Smart Sensors Measurement and Instrumentation: Shreesha Chokkadi Rajib Bandyopadhyay Editors

Lecture Notes in Electrical Engineering 957
Shreesha Chokkadi
Rajib Bandyopadhyay Editors
Smart Sensors
Measurement and
Instrumentation
Select Proceedings of CISCON 2021
Lecture Notes in Electrical Engineering
Volume 957
Series Editors
Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli
Federico II, Naples, Italy
Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán,
Mexico
Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India
Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany
Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China
Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore,
Singapore, Singapore
Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology,
Karlsruhe, Germany
Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China
Gianluigi Ferrari, Università di Parma, Parma, Italy
Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid,
Madrid, Spain
Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität
München, Munich, Germany
Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA,
USA
Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt
Torsten Kroeger, Stanford University, Stanford, CA, USA
Yong Li, Hunan University, Changsha, Hunan, China
Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA
Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra,
Barcelona, Spain
Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore
Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany
Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA
Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany
Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University,
Palmerston North, Manawatu-Wanganui, New Zealand
Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA
Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan
Luca Oneto, Department of Informatics, BioEngineering, Robotics and Systems Engineering,
University of Genova, Genova, Genova, Italy
Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy
Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University,
Singapore, Singapore
Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany
Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal
Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China
Walter Zamboni, DIEM—Università degli studi di Salerno, Fisciano, Salerno, Italy
Junjie James Zhang, Charlotte, NC, USA
The book series Lecture Notes in Electrical Engineering (LNEE) publishes the latest
developments in Electrical Engineering—quickly, informally and in high quality.
While original research reported in proceedings and monographs has traditionally
formed the core of LNEE, we also encourage authors to submit books devoted
to supporting student education and professional training in the various fields and
applications areas of electrical engineering. The series cover classical and emerging
topics concerning:
• Communication Engineering, Information Theory and Networks
• Electronics Engineering and Microelectronics
• Signal, Image and Speech Processing
• Wireless and Mobile Communication
• Circuits and Systems
• Energy Systems, Power Electronics and Electrical Machines
• Electro-optical Engineering
• Instrumentation Engineering
• Avionics Engineering
• Control Systems
• Internet-of-Things and Cybersecurity
• Biomedical Devices, MEMS and NEMS
For general information about this book series, comments or suggestions, please
contact leontina.dicecco@springer.com.
To submit a proposal or request further information, please contact the Publishing
Editor in your country:
China
Jasmine Dou, Editor (jasmine.dou@springer.com)
India, Japan, Rest of Asia
Swati Meherishi, Editorial Director (Swati.Meherishi@springer.com)
Southeast Asia, Australia, New Zealand
Ramesh Nath Premnath, Editor (ramesh.premnath@springernature.com)
USA, Canada
Michael Luby, Senior Editor (michael.luby@springer.com)
All other Countries
Leontina Di Cecco, Senior Editor (leontina.dicecco@springer.com)
** This series is indexed by EI Compendex and Scopus databases. **
Shreesha Chokkadi · Rajib Bandyopadhyay
Editors
Smart Sensors Measurement

and Instrumentation
Select Proceedings of CISCON 2021
Editors
Shreesha Chokkadi Rajib Bandyopadhyay
Department of Instrumentation and Control Department of Instrumentation
Engineering and Electronics Engineering
Manipal Institute of Technology Jadavpur University
Manipal Academy of Higher Education Kolkata, India
Manipal, India
ISSN 1876-1100 ISSN 1876-1119 (electronic)

Lecture Notes in Electrical Engineering
ISBN 978-981-19-6912-6 ISBN 978-981-19-6913-3 (eBook)
https://doi.org/10.1007/978-981-19-6913-3
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Singapore Pte Ltd. 2023
This work is subject to copyright. All rights are solely and exclusively licensed 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
Preface by Prof. Rajib Bandyopadhyay
The Control Instrumentation System Conference (CISCON) 2021 was held during
26–27 November 2021. This conference was for the eighteenth time organized by
the Department of Instrumentation and Control Engineering Department, Manipal
Institute of Technology. Due to the pandemic, it was organized in virtual mode. This
year it was arranged in partnership with the International Society of Automation
(ISA), Education and Research Division and Bangalore Section and with finan-
cial support from the Board of Nuclear studies and the Institution of Engineers,
Mangalore. Around 120 manuscripts were received, and the acceptance rate was
25%. Four plenary lectures were delivered. While on the first day, Prof. Soumyo
Mukherji of IIT Bombay and Prof. Rajib Bandyopadhyay the undersigned delivered
the keynote addresses, Prof. Russel Rhinehart of the Okalama State University, USA,
Dr. Jayesh Barve, Principal, General Electric, Bengaluru, and Prof. Karabi Biswas of
IIT, Kharagpur, were invited to present the keynote addresses on the second day. There
were four panel discussions on contemporary topics participated by eminent industry
experts, academicians, and doctors. There were thirty-four (34) oral presentations.
The papers were categorized into four groups—(a) Signals and Computational Tech-
niques, (b) Automation, Robotics and Machines, (c) Electronics and Instrumentation
and (d) Control Systems and were presented in three sessions. The proceedings of
CISCON 2021 is an electronic publication for the accepted and presented papers at
the conference. All the papers submitted went through a double-blind review process
by two or three reviewers prior to being accepted to the conference. On the whole, the
eProceedings presents a comprehensive overview of ongoing research in the field of
instrumentation, control and automation in India and abroad and the current interests
and research directions of the instrumentation and control community. I hope that
we will meet at the next CISCON conference in 2022 in physical mode.
Kolkata, India Prof. Rajib Bandyopadhyay
v
Preface by Prof. Shreesha Chokkadi
I am very glad that CISCON 2021 held during 26–27 November 2021, XVIII in the
series, is a grand success. Started in 2004 when the first batch of Instrumentation and
Control Engineering undergraduates of MIT Manipal are in the verge of entering final
year of engineering, with an objective to give them an opportunity to explore about
trends in instrumentation, control and systems engineering, this annual event has
grown year after year with support from MAHE management, industry and research
organizations of the country like ISRO, DRDO, VSSC, BRNS, DST, CSIR, etc.
Since 2020 due to the pandemic, we conducted this event as a virtual conference.
In spite of this, the response from researchers across the country is tremendous, and
this year, we also had an international contribution as well as invited talk. With the
proceedings of the presentation being published as Scopus-indexed Lecture Notes in
Electrical Engineering by Springer Nature, the value of the presentations has been
enhanced. The quality of the conference can be measured with the fact that only 34
of the submitted 128 articles have been accepted for presentation.
This years’ conference had representation from both academia with profes-
sors from reputed IITB, IITKGP, NIT Jadavpur, and experts from industry like
GE presenting invited talks. A feather on cap of this version of CISCON is two
debates where the panellists were industry experts from Honeywell, GE and Mylabs
Discovery Solutions along with experts from academia like IITB, NITJ and MIT
Manipal, wherein deliberations on contemporary issues like sensor design for
biomedical applications and relevance of AIML in Electronics and Instrumentation
curriculum were very critically introspected.
Overall, all deliberations and presentations are very informative and provided very
good input to all participants especially to our faculty and students. I congratulate
and acknowledge all participants, resource persons, conveners for taking out their
vii
viii Preface by Prof. Shreesha Chokkadi
time and expertize to organize and make the event a very successful one. I wish this
annual event of Department of ICE, MIT Manipal, would leave up to the expectations
of all key stakeholders and grow as a sought-after multidisciplinary conference in
contemporary domains.
Manipal, India Prof. Shreesha Chokkadi

About the Conference
Control Instrumentation System (CISCON) is the annual conference event organized

by the Department of Instrumentation and Control Engineering, Manipal Institute of
Technology. The department initiated CISCON in the year 2004 to provide a plat-
form for its first batch of B.E. in Instrumentation and Control Engineering students
to have interaction and exchange of ideas with their counterparts in and outside
the institution. This is the first of its kind in the institute and under the able lead-
ership of Dr. V. I. George. With very few institutes in the country offering this
specialized interdisciplinary course, people working in both Instrumentation and
Control Engineering sought after for this conference every year and have gained
lots of recognition. The conference has been sponsored by national research orga-
nizations like Defence Research and Development Organization (DRDO), Board
of Research in Nuclear Sciences (BRNS), Indian Space Research Organization
(ISRO), and Council of Scientific and Industrial Research (CSIR) to name a few.
The proceedings of CISCON has been brought out regularly since its inception.
In 2015, it was decided to bring out the published papers in Scopus-indexed jour-
nals to give additional incentive to authors who put forward their research articles
to CISCON, and the same trend has continued till 2017 with the rapid increase in
submission. Later, presented papers were published in Lecture Notes in Electrical
Engineering published by Springer Nature. The conference has attracted a large
number of papers in varied disciplines like process control, automation, renewable
energy, robotics, image processing, sensor and instrumentation, etc. Out of the total
120 papers submitted, 85 papers were sent for double-blind review after preliminary
inspection and plagiarism check. Out of these, 36 papers have been accepted and
presented in the conference and would be considered for publication in this book
as chapters. We believe that the proceedings of the conference will be well received
by researchers working in the domain and get inspiration for budding researchers to
explore more into the varied domains in which the papers are presented. The papers
presented in this proceedings are mainly in the domain of process control automa-
tion, instrumentation, robotics, image processing and many more. The readers of this
proceedings will get an insight into the varied areas in which contemporary research
is being carried forward in this domain and get started to go ahead. These papers will
ix
x About the Conference
give openings for beginners and also the direction for those who are working in these
specific domains already. We are confident that the proceedings will be accepted by
prospective researchers very well and give encouragement for us to go ahead with
organizing CISCON every year with many new ideas and scope. This event was made
possible by the utmost support from Chancellor of MAHE Padmashree Awardee Dr.
Ramadas M. Pai, Pro-Chancellor Dr. H. S. Ballal, Vice-Chancellor Lt. Gen. (Dr.) M.
D. Venkatesh, Registrar Dr. Narayana Sabhahit, Section Heads of finance and other
logistic services, and they deserve our heartfelt gratitude. Director of Manipal Insti-
tute of Technology Cdr. (Dr.) Anil Rana, Joint Director Dr. Somashekara Bhat and Dr.
Shreesha C. and Head of the Department, Instrumentation and Control Engineering
deserve lots of appreciation for their constant guidance and motivation. Our sincere
gratitude to Prof. Soumyo Mukherji, Prof. Rajib Bandyopadhyay, Prof. R. Russel
Rhinehart, Prof. Karabi Biswas and Dr. Jayesh Barve for sharing their knowledge
and views at the conference gathering. The convener of the conference, Dr. Kapil
Sadani, deserves special recognition for his several months of untiring work towards
this conference and a special mention to the administrative staff of Manipal Academy
of Higher Education (MAHE), Manipal Institute of Technology and also the Depart-
ment of Instrumentation and Control Engineering for their wholehearted support in
making the conference event. Our sincere acknowledgement to the unanimous tech-
nical reviewers, to all contributing authors for taking time and effort to send their
research work and adhering to all review comments and formatting requirements.
We also wish to place our gratitude to Springer Nature for accepting our request to
publish the accepted/presented papers in CISCON 2021 and finally our acknowl-
edgement for all who have directly or indirectly helped us in organizing this event
successfully and bringing out these proceedings.
Contents
Muffler Transmission Loss Optimization for a Vehicle Using

Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Riziyamaalisa Gavit and Kiran Wani
Design and Simulation of a Wireless Charging System for Electric
Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Nikhil Kadam and Archana Thosar
Recent Advances in Sensor Technology for Biomedical
Applications: A Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Niharika Karnik, Karan Bhadri, and Pankaj Dhatrak
Performance Analysis of Diode Clamped and Flying Capacitor
Multilevel Matrix Converter Used for DFIG-Based Wind System . . . . . . 59
G. Pandu Ranga Reddy, D. Mahesh Kumar, K. Rajesh,
Y. Chintu Sagar, and J. Nageswara Rao
Real Time Feedback System for Speech Dysfluency in Children . . . . . . . . 75
Jennifer C. Saldanha and Rohan Pinto
Nonlinear Model-Predictive Control Using First-Principles Models . . . . 93
R. Russell Rhinehart
DC Motor System Identification and Speed Control Using dSPACE
Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
S. Menaka and S. Patilkulkarni
Oil Quality Analysis Using Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . 129
Nivedita Daimiwal, Revati Shriram, Harish Shinde, Radhika Kulkarni,
and Apeksha Galewad
Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive
Moment for Washing Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
S. Elavaar Kuzhali, Kotha Manvitha, Anisha Singh, Lakshmi Pranathi,
Shreya Dhavule, and M. Poorvita
xi
xii Contents
System Identification, Stability Analysis and PID Controller

Design for PEM Electrolyzer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Aruna Rajaiah and Jaya Christa Sargunar Thomas
Sliding Mode Hybrid Control of PMSM for Electric Vehicle . . . . . . . . . . . 165
Ajay Pawar and S. V. Jadhav
Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch
Reactor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
M. Bala Abhirami and I. Thirunavukkarasu
Performance Prediction of Solar Cell Using Virtual Production
Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
B. Ashok Kumar, T. S. Bagavat Perumaal, S. Senthilrani,
and Parthasarathy Seshadri
Optimisation of FPGA-Based Designs for Convolutional Neural
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
P. L. Bonifus, Ann Mary Thomas, and Jobin K. Antony
Design and Implementation of an Automated Fuel Station . . . . . . . . . . . . . 223
M. Jyothirmayi, Vibha B. Raj, V. Lekhana, and P. Manjunath
Heart Disease Prediction Using Machine Learning Algorithms . . . . . . . . . 239
Rea Mammen and Arti Pawar
Design and Performance Evaluation of a Simple
Resistance-to-Digital Converter for Tunneling
Magneto-Resistance-Based Angular Position Sensor with 180°
Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Kishor Bhaskarrao Nandapurkar
The Use of LBP Features in Transform Domain for Object
Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
R. Ahila Priyadharshini and S. Arivazhagan
Design and Simulation of Capacitive Pressure Sensor
for Monitoring Lead-Acid Battery Charge . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
Yashwant Adhav, Dayaram Sonawane, and Chetankumar Patil
Development of Screw Press-Dewatering Unit for Biogas Slurry . . . . . . . 303
Madhuri More, Chitranjan Agrawal, and Deepak Sharma
The Use of Photoplethysmography for Blood Glucose Estimation
by Noninvasive Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Vandana C. Bavkar and Arundhati Shinde
Single-Stage Stand-Alone Induction Motor Driven Solar Water
Pumping System with Minimal Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Anup Shetty, K. Suryanarayana, and L. V. Prabhu
Contents xiii
Automation of Weight-Based Sorting System Using Programmable

Logic Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
P. Chenchu Saibabu, R. Anjana, Manisha Kumari, and C. R. Srinivasan
Design, Development and Verification of a Fault Injection Capable
Synchronous Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Sona Meiyappan, P. Chaithanyasai, S. Swetha, M. Vishnu Deepika,
and P. V. Sunil Nag
UKF/H-Infinity Filter for Low-Cost Localization in Self-driving
Cars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
K. Bipin and P. V. Sunil Nag
Design and Implementation of Efficient IoT-Based Smart Oil
Skimmer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
S. Rajesh Kannan, V. G. Rajagopalan, H. Ramakrishnan,
S. Sibi Selvan, and Sushanth Krishnamithran
Comparison of Discrete Time Sliding Manifold and Its Impact
on System Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
Shaktikumar R. Shiledar and Gajanan M. Malwatkar
Volkswagen Emission: An Analysis on the VW Vento Using
Automotive Network Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Suprava Sarkar and Nithin Mohan
Hand Gesture-Controlled Wheeled Mobile Robot for Prospective
Application as Smart Wheelchairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Leon Muli Suryavanshi, Ananth Jnana Chandraraj,
Kshetrimayum Lochan, and Pooja Nag
Manual Dexterity Assessment Using a Nine-Hole Pegboard Test . . . . . . . 449
K. Aneesha Acharya and Amartya Choudhary
Implementation of Indoor Navigation Control for Two-Wheeled
Self-balancing Robot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
B. Vignesh, Deepa Jose, and P. Nirmal Kumar
Application of NIR Spectroscopy with Chemometrics
for Discrimination of Indian Black Pepper Berries . . . . . . . . . . . . . . . . . . . . 475
Arnab Giri, Dilip Sing, Sudarshana Ghosh Dastidar,
Pallab Kanti Halder, Nanaocha Sharma, Pulok K. Mukherjee,
and Rajib Bandyopadhyay
xiv Contents
A Comparative Study Between Partial Least Squares

and Principal Component Regression for Nondestructive
Quantification of Piperine Contents in Black Pepper by Raman
Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
Dilip Sing, Sudarshana Ghosh Dastidar, Wasim Akram,
Sourav Guchhait, Shibu Narayan Jana, Subhadip Banerjee,
Pulok Kumar Mukherjee, and Rajib Bandyopadhyay
Power Quality Data Mining Using Hybrid Feature Extraction
Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
Vidhya Sivaramakrishnan, Balaji Mahadevan, and Kamaraj Vijayarajan
Low-Cost, IOT-Based Child Safety Monitoring Robot
with User-Friendly Mobile App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 503
Kalyan Kasturi, Rajani Dharanikota, Khaleelu Rehman,
Senthilkumar Meyyappan, and Akhil Kommineni
Secure Image Classification Using Deep Learning . . . . . . . . . . . . . . . . . . . . 513
K. Gururaj, Alaka Ananth, and Sachin S. Bhat
Editors and Contributors
About the Editors
Prof. Shreesha Chokkadi completed his graduation with B.E. in Electrical and
Electronics Engineering from Government BDT College of Engineering Davanagere
in 1988. He obtained M.E. from Walchand College of Engineering Sangli in Control
Systems (Electrical) during 1992. In 2002 he was awarded with Ph.D. from IIT
Bombay for his thesis entitled New Approaches to Control Relevant Identification.
He has published about 30 research articles in indexed International journals and more
than 40 technical papers in peer reviewed conferences. He has guided 8 doctoral
students with three on rolls. Professor Shreesha is member of multiple scientific
professional bodies such as IEEE, ISTE, ISLE and IEI.
Prof. Rajib Bandyopadhyay completed bachelors, masters and Ph.D. in Electronics

from Jadavpur University in 1984, 1987 and 2001 respectively. He has been serving
the department of Instrumentation and Electronics as a professor since 2001. Prof.
Bandopadhyay works on electronic olfaction, gas sensing and design and testing
of analytical instruments. He has successfully delivered over 13 externally funded
projects, guided over 20 PhD students and published over 92 research articles in
the domain of Instrumentation. Professor Bandyopadhyay is the president of ISA
Kolkata chapter.
Contributors
Yashwant Adhav Cummins College of Engineering for Women, Pune, India

Chitranjan Agrawal Department of Mechanical Engineering, CTAE, MPUAT,
Udaipur, Rajasthan, India
R. Ahila Priyadharshini Mepco Schlenk Engineering College, Sivakasi, India
xv
xvi Editors and Contributors
Wasim Akram Department of Instrumentation and Electronics Engineering,

Jadavpur University, Kolkata, India
Alaka Ananth NMAM Institute of Technology, Nitte, India
K. Aneesha Acharya Department of Instrumentation and Control Engineering,
Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal,
India
R. Anjana Department of Instrumentation and Control Engineering, Manipal Insti-
tute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka,
India
Jobin K. Antony Department of Electronics & Communication Engineering, Raja-
giri School of Engineering & Technology, Kochi, Kerala, India;
APJ Abdul Kalam Technological University, Thiruvananthapuram, Kerala, India
S. Arivazhagan Mepco Schlenk Engineering College, Sivakasi, India
B. Ashok Kumar Department of Electrical and Electronics Engineering, Thia-
garajar College of Engineering, Madurai, India
T. S. Bagavat Perumaal Department of Electrical and Electronics Engineering,
Thiagarajar College of Engineering, Madurai, India
M. Bala Abhirami Department of Instrumentation and Control Engineering,
Karnataka, India
Rajib Bandyopadhyay Department of Instrumentation and Electronics Engi-
neering, Jadavpur University, Kolkata, India
Subhadip Banerjee School of Natural Product Studies, Jadavpur University,
Kolkata, India
Vandana C. Bavkar Department of Electronics, Bharati Vidyapeeth (Deemed to
be University), College of Engineering, Pune, India
Karan Bhadri Dr. Vishwanath Karad, MIT-World Peace University, Pune, India
Sachin S. Bhat Shri Madhwa Vadiraja Institute of Technology and Management,
Bantakal, Udupi, India
K. Bipin Department of Electronics and Communication Engineering, Amrita
School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India
P. L. Bonifus Department of Electronics & Communication Engineering, Rajagiri
School of Engineering & Technology, Kochi, Kerala, India;
P. Chaithanyasai Department of Electronics and Communication Engineering,
Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India
Editors and Contributors xvii
Ananth Jnana Chandraraj Department of Mechatronics Engineering, Manipal

Institute of Technology, Manipal, Karnataka, India
P. Chenchu Saibabu Department of Instrumentation and Control Engineering,
Karnataka, India
Amartya Choudhary Department of Instrumentation and Control Engineering,
India
Nivedita Daimiwal Cummins College of Engineering for Women, Pune, India
Sudarshana Ghosh Dastidar Department of Instrumentation and Electronics
Engineering, Jadavpur University, Kolkata, India
Rajani Dharanikota ECE Department, Nalla Malla Reddy Engineering College,
Hyderabad, India
Pankaj Dhatrak Dr. Vishwanath Karad, MIT-World Peace University, Pune, India
Shreya Dhavule Electronics and Instrumentation Engineering, Ramaiah Institute
of Technology, Bengaluru, India
S. Elavaar Kuzhali Electronics and Instrumentation Engineering, Ramaiah Insti-
tute of Technology, Bengaluru, India
Apeksha Galewad Cummins College of Engineering for Women, Pune, India
Riziyamaalisa Gavit MTech Automotive Technology, College of Engineering,
Pune, India
Arnab Giri Department of Instrumentation and Electronics Engineering, Jadavpur
University, Kolkata, India
Sourav Guchhait Department of Instrumentation and Electronics Engineering,
Jadavpur University, Kolkata, India
K. Gururaj NMAM Institute of Technology, Nitte, India
Pallab Kanti Halder School of Natural Product Studies, Jadavpur University,
Kolkata, India
S. V. Jadhav Department of Electrical Engineering, College of Engineering, Pune,
India
Shibu Narayan Jana School of Natural Product Studies, Jadavpur University,
Kolkata, India
Deepa Jose Department of Electronics and Communication, KCG College of
Technology, Chennai, Tamil Nadu, India
M. Jyothirmayi Department of Electronics and Instrumentation Engineering,
Ramaiah Institute of Technology, Bangalore, India
xviii Editors and Contributors
Nikhil Kadam Department of Electrical Engineering, College of Engineering Pune,

Pune, India
Niharika Karnik Dr. Vishwanath Karad, MIT-World Peace University, Pune, India
Kalyan Kasturi ECE Department, Nalla Malla Reddy Engineering College, Hyder-
abad, India
Akhil Kommineni ECE Department, Nalla Malla Reddy Engineering College,
Hyderabad, India
Sushanth Krishnamithran Department of Electronics and Communication Engi-
neering, St. Joseph’s College of Engineering, Chennai, India
Radhika Kulkarni Cummins College of Engineering for Women, Pune, India
D. Mahesh Kumar Deparment of EEE, PVKK Institute of Technology, Anantapur,
India
Manisha Kumari Department of Instrumentation and Control Engineering,
Karnataka, India
V. Lekhana Department of Electronics and Instrumentation Engineering, Ramaiah
Institute of Technology, Bangalore, India
Kshetrimayum Lochan Department of Mechatronics Engineering, Manipal Insti-
tute of Technology, Manipal, Karnataka, India
Balaji Mahadevan Sri Sivasubramaniya Nadar College of Engineering, Chennai,
India
Gajanan M. Malwatkar Department of Instrumentation Engineering, Govern-
ment College of Engineering, Jalgaon, India
Rea Mammen ICAS, Manipal Academy of Higher Education, Manipal, India
P. Manjunath Department of Electronics and Instrumentation Engineering,
Kotha Manvitha Electronics and Instrumentation Engineering, Ramaiah Institute
of Technology, Bengaluru, India
Sona Meiyappan Department of Electronics and Communication Engineering,
S. Menaka JSS Science and Technology University, Mysuru, India
Senthilkumar Meyyappan ECE Department, Nalla Malla Reddy Engineering
College, Hyderabad, India
Nithin Mohan Influx Big Data Solutions, Influx Technology India, Domlur,
Bengaluru, Karnataka, India
Editors and Contributors xix
Madhuri More Department of Renewable Energy Engineering, CTAE, MPUAT,

Pulok K. Mukherjee School of Natural Product Studies, Jadavpur University,
Kolkata, India;
Department of Biotechnology, Institute of Bioresources and Sustainable Develop-
ment, Government of India, Imphal, India
Pulok Kumar Mukherjee School of Natural Product Studies, Jadavpur University,
Kolkata, India;
Institute of Bioresources and Sustainable Development, An Autonomous Institute
under Department of Biotechnology, Government of India, Imphal, India
Pooja Nag Department of Mechatronics Engineering, Manipal Institute of Tech-
nology, Manipal, Karnataka, India
Kishor Bhaskarrao Nandapurkar Indian Institute of Technology (Indian School
of Mines) Dhanbad, Dhanbad, Jharkhand, India
P. Nirmal Kumar Electronics and Communication Engineering, College of Engi-
neering, Anna University, Chennai, India
Chetankumar Patil Department of Instrumentation and Control Engineering,
College of Engineering, Pune, India
S. Patilkulkarni JSS Science and Technology University, Mysuru, India
Ajay Pawar Department of Electrical Engineering, College of Engineering, Pune,
India
Arti Pawar ICAS, Manipal Academy of Higher Education, Manipal, India
Rohan Pinto Department of Electronics and Communication, St Joseph Engi-
neering College, Mangaluru, India;
Affiliated to Visvesvaraya Technological University, Belagavi, India
M. Poorvita Electronics and Instrumentation Engineering, Ramaiah Institute of
Technology, Bengaluru, India
L. V. Prabhu HEXMOTO Controls Pvt. Ltd, Mysuru, India
Lakshmi Pranathi Electronics and Instrumentation Engineering, Ramaiah Insti-
tute of Technology, Bengaluru, India
Vibha B. Raj Department of Electronics and Instrumentation Engineering,
V. G. Rajagopalan Department of Electronics and Communication Engineering,
St. Joseph’s College of Engineering, Chennai, India
Aruna Rajaiah P.S.R. Engineering College, Sivakasi, Tamil Nadu, India
xx Editors and Contributors
K. Rajesh Deparment of EEE, RGM College of Engineering and Technology,

Nandyal, India
S. Rajesh Kannan Department of Electronics and Communication Engineering,
H. Ramakrishnan Department of Electronics and Communication Engineering,
J. Nageswara Rao Deparment of Electrical and Computer Engineering, Mizan-
Tepi University, Teppi, Ethiopia
G. Pandu Ranga Reddy Deparment of EEE, G. Pullaiah College of Engineering
and Technology (Autonomous), Kurnool, India
Khaleelu Rehman ECE Department, Nalla Malla Reddy Engineering College,
Hyderabad, India
R. Russell Rhinehart Oklahoma State University, Stillwater, OK, USA
Y. Chintu Sagar Deparment of EEE, Ashoka Women’s Engineering College,
Kurnool, India
Jennifer C. Saldanha Department of Electronics and Communication, St Joseph
Engineering College, Mangaluru, India;
Suprava Sarkar Department of Instrumentation and Control, Manipal Institute of
Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India
S. Senthilrani Department of Electrical and Electronics Engineering, Velammal
College of Engineering and Technology, Madurai, India
Parthasarathy Seshadri Department of Electrical and Electronics Engineering,
Thiagarajar College of Engineering, Madurai, India
Deepak Sharma Department of Renewable Energy Engineering, CTAE, MPUAT,
Nanaocha Sharma Department of Biotechnology, Institute of Bioresources and
Sustainable Development, Government of India, Imphal, India
Anup Shetty Department of Electrical and Electronics Engineering, NMAM Insti-
tute of Technology, Nitte, Karkala, Karnataka, India
Shaktikumar R. Shiledar Department of Instrumentation Engineering, Govern-
ment College of Engineering, Jalgaon, India
Arundhati Shinde Department of Electronics, Bharati Vidyapeeth (Deemed to be
University), College of Engineering, Pune, India
Harish Shinde Cummins College of Engineering for Women, Pune, India
Revati Shriram Cummins College of Engineering for Women, Pune, India
Editors and Contributors xxi
S. Sibi Selvan Department of Electronics and Communication Engineering, St.

Joseph’s College of Engineering, Chennai, India
Dilip Sing Department of Instrumentation and Electronics Engineering, Jadavpur
University, Kolkata, India
Anisha Singh Electronics and Instrumentation Engineering, Ramaiah Institute of
Technology, Bengaluru, India
Vidhya Sivaramakrishnan New Prince Bhavani College of Engineering and
Technology, Chennai, India
Dayaram Sonawane Department of Instrumentation and Control Engineering,
College of Engineering, Pune, India
C. R. Srinivasan Department of Instrumentation and Control Engineering, Manipal
Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka,
India
P. V. Sunil Nag Department of Electronics and Communication Engineering,
K. Suryanarayana Department of Electrical and Electronics Engineering, NMAM
Institute of Technology, Nitte, Karkala, Karnataka, India
Leon Muli Suryavanshi Department of Mechatronics Engineering, Manipal Insti-
tute of Technology, Manipal, Karnataka, India
S. Swetha Department of Electronics and Communication Engineering, Amrita
School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India
I. Thirunavukkarasu Department of Instrumentation and Control Engineering,
Karnataka, India
Ann Mary Thomas Department of Electronics & Communication Engineering,
Rajagiri School of Engineering & Technology, Kochi, Kerala, India;
Jaya Christa Sargunar Thomas Mepco Schlenk Engineering College, Sivakasi,
Tamil Nadu, India
Archana Thosar Department of Electrical Engineering, College of Engineering
Pune, Pune, India
B. Vignesh Electronics and Communication Engineering, College of Engineering,
Anna University, Chennai, India
Kamaraj Vijayarajan Sri Sivasubramaniya Nadar College of Engineering,
Chennai, India
xxii Editors and Contributors
M. Vishnu Deepika Department of Electronics and Communication Engineering,

Kiran Wani ARAI ACADEMY, Chakan, Pune, India
Muffler Transmission Loss Optimization
for a Vehicle Using Genetic Algorithm
Riziyamaalisa Gavit and Kiran Wani
Abstract Automotive reactive mufflers attenuate sound using interference which

is produced by engine exhaust sound waves that are partially or fully canceled. Its
performance depends on back pressure which depends on the structure. The indicator
of the performance of the muffler is a measurement of transmission loss (TL) which
depends on the geometry of the muffler. In this work, experimental as well as CAE
analysis is used to find TL; for experimental, validation test is carried on impedance
tube, and conical adapters are designed to accommodate different diameter openings
of impedance tube and muffler. Inlet, outlet, and muffler diameter and length are
optimized to maximize TL. Genetic algorithm is used for maximization of TL given
by CAE analysis of variable models with change in range of − 15% to + 15%
of original dimensions. With the help of this work, contribution of each part can
be studied. We achieve 10.24% maximization in TL suggested by this optimized
solution. The purpose of this paper is to reduce the complexity of the design.
Keywords Transmission loss · Muffler · Genetic algorithm
1 Introduction
Sound is a normal part of daily life, and there are lots of sounds pleasant and
unpleasant, whereas unpleasant and damaging sound is called noise. It can be gener-
ated out by humans or machines. Noise and vibration are both variations in air
and media pressure that affect the human body. Noise exceeding a certain value
can create a lot of damage to human hearing ability [1]. Noise causes discom-
fort in physical or mental performance. Humans who are in direct exposure to a
R. Gavit (B)
MTech Automotive Technology, College of Engineering, Pune 41100, India
e-mail: riziyamaalisa@gmail.com
K. Wani
ARAI ACADEMY, Chakan, Pune 410501, India
e-mail: wani.pga@araiinida.com
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 1
S. Chokkadi and R. Bandyopadhyay (eds.), Smart Sensors Measurement
and Instrumentation, Lecture Notes in Electrical Engineering 957,
https://doi.org/10.1007/978-981-19-6913-3_1
2 R. Gavit and K. Wani
high level of noise coming out from engine working of an automobile or farm trac-
tors are under mild but repetitive acoustic danger. Therefore, there is necessary to
reduce this unwanted emitted noise using some devices structures [2]. Mufflers are
sound-attenuating devices that reduce radiated noise from equipment such as engines.
Sound can be affected by some external influences like pressure, temperature, and
the density of the medium. There are four types of noise sources in vehicles. They are
engine noise, wind noise, road noise, and exhaust noise. To keep up with government
legislation [3]. The muffler needs to satisfy the compatibility of appropriate dB level
means accepted minimal transmission loss, also its robustness and structural integrity.
From the above, we say that searching for the best design within the restricted time
for the vibroacoustic optimization applications is always an issue, so to address this
issue, geometry optimization of automotive mechanical structure is used to reduce
radiated sound power level by different types of optimization methods [4]. Barbieri’s
work shows us combining finite element analysis and optimization methods for
mufflers’ geometry design [5]. The intention is to obtain the dimensions for the
acoustic muffler with the transmission loss (TL), being maximized in the desired
frequency range.
L.J Yeh’s paper has shown us that a genetic algorithm (GA) can be utilized in
the optimization of uproar noise reduction in duct systems by adjusting the muffler
geometry under area or space constraints [6]. Also, its extended work showed that
parameters like crossover, mutation, and elitism techniques play pertinent roles in GA
optimization. TL optimization in a double-chamber attenuator was applicable, and it
is validated with calculations used in conventional gradient approaches or performing
experiments in the laboratory [7]. Simulated annealing (SA) is also an alternate
approach for optimization of sound transmission loss (STL) [8], but research shows
that on comparing both simulated results, GA shows better acoustic performance
than SA.
The methodology of this paper is followed as per the work of Ranjbar in which they
have investigated the effect of variation of the dimension of a component individual
on overall performance; in this project, they are working on the single-chamber or
regular muffler proper. The experimentation is done for changes in between 5 and 30%
which is carried out by the method of three-point method which is carried out in MAP
software. Later, all the results are compared, analyzed, and optimized using a genetic
algorithm. From this work, we can get that inlet diameter is a critical parameter than
other variables; the results show a rise in TL by 9 dB [9, 10]. TL consists of audio
spectrum which is a function of frequency, so to obtain a unified global measure
of noise, TL behavior is provided in the desired frequency range. The root mean
square of TL over that frequency waveband is taken for optimization-independent
variable [11]. In their extended work, shape optimization of multichamber muffler
which carried also the effect of geometrical [12] properties of muffler components
on TL optimization is studied.
M. Vora checked the outcome result of muffler design on amplitude and tractor
noise sound frequency. Quantitative analysis of apex amplitude-frequency observa-
tion various muffler installation at distinct engine speed (RPM), the result shows
that present existence of higher frequencies in a surplus amount at engine speeds of
Muffler Transmission Loss Optimization for a Vehicle Using Genetic … 3
1000, 1500, 2000 RPM. At operator ear level, SPL observations on agricultural tractor
showed lowest 930.6 Hz and highest 1502.6 Hz peak frequencies for no and standard
muffler, but normal muffler peak frequency is at 1134.9 Hz. For no muffler condition
at 1000 and 1500 RPM, peak frequencies generated are (1525.8 Hz, 103.7 Hz) and
(665.4 Hz, 2016.5 Hz), respectively. This is taken at OEL and 10 m distance from
farm tractor condition, respectively. Now, noise generated from this is (20–55 dB,
50–80 dB) and (20–70 dB, 40–90 dB), respectively [13]. So, studying the acoustic
audio spectrum for the values between 930.6 and 1502.6 Hz produces more noise
than other frequencies.
The laboratory test shows that experimental validation for finding out TL [14] of
the multichamber muffler is done using modified four-microphone impedance tube
setup which is equipped with power analyzer and DAQ system, in which modified
cones do not affect the final TL result.
1.1 The Targets in the Project to Be Achieved Are as Follows
• Optimize the multichamber muffler lengths and diameters.

• Obtain above 5% maximization of transmission loss compared to base muffler
design.
• Suggest design modification for the muffler.
• Check the effectiveness of the use of impedance tube for experimental calculation
of transmission loss for a low cost.
2 Theory
2.1 Acoustic Filter Design Parameters
1. Adequate insertion loss (attenuation)

2. Backpressure
3. Structural.
2.2 Acoustic Filter Performance Parameters
1 Transmission loss
2 Insertion loss
3 Noise reduction.
Transmission loss (TL)

It is defined as a power incident on a muffler and that is transmitted downstream into
an exhaust outlet.
TL = Lw1 − Lw2 (dB) (1)
| |
| S2( A2)2 2 |
|
TL = 10 ∗ log| | (dB) (2)
2 S1( A1) |
2
( / )
TL = 20 ∗ log A2 A1 (dB) (3)
where S1, S2 are areas of the exhaust pipe and A1, A2 are associated with the incident
and transmitted acoustic power [2, 4].
2.3 Calculation of Noise Transmission Loss by Impedance

Tube
Another way to find out transmission loss at a low cost as well without full laboratory
setup at a high cost is that we can use the impedance tube method to calculate TL.
According to standards ASTM E2611, transmission loss of a muffler can be calcu-
lated by using an impedance tube setup. To measure the transmission loss, they use
the four-microphone transfer function method [12] (Fig. 1).
In the impedance tube, we can find out sound absorption coefficient and transmis-
sion loss of any material; frequency response at four microphones is calculated and
imported into the transfer matrix method. Using the impedance tube, we can find TL
at lower as well as higher frequencies.
Fig. 1 Diagram of transmission loss measurement setup using impedance tube

2.4 Genetic Algorithm (GA)
Its ideas are like the idea of Darwin’s theory of survival of the fittest. It is normally
able to find the global minimum and can neglect local minima, so it can operate at
all types of discontinuities.
This is an optimization method, and this algorithm shows the process of natural
selection where the fittest individuals are selected for procreation to generate
offspring of the latest generation, simple law of survival of the fittest. TL comes
in the audio spectrum which is a function of frequency, so to obtain a unified global
measure of noise, TL gives noise behavior in the desired frequency range, root mean
square of TL is done, and the calculation formula is given below for RMSTL.
/
∫ ff max
min TL ( f )d f
2
RMSTL = (4)
f max − f min
This method is considered to find the best optimum value of a simple muffler which
generates a maximum root mean square of transmission loss (RMSTL). Crossover is
a genetic guide used to mix the hereditary information of two parental chromosomes
to generate the latest offspring with new properties. The mutation is manipulation as
per other data.
Discrete variable optimizations are used for engineering problems where the
design variable is fixed from a set of discreet values. Solving discreet problems
have prone and cone to as we know we have limited range of value, but the combi-
nation explosion can occur lead to take much longer time to solve the equation is
simple to type of optimization where any function can be used as fitness function
and specified as per objective maximization or minimization. Both methods have
its advantages and disadvantages, so the combination of function-based and discreet
variable optimization can give us a fast and less complex solution.
Curve fitting is a method used to find possible best-fitted curves on basis of discreet
data. You can use different sorts of postprocessing methods for plotting, interpolation,
and extrapolation methods to estimate confidence in intervals and calculate integrals
and derivatives once you have created a fit for the curve.
TL comprises a spectrum, i.e., it is a function of (f ) frequency. So, to obtain
some unified global measure of the noise TL behavior of a muffler in each desired
frequency range, the root means the square level of noise transmission loss over that
frequency waveband spectrum, known subsequently RMSTL is measured Based on
software simulation is used in this project. By computing data of TL at each frequency
obtained in numerical software MATLAB [13, 15].
3 Methodology
See Fig. 2.
Fig. 2 Workflow of research thesis
4 Simulation and Experimentation
The muffler model taken for the study is Mahindra 575 Diesel Direct Injection engine.
DI: 4 strokes, 4-cylinder; diesel engine: 2730 cc, 45 Hp; engine rated RPM: 1900
(Table 1).
Table 1 Original dimensions

Component Dimensions (mm)
of muffler
Inlet diameter 48
Outlet diameter 42
Muffler diameter 93
Inlet pipe length 400
Outlet pipe length 272
Muffler length 503
Distance from first point to the upstream 25
baffle plate
Distance between two baffle plates 75
First baffle plate hole number 6
Second baffle plate hole number 3
5 CAE Simulation
Computer-aided engineering simulation is done in COMSOL Multiphysics software

in pressure acoustics. Frequency-domain interface is used for computation of the
pressure fluctuation for the propagation of acoustic waves in fluids in inert back-
ground conditions. It is suitable for all frequency-domain simulations with harmonic
variations of the pressure field (Fig. 3).
It is used in acoustics to calculate transmission loss versus frequency. The impacts
of all eigenmodes that are adequately resolved by the mesh, as well as how they
couple with the applied loads or excitations, are considered in a frequency-domain
research. A transfer function, such as magnitude or phase of deformation, sound
pressure, impedance, or scattering parameters versus frequency, is often displayed
as the result of a frequency-domain analysis (Fig. 4).
COMSOL recommends using a minimum of five elements per wavelength for
a 3D problem, and according to the acoustic meshing criteria, six elements per
wavelength are required to capture the acoustic sound wave, but there is a chance
that one may still need to confirm the best-suited mesh size for your model using a
mesh refinement study.
• Maximum frequency of wavelength = 1800 Hz
• Sound velocity = wavelength * frequency
• 343 = wavelength * 1800
• Wavelength = 0.1905 m
• Element size = wavelength/6 = 0.03175 m = 31.4 mm.
From the above, better accuracy will be found nearby 31.4 mm element size, but
for every model, we see that it does not apply, so we take in a range of 65–28.4 mm
(Fig. 5).
Fig. 3 Three-dimensional model of muffler

Fig. 4 Meshed model in COMSOL
Fig. 5 Variable used for calculation
5.1 Experimental Testing
In impedance tube, we can find out sound absorption coefficient and transmission
loss of any material; for practical validation, experimental results are necessary.
Experimental setup
To calculate transmission loss of muffler, frequency response at four microphones is
calculated and imported into transfer matrix method which will be conducted auto-
matically by LabVIEWsoftware; the DAQ system gives us the plot of transmission
loss.
The experiment is conducted using white noise and with the lid open and closed
alternatively. A random white noise signal is provided from the speaker at the inlet
end of the muffler because generally, white noise predicts a good result to evaluate
transmission loss performance. At outlet, sound pressure level is measured with
the use of microphones. The first three readings are taken keeping the end of the
Fig. 6 Actual experimental setup
impedance tube open, and the next three readings are taken with the end of the
impedance tube closed. Further, muffler transmission loss is computed with the use
of the transfer function method.
Figure 6 shows actual experimental setup on which experiments are conducted.
From Fig. 6, we can see that the changes in the actual setup are made. The tube
used to place the samples in the impedance tube is replaced with a muffler tube
with modified conical ends, so it can be easily held into the impedance tube, and the
remaining setup remains the same.
6 Result and Discussion
6.1 Transmission Loss Experimental Results for Base

Muffler
From Table 2, we can see that the dominant frequencies lie in the range of 800–
1600 Hz; in this range, we obtain the highest transmission loss (Fig. 7).
6.2 Transmission Loss CAE Simulation Results for Base

Muffler
Using the COMSOL tool, the transmission loss at each point with the interval of
5 Hz is within the range of 60–2000 Hz, and all 389 points’ data are imported in
MATLAB where root-mean-square transmission loss (RMSTL) is calculated which
is about 20.90 dB (Fig. 8).
Table 2 Transmission loss

Frequency (Hz) TL (dB)
from experimental results for
base muffler 63 7.8
80 6.9
100 7.3
125 7.5
160 11
200 9.6
250 13.3
315 0.5
400 11.5
500 9.7
630 0.3
800 15.1
1000 17.1
1250 18.3
1600 22.2
Fig. 7 Transmission loss

from experimental result
6.3 Modified Geometry
For input variables, we are taking the length and diameter of each component length,
muffler, and outlet.
Effect of varying lengths of components

See Table 3.
Fig. 8 Transmission loss

from the simulation result
Table 3 Effect of length variation of components on TL

Percentage of change Inlet pipe length Outlet pipe length Muffler length RMSTL
00 400 272 503 20.90
+ 5% 420 19.43
285.6 20.68
528.15 19.84
+ 10% 440 18.71
299.2 20.42
553.3 20.02
+ 15% 460 18.62
312.8 20.12
578.45 19.93
− 5% 380 20.06
258.4 19.36
477.85 18.97
− 10% 360 19.38
272.8 19.81
452.7 18.92
− 15% 340 16.74
231.2 21.04
427.55 18.89
Table 4 Effect of diameter variation of components on TL

Percentage of change Inlet pipe diameter Outlet pipe diameter Muffler diameter RMSTL
0 48 42 93 20.90
+ 5% 50.4 19.56
44.1 19.46
97.65 19.39
+ 10% 52.8 19.68
46.2 19.52
102.3 19.19
+ 15% 55.2 19.99
48.3 19.77
106.95 19.13
− 5% 45.6 18.98
39.9 19.49
88.35 19.50
− 10% 43.5 19.47
37.5 18.99
83.7 19.55
− 15% 40.8 18.77
35.7 18.99
79.05 18.89
Effect of varying diameters of components

See Table 4.
From the graph, we can see that the base muffler has optimized values than others,
but when we reduce output length till the 5% reduction, TL is decreasing, but after
that, it is rising; at 15%, we get maximum RMSTL of 21 dB. The change in diameter
of the components does not show any significant effect on TL (Fig. 9).
6.4 Curve Fitting to Obtain Polynomial Equation

from Varied RMSTL
The effective criteria for model selection are comparing all criteria in which the
coefficient of determinations, R2, has the highest value among others. It measures the
value for variation considered for the best-fitted model. It is often used to compare
models and analyze which model gives the best-suited fit to the data. R always
escalates with the model size. For example,
Fig. 9 Combined graph of

the effect of variation of
length and diameter on
RMSTL
%% Muffler diameter
% Goodness of fit:
% SSE: 0.03593
% R-square: 0.9861
% Adjusted R-square: 0.9164
% RMSE: 0.1895
All data obtained till now are in discreet form; by importing these data in MATLAB
curve fitting tool, we can get polynomial function. For example,
A(x) = a1 ∗ x 6 + a2 ∗ x 5 + a3 ∗ x 4 + a4 ∗ x 3 + a5 ∗ x 2 + a6 ∗ x + a7 (5)
Coefficient values are different for all six conditions.

Figure 10 displays curve fitting tool window of polynomial curve for input
diameter.
Fig. 10 Polynomial curve for input diameter

Fig. 11 Optimized inlet length
The MATLAB Optim tool is used to optimize the data, for which genetic algorithm
solver is used with the population of 100 and other parameters are adjusted as per
best for each equation separately (Figs. 11 and 12).
The above plot shows the optimized best fit for inlet length. It is like the way the
other six optimizations are done. Now, the result obtained is given in Table 5.
From the above data, we can see that the most significant parameter is length
of transmission loss. After getting all optimized values for each parameter using a
genetic algorithm, we reconstructed the 3D model to test it for transmission loss.
Now, the value of obtained RMSTL is 23.04 dB (Fig. 13).
If we observe the plot by comparing it with the plot of TL of the base muffler, we
can see that its peak muffler TL is higher than the base muffler which is 85 dB, and
for the optimized muffler, it is 90.05 dB. If we compare the peak frequency in the
base muffler, it is observed at 1000 Hz, but in the optimized muffler, it is at 800 Hz.
7 Conclusion
Obtained RMSTL for the optimized muffler is 23.04 dB, and for the base muffler, it
is 20.9 dB. So, there is a 10.24% increase in overall RMSTL. Variation of diameter
shows the least effect on the maximization of transmission loss, but length reduc-
tion of inlet and outlet pipe by − 2.02% and − 2.09% shows significant effect on
Fig. 12 Optimized inlet diameter
Table 5 Optimization result table

Part Original dimension (cm) Optimization percentage Optimized dimension
Inlet diameter 48 − 2.0260 46.992
Inlet length 400 − 15 340
Muffler diameter 93 7.3963 99.882
Muffler length 503 − 0.1055 502.497
Outlet diameter 42 − 2.0938 41.118
Outlet length 272 15 312.8
RMSTL. Increasing the length of the expansion chamber will help to increase the
T, and the increase of muffler pipe length area for perforation will also increase
flow enhancement. Peak transmission loss in the base muffler is 85 dB, and in the
optimized muffler, it is 90.05 dB.
8 Future Work
Optimization of the muffler with the approach of the stress analysis using Modal
analysis of muffler to check for dominant frequency. As a result, frequency at which
the peak occurred is 1000 Hz in the base muffler and 800 Hz in the optimized muffler,
which is shifted to a lower frequency, so it is needed to check its overall effect and
study the effect of changing of material on the following geometry.
Fig. 13 Transmission loss of optimized muffler in COMSOL
Acknowledgements This work is carried in the NVH laboratory under the support of ARAI,
Chakan, India.
References
1. Herrin DW Design of muffler and silencer. NVH short course, lecture notes 11
2. Munjal M (1987) Acoustic of ducts and mufflers, 2nd edn, vol 21. Wiley, New York (NY)
3. Motor vehicle act regulations last amended September 20, 2020, by B.C. Reg. 240/2020,
Section 7.03
4. Ranjbar M, Marburg S, Hardtke H-J (2013) Vibroacoustic optimization of mechanical
structures: a controlled random search approach. Adv Mater Res 622:158
5. Barbieri R, Barbieri N (2006) Finite element acoustic simulation-based shape optimization of
a muffler. Appl Acoust 67:346
6. Yeh LJ, Chang YC, Chiu MC, Lai GJ (2002) Computer-aided design on single expansion
muffler under space constraints. In: Proceedings of the nineteenth C.S.M.E. national conference
(Taiwan, 2002), pp 625–633
7. Yeh LJ, Chang YC, Chiu MC, Lai GJ, Her MG (2003) Shape optimization of constrained
double-chamber muffler with extended tubes by mathematical gradient methods. In: Proceed-
ings of the 3rd conference on S.M.E. (Taiwan, 2003)
8. Yeh LJ, Chang YC, Chiu MC (2005) Shape optimal design on double-chamber mufflers using
simulated annealing and a genetic algorithm. Turk J Eng Environ Sci 29:207
9. Ranjbar M, Kermani M (2014) Muffler design by noise transmission loss maximization on
narrow band frequency range. In: The 7th automotive technologies congress (OTEKON 2014),
26–27 May 2014, Bursa, Turkey
10. Kermani M, Ranjbar M (2015) Muffler design by noise transmission loss maximization.
Doctoral thesis, Easter Merriten University
11. Kermani M, Ranjbar M (2016) Comparative study of design optimization of mufflers by genetic
algorithm and random search method. J Robot Mechatron Syst 1(2). ISSN: 2399-1550
12. Kermani M, Orak (2018) On sound transmission loss maximation of a multi-chamber exhaust
system. In: OTECON 2018
13. Vora M, Swarnkar R (2016) Effect of muffler design on amplitude and sound frequency of
tractor noise. IJERT 5(09):227–235. ISSN: 2278-0181
14. Barnard AR, Rao MD (2004) Measurement of sound transmission loss using a modified four
microphone impedance tube. In: Noise-Con’04. The 2004 national conference on noise control
engineering, Baltimore Maryland, United States 19
15. MathWorks academy documentation on “Genetic Algorithm”, “Curve Fitting”, “RMS” (2021)
Design and Simulation of a Wireless
Charging System for Electric Vehicle
Nikhil Kadam and Archana Thosar
Abstract This paper presents design process and working of wireless power transfer
system. WPT system is designed to add convenience and sense of reliability to end
consumer’s life. WPT system has been developed for small electric city car which has
power requirement of 560 W. Lead acid battery having maximum terminal voltage
of 56 V and maximum charging current is 10 A. Depending on components used
to transfer power wirelessly, wireless power transfer (WPT) technology has been
divided into three parts, i.e., capacitive WPT, inductive WPT and magnetic gear
WPT. Among other methods, resonant inductive coupling method is advantageous.
Resonant inductive WPT is used to charge the electric vehicle (EV). Inductive power
transfer (IPT) mainly consists of the two coils separated by the air. One coil is placed
on ground and other on base plate of electric vehicle. Along with the coil, power
converters are also present in the circuitry. On the transmitting side, power supply
circuitry is connected to transmitting coil placed on the ground, and on receiving
side, power conditioning circuitry is placed on vehicle side.
Keywords Wireless power transfer · Plug-in electric vehicles · IPT · Spiral coil
1 Introduction
Wired charging has more power transfer efficiency still it has some disadvantages.
Wired charging system has some problems such are exposure of wire, problem of
electric shock and vulnerable to adverse weather conditions. Wireless charging tech-
nology provides advantages over wired charging system. In wireless charging tech-
nology, as there is no presence of physical connection risk of electric shock is reduced.
Evolution of wireless charging technique can make life of electric vehicle consumer
N. Kadam (B) · A. Thosar

Department of Electrical Engineering, College of Engineering Pune, Pune, India
e-mail: Kadamnb19.elec@coep.ac.in
A. Thosar
e-mail: agt.elec@coep.ac.in
https://doi.org/10.1007/978-981-19-6913-3_2
20 N. Kadam and A. Thosar
more convenient. Wireless charging can also introduce possibility of charging while
moving concept, which can eventually introduce concept of “infinite driving range”
for electric vehicle.
Wireless power transfer (WPT) technology is an emerging technology. It is finding
its applications in various fields such are wireless information transfer systems and
wireless chargers. Wireless charger is being promising feature of WPT, and it is been
developed across the world and spreading its area of implementation in E-vehicle
charging also. Wireless power transfer technology is a promising technology because
it offers lot more advantages over traditional wired power transfer technology.
Wireless power transfer systems (WPTS) are formed by two sections, transmit-
ting section, i.e., primary section and receiving section, i.e., secondary section. In
WPTS, coils are used to transmit power from primary to secondary, i.e., from trans-
mitter to receiver. Primary section transmits power to secondary section through
resonant coupling field technology. Three technologies can be used in the WPTS
which are inductive coupling WPT, capacitive WPT and magnetic gear WPT. From
these three inductive coupling, technology is most convenient because of its higher
power transfer efficiency [1]. To increase power transfer efficiency, further reso-
nance introduced to the system. Resonant inductive coupling technology provides
best performance for wireless charging of electric vehicle [2].
Adding resonance to the system makes WPT systems more efficient when they are
operated at higher power levels. Capacitors are used on transmitting side as well as
receiving side of the coupling coil to introduce resonance in the circuit. Position of the
capacitor is used to define the compensation topology. Series-series compensation
topology yields maximum efficiency when it comes to inductive coupled wireless
power transfer [3].
Efficiency of wireless power transfer in resonant WPT system does not depend
on self-inductance of the coil, but it depends on the mutual inductance between
transmitting coil and receiving coil. Transferring higher power levels requires higher
frequencies to be used. High frequency has impact on efficiency as well [4]. Power
transfer efficiency is affected by vertical distance between the coils as well as hori-
zontal misalignment between the coils. More the misalignment higher are the losses
[5]. Power electronics is used to support WPT system. Power electronics that is
used to develop wireless power transfer system consists of switching devices. Wide
band gap devices provide high capacity, high operating frequency range, wide oper-
ating range and good thermal stability. Wide band gap switching device such as
SiCMOSFET is intended to use in high frequency applications [6]. Bidirectional
inductive power transfer in WPT is possible between the two coils, separated by
distance, magnetic coupling [7]. A model has been developed along with controller
design. Khutwad [8] has implemented wireless IPT system with efficiency level 67%
and prototype was made.
This paper develops in two stages: at first, it gives wireless power transfer system
specifications for which WPT system is intended. Brief description about the wire-
less power transfer system is included in this stage. In second stage, design procedure
has been covered. More in detail, Sect. 2 gives overview of general scheme of wire-
less power transfer system. Section 3 describes wireless power system for electric
Design and Simulation of a Wireless Charging System for Electric Vehicle 21
vehicle. General block diagram of WPT for electric vehicle is included in it. Section 4
illustrates the battery charging profile which has been taken as a base to design the
WPTs. Specifications of the system have been derived from this charging profile
itself. Section 5 is devoted for study of coupling section and coils used in WPT
systems. Coil simulation and results got in ANSYS are included here. Section 6
deals with the designing of the power electronics part. Simulations and results has
been included in the same section. Section 7 concludes the paper.
2 Overview of WPT System
WPT can be supported by various techniques. Based on operating principles, these

techniques are identified. WPT techniques are capacitive, inductive and magnetic
gear. Classification is given in Fig. 1.
2.1 Magnetic Gear WPT
Magnetic gear technology consists of two magnets. Interaction between two synchro-
nized permanent magnets is used for coupling. Due to these moving parts, wear and
Fig. 1 Classification wireless power transfer (WPT) techniques

tear of the system can happen. In MG, current produced at the output is directly
proportional to speed of rotation of secondary magnet, which is very difficult to
control. This makes it not suitable for WPT of EV.
2.2 Capacitive WPT
In capacitive power transfer, two metal plates are used as medium to transfer power
wirelessly. The surfaces facing each other of these plates can act as conducting
plates can be used to transmit field. As electric field is confined between these two
metal plates only there will be less amount of loss of field. Though capacitive power
transfer has this advantage of very small loss of field, the capacitive wireless power
transfer technology poses a practical disadvantage of small coupling capacitance.
Capacitance depends on area between the plates. As distance between the plates in
wireless charging system for electric vehicle would be around 15–20 cm, capacitance
value will be small. This is the reason capacitive wireless power transfer technique is
useful for application with small air gap and not useful for large air gap application
such as EV wireless charging.
2.3 Inductive WPT
In inductive WPT to transfer power, wirelessly two magnetic coils are set to resonate
at the natural frequency. Two coils are placed with air gap in between. An electric
current is passed through the transmitter coil that will generate a magnetic field,
which causes receiver coil to induce voltage. The magnetic resonance results in the
transfer of electric energy through the air from transmitter coil to receiver coil.
Capacitive and magnetic gear WPT techniques are not suitable for the wireless
charging for the electric vehicle. Inductive WPT being the suitable for transferring
power over a distance of 10–20 cm it is useful for wireless charging of EV. Because
of which we have chosen inductive WPT to build wireless charging system.
WPT system consists of a coupling section. Coupling sections acts as core of the
WPT system. Coupling section consists of coil or capacitive plates based on the type
of the WPT technique used. In case of the inductive WPT along with inductive coils,
capacitor is used to add resonance to the circuit.
3 WPT System for Electric Vehicle
Wireless power transfer systems (WPTS) are formed by three sections, transmitting
section, receiving section and coupling section. Figure 2 shows all the components
of WPTS. Transmitting section aims to produce in phase AC voltage at the input
Fig. 2 General scheme of WPT system
of the coupling section. Current from grid supply has distortion in it which results
in less power factor. Loss of power factor can be overcome by using power factor
circuitry. In transmitter section, power conversion takes place as AC–DC–AC. This
double conversion stages are used to increase power frequency from 50 Hz to tens
of kHz.
Receiving section consists of AC–DC rectifier and DC–DC chopper circuit.
Receiving section uses DC–DC chopper circuit to maintain required voltage and
regulate charging current at the input of the battery.
Coupling section is used as connecting link between transmitting section and
receiving section. WPTS of electric vehicle uses inductive coils as coupling device.
When current passes through transmitting coil it produces magnetic field around
it. Produced magnetic field when interacts with receiver coil, induces current in
the receiving coil. This phenomenon is known as coupling. Amount of coupling
is defined by using coupling coefficient, more the coupling coefficient more is the
power transfer efficiency. To increase the efficiency and to transfer power of higher
rating, resonance is to be used.
Compensation network used in coupling sections is classified into four types based
on position of the capacitor, which are series-series, parallel-parallel, series-parallel
and parallel-series. Series-series compensation network yields maximum efficiency.
Power conversion cycles used in WPT system requires fast switching switches.
Modern wide band gap switches such as SiCMOSFET and SiCIGBT are used to
get high switching action. There are two main types of the wide band gap switches,
first is SiC, i.e., silicon carbide and second one is GaN, i.e., gallium nitride. Over
conventional Si, i.e., silicon material used in switches, SiC has multiple advantages.
SiC material is used to develop high capacity and low power loss transistors. In the
power conversion stages, power converters build with SiC material transistor helps to
reduce power losses. SiC material transistor provides high capacity, high operating
frequency range, wide operating range and good thermal stability over conventional
Si devices.
Fig. 3 a Green
line—charging current, b
blue line—terminal voltage,
c violet line—resistive load
4 WPT System Specification
Wireless power transfer system is being developed for small electricity car. Before
designing this circuits, one should now charging profile of the battery for which
system is being designed. From the charging profile, one can get to know about
nominal values that can be used to design the system.
In this charging profile in first stage, charging current is kept constant and terminal
voltage VB is increasing from 36 to 56 V. In second stage, VB is regulated at VB max , by
decreasing charging current. The constant current stage ends when charging current
falls below threshold value. During the constant current stage power also increases
from 360 to 560 W and then reduces to 0. Battery pack appears as a variable resistive
load. Which can be defined by VB and IB . Resistance of the load varies from 3.6 to
5.6 Ω during constant current stage. In constant voltage stage when current decreases
load resistance increases till 560 Ω (Fig. 3).
As this charging cycle is very slow and parameters, i.e., current, voltage and power
operates at nominal value for long time these values should be considered for while
designing. Specification of the supply voltage and supply frequency comes from the
single phase domestic socket connected to power grid.
The dimension of the city-car chassis dictates the receiving coil diameter D stan-
dard. SAE J294 standard gives standard useable frequency for operation of wireless
charging of small light weight vehicles. That frequency is 85 kHz, this is the oper-
ating frequency, and we have used to design this paper. The distance between the
coils is kept at 15 cm assuming Indian car ground clearance based on the deployment
of transmitting coil.
5 Coil Design for Electric Vehicle
Coupling section consists of transmitter coil with transmitter compensation network

and receiver coil with receiver compensation network. Compensation is used to
introduce resonance in the system. Compensation basically is process of adding
capacitor along with coil inductance on both sides, transmitter as well as receiver
Fig. 4 Circuital scheme of inductive WPTSs (with resonance)
side. The primary and secondary coils of a wireless power transmission system have a
high leakage inductance. The main coils require a lot of current to transfer electricity
properly. The WPTs system suffers from a large deal of deprivation as a result of the
high primary supply current, which leads to decreased effectiveness. The reactive
components should be compensated by using a capacitor to solve the problem. As a
result, the resistive element persists in the network, lowering losses while increasing
effectiveness.
To add resonance to the system, capacitors are used in compensation techniques.
Placement of capacitor indicates the name of compensation techniques, such as S-S,
S-P, P-S and P-P. From all of this S-S compensation, technique is easy to model and
efficiency analysis results show S-S helps to compute maximum efficiency (Fig. 4).
ω 2 M 2 RL
η= (1)
(RR + RL ) RS + ω2 M 2 (RR + RL )
2
The efficiency is increased by using high mutual inductance and operating

frequency values. Along with high M and ω values, reducing load resistance value
also helps to improve the efficiency.
5.1 Coil Types
1. Helix coil
In helical coil structure turns are packed in a rectangular section which has base b
and height h and their mean distance is Rm. The self-inductance of helix coil can be
calculated by using (2)
0.31(Rm N )2
L Helix = (2)
6Rm + 9h + 10b
2. Spiral coil
Spiral coil structure is a planer structure. In spiral coil structure, turns are saturated
at the outer side of the circle. The self-inductance of spiral coil can be calculated by
using (3)
c
2
L spi = C1 μo N 2 Ravg ln + c3 ∅ + c4 ∅2 (3)
∅
where Ravg and ∅ can be expressed as (4)
Ro + Ri Ro − Ri
Ravg = ,∅ = (4)
2 Ro + Ri
Between helix coil and spiral coil, spiral coil has more uniform field density as
compared helix coil which yields 18% more coupling coefficient. Because of which
in IPT spiral coil is used (Fig. 5).
In inductive WPT, inductive coils made with Litz wire are being used here in our
system. Physical standards of coil are given in Table 1. ANSYS is used to perform
the FEM analysis of the coil. Spiral shaped coil shown in Fig. 6 has been developed
in ANSYS, and performance of the coil is analyzed.
Fig. 5 Coils a helical type b spiral type
Table 1 WPT specifications

Parameter Value
Nominal supply voltage 230 V
Nominal output power 560 W
WPT operating frequency 85 kHz
Coil radius 0.19 m
Distance between coil 0.15 m
Number of turns (N) 18
Fig. 6 ANSYS model for FEM analysis of spiral coil model
5.2 Coupling Section Design
Transmitter and receiver coil are assumed to be of similar dimensions. Both are of
spiral type and has inner diameter of 0.16 m and outer diameter of 0.38 m (Table 2).
R0 = 0.19 m
Ri = 0.08 m
1
k= = 0.3563
2
2 23
1 + 2 3 √R R
D
o i
Ro + Ri 0.19 + 0.08
Ravg = = = 0.135 m
2 2
Ro − Ri 0.19 − 0.08 0.11
∅= = = = 0.4074 m
Ro + Ri 0.19 + 0.08 0.27
⎧
⎪ C1 = 1.00
c ⎪
⎨
C2 = 2.46
2
L spi = C1 μo N 2 Ravg ln + c3 ∅ + c4 ∅2
∅ ⎪
⎪ C = 0.00
⎩ 3
C4 = 0.20
L spi = 108.655 µH
Table 2 Coil design

Number of turns (N1, N2) 18
parameter
Coil radius 19 cm
Cross section of wire 2.6 cm
Fs 85 kHz
Self-inductance (L T , L R ) 118.655 µH
Capacitance (CT , CR ) 36 µF
M
k≙√ ; M = 35.86 µH
LT LR
1
CT = CR = = 36 µF
ω2 L
6 Power Converter Design
Figure 7 shows power flow that will be used in this wireless power transfer system.
For power conversion in primary, in stage 1, we have used full wave rectifier, followed
by in stage 2, we have used high frequency inverter to attain frequency requirement of
the system. In secondary stage, we have receiver rectifier circuit along with DC–DC
chopper circuit to get stable power output at the input of the battery of electric vehicle.
These power conversion stages require fast switching switches. Modern wide band
gap switches such as SiCMOSFET and SiCIGBT are used to get high switching
action.
6.1 Rectifier with PFC-Boost Circuit
See Fig. 8.
The design parameters of PFC boost chopper are as follows. The input design
data is taken from Table 3.
Preliminary values of currents and voltages are required to determine the inductor
and capacitor values of the boost chopper, and they may be determined as follows.
Fig. 7 Wireless charging system

Fig. 8 Rectifier with PFC-boost circuit
Table 3 Design parameters

VG,min 207 Vrms
of PFC boost chopper
VG,max 253 Vrms
Output voltage Vout 370 V
Vout,min 350 V
Grid frequency f G 50 Hz
Switching frequency f sw 85 kHz
Iripple 10%
Vripple 6%
Maximum output power Pout_ max 700 W
The following formula can be used to compute the maximum value of the output
current
Pout,max
Iout,max =
Vout
and value comes out to be 1.89 Arms that can be seen in Fig. 9.
The grid current can be calculated as
Pout,max
IG = = 3.71 Arms
ηVin P F
√
IG,pk = 2IG = 5.25 A
2IG,pk
IG,avg = = 3.34 A
π
Fig. 9 Output of AC-DC rectifier with PFC
By considering 10% current ripple of peak grid current, the ripple current will
0.525
√
VDC,R,min = 2VG,min = 294.7
By considering 6% voltage ripple of peak grid current, the ripple voltage will
17.56 V. Input filter capacitance can be obtained as
IRipple
CR = = 44 mF
8 f sw VG,ripple,max
D = (Vout − VDC,R,min )/V _out = 0.209
Vout .D(1 − D)
L> = 1.37 mH
Iripple . f sw
2ton Pout
C> = 0.597 mF = 0.6 mF
2
Vout − Vout,on
2
Capacitor at the out of the PFC is used to get smooth output at the output. To get
smooth output from the rectifier, capacitor is placed in parallel to the load.
6.2 High Frequency Inverter
Inverter used in this paper is full bridge inverter. Structure of full bridge inverter has
been showed in Fig. 10. Based on a phase shift full bridge, inverter with a phase shift
of as near to 100% is feasible. This inverter is designed to provide higher voltage
and power levels while maintaining a same load current level.
When compared to traditional pulse width modulation (PWM) waveforms, the
major advantage of phase shifted control is the ability to regulate not only the duty
cycle but also the switching losses to some extent. It is not possible to achieve a 100%
phase shift in the planned complete bridge, but it is conceivable to get extremely near,
about 99%.
The voltages of points A and B of Fig. 10 have the waveforms showed in the first
two plots of Fig. 11, while the voltage VAB has the alternate waveform reported in
the third plot. Because the supplied circuit resonates at the operating frequency, only
the first harmonic V S of VAB affects the coil current I S. Its peak amplitude VS,pk
is varied by adjusting φ according to
4 φ
VS,pk = VDC sin (5)
π 2
The maximum amplitude is VDC π4 , reached when φ is set at π and the output
voltage VAB of the H-bridge is a square wave.
∼ 9 A this can be visualized in the graphs (Fig. 12).
4
π VCH
IS,pk = ηωM min =
Using HFI inverter DC power from rectifier circuit has been converted to AC
power. Square AC voltage and current has given to coupling section.
Fig. 10 Full bridge high frequency inverter

Fig. 11 HFI output voltage
Fig. 12 Stage 2 output
6.3 Chopper Circuit
According to Fig. 8 which is the functional block diagram of the system at the end
of power conversion stage and just before the car battery chopper circuit is used.
A chopper is a static power electronic device that is used to convert fixed dc input
voltage to a variable dc output voltage (Fig. 13).
Fig. 13 Circuit diagram of chopper
The chopper circuit’s job is to regulate the current I B injected into the battery and
the voltage VB across its terminal according to the charging profile of the battery. Peak
current in the power switch during on-time is larger than ICH,N and equal to ICH,N
if current ripple is IB,N . Current ripple is maximum when the chopper duty cycle is
50%. This condition is never going to occur in this WPT system, as minimum duty
cycle that can occur here is VVB,min
CH
= 0.55, i.e., duty cycle never going to go down
below 55%. For chopper circuit to operate in the continuous mode during entire
process, ripple current should not go below IB,min value.
Data
1. Input nominal voltage, Vi: 65 V.
2. Output required voltage (charging voltage), Vo: 56 V.
3. Current ripple = 20%.
4. Voltage ripple = 2%
5. Switching frequency (Fs ) = 85 kHz
Calculations:
ΔIL = 20% of load current = 20% of 10 A = 2 A
ΔVc = 2% of 56 V = 1.12 V
Duty ratio = Vo
VS
= 56
64
= 0.875
D(1 − D)Vs 1
L CH = ∗ = 41.176 µH
ΔIL Fs
D(1 − D)Vs 2
COUT = ∗ T = 2.616 µF
8LΔVc
Fig. 14 Output after final stage
While converting DC–DC power chopper circuit has been developed such as
maximum charging voltage and current will not exceed value VB = 56 V and I B =
10 A. Graph shown in Fig. 14 shows that charging current is less than 10 A, charging
voltage is nearly equal to 56 V. Power transferred to the battery comes out to be
560 W.
7 Conclusion
The paper has given detailed survey which concludes that inductive WPT is best
suitable for electric vehicle application. And difference between spiral and helix coil
structure given with pictorial view. Spiral coil structure has advantage over helix
structure. This paper states design of the wireless power transfer system for electric
golf vehicle.
While designing the inverter, rectifier and converters they are assumed to have
efficiency of 95%, which in turn produce overall 77% efficiency.
References
1. Ching TW, Wong YS (2013) Review of wireless charging technologies for electric vehicles. In:
2013 5th International conference on power electronics systems and applications (PESA), Hong
Kong, pp 1–4. https://doi.org/10.1109/PESA.2013.6828235
2. Jegadeesan R, Guo YX (2012) Topology selection and efficiency improvement of inductive
power links. IEEE Trans Antennas Propag 60(10):4846–4854
3. Feng H, Cai T, Duan S, Zhang X, Hu H, Niu J (2018) A dual-sided tuned series–series compen-
sated resonant converter for wide charging region in a wireless power transfer system. IEEE
Trans Ind Electron 65(3):2177–2188
4. Buja G, Bertoluzzo M, Mude KN (2015) Design and experimentation of WPT charger for
electric city car. IEEE Trans Ind Electron 62(12):7436–7447. https://doi.org/10.1109/TIE.2015.
2455524
5. Mude KN, Outeiro MT (2017) Coil misalignment analysis under different radius of coil and
wire for wireless power transfer system. In: IECON 2017—43rd annual conference of the IEEE
industrial electronics society, Beijing, pp 5319–5323. https://doi.org/10.1109/IECON.2017.821
6921
6. Liu J, Zhang Y, Wang Z, Cheng M (2018) Design of a high-efficiency wireless charging system
for electric vehicle. In: 2018 1st workshop on wide bandgap power devices and applications
in Asia (WiPDA Asia), Xi’an, China, pp 40–44. https://doi.org/10.1109/WiPDAAsia.2018.873
4657
7. Swain K, Neath MJ, Madawala UK, Thrimawithana DJ (2012) A dynamic multivariable state-
space model for bidirectional inductive power transfer systems. IEEE Trans Power Electron
27(11):4772–4780
8. Khutwad SR, Gaur S (2016) Wireless charging system for electric vehicle. In: 2016 International
conference on signal processing, communication, power and embedded system (SCOPES),
Paralakhemundi, pp 441–445. https://doi.org/10.1109/SCOPES.2016.7955869
Recent Advances in Sensor Technology
for Biomedical Applications: A Review
Niharika Karnik, Karan Bhadri, and Pankaj Dhatrak
Abstract Sensor technology has become an integral part of the diagnosis, moni-
toring, therapeutic and surgical areas of medical science. Various sensors like glucose
biosensors for diagnosis of diabetes mellitus or fluorescent sensors for gene expres-
sion and protein localization have become a common part of the biomedical field.
Due to their widespread applications, various advances and improvements have taken
place in medical sensor technology which has led to an increase in the ease and
accuracy of diagnosis as well as treatment of diseases. This review article aims
at studying various novel and innovative developments in biosensors, fibre optic
sensors, sensors used for microelectromechanical systems, flexible sensors and wear-
able sensors. This article also explores new sensing methodologies and techniques
in different medical domains like dentistry, robotic surgery and diagnosis of severe
life-threatening diseases like cancer and diabetes. Various sensors and systems used
for rapid detection of the SARS-CoV-2 virus which is responsible for the COVID-19
pandemic have also been discussed in this article. Comparison of novel sensor-based
systems for detection of various medical parameters with traditional techniques is
included. Further research is necessary to develop low cost, highly accurate and
easy-to-use medical devices with the help of these innovative sensor technologies.
Keywords Biosensors · Fibre optic sensors · Flexible sensors · MEMS · Robotic

surgery · Sensors in dentistry
1 Introduction
Sensors are an integral part of the technical and technological advances that have
taken place in fields like Robotics, Industrial Automation, Internet of Things and
Mechatronics. The Biomedical world is another domain that has seen an expo-
nential rise in the use of sensor technology not just for diagnosis of diseases or
monitoring of patients but also for surgical procedures [1, 2]. One category of
N. Karnik (B) · K. Bhadri · P. Dhatrak

Dr. Vishwanath Karad, MIT-World Peace University, Pune, India
e-mail: niharika.karnik29@gmail.com
https://doi.org/10.1007/978-981-19-6913-3_3
38 N. Karnik et al.
sensors is used in the medical field measure and monitor physical aspects like
temperature of the body, pulse rate, heart sound, blood viscosity, respiratory rate
and flow [3]. Other sensors measure chemical and biological aspects like pH value,
oxygen/carbon dioxide concentration, enzymes, antigens, antibodies, receptors and
hormones. Due to this widespread use of sensors, various advances have taken place
in their technology. Due to the advances in biomedical microsystems, a large number
of ground-breaking microelectromechanical system (MEMS) devices which make
use of specialized sensors have been introduced in the fields of diagnosis, preci-
sion surgery and therapeutic systems [4]. Fibre optic sensors have both sensing and
communication functions and are hence used in biosensing applications such as
robotic surgery, endoscopy and thermal ablation [5]. A certain study investigated the
design and performance of an extrinsic fibre optic sensor for measuring the frequency
and amplitude of heart rate signal [6]. To overcome the drawbacks of rigid materials,
flexible sensors that have improved mechanical, electrical and thermal properties
have been presented for strain sensing in various body parts such as elbow, finger,
knee and neck [7]. Tissue-based, enzyme-based, DNA biosensors, immunosensors,
piezoelectric and thermal biosensors have applications in both the medical field and
metabolic engineering [8]. In the cardiac field, optical voltage sensors are applied
for cultured human stem cell-obtained cardiomyocytes as well as in whole hearts
in vivo [9]. In the urology domain, researchers have developed an innovative Zn2 +
sensor that can be applied to the clinical diagnosis of early prostate cancer [10]. These
sensor technologies have contributed towards inexpensive, quick and at times non-
invasive diagnosis of diseases. This review introduces the progress made by various
novel sensors and affiliated systems to improve the healthcare systems through fast
detection and treatment of diseases. The further sections in this paper will discuss
the various advances and improvements made by sensors in the medical domain.
2 Materials and Methods
Various sensors are used for diagnosing, treating and monitoring diseases. They
are generally used for sensing blood pressure, temperature, pH value, respiratory
rate, oxygen and carbon dioxide concentration and many such applications [2, 3].
Advances in sensor technology have made diagnosis of diseases easier and more accu-
rate [8]. The study of these advances is important to apply them to our daily healthcare
systems and to overcome any distribution and technological problems that might arise
while commercializing these products. Remote monitoring of patient health condi-
tions is also possible, and hence, the further sections discuss the introduction of novel
and innovative sensors in the biomedical field.
Recent Advances in Sensor Technology for Biomedical Applications: … 39
2.1 Biosensors
They are investigative devices that translate a biological response to an electrical

one, and they should be very specific, not dependent on physical parameters like
temperature and pH and must be reusable [8]. They have major benefits like quick
speed of determination, high sensitivity and decent selectivity and have applications
in safety of food, biomedical research and environmental detection [3]. A biological
factor which acts as a molecular recognizing agent and a physicochemical detection
transducer makes a biosensor [11]. Figure 1 shows the various transducers used in
biosensors and their principle of operation. Recognizing agents are immobilized
on the transducer surface, and these interact with target molecules while preventing
addition of reagents in the sample solution. For operative incorporation of recognizing
agent in the sensor, there are four major coupling mechanisms that are used for
obtaining the essential purpose: immobilization of membrane, entrapment of matrix,
covalent fabrication and encapsulation by physical adsorption as shown in Fig. 2
[12].
Due to the incessant advances in micro-electronics, micro-machining, opto-
electronics and biomedicine, biosensor technology has evolved to a great extent.
Due to innovative and useful nanomaterials and investigative technologies, there is
a potential for advancement of biosensor devices for various applications such as
biomedical, clinical, bio-technological, biological, food industries, health and envi-
ronmental monitoring [13]. Voltametric, Potentiometric, Conductometric, Optical,
Calorimetric, Enzymatic, Impedimetric, Piezoelectric, Immunosensors and DNA
sensors are the different types of biosensors in use today [14]. Further research
based on biosensors can prove useful in not just detecting cancer in patients but also
prevent in-stent restenosis and for study of genomes through detecting biomarkers.
Some of the recent advances in biosensors along with their principle and use and
Fig. 1 Various transducers used in biosensors and their operation principle

40 N. Karnik et al.
Fig. 2 Different coupling mechanisms for operative incorporation of recognizing agent in the
sensor: a covalent fabrication, b entrapment of matrix, c immobilization of membrane, d
encapsulation by physical adsorption [12]
their application are shown in Table 1. The various biosensors discussed in the table
will help future researchers to understand the principle and application of recently
developed sensors such as the exosome detection biosensor or the hydrogel-based
immunosensors.
2.2 Wearable Sensors
A wearable device used for biomedical applications consists of sensors, drives,

displays and computer elements that through Wireless Network connections gives us
a digital world and makes our lives convenient and comfortable [3]. Watches, rings,
wristbands, armbands, belts, running shoes, buttons and helmets make up a majority
in the market for medical devices that can be worn [19]. The quality of life of healthy
individuals and patients has improved drastically cause of implantable sensors and
bio-integrated electronics that can be worn [20]. Improvement in surgeries and their
outcomes has been achieved due to continuous and real-time monitoring of patho-
logical and physiological aspects for disease detection. Wearable sensors have the
potential to revolutionize the modern healthcare systems through their ability to
process data in real time. It thus becomes very important to understand the various
aspects of these sensors and carry out future research on integrating concepts like
neural networks into their processing systems. Uses of wearable sensors have two
major categories called biophysical tracking and biochemical monitoring which are
explained below in further detail [21].
Table 1 Recent advances in biosensors technology (2019–2020)
Name of author Biosensor Principle Application
Cheng et al. [15] Exosome detection biosensor Recognition is done with “lock and key” These biosensors can be used for quick and
aspect of antigen and antibody interaction. easy detection of cancer in regions that have
Electrical transduction is either optical or limited medical resources and where
electrochemical diagnostic facilities are not abundant
Hoare et al. [16] Resonator sensors based on capacitance in Sensors made use of passive capacitance A term called ISR (in-stent restenosis)
bioresorbable stents and inductance resonators which are causes narrowing or blocking of vessel
biodegradable where the stent is placed, and these sensors
are used for detection of stent occlusion
Hoare et al. [16] Magnoelastic resonant sensors in stents Resonance of magnoelastic materials who A term called ISR (in-stent restenosis)
change their magnetic susceptibility due to causes narrowing or blocking of vessel
mechanical stress application where the stent is placed, and these sensors
are used for detection of stent occlusion
Balahura et al. [17] Hydrogels based immunosensors Interactions between an antigen and an Immunosensor applications consist of
antibody on transducer surface produce a microbiological analyses, environmental
change in the hydrogel constitution which analyses, genomics and proteomics,
can be detected by a transducer microarrays, nanotechnology, agriculture,
quality of food and military applications
Recent Advances in Sensor Technology for Biomedical Applications: …
Bai et al. [18] Graphene-based biosensors Various approaches for outputting the Biomarkers related to cancer like miRNA,
signal such as electrochemistry, DNA, proteins and small molecules are
fluorescence, plasmon surface, Raman quantitatively detected
scattering which enhances the resonance of
graphene and surface are used
41
42 N. Karnik et al.
2.2.1 Biophysical Tracking
Human motion analysis, temperature and heart rate monitoring play a vital part in
the assessment of health, and thus, these are the major types of wearable sensors
for biophysical tracking [22]. Cardiovascular diseases are responsible for deaths
and can generally be identified by irregularities in the heart rate. A study made a
long-term wearable CNT (carbon nanotube) and PDMS (polydimethylsiloxane) and
ECG (Electrocardiogram) electrode that can be easily connected to the normal ECG
devices and is sweat and motion resistant [23]. Another study developed a biodegrad-
able, single-use pressure sensor useful in cardiovascular monitoring, and it displayed
a high sensitivity of 0.78 kPa−1 [24]. Many diseases are defined by change in body
temperature, and thus, measuring body temperature is a very important aspect that
can help doctors diagnose diseases and also analyse the efficiency of treatments [21].
Research has been done on temperature sensors that are both stretchable and wear-
able, and analysis of their capacitive response to temperature ranges of 25–290 °C
is shown [25]. Another study defined the fabrication of an extremely deformable
and wearable temperature sensor made up of an element of sensing, interconnec-
tions and the pads which can be used for post-surgery monitoring. [26]. Motion
sensing is another important factor of biophysical tracking as it can gauge the strain
in various joints and muscles when our body is in motion. Various piezoelectric
wearable sensors for tracking human motion have been defined with some of them
having energy harvesting capacities too [20].
2.2.2 Biochemical Tracking
Wearable sensors can be used for blood, pH and biomolecule monitoring and tracking
which can help in detection of cancer and diabetes and help in monitoring blood pres-
sure. Blood test comprises sample collection which is invasive and is followed with
centrifugation for separation of plasma. Analysis of chemicals is the next step, and all
this may avert people who need urgent medical help from getting the quick and accu-
rate medical attention needed [27]. Wearable biosensor offers health status tracking
in a real-time manner by monitoring non-invasively applicable biomarkers from
biofluids. A study presents a skin-worn microneedle sensing device which makes
use of a microneedle-based enzyme electrode for nominally invasive monitoring
of subcutaneous alcohol [28]. Without any compromise in stability, response time
or sensitivity this sensor provides free of interference ethanol detection in artificial
interstitial fluid. Another study presents the formation of a wearable and stretchable
array of microsensors that simultaneously monitors heavy metals like Cd, Pb, Zn, Cu
and ions of Hg via electrochemical SWASV (Square Wave Anodic Stripping Voltam-
metry) on Au and Bi microelectrodes in human body fluids [29]. Heavy metals like
Zn and Cu are important trace elements which may have extremely harmful effects
on the body if found in excess as high copper concentration can cause kidney and
heart failure or Wilson’s disease and low copper levels may cause anaemia and
osteoporosis. CuO-based electrochemical pH sensors that have flexible substrates
are wearable with interdigitated electrodes, and CuO nanostructures are presented
in a paper [30]. The pH value shows properties of various biological, physiological
and medical conditions and hence helps in the diagnosis and monitoring of diseases.
2.3 Microelectromechanical Sensors (MEMS)
These sensors have advantages like their ability to perform in harsh environments,
capability of multiplexing, tiny size and immunity to electromagnetic interference
and are applied in microsystems, environmental and biomedical domains [31]. Gener-
ally, MEMS devices need less energy to function, are less vulnerable to surrounding
vibration, have faster reactions and involve very less volume of fluids [32]. In treat-
ment of sleep apnoea, flow generators are required as patients need air flow into their
lungs in continuous cycles. Flow sensors are required in this disease to regulate flow
of air to avoid a feeling of breathlessness due to the upper airway being obstructed.
A study presented a MEMS LCP (Liquid Crystal Polymer) pressure sensor based on
membranes and performed experiments to examine the scope of this flow sensor for
sleep apnoea therapy [33]. Another study presented a MEMS pressure sensor based
on FPI (Fabry–Perot interferometer) with anodic bonding whose major principles
of design and rudimentary mechanical model were introduced [31]. Neural sensing
is the pillar of machine-brain interface, research in neuroscience and clinical neuro-
modulation. In a study, a flexible array of microelectrodes is proposed, which is used
for neural stimulation sensing on surface of the neuron, without tissue penetration
[34]. Triaxial MEMS tactile sensors placed on the tip of forceps for grasping are
used in laparoscopic surgery to evaluate stress. [35]. Despite their extensive appli-
cations, various limitations related to them restrained the use of MEMS sensors for
a while. These disadvantages included high fabrication cost related to the prerequi-
site of clean room facilities and brittleness that reduced their application spectrum
and saturation of their response over an extended time period [36]. Further for their
large-scale production, problems may arise due to their small size and the lack of
manufacturing facilities. To overcome these disadvantages and to increase the use
of MEMS sensors in various medical systems, it is important for researchers to
understand the fundamental principles that govern these sensors and increase their
subsequent applications. Table 2 shows some of the MEMS sensors, their sensing
methods and the parameter they detect.
2.4 Fibre Optic Sensors (FOS)
A growing need for non-stop and remotely monitoring medical parameters has
created an ever-increasing demand for fibre optic sensors because of their reduced
form and the advantage of spatial monitoring with a single or multiple sensors
[2]. Optical sensing mechanisms comprise Raman spectroscopy, fluorescence,
44 N. Karnik et al.
Table 2 MEMS sensors used in diagnosis [4]

