprocity theorem, integral surface can be set as S1, S2, S3, and S4 as
depicted in Fig. 1.
Reciprocity theorem is derived as Eq. (3)

f θA u θB + τrAθ u θB − τrBθ u θA dθ dr = 0 (3)
z =0
r =b S1
Fig. 1 Surface definition on hollow cylinder

614 J. Lee et al.
Fig. 2 Circular shape of corrosion defect
where superscripts A and B represent each elastomeric state. In the current situation,
state A is a forced generated field. The potential wave equation becomes much simpler
because it only considers θ. The two states of wave propagation are shown in Fig. 1.
State A is a wave generated by the point loads applied around the pipe. State B is a
virtual wave for mutual theorem.
3 Application to Scattering by a Surface Corrosion
As an example, scattering waves are considered by cylindrical cavities on the outer

surface of the cylindrical structure. The two-dimensional geometry of the surface
defect with depth D and width W is depicted in Fig. 2.
The cavity shape can be defined as
y = f (z) = D − M z 2 (4)
where coefficient M is defined as M = D

with boundary condition y = D at z =
( W2 )
2
0 and y = 0 at z = ±w/2.
4 Torsional Wave Scattered Wave Calculation
As an example of reciprocity theorem application on cylindrical structure is

performed. The scattered amplitude ratio is calculated based on the axisymmetric
circular and rectangular shape corrosion in this paper. The material properties are
79.87 GPa of shear modules, 0.2916 of Poisson’s ratio, and density as 7800 kg/m3 .
Figures 3 and 4 show the calculation results at the excitation frequency for the
lowest torsional wave mode. The values of A− sc / A i n are dependent on the excitation
frequency and the size of corrosion [5].
Theoretical Analysis of Scattering Amplitude Calculation … 615
Fig. 3 Scattered amplitude ratios for 20% thickness loss for tapered notches
Fig. 4 Scattered amplitude ratio when 20% thickness loss of the tapered notch for the range of 65
to 85
5 Conclusions
The reciprocity theorem is employed to solve the amplitude ratio of the scat-
tering wave induced by the equivalent force applied to the hollow cylinder for
torsional waves. The closed-form solution of scattered wave amplitude calculation is
presented as a function of frequency, geometry, and material properties. The proposed
theoretical approach provides a very simple calculation algorithm for the calculation.
Acknowledgements This work was supported by the National Research Foundation of Korea
(NRF) grant funded by the Korea Government (MSIT) (No. NRF-2019R1G1A1004577).
616 J. Lee et al.
References
1. Achenbach JD (2003) Reciprocity in elastodynamics. UK, Cambridge University Press,

Cambridge
2. de Hoop AT (1960) A reciprocity theorem for the electromagnetic field scattered by an obstacle.
Appl Sci Res Sect B 8(1):135–140
3. Achenbach JD (1999) Wave propagation in elastic solids, 8th edn. Elsevier, Netherlands
4. Achenbach JD (2005) Combination of a virtual wave and the reciprocity theorem to analyse
surface wave generation on a transversely isotropic solid. Phil Mag 85(33–35):4143–4157
5. Lee J, Achenbach JD, Cho Y (2018) Use of the reciprocity theorem for a closed form solution of
scattering of the lowest axially symmetric torsional wave mode by a defect in a pipe. Ultrasonics
84:45–52
Detection of a Micro-crack on a Rotating
Steel Shaft Using Noncontact Ultrasonic
Modulation Measurement
I. Jeon, P. Liu, and H. Sohn
Abstract Detection of a micro-crack on a rotating structure is limited because

conventional non-destructive testing (NDT) is developed based on contact-type trans-
ducers. By using air-coupled ultrasonic NDT system, micro-crack detection on a
rotating steel shaft is studied. Through air-coupled ultrasonic system, ultrasonic-
guided waves at two distinctive frequencies are generated. Then, their ultrasonic
modulation at the sum and difference of input frequencies are generated due to micro-
crack. Finally, the spectral correlation (SC) of ultrasonic modulation is computed for
diagnosis of micro-crack. This is the first study for the detection of a micro-level
fatigue crack on a rotating steel shaft using noncontact ultrasonic NDT system. The
developed system is experimentally validated using the steel shaft through cyclic
loading tests. At each cyclic loading case, the result was compared with conventional
penetrant test and metallographic analysis.
Keywords Ultrasonic-guided waves · Rotating steel shaft · Air-coupled · Spectral

correlation · Micro-crack
1 Introduction
A micro-fatigue crack is damage feature which is the most important feature in the
structural failure [1]. Continuous loading on the structure causes the crack growth to
critical failure point, and it could make unexpected failure [2]. Thus, several NDT
approaches have been suggested for characterizing the micro-fatigue cracks. Then,
micro-fatigue crack detection becomes highly important in many applications.
Recently, nonlinear characteristics of ultrasonic have been widely used for micro-
crack detection in mechanical, civil, and aerospace systems. As ultrasonic wave
propagates crack, creep, and delamination in the elastic material, then they exhibit
nonlinearity characteristics such as harmonics and modulations. These nonlinearity
I. Jeon · P. Liu · H. Sohn (B)

Department of Civil and Environmental Engineering, KAIST, Daehakro 291, Yuseong-gu,
Daejeon 34141, Republic of Korea
e-mail: hoonsohn@kaist.ac.kr

https://doi.org/10.1007/978-981-15-9199-0_59
618 I. Jeon et al.
features are highly sensitive to damage features, especially for early stage fatigue
cracks. Many studies showed the relation between fatigue crack and ultrasonic nonlin-
earity. However, conventional ultrasonic techniques for finding structural nonlin-
earity are developed using contact-type sensors, and a few of them are applied rotating
structures.
This paper develops a micro-crack detection in rotating steel shafts with noncon-
tact air-coupled technique. Section 2 provides the overview of air-coupled ultrasonic
based NDT system. Section 3 provides working principles of fatigue crack detection
techniques based on ultrasonic modulation. The performance of the proposed tech-
niques is examined by detecting fatigue cracks in rotating shaft in Sect. 4. Finally,
the result is discussed in Sect. 5.
2 Air-Coupled Ultrasonic System
Figure 1 shows a schematic diagram of the air-coupled ultrasonic NDT system devel-
oped in this study. The system is composed of an ACT sensing system and a data
acquisition system (DAQ). In the ACT sensing system, low-frequency (LF) and high-
frequency (HF) ACTs generate ultrasonic waves on the target specimen in a noncon-
tact manner and signals are measured by sensing ACT. The measured responses are
collected and analysed in the DAQ, which controls and synchronizes all signals.
Fig. 1 Schematic diagram of the air-coupled ultrasonic NDT system

Detection of a Micro-crack on a Rotating Steel Shaft Using Noncontact … 619
3 Spectral Correlation-Based Crack Diagnosis
Low signal-to-noise ratio is one of the biggest problems in noncontact nonlinear

ultrasonic inspection. Amplitudes of modulation components are several ten times
smaller than linear responses. Furthermore, amplitudes of modulation components
are affected by nonlinearity of experimental instruments or material anisotropy. To
overcome these problems, spectral correlation (SC) technique is developed for crack
diagnosis in this paper. The principle of SC technique is represented in Fig. 2.
Step 1: Both LF (ωa ) and HF (ωb ) sinusoidal inputs are applied to a specimen.
When LF input is applied to fatigue crack, it makes crack closing and opening.
Then, applied HF input produces modulation components (ωa ± ωb ). The first
graph in Fig. 2 shows modulation components generated by the fatigue crack.
Next, the correlation is calculated to obtain the crack-induced feature as follows:
Sωb ,ωa = E[X (ωb + ωa )X ∗ (ωb − ωa )] (1)
Based on the correlation value, the spectral correlation (SC) will be defined in the
next step.
Step 2: A noise in Sωb ,ωa is extracted in this step. First, correlation noise level
(n ωb ,ωa ,i , i = 1,2,…,n) is obtained by n times applied HF inputs. Note that the
noise correlation level should be zero in theory. Then, the exponential distribution
is fitted to series of noise responses (n ωb ,ωa ,i ). After that, noise index (N I ωb ,ωa ) is
obtained by specified confidential interval and the SC value is
SC ωb ,ωa = Sωb ,ωa − N I ωb ,ωa (2)
Multiple SC ωb ,ωa values are obtained by several combinations of input frequencies

(ωa , ωb ).
Step 3: The modulation amplitudes and noise level are changed by varying oper-
ational and environmental conditions at different loading cycles. Therefore, the
normal fitting-based outlier analysis is performed in the final step. If any of the
Fig. 2 Steps for the spectral correlation technique

620 I. Jeon et al.
SC ωb ,ωa ,l values exceed the threshold, it is defined as outliers, which indicate the
fatigue crack damage on the test specimen.
4 Detection of a Micro-crack on Steel Shafts
Figure 3 shows an experimental setup of the air-coupled NDT system for the exper-
imental validations. A torsion test rig produces a pulsating sinusoidal load, and the
initial torsional load (55,000 Nm) and amplitude (53,000 Nm) were calculated from
FKM guideline. From the calculation, crack initiation was expected at 75,000 cycles.
Figure 4 represents the frequency domain signal from the damaged case. Low-
frequency noises were caused by the shaft rotating condition. However, these were
below 30 kHz and distinguished with the HF modulation components. Figure 5
shows examples of SC measurements at intact and damaged case. Four outliers were
detected under damaged case, indicating fatigue crack. SC results at different cyclic
loading are summarized in Table 1.
At 20,000 cycles, the first fatigue crack indication was found. The outliers
increased until 26,000 cycles (except online inspection) and then decreased. It is
because crack opening and closing phenomenon is hard to perform with crack grew.
On the other hand, the outliers from online inspection well indicated crack status
compared to stationary condition.
For comparison, fluorescent penetrant inspection (FPI) was used. The results
are shown in Table 1. There was no micro-crack until 28,000 cycles by FPI, but
ambiguous crack indications were detected at 35,000 cycles. At 75,000 cycles, fatigue
Fig. 3 Setup for the air-coupled NDT system under torsional cyclic loading
Fig. 4 Frequency domain representation from the damaged case (22,000 cycles): a stationary and
b rotating condition
Fig. 5 Diagnosis of the steel shaft under rotating condition: a intact case and b damaged case
cracks were detected by a FPI. This result was finally checked by metallographic
method. The crack depth was 150 µm or less (Fig. 6).
622 I. Jeon et al.
Table 1 Results of micro-crack detection

Cycles SC FPI
(outliers)
0 0 No crack
0 online 0 N/A
10,000 0 N/A
20,000 2 N/A
20,000 online 1 N/A
22,000 6 Point indication
22,000 online 4 N/A
35,000 11 Line indication
75,000 2 Micro-crack
(a) 22,000 cycles (b) 35,000 cycles (c) 75,000 cycles (d) 75,000 cycles
Fig. 6 FPI and metallographic results
5 Conclusion
In this study, the fatigue crack detection technique is developed using noncon-
tact ultrasonic measurements. The proposed technique is applied to fatigue crack
detection under rotating conditions of a steel shaft based on ultrasonic modulation
components. Developed spectral correlation damage feature improves the robustness
of detection as minimizing false positive alarms. The proposed algorithm detected
crack formation earlier than conventional penetrant test or metallographic analysis.
The size of the micro-fatigue crack verified by the metallography was around 150 µm
or less.
References
1. Campbell SD (2008) Steel moment frame damage predictions using low-cycle fatigue. in: The
14th world conference on earthquake engineering
2. Symonds PS (1952) The interpretation of failure loads in the plastic theory of continuous beams
and frames. J Aeronaut Sci 19(1):15–22
Measurement of Axial Force of Bolted
Structures Based on Ultrasonic Testing
and Metal Magnetic Memory Testing
Siqi Yang, Laibin Zhang, and Jianchun Fan
Abstract Bolted structures are widely utilized in flange sealing of the petrochem-
ical high-pressure facilities, measurement of axial force of bolted structures can
contribute to monitor flange sealing status. In this paper, static tensile tests are
conducted in laboratory. The stress distribution of bolt with different axial tensile load
is analyzed by using finite element (FE) method. Two non-destructive testing (NDT)
methods including ultrasonic testing and metal magnetic memory testing are used to
measure axial force of two bolt materials (low-carbon steel 4.8 and medium-carbon
steel 8.8 bolts) during the whole tensile process. For the ultrasonic testing, the test
results show that there is a stable linear correlation between the axial force of the bolt
and the ultrasonic time of flight (TOF), and the TOF measurement is upgraded by
using cross-correlation analysis technique. For the metal magnetic memory testing,
magnetic signals increase with the increase of axial force because of the effect of
stress-magnetization. Furthermore, magnetic memory characteristic parameters have
a good linear relationship with axial force within the scope of a certain stress. There-
fore, metal magnetic memory testing can be used as a novel method to measure axial
force of bolted structures. Finally, the advantages and drawbacks of these two NDT
methods are analyzed, which can provide a reference for the engineering application
of effective NDT methods and technologies in bolt axial force measurement.
Keywords Bolts · Non-destructive testing · Axial force measurement · Time of

flight · Metal magnetic memory signal · Finite element
1 Introduction
In the field of petrochemical industry, high-pressure equipment such as fracturing

manifold, christmas tree (oil and gas), and blowout preventer may withstand up to
tens and even hundreds of MPa pressure in complex operating environment. Bolted
S. Yang · L. Zhang · J. Fan (B)

College of Safety and Ocean Engineering, China University of Petroleum-Beijing, Fuxue Road
18, Changping, Beijing 102249, People’s Republic of China
e-mail: 709944199@qq.com

https://doi.org/10.1007/978-981-15-9199-0_60
626 S. Yang et al.
flange joints as sealing components, which are diffusely used in connecting high-
pressure vessel and piping equipment. Facing such high operating pressure, in order
to prevent serious consequences caused by leakage failure, it is extremely important
to monitor flange sealing status [1]. Since bolts are crucial pressure-bearing structures
of bolted flange joints, the flange sealing status can be monitored by detecting the
axial force of the bolts on the flange plate.
At present, many researchers have paid great efforts on detecting the axial force of
bolted structures [2, 3]. Traditional axial force of bolted methods, such as strain gage
method, torque wrench method, angle control method, and direct tension indicator
method, and these traditional direct methods have obvious and unacceptable short-
comings, strain gage method lacks of accuracy, and efficiency due to the measurement
of elongationh; besides, it is not suitable for the field test. Torque wrench method is
commonly used in bolt axial force measurement, this method estimates axial force
by using torque-meters, however, the torque applied in this method is mostly used
to offset friction loss, which results in large errors. Angle control method predicts
axial force by measuring the rotation angle of bolts in the elastic range, however, it is
difficult to obtain the elastic range in field test. Direct tension indicator method can
measure axial force directly by special designed washers which are put either under
the bolt or the nut, but this method requires specific designs for each structure and is
almost impossible to monitor online [4].
Since traditional methods have low accuracy and are not suitable for engineering
application, more and more researches have focused on the application of indirect
NDT methods in bolt axial force measurement [5]. Ultrasonic testing method based
on acoustoelastic effect has been widely concerned. Although this method is consid-
ered as a potential alternative technique for measuring axial force of bolted structures,
accurate measurement of TOF is an obstacle needs to be solved, the usual method
of threshold-detection will lead to large error [4, 6].
As a novel NDT method, metal magnetic memory method is mainly used to
diagnose both stress concentration and early damage state [7]. Due to its high sensi-
tivity to stress variation, this technology has drawn increasing attention. Numerous
static tension and fatigue experiments show that there is an obvious correspondence
between magnetic memory signals and stress states. This technology can be regarded
as a novel solution to the measurement of axial force of bolted structures because of
the effect of stress-magnetization.
In order to study the application of above NDT methods to measure the axial force
of bolts, static tensile tests of bolted structures are conducted in laboratory based on
ultrasonic testing and metal magnetic memory testing. The stress distribution of bolt
with different axial tensile load is analyzed by using FE method. In the ultrasonic
testing, the cross-correlation analysis technology is applied in the measurement to
improve the accuracy of TOF extraction. The relationship between the axial force
of the bolts and the characteristic values for TOF is obtained. In the metal magnetic
memory testing, the variation of magnetic memory signal on the screw thread connec-
tion induced by tensile stress is given, which helps to take advantage of this NDT
method as a new technique for bolt axial force measurement. Finally, the advantages
Measurement of Axial Force of Bolted Structures Based … 627
and drawbacks of the two methods are compared, which can provide a reference for
the future engineering application in bolt axial force measurement.
2 Principle of Measurement
2.1 Acoustoelasticity Effect for Ultrasonic Testing
As shown in Fig. 1, on the basis of Hooke’s law and acoustoelastic effect, the
ultrasonic TOF in stressed state (t σ ) can be expressed by Eq. (1).

2L 1 2L 2 2L 1 1 + E −1 σ 2L 2
tσ = + = +
Vσ V0 V0 (1 + Aσ ) V0
2L 2L
∼
=
0
+
1
E −1 − A σ (1)
V0 V0
where L 0 is the initial bolt length (mm); L 1 is the stressed portion (mm); L 2 is the
unstressed portion (mm); V σ is the velocity of ultrasonic in stressed state (mm/s); V 0
is the velocity of ultrasonic in unstressed state (mm/s); σ is the tensile stress (MPa);
E is the Young’s modulus (MPa); A is the acoustoelastic coefficient.
Compare t σ with the ultrasonic TOF in unstress state (t 0 ). Then, the characteristic
values for TOF k can be shown as follow:
−1
tσ − t 0
2L 1
V0
E −A σ L 1 −1
k= = 2L 0
= E − A σ (2)
t0 V
L0
0
In this relation, the Young’s modulus E and the acoustoelastic coefficient A are depen-
dent of the materials, and the bolt length L 1 characterize the tightening configuration.
The linear correlation between the characteristic values for TOF and axial stress of
the bolt can be obtained.
Fig. 1 Bolt axial force measurement physical model

628 S. Yang et al.
Original signals usually contain noise, and the cross-correlation function is used to
accurately calculate the characteristic values for TOF, which is defined by Eq. (3).
N −1
1
Rx y (τ ) = lim [x(i)y(i − τ )] (3)
N →∞ N
i=0
where x(i) is the ultrasonic signal in unstressed state, y(i) is the ultrasonic signal in
stressed state, N means the signal length; τ means the time delay. Theoretically, y(i)
= x(i + τ ), and the result reach its maximum when i = −τ, then the time delay
between two ultrasonic signals can be obtained. Therefore, the characteristic values
for TOF can be calculated through cross-correlation analysis.
2.2 Stress-Magnetization Effect for Metal Magnetic Memory

Testing
The stress-magnetization effect can be explained according to the ‘Effective field

theory’ proposed by Jiles et al. [8]. The initial magnetic field (H) is changed by the
application of stress (σ ) through the reverse magnetostriction effect, which is equal to
add an equivalent magnetic field (H σ ), the equivalent effective magnetic field (H eff )
can be expressed by Eq. (4).
Heff = H + α M + Hσ
3 σ dλ
= H + αM +
2 μ0 dM
3σ
∼
= H + αM + γ1 (0) + γ1 (0)σ M + 2 γ2 (0) + γ2 (0)σ M 3 (4)
μ0
where α means the coupling intensity between the individual magnetic moment and
the magnetization, which is dimensionless; M means the magnetization (A/m); μ0
is vacuum permeability, which equals 4π × 10−7 (H/m); λ means the bulk magne-
tostriction constant; γ 1 and γ 2 are the magnetostrictive coefficients of empirical
formulas.
3 Experimental Systems
In order to simulate the real tightening of the bolted structure, a bolt axial tension
device has been designed. Two different grades of M20 stud bolts (grade 4.8, 8.8 bolts)
are used in this experiment, and four same grade test bolts can be simultaneously
Table 1 Material and

Bolt type Material σ s [MPa] σ b [MPa]
mechanical properties of test
bolts M20/4.8 Q235 steel 320 400
M20/8.8 45# steel 640 800
loaded by a hydraulic pump. The material and mechanical properties of those two bolt
types are given in Table 1. Due to the difference of mechanical properties between
two types of bolts, the axial tensile force of each bolt is increased from 0 to 30 kN
by 2.5 kN step for grade 4.8 bolts and 0 to 65 kN by 5 kN step for grade 8.8 bolts.
A load cell is used in this device to obtain the axial tension.
For the ultrasonic testing, the experimental system utilized in this paper is shown
in Fig. 2. An ultrasonic transducer is located on the head of the threaded end, which
frequency is 5 MHz. The pulse generator-receiver is connected with the ultrasonic
transducer, which can generate and receive an ultrasonic burst. A digital oscillo-
scope synchronized with the pulse generator-receiver is used to display and sample
the ultrasonic signals at maximum 1 GHz. Finally, the signals are transferred to a
processing PC with a flash memory.
For the metal magnetic memory testing, as shown in Fig. 3, four giant magneto
resistance (GMR) sensors are placed on the thread connection of bolts to obtain the
magnetic memory signals. A data acquisition device is used to collect signals, which
can be transferred to a computer for further processing.
Fig. 2 The experimental system for the ultrasonic testing
Fig. 3 The experimental system for the magnetic memory testing

630 S. Yang et al.
Fig. 4 Finite element model of the bolt
The FE model of test bolt is established by using FE software WORKBENCH. In

order to obtain the distribution of stress with different axial force, the static structural
analysis model has been built, as shown in Fig. 4. The object studied in this FE
simulation is a M20 grade 8.8 bolt, which is the same dimension as the bolt used
in the experiment. In accordance with the experimental setup, the bolt is applied an
axial pretension of 5 kN, 35 kN, 65 kN, respectively, and the load surfaces of nuts
on both sides are set to fixed support. The contact regions between screw and nut are
set as frictional and automatically controlled by software.
Figure 5 shows the distribution of stress with different axial force, in order to
facilitate comparative analysis, a line is choosed close to the thread in the longitudinal
section of bolt to plot the equivalent stress distribution in it. It can be clearly seen
that the distribution of stress is axisymmetric, and the whole bolt is divided into
the stressed portion and the unstressed portion. With the increase of axial force,
the equivalent stress in stressed portion of bolt increases. It is noteworthy that the
maximum equivalent stress locates at the first root of the bolt thread, which means the
plastic deformation happens here first [9]. This also explains why the GMR sensors
are placed on this position.
5 Test Results and Discusssion
5.1 Results for Ultrasonic Testing
Figure 6 shows the original ultrasonic signals obtained from the oscilloscope, which
includes the excitation signal and the echo signal. The noise signal occurs after the
1st echo signal, which is amplified locally, caused by the reflection at the uneven end
face of bolt and the screw thread.
In order to calculate the accurate characteristic value for TOF, the 1st echo signals
of being stress free and load imposed are processed by using cross-correlation anal-
ysis. Then, the characteristic time of TOF is extracted, as shown in Fig. 7. The
(a) axial force 5kN (b) axial force 35kN
(c) axial force 65kN
Fig. 5 Distribution of stress with different axial force
characteristic time t varies with the different axial tensile load, which characterizes
the difference of TOF in unstressed state and stressed state.
Thus, the relationship between the characteristic value for TOF k and the axial
force of the bolt F is obtained. Figure 8 shows the tension test results for the grade
4.8 and 8.8 bolts, respectively. During the whole tensile process, the characteristic
value for TOF shows a significant linear relationship with axial force. And, the linear
fitting results of two bolt materials as shown in Table 2, the slopes of fitted linear of
four bolt samples of the same bolt material are similar.
5.2 Results for Metal Magnetic Memory Testing
The tangential characteristic parameter Mean-H x is extracted from original signal,

which means the mean value of tangential magnetic memory signal in the same stress
state. Figure 9 shows the relationship between Mean-H x and the axial force for the
632 S. Yang et al.
Fig. 6 The original ultrasonic signals
Fig. 7 TOF measurement

based on the
cross-correlation analysis
grade 4.8 and 8.8 bolts, respectively. It seems obvious from the comparison results
that magnetic memory signals increase with the increase of axial force.
Due to the difference of heat treatment in manufacturing process, the initial
magnetic field of each specimen bolt is quite different, which can be clearly seen
from the comparison results. It is noteworthy that the magnetic memory signal of
grade 4.8 bolts shows a very stable linear trend during the whole tensile process, but
for grade 8.8 bolts, the linear region appears in a certain stress range, and the slopes
grade 4.8 bolts grade 8.8 bolts
Fig. 8 The comparison result for the two bolt materials based on ultrasonic testing
Table 2 The linear fitting result for the two bolt materials based on ultrasonic testing
Bolt grade Sample number linear fitting equation R2
4.8 1 k = 0.00958F + 0.00236 0.99884
2 k = 0.00991F + 0.00959 0.99682
3 k = 0.00939F + 0.00601 0.9984
4 k = 0.00884F-0.00039 0.99825
8.8 1 k = 0.00677F + 0.00929 0.99631
2 k = 0.00697F + 0.01244 0.9973
3 k = 0.0061F + 0.00533 0.99679
4 k = 0.00691F + 0.01558 0.99588
grade 4.8 bolts grade 8.8 bolts
Fig. 9 The comparison result for the two bolt materials based on magnetic memory testing
634 S. Yang et al.
Table 3 The slopes of the linear line of two bolt materials based on magnetic memory testing
Bolt grade Sample number The slope of the linear line R2
4.8 1 0.17384 0.99809
2 0.1823 0.96499
3 0.18128 0.97929
4 0.15834 0.99477
8.8 1 0.09872 0.9987
2 0.08943 0.92521
3 0.11274 0.94209
4 0.11355 0.96376
of the linear line of two bolt types is shown in Table 3. It can be found that the slopes
of the two bolt types are different because of the difference of material. Therefore,
the axial force of bolts can be inferred from the slope.
6 Conclusions
In this work, in order to monitor the axial force of bolted structures under tension
loading, static tensile tests of bolts are carried out in laboratory. The FE method
is used to analyze the stress distribution of bolt with different axial force, and the
maximum stress occurs in the first root of the bolt thread. During the whole tensile
process, the ultrasonic testing method based on the acoustoelastic effect and the metal
magnetic memory testing method based on the stress-magnetization effect are used
to measure axial force in two bolt materials: low-carbon steel 4.8 and medium-carbon
steel 8.8 bolts.
In these two NDT methods, the results of ultrasonic testing show that the character-
istic value for TOF is linear with the axial force of the bolt, and the cross-correlation
analysis technique is used to improve the accuracy of TOF calculation. The results
of the magnetic memory testing indicate that the magnetic memory characteristic
parameter Mean-H x , which is extracted form magnetic signal has a linear relation-
ship with the axial force of the bolt in a certain stress range, so it can be used as a
novel method to measure the the axial force of bolts by monitoring Mean-H x .
As shown in Table 4, the advantages and drawbacks of those two NDT methods
studied in this paper are listed for the reference of engineering application.
Table 4 The advantages and drawbacks of two NDT methods

Method Advantages Drawbacks
Ultrasonic testing Simple theory; less susceptible High sampling rate device cost
to outside interference expensive; the bolts need to be
dismantled
Metal magnetic memory Low equipment cost; suitable for Complicated theory; influenced
testing field monitoring by material and initial magnetic
field
Acknowledgements Funding: The authors are grateful to China Petroleum & Chemicals Corpo-
ration (SINOPEC) for financial support to this work.
References
1. Zhu L, Bouzid AH, Hong J (2018) Analytical evaluation of elastic interaction in bolted flange
joints. Int J Pres Ves Pip 165:176–184
2. Bickford JH (2008) Introduction to the design and behavior of bolted joints, 4th edn. CRC Press,
Boca Raton
3. Zhang Z et al (2016) Continuous monitoring of residual torque of loose bolt in a bolted joint.
In: 6th APWSHM, Hobart, pp 278–285
4. Nikravesh SMY, Goudarzi M (2017) A review paper on looseness detection methods in bolted
structures. Lat Am J Solids Struct 14(12):2153–2176
5. Heyman JS, Chern EJ (1982) Ultrasonic measurements of axial stress. J Test Eval 10(5):202–211
6. Kim N, Hong M (2009) Measurement of axial stress using mode-converted ultrasound. NDT&E
INT 42(3):164–169
7. Dubov AA (1997) A study of metal properties using the method of magnetic memory. Met Sci
Heat Treat 39(9):401–405
8. Jiles DC (1999) Theory of the magnetomechanical effect. J Appl Phys 28(8):1537–1546
9. Chen X et al (2015) fatigue failure analysis for bolt-nut connections having slight pitch
differences using experimental and finite element methods. Acta Polytech Hung 12(8):61–79
Ultrasonic Testing of Anchor Bolts
Y. M. Tong, A. R. K. Rajkumar, H. G. Chew, and C. K. Liew
Abstract Anchor bolts play a crucial role in supporting the power cables of the
overhead catenary system (OCS) in underground train tunnels. With exposure to
humid conditions and water seepage in the tunnel from wet weathers, the anchor
bolts can experience corrosion and damage at an accelerated rate. As anchor bolts
are embedded into the concrete tunnel, and their conditions are unknown without
proper inspection of their internal bodies. This paper aims to propose an inspection
process to allow technicians to perform ultrasonic testing (UT) on anchor bolts during
scheduled maintenance works in the tunnel. Experimental studies using ultrasonic
waves were carried out on the anchor bolts of different integrity conditions. Analytical
studies conducted were then used to determine the tolerance level and decision matrix
to categorize the different conditions of anchor bolts. The outcome of this study shows
that there is a feasible approach to perform UT inspections on the anchor bolts.
Keywords Maintenance · Scheduled inspections · Ultrasonic testing · Anchor

bolts · Train tunnels
1 Introduction
In Singapore, the North East Line (NEL) is the only mass rapid train line using
an overhead catenary system (OCS). The OCS is a system of overhead wires that
provides electrical power to operate trains through a pantograph. The power cables
are supported by vertical metal beams that are fastened by anchor bolts into the
concrete ceiling of the tunnel.
The current inspection process of the bolts occurs every three months which is
done primarily by visual inspection. The inspection is done by checking any form of
loose movements of the exposed ends of the anchor bolts and the securing nuts, as
well as the observation of any rust on the bolts which can indicate the presence of
corrosion. The present visual inspection method has the following disadvantages:
Y. M. Tong (B) · A. R. K. Rajkumar · H. G. Chew · C. K. Liew

Singapore Institute of Technology, 10 Dover Drive, Singapore 138683, Singapore
e-mail: Denise.Tong@SingaporeTech.edu.sg

https://doi.org/10.1007/978-981-15-9199-0_61
638 Y. M. Tong et al.
Fig. 1 UT of anchor bolts in the train tunnel during site trials: a scaffold for OCS maintenance
works; b a closed-up view of the placement of UT probe on anchor bolts at tunnel ceiling
• Unable to detect internal corrosion of the anchor bolts,