Sensor Sensing method Parameter detection
Micro-cantilevers Mechanical sensing Tuberculosis, human papilloma virus
(HPV)
Amperometric Electrochemical sensing Glucose
Potentiometric Electrochemical sensing Urea, penicillin, amino acids
Conductometric Electrochemical sensing Gases, non-conducting mediums
Impedimetric Electrochemical sensing Human immunodeficiency virus
(HIV)
Surface plasmon resonance Optical sensing Haemoglobin, pathogens
Interferometric sensors Optical sensing Herpes simplex virus (PHV)
colorimetry, absorption, luminescence, RI (refractive index) and SPR (surface

plasmon resonance) [5]. Light propagates through the core by total internal reflection
as the optical fibre is made of a core which is germanium-doped silica and is bound
by a cladding made of pure silica which has a somewhat lower RI. Light spreading
through fibres is divided into two parts: the “exponentially” decaying field in cladding
which is temporary and the directed field of the core. The field which is temporary
in cladding goes down to zero before it reaches the external medium. These sensors
are currently applied in the vascular, urologic, neural, and gastrointestinal diagnostic
fields. These sensors show promising potential for diagnosis of medical conditions
of the heart, kidneys and stomach. Due to their flexible properties and small size,
they show large number of applications as wearable sensors. A study presents design,
operation principle, performance and experiments done on an extrinsic FOS devel-
oped for measuring frequency and amplitude of heart rate [6]. The main application
areas of the fibre optic sensor are shown in Fig. 3. The fibre optic sensors are mainly
of two types: the Fabry–Perot interferometry (FPI) and Fibre Bragg grating (FBG)
sensors [5].
2.4.1 Fabry–Perot Interferometry (FPI)
A small cavity inside the optical fibre that produces large spectral variation having
a periodic pattern forms the basis of the FPI. They are one of the major aspects for
sensing of pressure in medical devices [37]. The cavity is enclosed by a diaphragm
that bends the cavity and controls the length when there is application of pressure
as shown in Fig. 4. Extrinsic FPI Pressure Sensors are advantageous as they have a
comparatively small size, less expensive and simple fabrication process, fairly high
sensitivity to pressure, are suited for operations in high temperature areas due to
their high mechanical strength [38]. A certain study proposed and demonstrated a
microcavity extrinsic FPI curvature sensor with enhanced sensitivity having a shift
sensitivity of −5.9607 nm/m−1 along with cross-sensitivity for lower temperature
of 9.9696 × 10−3 m−1 /°C in a range of curvature of 3–15 m−1 [39]. Another study
Fig. 3 Major application areas of fibre optic sensors in the biomedical field [2]
Fig. 4 General layout of the Fabry–Perot interferometer [2]
presents an extrinsic FPI pressure sensor based on a nano-thick silver diaphragm

[40]. The sensor thus designed exhibits a linear response within the range of pressure
variation up to 50 kPa. Pressure variations can help measure blood pressure, detect
variations in heartbeat and monitor the condition of stents. Further research in the
use of FPI sensors may prove to be beneficial in reducing risk of heart attacks and
low blood pressure and offer treatment for various cardiovascular diseases.
2.4.2 Fibre Bragg Grating Sensors (FBG)
These are the most widespread optical sensors and they are filters for reflection,
having a very fine bandwidth and a rapid response, while also being sensitive to
temperature and strain [37]. They have applications in oil and gas, structural health
monitoring and medical devices and are compact, less expensive devices. A study
presents manometry catheters made using FBG which gives a close to perfect method
46 N. Karnik et al.
Fig. 5 Schematic of sensor design used in the fibre optic manometry catheters [41]
for physiological monitoring of the gut in a nominally invasive way as shown in

Fig. 5 [41]. Another paper describes a new FBG pressure sensor comprising two
FBG sensors and a stretchable silicone cylinder made for monitoring of patients
with ventricular assist devices [42]. In another study, fabrication of a small acoustic
and flexible FBG sensor having improved sensitivity to pressure was shown which
used a side-polished arrangement of ∼15.5 mm length and ∼53 µm depth. Some
studies also use the Fibre Bragg grating sensors and the Fabry–Perot interferometer
in conjunction [2, 34].
2.5 Flexible Sensors
These sensors are mostly non-invasive, and detection of diseases can be done by
checking the composition, lack and abundance of constituent material in saliva,
tear, neural activity, blood and its pressure, skin, urine and numerous other param-
eters [32]. Sensors that change mechanical forces like tension, pressure, stress and
strain into electrical signals are vital amongst various flexible sensors [7]. Generally,
flexible physical sensors are made up of two separate building blocks, the template
material which is flexible and the active agents used for sensing, which have liquid or
solid form and can have a principle of mechanical deformation like twisting, pressing,
bending and stretching of the sensing device [32]. A study presented the fabrication,
design and application of an innovative graphite or PDMS (Polydimethylsiloxane)
flexible sensor for biomedical applications for detecting strain [1]. Another paper
presented a stretchable strain sensor built on ionic liquids and manufactured with
the help of Polydimethylsiloxane mould and the method of simple sealing [22].
Another study fabricated a flexible pressure sensor useful for applications of smart
skin using graphene oxide and its response to the various pressures assessed by
gauging the electric current [43]. Electronic skin is an artificial device that copies
the behaviour of the human skin with respect to carrying mechanical stimulus to the
inner layers and assembling and processing the received signals. Like human skin,
Table 3 Various flexible sensors with their biomedical applications [7]

Vitro/Vivo Body part Parameter evaluated Application
In vitro Hand Motion Recovery of hand monitoring
Wrist Pulse Cardiovascular diseases diagnosis
Upper arm Blood pressure Hypertensive treatment
Knee and elbow Strain distribution Posture and movement detection;
postoperative rehabilitation
Foot Pressure distribution Sport training and gait detection
Throat Sound vibration Diagnosis of damaged vocal cords
and voice recognition
Face Muscle movement Face paralysis diagnosis;
recognition of facial expressions
In vivo Cardiovascular ECG Treatment and prevention of
cardiovascular diseases
Heart Heart rhythm Sinus arrhythmia correction and
cardiac pacing
Bone Pressure Osteoporosis and fracture
treatments and nerve tissue repair
it needs to be flexible and hence latest advances in flexible sensors are encouraging
the developments of these electronic skins. Research was also carried out on the
fabrication of a 3D stretchable ECL (Electro-chemiluminescence) sensor built on a
GPE (graphene paper electrode) with large surface area, exceptional conductivity,
high porosity and an extensive electrochemical potential [44]. ECL is a procedure
where an electrochemical voltage triggers luminescent signals release and is used as
an operative investigative method in drug analysis and immune assay. Table 3 gives
a brief overview of the various flexible sensors, their placement in the body and their
applications.
2.6 Application of Sensors in Biomedical Fields
Generally, sensors are used for diagnosis and treatment of various cardiovascular,
urological and neurological diseases [2, 32]. However, recently sensors have also
been used in diagnosis of various kinds of cancers and diabetes mellitus [12, 45].
They have revolutionized the concept of robotic surgery with easy force feedback
for robotic forceps used in surgeries and are used extensively in the field of dental
implants for measuring the stress on the implant as well as the primary stability
of the implant [46, 47]. This paper highlights the use of novel sensors for devel-
oping monitoring and detection systems in the fields of dentistry and robotic surgery
while highlighting methods like biomarker detection and real-time monitoring of
48 N. Karnik et al.
cancer and diabetes. The following section highlights some of the advances in sensor
technology and their implementation in the above-mentioned fields.
2.6.1 Dentistry
The field of dentistry involves analysis of forces, stresses on the teeth as well as
bacteria concentration that can lead to cavities or plaque formation. Hence, various
sensors are used extensively in this field for quick and accurate diagnosis and analysis.
Maximum bite force present cause of joint action of mastication muscles, teeth and
joints is the main pointer for practical masticatory system condition. A study sheds
light on a procedure that uses a FBG sensor for dynamic bite force measurement
which is called a Bite Force Measurement Device (BFMD) [48]. Bacteria existing
on the tooth enamel dissolve the hard tissue of the teeth and can cause cavities.
A study designed a graphene-based wireless sensor for bacteria recognition on the
tooth enamel [49]. Dental implants find application majorly for replacing missing
teeth and mechanical factors play a key role in their failure. Therefore, it is very
important to evaluate key parameters like frequency of resonance, allowable stress
and strain (maximum), and fracture resistance for determination of implant dura-
bility [50]. A study uses a FBG sensor to evaluate strains at the surface of mandible
produced due to impact or static loads and the sensor is verified by loading and
recording uniaxial strain dynamically and statically [51]. Osseointegration assess-
ment is another vital factor to prevent premature failure of the dental implant. A
study performed the RFA (Resonance Frequency Analysis) for gauging the primary
stability of the implant using artificial bone blocks composed of polyurethane foam
and its FRF (Frequency Response function) characterized its structural resonance
[52]. Two Hall effect sensors were used to gauge the response and excitation signals
wherein a dual inductance device provided an excitement source to the bone-implant
assembly. Another study highlighted a Non-contacted RF (Radio Frequency) detec-
tion device made with a loudspeaker which generates excitation in the acoustic form
and for structure’s vibration analysis they used a capacitance sensor [46]. The setup
of this device is shown in Fig. 6. Accuracy check of each model was analysed with
“Osstell” Mentor device. Another study on RFA for dental implants talks about a
“L-shaped” transducer which is compact and is fixed to abutment or the implant by
a fastener [53]. Hence, use of sensors for measuring various mechanical parameters
of dental implants has caused advances in their technology. Dentists and implan-
tologists benefit greatly due to accurate and non-invasive techniques for monitoring
dental health that has developed due to integration of these sensors in biomedical
devices.
2.6.2 Robotic Surgery
Robotic surgery has replaced conventional surgical techniques, and thus, many tech-
nological advances have taken place in the sensors that assist in these surgeries. Force
Fig. 6 Non-contacted RF
(Radio Frequency) detection
device [46]
sensing is vital in increasing the competence of robots in multifaceted environments

like the medical field [54]. In micro-surgery of the retina, assistance from robots can
solve everyday problems encountered during the routine of micro-manipulations.
These problems include tool access, shaking of hands, limits to the resolution for
manipulation of tools, target motion, visualization of the target, resolution and use of
non-injurious but healing forces [2]. Several sensors for detecting force are used in
micro-manipulation, micro-surgery and nominally invasive surgery. In Vitreoretinal
surgery (VRS) which is robot-assisted, force feedback from tool-tip can deliver the
essential information to help a surgeon perform all manoeuvres, efficiently decrease
forces with enhanced precision and possibly increase the efficiency and safety of the
procedure as shown in Fig. 7 [47]. In a study, a novel cohesive system consisting
of a tremor-cancelling, active and hand-held micro forceps with motorized sensing
of force known as Micron was developed [55]. Another paper shows an enhanced
bimorph, piezoelectric tactile sensor which can distinguish between soft materials
having mechanical characteristics that are similar. The aim for this sensor is to
develop surgical robotic tools which are intelligent for resection of brain tumour
by means of integrated sensors [56]. Its function is to support neurosurgeons by
creating tactile tools that are intra-operative to advance tissue differentiation and
delineation. Another study presents the implementation of a tactile or force sensor
called HexaTactile which has a 3-DOF (Degree of Freedom) parallel mechanism
which is dissociated for Robot-Human interaction based on 6 MEMS barometers
array of soft tactile sensors [54]. Another research fabricated a novel MRE (Magne-
torheological) elastomer, which produces small movements beneath magnetic and
electric fields, and it is used to sense the angles of bending and movements for flexible
robots in biomedical applications [57].
2.6.3 Cancer Diagnosis
Recognition of cancer at an initial stage is crucial for cancer treatment and for
reducing mortality rate. Hence, new, quick and easy methods for detection of cancer
cells continue to be challenging [58]. Previously, the common approaches for analysis
of cancer contain biopsy examination, tumour imaging and recognizing biomarkers
of cancer. Fluorescent sensors based on Aptamers (single-stranded and short DNA
or RNA molecules capable of identifying many targets with specificity) have been
developed for detecting cancer biomarkers [59]. Another article developed a (PDA
50 N. Karnik et al.
Fig. 7 The sensor must be

near the tip of the tool so that
it can remain in the eye
during vitreoretinal surgery
[55]
NPs) Polydopamine nanoparticles sensor for free of label electrochemical (A-5490)

lung cancer cell detection, and its capability for detection was also investigated
[58]. Antigens associated with cancer have (CDE) Cancer-Derived Exosomes on the
surface, and exosome detection biosensors facilitate early detection in individuals
from resource-limited areas who formerly did not have access to innovative tools
for diagnosis [15]. Another paper presented an innovative microcantilever sensor for
detecting cancer prematurely with the help of a combined ring resonator [60]. In
this study, the reaction of the antibody and antigen will cause the mass of tumour
biomarkers on the microcantilever to increase. The sensing action will be performed
due to this increase. Thus, a lot of progress has been made in sensor technologies for
detecting cancer using cancer biomarkers.
2.6.4 Diabetes Mellitus Diagnosis
The growing number of people affected by diabetes mellitus each year has caused an
increase in the use of biosensors for glucose-sensing technologies [12]. This section
discusses the novel and innovative sensors that have been developed for non-invasive
and accurate glucose sensing. A certain study developed a nonenzymatic, dispos-
able, simple, fast and inexpensive sensor using a disposable PGE (pencil graphite
electrode) improved by [Cu(NP)] nanoparticles of copper which showed a peak of
absorption at 560 nm [45]. There is a variation in the composition of tears due to
metabolic and ocular factors and tears are thus used for the evaluation of health. In
another study, contact lenses made of microfluid were made for analysing in situ tear
glucose, protein, pH, and sensing of nitrite ions as shown in Fig. 8 [61]. Another
paper presents a monitoring system for glucose which is not just real time but can
also reference its electrochemical potential with an electrode made of wired glucose
dehydrogenase [62]. It evaluates changes in the glucose concentration between 3.5
Fig. 8 Microfluidic lens

developed for in situ tear
analysis [61]
and 8.5 mM. These recent advances have proved extremely useful in continuous and
real-time monitoring of glucose which has helped patients with diabetes.
2.6.5 Sensors for Battle Against COVID-19
The coronavirus disease presents many problems for its fast clinical diagnosis due
to its varied symptoms including loss of sense of smell and taste, headache, breath-
lessness, high fever, throat ache and diarrhoea and in some cases lack of any symp-
toms [63]. As of now, the sole established gold standard is the Reverse transcriptase
(RT-PCR) test that uses the mechanism of nucleic acid sequencing for definitive
detection of COVID-19 [64]. However, for rapid detection of the virus, various
biomedical sensors and systems have been developed in the last two years. An article
proposes the development of an optical fibre sensor for the detection of the virus
using the concept of Evanescent wave absorbance (EWA) technology [65]. This
sensor detects the elevated levels of immunoglobin M (IgM) and immunoglobin G
(IgG) biomarkers caused by an exposure to the SARS-CoV-2 virus. The infection can
also be detected by affinity-based biosensors such as the DPV-Based sensor which
uses calixarene functionalized graphene oxide for targeting RNA of SARS-CoV-2
and Plasmonic-based sensor which is built on the basics of surface plasmon reso-
nance (SPR) phenomenon [66]. Another study proposed the use of built-in smart-
phone sensors such as colour, humidity, cameras, temperature, proximity, inertial
and wireless sensors to develop an AI-based framework for rapid detection of the
virus [67]. Research was also conducted on detection of the virus in both symp-
tomatic and asymptomatic cases using available wearable sensors and integrating
them with Deep Neural Networks (DNN) [68]. The DNNs were trained from data
obtained from 87 subjects who were either healthy, COVID positive (symptomatic) or
COVID positive (asymptomatic) and were deployed on devices like smartphones and
smartwatches. Thus, various biosensors, wearable sensors and optical fibre sensors
have been developed for faster diagnosis of the coronavirus. However, these tools
52 N. Karnik et al.
are still not commercially available in the market and hence more research needs to
be carried out to overcome various technological challenges to make these sensors
and systems viable for mass production [69].
2.7 Advances in Sensor Technology
The use of sensors has helped in providing faster, non-contact and non-invasive
methods of diagnosis, monitoring and treatment of diseases and medical conditions
as compared to traditional methods. A comparative study between modern systems
comprising novel sensors and traditional methods has been made as shown in Table
4. Thus, in terms of technical parameters, the modern sensors that are developed
have a smaller size, better quality, are thermally and chemically stable and support
wireless communication and high rate of data transfer. They can also be implantable
and have improved noise cancellation techniques. Their sensitivity is higher, and
they have optimized timing which makes them perfect for real-time monitoring of
biomedical parameters.
Table 4 Comparison of traditional techniques and novel sensors developed for detection of various
medical parameters
Sr. No. Medical parameters Traditional technique Novel sensors developed
1 Heart rate Heart rate monitor, Wearable sensors [70], Fibre
electrocardiogram (ECG) optic sensors [71], Biosensors
[72], Flexible sensors [73]
2 Blood pressure Sphygmomanometer Wearable sensors [74], Fibre
optic sensors [75], Biosensors
[76]
3 Glucose levels Traditional glucometer MEMS sensors [4],
(Finger prick test) Biosensors [77]
4 Blood urea Blood urea nitrogen MEMS sensors [4]
(BUN) test
5 Temperature Thermometer Wearable sensors [22], Fibre
optic sensors [75]
6 Human motion analysis Gait analysis Wearable sensors [78], Fibre
optic sensors [79], Flexible
sensors [80]
References
1. Nag A, Afasrimanesh N, Feng S, Mukhopadhyay SC (2018) Strain induced graphite/PDMS

sensors for biomedical applications. Sens Actuators A Phys 271:257–269. https://doi.org/10.
1016/j.sna.2018.01.044
2. Tosi D, Poeggel S, Iordachita I, Schena E (2018) Fiber optic sensors for biomedical applications.
Elsevier Inc.,
3. Wan H et al (2020) Biomedical sensors. Elsevier Inc., https://doi.org/10.1016/B978-0-12-816
034-3.00002-X
4. Sonetha V, Agarwal P, Doshi S, Kumar R, Mehta B (2017) Microelectromechanical systems
in medicine. J Med Biol Eng 37(4):580–601. https://doi.org/10.1007/s40846-017-0265-x
5. Yin MJ, Gu B, An QF, Yang C, Guan YL, Yong KT (2018) Recent development of fiber-
optic chemical sensors and biosensors: mechanisms, materials, micro/nano-fabrications and
applications. Coord Chem Rev 376:348–392. https://doi.org/10.1016/j.ccr.2018.08.001
6. Yhuwana YGY, Apsari R, Yasin M (2017) Fiber optic sensor for heart rate detection. Optik
134:28–32. https://doi.org/10.1016/j.ijleo.2017.01.035
7. Cheng M et al (2020) A review of flexible force sensors for human health monitoring. J Adv
Res. https://doi.org/10.1016/j.jare.2020.07.001
8. Mehrotra P (2016) Biosensors and their applications—a review. J Oral Biol Craniofac Res
6(2):153–159. https://doi.org/10.1016/j.jobcr.2015.12.002
9. Acker CD, Yan P, Loew LM (2020) Recent progress in optical voltage-sensor technology and
applications to cardiac research: from single cells to whole hearts. Prog Biophys Mol Biol
154:3–10. https://doi.org/10.1016/j.pbiomolbio.2019.07.004
10. An Y et al (2020) A novel tetrapeptide fluorescence sensor for early diagnosis of prostate cancer
based on imaging Zn2+ in healthy versus cancerous cells. J Adv Res 24:363–370. https://doi.
org/10.1016/j.jare.2020.04.008
11. Nguyen HH, Lee SH, Lee UJ, Fermin CD, Kim M (2019) Immobilized enzymes in biosensor
applications. Materials (Basel) 12(1):1–34. https://doi.org/10.3390/ma12010121
12. Ali J, Najeeb J, Asim Ali M, Farhan Aslam M, Raza A (2017) Biosensors: their fundamentals,
designs, types and most recent impactful applications: a review. J Biosens Bioelectron 08(01):1–
9. https://doi.org/10.4172/2155-6210.1000235
13. Maduraiveeran G, Sasidharan M, Ganesan V (2018) Electrochemical sensor and biosensor
platforms based on advanced nanomaterials for biological and biomedical applications. Biosens
Bioelectron 103(Decem 2017):113–129. https://doi.org/10.1016/j.bios.2017.12.031
14. Singh S, Kumar V, Dhanjal DS, Datta S (2020) Biological biosensors for monitoring and
diagnosis. Microb Biotechnol 317–335. https://doi.org/10.1007/978-981-15-2817-0
15. Cheng N et al (2019) Recent advances in biosensors for detecting cancer-derived exosomes.
Trends Biotechnol 37(11):1236–1254. https://doi.org/10.1016/j.tibtech.2019.04.008
16. Hoare D, Bussooa A, Neale S, Mirzai N, Mercer J (2019) The future of cardiovascular stents:
bioresorbable and integrated biosensor technology. Adv Sci 6(20). https://doi.org/10.1002/
advs.201900856
17. Balahura LR, Stefan-Van Staden RI, Van Staden JF, Aboul-Enein HY (2019) Advances in
immunosensors for clinical applications. J Immunoassay Immunochem 40(1):40–51. https://
doi.org/10.1080/15321819.2018.1543704
18. Bai Y, Xu T, Zhang X (2020) Graphene-based biosensors for detection of biomarkers.
Micromachines 11(1). https://doi.org/10.3390/mi11010060
19. Kaewkannate K, Kim S (2016) A comparison of wearable fitness devices. BMC Public Health
16(1). https://doi.org/10.1186/s12889-016-3059-0
20. Zhou H et al (2020) Stretchable piezoelectric energy harvesters and self-powered sensors for
wearable and implantable devices. Biosens Bioelectr 168(May):112569. https://doi.org/10.
1016/j.bios.2020.112569
21. Nasiri S, Khosravani MR (2020) Progress and challenges in fabrication of wearable sensors
for health monitoring. Sens Actuators A Phys 312:112105. https://doi.org/10.1016/j.sna.2020.
112105
54 N. Karnik et al.
22. Zhang SH, Wang FX, Li JJ, Peng HD, Yan JH, Pan GB (2017) Wearable wide-range strain
sensors based on ionic liquids and monitoring of human activities. Sensors (Switzerland)
17(11):1–10. https://doi.org/10.3390/s17112621
23. Jung HC et al (2012) CNT/PDMS composite flexible dry electrodes for long-term ECG moni-
toring. IEEE Trans Biomed Eng 59(5):1472–1479. https://doi.org/10.1109/TBME.2012.219
0288
24. Boutry CM, Nguyen A, Lawal QO, Chortos A, Rondeau-Gagné S, Bao Z (2015) A sensitive and
biodegradable pressure sensor array for cardiovascular monitoring. Adv Mater 27(43):6954–
6961. https://doi.org/10.1002/adma.201502535
25. Park SJ, Jeon JY, Kang BC, Ha TJ (2020) Wearable temperature sensors based on lanthanum-
doped aluminum-oxide dielectrics operating at low-voltage and high-frequency for healthcare
monitoring systems. Ceram Int (Sept 2020). https://doi.org/10.1016/j.ceramint.2020.10.023
26. Wu X et al (2015) Thermally stable, biocompatible, and flexible organic field effect transis-
tors and their application in temperature sensing arrays for artificial skin. Adv Func Mater
25(14):2138–2146. https://doi.org/10.1002/adfm.201404535
27. Mohan AMV, Rajendran V, Mishra RK, Jayaraman M (2020) Recent advances and perspectives
in sweat based wearable electrochemical sensors. TrAC Trends Anal Chem 131:116024. https://
doi.org/10.1016/j.trac.2020.116024
28. Mohan AMV, Windmiller JR, Mishra RK, Wang J (2017) Continuous minimally-invasive
alcohol monitoring using microneedle sensor arrays. Biosens Bioelectron 91:574–579. https://
doi.org/10.1016/j.bios.2017.01.016
29. Gao W et al (2016) Wearable microsensor array for multiplexed heavy metal monitoring of
body fluids. ACS Sens 1(7):866–874. https://doi.org/10.1021/acssensors.6b00287
30. Manjakkal L, Sakthivel B, Gopalakrishnan N, Dahiya R (2018) Printed flexible electrochemical
pH sensors based on CuO nanorods. Sens Actuators B Chem 263:50–58. https://doi.org/10.
1016/j.snb.2018.02.092
31. Ge Y, Cai K, Wang T, Zhang J (2018) MEMS pressure sensor based on optical Fabry–Perot
interference. Optik 165:35–40. https://doi.org/10.1016/j.ijleo.2018.03.112
32. Vilela D, Romeo A, Sánchez S (2016) Flexible sensors for biomedical technology. Sens
Actuators A Phys 16(3):402–408. https://doi.org/10.1039/c5lc90136g
33. Abbasnejad B, Thorby W, Razmjou A, Jin D, Asadnia M, Ebrahimi Warkiani M (2018) MEMS
piezoresistive flow sensors for sleep apnea therapy. Sens Actuators A Phys 279:577–585. https://
doi.org/10.1016/j.sna.2018.06.038
34. Manikandan N, Muruganand S, Divagar M, Viswanathan C (2019) Design and fabrication of
MEMS based intracranial pressure sensor for neurons study. Vacuum 163(Nov 2018):204–209.
https://doi.org/10.1016/j.vacuum.2019.02.018
35. Kuwana K, Nakai A, Masamune K, Dohi T (2013) A grasping forceps with a triaxial MEMS
tactile sensor for quantification of stresses on organs. In: Proceedings of the annual international
conference of the IEEE engineering in medicine and biology society, EMBS, pp 4490–4493.
https://doi.org/10.1109/EMBC.2013.6610544
36. Afsarimanesh N, Nag A, Sarkar S, Sabet GS, Han T, Mukhopadhyay SC (2020) A review on
fabrication, characterization and implementation of wearable strain sensors. Sens Actuators A
Phys 112355. https://doi.org/10.1016/j.sna.2020.112355
37. Tosi D (2015) Advanced interrogation of fiber-optic Bragg grating and Fabry–Perot sensors
with KLT analysis. Sensors (Switzerland) 15(11):27470–27492. https://doi.org/10.3390/s15
1127470
38. Bremer K, Lewis E, Leen G, Moss B, Lochmann S, Mueller IAR (2012) Feedback stabilized
interrogation technique for EFPI/FBG hybrid fiber-optic pressure and temperature sensors.
IEEE Sens J 12(1):133–138. https://doi.org/10.1109/JSEN.2011.2140104
39. Liao N et al (2020) A sensitivity-enhanced micro-cavity extrinsic Fabry–Perot interfero-
metric fiber-optic curvature sensor. Optik 221(May):165310. https://doi.org/10.1016/j.ijleo.
2020.165310
40. Xu F et al (2012) High-sensitivity Fabry–Perot interferometric pressure sensor based on a
nanothick silver diaphragm. Opt Lett 37(2):133. https://doi.org/10.1364/ol.37.000133
41. Arkwright JW, Underhill I (2017) Fibre Bragg grating manometry catheters for in vivo
monitoring of peristalsis, 1005410. https://doi.org/10.1117/12.2255645
42. Stephens AF, Busch A, Salamonsen RF, Gregory SD, Tansley GD (2019) A novel fibre Bragg
grating pressure sensor for rotary ventricular assist devices. Sens Actuators A Phys 295:474–
482. https://doi.org/10.1016/j.sna.2019.06.028
43. Hosseindokht Z, Mohammadpour R, Asadian E, Paryavi M, Rafii-Tabar H, Sasanpour P (2020)
Low cost flexible pressure sensor using laser scribed GO/RGO periodic structure for electronic
skin applications. Superlattices Microstruct 140(Feb):106470. https://doi.org/10.1016/j.spmi.
2020.106470
44. Han Y, Fang Y, Ding X, Liu J, Jin Z, Xu Y (2020) A simple and effective flexible electrochemi-
luminescence sensor for lidocaine detection. Electrochem Commun 116:106760. https://doi.
org/10.1016/j.elecom.2020.106760
45. Pourbeyram S, Mehdizadeh K (2016) Nonenzymatic glucose sensor based on disposable pencil
graphite electrode modified by copper nanoparticles. J Food Drug Anal 24(4):894–902. https://
doi.org/10.1016/j.jfda.2016.02.010
46. Pan MC, Zhuang HB, Chen CS, Wu JW, Lee SY (2013) A noncontact resonance frequency
detection technique for the assessment of the interfacial bone defect around a dental implant.
Med Eng Phys 35(12):1825–1830. https://doi.org/10.1016/j.medengphy.2013.05.006
47. He X, Handa J, Gehlbach P, Taylor R, Iordachita I (2014) A submillimetric 3-DOF force sensing
instrument with integrated fiber Bragg grating for retinal microsurgery. IEEE Trans Biomed
Eng 61(2):522–534. https://doi.org/10.1109/TBME.2013.2283501
48. Padma S, Umesh S, Asokan S, Srinivas T (2017) Bite force measurement based on fiber Bragg
grating sensor. J Biomed Opt 22(10):1. https://doi.org/10.1117/1.jbo.22.10.107002
49. Mannoor MS et al (2012) Graphene-based wireless bacteria detection on tooth enamel. Nat
Commun 3. https://doi.org/10.1038/ncomms1767
50. Tatlisoz MM, Canpolat C (2017) Mechanical testing strategies for dental implants. https://doi.
org/10.1007/978-981-10-4166-2
51. Carvalho L, Silva JCC, Nogueira RN, Pinto JL, Kalinowski HJ, Simões JA (2006) Application
of Bragg grating sensors in dental biomechanics. J Strain Anal Eng Des 41(6):411–416. https://
doi.org/10.1243/03093247JSA191
52. Cong TM, Mou RZ, Chen CS, Pan MC (2015) Resonance frequency confirmation for osseoin-
tegration of dental implantation in vitro models with varied cortical thickness. J Med Devices
Trans ASME 9(3):2015–2017. https://doi.org/10.1115/1.4030599
53. Kittur N, Oak R, Dekate D, Jadhav S, Dhatrak P (2020) Dental implant stability and its measure-
ments to improve osseointegration at the bone-implant interface: a review. Mater Today Proc.
https://doi.org/10.1016/j.matpr.2020.08.243
54. Hamed A, Masouleh MT, Kalhor A (2019) Design and characterization of a bio-inspired 3-DOF
tactile/force sensor and implementation on a 3-DOF decoupled parallel mechanism for human-
robot interaction purposes. Mechatronics 66(May 2019):102325. https://doi.org/10.1016/j.mec
hatronics.2020.102325
55. Schmitt, Segert (2008) Towards robot-assisted vitreoretinal surgery: force-sensing micro-
forceps integrated with a handheld micromanipulator. Bone 23(1):1–7. https://doi.org/10.1109/
ICRA.2014.6907035
56. Oliva Uribe D, Schoukens J, Stroop R (2018) Improved tactile resonance sensor for robotic
assisted surgery. Mech Syst Signal Process 99:600–610. https://doi.org/10.1016/j.ymssp.2017.
07.007
57. Banerjee H, Ren H (2018) Electromagnetically responsive soft-flexible robots and sensors for
biomedical applications and impending challenges, 43–72. https://doi.org/10.1007/978-981-
10-6035-9_3
58. Bolat G, Vural OA, Yaman YT, Abaci S (2021) Polydopamine nanoparticles-assisted impedi-
metric sensor towards label-free lung cancer cell detection. Mater Sci Eng C 119(May
2020):111549. https://doi.org/10.1016/j.msec.2020.111549
59. Zhao X et al (2020) Aptamer-based fluorescent sensors for the detection of cancer biomarkers.
Spectrochim Acta Part A: Mol Biomol Spectrosc 119038. https://doi.org/10.1016/j.saa.2020.
119038
56 N. Karnik et al.
60. Upadhyaya AM, Srivastava MC, Sharan P (2020) Integrated MOEMS based cantilever sensor
for early detection of cancer. Optik 165321. https://doi.org/10.1016/j.ijleo.2020.165321
61. Moreddu R, Wolffsohn JS, Vigolo D, Yetisen AK (2020) Laser-inscribed contact lens sensors
for the detection of analytes in the tear fluid. Sens Actuators B: Chem 317(March):128183.
https://doi.org/10.1016/j.snb.2020.128183
62. Ramašauskas L, Meškys R, Ratautas D (2020) Real-time glucose monitoring system containing
enzymatic sensor and enzymatic reference electrodes. Biosens Bioelectron 164(May 2020).
https://doi.org/10.1016/j.bios.2020.112338
63. Dave PK, Rojas-Cessa R, Dong Z, Umpaichitra V (2021) Survey of saliva components and
virus sensors for prevention of COVID-19 and infectious diseases. Biosensors 11(1). https://
doi.org/10.3390/bios11010014
64. Behera S et al (2020) Biosensors in diagnosing COVID-19 and recent development. Sens Int
1(October):100054. https://doi.org/10.1016/j.sintl.2020.100054
65. Nag P, Sadani K, Mukherji S (2020) Optical fiber sensors for rapid screening of COVID-19.
Trans Indian Natl Acad Eng 5(2):233–236. https://doi.org/10.1007/s41403-020-00128-4
66. Drobysh M, Ramanaviciene A, Viter R, Ramanavicius A (2021) Affinity sensors for the
diagnosis of COVID-19. Micromachines 12(4):390. https://doi.org/10.3390/mi12040390
67. Maghded HS, Ghafoor KZ, Sadiq AS, Curran K, Rawat DB, Rabie K (2020) A novel AI-
enabled framework to diagnose coronavirus COVID-19 using smartphone embedded sensors:
design study. In: Proceedings—2020 IEEE 21st international conference on information reuse
and integration for data science, IRI 2020, pp 180–187. https://doi.org/10.1109/IRI49571.2020.
00033
68. Hassantabar S et al (2020) CovidDeep: SARS-CoV-2/COVID-19 test based on wearable
medical sensors and efficient neural networks. arXiv 1–11
69. Bhalla N, Pan Y, Yang Z, Payam AF (2020) Opportunities and challenges for biosensors and
nanoscale analytical tools for pandemics: COVID-19. ACS Nano 14(7):7783–7807. https://
doi.org/10.1021/acsnano.0c04421
70. Xiao N, Yu W, Han X (2020) Wearable heart rate monitoring intelligent sports bracelet based
on internet of things. Meas: J Int Meas Confederation 164:108102. https://doi.org/10.1016/j.
measurement.2020.108102
71. Samartkit P, Pullteap S, Seat HC (2021) Validation of fiber optic-based Fabry–Perot inter-
ferometer for simultaneous heart rate and pulse pressure measurements. IEEE Sens J
21(5):6195–6201. https://doi.org/10.1109/JSEN.2020.3041782
72. Coffen B, Scott P, Mahmud MDS (2020) Real-time wireless health monitoring: an ultra-
low power biosensor ring for heart disease monitoring. In: 2020 International conference on
computing, networking and communications, ICNC 2020, pp 626–630. https://doi.org/10.1109/
ICNC47757.2020.9049814
73. Kaidarova A et al (2020) Laser-printed, flexible graphene pressure sensors. Glob Challenges
4(4):2000001. https://doi.org/10.1002/gch2.202000001
74. Wang TW, Lin SF (2020) Wearable piezoelectric-based system for continuous beat-to-beat
blood pressure measurement. Sensors (Switzerland) 20(3):1–12. https://doi.org/10.3390/s20
030851
75. Poduval RK, Coote JM, Mosse CA, Finlay MC, Desjardins AE, Papakonstantinou I (2021)
Precision-microfabricated fiber-optic probe for intravascular pressure and temperature sensing.
IEEE J Sel Top Quantum Electron 27(4). https://doi.org/10.1109/JSTQE.2021.3054727
76. Jegan R, Jose PSH, Rajalakshmy P, Raj PAC, Kanmani HJ, Nimi WS (2021) Methodological
role of mathematics to estimate human blood pressure through biosensors. Int J Integr Eng
13(1):240–248. https://doi.org/10.30880/ijie.2021.13.01.021
77. Zhao L, Wen Z, Jiang F, Zheng Z, Lu S (2020) Silk/polyols/GOD microneedle based electro-
chemical biosensor for continuous glucose monitoring. RSC Adv 10(11):6163–6171. https://
doi.org/10.1039/c9ra10374k
78. Felix F et al (2020) Objective sensor-based gait measures reflect motor impairment in multiple
sclerosis patients: reliability and clinical validation of a wearable sensor device. Multiple
Sclerosis Relat Disord 39. https://doi.org/10.1016/j.msard.2019.101903
79. Monge J, Postolache O, Alexandre R, De Fatima Domingues M, Antunes P, Viegas V (2020)

Fiber bragg gratings solution for gait assessement. In: I2MTC 2020—international instrumen-
tation and measurement technology conference, proceedings, pp 1–6. https://doi.org/10.1109/
I2MTC43012.2020.9128421
80. Carbonaro N, Arcarisi L, Di Rienzo F, Virdis A, Vallati C, Tognetti A (2020) A preliminary
study on a new lightweight and flexible sensing sock for gait analysis. In: Proceedings of IEEE
sensors, vol 2020-Oct, pp 14–17. https://doi.org/10.1109/SENSORS47125.2020.9278682
Performance Analysis of Diode Clamped
and Flying Capacitor Multilevel Matrix
Converter Used for DFIG-Based Wind
System
G. Pandu Ranga Reddy, D. Mahesh Kumar, K. Rajesh, Y. Chintu Sagar,

and J. Nageswara Rao
Abstract It is impossible to meet the increasing demand for electrical energy with
the available deposits of fossils fuels that are already at the edge of depletion.
Harnessing energy by means of alternative energy sources has been initiated in order
to meet increasing energy demand in the future. Wind energy is found out to be
more potential resource than all other renewable sources. The converters that are
presented in the wind system play a major role for improving the quality of power.
The test system comprises matrix converter attached to DFIG machine and the rotor
of generator is coupled to the grid. The conventional matrix converter is then replaced
with the proposed multilevel matrix converters. THDs of the voltages and currents
at generating and load points are evaluated for the DFIG machine and it can run by
diode clamped and flying capacitor multilevel matrix converters. For managing the
switching devices in the converter circuit, the space vector pulse width modulation
approach is devised. MATLAB/Simulink software is used to graphically show and
evaluate the stator voltages and currents of the DFIG machine.
Keywords DFIG—double fed induction generator · DCMMC—diode clamped

multilevel matrix converter · FCMMC—flying capacitor multilevel matrix
converter · FFT—fast Fourier transformation · THD—total harmonic distortion ·
WECS—wind energy conversion system
G. P. R. Reddy (B)
Deparment of EEE, G. Pullaiah College of Engineering and Technology (Autonomous), Kurnool,
India
e-mail: panurangaeee@gpcet.ac.in
D. M. Kumar
Deparment of EEE, PVKK Institute of Technology, Anantapur, India
K. Rajesh
Deparment of EEE, RGM College of Engineering and Technology, Nandyal, India
Y. C. Sagar
Deparment of EEE, Ashoka Women’s Engineering College, Kurnool, India
J. N. Rao
Deparment of Electrical and Computer Engineering, Mizan-Tepi University, Teppi, Ethiopia
https://doi.org/10.1007/978-981-19-6913-3_4
60 G. P. R. Reddy et al.
1 Introduction
A renewable energy sources are now used as alternatives for conventional sources of
energy and employed into power grid. Distributed generation also known as decen-
tralized generation involves generation in lower capacities by conventional sources
in assistance with renewable sources [1]. Wind energy systems are widely used in
distributed generation in addition to conventional sources. Wind energy generation
systems are more advantageous than other renewable sources in terms of efficiency
and generational capabilities [2].
Static elements such as solar power systems result in PWM output voltage. But
wind energy system employed with dynamic machines such as synchronous and
induction machines result in nearly sinusoidal waveform. DFIG machine is usually
more preferable for a wind energy conversion system as it provides constant speed
irrespective of variations in the wind speed [1]. By controlling power electronic
converters are regulation of rotor input is controlled. Figure 1 shows the inter-
connection of DFIG machine with grid and it can operate by power electronics
converter.
Rotor of DFIG is usually fed by suitable two-stage VSIs having a common DC
link capacitor. Cycloconverters are another choice for a wind energy system based
on DFIG. Pulse width modulation is used to manage power electronic converters
(PWM). Conventional two or three level PWM for power electronic converters leads
to increased THD in rotor currents. This in turn results in increased harmonics in
stator currents thus injecting harmonics into the grid. Thus, harmonics at source side
has adverse affects on loads connected to grid and DFIG [3].
Performance of DFIG can be improved by reducing the harmonic content which
can be achieved by replacement of 2-level and 3-level voltage source inverters by
Fig. 1 DFIG interconnection to grid with power electronic converter

Performance Analysis of Diode Clamped and Flying Capacitor … 61
multilevel matrix converters. Several configurations of multilevel matrix converters

are available. It is important to choose the right configuration of converter that results
in lower harmonic distortion and is feasible with DFIG-based wind system. When
compared to other topologies, the diode clamped multilevel matrix converter and the
flying capacitor multilevel matrix converter have reduced harmonic distortion.
2 Concept of Wind Turbine and Wind Generator
2.1 Modeling of Wind Turbine
The wind turbine generates mechanical energy by converting into kinetic energy [1].
Equation (1) signifies mechanical power of the turbine [3].
1
Pm = ρCp (λ, β)Ar Vw3 (1)
2
where
ρ is the density of air (Kgm3 )

Vw is the speed of the wind (M/S)
Cp is the coefficient of power
λ is the ratio of tip-speed
β is the pitch angle of the rotor blades
and
Ar is the rotor swept area of a wind turbine (m2 ).Here,

116 −16.5 Pm
Cp (λ, β) = 0.5 − 0.4β − 5 e λi = (2)
λi PW
where
1
λi = (3)
1
λ+0.089β
− 0.035
β 3 +1
Theoretically, C p value is set to 0.59. But according to Betz’s Law, the practical
value is set to 0.2–0.4 and rotor pitch angle is a constant [4].
2.2 Modeling of DIF Wind Generator
See Fig. 2.
Fig. 2 Equivalent circuit diagram of DFIG
Their corresponding mathematical equations are [5]
dψds
Vds = Rs i ds + − ωs ψqs (4)
dt
dψqs
Vqs = Rs i qs + + ωs ψds (5)
dt
dψdr
Vdr = Rs i dr + − (ωs − ωr )ψqr (6)
dt
dψqr
Vqr = Rs i qr + + (ωs − ωr )ψdr (7)
dt
Here, ψ denotes the flux linkage and it can be expressed as:
ψds = L ss i ds + L m i dr (8)
ψqs = L ss i qs + L m i qr (9)
ψdr = L rr i dr + L m i ds (10)
ψqr = L rr i qr + L m i qs (11)
and also
L ss = L s + L m (12)
L rr = L r + L m (13)
Rotor swing equation expressed in terms of inertia constant J and damping factor
F is
dωrm
J + Fωrm = (Te − Tm ) (14)
dt
3 Wind Converter
Back-to-back, VSC converters are the most common converters used in technologies
for converting wind energy. This research paper focuses on performance of two
multilevel matrix converters: the flying capacitor multilevel matrix converter and
the diode clamped multilevel matrix converter, which were utilized in place of a
back-to-back converter [6].
3.1 Concept of Diode Clamped Multilevel Matrix Converter
Figure 3 depicts the conventional topology of diode clamped multilevel matrix

converter [7]. This topology is complex and voltage regulation in a 3-bus system
would complicate the modulation strategy. For obtaining higher levels of output
voltage, practical implementation of this configuration is not a promising solution
[8].
A modified DCMMC topology which does not exhibit aforestated problems is
introduced in this paper. As exposed in Fig. 4, this architecture consists of a traditional
2-level converter on the input side for AC–DC conversion and a multilevel converter
on the output side for DC–AC conversion.
3.2 Control Method for Diode Clamped Multilevel Matrix

Converter
Both output voltage and input current are adjusted via space vector pulse width modu-
lation. This strategy is efficient and beneficial in allowing input current and output
voltage switching vectors to be selected. This technique is helpful for decreasing
the harmonics presented in the system. For an adequate small interval of time, the
reference voltage vector may be anything you want it to be calculated with help of
stationary vectors that are produced by the matrix converter. This period of time is
called the converter sample time [9].When the reference voltage vector is rotated
to a different angular point at the next sampling moment, a novel set of voltage
vectors that are stationary is formed. The average output voltage would safely follow
Fig. 3 Conventional topology of DCMMC
the reference voltage if this procedure was extended beyond sampling the whole
waveform of the required voltage vector being produced in sequence.
On the other hand, the phase angle difference between input voltage and current
is calculated using specified stationary vectors [10]. The reference signal is a mix
of nonzero active switching vectors that are contiguous to the reference vector and
one or more zero vectors with proper sampling time in this SVPWM approach.
Figure 5 depicts the vector representation with space vectors, whereas Fig. 6 depicts
the reference vector switching duration (T Z ) calculation as a combination of adjacent
vector durations (T 1 , T 2 ) [11, 12].
Time duration of switching vector is calculated with the following equations
√ _

3.TZ .Vref
π n−1

T1 = sin −α+ π (15)
Vdc 3 3
√ _

3.TZ .Vref
n−1

T2 = sin α − π (16)
Vdc 3
T0 = Tz − T1 − T2 (17)
Fig. 4 Proposed diode clamped multilevel matrix converter
Fig. 5 Switching vector diagram of SVPWM technique

Fig. 6 Arrangement of adjacent vectors as a reference vector
3.3 Concept of Flying Capacitor Multilevel Matrix Converter
Conventional matrix converter with nine switches can be replaced by flying capac-
itor multilevel matrix converter [13]. Figure 7 shows the flying capacitor multilevel
matrix. As shown in Fig. 7 for achieving freewheeling operation, two MOSFETs are
connected antiparallel to diodes connected in antiparallel. Also, the switch count of
this topology is doubled as match up to usual matrix converter [14].
For attaining proper voltage division and to generate multilevel output voltage, it
is necessary to choose six capacitors having equal rating. PWM signals required for
the operation of proposed flying capacitor multilevel matrix converter are generated
by providing identical PWM signal to all the bidirectional switches [15, 16].
3.4 Control Method for Flying Capacitor Multilevel Matrix

Converter
The suggested converter uses a space vector pulse width modulation (SVPWM)
control approach to produce PWM signals for the 18-bidirectional switches [9].
Clarke’s transformation is used to the sinusoidal signal as a reference fundamental
signal to create space vector control signal via sector selection beginning with the
alpha and beta waveforms [17, 18].
⎡ i (t) ⎤
2 1 −√21 −√21 ⎣ a ⎦
i αβ (t) = i b (t) (18)
3 0 − 23 − 23
i c (t)
Fig. 7 Flying capacitor multilevel matrix converter topology
Alpha beta waveform represents the sectors to be taken for space vector trajectory.
Figure 8 shows the sector selection by alpha and beta components. A zero vector is
to be considered in addition to the six vectors for suitable pulse generation [13, 19].
Proper selection of sector aids to generation of suitable gating signals. The proposed
converter switching states are presented in Table 1. It corresponds to nine switching
states generating of Phase U resulting in multilevel inversion operation [20].
4 Simulation Results and Discussion
In this paper, the test system for evaluation and comparison of DCMMC and FCMMC
is given in Fig. 9. The test system consists of DFIG-based WECS employing matrix
converter and it is supplying power to load. Up to t = 0.2 s, the wind energy conversion
equipment was not linked to the source. After t = 0.2 s, the source is linked to the
wind plant’s DFIG machine-based power electronic converter. The load is powered
by a three-phase source up to t = 0.2 s. WECS uses DFIG with a power electronic
converter to supply power to the RL load after t = 0.2 s.
In the test system, it can be observed that conventional matrix converter is used
to induce voltage into the rotor circuit. This converter will be replaced with the
proposed multilevel matrix converters and THD calculations of the DFIG are carried
out with FFT analysis tool in MATLAB/Simulink Power GUI software. By running
the simulation with the space vector PWM technique, the THD values of voltages
and currents in percent at different places such as B1, B2 and B3 are determined
Fig. 8 Switching sector calculation with alpha and beta components
Table 1 Switching states of one of the phase in flying capacitor multilevel matrix converter
Changing mode (On: 1, Off: 0) Voltage value at phase ‘A’
SAa1 SAa2 SBa1 SBa2 SCa1 SCa2
1 1 0 0 0 0 VA
1 0 0 1 0 0 (VA + VB )/2
1 0 0 0 0 1 (VA + Vc )/2
0 1 1 0 0 0 (VA + VB )/2
0 0 1 1 0 1 VB
0 0 1 0 0 1 (VB + Vc )/2
0 1 0 0 1 0 (VA + Vc )/2
0 0 0 1 1 0 (VB + Vc )/2
0 0 0 0 1 1 Vc
using MATLAB/Simulink software. Here, B1 is the three-phase source point or grid

point, B2 is the DFIG machine connected point and B3 is the load point. In the test
system, initially DCMMC is replaced in place of conventional matrix converter after
FCMMC is replaced and compared the performance of the converters.
Simulink model of the test system obtained by replacing of the basic matrix
converter with diode clamped multilevel matrix is as exposed in Fig. 10. Figure 11
shows a test setup for a DIFG-based wind system developed by substituting the basic
matrix converter with an FCMMC.
Fig. 9 Test system with DFIG fed with conventional matrix converter topology
Fig. 10 Diode clamped multilevel matrix converter fed DFIG wind generator
4.1 Simulation Results of Diode Clamped MMC Based

on SVPWM for the DFIG
The subsequent figures demonstrate the Simulink model results of DFIG-based wind
system using SVPWM by diode clamped multilevel matrix converter. The %THD
of DFIG source currents and voltages in WECS driven by DCMMC is shown in
Fig. 12. It gives information of THD values in percent of source currents and voltages
as 2.55 and 0.46 correspondingly. The percent THD values of DFIG voltages and
currents under load in WECS operated by DCMMC are shown in Fig. 13. It gives
Fig. 11 Flying capacitor multilevel matrix converter fed DFIG wind generator
information of THD values in percent of load currents and voltages as 0.54 and 0.48
correspondingly.
Figure 14 demonstrates the THD values of DFIG source currents and voltages
in WECS employing FCMMC. Figure 14 gives the %THD values of DFIG source
currents and voltages as 1.83 and 0.40 correspondingly. Figure 15 depicts the %THD
rates of load currents and voltages of DFIG-based WECS by means of proposed
Fig. 12 Source voltage and current harmonic spectrum
Fig. 13 Load voltage and current harmonic spectrum

Fig. 14 Source voltage and current harmonic spectrum
Fig. 15 Load voltage and current harmonic spectrum
converter FCMMC. From Fig. 15, the load currents and voltages %THD are 0.41
and 0.39 correspondingly.
4.2 Simulation Results of a Flying Capacitor MMC Based

on SVPWM for the DFIG
THDs of voltages and currents in percentage are shown in Table 2 at various points
in the system where suggested converters are used.
Table 2 %THDs of
Parameter DCMMC FCMMC
DCMMC and FCMMC with
SVPWM Point B1: source voltage 0.46 0.40
Points B1: source current 2.55 1.83
Point B2: integrated voltage 5.89 5.66
Point B2:integrated current 2.72 1.76
Point B3:load voltage 0.48 0.39
Point B3:load current 0.54 0.41
5 Conclusion
The converters performance can be observed by modeling the system using

MATLAB/Simulink software. In this paper, the diode clamped and flying capac-
itor multilevel matrix converter models are designed for DFIG wind power produc-
tion systems, and the converters’ performance is evaluated. The voltage and current
%THD values at different points in wind energy conversion system operated by
DFIG machine are calculated with the help of FFT analysis. The flying capacitor
multilevel matrix converter architecture is thought to be more competent than the
diode clamped multilevel matrix converter, according to the FFT analysis tool.
References
1. Casadei D, Serra G, Tani A, Zarri L (2020) Matrix converter modulation strategies; a new
general approach based on space-vector representation of the switch state. IEEE Trans Ind
Electron 49(2)
2. Panwar V, Kaur T (2018) Overview of renewable energy resources of India. Int J Adv Res
Electr Electron Instrum Eng 3(2):7118–7125
3. Reddy GPR, Kumar MV (2015) Analysis of wind energy conversion system employing DFIG
with SPWM and SVPWM type converters. J Electr Eng 5(4):95–106
4. Reddy GPR, Kumar MV (2017) Two level versus matrix converters performance in wind energy
conversion system employing DFIG. J Inst Eng (India): Ser B. Springer Publications, ISSN:
2250-2106
5. Pena R, Clare JC, Asher GM (1996) Doubly fed induction generator using back-to-back PWM
converters and its application to variable speed wind-energy generation. IEE Proc-Electr Power
Appl 143(3):231–241
6. Jacob JT (2014) Review on high power multilevel-matrix converters. Int J Adv Res Electr
Electron Eng 3(1)
7. Khajehoddin SA, Bakhshai A, Jain P (2017) A sparse multilevel matrix converter based on
diode-clamped topology. IEEE Trans 0197-2618/07/$25.00 © 2017
8. Reddy GPR, Kumar MV Diode clamped multilevel matrix converter for DFIG based wind
energy conversion system. IOSR J Electr Electron Eng 44–52. ISSN: 2320-3331
9. Lee MY, Wheeler P, Klumpner C (2015) Space-vector modulated multilevel matrix converter.
IEEE Trans Ind Electron 57(10):3385–3394
10. Sun Y, Xiong W, Su M, Li X, Dan H, Yang J (2014) Topology and modulation for a new
multilevel diode-clamped matrix converter. IEEE Trans Power Electron 29(12)
11. Ghoni R, Abdalla AN (2014) Analysis and mathematical modeling of space vector modulated
direct controlled matrix converter. J Theor Appl Inf Technol 27–35
12. Gruson F, Le Moigne P, Delarue P, Videt A, Cimetiére X, Arpillière M (2013) A simple
carrier-based modulation for the SVM of the matrix converter. IEEE Trans Ind Inform 9(2)
13. Wheeler P, Xu L, Lee MY, Lee E, KIumpner C, Clare J (2019) A review of multi-level
matrix converter topologies. In: Conference on power electronics, machines and control group,
University of Nottingham, pp 286–290. Published in IEEE Digital Explore
14. Shi Y, Xu Y, He Q, Wang Z (2005) Research on a novel capacitor clamped multilevel matrix
converter. IEEE Trans Power Electron
15. Miura Y, Inubushi K, Ito M, Ise T (2014) Multilevel modular matrix converter for high voltage
applications. IEEE Digital Explore, 2014, pp 4690–4696
16. Shukla RD, Singh A, Singh SP (2012) Generators for variable speed wind energy conversion
systems: a comparative study. IJESCC 3(2):103– 117
17. Xu L, Clare JC, Wheeler PW, Lee E, Li Y (2012) Capacitor clamped multilevel matrix converter
space vector modulation. IEEE Trans Ind Electron 59(1):105–115
18. Odaka A, Sato I, Ohgushi H, Tamai Y, Mine H, Ito J (2005) A PAM: control method for matrix
converter. In: Proceedings of the 2005 Japan industry applications society conference
19. Hojabri H, Mokhtari H, Chang L (2011) A generalized technique of modeling, analysis, and
control of a matrix converter using SVD. IEEE Trans Ind Electron 58(3):949–959
20. Nguyen TD, Lee H-H (2014) Multilevel indirect matrix converter with carrier-based pulse
width modulation. Published in IEEE Digital Explore, 2014, pp 3318–3323
Real Time Feedback System for Speech
Dysfluency in Children
Jennifer C. Saldanha and Rohan Pinto
Abstract Stuttering and Cluttering is a speech dysfluency disorder found in people

which makes them to speak in a disordered way leading to prolongations, repetitions,
pauses, and blocks. The main aim of this work is to develop a real time system that
will help people with this disorder to improve their speech. The first phase of this
work deals with the three methods that can treat the stutterers, i.e., delayed audi-
tory feedback, frequency altered feedback, and metronome generation. The second
phase deals with stuttered speech analysis to find the severity of stuttering using
prolongation, repetition, and silence blocks. A model is developed using SoX audio
processing toolbox. The processor used in the system is Raspberry Pi B. For the
benefit of the user a GUI is implemented using Tkinter. A database is created with
a given Kannada passage with 80 samples which includes both the stuttered and
normal speech samples.
Keywords Delayed auditory feedback · Frequency altered feedback · k-nearest

neighbor · Mel frequency Cepstral coefficients
1 Introduction
Stuttering and cluttering is commonly known as stammering. It is a speech dysflu-

ency disorder. Here the flow of speech is interrupted by repetitions and prolongations
of sounds, syllables, words, and phrases without conscious control. There are even
involuntary silent pauses or blocks. In this disorder the person who has the problem
of stuttering is unable to produce sounds. The impact of this disorder on a person’s
emotional state of mind and normal functioning can be severe [1]. This may include
J. C. Saldanha · R. Pinto (B)

Department of Electronics and Communication, St Joseph Engineering College, Mangaluru, India
e-mail: rohanp@sjec.ac.in
J. C. Saldanha
e-mail: jennifers@sjec.ac.in
https://doi.org/10.1007/978-981-19-6913-3_5
76 J. C. Saldanha and R. Pinto
fears of enunciating specific vowels or consonants, there may also be a fear of being
caught stuttering in a social gathering, isolating one-self, not being able to express
ones feelings, shame, low self-esteem, etc. This disorder can be basically classified
into developmental, psychogenic, and neurogenic stuttering. Developmental stut-
tering is found in all people in their early ages which reduces on its own in most
cases but stuttering retains in some cases [2]. This is the most common type of stut-
tering. A normal person may start stuttering after being subjected to a shock or fear
or such situations which is called psychogenic stuttering. A person may start stut-
tering after a neurological trauma which is called neurogenic stuttering. Out of these,
developmental and psychogenic stuttering can be improved effectively by employing
different methods whereas neurogenic stuttering cannot be improved.
The objective of this work is to develop a real time device which facilitates the
people suffering from stuttering and cluttering to reduce the disorder and improve
their speech by using three techniques namely Delayed Auditory Feedback (DAF),
Frequency Altered Feedback (FAF), and metronome generation. DAF is a technique
used in cases where the speech of the stutterer is fed to his ears after a delay. In FAF
method the person is made to hear his own pitch shifted sound while speaking and in
metronome generation, a continuous impulse signal is generated at regular intervals
according to which the person is made to speak.
The paper is organized as follows: Sect. 2 gives a brief review of the previous
work with the benefits and short comings of the different methods used in the work.
The methodology of this work is described in Sect. 3. Section 4 gives a detailed
analysis of the obtained results. The paper is concluded in the Sect. 5.
2 Related Work
Various methods have been tried in the literature to implement DAF and FAF.
Bahadorinejad and Almasganj [3] discussed the implementation of anti-stuttering
device using AVR ATmega128 and hearing aid technology. In this work FAF is used
to feed once own pitch shifted voice to the user as an auditory feedback. Ramteke
et al. [4] focuses on detection of the repetition in the stuttered speech using shimmer,
MFCC, and formant features. These features are extracted from the isolated words
of 27 s speech data. The dynamic time warping technique with a suitable threshold
value is used for the detection of the repetitions in the stuttered speech. Macleod
et al. [5] discussed about the implementation of the combination of DAF and FAF in
stuttered speech therapy and found that the combination obtained better results than
DAF or FAF when used separately. DAF with 50 ms delay and FAF with a downward
shift of one half octave was used in this study. Kadam et al. [6] proposed a human
voice recognition system that uses Mel frequency cepstral coefficients (MFCC) with
artificial neural networks (ANN) to analyze and detect human voice in different appli-
cations such as medical, military, and telecommunication with Raspberry Pi as the
processor. MFCC features are extracted from the recorded voice samples. Built-in
ANN is trained with the training data-table created using MFCC feature. Finally, the
Real Time Feedback System for Speech Dysfluency in Children 77
test data are given to ANN which compares the test data with the training data and
recognizes the speaker. Chee et al. [7] discussed about the classification of stuttered
speech from the normal speech by acoustically modeling frequently occurring repeti-
tions and prolongations in stuttered speech using MFCC. The performance of MFCC
in modeling repetitions and prolongations in stuttered speech was examined using
linear discriminant analysis and k-NN classifiers. Czyzewski et al. [8] discusses about
the detection of the stop gaps, repetition, and prolongation of a syllable for counting
the stuttering events. ANN is used as a classifier for the automatic recognition of stut-
tering disorder in the extracted feature set. Sathya and Chandra [9] discussed about
using a feature extraction algorithm known as MFCC for the detection of repetitions
and prolongations in a stuttered speech. Here Linear Discriminant Analysis (LDA)
classifier and k-NN classifier are used to measure the effectiveness in recognition of
a stuttered speech. An average accuracy of 90% is obtained using these classifiers.
Ai and Yunus [10] introduced four different types of anti-stuttering devices which
are based on DAF, FAF, masked auditory feedback (MAF), and combined/multiple
feedback. These algorithms were developed to assist a speech language pathologist
for Malay speech therapy in Malaysia. Saranya and Ravi [11] proposed a low cost
portable device based on digital signal processing to alleviate the problem of stut-
tering. Authors implemented DAF with the delays in the range of 50–200 ms. FAF
was implemented with varying pitches such as a quarter, half or full octave shift
resulting in 55–74% decrease of stuttering in the short reading tasks. Yaruss [12]
discusses about the online real time data collection to analyze the frequency of the
different types of speech disfluencies produced by the people who stutter and also
proposes the frequency count values to be used by the clinicians in stuttered speech
therapy. Borsel et al. [13] investigated the effect of DAF on the people with the
stuttering disorder. The patients with the stuttering disorder were exposed to DAF
for a period of 3 months. Significant improvement was observed in the stuttering
after 3 months as the percentage of stuttered words in the people who were exposed
to DAF was reduced drastically. Metz et al. [14] discusses about the modified Van
Riperian procedure in the analysis of the stuttered speech to investigate its efficiency
in the post and pre stuttering therapy. The experiment was carried out by involving
nine adult stutterers. It was found that the stuttering frequency is reduced signifi-
cantly over the course of therapy using the modified Van Riperian procedure. Riley
[15] proposes a device for assessment of the stuttering severity by analyzing the
stuttered speech samples of 109 children and 28 adults. The assessment of severity
was mainly based on the frequency and duration of the prolongation and repetition
of the speech segments in a given passage. Leith and Chmiel [16] verified the effect
of DAF in improving the speech rate of a stutterer. The authors also focuses on two
major programs: rate control program and fluency program to increase the fluency
and rate of the speech in a stutterer. Chee et al. [17] investigated the performance
of artificial neural networks (ANN), hidden Markov models (HMM), and support
vector machine (SVM) classifiers for the detection of the stuttering. The highest
classification accuracy of 96% is obtained with HMM, where as SVM and ANN
yielded almost similar accuracies with 94.9% and 94.35%, respectively.
The usefulness of DAF and FAF in treating the stutter and clutter speech disorder
can be seen in the literature. Hence, in this study the effect of DAF and FAF are
observed for a Kannada passage in the age group of 10–12 years including both male
and female gender.
3 Proposed Method to Treat Stuttering
This work involves two phases in detection and treatment of the stuttering. The
first phase discusses speech therapy where three feedback methods for assisting the
stutterer to improve fluency are discussed. The second phase is on different techniques
used for speech analysis for assessment of severity of stuttering.
3.1 Speech Therapy
Speech therapy can help patients learn to speak clearly. This helps them to feel
confident and less frustrated while communicating. The general block diagram of
the speech therapy system is shown in Fig. 1. Raspberry Pi B is used as the processor
in this system. The three techniques namely DAF, FAF and metronome generation
are used in this work for speech therapy. The user is intended to choose one among
the three therapy techniques and follow the instructions as provided by the therapist.
Delayed Auditory Feedback: The speech input is given to the processor via a micro-
phone with a good quality reception. With the help of a library called the SoX, the
signal is shifted in the time domain causing a delay. The delay is kept variable which
will be different for different people.
Frequency Altered Feedback: The noise free signal is Fourier transformed to obtain
a frequency domain signal. The frequency can be up-shifted or down-shifted. The
output is then inverse Fourier transformed to get back the time domain signal. The
time domain audio signal is then sent back to the user as a feedback.
Fig. 1 Block diagram of speech therapy

Metronome Generation: A sine wave of a particular frequency is generated for

100 ms and the same sine wave is repeated after a specified interval of time. The user
is made to utter a word according to the metronome generated.
3.2 Speech Analysis
The analysis of the stutterer’s speech is carried out to classify the patients according
to their criticality and provide the treatment accordingly. The analysis is also done
to measure the improvement during the course of the therapy. The categorization in
presently done manually by the therapists by listening to the audio recorded during
the therapy sessions. The judgements may be different by different therapists and lead
to the lack of standardization. Moreover, listening to the recordings of the patients
and classifying them manually is a hectic task. The introduction of a standard analysis
method can help in getting the same results even if the patient is treated by different
therapists with the added advantage of reducing the work of the therapists in listening
to the audio sample and classifying. The block diagram of the speech analysis is
shown in the Fig. 2.
3.2.1 Preprocessing
Framing and Windowing: The time period for a frame is fixed and the number of
samples in a frame is obtained according to the time period and the sampling rate.
Each frame is multiplied with a window function of the length of frame size, giving a
finite length weighted version of the original signal to avoid a discontinuity between
any two frames. This work uses hamming window with 50% overlap and the audio
signals have sampling rate of 16 kHz. The time period of the frame is 25 ms. So the
frame length is equal to 400 samples. Hamming Window weight equation is given
by,

2π n
w(n) = 0.54 − 0.46 cos (1)
M
where M is the length of the window.

Pre-Emphasis: In speech processing, the original speech signal has a greater amount
of low frequency energy content compared to the high frequency energy content.
Fig. 2 Block diagram of speech analysis

Pre-emphasis is used to emphasize the high frequency content of a signal thereby

improving the performance of the system. From the spectrum of the speech sounds,
one can notice that less energy is present in the highest frequencies, with an overall
decreasing slope. The goal of the pre-emphasis filter is to counterbalance this, and
flatten the spectrum. It is also used to increase the high frequency formants in the
speech. Pre-emphasis can be implemented using a simple filter as shown in Eq. 2.
y(n) = x(n) − αx(n − 1) (2)
where α is a non-zero constant close to 1. (In speech processing application α is

commonly in the range of 0.95–0.97).
3.2.2 Feature Extraction
Mel Frequency Cepstral Coefficients: Fast Fourier Transform (FFT) is applied to

transform every frame with 400 samples from the time domain to the frequency
domain. Since there are 400 samples, 512 point discrete Fourier transform (DFT) is
applied. FFT is a fast and efficient method to calculate the DFT. The power spectrum
of each frame is calculated. Next step in the computation of MFCC is Mel frequency
warping. This step involves the creation of a filter bank with triangular band pass
filters placed at a specified intervals. It involves converting the Hertz scale to the
Mel scale because Mel scale represents the human ear perception [18]. Equally
spaced 13 bandpass filters are used in this work. The filter bank energies are obtained
by multiplying each filter bank with the power spectrum and then adding up the
coefficients. The 13 values obtained represent the energy in each bandpass filter. The
logarithm of all the 13 filter bank energies is calculated to obtain the log filter bank
energies. This is done so that the multiplication gets converted to the addition in the
logarithmic scale and thus, the envelope can be detected easily by applying a filter.
The discrete cosine transform is applied instead of the inverse Fourier transform
due to its computational efficiency. It reduces the steps by not calculating for the
redundant values and uses only real numbers instead of the complex numbers unlike
the inverse discrete Fourier transform (IDFT) which uses complex numbers. It is
thus, a better replacement for the inverse Fourier transform.
Prolongations: It is basically defined as involuntary lengthening of speech sounds.
It is also one of the types of speech dysfluencies [19]. Prolongation is found out by
finding the power of each frame and comparing the power with the next consecutive
frame. If the power difference between consecutive frames is less than a threshold
value for a certain number of frames, prolongation can be easily found out. The
number of prolongations occurred in the sentence as well as the duration of the
prolongations are noted for reference. The algorithm to analyze and determine the
prolongation is as follows:
Algorithm:
Step 1: Start
Step 2: Take the Pre-Processed output as an input.
Step 3: Compute the power of the individual frames and store the output in an array
a[].
Step 4: Find the power difference between consecutive frames (a[i + 1]) − a[i]) and
store the result in another array b[].
Step 5: If the power difference is less than 1 mW for 25 consecutive frames or more,
consider it as a prolongation else do not consider it as prolongation.
Step 6: Store and output the result.
Step 7: Stop.
Silence Blocks: They are said to occur when sound or air stops in the lungs, throat
or mouth, lips or tongue. This is said to occur when a person is repeating the same
sound and is unable to utter the first or the next sound, the person is said to have
been stuck. The word ‘stuck’ here basically defines the meaning of a ‘block’. The
silences are found at the low power in the periodogram for more than a specific
number of consecutive frames. The number of silence blocks as well as the duration
of the silence blocks are noted for further reference. The algorithm to analyze and
determine silence blocks is as follows:
Algorithm:
Step 1: Start
Step 2: Take the Pre-Processed signal as an input.
Step 3: Compute the power of the individual frames and store the output in an array
a[].
Step 4: If the power is less than 10 mW for 60 consecutive frames or more, consider
it as a silence block else do not consider it as silence block.
Step 6: Store and output the result.
Step 7: Stop.
Repetitions: A stutterer tends to repeat certain syllables or words or part of phrases.
The criticality of his stuttering depends on the number of repetitions he makes [19].
To count the number of repetitions, syllable counting method is used. The patient
is made to read a prescribed passage for which the number of syllables are known.
By segmenting the syllables in the person’s speech and counting them allows to
evaluate the severity of the stuttering. The number of syllables in the predefined
passage which is already known is compared with the number of syllables in the
speech sample. The results obtained can indicate the number of repetitions done.
The syllable segmentation is done by a sinusoidal signal modeling technique. The
position of the maximum of the amplitude trajectory is the same for each syllable.
This makes comparison between the different syllables easy.
3.2.3 Classifier
Classification is performed using k-NN classifier. This classifier works on the prin-
ciple of finding the nearest sample in the training dataset. In this work Euclidean
distance metric is used for the comparison between the test feature vectors with the
training feature vectors. The value of k is chosen to be 1, due to which the test sample
is simply assigned to the class of that single nearest neighbor. 50% of the total data
set is used for the training, and the remaining 50% is used for testing the classifier
accuracy.
3.3 Database
Stuttered speech and normal speech samples of the children of the age group 10–
12 years were recorded using a Sony ICDX140 voice recorder. A total of 40 samples
were collected of children suffering from dysfluency due to stuttering and 40 samples
were collected of children without any speech disorder. The speech samples were
collected for the given Kannada passage.
(Bengaluru is one of the biggest cities in our state. This city is called Mumbai of
Karnataka. This is one of the biggest cities in India. People come from different
states and different places to see this city. Other than this, people come to visit
Bellur, Jog, and Nandi in our state. Ragi is grown in this state).
4 Results and Analysis
4.1 Speech Therapy Implementation
In this work DAF was implemented with the delay of 100 ms. FAF was implemented
with 300 Hz frequency up-shift and metronome generation was carried out using
a sinusoidal signal with the time period of 1 ms using an audio processing library
called the SoX. The working interface was set up using shell script. Figures 3, 4, and
5 show the implementation of delayed auditory feedback, frequency altered feedback
and metronome generation using the shell script in SoX.
The user is intended to select one among the different therapy techniques as
suggested by the therapists along with the prescribed specifications. The delay,
frequency shift, and metronome speed is decided by the therapist. Figure 6 shows
the GUI implementation of these techniques using Tkinter.
For example, if the user selects ‘Metronome–0.8 s’ using the touch enabled TFT
display, the metronome is generated every 800 ms and the person is intended to utter
a word once in every 800 ms. Similarly, if the user selects ‘DAF—200 ms’, the delay
between the users speech and his perception is set to 200 ms.
Fig. 3 DAF with a delay of 100 ms
Fig. 4 FAF with 300 Hz frequency up-shift

Fig. 5 Metronome generation with time period of 1 ms
Fig. 6 GUI implemented

using Tkinter
4.2 Speech Analysis
During the preprocessing stage the speech samples are recorded at a sampling
frequency of 16 kHz and quantized at 16 bit pulse code modulation (PCM). The
samples are then passed through the feature extraction stage and later used as the
inputs for the classifier to train the model and test the samples.
Figures 7a and b show the stuttered speech and normal speech for the first sentence
in the passage given in the Sect. 3.3. From the
figures it is observed that the stuttered speech sample has many silence intervals when
compared with the normal speech sample. The speech samples are converted into
small segments of speech for further processing using hamming window of 25 ms
duration or 400 samples. Figure 8 shows the windowed speech segment with 400
samples.
Fig. 7 a Speech signal of a normal person. b Speech signal of a stutterer
Fig. 8 Windowed speech frame having 400 samples
The plot of the pre-emphasis of the windowed frame is as shown in Fig. 9. The
plot clearly shows that the amplitude of the lower frequency signals being attenuated
and that of the higher frequencies being amplified.
In order to compute MFCC, the power spectrum of the windowed speech segments
is obtained. Figure 10 shows the power spectrum of a frame of a stuttered speaker.
Figure 11 shows the Mel filter bank with 13 triangular filters used in the compu-
tation of MFCC. The filter bank has number of filters placed at the lower frequencies
and very few filters at the higher frequencies to mimic the human auditory response
Fig. 9 Emphasis plot
Fig. 10 Power spectrum of a frame of a stuttered speaker

which is more sensitive to the lower frequencies below 1 kHz than for the higher
frequencies. The log energies of these filters constitute MFCC’s. Figure 12 shows
plot of Mel frequency cepstral coefficients of a speech frame.
Fig. 11 Mel filter bank with 13 triangular filters
Fig. 12 MFCC of a frame

Fig. 13 Euclidean distance
The average of the MFCC is taken by calculating the average of each individual
coefficients with the corresponding coefficients in the other frames throughout the
audio which is used as the feature vector for the classification. These feature vectors
obtained are used during the classification stage to compute the Euclidean distance.
The Euclidean distance of the average MFCC of the test sample from the training
samples are calculated and is assigned with the label of the nearest neighbor. The
plot of the Euclidean distance of the test sample from the other training samples is
as shown in Fig. 13.
The results obtained above were not very satisfactory as they did not yield the
expected results. The accuracy of the classification using k-NN classifier was 22.23%
which is found to be very low to be used in the real time applications. Mel frequency
cepstral coefficients are found to be effective in speech/speaker recognition but these
features did not obtain significant results in the recognition of stuttering. Hence the
effectiveness of the alternate features such as prolongation, repetition and silence
intervals were investigated. Prolongation is found out by finding the power of each
frame and comparing the power with the next consecutive frame. The detection of
prolongation is as shown in Fig. 14.
The silences are found at the low power in the periodogram for more than a specific
number of consecutive frames. Figure 15 depicts the detection of the silence blocks
in a sample.
By segmenting the syllables in the person’s speech and counting them allows
to evaluate the severity of the stuttering. The number of syllables in the predefined
passage which is already known is compared with the number of syllables in the
speech sample. The results obtained can indicate the number of repetitions done.
Figure 16 shows the spectrogram of the speech sample as well as the segmented
syllables represented in a red color.
Fig. 14 Detection of prolongation
Fig. 15 Silence block detection
It was evident from the analysis of the stuttered speech that counting the number of
prolongations, silence intervals, and repetitions in a given passage can be very useful
in the recognition of stuttering, and to determine the severity level of stuttering.
Fig. 16 Detection of repetitions
5 Conclusion
The three techniques namely DAF, FAF, and metronome generation were imple-
mented using Raspberry Pi as the processor and audio processing library called the
SoX. Graphical User Interface (GUI) for the above application was implemented
using Tkinter–a Python-based GUI Toolkit. To measure the improvement in the
speech, the feature vectors of the speech samples were extracted using MFCC. K-
NN classifier was used to classify stuttered and normal speech using the MFCC
feature vector. Using only MFCC as a feature did not prove to yield the expected
results. Hence silence blocks, repetition, and prolongation were used as additional
features. The speech input was analyzed for the occurrence of prolongation, silence
interval and repetition and for the number of times these events occur in speech
utterances. These features found to be effective in the detection and assessment of
the severity level of stuttering.
References
1. Maguire GA, Yeh CY, Ito BS (2012) Overview of the diagnosis and treatment of stuttering. J
Exp Clin Med 4(2):92–97
2. Herder C, Howard C, Nye C, Vanryckeghem M (2006) Effectiveness of behavioral stuttering
treatment: a systematic review and meta-analysis. Contemp Issues Commun Sci Disord 33:61–
73
3. Bahadorinejad A, Almasganj F (2012) Delayed auditory feedback for speech disorders. In:
IEEE international conference on biomedical engineering (ICoBE), 2012, pp 585–588
4. Ramteke PB, Koolagudi SG, Afroz F (2016) Repetition detection in stuttered speech.
In: Proceedings of 3rd international conference on advanced computing, networking and
informatics, New Delhi. Springer, pp 611–617

5. Macleod J, Kalinowski J, Stuart A, Armson J (1995) Effect of single and combined altered
auditory feedback on stuttering frequency at two speech rates. J Commun Disord 28(3):217–228
6. Kadam M, Rao G, Sakshi AP (2018) MFCC feature extraction for speech recognition with
hybrid application. Int J Adv Res Ideas Innov Technol 4(3):1380–1384
7. Chee LS, Ai OC, Hariharan M, Yaacob S (2009) MFCC based recognition of repetitions and
prolongations in stuttered speech using k-NN and LDA. In: 2009 IEEE student conference on
research and development (SCOReD), Nov 2009, pp 146–149
8. Czyzewski A, Kaczmarek A, Kostek B (2003) Intelligent processing of stuttered speech. J
Intell Inf Syst 21(2):143–171
9. Sathya MAJ, Chandra E (2015) Types and tools available for fluency disorder—speech therapy.
Int J Adv Res Comput Sci Softw Eng 5(2)
10. Ai OC, Yunus J (2006) Overview of a computer-based stuttering therapy. In: Regional post-
graduate conference on engineering and science (RPCES 2006), Johore, 26–27 July 2006, pp
207–211
11. Saranya R, Ravi N (2016) A low cost portable device to alleviate the problem of stuttering. Int
J Emerg Technol Comput Sci Electron (IJETCSE) 21(3)
12. Yaruss JS (1998) Real-time analysis of speech fluency. Am J Speech-Lang Pathol 7(2):25–37
13. Van Borsel J, Reunes G, Van den Bergh N (2003) Delayed auditory feedback in the treatment
of stuttering: clients as consumers. Int J Lang Commun Disord 28(2):119–129
14. Metz DE, Onufrak JA, Ogburn RS (1979) An acoustical analysis of stutterer’s speech prior to
and at the termination of speech therapy. J Fluency Disord 4(4):249–254
15. Riley GD (1972) A stuttering severity instrument for children and adults. J Speech Hear Disord
37(3):314–322
16. Leith WR, Chmiel CC (1980) Delayed auditory feedback and stuttering: theoretical and clinical
implications. Speech Lang 3:243–281
17. Chee LS, Ai OC, Yaacob S (2009) Overview of automatic stuttering recognition system.
In: Proceedings of the international conference on man-machine systems (ICoMMS), Batu
Ferringhi, Penang, Malaysia, 11–13 Oct 2009, pp 1–6
18. Jhawar G, Nagraj P, Mahalakshmi P (2016) Speech disorder recognition using MFCC. In: IEEE
international conference on communication and signal processing (ICCSP), Melmaruvathur,
India, 24 Nov 2016, pp 0246–0250
19. Chee LS, Ai OC, Hariharan M, Yaacob S (2009) Automatic detection of prolongations and repe-
titions using LPCC. In: IEEE international conference for technical postgraduates (TECHPOS),
Dec 14 2009, pp 1–4
Nonlinear Model-Predictive Control
Using First-Principles Models
R. Russell Rhinehart
Abstract The common attribute in horizon-predictive control is the use of a dynamic

model of the process and optimization to plan a sequence of future control actions
to best make the model match a desired future path, while avoiding constraints.
This paper describes a structure to use first-principles process models in horizon-
predictive control, which contrasts the conventional use of linear empirical models
derived from step-testing the process.
Keywords Process control · Nonlinear · Horizon predictive
1 Introduction
Model-predictive control (MPC) has many alternate names and acronyms. Some are
DMC (dynamic matrix control), HPC (horizon-predictive control), IdCom (Identify
and Command), and APC (advanced process control). The common attribute is that
they use a dynamic model of the process and use optimization to plan a sequence
of future control actions to best make the model match a desired future path, while
avoiding constraints. The first control action (manipulated variable value, MV), in
the plan, is implemented. The entire plan is not executed, just the first step. Then after
one control interval, the future MV sequence plan is recalculated to accommodate
changes in set points, disturbances, and process-model mismatch. And, again, the
entire plan is not executed, just the first step.
The models in MPC can reveal both disturbance and interaction, as well as dead-
time and other ill-behaved dynamics. The optimization in MPC compensates for the
model features and also has override action and can choose best MVs to manage the
process when there are extra options. Although such features could be implemented
in classic ARC (advanced regulatory control) with feedforward, decouplers, ratio,
override, etc., once an implementation exceeds about 2 inputs and 2 outputs, the
R. R. Rhinehart (B)
Oklahoma State University, Stillwater, OK 74074, USA
e-mail: russ@r3eda.com; rrr@okstate.edu
https://doi.org/10.1007/978-981-19-6913-3_6
94 R. R. Rhinehart
complexity and tuning of the number of model coefficients become very compli-
cated to manage. MPC provides a unified structure, with only one tuning coefficient
for each controlled variable (CV).
A rule of thumb is that MPC (or any control strategy with equivalent features)
will halve the CV deviations associated with conventional regulatory control. This
reduction in variance permits assigning set points that are closer to specification
limits, which translates to any of many economic benefits—higher throughput, less
waste, lower energy and raw material use, etc. Although relatively expensive to
license an MPC product, and although the process step testing to develop empirical
models takes time and moderately upsets the process, MPC has become accepted as
a best control approach in continuous chemical processing.
Normally, the MPC models of the process response are Finite Impulse Response
(FIR) models, a vector of normalized responses to the steps in both control action
and disturbances representing a linear response. Alternately, models could be second-
order plus deadtime (SOPDT), also linear. And commonly, the optimization used to
plan a future sequence of manipulated variable (MV) moves is Linear Programming
(LP). Although many commercial products offer empirical models with nonlinear
gain (such as neural networks based on steady-state responses, or even heuristically
set gains from interviewing operators and engineers), most models remain linear in
the dynamics (lags and delays). Because LP seeks to place the control action on a
constraint, an extreme boundary, and requires linear objective function and constraint
models, many products use alternate optimizers, such as Successive Quadratic.
One issue is that chemical processes are nonlinear and non-stationary. Both gains
(sensitivities and interactions) and dynamics (time constants and delays) change with
operating conditions, product specification, internal tank levels, piping reconfigura-
tions, equipment switching, etc. When the process operating conditions or process
attributes change, linear models become significantly mismatched to the process and
require model recalibration (recurring process step testing). It would be desirable to
use models that remain consistent with the process.
Another issue is that the mathematics and style of the classic controller models
are substantially different from the phenomenological models that process engineers
commonly used for process design, analysis, troubleshooting, constraint forecasting,
optimization, and training. The mathematics of the linear models obscures process
understanding. It would be desirable if the models in control were consistent with
those used in all the process management actions.
This article is about the use of first-principles models in MPC. Since the models
are nonlinear, it is nonlinear model-predictive control (NMPC).
Phenomenological models are mechanistic, cause-and-effect models. For process
control, start by applying fundamental mass, momentum, and energy balances on a
control volume, to get a model of an element of the process. Then integrate (sequen-
tially combine) the individual models to describe the unit. If the models are elemen-
tary in their description of the process, if they use ideal relationships, they are termed
first-principles models. If the models seek to fully describe all phenomena, they are
rigorous models. First-principles models are characteristic of undergraduate courses,
Nonlinear Model-Predictive Control Using First-Principles Models 95
product bulletins, training courses, etc. First-principles models are relatively easy to
develop, understand, and adapt and are fully adequate for control.
Since about 1980, there have been many approaches to nonlinear process control.
It seems that the industrial acceptance of digital control devices and distributed
control systems (DCS) of the 70s, revealed a vision to many investigators that model-
based control could be practicable. There have been many concepts as to how best to
implement it. For example in increasing complexity: Generic process control (GMC)
provides a structure to use steady-state models in an output characterization of a PI
controller. If the process is SISO and the problem is nonlinear gains, this is an excel-
lent solution. Process model-based control (PMBC) is a single step-ahead controller
(it does not forecast a future plan or handle ill-behaved dynamics), but can handle
MIMO processes with constraints, disturbances, and nonlinear and non-stationary
features. It was developed to adapt the model to changing process attributes such
as catalyst reactivity, tray efficiency, heat exchanger fouling, and drag coefficients.
If delay, high-order response, or inverse action is strong, PMBC would not be a
good solution. Predictive functional control (PFC) was developed to handle both
nonlinearity and ill-behaved dynamics (delays, high order, inverse action). It uses
a process model to calculate a step-and-hold MV action, if implemented now, that
would make the model match its set point at some future time (the coincidence point)
well past the influence of the ill-behaved dynamics. These are all practical, indus-
trially proven techniques. Complying with the K.I.S.S. principle (Keep it Simple
and Safe) that guides applications, those strategies should be considered prior to the
topic of this article, first-principles models for multivariable, constraint handling,
and model-predictive control.
2 What Are First-Principles Models?
First-Principles models are the process engineer’s models which are used for process
design, analysis, optimization, data reconciliation, and troubleshooting. They are
based on elementary material and energy balances and ideal constitutive relations
and are characteristic of college classroom teaching models for heat exchange, fluid
flow, reaction engineering, separations, etc. Representing process mechanisms, they
are typically nonlinear.
Alternatively, one could use rigorous models attempting to perfectly explain every
nuance. But balancing sufficiency with perfection, first-principles models are fully
adequate.
2.1 How Can Elementary Models Be Adequate for Control?
Contrasting modeling perfection, consider that humans can catch a fly ball or steer a
car through a curve, without rigorous models, with just intuitive continual correction.
96 R. R. Rhinehart
And consider that PI controllers, based on linear, stationary, simplistic (FOPDT),

inexact models, are also adequate for control. Here is a qualitative explanation: If
there are 30 or so control intervals during a process transient, then even if the model
is only 70% correct, leaving a 30% error in its first action, at the next control action,
the remaining error is 70% corrected, leaving a 0.3 * 0.3 = 0.09 error. By the time
a few control actions have been taken, the error is smaller than the discrimination
interval for digital values.
2.2 Automobile Speed Illustration
As an example of first-principles models, I will choose an automobile speed illustra-

tion. Although not a process model (such as reaction of distillation or heat exchange),
the concepts are more universally familiar.
Start with Newton’s Law of motion and then expand the acceleration term to the
rate of change of velocity.
ma m dv
F= = (1)
gc gc dt
An elementary representation of the forces on the car is from: (1) The engine,
represented by the accelerator pedal position, u, which will be modeled as a linear
response, (2) aerodynamic drag, ideally represented as proportional to velocity
squared, and (3) gravitational impact of going up or down hill with an angle, θ ,
from the horizontal.
F = ku − bv 2 − sin(θ )mg/gc (2)
Combine Eqs. (1) and (2), to obtain the elementary model.
m dv
= ku − bv 2 − sin(θ )mg/gc (3)
gc dt
This can be converted into the classic control form of a 1st-order process by
dividing by bv then rearranging terms.
m dv k sin(θ )mg
+v = u− (4)
bvgc dt bv bvgc
And, now, it can be represented in generic terms for a first-order process with a
locally linearized disturbance model.
dv
τ + v = Kpu + Kdθ (5)
dt
Compare Eqs. (4) and (5) and notice that the time constant, τ , is nonlinear,
depending on the velocity. The time constant is also non-stationary because the
drag coefficient, b, will depend on air density; and because the mass, m, will depend
on fuel in the tank and other loading changes. The process gain is also nonlinear and
non-stationary in spite of the linear force model in Eq. (2). And the sine function
nonlinearity of the road angle impact is now also confounded by speed, mass, and
drag coefficient.
For model-based control, we will use an alternate rearrangement of Eq. (3) which
isolates the rate of change.
dv gc
= ku − av 2 − sin(θ )mg/gc (6a)
dt m
In generic terms, with y as the state variable, u as the MV value, d the vector of
disturbance variables, and p as the vector of model parameter values, Eq. (6a) can
be represented as
dy
= f y, u, d, p (6b)
dt
This represents a first-order process, a process stage with one inventory location
of mass, momentum, or energy. It could represent how fluid flow rate responds to the
controller signal to the valve actuator, how temperature in a well-mixed tank depends
on a heat source, how composition in a well-mixed tank depends on influent flow
rate or composition, how pressure in a vessel depends on in- or out-flow or internal
generation of gas, etc.
Some one-inventory-stage processes have more than one state variable. For
instance, in a continuous flow reactor with well-mixed contents, temperature, compo-
sition, and volume would be several interacting states, which would be dependent on
several influences input flow rates, heating/cooling rates, and discharge rates. Here,
each CV would have a first-order model such as Eq. (6a), but each state variable
would have an influence on the rate of change of the others.
In generic terms for a MIMO set of coupled first-order processes, with y as the
vector of state variables, u as the vector of MV values, d the vector of disturbance
variables, and p as the vector of model parameter values:
dy
= f y, u, d, p (7)
dt
However, the process might have several stages that sequentially influence each
other. For example, in a tray-to-tray distillation column, the liquid contents on each
tray have several interacting states (temperature and compositions), which are influ-
enced by the trays above and below. In a 10-tray binary distillation, there would be
20 states on the trays (temperature and x1 ) (the other liquid and vapor compositions
would be related to the states). But there would also be column pressure and reboiler
98 R. R. Rhinehart
and condensate receiver as states. Although there may be 23 states, only 5 would be
controlled (top and bottom composition, column pressure, and reboiler and receiver
levels). Equation (7) could represent 23 first-order, nonlinear, interactive models, but
there may only be 5 MVs (reboiler heat, bottom’s flow rate, cooling flow rate to the
condenser, reflux flow rate, and distillate flow rate).
As a note, regardless of the form of the model presentation, do not solve the
model with calculus for several reasons. The nonlinearity and non-stationarities may
preclude analytical solution. Any solution will be predicated on an idealized time-
dependent behavior of the MV and disturbances. And any solution will be predicated
on fixed values (or at least elementary models) of how the model coefficients might
change in time.
Alternately, to determine what the model predicts, use a simple numerical solution.
Here is Euler’s explicit method, by converting the differential to a forward finite
difference.

yt = yt−Δt + Δt · f y, u, d, p (8)
t−Δt
This is a recursive relation, the new values of the states, yt , are calculated from the

prior values of the states, and of all other variables, yt−Δt + Δt · f y, u, d, p .
t−Δt
Initialize this calculation with the prior modeled y-value.
In this car speed example, the model was a single variable and could be coded on a
single line. It appears that Eq. (8) could be coded on N lines, one for each of the N state
variables. Alternately, a model may have IF–THEN conditionals such as representing
disparate mechanisms such as due to the transition from laminar to turbulent flow, or
from turbulent to choked flow. And there may be many other constitutive relations,
such as thermodynamic models that relate state variables, even requiring root-finding
methods to solve them. Or steady-state or equilibrium relations for actions that have
very fast dynamics. Also, for numerical stability the model Δt might be 1/10th of the
control interval, requiring that Eq. (8) be recursively calculated 10 times within each
controller interval. Alternately, a preferred numerical solution might be a 4th-order
Runge–Kutta approach. For many reasons then, the future modeled values might be
coded as a separate procedure (a function or a subroutine). There is no need to have
a single line of code (or N single lines) as suggested by Eq. (8).
Note: The model is not the truth about the process. This model presumed that the
engine force is linear with pedal position and independent of engine speed, gear, and
intake air composition. This model presumed that the engine force instantaneously
responds to pedal position and that there are no lags in getting fuel–air mixture into the
engine or back pressure on exhaust as engine speed increases. This model ignores
rolling friction of the tires on the pavement. This model uses the ideal Bernoulli
exponent of 2 for the aerodynamic drag force, and the model also ignores that the
drag force would be a function of the differential velocity between wind and car
speeds, not just car speed.
First-principles models are fully adequate for control. Use ideal relations. Do not
seek fully rigorous models.
3 MPC Controller Structure
There are three functions within model-based control—predict, correct, and act. The
structure can be represented by the operations in Fig. 1.
3.1 Predict
The first MBC function is to use the model to predict or mimic what the process
output is expected to do. This is the lower right box in Fig. 1. This updates the
past model value (from the prediction) with past disturbance and control action, to
predict the updated (current or new) modeled output. With conventional practice
of 30 control samplings in an open loop transient (or 10 samplings within a time
constant), the control interval, Δt, is usually small enough to permit Euler’s explicit
finite difference method to solve the model numerically. Using Eq. (8) to represent
a SISO dynamic model, this model prediction calculation becomes:
ym,i = ym,i−1 + Δt · f (ym , u, d, p)i−1 (9)
where “i” is the time increment counter, representing the present value, and “i − 1”
the immediate past values for each item in the function list. The subscript “m”
refers to the modeled value, not the measured process value. However, “u” refers to
the actually implemented MV value, after any overrides change what the controller
initially desired. The symbol “d” refers to measured disturbances or feedforward
variables, and “ p” represents parameter values in the model.
Biased
SP
y-process
SP Act Override Process
MV
bias y-
model
Correct Predict
pmm
Fig. 1 MBC representation

100 R. R. Rhinehart
Note: The right-hand side of Eq. (8) uses the prior modeled output value, ym,i−1 ,
not the prior process value, to update the modeled value, ym,i . Because of process-
model mismatch, noise, and such, the process measurement is not the right value to
initiate the model update.
However, since computers use assignment statements, the i-subscript is not neces-
sary, and there is no need to store all of the past model values. The new modeled
y-value is based on the most recent prior ym and possibly delayed input conditions.
Here, the symbol “:=” will indicate a computer assignment statement. A delay array
with put and retrieve pointers would be needed to obtain the u θ and dθ values.
ym := ym + Δt · f (ym , u θ , dθ , p) (10)
If the process model is high order, then internally to a procedure representing

“Predict” there would be a set of equations similar to Eq. (10). Also, if the model has
conditionals, or root-finding algorithms, then again, the model function would be a
procedure, not just one equation.
3.2 Correct
The second function within model-based control is a correction for process-model

mismatch. Process-model mismatch, pmm, is the difference between modeled predic-
tion and process measurement, as represented by the circle difference operation on
the lower right side of Fig. 1. The modeled value will not exactly match the process
value for a variety of reasons including model error (as discussed in Sect. 2.2), noise,
or unmeasured disturbances. A simple correction approach is to bias the set point for
the model with the process-model mismatch as follows:
pmm = yp − ym (11)
ySPbias = ySP − pmm (12)
The logic is illustrated in Fig. 2. “If you aim at the target, but hit 3 cm low. Then,
next time aim 3 cm above the target.”
The box labeled “correct” in the lower left of Fig. 1 could provide other functions.
One related to model coefficient adjustment has some managerial information bene-
fits, but in this simple correction approach, it simply is a pass-through of the pmm
value to bias the set point for the model.
Note: The biased set point is the set point for the model, not the set point for the
process.
Note: pmm is not the actuating error. It is not the process deviation from the target.
If at first you aim at the bull’s eye, but hit 3 cm low. Next aim 3 cm high. If that shot
(modeled to hit 3 cm above) has a process response that is 0.5 cm above the bull’s
3) So aim here
instead
1) Modeled value, you
aimed here
Aim higher by 2) Actual process value
pmm. Bias the tar- that results
get by -pmm
Fig. 2 The concept of biasing setpoint for model by pmm
eye, the pmm is the difference between where you aimed (+3 cm) and what happened
(+0.5 cm). So, the pmm is 2.5 cm, next aim 2.5 cm above (not 0.5 cm below).
3.3 Act—An Elementary Choice—Simple PMBC
The third model-based control function is to calculate the control action, to deter-
mine the controller MV value plan. This is represented as the box labeled “Act” in
Fig. 1. This starts with a user-defined, desired performance objective for the controller
to achieve. A simple desire, in the strategy termed process-model based control
(PMBC), is to calculate a u-value that would push the model toward the biased set
point at a rate proportional to the deviation from biased set point. (This is an effective
strategy for low-order models with inconsequential delay. Section 3.5 discusses how
to include constraints and ill-behaved dynamics in the MV plan.)
3.4 A Caution About the Triple Model Use
The control model is used in two places. In both the Predict and Act functions.
The Predict function updates the past modeled value with the actual implemented
MV to predict what the CV is doing now. This is a one-time-step update using the
first step of past values. I use the name Past-to-Now which has the acronym P2N.
Initialize the one-time-step update with the prior P2N modeled value.
The Action (Act) function uses a model to forecast what might happen in the
future, associated with a trial solution for the sequence of MV values. Guess at an
MV plan and see what the modeled CV does. Re-guess at an MV plan until the future
modeled CV matches the desired rate path. I call this the Now-to-Future model, which
has the acronym N2F. Initialize the N2F prediction with the current P2N modeled
value. After each MV guess, the N2F model needs to be re-initialized with the current
P2N value.
102 R. R. Rhinehart
Note: Both the P2N and N2F models use the same Eq. (10). But they need to be
initialized with the correct ym values. Use ym,i−1 for the P2N model. Use ym,i for the
N2F model.
Note: The P2N and N2F models must be consistent with each other. Once when
I was clever, seeking a way to avoid the iterative search for the MV, I modeled the
inverse with empirical approaches. But at steady state, the inverse surrogate for the
N2F prediction did not exactly match the P2N model, and accordingly, control had
−1
a steady-state offset. The inverse could be represented as u = MN2F (y
SPbias ), and the
−1
model could be represented as ym = MP2N (u). Then, ym = MP2N MN2F (ySPbias )
−1
and the models must ensure that at steady state, ySPbias = MP2N MN2F (ySPbias ) .
In control simulations, the process function will be simulated by a model. If the
process model in the simulator is the same as the controller model, then there will be
no steady-state offset, zero pmm, and perfect control. This is cheating. I do not know
of any process that could possibly be exactly modeled. Be sure that the controller
models are not identical to the process model in the process function. This includes
both functionality and coefficient values. For example: If you use the Ideal Gas law
in the controller model, use a different Equation of State in the process. If you use
the ideal Bernoulli relation for fluid behavior in the controller model, use 1.8 rather
than 2 as the exponent in the process. If you use a catalyst reactivity of 0.8 in the
controller model, use a different and ever-changing value in the process simulator.
There should be many differences between the controller models and the process
simulator model.
3.5 The Act Function in Horizon-Predictive Constraint

Handling Control
In horizon-predictive control (HPC), the Act function forecasts several future MV

moves (3 will be used in this article) to:
1. Shape the future modeled CV response to best fit a reference trajectory (here,
first-order, from current model P2N value to the biased setpoint for the model).
2. Steer variables clear of constraints or limits, or balance violations when a
constraint is unavoidable.
3. Economically optimize which MVs to use when there are extra Degrees of
Freedom (extra MVs).
Calculating the future MV sequence is an optimization application.
HPC Objective 1—Following the Reference Trajectory—Figure 3 illustrates Item
1 in the list above. The vertical axis represents a PV or MV value, and the horizontal
axis represents time. The time of 40, marked by the vertical line, is the current time.
To the left are past events, and to the right extending beyond a time of 50 is a possible
future plan of three MV steps. The plan has not happened and might not happen.
Like a trip plan, which is revised when you hit a detour or decide to take a side visit,
Fig. 3 Forecast MV moves to make the model best match a desired path
the original might never exactly happen. In this example, the process started at an
initial steady state. At the left of the time maker, the black line (the lower of the three
lines) has been the CV, which is steady and at the set point. The red line (the upper of
the three) has been the MV. The turquoise line (the middle line) has been the model
predictions. The difference between the modeled and actual CV values is the pmm.
To the right of the time marker are the possible future MV outcomes. The lower
horizontal line is the biased setpoint, the current target for the model. The black line,
which makes a first-order transition from the current modeled CV to the biased set
point, is the reference trajectory, the desired path for the controller MV plan to move
the modeled CV value. This was chosen to be a first-order path. The plan is the three
steps in the red line, the possible MV sequence. The green line, the upper of the lines
early in the future, just after the now marker, is the modeled CV value response to the
three MV steps. It starts at the present modeled value and shows a bit of a delay, then
a higher order lag, and a subsequent path which substantially follows the reference
trajectory. To get this modeled CV path, the controller MV plan is to first make a
large change to get the modeled CV to move quickly; but, if held there, the modeled
CV would fall substantially lower than desired. Continuing the plan, raise the MV
to prevent overshoot from the reference trajectory, but this would cause the modeled
CV to remain above the reference trajectory longer into the future, so the third step
in the MV plan is a bit lower.
Note: The future forecast time is termed the control horizon. Hence, the name
is horizon-predictive control. The horizon does not extend to the settling time, to
steady state for either the modeled or process CV or for the reference trajectory, but
typically to 70–90% of the settling time. Accordingly, the reference trajectory does
not get to the biased set point, as Fig. 3 illustrates. This does not matter, because only
the first MV step will be implemented. Then at the next sampling, the future plan
is recalculated with information about new disturbance values, new pmm evidence,
and new set points.
Note: The control interval is not the length of the first MV step-and-hold. In
this simulation, there are about 30 control intervals in the future plan. The first MV
step-and-hold in the current plan lasts for about 15% of the horizon (about 5 control
104 R. R. Rhinehart
intervals), the second step-and-hold for about 25% (about 7 control intervals), and
the last for about 60% (about 24 intervals).
Note: This three-step forecast MV pattern is not essential. More steps would give
more precise matching of model to reference trajectory, but in my experience three
is fully adequate and minimizes the optimization burden. One could use alternate
spacing patterns for the MV transitions. Since the first step is what will be imple-
mented and it shapes the early CV response and since subsequent MV steps trim the
plan and are farther out into the future, I like a plan that has more MV moves earlier.
This plan roughly spaces MV moves according to the golden ratio, but any of many
options works just as well. You might be more comfortable with 1/4, 1/4, and 1/2.
The optimization objective is not to force the model to match the reference trajec-
tory exactly, but to minimize the forecast CV deviations from r , the reference trajec-
tory. If the model dynamics are not simply first order, then it would be impossible
to make model match the reference exactly. The objective function is a least squares
deviation from reference trajectory over the control horizon.
H
2
Min
J= r1i − Ỹ1i (13)
{u 1 , u 2 , u 3 }
i=1
Here, i is a counter for future intervals, and H is the number of intervals in the
control horizon.
If the reference trajectory is first order, the time constant, τ , is the controller tuning
factor:
dr
τ + r = SP, , tt=now ≤ t ≤ H, rt=now = ỸP2N,now (14)
dt
If Δt, the control interval is small relative to τ , the time constant for the reference
trajectory, as it should be, then using Euler’s explicit method to determine r(t):

Δt , Δt
rt+Δt = SP + 1 − rt (15)
τ τ
For a multivariable process, each CV will have its own reference trajectory, and
each MV will have several future values.
Note: Some MPC algorithms have as the objective to make the model match the
biased set point, which would create a very aggressive control action. For these,
the control action is tempered by a move suppression factor, a penalty for large
MV moves. This is termed MV damping. What I have described with a reference
trajectory target is termed CV damping. In a MIMO process, the move suppression
factor on any one MV affects all CV responses. Since tuning with MV damping is
interactive in its effect on CV responses and since it does not have an explicit relation
to any one of the CV dynamics, I prefer CV damping. CV damping choice of τ for
each CV has a direct relation to that CV response speed.
HPC Objective 2—Avoiding Constraints—Another desirable, Item 2 in the list

above, is to plan the future MV sequence to avoid constraints. The constraint might
be on a CV, or any other process variable (PV), often termed an Auxiliary Variable
(AuxV). The constraint might represent a safety or specification limit, or a desired
resting value, or a rate of change. There are diverse ways to penalize either a constraint
violation or proximity to the constraint. The “Soft” penalty method for a deviation
penalizes a violation, proportional to the square of the violation.
0, if AuxV1 is within the limit

Penalty1 = (16)
(PV1 − L 1 )2 , if AuxV1 exceeds the limit
But constraints could happen at any future time, so sum all of the penalties.
Min
H
Penalty1,i (17)
{u 1 , u 2 , u 3 }
i=1
If several variables are subject to constraints, calculate the penalties for each.
Constraints on the MV, such as rate-of-change limits or values beyond the 0–100%
range of feasible values, could be handled as either soft constraints in Eq. (17) or
as hard constraints, limiting MV choices in the optimization. I prefer to treat MV
constraints as hard, limiting feasible values for the optimizer.
CV or AuxV constraints should be soft. Near a constraint, a disturbance, noise,
or such could move the AuxV beyond the constraint. Then, being beyond now, it has
happened, and no future control action can undo that. In such a case, all attempts to
find the MV sequence will fail. To prevent such infeasible corrections, treat CV and
AuxV constraints as soft.
HPC Objective 3—Minimizing Utility Cost—The third objective in the list, Item 3,
is to minimize costs of using MVs, when there are extra MVs available. If there are
extra MVs, then an infinite number of MV actions can provide equivalent control. But
usually, some MVs are more expensive than others. For instance, in heating, you may
have available electrical resistance heating or low-pressure steam. You would prefer
to use the less expensive steam until supplemental heat is necessary. To minimize the
cost of MV use over the future horizon calculates the cost for each MV in the plan.
A simple model for cost rate for an MV is Ċ = c · MV. Summed over each future
time interval, for one MV the objective is:
Min
t=H
t=H
J= Δt · Ċt = Δt c · MVt (18)
{u 1 , u 2 , u 3 } t=now t=now
For a process with multiple MVs, the projected cost over the control horizon is:
106 R. R. Rhinehart
Min
L t=H
J = Δt cl · MVl,t (19)
{u 1 , u 2 , u 3 } t=now
l=1
Note: if there are extra MVs then, instead of this economic cost method, you could
define a resting value of zero for the high cost MVs and use a deviation from the
resting value as a penalty as described next.
Combining all Optimization Considerations—Each of the three considerations
represents undesirable aspects. The controller should determine the control action
that minimizes all. Now that the three considerations are quantified, a standard way
of combining all is to sum all the measures of badness, and seek the MV sequences
that minimize the sum of all undesirables.
2 2
H H
Min i=1 r1i − Ỹ1i i=1 r2i − Ỹ2i
J= +
u1, u2, u3, . . . EC21 EC22
H H
i=1 Penalty1,i i=1 Penalty2,i
+ ··· + +
EC23 EC24
L H
Δt i=1 cl · MVl,i
+ ··· + l=1
(20)
ECcost
Note: The decision variables, the u i factors, are underscored, representing vectors.
In a multivariable process, there are several MVs, and each would have its several
future possible values. (Three MV values for one MV have been illustrated here.)
Because each term in Eq. (20) will have unique units, and because some aspects
will be more meaningful than others, weighting factors are required to unify units,
and to balance relative importance. This representation uses Equal Concern (EC)
factors. They scale each quantity by a factor that has the same units as the numerator
and represent equal concerns for the violation of the desired value.
Note: If there are no constraint violations, then the penalty terms in Eq. (20) are
all zero. If the model CVs can exactly follow the reference trajectories (or if the
model is at steady state at the biased setpoint), then the first set of terms is also zero.
The MV cost term, however, will not be zero. If the optimizer changes MV values
so that the MV cost term is reduced, the new MV values will cause the model CVs
to deviate from the desired path or from the biased setpoint value. This would make
the first set of terms non-zero. It may be possible to reduce cost by a greater amount
than the penalty for not following the reference trajectory. This would mean that
the process CVs will have a steady-state offset. So, if the cost term is included in
the optimization objective function, it is important that the weighting factors make
the MV cost term much less important than either the CV deviation or the violation
penalty terms. Choose EC factors so that any steady-state offset is imperceptible.
See the next subsection. Alternately consider other ways to manage the economic
factors. See the second-next subsection.
Note: The EC weighting factors could be used for tuning aggressiveness, but
that is not their role. The time constants for the reference trajectories are the tuning
factors. One for each CV.
HPC 2—Setting the EC Factor Values—Here is a procedure for setting the EC
factors.
2
H
1. Choose the most critical term. One term to minimize is i=1 r1i − Ỹ1i , maybe
this is of greatest concern.
2. Choose a moderate Ỹ1 deviation from perfection. It might be the deviation from
the reference trajectory along the future time or the deviation from the biased
set point at steady state. This value is an EC factor, and EC1 is equal to the Ỹ1
deviation considered to be moderate.
3. Feel the concern this deviation creates. Use both qualitative and quantitative
aspects. Perhaps rate the concern level on a 1–10 basis. Concern aspects could
include your boss’s opinion, customer reaction consequences, the generation of
waste product, etc.
4. Assign EC values to each term in Eq. (20) representing each deviation from
perfection that has the same level of concern. As an example:
• Perhaps EC1 = 0.5 wt% product deviation from SP has a 1–10 concern rating
of a 3.
• To have the same concern rating of 3, perhaps the deviation from a target flow
rate is 15 lb/min. Then, EC2 = 15 lb/min.
• Also, to have the same concern rating of 3, perhaps the deviation from a
process temperature limit is 10 F. Then, EC3 = 10 F.
The EC values are situation-dependent, and they change with the situation.
Note: Traditional optimization uses Lagrange-type multipliers for the terms in a
sum. Here, those values are the reciprocal of the denominator values. I think that the
weighting method using EC factors is intuitive and easier to assign weighting values,
which provide the relative importance.
Choosing the economic EC factor may be a bit difficult because it must also
balance the consequential steady-state offset. To help see the way to include this
aspect, consider a single CV. Ideally, the deviation from set point is zero with some
particular MV value. If the optimizer chooses a less expensive MV value perturbed
by Δu, there will be a corresponding Δ ỹ = K m Δu, and the CV deviation factor
2
Δ ỹ
will be n EC y
, where n is the number of samplings in the horizon. The cost term
improvement will be Δt·n·cl ·Δu
ECcost
= Δt·n·cECl ·Δ ỹ/K m
cost
. We would like the CV deviation to
be imperceptible, perhaps small relative to CV noise or relative to the width of the
image on the display. We desire Δ ỹ ≤ ε. Equating the two factors and solving for
Δt·cl ·EC2y
ECcost , ECcost = ε·K m
.
Alternate Methods to Account for MV Costs—One option is to not include the
last term in Eq. (20) objective. It is not in standard linear empirical model-based
controllers.
108 R. R. Rhinehart
Another option is to recognize that two MVs may provide the same benefit, but
one is more expensive than the other. These may be two possible heating mechanisms
steam or the more expensive electrical resistance. In such a case, use the steam until
its capacity is inadequate and then add the electrical resistance to provide the extra
needed. As another example, there may be two sources of a reagent, dilute and the
more expensive concentrate. Use the dilute first, until a flow rate limit is reached or
until the dilution causes alternate issues, then trim with the concentrate. There are
several ways to handle this logic, and one is to split the 200% range of the optimizer
MV. If 0–100% use the less expensive source, then 101–200% also use the more
expensive source. Another method is to have a resting value for the more expensive
source (it actually may be non-zero to ensure it is immediately available) and add a
penalty for deviations from the resting value.
A third case might be one in which the extra MV does not have a cost, but affects
the values of the other MVs. As an example, pressure on a distillation column can
be adjusted by a throttling valve on the vapor leaving the column or on adjusting
the flow rate of coolant (process water, or air) in the distillate condenser. Pressure
is normally specified in an optimization exercise to stay within limits, maximize
column efficiency, and balance performance over a range of expected duties, then not
included in the column composition control. However, pressure could be considered
an MV for control, and the optimization may determine the pressure that results in
minimizing the cost of reboiler power and distillate condenser cooling.
Optimization to Determine the Possible Future MV Values—The optimization
statement is
2
H

N
i=1 r ji − Ỹ ji
M H
Penaltyk,i
i=1
min J = +
{U l,m } j=1
EC2j k=1
EC2k
L H
Δt i=1 cl · ṀVl,i
+ l=1
(21)
ECcost
S.T: hard constraints on the MVs (Value or Rate-of-Change).

The U Subscripts are l = # MVs, m = # future moves for each MV
N = # CVs, M = # soft constraints, L = # MVs
H = control horizon—the number of Δt calculations in the future time projection.
This should go to about 85% of open loop settling time.
One could include MV constraints, such as values or rate-of-change (RoC) in the
soft penalty terms. These are input variables to the process. However, one cannot
include process response variables (CVs and AuxVs) in the list of hard constraints,
because a disturbance may cause a response to exceed a limit, and the controller will
be unable to immediately remove it from the constraint.
Classic MPC products use Linear Programming or Successive Quadratic.
However, there are a number of other constraint-handling gradient-based tech-
niques that could be used. Contrasting these, my preference is direct search, multi-
player optimization methods. Multi-player methods are “global,” and direct search
methods do not get confounded by surface aberrations (non-idealities, disconti-

nuities, numerical striations, etc.) and can handle hard constraints, and nonlinear
responses. Leapfrogging is my choice, considering effectiveness, speed, simplicity,
and robustness [1].
DDC or Not?—Direct digital control (DDC) sends signals from the controller
directly to the final control element (valves, heaters, etc.). If you have a single loop
controller, SISO, and the model-based control calculation is within the PLC or DCS,
then this may be permissible.
However, model-based controllers are typically implemented external to the PLC
or DCS where greater computational power is available. Although the model-based
controller decisions regarding the signal to the final control element could be trans-
ferred to the DCS to be relayed to the process, experience has shown that it is best
to transfer set points for the process primitive variables (flow rates, temperatures,
levels, etc.) and let the PID algorithms in the DCS or PLC determine the signals
to the final elements. This is a cascade type of implementation. The primitive or
lower-level loops typically respond faster than the primary variables that need to be
managed by constraint-handling horizon-predictive control. So, this transfer of set
points does not realistically slow control response. The advantages are that the DCS
or PLC has many auxiliary functions related to signal processing and validation that
should be used, and it is designed to be robust to power and communication failures.
If for any reason the computer in which MPC is implemented gets confounded or
fails, the DCS or PLC will continue process operation with the last set of set point
values. The process will not shut down, if the MPC computer does.
4 MPC Example
This example will discuss nonlinear, horizon-predictive, constraint-avoiding model-

based control for the car example in Sect. 2.2. The process, representing the real
car, is a simulation. To make the illustration legitimate, the simulation equations
for the process are not the same as those for the controller model. The simulation
is more complex and has alternate functionalities for effects, and coefficient values
that are different from the controller model. One cannot perfectly model the reality
of a process. Simulations of control should reflect that reality.
For control, this is a SISO process with the controller output to the final control
element. However, the process also has an auxiliary variable, intended to represent
cabin noise, which has an upper limit constraint.
Figure 4 represents a deterministic and disturbance-free trend in simulated process
and controller model, subject to ideal step changes in the MV.
The red trace that makes step changes is the car accelerator pedal position, the
MV. It is attached to right-hand vertical axis. Each step is an increase of 20%. The
upper curve is both the car speed and the set point, which are attached to the left-hand
vertical axis. In the MAN mode, the set point tracks the CV for bumpless transfer
110 R. R. Rhinehart
Fig. 4 Trends in MAN mode
to AUTO. Note, the first MV step increases the CV by about 42 kph, but the last
equivalent MV step is only by about 6 kph. The gain change ratio is about 7:1, being
one evidence of nonlinearity. Also note that at the higher speed the settling time
is about 15 s, while at the first MV step at the lower speed, even after 25 s, the
process is still rising. This nearly 2:1 change in settling time is a second evidence of
nonlinearity of the process.
The model is represented by the blue dashed line, also attached to the left-hand
vertical axis. It too, is nonlinear, with dynamics that change with operating speed.
However, the model features do not match the process, exactly. The model is usually
lower than the process and the pmm progressively increases with speed. But because
the dynamics are different at the first MV step, the model actually rises, temporarily,
above the process.
The lowest curve, the blue dots, is an Auxiliary Variable; in this case, it is a very
simple representation of noise volume associated with the engine and car movement.
Figure 5 illustrates several features in AUTO mode. Like Fig. 3, the vertical line
toward the right of the graph marks the time now, separating the past and future.
Initially, at time = 0 the controller is in MAN mode and placed in AUTO at a time of
2 s. Because of set point tracking and P2N model calculation in MAN mode, there
is no bump in the transfer to AUTO. Set point changes occur at times of 5, 40, 60,
and 80 s indicated by the black line that makes steps. The solid blue curve represents
the process CV (car speed); and in spite of the 7:1 gain nonlinearity, the 2:1 settling
time ratio, and the changing pmm, the overshoot magnitude and rise time in the first
and third set point changes are very similar.
The second and fourth set point changes represent constraint conditions. The
AuxV (blue dots) has a soft constraint of 50 (right vertical axis) so even though
Fig. 4 shows that the speed set point of 120 kph is feasible at MV = 85%; at that MV
value, the AuxV has a value of about 70, exceeding the constraint. In Fig. 5, the EC
factors make constraint violations more important than meeting the speed set point,
and the optimizer limits the MV to about 60% preventing the speed from reaching
Fig. 5 Events in AUTO mode—1
its set point. When the setpoint is dropped to a speed of 100 kph, there is no delay in
the MV action (there is no wind-up in the controller). That was a soft constraint. A
second constraint happens at the final set point change, where the controller would
like to place the MV below zero. The MV stays at the constrained value limit, until
responding (again without wind-up). That is a hard constraint. The third constraint
is a rate-of-change constraint on the controller, another hard constraint, which is
evidenced by the linear ramp change in the MV immediately after each SP change.
Figure 6 illustrates the same events as in Fig. 5, except that the rate-of-change
constraint is removed, and the EC factor for the AuxV violation is increased to place
greater importance on the CV to SP response. Without the rate-of-change constraints,
note the spikes in the MV action just after the set point changes and the MV limit of
100% at a time of 40 s. Also, notice the lesser deviation of the CV to the set point
and associated larger deviation of the AuxV from the limit of 50 during the 40–60 s
period.
There is only one tuning factor for each CV, the time constant for the reference
trajectory. In Fig. 7, the time constant is significantly larger, making the controller
less aggressive. Compared to Fig. 6, also with no rate-of-change constraint and
the duplicate EC factors. Notice the smaller jumps in the MV action and nearly
overdamped response of the CV to the SP.
Finally, in Fig. 8 both measurement noise and random disturbance patterns are
included in the process. Compare this to Fig. 6, which has the same tuning and EC
values. The noise is evident in the perturbations on the CV value throughout. The
disturbance is evident on the trend in the MV during the 20–40 s period and the up-
and-down action in the MV during the 60–80 s period, even though the CV remains
about at the set point. The disturbance is unmeasured, not included in the controller
model, but affecting the process only. This additional source of the process-model
mismatch is also evident in the 20–40 s period where the pmm changes during the
20–40 time period.
112 R. R. Rhinehart

The pmm is filtered to temper the impact of measurement noise. The noisy CV
is not filtered, which has a benefit of being able to display the noise pattern and the
real state of the process. If the process CV were filtered creating a significant lag,
then the modeled CV should be equivalently filtered also.
5 End Notes
The model does not have to be perfect, great, very good, or even good. A fair model
works as this simulation reveals. As additional evidence that imperfection works:
We use linear constant FOPDT models for nonlinear high-order processes. We use
linear FIR models in APC (MPC, DMC). We steer cars through curves and catch fly
balls by continually correcting our mental intuitive “models.”
Implement advanced control in stages. Start with the model shadowing the
process to affirm model validity. Test bumpless transfer. Then observe the controller
suggesting action. If confident that it is making good decisions, let the controller
take action while being supervised. Finally, implement autonomous control, but
displaying goodness of control messages.
Implement HPC showing the operator the horizon-predictive constraint-avoidance
predictions. Otherwise, the decoupling and predictive consideration might lead to
unexpected decisions and operator discomfort.
There may be MVs removed for temporary maintenance. These should be removed
from optimization. Include on–off switches so that the controller understands the
temporary condition.
Be sure that measurements are valid. Faulty inputs lead to bad decisions. This is
not different for any level of complexity in control. You may need sensor redundancy
or data reconciliation for critical measurements.
When the process changes (perhaps piping or units are reconfigured), you may
need to change the model.
Provide a trap for any possible infeasible calculations and a Plan B action.
Cascading MPC outputs as set points to lower-level PLC or DCS controllers is one
strategy to keep the process running even if the supervisor MPC gets stuck.
K.I.S.S. Keep it Simple and Safe. If gain-scheduled PI works use it, if classic
advanced regulatory control solves the problem use it, if GMC-SS solves the
nonlinearity use it.
Consider incremental model parameter adjustment online to make the model adapt
to the process. Observation of the adapting model coefficient can be a clue for predic-
tive maintenance, and the model will be better at forecasting the constrained limits
of process operation.
The justifications for first-principles models in control include:
• One-model consistency for many uses.
• Preserving process mechanistic knowledge.
• Process analysis, monitoring, and constraint forecasting.
114 R. R. Rhinehart
• Handling nonlinearity, non-stationarity, and disturbances that invalidate linear

models.
• Ease of tuning.
• Ease and reduced process testing for model calibration (adjustment).
• Wider valid modeling range than with linear models.
How to economically justify using a nonlinear model? APC (linear multivariable
control) experience shows a halving (or better) of process variance, which permits
operating closer to constraints and specifications. This permits higher throughput,
reduced waste, higher quality, lower operating costs, etc., which provides the
economic benefit. In my assessment, any control scheme (ARC, Fuzzy, APC, NMBC)
that has equivalent disturbance-rejecting, feedforward, and decoupling features has
the same benefits as any other controller with those same features. The nonlinear
model-based controllers will have those benefits. As well, it will work over a wider
operating range, require fewer recalibrations, and be less expensive to recalibrate
than controllers with empirical models.
Reference
1. Rhinehart RR (2018) Engineering optimization: applications, methods, and analysis. Wiley, New
York, NY
DC Motor System Identification
and Speed Control Using dSPACE Tools
S. Menaka and S. Patilkulkarni
Abstract This paper presents the system identification and speed control technique
of DC motor using dSPACE tools. The two different methods for the speed measure-
ment of the DC motor using dSPACE encoder and the hardware interrupt blocks
are discussed. The system identification of DC motor is done using the dSPACE
Control Desk 5.3 and system identification toolbox. The DC motor speed controller
is implemented using dSPACE hardware tools by PWM technique, where the speed
of the DC motor is controlled by controlling the motor’s terminal voltage with the
help of duty cycle of PWM. The control algorithm is developed using MATLAB-
Simulink and the Real time interfacing (RTI) is done by using dSPACE DS1104
controller board and CP1104 controller panel. The data capturing, visualization and
layout designing is done by using Control Desk 5.3. The developed model can be
used for Rapid control prototyping (RCP) and testing purposes for DC motor related
applications.
Keywords DC motor · CP1104 · MATLAB Simulink · Control Desk 5.3 · System

identification · RTI · PWM · dSPACE · RCP
1 Introduction
DC motors in industrial control systems have many applications as they are straight-
forward to operate and model. Sometimes an accurate DC motor model utilized
in a control system may be necessary for analytical analysis and optimization of
the control system design. In this scenario, the reference values of the DC motor
characteristics as specified in the motor specification which is generally provided
by the manufacturer of the motor might not be regarded sufficient particularly for
S. Menaka (B) · S. Patilkulkarni

JSS Science and Technology University, Mysuru, India
e-mail: menaka2211@gmail.com
S. Patilkulkarni
e-mail: sudarshan_pk@sjce.ac.in
https://doi.org/10.1007/978-981-19-6913-3_7
116 S. Menaka and S. Patilkulkarni
the more affordable DC motors, whose electric and mechanical parameters tend to
be quite high. DC motor model identification may be accomplished using general
system identification methods [1–3]. In the absence of costly testing equipment and
a lengthy testing cycle, a rapid and effective technique of system identification [4–
7] based on the input and output values (i.e., the voltage and speed values) of the
motor is desirable and advantageous, particularly for field applications and quick
controller development [8]. It is necessary to load a specific input signal and identify
the system’s output [9] in order to make an approximation to obtain the mathematical
model of the system. This permits the design and insertion of PID controller to a
closed-loop system model [10, 11]. By the auto tuning process of the PID controller,
PID parameters are created [12]. The modified PID controller can then integrated
into the hardware in the loop (HIL) simulation of the control system [13], and the
performance of the system’s response can be assessed. From the literature review
for the system identification process different models such as Process model, ARX,
ARMAX, OE model, State space model [14], non-linear models were used. Among
all of the models used the process model is considered in our work since it is less
complex and is sufficient for second order system and the model coefficients have
an easy interpretation as poles and zeros. Other models can also be considered for
estimating transfer function to achieve more accuracy which is the future scope of
the work.
In this paper, a DC motor speed controller system using PWM technique is
proposed and implemented, and its performance is analyzed. The system identi-
fication process is also discussed. The controller is developed using MATLAB-
Simulink and implemented in real time using dSPACE Control Desk and hardware
tools (DS1104 controller board and CP1104 controller panel). The dSPACE tool
helps in developing a virtual controller for easy prototyping with making use of less
hardware components. The proposed system uses voltage, current and speed sensor
for measuring the input and output signal of the DC motor.
2 Methodology
Figure 1 describes the proposed block diagram of the experiment. The voltage sensor
is used to detect the voltage drawn by the motor. Output of voltage sensor is given to
ADC of ds1104 controller board and is then calibrated using the gain blocks to obtain
the motor voltage output. The current sensor is used to detect the current drawn by
the motor. The output of current sensor is given to ADC of ds1104 controller board
and is then calibrated using the gain blocks to obtain the motor current output. The
quadrature incremental magnetic encoder is used to obtain the speed feedback of the
motor. The encoder channel outputs is given to the encoder or interrupt pin of ds1104
controller board based on the speed technique used and is calibrated using the gain
blocks to obtain the output in terms of rpm value of the motor. The experimental
data of voltage, current and speed values are captured using Control Desk and are
used to obtain transfer function by system identification method. The speed control
DC Motor System Identification and Speed Control Using dSPACE Tools 117
Fig. 1 The proposed block diagram
of DC motor is implemented by using the pulse width modulation technique (PWM),

by applying variable duty cycle PWM signal to the motor driver circuit which also
controls the motor terminal voltage.
2.1 Voltage and Current Measurement
RTI Simulink model for voltage sensing is shown in Fig. 2. For analog inputs, the
dSPACE unit has an integrated gain of 0.1. Any signal read from a dSPACE ADC
block is multiplied by 0.1 in Simulink. Hence a gain value of 10 is applied on the
output of an ADC block to compensate for this. The 5 V range obtained across the
analog pin of voltage sensor lets recognizing of voltages up to 25 V with the help of
controller. Hence a gain value of 5 is multiplied to obtain the voltage value drawn
by the motor.
RTI Simulink model for current sensing is shown in Fig. 3. The dSPACE device
features an inbuilt gain of 0.1 for analog inputs. In Simulink, each signal received
from a dSPACE ADC block is multiplied by 0.1. Therefore, a gain value of 10 is
used to compensate for this in the output of an ADC block. The voltage observed
across VCC of ASC712 is around 4.86 V. When there is no current flow through the
terminals of the module, the output voltage will be half of the supply voltage, i.e.,
2.4325 V. This implies subtracting 2.4325 V from the voltage recorded at the ADC
pin. The current value is calculated by dividing this value by the sensitivity value of
the sensor. The below equation is used to calculate current drawn by the motor
Currentvalue = (adcvoltage − offsetvoltage)/Sensitivity (1)
Fig. 2 The RTI Simulink

model for voltage sensing
Fig. 3 The RTI Simulink model for current sensing
2.2 Speed Measurement of DC Motor Using Incremental

Magnetic Encoder
2.2.1 Speed Measurement Techniques
There are two techniques for measuring the speed using an incremental encoder they
are
– Pulse counting (Frequency measurement)
– Pulse timing (Period measurement).
Pulse Counting (Frequency Measurement)

The conventional and perhaps easiest way to determine rotor speed is by measuring
the frequency of the pulses of the encoder. In a given and constant time window,
the number of pulses observed is counted. Pulse counting method determines the
average pulse time. For the calculation of the angular velocity ω, say the count is n
pulse for a sample period T. Thus, the average pulse time is T /n. If N windows are
available on the disc, NT /n is the average time of one rotation. Therefore, the angular
velocity is given by
ω = 2π n/N T (2)
where
ω angular velocity
n pulse count
T sample period
N pulse per revolution.
Pulse Timing (Period Measurement)

The high frequency clock signal is measured with the pulse timing method during a
single encoder cycle. Assume the frequency of the clock is f Hz. If m clock cycles are
counted over an encoder cycle, m/f will be the time taken to complete one encoder
cycle. The average time for one disc revolution is Nm/f when there are a total of N
windows on the track. Therefore, the angular velocity is given by
ω = 2π f /N m (3)
where
ω angular velocity
f clock frequency
m number of clock cycles
N pulse per rotation.
2.2.2 The Implementation of Simulink RTI Model for the Speed

Measurement of DC Motor
Method 1: Speed measurement using encoder of ds1104 board (Pulse counting

method).
In this method speed is calculated based on the number of pulses obtained during a
certain amount of time. The block “DS1104ENC_POS_C1” gives read access to two
encoder interface channel’s position and delta position. The encoder master configu-
ration block defines the global characteristics for the encoder interface channels. The
encoder signal type is set to single ended (TTL) in the “Encoder master setup” block.
The encoder has a disc of 51.42° for each line, with 7 lines per revolution. The line
count is multiplied by the fixed number 51.42 to convert it to degrees. The “Encoder
delta position” output is used to compute the angular velocity of the motor in rad/s
by dividing it by the sampling time as in Fig. 4. The angular velocity is converted
to rpm by multiplying the rad/s value with 60/2π. The rpm values obtained in this
method was not accurate because in a fixed sample time the sufficient information
of the pulses is not obtained as a low-speed motor is used. This method is best suited
for high-speed motors.
Method 2: Speed measurement using hardware interrupt of ds1104 controller board
(pulse timing method).
The implementation of time measurement is being carried out by the function
call subsystem which is triggered by the DS1104MASTER_HWINT_I1 external
hardware interrupt block as shown in Fig. 5. The channel A of the incremental
magnetic encoder is connected to the external hardware interrupt pin of the ds1104
controller board. The input port of the function call subsystem, i.e., “Angle” denotes
the fixed angle between two successive pulses of the sensor which is calculated using
the equation 2π /(impulses per revolution).
Figure 6 depicts the proposed contents of the function call subsystem for the time
measurement. The previous time and the time difference value between two pulses
of the encoder are being stored in the data store memory and data store read blocks.
The system outputs block has been taken from the custom code library.
Fig. 4 The RTI Simulink model for speed measurement using encoder
Fig. 5 The RTI Simulink model for speed measurement using hardware interrupt
The following code is entered in the System Outputs block

// declare local variable
double currentTime;
// read timestamp
currentTime = ts_time_read();
// calculate difference to previous step
deltaT = currentTime – oldTime;
// save new timestamp
oldTime = currentTime;
The ts_time_read() function reads the absolute simulation time from the hard-
ware through which the currentTime and the oldTime values are obtained. The time
difference deltaT and the previous time of the encoder pulse oldTime is written to
the corresponding data store memory blocks. Thus the speed of the motor is calcu-
lated by dividing the fixed angle with the measured time difference deltaT, i.e.,
(2 * pi)/(impulses per revolution * deltaT).
Fig. 6 The Simulink model of function call subsystem block
3 System Identification of DC Motor
The experimental voltage and speed values of the DC motor captured using the
dSPACE control desk are exported to the MATLAB workspace. The transfer function
of the DC motor is identified by system identification toolbox as shown in Fig. 7.
Several models such as Process model, Polynomial models transfer function
model, state equation model, non-linear models are available to estimate the plant
model. In this work the process model structure was incorporated for the identifica-
tion process because the order of the model was easier to anticipate as the DC motor
model is a second order system. The reason why the ARX model is being ignored is
that the estimated model is a prediction model and is not directly dependent on the
input and output values which might cause difficulty in building the controller. The
obtained transfer function model has the best curve fit of 84% between the model
response and the measured output.
Figure 8 shows the flowchart of the system identification process which depicts
the steps involved during the identification process.
The obtained transfer function from the system identification process is given
below
Fig. 7 System identification toolbox
83.2
G(s) = (4)
(1 + 0.01534s)(1 + 0.00234s)
The Transient response, frequency response and the pole-zero plots of the obtained
transfer function from the system identification process is as shown in Figs. 9, 10
and 11, respectively.
4 Controller Implementation and Result
The Pulse Width Modulation (PWM) method can regulate the efficient armature
voltage using ON-OFF control technique. The greater duty cycle pulse turns ‘ON’
more long than the lower one. The duty cycle δ, is regulated by the following equation
Pulse on time or pulse width (ton )

Duty cycle (δ) = (5)
Pulse time period (T )
The ratio of ON to OFF time is known as the duty cycle, and it influences the
motor’s speed. By varying the duty cycle the desired speed can be obtained. PWM
pulses are used to control the duty cycle of a motor driver circuit. A square wave with
a constant voltage but variable pulse width or duty cycle is used to power the motor.
The controller implementation using pulse width modulation technique is as shown
in Fig. 12. The PWM pulse is generated using DS1104SL_DSP_PWM3 block. Since
Fig. 8 Flowchart of system

identification process
Fig. 9 Step response

Fig. 10 Bode plot
Fig. 11 Pole zero map
the frequency of the PWM pulse is kept constant to 1 kHz but the on-off duration
varies, the determination of duty cycle of PWM is based on the pulse width of the
pulse. As pulse width increases, the duty cycle of PWM also increases which in turn
increases the power delivered to the motor driver circuit. The duration of ON period
of PWM is longer at 100% duty cycle than at 50% duty cycle. This result in a faster
motor speed at 100% duty cycle compared to 50% duty cycle. The following equation
expresses the relationship between average voltage, supply voltage, and duty cycle.
δ
V(average) = (6)
V(supply)
Table 1 shows experimental results of the DC motor speed and voltage values
obtained at different values of duty cycle. The duty cycle was varied from 0 to 1
and the corresponding voltage and speed values were noted down. As the duty cycle
increases the DC motor’s speed also increases. The motor speed is 295 RPM at 30%
duty cycle, and the corresponding output voltage of the converter is 3.62 V. When
the duty cycle is set to 50%, the motor’s speed is 498 RPM, and the resulting output
voltage of the converter is 6.14 V. As the duty cycle approaches 70 and 90%, the
motor’s speed and the converter output voltage increases.
The graphs shown in Figs. 13 and 14 depicts the relation between duty cycle,
voltage and speed.
Fig. 12 The controller implementation model of DC motor
Table 1 Experimental results

Duty cycle (%) Voltage (V) Speed (RPM)
0 0 0
0.3 3.62 295
0.4 4.85 393
0.5 6.14 498
0.6 7.25 595
0.7 8.47 698
0.8 9.64 794
0.9 10.82 896
1 12.3 997
Fig. 13 Duty cycle versus

speed
Fig. 14 Duty cycle versus

voltage
Figure 15 shows the GUI developed using control desk for capturing the voltage,
current and speed data’s of the motor.
5 Conclusion and Future Scope
The two different methods for speed measurement of motor were implemented. It
was found that the pulse counting method using encoder block is best suited for
high-speed motors and the pulse timing method using hardware interrupt block is
best suited for low-speed motor. The process of system identification is discussed,
and the transfer function of the motor was achieved. The PWM method was used to
create a motor speed controller system. The relation of duty cycle against voltage and
speed parameters was investigated. It has been discovered that the greater the pulse
width, the higher the average voltage supplied to the motor terminals. As a result,
the magnetic flux inside the armature windings becomes greater. As a result, the
motor will revolve faster. By varying the duty cycle of the PWM pulse, the DS1104
controller gives flexibility in managing the speed. The impact of PWM pulse width
on DC motor voltage and speed values has been investigated.
The accuracy of the obtained transfer function from system identification process
can be increased by considering other models for estimation. The closed-loop
controller can be implemented for the transfer function obtained from the system
Fig. 15 The control desk layout designed for the experiment
identification process and can be implemented on hardware. The cascade control of

the motor can also be incorporated by considering current also as feedback along
with speed. The same model can be worked on and implemented using dSPACE
MicroAutobox II.
References
1. Awoda M, Ali R (2019) Parameter estimation of a permanent magnetic DC motor. Iraqi J Electr
Electron Eng 15(1):28–36
2. Sankardoss V, Geethanjali P (2017) Parameter estimation and speed control of a PMDC motor
used in wheelchair. Energy Procedia 117:345–352
3. Ramasubramanian D (2016) Identification and control of DC motors. Master’s thesis in
automatic control and robotics, Barcelona School of Industrial Engineering, ETSEIB
4. Bature A (2013) Identification and real time control of a DC motor. IOSR J Electr Electron
Eng 7(4):54–58
5. Becedas J, Mamani G, Feliu V (2010) Algebraic parameters identification of DC motors:
methodology and analysis. Int J Syst Sci 41(10):1241–1255
6. Othman K, Kamal M (2009) System identification of discrete model for DC motor positioning.
In: Recent advances in circuits, systems, electronics, control and signal processing, pp 212–216
7. Hassan MM, Aly AA, Rashwan AF (2007) Different identification methods with application
to a DC motor. J Eng Sci 35(6):1481–1493
8. Tang WJ, Liu ZT, Wang Q (2017) DC motor speed control based on system identification and
PID auto tuning. In: Chinese control conference, CCC, 61403422, pp 6420–6423
9. Ljung L (1998) System identification. In: Procházka A, Uhlíř J, Rayner PWJ, Kingsbury NG
(eds) Signal analysis and prediction. Applied and numerical harmonic analysis. Birkhauser,
Boston, MA
10. Joshi B, Shrestha R, Chaudhary R (2014) Modeling, simulation and implementation of brushed
DC motor speed control using optical incremental encoder feedback. In: Proceeding of IOE
graduate conference, Nov 2014, pp 497–505
11. Fliess M, Sira-Ramirez H (2008) Closed-loop parametric identification for continuous-time
linear systems via new algebraic techniques. In: Identification of continuous-time models from
sampled data. Springer, pp 363–391
12. Dewangan AK, Chakraborty N, Shukla S, Yadu V (2012) PWM based automatic closed loop
speed control of DC motor. Int J Eng Trends Technol V3(2):110–112
13. Wu W (2010) DC motor parameter identification using speed step responses. In: Proceedings
of the American control conference 2012, no 3, pp 1937–1941
14. Basilio JC, Moreira MV (2004) State-space parameter identification in a second control
laboratory. IEEE Trans Educ 47(2):204–210
15. Rahmani F, Quispe D, Agarwal T, Barzegaran M (2020) Speed control of brushless DC motor
by DC-DC boost and buck converters using GaN and SiC transistors for implementing the
electric vehicles. Comput Res Prog Appl Sci Eng 06(02):70–75
16. Amalrajan R, Gunabalan R, Tewari N (2020) DSPACE1103 controller for PWM control of
power electronic converters. In: Lecture notes in electrical engineering, vol 687, pp 307–322
17. Fatma MW, Hamid MI (2019) PWM speed control of DC permanent magnet motor using a
PIC18F4550 microcontroller. IOP Conf Ser Mater Sci Eng 602(1)
18. Hamzi B, Owhadi H, Koch J, Naung Y, Schagin A, Oo HL, Ye KZ, Khaing ZM, Simani S,
Alvisi S, Venturini M (2021) Implementation of data driven control system of DC motor by
using system identification process. Electronics (Switzerland) 8(2):1801–1804
19. dSPACE DS1104 control workstation & Simulink tutorial (2018) Utah University, pp 1–38
20. Tatenda Katsambe C, Luckose V, Shahabuddin NS (2017) Effect of pulse width modulation on
DC motor speed. Int J Students’ Res Technol Manag 5(2):42
21. Agung IGAPR, Huda S, Wijaya IWA (2014) Speed control for DC motor with pulse width
modulation (PWM) method using infrared remote control based on ATmega16 microcontroller.
In: Proceedings—2014 international conference on smart green technology in electrical and
information systems: towards greener globe through smart technology, ICSGTEIS, pp 108–112
22. Gunda KK (2008) Adjustable speed drives laboratory based on dSpace controller. MSc thesis,
Louisiana State University
23. Mamani G, Becedas J, Feliu-Batlle V, Sira-Ramirez H (2007) Open-loop algebraic identifica-
tion method for a DC motor. In: 2007 European control conference (ECC), pp 3430–3436
24. Petrella R, Tursini M, Peretti L, Zigliotto M (2007) Speed measurement algorithms for low-
resolution incremental encoder equipped drives: a comparative analysis. In: International
Aegean conference on electrical machines and power electronics and electromotion ACEMP’07
and electromotion’07 joint conference, pp 780–787
25. Azli NA, Bakar MS (2004) A DSP-based regular sampled pulse width modulation (PWM)
technique for a multilevel inverter. In: International conference on power system technology,
POWERCON, vol 2, pp 1613–1618
26. Lord W, Hwang JH (1977) DC servomotors-modeling and parameter determination. IEEE
Trans Ind Appl 3:234–243
Oil Quality Analysis Using Image
Processing
Nivedita Daimiwal, Revati Shriram, Harish Shinde, Radhika Kulkarni,

and Apeksha Galewad
Abstract Now-a-days, people are demanding good quality of cooking oil (edible
oil). For a healthy heart, quality of cooking oil we consume plays a very important
role. As the cost of good quality cooking oil is more, there is a possibility of adulter-
ation of oil by the traders. FSSAI also decided to crack down on the sale of adulterated
edible oil. The campaign was made by FSSAI all over India in August 2020. But, the
test results are expected in a month’s time. The aim of the project is to determine the
change in various parameters of an oil based on the presence of adulterants or some
other oil residues. Determination of the presence of adulterants would be helpful in
avoiding many health conditions. Changing the ratios of the oils in the mixture can
show a change in the parameters that are being measured as the chemical contents of
the oils are changed. The objective is to design a system to detect the adulteration and
its percentage using machine learning method in various types of oil using oil image.
So, the initiative is made to develop a machine learning image processing system. In
this method, oil images were used to detect the adulteration in the oil. Features are
collected for different ratios of adulteration, and these features obtained by image
processing are used for detection of adulteration and its percentage. Variation in the
features like RGB, mean, variance and entropy is measured and plotted in Minitab
for various percentage of adulterations.
Keywords Pure groundnut (wooden pressed edible oil) · Refined oil (chemically
treated) · Quality · Adulterants · RGB · Entropy · Mean · Variance
1 Introduction
Oil and its quality-related issues have been dealt with by researchers worldwide
for more than 50 years. Contact-based and non-contact-based many techniques are
tried and tested till date. Quality of edible oil is checked mainly for the reasons
like to look for adulteration in the oil, to decide its suitability from the point of
N. Daimiwal (B) · R. Shriram · H. Shinde · R. Kulkarni · A. Galewad

Cummins College of Engineering for Women, Pune 411052, India
e-mail: nivedita.daimiwal@cumminscollege.in
https://doi.org/10.1007/978-981-19-6913-3_8
130 N. Daimiwal et al.
view of cardiovascular health of a person, then use of reduced quality edible oil
after multiple frying sessions for the vehicles as a biodiesel, etc. Depending on
the objective to check the oil quality, various methods are tried and tested by the
researchers internationally and nationally. The oil is image is capture by using digital
camera and various statistical parameters are extracted from the oil image by using
image processing techniques. Results show the variation in the statistical parameters
for different percentages of adulteration. These features can be used for adulation
detection and checking oil quality from oil images [1, 2].
1.1 Methodology
The oils that were used for observing the results were—pure groundnut, pure palm,
refined groundnut and a 50–50 mixture of refined groundnut and pure palm. A total
of 10 ml was considered as 100%. Refined groundnut, pure groundnut and pure palm
were taken in a 10-ml measurer. For a 50–50 ratio of the oils, 5 ml of the measurer
was filled with refined groundnut oil, and the rest 5 ml was filled with pure palm oil.
A clear picture of the 100% oil and the 50–50 mixture was taken with the help of a
camera of resolution 12 megapixel (f /1.8) + 12 megapixel (f /2.4) + 12 megapixel
(f /2.0). The parameters that are being measured are—RGB, entropy, mean, variance.
1.2 Review of Literature
Babatunde proposed image processing based on a chemical method. Babatunde

proposed image processing based on chemical method for identification of sunflower
and groundnut oil. Using this method of inspection for oil is too long and hard to
satisfy the oil circulation demands. Also, application of this method for inspection is
expensive. Although this method has many limitations, it can give more exact results
[3].
Shariff performed analysis on Agilent 1260 Infinity Analytical SFC System
combined with an Agilent 1260 Infinity ELSD Evaporative Light Scattering Detector
and proposed an image processing technique based on that. The ELSD was coupled
to the SFC module. To obtain good sensitivity and reproducibility, the addition of
a make up flow before the back pressure regulator with additional heating at the
entrance of ELSD is necessary [4].
Ariffin proposed the estimation of quality analysis on oil on the basis of color and
texture with the help of image processing. A digital camera having high resolution
is used to capture an image. After getting the image, the noise from the image is
removed by applying preprocessing technique. From that source image, the color
and texture characteristics of oil are extracted. The color characteristics are analyzed
using RGB and HSI analysis. Texture features like entropy are extracted [5].
Oil Quality Analysis Using Image Processing 131
Pearson proposed a technique of supercritical fluid chromatography (SFC) in

combination with evaporative light scattering detection (ELSD) for image processing
which is important for the determination of triglyceride composition of vegetable oils.
SFC separates triglycerides at much lower temperatures, and SFC permits greater
selectivity with shorter analysis times as compared to GC and HPLC, respectively
[6].
Heinemann proposed a capacitive oil sensor technique for image processing. The
sensor from Testo can make measurements in hot oil because of its ergonomic design.
The user is protected from heat, protected against water jets without a protective
cover, and there is also a visual alarm indication. Frying with cooking oil is not
healthy after couple of turns of use. Using this sensor when the TPM (Total Polar
Material)/TPC (Total Polar Compound) value is reached, the cooking oil can be
changed [7, 8].
Work carried out for the oil quality checking at national level in the last decade
is as follows:
Choudhary et al., India: Study presented the conventional and non-conventional
edible oils in India with the objective of how difficult it is to choose the correct edible
oil [9].
Prema, Department of Commerce, Coimbatore, India: Empirical study of the oil
brand preference in rural India was presented by the author. Study did not present
any oil property-related analysis [10].
1.3 Digital Image Processing
Image processing involves transforming an image into digital form and performing
various operations on it in order to extract desired information or to get an enhanced
image. The input in this process is the image itself, but the output is the desired
part of the image. The images are treated as a two-dimensional signal, and then,
signal processing methods are applied on it by any image processing system. Image
processing can be used for pattern recognition, to define various objects in an image,
to detect objects in an image, to retrieve high quality images, to improve noisy images
in order to perform image sharpening and also to detect and visualize objects that
are difficult to see.
2 System Specifications
The tool used for image processing is MATLAB. MATLAB is a multi-paradigm

programming language and a numeric computing environment. It has a built-in
toolbox for image processing which helps in maintaining the efficiency of the code.
The advantages of using MATLAB for image processing are as follows:
• MATLAB is a scientific language and hence helps implementation of advanced

algorithms through mathematical and numerical support.
• The final result has maximum numerical precision.
• The image processing steps can be documented and replicated at any given time.
• Image processing algorithms are more advanced in MATLAB than any other
language.
Version—MATLAB R2021a
Software—Windows 10
Hardware—Intel core i5 processor, minimum 2 GB RAM.
3 Observations
3.1 RGB Analysis
RGB—values of red, green and blue hues in a particular cropped portion of an image.
Entropy—measures the degree of randomness of an image. Shown in Eq. 1.
Mean—gives the contribution of individual pixel intensity for the entire image.
Shown in Eq. 2.
Variance—variance tells us how each pixel is different from the neighboring pixel
and is used to classify different regions of the image.

Entropy = sum p.∗ log 2( p) (1)
where p contains the normalized histogram counts
Mean = Sum of pixel values/Total number of pixel values (2)
3.2 Methodology
A cropped portion measuring 10 mm * 10 mm having the same coordinates was

selected from the image. The analysis was performed on that cropped portion for
extracting statistical parameters like mean, variance and entropy. Analysis of pure
groundnut (Wooden pressed) and 50–50 mixture of pure palm and pure groundnut
oil. Figure 1 shows steps incorporated in the oil image analysis.
Steps of Image Processing: Technique used for feature extraction:
• Read the image from the location
• Display the read image
• Select the cropped portion from the image
Fig. 1 Flowchart of
methodology
• Display the cropped portion

• Get the information about individual pixels in the image
• Calculate the entropy of the cropped image
• Get red hues for RGB analysis
• Get green hues for RGB analysis
• Get blue hues for RGB analysis
• Find mean of the 2-dimensional cropped image
• Find average of the red pixels
• Find average of the green pixels
• Find average of the blue pixels
• Convert the type of cropped image into double
• Calculate the variance.
For calculating the variance, it is mandatory that the image is of the type double.
This same code was used for the analysis of both pure groundnut and the 50–50
mixture of two oils. Oil image of pure groundnut oil and 50% adulteration oil (wooden
pressed groundnut and palm oil) captured using mobile camera are shown in Figs. 2
Table 1 shows the statistical parameter values.
Analysis of pure palm and refined groundnut oil (chemically treated)
• Assume two inputs and take image of jpeg format through location
• If filename is not of suitable format then it will return null
• Assign filename and path to variable file path
• Read file and storing it in variable
• Show image on display
Fig. 2 Pure groundnut oil

(wooden pressed)
Fig. 3 50–50 mixture of

pure palm and pure
groundnut oil
Table 1 Observations table of pure groundnut oil (wooden pressed) and 50% adulteration with
palm oil
Parameter Pure groundnut 50–50 mixture
RGB 145.06, 135.48, 57.20 155.07, 130.37, 61.59
Mean 112.7033 115.5109
Variance 1.6397e + 03 1.5669e + 03
Entropy 5.4382 4.9049
• Crop given image with specified pixel values

• Calculate total entropy of given image
• Get RGB values from given image
• Calculate mean RGB values
• Calculate average mean
• Convert image to double values
• Show above details.
Oil image of refined groundnut oil (chemically treated) and pure palm oil is shown
in Figs. 4 and 5 respectively. Statistical parameters of refined ground nut oil and pure
palm oil are given in Table 2.
The work is carried out for different percentage of adulteration of palm oil in
refined ground nut oil. Table 3 shows the parameters (mean, variance and entropy)
for 10%, 20%, 50% and for pure ground nut oil and palm oil.
Fig. 4 Refined groundnut

oil (chemically treated)
Fig. 5 Pure palm oil
Table 2 Observations from the analysis

Parameter Refined groundnut Pure palm
RGB 138.311, 119.68, 6.337 119.290, 110.26, 7.32
Mean 78.1925 149.67
Variance 4.19e + 03 3.208e + 03
Entropy 6.5201 6.245
Table 3 Comparison of different percentages of impurities

Parameter 10% adulteration 20% adulteration 50% adulteration 100% pure
Mean 123.3436 107.3899 115.5109 149.67
Variance 3.7230e + 03 1.5592e + 03 1.5669e + 03 3.208e + 03
Entropy 6.672 5.9573 4.9049 6.245
4 Conclusion
Various adulterations like mixing of two different oils can be detected. System of
digital image processing extracts the feature from oil images. The features extracted
are RGB, mean, variance and entropy. The variations in the features were observed
for oil adulteration. It is observed that the mean of pure groundnut oil (wooden
pressed) is low as compared to the 50% adulteration with palm oil. The variance of
pure groundnut oil is greater than that of adulterated refined oil. The entropy of pure
groundnut oil (wooden pressed) is greater than oil adulterated with palm oil. For the
pure 100% palm and refined oil, the observations were made. The mean of pure palm
oil is greater than that of refined groundnut oil (chemically treated), and the variance
is also greater. The entropy of refined groundnut oil is slightly greater than palm oil.
Based on the database of oil adulteration and corresponding features, we can use the
system for preliminary detection of oil adulteration using a mobile camera. Mobile
application can be developed using the same.
References
1. Kabagambe EK, Baylin A, Ascherio A, Campos H (2005) The type of oil used for cooking
is associated with the risk of nonfatal acute myocardial infarction in Costa Rica. J Nutr
135(11):2674–2679
2. Rashvand M, Akbarnis A (2019) The feasibility of using image processing and artificial neural
network for detecting the adulteration of sesame oil. AIMS Agric Food 4(2):237–243
3. Babatunde OO, Ige MT, Makanjoula GA (1988) Effect of sterilization on fruit recovery in oil
palm fruit processing. J Agric Eng Res 41:75–79
4. Shariff AR, Nor AA, Mispan R, Shattri M, Rohaya H, Goyal R (2002) Correlation between oil
content and DN values. In: MAP Asia conference. Geospatial application papers, India
5. Ariffin AA (1991) Chemical changes during sterilisation process affecting strippability and
oil quality. In: Seminar on developments in palm oil milling technology and environmental
management. Palm Oil Institute of Malaysia (PORIM), Bangi
6. Pearson YT (1996) Machine vision system for automated detection of stained pistachio nuts.
J Food Sci Technol 19(3):203–209
7. Heinemann PH, Hughes R, Morrow CT, Sommer HJ, Beelman RB, Wuest PJ (1995) Grading
of mushrooms using a machine vision system. Trans ASAE 37(5):1671–1677
8. Zahir E, Rehana S, Mehwish AH, Anjum Y (2014) Study of physicochemical properties
of edible oil and evaluation of frying oil quality by Fourier transform-infrared (FT-IR)
spectroscopy. Arab J Chem
9. Choudhary M, Sangha JK, Grover K (2014) Conventional and non-conventional edible oils:
an Indian perspective. J Am Oil Chem Soc 91(2):179–206
10. Prema R (2013) Empirical study of the oil brand preference in rural area. Indian J Appl Res
3(3)
Automatic Fabric Classifier Using
Nesterov-Accelerated Adaptive Moment
for Washing Machine
S. Elavaar Kuzhali, Kotha Manvitha, Anisha Singh, Lakshmi Pranathi,

Shreya Dhavule, and M. Poorvita
Abstract There are diverse fabric materials available in today’s world, which need
various washing methods to make them last longer. However, present technology
does not have an efficient mechanism for managing numerous fabrics when it comes
to washing these fabrics in the washing machine. Therefore, it is important to classify
different fabrics and give the optimized wash according to the quality of the fabrics.
So, clothes quality and lifetime are increased. If a neural network model is trained
to learn these features that are inherent to the fabric structures, by exhibiting a set
of cotton and silk fabric images, then the neural network would be able to correctly
distinguish between different fabric images. To implement this a convolutional neural
network (CNN) algorithm is used. The hardware part of the project includes a Rasp-
berry Pi, which is connected to a camera sensor, a temperature sensor, a motor, and
a water level sensor for controlling the level of water in the washing machine. It is
also automated to supply the required amount of soap and water based on different
fabrics.
Keywords Fabric classification · Adam · Nadam · Automation · CNN · VGG16
1 Introduction
Fabric is one of the ancient human inventions [1] that has evolved from handcrafted
woven textiles to modern machine-based electronic textiles. In the textile industry,
the weave pattern, which is the most important factor for fabrics, plays an important
role in design and redesign, analysis of the texture for structure, and appearance of
the fabrics [2–5]. Clothes are made from different fabrics, and each fabric has its
preferred method of washing technique. Segregating clothes based on their fabric is
the first step to ensure that the fabrics are healthy and durable. It is essential to perform
recognition of different types of fabric for different wash method for each fabric type.
Currently, there is no effective method to identify fabrics and wash them according to
S. Elavaar Kuzhali (B) · K. Manvitha · A. Singh · L. Pranathi · S. Dhavule · M. Poorvita

Electronics and Instrumentation Engineering, Ramaiah Institute of Technology, Bengaluru, India
e-mail: kuzhalisubbnaiah@msrit.edu
https://doi.org/10.1007/978-981-19-6913-3_9
140 S. Elavaar Kuzhali et al.
their required conditions. This long-term practice is detrimental to the durability of

the fabric. Presently, fabric-type recognition is relied on the manual operation using
human eyes and touch. Traditionally, this manual inspection is performed by experts
who require knowledge and experience.
However, along with various inconveniences such as heavy workload, low effi-
ciency, and time consuming, it also leads to subjective human factors, such as phys-
ical and mental stress, dizziness, and fatigue, which ultimately affect the recognition
results. Therefore, it is required to develop a set of fabric-type recognition automation
to produce high-quality products that meet customer needs.
The traditional washing machine requires the user to manually set all of the criteria
required for optimized and efficient washing. Consequently, there is a need for a
clever washing device that can not only identify the given fabric type but also set the
washing parameters automatically based on the fabric type.
Many researchers have used neural networks and artificial intelligence (AI) to
build models which identify and classify the type of fabric. Few examples are fabric
defect inspection system [6–8] that uses image processing technique to detect good
and faulty fabrics using the distribution of the areas of super pixels, image retrieval
system using different methods like convolutional neural network [9], sketch-based
system and with the help of metadata such as tags [10–12], fabric classification
[13–16], yarn-dyed fabric identification [17], erosion and dilation-based system for
organic fabric [18], feature extraction of different fabrics has been proposed [19–22],
and so on. Identifying faults in a limited number of fabric types does not fulfill the
functionality of an automatic washing machine, it also needs to identify the type
of fabric. This requires a very efficient technique with better accuracy. Therefore,
research continues to create the best model which can identify the fabric type with
utmost accuracy. Some researchers, namely Kampouris et al. [23, 24] have used a
photometric stereo sensor to collect the database of images and used VGG-M neural
network. Based on the findings, they determined that cotton, terry cloth, denim, and
fleece can all be correctly identified with nearly no errors. Nylon and viscose, on the
other hand, are difficult to classify due to their glossy look and tiny array of samples.
Anami et al. [18] have presented a paper that focused mainly on the classification of
a variety of plants, animals, and mineral origin fabric through images. Erosion and
dilation are the morphological operations used. The classification rates are predicted
using the ANN classifier.
However, most of the existing techniques involve the classification of yarn-dyed
fabric, based on its color and woven pattern. But only by having the knowledge of
color for washing, a definite washing method cannot be proposed since two different
types of fabric with the same color cannot be washed in a similar way.
Drawbacks of the existing papers are as follows:
• Most of the existing papers concentrate on finding faults in the fabric but do not
identify the type of fabric.
• Some paper involves identification of the fabric using metadata like tags and labels
associated with the images which are unreliable for images without captions.
Automatic Fabric Classifier Using Nesterov-Accelerated Adaptive … 141
• Almost none of the research papers used the output of the model to create an
application for the washing machine.
The paper presents an application of implementing the neural network algorithm
in the washing machine. An effective classification technique for commonly used
fabrics (cotton and silk) is developed. Convolutional neural network (CNN) is a type
of neural network built for two-dimensional image processing, although it can be
used with one-dimensional and three-dimensional data.
VGG is a classical convolutional neural network architecture. It is based on an
analysis of how to increase the depth of such networks.
VGG16 is one of the earliest CNN models used for large-scale image recognition.
The model achieves 92.7% top-5 test accuracy in ImageNet, which is a dataset of
over 15 million images belonging to 1000 classes. VGG16 is distinguished by the
fact that it focuses on having convolution layers of 3 × 3 filter with stride 1 rather
than a large number of hyper-parameters, and it always uses the same padding and
max pooling layer of 2 × 2 filter with stride 2. This convolution and max pool
layer configuration are maintained throughout the architecture. Finally, two fully
connected FC (completely connected layers) and a SoftMax layers are included.
VGG16 architecture is implemented from scratch using TensorFlow and Keras for
the classification of fabrics. A dataset is created for each type of fabric, such as cotton
and silk. Fabric categorization is done using convolutional neural network technique.
The developed neural network model is put into Raspberry Pi attached to the camera
sensor after training, validating, and testing the performance of the convolutional
neural network framework on fabric photos. Images of fabrics are captured by the
camera sensor for classification.
Using a convolutional neural network, this research offers an excellent classifi-
cation strategy for regularly used materials (cotton and silk). The proposed method
makes use of the VGG16 network convolution model, which has a 92.7% accuracy,
and it is trained using Adam, SGD, and Nadam optimizer. Since Nadam optimizer
gave the best result, it has been used in the proposed model. After training, vali-
dating, and testing the performance of the convolutional neural network model on
fabric images, the model is deployed into Raspberry Pi which is connected to a
camera sensor. The paper presents an application of neural networks, where it is
integrated with the washing machine to make it smarter.
Major contributions of the paper are as follows:
• The proposed model identifies the type of fabric between cotton and silk with
Nadam optimizer having a learning rate of 0.00001 to obtain an improved
validation accuracy of 98.7% and training accuracy of 99.3%.
• The model is deployed with Raspberry Pi to automate the process of washing in
the washing machine.
The further sections are depicted as follows. Section 2 proposed an approach for
the washing machine which consists of two subsections explaining fabric classifica-
tion using CNN model and circuit design and process flow, respectively. Section 3
discusses the results obtained. Section 4 gives the conclusion of this paper.
2 Proposed Approach for Washing Machine
The proposed method aims at developing a prototype for classification of fabrics and
a suitable washing technique based on the load of fabric clothes given. As stated
before, the approach is composed of two main steps. In the first step, a CNN model
is developed for the classification of given fabric images. The second step consists
of design of washing techniques for different sets of loads and setting up of the
hardware prototype using Raspberry Pi and other required hardware components. In
the following sections, each of these steps has been explained in detail.
2.1 Fabric Classification Using CNN Model
To solve the problem of classification of types of fabrics, convolutional neural

network is used. Figure 1 shows the flowchart of CNN model. VGG16 is a network
convolution model for neural networks that can achieve a test accuracy of 92.7%
among the top five on ImageNet. This is a collection of over 15 million photos
divided into 1000 categories. The datasets created by image augmentation can be
separated into two types: training datasets and validation datasets. Image augmen-
tation is used to enlarge the training dataset and improve the model’s performance
and generalization capabilities by creating a new dataset from the input photos. The
validation set is only used to evaluate the model’s performance, while the training set
is utilized to train it. A dataset of 4331 images of cotton and silk was collected out of
which a training dataset comprised 3466 images and a validation dataset comprised
865 images.
The amount of data available usually improves the performance of convolutional
neural networks. VGG16 model with four convolutional blocks has been proposed
by slightly modifying the last three dense layers by including average pooling layer,
flatten layer, and ReLU activation layer. VGG16 has a convolutional layer of 3 * 3
filter and a max pool layer of 2 * 2 filter. To down sample the identification of
features in feature maps, pooling is required. The pooling methods used by VGG16
include max pooling and average pooling. The maximum value for each patch of
the feature map is calculated using max pooling. The average value for each patch
on the feature map is calculated using average pooling. Throughout the architecture,
VGG16 maintains this arrangement of convolution and pooling layers. It features
two fully connected layers at the end. After the convolution, the data is passed on
to the dense layer in order to smooth out the convolution from the emerging vector.
The images will be flattened to size 224 * 224.
The rectified linear activation function (ReLU) activation is used to stop the
transmission of negative values through the network.
Mathematically, ReLU activation function is expressed as
f (x) = max(x, 0) (1)

Fig. 1 Flowchart
The SoftMax level generates the value between 0 and 1 based on the class of
the image identified. SoftMax is the following function mathematically, where z is a
vector of inputs to the output layer and j is an index of output units from 1, 2, 3 … k.
ez j
σ (z) j = Σ K for j = 1 (2)
k=1 ez k
Different optimizers like stochastic gradient descent (SGD), adaptive moment

estimation, and Nesterov-accelerated adaptive moment estimation (Nadam) were
used to test the model, and their efficiency was compared. Adam optimizer can be
written as

α 1 − β1 ∂ L
wt+1 = wt − / β1 V̂t−1 + · (3)
1 − β1t ∂wt
Ŝt + ∈
Nadam employs Nesterov to update the gradient one step ahead by substituting
the current in the above equation for the previous:

α 1 − β1 ∂ L
wt+1 = wt − / β1 V̂t−1 + · (4)
1 − β1t ∂wt
Ŝt + ∈
where
Vt St
Δ
V̂t = t , St = (5)
1 − β1 1 − β2t
And
∂L
Vt = β1 Vt−1 + (1 − β1 ) (6)
∂wt