• Unable to detect defects like cracks on the section of the anchor bolts embedded
into the concrete tunnel,
• Difficulty in this inspection process for anchor bolts that are located high up at
the tunnel ceiling.
Previous experimental research was done on anchor bolts and rebar embedded
into concrete foundation by ultrasonic testing (UT). Through the exposed surface
areas, the anchor bolts were mostly found to be in good conditions, with 142 out
of 448 anchor bolts suspected to be flawed [1]. Another published research showed
UT conducted on a uniformly corroded rebar which was embedded into concrete,
where it was found that there was a decrease in signal strength with a higher level of
corrosion [2].
By conducting UT inspection on embedded anchor bolts in concrete, this will
increase the chances of identifying the anchor bolts’ true condition for making proper
maintenance and repair works. Figure 1 shows the inspection process using the
developed UT method for the anchor bolts on the OCS.
2 Proposed UT Inspection
This section presents the current UT tools and inspection framework proposed for the
anchor bolts during scheduled maintenance works in the train tunnel. The equipment
Ultrasonic Testing of Anchor Bolts 639
Fig. 2 UT inspection equipment: a Omniscan SX pulsar-receiver unit [3]; b fingertip probe (v110,
5 MHz, 0.25 in.) [4]
used is shown in Fig. 2. The contact probe can be installed in a holder with an
extension pole to inspect anchor bolts located at high ceilings, as shown in Fig. 1b.
The Omniscan SX is a compact, lightweight and user-friendly pulsar-receiver
unit which allows easy setup with automatic probe recognition and pre-configurated
application wizard. It allows for an extensive readings database with flaw statistics,
as well as an optimized preconfigured layout for quick and simple length, depth
and height sizing of flaws. The fingertip probe is a single element longitudinal wave
transducer which makes direct contact with inspected anchor bolts.
The proposed inspection process is presented in a flowchart that is split into three
different phases, namely test, normalization and evaluation. The inspection process of
the flowcharts was based on analyzing the graph as shown below. A series of different
tests of anchor bolts were used to analyze individual patterns for each anchor bolt’s
condition with the average signal results shown in Fig. 3.
In Fig. 3, the initial pulse and near-field zone (0 mm to 30 mm) are not taken
into consideration for evaluation as there are interferences with the sound energy
transmitted from the transducer to the sub-surface of the anchor bolt. Defect-free
anchor bolts are referred as the baseline result to aid with classifying and differenti-
ating the conditions of the anchor bolts. Defect-free anchor bolts are found to give
an amplitude of less than 4% before the first backwall pulse.
With comparison to the baseline, there are spike patterns for defective anchor bolts
occurring before the first backwall pulse while there are dissimilar signal patterns
observed after the first backwall pulse. The crack location is located around 38 mm
from the inspected surface of the anchor bolt with the amplitude of 7.8%, which is
higher than the 4% background amplitude.
The signal readings of rusted anchor bolts are less than the 4% background ampli-
tude before the first backwall pulse. In addition, readings after the first backwall pulse
display a clear declining signal amplitude pattern as compared to the baseline.
Average UT Results
100
90
80
70
Amplitude (%)
60
50
40
30
20
7.84%
10 2.35%
0
0 50 100 150 200 250 300
30 mm Depth (mm)
Baseline Rusted DefecƟve
Fig. 3 Average UT results for defect-free, rusted and defective anchor bolts
2.1 Process Flowchart for the Test Phase
The flowchart P1 in Fig. 4 shows the procedure of testing the anchor bolts using UT
equipment.
The inspection process starts off with setting up the UT equipment and probe,
followed by applying couplant on the anchor bolt surface end to facilitate the trans-
mission of ultrasonic waves from the probe to the anchor bolt. The probe can be
attached to an extension pole to inspect anchor bolts located at high ceilings. Once
the probe is in contact with the surface of the anchor bolt, signals can be observed
on the UT equipment display, unless the probe is not properly placed or insufficient
couplant is applied.
Invalid signal applies when amplitude is below 15% or no signal is observed,
which then requires the inspection to be redone. A few troubleshooting solutions are
provided in Table 1 if the inspection has to be redone because of invalid signals. A
valid signal is acquired when its amplitude is 15% and above, where these results
are collected or stored for further analysis in the subsequent phases.
2.2 Process Flowchart for the Normalization Phase
The flowchart P2 in Fig. 5 shows the procedure to normalize the collected results. The
purpose of normalizing the signals is to ensure that the collected signal possesses a
first backwall pulse of 80% ± 5% amplitude to allow for comparison to the baseline
or reference signal during the evaluation phase. The first backwall pulse is determined
Fig. 4 Process flowchart P1 for the test phase

Table 1 Troubleshooting solutions for invalid signals

Invalid Signal Cause Solution
(1) No signal shown Insufficient couplant Apply additional couplant
(2) Very low amplitude (less (couplant must cover the
than 15%) entire probe surface)
Misorientation of probe Ensure that the probe and the
bolt axial axes are closely
aligned
Poor surface contact of probe Adjust the probe surface to
with anchor bolt make full contact with the
anchor bolt’s surface area and
apply a greater amount of
pressure to the probe toward
the bolt
from the wave propagation distance that corresponds to the length of the bolt. In the
event that no significant backwall pulse can be found on the signal but other pulses or
spikes in the signal are observed, then the normalization process is skipped to move
straight to the evaluation phase.
2.3 Process Flowchart for the Evaluation Phase
The flowchart P3 in Fig. 6 shows the procedure to evaluate the condition of the anchor
bolt based on the signal collected by comparing it to a reference or baseline signal
collected earlier from the same bolt in pristine condition. Once the condition of the
anchor bolt is identified, it can then be determined if the bolt needs to be replaced.
Evaluation is done by categorizing the condition of the anchor bolt to defect-free,
rusted or defective. Based on previous ultrasonic testing experiments on the anchor
bolts in different conditions, a set of criteria has been determined under these three
categories of evaluation.
The criteria for a defect-free anchor bolt are that the amplitude of the signal
up to the first backwall pulse is less than 4% with similar spike patterns after the
backwall when compared to the baseline. For a rusted anchor bolt, when compared
with the baseline signal, the amplitude of the signal up to the first backwall is less than
4% with a gradual decrease in the amplitude of the signals after the first backwall.
The rusted anchor bolt may require replacement, depending on the severity of the
corrosion, which is the subject of further studies. For defective anchor bolts, when
compared with the baseline signal, the amplitude of the signal up to the first backwall
pulse is higher than 4% with spike pulses appearing before the first backwall. The
amplitude of the spikes can be higher than 15% amplitude when the backwall pulse
is absent. This anchor bolt is considered defective and requires replacement as this
is an indication of cracks present in the bolt which may increase in size over time
Fig. 5 Process flowchart P2 for the normalization phase
and eventually fail with continued loading especially from vibrations in the tunnel
during train operations.
Fig. 6 Process flowchart P3 for the evaluation phase
3 Conclusion
The current test method conducted on the anchor bolts of the OCS in the NEL train
tunnel is by visual inspection. Visual inspection only checks the surface condition
of the exposed anchor bolt section from the concrete tunnel, and thus, their true
condition cannot be identified. A proposed solution is developed using flowcharts
for three phases (test, normalization and evaluation) of an inspection process using
UT. Previous experimental studies show that the anchor bolts can be classified into
three different categories based on their conditions, i.e., defect-free, rusted or defec-
tive, with distinguishing features in the acquired signals. Understanding these distin-
guishing features for the categories of the anchor bolt condition has allowed the
development of a robust inspection procedure at the evaluation phase of the UT
signals.
References
1. A. A. Al-Shehry, UT Inspection Practice for Anchor Bolts, IDSCMT5145, United Kingdom,

Fifth International Conference on Sustainable Construction Materials and Technologies, 2013
2. Benamin L. Ervin, Daniel A. Kuchma, Jennifer T. Bernhard and Henrique Reis, Monitoring
Corrosion of Rebar Embedded in Mortar Using High-Frequency Guided Ultrasonic Waves,
Journal of Engineering Mechanics, 2009
3. Ultrasonic Flaw Detector Omniscan SX, Oylmpus Scientific Solutions Americas Corp., USA,
Waltham, 2010, P/N: 920–269-EN Rev.E
4. Panametrics Ultrasonic Transducers, Oylmpus Scientific Solutions Americas Corp., USA,
Waltham, 2010, Pana_UT_EN_201012
Weld Inspection of Piping Elbow Using
Flexible PAUT Probe
S. J. Lim, T. S. Park, M. Y. Choi, and I. K. Park
Abstract In order to optimize the operation of power generation equipment, struc-

tures, and facilities, and to maintain safety and reliability for a long period of time,
non-destructive testing (NDT) has to be conducted to secure the structural integrity.
Among the various NDT methods, radiographic testing (RT) is mainly used for the
production/installation and scheduled maintenance of power generation facilities
according to relevant codes and specifications, and they mostly have a significant
effect on process delay and economic loss that occur due to duplicate application
of RT and ultrasonic testing (UT). Therefore, the phased array ultrasonic testing
(PAUT) is actively introduced and utilized as a new volumetric examination tech-
nique, which can reduce the use of the RT as much as possible and replace the RT in
the industrial field. The PAUT is capable of high-speed inspection because not only
it can control the injection of the ultrasonic beam electronically, but it can also easily
test specimens with complicated shapes, which are difficult to be detected with the
conventional UT. It is not easy to obtain objective and reliable testing data because
the probe contact is difficult when the PAUT is used to examine the welded part with
curved shapes such as a piping elbow welding part. Keywords: Please list a maximum
of 6 keywords or phrases. Therefore, in this study, instead of using a wedge, which
was essentially used for the inspection of the pipe with curvature structures in the
conventional studies, a 5 MHz, 32 Ch Linear flexible PAUT probe that can examine
a piping elbow welding part was designed and manufactured, and the usefulness was
T. S. Park · I. K. Park (B)

NDT Research Center, Seoul National University of Science and Tehcnology, 232 Gongneung-ro,
Nowon-gu, Seoul, Korea
e-mail: ikpark@seoultech.ac.kr
S. J. Lim
Graduate School of Energy
and Enviroment, Seoul National University of Science and Tehcnology, Seoul, Korea
M. Y. Choi
Safety Measurement Center, Korea Research Institute of Standards and Science, Daejeon, Korea

https://doi.org/10.1007/978-981-15-9199-0_62
648 S. J. Lim et al.
verified. In order to verify the usefulness, welding specimens were manufactured

according to the ASME standard and artificial defects were processed. Finally, the
possibility to overcome the limitations of the existing PAUT technologies and to
improve the reliability of test results were confirmed through and signal acquisition
and analysis.
1 Introduction
It is necessary to ensure structural integrity through non-destructive testing (NDT)

in order to optimize the operations of power generation units and structures/facilities
while maintaining long-term safety and reliability. The main NDT technique that is
used for power generation unit construction, installation, and preventative mainte-
nance in South Korea is radiography testing (RT) owing to the relevant codes and
regulations. In most cases, RT and ultrasonic testing (UT) are performed together,
which contributes greatly to process delays and economic losses. Moreover, as RT
has been used in the field, many incidents have occurred in which testers have been
killed due to high levels of radiation exposure. As a result, safety regulations in
South Korea have been strengthened across the board, and the relevant laws have
been revised so that RT can only be performed under strictly limited conditions.
Because of these changes, RT consumes a great deal of time and causes construction
and manufacturing process delays, which lead to huge economic losses and weaken
exporting competitiveness. To resolve this, the most efficient method for industrial
sites around the world has been to reduce their dependence on RT where possible and
introduce and use new non-destructive volumetric examination technologies such as
phased array UT (PAUT) and time of flight diffraction (TOFD) that can replace
RT. Unlike conventional UT, PAUT normally arranges multiple vibration elements
and electronically controls them to create ultrasonic beams at the desired angle. It is
capable of fan-shaped scans, rear scans, etc., which allow a fixed area to be examined
without mechanical scanning. The scan results are depicted and stored as images, and
it has garnered attention as a technique that can replace RT. However, conventional
PAUT has limited access to welded parts when the testing target has elbows and
complex shapes, which makes testing difficult [1]. Moreover, when curved items are
tested, a wedge that is suitable for the curvature must be designed, manufactured,
and used. Therefore, to overcome these problems, this study has designed and manu-
factured a flexible probe that can test piping elbow welds in power generation units,
and the performance and usefulness of the probe was verified.
Weld Inspection of Piping Elbow Using Flexible PAUT Probe 649
2 Experimental Details
2.1 Design and Fabrication of Flexible Probe
Figure 1 shows a PAUT probe that was designed and manufactured as a 5 MHz, 32-
channel flexible linear type probe. Each element’s size was designed to be 0.4 mm
× 10 mm, and the gaps between the centers of the elements (pitch) was 0.5 mm.
The probe was designed to give the vibrator sufficient access to the welded area, and
the matching layer and damper components were designed so that they were located
on the end of one side’s surface [2]. The probe’s main features are flexibility and
bending in the probe cross section’s X and Y axes. With these features, it can perform
scans to obtain accurate test data by adhering to the curved welded area’s surfaces
and capturing good signals. It was also designed with a structure that allows for UT
using the PA technique [3]. The designed flexible probe as manufactured by Japan
Probe Inc., and the specific design specifications are shown in Table 1.
Fig. 1 Flexible probe design
Table 1 Design parameter of the flexible probe

Frequency 5 MHz
The number of element 32 CH
Element size 0.4 mm × 10 mm
Element pitch 0.5 mm
Case material POM(BK)
Connector type IPEX
2.2 Elbow Specimen
Figure 2 shows the shape of the specimen that was fabricated to verify the perfor-
mance and usefulness of the flexible PAUT probe that was designed and manu-
factured. The specimen’s curved welded area was designed and manufactured
so that its shape and specifications met ASME B16.9 (Factory-made wrought
buttwelding fitting). The specimen’s material was a normal ferritic steel, and it was
designed/manufactured with a 6-inch diameter, a 10.9 mm thickness, and a 90° bend.
Here, the welded area that was needed to verify the performance and usefulness of
the actual flexible probe was limited to the area marked in red at the inner side of the
elbow, as shown in Fig. 3. For the artificial defects, three EDM notches and one flat
bottom hole (FBH) were made, as shown in Fig. 3. The width, length, and depth of
each defect are shown in Table 2.
Fig. 2 Dimensions and shape of elbow specimen
Fig. 3 Location of artificial defects
Table 2 Artificial defect size

No. Width Length (mm) Depth (mm)
1 ∅1.5 – 2
2 0.3 mm 3 1
3 0.3 mm 5 1
4 0.3 mm 5 2
Fig. 4 Experimental setup of the PAUT System
2.3 Elbow Specimen
Figure 4 shows the equipment setup used in this study. An existing commercial
system made by Olympus (Model: Omni Scan MX, 32Ch-128 P/R) was used as the
PAUT system, and the data analysis was performed by using the Tomoview software
program to analyze the defects. To compare the proposed probe to an existing PAUT
probe, the same specimen was analyzed by a 7.5 MHz CCEV probe that uses a
wedge. Before testing, the equipment was calibrated using the calibration block that
is used in ISO 18563-3:2015.
3 Results
3.1 Characterization of the Flexible Probe
This study’s flexible probe was designed and fabricated to have a central frequency
of 5 MHz. A frequency spectrum analysis was performed to confirm that the probe
was fabricated to the designed frequency specifications. The signal had a central
frequency of 5.4 MHz, and the frequency bandwidth showed broadband performance.
The probe has 32 array elements, and each element must have a fixed amplitude
value and frequency characteristics. Therefore, the frequency and amplitude were
measured for each of the 32 elements. Figure 5 shows the frequency characteristics
and amplitude values measured for each element. It can be seen that the 32 elements
have very similar amplitudes and frequency values.
(a) Gain (b) Frequency
Fig. 5 Gain and frequency for each element
3.2 Detection of Artificial Defects by PAUT Probe
Figure 6 shows the results of testing the elbow specimen. The data were acquired
using the flexible probe and a PAUT probe that uses a wedge. It was confirmed that
the flexible probe detected all of the artificial defects at the accurate locations. When
the proposed probe was compared to the existing PAUT probe’s results, it was found
that the defects’ locations were detected accurately, and the signal’s sensitivity was
detected at similar levels.
4 Conclusions
Conventional PAUT has limited access to welded parts when the testing target has
elbows and complex shapes, which makes testing difficult. Furthermore, when curved
items are tested, a wedge that is suitable for the curvature must be designed, manu-
factured, and used. This study designed and manufactured a flexible probe that has
flexibility and bending features to resolve the problem of reduced reliability caused
by curves such as piping elbows in power generation units. Signals were obtained
from a welded specimen that contained artificial defects, and these signals were
analyzed. The following conclusions were obtained.
(1) The fabricated probe showed broadband characteristics and the same sensitivity
on all channels (32 channels). Through this, it was confirmed that the design
and fabrication of the probe met design specifications.
(2) A specimen with artificial defects was examined using the flexible probe and the
existing PAUT probe that uses a wedge. The results revealed that both probes
accurately detected the locations of all artificial defects. The defect detection
results confirmed that the developed flexible probe can be used as a substitute
for the existing probe that uses a wedge.
Defect Flexible PAUT Conventional PAUT
FBS
Notch
Notch
Notch
Fig. 6 Comparison of the conventional PAUT probe with flexible probe
(3) In the future, various empirical tests will be performed using mockup specimens,
and the proposed flexible probe will be used to replace RT, which can be used
to test welded areas in the piping elbows of thermal power generation units.
Acknowledgements His work was supported by the Korea Institute of Energy Technology Evalu-
ation and Planning(KETEP) and the Ministry of Trade, Industry & Energy(MOTIE) of the Republic
of Korea (No. 20181510102130).
References
1. Jayasinghe V (2015) Application of flexible PAUT Probes for small diameter flow assisted
corrosion elbow inspection. In: ASNT Annual Conference, pp 82–86
2. Bray DE (1989) Nondestructive evaluation. Array Transducers, pp 98–100
3. Asher RC (1997) Ultrasonic sensors. Piezoelectric Transducers, pp 84–122
Laser Ultrasonic Imaging of Wavefield
Spatial Gradients for Damage Detection
S. Y. Chong, Z. Wu, and M. D. Todd
Abstract Laser ultrasonic techniques (LUTs) are increasingly used in non-

destructive evaluation (NDE) applications because they can provide full-field ultra-
sonic wavefield data with high resolution in a three-dimensional (3D) space–time
domain. This paper investigates the properties of the wavefield spatial gradient as a
detector and localizer for defects that distort the local wavefield. A laser ultrasonic
interrogation method based on a 527-nm Q-switched laser scanning system was used
to interrogate 3D ultrasonic signals in a 3-mm aluminum plate. The plate was tested
with and without multiple artificial scatterers to test the proposed method. During
the scanning process, the ultrasonic waves were generated by the pulsed laser at an
n-by-m grid of sampling points and recorded by a single fixed PZT transducer. The
whole scan generates a 3D matrix which allows the accessibility of showing the wave
propagation over time in movie form with a high spatial resolution. After the scanning
process, the full-field ultrasonic data of the plate without and with artificial damage
were obtained, respectively. Then, spatial gradient vectors were computed from each
frame of the full-field ultrasonic data along the time axis. For a space-domain frame
{Ui }n×m at a time i ∈ t, the gradient vectors were determined by two components,
n×m n×m
G i,x and G i,y , which reflected the spatial derivatives of every pixel with
respect to the position coordinates. Once the gradient vectors of the full-field ultra-
sonic data (intact and damage cases) were determined, the spatial gradient ultrasonic
wavefield imaging was generated with the gradient orientation profile {θi }n×m and the
gradient magnitude profile {Ai }n×m as the features used for damage detection. The
wavefield spatial gradients of the defect-free case were used as a baseline to compare
with the wavefield spatial gradients of the damaged case. The gradient profiles of
the wavefield spatial gradients could detect and localize the damage by the inconsis-
tency of the wave pattern in the 2D space domain. The wavefield spatial gradients
also showed the ability to remove the complexity of identifying the damage among
multi-mode waveform. Future work will be focusing on applying the algorithm to
complex structures such as CFRP composites.
S. Y. Chong · Z. Wu · M. D. Todd (B)

Department of Structural Engineering, University of California San Diego, 9500 Gilman Drive,
La Jolla, CA 92093-0085, USA
e-mail: mdtodd@eng.ucsd.edu

https://doi.org/10.1007/978-981-15-9199-0_63
656 S. Y. Chong et al.
Keywords Guided waves · Laser ultrasonics · Wavefield imaging ·

Non-destructive evaluation · Spatial gradient analysis
1 Introduction
Laser ultrasonic techniques are increasingly used in non-destructive evaluation appli-

cations because they can provide full-field ultrasonic wavefield data with high spatio-
temporal resolution. The full-field ultrasonic wavefield data is generally formed via
3D imaging techniques, known as ultrasonic wavefield imaging (UWI). Obtaining
the UWI directly with such a high resolution (and extracting certain features) gives
unprecedented insight into how the waves interact with boundaries and how the wave-
field may be distorted by scattering from potential defects. With these advantages,
many feature-extraction algorithms have been developed based on the temporal- and
spatial-frequency transformation methods [1, 2] to detect damage such as fatigue
cracks, impacts, and corrosion. Temporal-correlation [3] and spatial covariance [4,
5] methods were developed as well to detect and localize the damage.
Beyond these features mined from the UWI, the wave propagation orientation of
the ultrasonic wavefield has not been considered in ultrasonic interrogation. Since the
UWI is formed in 3D space–time domain, 2D spatial gradient field can be determined
in each time frame. Since boundary or material changes are known to induce changes
in spatial ultrasonic field properties, it may then be reasonably further surmised that
defects may be detected by monitoring the change of the spatial gradient field as
time progresses.
Many spatial gradient field analyses are used in seismic and image recognition
applications. In seismic applications, the spatial gradient of the seismic wavefield is
analyzed to investigate important wave characteristics that can be directly related to
wave propagation in layered earth [6]. In image recognition applications, the spatio-
temporal gradient method is used to recognize a moving object in video data [7].
Hence, it suggests the hypothesis that the change of the local spatial gradient of the
UWI may be detected through the spatial gradient analysis method to the ultrasonic
wavefield imaging.
In this paper, we propose a wavefield spatial gradient analysis method to laser
ultrasonic imaging for damage detection. First, the spatial gradient analysis method
to the laser ultrasonic imaging is presented. Then, the experimental setup to obtain
the full-field ultrasonic wavefield data from the specimen with and without artificial
defects is discussed. Lastly, the detectability of the proposed method on multiple
simultaneous defects is discussed.
Laser Ultrasonic Imaging of Wavefield Spatial Gradients … 657
2 Ultrasonic Wavefield Spatial Gradients
The proposed method is based on acquiring the gradient orientation and the gradient
magnitude profiles along with time history. To evaluate the presence of potential
defects, the 2D spatial gradient vectors of each time frame (denoted as i-index) are
calculated and the corresponding gradient vector components may be expressed by
∂Ui ∂Ui
{G i,x }n×m = , {G i,y }n×m = , (1)
∂x ∂y
which represent the differences for each pixel of {Ui }n×m in the x (horizontal) and
y (vertical) directions, respectively. The spacing between pixels in each direction is
adjusted with the expected spatial resolution. Based on Eq. (1), the gradient orien-
tation and gradient magnitude profiles for each time frame{θi }n×m and{Ai }n×m , can

be determined by the relation of θi = tan−1 (G i,y G i,x ) and Ai = G i,x 2
+ G i,y
2
,
respectively. Thus, by integrating the variations of these spatial gradient’s compo-
nents with the sampled time, i, the gradient orientation and gradient magnitude maps
of the full-field ultrasonic wave can be generated

tend −1
tend −1

{G θ }n×m = {θi+1 } − {θi } , and {G A }n×m = [{Ai }], (2)
i=1 i=1
respectively. Consequently, these maps may detect and visualize the damaged area
with a change from approximately low accumulation to high accumulation. However,
the spatial gradient vectors are sensitive to the reflected waves, too, of course. The
reflected wave may make these maps fail to distinguish the defects located near the
boundary of the specimen. Given that, a scanning result of the defect-free case,
n×m
Ui,0 , is considered and pre-recorded as a baseline and the gradient orien-
n×m
tation and gradient magnitude maps obtained from Ui,0 are defined as the
n×m
residual gradient orientation map G θ,0 and the residual gradient magnitude
n×m
map G A,0 , respectively. For that, ideally, the effect of the reflected wave can
be removed by subtracting the baseline from the damaged data expressed in the
following form:
n×m n×m n×m n×m
G θ,sub = {G θ }n×m − G θ,0 , and G A,sub = {G A }n×m − G A,0 ,
(3)
n×m n×m
where the G θ,sub and G A,sub are the residual trajectories caused by the
defects. Both residual maps may have different damage detection patterns which
can be further interpreted to localize the damages according to their sensitivities to
different types of damages. In this paper, the circle scanning data was considered, and
the space-domain n-by-m frame (x- and y-directions) corresponds with the dimension
defined by total inspection scan angle points and total inspection scan radius points
in the Cartesian coordinate scheme. The angular direction of the damage can be
accurately determined in light of the spatial covariance method [4, 5].
Figure 1a shows a schematic diagram of a laser ultrasonic interrogation system

(LUIS). This system consists of a laser scanning system incorporated with a signal
conditioning device, a data acquisition (DAQ) module, a contact sensor, and a
computer used for signal processing and operation control.
The laser scanning system has a two-dimensional laser scanner and a diode-
pumped solid-state Q-switched Nd:YAG laser. The Q-switched laser is composed of
a laser controller and a laser head with an output beam of 527 nm wavelength and
pulse repetition rates (PRR) from single-shot to 1 kHz. The 2D laser scanner is used
to synchronize the 2-axis galvanometer scanner (Fig. 1b) with the Q-switched laser
to maneuver the laser impinging point rapidly on a target at a PRR according to the
preset scanning pattern coordinates. As for this paper, the circle scanning pattern was
considered, and the ultrasounds were generated and obtained at each coordinate of
(R, θ ) in the polar coordinate space. The scanning radius ranges were set from 20 to
220 mm with an interval of 0.5 mm, and the angular interval was set to 0.26°.
During the scanning process, a laser pulse impinges at a scan point, and the
corresponding ultrasound is generated at the local point based on the thermoelastic
principle. The generated ultrasound is then signal-conditioned through a contact
sensor and digitized in the DAQ module as shown in Fig. 1a. In the DAQ module, the
generated ultrasound is digitized synchronously with the PRR of the laser scanning
system by receiving a triggering signal from the laser controller after the laser pulse
is emitted. In the digitizing process, the ultrasound is sampled for K number of data
Fig. 1 a Laser ultrasonic interrogation system configuration, b 3-mm aluminum plate configura-
tions with simulating defects, and c ultrasonic wavefield image at 82 µs
points (each sample point is denoted as index i) at sampling time interval T S and
stored in a computer. The digitizing process is repeated as the scanning process is
performed. In this paper, the K was set to 2000 at the sampling interval of T S =
0.2 µs.
Lastly, all digitized ultrasonics are represented in the form of a three-dimensional
R-θ-T matrix, indexed by spatial-horizontal-direction (m), spatial-vertical-direction
(n), and i time samples, respectively, along with each dimension. In this paper, two
different notations of 2D Cartesian coordinates scheme were used for the {Ui }n×m
obtained from the polar scan. One was presented in the R-θ domain, and the other
one was presented in the X–Y domain. The corresponding coordinates of X and Y
axes were simply determined by the relationships X = R cos θ and Y = R sin θ ,
respectively. For example, Fig. 1c shows the snapshot of an ultrasonic wavefield
image at 82 µs, representing in X–Y domains.
Figure 1b shows the defect types and locations as considered for the multiple
damage scenarios. Four scanning processes were performed. First, the specimen
without any defects was considered and used as a baseline for the damage detection.
Second, the single-damage case was considered with plasticine at location 2. Third,
the multi-location damage case with plasticine at location 1 to 4 as shown in Fig. 1b.
Last, the multi-type damage case was considered by including a screw (inset of
Fig. 1b), where the red surface was glued on the rear scanning surface at location 5.
4.1 Single-Damage Case
The proposed method is firstly demonstrated for a single-damage case. A piece

of plasticine (20 mm × 20 mm) was bonded at location 2 of the aluminum plate
by grew acting as a damper to the guided waves. Figure 2a shows the ultrasonic
wavefield image at 60 µs. The attenuated local amplitudes of the A0 mode were
Fig. 2 a UWI at 60 µs, b gradient orientation map for the single-damage case, and c the
corresponding angle covariance of the gradient orientation map obtained from circle scanning
Fig. 3 a gradient orientation map for the single-damage case at 64 µs, b the corresponding angle
covariance of the gradient orientation map, and the residual gradient orientation map based on
baseline method at 90 µs
barely visible at the damage location. In contrast, the gradient orientation map (2),
as shown in Fig. 2b, was able to detect and localize the damage at location 2 with the
Cartesian coordinate of (−95, 125). Then, the angular direction of the damage was
also accurately determined at 125° using the spatial covariance method [4, 5] to the
gradient orientation map, as shown in Fig. 2c.
However, even though the damaged area was detected in the gradient orientation
map with a change in variation summation from approximately zero value to a higher
accumulation, the methodology is extremely limited by the incident wave because
the presence of the reflected wave, after the time of 60 µs, distorts the consistency of
the guided wave and make the undamaged area that near the boundary have a higher
accumulation, as shown in Fig. 3a at 64 µs. Figure 3b shows the false positive on
the covariance map which due to the reflected waves. Thus, the baseline subtraction
method was considered and showed the ability to detect the damage as showed in
Fig. 3c. Figure 3c shows the gradient orientation map is able to indicate the location
of the damage even furthermore the time span (>60 µs) where the reflected waves
were starting propagated toward the center plate.
4.2 Multi-location Damage Case
The multi-location damage case with three artificial plasticine pieces including one
near the plate boundary was conducted to evaluate the feasibility of the proposed
method after removing the reflected wave limitation through the baseline subtraction.
Figures 4a, b show that all three simulating defects were detected in the gradient orien-
tation map clearly at the time 120 µs. However, there were false-positive detected
regions (Fig. 4a) caused by the scattering effect of the damage (location 1). In this
case, the defect at location 4, which is outside the scanning area (Fig. 1b), was not
able to be detected.
Fig. 4 Multi-location damage case: a Residual gradient orientation map with baseline subtraction
(reflected wave removed), and b corresponding angle covariance of the residual gradient orientation
map obtained from circle scanning
4.3 Multi-type Damage Case
In this case, a screw was bonded in the mark location 5, as shown in Fig. 1b, acting
as a scattering source to evaluate the gradient magnitude map result. The gradient
magnitude map (3), as shown in Fig. 5a, was able to detect and localize the multi-
type and multi-location damages (at location 1, 2, 3, and 5) with the polar scan
coordinates. The magnitude-based map is converted into Cartesian coordinate as
a more intuitive view. The angular direction of the damages was again accurately
determined at 125°, 180°, 220°, and 295° using the spatial covariance method to
the gradient magnitude map, as shown in Fig. 5c. The four trajectories with distinct
patterns in this map indicate the change of the wavefield caused by the four artificial
defects in the experimental case. With the angle indicated by the corresponding
angle covariance map (Fig. 5c), the damages can be located by the peak value for
each trajectory.
Fig. 5 Multi-type damage case: Residual gradient magnitude map on a R-θ axes and b x–y axes
with baseline subtraction, and c corresponding angle covariance of the residual gradient magnitude
map
5 Conclusion
In this paper, the laser ultrasonic imaging of the wavefield spatial gradients for
damage detection was proposed. The ultrasonic wavefield spatial gradient analysis
method with the gradient magnitude and gradient orientation maps was presented and
used to detect the damages by monitoring the change in the components of the spatial
gradient vector. Then, the proposed method was tested on the ultrasonic wavefield
imaging that obtained from the aluminum plate with four different configurations
of the artificial damages. In the first case, the gradient orientation map showed the
ability to detect the single damage but only limited to using the incident waves.
The presence of the reflected waves made the map difficult to distinguish between
changes due to the scattering waves from the damage and the reflected waves from
the plate boundaries. To address that, the baseline subtraction method was considered
to suppress the effects of the reflected wave. Then, the residual gradient orientation
map obtained by the baseline subtraction was tested on the multiple damages case.
It showed the ability to detect the multiple damages, but there was false positive
that due to the scattered wave from the damage boundaries. In the last cases, the
residual gradient magnitude map was considered on the multi-type damages. The
map was able to identify the distinct patterns due to the damages which were different
from the residual gradient orientation map. The map showed four trajectories with
distinct patterns in the map that indicated the change of the wavefield caused by four
artificial defects. Upon that, the spatial covariance of the residual gradient orientation
and gradient magnitude maps showed the good agreement of the angular directions
of the damages, respectively. The proposed method showed the detectability of the
multiple and multi-type artificial damages even the damage was located at the end of
the scanning region (location 3). However, it was not able to detect the damage located
outside the scanning region. In future works, the spatial gradient analysis method will
be further tested on the complex wavefield patterns, e.g., in the anisotropic material
plate-liked structure.
Acknowledgements The authors acknowledge funding from UC Office of the President for partial
support of this work.
References
1. Lee J-R et al (2011) Laser ultrasonic propagation imaging method in the frequency domain
based on wavelet transformation. Opt Lasers Eng 49(1):167–175
2. Flynn EB et al (2013) Structural imaging through local wavenumber estimation of guided waves.
NDT E Int 59:1–10
3. Lee JR, Chia CC, Park CY, Jeong H (2012) Laser ultrasonic anomalous wave propagation
imaging method with adjacent wave subtraction: algorithm. Opt Laser Technol 44(5):1507–1515
4. Chong SY, Todd MD (2018) Dispersion curve estimation via a spatial covariance method with
ultrasonic wavefield imaging. Ultrasonics 89:46–63
5. Chong SY, Todd MD (2018) Statistical damage detection based on full-field covariance of
circumferential scan ultrasonic measurement. In: Health monitoring of structural and biological
systems XII, vol 10600. International Society for Optics and Photonics, p 106002J
6. Van Renterghem C et al (2017) Spatial wavefield gradient-based seismic wavefield separation.
Geophys J Int 212(3):1588–1599
7. Liou S-P, Jain RC (1989) Motion detection in spatio-temporal space. Comput Vis Graph Image
Process 45(2):227–250
Guided Waves Testing
Guided Ultrasonic Waves in Glass
Laminate Aluminium Reinforced Epoxy
M. Rennoch, M. Koerdt, A. S. Herrmann, N. Rauter, and R. Lammering
Abstract Knowledge on the structural health of hybrid composites that consist

of laminated layers of thin metal sheets and fibre reinforced adhesive films are of
increasing importance due to growing use in aircraft industry. For this purpose,
we propose damage detection techniques based on the propagation behaviour of
guided ultrasonic waves. Therefore, it is crucial to know the characteristic effect
of fibre metal laminates on the wave propagation. Accompanying, in this study, we
investigate experimentally and numerically how guided ultrasonic waves propagate
in a fibre metal laminate.
Keywords Guided ultrasonic waves · Structural health monitoring · Glass

laminate aluminium reinforced epoxy · Fibre metal laminate
1 Why Knowledge of Guided Ultrasonic Waves in Fibre