∂L 2
St = β2 St−1 + (1 − β2 ) (7)
∂wt
m and v are both set to zero. The following are the default values (taken from Keras):
α = 0.002, β1 = 0.9
β2 = 0.999, ε = 10−7
Out of all the optimizers mentioned, the best results were given by Adam and
Nadam. The training and validation accuracy for Adam came out to be 98.25 and
98.96, respectively. The training and validation accuracy for Nadam came out to be
99.3 and 98.7, respectively.
Out of these two, Nadam is used to train the VGG16 model since better results
were observed using this optimizer. The model is trained until the maximum epochs
are reached, and the best model is saved. The dataset ran for 100 epochs with steps
per epoch being 70 and a learning rate of 0.00001. The model was successful in
classifying the images between cotton and silk.
2.2 Circuit Design and Process Flow
The circuit architecture of fabric classification and washing machine prototype is

shown in Fig. 2.
Fig. 2 Circuit architecture

of washing machine with
fabric classification
The circuit is constructed with two main sections which are input and output. The
first section consists of the camera sensor, temperature sensor, float sensor, and IR
sensor connected with Raspberry Pi as input devices.
The second section consists of a DC motor, liquid crystal display, and water pump
connected with Raspberry Pi as output devices. DC motor is operated using a motor
driver for which a 9 V supply is given. For temperature sensors also, 9 V supply is
used. For water pump, relay is used as a driver or switches to produce the desired
voltage.
Figure 3 describes about the block diagram of the proposed model. In order to
have a longer and safer usage period, the camera sensor in the washing machine
that is placed near the rim of the door should be water resistant. The camera sensor
used for the prototype is not water resistant, so care was taken to implant the camera
sensor to avoid water contact. To get accurate predictions, it is suggested that the
camera sensor be placed such that the illumination is uniform and the fabric should
be between 4 and 6 cm from the camera sensor lens to capture clear images. The
output of the VGG16 model is deployed into the Raspberry Pi microcontroller using
TensorFlow Lite framework, where it is programmed to control the operations of a
temperature sensor, DC motor, float sensor, and IR sensor. Based on the identified
fabrics, the protocol is set for the temperature of the water, washing time, spin rate,
and detergent quantity. After the protocol is set, the system will check if the washing
machine door is closed or not using an IR sensor. If the door is not closed, it will
give an indication using LED light.
However, if the door is closed then water will be poured till the set level is reached
which will be indicated by the float switch which goes up and the water pump stops.
The detergent will be added according to the preset quantity. Simultaneously heater is
switched on, the water gets heated till the preset temperature for the particular fabric
is reached, and the DC motor will be ON for a specified time having a specified
spin rate. The standard specified water temperature, spin rate, and time of wash for
cotton are 45 °C, 900–1400 rpm, and 30 min, respectively. The standard specified
water temperature, spin rate, and time of wash for silk are 32 °C, 800–1000 rpm, and
35 min, respectively.
Fig. 3 Block diagram of

proposed method
3 Results and Discussion
The CNN model for the classification of fabrics is developed using a Python program-
ming language with TensorFlow framework and Keras API for image processing
tools. An open-source archive of images was available that was originally prepared
for research to acquire a dataset for training the neural network. Dataset consisting
of 356 cotton images and 168 silk images is taken from an open-source archive.
The images are of 400 × 400 resolution, and the feature size is of dimension 224
× 224. The dataset also consists of 2058 cotton images and 1749 silk images with
phone cameras and re-sized all images to 400 × 400. Training dataset consisted of
4331 images and a validation dataset of 153 images belonging to cotton and silk
classes. VGG16 network model is used to train the fabric image dataset. VGG16
model consists of 16 layers in the form of five convolutional blocks, and each block
has convolution layers of 3 × 3 filter followed by a 2 × 2 filter max pooling layer
and last three fully connected layers.
Results were observed by using the transfer learning VGG16 convolution neural
network model with different optimizers such as stochastic gradient descent (SGD),
adaptive moment estimation, Nesterov-accelerated adaptive moment estimation
(Nadam), Adamax which is a variation of Adam estimation, and root mean square
propagation. Nadam and Adam optimizers are preferred because they provide accu-
rate, quick results, are preferred for sparse data, and decrease noise issues. 94%
validation accuracy and 99.60% training accuracy are obtained using all five blocks
of the VGG16 model. But better result of 97% validation accuracy is seen by using
only four blocks of the VGG16 model. Hence, VGG16 model with four convolu-
tional blocks has been proposed by slightly modifying the last three dense layers
by including average pooling layer, flatten layer, and ReLU activation layer. From
Table 1, Adam and Nadam show higher accuracy with better results compared to
other optimizers. Training the model with Adam optimizer has achieved training
accuracy 98.25% and validation accuracy 98.96%. Similarly, with Nadam optimizer
there is a slight increase in training accuracy with 99.3% and validation accuracy
with 98.7%. Optimizers are used to adjust the characteristics of the convolutional
neural network, such as weights and learning rate, to reduce the losses. Optimization
algorithms or techniques are in charge of lowering losses and delivering the most
reliable performance.
Table 2 shows the summary of the VGG16 model used, model summary includes
complete information on types of layers in the model, output shape for each layer,
and parameters or weights of each layer which is updated during training the model
using optimization techniques to minimize the error. It is also observed from Table 2
that there are four convolutional blocks.
Figure 4 shows that model accuracy is increased with a decrease in the loss. Test
or validation accuracy is close to training accuracy. This represents that the model
has achieved good accuracy with silk and cotton datasets.
Table 1 Model accuracy with Adam, Nadam, and SGD optimizer—classification of cotton and
silk
Optimizer Learning rate Epochs Training accuracy % Validation accuracy %
Adam 0.00001 100 98.25 98.96
Nadam 0.00001 100 99.3 98.7
SGD 0.00001 120 69.97 65
Table 2 Model summary

Type of layer Output shape Parameter Blocks
conv2d_13 (Conv2D) (None, 224, 224, 64) 1792 Block1
conv2d_14 (Conv2D) (None, 224, 224, 64) 36,928
max_pooling2d_5 (MaxPooling2D) (None, 112, 112, 64) 0
conv2d_15 (Conv2D) (None, 112, 112, 128) 73,856 Block2
conv2d_16 (Conv2D) (None, 112, 112, 128) 147,584
conv2d_17 (Conv2D) (None, 56, 56, 256) 295,168 Block3
conv2d_18 (Conv2D) (None, 56, 56, 256) 590,080
conv2d_19 (Conv2D) (None, 56, 56, 256) 590,080
conv2d_20 (Conv2D) (None, 28, 28, 512) 1,180,160 Block4
conv2d_21 (Conv2D) (None, 28, 28, 512) 2,359,808
conv2d_22 (Conv2D) (None, 28, 28, 512) 2,359,808
average_pooing2d_1 (Average) (None, 7, 7, 512) 0 Output layers
flatten_1 (Flatten) (None, 25,088) 0
dense_2 (Dense) (None, 128) 3,211,392
dense_3 (Dense) (None, 2) 387
Fig. 4 Plot of model

accuracy and loss
Figure 5 shows the confusion matrix for the predictions made by the trained model.
336 cotton images and 248 silk images which are not included in the training dataset
are taken, out of which 305 and 225 images are correctly predicted as cotton and
silk, respectively, thus it can be stated that the model is predicted with 90% accuracy
for non-trained images and it has an error rate of 10%.
After the model is trained, it is saved in ‘.h5’ format. In the ‘.h5’ file, it is easy
to store structured data; hence, all the weight parameters are stored in this file. This
file is used to make predictions on the input image with a confidence rate mentioned.
Figure 6 represents the prediction and confidence rate on input images.
For estimating the accuracy of the model for non-trained images in the existing
system, 91 cotton fabrics and 70 silk fabrics are taken. These fabrics are captured
Fig. 5 Confusion matrix

Fig. 6 Prediction on silk and cotton images
Fig. 7 Predictions on fabric images using Raspberry Pi
by the camera sensor connected to the Raspberry Pi. The results obtained from the
prediction are that 83 cotton fabrics are predicted correctly as cotton and 51 silk
fabrics are correctly predicted as silk. Thus, it can be stated that the accuracy of
non-trained fabric images is 82.025% and the error rate obtained is 17.97%.
Figure 7 is the output of predicted images using Raspberry Pi and a camera module
displayed on a 16 × 2 liquid crystal display. Based on the results of fabric images,
predicted spin rate of the motor, the temperature of the water, and time of washing
are set.
4 Conclusion
A convolutional neural network model is used for classifying fabrics of cotton and silk
images. A VGG16 architecture model with network weights is used for training the
model. Four blocks of the VGG16 model with an average pooling layer, a flattening
layer, and a ReLU activation layer have been used for the proposed model. Adam,
SGD, and Nadam optimizers are used to evaluate the model for better performance.
69.97% and 98.25% training accuracy have been achieved using SGD and Adam
optimizers, respectively. The Nadam optimizer is used for the final model as it has
higher performance and is efficient in improving validation accuracy. 99.3% training
accuracy and 98.7% validation accuracy are obtained with the final model using the
Nadam optimizer with a learning rate of 0.00001. For non-trained data, the accuracy
is 82.025% in the existing system. The proposed model helps in improving the
quality of the wash and the longevity of the cloth since the amount of detergent and
water level are automatically set based on the load of the fabric and the type of fabric
detected. This AI model can also be applied to the textile industries, where automatic
classification of various fabrics is required instead of manual labor.
5 Future Work
Future work is to ensure that neural network model can classify more than two classes,
i.e., nylon, polyester, wool, linen, denim, etc., and to identify multiple images simul-
taneously. Further experiments can be conducted using different network models for
image classification like ResNet50, Inception v3, and EfficientNet since they have
an important influence on the results obtained at the end. To extend the result of
fabric classification to different applications like laundries and textile industries, this
AI model can be applied to the textile industry, where automatic classification of
various fabrics is required instead of manual labor.
References
1. Riello G, Roy T (2009) How India clothed the world: the world of South Asian textiles 1500–
1850. Brill
2. Osborne EF (2010) Weaving process for production of a full-fashioned woven stretch garment
with load carriage capability. U.S. patent 7,841,369
3. Guarnera GC, Hall P, Chesnais A, Glencross M (2017) Woven fabric model creation from a
single image. ACM Trans Graph (TOG) 36(5):1–13
4. Shih C-Y, Kuo C-FJ, Cheng J-H (2016) A study of automated color, shape and texture analysis
of Tatami embroidery fabrics. Text Res J 86(17):1791–1802
5. Jing J, Xu M, Li P, Li Q, Liu S (2014) Automatic classification of woven fabric structure based
on texture feature and PNN. Fibers Polym 15(5):1092–1098
6. Ouyang W, Xu B, Hou J, Yuan X (2019) Fabric defect detection using activation layer embedded
convolutional neural network. IEEE Access 7:70130–70140
7. Hussain I, Ather M, Khan B, Wang Z, Ding S (2020) Woven fabric pattern recognition and
classification based on deep convolutional neural networks. Electronics 9(6):1048
8. Garg M, Dhiman G (2020) Deep convolution neural network approach for defect inspection of
textured surfaces. J Inst Electron Comput 2(1):28–38
9. Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classification with deep convolutional
neural networks. Commun ACM 60(6):84–90
10. Xiang J, Zhang N, Pan R, Gao W (2019) Fabric image retrieval system using hierarchical
search based on deep convolutional neural network. IEEE Access 7:35405–35417
11. Lei H, Yi Y, Li Y, Luo G, Wang M (2018) A new clothing image retrieval algorithm based on
sketch component segmentation in mobile visual sensors. Int J Distrib Sens Netw 14(11)
12. Joshi KD, Bhavsar SN, Sanghvi RC (2014) Image retrieval system using intuitive descriptors.
In: ICIAME 2014. Elsevier, India. Procedia Technol 14:535–542
13. Chae YK (2020) Fabric identifying method, apparatus, and system. U.S. patent application
16/599,960
14. Da Silva Barros MAC, Ohata EF, da Silva SPP, Almeida JS, Filho PPR (2020) An innovative
approach of textile fabrics identification from mobile images using computer vision based on
deep transfer learning. In: International joint conference on neural networks (IJCNN) 2020.
IEEE, Glasgow, UK, pp 1–8
15. Varma M, Zisserman A (2008) A statistical approach to material classification using image
patch exemplars. IEEE Trans Pattern Anal Mach Intell 31(11):2032–2047
16. Xie T, Grossman JC (2018) Crystal graph convolutional neural networks for an accurate and
interpretable prediction of material properties. Phys Rev Lett 120(14)
17. Zhang H, Zhang L, Li P, Gu D (2018) Yarn-dyed fabric defect detection with YOLOV2 based
on deep convolution neural networks. In: IEEE 7th data driven control and learning systems
conference (DDCLS) 2018. IEEE, China, pp 170–174
18. Anami BS, Elemmi MC (2019) ANN approach for classification of different origin fabric
images. Int J Image Graph Signal Process 10(12):29–38
19. Makantasis K, Karantzalos K, Doulamis A, Doulamis N (2015) Deep supervised learning for
hyperspectral data classification through convolutional neural networks. In: IEEE international
geoscience and remote sensing symposium (IGARSS) 2015. IEEE, Italy, pp 4959–4962
20. Li Z, Meng S, Wang L, Zhang N, Gao W (2019) Intelligent recognition of the patterns of
yarn-dyed fabric based on LSRT images. J Eng Fibers Fabr 14
21. Benco M, Hudec R, Kamencay P, Zachariasova M, Matuska S (2014) An advanced approach
to extraction of colour texture features based on GLCM. Int J Adv Robot Syst 11(7):104
22. Wang X, Wu G, Zhong Y (2018) Fabric identification using convolutional neural network. In:
International conference on artificial intelligence on textile and apparel. Springer, Cham, pp
93–100
23. Kampouris C, Zafeiriou S, Ghosh A, Malassiotis S (2016) Fine-grained material classification
using micro-geometry and reflectance. In: European conference on computer vision. Springer,
Cham, pp 778–792
24. Sparavigna AC (2017) Image segmentation applied to the analysis of fabric textures. Philica.
HAL Id: hal-01633061
System Identification, Stability Analysis
and PID Controller Design for PEM
Electrolyzer
Aruna Rajaiah and Jaya Christa Sargunar Thomas
Abstract In a hybrid renewable energy system (RES), different types of energy

sources are integrated for meeting the continuous power demand. To overcome the
problem of intermittent availability of energy RES, suitable energy storage system
is required. The energy storage as hydrogen gas is more efficient and suitable for
long-term storage applications. Proton exchange membrane (PEM) electrolyzer is
used to store the surplus amount of energy from RES in the form of hydrogen gas. In
this paper, initially the mathematical modeling of PEM electrolyzer is developed to
know the influence of various parameters on the performance. It is noted that the cell
voltage and efficiency of the electrolyzer depend on its temperature. Based on this
observation, temperature is taken as input parameter, and the system model is identi-
fied for PEM electrolyzer using system identification technique. The different black
box approaches are used to identify the suitable system by comparing the response
on the basis of percentage of fitness, cost function and final prediction error. It is
found that Box–Jenkins method is appropriate for PEM electrolyzer with percentage
fitness of 90.5%. To check the system stability, Lyapunov stability analysis is applied
and it is found to be stable. For the stable system, conventional PID controller based
on Ziegler–Nichols method is used to maintain constant cell voltage by varying the
temperature of PEM electrolyzer. For the system with the PID controller, the rise
time is 3.8 s, settling time is 75.3 s, peak time is 9.3 s and overshoot is 57.4%.
Keywords PEM electrolyzer · System identification · Lyapunov stability

analysis · PID controller
A. Rajaiah (B)
P.S.R. Engineering College, Sivakasi, Tamil Nadu, India
e-mail: arunamuthukrishnan@gmail.com; aruna@psr.edu.in
J. C. S. Thomas
Mepco Schlenk Engineering College, Sivakasi, Tamil Nadu, India
e-mail: stchrista@mepcoeng.ac.in
https://doi.org/10.1007/978-981-19-6913-3_10
154 A. Rajaiah and J. C. S. Thomas
1 Introduction
The utilization of renewable energy sources is increased nowadays, to reduce the

usage of fossil fuels. While using renewable energy sources, the main limitation
is their irregular energy. There are several alternatives available to overcome this
drawback. Among those, saving the excess energy as hydrogen gas using electrol-
ysis method is considered as one of the options. Since hydrogen is an abundantly
available chemical element, environment friendly, effective and not dependent on
hydrocarbon-based fuels. Various electrolysis techniques are available, among them
PEM electrolyzer has the following merits such as compact design, high working
current densities, pure gas generation and production at high pressure. Many research
works are carried out in the modeling and analysis of PEM electrolyzer.
A review on recent development in hydrogen production using PEM electrolyzer
is presented in [1]. The challenges related to the components of PEM cell and
electro-catalyst are addressed. A case study on integrated system with solar PV
panel, hydrogen storage system and a geothermal heat pump is carried in [2]. On
analyzing the integrated system, it is found that the system is able to meet the require-
ment in agriculture field as greenhouse heating. Reference [3] provides a review on
modeling of PEM electrolyzer for the beginners including the basic principles of this
technology. In [4], the hydrogen generator system based on the PEM electrolysis is
developed using MATLAB Simulink tool. The cell and stack behavior under various
temperatures and pressures are presented.
The photovoltaic panel with electrolyzer and fuel cell is given in [5]. The system is
simulated to verify the effectiveness. It provides solution for energy storage problem
and is able to meet the power demand. In a standalone renewable hybrid power
system, the capacity of PEM electrolyzer for hydrogen production is evaluated in
[6]. A comprehensive analysis for PEM electrolyzer is done [7] for improving the
performance, reliability and to reduce cost. The literature [8] provides modeling of
electrolyzer with the controller for hydrogen production. In [9], a simplified model
of PEM electrolyzer is presented with a comparison of experimental data for various
temperatures. It is concluded that the electrolyzer cell voltage and efficiency depend
on the temperature of the system.
For controlling the system, a suitable system model is required. It can be found
by using system identification techniques. In literature [10], the black box and white
box approaches in system identification technique for cascade tank are applied and
compared to choose the most appropriate model. For a heating system, the best
model structure is found among the black box approaches based on the estimated and
validated test results [11]. A more realistic model structure is identified using black
box approach in system identification techniques for PEM fuel cell [12]. Similarly,
the model identification and controller design for metal hydride storage tank during
desorption process are presented in [13].
In literature [14], the PEM fuel cell is combined with PEM electrolyzer. A sliding
mode control is developed to control the fuel cell system and the rate of hydrogen
production in the electrolyzer. A sliding mode pulse width modulation controller is
System Identification, Stability Analysis and PID Controller Design … 155
System Identifi- Lyapunov PID

Modeling of Controller
cation for PEM stability
PEM Electrolyzer Design
Electrolyzer analysis
Fig. 1 Flow diagram presents the flow of work in this paper
designed to ensure the voltage stability and minimize the ripple current in electrolyzer
[15]. The stability analysis and design of conventional PID controller for various
linear and nonlinear systems are provided in [16, 17]. Based on the literature survey,
it is found that there is a research gap in model identification for PEM electrolyzer.
Hence, the novelty of this work is system identification for PEM electrolyzer and
design of PID controller to maintain a constant output voltage. The flow of this
research work is presented (see Fig. 1).
2 Modeling of PEM Electrolyzer
In PEM electrolyzer, water molecules dissociate into hydrogen ions and oxygen on
applying electrical energy. The PEM electrolyzer is compact in design and produces
pure hydrogen gas at high pressure and high current density [3–6]. It has four modules
such as anode, cathode, membrane and voltage modules and are interlinked to each
other. On applying the input current, an electrochemical reaction takes places where
the hydrogen ions developed from the anode and passed to the cathode through the
catalyst layer where the H2 gas is produced. Thus, the rate of hydrogen production
is depending on the current density.
The mathematical model of PEM electrolyzer is taken from [3]. A single cell with
100 cm2 active area is considered and the maximum current density is assumed as
1.8 A/cm2 . The simulation diagram of PEM electrolyzer is presented (see Fig. 2). The
electrolyzer works at a temperature range of 343–373 K and the pressure is considered
as 1 bar. Increasing the temperature of PEM electrolyzer from 343 to 373 K reduces
the cell voltage and increases the efficiency (see Figs. 3 and 4, respectively). The effi-
ciency is calculated based on the thermoneutral voltage, 1.48 V. Thus, on increasing
the temperature, the activation loss and ohmic over potential are decreased due to
the increase in reaction kinetics and improved ionic conductivity.
As on increasing the current density, the rate of power loss on the stack becomes
large and the efficiency starts to decrease. By manipulating the operating temperature,
the losses can be reduced and thus higher efficiency of PEM electrolyzer can be
maintained.
Fig. 2 Simulation diagram of PEM electrolyzer
1.8
Cell Voltage (V)
1.6 Temperature at 343K

1.4 Temperature at 353K
Temperature at 373K
1.2
1
0 0.5 1 1.5 2
Current density (A/cm2 )
Fig. 3 J–V characteristic curves obtained for PEM electrolyzer
100
80
Efficiency (%)
60
Temperature at 343K
40
Temperature at 353K
20 Temperature at 373 K
0
0 2 4 6 8 10
Current density (A/cm2 )
Fig. 4 Efficiency curve of PEM electrolyzer

3 System Identification
System identification is a tool to identify the numerical expression of the system

with the help of input and output real-time measurement data or experimental data
without a knowledge of what is happening inside the system [11]. The process of
model identification is to select the modeling technique first, then to choose the
structure of model and finally the model is estimated and validated. On the basis of
final prediction error, loss function and percentage of fitness, the best system model
is selected. The system identification technique has three common approaches such
as white box, gray box and black box modeling methods.
The white box and gray box modeling methods need knowledge of internal struc-
ture of the system along with large amount of data. They are very complex and
consuming time is also high. If the internal model structure and first principle equa-
tions are unknown, then the black box technique can be used to identify the model
structure only by having measured input and output data [12]. In this research work,
black box techniques such as Auto-Regressive Exogenous (ARX), Auto-Regressive
Moving Average Exogenous Model (ARMAX), Output Error (OE) and Box–Jenkins
(BJ) are used to find out the suitable model structure for PEM electrolyzer.
In discrete time, the general linear time invariant model [10] can be written as
y(t) = G(q)u(t) + H (q)e(t) (1)
where y(t) is signal output of the model and it corresponds to the input u(t) and
the disturbance e(t). Based on Eq. (1), the models are able to depict the dynamical
relations based on the linear difference equations. The ARX model is expressed [13]
as
y(t) + a1 y(t − 1) + · · · + ana (t − n a ) = b1 u(t − 1) + · · · + bn b u(t − n b ) + e(t)

(2)
where e(t) is direct error (white noise term). The variables A and B are stated in terms
of delay operator q, thus the expression is
A(q) = 1 + a1 q −1 + · · · + ana q −na (3)
B(q) = b1 q −1 + · · · + bn b q −n b (4)
By comparing Eqs. (1)–(4), we get
B(q) 1
G(q) = , H (q) =
A(q) A(q)
The ARMAX model structure contains the error equation as a moving average of
white noise. The model structure is expressed as
y(t) + a1 y(t − 1) + · · · + ana (t − n a ) = b1 u(t − 1) + · · · + bn b u(t − n b )

+ e(t) + c1 e(t − 1) + · · · + cn c u(t − n c )
(5)
where ai , bi and ci are model parameters and e(t) is white noise term.
In the Output Error (OE) model, the linear difference equation is correspondence
between the input and undisturbed output.

w(t) + f 1 w(t − 1) + · · · + f n f w t − n f = b1 u(t − 1) + · · · + bn b u(t − n b )
(6)
where f and b are model parameters.

In Box–Jenkins (BJ) model, the model structure corresponds to the disturbance
properties and is independent of system dynamics. It is expressed as,
B(q) C(q)
y(t) = u(t) + e(t) (7)
A(q) D(q)
where B(q), A(q), C(q) and D(q) are polynomials with respect to the delay q.
3.1 System Identification of PEM Electrolyzer
From the characteristics and efficiency curve of PEM electrolyzer shown in Figs. 3
and 4, it is found that the efficiency of the system can be maintained by controlling the
operational temperature. Hence for model identification, the operational temperature
is taken as input variable and the variation of cell voltage is taken as output variable
(see Fig. 5). The estimated output of PEM electrolyzer (see Fig. 6).
Input and output signals

2
Cell Voltage (V)
1.5
0 200 400 600 800 1000
360
Temperature (K)
355
350
0 200 400 600 800 1000
Time (seconds)
Fig. 5 Input and output variables of PEM electrolyzer

Measured and simulated model output

2
1.9
Cell Voltage (V)

1.8
1.7 BJ - 90.6%
OE - 89.05%
1.6 1.85 ARMAX - 77.62%
500 520 540 Cell Voltage
1.5
0 200 400 600 800 1000
Time (Seconds)
Fig. 6 Comparison of simulated output and estimated output by black box approach for PEM
electrolyzer
Table 1 Comparison of black box models for PEM electrolyzer

Model Final prediction error Cost function Fitness (%)
ARMAX 1.743 × 10−5 1.756 × 10−5 77.62
Output error 2.687 × 10−5 3.21 × 10−5 89.05
Box–Jenkins 1.826 × 10−5 2.364 × 10−5 90.6
From Fig. 6, it is found that the best fitness is provided by Box–Jenkins model,
followed by Output Error model and then by ARMAX model. The ARX model
provides less than 20% of fitness and it is neglected. For the above three models,
the cost function and final prediction values are listed in Table 1. Based on the
comparison table, it is identified as the Box–Jenkins model is best fitted compared
to other models.
4 Lyapunov Stability Analysis
Lyapunov stability analysis is the most suitable tool for analyzing nonlinear systems
[16]. A nonlinear autonomous system is represented by state equation as
ẋ = f (x); f (0) = 0 (8)
The sufficient stability condition for nonlinear systems is as referred in [16].

The Lyapunov function is represented as
V (x) = x T P x (9)
where P is a positive definite Hermitian matrix. The time derivative of V (x) along
any trajectory is
·
V (x) = ẋ T P x + x T P ẋ
= (x A)T P x + x T P Ax (10)
·
V (x) = x T A T P + P A x (11)
·
Since V (x) is selected as positive definite, for asymptotically stable V (x) be
negative definite. Therefore,
·
V (x) = −x T Qx (12)
where,

Q = − A T P + P A = positive definite. (13)
4.1 Stability Analysis for PEM Electrolyzer
From the Box–Jenkins model of PEM electrolyzer, the system matrices are obtained
as

−0.0564 −0.0002 1
A= B=
1 0 0

C = 5.08e−4 8.45e−7 D = [0]
For Lyapunov stability method, the Q matrix is assumed as positive definite matrix

10 P11 P12
Q= and P =
01 P12 P22
The Lyapunov function is
V (x) = x T P x
From Eq. (13), the positive definite P matrix is obtained as,

44.3e3 2.5e3
P= >0
2.5e3 149.87
Hence, P is positive definite. The equilibrium state at the origin is asymptotically

stable in large and a Lyapunov function is
V (x) = 44.3e3 x12 + 5000x1 x2 + 149.87x22
·
V (x) = −x22
·
V (x) is negative definite. Hence, the PEM electrolyzer is asymptotically stable.
5 Design of Controller
The conventional PID controllers more commonly used are electronic, hydraulic
and pneumatic types depending on the needs of industries [17]. In design of PID
controller, there are various possible techniques available for identifying the param-
eters of the controller. It will help to meet up the specification such as transient
and steady-state response of the closed loop system. This process of selecting the
controller parameters is known as controller tuning [16]. Ziegler–Nichols tuning
method is able to find the parameters such as proportional gain, integral time constant
and derivative time constant as K p , T i and T d , respectively. The block diagram of
PID controller is represented (see Fig. 7).
5.1 PID Controller for PEM Electrolyzer
In PEM electrolyzer, the cell voltage is controlled by manipulating the operational

temperature of cell to meet the specified cell voltage which is given as reference
voltage 1.8 V. The PID gain parameters identified based on the Ziegler–Nichols
tuning method are K p = 67.2, T i = 196.75 and T d = 49.18. The simulation result of
PID controller for PEM electrolyzer (see Fig. 8).
From Fig. 8, the time domain response is obtained with rise time as 3.8 s, settling
time as 75.3 s, peak time as 9.3 s and percentage of overshoot as 57.4%. For improving
the response, the controller tuning can be done by optimization techniques.
Fig. 7 Block diagram of

PID controller for a closed
loop system
3
Reference voltage
2.5 PID controlled output
Cell voltage (V)

2
1.5
1 Rise time : 3.8 seconds

Settling time : 75.3 seconds
0.5 Peak time : 9.3
Overshoot :57.4%
0
0 20 40 60 80 100
Time (seconds)
Fig. 8 PID controller responses for PEM electrolyzer
6 Conclusion
This paper presents the model structure, Lyapunov stability analysis and design
of PID controller for PEM electrolyzer. The system model is identified by using
black box approaches of system identification technique. On comparing the estimated
and validated output of ARX, ARMAX, Output Error and Box–Jenkins methods,
the Box–Jenkins method has obtained better percentage of fitness as 90.8%, final
prediction error as 1.826 × 10−5 and cost function as 2.364 × 10−5 . For the identified
Box–Jenkins model, the system stability is analyzed using Lyapunov technique and
found that the system is stable. PID controller is designed by using Ziegler–Nichols
tuning method. The desired cell voltage of 1.8 V is obtained with a rise time of 3.8 s,
settling time of 75.3 s, peak time of 9.3 s and overshoot of 57.4%. Further, to improve
the controller response, optimization techniques can be adopted for controller tuning.
References
1. Kumar SS, Himabindu V (2019) Hydrogen production by PEM water electrolysis—a review.
Mater Sci Energy Technol 2(3):442–454
2. Pascuzzi S, Anifantis AS, Blanco I, Mugnozza GS (2016) Electrolyzer performance analysis
of an integrated hydrogen power system for greenhouse heating. A case study. Sustainability
8(7):629
3. Falcão DS, Pinto AMFR (2020) A review on PEM electrolyzer modelling: guidelines for
beginners. J Clean Prod 261:1–38
4. Yigit T, Selamet OF (2016) Mathematical modeling and dynamic Simulink simulation of
high-pressure PEM electrolyzer system. Int J Hydrogen Energy 41(32):13901–13914
5. Lajnef T, Abid S, Ammous A (2013) Modeling, control, and simulation of a solar hydrogen/fuel
cell hybrid energy system for grid connected applications. Adv Power Electron 1–9
6. Dursun E, Acarkan B, Kilic O (2012) Modeling of hydrogen production with a stand-alone
renewable hybrid power system. Int J Hydrogen Energy 37(4):3098–3107
7. Ma Z, Witteman L, Wrubel JA, Bender G (2021) A comprehensive modeling method for proton
exchange membrane electrolyzer development. Int J Hydrogen Energy 46(34):17627–17643
8. Zhou T, Francois B (2009) Modeling and control design of hydrogen production process for
an active hydrogen/wind hybrid power system. Int J Hydrogen Energy 34(1):21–30
9. Liso V, Savoia G, Araya SS, Cinti G, Kær SK (2018) Modelling and experimental analysis
of a polymer electrolyte membrane water electrolysis cell at different operating temperatures.
Energies 11(12):1–19
10. Giordano G, Sjöberg J (2018) Black and white-box approaches for cascaded tanks benchmark
system identification. Mech Syst Signal Process 108:387–397
11. Rabbani MJ, Hussain K, Khan AUR, Ali A (2013) Model identification and validation for a
heating system using MATLAB system identification toolbox. IOP Conf Ser Mater Sci Eng
51(1):1–10
12. Chavan SL, Talange DB (2018) System identification black box approach for modeling
performance of PEM fuel cell. J Energy Storage 18:327–332
13. Aruna R, Jaya Christa ST (2020) Modeling, system identification and design of fuzzy PID
controller for discharge dynamics of metal hydride hydrogen storage bed. Int J Hydrogen
Energy 45(7):4703–4719
14. Sankar K, Jana AK (2021) Nonlinear control of a PEM fuel cell integrated system with water
electrolyzer. Chem Eng Res Des 171:150–167
15. Koundi M, El Fadil H (2019) Mathematical modeling of PEM electrolyzer and design of
a voltage controller by the SMPWM approach. In: 5th international conference on power
generation systems and renewable energy technologies. IEEE, Turkey, pp 1–6
16. Ogata K (2002) Modern control engineering, 5th edn. Prentice-Hall, New Jersey
17. Stephanopoulos G (1984) Chemical process control: an introduction to theory and practice, 1st
edn. Prentice-Hall, New Jersey
Sliding Mode Hybrid Control of PMSM
for Electric Vehicle
Ajay Pawar and S. V. Jadhav
Abstract Permanent magnet synchronous motor (PMSM) is becoming a favorable

motor for electric vehicle. This is because of its well-known features like better
torque to weight ratio, larger power factor, high efficiency, and reliability as against its
counterpart induction motor (IM). The paradigm of control of drives has been shifted
because of vector control (VC) and field-oriented control (FOC). Speed control using
proportional-and-integral (PI) controllers and many nonlinear controllers is popular
in the heavy performance application of PMSM. Applications like robotics, CNC, and
electric vehicle demand for dynamic control over a large operation range including
four-quadrant operations. This work envisages control of PMSM drive for electric
vehicles. Sliding mode control (SMC) algorithm is used in order to achieve higher
energy efficiency, in constant power and constant torque regions. Further, higher
order sliding mode control (HOSMC) algorithm is also tested for robustness against
disturbances and parameter variations, nonlinearity load patterns, and, importantly,
EV. HOSMC not only reduces the ripples caused by SMC but also offer dynamic
speed and torque response with optimized control input. Many simulations are carried
on MATLAB Simulink platform to prove the effectiveness of SMC and HOSMC.
Keywords Permanent magnet synchronous motor (PMSM) · Field-oriented

control (FOC) · Field weakening control (FWC) · Sliding mode control (SMC) ·
Higher order sliding mode control (HOSMC) · Integral time absolute error
(ITAE) · Four-quadrant operation · Simulation and result
A. Pawar (B) · S. V. Jadhav

Department of Electrical Engineering, College of Engineering, Pune, India
e-mail: pawarab19.elec@coep.ac.in
S. V. Jadhav
e-mail: svj.elec@coep.ac.in
https://doi.org/10.1007/978-981-19-6913-3_11
166 A. Pawar and S. V. Jadhav
1 Introduction
Advances in the control of electric drives have led to the replacement of the IC
engines by gear-less drive systems where electric motors drive the vehicles. Robust
and dynamic control in constant torque and constant power regions replicate the
lower and higher gears of a mechanical engine. Advanced control algorithm like
sliding mode control (SMC) or intelligent control has made the control operation
fast and accurate for a wide operating speed range [1]. PMSM is popularly used
in electric vehicles. On account of reduce copper losses and use of high energy
permanent magnets, PMSM possesses advantages like high efficiency, high torque
to weight ratio, etc. [2]. In past two decades, vector control (VC) and field-oriented
control (FOC) have dominated the drive applications [3, 4] due to the well-known
feature of decoupled speed and torque control. Thanks to the developments in the high
frequency power switches and DSP processors; because of which modern complex
control algorithms can be applied in real time. Various intelligent controllers and
variable structure control (VSC) algorithms are now preferred over simple PI control
algorithms [5]. Several algorithms like SMC, fuzzy control, artificial neural network
(ANN)-based control, and fuzzy-SMC control are comparable and offer satisfactory
drive performance [6]. The researchers have explored combination of SMC and PI
control to achieve required performance in steady-state and transient conditions [7].
In order to reduce the ripples in the response, higher order SMC (HOSMC) [8] is
proposed and is applied in drives applications [9–11]. Speed control of PMSM is
generally has cascaded control loop structure. It consists of an outer speed control
loop and two inner current loops. This gives good disturbance rejection and better
steady-state performance. Torque and speed decoupling is achieved by two inner
current loops out of which, magnetizing current is regulated to zero to obtain field-
oriented control (FOC) [12] improving the torque per ampere ratio. The EV is oper-
ated in constant torque region for normal speeds. High speed operation is obtained
using field weakening control (FWC) in constant power mode [13]. Thus, wide
speed range along with four-quadrant operation is obtained for EV. In this paper,
SMC and HOSMC, both algorithms are attempted separately and the performances
are compared. For better settling time and less steady-state error, super twisting algo-
rithm (STA) is employed [14]. For the EV application purpose TATA nano car, which
is currently IC engine based, is considered because of its low weight, good volume
space, and low price. It is tested for its conversion to EV. The paper is arranged as
follows: Modeling of PMSM is illustrated in Sect. 2. FOC and FWC as applied to
PMSM is explained in Sect. 3. Design of cascade SMC hybrid control is given in
Sect. 4. Simulation and results are shown in Sect. 5, and Sect. 6 concludes the paper.
Sliding Mode Hybrid Control of PMSM for Electric Vehicle 167
2 PMSM Model
2.1 Modeling of PMSM
Three-phase voltage model of stator of PMSM is given by following equations [15].
di a
Va (t) = i a R + L s + E a (t) (1)
dt
di b
Vb (t) = i b R + L s + E b (t) (2)
dt
di c
Vc (t) = i c R + L s + E c (t) (3)
dt
2.2 d-q Model of PMSM
The model described by Eqs. (1), (2), (3) is converted into time invariant model by
applying Clarke and Parke transformations. This results in a d-q-axis model in the
synchronously rotating stator flux reference frame.
di ds
Vds = Ri ds + L − ωe i qs L (4)
dt
di qs
Vqs = Ri qs + L + ωe i ds L + ωe Φ (5)
dt
Rearranging the equations
di ds 1
= Vds − Ri ds + ωe Li qs (6)
dt L
di qs 1
= Vq − Ri qs + ωe Li ds − ωe Φ (7)
dt L
The expression of electromagnetic torque in terms of inductance, direct axis, and
quadrature axis current is
3
Te = P ϕi qs + L d − L q i ds i qs (8)
2
In surface mounted PMSM, L d = L q and hence electromagnetic torque becomes
3
Te = P ϕi qs (9)
2
The expression for electromagnetic torque in terms of mechanical torque is
dωm
Te = TL + Bωm + J (10)
dt
Here, ωm = P
ω.
2 e
3 Field-Oriented and Flux Weakening Control
From Eq. (9), electromagnetic torque is represented as
3
Te = P ϕi qs (11)
2
dωe 1
= Te (12)
dt J
FOC gives better result below base speed when we require speed above the base
speed. Field weakening control is come into the picture. In FOC, we make ids current
equal to zero but for FWC we make ids current negative so that we can achieve speed
more than base speed similar to DC motor. For that purpose, we select i dref as
/ 2
∅ 1 Udc
i dref =− + − Li qs (13)
L L w
4 Design of Cascade SMC Hybrid Control
This is the cascade sliding mode hybrid control with current control as inner loop
and voltage control as an outer loop, with current we can control torque and with
voltage we can control speed. A switch is used for hybrid control as shown and gives
efficient control scheme (FOC and FWC) depending on the reference speed (Fig. 1).
As there are two separate current control loops (d-axis and q-axis currents), and
one outer speed loop, and three SMC controllers are designed. The surfaces are
designed on the basis of error in current and speed as below
∗

Sd = C T × ed = C T i ds − i ds =0 (14)
Fig. 1 Block diagram of SMC hybrid control
∗
Sq = C T × eq = C T i qs − i qs =0 (15)

Sw = C T × ew = C T ωe − ωe∗ = 0 (16)
4.1 Direct Axis Sliding Mode Control
We have
di ds −R P 1
= i ds + ωe i qs + u d
dt L 2 L
where u d = flux control input.
∗
Now sliding surface, sd = i ds − i ds .
By taking first derivative of sd , we get
di ds di ∗
ṡd = − ds
dt dt
R p 1 di ∗
Ṡd = − i ds + ωe i qs + u d − ds (17)
L 2 L dt
The presence of control input in Ṡd equation ensures that we choose right sliding
surface.
Here, u d = u d1 + u d2
where u d1 = equivalent control
u d2 = switching control.
So, u d1 is chosen as
∗
di ds p
u d1 = Ri ds + L − ωe Li qs (18)
dt 2
Now, u d2 can be defined with upper bound of uncertainties, i.e.,
| |
| di ∗ds p |
u d2 |
≥ |Δ Ri ds + Δ L − ωe Δ Li qs || (19)
dt 2
switching component with Lyapunov stability approach is
u d2 = −k sgn(sd ) (20)
Then, total flux control input becomes
p di ∗
u d = Ri ds + ωe Li qs + L ds + −k sgn(sd ) (21)
2 dt
4.2 Quadrature Axis Sliding Mode Control

di qs −R p p φ 1
= i qs − ωe Li ds − ωe + u q
dt L 2 2 L L
where V q = uq = control input.
Sliding surface
Sq = i qs − i qs ∗
Taking derivative
di qs di ∗qs
sq· = −
dt dt
−R p p φ 1 di ∗qs
= i qs − ωe Li ds − ωe + u q − (22)
L 2 2 L L dt
Then, u q1 is chosen as
∗
p p di qs
u q1 = Ri qs + ωe Li ds + ωe φ + L (23)
2 2 dt

u q2 = −k sgn sq (24)
4.3 Speed Sliding Mode Control
Similarly, here, control input is
1 dω∗
i qs1 = τm + Bωe + I e (25)
k dt
where k = 43 Pφ
i qs2 = −k sgn(sw ) (26)
4.4 Super Twisting Algorithm (STA)
Chattering phenomenon can be reduced by modifying the switching control with the
help of super twisting algorithm (STA) we choose uq as

u s = k1 |S|1/2 sgn(S) − k2 sgn(S)dt (27)
Replace control input according to selected sliding surface.
5 Simulations and Result
The simulations are carried out on MATLAB Simulink platform. The rating of PMSM
are 1.3 kW, 300 V, and 4.5 A. The parameters for simulation of PMSM motor is as
shown in Table 1.
Figure 2 shows transient no load speed response with PI, SMC, and HOSMC
(hybrid) controllers. Also, Fig. 3 shows steady-state responses. Settling time for PI
Table 1 Parameters of
Parameter Value
PMSM
R 1.3 Ω
Ld = Lq 0.0069 H
φ 0.157 Wb
Pole pair 6
Inertia 0.0027 kg m2
B 0.0001 Nm S
controller is 16 ms and that for SMC and HOSMC is 10 and 6 ms. No load steady-state
errors are 0.4 rpm, 0.002 rpm, and 0.0001 rpm, respectively.
Speed response for load of 5 Nm is applied at 0.025 s, and transients are shown
in Fig. 4.
Figure 4 shows the speed response, and Fig. 5 shows stator current with load
(5 Nm) at 0.025 s.
Speed response with change in resistance by 150% is shown in Fig. 6, and from
the figure, it is observed that SMC and HOSMC transient behavior is unaffected by
resistance change, but in case of PI control, settling time is increased from 15 to
19 ms also when load is applied, then transient behavior is changed in case of PI.
Speed response when disturbance of 0.25 sin(1000t) is inserted in the system and
Fig. 7 shows that HOSMC and SMC are unaffected with disturbance but there is
change in speed response with PI controller.
Fig. 2 No load speed transients with PI, SMC, and HOSMC
Fig. 3 Speed at no load (steady-state error)

Fig. 4 Speed response for step load (5 Nm) at 0.025 s
Fig. 5 Stator current for condition in Fig. 4
Fig. 6 Speed with 150% change in resistance and load (5 Nm) at 0.025 s
Fig. 7 Speed response with disturbance (0.25 sin(1000t))
Figure 8 shows speed response when nonlinear load applied, and Fig. 9 shows the
nonlinear load torque.
ITAE error with different values of inertia shown in Fig. 10.
Four-quadrant operation of PMSM is shown in Figs. 11 and 12.
Quadrant Mode of operation Time Speed Torque

First quadrant Forward motoring 0–0.025 525 +5
Fourth quadrant Forward regeneration 0.025–0.05 525 −5
Third quadrant Reverse motoring 0.05–0.075 −525 −5
Second quadrant Reverse regeneration 0.075–0.1 −525 +5
Constant torque and constant power mode of operation with PI, SMC, and
HOSMC are compared as in Figs. 13, 14, 15 and 16.
Fig. 8 Speed response when nonlinear load applied

Fig. 9 Nonlinear load torque with HOSMC
Fig. 10 ITAE for the normal operating conditions with step load applied at 0.025 s with J =
0.027 kg m2
5.1 Application on Electrical Vehicle
For an electric vehicle application, we have considered TATA nano vehicle because of
its low weight, less price, and sufficient volume space. As it has low weight requires
low power to drive the vehicle. For this purpose, we have considered 300 Vdc,
4500 rpm surface mounted PMSM motor (Tables 2, 3 and 4).
Figure 17 shows the drive cycle source (FTP75) for 50 s with HOSMC (Figs. 18,
19; Tables 5, 6, 7).
Fig. 11 Speed in all quadrant
Fig. 12 Torque in all quadrant with HOSMC
6 Conclusion
This paper deals with the cascade SMC hybrid control for PMSM. Four-quadrant
operation is achieved with this control strategy, and results are satisfactory. SMC
deals with parameter uncertainty (e.g., change in stator resistance) and external
disturbance. Smooth ripple-free operation of drive (TATA nano car) is a result of
HOSMC. It is observed that HOSMC offers better performance with respect to SMC
and PI which is proved by the values of ITAE, peak overshoot, and settling time.
HOSMC is the most suitable algorithm for EV applications.
Fig. 13 Speed in constant torque and constant power mode
Fig. 14 Direct axis current with PI

Fig. 15 Direct axis current with SMC
Fig. 16 Direct axis current with HOSMC (hybrid)
Table 2 PMSM parameters

Motor power (P) = 3 kW
Stator resistance (Rs) = 0.62
Stator inductance (Ls = Ld = Lq) = 0.002075
Rotor flux linkage (Øm) = 0.08627
Pole pair (P) = 4
Viscous friction coefficient (B) = 0.00009444
Wheel inertia (J) = 0.1
Table 3 TATA nano

Mass of TATA nano = 630 kg
specifications
Wheel radius of TATA nano = 0.3 m
Area of TATA nano = 2.5 m2
Table 4 Gear ratios

Gear ratio (R1) = 3.45
Fig. 17 Drive cycle source
Fig. 18 Stator current

Fig. 19 Stator current (steady-state error)
Table 5 Comparison of PI, SMC, and HOSMC is carried out, on the basis of different performance
parameters
Property Controller
PI SMC Hybrid-HOSMC
J= J= J= J= J= J=
0.027 kg m2 0.03 kg m2 0.027 kg m2 0.03 kg m2 0.027 kg m2 0.03 kg m2
Normal 0.011 0.012 0.0025 0.0028 0.0022 0.0023
operating
condition
Normal 0.0115 0.013 0.0027 0.003 0.0024 0.0024
operating
conditions
with step
load
applied at
0.5 s
Change in 0.0117 0.0134 0.0027 0.003 0.0024 0.0024
stator
resistance
Table 6 THD analysis

Controller (J = 0.027 kg m2 ) PI SMC Hybrid-HOSMC
Speed overshoot (%) 0.1 0.05 0.0001
Steady-state error in speed 0.09 0.002 0.0001
THD 0.776 0.7751 0.7747
Table 7 Comparative analysis

Property Controller
PI SMC Hybrid-HSMC
Starting transient performance Good Very good Excellent
Robustness Good Very good Excellent
Steady-state performance Good Very good Excellent
Computational effort Medium Low High
References
1. Maini C, Gopal K, Prakash R (2013) Making of an ‘all reason’ electric vehicle. In: 2013 world
electric vehicle symposium and exhibition (EVS27), pp 1–4. https://doi.org/10.1109/EVS.2013.
6915015
2. Loganayaki A, Kumar RB (2019) Permanent magnet synchronous motor for electric vehicle
applications. In: 2019 5th international conference on advanced computing & communication
systems (ICACCS), pp 1064–1069. https://doi.org/10.1109/ICACCS.2019.8728442
3. Mishra A, Makwana JA, Agarwal P, Srivastava SP (2012) Modeling and implementation of
vector control for PM synchronous motor drive. In: IEEE-international conference on advances
in engineering, science and management (ICAESM-2012), pp 582–585
4. Pindoriya RM, Mishra AK, Rajpurohit BS, Kumar R (2016) Performance analysis of control
strategies of permanent magnet synchronous motor. In: 2016 IEEE region 10 conference
(TENCON), pp 3224–3227. https://doi.org/10.1109/TENCON.2016.7848645
5. Jadhav SV, Srikanth J, Chaudhari BN (2010) Intelligent controllers applied to SVM-DTC based
induction motor drives: a comparative study. In: 2010 joint international conference on power
electronics, drives and energy systems & 2010 power India, pp 1–8. https://doi.org/10.1109/
PEDES.2010.5712400
6. Jadhav SV, Kirankumar J, Chaudhari BN (2012) ANN based intelligent control of induction
motor drive with space vector modulated DTC. In: 2012 IEEE international conference on power
electronics, drives and energy systems (PEDES), pp 1–6. https://doi.org/10.1109/PEDES.2012.
6484350
7. Levant A (2003) Higher-order sliding modes, differentiation and output-feedback control. Int J
Control 76(9–10):924–941. https://doi.org/10.1080/0020717031000099029
8. Laghrouche S, Harmouche M, Ahmed FS, Chitour Y (2015) Control of PEMFC air-feed system
using Lyapunov-based robust and adaptive higher order sliding mode control. IEEE Trans
Control Syst Technol 23(4):1594–1601. https://doi.org/10.1109/TCST.2014.2371826
9. Feng Y, Zheng J, Yu X, Truong NV. Hybrid terminal sliding-mode observer design method for
a permanent-magnet synchronous motor control system. IEEE Trans Ind Electron
Maximum Sensitivity-Based PID
Controller for a Lab-Scale Batch Reactor
M. Bala Abhirami and I. Thirunavukkarasu
Abstract In this article, a PID tuning rule is presented for a lab-scale batch reactor
model in order to minimize the impact of load disturbance on the process. The tuning
technique is solely determined from the maximum sensitivity (M S ) criteria, which
are dependent on the full-state feedback and also on the matching of frequency
response. The only tuning parameter for the controller is the maximum sensitivity
(M S ) which indicates the level of robustness. Hence, by adjusting the value of (M S ),
the values of PID tuning coefficients can be determined precisely. The suggested
method is suitable for integrating processes with dead time and also for other complex
systems by narrowing it down to FOPDT process. This article provides a controller
for a lab-equipped batch reactor using the proposed method for tuning the controller
parameters.
Keywords Maximum sensitivity (M S ) · PID controller · Load disturbance rejection
1 Introduction
Most of the industrial control system applications are based on conventional PID
controllers owing to its simplicity, cost-effectiveness, and easy maintenance. It must
be noted that about 90% of the industrial applications still choose PID controllers
over other modern control methods. Various approaches are employed to tune of the
parameters of a PID controller [1–5], some of them which include Ziegler–Nichols
method, internal mode control (IMC), pole placement method, etc.
The most popular method for tuning PID control parameter is the Ziegler–Nichols
method. This method is used widely in most of the industries due to its simplicity.
Meshram et al. [6] designed a controller to control the speed of a DC motor. They used
M. Bala Abhirami (B) · I. Thirunavukkarasu

Department of Instrumentation and Control Engineering, Manipal Institute of Technology,
Manipal Academy of Higher Education, Manipal, Karnataka, India
e-mail: mabhirami2206@gmail.com
I. Thirunavukkarasu
e-mail: it.arasu@manipal.edu
https://doi.org/10.1007/978-981-19-6913-3_12
184 M. Bala Abhirami and I. Thirunavukkarasu
Ziegler–Nichols methods to tune the parameters of the PID controller. Li et al. [7]
used Ziegler–Nichols method to reduce and validate PID parameters using MATLAB
for a model as an example. Ziegler–Nichols method is known for its simplicity, but the
tuning the parameters might be difficult in the presence of noise. In case of complex
systems, the process output might give oscillatory response with large overshoots.
Integral model control (IMC) design of PID is majorly known for achieving robust-
ness in the system design [8]. It is also employed for the tuning of model with delay.
It is also used for stable and integrating process models [9, 10]. In spite of their
advantages, IMC control design has one major setback. It sets the integral time of
the controller equal to that of the process time constant. Due to this, the controller
might have very long integral time which leads to slower recovery from the load
disturbances.
Pole placement method is used to design linear control systems. This method
makes use of the dominant poles of the system, and the closed loop poles are placed on
the desired location such that all the poles remain to the far left of the poles assigned.
Primary setback of this method is that it does not place the zeros at desired locations,
and this method requires more skill and knowledge to find the optimum locations
to place the poles. Examples of this method of controller design are presented in
[11, 12].
In this article, a simple PID controller is developed based upon the maximum
sensitivity (M S ) in order to improve the effectiveness of load disturbance rejection
[13]. The controller is to be designed for a lab-equipped batch reactor where the
system model is approximated to FOPDT model. Simulations have been carried out
for the FOPDT model using the proposed tuning method.
The article is structured accordingly: Sect. 2 demonstrates the design method of
the proposed method. Section 3 presents a controller design for a lab-scale batch
reactor on the basis of the proposed method to tune the controller gains, and Sect. 4
summarizes the paper with final remarks.
2 Design Method
The expression for FOPDT process can be written as:
1.4493 −0.1093s
G(s) = e (1)
5.832s + 1
where K, T, and L represent the steady-state gain, time constant, and time delay of
the FOPDT system.
The expression for a PID controller is can be written as:

KI 1
G(s) = K C + + KDs = KC 1 + + Td s (2)
s Ti s
Maximum Sensitivity-Based PID Controller for a Lab-Scale Batch Reactor 185
Table 1 M S -based PID tuning rule for a general FOPDT process [13]
Controller parameter Constants
ac = −0.3751MS2 + 1.875MS − 1.439

K K C = ac (L/T )bc + cc bc = −0.01233MS − 0.9926
cc = −0.1847MS2 + 0.9121MS − 0.7085
ai = −0.72721MS−3.083 + 1.587
bi = −0.5518MS−7.687 + 0.1663
Ti /T = ai ebi (L/T ) + ci edi (L/T )
ci = −0.01998MS3.535 − 1.073
di = −51.19MS−8.921 − 1.581
−0.0638MS2 + 0.659MS − 0.6299

ad =
MS − 1.143
bd = −1.359MS−7.717 + 0.1459
Td /T = ad ebd (L/T ) + cd edd (L/T ) −0.4453MS + 0.4476
cd =
MS − 1.129
−4.573
dd = 2
MS − 3.7MS + 10.41
where K C , K I , and K D are the proportional, integral, and derivative gains of the
controller. K I and K D are defined as K C /T i and K C T d where T i and T d denote the
integral time and derivative time constants, respectively.
Maximum Sensitivity M S is given by,

1

MS = max (3)
0≤ω≤∞ 1 + G( jω)C( jω)
The robustness of the system is measured by the value of M S and smaller value
of M S indicates larger stability margin of the control system. The value of M S fixed
in the range of 1.2–2.0 [13].
The PID control parameters can be determined directly by choosing the most
appropriate value for M S and by calculating the constants KK C , T i /T, T d /T using the
equations provided in Table 1.
3 Controller Design for Lab-Scale Batch Reactor
Batch reactor is highly nonlinear system with temperature profile as a set point instead
of constant set point. Batch reactors are mostly used in chemical process industries
having wide application in paint and pharma industries.
The objective is to maintain the reactor temperature in-line with the temp. profile.
The challenge is the reagents A and B reacts in exothermic way results with constant
rise in reactor temperature. The reactor temperature will be measured by the RTD.
The reactor will have coolant jacket with continues coolant flow with control valve.
The outlet of jacket coolant water will also be measured to control the valve action.
Another RTD is used to measure the coolant jacket outlet temperature. The batch
reactor setup is displayed in Figs. 1 and 2 shows the schematic diagram of the process.
The transfer function of a batch reactor setup is expressed by FOPDT model as:
1.4493 −0.1093s
G(s) = e (4)
5.832s + 1
where is the steady-state gain K = 1.4493, time constant T = 5.832, time delay L =
0.1093.
The controller parameters are obtained for several other FOPDT processes with
different L/T ratios. Figure 3 shows the controller gains of the batch reactor along
with several other FOPDT processes that are considered. The scaled controller gains
KK C are plotted against the system parameter L/T. The PID parameters are computed
for different M S values as shown in the figure. Table 2 shows the different values
of M S that are considered to calculate the constants ac , bc , cc . Figure 4 shows the
constants ac , bc , cc plotted against the M S values.
Figure 5 shows the scaled integral time constants T i /T plotted against the system
parameter L/T for different values of M S . Table 3 gives the values of ai , bi , ci , d i
for different M S values. Figure 6 shows the constants ai , bi , ci , d i plotted against
different M S values.
Figure 7 shows the scaled derivative time constants T d /T plotted against the system
parameter L/T for different values of M S . The values of ad , bd , cd , d d for the different
Fig. 1 Batch reactor setup

Fig. 2 Schematic diagram of a batch reactor
20
Ms=1.2
Ms=1.3
18
Ms=1.4
16 Ms=1.5
Ms=1.6
14 Ms=1.7
Ms=1.8
Gain KKc
12 Ms=1.9
Ms=2.0
10
0.1 0.15 0.2 0.25 0.3 0.35

L/T
Fig. 3 Plots of KK C versus system parameter L/T
M S values that are considered are given in Table 4. The constant values of ad , bd ,
cd , d d , plotted against M S values are shown in Fig. 8. As seen from the plots, the
parameters KK C , T i /T, and T d /T are dependent on M S . Hence, by assigning a suitable
value for M S , the PID control parameters are calculated as per the equations provided
in Table 1.
The PID values for M S = 1.2 are calculated from above-mentioned parameters as
K p = 9.90, K i = 01.633, and K d = −0.00639, whereas the PID values for lambda
method are K p = 0.383, K i = 0.285, and K d = 0.87.
Table 2 Constants values for

MS ac bc cc
the curves in Fig. 5
1.2 0.3418 −1.0074 0.1201
1.3 0.4489 −1.0086 0.1651
1.4 0.5497 −1.0099 0.2064
1.5 0.6441 −1.0111 0.2441
1.6 0.7322 −1.0123 0.2780
1.7 0.8139 −1.0136 0.3083
1.8 0.8893 −1.0148 0.3349
1.9 0.9583 −1.0160 0.3577
2.0 1.0210 −1.0173 0.3769
1.5
0.5
ac,bc,cc
ac vs Ms
0 bc vs Ms
cc vs Ms
-0.5
-1
-1.5
1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Ms
Fig. 4 Constants ac , bc , cc versus maximum sensitivity M S
4 Simulation and Results
The open-loop response of the lab-scale batch reactor is provided in Fig. 9. The
controller for the batch reactor is designed using the suggested method, and the
results are compared with the PID controller implemented using the lambda tuning
approach as proposed in [14]. The most suitable value of M S is 1.2 for the suggested
controller design.
Figures 10 and 11 show the real-time implementation of M S -based PID and
lambda PID on a batch reactor setup for trajectory tracking and respective control
signal. It is observed with oscillation on the reactor temperature as well on the manip-
ulated variable. In general, PID control performance will not be good for trajectory
tracking due to the absence of hard constraints engagement. This can be overcome
0.9
0.8
Ti/T
Ms=1.2
0.7 Ms=1.3
Ms=1.4
Ms=1.5
Ms=1.6
0.6
Ms=1.7
Ms=1.8
Ms=1.9
0.5 Ms=2.0
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45

L/T
Fig. 5 Plots of T t /T versus system parameter L/T
Table 3 Constants values for the curves in Fig. 6

MS ai bi ci di
1.2 1.1726 0.0304 −1.111 −11.646
1.3 1.2630 0.0929 −1.1235 −6.5093
1.4 1.3294 0.1248 −1.1386 −4.1254
1.5 1.3787 0.1419 −1.1568 −2.9559
1.6 1.4163 0.1514 −1.1782 −2.3540
1.7 1.4498 0.1570 −1.2034 −2.0311
1.8 1.4683 0.1603 −1.2326 −1.8513
1.9 1.4865 0.1623 −1.2662 −1.7479
2.0 1.5012 0.1636 −1.3046 −1.6866
with prediction-based controller, especially with MPC and NMPC categories with
nonlinear Kalman filters.
5 Conclusion
This article provides a simple method of PID controller design using maximum
sensitivity (M S ) which relies on the full-state feedback and frequency response model
matching. This technique is employed on FOPDT process models but is also suitable
for other complex models after converting it to FOPDT model. A controller design
is implemented on a lab-equipped batch reactor using MATLAB2021® platform.
-2
ai vs Ms
-4 bi vs Ms
ai,bi,ci,di
ci vs Ms
di vs Ms
-6
-8
-10
-12
1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Ms
Fig. 6 Constants ai , bi , ci , d i versus maximum sensitivity M S
0.26
0.24
0.22
Td/T
Ms=1.2
0.2 Ms=1.3
Ms=1.4
Ms=1.5
0.18 Ms=1.6
Ms=1.7
Ms=1.8
Ms=1.9
0.16 Ms=2.0
0.45 0.5 0.55 0.6 0.65

L/T
Fig. 7 Plots of T t /T versus system parameter L/T
The output response is compared with the controller designed using lambda tuning
approach. The real-time results show the tight tracking with M S -based PID than
lambda PID method. Oscillations can be reduced or nullified with the predictive-
based nonlinear controllers.
Table 4 Constants values for the curves in Fig. 7

MS ad bd cd dd
1.2 1.2110 −0.1869 −1.222 −0.617
1.3 0.7578 −0.0335 −0.7678 −0.627
1.4 0.6523 0.0446 −0.6488 −0.636
1.5 0.6024 0.0864 −0.5939 −0.643
1.6 0.5715 0.1098 −0.5624 −0.648
1.7 0.5494 0.1233 −0.5419 −0.652
1.8 0.5321 0.1313 −0.5275 −0.654
1.9 0.5577 0.1363 −0.5168 −0.654
2.0 0.5051 0.1394 −0.5086 −0.652
1.5
ad vs Ms
bd vs Ms
1 cd vs Ms
dd vs Ms
0.5
ad,bd,cd,dd
-0.5
-1
-1.5
1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
Ms
Fig. 8 Constants ad , bd , cd , d d versus maximum sensitivity M S

1.5
Time Series Plot:
1
Process Output
0.5
0
0 10 20 30 40 50 60 70 80 90 100
Time (seconds)
Fig. 9 Open loop response of temperature control in a batch reactor
Fig. 10 Implementation of M S -based PID (pink) and lambda PID (blue) controller on a pilot plant
batch reactor for trajectory tracking
Fig. 11 Manipulated variable (coolant control valve) for M S -based PID control
References
1. Hagglund T (2019) The one-third rule for PI controller tuning. Comput Chem Eng J
2. Astrom KJ, Hagglund T (2006) Advanced PID control. ISA-The Instrumentation, Systems and
Automation Society, Research Triangle Park, NC
3. Cvejn J (2011) Simple PI/PID controller tuning rules for FOPDT plants with guaranteed closed-
loop stability margin. Acta Montan Slov 16(1):17–25
4. Hayashi K, Yamamoto T (2011) Design of a PID tuner based on generalized-output errors. In:
SICE annual conference, Tokyo, pp 2614–2618
5. Kolaj W, Mozaryn J, Syfert M (2016) PLC-PIDTuner: application for PID tuning with SIMATIC
S7 PLC controllers. In: 21st international conference on methods and models in automation
and robotics (MMAR), Miedzyzdroje, pp 306–311
6. Astrom KJ, Hagglund T (1995) PID controllers theory design and tuning. Instrument Society
of America, Research Triangle Park, NC
7. Seborge DE, Edgar TF, Mellichamp DA (1989) Process dynamics and control. Wiley, New
York
8. Vanavil B, Krishna Chaitanya K, Sheshagiri Rao A (2013) Improved PID controller design for
unstable time delay processes based on direct synthesis method and maximum sensitivity. Int
J Syst Sci 1–18
9. Yang X, Xu B, Chiu MS (2011) PID controller design directly from plant data. Ind Eng Chem
Res 50:1352–1359
10. Shamsuzzoha M (2013) Closed loop PI and PID controller tuning for stable and integrating
process with time delay. Ind Eng Chem Res 52:12973–12992
11. Meshram PM, Kanojiya RG (2012) Tuning of PID controllers using Ziegler-Nichols method
for speed control of DC motor. In: IEEE-international conference on advances in engineering,
science and management (ICAESM), Nagapattinam, India
12. Li S, Jiang Q (2011) Study on PID parameters tuning based on Matlab/Simulink. In: IEEE 3rd
international conference on communication software and networks, Xi’an, China
13. Raza A, Pathak N, Anwar MN (2017) A PID controller tuning rule for FOPDT process
to achieve better load disturbance rejection based on maximum sensitivity. In: ICSPACE:
international conference on smart grids, power and advanced control engineering, pp 149–154
14. Pruna E, Sasig ER, Mullo S (2017) PI and PID controller tuning tool based on the lambda
method. IEEE
Performance Prediction of Solar Cell
Using Virtual Production Simulation
B. Ashok Kumar, T. S. Bagavat Perumaal, S. Senthilrani,

and Parthasarathy Seshadri
Abstract Recent advancements in renewable energy harvesting system have trig-

gered the photovoltaic research in different dimensions. It has enabled rapid devel-
opment of systems for renewable energy production. To yield an effective production
of solar energy, it has become mandatory to predict the solar cell performance virtu-
ally. Such prediction will lead to performance analysis before the commencement of
solar cell production. Virtual performance analysis of solar cell will support better
design, analysis and control of photovoltaic systems. By an appropriate choice of
software tool, performance prediction of solar cells will minimize time and cost
by virtual production analysis. The work presented in this paper provides successful
production rate of solar cell. For analysis, 10,000 samples have been taken into virtual
performance analysis, where 9337 successful samples have been yielded. Thus, elec-
trical yield of 92% have been obtained using photo luminescent (PL) imaging. It is
observe that analysis without photo luminescent (PL) imaging have electrical yield
of 86.2%. The defect rate of solar cells has been minimized through virtual perfor-
mance prediction. Around 5.8% of increase in electrical yield has been observed
through PL imaging.
Keywords Renewable energy harvesting system · Photovoltaic systems · Virtual

performance analysis · Electrical yield · Photo luminescent imaging
B. Ashok Kumar (B) · T. S. Bagavat Perumaal · P. Seshadri

Department of Electrical and Electronics Engineering, Thiagarajar College of Engineering,
Madurai, India
e-mail: ashokudt@tce.edu
T. S. Bagavat Perumaal
e-mail: bagavatperumaal123@gmail.com
P. Seshadri
e-mail: parthasarathy396@gmail.com
S. Senthilrani
Department of Electrical and Electronics Engineering, Velammal College of Engineering and
Technology, Madurai, India
e-mail: ssr@vcet.ac.in
https://doi.org/10.1007/978-981-19-6913-3_13
196 B. Ashok Kumar et al.
1 Introduction
India is one of the thickly populated countries in the globe. The technological
advancement and usage of more electronics and electrical gadgets by the people
in such densely populated region has created power scarcity [1]. One such problem
is satisfying consumer’s energy demand. Over the last few decades, the energy gener-
ation and energy demand couldn’t be met out effectively. The increase in demand
can be satisfied with fossil fuel-based energy generation alone [2]. The depleting
fossil fuel cannot support the increasing energy demand of the country in the next
two decades [3]. Hence, there arises the need for alternative sources of energy. The
shift toward alternative sources of energy has recently dominated by wind and solar
energy. Thus, energy demand can be met easily with zero emission system (ZES)
[4, 5].
Among the two types renewable energy-based systems, more preference is given
to solar based because in country like India sunshine is abundant throughout the year
[6, 7]. When compared to wind energy-based generation system, the installation and
maintenance expenditure of solar energy-based generation system is cost effective. To
promote clean energy system, Government offers subsidy to residential and industrial
consumers. Indian government has set a target to achieve 100 GW by 2022 and this
can be termed as solar revolution [8, 9].
Initially, the Indian government set its target as 20 GW capacity for 2022, which
was revised as 100 GW of solar capacity by 2022 during the year 2015 [10]. To
promote solar plant installation, India has established nearly 42 solar parks. Around
2.1 GW rooftop solar power generations satisfies 70% of commercial demand. In
India, apart from large-scale grid-connected solar photovoltaic initiative, to satisfy
local energy demand off-grid solar power systems are developed [8]. The rural energy
needs are met out through solar lanterns and thus the utilization of kerosene has been
drastically reduced. Similarly, under National program, 118,700 solar home lighting
systems were installed and 46,655 solar street lighting installations were executed
[11, 12].
2 Solar Cell Manufacturing
Thin crystalline silicon is the raw material for solar cell fabrication, where the conduc-
tivity of cell is decided by intentional doping level of boron into silicon. Higher
the level of boron increases the conductivity of solar cell. The doping process is a
controlled process in the solar cell fabrication [13]. The most common method of
silicon crystallization techniques are Czochralski (CZ) method and casting method
[14].
Single-crystal ingot with high level of purity is obtained through CZ method,
while in casting method multicrystalline silicon brick is obtained. Though casting
method provides silicon with low cost, it comprises more defects and impurities.
Performance Prediction of Solar Cell Using Virtual Production Simulation 197
Hence, the electrical characterization will deviate in casting method. In both cases,
silicon ingot is grown through controlled temperature chamber. The boron occupies
crystal lattice when it is solidified.
The resistivity of ingot is decided by the doped boron into silicon. During the
fabrication process, a variation in the concentration of impurities can lead to auto
doping process [15]. Such process occurs unintentionally in solar cell fabrication. The
resultant of these processes is defects that increase recombination, thereby reducing
the solar cell efficiency.
2.1 PV Factory
In this work, the production of solar cells are analyzed using PV Factory, which
a cloud-based simulation. The process of fabrication of PV cell in each step of
processing can be analyzed using PV factory simulation tool. Since, the batch size
of PV manufacturing is large in reality. Thus, analyzing the performance for such a
large dataset will be time consuming.
In order to perform simulation with smaller batch size, PV factory simulation tool
can be deployed. In this tool, the step by step processing of PV cell is similar to
actual manufacturing process of PV cell. PV Factory simulates the manufacturing of
photovoltaic (PV) solar cells for smaller batch size of 10 or 20 cells per batch. The
fabrication steps of PV cell from crystalline silicon wafers till electrical contacts are
simulated by PV factory.
The diffusing phosphorus into the silicon, deposition of silicon nitride film as
antireflection coating and screen printing-based electrical contacts on either side
of PV cells can be carried by this simulation tool. It is used to predict the factors
relating PV cell efficiency with fabrication parameters. It provides a platform improve
the efficiency of the solar cells by analyzing various parameters such as chemical
concentrations, the temperatures and the gas flows.
2.2 PL Imaging
The achievement of higher cell efficiencies, high yield at cost effective manner has
resulted in adoption of improved quality control and process monitoring tools. PL
imaging is well suited for such inline monitoring across the entire PV cell manufac-
turing process because of its minimum time of measurement and high repeatability.
For the characterization of silicon samples across almost the entire photovoltaic (PV)
value chain, Photoluminescence (PL) imaging is considered as a versatile technique.
PL imagining pays a way for inline process monitoring in PV cell production. As
fossil fuels are depleting in nature, there arises a need for search of alternate to fossil
fuels.
Solar PV cells were considered as best alternative energy. To make this a cost
effective technology, many researches are made toward cost reduction of PV cell
manufacturing. Though there are various ways through which cost reduction can be
attained, characterization of PV cell is considered to be the most essential method
of cost reduction in PV cell manufacturing.
An optical excitation is used by PL imaging, because it can be employed for wider
range of samples. The throughput of PV cell production can be increased by reducing
the risk of mechanical damage of samples. The defects in PV cell production can be
minimized by deploying contactless imaging of samples.
The silicon sample surface emits luminescence, when it is excited. This lumi-
nescent emission image is captured through a camera. Electroluminescence (EL)
imaging requires electrical contacts. Thus, it is deployed to fully processed solar
cells and modules alone. EL imaging is widely used for module inspection.
Besides doping level, the thickness of silicon ingot that is sliced into thin wafers
also decides the solar cell performance. During ingot slicing, the thickness of wafer
determines the amount of light absorbed by solar cell. Higher the light absorption
capability, higher the electron generations inside the solar cell [16]. To make high-
efficiency wafers from a wide range of wafers, manufacturers must produce more
good wafers. Any defective wafer rejection will lead to unnecessary expense of solar
cell production. For assisting this task, identification and removal of bad wafers at
the start of a production line is highly mandatory. The condition or status of wafers
produced can be examined with inline photo luminescent (PL) imaging [17].
For testing purpose, batch processing method has been adopted. Ingot is grown
using casting method and boron doping with concentration of 4.78 × 1015 cm−3
the batch size of casting method is considered as 10,000 wafers with thickness of
190 µm. Since casting method-based ingot growth is cost effective, the proposed
work deals with multicrystalline silicon ingot growth. At the end of PL imaging,
4.6% rejection has been obtained. Thus, 462 wafers are rejected from 10,000 wafers
during PL imaging. Defect fraction and impurity fraction are set as rejection criteria
during PL imaging [18].
The relation between number of wafers versus impurity fraction and defect frac-
tion is shown in Figs. 1 and 2, respectively. The inference from it states that the
rejection rate is minimum for each percentage of increase in impurity fraction. The
rejection criteria applied to PL imaging is tabulated in Table 1 and based on which
wafers are passed and rejected.
2.3 Chemical Etching
After wafer slicing process, the solar cell needs isotexturing, which is process that
chemically, etches the surfaces of multicrystalline silicon wafers [19]. Isotexturing
removes saw-damaged surface layer and induce a textured surface. Such textures
silicon cell can enhance the transmission of light into the solar cell [20].
Fig. 1 Statistics of passed

and rejected number of
wafers based on impurity
fraction

wafers based on defect
fraction
Table 1 Rejection criteria

Rejection criteria Level (%)
applied to PL imaging
Defect fraction 4.5
Impurity fraction 30
Table 2 Impact of
Solar cell parameters Wafer
isotexturing on wafers
Before isotexturing After isotexturing
Number of wafers 9532 9538
Thickness (µm) 190 169.7
Rate of removal 0 9.8
through etchant (µm)
Since etching is a chemical assisted process and fabricators need to have an etch
rate that is independent of the crystallographic orientation of the surface [21]. It
is also important to ensure that sufficient etching of silicon at saw-damaged layer
is performed. During the etching process, maintaining high optical transmission is
also demanded for solar cell production. The isotexturing process is performed at
a temperature of 25 °C with 25% of HNO3 and 20% of HF concentration. The
chemical reaction takes places for 25 min time duration. Table 2 shows the impact of
isotexturing on wafer, where the number of wafers placed for isotexturing are 9538
and wafers obtained after chemical reaction are 9532. The etchant removes 9.8 µm
silicon from each side of wafer and thus yielding an average wafer thickness around
169.7 µm. The optical transmission angle of texturing is 58.9°.
The textured solar cells are then loaded into a furnace subjected to high tempera-
ture in the presence of phosphorus gas. The diffusion of phosphorus atoms into silicon
surface leads to n-type emitter region formation. The electron transportation to nega-
tive contacts is supported by emitter, which inturn minimizes the recombination and
resistive losses in the solar cell [22].
3 Emitter Region Formation
The performance of solar cell is dependent on emitter properties. When the diffusion
of phosphorus concentration is high then resistive losses can be minimized [23]. The
impact of diffusion on resistivity decides the number of wafer passed and rejected
and is shown in Fig. 3.
The diffusion of POCl3 at a temperature of 800 °C for a duration of 20 min leads
to a sheet resistance of 126.8 Ω/sq. The average effective lifetime of electrons is
21.9 µs. The electron transportation provides an average emitter current density J0
as 841.8 fA/cm3 . The mechanical yield of POCl3 diffusion process provides 9529
wafers out of 9532 wafer supplied. Thus mechanical yield is 99.9% during POCl3

wafers based on resistivity
after diffusion of POCl3
diffusion process. The impact on number of wafers rejected and accepted after POCl3
diffusion based on impurity fraction and defect fraction are shown in Figs. 4 and 5.
A very thin layer of silicon nitride (SiNx ) is coated in the front surface of solar
cells is termed as antireflective coating (ARC) [24]. It reduces recombination by
‘passivating’ the surface and by providing a source of hydrogen that deactivates
some defects in the bulk of the wafer. The rate of defects after diffusion in terms of
wafers passed and rejected is shown in Fig. 6.
A combination of plasma of silane (SiH4 ) and ammonia (NH4 ) gases react to form
SiNx and hydrogen is supplied into a vacuum chamber were wafers are placed. The
ratio of these gases and temperature impacts on the deposition rate and on properties
of the film. A film thickness of 77.5 nm of SiNx is grown at a temperature of 375 °C.
Thus mechanical yield is 99.8% during ARC process.
4 Extraction of Electricity
A metal layer is formed to extract electricity from solar cell, where metal on front
surface contacts n-type emitter and forms negative electrode. While metal on rear
surface contacts p-type bulk of solar cell and forms positive electrode [25]. For

wafers based on impurity
fraction after POCl3
diffusion

wafers based on defect
fraction after POCl3
diffusion

wafers based on DAD factor
after POCl3 diffusion
preventing electrons from traveling to the positive contact and avoiding recombi-
nation at rear metal interface, a heavily doped p-type layer called a back-surface
field (BSF) is formed. To ensure good contact between metal and silicon, the firing
temperature and time must be properly selected. In case of more heat applied to paste,
then metal layer may fail. Apart from this, firing process also releases hydrogen from
the silicon nitride film and diffuses it into the bulk of the wafer where it can deacti-
vate recombination defects, particularly crystallographic defects. With 45 µm finger
width, 1500 µm spacing of metal grid, Ag paste thickness of 30 µm and Al paste
thickness of 30 µm, a mechanical yield of 99.95% is obtained. This yield is obtained
at a peak temperature of 810 °C.
The virtual production analysis is carried out with a batch size of 10,000 wafers. The
inference from analysis states that PL imaging-based production analysis provides
electrical yield 92%, while without PL imaging the yield is 86.2%.
The number of wafers failed during analysis is 745 for with PL imaging and 1376
for without PL imaging. Table 3 shows the comparison of virtual production analysis
Table 3 Comparison of
With PL imaging Without PL imaging
virtual production analysis
with and without PL imaging Median V oc (mV) 640.6 640.3
Median J sc 35.1 35.1
(mA/cm2 )
Median fill factor 79.7 79.6
(%)
Median efficiency 17.9 17.9
(%)
Cells failed 754 1376
Electrical yield (%) 92.0 86.2
Fig. 7 Comparison of with

and without PL imaging
based on J sc
with and without PL imaging. The comparison on number of cells based on J sc , Fill
factor, V oc and efficiency are shown in Figs. 7, 8, 9 and 10.
6 Conclusion
In recent days it has became mandatory to predict the solar cell performance in
prior to installation. Such prediction will support better design, analysis and control
of photovoltaic systems. In this work, performance prediction of solar cells using

based on fill factors

based on V oc

based on efficiency
PV factory emphasizes minimized time and cost by virtual production analysis.

10,000 samples under virtual performance analysis, yields 9337 successful samples.
Thus, electrical yield of 92% is obtained using photo luminescent (PL) imaging. An
electrical yield of 86.2% is obtained for analysis without photo luminescent (PL)
imaging. Thus by using PL imaging, the defect rate of solar cells can be predicted
before solar cell installation. It is inferred that using virtual performance prediction,
defect rate is minimized.
References
1. Andrew RM (2020) Timely estimates of India’s annual and monthly fossil CO2 emissions.
Earth Syst Sci Data 12:2411–2421. https://doi.org/10.5194/essd-12-2411-2020
2. Raghuwanshi SS, Arya R (2019) Renewable energy potential in India and future agenda of
research. Int J Sustain Eng 12(5):291–302. https://doi.org/10.1080/19397038.2019.1602174
3. Owusu PA, Asumadu-Sarkodie, S, Dubey S (reviewing ed) (2016) A review of renewable
energy sources, sustainability issues and climate change mitigation. Cogent Eng 3(1). https://
doi.org/10.1080/23311916.2016.1167990
4. Abbasi T, Abbasi SA (2012) Is the use of renewable energy sources an answer to the problems
of global warming and pollution? Crit Rev Environ Sci Technol 42(2):99–154. https://doi.org/
10.1080/10643389.2010.498754
5. Jurasz J, Canales FA, Kies A, Guezgouz M, Beluco A (2020) A review on the complementarity
of renewable energy sources: concept, metrics, application and future research directions. Sol
Energy 195:703–724. ISSN 0038-092X. https://doi.org/10.1016/j.solener.2019.11.087
6. Bhattacharjee S, Bhattacharjee R (2021) Comprehensive solar energy resource characterisation

for an intricate Indian province. Int J Ambient Energy 42(2):187–202. https://doi.org/10.1080/
01430750.2018.1531257
7. Tyagi V, Rahim N, Abd Rahim N, Selvaraj J (2013) Progress in solar PV technology: research
and achievement. Renew Sustain Energy Rev 20:443–461. https://doi.org/10.1016/j.rser.2012.
09.028
8. Goel M (2016) Solar rooftop in India: policies, challenges and outlook. Green Energy Environ
1(2):129–137. ISSN 2468-0257. https://doi.org/10.1016/j.gee.2016.08.003
9. India’s solar energy future: policy and institutions
10. India 2020—energy policy review
11. Sen R, Bhattacharyya S (2014) Off-grid electricity generation with renewable energy technolo-
gies in India: an application of HOMER. Renew Energy 62:388–398. https://doi.org/10.1016/
j.renene.2013.07.028
12. Feron S (2016) Sustainability of off-grid photovoltaic systems for rural electrification in
developing countries: a review. Sustainability 8:1326. https://doi.org/10.3390/su8121326
13. Hudedmani M, Soppimath V, Jambotkar C (2017) A study of materials for solar PV technology
and challenges. Eur J Appl Eng Sci Res
14. Kutsukake K (2019) Growth of crystalline silicon for solar cells: mono-like method. In: Yang D
(ed) Handbook of photovoltaic silicon. Springer, Berlin, Heidelberg. https://doi.org/10.1007/
978-3-662-56472-1_35
15. Koleske D, Wickenden A, Henry R, Twigg ME (2002) Influence of MOVPE growth conditions
on carbon and silicon concentrations in GaN. J Cryst Growth 242:55–69. https://doi.org/10.
1016/S0022-0248(02)01348-9
16. Bhoopathy R, Kunz O, Juhl M, Trupke T, Hameiri Z (2020) Outdoor photoluminescence
imaging of solar panels by contactless switching: technical considerations and applications.
Prog Photovolt Res Appl 28:217–228. https://doi.org/10.1002/pip.3216
17. Schindler F, Giesecke J, Michl B, Schön J, Krenckel P, Riepe S, Warta W, Schubert MC (2017)
Material limits of multicrystalline silicon from state of the art photoluminescence imaging
techniques. Prog Photovolt Res Appl 25:499–508. https://doi.org/10.1002/pip.2836
18. Michl B, Padilla M, Geisemeyer I, Haag S, Schindler F, Schubert M, Warta W (2014) Imaging
techniques for quantitative silicon material and solar cell analysis. IEEE J Photovolt 4:1502–
1510. https://doi.org/10.1109/jphotov.2014.2358795
19. Wu X, Li J (2020) Texturing technology on multicrystalline silicon wafer by metal-catalyzed
chemical etching: a review. J Inorg Mater 361. https://doi.org/10.15541/jim20200361
20. Chen K, Liu Y, Wang X, Zhang L, Su X (2014) Novel texturing process for diamond-wire-sawn
single-crystalline silicon solar cell. Sol Energy Mater Sol Cells 133. https://doi.org/10.1016/j.
solmat.2014.11.016
21. Kolasinski KW, Unger BA, Ernst AT, Aindow M (2019) Crystallographically determined
etching and its relevance to the metal-assisted catalytic etching (MACE) of silicon powders.
Front Chem 6:651. https://doi.org/10.3389/fchem.2018.00651
22. Richards B, Richards S, Boreland M, Jamieson D (2004) High temperature processing of TiO2
thin films for application in silicon solar cells. J Vac Sci Technol A Vac Surf Films 339–348.
https://doi.org/10.1116/1.1647594
23. Chen N, Ebong A (2015) Generalized analysis of the impact of emitter sheet resistance on
silicon solar cell performance. Jpn J Appl Phys 54:08KD20. https://doi.org/10.7567/JJAP.54.
08KD20
24. Shanmugam N, Pugazhendhi R, Madurai Elavarasan R, Kasiviswanathan P, Das N (2020) Anti-
reflective coating materials: a holistic review from PV perspective. Energies 13:2631. https://
doi.org/10.3390/en13102631
25. Lu Z, Lu PH, Cui J, Wang K, Lennon A (2013) Self-patterned localized metal contacts for
silicon solar cells. J Mater Res 28:1984–1994. https://doi.org/10.1557/jmr.2013.204
Optimisation of FPGA-Based Designs
for Convolutional Neural Networks
P. L. Bonifus , Ann Mary Thomas, and Jobin K. Antony
Abstract Convolutional neural networks (CNN) is a widely known deep learning

architecture and achieves higher accuracies in speech processing and computer vision
applications. CNN has arisen as best option in the fields such as video surveillance,
mobile robot vision, smart factories, and medical diagnostics. CNN has shown its
tremendous potential for image understanding in Cardio-Vascular (CV) systems. But,
the tremendous potential of CNNs came up with additional computational burden. To
get an outstanding performance of CNN-based algorithms, it requires large amount
of enormous applications and memory resources. Matrix multiplication (MM) is
the most fundamental computational operation taking place in CNN. Researchers
in this field found that multiplication operation is the most resource-intensive and
power-hungry operation in CNN. This work aims to design an FPGA-based CNN
using systolic array which can improve both the accuracy and hardware efficiency of
convolutional neural network. Here, we made design optimisation in multiplier unit
which brings a significant impact on overall performance of convolutional neural
networks.
Keywords Deep learning · CNN · FPGA · Bfloat-16
P. L. Bonifus (B) · A. M. Thomas · J. K. Antony

Department of Electronics & Communication Engineering, Rajagiri School of Engineering &
Technology, Kochi, Kerala, India
e-mail: bonifus@rajagiritech.edu.in
A. M. Thomas
e-mail: annmaryoliapuram14@gmail.com
J. K. Antony
e-mail: jobinka@rajagiritech.edu.in
https://doi.org/10.1007/978-981-19-6913-3_14
210 P. L. Bonifus et al.
1 Introduction
1.1 General Overview
Hardware implementation of deep learning (DL) by means of convolutional neural

networks (CNNs) has received wide acceptance recently. It has been widely used
in various applications ranging from handwriting recognition to medical diagnosis
and much more. Structure of optic nerves in living beings was the inspiration for
CNNs. Approximately 86 billion (86 × 109 ) neurons are there in a human brain.
Likewise, artificial neuron is the basic building block of neural network. Lot many
artificial neurons are interconnected to form a Neural Network. Hence, CNN model
demands a lot of computation. This makes the training and deployment phase to
consume more power. In real-world applications, graphic processing units (GPUs)
are commonly used as a platform for CNN. But, GPUs are inherently power-hungry
devices; this will limit their usage in embedded applications like transportable gadgets
and wearable devices. In low-power budget systems like portable devices and in other
embedded systems, field programmable gate arrays (FPGAs) or application specific
integrated circuits (ASICs) could be considered as an alternative for GPUs because of
their lesser power consumption compared to GPUs. Among FPGA and ASIC-based
designs, FPGA-based accelerators offer low non-recurring engineering (NRE) cost,
less deployment time, and reconfiguration capability. Reconfiguration capability of
FPGAs makes them a better choice for neural networks.
FPGAs are configurable hardware devices that have adaptable components. When
comparing with other processor-based CNN accelerators, accelerator implementa-
tions as soft logic on FPGA have the following advantages [1].
• Because of rich parallel computing resources and high energy efficiency, FPGA-
based accelerators are considered as perfect for deep convolutional neural
network.
• Reprogramming property of FPGA reduces their electric requirements and offers
them higher overall performance in terms of acceleration and throughput.
• FPGA are more resistant to rugged settings and environmental factors and have
longer life span than that of GPU.
• FPGAs are already using in AI products by some companies. An example is
Microsoft, which provides its FPGA-based powered machine learning technology
for its Azure cloud services.
• FPGA offers high energy efficiency compared to CPU or GPU.
Therefore, FPGAs can be optimised for specific styles of architectures inclusive
of CNN.
Optimisation of FPGA-Based Designs for Convolutional Neural Networks 211
1.2 Motivation
Although the basics of CNN were proposed way back in the 1980s, hardware imple-
mentation was not popular because of the lack of enough computational resources.
Because of the advancements in VLSI technology, such as 3D stacking and Finfets,
now transistor count and hence processing power is not a big concern. Due to this,
hardware implementation of CNNs is again in limelight. Deeper and larger CNNs
require more computational resources. Reducing the complexity required for these
computational resources could reduce the overall design complexity and could reduce
the computational delay and computational power required.
1.3 Objective
This work aims to improve the hardware implementation efficiency of CNN by

reducing its computational complexity in the multiply and accumulation (MAC)
unit.
2 Literature Review
The aim of this literature review is to study the existing FPGA-based CNN archi-
tecture and its design optimisations which improves the hardware implementation
efficiency and prediction accuracy of Neural Networks.
There are several researches were done to optimize or reduce the computational
complexity of CNN by applying certain transformations to the matrix multiplication
operation. Several methods proposed by different researchers are explained here.
A most commonly observed issue in FPGA-based accelerator design is that the
computations required may not match with the memory bandwidth available in an
FPGA platform; this may cause either underutilisation of logic resources or memory
bandwidth. In order to address this issue, a roofline model-based scheme using
analytical design was proposed [1]. This implementation was having a throughput
of 61.62 GFOP/s under 100 MHz working frequency. Limitation of this technique is
that it only considered small CNN models for simple task.
In [2], describes a dynamic-precision method which quantises the data to have
a visible reduction in bandwidth required and reduction in the number of resources
utilised. This research says that implementation of computational units on FPGAs
using fixed point logic is comparatively more efficient than floating point computa-
tional units. This system offers performance at 187.8 and 137 GOP/s under 150 MHz
frequency. Main drawback of this method is time-consuming when we retrain the
model.
In [3], explains about a FPGA-based accelerator in which all convolutional layers

were mapped onto a single chip and thereby concurrent processing can be done
by different layers. This technique of computations on batches helps to increase
bandwidth utilisation required for memory access. This system achieves a maximum
performance of 565.4 GOP/s under 156 MHz clock frequency and 391 GFPO/s under
156 MHz clock frequency. Main drawback is the sufficiently high time required to
retrain the model.
Depth-wise separable convolution design technique [4] studied the possibilities
of reducing computational complexities in a standard convolution. Reduction in
computational complexities could make neural network applications more suitable
for embedded system platforms such as portable gadgets and wearable systems. This
system is able to do each classification job 3.75 ms using ImageNet database and
have 20× times speed if compared to CPU. Limitation of this technique is time
consuming when we retrain the model.
The method proposed in [5] can solve the issue of time-consuming when retraining
a CNN model. Here refers a design technique called block-floating point arithmetic
which increased hardware implementation efficiency and energy efficiency by three
times. In this method, a highly optimised floating point unit makes use of block-
floating point concept is first developed to reduce the hardware implementation cost
and off-chip data traffic. To improve the throughput of the hardware, a memory
access pattern, optimised off-chip buffer memory, and a 3-level parallel convolution
processing units were adopted.
Researchers like Guo et al. and Han et al. found that computations with low-
precision arithmetic are sufficient for doing inference part in deep learning. They
were able to cut down the implementation complexity for modelling the neural
network, by reducing the arithmetic precision and found that accuracy reduction was
within tolerance limit. Compared to the reduction in hardware complexity accuracy
reduction was very minimal.
Fast Fourier Transforms (FFTs) can be used to reduce the computational
complexity in CNNs [6]. Accuracy loss can be prevented but found useful only
for networks having large kernels.
Toom and Cook proposed an algorithm based on minimum filtering called as
Winograd’s algorithm. Winograd’s algorithm is able to reduce the count of multi-
plication (floating point) operations up to 2.25 times inside a CNN model. Here,
instead of dot product, calculating the resultant matrix using specific formula. Main
drawback of this algorithm is that it can only work in neural networks with small
kernels [7].
3 Impact of Matrix Multiplier on CNN
Convolutional neural network is basically made up of three different types of layers:

convolutional layer, pooling layer, Rectified Linear Units (ReLU) correction layer,
and fully connected layer. The convolutional layer can successfully record the spatio-
temporal dependencies of an image by applying relevant filters. Matrix multiplication
is the fundamental operation inside a convolutional layer. That means convolution
is nothing but a matrix operation between ‘k’ of filter and portion ‘p’ of image over
which the filter is hovering.
Consider an example of recognizing of a 224 × 224 RGB image using CNN when
we are applying in a VGG-16 model, the complete process contains 138 billion (138
× 109 ) parameters. To do the feed-forward operation for an RGB image of size 224
× 224, the whole network requires 30 billion (30 × 109 ) addition and 30 billion
(30 × 109 ) multiplication operations. Hence, we can summarize from this example
that, CNN is a computationally complex network. Researchers in this field found that
multiplication operation is the most resource-intensive and power-hungry operations
in CNN. Thus, any improvement made on multiplier unit brings a significant impact
on overall performance of CNN [5].
4 Proposed System
To address such challenges dealt in previous sections, here we proposed a suitable

architecture called systolic array architecture using bfloat16. This system is able to
accept inputs one by one, perform multiplication of two inputs, and accumulate the
result with previous multiplication results as required by the CNN architecture. This
section discusses on systolic array and its advantages to develop a scalable CNN
implementation.
4.1 Systolic Array Architecture
In systolic array architecture (SA), processing elements (PEs) are arranged as a 2-

dimensional array arrangement with pipelining capability in order to reduce delay
for data movement. This architecture is capable of accelerating general matrix multi-
plication (GEMM) and being accepted widely in academia as well as in industry.
The systolic array family can reuse intra-PE operand and have high data path effi-
ciency results in reduction of circuit area and power dissipation. The basic idea of
systolic array architecture is mapping high-level computation into hardware struc-
tures. Systolic architecture allows multiple computations for each memory access.
Hence, it can execute compute-bound problems at high speed without increasing
input–output requirements. Main attractive features are regularity, re-configurability,
and scalability. Because of the high level of parallelism incorporated, systolic array
architecture can offer high rate of throughput. Data flow through this architecture is
having a regular layout pattern and simple to understand. Systolic array features low
data transfer latency and enables higher clock frequencies because of its own regular
layout pattern and support for local communication. Thus, this architecture is suit-
able for complex parallel designs on FPGAs. Because of these advantages, systolic
array architecture is being widely accepted for using in hardware implementation of
matrix multiplications, especially in reconfigurable devices like FPGA [8].
4.2 Systolic Array Design
Figure 1 shows the two-dimensional 4 × 4 systolic array representation. The 4 × 4

refers the arrangement of processing element in both row and column wise. There
are total of 16 processing elements are present inside a 4 × 4 systolic array. Multiply
and accumulate operation (MAC) is taking place inside each processing element [9].
Let’s look on how matrix multiplication operation carried out inside the systolic
array which is illustrated in Fig. 2. At first, here it is Matrix B, fed in row-wise which
is then moved down in the array. Next, the other matrix, in this case, Matrix A will
be fed in a column-wise order and move to right from left. Till each processor is
packed with a complete row and a complete column, dummy values will be issued.
Once, each processor is packed with a complete row and a complete column, the
final result of multiplication can be saved in an array and could be taken as output
either as single row in one time instant or single column in one time instant.
Fig. 1 4 × 4 systolic array representation

Fig. 2 Working of 4 × 4 systolic array
In this systolic array, each rectangular box represents the processing element (PE).
Processing elements placed on the periphery of the design can only communicate
with the exterior environment. Processing steps in a single processing element can be
summarized as it receives an input, multiply and accumulate (MAC) unit inside the
processing element compute with the inputs, and finally transmits the corresponding
output to neighbouring cell [9].
Figure 3 shows a single MAC unit. Systolic array is interconnection of array of
processing elements. Each processing elements composed of multiply-accumulate
unit (MAC) unit. MAC unit performs the operation of both multiply and accumulates.
Principle operation of MAC unit is evaluating partial products and then partial prod-
ucts result is added by previous stage result. Here, the single element from matrix
A is multiplied with single element from matrix B. Then the multiplication result is
added with previous cell’s output [4]. As the first phase of project, implemented 32 ×
32 integer-based systolic array in Verilog RTL. Vivado 2020.1 is used for simulation.
4.3 IEEE 754-Based Floating Point
Advantage in using any fixed point denotation is performance and disadvantage is

that we can only represent a very few range of values. Thus, it is usually inadequate
for numerical analysis it does not permit sufficient numbers and accuracy. Currently,
IEEE 754 is the widely accepted single precision standard used by personal computers
Fig. 3 Single MAC unit
using MACs, Windows, and most of the Unix platforms. Float-based number repre-
sentation is used when more precision is needed. To design our CNN system with
more precision and accuracy, this project work opting floating point-based arithmetic
operation in our architecture [10].
4.4 Bfloat-16
From the very beginning itself, the workloads for machine learning tasks may take
hours or even days to run. In order to help organisations, by reducing the workload
running time, recently Google Inc., designed and developed custom processors for
deep learning applications named tensor processing units (TPUs). And this cloud
TPU achieves high performance because it is working with innovative floating point
format called as Bfloat-16 or Brain Float-16. The high-performance capability of
cloud TPUs was achieved with the help of Bfloat-16. And also it helps to improve
hardware efficiency with minimal effort for switching from FP-32. Thus, we can
achieve even performance with Bfloat-16 [11].
4.5 Floating Point Versus Bfloat-16
There are many floating point formats you can hear about in the context of deep
learning. In this work, we are only dealt with IEEE 754 Floating point-32 and Bfloat-
16.
As you can see in Fig. 4, single precision floating point format defined in IEEE 754
standard (FP-32) is divided into 1 bit for sign (S) representation, 8 bits for exponent
part (E) representation, and 23 bits for mantissa (M) or fractional part representation.
FP-32 is a 4-byte 32-bit floating point number ranges from ~1e−38 to ~3e38 . When
it comes to fp-16 format, Mantissa bits are reduced to 10 bits but precision of FP-16
is reduced by half when compared to FP-32. As well, FP-16 can only represent less
range of numbers. Here comes the relevance of newly developed innovative floating
point format called bfloat-16.
Bfloat-16 is a custom-tailored single-precision floating-point representation using
only 16 bits, proposed for machine learning applications. Bfloat-16 consists of 1 sign
bit (S), 8 exponent bits (E), and 7 mantissa bits (M). Even with this 16-bit represen-
tation, Bfloat-16 can represent same range of values as that of fp-32 floating point
format. And also, there is no loss of accuracy and precision. Thus, artificial intelli-
gent groups at Google found that bfloat-16 format provides same results similar to
the fp-32 format while reducing the hardware implementation complexity, reducing
the memory usage, and thus increasing the overall performance [11].
In this project work, simulated and synthesised 4 × 4, 8 × 8, 16 × 16, and 32 ×
32 systolic arrays using both IEEE 754 floating-point standard and bfloat-16 format.
Fig. 4 fp32 versus fp16 versus bfloat16

Output waveform obtained after the simulation of 32 × 32 IEEE 754 floating point-
based systolic array is shown in Fig. 5. Similarly, designed 32 × 32 bfloat-16-based
systolic array and its corresponding waveform is shown in Fig. 6. Finally, compares
the results with respect to power consumption, area utilisation, and critical path delay.
Fig. 5 IEEE 754 standard float-based 32 × 32 systolic array
Fig. 6 Bfloat-16-based 32 × 32 systolic array

5 Results and Discussions
This project work includes the simulation and synthesis of 4 × 4, 8 × 8, 16 × 16, and
32 × 32 systolic arrays using both IEEE 754 standard floating point and Bfloat-16.
The systolic array is modelled using Verilog HDL. For simulation and synthesis,
Vivado 2020.1 is used. From the synthesis report targeted for Zynq xc7z020clg484-
1 device, obtained the information on how much power each type of systolic array
(whether it is IEEE 754 standard float standard or Bfloat-16) is consumed, area
utilization, and critical path delay. The exact values of power consumption, hardware
utilisation, and critical path delay obtained after synthesising Verilog models of
different systolic array sizes based on IEEE 754 implementation against Bfloat-16
are shown in Table 1. Based on that information, reached in an inference that Bfloat-
16-based systolic array architecture requires less power, less number of hardware
resource, and hence less value of critical path delay when compared with IEEE
754 standard floating point. Hence, we can say that implementing bfloat-16-based
systolic array architecture on CNN network can improve its hardware efficiency and
gives us high throughput.
5.1 Analysis Table
The comparison of power consumption and area utilisation by different size systolic
array with respect to both IEEE 754 standard float standard or Bfloat-16 is plotted
on Figs. 7 and 8.
Figure 7 shows graphical representation of power consumed by different size of
systolic array designed with respect to IEEE 754 standard floating point and Bfloat-
16. Here, x-axis depicts different size of systolic array, which are 4 × 4, 8 × 8,
Table 1 Analysis of systolic arrays implemented using IEEE 754 standard floating point and Bfloat-
16
IEEE 754 BFloat16
4×4 8×8 16 × 16 32 × 32 4×4 8×8 16 × 32 ×
array array array array array array 16 32
array array
LUT as 19.06 45.8 71.3 86.2 8.6 34.2 63.9 70.09
logic (%)
Registers 2.43 9.6 24.3 47.1 1.1 4.67 16.9 33.3
as flip
flops (%)
Power (W) 607.17 3160.5 13,588.1 24,279.6 247.6 43.4 740.2 999.6
Critical 13.73 20.725 30.15 45.4 11.03 12.61 15.22 20.3
path delay
(ns)
Fig. 7 Power consumption—floating point versus Bfloat
Fig. 8 Area utilization—floating point versus Bfloat
16 × 16, 32 × 32, and y-axis depicts power consumption in Watts (W). Blue line
represents systolic array designed with respect to IEEE 754 standard float. While red
line represents systolic array designed with respect to Bfloat-16 format. By analysing
this graph, we can infer that systolic array based on bfloat-16 consumes less power
when compared to IEEE 754 standard float standard format.
Figure 8 shows graphical representation of area utilised by different size of systolic
array designed with respect to IEEE 754 standard floating point and Bfloat-16. Here,
x-axis depicts different size of systolic array, which is 4 × 4, 8 × 8, 16 × 16, 32 ×
32, and y-axis depicts the percentage of area utilization. Red line represents systolic
array designed with respect to IEEE 754 standard float. While violet line represents
systolic array designed with respect to Bfloat-16 format. By analysing this graph, we
can infer that systolic array based on bfloat-16 requires less area when compared to
IEEE 754 single precision floating point standard.
6 Conclusion
This paper aims to improve the hardware efficiency of CNN by reducing its compu-
tational complexity in multiplication unit of convolutional neural network. For that,
here proposed a systolic array architecture to carry out matrix multiplication oper-
ation inside the convolutional layer of CNN. This architecture offer us to gain the
advantages of parallel computing. Then, designed the same systolic array architec-
ture implementation in two different floating-point number representations. They
are IEEE standard floating point and Bfloat-16-based format. Finally, we arrived at a
conclusion that bfloat-16-based systolic array architecture can reduce the area utilisa-
tion, reduce the power consumption, reduce the critical path delay and hence improve
the overall performance metrics of CNN implementations targeted on FPGA.
References
1. Zhang C, Li P, Sun G, Guan Y, Xiao B, Cong J (2015) Optimizing FPGA-based accelerator

design for deep convolutional neural networks. In: Proceedings of the 2015 ACM/SIGDA
international symposium on field-programmable gate arrays, Feb 2015, pp 161–170
2. Lian X, Liu Z, Song Z, Dai J, Zhou W, Ji X (2019) High-performance FPGA-based CNN
accelerator with block-floating-point arithmetic. IEEE Trans Very Large Scale Integr (VLSI)
Syst 27(8)
3. Li H, Fan X, Jiao L, Cao W, Zhou X, Wang L (2016) A high performance FPGA-based accel-
erator for large-scale convolutional neural networks. In: 2016 26th international conference on
field programmable logic and applications (FPL), Sept 2016
4. Bai L, Zhao Y, Huang X (2018) A CNN accelerator on FPGA using depthwise separable
convolution. IEEE Trans Circuits Syst II 65(10)
5. Qiu J et al (2012) Going deeper with embedded FPGA platform for convolutional neural
network. In: Proceedings of the ACM/SIGDA international symposium on field programmable
gate arrays, pp 26–35
6. Zhao Y, Wang D, Wang L, Liu P (2018) A faster algorithm for reducing the computational
complexity of convolutional neural networks. www.mdpi.com/journal/algorithms. Received:
10 Sept 2018; Accepted: 16 Oct 2018; Published: 18 Oct 2018
7. Winograd S (1980) Arithmetic complexity of computations. SIAM, Philadelphia, PA
8. Liu Z-G, Whatmough PN, Mattina M (2020) Systolic tensor array: an efficient structured-sparse
GEMM accelerator for mobile CNN inference. IEEE Comput Archit Lett 19(1)
9. Yang Z, Wang L, Ding D, Zhang X, Deng Y, Li S, Dou Q. Systolic array based accelerator and
algorithm mapping for deep learning algorithms
10. https://www.sciencedirect.com/topics/engineering/floating-point-number
11. https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-per
formance-on-cloud-tpu
Design and Implementation
of an Automated Fuel Station
M. Jyothirmayi, Vibha B. Raj, V. Lekhana, and P. Manjunath
Abstract Currently, fuel stations are operated manually which requires a lot of
time and staff to operate all the system. Customers are defrauded while dispensing
the fuel by not delivering the exact amount of fuel, and during payments, invalid
notes may be exchanged. So, an automated fuel station can be a solution to all the
above-stated problems. In this project, an automated fuel station is designed and
implemented which can work 24/7 conveniently. An automated fuel station consists
of an automated fuel dispensing system, automatic level indicator of the fuel in the
storage tank, automatic vehicle washing system, smart lighting system, and automatic
fire extinguisher. An automatic fuel dispensing system is realized by interfacing
RFID, GSM module with Arduino microcontroller. Level indication of the fuel in
the storage tank is made automatic using an ultrasonic sensor. Vehicle washing is
done by spraying water controlled automatically through the movable frame. Light
intensity is automatically controlled based on the surrounding ambiance light and
presence/absence of a vehicle. The accidental fire in the station is handled by an
automatic fire extinguisher.
Keywords Automatic fuel dispensing · Fuel level monitoring · Vehicle washing

system · Arduino microcontroller · Lighting system · Fire extinguisher
1 Introduction
In recent years, fuel stations have gone through major changes in terms of its func-
tioning. The primary requirement of the fuel station is not only to supply petrol,
gas, air, and other gasoline products, it also caters to other facilities like cafeteria,
child play area, and other essentialities. To maintain these facilities for 24/7, huge
M. Jyothirmayi · V. B. Raj (B) · V. Lekhana · P. Manjunath

Department of Electronics and Instrumentation Engineering, Ramaiah Institute of Technology,
Bangalore, India
e-mail: vibha.raj@msrit.edu
M. Jyothirmayi
e-mail: jyothirmayi@msrit.edu
https://doi.org/10.1007/978-981-19-6913-3_15
224 M. Jyothirmayi et al.
manpower is required and more power is utilized. An automated fuel station is a

solution to reduce the man power and save energy.
Currently, most of the fuel stations in India are operated manually, though there
is the use of a microcontroller for fuel dispensing. The main aim of this work is to
develop a fully automatic fuel station by reducing the manpower used for filling the
fuel, collecting cash, vehicle washing and checking the level of fuel in the storage
tank and also power saving. Along with minimal usage of manpower, malpractices
in dispensing of fuel should be prevented and timely service needs to be provided.
The primary part of any fuel station is dispensing. An automated fuel station
needs to have an automatic fuel dispensing system to get the exact amount of fuel
as per the user requirement. This is achieved by providing a smart card (RFID tag)
[1, 2] to all the customers using which they can operate individually for the facilities
provided in the station and then enter the quantity of fuel to be dispensed. After the
dispensing of fuel as per the customer’s requirement, the amount is deducted from the
preloaded money in the RFID tag and the transaction details are sent to the customer’s
mobile phone through the GSM module. Continuous level monitoring of the fuel in
the storage tank is necessary to ensure that there is no shortage of fuel and provide
24/7 service. Therefore, an automatic level indication of the storage tank is designed
which displays the fuel level using an ultrasonic sensor [3, 4], which is presently
done manually using a dipstick/dip rod. An ultrasonic sensor is interfaced with the
Analog-to-Digital Controller (ADC) which converts analog output from the sensor to
the digital form and feed it to the microcontroller. Then, the microcontroller calculates
the level according to the digital value obtained and displays the results. Providing
quick services are essential in a fuel station. Hence, an automatic vehicle washing
system is realized, which can save a lot of time for the customers. The washing
process takes place in six steps, i.e., check if the vehicle is present on the conveyor,
soapy water spraying, brushing, cleaning with normal water, drying, exit from the
conveyor. IR sensors [5] are used to detect the vehicle in each step. The vehicle is
parked on the conveyor and washed along with the movement of the conveyor [6].
Energy-saving is a prime objective of any system, therefore implementing a smart
lighting system in the fuel station can reduce almost half of the energy usage. The
implementation of smart lighting system in street lights is carried out using a motion
sensor and a brightness sensor [7, 8]. The street lamps turn on automatically only
when it is dark and a movement is detected. When there is no movement, the lamps
are automatically turned off or adjusts intensity based on the ambient light in the
environment. The sensor outputs are sent to the microcontroller, which analyses the
data received and sets the intensity of the lights accordingly. Wireless technology
is implemented using the ZigBee module for communication between the sensors,
LED, and microcontroller. Thus, making it an energy-efficient system. Safety being
an important parameter in fuel stations, a check for fire is a major requirement as
fuel is a highly flammable. Most disasters can happen if proper precautions are
not taken. Hence, an automatic fire extinguishing system [9, 10] is realized using a
temperature sensor, smoke sensor, and solenoid valve. The sensors are used to detect
the presence of fire and the valve regulates the opening and closing of the nozzle of
the fire extinguisher.
Design and Implementation of an Automated Fuel Station 225
Since the advancement of technology has led to an increase in automation of all

industries, our idea is to automate not just the dispensing pump, but also automate
other systems such as level indication of the fuel storage tank, vehicle washing
system, lighting system, and safety of the fuel station. Hence, this work attempts to
integrate different sub-systems and to automate its use so that the petrol dispensing
system works independently with very less manpower and to understand the working
of different sub-systems for use as an automated system for managing the petrol
station.
2 Methodology
The suggested system is depicted in Fig. 1. The design of automated fuel station
includes an automatic fuel dispensing pump, automatic fuel level indicator, automatic
vehicle washing system, smart lighting system, and automatic fire extinguisher. The
next sections go through the details of each system block.
2.1 Automatic Fuel Dispensing Pump
Fuel dispensing in fuel stations demands more staff and is time-consuming. In addi-
tion to that, malpractices and fuel thefts are happening. To overcome these issues, it
Fig. 1 Sub-systems of an
automated fuel station
Fig. 2 Block diagram of an automatic fuel dispensing pump
is necessary to automate the dispensing pump. RFID and GSM technology are inte-
grated with the fuel dispensing system to make it completely automatic. A detailed
block diagram of the automatic fuel dispensing pump is depicted in Fig. 2.
The block diagram shows various components interfaced with the microcontroller.
Arduino Mega 2560 is the microcontroller used, which controls all the other compo-
nents. A smart card (RFID tag) is issued to all the customers, which needs to be
preloaded with money and it is scanned by the RFID reader. If the card is valid,
the LCD displays a message that the access is granted and the process continues.
Next, the customer has to enter the password using the keypad and if the password is
correct, the option for entering the amount of fuel required in terms of rupees (Rs.) or
milliliters (mL) is displayed on the LCD. The user has to select one of the options and
enter the amount accordingly. The amount of fuel to be dispensed and money to be
paid is calculated by the microcontroller using the input given by the customer. The
payment amount is checked with the balance available in the card, and if it is within
the balance the amount is deducted. The LCD shows the remaining balance in the
card and the amount of fuel to be dispensed. The microcontroller gives a command
to the relay to run the pump. The pump dispenses the exact amount of fuel as per the
inputs given by the user. Every process taking place is displayed on the LCD for the
user’s convenience. Also, a message is sent to the user’s mobile phone regarding the
money deducted and the remaining balance using the GSM module. The data about
the money collected and the amount of fuel dispensed to every customer is stored in
the workstation.
Fig. 3 Block diagram of an automatic fuel level indicator
2.2 Automatic Fuel Level Indicator
The level measurement has been a major issue especially when it comes to flammable
substances. So, a new method has to be adopted to measure and monitor the level.
Continuous monitoring of fuel level in the storage tank can be done by implementing
an automatic fuel level indicator (Fig. 3).
In this system, an ultrasonic sensor is used to detect the object and to measure
the distance between the object and itself. This sensor has two transducers one is
the transmitter and the other is the receiver. The ultrasonic sensor will be placed on
the top of the storage tank. The transmitter will generate ultrasonic waves for every
microsecond and the receiver will absorb the micro sound waves which bounce back
from the surface of the fuel level. Arduino will calculate the level of the fuel based
on the time taken by the sound waves to touch the surface of fuel and bounce back
to reach the receiver. The calculated result of the fuel level will be displayed on the
LCD. If the level of the fuel is less than the minimum level, then the dispensing of
the fuel will stop.
2.3 Automatic Vehicle Washing System
In this developing world, automation has become a demand. It allows us to save time,
money, and manpower. It’s necessary to have a well-functioning and effective method
for keeping the vehicles clean. A vehicle washing system is a simple maintenance
technique for keeping the vehicle’s exterior clean (Fig. 4).
The exterior of a vehicle must be kept clean to prevent rust, oxidation, and
scratches. With the help of a microcontroller, this technology cleans the vehicles
automatically. This process is divided into two parts: washing and drying. To begin
with, the user will scan an RFID tag on the reader, and the associated service charge
will be displayed on the LCD. To begin the vehicle washing procedure, the user
must press the start button, after pressing the button the appropriate amount will be
deducted and the Arduino is initialized. Car washing consists of three major steps:
first, clean water is sprayed over the vehicle, then detergent water is sprayed, and
Fig. 4 Block diagram of an automatic vehicle washing system
last, regular water is sprayed. To wash the car, the spraying frame will move forward
and backward at a set speed and time for each step. After that, the vehicle is dried
by switching the DC fans, fixed around the washing area. The traditional way of car
washing takes a long time and uses a lot of water. This may be prevented by adopting
this automatic vehicle washing system, which allows the vehicle to be washed in a
shorter amount of time while using less water. Further the used water is also filtered
and recycled in a water recycling thereby saving the water used.
2.4 Smart Lighting System
Automatic light operation with power saving. Using the Light dependent resistor
(LDR) sensor the ambient light is measured and the sensed value or the information
is sent to the Arduino. Arduino accordingly calculates and sets the brightness of the
LEDs (these LEDs are for the whole fuel station) as per the requirement of the station.
Another set of LEDs are placed near the functional areas. Infrared (IR) sensor is used
to detect object. As a vehicle approaches toward the station the IR sensor detects the
vehicle and sends the information to the Arduino. Arduino checks for ambient light
in the area where the vehicle has arrived, if the ambient light in that area is dark, then
it makes the light to glow only in that part of the station. This makes the station use
a minimum amount of power. Hence making it an energy-efficient system (Fig. 5).
Fig. 5 Block diagram of a smart lighting system
2.5 Automatic Fire Extinguisher
Fire causes damage to both life and property, therefore a fuel station’s assurance
of safety is very necessary. An automatic fire extinguisher can prevent any type of
losses by timely detection and quick extinguishing of the fire. The block diagram of
an automatic fire extinguisher is given in Fig. 6.
Fire in the fuel station is mainly caused due to fuel leakage or short circuit. Water
cannot be used to extinguish this type of fire, instead CO2 -based fire extinguishers
should be utilized.
To realize a prototype of a compressed CO2 type of fire extinguisher, a balloon is
used as a dupe to the compressed CO2 cylinder as shown in Fig. 7. A flame sensor and
a smoke sensor are fixed in areas where there is a higher risk of fire. For extinguishing
the fire, we need to burst the balloon (similar to the opening the nozzle of the fire
extinguisher cylinder), for which servo motor is used. Buzzers are used to alert and
alarm the surroundings. The sensors detect the fire and send the information to the
microcontroller, which then gives command to the servo motor to rotate and turn on
the buzzers. A needle is fixed to the servo motor so that the needle rotates along with
the servo motor and bursts the balloon. The air inside the balloon rushes out when
burst and blows off the fire.
Fig. 6 Block diagram of an automatic fire extinguisher

Fig. 7 Balloon used as a

dupe to the compressed
cylinder
3 Results
An automated fuel station is designed and implemented to a project prototype, which

is tried, executed and accurate results are obtained. In this section, we evaluate
the characteristics of the components utilized in realizing the sub-systems of the
automated fuel station.
3.1 Automatic Fuel Dispensing Pump
Calibration of the pump is critical in order to receive the correct amount of fuel,
delivered by the dispensing system. A non-submersible DC pump is used in imple-
menting the automatic fuel dispensing system. The effect of the fuel level in the
storage tank on the quantity of fuel being dispensed by the pump is plotted in Fig. 8.
Table 1 contains the quantity of fuel delivered by the pump at different fuel level
in the storage tank for various time intervals. The table and plot indicate that the
amount of fuel delivered is unaffected by the level of fuel in the storage tank.
It is also important to know if the pump dispenses the fuel accurately, Fig. 9 depicts
the plot and Table 2 contains the reading taken for the quantity of fuel dispensed by
the pump versus the time interval it runs. According to the results, the pump can
dispense 30 mL of fuel in 1 s and the relationship between the quantity and the time
is linear.
Figure 10a shows the message delivered to the customer’s mobile phone through
the GSM module, and Fig. 10b depicts the data logged in the workstation for the
quantity of fuel dispensed and money received.
Fig. 8 Plot depicting the effect of fuel level in the storage tank on the fuel quantity dispensed
Table 1 Level versus

Level Time (s)
quantity for different time
interval 2 4 5
Full (mL) 60 120 150
Half (mL) 60 120 150
Minimum (mL) 60 120 150
Fig. 9 Plot depicting the quantity of fuel dispensed versus time
Table 2 Quantity of fuel

Time (s) Quantity (mL)
dispensed
1 30
2 60
3 90
4 120
5 150
10 300
Fig. 10 a Messages received by the customer, b data logging

3.2 Automatic Fuel Level Indicator
The experiment was to calibrate level measurement from the reading obtained from
the ultrasonic sensor to relate liter and distance measured in cm. The height of the
storage tank is 22 cm (prototype) and width is 10 cm based on the distance the level
is calculated. The ultrasonic sensor is fixed on top of the tank below the lid. When
the distance is read as 2 cm, it means that the fuel level is full and when it shows
20 cm it means that storage tank is almost empty (0.2 L). Hence, the distance to the
level of the fuel is inversely related as shown in Fig. 11 (Table 3).
Fig. 11 Plot depicting the level-distance
Table 3 Fuel level

Distance in cm Level in liter
measurement
2 2.0
4 1.8
6 1.6
8 1.4
10 1.2
12 1.0
14 0.8
16 0.6
18 0.4
20 0.2
Fig. 12 Process of vehicle washing system
3.3 Automatic Vehicle Washing System
The automatic vehicle washing system is demonstrated as a prototype, in which a

vehicle enters a washing station and is cleaned automatically. The entire system is
automated, so no human intervention is necessary, the system is built on Arduino
automation, so all work is done digitally, and only mechanical assembly is required.
Vehicle is parked inside the washing area as shown in picture 1 (Fig. 12). Then, the
regular water will be sprayed on the vehicle in the forward to backward movement
of the frame attached with the sprayers as in picture 2 (Fig. 12). In the next forward
to backward movement of the frame, soap water is sprayed. Again regular water
is sprayed to wash off the soapy texture on the vehicle. At the end, the vehicle is
dried by the dryers which are placed around the frame. By adopting this automation
system, benefits such as decreased operating time and manpower can be achieved.
3.4 Smart Lighting System
The IR sensor detects the objects and sends a signal in the form of voltage to the
microcontroller through the analog pins. The data received by the microcontroller
are simply voltage equivalent analog values, i.e., 5 V equals 1024 units. Sensor
calibration of IR sensor is shown in Fig. 13. The analog value received by the sensor
Fig. 13 IR sensor calibration
when there is no vehicle in fuel station is less than or equal to 140 units, and when
vehicle is present the values obtained is greater than or equal to 865 units.
The LDR sensor measures the ambient light intensity and sends the data in the
form of voltage to the microcontroller through the analog pins. The data received by
the microcontroller are simply voltage equivalent analog values, i.e., 5 V is equivalent
to 1024 units. Sensor calibration of LDR sensor is shown in Fig. 14. The analog value
received by the sensor when there is low ambient light intensity is less than or equal
to 21 units, medium ambient light intensity the values are between 41 and 46 units,
and under high ambient light intensity the values obtained is greater than or equal to
130 units.
3.5 Automatic Fire Extinguisher
The analog pins on the microcontroller receive various readings from the sensor,
which are voltage equivalent analog values, i.e., 5 V equals 1024 units or 0.0049 V
(4.9 mV) per unit. The calibration graph of the flame sensor is shown in Fig. 15. The
analog value received by the sensor when there is no flame detected is 1023 units,
and when there is a flame detected, the analog value obtained is less than or equal to
33 units.
As illustrated in Fig. 16, the smoke sensor gets rapidly shifting analog readings.
The smoke is recognized when the resultant analog value exceeds the threshold value
of 400 units.
Fig. 14 LDR sensor calibration
Fig. 15 Flame sensor

calibration
4 Conclusion
In this work, it is possible to operate all the systems conveniently 24/7 with reduced
manpower and also the fuel is secured and malpractices can be avoided. The lighting
system of whole fuel station is made completely automatic. Automatic Vehicle
washing system saves the time of the customer. The level of the fuel in the storage
Fig. 16 Smoke sensor

calibration
tank is monitored and indicated automatically. In case of fire in the fuel station,
automatic fire extinguisher reacts immediately. Fire extinguishing system is also a
unique design that was used to extinguish the fire. This paper attempts to integrate
different systems that work as a single system that is ideal for a complete automation
system. The advancement of technology leads everything to get upgraded which led
to this idea.
References
1. Janani G (2018) Petrol bunk automation with prepaid card using GSM identification. Int J Res
Appl Sci Eng Technol (IJRASET)
2. Baqir ZM, Motlak HJ (2021) Smart automatic petrol pump system based on internet of things.
Int J Electr Comput Eng (IJECE)
3. Rai N, Jagtap A, Rajguru PD (2020) Digital fuel level and battery life indicator. Int Eng Res J
(IERJ) 3
4. Sunmonu RS, Sodunke MA, Abdulai OS, Agboola EA (2018) Development of an ultrasonic
sensor based water level indicator with pump switching technique. Int J Res Electron Electr
Eng
5. Singh RD, Nigam S, Aggrawal S, Neelgar MR, Kaura S, Sharma K (2018) Design and
implementation of automatic car washing system using PLC. Int Res J Eng Technol (IRJET)
05(05)
6. Lin KK (2019) Arduino based automatic car washing system. Int J Trend Sci Res Dev
7. Fujii Y, Yoshiura N, Takita A, Ohta N (2013) Smart street light system with energy saving
function based on the sensor network. Grant-in-aid for scientific research
8. Khachane MY (2018) Intelligent street lighting system. Int J Eng Res Comput Sci Eng
(IJERCSE)
9. John AJ, Ashik K, Avinash Vishnu KS, Fahmi P, Henna P (2016) Automatic fire extinguishing
robotic vehicle. Int J Sci Eng Res
10. Dey P, Das S, Das C, Ahmed T, Reza CMFS, Rahman M (2017) Design and implementation of
an automatic fire extinguishing system based on fault secure multi-detectors. In: International
conference on mechanical engineering and renewable energy
Heart Disease Prediction Using Machine
Learning Algorithms
Rea Mammen and Arti Pawar
Abstract Heart disease is synonymous with heart attacks and strokes. But, cardio-
vascular disease also includes maladies like coronary artery disease (CAD), heart
arrhythmias, hypertension, congenital heart disease, etc. Heart disease plagues a
majority of the population today and is the leading cause of death globally. Efficient
prediction systems to diagnose heart diseases are a must in the health care industry.
Such systems are already in use but there is scope for improvement and with tech-
nological advancement over the years, the accuracy of disease prediction has been
improved. Machine learning is a branch of artificial intelligence that predicts several
naturally occurring events by training a model with some data and then using unseen
data to test it. This paper seeks to analyze a few machine learning algorithms and
tests their accuracy in predicting heart diseases.
Keywords Machine learning algorithms · Coronary artery disease · Logistic

regression · SVM · Decision tree · Naïve Bayes
1 Introduction
Heart disease patients experience different symptoms like shortness of breath, dizzi-
ness, chest pain, or discomfort, etc. Predicting the heart failure risk of patients
suffering from these symptoms is vital in minimizing the risk of peracute heart
ailments, as an early diagnosis can aid the afflicted in seeking immediate health care.
The dearth of proper health care facilities/personnel in many areas and the increasing
population may lead to delayed diagnosis of heart diseases. Hence, computational
methods can help overcome these issues by providing a timely diagnosis. Machine
learning algorithms play a significant role in predicting the presence of heart diseases
R. Mammen · A. Pawar (B)

ICAS, Manipal Academy of Higher Education, Manipal 576104, India
e-mail: arti.pawar@manipal.edu
R. Mammen
e-mail: rea.mammen@learner.manipal.edu
https://doi.org/10.1007/978-981-19-6913-3_16
240 R. Mammen and A. Pawar
in a patient. This paper examines the accuracy of a few machine learning algo-
rithms, namely, support vector machine (SVM), logistic regression, Naïve Bayes,
and decision tree.
2 Background
2.1 Literature Survey
Machine learning techniques are helpful to detect heart disease in patients based
on their medical records. Different machine learning algorithms like SVM, Naïve
Bayes, decision tree, and KNN, applied on a heart disease Framingham dataset, are
analyzed. Using the ROC curve and computing the area under the curve, the accuracy
of these algorithms is calculated and tabulated. KNN algorithm scored the highest
accuracy (83.60%), followed by Naïve Bayes (80.66%), decision tree (75.58%), and
lastly, SVM (65.56%) [1].
WEKA, an open-source software containing machine learning algorithms for
data mining tasks, is used for comparing the accuracy of different decision tree
classification algorithms. Logistic model tree, J48, and random forest algorithms are
the classification tree algorithms under analysis. The Cleveland database in the UCI
repository provides the required heart disease dataset. The J48 algorithm showed the
best overall performance as it resulted in high accuracy (56.76%) and less build time.
In contrast, the logistic model tree algorithm showed the lowest accuracy (55.77%)
and a build time of 0.39 s [2].
SVM, ANN, KNN, logistic regression, classification tree, and Naïve Bayes are
implemented and applied on the StatLog heart disease data set. The study involved
partitioning the given data samples in tenfold with each fold being used exactly once
as the validation data for testing and the remaining folds used as the training data.
Logistic regression proved to have the highest classification accuracy of about 85%,
followed by ANN with an accuracy of 84%. SVM emerged as the most suitable for
detecting the absence of heart disease in patients. It has a high specificity of around
89%, i.e., it is good in identifying true negatives. Logistic regression and ANN have
high sensitivities (89% and 87%, respectively), which means that they are good at
identifying true positives, i.e., people with heart disease [3].
The expanding data analytics industry increases the necessity of more accurate
disease prediction systems, especially cardiovascular disease prediction. Efficient
techniques to forecast such fatal diseases need to be implemented to help detect
and treat these diseases well in time. Hence, this study involves studying machine
learning algorithms like Naïve Bayes, random forest, logistic regression, decision
tree, SVM, and KNN, applied on the heart disease dataset taken from the UCI Repos-
itory. Analyzing the comparison of their performance accuracies helps to see which
methodology performs best in accurately classifying the presence or absence of heart
disease in a patient. After pre-processing, the data was split into the training set (70%)
Heart Disease Prediction Using Machine Learning Algorithms 241
and the testing set (30%). Random forest was the most accurate (91.80%), followed
by Naïve Bayes (88.52%), decision tree (76.68%), SVM (88.52%), and lastly, logistic
regression and KNN with an accuracy of 86.88% each. The study also shows more
people afflicted with heart disease in the 50–60 age bracket [4].
SVM, a supervised machine learning algorithm, is used to predict cardiovascular
disease in a patient. The model considers factors like age, sex, and pulse rate, etc.,
to predict the disease. The SVM algorithm results in greater prediction accuracy
when compared to other machine learning algorithms. It has an average precision of
around 91%. SVM accuracy proved to be 85.97%. It has high sensitivity (90.10%)
and specificity (77.20%), proving it to be a good model overall compared to other
models [5].
The proposition of a hybrid method for coronary artery disease prediction seeks to
find an alternative to angiography. With the help of the genetic algorithm, the hybrid
method improves the performance of neural networks by approximately 10%. The
feature selection process involved popular ranking methods like Gini Index, weight
by SVM, information gain, and principal component analysis. The proposed method
tested on coronary artery disease (CAD) patients of the Z-Alizadeh Sani dataset
achieved an accuracy of 93.85%, sensitivity of 97%, and specificity of 92% [6].
Supervised machine learning algorithms (Bayesian networks, logistic regression,
decision trees, random forests, SVM, k-NN, and neural networks) are applied on the
Titanic data set obtained from Kaggle. Factors such as accuracy, comprehensibility,
and computation time form the basis of comparison among the supervised machine
learning algorithms under analysis. The mean replaces missing values in the dataset if
the data is continuous and the mode if the data is categorical. It is noteworthy that tree-
based algorithms like decision trees and random Forests, and discriminant analysis
outperformed the other techniques. Combining two or more efficient algorithms to
form a multimodal system can improve the accuracy of the predictive model [7].
In a PCA-SVM hybrid model, PCA is used to extract features and reduce the
dimensionality of the dataset, and SVM is successively employed for classification.
This hybrid model is effective in the early diagnosis and risk assessment of breast
cancer. The model presents an accuracy of 97.62%, a sensitivity of 95.24%, and a
specificity of 100% when tested on the dataset obtained from Lagos State University
Teaching Hospital (LASUTH). These improved values confirm the viability of this
hybrid model approach [8].
The C4.5 algorithm, k-NN, SVM, and Naïve Bayes applied on the Wisconsin
Breast Cancer datasets show that SVM has the highest performance accuracy
(97.13%) and lowest error rate in breast cancer risk prediction and diagnosis. The
comparison drawn among the algorithms mentioned above is related to the accu-
racy, sensitivity, specificity, and precision in classifying the breast cancer data set.
Both C4.5 and k-NN resulted in the highest error rate and consequently led to many
incorrectly classified classes [9].
The heart disease dataset tests the performance of three Bayes classifiers, namely,
Naïve Bayes, Bayes Net, and Naïve Bayes Multinomial. This performance estimation
utilizes the cross-validation parameter. The experimental results indicate that the
Naïve Bayes algorithm has the highest performance or classification accuracy and
the lowest error rate than Bayes Net and the Naïve Bayes Multinomial algorithms.
The Naïve Bayes algorithm correctly classified 83.7037% of the instances, while
the Bayes Net method correctly classified 83.3333%. On the other hand, the Naïve
Bayes Multinomial algorithm correctly classified only 73.7037% of the cases and
incorrectly classed nearly 26.30% of the cases [10].
Based on factors like sex, age, BMI, pulse rate, etc., an application to predict the
presence of heart disease in a patient is developed using neural networks because
of its accuracy and reliability. The dataset used in this study is the Cleveland heart
disease dataset from the UCI repository. The proposed system uses the multi-layer
perceptron (MLP) algorithm, a type of neural network, to predict the presence of
CAD in an individual before its onset. The system obtained an overall precision of
91% and predicts if a person will have CAD in the future [11].
Machine learning algorithms k-NN, decision trees, linear regression, and SVM
are applied to the heart disease dataset obtained from the UCI repository to compare
their accuracies. About 73% of the collected data was used as the training set, and
37% as the testing set. The performance accuracy of k-NN turns out to be the best
among the four algorithms under comparison, with an accuracy of 87%. SVM has
an accuracy of 83%, followed by decision tree (79%) and linear regression (78%).
Confusion matrices support the calculation of the accuracies mentioned above [12].
Dealing with large datasets involves a certain degree of complexity. The devel-
opment of an Enhanced New Dynamic Data Processing (ENDDP) algorithm helps
detect, more accurately, the existence of heart disease in its early stages. This study
uses the processed UCI Cleveland dataset. The proposed system, which amalgamates
several Bayesian functions that are independent of each other, has higher accuracy
(97.98%), sensitivity (97.45%), and specificity (98.54%) as compared to the Naïve
Bayes and Random Forest algorithms [13].
Analysis of different algorithms like logistic regression, k-NN, decision tree,
random forest classifier, and SVM helps identify the appropriate technique to build a
heart disease prediction system. This system could, in turn, aid the healthcare industry
make predictions with a low error rate and high accuracy. The collected data was
split into a training and testing set, with 90% of the data used for training and 10%
for testing. SVM and the Random Forest classifier received a testing accuracy of
90.32% each, which proved to be the best among the algorithms under analysis.
K-NN acquired the least testing accuracy of the lot (70.96%) [14].
Analysis of algorithms like k-NN, K-means clustering, neural networks, Naïve
Bayes, logistic regression, and hybrid techniques like K-means clustering with Naïve
Bayes classifier or Fuzzy k-NN help determine which of these machine learning
algorithms would be the most helpful in predicting the presence of cardiovascular
disease in a person. These algorithms are applied and tested on the Cleveland heart
disease dataset. Experimental results show that hybrid approaches such as neural
networks and Fuzzy k-NN performed better than the other algorithms, with neural
networks resulting in an accuracy of 98% and Fuzzy k-NN with an accuracy of
94.19% [15].
Diabetes is a condition/disease in which the blood sugar levels rise above the
ordinary level, and if not treated in time, it can prove fatal. Machine learning can
help with the early prediction of this disease. Naïve Bayes, k-NN, decision tree clas-
sifier, SVM, and random forest are the five supervised machine learning algorithms
applied on the Pima Indians Diabetes dataset. The experiment involves five phases:
Dataset collection, data preprocessing, setting classification metrics, applying ML
classifiers, and K-fold cross validation. The dataset includes eight features. In the data
preprocessing step, the missing values in the dataset are filled in with the calculated
standard deviation of that particular feature. Precision, recall, F1 score, and accu-
racy are the classification metrics used. After performing K-fold cross-validation
(K = 10 in this case), the k-NN classifier emerged as the most accurate algorithm,
achieving an accuracy of 76%. SVM, Naïve Bayes, decision tree, and random forest
also achieved accuracies above 70%, but k-NN received the highest [16].
Data mining and machine learning help analyze large data sets and can hence
be used in the medical field to predict diseases like heart disease. To perform a
comparative analysis to conclude which algorithm would be the best choice for a
prediction system, a study of supervised machine learning models like Naïve Bayes,
logistic regression, SVM, k-NN, random forest, decision tree, and ensemble machine
learning technique XGBoost are implemented and tested on the Cleveland heart
disease dataset. The dataset was divided into a training set (80%) and a testing set
(20%). Each classifier was trained using the training set and their efficiency was tested
using the testing set. Random forest obtained an accuracy of 86.89%, followed by
XGBoost with 78.69%. These proved to be the most accurate, whereas k-NN achieved
an accuracy of only 57.83% [17].
Improved classification and regression systems provide valuable data that help
medical personnel in the early detection of diseases which, in turn, improves the
chances of survival of patients. Classification algorithms are applied on three separate
disease databases (heart, diabetes, and breast cancer). Any kind of missing data is
replaced with the mean value (for continuous variables) or the mode value (for
categorical variables). Backward elimination (p-value test) accomplishes the task
of feature selection. For the breast cancer dataset, the ideal algorithm is AdaBoost,
with an accuracy of 98.57%. SVM (linear kernel) achieved the highest accuracy
of 85.71%, while SVM (radial basis function) attained the lowest (66.23%) when
applied on the Pima Indians Diabetes dataset. Logistic regression scored the highest
accuracy (87.10%) and SVM (RBF) the worst (54.84%) when applied to the heart
disease dataset [18].
The Naïve Bayes classification model forms the basis of an application for
predicting heart disease in a patient. Attributes such as the patient’s age, sex, BP,
etc., fed to the Naïve Bayes classifier. About 80% of the collected data (obtained
from the UCI repository) was used for training, and the remaining 20% for testing.
An encryption algorithm (AES) is also used to securely transfer patient records to the
database. The Naïve Bayesian technique of classification achieved a greater accuracy
(89.77%) and lesser build time (0.01 s) in comparison to other methods like Bayes
Net, MLP, and sequential minimal optimization (SMO) [19].
Diabetes is predictable with the help of supervised machine learning models like
SVM and random forest. These models are tested on the dataset from Security Force
Primary Health Care, Tabuk. The performance of the algorithms was deduced from
observing the ROC curve and the confusion matrix for each algorithm. SVM emerged
with a prediction accuracy of 97%, while random forest performed better and scored
an accuracy of 98.9% [20].
3 Dataset Description
This project utilizes the Cleveland heart disease dataset, which has been taken from
the UCI (UC Irvine) machine learning repository. The dataset comprises 303 rows
and 14 columns or attributes (Figs. 1, 2 and 3).
4 Methodology
This study analyses four machine learning algorithms:

1. Support vector machine (SVM)
2. Naïve Bayes
3. Logistic regression
4. Decision tree.
(1) Support Vector Machine (SVM)
It is a supervised machine learning technique utilized for both binary classification

(predicting a categorical value) and regression (the method of predicting a continuous
value) purposes. SVM works by mapping data points from a lower dimensionality
feature space to a higher dimensionality feature space to categorize the data points
easier. The main aim of an SVM is to find an optimal hyperplane that separates data
points into two discrete classes.
The plus-points of support vector machines are that they are accurate in high
dimensional spaces and are memory-efficient because they use a subset of the training
points in the decision function, called support vectors (these are data points closest
to the hyperplane, which are responsible for building the SVM as they influence the
orientation of the hyperplane).
The drawbacks of this algorithm are that they are prone to overfitting when the
number of features is more than the number of samples, and they are not very efficient
on large data sets.
The confusion matrix for the training data (Fig. 4) shows that of the 176 normal
patients, 162 (92.05%) were correctly classified, and only 14 (7.95%) were incor-
rectly classified. Of the 66 heart disease patients, 54 (81.82%) correct classifications
were made, and only 12 (18.18%) were incorrectly classified. The accuracy of the
SVM for the training data set is 89.26%.
In the confusion matrix for the testing data (Fig. 5), we see that of the 44 normal
patients, 38 (86.36%) were correctly classified, and only 6 (13.64%) were incorrectly
Fig. 1 Brief description of the dataset attributes

Fig. 2 Distribution of the target variable
Fig. 3 Distribution of heart

disease patients based on
gender
classified. Of the 17 heart disease patients, 14 (82.35%) correct classifications were

made, and only 3 (17.65%) were incorrectly classified. The accuracy of the SVM for
the testing dataset is 85.26%.
(2) Naïve Bayes
It is a supervised machine learning technique and is “Naïve” because it assumes that

each attribute/feature is independent of all others. It is a conditional probability model
that assumes that for all the instances of a given class, the features bear little to no
correlation with each other and that each attribute contributes equally to the outcome.
The Naïve Bayes classifier is built on Bayes Theorem. Real-time applications include
recommendation systems, spam filtering, and so on.
Fig. 4 Confusion matrix for the training data
Fig. 5 Confusion matrix for the testing data
The plus-points of Naïve Bayes classifiers are that they are suitable for high-
dimensional data like text since they are fast to train and use for prediction purposes.
Since it is fast, it can help save computational time. It is also ideal for applications
involving large datasets. Naïve Bayes assumes that each attribute/feature is inde-
pendent of all others, which is not always correct in real-time problems. Another
Fig. 6 Performance accuracy of the Naïve Bayes algorithm on the training data
drawback of this Naïve Bayes is that it assigns a zero probability to a categorical

variable present in the test dataset but absent in the training set.
were made, and only 9 (13.64%) were incorrectly classified. The accuracy of Naïve
Bayes for the training data set is 88.84%.
In the confusion matrix for the testing data, we see that of the 44 normal patients, 34
(77.27%) were correctly classified, and only 10 (22.73%) were incorrectly classified.
Of the 17 heart disease patients, 14 (82.35%) correct classifications were made, and
only 3 (17.65%) were incorrectly classified. The accuracy of Naïve Bayes for the
testing data set is 78.69% (Fig. 7).
(3) Logistic Regression
A classification technique for categorical variables, logistic regression requires the

explanatory variables to be continuous. In this machine learning method, one or more
independent variables predict the target/dependent variable. The logistic regression
model is useful when data is binary, linearly separable, and there is a requirement for
probabilistic results. Logistic regression is identical to linear regression, but instead
of predicting a discrete value, it predicts a categorical value. It cannot predict non-
linearly separable data.
were made, and only 17 (25.76%) were incorrectly classified. The accuracy of logistic
regression for the training data set is 88.02%.
Fig. 7 Performance accuracy of the Naïve Bayes algorithm on the testing data
Fig. 8 Representation of the number of correct and incorrect predictions made by the logistic
regression algorithm on the training dataset
patients, 39 (88.64%) were correctly classified, and only 5 (11.36%) were incorrectly
made, and only 3 (17.65%) were incorrectly classified. The accuracy of logistic
regression for the testing dataset is 86.89%.
(4) Decision Trees
As the name suggests, decision trees are “tree-shaped” algorithms/diagrams that lay
down a course of action. Each branch of the tree indicates a decision based on a test. A
Fig. 9 Representation of the number of correct and incorrect predictions made by the logistic
regression algorithm on the testing dataset
decision tree comprises a root node, internal nodes, and leaf nodes. They are used for
classification and regression purposes. The training dataset is input to the decision
tree, from where it splits into nodes. Hence, a feature is tested, and subsequent
branching occurs depending on the test outcome. Decision trees are relatively easy
to decipher and conceptualize. Moreover, they can be used with both numeric and
categorical data. However, decision trees are susceptible to overfitting.
patients, 176 (100%) were correctly classified, and none were incorrectly classified.
Of the 66 heart disease patients, 66 (100%) correct classifications were made, and
none were incorrectly classified. The accuracy of the decision tree algorithm for the
training data set is 100%.
patients, 33 (75%) were correctly classified, and only 11 (25%) were incorrectly
made, and only 7 (41.18%) were incorrectly classified. The accuracy of the decision
tree algorithm for the testing data set is 70.49%.
5 Results
We measured the testing and training accuracies for each of the algorithms imple-
mented to examine if the problems of underfitting or overfitting occur. If the training
accuracy is very high, the model is facing the problem of overfitting. On the other
hand, a high testing accuracy indicates model underfitting. Therefore, an ideal clas-
sifier is one with comparable, testing, and training accuracies. For the Naïve Bayes
and the decision classifiers, the training and testing accuracies differ by over 10%,
Fig. 10 Confusion matrix for the decision tree algorithm applied on the training dataset
Fig. 11 Confusion matrix for the decision tree algorithm applied on the testing data
with the training accuracy being much higher, indicating model overfitting. SVM
performed reasonably well. Logistic regression performed the best, with a testing
accuracy of 86.89% (Table 1).
Table 1 Training and testing

Machine learning Testing accuracy Training accuracy
accuracies
algorithm (%) (%)
(1) Support vector 85.26 89.26
machine
(2) Naïve Bayes 78.69 88.84
(3) Logistic 86.89 88.02
regression
(4) Decision tree 70.49 100
6 Conclusion
Machine learning can improve the accuracy of disease prediction exponentially. Early
disease detection is vital to improve the chances of survival of a person who would
otherwise succumb to a disease due to a delayed diagnosis. Algorithms like SVM,
Logistic Regression, Decision Trees, and Naïve Bayes help build effective prediction
systems like a heart disease prediction system. Cardiovascular diseases can affect all
age groups and, if not detected in time, can be fatal. In this study, Logistic Regression
has proven to be an ideal algorithm for a heart disease prediction system owing to
its high accuracy of 86.89%.
Future studies will address the efficiency and accuracy of these machine learning
algorithms on other data sets, such as the Wisconsin Breast Cancer dataset. Applying
the same machine learning techniques on different datasets will help formulate an
interesting comparative analysis among the different datasets used. Principal Compo-
nent Analysis (PCA), a dimensionality reduction algorithm, can also be applied to
large datasets to reduce the complexity level. By reducing the feature space of large
datasets, analysis of the features is more effective.
References
1. Marimuthu M, Deivarani S, Gayathri R (2019) Analysis of heart disease prediction using

various machine learning techniques. In: Advances in computerized analysis in clinical and
medical imaging, pp 157–168
2. Patel J, Upadhyay T, Patel S (2016) Heart disease prediction using machine learning and data
mining technique. Int J Comput Sci Commun 7(1):129–137
3. Dwivedi A (2016) Performance evaluation of different machine learning techniques for
prediction of heart disease. Neural Comput Appl 29(10):685–693
4. Malavika G (2020) Heart disease prediction using machine learning algorithms. Biosci
Biotechnol Res Commun 13(11):24–27
5. Sandhya Y (2020) Prediction of heart diseases using support vector machine. Int J Res Appl
Sci Eng Technol 8(2):126–135
6. Arabasadi Z, Alizadehsani R, Roshanzamir M, Moosaei H, Yarifard A (2017) Computer-aided
decision-making for heart disease detection using hybrid neural network-genetic algorithm.
Comput Methods Programs Biomed 141:19–26
7. Singh A, Thakur N, Sharma A (2016) A review of supervised machine learning algorithms.

In: 2016 3rd international conference on computing for sustainable global development
(INDIACom), pp 1310–1315
8. Akinnuwesi B, Macaulay B, Aribisala B (2020) Breast cancer risk assessment and early diag-
nosis using principal component analysis and support vector machine techniques. Inform Med
Unlocked 21:100459
9. Asri H, Mousannif H, Moatassime H, Noel T (2016) Using machine learning algorithms for
breast cancer risk prediction and diagnosis. Procedia Comput Sci 83:1064–1069
10. Nithya R, Ramyachitra D, Manikandan P (2016) An efficient Bayes classifiers algorithm on
10-fold cross-validation for heart disease dataset
11. Gavhane A, Kokkula G, Pandya I, Devadkar K (2018) Prediction of heart disease using
machine learning. In: 2018 second international conference on electronics, communication,
and aerospace technology (ICECA), pp 1275–1278. https://doi.org/10.1109/ICECA.2018.847
4922
12. Singh A, Kumar R (2020) Heart disease prediction using machine learning algorithms. In:
2020 international conference on electrical and electronics engineering (ICE3), pp 452–457.
https://doi.org/10.1109/ICE348803.2020.9122958
13. Rao J, Prasad R (2021) An enhanced novel dynamic data processing (ENDDP) algorithm for
predicting heart disease in machine learning. Int J Sci Res Comput Sci Eng Inf Technol 94–104
14. Bhunia PK, Debnath A, Mondal P, Monalisa DE, Ganguly K, Rakshit P (2021) Heart disease
prediction using machine learning. Int J Eng Res Technol (IJERT) NCETER—2021 09(11)
15. Yadav SS, Jadhav SM, Nagrale S, Patil N (2020) Application of machine learning for the
detection of heart disease. In: 2020 2nd international conference on innovative mechanisms
for industry applications (ICIMIA), pp 165–172. https://doi.org/10.1109/ICIMIA48430.2020.
9074954
16. Lyngdoh AC, Choudhury NA, Moulik S (2021) Diabetes disease prediction using machine
learning algorithms. In: 2020 IEEE-EMBS conference on biomedical engineering and sciences
(IECBES), pp 517–521. https://doi.org/10.1109/IECBES48179.2021.9398759
17. Anbuselvan P (2020) Heart disease prediction using machine learning techniques. Int J Eng
Res Technol (IJERT) 09(11)
18. Kohli PS, Arora S (2018) Application of machine learning in disease prediction. In: 2018
4th international conference on computing communication and automation (ICCCA), pp 1–4.
https://doi.org/10.1109/CCAA.2018.8777449
19. Repaka AN, Ravikanti SD, Franklin RG (2019) Design and implementing heart disease predic-
tion using Naïve Bayesian. In: 2019 3rd international conference on trends in electronics and
informatics (ICOEI), pp 292–297. https://doi.org/10.1109/ICOEI.2019.8862604
20. Alanazi AS, Mezher MA (2020) Using machine learning algorithms for prediction of diabetes
mellitus. In: 2020 international conference on computing and information technology (ICCIT-
1441), pp 1–3. https://doi.org/10.1109/ICCIT-144147971.2020.9213708
Design and Performance Evaluation
of a Simple Resistance-to-Digital
Converter for Tunneling
Magneto-Resistance-Based Angular
Position Sensor with 180° Range
Kishor Bhaskarrao Nandapurkar
Abstract Numerous applications require angular position measurement of a rotating

shaft up to 180° range. Some examples include engine throttle-valve, joint angle of
robotic arms, skid loaders, etc. Angle sensors based on tunneling magneto-resistance
(TMR) effect can be used for many such applications due to their features such as
compact and cost-effective design, low powered operation, and high sensitivity. The
resistances of TMR angle sensor vary as sine and cosine function of the shaft angle. In
this paper, a simple Resistance-to-Digital Converter for TMR angle sensor (RDCT)
is presented. The proposed RDCT processes the nonlinear natured TMR resistances
and render a linear digital output for 180° range. The digital indication of the shaft
angular position is achieved without using any explicit analog-to-digital converter
(ADC) and look-up table (LUT). The RDCT is based on standard dual-slope principle
and hence possesses all the merits of dual-slope-based measurement method. It can
also be employed to detect angular position of the target shaft rotating between 0
and 180° at slow speeds. The working principle of the RDCT is explained first in
a detailed manner. Later, the effects of practical parameters of TMR angle sensor
and the RDCT circuitry are analyzed. An effective offset-error reduction method is
then discussed that reduces the output error due to the mismatch in TMR resistances.
Extensive simulation studies are carried out to quantify the RDCT performance
under ideal as well as non-ideal conditions. The effectiveness of the offset-reduction
method is also proved in simulation studies.
Keywords Angle sensor · Dual-slope-based technique · Error analysis ·

Resistance-to-digital converter · Tunneling magneto-resistance
K. B. Nandapurkar (B)
Indian Institute of Technology (Indian School of Mines) Dhanbad, Dhanbad, Jharkhand, India
e-mail: kbn.197@gmail.com
https://doi.org/10.1007/978-981-19-6913-3_17
256 K. B. Nandapurkar
1 Introduction
Angular position of a rotating shaft is a variable of interest for fields such as manufac-
turing and process industries, automobiles, medicine, military, aerospace, and several
other. Some of the specific applications that require knowledge of angular position
of the target shaft include detection of position of knobs, rotary valves, solar panels,
vehicle engine throttle-valve, and motor shaft [1, 2]. In many of these applications,
the target shaft rotates for angles up to 180° [3, 4]. Different types of angle sensors
are commercially available that can be employed in such applications. These sensors
can be broadly categorized as resistive potentiometers, optical, capacitive, inductive,
hall-effect-based encoders [5]. Angle sensors based on tunneling magneto-resistance
(TMR) effect are relatively new in the market, and they have many advantages over the
aforementioned (conventional) angle sensors. The TMR angle sensors, as opposed to
resistive potentiometers, offer non-contact measurement. In contrast to RVDTs, they
are compact in design and consume less power. The output sensitivity of TMR angle
sensors is high when compared with hall-effect-based encoders [5]. The availability
of TMR angle sensors at low price, together with the advantages discussed above,
makes them a favorable choice for shaft angle sensing. The TMR-based angle sensors
yield two outputs which possess sine and cosine function dependency on the shaft
angular position. It is to be interestingly noted that resolvers and hall-effect based
encoders also possess such sine–cosine transfer characteristic [1, 5]. It is because of
the two quadrature-shifted sine–cosine outputs, and shaft angles up to 360° range
can be accurately estimated. However, such sine–cosine response of the sensor is
nonlinear. So, despite having several merits, such nonlinear response sometimes
makes an end user give a second thought on the selection of TMR angle sensor for
the intended application. This issue can be easily addressed by designing a suit-
able interfacing circuit for the TMR angle sensor that linearizes the sensor response
over the ‘desired measurement range’. It is to be noted that the phrase ‘desired
measurement range’ is very wisely used in the previous statement. This is because
the complexity of the interfacing circuit can be significantly reduced if the measure-
ment range is reduced to 180° from 360°, which, as mentioned earlier, is sufficient
for many applications. Some of these applications cover joint angle of robotic arms,
skid loaders, solar panels, operating tables and dental chairs in hospitals, etc. [3, 4,
6, 7]. Researchers have designed several linearizing interface circuits for TMR-like
sine–cosine angle sensors [8–13]. The features of these circuits are briefed next.
Lopez-Martin and Carlosena have discussed three analog and digital interfacing
circuits for GMR-based angle sensor in [8]. The maximum operating range achieved
in this work is 10–170°. These circuits are complex as they either use programmable
gain amplifier, encoder, multiplexers, and/or dedicated ADCs. A simple ADC-based
digitizer for GMR angle sensor is presented in [9]. However, the angular resolution
of the circuit is 1° which is quite poor. An interface circuit which requires flash ADC
with nonlinear and non-standard resistors has been designed for MR angle sensor
in [10]. Anoop and George have proposed an electronic scheme that estimates the
shaft angle from the outputs of MR angle sensor using inverse-cosine operation.
Design and Performance Evaluation of a Simple Resistance-to-Digital … 257
This circuit requires a costly instrumentation amplifier and a sine wave oscillator
for linearization and provides analog output [11]. A resolver-to-DC converter has
been developed and discussed by Benammar et al. [12]. This circuit is signifi-
cantly complex and employs costly analog multipliers for linearization. The direct-
digital interfacing circuits, described in [13, 14], also require a sine wave/quadrature
oscillator and instrumentation amplifier.
Thus, from the above brief discussion, it can be concluded that an efficient inter-
facing circuit is essential for TMR-like sine–cosine natured angle sensors. Consid-
ering the increasing demand for the digital measurement systems, the interfacing
circuit should render a linear and digital output, preferably, without using any dedi-
cated ADC(s) and LUTs. Such a digitizer should be simple-to-design, utilizing low-
cost electronic components. The use of costly components such as instrumenta-
tion amplifier and analog multiplier should be discouraged. In this paper, a simple
Resistance-to-Digital Converter for TMR angle sensor (RDCT) is presented that
achieves all the aforementioned features. The proposed RDCT is a simplified version
of the linearizing signal conditioner (LSC), presented in [15]. Similar to the LSC,
the RDCT employs a DC excitation for TMR angle sensor, and hence, the parasitic
capacitances of the sensor do not affect the measurement. The merits of the RDCT
over [15] are highlighted below.
1. The LSC processes both sine and cosine half-bridges of the TMR sensor. On
the other hand, the RDCT processes only cosine half-bridge and renders a linear
output for half-circle range. As clarified earlier, 180° range of measurement is
sufficient for many applications.
2. The LSC requires two integration phases and two de-integration phases for angle
estimation, and hence, it is slower in operation. The proposed RDCT, however,
requires one integration and one de-integration phase for the angle measurement.
Since the RDCT measurement time is faster, intuitively, it can track dynamic
inputs with a faster speed compared with the LSC.
3. The RDCT requires one switch and one control signal less compared to that of
LSC.
The next section covers the basic principle of the TMR angle sensor and the
working of the proposed RDCT. In the subsequent sections, factors affecting the
RDCT performance and the verification of RDCT working are covered.
2 Resistance-to-Digital Converter for TMR Angle Sensor
The working principle of TMR angle sensor is explained first before heading toward
the RDCT.
2.1 TMR Angle Sensor—Construction, Operating Principle,

and Transfer Characteristic
The TMR-based shaft angle sensing setup is shown in Fig. 1a. This sensing arrange-
ment demands a diametrically magnetized permanent magnet (PM) to be attached to
the rotating shaft. Such a PM produces magnetic field whose plane will be in parallel
to the plane of the PM. The TMR angle sensor, which is available in a compact
IC form [16], is placed below the PM. The TMR angle sensor requires saturating
magnetic field of specific values for its proper operation. The range of this field
depends on the manufacturer. For an instance, the AAT00X series of half-bridge
TMR angle sensor, from NVE Corporation, requires magnetic field between 30 and
200 G [16]. This can be easily achieved by carefully adjusting the air gap between
the TMR IC and the PM. Once the PM and the TMR IC are positioned properly, the
sensing setup is ready to be used for angle measurement purpose. From the figure
and the above explanation, it is clear that the TMR-based sensing arrangement is
non-contact type and, hence, offer long life sensing solution. It is to be noted that
this sensing arrangement need not be vertical, as shown in Fig. 1a. In applications
such as motor shaft positioning, camera surveillance, the target shaft may be orien-
tated in any direction. The TMR-based arrangement can be used in such scenarios
as well.
Commercially available TMR angle sensors are categorized as full-bridge type
and half-bridge type. The full-bridge TMR sensors have eight ‘active’ TMR elements,
embedded in them. On the other hand, as shown in Fig. 1b, a half-bridge-type TMR
angle sensor internally consists of four active resistances. The word active in the
previous two statements signifies that the resistance varies with the change in the
direction of the biasing magnetic field of the PM. The proposed RDCT essentially
requires a part of half-bridge TMR angle sensor for accurate shaft angle measurement.
Fig. 1 a TMR-based shaft angle sensing setup. b Internal circuit diagram of a half-bridge TMR
angle sensor, comprising four TMR elements: RC1 , RC2 , RS1 , and RS2 . c TMR resistances vs shaft
angle characteristic. Four TMR resistances have sine–cosine natured relationship with the shaft
angle, θ
As shown in Fig. 1b, the resistances RS1 , RS2 , RC1 , and RC2 form half-bridge TMR
angle sensor.
To understand the working principle of TMR angle sensor, let us first discuss
the structure of individual TMR resistances. A single TMR element (say, RS1 ) is
a three-layered structure. The top layer is formed using a ferromagnetic material,
and hence, its magnetization can align with the direction of magnetic field of the
PM. The bottom layer is commonly called as pinned layer due to the fact that its
magnetization is fixed or pinned in a particular direction [16]. These two layers are
separated by an ultra-thin insulator layer (usually made of Al2 O3 or MgO). The
nominal device resistance is mainly governed by the thickness of this insulator layer.
For angle sensing applications, as shown in Fig. 1a, when the shaft rotates by some
angle (say, θ ), the PM also rotates by the same angle. Thus, there is a rotation of the
magnetic field ‘seen’ by the TMR sensor. The latter is powered using AC/DC voltage
or current source of suitable magnitude. The resistance of the TMR element changes
when the direction of the magnetic field changes. This is because of the change in
the relative angle between the magnetization of free layer and pinned layer. For an
instance, when the magnetizations of both the layers are parallel, then electrons are
easily tunneled. This results in minimum resistance. On the contrary, when the two
layers have their respective magnetization antiparallel, the device resistance is the
largest.
Since the resistance varies with the external magnetic field, this effect is known
as ‘magneto-resistance’ effect. In particular to the TMR sensors, due to the presence
of ultra-thin insulating layer, the electrons literally tunnel from the bottom layer to
the top layer (or, vice versa). This is the reason of calling this overall phenomenon
as ‘tunneling magneto-resistance’. More details on the fabrication and working of
the TMR angle sensors can be found in [16–18]. The physical arrangement of four
TMR elements RS1 , RS2 , RC1 , and RC2 , within the sensor, is different, and hence,
their response to the change in the direction of magnetic field is also different. The
mathematical expressions of these resistances are given in (1).
RS1 = RN − ΔR sin θ ; RS2 = RN + ΔR sin θ ;