Metal Laminates Is Important
In pure metal cracks grow easily, whereas for structures made of fibre reinforced
composite impact damages are of significant risk, causing delamination or fibre
crack. Combining both materials in a layered structure, so called fibre metal laminate
(fml), minimises the potential risks: on the one hand, the composite material stops
the crack propagation in the metal layer [1]. On the other hand, the metal layers are
mainly used at top and bottom layer to absorb impact energy [2]. Nevertheless, the
M. Rennoch (B) · M. Koerdt

Measurement Systems & Monitoring, Faserinstitut Bremen e.V., Am Biologischen Garten 2,
Bremen, Germany
e-mail: rennoch@faserinstitut.de
A. S. Herrmann
Faculty Production Engineering—Mechanical Engineering & Process Engineering, Materials
Engineering/Fibres and Fibre Composites, University of Bremen, Bremen, Germany
N. Rauter · R. Lammering
Institute of Mechanics, Helmut-Schmidt-University/University of the Federal Armed Forces,
Hamburg, Germany

https://doi.org/10.1007/978-981-15-9199-0_64
668 M. Rennoch et al.
energy introduced can lead to plastic deformation caused by the metal layers [3].
However, fml is extensively used by aerospace industry, e.g. in the fuselage of the
Airbus A380 aircraft, lower wing skin panels of previous Fokker 27 aircraft and
cargo door of Boeing C-17. Therefore, knowledge on impact related damage, e.g.
due to bird strikes, debris on the runway or similar is crucial. Thus, the behaviour of
crack growth in fml is investigated [4]. However, due to the very complex structure
and the combination of isotropic and anisotropic materials with different material
properties, its use remains very challenging. This does not only refer to production
and design, but also to the damage detection.
For damage detection in composite structures, the use of guided ultrasonic waves
(guw) was profoundly investigated over the last decade [5]. It has been shown that
guw are suitable for damage detection as their propagation behaviour changes when
interacting with damage [6]. Despite this, so far, only few publications investigate
wave propagation in fml, and yet, the effects of delamination between fibre rein-
forced epoxy and metal layers as well as cracks in the metal layers on guw propaga-
tion are studied with finite element methods (fem) [7]. This indicates how important
numerical methods are for the well-founded analysis of wave propagation and its
suitability for damage detection. With their help, a variety of phenomena can be
efficiently investigated and fundamentally examined.
From a computational mechanics point of view, 2D and 3D forward models are the
fundament of the numerical simulation of guw in thin-walled fml structures. In this
study, the wave propagation will be simulated by use of a 2D finite element model
based on the plain strain assumption. Therefore, the wave propagation is analysed
through the thickness of the waveguide and along one propagation direction across
the plate. The main challenge lies in a proper spatial discretization of the structure
as well as a sufficiently fine time discretization to ensure an exact approximation of
the displacement field.
Although fml became popular by aramid reinforced aluminium laminate (arall)

[8], it was replaced by the development of glass laminate aluminium reinforced
epoxy (glare), partly because of the lower strength of the aramid fibres [2].
Additionally, glare is characterised by excellent fatigue strength, high specific
strength, excellent impact strength, and low weight [9]. For this study, we manu-
factured a plate of 500 mm × 500 mm × 4 mm in size, made from primed
aluminium 2024 T3 and aerospace qualified bidirectional glass-fibre woven fabric
prepreg EHG250-68–37 T2 100 in 8-harness satin weave. The total thickness
results from 0.25 mm thick glass-fibre layers (gf), aluminium layers (al) of
0.4 mm thickness each, and a layup according to impact robust glare-5, in fact
[al/gf/gf/al/gf/gf/al/gf/gf/al/gf/gf/al]. We fabricated the plate with a hotpress
vacuum bag process, i.e. curing for 40 min at 155 °C under a pressure of approx.
4 bar and slowly cooling down to 60 °C under continuous pressure. Afterwards, we
Guided Ultrasonic Waves in Glass Laminate … 669
trimmed and sealed the edges. In the upper corners of one side, we drilled holes for
the further hanging storage. Figure 3 provides a microsection from an edge of the
manufactured plate.
At the surface of the specimen, we adhesively bond in a line three circular trans-
ducers, which components are made of ferroelectric soft piezo material PIC255,
modified lead (Lat.: plumbum) zirconium titanate (pzt) that is the discs, 0.2 mm
thick with a diameter of 10 mm, are bonded equidistantly at 7.6 cm with cyanacry-
late adhesive Z70, such that the second transducer is at the centre of the plates surface.
Therefore, the thickness of the adhesive layers was measured via laser triangulation
and is about 50 µm for every disc, see Fig. 1, showing even bonding. Likewise, three
transducers are bonded with an epoxy at similar positions on the backside of the plate,
for which height profile measurements indicate non-uniform bondings. Further, the
transducers are soldered to electrically shielded cables. Then, the sample is placed
freely suspended in a climate test chamber and connected to the outside measurement
devices.
The central piezoelectric wafer active sensor (pwas) on the front is used as actu-
ator, whereas the lower pwas serves as sensor. A Hanning windowed sinus burst of
five-cycles is excited with a peak-to-peak voltage of 1.3 V and carrier frequencies
ranging from 40 to 260 kHz in steps of 20 kHz by an arbitray waveform generator HP
Agilent 33120A and amplified by a high-voltage amplifier WMA-300 from Falco
systems that results in a peak-to-peak voltage of around 60 V. The guw, hence,
propagate at 20 °C and 50% relative humidity and were recorded by a digital storage
oscilloscope Tektronix DPO5054, which averages the signal for denoising for 30 s.
Additionally, we apply a 33 k resistor as high-pass filter to get rid of low-frequency
noise. Figure 2 depicts the recorded signals.
Fig. 1 Height profile from laser triangulation of top pwas. The sensor is bonded parallel to the
plates surface (blue, zero height) and due to its 0.2 mm thickness, indicates 0.1 mm thickness of
the adhesive layer (Heigths above 1.3 mm correspond to soldered wires)
Fig. 2 Recorded guw signals according to carrier frequencies from 40 to 260 kHz in steps of
20 kHz
Fig. 3 Micro cross-section from an edge of the manufactured glare (right), scheme (centre) and
discretization (left) of the numerical model (The al layers are 0.4 mm thick)
To determine the group velocities cg of the recorded wave packages, we calculate

the absolute value of the Hilbert envelope of the signal and identify the moment of its
maxima at the corresponding wave packages. The same is done with the excitation
burst. As the distance is known, the group velocities at 100 kHz are given as 4252 m/s
for the first wave package and 1410 m/s for the second wave package. Here, we
provide explicit values only for that frequency throughout this work, because of that
the group velocities for analytical and numerical solutions are given in Table 3.
The numerical simulations are carried out using a 2D model under plane strain
assumption. This allows one to reduce the degree of freedom significantly. Below,
the numerical model of the waveguide, the specifications of the numerical simulations
as well as the results are described in detail.
For the numerical model of a 2D section of the specimen, symmetry properties
are used to reduce the degree of freedom even further. As the actuator is placed in the
middle of the specimen, only one propagation direction is modelled and symmetry
boundary conditions are applied to one end. To achieve a sufficient displacement field
of the wave propagation, each layer is modelled individually in this study. Because of
that, the bidirectional gf is modelled from two orthogonally individual unidirectional
(ud) layers via the mosaic model. The used material properties are given in Table 1.
To analyse the wave propagation in more detail, instead of representing the
250 mm half length of the specimen, the numerical model is set to 750 mm in length.
Table 1 Material properties of aluminium (Al 2024 T3) and unidirectional E-glass-fibre (ud gf)
[10]
Material E 1 (GPa) E 2 (GPa) ν12 G 12 (GPa) G 23 (GPa) ρ (kg/m3 )
Al 2024 T3 73 73 0.33 26.6 26.6 2700
ud gf 41 10.4 0.28 4.3 3.5 1850
This allows one to reduce the interaction of the wave modes with waves reflected at
the end of the waveguide. In accordance to the specimen, the model has a thickness
of 4 mm.
For the discretization of the numerical model, 2D plane elements with septic
shape functions are used. Each ud gf layer is discretised with one element over the
thickness. For an evenly distribution of the element length in thickness direction,
four elements over the thickness are used for the aluminium layers. Due to the higher
order shape function, the element length along the wave guide can be significantly
enlarged. In this study, elements with a size of 0.125 mm × 1 mm are used. In Fig. 3
, the structure and its discretization are shown.
For the numerical simulation, the wave is excited by an out-of-plane force at the
left corner point of the upper waveguide surface. The used signal is identical to the
excitation signal that is used in the experiments. Therefore, a Hanning windowed
sinus burst of five-cycles is applied to the node. The centre frequency is set to 100 kHz,
which is inside the frequency range used during the experiments. This excitation
frequency is chosen because it offers a good compromise between the expected
wavelengths of the individual wave modes and the complexity of the wave field.
For lower frequencies, the wavelength becomes so large that the model has to be
extended considerably in order to perform a well-founded analysis of the individual
wave packages. If a significantly higher excitation frequency is selected, further wave
packages are excited in addition to the fundamental modes. Based on the dispersion
curves as a solution of the equation of motion, considering the boundary conditions
for guw, two wave modes are expected for the used material. These modes have
a phase velocity of 4887 m/s and 1287 m/s at a frequency of 100 kHz and can
be assigned to a symmetric and an antisymmetric Lamb wave mode, respectively.
Finally, the time stepping t is set to 0.19 µs, which equals T /20 in accordance
with [11]. Here, T is the period.
In Fig. 4 the in- and out-of-plane displacement fields of the wave propagation are
shown after 125 µs as well as in a distance of 0.2 m from the excitation point. As
expected, the excitation at 100 kHz generates two wave packages. However, for the
out-of-plane displacement component, the first wave package is hardly detectable.
Beside this, the overall behaviour of the wave packages can be clearly assigned to a
symmetric and an antisymmetric Lamb wave mode. The first wave package shows a
dominant in-plane displacement component, whereas the out-of-plane displacement
is dominant for the second wave package. Furthermore, the in-plane displacement of
the second wave package as well as the out-of-plane displacement of the first wave
package at the upper and lower surface are oppositely orientated. Based on these
characteristics, the first wave package shows symmetric and the second one anti-
symmetric Lamb wave mode properties. To verify this conclusion more in detail, the
wave lengths λ and frequencies f are determinded by a wavelet transform using the
Morlet wavelet as mother wavelet. Afterwards, these parameters are used to calcu-
late the corresponding phase velocities c p . Furthermore, the wavelet transform of the
spatial data can be used to decompose the numerically obtained wave propagation
data of the two wave packages as their wave lengths differ significantly. Based on the
separated wave data, the group velocities c g are determined by generating the signals
Fig. 4 In- and out-of-plane displacement field in the spacial and time domain at upper and lower
surface in a distance of 0.2 m to the actuator (top) and after 125 µs (bottom)
envelope using the Hilbert transform at two different points in time. The difference
of the two envelope maxima equals the propagation distance of the wave package.
The results for the upper surface are given in Table 2.
Due to the hardly detectable out-of-plane displacement component for the
symmetric wave mode, the determination of the group velocity is inaccurate. There-
fore, this value is neglected. The calculated phase and group velocities are finally
compared with the analytically obtained values by using the dispersion relation. Table
3 summarises the deviations between numerical and analytical results.
Table 2 Results of the numerical simulation based on the displacement field at the upper surface
Wave package f (kHz) λ (mm) c p,num (m/s) cg,num (m/s)
WP 1 (symmetric) in-plane 101 45.37 4582 4700
WP 1 (symmetric) out-of-plane 101 45.37 4582 –
WP 2 (antisymmetric) in-plane 101 13.97 1411 1680
WP 2 (antisymmetric) out-of-plane 101 13.97 1411 1650
Table 3 Deviation of numerical simulation and analytical solution

Phase velocity c p Group velocity cg
Wave package Numeric Analytic Deviation Numeric Analytic Deviation
(m/s) (m/s) (%) (m/s) (m/s) (%)
WP 1 4582 4887 6.2 4700 4820 2.5
(symmetric)
in-plane
WP 1 4582 4887 6.2 – 4820 –
(symmetric)
out-of-plane
WP 2 1411 1287 9.0 1680 1581 6.3
(antisymmetric)
in-plane
WP 2 1411 1287 9.0 1650 1581 4.4
(antisymmetric)
out-of-plane
4 Résumé
The numerical simulations meet the analytic results with a deviation of approximately
2% to 9%. However, there are no remarkable differences between the in- and out-of-
plane displacement component. These indicate that the chosen simulation parameters
are suitable for the numerical simulation of guw in fml. Furthermore, the results lead
to the conclusion that the wave propagation in fml meet the analytical description
of the Lamb waves for a propagation along a symmetry line in a composite mate-
rial. Additionally, the group velocities from the experimental data meet the analytic
and numerical results for the first wave package with a deviation of approximately
12% and 9.5%, respectively. For the second wave, package group velocities deviate
approximately 11% from analytic results and 16% from the numerical results.
Acknowledgements M. Rennoch gratefully acknowledges the financial support of his research by

the Federal Ministry for Economic Affairs and Energy on the basis of a decision by the German
Bundestag under the grant number 49VF180016.
References
1. Beumler T (2001) Damage tolerance aspects. In: Gunnik JW, Vlot A (eds) Fibre Metal
laminates. An introduction, 1st edn, Springer, Netherlands, pp 219–233
2. Chai GB, Manikandan P (2014) Low velocity impact response of fibre-metal laminates—a
review. Compos Struct 107:363–381
3. Bienias J et al (2014) Experimental and numerical investigations of low-velocity impact
behavior of selected fibre metal laminates. In: 16th European conference on composite
materials, Seville, pp 1–6
4. Asghar W et al (2017) Investigation of fatigue crack growth rate in carall, arall and glare.
Fatigue Fract Eng M 40(7):1086–1100
5. Su Z, Ye L, Lu Y (2006) Guided Lamb waves for identification of damage in composite
structures: a review. J Sound Vib 295(3–5):753–780
6. Su Z, Ye L (2009) Algorithms for damage identification—fusion of signal features. In: Identi-
fication of damage using lamb waves—from fundamentals to applications, 1st edn. Springer,
Berlin, pp 195–254
7. Maghsoodi A et al (2016) Damage detection in multilayered fiber-metal laminates using guided-
wave phased array. J Mech Sci Technol 30(5):2113–2120
8. Marissen R (1983) arall (Aramid Fibre Reinforced Aluminum Laminate)—a new fatigue
resistant hybrid composite. Z Werkstofftech 14(8):278–283
9. Sadighi M, Tooski MY, Alderliesten RC (2017) An experimental study on the low velocity
impact resistance of fibre metal laminates under successive impacts with reduced energies.
Aerosp Sci Technol 67:54–61
10. Daniel IM, Ishai O (2006) Appendix A: material properties. In: Engineering mechanics of
composite Materials, 2nd edn. Oxford University Press, Oxford, pp 380–382
11. Moser F, Laurence LJ, Qum J (1999) Modeling elastic wave propagation in waveguides with
the finite element method. NDT E Int 32:225–234
Ultrasonic Guided Waves
in Unidirectional Fiber-Reinforced
Composite Plates
Ductho Le, Jaesun Lee, Younho Cho, Duy Kien Dao, Truong Giang Nguyen,
and Haidang Phan
Abstract The article is concerned with the motion of Lamb-type guided waves in
a unidirectional fiber-reinforced composite under ultrasonic sources. This kind of
composite is much stiffer along its axis than across its transverse direction and there-
fore shows orthotropic material properties. The dispersion relation of free guided
waves propagating in a unidirectional composite plate is derived that results in disper-
sion curves. Using reciprocity in elastodynamics, closed-form solutions of elastic
wave motions subjected to time-harmonic loads in a unidirectional fiber-reinforced
composite are in a simple manner computed. The obtained computations can be
very useful for developing ultrasound-based methods for nondestructive evaluation
of composite structures.
Keywords Guided waves · Fiber-reinforced composite plates · Dispersion

equations · Nondestructive evaluation · Reciprocity theorems
D. Le
Faculty of Mathematics, Mechanics and Informatics, University of Science, VNU, Hanoi, Vietnam
J. Lee
School of Mechanical Engineering, Changwon National University, Changwon, South Korea
Y. Cho
School of Mechanical Engineering, Pusan National University, Busan, South Korea
D. K. Dao
Faculty of Civil Engineering, Ho Chi Minh City of Technology and Education, Ho Chi Minh City,
Vietnam
T. G. Nguyen
Institute of Mechanics, Vietnam Academy of Science and Technology, Hanoi, Vietnam
H. Phan (B)
Institute of Theoretical and Applied Research (ITAR), Duy Tan University, Hanoi, Vietnam
e-mail: haidangphan.vn@gmail.com

https://doi.org/10.1007/978-981-15-9199-0_65
678 D. Le et al.
1 Introduction
Composite materials are being widely used in various components of engineering

structures such as automotive parts, civil infrastructures and especially aerospace
components. The aerospace usage of composites has experienced a continuously
growing over several decades. As an example, the Airbus A350 XWB contains up
to 83% of total composite materials by volume and 52% by weight [1]. Aerospace
composite structures are made by overlapping several unidirectional fiber-reinforced
composite layers with different angle orientations. These composite materials have
orthotropic material properties as they are much stiffer along the fiber direction than
across the transverse direction.
Guided waves can propagate in composite plates with little loss of energy and
inspect large areas from a single location; thus, they are critical to the applica-
tions of nondestructive evaluation (NDE) and structural health monitoring (SHM)
of aerospace structures. Understanding of wave interaction with composites is still,
however, far from complete due to the nature of multi-layer and anisotropic behav-
iors [2]. The study of free guided wave propagation in unidirectional composites
can be found, for example, in textbooks [2–4] and in numerous research articles,
see [5, 6] for example. The wave motion due to ultrasonic sources in the plates is
much more complicated but unquestionably one of the most fundamental problems
of elastodynamics. Conventionally, wave motion generated by a loading is solved
using integral transform technique [7]. However, it is more difficult to use this clas-
sical approach for anisotropic materials and impossible to apply the method to the
wave motion in inhomogeneous media. In order to avoid these complications, another
analytical approach based on reciprocity in elastodynamic, strictly to determine the
guided waves, has been proposed in recent years, see [8–13]. The applications of
elastodynamic reciprocity theorems in order to obtain favorite information about the
scattered field of surface waves generated by ultrasonic loads were studied in [12, 14,
15]. Verification of the reciprocity approach used in the current study was presented
in [10, 16] for the case of surface waves.
In elastodynamics, reciprocity theorems offer an integral relation between two
loading states of an elastic body [8]. The relation can be used to derive systems
of integral equations. In the current work, we aim to suggest direct applications
of the reciprocity theorems to compute guided wave fields in a unidirectional fiber-
reinforced composite plate subjected to a time-harmonic load for further development
of ultrasound-based NDE and SHM techniques. By appropriately choosing the virtual
state as a free guided wave, the scattered amplitudes of guided waves generated by
the load (the actual state) are obtained through the reciprocity relations.
Ultrasonic Guided Waves in Unidirectional Fiber-Reinforced … 679
2 Guided Waves in a Unidirectional Composite
A thin unidirectional lamina of thickness 2h in the Cartesian coordinate system

(x1 , x2 , x3 ) is illustrated in Fig. 1. Consider the propagation of guided waves in the
lamina along the x1 -direction with the phase velocity c and the wave number k. The
displacement components in the (x1 , x2 ) plane may be written in the forms of

u 1 = A1 eis1 kx2 + A2 eis2 kx2 + A3 e−is1 kx2 + A4 e−is2 kx2 eik(x1 −ct) (1)

u 2 = α1 A1 eis1 kx2 + α2 A2 eis2 kx2 − α1 A3 e−is1 kx2 − α2 A4 e−is2 kx2 eik(x1 −ct) (2)
where A1 , A2 , A3 , A4 are unknown constants, and t indicates time.

In Eqs. (1) and (2), we define
(C12 + C66 )s j C11 − ρc2 + C66 s 2j

αj = − = − , j = 1, 2 (3)
C66 − ρc2 + C22 s 2j (C12 + C66 )s j
where
√ √
S+ S 2 − 4P S− S 2 − 4P
s1 = , s2 = (4)
2 2
with

C11 − ρc2 C22 + C66 − ρc2 C66 − (C12 + C66 )2
S=− (5)
C22 C66

C11 − ρc2 C66 − ρc2
P= (6)
C22 C66
Fig. 1 Unidirectional composite lamina

680 D. Le et al.
In Eq. (3), ρ is the mass density, and C11 , C12 , C22 , C66 are four elastic constants.
It is noted that for axisymmetric or asymmetric deformations, in general, there are
five constants. For plane-strain deformations, only four constants are involved since
the governing equations are independent of the material constant C13 .
Based on the stress-strain relations for an orthotropic medium, stress components
are computed as

σ11 = ikc66 δ1 A1 eis1 kx2 + δ2 A2 eis2 kx2 + δ1 A3 e−is1 kx2 + δ2 A4 e−is2 kx2 (7)

σ12 = ikc66 β1 A1 eis1 kx2 + β2 A2 eis2 kx2 − β1 A3 e−is1 kx2 − β2 A4 e−is2 kx2 (8)

σ22 = ikc66 γ1 A1 eis1 kx2 + γ2 A2 eis2 kx2 + γ1 A3 e−is1 kx2 + γ2 A4 e−is2 kx2 (9)
where
C11 C12 C12 C22
δj = + αjsj, βj = αj + sj, γj = + α j s j , j = 1, 2 (10)
C66 C66 C66 C66
There are totally four boundary conditions, two on the top and the other two at
the bottom of the plate, expressed as
σ12 = 0; σ22 = 0 at x2 = −h (11)
σ12 = 0; σ22 = 0 at x2 = h (12)
Equations (11) and (12) result in
DA = 0 (13)
where
T
A= A1 A2 A3 A4 (14)
and
⎡ ⎤
β1 e−is1 kh β2 e−is2 kh −β1 eis1 kh −β2 eis2 kh
⎢ γ1 e−is1 kh γ2 e−is2 kh γ1 eis1 kh γ2 eis2 kh ⎥
D=⎢
⎣ β1 eis1 kh
⎥ (15)
β2 e is2 kh
−β1 e−is1 kh
−β2 e−is2 kh ⎦
γ1 eis1 kh γ2 eis2 kh γ1 e−is1 kh γ2 e−is2 kh
For nontrivial solutions, the determinant of D is required to be zero that results

in a complicated equation with an unknown phase velocity. This equation can only
be solved by numerical methods which result in dispersion curves. Similar to Lamb
wave in an isotropic medium, Lamb-type guided wave in a unidirectional composite

plate is dispersive.
For guided waves propagating in a laminated composite plate which includes
several unidirectional layers with different angle orientations, the characteristic equa-
tions may be found by either global matrix method (GMM), transfer matrix method
(TMM), or stiffness matrix method (SMM). Curious readers can refer to the text-
books by Giurgiutiu [1], Rose [2] and Nayfeh [17] for more details. For this study, a
computer program is made to obtain dispersion curves of phase and group velocities.
Since Lamb-type wave modes and shear horizontal (SH) wave modes are inseparable
in general anisotropic materials, our computer code is intended for solutions of all
guided wave modes. Therefore, SH wave modes will be appeared in the representa-
tion of results although they are not of interest in the current work. For the case study,
we use material properties of T300/914 CFRP which has a density ρ = 1560 kg/m3
and a stiffness matrix given in Eq. (16). Examples of phase and group velocity
solutions are shown in Fig. 2 (unidirectional fiber composite with 0° fiber), Fig. 3
Fig. 2 Dispersion curves of unidirectional fiber composite with 0° fiber, Lamb-type guided waves
(solid), shear horizontal waves (dash)
Fig. 3 Dispersion curves of unidirectional fiber composite with 45° fiber, Lamb-type guided waves
(solid), shear horizontal waves (dash)
682 D. Le et al.
Fig. 4 Dispersion curves of multi-layer laminated composite [0/45/90/135], Lamb-type guided

waves (solid), shear horizontal waves (dash)
(unidirectional fiber composite with 45° fiber) and Fig. 4 (four-layered laminated
composite [0/45/90/135]).
⎡ ⎤
143.8 6.2 6.2
⎢ 6.2 13.3 6.5 ⎥
⎢ ⎥
⎢ ⎥
⎢ 6.2 6.5 13.3 ⎥
C=⎢ ⎥ (16)
⎢ 3.6 ⎥
⎢ ⎥
⎣ 5.7 ⎦
5.7
3 Guided Wave Motions in a Unidirectional Composite

Plate by Reciprocity Considerations
Reciprocity theorem is a relation between displacement, traction components, and

forces of two different loading statukhh,es of a body that may be expressed in the
form of [8]

A B B A
f j u j − f jB u Aj d = σi j u j − σiAj u Bj n i d (17)

where describes a contour around domain , n i is normal vector and superscripts A,

and B indicates two elastodynamic states. Discussed in this section is the application
of reciprocity relations to obtain the solutions of guided waves in a unidirectional
composite plate under a time-harmonic source. This is called the actual state, say A.
The other one, say B, playing as the virtual state, is a free guided wave. By choosing
state B suitably, state A can be determined through the reciprocity theorems.
In order to conveniently perform the application of reciprocity theorems, the

expressions of displacement and stress components given in Eqs. (1–2) and (6–8),
respectively, are desired to be presented through only one unknown constant, say A,
instead of A1 , A2 , A3 , A4 . Note that Eq. (13) contains four unknowns while it has
only three independent equations due to the requirement of the determinant of matrix
D to be vanished. Thus, it has an infinite number of solutions. The expressions of
displacements and stresses may now be rewritten as

u 1 = A d1 eis1 kx2 + d2 eis2 kx2 + d3 e−is1 kx2 + d4 e−is2 kx2 eik(x1 −ct) (18)

u 2 = A α1 d1 eis1 kx2 + α2 d2 eis2 kx2 − α1 d3 e−is1 kx2 − α2 d4 e−is2 kx2 eik(x1 −ct) (19)

σ11 = ikC66 A δ1 d1 eis1 kx2 + δ2 d2 eis2 kx2 + δ1 d3 e−is1 kx2 + δ2 d4 e−is2 kx2 eik(x1 −ct)
(20)

σ12 = ikC66 A β1 d1 eis1 kx2 + β2 d2 eis2 kx2 − β1 d3 e−is1 kx2 − β2 d4 e−is2 kx2 eik(x1 −ct)
(21)
where d1 , d2 , d3 , d4 are a general solution of Eq. (13), see our forthcoming work for
detail [18].
Consider a time-harmonic vertical source at the origin of the coordinate system
that may be written in the form of
f 2A = Pδ(x1 )(x2 )e−iωt (22)
The load generates far-field Lamb-type guided wave modes in the composite plate
in the negative and positive directions. Their corresponding unknown amplitudes
AnP− (the negative) and AnP+ (the positive), where n indicates the mode and will be
computed. The elastic wave motion in far field of the plate is a summation over the
guided modes. In the positive direction, we may write
∞
∞

u1 = u n1 = AnP+ V1n (x2 )eikn (x1 −cn t) (23)
n=0 n=0
∞
∞

u2 = u n2 = AnP+ V2n (x2 )eikn (x1 −cn t) (24)
n=0 n=0
We first choose B to be a free guided wave of mode p in the negative direction.

Substitution the expressions of the states A and B into Eq. (17), after manipulation,
yields the amplitude of the scattered wave field as
p
P V2 (0)
p =
A P+ (25)
2C66 I pp
684 D. Le et al.
where
h
p p p p
I pp = k p T11 V1 − T12 V2 dx2 (26)
−h
The amplitude of scattered surface waves in the negative direction is obtained as

state B is chosen in the negative direction
p
P V2 (0)
p =
A P− (27)
2C66 I pp
Similarly, for a horizontal load of magnitude Q, we have

p p
−QV1 (0) QV1 (0)
p =
A Q+ , p =
A Q− (28)
2C66 I pp 2C66 I pp
4 Conclusions
It has been discussed in this article the guided wave motions in a unidirectional
fiber-reinforced composite plate subjected to time-harmonic sources. The dispersion
relation of Lamb-type wave guided waves has been derived, and the dispersion curves
have been found. We have then obtained the closed-form expressions of guided waves
in a unidirectional composite lamina under an ultrasonic load. The amplitudes of
guided waves in positive and negative x1 -direction due to the load have derived.
Acknowledgements This research has been supported by the Vietnam National Foundation for
Science and Technology Development (NAFOSTED) under Grant reference 107.02-2019.21.
References
1. Giurgiutiu V (2015) Structural health monitoring of aerospace composites. Elsevier Science,

Amsterdam
2. Rose JL (2014) Ultrasonic guided waves in solid media. Cambridge University Press,
Cambridge
3. Rokhlin S, Nagy SRDCP, Chimenti D, Nagy P (2011) Physical ultrasonics of composites.
Oxford University Press, Oxford
4. Datta SK (2009) Elastic waves in composite media and structures: with applications to
ultrasonic nondestructive evaluation. CRC Press, Boca Raton
5. Nayfeh AH, Chimenti DE (1988) Propagation of guided waves in fluid-coupled plates of
fiber-reinforced composite. J Acoust Soc Am 83:1736–1743
6. Habeger CC, Mann RW, Baum GA (1979) Ultrasonic plate waves in paper. Ultrasonics 17:57–
62
7. Achenbach JD (1973) Wave propagation in elastic solids. North-Holland Publishing Company,
Amsterdam
8. Achenbach JD (2003) Reciprocity in elastodynamics, Cambridge University Press, Cambridge
9. Phan H, Cho Y, Achenbach JD (2013) Application of the reciprocity theorem to scattering of
surface waves by a cavity. Int J Solids Struct 50:4080–4088
10. Phan H, Cho Y, Achenbach JD (2013) Validity of the reciprocity approach for determination
of surface wave motion. Ultrasonics 53:665–671
11. Balogun O, Achenbach JD (2013) Surface waves generated by a line load on a half-space with
depth-dependent properties. Wave Motion 50:1063–1072
12. Phan H, Cho Y, Le QH, Pham CV, Nguyen HTL, Nguyen PT, Bui TQ (2019) A closed-form
solution to propagation of guided waves in a layered half-space under a time-harmonic load:
an application of elastodynamic reciprocity. Ultrasonics 96:40–47
13. Phan H, Cho Y, Li W (2018) A theoretical approach to multiple scattering of surface waves by
shallow cavities in a half-space. Ultrasonics 88:16–25
14. Phan H, Cho Y, Pham CV, Nguyen H, Bui TQ (2019) A theoretical approach for guided waves
in layered structures. AIP Conf Proc 2102:050011
15. Phan H, Bui TQ, Nguyen HTL, Pham CV (2018) Computation of interface wave motions by
reciprocity considerations. Wave Motion 79:10–22
16. Phan H, Cho Y, Achenbach JD (2014) Verification of surface wave solutions obtained by the
reciprocity theorem. Ultrasonics 54:1891–1894
17. Nayfeh AH (1995) Wave propagation in layered anisotropic media: with applications to
composites. Elsevier, Amsterdam
18. Phan H (2019) A theoretical study of guided waves in an orthotropic plate. Unpublished results
Mode Selectivity and Frequency
Dependence of Guided Waves Generated
by Piezoelectric Wafer Transducers
in Rebars Embedded in Concrete
Rajeshwara Chary Sriramadasu, Ye Lu, and Sauvik Banerjee
Abstract Guided waves have the potential to recognise the corrosion damage in
reinforced concrete (RC) at its nascent stages. The latest approaches to generate
guided waves deploy contact acoustic transducers and piezoelectric wafer transducers
(PWTs). Experimental studies show that excitation of guided waves using contact
transducers causes multiple longitudinal guided wave modes in embedded rebars and
these modes can be used to assess various features of the damage. In contrast, the
excitation with piezoelectric transducers introduces only a dominant wave mode. In
this study, mode selectivity of guided waves in embedded rebars is investigated. Finite
element simulations are adopted to calculate the response of the RC beam to PWT
excitation and are verified with the experimental signal. The frequency dependence of
the dominant guided wave mode is studied by examining the displacement profile of
the RC beam. The displacement pattern is compared with the mode shape of various
wave modes that are calculated using theoretical dispersion curves to recognise the
possible wave modes. It is found that the lower frequencies cause F(1, 1) mode,
whereas the higher frequencies trigger higher-order flexural modes. The results of
this work assist in selecting the appropriate frequency to assess damages of debonding
or diameter reduction.
Keywords Guided waves · Dispersion · Piezoelectric transducers · Structural

health monitoring · Pitting corrosion · Reinforced concrete
R. C. Sriramadasu
IITB-Monash Research Academy, Mumbai, India
R. C. Sriramadasu · Y. Lu (B)
Department of Civil Engineering, Monash University, Clayton, VIC 3800, Australia
e-mail: ye.lu@monash.edu
R. C. Sriramadasu · S. Banerjee
Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai, India

https://doi.org/10.1007/978-981-15-9199-0_66
688 R. C. Sriramadasu et al.
1 Introduction
A major part of the rapidly expanding civil infrastructure is built by using reinforced
concrete (RC). The RC structures are cheaper to construct in comparison to metallic
structures, but their maintenance is expensive. The current methods of damage detec-
tion in RC like half-cell potential and visual inspection are limited in their capacity
to estimate the beginning of corrosion in embedded rebars [1–3]. Guided waves have
found their utility in damage prognosis of metallic and composite structures and are
increasingly deployed to study corrosion damage in RC [4–8]. Guided wave-based
techniques use high frequencies to estimate the location and extent of small defects
with good accuracy. The guided waves disperse and attenuate when they interact
with constituent materials of RC.
Guided waves are commonly generated and received with embeddable smart
aggregates, piezoelectric wafer transducers (PWTs), and contact transducers [9–
12]. Contact transducers can actuate and sense pure longitudinal modes. However,
it is not practical to use the contact transducers in RC structures as they require
rebar ends to be accessible and are expensive. Previous studies by the authors have
shown that guided waves excited with PWT result in a dominant flexural mode. The
investigations indicate that the surface bonded PWTs are cheaper and sensitive to
track several stages of corrosion, including the corrosion initiation [13, 14]. However,
the studies could not elaborate on the methods to distinguish the stages in case the
continuous data is missing. The sensitivity of a particular guided wave mode to
damage is determined by the energy profile of that mode. For example, the wave
modes with maximum energy in the core of rebar can detect the reduction in the
diameter and the wave modes with maximum energy on the surface of rebar can
detect the debonding type of damage [15]. A thorough understanding of the modes
is thus necessary to deploy the guided waves for corrosion detection.
PWTs can excite selective modes, and the excitability of a guided wave mode
depends on the PWT dimensions and resonant frequencies [16]. In this study, a
detailed numerical examination of the guided waves in the RC beam is carried out
to understand their propagation and utility in non-destructive evaluation. Dispersion
curves are generated for the embedded rebars using an approximate model that is an
excellent alternative to the actual model. The wave modes produced by the PWTs at
various frequencies are recognised and their potential to audit the corrosion damage
is discussed.
2 Numerical Modelling
A co-simulation approach is followed to generate and capture guided waves using

PWTs. The numerical models are created in the commercial finite element analysis
(FEA) software Abaqus/CAE. The mechanical response of the structural component
is simulated using the dynamic explicit code, and the piezoelectric behaviour is
Mode Selectivity and Frequency Dependence of Guided Waves … 689
Fig. 1 The co-simulation model of, a RC beam in explicit code, and b PWT in implicit code
simulated using the dynamic implicit process. The RC beam and PWT models are
shown in Fig. 1. The guided waves are generated when the stress from PWT is
transferred to the explicit model through their interaction. The stress transfer from
the rebar to PWT effectuate the build-up of charge in the piezoelectric element. The
difference in the electric potential of the top and bottom faces of the PWT is measured
to calculate the guided wave response of the structure.
The bare-rebar model is used to validate the current co-simulation approach.
The rebars used in the study are high yield strength deformable bars with Young’s
modulus of 205 GPa, Poisson’s ratio of 0.29 and density of 7850 kg/m3 . The length
and diameter of the rebar are 700 mm and 12 mm, respectively. The rebars are
modelled with circular cross-section without any ribs as their effect is negligible for
the current frequency of guided wave testing. PWTs, supplied by PI Ceramic, with
material conforming to PIC 151 are used in this study. The material properties of the
PWT are listed in Table 1.
A three-dimensional FEA model of the concrete beam is generated in
Abaqus/CAE. The RC beam is 500 mm-long, 100 mm-deep, and 100 mm-wide. The
embedded rebar has the same properties as that of the bare-rebar and is projected
by 100 mm on either side of the beam. The concrete used in the simulation has a
density of 2238 kg/m3 , Young’s modulus of 19 GPa, and Poisson’s ratio of 0.15. A
mesh size of 2 mm and a step interval of 0.1 µs are used. A five cycle Hanning pulse
at various central frequencies is input as the electric potential difference in the PWT
that is located at one end, and the response potential is captured at the PWT that is
located at the opposite end of the rebar.
The response signal of the bare-rebar to the Hanning pulse that is excited at a
frequency of 60 kHz is shown in Fig. 2. The guided wave modes are identified
Table 1 Material properties

Coefficient Value Unit
of the PWT used to actuate
and sense guided waves in the Density 7760 kg/m3
RC beam Poisson ratio 0.34 –
S11 1.683 × 10−11 m2 /N
S33 1.900 × 10−11 m2 /N
S55 5.096 × 10−11 m2 /N
S12 −5.656 × 10−11 m2 /N
S13 −7.107 × 10−11 m2 /N
S44 5.096 × 10−11 m2 /N
S66 4.497 × 10−11 m2 /N
d31 −2.140 × 10−10 m/V
d33 4.230 × 10−10 m/V
d15 6.100 × 10−10 m/V
Fig. 2 The response of a

bare-rebar to excitation pulse
with frequency 60 kHz
using the dispersion curves for the 12 mm rebar. The two modes observed in the
signal are L(0, 1) and F(1, 1) modes with group velocities 4640 m/s and 3014 m/s,
respectively. The signals illustrate that a dominant flexural mode is excited when a
PWT is bonded to the surface of rebar.
The guided wave response of an RC beam is computed by using the model shown
in Fig. 1 for various frequencies. The response of the beam excited by 20 mm-long,
5 mm-wide, and 1 mm-thick PWT at the frequency of 60 kHz is shown in Fig. 3.
The FEA signal matches well with the experiments and signals show a dominant
wave mode. The numerical calculations are repeated for the excitation frequencies
45 and 75 kHz. The response signals for these frequencies also show a dominant
mode (Fig. 4). Dispersion curves are generated to recognise these modes. These
curves are calculated using the multiple layer matrix of rebar and concrete in the
Matlab code PC Disp [17]. The cross-section of the concrete part in the RC beam
is assumed as circular instead of rectangular for the simplicity in calculations. The
dispersion curves of the RC beam are shown in Fig. 5.
Fig. 3 The guided wave

response of RC beam excited
with a frequency of 60 kHz
Fig. 4 Dominant wave

modes triggered by a PWT at
various frequencies
Fig. 5 Dispersion curves of

a 12 mm rebar embedded in
100 mm-diameter concrete
cylinder
Continuous wavelet transform is used to estimate the times of flight (ToF) of

the dominant mode of the signals that are shown in Fig. 5. The ToF is computed
as 315.5 µs, 308.4 µs, and 313.0 µs for the input frequencies of 45 kHz, 60 kHz,
and 75 kHz, respectively. Several modes have the same group velocity; thus, it is not
possible to identify modes based on the group velocity alone. The displacement values
obtained from the FEA are transformed from the Cartesian coordinates to cylindrical
coordinates and are correlated with the theoretical dispersion calculations. The centre
of the rebar is fixed as the origin of the coordinate system. The z-axis is along the
Fig. 6 Displacement of the beam noted at the time of arrival of the dominant mode for the excitation
frequencies, a 45 kHz, b 60 kHz, and c 75 kHz
length of rebar, r-axis is away from the centre, and θ-axis is in the circumferential
direction. The FEA displacement of the RC beam near the receiver end is captured
at the time of arrival of dominant mode and is shown for various frequencies of
excitation in Fig. 6.
The total displacement values for the input frequency of 45 kHz are shown in
Fig. 6a. The changes in radial and circumferential displacements are insignificant
for this case. The dominance of green at the centre and blue at the outer edges in
Fig. 6a reveals that the displacement is high at the centre and low at the outer edges.
The mode excited by 45 kHz is recognised as F(1, 1) as the radial displacement is
small. The F(1,1) mode has a substantial axial displacement apart from the radial
component; thus, this mode can also indicate the changes in the rebar diameter.
Likewise, the displacement at the receiving end for the input frequencies of 60
and 75 kHz are plotted in Fig. 6b, c, respectively. The variations in radial and circum-
ferential displacements are more pronounced for 60 and 75 kHz cases that include
alternate cusps in the displacement profile. The values of displacements are higher at
the outer edges and lower at the centre of the RC beam. The contribution of radial and
circumferential displacement increases as the frequencies increases. These observa-
tions indicate the possibility of higher-order flexural mode which are sensitive to
debonding type of damage for the higher input frequency. Further investigations are
required to ascertain the exact mode which depends on the PWT dimensions and
properties.
4 Conclusions
The co-simulation model deployed for guided wave triggering and capturing can
accurately simulate the actual experimental conditions. The results of the FEA show
a good correlation with those of the experiments. The dominant guided wave mode
generated by PWTs in the bare-rebars is recognised as F(1, 1) mode. Dispersion
curves show that multiple wave modes are possible for rebars embedded in concrete.
The input frequency of 45 kHz demonstrates that F(1, 1) mode is the dominant
mode. The input frequencies of 60 and 75 kHz show the possibility of higher-order
flexural mode as the dominant mode. PWTs show a strong selectivity of mode that
is reliant on its dimensions and excitation frequency. The future studies will focus
on the selection of PWT dimensions and frequency to detect the type of damage in
various rebars.
References
1. Song HW, Saraswathy V (2007) Corrosion monitoring of reinforced concrete structures—a

review. Int J Electrochem Sci 2:1–28
2. Otieno M, Beushausen H, Alexander M (2011) Prediction of corrosion rate in RC structures—a
critical review. RILEM Bookseries
3. Regier R, Hoult NA (2015) Concrete deterioration detection using distributed sensors. Proc
Inst Civ Eng Struct Build 168(2):118–126
4. Na WB, Kundu T (2003) Inspection of interfaces between corroded steel bars and concrete
using the combination of a piezoelectric zirconate-titanate transducer and an electromagnetic
acoustic transducer. Exp Mech 43(1):24–31
5. Dongsheng L, Tao R, Junhui Y (2012) Inspection of reinforced concrete interface delami-
nation using ultrasonic guided wave non-destructive test technique. Sci China Technol Sci
55(10):2893–2901
6. Ervin BL et al (2006) Estimation of general corrosion damage to steel reinforced mortar
using frequency sweeps of guided mechanical waves. Insight Non-Destr Test Cond Monit
48(11):682–692
7. Li J, Lu Y, Guan R, Qu W (2017) Guided waves for debonding identification in CFRP-reinforced
concrete beams. Constr Build Mater 131:388–399
8. Wang Y, Guan R, Lu Y (2017) Nonlinear Lamb waves for fatigue damage identification in
FRP-reinforced steel plates. Ultrasonics 80:87–95
9. Miller T, Hauser CJ, Kundu T (2002) Nondestructive inspection of corrosion and delamination
at the concrete-steel reinforcement interface. Nondestruct Eval 2002:121–128
10. Miller TH et al (2013) A new guided wave-based technique for corrosion monitoring in
reinforced concrete. Struct Health Monit 12(1):35–47
11. Sharma A et al (2018) Investigation of deterioration in corroding reinforced concrete beams
using active and passive techniques. Constr Build Mater 161:555–569
12. Sriramadasu RC, Banerjee S, Lu Y (2019) Detection and assessment of pitting corrosion in
rebars using scattering of ultrasonic guided waves. NDT E Int 101:53–61
13. Sriramadasu RC, Lu Y, Banerjee S (2019) Identification of incipient pitting corrosion in rein-
forced concrete structures using guided waves and piezoelectric wafer transducers. Struct
Health Monit 18(1):164–171
14. Rajeshwara CS, Banerjee S, Lu Y (2017) Identification of zero effect state in corroded RCC
structures using guided waves and embedded piezoelectric wafer transducers (PWT). Procedia
Eng 188:209–216
15. Sharma S, Mukherjee A (2010) Longitudinal guided waves for monitoring chloride corrosion
in reinforcing bars in concrete. Struct Health Monit 9(6):555–567
16. Nowak LJ, Zieliński TG (2013) Modal sensitivity and selectivity of small, rectangle-shaped
piezoelectric transducers used as sensors and actuators in active vibroacoustic control systems.
J Low Freq Noise Vib Act Control 32(4):253–271
17. Seco F, Jimenez AR (2012) Modelling the generation and propagation of ultrasonic signals in
cylindrical waveguides. In: Ultrasonic waves, chap 1. Intech Open Access Publisher, London,
pp 1–28
How Temperature Variations Affect
the Propagation Behaviour of Guided
Ultrasonic Waves in Different Materials
M. Rennoch, J. Moll, M. Koerdt, A. S. Herrmann, T. Wandowski,
and W. M. Ostachowicz
Abstract As guided ultrasonic waves are sensitive to changes in the acoustic

impedance of the structure, knowledge of the waves behaviour with respect to
different materials and varying temperatures is crucial for reliable damage detection.
Hence, we present how specific wave characteristics, such as velocity and amplitude,
behave for typical materials at various frequencies and temperatures, and how this
knowledge can be integrated into a damage detection technique. We also provide an
analysis of the thickness of the adhesive layer since we have experimentally observed
a strong temperature-related effect on guided wave signals.
Keywords Guided ultrasonic waves · Structural health monitoring · Temperature

variations · Composite structures
1 Introduction
While techniques of condition monitoring to predict the potential breakdown of a

machine circulate in industry and academia for many decades [1], research to assess
structural integrity came up in the 1970s [2]. The main goal in such a structural health
M. Rennoch (B) · M. Koerdt

Measurement Systems & Monitoring, Faserinstitut Bremen e. V., Am Biologischen Garten 2,
Bremen, Germany
e-mail: rennoch@faserinstitut.de
J. Moll
Department of Physics, Goethe University of Frankfurt, Frankfurt, Germany
A. S. Herrmann
Faculty Production Engineering—Mechanical Engineering & Process Engineering, Materials
Engineering/Fibres and Fibre Composites, University of Bremen, Bremen, Germany
T. Wandowski · W. M. Ostachowicz
Institute of Fluid-Flow Machinery, Polish Academy of Science, Warsaw, Poland

https://doi.org/10.1007/978-981-15-9199-0_67
monitoring (shm) system is detecting defects in a highly automatic manner. Many

systems available today base on guided ultrasonic waves (guw) as they travel along
the boundaries of the structure over relatively long distances and interact sensitivity
with structural defects [3]. Accordingly, there are many damage detection techniques
either available [4] or still in development [5].
Given by the complex multimodal and dispersive nature of guw, it is not straight-
forward to transfer existing analysis tools to different material systems [6]. Hence,
knowledge of the waves behaviour with respect to the material in which they prop-
agate becomes crucial. Especially, variations caused by the temperature effect have
to be properly addressed for practical applications [7]. Because of that, in this work,
we describe a study that investigates the propagation behaviour of guw in plate
structures made from different materials, i.e. aluminium (al), glass fibre-reinforced
plastic (gfrp), carbon fibre-reinforced plastic (cfrp) and a hybrid material system,
called glass laminate aluminium-reinforced epoxy (glare). Therewith, we provide
a representative synopsis on the most commonly used materials. We experimen-
tally show how specific wave characteristics, such as velocity and amplitude, behave
for typical materials at various frequencies and temperatures. This holds true for
the symmetric and antisymmetric wave modes. Finally, as this is essential for the
development of damage detection techniques [9], we will provide first ideas of how
damage detection algorithms benefit from incorporating such knowledge.
guw-based inspection has most applications in one-dimensional structures such as

metallic pipes or metallic rails with isotropic material properties. Far more difficult to
analyse are two-dimensional fibre-reinforced polymer composites waveguides due
to the anisotropic properties of the laminate. The latter is a good example for the
usefulness of guw techniques, as inner defects are seldom visible from the outside.
Hence, for this study, we chose four plates as depicted in Fig. 1 all having the
dimensions of 500 × 500 × 4 mm.
Both fibre-reinforced specimens are made from non-crimp fabrics in a vacuum
infusion process. According to impact robust glare-5, the fibre metal laminate (fml)
is made from five layers of primed aluminium 2024 T3 and eight plys of aerospace
qualified bidirectional glass fibre prepreg EHG250-68-37 T2 100 with a hot press
vacuum bag process. All edges were trimmed, and holes were inserted in the upper
corners of one side for the further hanging storage.
At the surface of each specimen, we adhesively bonded three circular PIC255
transducers with a diameter of 10 mm and thickness of 0.2 mm. The spacing between
the transducers is 7.5 cm each. Transducers T1 and T2 , as shown in Fig. 1, are
placed, such that the centre of the plate surface lies halfway in between. Afterwards,
we measure the thickness of the adhesive layer by laser triangulation, see Fig. 2, to
provide plausible causes in the case that the received signal strength behaves notably.
How Temperature Variations Affect the Propagation … 697
Fig. 1 Four plates in a climate chamber (left) under the same temperature conditions. A temperature
probe is placed on the top left of each specimen. Illustration of the transducer positions (right) where
all actuator–sensor combinations are recorded in a round-robin manner
We then connect all transducers by shielded coaxial cables to a multiplexing

unit that is able to measure all actuator– sensor combinations [10]. A Handyscope
HS5 (TiePie Engineering) is used to generate arbitrary waveforms and to record
the signals by analogue-to-digital conversion (14-bit resolution). Further, we mount
a temperature sensor on each plate and place all samples at once in a climate test
chamber. We record pointwise the propagating guw excited by a Hanning windowed
sine wave of five cycles with a peak-to-peak voltage of 200 V and carrier frequencies
ranging from 40 to 260 kHz in steps of 20 kHz. Additionally, temperature varies
from 20 to 60 °C in steps of 0.5 °C, while the relative humidity is constant at 50%.
3 Data Analysis
Figure 3 illustrates the temperature dependency for the different carrier frequencies
for the four plates in this study. This visualisation includes measurement signals for
every temperature level ranging from 20 to 60 °C in steps of 0.5 °C. As expected,
the waves attenuation in the aluminium plate is relatively weak compared to the
composite plates. Moreover, the dispersion effect can be clearly seen, especially for
the antisymmetric wave mode. In every specimen, the mode tuning effect described
in [11] and [12] can be identified, which stands for the excitability of guw modes as
a function of frequency. Here, we observe a novel, superimposed effect that shows
that mode tuning behaviour is modified by temperature variations.
A more detailed analysis of this data is depicted in Fig. 4 that shows selected
results for the cfrp and glare specimens, respectively. Two frequencies have been
Fig. 2 Height profiles from laser triangulation of adhesively bonded 0.2 mm thick pwas. From top
to bottom, the sensors are bonded on al, cfrp, gfrp and glare. From left to right, the sensors are
at the locations T1 , T2 and T3 . The y-axis is labelled in scan points, which by factor 0.05 roughly
correspond to mm (heights above 1.3 mm correspond to soldered wires and thus are blanc. Gaps
inside the pwas, given by blue dots, are due to measurement errors)
Fig. 3 Dependence on temperature as function of frequency. From left to right, top to bottom: al,
cfrp, gfrp and glare
selected, namely 140 kHz and 260 kHz. A semi-automated procedure has been imple-
mented based on the waterfall representation in Fig. 4. Therewith, it is possible to
extract the maximum amplitudes of the symmetric and antisymmetric wave modes,
respectively. The amplitudes can be plotted as a function of the number of measure-
ments (proportional to time) and also temperature. It can be observed that the wave
mode amplitudes vary slightly for the case of the cfrp sample, but strongly for the
glare sample. This effect is particularly strong for the antisymmetric wave mode
in the case of 140 kHz leading to an amplitude reduction of about 85%. This result
is important, because it means that through the temperature effect, the propagation
distance is strongly reduced. Similar results can be observed for the symmetric and
antisymmetric wave mode at 260 kHz.
The results presented above are systematically investigated for several guw
frequencies as shown in Fig. 5 for the glare sample. The symmetric and the antisym-
metric wave mode have a significant temperature-modulated behaviour in addition
to the conventional mode tuning effect. In the present case, the effect related to
temperature is even stronger and must be taken into account in the design of a shm
system. Finally, an exemplary damage detection result for the cfrp plate at 60 kHz is
presented in Fig. 6 using a similar approach as described in [13], combining optimal
baseline selection (obs) and baseline signal stretch (bss), hence, providing a damage
Fig. 4 Dependence on temperature with respect to frequencies 140 kHz (top) and 260 kHz (bottom)
for cfrp (left) and glare (right)
detection technique that takes the temperature effect into account. Therefore, signals
are recorded for the different states of the specimen—undamaged as well as damage
emulating via reversible pseudo defects (cf. [14]) and both exposed to the same
temperature cycle. Hence, baseline subtraction reveals remaining amplitudes in the
signal recorded for the damage state. A damage indicator relying on the root mean
square (rms) shows that an increased number of measurements provide a sustainable
damage detection.
4 Conclusion
We showed that specific wave characteristics, such as velocity and amplitude, change
considerably at various temperatures and quite depend on the material of the waveg-
uide. Here, aluminium (al), glass fibre-reinforced plastic (gfrp), carbon fibre-
reinforced plastic (cfrp) and glass laminate aluminium-reinforced epoxy (glare)
Fig. 5 Analysis of dependence on temperature in detail for glare
were tested. Thereof, gfrp and glare showed to be notably sensitive for temperature
variations leading to strong changes in amplitude. Further, an example of a damage
detection technique is given that compensates the temperature effect.
For further research, we think it is worth to investigate the temperature effect on
the adhesive layer, as we have observed a strong temperature-related effect on the
wave signals. This may correlate to fluctuations in thickness.
Fig. 6 guw signals (top left) at 60 kHz recorded from the undamaged and damaged structure.
Temperature profile (top right) in the climate chamber. Differential signal representation (bottom
left). Normalised damage indicator (bottom right) defined as the rms of the differential signal
Acknowledgements M. Rennoch and M. Koerdt gratefully acknowledge the financial support of

this research by the Federal Ministry for Economic Affairs and Energy on the basis of a decision by
the German Bundestag under the grant numbers 49VF180016 and 20W1703B. J. Moll acknowledges
support from the Federal Ministry for Economic Affairs and Energy (Grant 03SX422B).
References
1. Mitchell JS (2007) From vibration measurements to condition based maintenance: seventy

years of continuous progress. J Sound Vib 40:62–75
2. Cawley P (2018) Structural health monitoring: closing the gap between research and industrial
deployment. Struct Health Monit 17(5):1225–1244
3. Su Z, Ye L, Lu Y (2006) Guided lamb waves for identification of damage in composite
structures: a review. J Sound Vib 295(3–5):753–780
4. Su Z, Ye L (2009) Algorithms for damage identification—fusion of signal features. In: Iden-
tification of damage using lamb waves—from fundamentals to applications, 1st ed. Springer,
Berlin, pp. 195–254
5. Rose JL (2011) The upcoming revolution in ultrasonic guided waves. Proc of SPIE 7983,
798302
6. Moll J et al (2010) Multi-site damage localization in anisotropic plate-like structures using an

active guided wave structural health monitoring system. Smart Mater Struct 19(4):045022
7. Croxford AJ et al (2007) Strategies for guided-wave structural health monitoring. Proc R Soc
A 463(2087):2961–2981
8. Attarian VA, Cegla FB, Cawley P (2014) Long-term stability of guided wave structural health
monitoring using distributed adhesively bonded piezoelectric transducers. Struct Health Monit
13(3):265–280
9. Cross EJ et al (2012) Features for damage detection with intensity to environmental and
operational variations. Proc R Soc A 468(2148):4098–4122
10. Neuschwander K et al (2019) Simultaneous load and structural monitoring of a carbon fiber
rudder stock: experimental results from a quasi-static tensile test. J Intell Material Syst Struct
30(2):272–282
11. Giurgiutiu V (2005) Tuned lamb wave excitation and detection with piezoelectric wafer active
sensors for structural health monitoring. J Intell Material Syst Struct 16(4):291–305
12. Moll J, Fritzen C-P (2010) Advanced aspects of mode-selective excitation of ultrasonic guided
waves. In: 24th international conference on noise and vibration engineering. Leuven, Belgium,
pp. 969–984
13. Croxford AJ et al (2010) Efficient temperature compensation strategies for guided wave
structural health monitoring. Ultrasonics 50(4–5):517–528
14. Moll J et al (2018) Open guided waves—online platform for ultrasonic guided wave
measurements. Struct Health Monit 12
Nonlinearity Study of Damaged Spring
Materials Using Guided Wave
Jeongnam Kim, Yonghee Lee, Sungun Lee, and Younho Cho
Abstract Nonlinear ultrasonic techniques that can detect micro crack or defects
before use of ultrasonic nonlinearity are emerging as a way to predict the life of
materials. Guided wave propagates according to the shape of the material, thus
providing great potential for non-destructive evaluation applications. However, in
the non-destructive evaluation method using nonlinear guided waves, it is difficult
to get necessary information due to the dispersion and multi-mode characteristics
of guided wave. In this study, the nonlinearity of spring materials and the degree of
fatigue was investigated using nonlinear guided wave. The specimens were prepared
with different degree of fatigue through Nakamura fatigue test. The guided wave
has generated a signal at the point that 460 mm far from the left end of the spring
specimen and was received at a distance of 50 mm from the transmitted point. In
addition, the experiment carried out for six sections at intervals of 10 mm. The signal
was generated using a tone buster (RPR-4000), and the vicinity of the point, where
the fatigue load was largest (the middle point of the specimen), was investigated.
This study presents a quantitative evaluation method for the health of spring mate-
rials using nonlinear guided wave. Histological examinations, such as SEM imaging,
can provide more accurate and meaningful results, and further, studies such as these
must identify the cause.
Keywords Nonlinear guided wave · Spring material · Nakamura fatigue test
J. Kim · Y. Lee · Y. Cho (B)

School of Mechanical Engineering, Pusan National University, Busan 46241, Republic of Korea
e-mail: mechchy@pusan.ac.kr
S. Lee
Institute of Nuclear Safety and Protection, Pusan National University, Busan 46241, Republic of
Korea

https://doi.org/10.1007/978-981-15-9199-0_68
706 J. Kim et al.
1 Introduction
The conventional ultrasonic inspection techniques show excellent detection capa-

bility for the detection of defects over a certain size of ultrasonic waves inside the
material. However, it has the limit in the detection for micro-damage and early-stage
damage.
As an alternative, the study of nonlinear ultrasonic techniques, which can detect
conditions before microcracks or defects are generated using ultrasonic nonlinearity,
has emerged as an alternative to the life prediction technique of elastic materials.
Nonlinear ultrasonic techniques utilize the properties of secondary harmonics gener-
ated by materials to detect micro-damage, such as corrosion, degradation and fatigue
damage, rather than simple defects. In addition, it has the advantage of diagnosing
microscopic defects or changes in microstructures within a short time and research
on this continues [1, 2]. In particular, nonlinear guided ultrasound has the advantage
of long range detection, and higher sensitivity can be expected through various mode
selection [3].
The study of nonlinear guided ultrasonic techniques has long ongoing. A study was
conducted comparing the trends of fatigue and nonlinearity in materials, and Jacobs
and Jianmin studied nonlinear guided ultrasound using nonlinearity for nickel alloy
steels using Rayleigh surface wave [3, 4]. Gazis performed basic guided ultrasonic
analysis of cylindrical structures [5]. Deng used nonlinear ultrasonic evaluation in that
only the cumulative second-harmonic wave, which is a component that grows linearly
with the increase of the propagation distance, is generated [6, 7]. In addition, he
contributed to the study of phase matching, which is a method of selecting the induced
ultrasound mode that can generate the second-harmonic expression more actively in
the nonlinear ultrasound technique, and Weibin Li conducted an experimental study
[8].
In this study, the second-harmonic wave generation of the longitudinal wave was
investigated in specimen and performed that correlation between the fatigued spring
material and the nonlinearity using the nonlinear guided wave test. A non-destructive
method for quantitative evaluation of material integrity of specimens was provided.
2 Theory
2.1 General Solution of Nonlinear Wave Equation

in the Cylindrical Rod
Elastic fracture with arbitrary cross-sectional area has been studied theoretically,
and de Lima et al. report the generation of ultrasonic wave propagation harmonics in
cylindrical rod. Indeed, in a tube-type structure, induced wave propagation provides a
more efficient approach to detect material nonlinearity by reducing energy diffusion.
[9].
Nonlinearity Study of Damaged Spring Materials Using Guided Wave 707
The nonlinear wave equation of the Lagrange coordinates is expressed as follows:
∂ 2u
(λ + 2μ)∇(∇ · u) − μ∇ × (∇ × u) + f = ρ0 (1)
∂t 2
When a secondary order solution is configured using model extensions, the total
secondary harmonic fields can be written as [11]
∞
1
u2 = Am (z)u (2,m) (r )e−i2ωt
2 m=1
∞
1
v2 = Am (z)v(2,m) (r )e−i2ωt
2 m=1
∞
1
s2 = Am (z)s(2,m) (r )e−i2ωt (2)
2 m=1
The primary wave field of particle displacement component is assumed as follows

[10]:
u r1 = u r (r ) cos(nθ )ei(kz−ωt)
u θ1 = u θ (r ) sin(nθ )ei(kz−ωt)
u 1z = u z (r ) cos(nθ )ei(kz−ωt) (3)
where the circumferential order is n = 0, 1, 2, …. In the case of axi-symmetric

modes, n = 0 and n ≥ 1 for the flexural modes. In general, terms (u r1 , u θ1 and u 1z )
vary depending on the type of guided wave and family of modes.
Indeed, second-harmonic wave generation with cumulative effect is of great
interest because the cumulative effect of the second-harmonic wave amplitude is
of great interest to be detected in experimental work to measure the nonlinearity at
sufficient signal-to-noise ratios. From the above analysis, certain conditions shall be
met for secondary harmonics with cumulative effects [12].
➀ Nonzero power flux from the primary to the second-harmonic wave. f nsurf +
f nvol = 0
➁ Phase and group velocities matching require equal phase and group velocities of
the primary and the second-harmonic wave. When kn∗ = 2k (synchronism).
When the wave mode satisfy above two conditions, the second-harmonic wave
generation of this guided wave mode will have cumulative effect with propagation
distance increase.
708 J. Kim et al.
3 Experiment Setup and Experiment Result
3.1 Experiment Setup
To investigate the behavior of nonlinear induced ultrasonography, the experiment

was conducted as follows. Figure 1 is a schematic of the experimental setup, and
Table 1 is the sensor information of the nonlinear experiments.
The signal was generated at 460 mm from the end of the material and received
the signal at a distance of 60 mm with the interval of 20 mm and examined for 7
sections including the center of the specimen. The total length of detection length
is 180 mm, and the nonlinear correlation to investigate the performance of up to
fatigue, according to the fatigue, stress is measured a signal near the area of research
that occurred.
A specimen was constructed to investigate the relationship between fatigue degree
and nonlinearity. Figure 2 is schematic of Nakamura Fatigue Test specimens and it
used to make fatigue step. The number of revolutions per minute was equal to that of
all specimens at 5000 rpm and the number of fatigue was 0, 6, 10, 13 million times
for each specimen.
In this study, 0, 6, 10, 13 million times of fatigue is estimated to fatigue level as
0, 60, 100, 130%, respectively.
The other specimen is shot peening specimen. Shot peening is the most econom-
ical and effective way to increase the product life of treated metal parts by creating
Fig. 1 Schematic of second-harmonic wave testing method
Table 1 Transducer information of the nonlinear experiments