(1)
RC1 = RN − ΔR cos θ ; RC2 = RN + ΔR cos θ
In the above expressions, RN stands for the nominal resistance of the TMR element
and ΔR is the maximum change in the device resistance. Typically, RN is significantly
greater than ΔR. For an ideal TMR angle sensor, the value of RN for all four resistances
is equal. Also, ΔR is same for all four elements. Figure 1c shows the nature of
variation of resistances of half-bridge TMR angle sensor with respect to the shaft
angle, θ. From (1) and the graph, it is clear that the resistances possess unique sine and
cosine relationship with θ. Of these four resistances, RS1 and RS2 form half-bridge
HBS and RC1 and RC2 form HBC . This can be better visualized in Fig. 1b. From
(1), we can also say that RS1 and RS2 (and RC1 and RC2 ) are arranged in push–pull
configuration. Such arrangement keeps the total half-bridge resistance (RS1 + RS2
and RC1 + RC2 ) constant and independent of the sensing angle, θ.
From the sensor characteristic, shown in Fig. 1c, it is straightforward that the
resistances RC1 and RC2 of HBC are unique for every value of θ ∈ (0◦ , 180◦ ). It
is due to this property, the shaft angle, θ, can be accurately determined if just the
RC1 and RC2 are processed using a suitable interfacing circuit. The proposed RDCT
achieves this objective. Please note that RS1 and RS2 of HBS have distinct values for
θ ∈ (90◦ , 270◦ ). The RDCT can process these resistances as well to determine θ.
However, an offset of 90° is required in the sensing unit, i.e., between the sensor and
the PM to render a linear indication from 0 to 180°. Since the measurement accuracy
is independent of choice of resistances, HBC is chosen to avoid need of mechanical
offset.
Let us now slightly digress from the concept of processing the resistance variations
to measure the shaft angle. The half-bridge-based TMR angle sensor, when powered
using a suitable excitation voltage (say, V EX ), renders two sine–cosine natured
outputs. These outputs (say, vSIN and vCOS ) can be given as vSIN = (0.5 + k sin θ ) ×
VEX and vCOS = (0.5 + k cos θ ) × VEX . Here, the term k = ΔR/(2RN ) refers to the
transformation constant of the TMR sensor. The typical value of k is 80 mV/V. Thus,
we can say that the half-bridge sensor outputs contain an offset term (= 0.5V EX )
which is significantly larger than the modulating term ‘k cos θ ’ or ‘k sin θ ’. In
order to extract a linear output from such sine–cosine natured voltages, the offset
term has to be eliminated. Later, the offset-free signals can be suitably processed by
ADC and a processor for linearization and digitization. However, please note that
the offset-free signals are small in magnitude and bipolar in nature. The strength
of the signals can be enhanced using additional amplifier stage. However, the ADC
must be bipolar to accept such signals or suitable demodulation stage is required.
Considering factors of the ADC such as resolution, gain and offset error, and differ-
ential and integral nonlinearity, the ADC-based angle measurement system may not
be the best and cost-effective choice. Therefore, an alternative to such ADC-based
digital angle measurement system must be searched. The proposed RDCT targets
this particular issue efficiently. Having discussed about the TMR angle sensor in
detail and the limitations of the ADC-based measurement, let us now discuss the
working principle of the proposed RDCT in an elaborated manner.
2.2 Working of RDCT
The simplified schematic of the RDCT is shown in Fig. 2. As mentioned earlier,

the RDCT, unlike many dual-slope-based digitizers, requires a single DC reference
voltage (here, marked as V D ) for its operation. The RDCT is designed using three
operational amplifiers: OP1 , OP2 , and OC, one single-pole double-throw switch SW,
and a control and logic unit (CLU). Any pre-programmable microcontroller with
sufficient number of digital input/output pins can be used as the CLU. The op-amp
OP1 together with TMR resistances RC1 and RC2 forms an inverting amplifier. This
inverting amplifier is supplied with V D as its input signal. The integrator circuit is
designed using OP2 and passive components RI and C I . The selection of RI and C I
Fig. 2 Simplified schematic of the proposed RDCT. The half-bridge TMR sensor resistances RC1
and RC2 are processed to obtain linear digital indication for 180° range
is covered in the performance verification section. The op-amp OC is configured

as a comparator to facilitate digital measurement. The components SW, integrator
circuit, and comparator OC form the base for dual-slope-based measurement.
As shown in Fig. 2, the inverting amplifier output, vS = −(RC2 /RC1 ) × VD is fed
to position-0 of SW. The expressions of RC1 and RC2 are already given in (1). Since
the value of RN is larger than ΔR, vS is negative irrespective of value of θ. The signal
V D is applied at position-1 of SW. The SW position is controlled using the digital
signal, DC1 , obtained from the digital port of the CLU. Note that the signal vS acts
as a measuring input for the dual-slope-based circuit and the voltage V D acts as a
reference signal.
A typical measurement cycle of the RDCT consists of an integration phase which
runs for a preset period of time (say, T I ) and a variable de-integration phase. The time
duration (say, T V ) of this time-variable phase is dependent on the input signal, vS ,
which in turn depends on θ. The operation of the RDCT is explained assuming that
the shaft is at a steady position for at least one measurement cycle. Before the start
of the measurement cycle, it is important to ensure that the integrator output, vI , is
zero. This is done to avoid saturation of OP2 which will result in huge measurement
error. The process of making vI = 0 is commonly called as ‘auto-zeroing’ and can be
executed easily using simple CLU-based programming. The auto-zeroing step does
not require additional components. Also, it is to be noted that the auto-zeroing phase
is required only once after powering on the RDCT. Once vI is ensured to be zero,
RDCT is ready for the measurement. Initially, the CLU sets DC1 = 0 to feed the
signal vS to the integrator circuit, as shown in Fig. 3a. This signal, being constant,
causes the integrator output to increase linearly with time, t as can be seen in Fig. 3b.
The value of signal vI at the end of the integration phase, i.e., at time t = T I , can be
expressed using (2).
Fig. 3 Important waveforms of the RDCT. The integrator input and output are shown in a and b,
respectively. The digital output, DP , of OC is shown in c. Here two measurement cycles are shown
for better understanding. The input angle can be estimated using one integration phase (T I ) and one
de-integration phase (T V ) alone
{TI {TI ( )
1 1 RC2
vI (TI ) = − vS · dt = − − × VR · dt
RI C I RI C I RC1
0 0 (2)
( )
1 RN + ΔR cos θ
vI (TI ) = × × VR × TI
RI C I RN − ΔR cos θ
Once the integration phase is over, the CLU executes the de-integration phase. For
this purpose, it places SW at position-1 by making DC1 = 1. The reference signal V D
is fed to the integrator. The output of latter, as shown in Fig. 3b, starts falling linearly.
The signal vI is supplied to the non-inverting terminal of OC. The latter’s inverting
terminal is grounded. In the absence of feedback, for such a connection, OC acts as
zero-crossing detector. Thus, the OC output, DP , is digital in nature. For vI ≥ 0, DP
= 1, otherwise 0. In the de-integration phase, when the decreasing vI crosses zero
(vide Fig. 3b), the signal DP changes its state from logic HIGH (or, 1) to logic LOW
(or, 0), as can be seen in Fig. 3c. This signal is fed to the digital input pin of the
CLU. The state change of DP is detected by the CLU, and this marks the end of the
de-integration phase. Note that, as shown in Fig. 3, at the end of the de-integration
phase, vI = 0. Thus, the charge-balance equation can be given as per (3).
( )
1 RN + ΔR cos θ VR
× × VR × TI − × TV = 0 (3)
RI C I RN − ΔR cos θ RI C I
The time period T V of the de-integration phase is measured by the CLU with the
help of pulse DP . The mathematical expression of T V can be obtained from (3) as
per (4).
( )
RN + ΔR cos θ TV RN + ΔR cos θ
TV = × TI which yields = (4)
RN − ΔR cos θ TI RN − ΔR cos θ
The CLU then performs a simple ratiometric operation as per the following
expression to obtain a variable x.
TV − TI RN + ΔR cos θ − RN + ΔR cos θ ΔR
x= = = × cos θ (5)
TV + TI RN + ΔR cos θ + RN − ΔR cos θ RN
Thus, it can be said that processing the time periods, as per (5), is equivalent of
processing the resistances RC1 and RC2 of the TMR angle sensor. The variable x, so
obtained, varies as a cosine function of θ, which is again nonlinear in nature. The
simplest method of linearizing any nonlinear function is to perform inverse of the
nonlinear operation. Thus, here, inverse-cosine operation is suitably performed to
obtain a linear indication (say, Dθ ) of θ for θ ∈ (0◦ , 180◦ ). This transformation is
defined in (6).
( )
RN
Dθ = cos−1 ×x = θ for θ ∈ (0◦ , 180◦ ) (6)
ΔR
Therefore, the proposed RDCT assists to obtain a linear estimate of shaft angular
position for half-circle range. Please note that instead of inverse-cosine transfor-
mation, any sufficiently accurate approximation function can also be used to fetch
linear output. This selection will depend on the computation time of the CLU and
the accuracy of the approximation function. As mentioned in the previous section,
the RDCT does not require costly instrumentation amplifier for the offset reduction
and/or linearization. It does not use dedicated ADCs and LUTs for the output digi-
tization. The DC excitation for the sensor ensures that the output is least influenced
by the parasitic capacitances of the TMR sensor. The conversion time (say, T CONV )
of RDCT is given by T CONV = T I + T V . The maximum value of T CONV can be
expressed as per (7).
( ) ( )
2RN 2RN
max(TCONV ) = TI + TV = TI × = TI ×
RN − ΔR cos θ θ =0◦ RN − ΔR
(7)
Thus, for a given TMR sensor, T CONV can be tuned by adjusting T I alone. Please
note that a large value of T I will offer very high resolution. On the other hand, lower
value of T I will facilitate dynamic input measurement. Therefore, the integration
time should be chosen as per the application demand and the time resolution of the
CLU employed in the RDCT. Next paragraph summarizes the working of the RDCT.
The RDCT, during its integration phase, processes signal vS for T I s. This signal
is dependent on the values of RC1 and RC2 and hence on θ. Later, the de-integration
phase is executed, and the time period T V is measured by the CLU. The ratiometric
operation, as given in (5), is then performed to obtain variable x. The linear indication
of θ is finally obtained using the transformation defined in (6). As explained earlier,

the linear response of the RDCT, obtained for half-circle range, is sufficient for many
applications such as engine throttle-valve, accelerator-pedal position, joint angle of
robotic arms, and skid loaders [2–7]. Therefore, the RDCT can be well used in such
applications.
3 Error Analysis of the RDCT
The working of the RDCT and the measurement methodology assumes that both
the TMR angle sensor and the electronic components employed in RDCT are ideal
in nature. However, they may not behave ideally always. The effect of non-ideal
parameters of sensor and circuit components is analyzed and presented in this section.
3.1 Error Sources of TMR Angle Sensor
Mismatch in ΔR of RC1 and RC2 : Let δ be the mismatch between ΔR of RC1 and RC2
such that ΔRC2 = (1 + δ) × ΔRC1 . Here, ΔRC1 and ΔRC2 are the maximal change
in resistance of RC1 and RC2 , respectively. For a nonzero value of δ, the variable x
gets modified to x δ as per (8).
xδ = [ΔRC1 (2 + δ) cos θ ]/[2RN + δ ΔRC1 cos θ ] (8)
Since x /= x δ , the output angle, Dθ , gets affected for a nonzero δ. The quantity Dθ
was mathematically computed using (6) and (8) for every value of θ ∈ (0◦ , 180◦ ).
For δ = 0.05% (or, 0.0005), the maximum error in the measurement of shaft angle is
close to 0.78%, and it occurs at θ = 180°. For the range 5–175°, this error always stays
within 0.1%. It is to be noted that several industrial applications do not have tight
requirements on the accuracy. In such cases, the proposed low-cost, less-complex
RDCT is certainly a good choice.
Mismatch in RN of RC1 and RC2 : Similar to the above non-ideality, the mismatch
in RN value of two resistances can be modeled as follows: RNC2 = (1 + τ ) × RNC1 .
Here, τ represents the mismatch extent. The new ratiometric quantity, x τ , in this
case, can be described as per (9). Here, the term ‘τ × RNC1 ’ can be considered as an
undesired offset.
2ΔR cos θ + τ RNC1
xτ = (9)
(2 + τ )RN
For τ = 0.05%, the worst-case error in the measurement of θ for half-circle range
is 1.5%, whereas, for 5–175° range, this error is 0.56%. Thus, this non-ideality can
be considered as a major error source. The effect of nonzero τ can be reduced using
a simple offset-error reduction technique, described next.
Let us perform a new ratiometric operation as described using the following
equation.
TV − (1 + τ )TI (2 + τ )ΔR cos θ

x, = =
TV + (1 + τ )TI 2 × (1 + τ )RNC1 − τ × ΔR cos θ
∼ (2 + τ )ΔR
= × cos θ (10)
2 × (1 + τ )RNC1
It can be seen that the newly obtained x , does not contain the offset term ‘τ × RNC1 ’
as in (9). The quantity x , can be easily realized in the CLU as the value of τ can be
found from the sensor characteristic. The maximum error in Dθ was computed using
x , for τ = 0.05%, and it was 0.72% for 180° range as opposed to 1.5% obtained using
x τ . For the range of 5–175°, this error has also been reduced to 0.09% from 0.56%.
Thus, it can be concluded that the offset-reduction technique is notably effective
to minimize the output error due to nonzero τ. It is very important to note that the
mismatch extent remains constant even when the surrounding temperature fluctuates.
This is because both RC1 and RC2 increase or decrease with respect to the change in
temperature equally. Thus, the proposed offset-error reduction technique needs no
modification when the RDCT is employed in temperature fluctuating environments.
Phase imbalance in RC1 and RC2 : The resistances RC1 and RC2 with a phase imbal-
ance (say, β) can be given as RC1 = RN − ΔR cos θ and RC2 = RN + ΔR cos (θ +
β). This factor also affects the ratio x. The modified expression of x β is given in (11).
ΔR × [cos θ + cos(θ + β)]

xβ = (11)
2RN + ΔR × [cos(θ + β) − cos θ ]
The typical value of β is 0.2°, which results in the worst-case error of 0.084% for
0–180° range and 0.064% for 5–175° range.
Non-sinusoidal nature of resistance variations: Apart from the above-discussed
non-idealities, the TMR resistances may not follow the expected sine–cosine natured
behavior. The degree of non-sinusoidality is usually quantified in terms of percentage
root mean square error (RMSE) [16]. The percentage RMSE can be computed
from the sensor characteristic and curve-fitting technique. The effect of such non-
sinusoidal nature of TMR sensor can then be found using a suitable simulation tool.
The exact reason for the non-sinusoidal behavior of the sensor may not be known.
Therefore, the error arising due to this non-ideality may not be fully nullified.
3.2 Error Sources of Circuit Components of RDCT
Offset voltage and bias current of OP1 : The offset voltage (say, vOF1 ) and bias
current (say, ib1 ) of inverting amplifier op-amp OP1 will alter its output, vS , as per
(12).
( )
RC2 RC2
vS, =− × VD + vOF1 × 1 + − i b1 × RC2 (12)
RC1 RC1
The typical value of vOF1 and ib1 is around 60 µV and 2.5 nA, respectively. For
these values, vS , was mathematically computed for every θ ∈ (0◦ , 180◦ ). Then, the
intermediate ratio x and the final output, Dθ , were calculated. The maximum error
observed for 180° range is 0.62%, whereas it is 0.07% for 5–175° range.
ON resistance of SW: The ON resistance of SW (say, RON ) increases the input
resistance of the integrator to (RI + RON ). This modified resistance remains same for
both integration and de-integration phase. Therefore, this non-ideality does not affect
the variable T V and hence the RDCT output, Dθ . This fact is further established by
carrying out detailed simulation studies for different values of RON . The results will
be discussed in the next section.
Bias current of OP2 : In the proposed RDCT, OP1 and OP2 are realized using two ICs
of same manufacturer. This suggests that the bias current of OP2 is also in nanoampere
range. The modified charge-balance equation for this non-ideality is given using the
following expression.
[( ) [ [ [
RN + ΔR cos θ i b2 TI VR i b2
× VR − × − + × TV = 0 (13)
RN − ΔR cos θ CI RI C I RI C I CI
For ib2 = ib1 = 2.5 nA, the intermediate ratio x and the final output Dθ are computed
using the suitable equations. The maximum error spotted for this non-ideality is
0.19% for half-circle range and 0.007% for 5–175° range.
Offset voltage of OC: The nonzero offset voltage (say, vOF2 ) of OC affects its zero-
crossing operation. The comparator OC will change its state from logic zero to logic
one (or vice versa) when its input (vI ) crosses vOF2 (instead of 0 V). Thus, this non-
ideality will only result in a vertical shift of integrator output waveform (vide Fig. 3b)
by vOF2 . The typical value of vOF2 is in millivolts. Thus, the vertical shift does not
cause saturation of OP2 output and hence do not affect the angle measurement.
Errors due to the time delay of SW, comparator response time, and the CLU
accuracy: In the RDCT, SW can be realized with the help of a commercially available
CMOS switch. These switches typically have a delay of 10 ns, and their response
time is around 60 ns. Similarly, comparator also takes on an average 170 ns of time to
respond to the step change in its input. On the other hand, the RDCT conversion time
is in milliseconds. Therefore, the effects of these two parameters can be conveniently
ignored. However, the time resolution of the CLU plays a crucial role in deciding the
integration time (T I ) which in turn governs the maximum speed limit for dynamic
input tracking. A CLU with time resolution of 1 µs is a good choice in terms of its
cost and dynamic input tracking capability.
Other non-idealities: The passive components (RI and C I ) of the integrator may drift
with time and/or temperature. However, such a drift does not affect the measurement
as both the integration and de-integration phases are equally influenced. The assump-
tion that drift in RI and C I is rapid and is in comparison with the conversion time is
very impractical. Therefore, we can confidently claim that variations in RI and C I do
not affect the measurement. Likewise, slow and minor variations in V D do not intro-
duce any error in the RDCT output. Apart from all these error sources, the RDCT
performance may get affected due to some random effects. The effect of random
errors can be minimized by taking average of multiple readings of Dθ .
4 Extensive Performance Evaluation of the RDCT
In the previous two sections, the working principle of TMR angle sensor and the
proposed RDCT, and detailed error analysis have been discussed. In this section,
simulation studies for ideal as well as various practical/non-ideal cases are described.
4.1 Simulation Studies for Ideal TMR Angle Sensor

and Ideal RDCT
The performance of the proposed RDCT was studied in simulation using the
LTSPICE software from Linear Technology Corporation. The TMR resistances RC1
and RC2 were made to vary as per (1) so as to mimic the shaft rotation for half-circle
range. This was achieved using SPICE programming. The values of RN and ΔR
were chosen as 26 kΩ and 4.5 kΩ, respectively, which are true for the AAT003-10E
half-bridge TMR IC [16]. Note that the RDCT will work equally well for any other
TMR angle sensor with different value of RN and ΔR [19]. The components OP1 and
OP2 were realized using the SPICE model for the OP07 IC from Texas Instruments.
The switch SW was built in a similar way using equivalent model of MAX4053
IC. The value of RI and C I was chosen as 22 kΩ and 470 nF, respectively. These
values ensure that the integrator output, vI , does not reach saturation when T I is set
as 20 ms. As mentioned earlier, the time T I should be chosen sufficiently small in the
case of dynamic input tracking. Therefore, the values of RI and C I must be selected
such that the saturation of OP2 output is avoided as well as the de-integration time is
measured accurately. The reference voltage V D is set as 2 V so as to restrict the net
excitation voltage for the TMR sensor (V D + vS ) within the specified limit [16]. The
programming part of the RDCT was achieved with the help of a precision monostable
Fig. 4 RDCT response obtained in LTSPICE-based simulation studies for ideal TMR angle sensor.
As desired, the output angle varies linearly with the shaft angle, θ, for half-circle range
multivibrator that provides a pulse of 20 ms duration for the integration phase. The
measurement of de-integration time T V was done using .meas command in LTSPICE.
For the ideal simulation, different non-idealities, discussed in the previous section,
were set to their respective ideal values. The resistances RC1 and RC2 were varied to
mimic the shaft rotation over half-circle range, in steps of 3°, and the variable T V was
recorded, for each step. Later, the intermediate variable, x and the output angle, Dθ
were computed offline using (5) and (6), respectively. The obtained Dθ was plotted
against θ and is shown in Fig. 4. As can be seen in this graph, Dθ varies linearly
with θ for 180° range. The error in Dθ , for each step of θ, is shown in Fig. 4 using
green-colored scatter plot. The maximum percentage error spotted in Dθ is 0.34%,
which amounts to an angular error of 0.62°. For 5–175° range, the worst-case error
in was found to be just 0.042% (= 0.08°). The high output error at θ = 0 and 180° is
due to the combined effect of undesired mismatch in actual T I and expected T I and
various parameters of RDCT which could not be set as ideal due to limitations.
4.2 Simulation Studies for a Practical TMR Angle Sensor
In this part of the study, the TMR resistances RC1 and RC2 were modeled to have
mismatch δ = τ = +0.05% and phase imbalance, β = +0.2°. Keeping all the other
parameters same as that for the previous case, the simulation program was run and
new T V , was recorded for every step of θ. The output angle was estimated from
T V , . Results obtained in this part of the study are shown in Fig. 5a. It can be seen
that the RDCT output, Dθ , fairly varies linearly with θ. However, the maximum
percentage error for such a practical case is increased to 1.8% for 180° range. On
the other hand, for 5–175° range, the maximum error is 0.49%. The offset-reduction
technique, discussed in the error analysis section, was then applied and the output
angle, Dθ , was estimated from x , . As expected, the output error is reduced greatly
Fig. 5 RDCT response for a practical TMR angle sensor with three non-idealities is shown in
a. The output angle has significant error, mainly due to mismatch in RN . After implementing the
offset-reduction strategy, this error, as shown in b, gets reduced significantly
to 0.41% for half-circle range. This proves the efficacy of the offset-compensation
technique in simulation environment. Later, the effect of different RON of SW was
also studied. The RON was varied from 50 to 300 Ω, in steps of 50 Ω, and the worst-
case error in Dθ was found for each case. The maximum error was observed quite
close for all the cases considered. Thus, as expected, RON was found to affect the
output negligibly.
Based on these simulation studies, it can be commented that the proposed RDCT
indeed works as a resistance-to-digital converter for TMR angle sensor. The RDCT,
as highlighted earlier, does not require costly instrumentation amplifiers, sine wave
oscillator, and two reference voltage supplies for its operation. In real-time environ-
ment, when realized using suitable components, the RDCT will yield linear direct-
digital output for 180° range, without utilizing any separate ADC(s) and LUTs.
The approximate component cost of RDCT (TMR angle sensor + RDCT circuit
components) is just 15 USD. Thus, a low-cost angle sensing solution with sufficient
accuracy is provided in this work for wide range of applications.
5 Conclusions and Future Work Plan
In this work, a simple RDCT, suitable for measuring shaft angle over half-circle
range, has been presented. The RDCT, in contrast to several existing interface circuits
[8–15], is much simpler in design. The RDCT was designed with an objective of
realizing an ADC-free, low-cost angle measurement system. This objective has been
successfully achieved. Factors affecting the RDCT performance have been identified
and analyzed. A simple and efficient offset-error reduction method was presented that
drastically reduces the output error due to the mismatch in the nominal resistances
of the TMR angle sensor. The performance of the RDCT was evaluated for ideal as
well as non-ideal TMR angle sensor using detailed simulation studies. The results
obtained in these studies showcase the capability of RDCT to linearize and digitize
the response of an inherently sine–cosine and hence, nonlinear natured response
of the TMR angle sensor. The RDCT can be used in numerous applications such
as tracking of solar panels, positioning of engine throttle-valve, accelerator-pedal,
operating tables, and dental chairs in the hospitals.
The future work is planned in the following manner. The experimental evaluation
of the RDCT will be done to verify the methodology in real-time scenario. For this
purpose, the TMR-based shaft angle sensing unit needs to be built. With the sensing
setup, we plan to do several experiments for static as well as dynamic measurement.
Since the RDCT is basically a digital instrumentation system, parameters such as
signal-to-noise ratio (SNR), angular resolution, standard deviation, effective number
of bits (ENoB), and repeatability error can be estimated. Note that these parameters
can be suitably adjusted by tuning the integration time (T I ) of the RDCT carefully. It
will also be interesting to verify the efficiency of the offset-error reduction technique
in actual experimentation with the TMR-based sensing setup. The findings of these
studies will be reported later.
References
1. Doebelin EO (2013) Measurement systems: application and design, 6th edn. McGraw Hill
Publishing Co., New Delhi
2. The sixth sense for automotive applications. http://www.nxp.com/documents/brochure/750
15728.pdf. Accessed 05 Aug 2021
3. Zhang J, Hua S, Qi Z, Feng W (2011) The design and implementation of the shaft angle
acquisition system used in the solar panel. In: Fourth international symposium on computational
intelligence and design. IEEE, Hangzhou, pp 287–290
4. Robot sensor and actuators. http://robotics.sjtu.edu.cn/upload/course/5/files/Robot%20Sens
ors%20and%20Actuators-new.pdf. Accessed 25 July 2021
5. Pallas-Arney R, Webster JG (2012) Sensors and signal conditioning, 2nd edn. Wiley, New
Delhi
6. Arc sensor applications. https://cambridgeic.com/applications/arc-sensor-applications.
Accessed 12 June 2021
7. Application note—transportation. https://sensing.honeywell.com/honeywell-value-added-
smart-rotary-boom-sprayer-arm-angle-position-sensing-000757-1en.pdf. Accessed 02 Aug
2021
8. Lopez-Martin AJ, Carlosena A (2009) Performance tradeoffs of three novel GMR contactless
angle detectors. IEEE Sens J 9(3):191–198
9. Farcas C, Cosma A, Palaghitia N, Grama A, Gabor MS, Tiusan C (2014) Development of a non-
contact angular transducer. In: International symposium on design technology and electronic
packaging. IEEE, pp 187–191
10. Leitis K, Bonath W (2002) Magnetoresistive sensors and a new hardware-based interpolation
method for length and angle measurements. In: Sensors conference. IEEE, Orlando, pp 1432–
1435
11. Anoop CS, George B (2012) Electronic scheme for computing inverse-cosine and its application
to a GMR based angle sensor. IEEE Trans Instrum Meas 61(7):1991–1999
12. Benammar M, Ben-Brahim L, Alhamadi MA (2005) A high precision resolver-to-DC converter.
IEEE Trans Instrum Meas 54(6):2289–2296
13. Bhaskarrao NK, Anoop CS, Dutta PK (2017) A simple and efficient front-end circuit for
magneto-resistive angle sensors. In: International instrumentation and measurement technology
conference. IEEE, Turin, pp 1–6
14. Kishor Bhaskarrao N, Sreekantan AC, Dutta PK (2018) Analysis of a linearizing direct digitizer
with phase-error compensation for TMR angular position sensor. IEEE Trans Instrum Meas
67(8):1795–1803
15. Bhaskarrao NK, Anoop CS, Dutta PK (2021) A novel linearizing signal conditioner for half-
bridge-based TMR angle sensor. IEEE Sens J 21(3):3216–3224
16. AAT003 TMR angle sensor kit. https://www.nve.com/Downloads/ag931-07.pdf. Accessed 10
Aug 2021
17. Wang D, Brown J, Hazelton T, Daughton JM (2005) Angle sensor using spin-valve with SAF
structure. IEEE Trans Magn 41(10):3700–3702
18. Wang WG et al (2013) Parallel fabrication of magnetic tunnel junction nanopillars by
nanosphere lithography. Sci Rep 3(1):1–6
19. AAT00X TMR angle sensor. https://www.nve.com/Downloads/AAT00X.pdf. Accessed 10 Aug
2021
The Use of LBP Features in Transform
Domain for Object Recognition
R. Ahila Priyadharshini and S. Arivazhagan
Abstract Local binary pattern (LBP) is very useful in various applications of

machine vision problems. It is an effective multi-resolution texture descriptor.
Various types of local binary patterns have been proposed till now. Even though
deep learning framework has gained more attention in generic object recognition
nowadays, in this article, we studied the effectiveness of LBP features in the trans-
form domain for the task of object recognition. The comparison was done based on
the performance measure such as recall, precision, equal error rate and F-measure.
Images have been pre-processed using Log Gabor filter, and then, different types of
LBPs are applied on the transformed images, and the statistical features are extracted
and classified with support vector machine (SVM). The evaluation is done on the
Graz databases. It is observed that rotation invariant LBP and direction coded LBP
perform better with lesser error rate and higher recall, precision and F-measure.
Keywords Object recognition · Local binary pattern · Log Gabor filter · Equal
error rate
1 Introduction
Object recognition is extensively used in the machine vision industry for the commit-
ments of inspection, registration quality control, etc. The images are dependent on
various parameters such as illumination, camera parameters and location. As an
object should be recognized from the images of a scene which contains multiple
objects, the complexity level of object recognition model relies on many aspects
such as scene constancy, number of total objects in an image, image-models spaces,
number of objects in the model database and possibility of occlusion. In the object
recognition task, an input image is assigned with one of the N class labels. The
main steps involved in this process are (i) preprocessing, (ii) feature analysis and
(iii) classification. In the preprocessing, basic transformation is applied on the image
R. Ahila Priyadharshini (B) · S. Arivazhagan

Mepco Schlenk Engineering College, Sivakasi, India
e-mail: rahila@mepcoeng.ac.in
https://doi.org/10.1007/978-981-19-6913-3_18
274 R. Ahila Priyadharshini and S. Arivazhagan
[1]. In feature analysis, set of feature vectors are derived that represent the image
highlighting the significant aspects.
Most of the machine vision systems use either global features or local features.
Global features represent the complete image as a single vector [2]. Local features
are derived at several locations in the image and are robust to clutter and occlusion.
In this work, local features are computed by applying uniform grid on the Log Gabor
transformed images.
The Gabor function is a prevailing tool in the field of machine vision. Gabor func-
tion is optimally localized in spatial and frequency domain. So, it is predominantly
used for texture analysis [3]. Log Gabor filter has Gaussian response on logarithmic
frequency scale. It captures additional information in high frequency regions [4, 5].
Hu et al. [6] used Log Gabor filter for fingerprint segmentation. Lajevardi et al. recog-
nized facial expression using Log Gabor filter and LBP [7]. The performance based
on Log Gabor filter gives better classification accuracy than LBPs. Seif et al. used
Log Gabor filter for iris recognition [8]. The traditional LBP was proposed by Ojala
et al. [9]. Fathi et al. [10] introduced a new local binary operator called rotational
invariant local binary operator (LBPriu2 P,R ) for automatic blood vessel segmentation.
Trefny et al. [11] presented two different encoding schemes for local binary pattern
to represent the intensity function of local neighborhood. Wu et al. [12] proposed a
novel local binary operator named as improved center symmetric LBP (ICSLBPP,R )
to extract more texture information from an image to increase the recognition rate.
Martins et al. [13] demonstrated the classification of different forest species using
original LBP. Nguyen et al. [14] proposed a modification in the LBP named as non-
redundant LBP (NRLBPP,R ) and used it for detecting the objects. Liu et al. [15] used
a set of six LBP-like features derived from local intensities and differences for face
recognition. Arivazhagan et al. [16, 17] used Log Gabor transform for automatic
target recognition in SAR images and vehicle recognition.
The main focus of this work is to compare the performance of various types of
LBPs for the object recognition task and to find the best binary pattern for each
category of objects like bike, bike and person, person and car based on the perfor-
mance measures such as recall, precision, F-measure and equal error rate. Instead of
applying local binary pattern to the image directly, here, LBP is taken on the trans-
formed image. Local features are derived here to capture various regions of complex
images and are categorized using SVM classifier. The proposed method is evaluated
on Graz database.
First, the input is transformed using Log Gabor filter with four scales and six orienta-
tions. Then, a uniform rectangular grid is applied on the transformed images. After-
ward different types of LBPs such as original LBP (o_LBP), rotation invariant LBP
(ri_LBP), improved center symmetric LBP (ICS_LBP), transition LBP (tLBP), direc-
tion coded LBP (dLBP) and non-redundant LBP (NRLBP) are computed over the
The Use of LBP Features in Transform Domain for Object Recognition 275
Fig. 1 Proposed method’s block diagram
non-overlapping patches. Finally, mean and standard deviation are computed and
classified using SVM classifier. The proposed method’s block diagram is displayed
in Fig. 1.
2.1 Log Gabor Transform
Gabor filter, a linear filter which has impulse response as a sine modulated Gaussian
function. Due to the limitations of Gabor filter [18], Log Gabor function, an alternate
to Gabor is suggested by D Field [19]. Log Gabor filters can be built with arbitrary
bandwidth. The radial component G( f ) and the angular component G(ϕ) of Log
Gabor function are defined in Eq. 1.
2
f
log f
− 20
−(
σ ϕ−ϕ0 )2
2 log f f
G( f, ϕ) = G( f ) × G(ϕ) = e 0 ×e 2σϕ2
(1)
where f 0 denotes the center frequency, σf denotes the bandwidth scaling factor, ϕ0
denotes orientation angle and σϕ represents angular bandwidth.
The vital characteristics of Log Gabor functions are (i) DC component is not
present in Log Gabor functions so that the contrast ridges and edges of images are
improved. (ii) It has a prolonged tail transfer function, providing wide spectral infor-
mation with localized spatial extent. Here, Log Gabor functions with six orientations
and four scales are used. Figure 2 depicts 24 sub-bands of the transformed images.
2.2 Local Binary Patterns
LBP is a powerful descriptor for texture classification [20]. LBP is robust to

illumination, and it is computationally simple.
Original Local Binary Pattern (LBPP,R )
Fig. 2 Output of Log Gabor transform with four scales and six orientations for a sample image
The original LBP is computed by thresholding a N × N neighborhood with the

center value and is depicted in Fig. 3. The LBP has a limitation while using small
neighborhood as it is unable to capture the prevailing features in huge structures. This
can be overcome by extending the operator to cope with larger neighborhoods [9].
With the help of circular neighborhoods and bilinear interpolation, the neighborhood
with arbitrary radius can be allowed.
The LBPP,R generates 2 p dissimilar binary patterns. The mathematical represen-
tation of the LBPP,R is given by Eq. 2
Fig. 3 LBPP,R computation


p−1

LBPP,R = s g p − gc 2 p (2)
p=0
Rotational Invariant Local Binary Pattern (LBPP,R riu2 )

In traditional LBP, if the size of neighborhood increases, the number of prevailing
texture patterns also grow extensively. To diminish the count of texture patterns, an
extension, rotation invariant method, named LBPriu2 , was proposed [10].
The LBPriu2 is represented in Eq. 3
⎧ p−1
⎨
BCP,R (i ) U BCP,R ≤ 2
P,R =
LBPriu2 (3)
⎩ i=0
p+1 Otherwise
The LBPriu2 successfully extracts texture primitives in flat surfaces and straight
edges [9].
Improved Center Symmetric Local Binary Pattern (ICS_LBP)
The improved version of CS-LBP [12] named ICS-LBP is illustrated in Fig. 4 for 8
neighbors. The LBP yields 256 (28 ) dissimilar binary patterns, however, ICS-LBP
gives only 16 (24 ) dissimilar binary patterns.
The mathematical representation of ICS-LBP is given in Eq. 4
2 )−1
(
p
ICS-LBPP,R (x, y) = sICS-LBP pi , pc , pi+( 2p ) × 2i (4)

i=0
Transition Local Binary Pattern (tLBP)

Transition coded LBP gives the neighborhood pixel comparison in rightward direc-
tion for all pixels except the center. Each sequence with same binary values indicates
ordered sequence of pixel intensities. Thus, the tLBP is gray-scale invariant.
Neighborhood Binary Pattern
Fig. 4 ICS-LBP features for 8 pixel neighborhood

The mathematical representation for transition LBP is given in Eq. 5

P−1

tLBPP,R = s(go − g P−1 ) + s g P − g p−1 2 P (5)
p=1
Direction Coded Local Binary Pattern (dLBP)

Motivation of dLBP is to provide improved information of local pattern in terms of
direction functions similarly to CS-LBP [11]. In dLBP, intensity variation toward
four base directions are encoded into two bits.
The expression for dLBP is given in Eq. 6

−1
P

dLBPP,R = s g p − gc g p + p − gc 22 p + s g p − gc − g p + p − gc 22 p +1
p =0
(6)
Both the LBP and the dLBP rules encode if center pixel is an extrema. Unlike the
LBP rule, the dLBP does not encode it as maximum or minimum but encodes if sign
of first and second differential is the same.
Non-redundant LBP (NRLBP)
The two notable disadvantages of original LBP to encode object’s appearance [14]
are storage requirement and discriminative ability. LBP codes indicated by red boxes
in both the images in Fig. 5 are different although they characterize same structure.
Fig. 5 LBP code of a sample image and its inverted image

Non-redundant LBP (NRLBP) overcome both of the above issues and is given in
Eq. 7

NRLBPP,R (xc , yc ) = min LBPP,R (xc , yc ), 2 P − 1 − LBPP,R (xc , yc ) (7)
The NRLBP treats the LBP code and its complement as equal. The number of
histogram bins in NRLBP is decreased by half.
2.3 Performance Measures
Recall gives the proportion of the objects that are detected correctly, and precision
indicates how often the detections the system makes are false. These two quantities
are given in Eqs. 8 and 9, respectively.
TP
Recall = (8)
TP + FN
TP
Precision = (9)
TP + FP
F 1 measure is shown in Eq. 10. Accuracy is the ratio of right predictions to the
total predictions and is given in Eq. 11.
2 × Recall × Precision
F-measure = (10)
Recall + Precision
TP + TN
Accuracy = (11)
TP + FN + TN + FP
3 Experimentation Results
The experiments are done on Graz database [21, 22]. The images in Graz databases
are resized to 300 × 300 for experimentations. All categories of Graz 01 database
and one category (car) from Graz 02 database are considered for experimentation.
Every image in Graz database is convolved with Log Gabor function with six
orientations and four scales and resulting in 24 sub-bands. For every sub-bands, a
uniform grid of size 3 × 3 is applied. Then, for every 3 × 3 non-overlapping patches,
six different types of LBPs such as original LBP, rotation invariant LBP, ICS-LBP,
transition LBP, dLBP and non-redundant LBP are applied. The statistical features
are derived and classified using SVM classifier (RBF kernel). The different types of
LBP patterns for a sample image are shown in Fig. 6.
i ii iii iv
v vi vii viii
Fig. 6 Different types of LBP pattern (i) sample image, (ii) filtered image, (iii) LBPP,R , (iv)
LBPP,R riu2 , (v) ICS_LBP, (vi) tLBP, (vii) dLBP, (viii) NRLBP
The training and testing ratio are maintained as in [23]. The training and testing are
repeated for 10 times with non-overlapped random split. The performance measures
and accuracy obtained using six different types of LBPs separately for Graz 01
database and Graz 02 databases are shown in Tables 1 and 2, respectively.
Table 1 depicts that rotation invariant LBP gives highest recall, precision and F-
measure for two categories such as bike and bike and person. For categories such as
person and car, direction coded LBP yields the highest measure. Figure 7 depicts the
average ROC curves of Graz 01 database and Graz 02 database using six different
LBPs.
The obtained AUC values are given in Table 3. The largest AUC values are high-
lighted. Equal error rates (EER) for the Graz databases are displayed in Fig. 8. From
Fig. 8, it is understood, riLBP accomplishes well for two categories. For person
and car tasks, direction coded LBP yields better performance. Table 3 shows the
computed AUC values for Graz database.
Table 4 shows the average EER of 10 runs for the Graz database. Our proposed
method is comparable to the state-of-the-art methods.
Table 1 Recall, precision and F-measure of Graz database
Category Recall (%) Precision (%) F-measure (%)
O_LBP ri_LBP ICS_LBP tLBP dLBP NRLBP O_LBP ri_LBP ICS_LBP tLBP dLBP NRLBP O_LBP ri_LBP ICS_LBP tLBP dLBP NRLBP
Bike 86.8 92.4 85.8 87.8 70.4 87.6 86.6 91.8 84.4 87.1 73.9 83.9 86.7 92 85.1 87.4 72.1 85.7
Person 75.8 82.8 76 72.4 97.4 76 52.2 62.4 57.1 55.3 91.4 59.7 61.8 71.2 65.2 62.7 94.3 66.9
Bike and 92 99.2 95 97.4 86.4 97 62.9 85.96 75.2 73.5 73.3 78.8 74.7 92 83.9 83.8 79.3 87
person
Car 79.7 95.6 79.2 84.9 97.3 88.4 61.3 88.96 61.9 66.8 94.7 78.3 69.3 92.2 69.5 74.8 95.98 83
o_LBP—original LBP, ri_LBP—rotation invariant LBP
Bold specifies the highest performance measure
The Use of LBP Features in Transform Domain for Object Recognition
281
Table 2 Accuracy for Graz databases

Database AUC
OLBP riLBP ICS-LBP tLBP dLBP NRLBP
Bike 86.70 92.10 85.00 87.40 72.80 85.40
Person 58.20 66.50 59.40 56.90 94.10 62.50
Bike and person 68.90 91.50 81.80 81.10 77.50 85.20
Car 64.70 91.90 65.20 56.90 95.90 80.90
ROC curve for Bike in Graz01-LBP ROC curve for Person in Graz01-LBP
1 1
0.9 0.9
0.8 0.8
True positive rate
0.7 True positive rate 0.7
0.6 0.6
0.5 0.5
0.4 RotLBP 0.4 RotLBP

OriLBP OriLBP
0.3 ICSLBP 0.3 ICSLBP
TransLBP TransLBP
0.2 0.2
DirLBP DirLBP
0.1 NRLBP 0.1 NRLBP
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False positive rate False positive rate
(i) (ii)
ROC curve for Bike and Person in Graz01-LBP ROC curve for Car in Graz02-LBP
1 1
0.9 0.9
0.8 0.8
True positive rate
True positive rate
0.7 0.7
0.6 0.6
0.5 0.5
0.4 RotLBP 0.4 RotLBP
OriLBP OriLBP
0.3 ICSLBP 0.3 ICSLBP
TransLBP TransLBP
0.2 0.2
DirLBP DirLBP
0.1 NRLBP NRLBP
0.1
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False positive rate False positive rate
(iii) (iv)
Fig. 7 ROC curves of Graz databases. (i) Bike, (ii) person, (iii) person and bike, (iv) car
4 Conclusion
LBP extracts structural primitive textures excellently. In this paper, the task of object
recognition is investigated by using six different types of LBPs in the transformed
Table 3 Computed AUC values for categories of Graz database

Database AUC
riLBP OLBP ICS-LBP tLBP dLBP NRLBP
Bike 0.9210 0.8670 0.8500 0.8740 0.7280 0.8540
Person 0.6650 0.5320 0.5940 0.5690 0.9410 0.5870
Bike and person 0.9150 0.6890 0.8180 0.8110 0.7750 0.8520
Car 0.9187 0.6467 0.6520 0.7133 0.9593 0.8093
Fig. 8 Comparison of ROC EER for categories of Graz databases
Table 4 Average EER on the Graz database

Database [23] [24] [25] [26] [27] ri_LBP d_LBP
Bike 16.5 13.7 10 – 19.8 8.2 24.8
Person 23.5 17.7 13 – 16 49.8 9.2
Bike and person – – – – – 16.2 31.4
Car 29.8 – – 31.61 24.4 11.87 5.47
domain on the complex Graz database. It can be concluded that rotation invariant
LBP and direction coded LBP are effective for object recognition task. The improved
performance is due to the intensity inversion invariance property of direction coded
LBP.
References
1. Kolsch T (2003) Local features for image classification. Diploma thesis, Lehrstuhl fur
Informatik VI, RWTH Aachen University, Aachen, Germany
2. Hegerath A (2006) Patch-based object recognition. Diploma thesis, Human Language
Technology and Pattern Recognition Group, RWTH Aachen University, Germany
3. Kanagalakshmi K, Chandra E (2012) Frequency domain enhancement algorithm based on
Log-Gabor filter in FFT domain. Glob J Comput Sci Technol 12(7), Version 1.0
4. Cook J, Chandran V, Sridharan S, Fookes C (2005) Gabor filter bank representation for 3D
face recognition. In: Proceedings of the digital image computing techniques and applications
(DICTA’05). IEEE. ISBN 0-7695-2467-2
5. Liu C, Wechsler H (2003) Independent component analysis of Gabor features for face
recognition. IEEE Trans Neural Netw 14(4):919–928
6. Hu C, Yin J, Zhu E, Chen H, Li Y (2010) A composite fingerprint segmentation based on Log-
Gabor filter and orientation reliability. In: IEEE international conference on image processing
(ICIP), pp 3097–3100
7. Lajevardi SM, Hussain ZM (2009) Facial expression recognition using Log-Gabor filters and
local binary pattern operators. In: Proceedings of international conference on communication,
computer and power (ICCCP’09), Feb 2009, pp 349–353
8. Seif A, Zewail R, Saeb M, Hamdy N (2003) Iris identification based on log Gabor filtering. In:
Proceedings of IEEE 46th Midwest symposium on circuits and systems, vol 1, pp 333–336
9. Ojala T, Pietikainen M, Maenpaa T (2002) Multi resolution grayscale and rotation invariant
texture classification with local binary patterns. IEEE Trans Pattern Anal Mach Intell
24(7):971–987
10. Fathi A, Naghsh-Nilchi AR (2014) General rotation-invariant local binary patterns operator
with application to blood vessel detection in retinal images. Pattern Anal Appl 17(1):69–81
11. Trefny J, Matas J (2010) Extended set of local binary patterns for rapid object detection. In:
Computer vision winter workshop, vol 1, pp 1–7
12. Wu X, Sun J (2009) An effective texture spectrum descriptor. In: Fifth international conference
on information assurance and security (IAS), vol 2, pp 361–364
13. Martins J, Oliveira LS, Nisgoski S, Sabourin R (2013) A database for automatic classification
of forest species. Mach Vis Appl 24(3):567–578
14. Nguyen D, Zong Z, Ogunbona P, Li W (2010) Object detection using non-redundant local
binary patterns. In: IEEE 17th international conference on image processing, pp 4609–4612
15. Liu L, Fieguth P, Zhao G, Pietikäinen M, Hu D (2016) Extended local binary patterns for face
recognition. Inf Sci 358–359:56–72
16. Arivazhagan S, Priyadharshini RA, Sangeetha L (2017) Automatic target recognition in SAR
images using quaternion wavelet transform and principal component analysis. Int J Comput
Vis Robot 7(3):314–334
17. Priyadharshini RA, Arivazhagan S, Sangeetha L (2014) Vehicle recognition based on Gabor
and Log-Gabor transforms. In: IEEE international conference on advanced communications,
control and computing technologies, pp 1268–1272. https://doi.org/10.1109/ICACCCT.2014.
7019303
18. Yu Q, Tan X, Wang H (2015) A face recognition method based on total variation minimization
and Log-Gabor filter. In: International conference on electromechanical control technology
and transportation
19. Field D (1987) Relations between the statistics of natural images and the response properties
of cortical cells. J Opt Soc Am A 4(12):2379–2394
20. Oulu (2003) The local binary pattern approach to texture analysis—extensions and applications.
Diploma thesis, Infotech Oulu, University of Oulu
21. http://www.emt.tugraz.at/~pinz/data/Graz-01/
22. http://www.emt.tugraz.at/~pinz/data/Graz-02/
23. Opelt A, Pinz A, Fussenegger M, Auer P (2006) Generic object recognition with boosting.
IEEE Trans Pattern Anal Mach Intell (PAMI) 28(3):416–431
24. Lazebnik S, Schmid C, Ponce J (2006) Beyond bags of features: spatial pyramid matching for
recognizing natural scene categories. In: CVPR, vol 2, pp 2169–2178
25. Boiman O, Shechtman E, Irani M (2008) In defense of nearest-neighbor based image
classification. In: CVPR 2008, pp 1–8
26. Aldavert D, Ramisa A, Toledo R, de Mantaras RL (2009) Efficient object pixel-level catego-
rization using bag of features. In: Advances in visual computing. Lecture notes in computer
science, vol 5875, pp 44–54
27. Ghodrati M, Khaligh-Razavi S-M, Ebrahimpour R, Rajaei K, Pooyan M (2012) How can selec-
tion of biologically inspired features improve the performance of a robust object recognition
model? PLoS ONE 7(2):e32357
Design and Simulation of Capacitive
Pressure Sensor for Monitoring
Lead-Acid Battery Charge
Yashwant Adhav, Dayaram Sonawane, and Chetankumar Patil
Abstract In the present study, we have proposed a design of a novel, graphene-

based micro-capacitive pressure sensor to measure minute variation in differential
pressure developed in the air-purge system of lead-acid battery. Online state of charge
(SOC) monitoring of lead-acid batteries using a sensor is a critical problem. The key
principle is that specific gravity of the acid solution interlinks the amount of charge
or discharge of the acid solution and the differential pressure of the air-purge system.
We have obtained results from the analytical model, which are benchmarked with
the numerical model of Comsol Multiphysics commercial software. The proposed
graphene-based capacitive MEMS sensor has been designed for the range of 0–
1 kPa differential pressure variation. The gap between the two capacitive plates is
maintained at 3 µm and the thickness of the circular diaphragm is 10 µm. It is found
in this study that the minuscule range of small changes in the specific gravity (or
pressure) is consistent with the sensitivity of the newly designed MEMS sensor. In a
nutshell, it can measure a tiny pressure drop in the air-purge system. On the contrary,
the micro-piezo resistance sensor is observed to be less accurate on the account
of its temperature sensitivity. Through this study, it is observed that the graphene-
based MEMS sensor shows greater accuracy compared to the silicon-based sensor.
The change in capacitance of graphene is linear compared to silicon. In case of the
graphene, capacitance changes from 4.875 × 10−12 to 4.955 × 10−12 pF whereas in
silicon, capacitance changes from 2.3 × 10−12 to 3.5 10−12 pF.
Keywords State of charge · MEMS capacitive sensor · Graphene · Air-purge

tube · MEMS sensor design
Y. Adhav (B)
Cummins College of Engineering for Women, Pune 411052, India
e-mail: yashwant.adhav@cumminscollege.in
D. Sonawane · C. Patil
Department of Instrumentation and Control Engineering, College of Engineering, Shivajinagar,
Pune 411005, India
e-mail: dns.instru@coep.ac.in
C. Patil
e-mail: cyp.instru@coep.ac.in
https://doi.org/10.1007/978-981-19-6913-3_19
288 Y. Adhav et al.
1 Introduction
This study is about knowing the SOC of a lead-acid battery. Lead-acid batteries
have stable outputs, low costs and capability to withstand at different temperature
ranges. Such battery has a lot of scope in photo voltaic power systems, wind power
plants, uninterpreted power systems, smart grids and submarine [1–6]. The hydro-
static principle is used to see the state of charge of the battery. When the battery gets
charged or discharged, the concentration of acid changes. The specific gravity (SG)
of acid changes due to a change in the concentration of the sulphuric acid. The state
of the battery is monitored by measuring SG [3]. As per the hydrostatic pressure
principle, if the hydrostatic head is constant, then pressure is directly proportional
to the density of sulphuric acid. In a lead-acid battery, the bottom portion is covered
with lead sheets. At the top side of lead sheets, a very little space is available for
acid where these air-purge sensing tubes are immersed. Back pressure developed in
this air tube is directly proportional to the specific gravity of acid in the battery. It
is the most important point to design a sensor that can measure this back pressure
developed in the bubbler tube. As the head is very low, it is challenging to measure a
specific gravity of the solution having the head as almost 2–3 cm above lead sheets in
the battery. A small air-purge system has already been developed [7] to know the SG
of the acid solution. In this patent [7], MEMS-based piezoresistive pressure sensor is
used to measure differential pressure. The main disadvantage of this pressure sensor
is being more sensitivity to temperature. As altitude changes, the sensor can produce
errors as pressure change is directly proportional to the altitude.
To overcome this limitation in the current study, a graphene capacitive pressure
sensor is designed and simulated using COMSOL Multiphysics simulation soft-
ware. When load is applied to a thin diaphragm, it deflects. Pressure-deflection
proportional relation can be simulated using a numerical method known as the finite
element method. An analytical method based on large-small deflections is used to
predict sensitivity and linearity [4]. The thickness of diaphragm can produce errors
in linearity. Linearity error of ± full-scale span (FSS) is accomplished. Both circular
as well as the rectangular diaphragm are examined for optimization. The analytical
method in this study is used to compare the results obtained by the simulation finite
element method. When pressure is applied to diaphragm, the gap between the plates
changes non-uniformly. FEM is performed to detect that deformation and detect a
change in capacitance of polysilicon plates [8]. For 0.1–1 MPa pressure displacement,
the change in capacitance of 30–44 pF has been noted.
The finite element method (FEM) used as the gap between plates varies non-
uniformly when pressure is applied to one of the plates. PDMS (polydimethyl-
siloxane) deformable dielectric is used between the two plates. When the pres-
sure is applied, dielectric layer separates the two electrodes [2], hence changes the
sensor’s capacitance. Graphene is an emerging material in the MEMS field having
high mechanical strength. This material consists of a thin layer of carbon atoms that
can make thickness up to nanometers [9]. The sensitivity of this low-pressure sensor
is 288 pF/Pa of 4 mm × 4 mm size.
Design and Simulation of Capacitive Pressure Sensor for Monitoring … 289
Fibre optical sensor, multiple point fibre sensor, ultrasonic, quartz crystal micro
balance technique, the piezoelectric sensor are used to estimate SOC which is
discussed in [3, 5, 6, 10, 11] in this paper; micro technology is superior in performance
over these macro techniques in terms of accuracy, size, strength and cost. Capacitive
pressure sensor with novel liquid crystal polymer is designed to have circular thin
plate [10]. Diaphragm having radius 0.0016 µm provides a total capacitance change
of 0.277 pF for the pressure range of 0–100 kPa.
In this study, circular diaphragm sensor is designed for the 0–6 kPa pressure range.
The circular Silicon diaphragm and graphene diaphragm results are compared in
terms of displacement of thin plate and capacitance change. The change in capaci-
tance is non-linear in Silicon in comparison with the graphene. Capacitive sensors
have lower noise and as well it requires lower power which is an added advantage.
The fabrication process of the capacitive pressure sensor is comparatively more
straightforward than the piezoresistive pressure sensor. For the measurement of low
pressure, a more sensitive pressure sensor is required. With an increase in sensitivity
of diaphragm, thickness of the sensor decreases, which in turn leads to the non-linear
relationship between applied pressure and deflection of the diaphragm. In capacitive
pressure sensor design, the diaphragm’s thickness, linearity and deflection are equally
essential parameters [4]. As silicon thickness is reduced, it is difficult to fabricate the
thin plate, and hence the sensor becomes bulky. CMOS (Complementary metal oxide
silicon) material is used for the thin plate to reduce the noise and parasitic capaci-
tance [12]. This pressure sensor has a sensitivity to 4.2 pF/bar. Flexible capacitive
pressure sensor is developed for different applications like sitting posture and tongue
pressure. The surface sensor is made with polyimide substrate stacked between ethy-
lene vinyl acetate (EVA) and polyvinyl silicone (PVS) [2]. The circular diaphragm
shows maximum deflection compared to rectangular and square diaphragm [13]. The
circular shape is selected for the diaphragm to get more sensitivity and stiffness.
In the past-related research, no researcher carried out work on designing a circular
capacitive pressure sensor whose diaphragm is made from graphene material for an
acid battery’s charge monitoring. Graphene material with high ultra toughness, high
elasticity and electron mobility is also high, transparency and conductivity are also
superior. As the hydrostatic head of sulphuric acid is minimal in the lead-acid battery,
this sensor is designed for 0–1 kPa pressure range. In this study, we can observe and
measure the diaphragm’s deflection when the battery is discharged and the displace-
ment of the diaphragm when the battery is charged. Graphene diaphragm is one plate,
and the other plate is placed with a gap of 3 µm. As for battery charge–discharge
due to change in differential pressure, the graphene diaphragm deflects; hence, the
capacitance of the sensor changes. By measuring capacitance, the state of the battery
can be estimated. In this simulation, electromechanics physics is encompassing to
design sensor available under structure mechanics. Structure mechanics integrate
Solid Mechanics and Electrostatics and also allows to move mesh. The deformation
can be modelled using electro statically actuated structures.
This paper consists of three main parts. In Sect. 2, the design of MEMS capaci-
tive pressure sensor using analytical method and design with Comsol Multiphysics
290 Y. Adhav et al.
are described. In Sect. 4, the results of analytical and finite element methods are
compared. Section 5 is about the conclusion and future scope of research.
2 Design of MEMS Capacitive Pressure Sensor Using

Analytical Method
As shown in Fig. 1, the state of charge in terms of battery percentage and SG gravity
is measured using a short air-purge system with integrated MEMS differential silicon
pressure sensor. Minute pressure is developed from 215 to 246 Pa in the battery. We
can use Pascal’s hydrostatic principle to measure SG of the acid solution and estimate
battery SOC [7]. In the present study, graphene-based capacitive sensor is simulated
to which pressure sensing tube can be connected and can get an added advantage
over previous differential MEMS sensor.
2.1 Analytical Design of Graphene Capacitive Sensor
A capacitive pressure sensor is designed analytically and this design is afterwards

simulated in COMSOL. When pressure is applied to the diaphragm, it deflects at the
centre. The change in the displacement of a thin plate causes a change in capacitance.
This capacitive sensor consists of two plates made from graphene material. The
thickness of the circular graphene diaphragm is 10 µm. The gap between the two
Fig. 1 Schematic diagram of SG measurement system

plates is maintained at 3 µm. The Poisson’s ratio of graphene is 1000 GPa. To

get linearity, the diaphragm’s deflection should be 25–30% of the gap between the
two plates. The sensor is designed for 0–1 kPa range. When a battery gets charged
or discharged, the change in pressure is 215–246 Pa as per the hydrostatic pressure
principle. The sensor is designed for the measurement of tiny variation in differential
pressure. The radius of the sensor is calculated using the equation:
/
Wmax × 4.13 × Eh 3
R= 4
( ) (1)
1 − v2 × p
In the above Eq. (1), R denotes radius of the circular diaphragm (µm). Wmax is the
maximum displacement of diaphragm at centre (µm), E represents Young’s modulus
of elasticity (N/m2 ), h is the thickness of the diaphragm (µm), v denotes Poisson’s
ratio and p represents applied pressure measured in the unit (bar).
This sensor is designed for 0–1 kPa using graphene material. The linearity of the
deflection of the diaphragm and the input applied pressure have prime importance.
For linear displacement of the diaphragm, its maximum displacement at the centre
should be 25–30% of the gap between the two plates (3 µm). In this design, 28%
deflection of the diaphragm has been considered.
Wmax indicates maximum displacement of diaphragm at the centre which is equal
to 0.84 (µm), E stands for Young’s modulus of elasticity of 1000 GPa. h represents
thickness of the diaphragm of 10 µm and v is Poisson’s ratio having value 0.17. p
signifies pressure applied to the thin plate of 0.01 bar. The radius of the thin graphene
plate is calculated using Eq. (1), the radius is 0.002432 µm for the pressure 0.01 bar.
The displacement of the diaphragm is calculated using the equation:
( )
1 − v2 × R4 × p
Wmax = (2)
4.13 × E × h 3
The maximum stress is calculated using the equation:
1.25 × p × R 2
σmax = (3)
h2
In the above Eq. (3), σmax represents maximum stress that can be sustained by thin
diaphragm. R denotes radius of the circular diaphragm of 0.002432 µm. h means
thickness of the diaphragm of 10 µm. P signifies pressure applied of 0.01 bar. As per
the above Eq. (3), the maximum stress calculated is 7.54 MPa. The voltage applied
between the two capacitive plates is 1 V. As shown in Fig. 2, voltage is distributed
across the thin plate. For different applied loads, the voltage distribution in a thin
plate is different.
The change in the capacitance is calculated using an equation
∈o ∈r A
C= (4)
d
292 Y. Adhav et al.
Fig. 2 Voltage distribution in the thin plate
In the above Eq. (4), ∈o denotes relative permittivity of air (8.854×10−14 F/cm). ∈r
represents relative dielectric constant of medium. A means overlapping area between
the plates and d stands for distance between the plates. The change in the capacitance
is 2.945 × 10−11 pF for 100 Pa pressure. For 1000 Pa pressure, the change in the
capacitance is 3.04 × 10−11 pF.
3 Design of MEMS Capacitive Pressure Sensor Using

a Finite Element Method in Comsol Multiphysics
Comsol Multiphysics uses the following main component to simulate the sensor
environment.
Parameters: In the Comsol simulation, initially, parameters have to be set where the
operating condition of the sensor has to be mentioned.
Geometry: To make the geometry, a 3D block is used with four layers of dimensions of
0.7 mm, 0.397 mm, 0.003 mm, 0.01 mm, respectively, to get an input port, sensor plate
thickness, the gap between two capacitive plates and the base of sensor respectively.
The cylinder is used to get a circular diaphragm of the capacitive sensor and circular
port at one end.
An average component coupling is used to compute the average of an expres-
sion over the selected geometric entities, which is the underside area of the circular
graphene plate. Integration component coupling is used with a point as the source
having maximum displacement at the centre. The collection of the boundaries done
by XZ and YZ symmetries is used to specify the symmetry of boundary condition in
physics interface. The first explicit domain is used for cavity selection, which is the
area between two main plates of the capacitive sensor. The second explicit is used to
select all domains to expect the cavity to which linear elastic properties are assigned.
Physics: Electromechanics physics is used to apply pressure and provide excitation
to the plates and assign boundary conditions. Electromechanics physics equation
form for stationary analysis is used as mentioned below.
−∇ · σ = Fv (5)
∇ · D = eq (6)
The prescribed displacement is the constraint in Z-axis. The prescribed displace-

ment node adds a condition prescribed in one direction and solid free to deform in
the other directions. Boundary load is used to apply pressure to the top plate of the
sensor. Terminal and ground are used to provide excitation to the sensor.
Mesh: When any sensor is fabricated, it has an irregular shape; hence to solve the finite
element equation, meshing is done by creating a mesh of different sizes and shapes.
Meshing decides the area under tension and displacement that is the membrane of the
sensor. The mapped mesh is used to lower the surface of the model. The maximum
mesh element size is 4.88 mm. Auxiliary sweep is used to apply a pressure range.
Study: This is used to observe stress and displacement and change in capacitance of
the diaphragm due to the change in pressure.
3.1 Geometry Building of Sensor and Material Selection
As shown in Fig. 3, cut section of the capacitive pressure sensor consists of a thin
plate having a thickness of 10 µm. The radius of the diaphragm is 0.002432 µm. The
second plate is made of graphene material having a thickness of 0.397 mm. Between
the two plates, space is evacuated by vacuum. This cavity has a relative permittivity
equal to 1. This pressure sensor is supported by borosilicate glass. The thickness of
the borosilicate glass base is 0.7 mm. The circular port is etched out of 0.41 mm to
enter air pressure towards the diaphragm. As shown in Table 1, the properties of the
material are entered in the material contents.
3.2 Selection of Physics and Meshing
This simulation has been undertaken under the electromechanics interface. It comes
under structural mechanics. It combines solid mechanics and electrostatics. Electro-
statically actuated structures deformation is modelled with moving mesh. As shown
294 Y. Adhav et al.
Fig. 3 Sensor geometry
Table 1 Material properties

Sr. No. Parameters Silicon Graphene Borosilicate Unit
1 Relative permittivity 11.7 2.14 4.8
2 Density 2330 2000 2230 kg/m3
3 Young’s modulus 170 1000 63 GPa
4 Poisson’s ratio 0.06 0.17 0.20
in Fig. 4, free tetrahydral meshing has been carried out for the capacitive pressure
sensor.
Fig. 4 Free tetrahydral

meshing
Fig. 5 Pressure versus capacitance-silicon
4.1 Change in Capacitance for 0–6 kPa Pressure Range
As shown in Fig. 5, when 0–6 kPa pressure is applied to silicon diaphragm, the
change in capacitance changes from 2.3 × 10−12 to 3 × 10−12 pF. In Fig. 6, when
input is applied to graphene diaphragm, the change in capacitance is from 4.875 ×
10−12 to 4.95 × 10−12 pF. The change in graphene material finds more linear as
compared to silicon.
4.2 Change in the Displacement of Diaphragm for Pressure

Range 0–6 kPa
As shown in Fig. 7, a thin plate deflects when the applied pressure is increased and
vice versa. As shown in Fig. 8, red colour indicates more deflection of thin plate and
blue colour indicates less deflection than the central deflection.
The first sensor model is designed for pressure range 0–6 kPa. Change in the
displacement of graphene and the silicon is linear. As shown in Fig. 9 in the case
of graphene, the displacement changes from 0 to 0.32 µm. As shown in Fig. 10
in the case of silicon, the displacement changes from 0 to 1.8 µm. The difference
in displacement in silicon is more than the graphene diaphragm displacement. The
296 Y. Adhav et al.
Fig. 6 Pressure versus capacitance-graphene
Fig. 7 Displacement of sensor
deflection of the graphene material is less than that of silicon because the graphene
strength is more, so that this material can handle more substantial applied pres-
sure. Also when the range changes, the radius of sensor differs and shows different
displacement.
Fig. 8 Graphene thin plate
Fig. 9 Applied pressure versus graphene displacement
4.3 Capacitance and Diaphragm Displacement of Graphene

for 0–1 kPa
Finally, the sensor model is designed for pressure range 0–1 kPa. Change in the
displacement of graphene diaphragm is linear. As shown in Fig. 12 in the case of
graphene, the displacement changes from 0 to 0.6 µm. The displacement of the
298 Y. Adhav et al.
Fig. 10 Applied pressure versus silicon displacement
diaphragm is different for different ranges because the radius of the diaphragm
changes as per the change in range. As shown in Fig. 11 in the case of graphene, the
capacitance changes from 2.935 × 10−12 to 3.025 × 10−12 pF.
Fig. 11 Pressure versus capacitance

Fig. 12 Pressure versus displacement
4.4 Comparison of the Displacement of the Diaphragm

by Analytical Method and Finite Element Method
The load is applied to the thin material in the range of 0–1 kPa in the step of 100.
The deflection of the diaphragm is calculated using Eq. (7) and calculated values and
observed values of displacement and capacitance are entered in Table 2.
( )
1 − v2 R4 p
Wmax = (7)
4.13Eh 3
As shown in Fig. 13, the analytical results are compared to the results obtained
by Comsol Multiphysics software. The displacement of the diaphragm is linear with
the applied pressure. The sensitivity of the designed sensor is 8.3 × 10−4 as per the
finite element method and as per the analytic approach, it is 6.4 × 10−4 .
4.5 Diaphragm Displacement at the Different Thicknesses
As shown in Fig. 14, pressure is applied to the diaphragm having thickness 0.010–
0.001 mm. To increase the sensitivity, thickness of the diaphragm is minimised and
readings are entered in Table 3, but the linearity between the applied pressure and
diaphragm decreases. In this design, the diaphragm’s thickness has been fixed and
kept 0.010 mm to get good linearity and sensitivity.
300 Y. Adhav et al.
Table 2 Displacement of
Pressure (Pa) Calculated Observed Capacitance
diaphragm and capacitance
displacement displacement (pF) × 10−11
(µm) (µm)
100 0.08 0.06 2.945
200 0.16 0.13 2.955
300 0.25 0.19 2.965
400 0.33 0.25 2.975
500 0.41 0.31 2.985
600 0.50 0.37 2.995
700 0.58 0.44 3.005
800 0.66 0.5 3.02
900 0.75 0.56 3.03
1000 0.83 0.64 3.04
Fig. 13 Analytical method

versus finite element method
Fig. 14 Diaphragm
displacement at different
thickness
Table 3 Displacement of diaphragm and applied pressure at different thickness

Pressure (Pa) Thickness (mm)
0.001 mm 0.003 mm 0.005 mm 0.007 mm 0.010 mm
100 2.97 2.96 2.95 2.95 2.94
200 3.02 2.99 2.97 2.96 2.95
300 3.07 3.02 2.99 2.98 2.96
400 3.13 3.06 3.02 2.99 2.97
500 3.2 3.1 3.04 3.01 2.98
600 3.28 3.15 3.07 3.03 2.99
700 3.38 3.2 3.09 3.04 3.00
800 3.52 3.25 3.12 3.06 3.01
900 3.7 3.32 3.16 3.08 3.02
1000 3.99 3.40 3.19 3.1 3.03
5 Conclusion
The first sensor model is designed for the pressure range of 0–6 kPa. The change in the
displacement of graphene and the silicon is linear. The difference in displacement
in silicon is more than the graphene diaphragm displacement. Finally, the sensor
model is designed for the pressure range of 0–1 kPa. The change in the displacement
of graphene diaphragm is linear. In the case of graphene, the displacement changes
from 0 to 0.6 µm whereas in graphene, the capacitance changes from 2.935 × 10−11
to 3.025 × 10−11 pF. The sensitivity of the designed sensor is 8.3 × 10−4 as per the
finite element method and as per the analytic approach, it is 6.4 × 10−4 pF. In this
design, the diaphragm’s thickness has been kept 0.010 mm to get good linearity and
sensitivity.
In this capacitive sensor design, only one port has been made to allow the applied
pressure to a thin plate. For the lead-acid battery to measure differential pressure,
a sensor having two ports is needed. We can use a single port sensor, but this will
measure the gauge pressure. As long as the battery’s acid level is constant, this sensor
shows output proportional to the electrolyte’s specific gravity. After a few months
in a lead-acid battery, the level of acid changes due to acid evaporation, so it needs
a differential pressure sensor to measure specific gravity. As per the hydrostatic
principle, if the head is constant, the differential pressure is directly proportional to
the specific gravity electrolyte. The output of this sensor is the change of capacitance.
Most of the microcontrollers require the input signal in voltage form. So, if the
capacitance to voltage converter is designed on the same platform, it could be easier
to present the signal in the form of a state of charge battery.
The first sensor model is designed for the pressure range of 0–6 kPa. Change in
the capacitance of graphene is linear as compared to the silicon. As shown in Fig. 6
in the case of graphene, the capacitance changes from 4.875 × 10−12 to 4.955 ×
302 Y. Adhav et al.
10−12 pF. As shown in Fig. 5 in the case of silicon, the capacitance changes from 2.3
× 10−12 to 3.5 × 10−12 pF.
Acknowledgements The authors thank for the financial support provided by the Instrumentation
and Control Department, College of Engineering, Pune (C.O.E.P) for patent filing and further
process on the research work.
References
1. Degla A, Chouder MCA, Bouchafaa F, Taallah A (2018) Update battery model for photovoltaic
application based on comparative analysis and parameter identification of leadacid battery
models behaviour. IET Renew Power Gener 12(4):484–493
2. Dinh THN, Joubert P, Martincic E, Dufour-Gergam E (2014) Flexible 3-axes capacitive pressure
sensor array for medical applications. In: SENSORS, 2014. IEEE, pp 855–858
3. Hancke GP (1990) A fiber-optic density sensor for monitoring the state-of-charge of a lead
acid battery. IEEE Trans Instrum Meas 39(1):247–250
4. Lin L, Chu H-C, Lu Y-W (1999) A simulation program for the sensitivity and linearity of
piezoresistive pressure sensors. J Microelectromech Syst 8(4):514–522
5. Marcos-Acevedo J, Cao-Paz A, del Rio-Vazquez A, Quintans-Grana C, Martinez-Penalver C
(2009) Density measurement into lead-acid batteries with multi-point optical fiber sensor. In:
2009 IEEE instrumentation and measurement technology conference, pp 715–718
6. Neto PBL, Saavedra OR, de Souza Ribeiro LA (2018) A dual-battery storage bank configuration
for isolated microgrids based on renewable sources. IEEE Trans Sustain Energy 9(4):1618–
1626
7. Patil CY, Adhav YG (2020) A lead acid battery charge monitoring system. The patent office
journal no. 15/2020 (application no. 201821037211 A), P–16896, 10 Apr 2020
8. Sathyanarayanan S, Juliet AV (2010) Design and simulation of touch mode MEMS capacitive
pressure sensor. In: 2010 international conference on mechanical and electrical technology, pp
180–183
9. Zhang Y, Gui Y, Meng F, Gao C, Hao Y (2016) Design of a graphene capacitive pressure
sensor for ultra-low pressure detection. In: 2016 IEEE 11th annual international conference on
nano/micro engineered and molecular systems (NEMS), pp 192–195
10. Palasagaram JN, Ramadoss R (2006) MEMS-capacitive pressure sensor fabricated using
printed-circuit-processing techniques. IEEE Sens J 6(6):1374–1375
11. y Paz AMC, Acevedo JM, Gandoy JD, del Rio Vazquez A (2006) Multisensor fibre optic
for electrolyte density measurement in lead-acid batteries. In: IECON 2006—32nd annual
conference on IEEE industrial electronics, pp 3012–3017
12. Unigarro E, Bohrquez JC, Achury A, Ramirez F, Sacristn J, Segura-Quijano F (2017)
Differential capacitive pressure sensor design based on standard CMOS. Electron Lett
53(11):737–739
13. Sharma A, Singh J (2013) Design and analysis of high performance MEMS capacitive pres-
sure sensor for TPMS. In: 2013 international conference on control, automation, robotics and
embedded systems (CARE), pp 1–5
Development of Screw Press-Dewatering
Unit for Biogas Slurry
Madhuri More , Chitranjan Agrawal, and Deepak Sharma
Abstract The present research work aims to develop the screw press-dewatering
unit of 50 kg per hour capacity for biogas slurry as waste and produces organic
solid–liquid fertilizers. The performance effect on moisture content in biogas slurry
and shaft speed of developed unit in terms of dewatering efficiency and dewatering
rate was evaluated. The result showed the higher dewatering efficiency of 81.82% and
dewatering rate of 49.38 kg per hour. The properties of separated solid fertilizer from
raw biogas slurry were determined such as total solids content of 10.7%, ammoniacal
nitrogen of 2.6 g per kg, total nitrogen content of 4.9 g per kg, phosphate content of
5.0 g per kg, and potassium content of 3.0 g per kg, respectively. In liquid, fertilizers
were obtained such as total solids content of 4.07%, ammoniacal nitrogen of 3.1 g
per kg, total nitrogen content of 5.8 g per kg, phosphate content of 2.2 g per kg,
and potassium content of 5.8 g per kg, respectively. The organic solid fertilizer
can be used as phosphorous-rich fertilizer for improving soil characteristics, and
liquid fertilizers can be used as pesticide and fungicide to control pests on crop.
The economic analysis shows that the manufacturing cost of whole process was
2325.35 US$ and the payback period was 0.88 years.
Keywords Screw press-dewatering · Biogas slurry · Solid fertilizer · Liquid

fertilizer
1 Introduction
Biogas is efficient renewable source to reduce greenhouse gas (GHG) emissions

and waste disposal [1]. In recent years, biogas plant (132,000 BP) is installed in
the world, having a hug potential for biogas production. These biogas plants are
M. More (B) · D. Sharma

Department of Renewable Energy Engineering, CTAE, MPUAT, Udaipur, Rajasthan 313001,
India
e-mail: moremadhuri1992@gmail.com
C. Agrawal
Department of Mechanical Engineering, CTAE, MPUAT, Udaipur, Rajasthan 313001, India
https://doi.org/10.1007/978-981-19-6913-3_20
304 M. More et al.
used to produce biogas as well as biogas slurry. Biogas slurry production increased
toward the years due to increasing demand in developing countries. From all available
manure, 10 billion tons of nutrient-rich biogas slurry is produced under the process
of anaerobic digestion and used as organic manure for production of crops [2]. In
India, Ministry of New and Renewable Energy (MNRE) installed 12,019 number of
biogas plants in the year 2019–20 and up to 78 MT of biogas slurry was produced
per year [3]. Currently, increasing use of chemical fertilizers in farming systems
creates environmental issues due to improper use of fertilizers, which includes an
unacceptable percentage of nitrogen and phosphorous being released into soil and
ammonia losses to the atmosphere [4].
Biogas slurry produced from the biogas plant has good amount of nutrients, such
as macro (NPK) and micronutrients; therefore, it is very useful to the crop production
as a valuable organic fertilizer [5, 6]. Biogas slurry is an inexpensive way to reduce
the use of inorganic fertilizers in the soil and hence reduces the nature pollution
[7]. Nowadays, the production of biogas slurry is increasing; however, there is need
to store biogas slurry. Most of the farmers does not have sufficient field for proper
storage of biogas slurry. Therefore, they directly spread biogas slurry on their own
field [8]. The higher production of biogas slurry would pollute water bodies and
environmental issues, i.e., nitrogen losses and ammonia losses because of improper
management of biogas slurry waste [9]. The processing of biogas slurry is very costly
and energy consuming [10]. For the proper utilization of biogas slurry, dewatering
is a common method, which helps to reduce transportation, handling, and storage
costs. Long-distance transportation justifies economically investment on slurry up-
gradation for which dewatering is commonly adopted process [11]. This process
helps to separate solid–liquid fertilizers from biogas slurry.
Dewatering of biogas slurry is a process to dewatered liquid fraction from biogas
slurry and to separate solid–liquid fraction. This process can reduce nutrient losses
such as nitrogen (N), phosphorus (P), and potassium (K) to the environment and
manage the use of organic fertilizer. The separated solid fraction can be managed
differently on-farm for slurry utilization and crop fertilization [12, 13]. Application
of separated fractions to the soil can reduce the risk of the polluting soil surface
and groundwater along with greenhouse gas emissions [10]. There are several tech-
niques for the separation of biogas slurry like screw press, centrifugation, sieving or
screening, belt filters, and sedimentation or chemically enhanced sediment settling
[8]. Generally, screw press technology can be used to produce high dry matter (solid
fraction) and low dry matter (liquid fraction). Al Seadi et al. [14] found that the
solid and liquid partitioning divides most of the nitrogen in liquid fraction and phos-
phorous in solid fraction, which helps the plant nutrients management for better
crop production. The phosphorous-rich organic solid fertilizer can be applied as
phosphorous-rich fertilizer for improving soil characteristics. The major fraction
deriving from the dewatering process is organic liquid fertilizers, which contain the
part of nitrogen and potassium. This organic liquid fertilizer can be used as pesti-
cide and fungicide [15]. In the previous studies, the dewatering of biogas slurry was
carried out by traditional methods with the help of wheat straw, paddy straw, and
agricultural waste [16]. However, these traditional methods are labor intensive and
Development of Screw Press-Dewatering Unit for Biogas Slurry 305
time-consuming. Only few investigations have focused on screw press technology,

which is used in medium (2 m3 ) to large size (140 m3 ) biogas plants [8, 14]. The
screw press technology has gained popularity, especially those farmers who need to
export their extra fertilizer to other farms.
By considering the scope, biogas slurry has abundant availability and great poten-
tial, but slurry utilization is very difficult. The research on utilization of biogas is
slurry to produce organic solid–liquid fertilizers as a waste management. There-
fore, the present study investigated production of organic solid–liquid fertilizers
from biogas slurry in a screw press-dewatering unit of capacity of 50 kg/h. The
effect of moisture content of biogas slurry and shaft speed of unit in terms of dewa-
tering efficiency and dewatering rate was evaluated. The detailed characterization of
produced organic solid–liquid fertilizers was determined by using standard proce-
dure. The economic analysis of the present work was determined in terms of net
present worth, payback period, benefit–cost ratio, and internal rate of returns for the
commercialization.
2 Experimental Methods
The fresh biogas slurry was collected from available biogas plants such as KVIC
and Deenbandhu from the DREE, CTAE, MPUAT, Udaipur, Rajasthan (India). The
experimental method was carried out, which is shown in Fig. 1.
2.1 Characterization of Biogas Slurry
The properties of fresh biogas slurry were determined using ASTM standards Method
8075, Method 8190, and Method 8049 [17, 18]. The composition of biogas slurry in
terms of pH, density, moisture content (MC), volatile solids content (VS), ash content
(Ash), total solid content (TS), electrical conductivity (EC), nitrogen (N), phospho-
rous (P), and potassium (K) content was determined by using Kjeldahl apparatus and
spectrophotometer.
2.2 Screw Press-Dewatering Unit
The screw press-dewatering unit of capacity of 50 kg per hour was developed for the
present research work (Fig. 2). The developed unit consists of hopper, screw press,
outlet system, pumping system, and separate solid–liquid collector. The screw press-
dewatering unit was made by mild steel having barrel length and inside diameter
of 0.64 m and 0.16 m, respectively. The unit was operated with automatic power
connection of screw press and slurry feed pump. Helical gearbox and slurry pump
306 M. More et al.
Fig. 1 Experimental
methodology for dewatering
process Biogas
Slurry
Screw
Press
Unit
Solid
Fertilize
Fertilizer
Liquid
Fertilize
Fertilizer
are connected to the panel system of the developed unit. The helical gear mechanism
was used to reduce the speed of rotation. Slurry pump was used to transfer biogas
slurry from outlet of biogas plant to the hopper. A motor was used to supply the
power to rotate the shaft at variable speed controlled by a speed controller. The
specifications of the screw press unit are shown in Table 1.
2.3 Experimental Procedure
The performance evaluation of the developed unit was analyzed by moisture content
of biogas slurry and shaft speed. Total nine sets of experiment with three replications
of each set were performed to evaluate the performance of screw press unit in terms
Fig. 2 Screw press-dewatering unit
of dewatering efficiency and dewatering rate. The screw press-dewatering unit test
was performed at different moisture content (MC) of biogas slurry, i.e., 90–95%,
85–90%, and 80–85%, and shaft speed of 30 rpm, 35 rpm, and 40 rpm, respectively.
The experiment was conducted with the following set of treatments: T 1 —90–
95% MC in biogas slurry and 30-rpm speed; T 2 —85–90% MC in biogas slurry and
30-rpm speed; T 3 —80–85% MC in biogas slurry and 30-rpm speed; T 4 —90–95%
MC in biogas slurry and 35-rpm speed; T 5 —85–90% MC in biogas slurry and 35-
rpm speed; T 6 —80–85% MC in biogas slurry and 35-rpm speed; T 7 —90–95% MC
in biogas slurry and 40-rpm speed; T 8 —85–90% MC in biogas slurry and 40-rpm
speed; and T 9 —80–85% MC in biogas slurry and 40-rpm speed, respectively.
Dewatering efficiency was determined by the ratio of the total weight of liquid
dewatered from the biogas slurry to the total weight of biogas slurry by using equation
[4, 19];
W1
ED = × 100 (1)
W2
Dewatering rate represents the quantity of liquid dewatered from the fresh biogas
slurry per unit time expressed by using equation [4, 11];
W1
DR = × 100 (2)
T
where E D is the dewatering efficiency in percentage, DR is the dewatering rate in kg
per hour, W 1 is the weight of liquid dewatered from the slurry in kg, W 2 is the total
weight of slurry (kg), and T is the time taken for dewatering (h).
308 M. More et al.
Table 1 Specification of
Parameters Values
screw press-dewatering unit
Capacity 50 kg/h
Diameter of the first screw 0.15 m
Diameter of the last screw 0.08 m
Pitch of screw 0.09 m
Tapered angle 3°
Speed of rotation 30, 35, and 40 rpm
Motor 0.746 kW, 1440 rpm
Slurry pump 0.746 kW, 6 m head
Bulk-density raw material 1220 kg/m3
Theoretical screw volume 0.0015 m3
Mass flow rate 7.10 kg/min
Helix angle 10.81°
Screw conveyor length 0.635 m
Number of threads 7
Volume per pitch 0.042 m3 /h
Torque on the screw 402.78 Nm
Power on shaft 0.70 kW
Pressure created on a barrel 7.24 MPa
Volume of a barrel on a screw 0.012 m3
Diameter of the screw shaft 40 mm
Load lifted by the screw 42.82 kN
Pressure to be developed by the screw 2.59 MPa
Pressing area 0.1649 m2
Compression ratio 20:1
Capacity of screw press 531.9 g/h
Power to drive screw press 0.75 kW
Specific energy consumption 0.0022 kWh/kg
2.4 Statistical Analysis
Analysis of variance (ANOVA) was applied to analyze the data using randomized
block design (RBD) and compared by Design Expert-11 software at the 5% level of
significance in the Department of Statistics, RCA, MPUAT, Udaipur.
2.5 Characterization of Separated Solid–Liquid Fertilizer
The organic solid–liquid fertilizers were selected for the property analysis. The
total solids content (TS), ammoniacal nitrogen (NH4 -N), total nitrogen (TN-Method
8075), phosphate content (P2 O5 -Method 8190), and potash content (K-Method 8049)
were determined by using Kjeldahl apparatus and spectrophotometer [14, 18].
2.6 Economic Assessment
The economic feasibility of production of separated solid–liquid by dewatering

process using 50 kg per hour capacity unit was analyzed by the following parameters
[20–22]
i. Net present worth (NPW)
ii. Benefit–cost ratio (BCR)
iii. Payback period (PP)
iv. Internal rate of return (IRR).
3.1 Characterization of Biogas Slurry
The characterization of biogas slurry including physicochemical properties is

presented in Table 2. The biogas slurry contains 85.22% of moisture and high
percentage of volatile matter content of 73.53%, which is suitable for the dewatering
process. The physicochemical analysis of biogas slurry has good NPK content, which
is much suitable for increasing crop productivity. The overall amount of nutrient
content available in biogas slurry is sufficient for good crop growth [17].
3.2 Effect of Moisture Content in Biogas Slurry and Shaft

Speed on Dewatering Efficiency (ED )
The performance of developed unit was evaluated in terms of moisture content of

biogas slurry and shaft speed (Table 3). The experiment was carried out with nine
treatments of biogas slurry having moisture content (MC) in the range of 90–95,
85–90, and 80–85% and speed of rotation as 30, 35, and 40 rpm. All treatments
were arranged randomly with three replications for each treatment. Observations
were recorded for dewatering efficiency at various level of moisture content and
310 M. More et al.
Table 2 Physicochemical
Sr. No. Parameter Value
analysis of biogas slurry
1. pH 8.5
2. Electrical conductivity (dS m−1 ) 3.0–3.2
3. Moisture content (%) 85.22
4. Total solid content (%) 14.77
5. Volatile solid content (%) 73.53
6. Ash content (%) 26.47
7. Nitrogen content (%) 1.51
8. Phosphorus content (%) 0.77
9. Potassium content (%) 0.81
different screw shaft speed. Figure 3 shows that the dewatering efficiency of devel-
oped unit was increased with decreasing moisture content and increasing shaft speed.
Overall results concluded that Treatment 9 obtained maximum dewatering efficiency
of 81.82% at 80–85% MC in biogas slurry and 40-rpm speed. The results’ data
compared to Oladipo et al. [19] tested screw press unit with cassava mash, and they
have founded an average dewatering efficiency of 38.15% at a 66.75% MC (w.b.) for
fermented mash and 37.68% at a 70% MC (w.b.) for unfermented mash.
The data were analyzed to check effect on dewatering efficiency of performance
of developed unit by using ANOVA and the data compared to Design Expert-11
software. The observation shows that the 80–85% moisture content of biogas slurry
and 40-rpm shaft speed came out to be significant treatment at 5% level of significance
for evaluation of dewatering efficiency. The generated response surface diagram
between moisture content (A) and shaft speed (B) clearly indicates that the decrease
in moisture content and increase in speed of shaft gave better dewatering efficiency
as shown in Fig. 4. It was observed that the developed screw press-dewatering unit is
best suited for the performance and dewatering efficiency varies in the range of 78.01–
81.82%. The range being higher than the CD value points toward the significant effect
in dewatering efficiency due to the moisture content and speed of shaft (Table 4).
Table 4 shows that the Model F-value of 14.24 suggest that the model is significant
and P-values less than 0.0266 indicate model terms are significant. Figure 5 indicates
that the model is fit to relate the predicted and actual values and that the number of
experimental combinations formed in the design of experiment was enough to find out
the effect of independent variables (i.e., moisture content and speed) for dewatering
efficiency.
Table 3 Performance evaluation of screw press-dewatering unit at different moisture content and speed of shaft
Sr. Parameters Treatments Max. Min. Range Mean S.D. SEM CD CV
No. T1 T2 T3 T4 T5 T6 T7 T8 T9 ± (5%) (%)
1. DE (%) 81.01 81.72 81.53 78.01 78.94 80.42 79.49 81.19 81.82 81.82 78.01 3.81 80.46 1.35 0.42 1.24 0.9
2. DR 48.84 49.38 39.26 33.25 32.66 33.94 32.3 32.7 30.24 49.38 30.24 18.84 36.99 7.27 0.7 2.42 19.66
(kg h−1 )
DE: dewatering efficiency, DR: dewatering rate, Max: Maximum value, Min: Minimum, S.D.: standard deviation, SEM ±: standard error, CD: critical difference,
CV: coefficient of variation
Development of Screw Press-Dewatering Unit for Biogas Slurry
311
312 M. More et al.
82
81
Dewatering Efficiency, %
80
79
78
77
76
T1 T2 T3 T4 T5 T6 T7 T8 T9
Treatments
Fig. 3 Performance evaluation of moisture content and shaft speed on dewatering efficiency of
developed unit
Fig. 4 Statistical presentation of generated response surface diagram between moisture content (A)
and shaft speed (B) on dewatering efficiency of developed screw press-dewatering unit
Table 4 Test analysis of dewatering efficiency (E D ) of developed screw press-dewatering unit

Source Sum of squares df Mean square F-value p-value
Quadratic model 14.11 5 2.82 14.24 0.0266 Significant
A-MC 4.61 1 4.61 23.27 0.0170
B-Speed 0.5087 1 0.5163 2.61 0.2049
AB 0.8190 1 0.8190 4.13 0.1350
A2 0.1089 1 0.1089 0.5495 0.5053
B2 8.05 1 8.05 40.64 0.0078
Residual 0.5945 3 0.1982
Cor. total 14.70 8
Fig. 5 Plot diagram of predicted versus actual on dewatering efficiency of developed screw press-
dewatering unit
3.3 Effect of Moisture Content in Biogas Slurry and Shaft

Speed on Dewatering Rate (DR)
Observations were recorded for dewatering rate at various levels of moisture content
and shaft speed. The dewatering rate of the developed unit was determined by the ratio
of weight of raw material to the processing time. It can also be observed that the vari-
ation in dewatering rate of developed unit occurs due to variation in processing time.
314 M. More et al.
50
45
40
35
30
DR, kg/h
25
20
15
10
5
0
T1 T2 T3 T4 T5 T6 T7 T8 T9
Treatments
Fig. 6 Performance evaluation of moisture content and shaft speed on dewatering rate of developed
unit
The processing time was decreased with increasing dewatering rate. Figure 6 shows
that the dewatering rate of developed unit was increased with increasing moisture
content in biogas slurry and decreasing shaft speed. From Fig. 6, the mean dewa-
tering rate ranged from 30.54 to 49.38 kg h−1 . It can be observed that Treatment
2 can achieve the maximum dewatering rate of 49.38 kg h−1 for 85–90% moisture
content of biogas slurry and 30-rpm shaft speed. The data were analyzed by ANOVA,
compared by Design Expert-11 software, and concluded that the moisture content
of biogas slurry and speed of shaft has a significant effect on dewatering rate at 5%
level of significance.
Table 5 shows that the Model F-value of 8.43 suggest that the model is significant
and P-values less than 0.0181 indicate model terms are significant. Figure 7 shows
that the generated response surface diagram between moisture content (A) and shaft
speed (B) clearly indicates that the increase in moisture content and decrease in speed
of shaft gave better dewatering rate. Figure 8 indicates that the model is fit to relate
the predicted and actual values and that the number of experimental combinations
formed in the design of experiment was enough to find out the effect of independent
variables (i.e., moisture content and speed) for dewatering rate.
3.4 Characteristics of Organic Solid–Liquid Fertilizers
The properties of organic solid–liquid fractions were carried out by respective

ASTM standard methodology of total solids content (TS), ammoniacal nitrogen
Table 5 Test analysis of dewatering rate of developed screw press-dewatering unit