Calculated central frequency Sensor selection for experiments
Incident frequency 1.4 MHz 1 MHz
Angle of wedge 25° 25°
Receiver frequency 2.8 MHz 2.25 MHz
Fig. 2 Schematic of second-harmonic wave testing method. Reference: Robson Bussoloti [13]
and creating surface residual compression stress. Shot peening is also used for curing
work to improve wear characteristics, correct distortion, texture surface, and generate
aerodynamic curvature for aerospace design. Improved strength of treated parts
enables lighter parts with high wear resistance and fatigue resistance.
3.2 Experiment Result
The Fig. 3 is a graph comparing of each fatigue damaged specimen on the same
scale. Comparison of the graph above, with increasing fatigue more than a higher
rate of increase in nonlinearity was confirmed that, constant fatigue, the primacy of
fatigue loads. Nonlinear is reduced to near the point.
Figure 4 is the result of the comparing of the normal and shot-peened specimen.
As a result, the nonlinearity of shot-pinned specimens was measured higher than that
of normal specimens.
Fig. 3 Nonlinear trend of 0,

60, 100 and 130% specimen
710 J. Kim et al.
Fig. 4 Nonlinear trend of

normal and shot-peened
specimen
4 Conclusion
In this study, an experimental study was carried out on the relation between nonlin-
earity and plastic deformation of wire specimens and fatigue—induced early damage.
Nonlinearity tests were conducted on samples with different fatigue levels through
Nakamura fatigue test to investigate the correlation between fatigue and nonlin-
earity. Nonlinearity is increased as the degree of fatigue increased. On the other
hand, the nonlinear tendency of specimens with fatigue levels of 130% is shown to
decrease around 120 mm, the largest fatigue load. Nonlinearity of specimens is seen
as a tendency to decrease deformation of specimens due to amplitude attenuation or
scattering patterns due to changes in the tissue.
Moreover, the comparison of normal and short-finned specimens showed higher
nonlinearity of the short-finned specimens, which is shown to increase in nonlinearity
due to the surface residual stresses accompanying the shot peening.
In both types of specimens, it is difficult to see damage with the naked eye, but
the damage can be observed using ultrasonic nonlinear signals.
In this study, non-destructive methods were provided for quantitative evaluation
of the integrity of linear materials and initial damage using nonlinear ultrasound
techniques. It is expected that more accurate and meaningful results can be obtained
through historical examinations such as SEM imaging in the future, and further,
studies should be made to identify the cause.
Acknowledgements This work was supported by National Research Foundation of Korea(NRF)

grant funded by Korea government(MSIT) (No.2016M2A2A9A03913295).
References
1. Bermes C, Kim JY, Qu J, Jacobs LJ (2007) Experimental characterization of material

nonlinearity using Lamb waves. Appl Phys Lett 90(1):021901
2. Nagy PB (1998) Fatigue damage assessment by nonlinear ultrasonic material characterization.
Ultrasonics 36:275–381
3. Kim JY, Qu J, Jacobs LJ (2006) Acoustic nonlinearity parameter due to micro plasticity. J
Nondestr Eval 25(1):29–37
4. Buck O (1976) Harmonic generation for measurement of internal stress as produced by
dislocation. IEEE Trans Sonics Ultrasonics SU-23:346–350
5. Gazis DC (1959) Three dimensional investigation of the propagation of waves in hollow circular
cylinders. 1. Analytical foundation. J Acous Soc Am 31(5):568–573
6. Deng M (1998) Cumulative second-harmonic generation accompanying nonlinear shear
horizontal mode propagation in a solid plate. J Appl Phys 84:3500
7. Deng M, Wang P, Lv X (2005) Experimental observation of cumulative second-harmonic
generation of Lamb-wave propagation in an elastic plate. J Phys 38:344–353
8. Li W, Cho Y, Achenbach JD (2012) Detection of thermal fatigue in composites by second
harmonic Lamb waves. Smart Mater Struct 21(8):085019
9. Li W, Deng M, Xiang Y (2017) Review on second-harmonic generation of ultrasonic guided
waves in solid media (I): theoretical analyses. Chin Phys B 26(11):114302
10. de Lima WJN, Hamilton MF (2005) Finite amplitude waves in isotropic elastic waveguides
with arbitrary constant cross-sectional area. Wave Motion 41(1):1–11
11. Auld BA (1990) Acoustic fields and waves in solids, vol. 1, 2, Wiley
12. Li W, Deng M, Cho Y (2016) Cumulative second harmonic generation of ultrasonic guided
waves propagation in tube-like structure. J Comput Acous 24(3):1650011
13. Bussoloti R (2015) Delta ferrite: cracking of steel fasteners
Transducer Arrangement
for Baseline-Free Damage Detection
Methods With Guided Waves
U. Lieske, T. Gaul, and L. Schubert
Abstract Structural Health Monitoring (SHM) refers to methods for the early detec-
tion of damage in components in order to initiate planned repair. Acoustic methods
using guided waves are widely established for the monitoring of large-area, thin-
walled structures. Current SHM approaches based on active ultrasonic techniques
use the knowledge of a known reference condition of a considered transmit–receive
path, with which all subsequent conditions are compared, in order to obtain a state-
ment about the local component condition. In most cases, these methods evaluate
amplitude or phase changes with respect to the reference signal. The dependency of
the signal evaluations on environmental influences such as temperature or humidity
changes makes it necessary to explicitly consider these influences through statistical
analyses and background knowledge on the material side or to provide different refer-
ence states known as baselines for different environmental conditions. The subject
of this study is the development of baseline-free methods. The focus is on the iden-
tification of the damage interaction of the guided waves and the mode identification
by suitable transducer arrangements for the determination of structural damage on
reciprocal sensor transmission paths. For this purpose, parameterized simulation
experiments were performed to investigate different arrangements of piezoelectric
transducers for wave mode identification and to validate their results in laboratory
measurements. It could be shown that stacked transducers can be used for wave-
mode identification. Further more an additional effect of phase shifts in between two
wavemodes was observed using a piezoelectric ring transducer with two electrode
pairs.
Keywords Baseline-free · Piezoelectric transducer · SHM · Guided waves
U. Lieske (B) · T. Gaul · L. Schubert

Condition Monitoring, Fraunhofer IKTS, Maria-Reiche-Str. 2, 01109 Dresden, Germany
e-mail: uwe.lieske@ikts.fraunhofer.de

https://doi.org/10.1007/978-981-15-9199-0_69
714 U. Lieske et al.
Fig. 1 Comparison of the temperature influence and damage interaction on guided waves [4]
1 Introduction
In recent years, guided waves have established in research and development for the
assessment of structural damage to thin-walled structures [1, 2] and are increasingly
being used for industrial solutions. The most frequently used methods for excitation
of guided waves are electromagnetic methods, piezoelectric ultrasonic transducers, or
magnetostrictive methods. The underlying methods are mostly used for inspection
solutions and frequently apply shear waves in the pulse-echo method to identify
structural damage [3].
For monitoring methods, i.e., for sensor systems permanently remaining on the
structure, ambient conditions also have an influence on the wave propagation, which
often shows signal changes of the same magnitude as damage interactions of the
ultrasonic waves (Fig. 1) and can thus make the identification of damages more
difficult [4–6].
To compensate for environmental conditions, various methods have been devel-
oped in which historical data of the same transmit–receive paths, the so-called base-
lines, are used and compared to the actual conditions [7, 8]. However, other envi-
ronmental parameters, such as the condition of the transducer coupling, cannot be
mapped and can lead to erroneous results. For this reason, the development and
application of methods that only use instantaneously available comparison datasets
without the need of compensation for environmental changes is desirable.
2 Structural Health Monitoring with Guided Wave
2.1 Baseline Methods
Baseline methods as described above utilize the acoustical fingerprint of a measure-

ment path to describe the state of the inspected structure. Therefore, an initial
measurement, which could be an impulse-echo measurement as well as an actuator-
sensor (pitch-catch) measurement, will be stored as a historical baseline for signal
Transducer Arrangement for Baseline-Free … 715
comparison with actual measurement data. Since the magnitude and propagation
speed of guided waves vary with temperature and strain, a set of different base-
lines need to be collected prior to damage evaluation. Later, measurements along
a specific measurement path can reference these measurements to determine the
structural state. Baseline-based measurements have the advantage that they can also
be applied to complex shaped structures without having to take into account the
structure’s complexity. Many examples of signal processing for damage detection
are described in [9, 10]. Throughout the following, baseline based monitoring will
be considered to be non-time coincident (historical) while baseline free monitoring
will be considered to be time coincident (reference measured at the incident of
monitoring).
2.2 Simultaneous Baseline Methods
Simultaneous sensor paths were proposed by overly [11] by using equally spaced
piezoelectric sensors for pitch-catch measurements on a structure. The approach
could be compared to conventional impulse-echo or transmitting techniques, where
the magnitude of a back wall and potential damage echo is compared to an A-scan
of a nearby coordinate. The advantage of comparable sensor paths is that no histor-
ical baseline is needed to determine structural damage, because all sensor paths are
assumed to be exposed to the same temperature and structural strain state. Further-
more, aging of the transducers will also have less influence if equivalent aging of those
can be assumed. However, setup of such comparable paths is limited to geometry
conditions and even more complicated on anisotropic materials.
2.3 Lamb Wave Decomposition
A special case of the simultaneous baseline method is the use of the polarization
behavior of two piezoelectric transducers collocated at the same position at both sides
of a thin-walled structure [12, 13]. The polarization for both transducers is along the
normal of the each surface. The guided waves generated by both transducers are
reflected from the boundaries of the structure or/and from damage and finally sensed
either by the same transducers or by a second transducer set on the inspection target.
If an actual damage is located inside the structure, then a mode conversion will
happen if the initial wave packages encounter the damaged area. Due to the different
compression and tension behavior of symmetric and antisymmetric Lamb waves at
the upper and lower surface of the structure, the mode converted signal parts will
have the opposite sign on the upper and lower measurement paths. Another advantage
of this transducer arrangement is the possibility for mode selective excitation by
the simultaneous excitation of the transducers with an appropriate input voltage
(symmetric modes same sign, antisymmetric modes opposite sign). This method can
also be used for a pitch-catch measurement scenario. The need to attach transducers
to both sides of the structure under inspection may however be disadvantage of this
method in specific cases like pipes.
To avoid this disadvantages, a transducer arrangement with two piezoelectric
transducers was proposed [14]. Here, the mode tuning of piezoelectric transducers
is used. If a multiple of half the wavelength of a passing Lamb wave is identical
with the transducer dimension, then the received amplitude is maximized. If the
transducer dimension is identical to a multiple of the entire wavelength, then the
receive amplitude is minimal. The mode conversion of a structural damage and the
associated generation of a new wave package with a different wavelength allows
damage to be identified.
Other methods are based on interdigital or multi-element converters for the iden-
tification of wave modes which can be used for the identification of structural flaws
[15, 16].
The subject of this paper is the investigation of arrangements of piezoelectric
transducers for the identification of wave modes. Based on the identified wave modes,
an assessment can then be made about damages in the transmission path.
3 Simulation
The underlying objective of the work is the identification of wave modes with stacked
piezoelectric transducers. In the following, the processes that occur when the wave
modes pass through a stacked transducer arrangement will be investigated. For this
purpose, simulations with Lamb wave generated at the right transducer B1 propagated
in the structure and received at the left transducer stack were carried out (Fig. 2).
The 2D simulation was performed with Ansys.
Adhesive thickness: 0.01mm PWAS height: 0.2mm

PWAS size: 10mm
1m 1m
A2 y B2
A1 B1
x 2mm
0.1m
0.2m 0.2m
• damage depth: 1mm, damage size: 1,2,5 mm
• Plate: aluminum (E: 70 GPa, n: 0.3, r: 2700kg/m≥)
• Adhesive: Delo Rapid (E: 2 GPa, n: 0.3, r: 1160kg/m≥), linear elastic approach
• PWAS: PZT-21, Polarisation: y-Direction

• ANSYS Multiphysics
Fig. 2 Setup of the simulation experiments

Fig. 3 Deflection at the transducer stack during the S0 wave mode passes the transducer (left) and
associated signal (right)
The S0 wave mode mainly generates in-plane vibration components in the consid-
ered frequency ranges and therefore resembles a classical compression wave. When
the wave mode hits the transducer arrangement, the strain is transferred to the trans-
ducers. However, due to the coupling of the upper and lower transducers via an
adhesive layer, the deflection of the upper transducer A2 is slightly lower than that
of the lower transducer A1 (Fig. 3).
The processes take place analogously with the maximum compression of the trans-
ducers by the S0 wave mode. The deformation of the lower transducer is again larger
(green) than that of the upper transducer (red), which experiences less compression
of the transducer dimension caused by the wave mode due to the additional adhesive
layer.
The A0 wave mode generates significant out-of-plane vibration components in
the considered frequency range and thus causes the bending of the aluminum plate
during propagation. The bending causes the upper transducer, which has the greater
distance from the neutral fiber, to experience the greater compression or elongation.
Due to the greater elongation of the upper transducer A2 and the lower proportion of
the in-plane deflection of the A0 wave mode, the resulting electrical voltage is larger
at the upper transducer A1 of the transducer stack than at the lower transducer A2
when the A0 wave mode passes (Fig. 4).
In summary, it can be stated as a result from simulations:
1. The mechanical stress/strain of stacked transducers is different due to the adhesive
and the distance to the neutral fiber when passing through the same wave mode.
2. The different mechanical deformations generate correspondingly proportional
electrical voltages at the electrodes of the transducers.
3. The amplitude ratio of the electrical voltages can be used for mode identification
(Fig. 5).
It should be noted that the observed effect can be enhanced by varying the trans-
ducer dimensions and thickness of the transducers. In addition, it should be noted
that changes in adhesive thickness can have a very large influence, which must be
taken into account in the experimental investigation.
Fig. 4 Deflection at the transducer stack during the A0 wave mode passes the transducer (left) and
associated signal (right)
Fig. 5 Signal response of the transducer electrodes in simulation experiments (green A1, red: A2),
RC4 100 kHz
4 Measurement Setup
Commercial stacked transducer arrays were used to minimize deviation from ideal
coupling conditions due to the manual assembly of stacked transducers. The stacked
transducers (QuickPack QP21B, Midé) were mounted 200 mm apart on an aluminum
plate, and transmit-receive measurements were performed (Fig. 6). Figure 7 shows
the measurement signals for two transmit/receive paths and clearly display the char-
acteristics that were also demonstrated in the simulation investigations. The lower
transducer has a more direct coupling to the structure and thus the higher amplitude
for the S0 wave mode with its higher in-plane vibration component.
The upper transducer shows the higher signal amplitude for the A0 wave mode
with its higher out-of-plane component and the resulting bending of the structure.
Fig. 6 Measurement setup with stacked piezoelectric transducers

Fig. 7 Measurement response and signal envelope of the stacked transducers
However, due to the very thin stacking arrangement, the overall effect is compar-
atively small. In principle, however, this transducer arrangement is suitable for
identifying the wave modes via the amplitude ratio.
In order to minimize the influence of coupling and adhesive layers, a transducer
arrangement consisting of a PZT ring was investigated. For this purpose, electrodes
were attached to a ring-shaped piezoelectric transducer (12 mm diameter, 6 mm inner
diameter, 2 mm thickness, polarized in thickness direction, PI Ceramic) on the inside
and outside of the ring in addition to the top and bottom electrodes. Lamb waves were
excited at a second transducer (P.876-SP1, PI Ceramic), and the received signal was
evaluated at the electrode pairs of the ring transducer (inner-external vs. top-bottom).
The measurements were conducted on a 2 mm thick aluminum plate. Figure 8 shows
the test setup.
When looking at the received signals of the inside-outside and upper–lower elec-
trodes, it is noticeable that the received signals are partly phase shifted. The signals
of the S0 wave mode are in phase; those of the A0 wave mode are nearly in phase
opposition (Fig. 9).
To verify this result, a ring transducer was used as transmitter, and a B-Scan
was recorded in the direction of propagation with a scanning vibrometer (PSV-500,
Polytec). The measurements were performed for both electrode pairs. Figure 10
shows the laser vibrometry results for two different distances. The amplitude of the
velocity can also be observed to be in the phase opposition for S0 and A0 modes
by laser vibrometry, which confirms the results of the transceiver measurements
between the ring transducer and the P.876-SP1 transducer.
Fig. 8 Measurement setup with piezoelectric ring transducer

Fig. 9 Signal response at electrode pairs at 160 and 380 kHz
Fig. 10 Results of laser vibrometry on two different positions (60.1, 100 cm) in propagation
direction
5 Summary and Conclusions
The concept of baseline-free damage detection based on Lamb wave identification

was transferred to stacked piezoelectric transducers using FEM simulations. The
identification of Lamb wave modes by stacked piezoelectric transducers was verified
by laboratory measurements. The transducer stack can be seamlessly combined with
baseline techniques and also be used for mode identification in baseline applications
or acoustic emission measurements.
The ring transducer shows potential for the wave mode identification with lower
transducer coupling requirements. Since no explanation for this results have been
found yet, further investigations on the transmit–receive behavior of ring transducers
with two electrode pairs will be carried out in the future and supported by simulation
studies.
References
1. Rose JL (1999) Ultrasonic waves in solid media. Cambridge University Press, U.K
2. Giurgiutiu V (2008) Structural health monitoring with piezoelectric wafer active sensors.
Academic Press
3. Demma A et al (2004) The reflection of guided waves from notches in pipes: a guide for
interpreting corrosion measurements. NDT and E Int 37(3):167–180
4. Schubert L (2012) Zustandsüberwachung an Faserverbundwerkstoffen mit geführten Wellen.
TU-Dresden, Dissertation, TUD-Press
5. Chen F, Wilcox PD (2007) The effect of load on guided wave propagation. Ultrasonics 47:111–
122
6. Moll J, Eremin AA, Golub MV (2018) The influence of global and local temperature variation
on elastic guided wave excitation, propagation and scattering. In: 9th European Workshop
Structural Health Monitoring, July 10–13. Manchester, United Kingdom
7. Croxford AJ et al (2010) Efficient temperature compensation strategies for guided wave
structural health monitoring. Ultrasonics 50(4–5):517–528
8. Salmanpour M, Sharif Khodaei Z, Aliabadi M (2017) Guided wave temperature correction
methods in structural health monitoring. J Intell Mat Sys Struct 28(5):604–618
9. Staszewski W, Boller C, Tomlinson GR (2004) Health monitoring of aerospace structures:
smart sensor technologies and signal processing. Wiley
10. Su Z, Ye L, Lu Y (2006) Guided Lamb waves for identification of damage in composite
structures: a review. J Sound Vib 295:753–780
11. Overly TG et al (2009) Piezoelectric active-sensor diagnostics and validation using instanta-
neous baseline data. IEEE Sens J 9(11):1414–1421
12. Kim SB, Sohn H (2007) Instantaneous reference-free crack detection based on polarization
characteristics of piezoelectric materials. Smart Mater Struct 16:2375–2387
13. Dutta D (2010) Ultrasonic techniques for baseline-free damage detection in structures, Thesis
(Ph.D.). Carnegie Mellon University
14. Sohn H, Kim SB (2010) Development of dual PZT transducers for reference-free crack
detection in thin plate structures. IEEE Trans Ultrason Ferroelectr Freq Control 57(1):229–240
15. Stepinski T, Mańka M, Martowicz A (2017) Interdigital Lamb wave transducers for applications
in structural health monitoring. NDT and E Int 86:199–210
16. Ren B, Cho H, Lissenden CJ (2017) A guided wave sensor enabling simultaneous wavenumber-
frequency analysis for both Lamb and shear-horizontal waves. Sensors 17:488
Mode Switchable Guided Elastic Wave
Transducer Based on Piezoelectric Fibre
Patches
Y. Kim, K. Tschöke, L. Schubert, and B. Köhler
Abstract An improved transducer for shear horizontal guided waves is presented.

Two ideas are combined: (1) the generation of pure shear tractions by stapling two
piezoelectric fiber patches and (2) matching the excitation pattern with the spatial-
time behaviour of the intended wave. The spatial pattern is determined by the structure
of the interdigital electrodes. One possible realization of the electrode layout is
selected, and the performance of this type of transducers is demonstrated by FEM
modelling of the generated wave field. The results are discussed and scheduled further
work which is indicated.
Keywords Shear horizontal guided waves · Structural health monitoring ·

SHPFP · Piezoelectric fibre patch · Directivity · Patterned excitation
1 Introduction
Guided elastic waves (GEWs) play an important role in both long-range ultrasonic
testing and the structural health monitoring of plate- and shell-like structures. Objects
can be tubes for fluids, such as oil, or pipes for hot gases in power plants. Other
interesting objects are pillows for offshore wind turbines or the skin of an aircraft’s
fuselage. We can distinguish symmetric and antisymmetric Lamb waves and horizon-
tally polarized shear waves in plate-like structures with isotropic material properties.
For each type, there is an infinite series of modes that are characterized by their
dispersion, which is the dependence of the phase velocity on the frequency.
Most GEW applications are based on Lamb waves, which have particle displace-
ments in the sagittal plane. They are easily excited by piezoelectric wafer sensors
and actuators (see [1] for an overview). Also flexible interdigital PVDF transducers
Y. Kim
Institute of Advanced Composite Materials, KIST, Seoul, South Korea
K. Tschöke · L. Schubert · B. Köhler (B)
Fraunhofer IKTS, Maria Reiche Str. 2, 01109 Dresden, Germany
e-mail: Bernd.koehler@ikts.fraunhofer.de

https://doi.org/10.1007/978-981-15-9199-0_70
724 Y. Kim et al.
have been proposed [2]. These sensors have many advantages, such as a low weight
and a low price. However, it is difficult to excite a single mode with high purity.
Additionally, the major challenge in the application of Lamb waves is their disper-
sive characteristics, which render it complicated to use localization methods based
on the signal travel time. While there are methods to compensate for the dispersion
[3], it is best to use wave modes, which are nondispersive. The fundamental order of
the shear horizontal mode (SH0) fulfils this condition. Moreover, the SH0 is the only
horizontally polarized plate mode for frequencies below the cut-off frequency of the
SH1 mode. That is why, there is increasing effort to find efficient ways to generate
and detect SH0 waves.
There are several options to generate surface tractions by piezoelectric elements
[4]. For generation of SH waves, the face shear and the thickness shear modes are
appropriate (see Fig. 4 of [4]). There are a number of recent papers about transducers
based on these principles. The thickness shear mode is studied for example in [5–7]
and the face shear mode in [8–12]. The directivity characteristic of SHM transducers
depends on the intended application. For surveying large areas with a limited number
of transducers, an omnidirectional characteristic is advantageous and therefore aimed
for in many papers [12, 13]. On the other hand, a very high directivity can be beneficial
for the investigation of a hot spot with good selectivity against noise coming from
other directions. One recent publication proposes a novel way to make a SH0 wave
transducer with high directivity using two adjoining face shear PZT wafers [14].
By dividing solid piezoelectric material into fibres and forming patches, very flex-
ible and lightweight piezoelectric transducers can be fabricated. Invented originally
for adaptronic applications, they are studied today rather extensively as Lamb wave
transducers [15, 16] where 3D scanning Doppler laser vibrometry was used [17, 18].
In the present contribution, we combine two ideas presented in separate previous
works: (1) as shown in [4], stacked piezoelectric fibre patches (PFPs) of special
design can be used to introduce pure in-plane tractions for generating SH waves and
(2) grouping the comb-like PFP electrodes allow forming a patterned excitation for
mode selectivity [19]. The transducer performance is demonstrated by modelling
only. In some sense, this concept extends the design of the dual SHPFP [20].
The paper is organized as follows. In Sect. 2, we describe the principle of PFPs
and the set-up of PFP for shear horizontal excitation (SHPFP). In Sect. 3, we explain
the type of modelling of the PFP which is exactly what we did in recent work [20].
Based on that, we give some examples of wave fields for the proposed transducer. In
Sect. 4, we discuss the results and ideas for the continuation of the research.
2 Basics of Piezoelectric Fibre Patches (PFPs) and the New

PFP Design
PFPs have been used for guided wave generation and detection in SHM. A PFP
consists of piezoelectric fibres sandwiched between layers of adhesive, electrodes and
Mode Switchable Guided Elastic Wave … 725
polyimide film. The electrodes are perpendicular to the piezoelectric fibre orientation
and are in an interdigitated pattern (Fig. 1).
The assembly enables in-plane poling, which can utilize the d33 effect for actu-
ation. The polarization in a piezoelectric fibre is along the fibre direction with an
alternating orientation (Fig. 2). Considering that the applied electric field also has
an alternating orientation, generated forces along the fibre cancelling each other at
every position where polarization sign is changing. Therefore, only the both ends of
the fibre generate external forces along the fibre orientation to excite elastic waves.
If groups of electrodes are connected individually and thus made accessible from
outside, as indicated in Fig. 3, then there is much more freedom in controlling the
excited field. If the electrodes drawn in bold in Fig. 3 are assumed to take voltage while
the others are on ground, then only the area coloured brown is elongated/contracted.
This individual connections make all the options for spatially patterned excitation
available for such a transducer. One of the options is the sequential excitations of
Fig. 1 Scheme of a piezoelectric fibre patch (reprinted from [4] with a friendly permission of
Elsevier). The finger-like electrodes are used for both, the poling of the fibres and the generation of
the electric fields in active operation
Fig. 2 Cross section of Fig. 1 along the line A-A [20]. Upper row: a fibre with electrodes; lower
row: magnification of the left fibre end; the application of the electric voltage in two orientations
lead to electric fields (E) antiparallel (left) and parallel (right) to the polarization (P). Under clamped
conditions, a force fx is generated only at the end of the active area (where both the electric field
and the polarisation are present)
726 Y. Kim et al.
Fig. 3 PFP with external connection to groups of electrode fingers for Lamb wave excitation
according [19]
all eight groups from left to right with certain delays that the phase velocity of the
corresponding lamb wave is matched [19]. It was demonstrated that a Lamb wave
with a strong directivity and mode selectivity (S0 versus A0) can be generated.
The main idea of the present contribution is to apply the idea of the matched
excitation to the SHPFP introduced in [4]. Figure 4 presents the main idea of SHPFP
in one of the possible variants. The fibres are shown in yellow, and they are arranged
at an angle of 45° to the edges of the patch. While the electrodes are no longer
perpendicular to the fibres as usually [21] and mostly applied in commercial PFP,
they still generate electrical fields, which are nearly parallel to the fibres. That is
because the dielectric constant of the piezoelectric ceramics is by about three orders
of magnitude higher when compared to the plastic material in the gaps between
the fibres [22]. If we stack two patches with orthogonal fibre orientation, then the
combined patch will generate the surface tractions as indicated in the right hand side
of Fig. 4. This surface traction generates pure shear in the surface. Therefore, we talk
about in-plane shear transducer.
One of the options to generate a variable spatial excitation pattern by such a SHPFP
is by forming electrodes as indicated in Fig. 5. In this drawing, the black frames
indicate the excitation area, when only the most forward located electrodes carry a
nonzero voltage. Figure 6 illustrates three of several options for spatial excitation.
The left side just reproduces the case of a SHPFP according to [4]. All groups are
excited with the same voltage. Where the groups join each other, the tractions from
both sides of the (virtual) interface are equal in magnitude and opposite in the sign.
Thus, they cancel out each other and the PFP acts as a normal SHPFP. In the middle
figure, adjacent group of electrodes is controlled with alternating voltage. Thus, the
surface traction at the (virtual) interface is equal in sign and thus sums up to twice
Fig. 4 Principle of operation of an SHPFP in one of its possible electrode configuration (modified
version, Fig. 7 of [4], with friendly permission of Elsevier)
Fig. 5 Scheme of a SHPFP with separate access to electrode groups. The upper fibre layer and the
lower layer are accessible separately by their individual electrodes. The black frames indicate the
area of the fibres, which are excited by one of the corresponding electrode pares for each layer
Fig. 6 Three operating modes of the array PFP of Fig. 5. The green solid arrows indicate the
tractions of the upper fibre layer (compare Fig. 5) and the green dotted arrows those of the lower
layer. The red arrows indicate the sum of both. Left: All groups are excited in the same way. Middle:
Adjacent groups are operated in opposite way; thus, the tractions at the contact line sum up. There
is a pattern of surface traction inscribed. Right: At the given time, only the lowest group is excited.
The excitation is shifted from group to group with time delay corresponding the phase velocity of
the shear wave mode
the traction of one part. If the periodicity of the traction pattern matches the shear
wavelength, then a strong shear horizontal wave is excited. If we finally excite each
individual group with a time delay corresponding to the shear wave travel time to its
position, then a strong shear horizontal wave is emitted only in the direction of that
delay. This way a unidirectional wave will be generated.
728 Y. Kim et al.
3 Modelling of the Guided Waves Generated by the Mode

Switchable PFP
3.1 The Numerical Model
Numerical simulations were performed with the commercial FEM software ANSYS.
The procedure of modelling is described in [20]. The essential part is the description
of the heterogeneous active piezoelectric fibre structure by an effective homogeneous
medium [23, 24]. This helps to keep the computer memory and computation time
expenses to a level that the computation could still be done on a PC or a reasonable
workstation. For details, we refer to [20].
3.2 Modelling Results
At this moment, we can present only first preliminary results of this ongoing work.
The excitation scheme indicated in the RHS of Fig. 6 was used. The transducer size
is 48 × 15.5 mm2 , and it consists of 31 individually controllable regions of a width
of 0.5 mm each. The time delay between the excitation of adjacent electrode groups
corresponds to the travel time of the shear wave for travelling a group width distance.
Since we excite from bottom to top, we expect a strong ’upward’ unidirectional SH
wave and only minor waves in other directions. In Fig. 7, a snapshot of the tangential
displacement at time t = 50 µs is shown. Since the tangential displacement is normal
to the propagation direction and to the plate normal, we see essentially the SH wave
in this image. That snapshot confirms the expectation of unipolar excitation and the
displayed wavelength matches the expected wavelength of a shear wave in aluminium
at 100 kHz which is 31.5 mm.
4 Discussion and Further Work
We presented an improved shear horizontal guided wave transducer named ’mode

switchable PFP’. Its working principle is based on the combination of two different
ideas previously published. By using a numerical simulation model of this transducer,
its performance in one of its SH0 operation modes could be demonstrated here.
Examples of the other operation modes for generation of SH0 will be given in the
oral presentation. It has to be mentioned that principles of operation discussed for
the case of transmission of guided waves can be transferred one to one to the case of
reception. This follows from the general principle of reciprocity of all transducers
operating purely on the piezoelectricity [25].
As the name already suggests, the mode switchable PFP cannot only excite the
shear horizontal wave SH0 but also other modes including Lamb wave modes. The
Fig. 7 Snapshot at t = 50 µs of the wave field generated by an SHPFP excited sequentially with
time delay matched to the shear wave velocity in the aluminium plate. The tangential displacement
component visualized by colour coding. The size of the plate is 400 × 600 mm2 . The excitation is
by a 3-cycle Hanning windowed signal of 100 kHz
switching between the modes requires only change in the control of several channels
of the electrode groups. This could not be discussed here due to page limitations
and time restrictions. Further work will be focused on the verification of the mode
switchable transducer for different guided wave modes with FEM simulation and the
realization of the transducer for experimental demonstrations.
Acknowledgements One of the authors, Yongtak Kim, has received funding from the European
Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie
grant agreement N. 722134—NDTonAIR. The authors thank Jan Kunzmann from Smart Material
GmbH for helpful discussions and for supplying the MFC patches.
References
1. Giurgiutiu V (2007) Structural health monitoring: with piezoelectric wafer active sensors.
Academic Press, Burlington, USA
2. Monkhouse R, Wilcox P, Cawley P (1997) Flexible interdigital PVDF transducers for the
generation of Lamb waves in structures. Ultrasonics 35:489–498
730 Y. Kim et al.
3. Hall JS, Michaels JE (2012) Adaptive dispersion compensation for guided wave imaging. AIP
Conf Proc 1430(1)
4. Köhler B, Gaul T, Lieske U, Schubert F (2016) Shear horizontal piezoelectric fiber patch
transducers (SH-PFP) for guided elastic wave applications. NDT and E Int 82:1–12
5. Kamal A, Giurgiutiu V (2014) Shear horizontal wave excitation and reception with shear
horizontal piezoelectric wafer active sensor (SH-PWAS). Smart Mat Struct 23(085019):17
6. Boivin G, Viens M, Belanger P (2016) Development of a low frequency shear horizontal
piezoelectric transducer for the generation of plane SH waves. AIP Conf Proc 1706:030019
7. Huan Q, Miao H, Li F (2017) A nearly perfect omnidirectional shear-horizontal (SH) wave
transducer based on a thickness poled, thickness-shear (d15) piezoelectric ring. arXiv preprint
arXiv:1704.03629
8. Zhang S, Jiang W, Meyer RJ Jr, Li F, Luo J, Cao W (2011) Measurements of face shear
properties in relaxor-PbTiO3 single crystals. J Appl Phy 110(064106):6
9. Zhou W, Li H, Yuan F-G (2015) Fundamental understanding of wave generation and reception
using d36 type piezoelectric transducers. Ultrasonics 57:135–143
10. Miao H, Dong S, Li F (2016) Excitation of fundamental shear horizontal wave by using face-
shear (d36 ) piezoelectric ceramics. J Appl Phy 119(174101):7
11. Miao H, Huan Q, Li F (2016) Excitation and reception of pure shear horizontal waves by using
face-shear d24 mode piezoelectric wafers. arXiv preprint arXiv:1604.03765
12. Miao H, Huan Q, Wang Q, Li F (2017) A new omnidirectional shear horizontal wave transducer
using face-shear (d24 ) piezoelectric ring array. Ultrasonics 74:167–173
13. Belanger P, Boivin G (2016) Piezoceramic omnidirectional transduction of the fundamental
shear horizontal guide wave mode. Nondestructive characterization and monitoring of advanced
materials, aerospace, and civil infrastructure. Int Soc Opt Photonics 98040M:6
14. Miao H, Huan Q, Li F, Kang G (2018) A variable-frequency bidirectional shear horizontal (SH)
wave transducer based on dual face-shear (d24 ) piezoelectric wafers. Ultrasonics 89:13–21
15. Mańka M, Rosiek M, Martowicz A, Stepinski T, Uhl T (2013) Lamb wave transducers made
of piezoelectric macro-fiber composite. Struct Control Health Monit 20(8):1138–1158
16. Köhler B, Schubert F, Barth M, Frankenstein B (2008) Selective excitation and detection of
Lamb waves for SHM applications. Proc Fourth Eur Workshop Struct Health Monit. Krakow,
Poland
17. Schubert L, Barth M, Klesse T, Köhler B, Frankenstein B (2008) Guided elastic waves and
their impact interaction in CFRP structures characterized by 3D laser scanning vibrometry.
Proc 15th Int Sym Smart Struct Mat Nondestructive Eval Health Monit 69350G
18. Barth M, Köhler B, Schubert L (2008) 3D-Visualisation of Lamb waves by laser vibrometry.
Proc 4th Eur Workshop Struct Health Monit:641–648
19. Koehler B, Frankenstein B, Schubert F, Barth M (2009) Novel piezoelectric fiber transducers
for mode selective excitation and detection of Lamb waves. AIP Conf Proc 1096(1):982–989
20. Yongtak K, Gaul T, Köhler B (2019) Improved SH0 guided wave transducers based on
piezoelectric fiber patches. Sensors 19:2990
21. Bent AA, Hagood NW (1997) Piezoelectric fiber composites with interdigitated electrodes. J
Intell Mat Syst Struct 8:903–919
22. Ernst D, Bramlage B, Gebhardt S, Schönecker A (2015) High performance PZT thick film
actuators using in plane polarisation. Adv Appl Ceram 114:237–242
23. Williams RB, Inman DJ, Wilkie WK (2006) Nonlinear response of the macro fiber composite
actuator to monotonically increasing excitation voltage. J Intell Mater Syst Struct 17:601–608
24. Bowen CR, Giddings PF, Salo AI, Kim HA (2011) Modeling and characterization of piezo-
electrically actuated bistable composites. IEEE Trans Ultrason Ferroelectr Freq Control
58:1737–1749
25. Auld B (1979) General electromechanical reciprocity relations applied to the calculation of
elastic wave scattering coefficients. Wave Motion 1:3–10
Lamb Wave Excitation Using a Flexible
Laser Ultrasound Transducer
for Structural Health Monitoring
Wei Li, Jitao Xiong, and Wenbin Huang
Abstract Using Lamb waves for detection is a desirable method for the structural
health monitoring (SHM) and has been widely used in plate-like structure. In this
paper, an originally flexible ultrasonic transducer with high photo-acoustic conver-
sion efficiency was proposed by using candle soot nanoparticles (CSNPs) and poly-
dimethylsiloxane (PDMS). Experiments results demonstrate that the developed trans-
ducer can generate a longitudinal wave in the aluminum block with the amplitude
of 10 times to that of the conventional laser ultrasound technique, demonstrating the
high energy conversion efficiency of the proposed transducer. Wedge-shape trans-
ducers were developed to excited Lamb wave in the 1.5 mm aluminum substrate
by the oblique incident method. The specific A0 and S 0 modes of the Lamb wave
with the central frequency of 647 kHz were successfully excited in the aluminum
plate of 1.5 mm thickness. A0 mode Lamb wave was selected for damage detection
application. According to the synthetic aperture focusing imaging technique, a delay-
and-sum signal processing method was adopted for damage location in the plate. A
3.5 mm defect was well imaged and the results demonstrate that the developed flexible
photo-acoustic transducer can be a promising method for the SHM.
Keywords Laser ultrasonic transducer · Photo-acoustic effect · Lamb waves ·

Synthetic aperture focusing imaging · Structural health monitoring
1 Introduction
Bulk wave ultrasound, infrared thermal imaging, X-rays, eddy current and other
health assessment methods have been applied effectively to structural health moni-
toring (SHM). However, under some specific conditions, most of these techniques
are often tardy and redundant for the inspection of large structures, which including
pipelines, marine, ships, and aerospace [1]. The guided waves ultrasonic testing
W. Li · J. Xiong · W. Huang (B)

State Key Lab of Mechanical Transmission, Chongqing University, No. 174 Shazhengjie,
Chongqing, China
e-mail: whuang@cqu.edu.cn

https://doi.org/10.1007/978-981-15-9199-0_71
732 W. Li et al.
potentially provides a wise replaceable solution with traditional approach toward

SHM. The guided wave has some unique abilities as follows [2]: (a) The guided
wave can realize long-distance detection with low energy loss; (b) The guided wave
has dispersion characteristics and multi-mode characteristics, and can identify and
locate different types of defects by selecting the appropriate mode and frequency of
the guided wave; (c) wide application range and good sensitivity; (d) simple and fast,
and low cost. It is these advantages that make it the most widely used technology in
SHM.
Lamb waves are commonly used as detection signals in current structural health
monitoring studies [3]. Senyurek [4] successfully detected two common damage
defects on the wing slats of Boeing 737 aircraft through Lamb waves. Rucka [5]
used the finite element analysis to simulate the failure caused by the bolting of the
plate beam and beam on the railway bridge, and successfully detected these two
damages by Lamb wave.
The excitation of the Lamb wave can be activated based on various physical
effects, which can be categorized as the contact and non-contact types. Non-contact
type mainly includes laser ultrasonic and electromagnetic ultrasonic [6], and contact
type mainly includes the piezoelectric effect. Contact excitation has some limitations
due to its direct contact with the sample [7, 8]: (a) The bandwidth of the ultrasonic
signal excited by the resonant system is too narrow; (b) Limited by the size of the
sensor, the spatial resolution is low; (c) The coupling agent will chemically react with
the sample under high temperature and high pressure conditions; (d) It is difficult to
move the sensor while maintaining the coupling condition; and (e) The stress of the
contact sensor on the sample will affect the propagation of the ultrasonic wave.
For non-contact excitation, electromagnetic ultrasonic is an ideal choice for direct
excitation and detection of horizontal shear waves. However, transducers are very
large in size and their practicality is limited to conductive material [9, 10]. On the
other side, the laser ultrasonic technology has the characteristics of high energy, small
light divergence angle, short duration, and can excite broadband guided signals [11,
12]. However, the conventional laser ultrasound suffers from two shortcomings. The
excitation of the detectable Lamb wave always involves the ablation of the materials,
leading to the damage of the structure integrity. Meanwhile, the mode and frequency
of the excited Lamb wave by the conventional laser ultrasound are difficult to control,
thus lacking of the adaption for detecting different defects size and structures.
Recently, a new type of composite laser ultrasound transducer was developed
for achieving the high efficient laser to acoustic energy conversion. It composes of
a laser absorption component and a thermal expansion component. The absorbed
short period laser pulse converts into the heat and causes the thermal expansion. The
short duration of the thermal expansion corresponds to an acoustic pulse propagating
outward. Due to the high photo-acoustic conversion efficiency and short duration of
the laser pulse, the laser ultrasound transducer could generate a high pressure acoustic
signal with wide frequency bandwidth. Therefore, it has shown broad application
prospects in biomedical drug delivery, high-frequency ultrasound focused targeted
Lamb Wave Excitation Using a … 733
therapy, non-invasive high-precision surgery, and all-optical ultrasound imaging [13–

15]. However, there is no report about the Lamb wave excitation using this laser
ultrasound transducer.
In this paper, we described a new Lamb wave excitation technique by using a
flexible laser ultrasonic transducer with high photo-acoustic conversion efficiency.
The transducer is prepared from two materials: candle soot nanoparticles (CSNPs)
and polydimethylsiloxane (PDMS). The photo-acoustic performance was tested and
compared with the conventional laser ultrasound technique. According to Snell’s law,
Lamb waves of the symmetric and asymmetric modes are excited in the aluminum
plate by the oblique incidence of longitudinal waves. Finally, A0 mode is selected for
a damage detection demo using the synthetic aperture imaging technique. It is proved
that the Lamb wave excited by this kind of flexible laser ultrasound transducer has a
good application prospect for SHM.
2 Photo-acoustic Performance of the Transducer
The preparation of the CSNPs-PDMS transducer has been described in our previous
paper [16]. To illustrate the high photo-acoustic conversion efficiency of the new
CSNPs-PDMS ultrasonic transducer, the performance of the laser ultrasound trans-
ducer for exciting the longitudinal wave in a 15 mm thick aluminum block was
examinated and compared with that of the direct laser radiation without using the
transducer.
The experiment setup for the transducer characterization is shown in Fig. 6. A laser
beam about 6 mm in diameter was excited by a 532 nm Q-switch Nd:YAG pulsed
laser (CLASS-AI, Beijing Raycotech Optronic Corp) with a 10 Hz repetition rate
and a 6 ns pulse duration. The energy adjustment from laser beam was conducted by
an energy attenuator so as to ensure its stability. The received signal was amplified
by 20 dB using a pulse transceiver (5073PR, Olympus) and then acquired by a
digital phosphor oscilloscope (TDS2024C, Tektronix). Two sets of longitudinal wave
excitation experiments of an aluminum block were performed in this part. The first
part was the excitation by the proposed transducer and the second part was the direct
laser radiation excitation without using the transducer. A longitudinal wave probe
(DL10P6L, Guangzhou Doppler Corp) was placed on the backside of the aluminum
block to receive the longitudinal wave signal. We acquired each waveform after
averaging 16 signal sources in the time domain. The laser generator produced the
trigger signal. The input laser energy was kept within 15 mJ to avoid damaging the
transducer under excessive power (Fig. 1).
The acoustic signals generated with and without using the laser ultrasound trans-
ducer under a 5.1 mJ laser pulse are displayed in Fig. 2. Two signals have the similar
time duration, which have the center frequency of 19.68 MHz and −6 dB band-
width of 13.89 MHz. It indicates that the laser ultrasound transducer causes little
distortion to the acoustic signal. In addition, under the same laser energy excitation
condition, the average value of the amplitude of the excited longitudinal wave by the
734 W. Li et al.
Fig. 1 Experimental setup to characterize the transducer’s performance for exciting the longitudinal
wave in an aluminum block
Fig. 2 Comparison of the longitudinal wave signals in a 15 mm thick aluminum block excited by
a 5.1 mJ laser pulse with and without the laser ultrasound transducer
proposed transducer is 10 times to that of the signal generated by direct exciting in

the aluminum block without using the transducer. This shows that the photo-acoustic
conversion efficiency of the laser ultrasonic is greatly improved by using the proposed
transducer.
3 Lamb Wave Excitation
As the center frequency of acoustic pulse generated by a thin laser ultrasound trans-
ducer is above 10 MHz, it is too high for the Lamb wave excitation in a millimeter-
thick metal plate. Since the high-frequency ultrasound has a much higher attenuation
in the composite material, a thick transducer can be utilized for attaining an acoustic
wave with the center frequency about hundreds of kHz. A 20 mm thick CSNPs-
PDMS was fabricated and the center frequency was found to be 647 kHz and the −
3 dB bandwidth is 653 kHz. The Lamb waves are excited using the oblique incidence
method using a wedge shape transducer as shown in Fig. 3. The critical angles for
generating the modes of A0 and S 0 are given in Table 1.
A circular PZT piezoelectric patch (SMD05T04R111WL, STEMiNC) with a
center frequency of 450 kHz was served as a receiver to test Lamb wave signals
produced by the wedge shape transducer. After averaging 64 signal traces in the
time domain, each waveform was obtained and denoising by dB4 wavelet in the
subsequent data processing procedure.
Figure 4 shows the excitation effect of the wedge laser ultrasonic transducer on
A0 and S 0 . The signal waveform is received by the PZT sensor at the distance of
200 mm. In Fig. 4a, there is only one distinct wave packet with a group velocity of
3.138 mm/us. Basing on the dispersion curve of the A0 mode, the group velocity at
647 kHz is 3.177 mm/us. The error between the measured velocity and the theoretical
value is 1.22%, proving the successful excitation of pure A0 mode. In Fig. 4b, a
plurality of wave packets appears, wherein the group velocity of the first wave packet
is 5.311 mm/us, which is similar to the 647 kHz S 0 mode with the group velocity of
5.236 mm/us. The reason for the subsequent wave packets lies in the reflections of the
incident wave within the transducer. After the incident wave reflects at the interface
Fig. 3 Lamb wave excitation mechanism using the laser ultrasound transducer
Table 1 Wedge angles

Central frequency θ A0 θ S0
corresponding to different
Lamb wave modes excitation 647 kHz 27.07° 11.32°
736 W. Li et al.
Fig. 4 Performance of the wedge-shape laser ultrasound transducer for exciting the A0 (left) and
S 0 (right) mode Lamb wave
between the transducer and the metal plate, it could reflect again at the oblique
surface of the wedge and reach the interface with a double incident angle, thereby
exciting other low-speed modes. Therefore, based on the comparison between the
above experimental results, the A0 mode Lamb wave can meet the requirement for
the single mode excitation and is used in subsequent imaging experiments.
4 Damge Detection Using the Laser Ultrasound

Transmitter
The experimental setup for the damage detection using the laser ultrasound transducer
is shown in Fig. 5. Nine piezoelectric sensor elements were places as a linear array
with the pitch of 8.4 mm. The simulated defect is prefabricated at the center of the
aluminum plate with the size of 6 × 3.5 mm in an elliptical shape. The center position
of the array is set as the coordinate origin. The relative coordinate of the defect center
is (−5 mm, 160 mm), corresponding to r = 160.08 mm and ϕ = 91.79° in the polar
coordinate. The signal of each receiving element was recorded separately by the
digital oscilloscope and transmitted to a computer for its analysis.
For the experimental data processing, the scanning area is discretized within
the polar coordinates (r, ϕ). The distances between the interested point with the
excitation position and with the receiving element are calculated and converted into
the time delays on the basis of the group velocity of the guide wave. Based on the
SAFT method, the defect reflection signal strength can be obtained at that point. The
scanning half-plane image is shown in Fig. 6. The regions with the brightest color
represent the higher amplitude reflected echoes which occur around the defects. It
can be seen that the defect position is in the direction of 93° and with the radial
position of 161.72 mm. It is close to the actual position of the defect on the test
Fig. 5 Experimental setup for crack imaging by the laser ultrasound transmitter and piezoelectric
receiver array system
Fig. 6 Scanning image by the laser ultrasound transmitter and piezoelectric receiver array system
738 W. Li et al.
sample (r = 160.08 mm and ϕ = 91.79°) with the error down to 1.07%, indicating
that the above synthetic aperture imaging results could locate the defect accurately.
In addition, there are other bright areas on the same circumference of the defect
center. This can be attributed to the grating lobe of the piezoelectric sensing array and
the insufficient signal-to-noise ratio due to the limited number of the array elements.
Nevertheless, the amplitudes of these areas are much lower compared with that of
the real defect location. This phenomenon can be suppressed by using a denser
piezoelectric array distribution and a larger synthetic aperture.
5 Conclusions
In this paper, a flexible laser ultrasound transducer is designed to excite Lamb waves
and apply them to defect detection in aluminum panels. Comparing to the perfor-
mance of the conventional laser ultrasound technique for the longitudinal wave exci-
tation in a 1.5 mm thick aluminum plate, we found that the proposed transducer
could generate an acoustic wave with a 10 times amplitude magnification, proving
the high photo-acoustic conversion efficiency of the transducer. The laser ultrasonic
transducer was further used to excite the Lamb wave signal of a single A0 mode by
oblique incidence, which indicates that the interference of other modal signals can
be effectively avoided in actual application. Finally, the one-dimensional linear array
piezoelectric sensor is used to receive the reflected defect signal. According to the
principle of synthetic aperture focusing imaging, the Lamb wave successfully locates
the defect of the aluminum plate. It is verified that Lamb wave excitation based on
this new laser ultrasonic transducer is feasible for SHM.
Acknowledgements The authors gratefully acknowledge the support of the National Natural
Science Foundation of China (Nos. 51605052), the Fundamental Research Funds for the
Central Universities (2019CDQYJX008) and Natural Science Foundation of Chongqing
(cstc2017jcyjAX0188).
References
1. Mitra M, Gopalakrishnan S (2016) Guided wave based structural health monitoring: a review.
Smart Mater Struct 25(5):53001
2. Wang L, Yuan FG (2007) Group velocity and characteristic wave curves of Lamb waves in
composites: modeling and experiments. Compos Sci Technol 67(7–8):1370–1384
3. Su Z, Ye L (2009) Identification of damage using Lamb waves: from fundamentals to
applications, vol. 48. Springer Science & Business Media
4. Senyurek VY (2015) Detection of cuts and impact damage at the aircraft wing slat by using
Lamb wave method. Measurement 67:10–23
5. Rucka M, Zima B, K˛edra R (2014) Application of guided wave propagation in diagnostics of
steel bridge components. Arch Civ Eng 60(4):493–516
6. Silk MG (1984) Ultrasonic transducers for nondestructive testing
7. Dewhurst RJ, Edwards CE, McKie ADW, Palmer SB (1987) Comparative study of wide-band
ultrasonic transducers. Ultrasonics 25(6):315–321
8. Drain LE (2019) Laser ultrasonics techniques and applications. Routledge
9. Khalili P, Cawley P (2018) Relative ability of wedge-coupled piezoelectric and meander coil
EMAT probes to generate single-mode Lamb waves. IEEE Trans Ultrason Ferroelectr Freq
Control 65(4):648–656
10. Ribichini R, Cegla F, Nagy PB, Cawley P (2011) Study and comparison of different
EMAT configurations for SH wave inspection. IEEE Trans Ultrason Ferroelectr Freq Control
58(12):2571–2581
11. Liu Z, Chen H, Sun K, He C, Wu B (2018) Full non-contact laser-based Lamb waves phased
array inspection of aluminum plate. J Vis 21(5):751–761
12. Lee SE, Liu P, Ko YW, Sohn H, Park B, Hong J-W (2019) Study on effect of laser-induced
ablation for Lamb waves in a thin plate. Ultrasonics 91:121–128
13. Kim J, Kim HW, Chang WY, Huang W, Jiang X, Dayton P (2019) Candle-soot carbon nanopar-
ticles in photoacoustics: advantages and challenges for laser ultrasound transmitters. IEEE
Nanotechnol Mag 13(4):13–28
14. Chen S-L (2016) Review of laser-generated ultrasound transmitters and their applications to
all-optical ultrasound transducers and imaging. Appl Sci 7(1):25
15. Colchester RJ et al (2019) All-optical rotational ultrasound imaging. Sci Rep 9(1):1–8
16. Chang W-Y, Huang W, Kim J, Li S, Jiang X (2015) Candle soot nanoparticles-
polydimethylsiloxane composites for laser ultrasound transducers. Appl Phys Lett
107(16):161903
Lamb Wave Propagation of S0 and A0
Modes in Unidirectional CFRP Plate
for Condition Monitoring: Comparison
of FEM Simulation with Experiments
Subin Kim, Joonseo Song, Seungmin Kim, Younho Cho, and Young H. Kim
Abstract CFRP is a material widely used in a variety of industries, including

aerospace, wind blade, architecture, automobile and other fields of engineering.
Ultrasonic waves are widely used for condition monitoring (CM) of CFRP struc-
tures, so quantitative understanding of wave propagation on anisotropic material is
essential. Conceptual ambiguities were found in previous researches and were clar-
ified by introducing new perspective. FEM simulation was carried out for Lamb
wave propagation on unidirectional CFRP plate. The wave speed of S0 mode from
the simulation showed good agreement with that from experiment, which guarantees
the accuracy of the simulation. For the simulation results, S0 and A0 mode wave can
be separated by adding and subtracting displacements at upper and lower surface.
Group velocity of S0 mode is highly affected by the propagating direction whereas
A0 showed lower dependence, and this was shown in our simulation data. Also, an
unexpected wave was observed in the simulation and was identified as SH0 mode.
Keywords CFRP · Condition monitoring · Lamb waves · Group velocity · FEM

simulation
1 Introduction
Carbon Fibre Reinforced Plastic (CFRP) is a material widely used in many industries
including aerospace, tanks and automobiles. It is also known to be the material used
to make the blades of the wind power generator. This is because of its high strength
S. Kim · J. Song · S. Kim

Applied Acoustics Laboratory, Korea Science Academy of KAIST, Busan 47162, South Korea
Y. Cho
School of Mechanical Engineering, Pusan National University, Busan 46241, Republic of Korea
Y. H. Kim (B)
Institute of Nuclear Safety and Management, Pusan National University, Busan 46241, Republic
of Korea
e-mail: yhkim627@pusan.ac.kr

https://doi.org/10.1007/978-981-15-9199-0_72
742 S. Kim et al.
compared to its extremely lightweight. Condition monitoring (CM) is essential to

prevent failures, and ultrasonic method is one of the powerful tools for this purpose.
However, CM in CFRP is a challenge because of two reasons. Firstly, CFRP is
mainly used as plates which are a dispersive medium, and wave velocity changes
for different frequencies. Secondly, CFRP is anisotropic along its fibre direction,
causing wave velocity to change by direction. These two obstacles make prediction
of Lamb wave propagation difficult.
In the present work, we carried out both experiment and FEM simulation on Lamb
wave propagation on unidirectional CFRP plate. Through this research, we showed
the credibility of FEM simulation, by comparing it with our experimental results.
Based on this credibility, group velocity of S0 and A0 modes was obtained from
FEM simulation.
2 Conceptual Clarification of Group Velocity
When some certain conditions are applied on the solid, the wave which propagates
through the solid can have energy flow which is different from phase wave direction
and velocity; this is ‘group velocity’ or ‘energy velocity’. When analysing the group
velocity inside the anisotropic bulk media, we usually draw slowness surface, which
is the reciprocal of the phase velocity on a polar coordinate as shown in Fig. 1. Then,
the perpendicular direction of certain phase direction is the group velocity direction,
and the magnitude of the group velocity is
cp
cg = (1)
cos θ
where cg is group velocity, cp is phase velocity, and θ is skewed angle. In this case,
the group velocity is bigger than phase velocity.
Fig. 1 Direction of the

group velocity derived by the
slowness surface [1]
Lamb Wave Propagation of S0 and A0 Modes in Unidirectional … 743
In the other hand, the Lamb wave dispersion curves in isotropic media show that
the phase velocity is generally bigger than the group velocity. In this case, group
velocity is defined by
∂ω
cg = (2)
∂k
in one dimension. Because the magnitudes of the phase velocity and group velocity
are different in the above-stated cases; the group velocity in an anisotropic plate is
hard to derive. In fact, asymptotic wavefront (the set of group velocity of each direc-
tion) can be obtained by getting group velocity from certain direction and modifying
the direction by slowless surface [2].
However, the definition of group velocity in multi-dimension space is
cg = ∇k ω (3)
where ω is the angular frequency, and k is the wave number [3]. This definition
includes that the direction of the group velocity is perpendicular to the slowless
surface, but the value can be different with the ∂ω/∂k in certain direction. Therefore,
group velocity is given by gradient.
From the phase velocity dispersion curve, we can draw ω versus k graph for each
direction and combine them on two-dimensional wave vector domain, k x and k y .
Then, the group velocity of certain phase velocity direction and angular frequency
is the gradient ∇k ω of that point. When the point on the three-dimensional graph is
not continuous,

avx + bv y = ωab − ω0
(4)
cvx + dv y = ωcd − ω0
then solve this simultaneous equation about each partial derivatives. a, b and c, d
are the x and y components of two different small vector k from the point we are
investigating. This means that we check the difference of angular velocity between
the investigating point and two other adjusting points. vx , v y are solution of the
equation, and they are partial derivatives. ωab , ωcd is angular frequency of each
point, and (vx , v y ) is the group velocity.
3 Methodology
3D FEM simulation that completely matched the experimental conditions was

devised to get a theoretical account of the wave velocity of the Lamb wave on
unidirectional plate. The thickness of the plate, wave frequency and cycle were same
as the experimental conditions. The unidirection CFRP material was modelled by
entering the stiffness tensor of the CFRP plate.
744 S. Kim et al.
A benefit of using the FEM simulation is that it allows us to separate the symmetry
mode and antisymmetry mode. In experimental data analysis, we are only able to
find the velocity of S0 mode, which is the fastest mode, since we only can track the
velocity of the wave that comes first in the oscilloscope reading. Others are ambiguous
since the waves interfere. However, by using the characteristic of symmetry and
antisymmetry, for FEM simulation results, we can add or subtract the data from the
top and bottom of the plate of the same point on CFRP plate. By this method, we can
get a more reliable data analysis method, by being able to separate the two modes.
In our experiment, we prepared an unidirectional CFRP whose thickness is 1.4 mm
and used 1 MHz oscilloscope, wave generator and 1.0 MHz transducers to generate,
receive and record the wave data. The CFRP plate was eight-ply ([0]8 ) with thickness
of 1.4 mm. A five-cycle tone burst sine wave signal with 0.36 MHz frequency was
used as an input signal to the plate. Experiment was also carried out for fd value of
0.50 MHz·mm in order to only allow S0 and A0 mode to be observed.
The distance and angle between the receiver and the source were varied. Distance
from the source was from 3 to 12 cm with 1 cm interval, and the direction of placement
was from 0° to 90°, with five intervals from 0° to 30° and 10° interval from 30° to
90°. The 0° line represented the fibre direction (x-axis). Transducers and oscilloscope
were used to collect the signals.
Figure 2 shows typical waveforms along the radial lines, from 3 to 12 cm. The
waveforms were shifted with increasing propagation distance, and the group velocity
of the Lamb wave can be determined. For simulation, we could get more accurate
result by using the symmetric and antisymmetric property of the Lamb wave. By
adding and subtracting the out-of-plane displacements obtained at the upper and
lower surfaces, S0 and A0 mode can be obtained as shown in Fig. 3.
Even though only group velocity of S0 mode was obtained through experiment,
both those of S0 and A0 modes were obtained through simulation. Video snapshot
of the experiment results is shown in Fig. 4. For S0 mode, the experimental results
were compared with the simulation results and shown in Fig. 5a, which showed a
very good agreement. From this, the credibility of the FEM simulation is confirmed
in determination of the group velocity of the Lamb wave. This allows us to reliably
obtain the group velocity of the A0 mode from the simulation results as shown in
Fig. 5b. It is clearly shown that S0 mode was seen to have very high dependence for
direction, while A0 had much smaller dependence.
Unexpected wave was found in the FEM simulation video, with its vibration
displacement dominantly on y-direction (90° direction). Figure 4b shows the video
snapshot of y-displacement. This wave was seen to be nearly isotropic, with its
wavefront being almost in circular shape. From the vibrational direction of the wave
and the almost isotropic property of the wave velocity, we deduced that this wave
is the shear horizontal mode. Further verification of the wave mode was done by
Fig. 2 Time delay of successive graphs along a radial line
Fig. 3 Displacements a at upper the surface, b at lower the surface. Obtained c A0 and d S0 modes
by adding and subtracting two displacements, respectively,
comparing its velocity to theoretical SH mode velocity, and they were seen to be
similar, around 2000 m/s. It implies that in-plane displacement is also generated for
the normal excitation [4].
746 S. Kim et al.
Fig. 4 Video snapshot of vibrational displacement in a x-axis direction, b y-axis direction and
c z-axis direction (out-of-plane). The snapshot shows the image at time t = 0.0247 ms. The most
dominant oval wave is the A0 mode, whereas the straight waveform parallel to x-axis is S0 mode.
It is also noted that S0 mode is much weaker in z-direction displacement, which is expected, since
S0 mode consists of in-plane displacements
Fig. 5 a Comparison of group velocity of S0 mode Lamb wave on the CFRP, measured by FEM
simulation and experiment. b The group velocity of A0 mode of Lamb wave on CFRP, measured
by FEM simulation
5 Conclusion
In order to obtain quantitative information of wave propagation in CFRP plate, both

experiment and 3D FEM simulation have been carried out. The simulation and exper-
imental results showed good agreement with each other, which proves that the FEM
simulation results are credible. S0 and A0 mode wave can be separated by adding
and subtracting displacements at upper and lower surface. It is shown that group
velocity of S0 mode is highly affected by the propagating direction whereas A0
showed lower dependence. An unexpected wave was also observed in the simulation
and was identified as SH0 mode.
Acknowledgements This work was supported by the Korea Science Academy of KAIST with
funds from the Ministry of Science, ICT and Future Planning. Thanks to Dr. Jeong-Ki Lee for his
valuable comments.
References
1. Carcione JM (2015) Anisotropic elastic media. In: Wave fields in real media, 3rd edn. Elsevier
Science, Amsterdam, pp 1–62
2. Rhee S-H, Lee J-K, Lee J-J (2007) The group velocity variation of Lamb wave in fiber reinforced
composite plate. Ultrasonics 47:55–63
3. Auld BA (1973) Group velocity and energy velocity. In: Acoustic fields and waves in solid, vol.
2. Interscience, Geneva, p 202
4. Kim YH, Song S-J, Lee JK, Kim HC (2003) Transverse-wave modes in the pulse-echo signal
of a normal-beam longitudinal-wave mode transducer. J Korean Phys Soc 42:111–117
Thermographic Testing
Three-Dimensional Reconstruction
of Leaked Gas Cloud Image Based
on Computed Tomography Processing
of Multiple Optical Paths Infrared
Measurement Data
M. Uchida, D. Shiozawa, T. Sakagami, and S. Kubo
Abstract The current gas leakage source detection was conducted by the human
senses and experience. The development of remote gas leakage source detection
system is required. In this research, infrared camera was used to detect gas leakage.
The gas can be detected by the absorption of infrared rays by the gas, and the infrared
rays are emitted from the gas itself. Three-dimensional reconstruction of leaked gas
cloud was performed to identify the gas leakage source. The 3D reconstruction of
leaked gas cloud was performed by the multiple optical paths of infrared measurement
and inverse tomography analysis. Algebraic reconstruction techniques (ART) method
was applied to reconstruction. In the experiments, the gas concentration distribution
was simulated by the arrangement of gas cells. It was found that the gas concentration
distribution composed by gas cell could be estimated by infrared images obtained
with few optical path and ART method.
Keywords Infrared camera · Leaked gas cloud · Computed tomography
1 Introduction
One of the serious accidents at gas plants is gas leakage explosion. The current
gas leakage source detection is conducted by the human senses and experience.
The retirement of highly skilled workers and aged plant facilities is increasing. The
development of continuous and remote gas leakage monitoring and gas leakage
source detection is required. The gas can be detected by the absorption of infrared
rays by the gas and the infrared rays emitted from the gas itself. A wide range of gas
M. Uchida
Department of Mechanical Engineering, Kobe University, 1-1 Rokkodai, Nada-Ku, Kobe, Japan
D. Shiozawa (B) · T. Sakagami · S. Kubo
Department of Mechanical Engineering, Kobe University, Kobe, Japan
e-mail: 181t310t@stu.kobe-u.ac.jp
S. Kubo
Osaka University, Suita, Japan

https://doi.org/10.1007/978-981-15-9199-0_73
752 M. Uchida et al.
leak monitoring can be performed remotely using an infrared camera. If the three-
dimensional gas flow is measured, then it is useful for identifying the gas leakage
source. In this study, the computed tomography technique from infrared measurement
was applied to obtain the three-dimensional gas cloud. It is difficult to measure
infrared image of gas cloud from multiple optical paths. To reconstruct 3D gas
cloud from few infrared image, algebraic reconstruct technique (ART) was applied.
The applicability of 3D gas cloud identification using the infrared measurement
and ART method was experimentally investigated. In the experiment, the gas cloud
identification experiment was performed by simulating the gas distribution using
plural gas cells filled with fluorocarbon gas.
2 Measurement Principle
2.1 Gas Visualization Principle
The gas visualization principle is shown in Fig. 1. When gas is observed with an
infrared camera, the infrared ray coming from the background of the gas is partly
absorbed by the gas, and the intensity of the infrared rays is weakened. In addition,
infrared rays due to the temperature of the gas itself are observed. The gas is not
visible on the optical image, but it can be visualized on the infrared image. The
gas absorption characteristics depend on the type of gas. For example, since the
fluorocarbon absorption band exists in the range of 3.2–3.4 μm, it can be detected
with a mid-wavelength infrared camera. In order to emphasize the absorption by gas,
a filter in this wavelength band is used. A filter attached to the infrared camera is
used according to the type of gas to be detected.
Fig. 1 Illustration of gas visualization using an infrared camera