Source Sum of squares df Mean square F-value p-value
Linear model 311.36 2 155.68 8.43 0.0181 Significant
A-MC 18.76 1 18.76 1.02 0.3524
B-Speed 292.60 1 292.60 15.84 0.0073
Residual 110.82 6 18.47
Cor. total 422.18 8
Fig. 7 Statistical presentation of generated response surface diagram between moisture content (A)
and shaft speed (B) on dewatering rate of developed screw press-dewatering unit
(NH4 -N), total nitrogen (TN), phosphate content (P2 O5 ), and potassium content (K),
respectively. Ammoniacal nitrogen (NH4 -N) was determined for decreasing nitrogen
losses in separated solid and liquid fractions through ammonia volatilization and
nitrogen leaching [23]. Total phosphate content of separated fractions was higher in
organic solid fertilizer as compared to liquid fertilizer, which is very helpful for crop
productivity. The ammoniacal nitrogen content, total nitrogen content, and potassium
content were higher in organic liquid fertilizers due to increased ammonia content in
liquid followed by dewatering process. The dewatered organic liquid fertilizers were
used as pesticide and fungicide for controlling pest and fungus on crop growth [15].
316 M. More et al.
Fig. 8 Plot diagram of predicted versus actual on dewatering rate of developed screw press-
dewatering unit
It can be also recycled into biogas plant for increasing biogas production in short
period.
Table 6 represents the properties of organic solid–liquid fertilizers produced from
the dewatering of biogas slurry by using screw press-dewatering unit. The properties
of organic solid fertilizers were determined such as total solids content of 10.7%,
ammoniacal nitrogen of 2.6 g/kg, total nitrogen content of 4.9 g/kg, phosphate content
of 5.0 g/kg, and potassium content of 3.0 g/kg, respectively. In the separated solid
fertilizer, agriculture waste, neem leaves, and earthworm are added for increasing
nutritive values and good quality compost. In organic liquid, fertilizers were obtained
such as total solids content of 4.07%, ammoniacal nitrogen of 3.1 g/kg, total nitrogen
of 5.8 g/kg, phosphate content of 2.2 g/kg, and potassium content of 5.8 g/kg, respec-
tively. The dewatered liquid is mostly used as pesticide after the up-gradation and
value addition process. Moller et al. [21] found that the dewatering of slurry reduces
transport cost by separating the manure into a nutrient-rich solid–liquid fraction when
the solid nutrient-rich phase is over high distances and liquid fraction is transported
over low distances. The result compared to Table 9 concluded that the production of
organic solid–liquid fertilizers is suitable for increasing crop productivity.
Table 6 Distribution of nutrients in organic solid–liquid fertilizers

Separated fractions TS (%) TN (g/kg) NH4 -N (g/kg) P2 O5 (g/kg) K (g/kg)
Solid fraction 10.7 4.9 2.6 5.0 3.0
Liquid fraction 4.07 5.8 3.1 2.2 5.8
3.5 Economic Analysis
To evaluate the economic analysis of dewatering process, some assumptions were

made. The working life of developed screw press-dewatering unit was assumed
10 years, and the operation periods were considered 300 days in 365 days. The
capital cost of developed system is 1722.48 US$ (1 US$ = 69.67 INR as of June 30,
2019). The capital cost includes the price of helical gear and mud pump. The repair
and maintenance cost of developed unit is considered as 3% of capital cost. The
developed unit can process 50 kg per hour of biogas slurry and calculate production
cost of separated solid–liquid fractions. By selling the separated solid and liquid, the
capital investment of developed screw press unit can be recovered within 0.88 years.
The benefit–cost ratio of 1.59 specifies that the developed unit is economical for
dewatering of biogas slurry. However, the unit capacity, types of raw material, and
the dewatering efficiency of developed unit have direct impact on the production cost
of organic solid–liquid fertilizers.
The result of the present study compares with Kumar et al. [22], and it was found
that the unit screw press unit advantage is that it is economically cost-effective than the
rotary screen separator. Sievers et al. [24] performed techno-economic assessment of
solid–liquid separator, and it was performed for 2000-T slurry per day and production
cost is very high as compared to the present study. Tables 7 and 8 show the economic
analysis and economic measures of the developed screw press-dewatering unit for
production of organic solid–liquid fertilizers by dewatering process.
3.6 Comparison of Present Work and Literatures
This section presents the comparison of literature related to the solid–liquid sepa-
ration of biogas slurry and presents research work. Many researchers reported that
the biogas slurry has potential to produce organic manure, which is rich in nutritive
values. Khan et al. [25] studied on characterization of biogas slurry and concluded
that the use of biogas slurry is good for increasing crop yield. Yu et al. [5] also
studied biogas slurry and its properties for soil fertility for tomato crop. Hjorth et al.
[11] discussed various slurry separation techniques such as sedimentation, centrifu-
gation, and pressurized filtration (i.e., screw press) for solid–liquid separation and
production of organic fertilizers. It was observed that the screw press technology
for solid–liquid separation is good for recycling of organic matter and produced
plant nutrients could mitigate environmental pollutions. The output of separation is
318 M. More et al.
Table 7 Economic
Operating parameter/costs Output
feasibility for the developed
screw press-dewatering unit Production capacity 50 kg h−1
Operating days per year 300 days
Operating hours per day 8 h day−1
Capacity of utilization 80%
Raw material Biogas slurry
Output 96 T year−1
The installed cost of screw press unit (A) 1722.48 US$
Cost
Power @ 0.087 US$/kWh/year 258.37 US$
Maintenance includes @ 3% of (A) per year 15.93 US$
Manpower (1 labor) @ 5.74 US$/year 1722.48 US$
Discount rate @ 10% for screw press unit per 172.25 US$
year
Total cost (B) per year 2325.35 US$
Return
The net sale price of solid manure per tons 71.77 US$
The net sale price of solid manure per year (C) 6889.93 US$
Net annual saving per year (D = C − B) 4564.58 US$
Payback period 0.88
US$ 1.00 is equivalent to approximately 69.67 Indian rupees (cited
on, 03 July 2019)
Table 8 Economic measures

Economic measures Value
of the dewatering system
Net present worth (NPW) after 10 years 9390.62 US$
Benefit–cost ratio (BCR) 1.59
Payback period (PP) 0.88 years
Internal rate of return (IRR) 105%
varied with the type of separation technology, type of raw material, and speed of
unit applied. So far, very limited literature (small scale) is available on separation or
dewatering of biogas slurry and its by-product characterization. The research aimed
to study the production of organic solid–liquid fertilizers from biogas slurry by using
screw press-dewatering unit. Al Seadi et al. [14] studied characterization of sepa-
rated organic solid–liquid, which is nearly same as our results, but they studied all
dewatering technologies. Aguirre-Villegas et al. [26] concluded that the solid–liquid
separation of biogas slurry after digestion reduces GHG emission and increases
separation efficiency for manure management. The present research work investi-
gates production of organic solid–liquid fertilizer by using screw press-dewatering
technology and characterization of organic solid–liquid fertilizer. Table 9 shows
the comparative study of characterization of biogas slurry and separated organic

solid–liquid fertilizers.
4 Conclusions
In this study, biogas slurry is produced under the process of anaerobic digestion.
The biogas slurry is used as organic manure, but it has higher transportation cost
and storage cost. Therefore, the screw press unit was developed to separate solid–
liquid fraction from biogas slurry by dewatering process. Performance evaluation of
developed unit was carried out in terms of dewatering efficiency and dewatering rate
of developed unit. The treatment (T 9 ) obtained maximum dewatering efficiency of
81.82% at 80–85% MC in biogas slurry and 40-rpm shaft speed, and treatment (T 2 )
can achieve the maximum dewatering rate of 49.38 kg h−1 for 85–90% moisture
content of biogas slurry and 30-rpm shaft speed. It was found that the dependent
parameters such as moisture content and shaft speed had major impact on solid–
liquid separation. The properties of produced organic solid–liquid fertilizers indicate
that it has potential to be used as organic fertilizer and can be up-gradated by using
new technologies. Hence, the study concludes that the biogas slurry is utilized by
producing organic solid–liquid fertilizers by dewatering process and beneficial as
an economic and environment point of view. These produced organic solid–liquid
fertilizers can be utilized to provide healthy environment, and ultimately, it will help
to reduce nutrient losses and transportation problems. The economy of dewatering
process is an important scope to produce organic fertilizers from biogas slurry waste
of good quality and quantity in minimum production cost.
320
Table 9 Comparison of present work with literature

Characteristics pH EC TS (%) VS Ash Total C TN NH4 -N P2 O5 K References
Biogas slurry 7.9 3.1 ds/m 14.77 73.53% 26.47% – 1.51% – 0.77% 0.81% Present study
7.7 1468 µs/cm 5.1 0.78% – 1.6% 0.23% – – – [27]
– – 6.5 – – – 5.1 g/kg 3.2 g/kg 2.3 g/kg 5.5 g/kg [8]
7.64 – 4.85 3.39% 1.46% – 4.6 kg/t 3.1 kg/t 0.9 kg/t 3.5 kg/t [14]
Solid fertilizer 8.0 – 10.7 – – – 4.9 g/kg 2.6 g/kg 5.0 g/kg 3.0 g/kg Present study
9.2 52 µs/cm 38.1 0.91% – 23.3% 1.34% – – – [27]
– – 24.3 – – – 5.8 g/kg 2.7 g/kg 5.0 g/kg 5.8 g/kg [8]
– – 27.66 21.20% 6.46% – 8.15 kg/t 4.50 kg/t 6.52 kg/t – [14]
Liquid fertilizer 7.8 – 4.07 – – −5.8 g/kg 3.1 g/kg 2.2 g/kg 5.8 g/kg Present study
7.8 1620 µs/cm 2.9 0.98% – 1.1% 0.24% – – – [27]
– – 5.7 – – – 4.9 g/kg 3.0 g/kg 2.3 g/kg 6.2 g/kg [8]
– – 2.31 1.49% 0.082% 3.49 kg/t 2.63 kg/t 0.31 kg/t – [14]
M. More et al.
References
1. Martinov M, Scarlat N, Djatkov D, Dallemand JF, Viskovic M, Zezelj B (2020) Assessing

sustainable biogas potentials—case study for Serbia. Biomass Conv Bioref 10(2):367–381
2. World Biogas Association (WBA) (2019) Global potential of biogas. Available from: https://
www.worldbiogasassociation.org/wpcontent/uploads/2019/09/WBA-globalreport-56ppa4_
digital-Sept-2019.pdf
3. Ministry of New and Renewable Energy. Potential of renewable energy in India. http://mnre.
gov.in/file-manager/annual-report/2019-2020. Accessed 17 Sept 2020
4. Cocolo G, Hjorth M, Zarebska A, Provolo G (2016) Effect of acidification on solid–liquid
separation of pig slurry. Biosyst Eng 143:20–27
5. Yu FB, Luo XP, Song CF, Zhang MX, Shan SD (2010) Concentrated biogas slurry enhanced
soil fertility and tomato quality. Acta Agric Scand Sect B Soil Plant Sci 60(3):262–268
6. Koszel M, Lorencowicz E (2015) Agricultural use of biogas digestate as a replacement
fertilizers. Agric Agric Sci Procedia 7:119–124
7. Islam R, Rahman SME, Rahman MM, Deog Hwan OH, Chang RA (2010) The effects of biogas
slurry on the production and quality of maize fodder. Turk J Agric 34:91–99
8. Drosg B, Fuchs W, Al Seadi T, Madsen M, Linke B (2015) Nutrient recovery by biogas digestate
processing. IEA Bioenergy 711
9. Du H, Gao W, Li J, Shen S, Wang F, Fu L, Zhang K (2019) Effects of digested biogas slurry
application mixed with irrigation water on nitrate leaching during wheat-maize rotation in the
North China Plain. Agric Water Manag 213:882–893
10. Fangueiro D, Lopes C, Surgy S, Vasconcelos E (2012) Effect of the pig slurry separation
techniques on the characteristics and potential availability of N to plants in the resulting liquid
and solid fractions. Biosyst Eng 113(2):187–194
11. Hjorth M, Christensen KV, Christensen ML, Sommer SG (2011) Solid–liquid separation of
animal slurry in theory and practice. Sustain Agric 2:953–986
12. Fangueiro D, Hjorth M, Gioelli F (2015) Acidification of animal slurry—a review. J Environ
Manage 149:46–56
13. Hupfauf S, Bachmann S, Juárez MFD, Insam H, Eichler-Löbermann B (2016) Biogas digestates
affect crop P uptake and soil microbial community composition. Sci Total Environ 542:1144–
1154
14. Al Seadi T, Drosg B, Fuchs W, Rutz D, Janssen R (2013) Biogas digestate quality and utilization.
In: The biogas handbook, pp 267–301
15. Groot LD, Bogdanski A (2013) Bioslurry: brown gold. A review of scientific literature on the
co-product of biogas production. Food and Agriculture Organization of the United Nations
(FAO)
16. Kadam ER, Sharma D, Pawar EA (2017) Filtration of biogas spent slurry and its chemical
analysis. IJCS 5(3):405–408
17. Dahiya AK, Vasudevan P (1986) Biogas plant slurry as an alternative to chemical fertilizers.
Biomass 9(1):67–74
18. Yadav N, Kumar R, Rawat L, Gupta S (2014) Physico-chemical properties of before and after
anaerobic digestion of Jatropha seed cake and mixed with pure cow dung. J Chem Eng Process
Technol 5(2):1
19. Oladipo NO, Olotu FB, Adamade CA, Agaja MO (2015) Development of NCAM manually
operated single pole dewatering press
20. Mukhtar S, Sweeten JM, Auvermann BW (1999) Solid-liquid separation of animal manure and
wastewater
21. Moller HB, Lund I, Sommer SG (2000) Solid–liquid separation of livestock slurry: efficiency
and cost. Bioresour Technol 74(3):223–229
22. Kumar AK, Sharma S, Dixit G, Shah E, Patel A (2020) Techno-economic analysis of microalgae
production with simultaneous dairy effluent treatment using a pilot-scale high volume V-shape
pond system. Renew Energy 145:1620–1632
322 M. More et al.
23. Risberg K, Cederlund H, Pell M, Arthurson V, Schnurer A (2017) Comparative characterization

of digestate versus pig slurry and cow manure—chemical composition and effects on soil
microbial activity. Waste Manage 61:529–538
24. Sievers DA, Tao L, Schell DJ (2014) Performance and techno-economic assessment of
several solid–liquid separation technologies for processing dilute-acid pre-treated corn stover.
Bioresour Technol 167:291–296
25. Khan SA, Malav LC, Kumar S, Malav MK, Gupta N (2014) Resource utilization of biogas
slurry for better yield and nutritional quality of baby corn. Adv Environ Agric Sci 32:382–394
26. Aguirre-Villegas HA, Larson RA, Sharara MA (2019) Anaerobic digestion, solid-liquid sepa-
ration, and drying of dairy manure: measuring constituents and modeling emission. Sci Total
Environ 696:134059
27. Holly MA, Larson RA, Powell JM, Ruark MD, Aguirre-Villegas H (2017) Greenhouse gas and
ammonia emissions from digested and separated dairy manure during storage and after land
application. Agric Ecosyst Environ 239:410–419
The Use of Photoplethysmography
for Blood Glucose Estimation
by Noninvasive Method
Vandana C. Bavkar and Arundhati Shinde
Abstract Noninvasive blood glucose estimation system based on photoplethysmog-

raphy is proposed in this paper. Photoplethysmography (PPG) is an optical tech-
nique which uses infrared light to determine the fluctuations in blood stream in the
human body with every pulse. However literature reveals that, a functional corre-
lation exists among pulse signal waveform and the glucose contents in the blood.
Many researchers have shown estimation of glucose using different spectroscopy
techniques which involves an optical sensor. The key involvement of this paper is
withdrawal of different features with respect to time domain as well as frequency
domain. Exploration of single pulse is also carried out with different features. A
PPG data of 182 random people with varied health conditions like diabetic, blood
pressure condition is recorded for about 1 min duration. The glucose estimation is
done in two ways: the first one is feature extraction from complete PPG signal of
a patient and the second one is single pulse wave analysis. All the features from
above-mentioned methods are transformed into feature vector which acts as input
to system and estimation of blood glucose is performed. We used Clarke Error Grid
Analysis for blood glucose estimation which is clinically accepted. With time and
frequency domain features we got 43.75% samples in region A and 50% samples in
region B, so total 93.75% samples are in regions A and B. Using single pulse wave
analysis we got 59.66% samples in region A and 34.61% samples in region B, so total
94.27% samples are in clinically accepted regions (A and B). As single pulse wave
analysis shows significant improvement in region A (20% of actual blood glucose
value), this method is good selection for glucose measurement.
Keywords Blood glucose · Noninvasive · Photoplethysmography · Feature

extraction · Neural network
V. C. Bavkar (B) · A. Shinde

Department of Electronics, Bharati Vidyapeeth (Deemed to be University), College of
Engineering, Pune 411043, India
e-mail: cvaidehi@gmail.com
A. Shinde
e-mail: aashinde@bvucoep.edu.in
https://doi.org/10.1007/978-981-19-6913-3_21
324 V. C. Bavkar and A. Shinde
1 Introduction
Diabetes mellitus which is also known as diabetes is a group of metabolic disorders

caused due to high glucose level for a long period of time. Untreated diabetes may
lead to serious damages to heart, eyes, blood vessels, and nerves. Diabetes occurs due
to one of the two mechanisms. The first one is insufficient creation of insulin or insuf-
ficient sensitivity of cells to the action of insulin. Here, insulin is a type of a hormone
released by pancreas. Inside the cells, glucose is converted to energy. Depending
upon insulin use, Type1 and Type2 diabetes was found in people. Type1 diabetes is
also called as insulin-dependent diabetes. In Type1 diabetes pancreas doesn’t make
insulin, because of this glucose is not utilized by the body’s cell to get energy. Treat-
ment for Type1 diabetes involves injecting insulin into fatty tissues of your body.
Type2 diabetes is non-insulin-dependent diabetes. In this case, pancreas make insuf-
ficient insulin treatment for Type2 diabetes which includes oral medications [1–4].
Currently existing techniques for blood glucose measurements are invasive in nature,
and these methods are very painful and having discomfort in nature. This is the main
cause of enormous research in noninvasive glucose measurement technique which is
very comfortable and mainly no need of blood. Infrared spectroscopy is one of the
popular techniques for this purpose. NIR spectroscopy is used for proposed system
[5–11].
Photoplethysmography (PPG) is simple optical method to observe blood volume
changes in circumferential circulation in cardiovascular system. It is low cost, nonin-
vasive, and portable method because of which in last few decades this technique is
commonly used in various modern health care systems. PPG signal also gives infor-
mation about respiratory systems [12–16]. Researchers found that there is a correla-
tion among PPG signal and glucose level [17]. Yadav describes a PPG signal-based
system used to measure blood glucose. Galvanik skin response, temperature, and PPG
data are used as input to machine learning system. The multisensory system consid-
erably improves prediction error and correction coefficient in contrast to previous
methods in literature [18]. Vahlsing describes noninvasive blood glucose measure-
ment technique based on color-coded photoplethysmographic images of finger in
NIR and visible range [19]. Cruz suggests application of PPG signal for analysis
of glucose with the use of moving average filter. He found the correlation between
glucose and peak-to-peak voltage. For non-diabetic patient, it is direct proportion, and
for diabetic patient, it is indirect proportion [20]. Haroon developed a prototype for
noninvasive measurement of blood glucose, pulse rate, and oxygen saturation level
[21]. Philip developed a blood glucose monitoring system with acquired PPG signal
along with bioimpedance. Prediction error gets reduced with bioimpedance method
[22]. Venkataramanan proposed a blood glucose prediction system with NIR sensor
output voltage. With this sensor, voltage in regression model, prediction of glucose
is done with mean error of 0.8558 [23]. Chowdhury describes a blood glucose esti-
mation using smart phone video which collects data from measurement site. These
video frames are later transformed into PPG waveform. Principle component regres-
sion is useful for estimation of glucose from extracted features. The projected model
The Use of Photoplethysmography for Blood Glucose Estimation … 325
could estimate the glucose getting a standard error of 18.31 mg/dL for testing dataset
[24].
2 Methodology
2.1 System Diagram
The workflow for anticipated system is given in Fig. 1. It consists of patient dataset
module which created using pulse signal acquisition system. This data is given to
preprocessing module to filter the noisy signal. This signal is given to different feature
extraction modules. Extracted features are converted to feature vector and given to
system where the blood glucose values are predicted.
Fig. 1 System diagram

Fig. 2 NIR signal acquisition system
2.2 Database
Data is collected from 182 people having 25–85 age groups chosen arbitrarily. These
people are having various medical circumstances like high or low blood pressure,
healthy, diabetic, and non-diabetic. The different attributes like age, gender, sex, skin
tone, blood pressure level are considered for creation of in vivo database.
2.3 Data Acquisition System
Near infrared (NIR) optical sensor is used for recording of PPG signal from finger
as body site. When NIR light is passed through body site, a pulse signal is acquired
by enlightening the skin and calculating variations in the absorption of light. The
absorption depends upon the heart rate as the blood vessel through finger contract
and expands in accordance with heart beat. At the output of optical detector, we get
reflected signal in the form of pulse. Different parameters like arterial stiffness, pulse
rate, blood pressure, and glucose level can be analyzed using PPG signal analysis
[25]. Figure 2 shows NIR signal acquisition system.
2.4 Preprocessing of Signal
The output of detector contains some noise due to motion artifacts, power line inter-
ference. The quality of detector output depends on skin thickness, i.e., length of an
optical path, blood flow, and skin tone at the measurement. Noisy signal will affect the
feature extraction and blood glucose measurement. Therefore, preprocessing of this
noisy signal is necessary step. The input sequence x(n) for n samples is represented
as
x(n) = {x1 , x2 , x3 . . . , xn } samples.
The filter used for preprocessing is moving average filter which removes the
baseline drifting from the signal. The output of moving average filter is given by the
following equation:
N −1
1
y(n) = x(n − i ) (1)
N i=0
where
y(n) output of the filter

x(n − i) previous input
N length of the filter (10 in our case).
2.5 Feature Extraction
Time and frequency domain analysis is done for feature extraction. Analysis of phys-
ical signal with respect to time is time domain analysis and when signal is analyzed
with frequency, it is frequency domain analysis. For these features, we have calcu-
lated statistical measures like mean (μ), quartiles (interquartile range) (iqr), variance
(σ ), and skewness (skew). Statistical measures quantitatively describe or summarize
features of collection of data or information. Statistical measure ‘mean’ measures the
central tendency of the data. Interquartile range is used for measuring dispersion of
data. It is the difference between 25 and 75th quartile. Variance shows how individual
number relates to each other within dataset. It is calculated by arithmetic mean of the
squares of deviations of all values in a set of numbers form their arithmetic mean.
Skewness gives distortion or asymmetry in a normal distribution [17, 18]. Figure 3
shows PPG signal showing features like pulse interval, peak interval, etc.
Different frequency domain features are extracted and calculated the statistical
measures for each feature.
Regression Coefficients
Regression coefficient basically used to forecast upcoming values from past values.
These coefficients describe the envelope of the spectrum. When we apply this on
PPG signal, we can track the variations in shape of the signal because of blood flow.
The coefficients are denoted as A0 , A1 , A2 , A3 , A4 , A5 .
Power Spectral Density (PSD)
This feature gives information about power present in the signal with respect to
frequency. It also evaluates damping of signal, occurrence of noise, and the spectral
Fig. 3 PPG signal
harmonic components. Statistical features mean PSDμn , interquartile range PSDiqr

n ,
variance PSDσn , and skewness PSDskew
n are calculated.
Kaiser–Teager Energy (KTE)
This is popular technique used for measuring energy of signal. For differentiating
clean and noisy signal, this property is very useful. High energy signifies high ampli-
tude signal with good quality and low energy signal signifies noisy signal with low
amplitudes [17]. Statistical features like skewness, variance, interquartile range, and
mean are calculated.
Time domain features show signal variation over the time. Heart rate, pulse transit
time, and peak interval these features are calculated from PPG signal.
Heart Rate (HR)
Heart rate is one of the factors affecting blood glucose level. Hence, this can be
calculated from peaks of the PPG signal. Statistical features mean HRμn , interquartile
σ
range HRiqr
n , variance HRn , and skewness HRn
skew
are calculated.
Pulse Transit Time (PTT)
The travel time of pulse between two arterial points is referred to as pulse transit time.
For this measurement, we have to consider one frame. Statistical features mean PTTμn ,
σ
interquartile range PTTiqr
n , variance PTTn , and skewness PTTn
skew
are calculated.
Peak Interval (PPInt)
Peak interval is the width between consecutive peaks of PPG signal considering one
frame. Statistical features mean PPIntμn , interquartile range PPIntiqr σ
n , variance PPIntn ,
skew
and skewness PPIntn are calculated [26–29].
Table 1 Time and frequency domain features

Feature Statistical measures (mean, interquartile range, No. of features
variance, and skewness)
Auto-regression coefficients A0 , A 1 , A 2 , A3 , A 4 , A 5 6
μ
KTEn , KTEn , KTEσn , KTEskew
iqr
Kaiser-Teager energy n 4
μ
PSDn , PSDn , PSDσn , PSDskew
iqr
Power spectral density n 4
μ
HRn , HRn , HRσn , HRskew
iqr
Heart rate n 4
μ
PTTn , PTTn , PTTσn , PTTskew
iqr
Pulse transit time n 4
μ iqr σ
Peak-to-peak interval PPIntn , PPIntn , PPIntn , PPIntskew
n 4
μ
PAmpn , PAmpn , PAmpσn , PAmpskew
iqr
Peak amplitude n 4
Total 30
Peak Amplitude (PAmp)

Peak amplitude is the systolic pulse of PPG signal which shows rhythmic changes in
blood because of arterial blood flow. Statistical features mean PAmpμn , interquartile
σ
range PAmpiqr
n , variance PAmpn , and skewness PAmpn
skew
are calculated.
Till now we have calculated 30 different features listed in Table 1.
3 Single Pulse Wave Analysis
PPG signal is divided into single pulses, and each single pulse used for feature
extraction is shown in Fig. 4.
Features of single pulse:
• Width period (T )
Fig. 4 Single pulse wave
T0 B
T2 C
T1 T
Width period
• Highest peak value (A)

• Time of highest peak value (T 0)
• Diastolic peak amplitude (B)
• Time of Diastolic peak (T 1)
• Notch amplitude (C)
• Time of notch (T 2)
• Time Difference Start-Peak (T 0)
• Time Difference Peak-Notch (T 2 − T 0)
• Time Difference Notch-Diastolic Peak (T 1 − T 2)
• Time Difference Diastolic Peak-End (T − T 1)
• Mean amplitude
• Standard deviation of single pulse
• Length of Single period
• Mean Start-Max
• Mean Max-Notch
• Mean Notch-Diastolic Peak
• Mean Diastolic Peak-End.
We have calculated 18 features from single pulse wave. Feature dimension method
is used to reduce number of features as data samples are huge in number. Only seven
features are inputted to system. The network configuration details are summarized
in the following section.
3.1 Neural Network Configuration
In last few decades, neural networks are recursively used in medical diagnosis and
healthcare systems [30, 31]. All network structure is built in neural network toolbox of
MATLAB software. At start, the testing is performed with different neurons present
in hidden layer which affect the accuracy of the system. The hidden layer neurons
are calculated by following equation.
I+O √
N= + S (2)
2
where
N Number of neurons
I Number of inputs
O Number of outputs
S Total samples.
The neurons considered in the experimentations are ±5 of neurons calculated as
per above equation.
3.2 Clarke Error Grid Analysis
Clarke Error Grid Analysis is a crucial tool for examining the scientific accuracy
of blood glucose monitors. It shows the graphical relationship between reference
glucose value versus predicted glucose value. Error grid divides graph into five
different regions. In region A, samples are within 20% range of reference value.
In region B, samples are outside of Region A but this does not lead to inappro-
priate medication, also called ‘Clinically uncritical decisions’. Region C samples are
directing to redundant medication. Region D samples indicate hazardous malfunc-
tion to identify hypoglycemia or hyperglycemia, and lastly Region E samples would
mystify medication of hypoglycemia for hyperglycemia. Region A and Region B are
the clinically accepted regions [32].
3.3 Performance Measure
Clarke Error Grid Analysis is used as performance measure for different neurons in
hidden layer. The amount of samples in region A and region B varies according to
deviation or change in number of neurons. MATLAB software is used to build neural
networks and Clarke error grid graphs.
4 Results
4.1 Time and Frequency Domain Features
Thirty features were taken out from pulse signal and given as input to network. At
the network output, we get predicted blood glucose concentration level. The total
data samples are separated to train and test dataset. 80% samples (150 samples) are
applied to train the network and 20% (32 samples) samples are for testing purpose.
According to Eq. (1), hidden layers were calculated in the range of 24–34. With
different neuron configuration, (30, 1) as shown in Fig. 5, we get better accuracy
than other neuron configurations. Clarke error grid graph is shown in Fig. 6 for this
configuration.
With time and frequency related features, 43.75% samples in region A and 50%
samples in region B, so total 93.75% samples are in region A and region B.
Fig. 5 Two layer neural network detailed architecture
Fig. 6 Clarke Error Grid Analysis using time and frequency domain features
4.2 Single Pulse Wave Features
Four hidden layers are used as shown in Fig. 7. Total 12,196 single wave pulses
of 182 patients are there. 80% samples (10,012 samples) are applied to train the
neural network and 20% (2184 samples) samples are applied for testing purpose.
Starting with two layers (15, 10) the network gives optimum results. Hence, we have
Fig. 7 Four layer neural network architecture
to increase one layer with (15, 10, 5) neurons. Further testing is done with different
neuron combinations, and (20, 15, 10) configuration gives better accuracy with error
grid graph samples as shown in Fig. 8.
Using single pulse wave analysis, 59.66% samples are in A region and 34.61%
samples in B region, so total 94.27% samples are in A and B regions.
Fig. 8 Clarke Error Grid Analysis using single wave features

5 Conclusion
Proposed system gives accuracy with 94.27% with single pulse wave analysis and
that with time and frequency domain feature set is 93.75%. During this research work,
different time domain and frequency domain features are taken out from PPG signal
and from pulse wave. As single pulse wave analysis shows significant improvement
in region A (20% of actual blood glucose value) of Clarke Error Grid Analysis, this
method is good selection for noninvasive blood glucose measurement. Extraction of
derivative features is the extended task for our research work for better results.
Acknowledgements We would like to express thanks to Bharati Hospital and Research Center,
Pune, India for allowing us to gather data from people in pathology laboratory.
References
1. International Diabetes Federation. About diabetes. http://www.idf.org/about-diabetes

2. Kovatchev BP (2012) Diabetes technology: markers, monitoring, assessment, and control of
blood glucose fluctuations in diabetes. Scientifica 2012, 14 pages. Article ID 283821
3. Migdalis I, Leslie D, Papanas N, Valensi P, Vlassara H (2014) Diabetes mellitus. Int J Endocrinol
2014, 6 pages. Article ID 108419
4. Chen C, Zhao X-L, Li Z-H, Zhu Z-G, Qian S-H, Flewitt AJ (2017) Current and emerging
technology for continuous glucose monitoring. Sensors
5. Senthi Kumar A, Kavitha S, Muthubharathi R (2017) Design and development of absorption
spectrophotometric based non-invasive blood glucose measuring device. Res J Pharm Technol
10(1):91–97. https://doi.org/10.5958/0974-360X.2017.00022.1
6. Gamessa TW, Suman D (2019) Non-invasive blood glucose monitoring using visible laser
light. Res J Pharm Technol 2(2):831–840
7. Premalatha J, Grace Kanmani Prince P (2019) Analysis of non-invasive methods to diagnose
blood glucose level—a survey. Res J Pharm Technol 12(6):3105–3108
8. Thakur J (2019) Effect of maternal diabetes on fetus and newborn. Asian J Nurs Educ Res
9(3):463–465
9. Sirisha (2020) Recent advancements on colon targeted drug delivery systems. Asian J Pharm
Res 10(4):268–274
10. Saudagar RB, Samuel S (2016) Ethosomes: novel noninvasive carrier for transdermal drug
delivery. Asian J Pharm Technol 6(2):135–138
11. Ravindra HN, Christian K, Pooja G, Prem P, Priyal J, Rajat A, Riya R (2016) Knowledge
and attitude on self monitoring of blood glucose (SMBG) among diabetic patients belongs to
Waghodia taluka. Int J Adv Nurs Manage 4(4):398–403
12. Awasthi A, Sharma N (2014) Possibilities of using neural network for ECG classification. Res
J Eng Technol 5(1):13–16
13. Walia MS (2017) Performance analysis of feature extraction techniques for iris pattern
recognition system. Res J Eng Technol 8(4):431–435
14. Rajaguru H, Prabhakar SK (2018) Wavelet neural networks, Elman backpropagation and
multilayer perceptrons for epilepsy classification from EEG signals. Res J Pharm Technol
11(4):1301–1306
15. Meenakshi K, Maragatham G (2019) Computational intelligence in diagnosis and prognosis
of gestational diabetes using deep learning. Res J Pharm Technol 12(8):3891–3895
16. Moraes JL, Rocha MX (2018) Advances in photoplethysmography signal analysis for
biomedical applications. Sensors
17. Habbu S, Dale M, Ghongade R (2019) Estimation of blood glucose by non-invasive method
using photoplethysmography. Sadhana
18. Yadav J, Rani A, Singh V, Murari BM (2016) Investigations on multi-sensor based non-invasive
blood glucose measurement system. J Med Devices
19. Vahlsing T, Delbeck S, Leonhardt S, Michael Heise H (2018) Noninvasive monitoring of
blood glucose using color-coded photoplethysmographic images of the illuminated fingertip
within the visible and near-infrared range: opportunities and questions. J Diabetes Sci Technol
12(6):1169–1177
20. Cruz FRG, Paglinawan CC, Catindig CNV (2019) Application of reflectance mode photo-
plethysmography for non-invasive monitoring of blood glucose level with moving average
filter. In: International conference on biomedical engineering and technology, ICBET 2019
21. Haroon N, Tiwana MI (2017) Design and development of non-invasive prototype to measure
pulse rate, blood glucose and oxygen saturation level in arterial blood. In: Future technology
conference (FTC), Canada, Nov 2017, pp 226–233
22. Philip LA, Rajasekaran K, Jothi E (2017) Continuous monitoring of blood glucose using
photophlythesmograph signal. In: Proceedings of IEEE international conference on innovations
in electrical, electronics, instrumentation and media technology, ICIEEIMT 2017
23. Venkataramanan S, Kamble D, Bairolu A, Singh A, Rao R (2017) A novel heart rate and non-
invasive glucose measuring device. In: International conference on communication and signal
processing, ICCSP 2017
24. Chowdhury TT, Mishma T, Osman MS, Rahman T (2019) Estimation of blood glucose level
of type-2 diabetes patients using smartphone video. In: Proceedings of the 6th international
conference on networking, system and security, Dec 2019, pp 104–108
25. Habbu SK, Joshi S, Dale M, Ghongade R (2019) Noninvasive blood glucose estimation using
pulse based cepstral coefficients. In: IEEE international conference on signal processing and
information security (ICSPIS)
26. Zhang Y, Zhang Y, Siddiqui SA, Kos A (2019) Non-invasive blood-glucose estimation using
smartphone PPG signals and subspace KNN classifier. Electrotech Rev 86(1):68–74
27. Jubadi W, Sahak S (2009) Heartbeat monitoring alert via SMS. In: Proceedings of IEEE
symposium on industrial electronics & applications, vol 1, pp 1–5
28. Fu T, Liu S, Tang K (2008) Heart rate extraction from photoplethysmogram waveform using
wavelet multi-resolution analysis. J Med Biol Eng 28(4):229–232
29. Linder S, Wendelken S, Wei E, McGrath S (2006) Using the morphology of photoplethysmo-
gram peaks to detect changes in posture. J Clin Monit Comput 20(3):151–158
30. Bukhari MM, Alkhamees BF, Hussain S, Gumaei A, Assiri A, Ullah SS (2021) An improved
artificial neural network model for effective diabetes prediction. Complexity 2021, 10 pages.
Article ID 5525271
31. Bavkar VC, Shinde AA (2021) Machine learning algorithms for diabetes prediction and neural
network method for blood glucose measurement. Indian J Sci Technol 14(10):869–880. https://
doi.org/10.17485/IJST/v14i10.2187
32. Clarke WL, Cox D, Gonder-Frederick LA, Carter W, Pohl SL (1987) Evaluating clinical
accuracy of systems for self-monitoring of blood glucose. Diabetes Care 10(5):622–628
Single-Stage Stand-Alone Induction
Motor Driven Solar Water Pumping
System with Minimal Sensors
Anup Shetty , K. Suryanarayana , and L. V. Prabhu
Abstract Water pumps are essential to extract ground water for agricultural purpose
and majority of these pumps are driven by electric motors. However, in remote
places due to power scarcity, people rely upon diesel-powered pumps. Solar-powered
water pumping has gained attention in the recent years because of increased aware-
ness about global warming, caused by the greenhouse gases released by the fossil
fuel-based systems and encouraging government policies to reduce carbon foot-
print. Efforts are made to develop a cost effective and high performance solar water
pump drives. A single-stage induction motor driven photovoltaic water pumping
system with minimal sensors is presented in this paper. Proposed system consists
of PV array, power inverter stage and scalar (V /f ) controlled induction motor (IM)
to drive water pump. In the proposed system, PV array stage is connected to the
DC bus of power inverter stage eliminating intermediate DC-DC converter stage to
reduce overall cost and size. Optimal utilization of the PV system is assured with
the help of variable step size-incremental conductance (VSS-INC) maximum power
point tracking (MPPT) algorithm. Elimination of feedback signals in scalar control
approach reduces the sensor count. The proposed system has been modeled and simu-
lated using MATLAB/Simulink tool. System performance and effectiveness under
atmospheric parameter fluctuations are simulated in detail.
Keywords Maximum power point tracking · Photovoltaic system · Voltage source

inverter · Variable step size-incremental conductance · Solar water pumping · V/f
control
A. Shetty (B) · K. Suryanarayana

Department of Electrical and Electronics Engineering, NMAM Institute of Technology, Nitte,
Karkala, Karnataka 574110, India
e-mail: anup.shetty@nitte.edu.in
K. Suryanarayana
e-mail: suryanarayana@nitte.edu.in
L. V. Prabhu
HEXMOTO Controls Pvt. Ltd, Mysuru 570018, India
e-mail: prabhu@hexmoto.com
https://doi.org/10.1007/978-981-19-6913-3_22
338 A. Shetty et al.
1 Introduction
Renewable energy conversion and utilization is gaining interest due to the rise in
adverse effects of environmental pollution caused by burning fossil fuels and esca-
lating worldwide energy crisis [1]. Solar radiation energy is available in the cleanest
form and abundant even in remote places making it favorite among all other renew-
able energy sources [2]. The recent advancements in photovoltaic cell technology
and rapid growth in the semiconductor industry have empowered development of
efficient solar energy harvesting solutions with minimal maintenance. Mass produc-
tion of PV cells has led to steady drop in cost per watt-peak over the years. Possibility
of increasing installed power capacity by adding panels to the existing system is one
of the attractive features of solar energy harvesting. However, intermittence, high
initial cost and low efficiencies are some of the drawbacks of current PV systems.
Water pumping is a vital requirement for agriculture. Solar water pumping is
most useful for remote areas with ample solar irradiance [3]. In solar water, pumping
storage features such as batteries are not essential since water can be pumped and
stored in tanks whenever the irradiance is available [4]. Based upon the water
pumping requirement, a solar PV array with an appropriate rating could be selected.
In the case where the water pump needs to be operated throughout the day with full
power, a solar PV array with a battery system could be used. Energy stored in the
battery even could be used to energize household appliances in remote areas. The
natural connection present among the necessity of water pumping and the availability
of solar irradiance, depending upon season requirement of water pumping changes.
Solar water pumping (SWP) systems offer several advantages over the traditional
fossil fuel-based generator powered pumps in terms of enhanced reliability, reduced
operation and maintenance costs [5]. Solar energy is available only in the daytime
and it is affected by environmental situations, weather changes and clouds. To maxi-
mize the energy extracted, the PV array should be operated at its point of maximum
efficiency, called as maximum power point (MPP) [1, 6]. To locate maximum power
point, tracking algorithms are being used based on the requirement. Several algo-
rithms for MPP tracking have been presented by the researchers over the times.
Perturb and Observe algorithm (P&O) is one of the simplest and effective tracking
methods but has disadvantages such as oscillating around maximum power point,
low response speed and low accuracy [7]. Incremental conductance (INC) MPPT
algorithm overcomes the disadvantages of P&O algorithm and dynamic response
could be improved with the help of variable step size technique [2].
Induction motors are commonly used in water pumps for their robustness and
low cost despite low efficiency when compared to advanced motors [8]. Perma-
nent magnet synchronous motors (PMSMs) could be an alternative for induction
motors. PMSM has advantages of high efficiency, low noise, high air gap flux
density, high power-to-weight ratio, swift acceleration and deceleration capability,
high power factor and compact design [2] but suffers from high initial cost and control
complexity. Benefits of advanced motors are surpassed by the cost effectiveness and
Single-Stage Stand-Alone Induction Motor Driven Solar Water … 339
easy controllability of induction motor and is suitable for economical water pumping
solution.
Solar PV array generates DC voltage which needs to be converted to AC form with
the help of inverter. Single-stage IM drives suffer from DC voltage fluctuation [9, 10].
Feed forward-based reference speed control with VSS-INC is used to minimize the
DC voltage fluctuations and to improve dynamic performance. Speed of the motor
is varied based upon the maximum power point tracker algorithm. Motor control
in photovoltaic pumping must be focused on captivating maximum power. Power
control of the induction motor from the PV is demanding because the motor mechan-
ical time constant and electrical time constant of the system are different. Dynamic
power flow control is required for effective operation of the system. Several IM
control methods including scalar control, V/f control, vector control, field-oriented
control and direct torque control are present in the literature [7]. Field-oriented control
strategies are employed to adapt high performance induction motor control [5] with
additional current sensors. Vector control methods involve complex transformations
and additional current and position sensors. However, scalar control is good enough
as accurate speed control is not required for water pumps. Scalar control is chosen
for the system development because of its simplicity.
This paper is organized as follows: Sect. 1 presents a brief introduction to the
induction motor driven solar water pump. Section 2 discusses the system architecture
of the proposed methodology. Control scheme of the solar-powered pump is discussed
in Sect. 3. Section 4 presents the system modeling and simulation aspects. Section 5
gives the simulation results and observations and Sect. 6 presents the conclusion.
2 System Architecture
The system configuration of the proposed scheme for stand-alone solar-powered

water pumping is as in Fig. 1. The water pumping mechanism is driven by an induc-
tion motor controlled with conventional scalar control technique. The system consists
of a PV array and three-phase voltage source inverter to control IM. Suitable number
of PV modules are connected in series and parallel combination to obtain desired
voltage and power ratings. The PV array is connected to the DC bus of VSI, elimi-
nating intermediate DC-DC converter which is commonly used to implement solar
MPPT algorithms. Power captivated from the PV array is being controlled with the
help of variable step size-incremental conductance MPP tracking technique to draw
maximum power with available radiation. Speed reference for the scalar control
structure of IM drive is generated using MPPT algorithm.
Fig. 1 System configuration
2.1 Solar PV Array Design
Photovoltaic cells are used to convert solar radiation energy into electrical energy.
Power rating choice of the PV array depends upon the load requirement. PV array
rating is designed by considering the losses in power converter stage and maximum
irradiance available in the area of system installation. PV cells with power rating
slightly greater than the peak load power could be selected to compensate for the
losses. The proposed system makes use of a 4 kW induction motor; hence, PV array
with peak capacity of 4.4 kW under standard test condition (1000 W/m2 , 25 °C) is
preferred.
PV array is formed by connecting solar cell modules in series and parallel manner
to obtain desired voltage and power level. Standard solar module A10J-M60-220 by
A10 Green Technology is considered for the design.
PV array maximum power,
PMP = 4.4 kW (1)
Maximum power of single module,
PM = 219.87 W (2)
Total number of modules required is calculated as,
PMP 4.4 k ∼
n= = = 20 (3)
PM 219.87
Table 1 Solar module and

Parameters Value
PV array specifications
Maximum power of single module 219.876 W
Open-circuit voltage of module (V OC ) 36.06 V
Short-circuit current of module (I SC ) 7.95 A
Voltage at maximum power point (V MPP ) 30.12 V
Current at maximum power point (I MPP ) 7.3 A
Number of modules connected in array 20
PV array open-circuit voltage (V OC ) 721.2 V
PV array short-circuit current (I SC ) 7.95 A
PV array maximum power (PMP ) 4.4 kW
PV array voltage at maximum power (V MP ) 602.4 V
PV array current at maximum power (I MP ) 7.3 A
For a single-stage topology, V MP is selected near to V DC voltage level. Rated voltage

of motor is 400 V; hence, DC link voltage could be chosen as,
√
VDC = 2 ∗ VL−L ∼
= 566V (4)
VDC is taken as 600 V. To obtain the desired DC link voltage, 20 solar modules
are required to be connected in series. Specifications of the solar module and PV
array are as in Table 1.
2.2 DC Link Capacitor Selection
DC link capacitor is used to support the DC voltage level under transient conditions
and the value of CDC is calculated as [2],
IO 4400/600 ∼
CDC = = = 390 µF (5)
2 ∗ ω ∗ VDC 2 ∗ 314.16 ∗ 0.05 ∗ 600
where Io = PMP /VDC , ω = 2 ∗ π ∗ 50 = 314.16 rad/s and DC bus voltage ripple

is taken as 5% of VDC . Standard capacitor of 470 µF is considered for the system
design.
3 Control Methodology
Control section of the proposed system primarily consists of variable step size-
incremental conductance MPPT algorithm and V /f control for the IM drive. Voltage
and current signals of PV array are sensed and fed to the MPP tracker to extract
maximum power. Induction motor is driven with a scalar control technique which
greatly reduces the control strategy complexity. VSS-INC algorithm generates the
DC link voltage reference (V DCref ). Error between the actual DC link voltage (V DC )
and reference DC link voltage (V DCerror ) is fed to a PI controller to reduce the voltage
error by adjusting the speed reference value.
Assuming lossless system and steady state operation, output and input power
could be related as,
PPV = PCDC + Pm (6)
where PPV , PCDC and Pm are the power of PV array, DC link capacitor and motor,
respectively.
Under steady state, PCDC will be zero thus,
PPV = Pm (7)
where
Pm = τm ∗ ωm (8)
For a pump-type load using pump affinity law, speed and torque relationship for the
motor is given as [1],
τm ∝ ωm 2 (9)
Considering K m as the proportionality constant, Eq. (9) becomes,

τm
Km = (10)
ωm 2
where τm is the pump load torque under steady state condition in N-m and ωm is the
motor speed in rad/s. Using (10), Eq. (7) could be written as,
PPV = K m ∗ ωm 3 (11)
From Eq. (11), it is evident that the power absorbed from the PV array (PPV ) is
directly proportional to the cube of the motor speed (ωm ). Speed reference could be
generated with the knowledge of power supplied by the PV array for effective MPP
tracking. Accordingly, two speed reference components are generated as in Fig. 2.
ωref1 from the DC link voltage regulator and ωref2 from the speed feedforward term
employing the relation mentioned in Eq. (11). Speed feed forward term improves
the system performance under dynamic conditions. DC link voltage regulating PI
controller is used to fine-tune the ωref1 . Reference frequency for the V /f control is
Fig. 2 Control diagram
obtained from ωref1 and ωref2 . Sinusoidal pulse width modulation (SPWM) technique
is used to generate the control pulses for the VSI.
3.1 Variable Step Size-Incremental Conductance MPPT
Solar PV array PPV versus VPV curve is nonlinear in nature as in Fig. 3. For a
fixed irradiance and temperature, power delivered by the PV cells greatly depend on
the load impedance. To operate the solar array at the MPP, a tracking algorithm is a
mandatory requirement. DC-DC converters in conjunction with the MPPT algorithms
are universally used to match the load and PV array impedances for maximum power
absorption. P&O technique suffers from the steady state oscillation about MPP and
poor performance under irradiance changes. INC is another MPP tracking technique
which is based on the slope of PV characteristics curve. INC basically monitors the
slope of PV characteristics curve and makes use of the fact that at MPP, slope of the
PV curve is zero to track MPP. Dynamic performance under irradiation changes is
better in case of INC.
Incremental conductance method is conventionally implemented with fixed step
size for MPP tracking. Fixed step sized INC suffers from the possibility of oscillation
about MPP under steady state, leading to loss of power. Even though slope of PPV
versus VPV curve is tracked, steady state oscillation depends on the chosen step size.
Step size could be made small to avoid power oscillation but dynamic performance
of INC with small step size is poor and slower. These problems could be minimized
by employing variable step size, depending upon transient conditions, step size is
recalculated.
Flow diagram of the VSS-INC is as in Fig. 4. MPPT algorithm of the proposed
system controls the V DCref as,
VDCref (k) = VDCref (k − 1) ± step (12)
Variable step size is calculated in each iteration and is given as,

Fig. 3 PV characteristics

dPPV
step = K s ∗ (13)
dVPV
where d PPV and dVPV are the change in PV power and PV voltage from present
sampling instant (k) to previous instant (k − 1), respectively. Here, K s is the step size
constant which is chosen appropriately by considering system stability. As system
operating point reaches MPP, value of step tends toward zero, eliminating steady
state oscillations. Maximum value of step size should be limited to stepmax so that
under starting and transient conditions, speed of the tracking algorithm is limited.
To ensure convergence, step size of VSS-INC should follow the condition as in

d PPV

K s < stepmax / (14)
dVPV step=stepmax
3.2 Induction Motor V/f Control
Scalar or V /f is the common open-loop speed control technique used to drive induc-
tion motors. IMs are designed to operate at rated torque and speed for rated terminal
voltage and frequency. When the variable speed is desired, voltage needs to be
reduced and to get the rated torque performance of the IM, air gap flux needs to
be maintained constant. This is achieved by varying both magnitude and frequency
of terminal voltage so that IMs could perform best at all speed ranges. Air gap flux
(ϕ) is related to voltage (V ) and frequency ( f ) as,
Fig. 4 VSS-INC algorithm flowchart
V
ϕ∝ (15)
f
SPWM-controlled VSI is used to achieve scalar control of IM. Reference

frequency is given based upon speed requirement neglecting the slip speed. Since V /f
constant for a motor is known, corresponding modulation index and hence voltage
magnitude is calculated. Three sinusoidal reference signals with desired frequency,
magnitude and 120 ◦ apart from each other are generated. Reference signals are
compared with the carrier wave to generate desired SPWM control signals.
4 Modeling and Simulation
Individual modules are modeled to form overall system model. PV array, induction
motor and water pump load are modeled to match the practical system behavior.
Static single-diode model is used for the system analysis [11, 12]. PV voltage and
current relationship is given as,

VPV + IPV Rs
IPV = Iph − Id − (16)
Rsh

VPV + IPV Rs
Id = IS exp −1 (17)
ηn c VT
where Iph is the PV generated current, Id is the PV diode’s forward bias current, dark
saturation current is represented by IS , n c indicates number of series connected cells,
η indicates diode’s quality factor, Rs and Rsh are the series and shunt resistances,
respectively, VT is the thermal voltage.
Induction motor model in d-q reference frame rotating with motor speed is
considered for system simulation. Dynamic equations of IM [13–15] are given as,
dϕds
Vds = Rs i ds + + ωm ϕqs (18)
dt
dϕqs
Vqs = Rs i qs + − ωm ϕds (19)
dt
dϕdr
Vdr = Rr i dr + (20)
dt
dϕqr
Vqr = Rr i qr + (21)
dt
where ωm is the rotor speed, Rs is the stator phase resistance, Rr is the rotor phase
resistance, ϕds , ϕqs and ϕdr , ϕqr are the stator and rotor flux linkages in d-q frame,
respectively. Vds , i ds , Vqs , i qs and Vdr , i dr , Vqr , i qr are the stator and rotor voltages and
currents in d-q frame, respectively.
A three-phase 4kW, 400V induction motor is used to drive the pump load. To
model the load, pump affinity law is considered. Rated torque of the selected motor
is 26.71 N-m and rated speed is 1430 rpm (149.75 rad/s). From (10), the motor
Fig. 5 Simulink model of the system
constant K m could be obtained as,
26.71
Km = = 1.191 × 10−3 N-m/(rad/s)2 (22)
(149.75)2
Proposed system model is developed and simulated to evaluate the performance of

the control scheme. Simulation model developed is as in Fig. 5. PV array is formed by
connecting 20 solar modules in series, and variable irradiance and temperature signals
are given to the PV model to emulate real-world scenario. VSI stage is modeled with
the help of power semiconductor device model. Load torque for the motor model
is fed using the relationship given in Eq. (10). DC link voltage controller tuning is
carried out based on the desired performance of the system under momentary change
in operating conditions. Control strategy as explained in the control methodology
section is modeled and VSI control pulses are generated. PV voltage and PV current
signals are used to calculate reference speed for IM drive. Power delivered by the
PV array is monitored to ensure faithful tracking of the MPP.
Irradiance and temperature variation are simulated to analyze effectiveness of

proposed methodology. Nominal steady state irradiance is taken as 900 W/m2
at temperature of 25 ◦ C. PV dynamics are emulated by changing irradiation to
100W/m2 and temperature to 40 ◦ C as in Fig. 6.
PV voltage and current level vary during irradiation and temperature change.
Voltage and current sense signals of the PV array for an irregular irradiation and
temperature are as in Fig. 7. It is observed that the PV voltage is well regulated under
Fig. 6 Irradiance and temperature variation with time
steady state conditions and transient condition overshoots are limited well within the
acceptable range.
When the fall in irradiation is initiated at t = 2 s and temperature of the PV array
is increased at t = 6 s, PV power starts to decrease as in Fig. 8. Fall in the power
should be detected by the MPP tracker to find new MPP as operating characteristics
of the PV changes with irradiance and temperature fluctuation. Accordingly, speed
of the motor is adjusted to compensate for dynamic power variation.
Figure 9 depicts the induction motor speed and electromagnetic torque variation
for a given disturbance pattern.
Induction motor current and voltages are monitored to assess the need of soft
starting of the induction motor. Large starting current (34 A peak) is drawn when
Fig. 7 PV voltage and current

Fig. 8 PV power
Fig. 9 Motor speed and torque
the motor is operated without soft start as in Fig. 10 when MPP tracking is enabled
all the time. To avoid high starting current, soft start technique is used in the system.
Motor is driven by V /f control with slow ramp of speed reference signal without MPP
tracking algorithm till the motor reference speed reaches a predefined limit. Once
the reference speed attains set speed threshold, MPPT is engaged. Implementing soft
start improves the life of the induction motor. Motor current and voltages with soft
start are as in Fig. 11. It is observed that starting current drawn by the IM is reduced
to 7.5 A peak.
Fig. 10 Motor startup voltage and current
Fig. 11 Motor voltage and current with soft start
6 Conclusion
Model of the single-stage photovoltaic water pumping system with minimal sensors
is analyzed in this paper. Simulation of the proposed model is carried out to assess
the performance of the control strategy under various real-world conditions such
as motor startup, irradiance and temperature variation. Steady state and transient
performance of the VSS-INC algorithm are verified and it is observed that maximum
power point tracking is carried out with greater efficacy. Overall system size and cost
could be greatly reduced with the elimination of intermediate DC-DC converter stage.
Simulation results of the modest control structure is found to be satisfactory.
References
1. Singh B, Sharma U, Kumar S (2018) Standalone photovoltaic water pumping system using
induction motor drive with reduced sensors. IEEE Trans Ind Appl 54(4):3645–3655. https://
doi.org/10.1109/TIA.2018.2825285
2. Murshid S, Singh B (2020) Single stage autonomous solar water pumping system using PMSM
drive. IEEE Trans Ind Appl 56(4):3985–3994. https://doi.org/10.1109/TIA.2020.2988429
3. Montorfano M, Sbarbaro D, Morán L (2016) Economic and technical evaluation of solar-
assisted water pump stations for mining applications: a case of study. IEEE Trans Ind Appl
52(5):4454–4459. https://doi.org/10.1109/TIA.2016.2569415
4. Antonello R, Carraro M, Costabeber A, Tinazzi F, Zigliotto M (2017) Energy-efficient
autonomous solar water-pumping system for permanent-magnet synchronous motors. IEEE
Trans Ind Electron 64(1):43–51. https://doi.org/10.1109/TIE.2016.2595480
5. Maddalena ET, Moraes CGDS, Bragança G, Junior LG, Godoy RB, Pinto JOP (2019) A
battery-less photovoltaic water-pumping system with low decoupling capacitance. IEEE Trans
Ind Appl 55(3):2263–2271. https://doi.org/10.1109/TIA.2019.2900412
6. Caracas JVM, Farias GdC, Teixeira LFM, Ribeiro LAdS (2014) Implementation of a high-
efficiency, high-lifetime, and low-cost converter for an autonomous photovoltaic water pumping
system. IEEE Trans Ind Appl 50(1):631–641. https://doi.org/10.1109/TIA.2013.2271214
7. Elgendy MA, Zahawi B, Atkinson DJ (2015) Operating characteristics of the P&O algo-
rithm at high perturbation frequencies for standalone PV systems. IEEE Trans Energy Conver
30(1):189–198. https://doi.org/10.1109/TEC.2014.2331391
8. Jain S, Karampuri R, Somasekhar VT (2016) An integrated control algorithm for a single-stage
PV pumping system using an open-end winding induction motor. IEEE Trans Ind Electron
63(2):956–965. https://doi.org/10.1109/TIE.2015.2480765
9. Short TD, Mueller MA (2002) Solar powered water pumps: problems, pitfalls and potential.
In: 2002 international conference on power electronics, machines and drives (conf. publ. no.
487), pp 280–285. https://doi.org/10.1049/cp:20020129
10. Vitorino MA, de Rossiter Correa MB, Jacobina CB, Lima AMN (2011) An effective induction
motor control for photovoltaic pumping. IEEE Trans Ind Electron 58(4):1162–1170. https://
doi.org/10.1109/TIE.2010.2054053
11. Venkatramanan D, John V (2019) Dynamic modeling and analysis of buck converter based solar
PV charge controller for improved MPPT performance. IEEE Trans Ind Appl 55(6):6234–6246.
https://doi.org/10.1109/TIA.2019.2937856
12. Costa de Souza A, Cardoso Melo F, Lima Oliveira T, Eduardo Tavares C (2016) Performance
analysis of the computational implementation of a simplified PV model and MPPT algorithm.
IEEE Latin Am Trans 14(2):792–798. https://doi.org/10.1109/TLA.2016.7437224
13. MacDonald ML, Sen PC (1979) Control loop study of induction motor drives using DQ model.
IEEE Trans Ind Electron Control Instrum IECI-26(4):237–243. https://doi.org/10.1109/TIECI.
1979.351593
14. Lee RJ, Pillay P, Harley RG (1984) D, Q reference frames for the simulation of induction
motors. Electr Power Syst Res 8(1):15–26. ISSN 0378-7796. https://doi.org/10.1016/0378-
7796(84)90030-0
15. Krishnan R (2001) Electric motor drives: modeling analysis and control. Prentice-Hall,
Englewood Cliffs, NJ, USA, pp 196–213. ISBN 0130910147
Automation of Weight-Based Sorting
System Using Programmable Logic
Controllers
P. Chenchu Saibabu, R. Anjana, Manisha Kumari, and C. R. Srinivasan
Abstract Material sorting is an integral part of modern-day industries. In these

industries, the production are made in batches ranging from several dozen to several
thousand units, and thus, the raw material or any material used is also in abundance
for example in case of coal industry, scrap industry, or in any industry that it’s quantity
based or weight based factor that effects the production can be implemented. Using
manual labour for sorting of material is expensive, time consuming and inefficient.
This problem can be solved by automating the material sorting. In this paper, we
have discussed how we have developed a system that effectively sorts the material
according to weight using PLC as the controller. Sensor that is used for implementing
the idea is a load cell which was kept under the specially designed conveyor to
measure the weight of the object and then that in turn gives the analogous signal
which is amplified and given to the PLC unit which gives signals to the respective
solenoid valve and conveyor for sorting process.
Keywords Sorting system · ANSYS · Programmable logic controller · S7-1200
1 Introduction
Material sorting is a main concern in industries that deal with weight or quantity of
the materials. For example, in mining industries the raw material mined is of different
weights and sizes and requires different type of crusher according to their weight or
size, without a proper sorting system before going to the crusher the raw materials of
all sizes goes through all different types of crushes; thus, the material in smaller size
are crushed to powder and is removed as waste [1]. This causes loss of a huge amount
of natural resources in this case, thus using an automated sorting system to separate
these raw material will make the process faster and reduce the amount of wastage.
In other industries like the packaging industries or material distribution centres use
P. Chenchu Saibabu · R. Anjana · M. Kumari · C. R. Srinivasan (B)

Manipal Academy of Higher Education, Manipal, Karnataka, India
e-mail: cr.srinivasan@manipal.edu
https://doi.org/10.1007/978-981-19-6913-3_23
354 P. Chenchu Saibabu et al.
of automated sorting system can make the process more efficient, cost effective and
less time consuming thus making it more economical, where manual labour is time
consuming and can cause human error. In the postal industry, sorting of the mails
require a lot of labour. The automation of the sorting of the mails reduces the labour
charges and human interfere [2]. Thus, in many industries like coal industry where the
problems like wastage could be identified, and this could be proposed as an effective
solution and also in scrap industries where the weight of the scrap determines the
cost could be monitored by using this simple yet noticeable factor that is determining
the weight while offloading the material from the vehicle of transport via a conveyor
where the materials could be sent batch by batch to the special conveyor one at a time
and then to the recycling plant thus making the process more elementary. Using a PLC
that is rugged can be used for various industrial purposes as it is safe and reliable and
precise. Today, programmable controllers are a rudimentary segment of best in class
robotic arrangements. Smart arrangements are not founded on a contention between
these frameworks, but rather on solidarity. With vital preferences like, robustness
and simple taking care of, the PLC keeps on ensuring its prosperity [3]. Its further
improvement will be reliant on more and more undaunted mix of data and innovation.
For the designing procedure, this implies an incorporated work process that covers
all aspects. Automating the process reduces the human intervention, thus making it
free of human errors and effective in time management. Here in the paper we discuss
how we made the prototype that is responsible for sorting is made possible by the
use of pneumatic machinery and PLC as a controller for automating the process
where we have used a sensor that is load cell to sense the weight of the object [4].
Advance-based sorting system includes the vision-based sorting in order to sort the
objects with respect to the colour and shape [5, 6]. In this paper work is carried out as
three parts, first section deals about the structures which includes control elements,
sensors and pneumatic components. Second section is the design of prototypes of
stated system by using ANSYS tool [7]. Third section states the implementation of
the designed prototype with the programmable logic controllers.
2 Methodology
This work is divided into three parts namely structure, design and implementation.
The structure is again divided into three blocks. The first is the bench, where we
are mounted the sensors of the system. The second is the pneumatic valve desk,
where are mounted the valves, which are handling the actuators. The third part is
the controller, what is located under the bench and controlling the pneumatic valves,
evaluating the sensory data. In the sorting system, the following types of sensors are
used: (1) reflection light scanner (OBT200-18GM70-E5-V1), (2) inductive sensor
(NBN8-18GM50-E2) and capacitive sensor (CJ8-18GM-E2). The reflection light
sensor is acting in specimen reckoning and counting, while inductive and capacitive
sensors are responsible for selections. The training system comprises components
Automation of Weight-Based Sorting System Using Programmable … 355
Fig. 1 NPN proximity

sensor
that perform pneumatic movement and manipulation and control of them; these are
controlled by a programmable logic controller.
2.1 Structure
2.1.1 Sensors
Proximity Sensor
Capacitive-type NPN proximity sensors [8] with a range of 5–34 Vdc can be used to
detect metallic and also non-metallic targets like paper, wood, plastic, glass, wood,
powder, liquid, etc., without physical contact. The capacitive proximity sensor works
on the capacitor principle. The main components of the capacitive proximity sensor
are plate, oscillator, threshold detector and the output circuit (Fig. 1).
Strain Gauge
It is an industrial strain gauge with input voltage range of 5 V–12 V. It can sense
weight from 50 g to 5 kg [9]. When weight is applied on the free end of the instrument.
The strain gauge consists of an insulating flexible backing which supports a metallic
foil pattern. The gauge is attached to the object by a suitable adhesive, such as
cyanoacrylate. As the object is deformed, the foil is deformed, causing its electrical
resistance to change. This resistance change is usually measured using a Wheatstone
bridge. This output is in millivolts. At a time, we can weight only one object weight
using this strain gauge (Fig. 2).
Fig. 2 Strain gauge
Fig. 3 Pneumatic cylinder
Fig. 4 Solenoid valve
2.1.2 Pneumatic Valve Desk
Pneumatic Cylinder
These are basic pneumatic cylinders [10] with stroke length of 5 cm. The air is filled
from the inlet valve at the back. The air is filled inside the cylinder and it apply force
on the piston, and hence, it moves forward. Then, the air exits the cylinder through
the exit valve and the piston retreats. The speed of the stroke can be adjusted by
adjusting the flow of air which is done by the solenoid valve (Fig. 3).
Solenoid Valve
A solenoid valve is an electromechanically operated valve. The valve is controlled

by an electric current through a solenoid. Here, we have solenoid valves with 24 V
coil. This is an electromagnetic coil. When we supply current, it magnetizes the coil.
There is a rod inside the valve that moves up and down due to the magnetization
and thus create in and out flow pressure gauge used. The maximum air pressure this
system could handle is 10 kg pressure lesser than that is adequate for the system
operation (Fig. 4).
Fig. 5 Conveyor belt
2.1.3 Control Elements
Programmable Logic Controller
Programmable logic controller used in this work is Siemens S7 1200 series—

CPU1214C, which is capable of handling analogue inputs/outputs and digital input
and output signals. Step 7 software is used for the implementation of ladder logics
in the programmable logic controllers.
Conveyor Belt
We purchased a 75-cm-long conveyor belt, and it’s the structure. There is roller
placed on both the ends with marks that prevents them from slipping one of the roller
is aligned with a motor, and the other is the free end that can be adjusted according
to the strain. By trial and error method, we have observed that it can hold weight up
to 1 kg. Depending on the application weight range, we can choose the appropriate
conveyor belt (Fig. 5).
2.2 Design
2.2.1 Electrical Circuit Design
Power Supply Circuits
Different components that we are using in the model have different input ranges
such as the proximity sensors require input of 5 Vdc, the solenoid valves requires
input of 12 Vdc and the motor requires input of 6 Vdc. Due to this, we rigged up the
conversion circuit which converts the 230 V ac supply to the regulated dc supplies
of 5 V, 6 V and 12 V. For the supply conversions in this work, 7805, 7806 and 7812
ICs are used (Fig. 6).
Fig. 6 Circuit for 5, 6 and

12 V
Fig. 7 Amplifier circuit
Amplifier Circuit
For measuring weights, we connected an industrial strain gauge. The output of this
strain gauge is in terms of millivolts. The analogue input of the PLC cannot read
such small changes; thus, we used an amplifier circuit to get the output in terms of
volts (Fig. 7).
2.2.2 Mechanical Design
We first designed the model of the prototype deciding the type and the position of
the components to be used to build the prototype and made a schematic diagram for
the same as shown in Fig. 8. To build the structure of the prototype of the model, we
first made a virtual model of the model of the structure using ANSYS tool. We made
a 3D model of the structure to make it easier to make the structure. We designed the
types of structure we require to support the conveyor belt, the proximity sensors, the
strain gauge, the cylinder and the solenoid valves, and we made the model shown in
Fig. 9. We considered a conveyor belt of length as 75 cm and breadth as 5 cm. There
are total five number of capacitive-type proximity sensors, which have a supply of
24 V DC and give a signal of 24 V DC to PLC inputs when an object is detected.
Fig. 8 Prototype model
Fig. 9 Model developed by

using the ANSYS
2.3 Installation
2.3.1 Process Installation
The first one, i.e. proximity sensor 1 detects whether the object is there on the belt.
Once it detects the object, it sends a signal to the PLC which in turn sends the signal
that starts the motor of the conveyor which needs 6 V to drive the motor; thus, signal
turns the motor on and it starts moving. 24 V DC is the output of the PLC and it’s
given to a voltage divider with heat sinks to reduce it to 6 V supply as per requirement
of the motor. The next proximity sensor is placed just before the strain gauge that we
are using. This sensor detects whether the object reached near the strain gauge. This
sensor senses the object and sends a signal to the PLC. The object takes some time
to reach from proximity sensor to the strain gauge which may be in milliseconds.
2.3.2 Weight Determination
When the proximity sensor 2 gives a signal to the PLC, the PLC is programmed
to have a delay that is required by the object to reach form proximity sensor to the
strain gauge and then sends a signal to the strain gauge to turn it on which is at 6 V
and also turns the motor off of the conveyor. The strain gauge measures the weight
of the object. We have decided some ranges of weight for this prototype. Here, we
are sorting the material in two different ranges of weights. If the object is of 100 g
it goes to rack 1, if it is 200 g it goes to rack 2, the last one is the reject rack. In our
prototype, we are sorting the material for 2 different weights. We can increase the
number of weights according to our requirement and desire. The strain gauge gave
different output for different weights as AC input to the PLC unit. PLC determines
the range of weight by the time the delay has got over the motor starts and conveyor
starts working, and the signal from PLC is sent to the respective solenoid valve once
the respective proximity sensor is turned on.
3 Results
3.1 Sorting
For pushing the object into the rack, we are using pneumatic cylinder which are
attached to the pneumatic solenoid valve. The pneumatic cylinders have two holes,
one is inlet and the other is outlet. The electrical signal from the PLC is sent to the
coil of the solenoid valve which is 24 V DC. The valve has a regulated air supply
4 kg which comes from a filter which is used to get a proper supply if air that is free
of moisture and dust. The solenoid valve supply air to the cylinder inlet, thus causing
the stroke to do the pushing action. Then, the air comes out of the cylinder through
the outlet causing the retreating motion of the piston, and the object is pushed into
the rack. Once the process is done, the value is reset for the next object (Figs. 10 and
11; Table 1).
Fig. 10 Sorting of first

weight
Fig. 11 Sorting of the

second weight
Table 1 Result analysis

Case Weight (g) Pneumatic cylinder 1 Pneumatic cylinder 2
1 100 Activated No action
2 200 No action Activated
4 Discussion
In the automatic sorting machine [10], they have sorted the materials based on length
of the materials. In this technique, they have sorted only 5- and 10-cm-length objects.
Expansion of this project based on the object lengths is going to increase complexity
in the design, and the structure becomes very large. After the system design and
commission, it is not possible to change the sorting parameters. In our project, we
have used the sensors to measure the weight and object detection. Control actions are
driven by the programmable logic controllers. In our project, we can easily expand
the sorting parameters by incorporating the changes in programmable logic controller
programming, addition of pneumatic cylinders and sensors.
5 Conclusion
This concept makes the sorting process simpler and faster and thus more efficient.
Using a programmable logic controller (PLC) as the device used in many industries
to monitor and control various activities in production processes is a safe and reli-
able choice. Our objective was to propose the idea through the paper and make a
prototype have been successful as we sorted the material according to the weight of
the object. Also using a capacitive-type sensor gave us the chance to sort any type of
material regardless metallic or non-metallic making the process feasible for any type
of objects. Also the time required by the object to travel from one sensor to another
may vary with changes in the weight of the object, thus changing the delay provided
in the programming which in turn determines the position of the object as required.
This gave us the idea that sorting of the material has been effectively done and has
gave us an alternative effective method.
References
1. Bencsik AL, Lendvay M (2014) Sorting system for practical training of mechatronics. In: 9th
IEEE international symposium on applied computational intelligence and informatics
2. Manjunatha (2015) Postal automation system for mail sorting. Int J Emerg Technol Adv Eng
5(3). ISSN 2250-2459
3. Bencsik AL, Lendvay M (2002) Industry-institute partnership for PLC education and training.
In: IEEE international conference on information technology based higher education and
training (ITHET 2002), Budapest, Hungary, 4–6 July 2002
4. Aruna YV, Beena S (2015) Automatic convey or system with in-process sorting mechanism
using PLC and HMI system. Int J Eng Res Appl 5(11):37–42, ISSN: 2248-9622
5. Sheth S, Kher R, Shah R, Dudhat P, Jani P, Automatic sorting system using machine vision.
In: Multi-disciplinary international symposium on control, automation & robotics, At DDIT,
Nadiad, vol 1. https://doi.org/10.13140/2.1.1432.1448
6. Wahab DA, Hussain A, Scavino E, Mustafa MM, Basri H (2006) Development of a prototype
automated sorting system for plastic recycling. Am J Appl Sci 3(7):1924–1928
7. Sen S, Kumar A, Srinivasan CR, Saibabu C (2014) Efficient and low cost flow measurement
using bend sensor flowmeter. Int J Electron Electr Eng 7(8):879–885. ISSN 0974-2174
8. Oladapoa B, Baloguna VA, Adeoyea AOM, Ijagbemib CO, Oluwolea AS, Daniyana IA, Esoso
A, Asanta A, Simeona P (2016) Model design and simulation of automatic sorting machine
using proximity sensor. Eng Sci Technol Int J 19(3):1452–1456
9. Sadani K, Prabhakar DA, Nag P (2018) A cardio pulmonary resuscitation device for stretchers.
Int J Eng Technol 7(2.21):62–65
10. Prasad A, Gowtham M, Mohanraman S, Suresh M (2020) Automatic sorting machine. Int Res
J Multidisc Technovation 7–12
Design, Development and Verification
of a Fault Injection Capable Synchronous
Generator
Sona Meiyappan, P. Chaithanyasai, S. Swetha, M. Vishnu Deepika,

and P. V. Sunil Nag
Abstract Virtual models capable of fault injection are crucial for Fault Diagnosis
of Electrical machines like the Synchronous Generator (SG) and are researched
widely for this purpose. The SG is used in many significant industries for power
generation applications. Inter-turn stator faults account for the majority of SG faults.
Multiple methods are being developed to detect these faults in the early stages to
avoid industry shutdowns, reduce machine downtimes, and also as a cost-effective
maintenance technique. A virtual fault injection capable model of SG is designed
and developed to simulate and analyze the effects of inter-turn stator faults. The
efficacy of the proposed model was checked by using a model of 10 kVA SG. The
dimensions of this SG were calculated theoretically, and the resulting values were
fed to ANSYS Maxwell software, where the 2D design of the SG was developed, and
faults were introduced to replicate a hardware model closely. Faults were introduced
in all three phases under two different load conditions. The system was verified to be
fault injection capable with the help of fault signatures found in the stator currents and
field current. The second and third harmonic components of the above-mentioned
quantities are characteristic of turn-to-turn short circuit (TTSC) faults, and hence
were used to validate the developed model.
Keywords Fault injection · Synchronous generator · Fault diagnosis
1 Introduction
Over the years, the consumption of electric power has grown exponentially. As new
devices are added to the grid every year, a need for dependable electrical machines
has risen [1]. One such important machine is the Synchronous Generator (SG). A SG
S. Meiyappan · P. Chaithanyasai · S. Swetha · M. Vishnu Deepika · P. V. Sunil Nag (B)

Department of Electronics and Communication Engineering, Amrita School of Engineering,
Amrita Vishwa Vidyapeetham, Coimbatore 641112, India
e-mail: pv_sunil@cb.amrita.edu
https://doi.org/10.1007/978-981-19-6913-3_24
366 S. Meiyappan et al.
is an extensively used generator for power generation [2]. Any faults in the SG can
lead to a complete industry shutdown and compromise other systems and crew safety.
Winding faults are the most common and significant fault in synchronous generators
[3] and are mainly caused due to insulation failures [4]. The stator winding faults
occur more frequently compared to rotor winding faults. The inter-turn faults occur
in the initial stage of the stator winding failures [5] and can lead to complete damage
to the machine. Thus, fault detection in the incipient stage has become exceedingly
necessary.
A crucial step in detecting a fault is to develop a test rig where the behavior of
a faulty machine can be studied to develop methods capable of detecting the faults
in early stages. Fault injection models are required for this purpose. Fault injection
models are of two types: (1) Custom made Hardware setup [6], (2) Virtual machine.
A SG has additional components compared with other electrical machines and
hence is considered a complex machine. Researchers face difficulties in manufac-
turing and monitoring such a complex machine. A SG has additional components
compared with other electrical machines and hence is considered a complex machine.
A fault injection capable model of a SG is designed and implemented in ANSYS
Maxwell to simulate both healthy and faulty generator data. Our work in this paper
attempts to validate our virtual machine against a hardware setup. If verified, this
model gives researchers the freedom to introduce different types of faults, those that
may not be practically realizable with a hardware setup. Further, the design and
implementation of this virtual setup discussed in this work is cost-effective and any
aspiring researcher can easily design and develop a SG for research and exploration
in the field of fault diagnosis.
Analysis of the fault signature in the third harmonic component of stator current
for a SG with R Load has been discussed in [7]. The main contribution of this paper
is that it further builds on the mentioned work, by verifying the same for a SG with
RL Load. Additionally, the second and third harmonic components of both stator and
field currents were plotted and shown to be characteristics of a short circuit fault.
The layout of the paper is as follows. The flowchart is displayed in Sect. 2.
In Sect. 3, the design of SG is presented. Section 4 provides the Maxwell model
obtained from the design data. Section 5 discusses about the injection of fault in
ANSYS Maxwell. The plots of the induced currents and voltages obtained from the
Maxwell model are given in Sect. 6. The results of the frequency analysis of various
currents are discussed in Sect. 7.
2 Flowchart of the Working Model
See Fig. 1.
Design, Development and Verification of a Fault Injection Capable … 367
Fig. 1 Flowchart
3 Design of a Synchronous Generator
A Synchronous Generator (SG) was designed with the basic specifications of 10 kVA
power, a Y connected load with a 36 slot stator, a 4 pole cylindrical rotor, a line to
line voltage of 628 V rms , with a maximum load current of 4.5 A and a power factor of
0.8. The synchronous speed of the generator is 25 r.p.s. and has 220 V DC excitation
and 50 Hz frequency.
The design parameters required for the RMxprt are calculated referring chapter-
12 in [8]. A 10 kVA Synchronous Generator was designed theoretically and the
dimensions were obtained with suitable assumptions for average mechanical loading
(Bav = 0.4 Wb/m2 ) and average electrical loading (ac = 13,000 A/m).
4 Design of SG Using RMxprt and ANSYS Maxwell
RMxprt is a template-based electrical machine design tool developed by ANSYS that

is used to evaluate a design. After designing the analytical model of the generator,
the performance of the design can be verified using ANSOFT RMxprt. This RMxprt
Fig. 2 RMxprt model of the

SG
model can then be converted in 2D or 3D geometric creation in ANSYS Maxwell to

perform finite element calculations [9].
Combined together RMxprt and Maxwell fully meet the electric machine design
and analysis needs [10]. In the presented work, the analytical model of the generator
is simulated in RMxprt (Fig. 2).
After the design requirements are met and verified using RMxprt, this model is
then used for 2D transient analysis using ANSYS Maxwell. A load of 1.5 A is used
in all cases presented in this paper (Fig. 3).
The SG’s symmetry is a significant component in obtaining a desirable output
voltage [11]. Hence, the stator slot coils are split equally in the ANSYS Maxwell 2D
model (Fig. 4).
Fig. 3 ANSYS Maxwell

model of the SG
Fig. 4 Circuit before splitting of coils
5 Fault Injection in ANSYS Maxwell
Fault is introduced by splitting the stator coils of all three phases. Stator winding of
each phase consists of 30 turns and a split in coil is introduced between 7 and 8 turns
of the same phase [7] (Fig. 5).
Split coil
Unsplit coil
Fig. 5 Enlarged view of a spilt and unsplit coil

Fig. 6 Circuit after splitting of coils
An equivalent circuit of the Maxwell 2D model with the spilt in coils is developed
in the circuit editor of ANSYS Maxwell software (Fig. 6).
The fault is introduced by triggering a short circuit by adding a switching device to
the external circuit, which is designed to close after a certain time [12]. Simulations
are run for 1.5 s at a sampling frequency of 1000 Hz. It was observed that 0.6 s
was necessary for the SG to attain the steady state and hence fault was introduced at
0.8 s, resulting in 0.2 s of healthy data and 0.7 s of faulty data. This external circuit
is then manipulated to introduced fault into three phases each under two different
load conditions namely, Resistive Load (R Load), and Resistive Inductive Load (RL
Load), resulting in a total of 6 scenarios. AC motors like induction motors are the most
common loads associated with SGs and these loads can be modelled as a combination
of resistances and inductances. Hence, a R Load and RL Load cases were chosen for
this study. A total of 6 simulation cases are run on a High-performance computing
(HPC) machine. Induced and field current plots are obtained from each case for
further analysis (Figs. 7, 8, 9 and 10).
6 ANSYS Maxwell Results
Simulation values from Induced Current plot (stator currents) and Field Current plot
after simulation are exported to MATLAB for FFT Analysis.
Switch
R Load
Trigger
Fig. 7 External circuit of R load case: fault in phase A
Switch
RL Load
Trigger
Fig. 8 External circuit of RL load case: fault in phase A
Fig. 9 Induced and field current plots (R load with fault in phase A)
Fig. 10 Induced and field current plots (RL load with fault in phase A)
7 Fast Fourier Transform
7.1 Results from FFT for Analysis of Stator Current
The frequency analysis of stator current shows that the third harmonic component
proves to be a suitable indicator for the detection of Turn-to-Turn short circuit (TTSC)
fault. Figure 11 displays the stator currents of all the three phases for a Synchronous
Generator under healthy operation for R Load condition. Figure 13 represents the
plot of the stator currents for a machine under TTSC fault for R Load Condition.
Comparing the two figures, it can be observed that the third harmonic component
occurring at 150 Hz appears only when a fault is introduced. It can be inferred from the
above analysis that the third harmonic component of stator current is a characteristic
of TTSC fault, as stated in [13]. Figures 12 and 14 show the same frequency spectrum
under RL Load condition for healthy and faulty conditions, respectively, and the same
results obtained above were observed.
7.2 Results from FFT for Analysis of Field Current
Fast Fourier Transform is performed on the field current to study the harmonic
components. Figures 15 and 17 show the frequency analysis plots for a Synchronous
Generator under healthy and faulty conditions, respectively, for R Load condition.
Analysis of the plots shows that the second harmonic component occurring at 100 Hz
arises only when a fault is injected in the machine. According to [13], this is a verified
signature for the presence of fault. Hence, it is proved that the fault has been injected
Fig. 11 Frequency spectrum of stator current of a healthy machine under R load condition for
phases A, B and C
Fig. 12 Frequency spectrum of stator current of a healthy machine under RL load condition for
phases A, B and C
in the developed machine. As done previously, the machine was again simulated for
RL Load condition to further verify that the fault signatures function for all linear
load cases. Figures 16 and 18 show the frequency spectrum of field current of the
machine under RL Load condition for healthy and faulty cases, respectively.
Fig. 13 Frequency spectrum of stator current under TTSC fault for R Load condition in Phases A,
B and C
Fig. 14 Frequency spectrum of stator current under TTSC fault for RL load condition in phases
A, B and C
Fig. 15 Frequency spectrum of field current of a healthy machine under R load condition for phases
A, B and C
Fig. 16 Frequency spectrum of field current of a healthy machine under RL load condition for
phases A, B and C
Fig. 17 Frequency spectrum of field current under TTSC fault under R load condition in phases
A, B and C
Fig. 18 Frequency spectrum of field current under TTSC fault under RL load condition in phases
A, B and C
Table 1 Comparison of results with prior art

Results obtained in this paper Results from existing work
Appearance of third harmonic component of stator Variation of 180 Hz (third harmonic)
current under TTSC fault refer Figs. 13 and 14 component of terminal voltage under TTSC
fault refer Figs. 4 and 5 of [14]
Appearance of 2nd harmonic component of field Increase in magnitude of 2nd harmonic
current when a 7% stator winding fault is component after introduction of 15%, as
introduced refer Figs. 17 and 18 compared to the healthy machine. Refer
Fig. 23a, b of [13]
7.3 Comparison of Results Obtained
Table 1 compares the frequency spectrum obtained in this work with existing liter-
ature. From the comparison it can be concluded that the fault injection capable
synchronous generator can simulate TTSC fault.
8 Conclusion
In this paper, the development of a Fault Injection capable model of a Synchronous

Generator is presented. A Synchronous Generator was designed for a power of
10 kVA and was simulated using the ANSYS Maxwell software. Simulation was
done for various load conditions, with the introduction of fault into the machine. To
verify the presence of fault after its introduction in the machine, frequency analysis
was done on the stator current and on the field currents obtained from the simulation.
The appearance of second and third harmonic components of the stator current and
field current confirm the presence of fault in the developed machine.
Conflicts of Interest The authors declare that they have no conflicts of interest to report regarding
the present study.
References
1. DKG, Menon SR, Sreevathsava N, PM, Nag PVS, Kumar CS (2020) Stator inter-turn fault
diagnosis and fault location in synchronous generator using dual extended Kalman filter and
linear regression analysis. In: 5th international conference on communication and electronics
systems (ICCES) 2020, Coimbatore, India, pp 52–57
2. Bruzzese C (2014) Diagnosis of eccentric rotor in synchronous machines by analysis of split-
phase currents–part II: experimental analysis. IEEE Trans Ind Electron 61(8):4206–4216
3. IEEE Recommended Practice for the Design of Reliable Industrial and Commercial Power
Systems—Redline. In: IEEE Std 493–2007 (Revision of IEEE Std 493–1997)—Redline, pp
1–426
4. Jee-Hoon J, Jong-Jae L, Bong-Hwan K (2006) Online diagnosis of induction motors using

MCSA. IEEE Trans Ind Electron 53(6):1842–1852
5. Riera-Guasp M, Antonino-Daviu JA, Capolino GA (2015) Advances in electrical machine,
power electronic, and drive condition monitoring and fault detection: State of the art. In: IEEE
Trans Ind Electron 62(3):1746–1759
6. Gopinath R, Kumar C, Ramachandran K (2016) Scalable fault models for diagnosis of
synchronous generators. Int J Intell Syst Technol Appl 15(1):35
7. Kamala KS, Induvadhani VV, Lakshmi VI, Mithra P, Sunil Nag PV, Kumar CS (2020) Electrical
signature analysis (ESA) of a fault injection capable synchronous generator for inter-turn
stator faults. In: 2020 5th international conference on communication and electronics systems
(ICCES) Coimbatore, India, pp 171–175
8. Sawhney AK (1984) Electrical machine design, 1st edn. Dhanpat Rai, India, pp 702–833
9. MK, Warrier GS, Pathivil P, Kanagalakshmi S, Archana R (2019) Design and performance
analysis of brushless DC motor using ANSYS Maxwell. In: 2019 2nd international conference
on intelligent computing, instrumentation and control technologies (ICICICT), Kannur, Kerala,
India, pp 1049–1053
10. Allirani S, Vidhya H, Aishwarya T, Kiruthika T, Kowsalya V (2018) Design and performance
analysis of switched reluctance motor using ANSYS Maxwell. In: 2018 2nd international
conference on trends in electronics and informatics (ICOEI), Tirunelveli, pp 1427–1432
11. Dallas SE, Safacas AN, Kappatou JC (2011) Interturn stator faults analysis of a 200-MVA
hydrogenarator during transient operation using FEM. IEEE Trans Energy Convers 26(4):1151–
1160
12. Prasob K, Kumar NP, Isha TB (2017) Inter-turn short circuit fault analysis of PWM inverter fed
three-phase induction motor using finite element method. In: 2017 international conference on
circuit, power and computing technologies (ICCPCT), pp 1–6
13. Nadarajan S, Panda SK, Bhangu B, Gupta AK (2015) Hybrid model for wound-rotor
synchronous generator to detect and diagnose turn-to-turn short-circuit fault in stator windings.
IEEE Trans Ind Electron 62(3):1888–1900
14. Sottile J, Trutt F, Leedy A (2006) Condition monitoring of brushless three-phase synchronous
generators with stator winding or rotor circuit deterioration. IEEE Trans Ind Appl 42(5):1209–
1215
UKF/H-Infinity Filter for Low-Cost
Localization in Self-driving Cars
K. Bipin and P. V. Sunil Nag
Abstract A self-driving car is one of the engineering marvels in recent time due
to autonomous capabilities. An autonomous car can sense its surroundings through
various integrated sensors and navigate the vehicle based on sensor data with Model
Predictive Control System. Vehicle Navigation in self-driving car helps to deter-
mine the vehicle position by localization through high precision Global Naviga-
tion Satellite System and Inertial Navigation System (GNSS/INS) technologies with
various sensors like Light Detection and Ranging (LIDAR), Radio Detection and
Ranging (RADAR) and Global Positioning System (GPS) Sensor. Vehicle Navi-
gation is completely depending up on the localization of the vehicle. The most
popular method like particle filter and Kalman filter-based approach lacks stability
with random particle distribution, and some of the GNSS-based methods are very
expensive in terms of implementation. So here introducing an unscented Kalman
filter and H-Infinity filter (UKF/H-Infinity)-based novel approach for the low-cost
localization technique in self-driving cars. This robust control model helps to handle
the position error from the GPS sensor with H-Infinity filter and UKF by the help
of Constant Turn Rate Acceleration (CTRA) motion model. This work includes
the detailed discussion of pose estimation of vehicle with better yaw stability and
minimum computational overhead in estimation.
Keywords CTRA · GPS · H∞ · LIDAR · PF · UKF
1 Introduction
Vehicle positioning or localization is the primary information needed for any

Autonomous Transport System. This information can be collected by using various
integrated sensors and methods. Self-driving cars use sensors and maps for vehicle
positioning and understanding its surrounding infrastructure. The existing methods
K. Bipin · P. V. Sunil Nag (B)

Department of Electronics and Communication Engineering, Amrita School of Engineering,
Amrita Vishwa Vidyapeetham, Coimbatore 641112, India
e-mail: pv.sunil@cb.amrita.edu
https://doi.org/10.1007/978-981-19-6913-3_25
380 K. Bipin and P. V. Sunil Nag
available for determining the vehicle position in self-driving cars have numerous
problems and challenges. GNSS suffers from accuracy problem [1] in Urban Envi-
ronment, and Simultaneous Localization and Mapping (SLAM) methods have
error accumulation problem [2]. Another state-of-the-art localization technique 3D-
LIDAR also has same problem [3]; this technique observes the surrounding infras-
tructure and compares it with prior known 3D point cloud for estimating the position
of the vehicle in the map [4]. Here, we need massive storage space on the vehicle, or
simultaneous cloud access is necessary. Even though GPS has low accuracy problem
[5], we can improve the performance by using various algorithms.
2 Background
Understanding the localization process [6] along with various models is a time-
consuming process, because each model depends on different dynamics of the
vehicle. GPS-based localization technique is commonly used, but we cannot
completely rely on GPS as the only solution due to noise and its poor performance in
indoor and urban environments. Accuracy in any localization technique is crucial for
self-driving cars, because lack of accuracy leads to road accidents. We went through
the study of various approaches used in vehicle localization. Some of the methods
referred for understanding the problem statement of the work are as follows. Odom-
etry is a technique in localization where wheel displacement calculation is used for the
estimation of vehicle position, but this technique is inaccurate [7] since some vehicle
parameters are not taken into account for prediction. Another popular method called
SLAM uses the MAP [8]. The reference paper taken here for the implementation
has Kalman and particle filter-based approach [9] that compares the observations of
range sensors with the surrounding infrastructure map and creates particles around
the ego vehicle by using an iterative approach to make similar observation as that of
map.
3 UKF/H-Infinity Filter
We are proposing a new approach called UKF/H-Infinity robust control model with
CTRA motion model. The H-Infinity filter is commonly used in control theory due to
its robust nature with better stability; it can also be used in GPS/INS [10] and here we
are using UKF [11], as optimal control. The H-Infinity filter estimates better position
from noisy GPS sensor data by considering the acceleration and velocity noise from
the Inertial Measurement Unit (IMU) fusion filter. The UKF motion prediction can
be achieved in process space by using the CTRA motion model and the state position
input from the H-Infinity algorithm.
When these two algorithms are combined, we can call this as mixed model
or UKF/H-Infinity robust control model. H-Infinity algorithm helps to reduce the
UKF/H-Infinity Filter for Low-Cost Localization in Self-driving Cars 381
estimation error and gives better performance with constrained environment, but
sometimes outliers affect the performance of the algorithm. But choosing maximum
bounded value that can tolerate the outlier will help to avoid this problem. H-Infinity
filter’s robustness nature of handling [12] unknown noise resources help to reduce
noise from the localization system.
Figure 1 is the architecture of UKF/H-Infinity robust model controller, where we
can see the IMU fusion filter with GPS, UKF/H-Infinity filter model and LIDAR
sensor module in it. From the IMU/GPS fusion filter, we get the ego vehicle’s
lateral and longitudinal parameters like linear velocity, angular velocity, orientation
of the vehicle and noisy GPS position of the vehicle. While implementing model in
MATLAB, we can add Gaussian and non-Gaussian noise in the IMU fusion filter
for simulating the real-world scenario. The input of H-Infinity controller is the noisy
GPS position from the fusion filter, and it consists of the three states (x, y, heading
angle), angular velocity, linear velocity and acceleration noise. After the H-Infinity
process estimation, a new state vector with size 3 is given as the input to UKF filter.
The UKF filter again does the estimation by using the measurement space value from
the LIDAR, and the final estimation result will be the same size as input with best
accuracy and less computational overhead. The detailed description of H-Infinity
filter is shown below.
Fig. 1 UKF/H-Infinity model architecture

3.1 H-Infinity Filter
The H-Infinity filter [13] controller shown in Fig. 2 and general equation can be
written (1) as follows, where A, B are known matrices, and we are estimating the
state considering the unknown noise.
X k+1 = A ∗ X k + B ∗ U + K ∗ Mv (1)
If we consider the velocity, acceleration and angular velocity in model, it can

be written in the following form. Here, state vector “X” has the position of vehicle
from GPS sensor with the external noise. “U” is the unknown vehicle acceleration
(without the noise). M v is the measured velocity affected by the measurement “z”
noise from the plant. K has two elements, and it is considered as H-Infinity gain and
is constantly updated through the process. The system model can be formulated (2)
with some uncertainty.
x(k+1) = (A + Aunc ) ∗ xk + (B + Bunc ) ∗ u k + wk (2)
xnoise,max = xk + xoffset (3)
ynoise,max = yk + yoffset (4)

xk+1 = xk + xnoise,max . cos(θh ) (5)

yk+1 = yk + ynoise,max . sin(θh ) (6)
Fig. 2 H-Infinity control block

vk

x θk t
[sin(θk + θk (t)) − sin(θk )]
X h f,k+1 = = vk (7)
y θk t
[cos(θk ) − cos(θk + θk (t))]
The motion model (7) predicts the position of the vehicle by considering the
reduced positional error in the state space x and y with H-Infinity filter. This is
considered as one of the simplest vehicle models generally used for the reduction of
computational expense in vehicle motion model. The advantage of using this model
is that it is only considering few parameters (8) for lateral and longitudinal motion,
and hence, it does not affect the yaw rate stability of the vehicle. The motion model
will be added to the state space matrix of the general equations with acceleration and
velocity noises. The dynamics of Bicycle Vehicle Motion Model [14] is shown in
Fig. 3.
The outliers are the boundaries of the noise; x noise,max is the maximum noise in
x-direction (3) and ynoise,max is the maximum noise in y-direction (4). The x offset and
yoffset are the outlier offset parameters and Eqs. (5) and (6) are constraints with this
offset.