Three-Dimensional Reconstruction of Leaked Gas Cloud … 753
2.2 ART Reconstructed Method
In a gas plant, it is difficult to acquire data in all directions, such as X-ray CT. There-
fore, ART reconstructed method was applied to the identification of 3D gas distri-
bution. In particular, 3D gas cloud images are reconstructed from several measured
infrared images. The ART method is one of the image reconstruction algorithms.
The reconstructed cross-sectional image x̂ obtained by ART is expressed as follows;
x̂ = arg minC(x) (1)

M
C(x) = (yi − (Ax)i ) (2)
i=1
x is the assumed cross-sectional image in the analysis process, A is called as system

matrix, Ax is the forward projection of the assumed cross-sectional image x, y is the
measured projection image, and the subscript i indicates the ith projection image,
and M is the total number of projected image. The forward projection image Ax
for x is calculated, and this forward projection image Ax is compared with the
measured projection image y. x is updated based on the comparison. By repeating
these calculations, x that minimizes C(x) which is defined as the output image x̂ of
ART.
3 Experimental Equipment
To simulate a gas concentration distribution in the experiment, PFA bottles (gas cell)
containing gas (Fluorocarbon gas) are arranged on a rotating table as shown in Fig. 2.
The infrared images from multiple directions were obtained by rotating the table. An
infrared camera A6751sc (FLIR Inc.) was used for infrared measurement (Fig. 3).
The properties of infrared camera are shown in Table 1. The temperature of bottle and
gas was kept at room temperature. The temperature of a pseudo blackbody furnace
was kept constant at 303.15 K, is same to the room temperature, and the furnace was
placed on the background of the gas cell. So the infrared radiation emitted from gas
itself is small as to be ignorable as compared with an absorption of infrared ray from
background. In the experiments, the gas absorption of infrared radiation from the
background was measured. A filter with a transmission wavelength of 3.2–3.4 μm
was attached to an infrared camera. Infrared of this wavelength range is absorbed by
fluorocarbons. The rotation table was rotated by an angle θ°, and the infrared images
from various direction were captured. Example of infrared images by experimental
equipment is shown in Fig. 4. The gray bottles in the figure are filled with the gas,
and a bottle with the white bottles is empty.
Fig. 2 Illustration of the measurement equipment
Fig. 3 Infrared camera with InSb sensor (FLIR: A6751sc)
Table 1 Specification of infrared camera (FLIR: A6751sc)

Sensor type InSb
Number or pixels 640 × 512
Spectral ranges 3–5 μm
NETD 20 mK
Three-Dimensional Reconstruction of Leaked Gas Cloud … 755
(a) θ=0° (b) θ=30° (c) θ=60°
Fig. 4 Infrared images obtained by experimental equipment. θ is the angle at which the turntable
is rotated
An example of simulated gas concentration distribution is shown in Fig. 5. The gas

distribution was composed of arrangement of the gas cell with four empty bottles
and three fluorocarbon gas filled bottles as shown in Fig. 5. Infrared images were
obtained from several directions at equal interval angle. N is total number of images.
The cross-sectional two-dimensional images at the middle height of the bottle were
reconstructed from these N images using ART method. Figure 6a–c shows the recon-
structed cross-sectional image, by using infrared images taken from 4 directions at
45° angular intervals, 6 directions at 30° angular intervals, and 12 directions at 15°
angular intervals, respectively. The gray value of reconstructed cross section image
indicates the intensity of absorption by gas. The dark color means the high absorp-
tion by gas. It is found from Fig. 6d that the dark area coincides with the location
of bottles filled with gas. The distribution of gas cell can be clearly identified in the
reconstructed image using infrared images taken from 12 directions. From Fig. 3d,
the gas distribution can be recognized with sufficient accuracy even in a reconstructed
image using infrared images taken from 6 directions.
Fig. 5 Arrangement of PFA bottles

(a) N=4 (b) N=6 (c) N=12
Fig. 6 Reconstruction images of the cross-sectional 2D image. N is the number of projection

image (Infrared image). Gray value indicates the intensity of infrared absorption, and black is high
absorption intensity
5 Conclusion
The two-dimensional cross-sectional image of the gas concentration distribution

could be reconstructed by the ART method using the infrared images of multiple
optical path. It was confirmed that the presence of gas can be identified from the
number of projections of six or more paths.
Reference
1. Kudo H (2005) Iterative methods for tomographic image reconstruction: foundations and
surprizing examples. Med Imaging Technol 23(1):23–29
Experimental Analysis on Fault
Prediction of AC-Contactor by Infrared
Thermography
Yue Yu, Wei Li, Chao Ye, and Bin Hu
Abstract Aiming at the problems of poor contact, aging and dampness of elevator
AC-contactors, the infrared thermal imaging technology is proposed to detect the
faults of elevator AC-contactors. By simulating the operation of elevator, an experi-
mental platform is built to realize real-time monitoring of the operation of contactor.
Through a lot of experiments, the infrared thermal characteristic maps of contactors
under dust, loosening of terminals, oxidation of contacts and overload operation are
obtained, which can better detect the abnormal situation. At the same time, in order to
predict the remaining life of the contactor in advance, accelerated life experiment was
designed, and the experimental data were analyzed by the median iteration method
and the least square method. The reliability function of contactor under experimental
temperature is obtained by Weibull distribution model. The 90% reliability is taken as
the residual life of the contactor at this temperature, and the relationship between the
residual life of the contactor and the temperature is obtained by curve fitting. It can
be used to predict the residual life of contactors at different operating temperatures.
The experimental results can provide basis for inspection of elevator maintenance
personnel.
Keywords Contactor · Infrared thermal signature · Weibull distribution · Least

square method · Median iteration method · Residual life prediction
1 Introduction
Failures of elevator will directly endanger people’s life safety, which safety problem
is becoming more and more important. Through the analysis and statistics of elevator
fault cases, it was found that more 80% of the total faults are electrical control system
Y. Yu (B) · C. Ye · B. Hu
e-mail: yuyue_218@126.com
W. Li
School of Mechanical and Electrical Engineering, Jiangxi University of Science and Technology,
Ganzhou, Jiangxi, China

https://doi.org/10.1007/978-981-15-9199-0_74
758 Y. Yu et al.
faults, in which the AC-contactor faults accounted for the highest proportion. AC-
contactor were prone to aging, loose wiring, insulation failure and other problems
due to the influence of external environment after long time of use. If these problems
were not discovered and solved in time, they would lead to circuit breaking and
short-circuit fault, and ultimately affected the work of the entire electrical system,
thereby threatening people’s life safety. Therefore, if the abnormal condition could
be detected in advance, the elevator safety accident caused by AC-contactor failure
would be avoided.
The common life prediction methods of AC-contactors at home and abroad
includes fuzzy diagnosis method, bivariate prediction method based on overrange
time and suction time, diagnosis method based on N-Ib life curve and signal feature
extraction method. These methods mainly collect the relationship between operation
time and some parameters such as arcing energy, arcing time, contact resistance,
contact wear amount, voltage, current and take life prediction [1, 2]. The use of
above method is not very good actually. Infrared thermography test could be a more
direct method to predict the residual life with the analysis of operating temperature
of AC-contactor.
In recent years, infrared thermography detection technology has been widely used
in all walks of life due to its advantages of non-contact, no power failure and intuitive
detection results [3–5]. Mi weijian proposed to use infrared thermography to diagnose
four common faults of engine, such as water leakage, oil leakage, abnormal start and
cylinder shortage and obtained the corresponding relationship between internal faults
of the engine and surface temperature [6]. Wang li used infrared thermography to
collect temperature of airborne circuit board, proposed fault diagnosis and prediction
algorithm of airborne circuit board based on multilevel SVM, which improved fault
diagnosis accuracy [7]. Wu yingchang proposed to combine infrared thermography
and neural network to detect electrical contact thermal fault [8]. Zhang wenfeng
placed infrared sensors on the UAV to obtain infrared images of power lines, realized
automatic diagnosis and positioning of power line fault and defect location through
the algorithm process [9].
This paper proposes the use of infrared thermography to obtain contact surface
temperature, through the design of accelerated life test, bulids the reliability func-
tion of the AC-contactor with different surface temperature by the median repeat-
edly method, least square method and Weibull distribution model and acquires the
relationship between the surface temperature and acquired life.
2 Residual Life Prediction Model
The fault of AC-contactor is usually concentrated at the contact and its life is usually
evaluated by its electrical life. It is known that the electrical life of most AC-contactor
is hundreds of thousands of times or even more than a million times. But the on–off
frequency of AC-contactor in the actual usage is low in general; if calculated by this
parameter, the AC-contactor will be able to use for decades, which is not consistent
Experimental Analysis on Fault Prediction of AC-Contactor … 759
with the actual situation. In fact, most of the equipment failure rate is a function
of time, following the “bath curve,” while the Weibull distribution can completely
describe the entire “bathtub curve.” This paper adopts the two-parameter Weibull
distribution [10] to evaluate the life of the AC-contactor.
2.1 Weibull Distribution
The two-parameter Weibull distribution probability density function is as follows:

m
m t m−1 t
g(t) = exp − (1)
n n n
Cumulative distribution function:

m
t
G(t) = 1 − exp − (2)
n
Reliability function:
m
t
R(t) = exp − (3)
n
In the above equation:

t is the time from normal to complete damage of AC-contactor;
m is the shape parameter, reflecting the shape of Weibull distribution function;
n is the scale parameter, reflecting the average time between failures;
m and n are two important parameters of Weibull distribution to be calculated.
In order to calculate the Weibull cumulative distribution function of AC-contactor,
take the derivative of both sides of Eq. (2), and then take the logarithm of both sides,
as shown in Eq. (4).

1
ln ln = mlnt − mlnn (4)
1 − G(t)

1
Set ln ln 1−G(t) = y, lnt = x, the Weibull cumulative distribution function in
Eq. (4) could be y = mx − mlnn, and the parameters m and n would be got.
760 Y. Yu et al.
2.2 Median Rank Method
In the selection of Weibull distribution function value, the median rank method is
adopted. The idea is as following: arrange the fault time of the AC-contactor obtained
in the experiment from the beginning to the complete damage from the smallest to
the largest and mark the sequence number, which is the failure rank. The formula is:
i − 0.3
G(ti ) = (5)
k + 0.4
In the above equation:

i is the failure rank;
k is the total number of experimental samples;
G(ti ) is the Weibull distribution function value of AC-contactor in the rank i.
3 Experimental Design
3.1 Platform Building
The equipment to be used in the experiment mainly includes: UPC640 thermal

imaging instrument (Guangzhou Sate Infrared co., LTD), with ±2 °C or ±2% thermal
resolution and 0.65 mrad 640 × 480 spatial resolution, CJX2-12 AC-contactor
(Zhengtai co., LTD), NBE7C16 circuit breaker (Zhengtai co., LTD), DH48s-s time
relay, resistors of different powers (1000 W, 900 W, 800 W, etc.).
In order to fully capture the running state of AC-contactor, the thermal imager
needs to be in real-time monitoring state all the time. The infrared camera is placed
directly above the contactor. In this experiment, three contactors were used together
as one experiment at the same time, so as to shorten the experimental time. As shown
in Fig. 1, the experimental platform was established.
3.2 Experimental Steps
The ambient temperature in the experiment is about 25 °C, the temperature range
measured by the thermal imaging instrument is −100 to 300 °C, the shooting distance
is 0.5 m, and the dynamic range is taken from the adaptive mode to adapt to the
temperature change.
Since the contactor is affected by various factors during operation, the temperature
will vary greatly. In order to control the initial temperature of the contactor at the set
temperature, the experiment is mainly realized by changing the on and off frequency
of the contactor, the tightening force of the wiring terminal screw, and the high
Fig. 1 Experimental platform
power resistance with different powers in parallel. The experiment is divided into
two stages:
1. For the first stage, it is meaningless to study the remaining life of the contactor at
low temperature, since it is difficult to damage the contactor at low temperature.
Therefore, the temperature threshold that contactor is prone to failure should be
found through experiments.
The specific experimental method is: The contactor starting temperature is
controlled at 40 °C, and the temperature value is increased successively at 10 °C
for the experiment, and the contactor on and off times are used to judge whether
the contactor is prone to failure at this temperature. If the contactor has been
turned on and off more than 500,000 times and has not been damaged, it means
that the contactor is not prone to failure at this temperature.
2. For the second stage, three sets of temperature values were set above the temper-
ature threshold, eight contactors for each set were used for experiments, and the
running time from normal state to complete damaged state was recorded under
the temperature T.
4 Experiment and Analysis
4.1 Analysis of Heating Mechanism
Figure 2 is a model diagram of the main working part of the contactor. The
contactor generates heat due to the current effect, and the heating follows Joule’s
law Q = I 2 Rt. When the internal heat of the contactor and the external heat dissi-
pation reach thermal equilibrium, the temperature distribution will not change with
time. The main source of heat inside the contactor is the internal contact resistance
762 Y. Yu et al.
Fig. 2 Contactor working model
heating, and the contact resistance is mainly divided into shrinkage resistance and
film resistance. When the contactor is working, the upper/lower contacts will have
a shrinkage resistance due to frequent on–off, resulting in a relatively high contact
temperature. At the same time, due to long-term continuity, the contact will gradually
oxidize to form an oxide layer, resulting in serious heat.
According to field investigation and previous contactor failure analysis, the
contactor failure is mainly caused by loose terminals, contact oxidation, dust leakage
and overload operation. Figure 3 shows the infrared heat when the contactor is
(a)Terminals loose (b) Contact oxidation
(c) falling into dust (d) overload operation
Fig. 3 Thermal characteristic diagram of contactor abnormal conditions

abnormal. The characteristic map, through its infrared thermal characteristic map, can
easily find the abnormal condition of the contactor and the cause of the abnormality.
4.2 Remaining Life Prediction
4.2.1 Weibull Reliability Function
Through the first stage of experiments, it is found that the contactor will not be
damaged if the number of switching times is 500,000 when the contactor temperature
is below 90 °C, and the contactor will be damaged with 500,000 switching times when
the contactor temperature is higher then 90 °C. Therefore, the temperature threshold
at which the contactor is prone to failure is taken to be 90 °C.
According to the temperature threshold obtained in the first stage, the temperature
values of the second stage of the design are 110 °C, 130 °C and 150 °C in sequence,
and the statistical results of the experimental data are shown in Table 1. The values
of the Weibull distribution function obtained using the median rank method are as
shown in Table 2.
The data in Table 2 are fitted
using least squares method, and the functional
1
relationship between ln ln 1−G(t) and lnti is shown in Fig. 4.
The fitting equations are shown in Table 3, in which R 2 are all greater than 0.9,
indicating that the fitting curve works well.
According to the least squares fitting equation, the Weibull distribution function
of the contactor at different operating temperatures can be obtained. It can be seen
that the Weibull distribution reliability function is:
⎧ 5.963
⎪
⎪ exp − t
, T = 110 ◦ C
⎪
⎨ 594.50
6.824
R(t) = exp − 304.76
t
, T = 130 ◦ C (6)
⎪
⎪
⎪
⎩ exp − t 3.530 , T = 150 ◦ C
154.42
Table 1 Statistics of failure time of contactors

Contactor number 1 2 3 4 5 6 7 8
Operation time ti /min T = 110 °C 500 455 600 720 420 515 590 625
T = 130 °C 354 230 327 240 260 263 316 295
T = 150 °C 215 91 146 121 88 132 158 160
Failure rank number T = 110 °C 3 2 6 8 1 4 5 7
T = 130 °C 8 1 7 2 3 4 6 5
T = 150 °C 8 2 5 3 1 4 6 7
764
Table 2 Calculation results of median rank method

i k ti /min G(ti ) lnti 1
ln ln 1−G(t)
T = 110 °C T = 130 °C T = 150 °C T = 110 °C T = 130 °C T = 150 °C
1 1 420 230 88 0.0833 6.0403 5.4381 4.4773 −2.4421
2 2 455 240 91 0.2024 6.1203 5.4806 4.5109 −1.4866
3 3 500 260 121 0.3214 6.2146 5.5607 4.7958 −0.9475
4 4 515 263 132 0.4405 6.2442 5.5722 4.8828 −0.5435
5 5 600 295 146 0.5595 6.3970 5.6870 4.9836 −0.1986
6 6 590 316 158 0.6786 6.3801 5.7557 5.0626 0.1267
7 7 625 327 160 0.7976 6.4378 5.7900 5.0752 0.4684
8 8 720 354 215 0.9167 6.5793 5.8693 5.3706 0.9104
Y. Yu et al.
0 T=110 T=130 T=150
-1
-2
-3
4 4.5 5 5.5 6 6.5 7
x
Fig. 4 Least squares fitting curve
Table 3 Least squares fitting equation

Contactor temperature T /°C Fitting equations R2
110 y = 5.963x − 38.09 0.9485
130 y = 6.824x − 39.03 0.9339
150 y = 3.53x − 17.79 0.9346
Table 4 Residual life at

Contactor temperature T /°C Remaining life t/min
different temperatures
110 408
130 219
150 82
4.2.2 Remaining Life Prediction of Contactors
The article selects 90% reliability as the remaining life of the contactor when it
reaches this temperature. The calculation results are shown in Table 4.
The relationship between the remaining life of the contactor and the different
temperatures is obtained by using the least squares method again as shown in Fig. 5.
The remaining life prediction function of the contactor obtained by the fitting is:
t = −8.15T + 1296 (7)
Among them, T should be bigger than 90 °C, because the contactor temperature is
basically no damage below 90 °C. It can be obtained from the residual life prediction
formula that the contactor temperature is close to 160 °C, which is quickly damaged,
which is consistent with the actual experimental situation. Figure 6 shows the physical
map of the contactor damage.
766 Y. Yu et al.
600
400
Δ t/min
200
0
90 100 110 120 130 140 150 160
T/
Fig. 5 Residual life prediction curve
Fig. 6 Damage physical map
5 Conclusion
(1) Using infrared thermography detection technology to detect the abnormal condi-
tion of the contactor could effectively overcome the shortcomings of the tradi-
tional detection method that requires downtime and low efficiency. Through a
large number of experiments, the infrared thermal characteristics of the contactor
under the environment of dust, loose terminals, contact oxidation and overload
operation are obtained, in which the abnormal situation can better detect.
(2) Using the median-order method and the least squares method to analyze the
experimental data, the Weibull distribution model is obtained, and the reliability
function of the contactor at different temperatures is obtained, and the contactor
is used at this temperature with 90% reliability. The remaining life, and thus the
curve fitting, obtains the relationship between the remaining life of the contactor
and the temperature. The residual life prediction function can be used to predict
the remaining life of the contactor at different operating temperatures.
(3) In the prediction of the remaining life of the contactor, the number of experiments
is still too small due to time. At the same time, the contactor in this experiment is
Chint CJX2-1210. Other contactors are different in materials and construction.

There may be some errors in the experimental results. Therefore, the infrared
detection of contactors needs further research and improvement.
Acknowledgements National key research and development plan (2017YFC0805704).
References
1. Wang Z (2012) Study of electrical lifetime prediction of contactor’s contact based on wavelet
analysis. Hebei University of Technology
2. Di M, Li Z, Wu X (2004) Review on diagnostic techniques for electrical endurance of switchgear
contact. High Voltage Apparatus 40(3):201–204
3. Wang W, Fu Y, Dong F et al (2018) Research on infrared on-orbit ship target detection method.
Mod Electron Tech 41(14):92–95
4. Yu Y, Zhao H, Han L (2017) Research on forest fire identification technology based on infrared
thermal imaging. Mod Electron Tech 40(24):77–79
5. Li Y, Xiao J, Zhu L et al (2017) Applications of the infrared thermal wave technology in thermal
barrier coating thickness testing. Infrared Technol 39(7):669–674
6. Mi W, Shen Q (2016) Engine fault diagnosis based on infrared thermal imaging technology. J
Shanghai Marit Univ 37(4):65–69
7. Wang L, Li S, Wu H et al (2017) Research of aviation circuit board fault diagnosis and prediction
based on infrared technology. Comput Simul 34(07):89–93
8. Wu Y (2015) Research on fault detection of electrical contacts based on infrared thermal
imaging technology. Hunan University
9. Zhang W, Peng X (2014) Intelligent diagnostic techniques of abnormal heat defect in transmis-
sion lines based on unmanned helicopter infrared video. Power Syst Technol 38(5):1334–1338
10. Xu W, Hu W, Sun P (2013) Equipment reliability prediction based on dual-parametric Weibull
distribution. Chin J Constr Mach 11(2):112–116
Research on the Detecting Method
of Motor Running State of Amusement
Device Based on Infrared Thermography
Technology
C. Ye, Y. Yu, J. J. Zhang, R. Liu, and G. T. Shen
Abstract The motor is the source power of the amusement device. Once a failure
occurs, it will not only cause damage to itself, but also lead to high-altitude retention
and even casualties. The traditional method of monitoring the operating state by
installing sensors is costly and complicated to implement, and a large number of
existing motors cannot be installed. Therefore, there is an urgent need for quick
and convenient technical means and methods for detecting the operating state of
the motor. To address this issue, this study simulated various operating conditions
of the amusement devices and conducted comparative tests of different load rates,
different duty ratios, and different cycle durations. By comparing the temperature
rise data collected by the embedded sensor in the motor with the temperature rise
data collected by the external infrared thermography of the motor, the main factors
affecting the difference of temperature rise between the inner and outer parts of the
motor casing are obtained, and the correction value of the temperature rise difference
between the inside and outside of the casing is also obtained. A way of testing motor
running state by infrared thermography is proposed, which can be applied in site
detection with ease.
Keywords Infrared thermography · Amusement ride · Motor temperature rising
1 Introduction
In recent years, with the improvement of people’s living standards, amusement

industry as an important part of tourism and culture industry has also developed
rapidly, and its proportion in the national economy continues to increase. By the
end of 2018, there were 25,100 sets of large-scale amusement device that was used
in the country [1]. At present, the inspection of amusement device is limited by
the scale of equipment, detachability, and inspection time. The inspection project is
C. Ye · Y. Yu · J. J. Zhang · R. Liu · G. T. Shen (B)

China Special Equipment Inspection and Research Institute, Building 2, Hepingjie, Xiyuan,
Chaoyang District, Beijing, China

https://doi.org/10.1007/978-981-15-9199-0_75
770 C. Ye et al.
only for some detachable and easy-to-observe important parts, and static detection
and conventional non-destructive testing of some specific subsystems or component
indicators [2], lacking technical means to effectively grasp and evaluates the online
operation status of equipment.
The motor is the power source for the entire amusement device. Once the fault
occurs, it will not only cause damage to itself, but also cause high-altitude deten-
tion of tourists and even cause casualties. The motor of the wonder wheel of an
amusement park caused equipment failure due to overheating, causing more than ten
tourist detention. The online condition monitoring of the motor mainly monitors and
collects the current, vibration, temperature, power, and other parameters by installing
and embedding the sensor on the monitored motor in advance and then storing and
analyzing the obtained information to realize the evaluation for its running state [3].
The operating state of the motor can be grasped without interruption in real time, but
the cost is high, the implementation is complicated, and a large number of existent
motors cannot be installed. There is an urgent need for technical means and methods
for quickly and conveniently detecting the operating state of the motor.
Infrared thermography technology has been widely used in many fields.
Researchers from Tsinghua University, Xi’an Jiaotong University, Harbin Institute
of Technology, and other universities have done a lot of study on the application of
infrared thermal imaging non-destructive testing technology to the defect detection
and evaluation of material components [4]. Application guidelines have been issued
in the field of electric power to standardize and guide the state monitoring of elec-
tric power equipment by infrared thermography technology. Wang et al. [5] applied
infrared thermography technology to detect the brake shoe of elevator, enabling
quantitative evaluation of brake performance. However, there is currently no report
on the application of infrared thermography technology in detecting the operating
state of the motor.
The aim of this work is to analyze and compare the temperature rise data collected
by the embedded temperature sensor inside the motor with the temperature rise data
collected from the outside of the motor by infrared thermography and explores the
feasibility of infrared thermography technology to detect and evaluate the running
state of the amusement equipment motor.
2 Experiment Object and Equipment
The test used an AC motor (Y2-200L-4) commonly applied in amusement device,

with the insulation grade of the heat resistance used in the motor windings is F. The
rated power and rated current of the motor are 30 kW, 58 A, respectively, and the
corresponding codes of the front bearing and the rear bearing are 5312 and 6212,
respectively. The test system consists of an Infrared camera (FLIR E50) and in coop-
eration with other instruments involving paperless recorder, electronic thermometer
and hygrometer, contact thermometer, torque tachometer, eight coil pre-embedded
types, and an end cover embedded platinum resistor, etc.
Research on the Detecting Method of Motor Running State … 771
3 Experiment Schemes
The number of passengers and running cycle of amusement device changes frequently
in operation, and the motors are often operated under variable load and frequent start
and stop conditions. This test simulates different operating conditions of the motor,
including different load rates (the ratio of actual power to rated power), different duty
ratios (also known as load continuation rate, which is the percentage of working time
in cycle), different cycle durations, and different working environments. In this work,
the comparative test of the motor under different load rates, different duty ratios, and
different cycle durations was designed. Due to the high requirements of the different
working environment comparative test, no design has been made yet.
1. Comparative test of different load rates. Experimental scheme is shown in

Table 1.
2. Comparison test of different duty ratios. Experimental scheme is shown in Table
2.
3. Comparison test of different cycle durations. Experimental scheme is shown in
Table 3.
The design of the test system is shown in Fig. 1. The load motor is driven by
the test motor. The torque tachometer measures the output power of the test motor,
adjusts the load motor resistance and capacitance to change the power of the load
motor, adjusts the load motor voltage to fine-tune the load, and the switch button can
start and stop frequently for the test motor.
Table 1 Different load rates comparison test scheme

Test purposes Duty ratios (%) Operation plan Load rates (%) Time (h)
Effect of different load 100 Continuous operation 25 4
rates on motor 100 50 4
temperature rise
100 75 4
100 100 4
Table 2 Different duty ratios comparison test scheme

Test purposes Duty ratios (%) Operation plan Load rates (%) Time (h)
Effect of different duty 30 Run for 3 min and stop 75 4
ratios on motor for 7 min
temperature rise 50 Run for 5 min and stop 75 4
for 5 min
80 Run for 8 min and stop 75 4
for 2 min
772 C. Ye et al.
Table 3 Different cycle durations comparison test scheme

Test purpose Duty ratios (%) Operation plan Load rates (%) Time (h)
Effect of different cycle 50 Run for 3 min and stop 75 4
durations on motor for 3 min
temperature rise 50 Run for 5 min and stop 75 4
for 5 min
50 Run for 8 min and stop 75 4
for 8 min
Fig. 1 Schematic diagram of the test platform design
4 Experiment Preparation
4.1 Calibration for Thermal Imagers
The emissivity is the most important factor affecting the accuracy of temperature
measurement. The emissivity is calibrated before the experiment. For purpose of
adjust the emissivity, select three point on the motor casing to measure the temper-
ature by using contact thermometer and then measure the temperature by infrared
thermography at these point. After running for 10 min at full load, select the three
points on the motor casing again for calibration. The motor emissivity is finally set
to 0.95.
4.2 Thermal Resistor Embedded
The main heating components of the motor are coil winding and bearing. Therefore,
the six thermal resistors are, respectively, embedded in the three-phase copper wire
at the front and the rear of the stator coil to collect the temperature. In addition, two
holes are processed on the front and rear end covers of the motor, and one thermal
Fig. 2 Position map of embedded thermal resistance temperature measurement point
resistor is embedded in each hole. A total of eight thermal resistors are embedded in
the motor.
The location of the embedded thermal resistor temperature measurement point is
shown in Fig. 2.
4.3 Platform Construction
The thermal resistor in the three-phase coil at the front of the motor is connected to
the 1, 2, and 3 channels of the recorder. The thermal resistor in the three-phase coil
at the rear of the motor is connected to the 4, 5, and 6 channels, and the front and rear
cover thermal resistors are, respectively, connected to channels 7, 8. The recorder
sampling interval is set to 10 s. The platform is built as shown in Fig. 3.
Fig. 3 Diagram of experimental platform construction

774 C. Ye et al.
5 Data Collection
5.1 Motor Internal Temperature Rise Data Acquisition
The pre-embedded thermal resistor has high-temperature measurement accuracy, and

the signal quality is good in the comparison test. Temperature signals are collected
in the whole test during continuous operation. During intermittent operation, as the
maximum temperature rise has the greatest damage to the insulation performance of
the motor, the test records the variation trend of the maximum temperature rise per
cycle.
5.2 Motor External Temperature Rise Data Acquisition
The main heating parts outside the motor are bearing, the end cover, and the stator
casing. The image capture method is shown in Fig. 4. The first and fourth positions
mainly collect the temperature rise of the stator housing of the motor, and the second
and third positions mainly collect the temperature rise of the front bearing and the
cover. Since the rear end cover is close to the motor fan and the fan has a metal outer
casing, the thermal imager cannot take images at this position.
To accurately and comprehensively record the temperature rise process of the
motor casing, infrared thermal imaging was performed every 10 min in the continuous
operation test. During the intermittent work test, the acquisition was performed at
the last minute of each cycle. The temperature rises value of the hottest spot of the
Fig. 4 Location diagram of

infrared image acquisition
motor casing; the front end cover and the bearing are extracted from the thermal
image, respectively, and corresponding to the acquisition time.
6 Analysis of Infrared Thermography Method for Testing

Motor Running State
6.1 Analysis of Methods for Testing Temperature Rise

of Bearing
The motor bearing rotates at high speed during operation, and the temperature rise
value is large, and the rising speed is fast. Excessive temperature rise affects the
performance of the lubricating oil, causing the bearing wear to increase or even fail.
In the past, the thermal resistor method can only be pre-embedded in the end cover
close to the bearing, and the bearing temperature rise is indirectly understood by
monitoring the temperature rise of the end cover. In this work, the temperature rise
of the motor bearing and the end cover was recorded using an infrared thermography.
6.1.1 Application Analysis of Front End Bearing
Taking the 7.5 kW continuous operation test as an example, the temperature rise
process of the motor front end bearing is analyzed (see Fig. 5). It can be seen
that during the continuous operation of the motor for 4 h, the maximum temper-
ature of the front end of the bearing collected by the infrared thermography ≥ the
Fig. 5 Temperature rise curve of front end cover and bearing part in continuous operation of 7.5 kW
776 C. Ye et al.
maximum temperature of the front end cover collected by the infrared thermography
≥ the temperature of the front end cover collected by the thermal resistor is always
established.
This phenomenon reappears after all the comparative test of the motor under
different load rates, different duty ratios, and different cycle durations.
This phenomenon occurs because the front end of the motor bearing is integrally
wrapped by the front end cover, and the temperature rise at the end cover is mostly
caused by the heat generated by the rolling of the bearing. Infrared thermography
technology has the advantages of non-contact, large area, no need for pre-buried,
etc. The maximum temperature of the motor bearing front end can be directly and
accurately collected, which has obvious advantages compared with other traditional
contact temperature measurement methods.
6.1.2 Application Analysis of Bearing Back End
The rear end of the bearing is located in the back cover of the motor. The infrared
thermal imager fails to directly capture the thermal image of this part, and the temper-
ature rise also can not directly collect. The temperature rise at the rear end of the
bearing can be indirectly understood through the signal collected by pre-embedded
thermal resistor of the back end cover.
Figure 6 shows the pre-embedded thermal resistor signal diagram for the 22.5 kW
continuous operation test. It can be found that during the whole operation, the temper-
ature rise of the rear bearing is smaller than the temperature rises of the front end
bearing and the stator coil. Only after the motor stops running, the temperature signal
of the pre-embedded thermal resistor of the rear end cover will accelerate to rise (i.e.,
at the red circle mark in the figure), which will exceed the temperature of the front
Fig. 6 Diagram of pre-embedded thermal resistance signal for 22.5 kW continuous operation test
end cover. This was also confirmed during comparative test in all conditions: After
each set of tests was stopped, the back end cover had a rapid temperature rise of
about 3 °C.
The reason for this phenomenon is that since the rear end cover is connected to
the motor fan, the fan rotation makes the rear end cover the best position for the
entire motor cooling condition during operation, so the temperature rise is always
the smallest. However, when the motor stops running, the fan stops immediately. At
this time, the temperature of the rear end cover is the lowest, and the heat of other
positions of the motor is quickly transmitted to the rear end cover part, causing the
temperature to rise slightly. Since the temperature rise of the rear end of the bearing
is much smaller than the temperature rise of the front end of the bearing during the
whole operation of the motor, it only rises slightly after the stop of the operation
and then gradually cools down. Therefore, the detection of bearing temperature rise
should focus on the front end.
6.2 Analysis of Method for Inspecting Stator Temperature

Rise by Infrared Thermography
In the process of motor operation, the heat is mainly transferred to the outer casing
through the stator to dissipate heat, and excessive temperature rise will damage
the insulation performance of the motor. At present, the temperature rise of the
stator is mainly detected by the pre-embedded thermal resistor, but the operation is
complicated. In this test, the method of rapid stator temperature rise detection using
infrared thermal imaging technology is studied by comparing the corresponding
temperature rise inside and outside the casing measured by thermal resistor inside
casing and infrared thermography outside casing.
6.2.1 Comparison of Temperature Rise Inside and Outside the Casing
Figure 7 shows the temperature difference between the inside and outside of the
casing during continuous operation of the experimental motor at different load rates.
It is observed that: (1) T shows a trend of rapid increase—deceleration increase—
balance with the growth of running time. (2) T increases with the improve of the
load rate, and the variation range is large. (3) After continuous operation for 50 min,
T variation range gradually decreases.
Figure 8 shows the temperature difference between the inside and outside of
the casing during the intermittent operation of the experimental motor at different
duty ratios. It can be seen that: (1) T shows a trend of rapid increase—decelera-
tion increase—balance with the growth of running time. (2) T increases with the
improve of the duty ratio, but the increase is not large. (3) After the motor runs for
50 min, the amplitude of T changes gradually decreases.
778 C. Ye et al.
Fig. 7 Temperature difference between the inside and outside of the casing after continuous
operation of the motor at different load rates
Fig. 8 The load rate is the same, and the duty ratio is 30%, 50%, 80%, respectively,
Figure 9 shows the temperature difference between the inside and outside of
the casing during the intermittent operation of the experimental motor at different
cycle times. It can be found that: (1) T shows a trend of rapid increase—decelera-
tion increase—balance with the growth of running time. (2) T increases with the
increase of the cycle durations, but the increase is small. (3) After the motor runs for
50 min, the amplitude of T changes gradually decreases.
Fig. 9 The duty ratio and load rate are the same; the cycle duration is 6 min, 10 min, 16 min
6.2.2 Application Analysis
1. The motor load rate is the most important factor affecting the temperature differ-
ence between the inside and outside of the casing, followed by the duty ratio,
and the cycle duration is the smallest.
2. After the motor runs for 50 min, the temperature difference between the inside
and the outside tends to be stable. Therefore, the temperature rise of the motor
stator coil can be estimated by the temperature rise of the motor casing detected
by the infrared thermography and combined with the temperature rise correction
value.
Summarize the temperature rise data of the motor in the comparative working
condition test, and the temperature rise correction value inside and outside the
casing is obtained, see Table 4.
3. After the motor runs for 50 min, the temperature difference between the inside
and outside of the motor is relatively stable, but it is still in the rapid temperature
rise phase, and it has not reached the steady state. Therefore, the motor running
state cannot be judged by the result at this time. Combined with the temperature
rise curve of the running process, it is recommended to select the detection time
after 150 min of operation. At this time, the temperature rises and temperature
difference between the inside and outside of the casing are basically stable.
Table 4 Temperature rise correction value inside and outside the casing
Load rate 30/kW 22.5/kW 15/kW 7.5/kW
Temperature difference 34 22 16 10
780 C. Ye et al.
7 Conclusion
This study mainly explored the feasibility of infrared thermography technology to

test the motor running state of amusement device by simulating running experiments
in the laboratory; some important conclusions are drawn as follows:
1. Infrared thermography technology can quickly and directly obtain the temper-
ature rise of the motor bearing, which has obvious advantages compared with
other contact temperature measurement methods.
2. Infrared thermography technology can detect the temperature rise of the motor
casing. Combined with the temperature difference correction value, the temper-
ature rise of the stator and rotor inside the casing can be easily grasped, which is
convenient to use and has strong practicability.
3. A method for detecting the motor of amusement rides by infrared thermography
technology is proposed, which can realize the rapid detection of its running state.
Acknowledgements The authors gratefully acknowledge the financial support of the project of
China Special Equipment Inspection and Research Institute (grant no. 2018qingnian04).
References
1. Notice of the state administration for market regulation on the state of safety of special equipment
in the country in 2018. State administration for market regulation, 2019
2. Shen GT, Yao ZH, Wu Y (2006) Nondestructive testing of amusement ride. Nondestr Test
28(12):652–655
3. Li L (2016) Research on monitoring technology of motor online status. Electron Technol Softw
Eng 138
4. Mei L, Chen ZQ (2000) Pulse thermography analyzed by the finite element method for
nondestructive testing. J Xi’an Jiaotong Univ 34(1):66–70
5. Wang YM (2018) Study on the monitoring and evaluate method of elevator braking performance.
Southeast University, Nanjing
Application of Ultrasound Thermometry
to Condition Monitoring of Heated
Materials
I. Ihara, R. Sawada, Y. Ogawa, and Y. Kawano
Abstract Monitoring of thermal information of heated media is important for

ensuring the reliability of high temperature processes in various industries. Ultra-
sound thermometry, which is a temperature measurement method by ultrasound, is
expected as a potential candidate for such monitoring. In this work, an effective
ultrasound thermometry utilizing both pulse-echo measurements and unsteady heat
conduction analysis is applied to real-time monitoring of temperature profiles and
heat fluxes of a heated steel. This method provides non-invasive measurements of
such thermal information. To examine the feasibility of the method to condition moni-
toring, the internal temperature profile of a steel plate whose back surface is being
heated, and cooled is monitored every 70 ms. Furthermore, the ultrasonic method
is utilized to monitor heat fluxes of bonded materials of aluminum and copper, and
reasonable results on transient variation in heat flux are then obtained. Thus, it is
considered that the ultrasonic method could be applicable to condition monitoring
of various heated materials.
Keywords Ultrasonic monitoring · Thermometry · Heated material · Internal

temperature · Heat flux
1 Introduction
In various industries associated with high temperature processes, e.g., material

processing industries, chemical industries, and power plants, there are increasing
demands for monitoring thermal information of heated materials and structures, for
ensuring the reliability of the processes. In particular, it is necessary to know the quan-
titative information on transient variations of the interface and internal temperatures
of the materials and structures. This is because mechanical, electrical, and chemical
I. Ihara (B) · R. Sawada · Y. Ogawa · Y. Kawano

Department of Mechanical Engineering, Nagaoka University of Technology, Nagaoka, Niigata
940-2188, Japan
e-mail: ihara@mech.nagaokaut.ac.jp

https://doi.org/10.1007/978-981-15-9199-0_76
782 I. Ihara et al.
properties of a material are affected by its temperature and generally have a temper-
ature dependence. Therefore, an on-line monitoring technique to obtain temperature
information of materials during high temperature processing is quite attractive for
ensuring the reliability of the processes. Such monitoring technique is also useful
for making an practical process control in materials fabrications. Thus, temperature
measurements are an important issue in materials engineering and industries.
Although there exist some conventional methods for temperature measurements,
such as thermocouples or infrared radiation methods, they are not always acceptable
for the temperature monitoring mentioned above. Thermocouple techniques are not
suitable for measuring temperatures inside material in principle because it is required
to make a hole for inserting the sensors into the material. Moreover, the response time
in measuring temperature using a thermocouple is relatively slow because of a certain
amount of thermal capacity of the thermocouple itself. An infrared radiation method
provides noncontact temperature measurements, and it is useful for real-time obser-
vations. However, temperature observations with this method are adversely affected
by the reflection and emissivity of infrared radiation from any heat sources. Crucial
disadvantage of this method is that its application is limited to surface temperature
measurements only.
To overcome the problems with the conventional methods, ultrasound is consid-
ered as an alternative method. Since the ultrasonic velocity in a medium is sensitive
to temperature, ultrasound has the potential to be a useful tool for measuring temper-
atures. Since ultrasonic measurements have advantages such as nondestructive and
internal permeability, several studies on ultrasound thermometry which is a tempera-
ture measurement method based on ultrasound were conducted [1–5]. In our former
works, some techniques with effective ultrasound thermometry were developed and
applied to several heated materials [6–15]. The method provides surface or internal
temperature profiling by measuring ultrasound waves propagating through a mate-
rial in pulse-echo or pitch-catch configuration. Thus, the ultrasound method could be
a potential method applicable to condition monitoring of various heated materials,
mechanical parts, and structures associated with internal heating. Figure 1 shows
a typical concept of the application of ultrasound thermometry to such condition
monitoring, where the ultrasonic wave in the heated material is employed to obtain
the internal temperatures of the material. The quantitative and transient information
reflecting the thermal state of the structure are then utilized for controlling the heating
process. It should be noted here that the thermal condition at the inner surface of the
heated material is generally unknown and unexpectable. In this work, to examine the
practicability of the ultrasound thermometry to condition monitoring, the transient
variations of internal temperature and heat flux in a steel plate are monitored during
heating and cooling.
Application of Ultrasound Thermometry to Condition … 783
Ultrasonic wave reflected

from inner surface.
Ultrasonic transducer
Data acquisition
&
Analysis
Heated material
Structure with having temperature gradient
internal heating
Quantitative and transient
Process control
information on thermal state
Fig. 1 Concept of the application of ultrasound thermometry to condition monitoring of an inner

heated structure
2 Determining Temperature Profile and Heat Flux

by Ultrasound
It is known that the ultrasonic velocity in a medium changes with temperature.

Therefore, if the temperature in the medium is uniform, then the temperature can
simply be determined from the ultrasonic velocity in the medium as long as the
temperature dependence of the velocity is known. In this work, we consider a material
having a temperature distribution due to single side heating as shown in Fig. 2. When
ultrasonic wave in the material is measured by pulse-echo method, the traveling time
of the ultrasound, t L , in the temperature distribution direction is given by [6, 12]
L
1
tL = dx (1)
v(T (x))
0
where L is the traveling distance of ultrasonic wave; v(T (x)) is the ultrasound velocity
that is a function of the temperature distribution T (x). It should be noted here that our
purpose is to obtain the distribution T (x) from the traveling time t L . Unfortunately,
even if the time t L is precisely measured, based on the relationship of Eq. (1), it is
almost impossible to uniquely determine the T (x) from the traveling time t L . This is
because the solution is not unique in this inverse problem. In other words, the T (x)
cannot be determined uniquely from Eq. (1) because of lack of boundary conditions.
In order to solve the problem, an effective method utilizing ultrasonic pulse-echo
measurements and a finite differential calculation for unsteady heat conduction anal-
ysis were developed [6, 12]. The schematic of such ultrasound thermometry is shown
in Fig. 2 where T i n denotes the temperatures of the heated material at a certain time.
In this method, the temperatures (i = 2, …, N − 1) inside material are obtained from
finite differential calculations, the temperature at unheated surface (i = N) is a known
value, and the heating surface temperature (i = 1) is then determined from both the
784 I. Ihara et al.
Ultrasonic pulse wave
Ultrasonic transducer
Heating
Heat flux q
Temperature T
Temperature profile T(x)
Distance from heating surface x

Location
i=1 2 䞉䞉䞉䞉䞉䞉 N-1 N
n=1 Initial condition
Time step
2
τ
3
h
Fig. 2 Schematic of a material having a temperature gradient due to single side heating and a finite
differential model of unsteady heat conduction analysis
ultrasonic propagation time t L and the previously determined internal temperatures

[6, 12]. In the procedure for determining the temperatures, it is often assumed that
the ultrasonic velocity varies with the temperature proportionally. Actually, such
linear relationship between them is confirmed experimentally [6–8]. It is also noted
that the temperatures at unheated surface are obtained using any common technique
like a thermocouple because there is no difficulty in measuring it. Thus, the heating
surface and internal temperatures at a certain moment can be determined when the
propagation time t L is measured [6, 12].
Heat flux entering from the heating surface can be calculated from the temperature
gradient at the surface. The heat flux in the x direction, q, is given by Fourier’s law
as follows [13, 15]

dT λ 2 T1n+1 − T1n
q = −λ = h + T1 − T2
n n
(2)
dx h 2ατ
where λ is thermal conductivity, dT /dx is temperature gradient at the heating surface,

α is thermal diffusivity, τ is the time step, h is the interval between grid points in
the finite differential analysis, and T 1 and T 2 are temperatures at the heating surface
and its subsurface, respectively. Because we have already obtained the temperature
profile inside the material, the heat flux q can be determined from Eq. (2). Thus, the
heat flux as well as the temperature profile is being determined quantitatively as long
as the ultrasonic measurements are being conducted.
3 Experiment and Result
3.1 Temperature Profiling
Figure 3 shows a schematic of the experimental configuration for ultrasonic moni-

toring of a heated material. Specimen used here is a 30 mm thick steel plate. For
comparing with ultrasonically measured temperatures, thermocouples B are inserted
into the plate at different locations from the heating surface. The location of the
probing end of each thermocouple is 15 mm from the side surface. Thermocouple
A is used for measuring surface temperature that is an important data for deter-
mining temperature profile by the ultrasound thermometry. Ultrasonic pulse-echo
measurements are conducted using a longitudinal wave transducer of 5 MHz, with
100 MHz sampling rate for signal acquisition. The ultrasound thermometry system
using a conventional ultrasonic pulser/receiver provides the time resolution of 15 ms
and spatial resolution 0.5 mm in temperature profiling. It is noted in this experiment
that the time resolution τ and spatial resolution h were 70 ms and 1.1 mm, respec-
tively. This is because it took a certain time to average the fifty acquired waveforms
to improve the signal-to-noise ratio. In the experiment, the pulse echo measure-
ments are being conducted while the one side of the plate is directly heated by a
gas burner and subsequently being cooled by an air coolant. The first and second
echoes reflected from the heating side were successfully measured. The propagation
time of ultrasound, t L , is estimated by taking a cross-correlation of the two echoes
and is employed for determining temperature profile. In the analysis, the following
linear relation of the temperature dependence, v(T ) = −0.5350T + 6047.6 (m/s), is
Ultrasonic
transducer
PC
30 mm
A/D board
Steel
Heating
or Ultrasonic
pulser/receiver
Cooling
Thermocouple A
Amp.
Thermocouples B
Fig. 3 Schematic of the experiment configuration for ultrasonic measurements for providing
temperature profiling
786 I. Ihara et al.
Ultrasound Ultrasound Ultrasound

Thermocouple Thermocouple Thermocouple
Heating surface Non heating surface

(a) Before heating (b) 1 sec after heating has started (c) 3 sec after heating has started

(d) 16 sec after heating has started (e) 1 sec after heating has stopped (f) Immediately after forced air cooling

(g) 15 sec with forced air cooling (h) During natural air cooling (i) During reheating by a gas burner
Fig. 4 Monitoring results (snapshots) showing the variation of temperature profile in the steel
during single side heating and cooling
employed, where T is the temperature given in degrees Celsius. Such temperature

dependence is experimentally obtained in the range from −30 to 100 °C. The value
of thermal diffusivity α used here is 7.53 × 10–6 m2 /s.
Figure 4 shows the variation in the temperature profile during heating and cooling.
We can see that the internal temperature profile is markedly and quickly changed by
the single side heating and cooling. The temperature profiles determined by ultra-
sound and using thermocouple almost agree each other. It is interesting that there
are certain discrepancies between the results by the ultrasound and thermocouple in
the beginning of heating or cooling (Fig. 4b, e, f). It is reasonable to consider that
the discrepancy is caused by a relatively slow time response of the thermocouples.
In other word, the ultrasonic method may give proper results due to its faster time
response in measuring temperature. It is also noted that the heating surface temper-
ature can be obtained in spite of no use of any thermal boundary condition at the
heating surface. This will be a great advantage for practical applications in industries
because such thermal boundary condition is normally difficult to obtain.
3.2 Heat Flux Monitoring
Based on the feasibility study on heat flux measurement [13], an attempt to monitor
the heat fluxes entering to heated materials is made. The schematic of the experi-
ment configuration for such heat flux monitoring is shown in Fig. 5. This setup is
basically similar to that used for temperature monitoring shown in Fig. 3. The same
ultrasonic measurement system as shown in Fig. 3 is used for this experiment. In
this experiment, bonded materials consisting of aluminum (Al) and copper (Cu) are
employed for the specimen as shown in Fig. 5. The top surface of Cu is heated by
a gas burner, and consequently, Al is also gradually heated due to heat conduction
and transfer. Ultrasound measurements in pulse-echo configuration are conducted
for the bonded materials during the heating, and the internal temperature profiles in
both Al and Cu are then determined by the same procedure used in 3.1. Two kinds
of heat fluxes, qCu entering from the heating surface of Cu and qAl entering to Al
through the interface between Cu and Al, are then determined from Eq. (2). Thermal
conductivities λ used in the calculation are 237 W/m K for Al and 398 W/m K for
Cu, respectively. The variations of heat fluxes with the elapsed time are shown in
Fig. 6. Although the measured heat fluxes have a noisy deviation, it seems that the
obtained results show a reasonable variation of each heat flux corresponding to the
given heating. In particular, it is interesting that a sudden rise in the heat flux entering
to Cu occurs immediately after heating. Although further investigations are neces-
sary to understand such unclear result, it is considered that the present ultrasound
method could also be available for non-invasive monitoring of heat flux.
Heating
PC
Cu
qCu
28 mm A/D board
Al
15 mm
qAl
Thermocouple
Ultrasonic
transducer Ultrasonic
pulser/receiver
Fig. 5 Schematic of the experiment configuration for heat flux monitoring

788 I. Ihara et al.
Heating
500
Heat flux (kW/m2)

400
300 Cu
200
100
Al
0
-100
0 20 40 60
Elapsed time (s)
Fig. 6 Variations in heat fluxes entering to the Cu and Al during heating
4 Conclusions
A feasibility of the ultrasound thermometry to condition monitoring of heated mate-

rials has been investigated. It is demonstrated through the experiments that the present
ultrasonic method provides an effective monitoring of temperature profiles for a
single-side heated material. In addition, the method is also available for monitoring
heat flux at the interface of bonded materials. This method is available for any mate-
rial as long as the temperature dependence of the ultrasonic velocity for the objective
material is known. Thus, the ultrasound thermometry is expected to be a promising
technique applicable to condition monitoring of various heated materials, mechanical
parts, and structures associated with high temperature processes and environments.
Acknowledgements This work is supported by Grant-in-Aid for Scientific Research

(B)(19H02477) from Japan Society for the Promotion of Science.
References
1. Simon C et al (1998) Two-dimensional temperature estimation using diagnostic ultrasound.

IEEE Trans Ultrason Ferroelectr Freq Control 45(4):1088–1099
2. Mizutani K et al (1999) Acoustic measurement of temperature distribution in a room using a
small number of transducers. Jpn J Appl Phys 38:3131–3134
3. Huang KN et al (2002) High precision fast ultrasonic thermometer based on measurement of
the speed of sound in air. Rev Sci Instrum 73(11):4022–4027
4. Gulik G et al (2002) Measurement of 2D-temperature distributions in a pervaporation

membrane module using ultrasonic computer tomography and comparison with computational
fluid dynamics calculations. J Membr Sci 204:111–124
5. Wang S, Harada J, Uda S (2003) A wireless SAW temperature sensor using langasite as substrate
material for high temperature applications. Jpn J Appl Phys 42:6124–6127
6. Takahashi M, Ihara I (2008) Ultrasonic determination of temperature distribution in thick plates
during single sided heating. Mod Phys Lett B 22(11):971–976
7. Takahashi M, Ihara I (2008) Ultrasonic monitoring of internal temperature distribution in a
heated material. Jpn J Appl Phys 47(5):3894–3898
8. Ihara I, Tomomatsu T (2011) In-situ measurement of internal temperature distribution of
sintered materials using ultrasonic technique. IOP Conf Ser Mater Sci Eng 18:022008
9. Yamada H, Kosugi A, Ihara I (2011) Non-contact monitoring of surface temperature distribution
by laser ultrasound scanning. Jpn J Appl Phys 50:07HC06-1-5
10. Ihara I et al (2011) New ultrasonic thermometry and its applications to temperature profiling
of heated materials. In: Proceedings of the 5th international conference on sensing technology,
Palmerston North, New Zealand, pp 65–70
11. Kosugi A, Ihara I, Matsuya I (2012) Accuracy evaluation of surface temperature profiling by
a laser ultrasonic method. Jpn J Appl Phys 51:07GB01-1-8
12. Ihara I et al (2013) Ultrasonic thermometry for temperature profiling of heated materials. In:
Mukhopadhyay SC et al (eds) Advancement in sensing technology, new developments and
practical applications. Springer, Berlin, pp 211–236
13. Ihara I et al (2015) Simultaneous measurements of temperature and heat flux using ultrasound.
In: Proceedings of the 2015 ninth international conference on sensing technology, Auckland,
New Zealand, pp 254–257
14. Aoki S, Ihara I (2015) Feasibility study on ultrasonic in-situ measurement of friction surface
temperature. Mech Eng J 2(1):14–00431
15. Myers GE (1971) Analytical methods in conduction heat transfer. McGraw-Hill, New York, p
262
Gas Leak Source Identification
by Inverse Problem Analysis of Infrared
Measurement Data
H. Nishimura, D. Shiozawa, T. Sakagami, and S. Kubo
Abstract With the retirement of skilled workers and the aging of gas plant facili-
ties, interest in the development of automated and remote gas detection methods is
increasing. Therefore, a gas leak detection method using infrared camera has been
proposed. Invisible gas can be detected from infrared absorption of gas and the
emitted infrared ray from the gas itself. The piping of a gas plant is very compli-
cated, so it is very useful that the gas leak source can be identified. In this study, the
gas leak source identification is performed by the inverse problem analysis using the
infrared measurement data. The least residual method is one of the typical inverse
problem analysis methods, and it estimates the position of the gas leak source from
the sequential infrared images of the gas cloud distribution and flow. It was found
from numerical simulation and laboratory experiments that it is possible to estimate
the position of the gas leak source even when the leak source is hidden by piping in
the infrared image.
Keywords Infrared thermography · Gas leak source identification · Inverse

problem
1 Introduction
With the retirement of skilled workers and the aging of gas plant facilities, interest
in the development of automated and remote gas detection methods is increasing.
Therefore, a gas leak detection method using infrared camera has been proposed.
Invisible gas can be detected from infrared images because infrared light is absorbed
by gas, and the infrared ray is emitted from the gas itself [1]. The piping of a gas
plant is very complicated, so it is useful that the gas leak source can be identified.
H. Nishimura · D. Shiozawa (B) · T. Sakagami · S. Kubo

Department of Mechanical Engineering, Kobe University, 1-1 Rokkodai, Nada-ku, Kobe, Japan
e-mail: shiozawa@mech.kobe-u.ac.jp
S. Kubo
Osaka University, Osaka, Japan

https://doi.org/10.1007/978-981-15-9199-0_77
792 H. Nishimura et al.
The purpose of this study is to estimate the position of the gas leak source using
inverse problem analysis from the movement of gas cloud captured by an infrared
camera, even when the gas leak source is hidden by piping. In this paper, the gas
leakage source is identified from the experimental data obtained in laboratory using
the least residual-based inverse problem analysis. The applicability of this method
is discussed for the situation when the gas leakage source is hidden by the piping.
2 Identification of Gas Leakage Source by Infrared Image

and Inverse Analysis
When gas is observed with an infrared camera, the infrared ray coming from the
background of the gas is partly absorbed by the gas, and the intensity of the infrared
rays is weakened. In addition, infrared rays due to the temperature of the gas itself
are observed. The gas is not visible on the optical image, but it can be visualized
in the infrared image. The gas absorption characteristics depend on the spectrum
absorption of the target gas. In order to emphasize the absorption by the target gas,
a band-pass filter according to the spectrum absorption of the target gas is used.
The least residual method is used as an inverse problem method to identify the gas
leak source. In this method, the square sum of residuals Rs between the measurement
data and the computed data assuming the parameters to be identified is obtained. The
parameter that minimizes Rs is taken as the estimated one. In this study, time series
sequential infrared images of the gas cloud, which was described by the infrared
intensity value, are used. Rs defined by the following equation is calculated;

L
M
N
2
Rs = Pimjt − Picjt (x G , yG ) (1)
t i j
The subscript t is the frame number in the total frame number of L. The subscript
i and j are the x and y coordinates of the pixels in the M × N image, respectively. x G ,
and yG are the assumed position of gas leak source. Pm is the measured data of the
infrared intensity value, and Pc is the simulated data where the position parameter
of the gas leak source is assumed.
The laboratory experiment was conducted as shown in Fig. 1. The gas absorption
of infrared radiation from the background was measured. The gas leak source was
set in front of a pseudo black body furnace, whose temperature was kept constant
at 303.15 K. The type of gas was methane, the gas flow rate was 500 mL/min, the
direction of gas flow was upward, and there is no wind in the laboratory.
Gas Leak Source Identification by Inverse Problem Analysis … 793
Fig. 1 Schematic illustration of measurement and numerical simulation set up of leaked gas cloud
The measured infrared image of gas leakage is shown in Fig. 2. The simulation
data computed under the same conditions as in Fig. 1 is shown in Fig. 3. The gray
value in Figs. 2 and 3 indicates the absorption intensity of infrared ray by gas. The gas
cloud and its flow can be measured by the infrared camera with a spectrum band-pass
filter. The calculated gas cloud is in good agreement with the measured one.
At first, the influence of the time fluctuation of the gas cloud on the evaluation of the
Rs in the identification of the gas leak source was investigated. Rs(T ) was calculated
by changing the timing of the evaluation frame section of the measurement data, as
following;

L+T M
N
2
Rs(T ) = Pi jtm − Pi jtc x̂ G , ŷG (2)
t+T i j
Fig. 2 Infrared image of the leaked gas

794 H. Nishimura et al.
Fig. 3 Calculated infrared of the leaked gas
(x̂ G , ŷG ) is the actual location of gas leak source. T is the starting point of the
evaluation frame. When the total number of frames L is 30 and frame rate is 10
frame/s, the coefficient of variation of Rs was 0.0133, and the mean of Rs is 2.89
× 1011 . It was found that the fluctuation of gas cloud with respect to time had little
effect on Rs when there was no wind.
Figure 4 shows the result of identifying the gas leakage source for the measurement
data shown in Fig. 1. The gas leak source is located at (x, y) = (0, 0). The color dot
in the figure means the location assuming a leakage source, and the value of Rs
is obtained from the simulation result of the assumed gas leak source. Rs shows
small value at a position close to the actual gas leak source, and it is found that the
approximate gas leak source can be estimated.
Figure 5 shows the result of the identification of the leak source when the gas leak
source is hidden by the piping. The identification of gas leak source was conducted
Fig. 4 Estimated results of gas leak source (the distribution of Rs between measured and calculated
infrared image)
Gas Leak Source Identification by Inverse Problem Analysis … 795
Fig. 5 Estimated results of gas leak source hidden by the piping (the distribution of Rs between
measured and calculated infrared image)
using the measurement data where the gas cloud is hidden in the area surrounded by
the white solid line, as shown in Fig. 5. From the figure, Rs is small at a position
next to the piping, but the estimated position in the y direction agrees well with the
actual leak source position. The approximate location of the gas leak source can be
estimated even if the gas cloud near the leak source is hidden by piping.
4 Conclusions
The infrared measurement for gas cloud and inverse analysis was applied to the
identification of the gas leak source. The measurement data is the time series data of
infrared image, and the least residual method is used as an inverse analysis technique.
It was found that gas leak source can be identified from experimental data when there
was little wind or wind of constant speed.
Reference
1. Kasai N, Tsuchiya C, Fukuda T, Sekine K, Sano T, Takehana T (2011) Propane gas leak detection
by infrared absorption using carbon infrared emitter and infrared camera. NDT&E Int 44(1):57–
60

2021 Book AdvancesInConditionMonitoringA

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2021 Book AdvancesInConditionMonitoringA

Uploaded by

Copyright:

Available Formats

Lecture Notes in Mechanical Engineering

More information about this series at http://www.springer.com/series/11236

Andrew A. Malcolm Chin Kian (Edmund) Liew

ISSN 2195-4356 ISSN 2195-4364 (electronic)

Signal Processing: Fault Diagnosis for Complex System Monitoring

Measuring the Effectiveness of Descriptors to Indicate Faults

Recent Advances on Anomaly Detection for Industrial Applications

Classiﬁcation/Analysis for Fault or Wear Diagnosis

Fault Diagnosis of Single-Stage Bevel Gearbox by Energy Operator

System Model for Condition Monitoring and Prognostic

Recent Technologies for Smart and Safe Systems

Condition Monitoring of Electromechanical Systems

SHM/NDT Technologies and Applications

Guided Waves Testing

Experimental Analysis on Fault Prediction of AC-Contactor

Len Gelman, Ph.D., Dr. of Sciences (Habilitation) joined Huddersﬁeld University

Nadine Martin is a senior researcher at the CNRS, Nat. Centre of Scientiﬁc

X. Laval, N. Martin, P. Bellemain, and C. Mailhes

Abstract In the context of automatic and preventive condition monitoring of rotating

Keywords Condition monitoring · Signal processing · Hilbert demodulation ·

Gearboxes are very important parts of mechanical systems. An unforeseen failure

X. Laval (B) · N. Martin · P. Bellemain

© Springer Nature Singapore Pte Ltd. 2021 3

mode decomposition [4], Teager-Kaiser energy operator [5]. A common method is

2 Demodulation Process in AStrion and Physical Meaning

3 Data from the GOTIX Experimental Bench

The GOTIX bench, hosted in the GIPSA-lab, is an experimental setup consisting in a

4 Monitoring of the Modulation Functions

Fig. 4 AM for H1 and H3 for sensors 1 and 2 from 2 different accelerometers

5 Demodulation Function Features

the automatic classification, a gliding feature is computed on each function with a

6 Limitations of the Method

needs to be taken into account in the context of automatic monitoring, by a feed of

7 Conclusions and Perspectives

8. Firla M et al (2016) Automatic characteristic frequency association and all-sidebands demod-

C. Peeters, J. Antoni, and J. Helsen

Keywords Vibrations · Instantaneous speed estimation · Tacholess · Harmonics ·

C. Peeters (B) · J. Helsen

© Springer Nature Singapore Pte Ltd. 2021 15

knowledge of the rotation speed is an integral part of almost every vibration-based

2.1 Multi-harmonic Demodulation (MHD)

xk (n) = Ak (n)e j(αk θ(n)+φk ) + νk (n) (2)

ẋk (n) = βk (n)xk (n) (4)

where βk (n) is equal to:

ẋk (n) = βk (n)xk (n) + νk (n) (6)

where {.} denotes the imaginary part.

2.2 Maximum Likelihood Estimation of Harmonic Weighting

Equation 8 gives a uniform weighting to every harmonic. In an attempt to make

Solving the derivative for θ̇ˆ gives, after some algebra:

Deriving the maximum log-likelihood gives:

2.3 Practical Considerations

Fig. 1 Four examples of the

convergence or the maximum number of iterations has been reached.

In order to highlight the influence of the maximum likelihood weighting, a simple

Fig. 2 Spectrogram of the simulated signal

Fig. 5 Spectrogram of the aircraft engine data

Y. Marnissi, D. Abboud, and M. Elbadaoui

Abstract Order tracking is important in rotating machine signal processing. It

Keywords Order tracking · Vold-Kalman filter · Bayesian framework ·

In rotating machines, each mechanical source generates a unique pattern as the

Y. Marnissi · D. Abboud (B) · M. Elbadaoui

© Springer Nature Singapore Pte Ltd. 2021 25

as order tracking, is of high importance in rotating machine signal processing. On

2 A Brief Review of the Vold-Kalman Filter

where C = [C1 , . . . , C K ], R is the diagonal matrix formed with r1 , . . . , r K and is

where . denotes the 2 -norm and C t e is a constant independent of r k , τ and

∀r ∈ qr (r ) ∝ exp E qi (i )

It should be noticed that m = 7 (weekly seasonality), and q(≤ m/2)

where ψ refers to the Digamma function defined by ψ(a) = ∂ log (a)/∂a =