Vpos = x y θ (8)
x Ego vehicle position in x-direction

y Ego vehicle position in y-direction
Fig. 3 Bicycle vehicle motion model

θ Heading angle
v Velocity of ego vehicle.
“A” is the transition matrix with dimension “m × m” having the time space between
the previous and next position, “B” is the input matrix with dimension “m × n” having
the position estimation through the motion model, “C” is the measurement matrix
with dimension “n × m”, which keeps updated by the previous measurement state
and X hf is the initial state vector with dimension m × n; where m = 2 and n = 1.
The initial measurement state “Y ” is initialized below with state and measurement
update.
Y = C ∗ X hf (9)
The measurement can be initialized by measurement matrix C and initial state

vector X hf by using Eq. (9). The H-Infinity gain controls the algorithm by using the
below equations, we can apply the steady state gain if the system performance is poor
or else, and we can use time-invariant gain to get better control over the computation.
This can be achieved by Eqs. (10) and (11).
L = (I − g ∗ Q ∗ P + C T ∗ V −1 ∗ C ∗ P) (10)
K = A.P. L −1 . C T . V −1 (11)
Here, “I” is identity matrix, M v is the measurement variable, M Noise measurement

noise, PNoise process noise, Q weight matrix, P covariance matrix, K is H-Infinity
gain and mNoise nominal measurement values. Measurement state (14) and velocity
update (15) initialized with previous state. Equation (13) forces the matrix P (12) to
be symmetric.

10
P= (12)
01

P = P + P T /2 (13)
Mx = Y − C ∗ X k (14)
Mv = Y − C ∗ Vk (15)
Process noise and measurement noise can be calculated by Eqs. (16), (20), where
we are considering the acceleration noise and “B” matrix.

R1
PNoise = 2 ∗ aNoise ∗ B ∗ (16)
R2
X k+1 = A ∗ X k + B ∗ U + K ∗ Mx (17)
Vk+1 = A ∗ Vk + B ∗ U + K ∗ Mv (18)
X k+1 = A ∗ X k + B ∗ U + K ∗ PNoise (19)
MNoise = m noise ∗ (rand − 0.5) (20)
Here in Eq. (17), we can see two known and two unknown matrices, and we can
consider the known matrices as the acceleration and velocity of the ego vehicle and
unknown matrices as the noise of both. The noise is always coming with acceleration
and velocity of ego vehicle from the fusion filter module due to the external or
internal disturbance. We have taken acceleration and velocity as the system model
parameter and angular velocity as the external parameter to update the state (19) and
velocity (18) with measurement state. This can be applied in adaptive filter modelling,
and it helps the time invariant system to adapt the relative noise with constrained
boundaries. This method outperforms the particle filter due to its robust nature.
3.2 UKF
The UKF [15] is suitable for nonlinear space motion model. UKF takes the points
from the whole distribution called sigma points and Gaussian distribution given into
a nonlinear function [16] called unscented transformation, and this will generate
the corresponding sigma points in the prediction and measurement model with the
state space vector [17]. Then, we calculate the mean of state vector and covariance
matrix from these corresponding sigma points to find the Gaussian distribution of
the predicted or measurement state space. By taking the correlation of both state and
covariance matrix, we get the Kalman gain and using this and LIDAR measurement
space, the state and covariance matrix will be updated. The overall workflow of the
UKF algorithm is shown in Fig. 4 for better understanding.
UKF Prediction Generating sigma points from the state vector (21) is the first step
involved in the UKF algorithm, and it is considered as the symmetrical region around
the mean value of state. We augment this state matrix with acceleration and velocity
noise, then the resultant state vector is given to the UKF prediction. In the UKF
prediction, CTRA motion model is used for the prediction of new state. The CTRA
model is a nonlinear motion model used for vehicle state estimation, and this model
has six state vectors, and it can be converted to eight by adding additional augmented
state to prediction model.
T
X CTRA = x, y, v, a, h, w (21)
Fig. 4 UKF workflow

Here, x is the ego vehicle position in x-direction, y is the ego vehicle position in
y-direction, v and a are the velocity and acceleration of the vehicle and w are the
heading angle and heading turn rate of the vehicle.
⎡ ⎤ ⎡ ⎤
xhfukf,k (vω + aωT ) · sin(h + ωT )
⎢ yhfukf,k ⎥ ⎢ (−vω − aωT ) cos(h + ωT ) ⎥
⎢ ⎥ ⎢ ⎥
⎢ v + aT ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ a ⎥ 1 ⎢ 0 ⎥
X hfukf,k+1,aug = ⎢ ⎥+ 2⎢ ⎥
⎢ h + ωT ⎥ ω ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ω ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 ⎦ ⎣ 0 ⎦
0 0
⎡ ⎤
a ∗ cos(h + ω.T ) − vω sin(h) − a ∗ cos(h)
⎢ a ∗ sin(h + ωT ) + vω cos(h) − a ∗ sin(h) ⎥
⎢ ⎥
⎢ ⎥
⎢ 0 ⎥
1 ⎢⎢ ⎥
0 ⎥
⎢ ⎥ (22)
ω ⎢
2 0 ⎥
⎢ ⎥
⎢ 0 ⎥
⎢ ⎥
⎣ 0 ⎦
0
The motion model can be written as (22), where acceleration parameter is added in
the model to achieve the best performance in cornering. Here, we have four-velocity
models for the ego vehicle. This makes the system to work more efficient in handling
noise in cornering. Usually, CTRA model gives better results than CTRV in urban
environment [18, 19] due to handling the acceleration parameter in the motion model
equations.
UKF Updates Here, both process space and measurement space state mean and
covariance will be updated by taking the correlation of both spaces. By taking the
correlation of these two spaces we compute the Kalman gain. The state and covariance
will be updated using LIDAR measurement values and the Kalman gain. Finally, we
are getting the processed state estimation from UKF/H-Infinity filter with reduced
positional error and better stability.
4 Implementation
The implementation of UKF/H-Infinity filter is done in MATLAB by using various

MATLAB toolboxes. Some of the MATLAB toolboxes are described below with
respective area of development in this approach.
4.1 IMU–GPS Fusion Filter
The IMU fusion filter [20] is created by using MATLAB sensor fusion and tracking
toolbox for generating the true and noisy GPS data and other dynamics parame-
ters. For understanding these parameters with noise, we can add Gaussian and non-
Gaussian noise into the virtual sensor. Tables 1 and 2 show IMU and GPS noise
parameter used in the work.
5 Simulation Result
MATLAB automated driving toolbox and driving scenario toolbox are used for
creating road models for testing different velocity profiles with ego vehicle. I-Road
model has a vehicle profile “Fig. 5a” with virtual LIDAR/RADAR sensor embedded
in it and have road length of 60 m. Here in S-Road model “Fig. 6b”, the vehicle
profile is same as the I-Road model except the road length of 4 km.
The accuracy of algorithm is given in Table 3 and can observe that this approach
[20] works well in Gaussian noise. In real-world applications, we can’t expect
Gaussian noise, so here we had concentrated on non-Gaussian noise. The UKF/H-
Infinity approach gives better result in non-Gaussian noise with CTRA motion model.
Apart from UKF/PF uses the random distribution of particle in state prediction,
hence the accuracy of estimation changes on each test cycle, while the ego vehicle
taking cornering. UKF/H-Infinity filter circumvent this problem by considering the
H-Infinity filter in this approach.
Table 1 IMU noise parameter table

S. No. Noise parameter Accelerometer Gyroscope Magnetometer
1 Measurement (m/s2 ) 19.62 4.363 1200
2 Noise density 0.0012361 0.0000872 [ 0.6 0.6 0.9 ]
3 Resolution 0.0023965 0.00013323 0.1
Table 2 GPS noise parameter

S. No. GPS sensor parameter Non-Gaussian noise
1 Horizontal position accuracy Normrnd (9.65, 12.2)
2 Vertical position accuracy Normrnd (8.34, 12.33)
3 Decay factor 0.80
5 Reference location [ 40 − 50 30 ]
6 Random stream mt19937ar with 100 seed
(a) I-Road Model (b) S-Road Model
Fig. 5 Road models
(a) H-Infinity Filter Estimation (b) UKF Estimation
Fig. 6 I-road model vehicle profile
Table 3 UKF/H-infinity filter position accuracy table

Velocity Noise UKF/PF UKF/PF UKF/H-infinity UKF/H-infinity
km/h m GA (m) Non-GA (m) GA (m) Non-GA (m)
60 29 1.4510 1.6552 0.9822 1.2182
80 29 1.6519 1.4402 0.9713 1.2477
100 29 1.5012 1.6167 0.9408 1.2938
120 29 1.7721 1.4541 0.9900 1.2626
5.1 I-Road: Ego Vehicle at 60 km/h
See Fig. 7.
5.2 S-Road: Ego Vehicle at 60 km/h
See Figs. 8 and 9.

Fig. 7 I-road model—UKF/H-infinity position estimation of ego vehicle
(a) H-Infinity Filter Estimation (b) UKF Estimation
Fig. 8 S-road model vehicle profile
6 Conclusion
The UKF/H-Infinity robust model is implemented in MATLAB with four different

velocity profiles and two road models. From the simulation result, we can conclude
that the accuracy of the algorithm increases marginally when compared with UKF
without H-Infinity filter and from the figures we can view the stability in the position
estimation in cornering. We can further improve the accuracy by doing rigorous
tuning of the filter. This approach uses GPS and 2D/360/LIDAR for localization but
the implementation of UKF/H-Infinity filter with 3D-LIDAR measurement can be a
future enhancement of this approach.
Fig. 9 S-road model—UKF/H-infinity position estimation of ego vehicle
References
1. Lohan ES, Borre K (2016) Accuracy limits in multi-GNSS. IEEE Trans Aerosp Electron Syst
52(5):2477–2494
2. Jiang G, Yin L, Jin S, Tian C, Ma X, Ou Y (2019) A simultaneous localization and map-ping
(SLAM) framework for 2.5D map building based on low-cost LiDAR and vision fusion. Appl
Sci 9(10):2105
3. Chen CP, Xu L, Danping C, Qi Z, Zhu Y, Li Y, Tao (2019) Trajectory optimization of
LiDAR SLAM based on local pose graph. China satellite navigation conference (CSNC) 2019
proceedings, p 360370
4. Yin H, Berger C (2017) Mastering data complexity for autonomous driving with adaptive point
clouds for urban environments. In: 2017 IEEE intelligent vehicles symposium (IV) (2017), pp
1364–1371
5. Pawlowski E (2015) Experimental study of a positioning accuracy with GPS receiver. In:12th
conference on selected problems of electrical engineering and electronics WZEE’2015, Kielce,
Poland, vol 12
6. Woo A, Fidan B, Melek WW, Zekavat S, Buehrer R (2019) Localization for autonomous driving.
In: Handbook of position location: theory, practice, and advances, IEEE, pp 1051–1087
7. Goh V, Fischbeck P, Gerard D (2007) Identifying and correcting errors with odometer read-
ings from inspection and maintenance data: rollover problem for estimation of emissions and
technical change. Transportation research record
8. Frintrop S, Jensfelt P, Christensen HI (2006) Attentional landmark selection for visual SLAM.
In: IEEE/RSJ international conference on intelligent robots and systems, pp 2582–2587. https://
doi.org/10.1109/IROS.2006.281711
9. Lin M, Yoon J, Kim B (2020) Self-driving car location estimation based on a particle-
aided unscented Kalman filter. Sensors (Basel) 20(9):2544. https://doi.org/10.3390/s20092544.
PMID: 32365721; PMCID: PMC7249166
10. Tsogas M, Polychronopoulos A, Amditis A (2005) Unscented Kalman filter design for
curvilinear motion models suitable for automotive safety applications. In: 7th international
conference on information fusion, p 8. https://doi.org/10.1109/ICIF.2005.1592006
11. Jiang C, Zhang S-B, Zhang Q-Z (2016) A new adaptive H-infinity filtering algorithm for the
GPS/INS integrated navigation. Sensors 16(12):2127
12. Nag S, Silla G, Gummadi V, Harishankar B, Ray V, Kumar C (2016) Model based fault
diagnosis of low earth orbiting (LEO) satellite using spherical unscented Kalman filter, IFAC-
PapersOnLine 49, pp 635–638. https://doi.org/10.1016/j.ifacol.2016.03.127
13. Warrier E, Nag S, Kumar C (2019) A novel H-infinity filter based indicator for health monitoring
of components in a smart grid. Energy transfer and dissipation in plasma turbulence, pp 221–
230. https://doi.org/10.1007/978-981-13-59538
14. Polack P, Altch´e F, Novel B, de La Fortelle A (2017) The kinematic bicycle model: a consistent
model for planning feasible trajectories for autonomous vehicles? In: 2017 IEEE intelligent
vehicles symposium (IV), pp 812–818. https://doi.org/10.1109/IVS.2017.7995816
15. Sudheesh P, Jayakumar M (2018) Nonlinear tracking using unscented Kalman filter. In:
Advances in intelligent systems and computing, vol 678, pp 38–46, Springer Internal, Advances
in Signal Processing and Intelligent Recognition Systems
16. Sudheesh P, Jayakumar M (2018) Tracking of nonlinear variations of the parameters of high
mobility systems. Int J Pure Appl Math 118(7):221–226
17. Schubert R, Adam C, Obst M, Mattern N, Leonhardt V, Wanielik G (2011) Empirical evaluation
of vehicular models for ego motion estimation. IEEE Intell Veh Symp (IV) 534–539. https://
doi.org/10.1109/IVS.2011.5940526
18. Schubert R, Richter E, Wanielik G (2008) Comparison and evaluation of advanced motion
models for vehicle tracking. In: 11th international conference on information fusion, pp 1–6.
https://doi.org/10.1109/ICIF.2008.4632283
19. Shapovalov I, Maximychev E, Gafurov S, Ostankovich V, Fedorenko R (2020) Robust local-
ization of a self-driving vehicle in a lane. In: 4th scientific school on dynamics of complex
networks and their application in intellectual robotics (DCNAIR), pp 210–213
20. Caron F, Duflos E, Pomorski D, Vanheeghe P (2006) GPS/IMU data fusion using multisensor
Kalman filtering: introduction of contextual aspects. Inf Fusion 7:221–230. https://doi.org/10.
1016/j.inffus.2004.07.002
Design and Implementation of Efficient
IoT-Based Smart Oil Skimmer
S. Rajesh Kannan , V. G. Rajagopalan , H. Ramakrishnan ,

S. Sibi Selvan , and Sushanth Krishnamithran
Abstract An oil spill is the release of liquid hydrocarbon into the environment
especially in oceans and is a form of man-made disaster. 35.7% of oil spilt in oceans
is from oil tanker vessels. This hugely affects the marine ecosystem and also tourism,
marine economy and the human health. Thus, oil spill is a biological, economical and
a social disaster. This paper idea aims to detect the oil spilt from vessels accurately
and also as early possible with the help of different sensors and mechanisms. This
idea not only stops at that but also makes use of an oil skimming that will be pressed
into action as soon as oil spill is detected. An ESP8266 microcontroller, a GPS
module and also Wi-Fi signals connected to an app interface to control the entire
process. This idea is a true and unique solution that employs modern technologies
like IoT as well as an app interface to detect and clean up the oil spilt making it
highly stable, efficient and also economical. The bot which is used to clean up the
oil spilt makes use of a driver module as well as belt-based skimming mechanism
to do its job efficiently. We have designed this project to be controlled manually in
case of any exegesis without much strain making the response without any time lag
and making recovery as soon as possible.
Keywords Load cell · GPS · L298n · Belt-type oil skimmer · IoT · GUI
1 Introduction
Pollution has become a major threat to the environment. There are various kinds of
pollution such as water pollution, air pollution and land pollution, out of which there
is another kind of major pollution that harms aquatic animals and their ecosystem
called oil spill. Oil spill causes a thick dense layer of oil to be formed on the surface
of ocean which prevents the sunlight from entering the surface and thereby leading
S. Rajesh Kannan (B) · V. G. Rajagopalan · H. Ramakrishnan · S. Sibi Selvan ·

S. Krishnamithran
Department of Electronics and Communication Engineering, St. Joseph’s College of Engineering,
Chennai 600119, India
e-mail: rajeshkannans@stjosephs.ac.in
https://doi.org/10.1007/978-981-19-6913-3_26
394 S. Rajesh Kannan et al.
to decrease in dissolved oxygen content in water bodies. Oil spill caused by crude
oil is worse, and the crude oils have very high viscosity and are transported in
ships in heated temperature which causes thermal pollution and increases the water
temperature in the event of oil spill. According to the study made by Center for
Biological Diversity, the deep horizon oil spill which happened in New Mexico
caused harm or killed more than 82,000 birds of 102 species, 6165 sea turtles and as
many as 25,900 marine species. It also resulted in over 25,000 job loss, $2.3 billion
loss in industrial output and an overall $1.3 billion loss in gross regional product,
effects of which are even felt now. The latest oil spill in Mauritius resulted in the
economic shrink of $2.1 billion of the total $15 billion tourist-dependent economy.
Even the small level of oil spill in Ennore, India, caused severe damage to marine
economy for weeks in surrounding regions. The oil spill incidents lead to much
harmful impact due to the non-availability of portable equipment and a matured
industry. We have tried to address these problems with the help of our device. Most
of the oil is transported in oceans and most of which are hard crude oil. These oils
are transported in the hull of the ships, which on damage lead to the spillage of oil in
the oceans. We have designed a primary detection system that measures the volume
of oil stored in the hull of the ship at regular time intervals and alerts in case of any
spillage of oil into the oceans. By this way, we can immediately predict the disaster so
that we will be getting the most required precious moments to prevent a mishap. As
soon as the oil spill is detected, swarm of autonomous oil skimmers will be pressed
into action. These devices can be controlled remotely with the help of IoT and cloud
technologies, thus increasing the speed of action. These devices have installed GPS
modules so that each and every movement of these bots can be predicted accurately.
The devices can be pressed into action either by paradropping, remote release from
land or sea or immediate deployment from the vessel itself. They are propelled with
the help of rotors, and their stability is maintained with the help of flaps. These
rotors are in turn connected to controllers that have an established connection with
command centers. Once these bots reach the desired location, the flaps are deployed
to maintain their stability and the skimming process will be initiated. Since these
devices are small in size and have less weight, they can be easily mobilized and
can be quickly deployed. We make use of belt-type oil skimmers instead of other
types of oil skimmers. We have preferred belt-type oil skimmer because it has fewer
mechanical parts, high efficiency and less weight than others. These skimmed off
oils are then collected in a container whose size depends upon the number of belts
used in the skimming process. As soon as the container is filled, the bot is propelled
to a collection place where these oils can be collected. The typical capacity of our
device is 5–7 L per hour whose capacity can be further increased with the help of
addition of belts. Swarm of these bot deployed sooner based on the detection system
alert can immensely save the oil spill from spreading and can be arrested quickly
thus saving both money and ecosystem.
Design and Implementation of Efficient IoT-Based Smart Oil Skimmer 395
2 Literature Review
In “Oil recovery with novel skimmer surfaces under cold climate conditions”
authored by Keller and Clark [1], the skimming rate of skimmers at cold climate
is tested. Various tests were conducted on the skimming capacity with emulsions
and cold conditions. It was concluded that the cold climatic conditions reduce the
skimming rate only to a very small extent and adjusting the inclination angle would
be able to suffice those declining rate.
In “Aspects of modeling an electric boat propulsion system” authored by Freire
et al. [2], the design of a propulsion system with the help of electric motors for large-
scale boats is mentioned. The paper mentions the hardware in line methodology of
getting the output from a controller and driving the motors with the help of speed
regulator and providing a stable system.
In “Design of a load cell with large overload capacity” authored by Aghili [3],
the application of load cell in large scale of operations is explained. The paper
proposes a methodology to use a special deign structure that ensures the rigidity of
the structure without compromising in senor’s sensitivity. By the development of this
methodology, a sensitivity of 0.2% is achieved.
In “Implementation of GPS for location tracking” authored by Ariffin et al. [4],
the implementation of a portable GPS module for live tracking systems is discussed.
The paper discusses on how to implement the GPS live tracker for various portable
applications and to get the results with at most accuracy.
In “Design and efficiency comparison of various belt type oil skimmers” authored
by Patel [5], the efficiency of various types of belt-type oil skimmers is discussed.
It was concluded that adding more shafts and belts increases the efficiency of the
skimmers greatly and even the small design changes, and perfection of design can
make a significant impact in the efficiency of belt-type oil skimmers.
In “Optimized Design of Electric Propulsion System for Small Crafts using the
Differential Evolution Algorithm” authored by Lee et al. [6], the design of electric
propulsion system is using an algorithm to determine the optimal electric motor and
battery specifications for the basic requirements of a small boat implemented with
an EPS.
In “Development of a handy oil-skimmer” authored by Tatsuguchi et al. [7], the
development of a suitable portable oil skimmers for vulnerable spots is discussed. The
paper discusses the performance of portable skimmers when compared to traditional
ones and also the usefulness of these skimmers at places not suitable for traditional
skimmers.
In “Analysis of belt type skimmer” authored by Thombare Babasaheb et al. [8],
the size and type of material to be used in belt-type oil skimmer for highly efficient
process are discussed.
In “Review on analysis of oil skimmer” authored by Pund et al. [9], the design
of efficient belt-type oil skimmer is discussed. It is concluded that the best design is
capable of making the process highly efficient.
In “A Review of Mobile Oil Skimmer” authored by Sathiyamoorthy et al. [10],

the oil spill response by utilization of belt-type skimmer which is proved to be faster
and highly efficient in oil spill recovery is discussed.
In “Research on shipborne aided navigation system based on enhanced traffic
environment” authored by Fan et al. [11], the bot on its movement through oceans
is exposed to meteorological factors, climatic conditions which can be forecasted
by consistent live tracking using GPS for allowing the bot to operate in safer ocean
conditions which has been illustrated.
In “Real-time Location Tracker for Critical Health Patient using Arduino, GPS
Neo6m and GSM Sim800L in Health Care” authored by Kanani and Padole [12],
the location and identification of patients through GPS location are identified, which
can be modified to be used as in the same case for identification of bot GPS location.
In “Microcontroller Based Controlling of Electric Vehicle” authored by Channi
[13], controlling of DC motors using a microcontroller through a cloud-based plat-
form that could be accessed using android is emphasized. The DC motors are
controlled by L298N driver module, and it is concluded that it was efficient enough
to operate the vehicle.
3 Smart Oil Skimmer
With the increase of oil spill incidents in the world, it becomes an indomitable task
to clean up the oil spill as soon as possible and prevent the further damage. For
this purpose, we have used a detection system that detects the leakage of oil spill in
oceans and a device that can be deployed immediately as soon as possible to clean
up the oil spill effectively which can be controlled from a remote location preventing
the physical contact with the heavy oil. The flow diagram of our concept is illustrated
in Fig. 1.
3.1 Detection System
The primary detection system makes use of a load cell connected with controller
and a HX711 amplifier to amplify the values from load cell. A load cell is used to
measure the weight of an object placed on its. The load cell works similar to a sensor.
When load is applied to the load column, it undergoes compression and there is a
change in its length. This load cell is used as a transducer as it converts one form of
energy into other form, and it converts the force applied on it into change in length.
This change in length cannot be known as it is. A strain gauge is attached to load
cell which acts as a sensing element.
The electrical resistance of the strain gauge is dependent on the length and cross-
sectional area. Therefore, when length increases and its area decreases, the resistance
will increase. This is established by Ohm’s law for resistivity, R:
Fig. 1 Flow diagram of smart oil skimmer
R = ρl/ A (1)
where ρ is the resistivity of load cell element (Ω m), l is the length (m), A is the cross-
sectional area (m2 ) and R be the electrical resistance of spring gauge [14]. A strain
gauge is also a passive transducer which converts a mechanical deformation into a
change of resistance. When strain gauge gets compressed, its length changes, which
depends on the magnitude of force applied on top of the load cell. The resistance of
the connected strain gauge undergoes a change when the length of cell varies. They
are designed to make the two surfaces of the cell body bend depending on applied
load. Strain placed on convex surface will stretch, while placed on concave surface
will contract. The change in resistance is measured in terms of change in the output
voltage, and it is very small so it needs to be amplified using a module that performs
amplification. Thus, the load cell provides a voltage level equivalent to the weight
placed on the load cell.
The change in resistance is achieved by the Wheatstone bridge technique. Wheat-
stone bridge is implemented by arranging four strain gauges. The change-in-electrical
resistance and impedance are measured by Wheatstone bridge with very small vari-
ations. The detected differential voltage in the center of the bridge is measured.
The circuit converts variation in resistance into voltage change, and this eliminates
offset voltage that appears in the voltage divider. The signal obtained from the strain
gauge requires amplification. Amplification of the voltage signal from the Wheat-
stone bridge is necessary because the strain gauge’s Wheatstone bridge has only
2 mV/V of sensitivity. It gives a typical 3 V of cell excitation, and voltage available
at full scale is 6 mV which is very low and hence this requires amplification. This is
achieved by an instrumentation amplifier with characteristics like high differential
gain, high input impedance, low offset, low input currents and low drift.
Let us see the need of load cell for knowing the volume of oil leaked. We know
that
1 cm2 = 1 × 10−4 m2 1 kgm2 = 0.1 Pa (2)
Let the weight of tank filled with oil that was found using load cell be W = 5 kg
(for example) and density of crude oil given as
D = 820 kg/m3
that has specific gravity of 0.82. Let the assumed mass of oil tank to be nearly the
same as weight, M = 5 kg.
Assuming the oil-carrying container to be a perfect cylindrical carrier of radius,
R = 1 m. Therefore,
Area of tank be A = π × R × R = 3.14 m2 (3)
Now we have pressure exerted by liquid of mass M be
P = mass of liquid/area of tank base = 1.59 kg/m2 (4)
As we already know that 1 kg/m2 = 0.1 Pa, then for pressure exerted by liquid it
will be 0.159 Pa. We know that
1 Bar = 100, 000 Pa.
So, in terms of bar for pressure exerted, P = 0.0159 m bar, and this is actually the
hydrostatic pressure exerted by 5 kg oil.
Now we have the formula to calculate the height [14]
H = P/SG × D × g = 1.59/0.82 × 820 × 9.8 = 0.24 × 10−3 (5)
So, height is obtained; hence, volume can be easily calculated
V = π × R × R × H = 3.14 × 1 × 0.00024 = 0.00075 m3 (6)
Now due to some unavoidable circumstances, the hull is broken and hence the
carrying tank is damaged due to which oil is spilled into the ocean, so immediate
action will be taken to fix the hull. By that time, say 2 kg of oil is lost (load cell will
show 3 kg as it weighs only the entire container).
Now we have the same parameters of SG, D and mass M , of 2 kg.
Hence, the hydrostatic pressure exerted by M , be
P , = M , /A = 2/3.14 = 0.636 kg/m2 . (7)
Height this oil can reach be
H , = P , /SG × D × g = 0.636/6589.52 = 0.0000965 m. (8)
Volume of oil after leak be
V , = π × R × R × H , = 3.14 × 1 × 1 × 0.0000965 = 0.00030 m3 (9)
So, this is the volume of oil left; therefore, the oil leaked into the waters will be
V − V , = 0.00075 − 0.00030 = 0.00045 m3 . (10)
Hence, the volume of oil leaked is found to be 0.00045 m3 .

This calculation can be coded into any programming language using any software
available [15]. The generated code is then used to know the volume spontaneously
at any instant of time. This data is useful for oil spill response and recovery crew to
estimate the amount of oil lost which had to be recovered back immediately.
The need to find weight involves tedious calculation and looks to contemplate
us about the indirect approach in finding the volume of oil lost. Whenever the oil
containers are filled with oil some space which is left unoccupied, say like 25–30%
is left vacant which is indicated in Fig. 2. This is because the vessel might have
numerous vigorous movements and halts throughout the course of its journey. When
the ship suddenly stops, the oil inside the carrier remains to be in motion state as a
result of inertia of motion. This in turn pushes or rises the oil in forward direction.
Likewise, when ship sets into sudden motion after in rest, this pushes oil backward.
This happens then and there in its journey, and if no space is left inside the container,
the oil might leak out or overflow. Crude oil filled into containers might also exert
some gases. This causes pressure inside the container; if the tank is full, the tank may
not bear high pressure, and hence, oil might leak out. Partially filling the oil carriers
helps to reduce this effect caused by pressure. The load cell has one end that needs
to be fixed on a base, while the other end is let free. This setup is very important only
then the load cell can undergo deformation.
HX711 module is an amplifier breakout works to read the measured weight data
from the load cell. This is a 24-bit precision A/D converter chip [3]. It is designed
Fig. 2 Illustration of oil tank
Oil filled till 70%

Fig. 3 Load cell connected

with HX711 module
for high precision scale designs in electronics. The internal integration provided by
this module is 128 times the integration available in gain amplifier. The input circuit
in this case is pressure sensor load cell; this acts as the front end for the module. This
setup along with load cell is interfaced to Arduino Uno microcontroller to know the
weight. Communication is availed in HX711 by two wire interface called the clock
and data which is shown in Fig. 3. This module provides higher integration and faster
response when compared to any other modules. The features provided by HX711
are as follows:
• The module provides two differential input channels that are selectable.
• It gives on-chip power supply regulator for ADC power supply and load cell.
• On-chip power-on-reset is available.
• Provides easy digital control and simple serial interface.
• Choosable output data rate is available, either 10SPS or 80SPS.
• Parallel 50 and 60 Hz supply rejection is also provided.
• It provides operational supply voltage at a range of 2.6–5.5 V.
• The range of operational temperature is − 40 to + 85 °C.
With the help of this system, we are able to measure any leakage of oil into the
ocean and alert the crew so as to act accordingly.
3.2 Bot Propulsion System
The movements of a system in different directions are controlled using a particular

module called L298n motor driver module. This module is mainly used to commu-
nicate between the hardware motor component and the controller board. When the
program is uploaded using IDE for controlling motors using a single L298n motor
driver module and the GUI app. When uploading is completed, we can use the GU
interface in our phones to control the motors movements and also speed of the motor.
The motors are connected to the propeller for the movement in water. Our project
involves the use of L298n module and GPS module, where we get the location of
the oil spill using the GPS module and then these coordinates are processed through
the controller. After getting the coordinates, the bot moves to the particular location
using the L298n motor driver module which communicates with motors to move
in the specified direction [2]. As we can see, the driver module is working on the
principle of H-Bridge control for directional movement and PWM for speed control.
There are different types of L298 driver modules, but the L298n module has more
stability compared to the driver modules in terms of speed and durability. The L298
driver module is a high-voltage, high-current dual full-bridge driver designed to
accept standard TTL logic levels and drive inductive loads such as solenoids, DC
and stepping motors. The L298N driver module communicates with the Arduino
Uno microcontroller. Just by connecting the motors directly to the microcontroller
will not work as it could even damage the microcontroller because Arduino pins are
restricted to 50 mA current, but this is very less than 200 mA which is the requirement
of a small motor. Hence, when the microcontroller is connected to the L298N driver
module runs the motors as its output current is 2A. This is due to the function of
driver module to convert low-current signal into high-current signal [13]. In Fig. 4,
two enable inputs are provided to enable or disable the device independently of the
input signals as shown. Instead of using the diesel engines for propulsion, the electric
propulsion greatly reduces the pollution while maintaining the same degree of perfor-
mance [16]. Electric propulsion system is needed to be designed for every vehicle
for its transportation. Similarly, we have developed battery management system to
power up the propulsion motors. It offers us the feature to recharge the bot whenever
if required, and also it has been designed in such a way that battery lasting capacity
is higher and drain is minimalized. Once charged, the propulsion system with other
associated systems can withstand for 8 h. Algorithms were followed in our battery
management system to attain the best power up ability to charge our bot [6].
This type of driver module is used only when:
1. High-power motor driver is required.
2. It is needed to operate different loads like motors and solenoid, etc., where an
H-Bridge is required.
Fig. 4 Circuit diagram of

propulsion system
The L298n module is quite different from the other modules since it can communi-
cate faster than L293d module which has a quite a low communicating and processing
speed. L293d drivers operate between 4.5 and 36 V, whereas L298n can operate at
up to 46 V, L293d output current for each channel is 650 mA and it is 2 A for L298.
Hence, the heat sink is provided in L298 motor drivers. The batter draining capacity
of both modules is almost same, but L298n driver module has a heat sink which
dissipates heat can save battery life and the voltage regulator regulates voltage in
case of low supply of voltage. The double H-Bridge can handle up to:
This L298N motor driver module uses two techniques for the control speed and
rotation direction of the DC motors. These are pulse width modulation (PWM)—for
controlling the speed. Halving the needed supply voltage can generally reduce the
speed of motors due to proportionality. However, practically it is not possible to
change the supply voltage as we want. The speed control can be achieved by PWM
signal which is a high-frequency signal that controls the switching at a faster rate.
The switching changes the average value of voltage required by the DC motors. Duty
cycle of the PWM decides the value of average voltage. Duty cycle is the ratio of
duration of signal HIGH to the total time period of PWM’s duty cycle, and H-Bridge
is shown in Fig. 5—for controlling rotation direction. This module can control two
DC motors or one stepper motor at the same time. L298n motor driver module uses
the H-Bridge principle to control the direction of rotation of a DC motor as it contains
two enable inputs to initiate or non-initiate further devices like motors attached to the
output. In this technique, H-Bridge controls DC motor rotating direction by changing
the polarity of its given input (input voltage). The H-Bridge is controlled by logic
level signals. Since our designed bot needs all possible direction movements for
its propulsion which can be achieved by standard TTL logic signals. Each possible
logical combinations is responsible for varying operations of the DC motors used.
On applying different polarities of voltages on two sides of the motors, these actions
control switching to obtain HIGH or LOW logic signals to operate the motors in our
desired modes. Transistors (BJT or MOSFET), with the motor at the center forming an
H-like configuration. Input pins actually control the switches of the H-Bridge circuit
inside L298N module. Two inputs pair up to represent the logic signals HIGH and
LOW or HIGH and HIGH or LOW and LOW. Each representation is responsible for
operations like brake, clockwise operation and anti-clockwise direction. We change
the direction of the current flow by activating two particular switches at the same time,
this way we can change the rotation direction of the motor. A small representation
of working is given below.
The L298N driver module can control up to four DC motors without speed control
or two DC motors with speed control as illustrated in Fig. 6.
3.3 Live Tracking System
The GPS module is connected with a controller that in turn is connected with the
command center with the help of IoT and cloud technologies. An app is developed
Fig. 5 Implementation of
propulsion system
Fig. 6 Representation of
H-Bridge
to control and track the device. For prototyping purpose, we have used NEO 6 M
UBLOX module and Blynk IoT platform to integrate the operations of the device,
and the control is simple and effective with the use of GUI interfaces [2]. With the
help of the tracking system, the live location of the device can be known and can
be propelled to the desired location easily. Without the actual virtual representation
of bot location, the monitoring purpose is failed and could lose track of it. Just by
fetching only the geographical coordinates of the current bot location would not be
an accurate information. So the GPS module is interfaced with cloud platform via
Blynk that projects the bot location very accurately with all coordinates and direction-
related information [12]. Our idea is to focus on the movement of bot in the middle
of the ocean or it can be unaccessible places to reach out for man power. There
are several factors that influence the bot movement, and they include hydrological
factors and meterological factors. With the help of the aforementioned GPS module
which is interfaced with cloud technology like the Blynk platform, such influential
factors are forecasted to redirect and manually operate the bot to safer conditions.
This data obtained from the abovementioned factors can be studied so that the bot
can be further improved and reengineered to operate in such difficult conditions [17].
This retrospective study is possible because of our live tracking system [11].
3.4 Skimming System
After the device is driven to the spilt location on the ocean, then the skimming
mechanism is initiated. Belt-type skimming mechanism is employed in our device.
The rate of skimming depends on various factors such as the oil viscosity, density,
temperature, belt material, the alignment and the skimming angle. We know that oil is
a non-polar substance and water is a polar substance and also oil being less dense than
water and that is the reason why oil floats on the surface of water. Hence, a non-polar
substance type of belt is required to attract the oil. Oleophilic type of belt is used so
that it offers a maximum cohesive property and a maximum skimming rate. Oil spill
recovery is attained at its best efficiency when perfect type of skimmer is chosen.
Belt-type skimmer is highly recommended and capable enough of skimming the oil
at a much faster rate than any other types of skimmers. In our model, the belt type is
equipped with pulleys that are operated by DC motors. After many revolutions, oil
gets stuck to the material that is coated on the belt, and it is also easy to replace [18].
It was concluded that polyurethane is the best material where less clamor activity is
required. It was concluded that the oil recovery rate is inversely proportional to the
viscosity and the temperature of the surface. It was also mentioned that the pH values
of the water have an effect in the recovery rate. The oil recovery rate is proportional to
the speed of the belt and the surface area of the belt [19]. Thus, considering all these
proofs of concepts we have designed our skimmer out of polyurethane material of
area 80 cm2 and the skimming motor of RPM of about 40. The belt is inclined at a 90
◦
angle so as to facilitate proper contact with the surface. We have designed a BMS
system that is capable of supplying the skimming motor and propelling motors for up
to 8 h. The BMS is made as a rechargeable type so as to reuse the batteries again and
again. The life of the device can be further increased with the addition of batteries.
Our design is capable of skimming about 5–7 L of oil from water depending on the
above-said conditions such as temperature, density and viscosity. Much emphasis
was made to increase the efficiency of the skimmers in order to deliver a perfect
solution. Even though these kinds of skimmers are available in the market, they lack
credible performance and are mostly manually operated. With our device, we come
to a remotely operated skimmers so that humans are not required to risk their health
by manually going to the hazardous locations and carrying out the task of removing
the oil.
With the inclusion of new technologies like IoT and cloud, we have come up with
a unique solution that comes in handy at the times where the risk of oil spill has
significantly increased and at times where people are increasingly concerned about
their health.
The experimental results after prototyping process have revealed that our proto-
type consumed very less amount of power but was able to deliver great results. With
consuming just around 28.8 mAh, it was able to skim off about 5–7 L of oil from still
water at normal conditions. Our battery management system was sufficient to power
up the bot for up to 6 h. Our bot was able to deliver a peak RPM of about 6400,
marking a max speed of about 27.358 knots and the GPS was accurate to about
2.5 m which can be further increased with the use of marine GPS. We have used
polyurethane as the skimming belt material as it is proved that it is the most efficient
material but we can also employ other materials depending on the conditions.
Figure 9 illustrates oil spill that is being detected with the help of load cell and
the abovementioned calculation which are programmed, and in case of any leakage
of oil, it immediately indicates the position so that we can start to act on it. Figure 7
shows the smart oil skimmer prototype that we developed for the purpose of testing
our idea, and in Table 1, the results are given. A polyurethane material is used as
the skimming belt in our prototype, and Li ion batteries were used to power the
entire propulsion and skimming process. It was observed that the prototype was able
to recover 5–7 L of heavy oil from water every hour and capable of traveling at
27.358 knots at calm waters. The main idea of this device to carry out the task of
skimming without physical involvement was a success with the prototype being able
to be tracked with GPS as shown in Fig. 10 in 2.5 m of accuracy. The GPS data are
obtained from the GPS module and processed and entered in a cloud-based platform
so that it can be viewed from a remote place too. Figure 8 illustrates that the control
of the prototype was smooth with the help of a custom-made GUI interface which
enables even a semi-skilled person to control the whole process. The GUI has two
different parts: one to track the location and other one to control the device. Figure 8
shows that the device can be easily controlled with the help of sliders and movable
circles. Many such devices can be synchronized and can be controlled by just one
person. Thus, we have made it possible to detect the oil spill as early as possible and
thereby deploy the devices that can be propelled and tracked from a remote location
and execute the skimming action.
5 Conclusion
The oil skimmers find number of applications not only restricted to that of oil spill in
oceans. The oil skimmers can be employed at various industries including thermal
plants, desalination plants, offshore and onshore oil platform. The major application
Fig. 7 Prototype developed
Table 1 Results of prototype

Parameters Inferred value
model
Battery consumed 28.8 mAh
Hours lasted 6h
Max RPM delivered by skimming 6400 RPM
motors
Max operational of driver module 46 V
Max speed attained in knot 27.358
GPS accuracy 2.5 m
Belt-type skimmer material Polyurethane material
Oil skimming capacity 5–7 L
apart from oil spill in oceans is in sewage treatment and desalination plants, instead
of purifying the water at one end, simultaneous treatment of water along the flow of
water bodies can be made so as to maintain the same quality of water everywhere.
Based on the effluent that needs to be separated, different kinds of material belts can
be used while using the same mechanism.
The same mechanism can be employed to make the skimming process easy
and with no physical contact, thereby preventing harm to humans involved. With
the application of IoT, cloud and GPS, the traditional oil skimmers can be truly
synchronized with Industry 4.0 tech.
Fig. 8 GUI for the control

of prototype
Fig. 9 Implementation of
load cell concept
Fig. 10 GPS live tracking

implementation
References
1. Keller AA, Clark K (2008) Oil recovery with novel skimmer surfaces under cold climate
conditions. Int Oil Spill Conf Proc 2008(1):667–671
2. Freire T, Sousa DM, Branco PJC (2010) Aspects of modeling an electric boat propulsion
system. In: 2010 IEEE region 8 international conference on computational technologies in
electrical and electronics engineering (SIBIRCON), pp 812–817
3. Aghili F (2010) Design of a load cell with large overload capacity. Trans Can Soc Mech Eng
34:449–461
4. Ariffin AB, Aziz NHA, Othman KA (2011) Implementation of GPS for location tracking. In:
2011 IEEE control and system graduate research colloquium, pp 77–81
5. Patel M (2013) Design and efficiency comparison of various belt type oil skimmers. IJSR
4(1):998–1002
6. Lee DK, Jeong Y-K, Shin JG, Oh D-K (2014) Optimized design of electric propulsion system
for small crafts using the differential evolution algorithm. Int J Precision Eng Manuf Green
Technol 1(3):229–240
7. Tatsuguchi M, Mizutani M, Sano M, Fudo M, Ishida H, Fujita I (2014) Development of a handy
oil-skimmer. IEEE Techno-Ocean’04 3:1464–1469
8. Dawood MA, Algawi RJ (2017) Study of operating conditions for oil skimmer apparatus from
water. In: 2017 international conference on environmental impacts of the oil and gas industries:
Kurdistan region of Iraq as a case study (EIOGI), pp 65–70
9. Thombare Babasaheb B, Barse Babasaheb N, Barhat Ganesh B, Kolhe Sani M, Jagtap Harshal
B (2018) Analysis of belt type oil skimmer. IJARIIE 4(10):2917–2922
10. Pund R, Mhaske R, Rahane S, Rajput S (2018) Review on analysis of oil skimmer. IRJET
05(10):680–681
11. Fan Y, Huang L, Jiang D, Xu X (2018) Research on shipborne aided navigation system based
on enhanced traffic environment perception. ONE 13(10):e0206402
12. Kanani P, Padole M (2020) Real-time location tracker for critical health patient using Arduino,
GPS Neo6m and GSM Sim800L in health care. In: 2020 4th international conference on
intelligent computing and control systems (ICICCS)
13. Channi HK (2021) Microcontroller based controlling of electric vehicle. LAP Germany Editor:
LAPISBN: 978-620-3-47186-1
14. Müller Ivan A, de Brito RM, Pereira CE, Brusamarello VJ (2010) Load cells in force sensing
analysis—theory and a novel application. Instrum Meas Mag IEEE 15–19
15. John W, Eren H (1999) The measurement instrumentation and sensors handbook, 1st edn.
Springer
16. Son Y, Lee S, Sul S (2018) DC power system for fishing boat. In: 2018 IEEE international
conference on power electronics, drives and energy systems (PEDES), pp 1–6
17. Lee S, Tewolde G, Kwon J (2014) Design and implementation of vehicle tracking system
using GPS/GSM/GPRS technology and smartphone application. In: 2014 IEEE world forum
on Internet of Things (WF-IoT), pp 353–358
18. Sathiyamoorthy V, Arumugam K, Arun Pragathish M, Barath BN, Baskar M, Balamurugan S
(2018) A review of mobile oil skimmer. IJET 7(3):58–60
19. Khan MA, Mostafa SMG, Rahman A, Hoque MM, Islam MT, Kalam M (2021) Implementing
and improving the IoT based weather monitoring and controlling double discs type oil skimmer.
In: 2021 2nd international conference on robotics, electrical and signal processing techniques
(ICREST), pp 499–503
Comparison of Discrete Time Sliding
Manifold and Its Impact on System
Dynamics
Shaktikumar R. Shiledar and Gajanan M. Malwatkar
Abstract This paper presented the comparative study on development in the discrete
time sliding mode control (DTSMC) especially concern with the third-order process
to analysis of the system state variables. In the literature study, it was found that
many researchers were fascinated by the study of DTSMC strategies implemented
for control engineering applications. In this paper, different DTSMC approaches
such as second-order DTSMC, DTSMC for the higher order system, and DTSMC
with discrete PID sliding manifold have been included to find its effect on system
dynamics such as interval state variables of the systems. The presented controller
design depends on discrete controller techniques which are obtained from the laws
available in literature in recent years. In this paper, the presented techniques give
minimum value for the control input to reduce the effort for a controller, and it is
implemented on the proposed model of the system. In the formulation of DTSMC
law, the calculated values of gains from the constant matrix or gain tuning parameters
are used to derive one of the laws of DTSMC. In the presented work, stable poles of
the system have been given the desired value for the controller input signal. The major
advantages of the presented techniques included dynamics of the state variables and
effect of variables arises difficulty in implementation. The obtained simulation results
show the comparative study for DTSMC the approach in the form of stabilization of
the system states and its output responses.
Keywords Performance · Comparison · Discrete time sling mode · Simulation ·

State variables · Controller action
S. R. Shiledar (B) · G. M. Malwatkar

Department of Instrumentation Engineering, Government College of Engineering,
Jalgaon 425002, India
e-mail: shaktikumar.shiledar@gcoej.ac.in
https://doi.org/10.1007/978-981-19-6913-3_27
412 S. R. Shiledar and G. M. Malwatkar
1 Introduction
In recent days, all advance and optimized control strategies are based on digital
devise. In actual, implementation of the controller in discrete time domain becomes to
use the microprocessor or personal computer devices for computing the algorithms. It
is well known that control system design shifted from continuous to digital becomes
unstable after the sampling process. In the literature study, it was found that first
time work presented in [1] studied the discrete variable structure controller theory
and originated this from modification of the equivalent continues SMC reaching
condition applicable to solution for reaching condition in discrete [1, 2]. However,
in discrete sliding mode control (DSMC) structure case, all states do not stay on the
sliding manifold instead of it make a zigzag motion around the manifold with a certain
band becomes given the quasi sliding mode form. The chattering is arising mainly
due to the finite sampling time, so dynamics of the system perform zigzag motion
in the neighborhood of the sliding surface is unavoidable. A second-order DTSMC
law or a non-switching type reaching law is used for application to communication
networks [3]. Gao et al. is investigating the quasi-sliding mode design controller
technique and achieving the rule of thumb for a single input single output system
[4]. There were two types of state trajectories explained with sufficient conditions
to fulfill some attributes to design sliding mode and reaching law [2]. The chat-
tering effect occurs in the CSMC, when the sampling rate is high which has been
minimized by selection of proper quasi-sliding mode band around the switching
manifold. Misawa also inside into design the DSMC for a single input linear system
[5]. In the work of [5], Lyapunov stability theory is applied to design the DSMC
and Lyapunov function consider as V (k) = 21 s 2 (x). The DSMC presentation shows
the trade-off between the sliding band, system modeling errors, and control of gain
parameters. The performance of the Misawa [5] DSMC is compared with Furuta’s
et al.’s and Chan et al.’s results [5, 6]. On the other hand, the discrete time slide mode
control strategies introduced by Bartoszewicz et al. redefine the concept of slide
mode show that the strategy requires less control effort to ensure better performance
and system stability than regulation of Gao law [7]. In [7], designed algorithm no
need to cross system states successively in each steps or iteration so that ultimately
the effect of chattering was reduced as compared to results given in [2]. In light of
the paper flow is in Sect. 2, included introduction of DTSMC, in Sect. 3, describe the
problem formulation, Sect. 4 discussed the different DTSMC techniques, in Sect. 5,
contains simulation and discussion of outcomes. Section 6 included the conclusion
and comments of the newly developed prevalent controllers.
2 Preliminaries
The discrete time control system analysis can be defined into two ways: first is the
system derived from continues time output at the discrete instant, i.e., sampled data
Comparison of Discrete Time Sliding Manifold and Its Impact … 413
system. Second way is the system which are in the discrete form and where system
states are introduced only at discrete time instant.
Let us consider the representation of a CT system converted to a DT system with
a defined sample time (T ) and quantization rate for the given continues signal to
becomes discrete signal.
ẋ = Ac x̂(t) + Bc u(t)
(1)
y = Cc x̂(t) + Dc u(t)
where x̂(t) = [x1 , x2 , . . .] is the dynamics, matrix Ac , Bc , Cc , and Dc is the system

constant matrices with proper dimensions for the system input, control input, and
system output, respectively. The pair (Ac , Bc ) must be controllable in this case the
solution for Eq. (1) given as follows:
The solution for Eq. (1) given as follows:
t
x(t) = ϕ(t − tinitial )x̂(tinitial ) + ϕ(t − τ ) ⎡u(τ )dτ
tinitial
In discrete signal formation process consider the input signal constant during the
two sampling interval to form the discrete signal. The solution of Eq. (1) exists in
case of tinitial ≤ τ ≤ (k + 1)T where assume the initial interval tinitial = kT the term
t
tinitial ϕ(t − τ ) ⎡u(τ )dτ in Eq. (1) denoted by θ (t − kT )u(kT )
x((k + 1)T ) = ϕ(T )x(kT ) + θ (T )u(kT ) (2)
from Eq. (2) can be written for T = 1 and put θ (T ) = ⎡
x̂(k + 1) = ϕx̂(k) + ⎡u(k)

y(k) = H x̂(k) + Du(k) (3)
The most of the researchers are worked on continuous SMC, and the discrete
counterpart of it are attempted for some of the applications. The VSC theory imple-
mentation on the computer-based system with a relatively low sampling period lead
to discrete SMC. The availability of significant algorithm for implementation of
DTSMC on various types of systems are well studies in literature [1, 4, 6].
3 Problem Formulation
In the problem statement, consider the nth order form of the discrete time process
represented in form of transfer function

(b0 ± δb0 )z P + (b1 ± δb1 )z p−1 + · · · + b p ± δb p ld
z , p≤n (4)
z n + (a1 ± δa1 )z n−1 + · · · + (an ± δan )
where the term δ representing the parametric uncertainty in system. The term ld
represents the delayed samples and it written as d = lTd , where T represents the
sampling period. Equation (4) shows in state space discrete form as
x(k + 1) = (ϕ + δϕ)x̂(k) + ( ⎡ + δ ⎡)u(k) + f (k)

y(k) = H x̂(k) (5)
where the input state variable, input control signal, and system output are represented
Δ
by x (k) = Rn , u(k) = R and y(k) = R1×n , respectively. ϕ ∈ Rn×n , ⎡ ∈ Rn×1

and H ∈ R1×n are constant system matrices. The system parametric uncertainties
are represented as δϕ and δ ⎡, respectively. The nonlinear external disturbances are
represented by (k). In the problem formulation consider the stability and observ-
ability condition satisfied the system with (ϕ + δϕ) and (ϕ, H ). Equation (4) can
be represented in form of single input and single output, linear discrete time without
considering parametric uncertainties and disturbances in system as
Δ Δ
x [(k + 1)T ] = ϕx (kT ) + ⎡u(kT ) (6)
The dynamic response of the system error is calculated using the difference
between reference or input state with the desired value of the signal. Here, suppose
function error e(k), reference signal r(k), and system output y(k) represent as
e(k) = r (k) − y(k)

Δ
e(k) = x̂(k) − xd (k) (7)

Δ Δ
where x (k) is the system states, and x d (k) is the system desired states
4 DTSMC Techniques
The following DTSMC laws have been taken for comparison and simulation studies.
4.1 Second Order Discrete DTSMC
The work given in [8] design of second-order DTSMC with observer and equivalent
control. Assume the system equation fulfill the condition of matrix relation.
(A, H) is controllable and observable
x(k + 1) = ϕx̂(k) + ⎡u(k) + ⊓(k)

y(k) = H x̂(k) (8)
where ⊓(k) = δϕx̂(k)+δ ⎡ u(k)+d(k). On the basis of observer theory formulation

of first-order discrete time sliding mode observer (DSMO) formulated as

x(k + 1) = ϕx(k) + ⎡u(k) + L y(k) − ŷ(k)

ŷ(k)
+ Msat (9)
K
Δ Δ Δ Δ
y (k) = H x (k) where M ∈ Rn , y (k) = y(k) − H x (k) is chosen as the sliding

manifold, L ∈ Rn×1 is the observer gain matrix, and K denotes the constant. The
formulation of the total controller is written in the following form:
u(k) = u eq (k) + u sw (k) (10)
where u eq denoted as the equivalent control law and usw as the switching control.
The combined law can be represented as
−1
u eq (k) = c, ⎡ −βs(k) − c, ϕx̂(k) + c, (xd (k + 1))

s(k)
− Msat
σ
where s(k) are the sliding functions and xd desired state vector trajectory. The choice
for selection of M, β, σ and c, are given in [8].
4.2 DTSMC for Higher Order Systems
In the contribution by Khandekar et al. [9], a robust controller design for the higher
order time delay system was demonstrated. A sliding mode surface designed using
the function of the state variables and generated error from the reference signal. The
gain matrix or tuning parameter for the controller selected using the pole placement
strategy. The stability of the proposed sliding mode was validated through the pole
placement method [10]. The controller has satisfactory implemented on oscillatory
behaviors type of system with minimization of the chattering of the controller. The
sliding surface designed in the work includes of [9]
Δ
s(k) = cx (k) − K t e(k − 1) (11)
where c = [c1 , c2 , . . . cn ] and K t are gain matrix and it used as the tuning parameter
for the designed controller. The guideline for selection of tuning parameter and
constant are given precisely for system with higher order. The representation of
sliding surface condition as
s(k)[s(k + 1) − s(k)] < 0. (12)
The equivalent control law presented in [9] can be a modified form of basic
DTSMC.
u eq (k) = (c ⎡)−1 [−cϕx(k) − cd(k) − K t e(k)] (13)
The property of robustness to sustain to parameter variation and perturbations

achieved by the introducing of a discontinues with the equivalent control law
u sw (k) = αsgn(s(k)) (14)
The signum function used to switching for states on designed hyper plane. So,
total control law written as
u(k) = u eq (k) + u sw (k) (15)
In discrete sliding mode controller, it is found that the proper tracking of the
reference signal with improved results were obtained compared to continues sliding
mode algorithm were achieved in [9, 11, 12].
4.3 DTSMC with Modified Law
In work of [13], Ghabi and Dhouibi designed of the disturbance rejection estimation
algorithm presented to overcome the effect of uncertainties without modification of
the control law. This control strategy provided stability of close loop dynamics of
the system by reducing chattering. The design of the slide mode control ensures to
fast access and stability of the system (k + 1) provided as
s(k + 1) = c, x̂(k) + c, x̂(k)u(k) + ζ (k) − c, x̂d (k + 1) + λs(k) (16)
According to the reaching law in Eq. (16), the control law for the discrete uncertain
system given by the expression
u(k) = (c, x̂(k)) − 1[−c ⎡ x̂(k) + c, x̂d (k + 1) + s(k) − λs(k)

− βsign(s(k))] (17)
where λ is the positive constant, and ζ (k) ∈ R is equivalent disturbance included

with parametric uncertainty.
4.4 DTSMC Using Discrete PID Sliding Surface
In work given by Shiledar et al. [14] designed the sliding manifold using discrete PID
law with the estimation of the tuning parameter gain constants. The derived reaching
law shown the guaranty of the stability also, it reduces the reaching time of states
variable on sliding manifold. The sliding surface at instant s(k + 1) represented as

s(k + 1) = cϕx̂(k) + c ⎡ u(k) − K t ê(k + 1) + k p ê(k + 2) − ê(k + 1)

+ ki ê(k + 1) + kd ê(k + 2) − 2ê(k + 1) + ê(k) (18)
From Eq. (18), the design of control law as

u eq (k) = (K ⎡)−1 −K ϕx̂(k) + K t (ê(k + 1) + k p ê(k + 2) − ê(k + 1)

+ ki ê(k + 1) + kd ê(k + 2) − 2ê(k + 1) + ê(k) (19)
Δ
c is the sliding gain parameter, e(k) is the error function, kt is the gain constant, and
k p , ki , kd , are the tuning parameter constants.
Consider the transfer function of higher order time delay system given in [15].
1
G(s) = e−0.5s
s3 + 11s 2 + 35s + 25
The discrete form of model obtained by pole-zero method with chosen sampling
interval Ts = 0.1s and representation of continues transfer function equation in
z-domain as
0.0001473z 2 + 0.0002947z + 0.0001473 −5

G(z) = z
z 3 − 2.1182z 2 + 1.466z − 0.3329
The state space matrices of z-transform function represented in controllable

canonical form
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
0 1 0 0 0
ϕ=⎣ 0 0 1 ⎦, ⎡ = ⎣ 0 ⎦, H T = ⎣ 0.002412 ⎦
0.3329 −1.466 2.118 1 0.003481
and D = [0]
The change in the state x1 , x2 and x3 and control signal given by Mihoub et al.
[8], Khandekar et al. [9], Ghabi and Dhouibi [13] and Shiledar et al. [14] are shown
in Fig. 1a–c, respectively. From Fig. 1a–c, it is clear that DTSMC using discrete PID
sliding surface law [14] initially provides higher rate of change of states, and after
some instant it states becomes stable with low gain, while the DTSMC for higher order
systems law [9] has initially low rate of change of states as compared to [14], but states
are stable with high gain. The laws presented by Mihoub et al. [8] state variables are
started from the initial point and all states signals stable with smooth response of gain
value one as shown in Fig. 1. The DTSMC reaching algorithm suggested by Ghabi
and Dhouibi [13] provides initial phase shows the similar and comparable values of
the state while the after some time instant developed law of Ghabi and Dhouibi [13]
state variables gain shifted from gain value less than one as shown in Fig. 1. In Fig. 1d
and Fig. 2, the simulation results involved the responses of input control signals and
system outputs with setting of set point value (SP) is 1 and allowable band of error
consider as 2% of set point value. The sliding gain parameters for Mihoub et al. [8]
are as c matrix chosen as c = [0.3005, −0.5027, 0.4979], the value of sampling step
constant is chosen as Ts = 0.01, and switching gain constant M = 0.9. The constant
terms for [9] included the gain matrix c = [0.3005, −0.5027, 0.4979] which is same
as c matrix defined in [8] and switching constant α = 0.1 is chosen. In the algorithm
of [13], the parameters are chosen as, λ = 0.025 and β = 0.014. For the work of
[14], the sliding parameters include k p , ki and kd as 0.3, 0.01 and 1.65, respectively,
with gain constant kt = 11.15.
6 Conclusions
The development of advanced control algorithms, especially DTSMC makes atten-

tion due to its merits and feasibility of implementation for linear, nonlinear, higher
order, and delay time systems. The design of the DTSMC is the basic requirement
to develop a sliding surface or reaching law and formulation of the relation between
sliding surface and the control input signal are presented. In recent years, DTSMC
strategy is implemented to achieve the goal for high performance tracking with mini-
mization of the high frequency oscillations. So, suppression of chattering becomes
a help to reduce the wear and tear action at the final control element in the process
loop. The prevalent DTSMC laws were studied, which minimize the bandwidth of
the quasi-sliding mode domain and suppresses the effect of chattering. The compar-
ative responses of the state variables, control input, and system output response
were studied and simulated for different recent techniques reported in the literature.
In this paper, comparative study of the second-order DTSMC, higher order systems
DTSMC, DTSMC with modified law, and DTSMC using discrete PID sliding surface
law presented in terms of systems dynamics state variables, control input, and system
output presented with simulation results. It is remarked that there should be a low
rate of change of states so that the control action shall be smooth. From the simulated
result, it is observed that up to some extent all responses presented in the paper are
allowable for smooth control action for higher order time delay system.
Fig. 1 System states x1 , x2 , x3 and control input signal u

Fig. 2 System output y(k)
References
1. Dote Y, Hoft RG (1980) Microprocessor based sliding mode controller for dc motor drives. In:
Presented at the industrial applications society annual meeting, Cincinnati, OH
2. Gao W, Wang Y, Homaifa A (1995) Discrete-time variable structure control systems. IEEE
Trans Ind Electron 42(2):117–122. https://doi.org/10.1109/41.370376
3. Veselic B, Perunicic-Drazenovic B, Milosavljevic C (2010) Improved discrete-time sliding-
mode position control using Euler velocity estimation. IEEE Trans Ind Electron 57(11):3840–
3847. https://doi.org/10.1109/TIE.2010.2042416
4. Gao W, Hung JC (1993) Variable structure control of nonlinear systems: a new approach. IEEE
Trans Ind Electron 40(1):45–55. https://doi.org/10.1109/41.184820
5. Misawa EA (1997) Discrete-time sliding mode control: the linear case. ASME J Dyn Sys Meas
Control 119(4):819–821. https://doi.org/10.1115/1.2802396
6. Furuta K (1990) Sliding mode control of a discrete system. Syst Control Lett 14(2):145–152.
ISSN 0167-6911. https://doi.org/10.1016/0167-6911(90)90030-X
7. Bartoszewicz A, Zuk J (2010) Sliding mode control—basic concepts and current trends. In:
2010 IEEE international symposium on industrial electronics, pp 3772–3777. https://doi.org/
10.1109/ISIE.2010.5637990
8. Mihoub M, Ahmed Said N, Abdennour B (2011) A second order discrete sliding mode observer
for the variable structure control of a semi batch reactor. Control Eng Practice 19(10):1216–
1222
9. Khandekar AA, Malwatkar GM, Patre BM (2013) Discrete sliding mode control for robust
tracking of higher order delay time systems with experimental application. ISA Trans 52(1):36–
44. ISSN 0019-0578
10. Ogata K (2003) Discrete time control systems, 2nd edn. Prentic Hall, New Jersey
11. Eker I (2010) Second-order sliding mode control with experimental application. ISA Trans
49(3):394–405. ISSN 0019-0578
12. Malwatkar GM, Khandekar AA, Asutkar VG, Waghmare LM (2008) Design of centralized
PI/PID controller: interaction measure approach. In: 2008 IEEE region 10 and the third
international conference on industrial and information systems
13. Ghabi J, Dhouibi H (2018) Discrete time sliding mode controller using a disturbance compen-
sator for nonlinear uncertain systems. Int J Control Autom Syst 16:1156–1164. https://doi.org/
10.1007/s12555-017-0185-0
14. Shiledar SR, Malwatkar GM, Jadhav IS, Lakhekar GV (2021) Design of discrete sliding mode
controller for higher order system. Reliab Theory Appl SI 1(60)
15. Wang QG, Lee TH, Fung HW, Bi Q, Zhang Y (1999) PID tuning for improved performance.
IEEE transactions on control systems technology, vol 7, pp 457–465
Volkswagen Emission: An Analysis
on the VW Vento Using Automotive
Network Data
Suprava Sarkar and Nithin Mohan
Abstract The advancement in automotive technology has surged the intricacy, cost,
and size of the governing elements within its system. The demand for more and more
attributes has led automation to reach its pinnacles of modernity. The automation
within the network produces tons of data every second for efficient ECU (Electronic
Control Unit) response. Data extraction and research play a significant role during the
vehicle development processes for better evaluation of the vehicle under trial. Addi-
tionally, the central government of India has ordered that automotive manufacturers
only build, market, and register BS-VI (BS6) automobiles. The rise in automobile
pollution has a severe impact on global warming. Harmful chemical compounds like
NOx , CO, and HC have caused severe health issues among the general public. Hence,
it is essential to regulate these exhaust compounds until a permanent solution is
found. This research paper seeks to analyze CAN Bus data of the Volkswagen Vento,
TSI Highline Plus, petrol engine and determine its emission trail when subjected to
city/highway conditions. Various ECU parameters of the vehicle have been studied,
compared, and scrutinized to extract the desired data that can provide the emission
details of the vehicle under trial.
Keywords Volkswagen emission · Volkswagen Vento analysis · Lambda/oxygen

sensor · Air/fuel ratio · ECU parameters · CAN bus network · OBD II · Telematics
and fleet management
S. Sarkar
Department of Instrumentation and Control, Manipal Institute of Technology, Manipal Academy
of Higher Education, Manipal, Karnataka 576104, India
e-mail: suprava.sarkar@learner.manipal.edu
N. Mohan (B)
Influx Big Data Solutions, Influx Technology India, #2, Krishvi, Ground Floor, 2nd stage
Indiranagar, BDA Colony, Domlur, Bengaluru, Karnataka 560071, India
e-mail: nithin@influxtechnology.com
URL: http://www.influxbigdata.in
https://doi.org/10.1007/978-981-19-6913-3_28
424 S. Sarkar and N. Mohan
1 Introduction
The automotive industry’s major factors are the growing integration of technologi-
cally advanced solutions in the automotive sector to cater to the ever-increasing need
to optimize vehicle and passenger safety, reduce emissions, and the widespread elec-
trification of vehicles. The growing demand for autonomous vehicles has increased
the need to use advanced electronic control units in modern cars, ensuring that
contemporary automobiles achieve maximum efficiency of fossil fuel-powered vehi-
cles and electric vehicles, particularly in developed regions. The installation of a
small-sized onboard data logger helps to store and transmit vehicle data and vehicle
routes using the Global Positioning System (GPS). The recorded data is used for
diagnostics of a vehicle or the entire fleet and for understanding the driver usage
patterns, which helps OEMs produce better cars. Additionally, the government’s
regulations have become more stringent since the automotive sector was a signifi-
cant contributor toward Global warming; the strict norms have created an immense
demand for automotive testing and analysis.
2 Background Theory
2.1 CAN Bus and OBD II Overview
CAN Bus is a vehicle communication network protocol that facilitates communica-

tion between all the electronic control units within the vehicle. Other communication
protocols such as SPI and I2C communicate point to point, master and slave configu-
ration. Whereas CAN Bus is a multi-master setup, any ECU connected to the network
can take over the bus. The CAN bus broadcasts the message to all the ECUs present in
the network, and only the interested ECU pulls in the required message. CAN allows
the flexibility to install high-tech features within the car. For example, the parking
assists system allows multiple sensors and ECUs, such as the parking sensors, trans-
mission control unit, side door mirror motor, etc., to effectively communicate via the
CAN Bus to park your vehicle safely. More examples include the auto brake wiping
technology used in high-end cars, wiping off moisture from brake rotors during a
rainy day. In this case, the rain sensor communicates via the CAN network with the
breaking ECU to perform this particular function. These are the rudimentary purpose
of CAN, allowing effective and efficient communication between all the ECUs within
the vehicle.
The On-Board Diagnostics II or OBD 2 is the self-diagnostic and reporting capa-
bility of a vehicle. The OBD port gives access to the vehicle’s subsystem data for
analysis and error diagnostics. The Electronic Control Units (ECU) or nodes in a car
are the electronic subsystem units responsible for specific tasks within the vehicle.
For example, to detect if the car is stopped on an incline, the “hill hold” capability
uses input from the vehicle’s tilt sensor (also used by the burglar alarm), and the road
Volkswagen Emission: An Analysis on the VW Vento Using … 425
speed sensors (also used by the ABS, engine control, and traction control) through
the CAN bus. Similarly, the CAN bus receives data from seat belt sensors (part of the
airbag controls) to verify whether seat belts are fastened, etc., to effectively commu-
nicate via the CAN Bus so you can park your vehicle safely. More examples include
the anti-lock braking systems, which work with speed sensors in the vehicle to miti-
gate skidding on the road effectively. The OBD II port can retrieve the transmitted
ECU message via a data logger. The OBD 2 is a critical aspect of the Telematics
and Fleet management system that is used to measure vehicle health and condition.
The fleet management system has multiple advanced, high-tech features, such as
the vehicle can track down wear and tear of vehicle parts. More examples include
measuring driver behavior patterns, idling time, etc.
2.2 Telematics and Data Logging
A telematics system consists of a vehicle data logger that allows for the transmis-
sion and storage of telemetry data. It uses a built-in 4G modem with SIM Card
to communicate over a wireless network. The data logger captures vehicle-specific
data by interfacing with the onboard diagnostics (ODBII) system via CAN-BUS
connector and sends the data to a centralized server along with real-time location
information collected using a GPS module. The server interprets the data and makes
it available to end-users via secure websites and apps designed for smartphones
and tablets. Few recordable parameters include location, speed, idling time, hard
acceleration or braking, fuel consumption, vehicle issues, and other telemetry data
sets. When evaluated for specific events and patterns, this data can provide in-depth
insights across an entire fleet (Fig. 1).
2.3 ECU Parameter IDs
The On-Board Diagnostics II parameter IDs are a set of codes used to request data
from the CAN Bus network of the vehicle by utilizing the Mode $01 (request live
data) of the OBD II service. An automotive ECU communicates via the CAN Bus
network to external test equipment to request data using the parameter IDs. An OBD
II live data parameter usually consists of a unique hexadecimal identifier, the number
of bits returned, start bit, size, the minimum and maximum range of physical data,
and the unit (Fig. 2).
Vehicle Parameters Analyzed
Vehicle Speed Sensor (VSS)—The transmission output or wheel speed is measured
by the Vehicle Speed Sensor (VSS). The ECM uses this data to change vehicle
engine operations like ignition timing, air/fuel ratio, transmission shift points, and
Fig. 1 The telematics and fleet management connection, configuration, and data logging steps
Fig. 2 The number of parameter IDs that were tracked during this test
start diagnostic programs. The output of the Vehicle Speed Sensor is in terms of
Voltage and is communicated directly via the CAN Bus signals.
Vehicle Gear Sensor—The vehicle’s transmission uses a conjunction of gears to
keep the engine’s rotational speed within a safe range. The gear position sensor
detects the gear position, such as first or second, and communicates the information
to the ECU through the vehicle network.
Engine RPM Sensor—The rotational speed or position of the crankshaft is moni-
tored by an Engine RPM Sensor, which is an electrical device used in an internal
combustion engine. The ECM uses this information to manage fuel injection, ignition
timing, and other engine parameters.
Fig. 3 The relationship

between a car’s air/fuel
mixture and its
oxygen/lambda sensor. The
X-axis represents the air/fuel
ratio, where 1 being a
balance between the ratios.
The Y-axis represents the
corresponding equivalent
voltage output of the oxygen
sensor
Oxygen Sensor Voltage—The oxygen sensor is placed at the exhaust manifold and
is used to determine the amount of unburned oxygen present in the exhaust. The fuel
mixture is determined by monitoring the oxygen levels in the exhaust. The output
voltage, which ranges from 0.1 to 1 V, indicates whether the fuel mixture is burning
rich (with less oxygen) or lean (with more oxygen).
2.4 Relationship Between Oxygen Sensor and Air/Fuel

Mixture
The difference in oxygen levels across the sensing element generates a voltage across
the sensor’s platinum electrodes when the air/fuel mixture is rich. There is minimal O2
in the exhaust: typically, 0.8–0.9 V. The sensor’s voltage output reduces to 0.1–0.3 V
when the air/fuel mixture is lean and more oxygen is in the exhaust. The sensor’s
output voltage is roughly 0.45 V when the air/fuel mixture is balanced (Fig. 3).
2.5 Relationship Between the Fuel/Air Equivalence Ratio

and Relative Emissions
It is seen that NOx emission gradually increases as the mixture moves toward the
balanced point. The NOx production peaks when the mixture is balanced and drops
down when the mixture gets leaner or richer. The NOx emission varies from approx.
1900 ppm and peaks to approx. 3100 ppm at the balanced point. Additionally, Carbon
Monoxide (CO) is at its peak (approx. 40,000 ppm) when the mixture is rich and
Fig. 4 The relationship

between a car’s fuel/air
mixture and its relative
emissions. The X-axis
represents the fuel/air
equivalence ratio, where 1 is
a balance point. The Y-axis
represents the typical HC and
NOx concentration in PPM
gradually decreases near approx. 3000 ppm as it gets closer to the balanced point
(Fig. 4).
Further, we see that the CO production increases as the mixture gets leaner, but
the rate of change in increment is minor. Furthermore, we see that the Hydrocarbon
production is at its peak (approx. 28,000 ppm), initially, when the air/fuel mixture is
rich and drops down as it lets leaner. We also see that the production of HC gradually
increases before the mixture balanced point.
Additionally, at a 0.9 air to fuel ratio value, all the three emissions produced are
relatively at their minimum. Hence at 0.9 V, the vehicle is at its efficient point, and
a 0.9 V value should be maintained to emit the least emission relatively.
3 Results and Analysis
3.1 Test Route
The test was conducted for 3 h 30 min with a total distance of 193 km. The vehicle was
subjected to both city and highway conditions for maximum clarity in results. The
data acquisition system was configured to log OBD II data and GPS and interfaced
with the vehicle CAN network. The data logger acquired 140,000 data point samples
from each ECU sensor under observation. Electronic Control Unit (ECU) parameters
such as lambda sensor (O2 sensor) voltage, vehicle speed, RPM, vehicle power, gear,
etc., were observed throughout the test (Fig. 5).
Fig. 5 The testing location and the path taken by the car
3.2 Oxygen Sensor and Vehicle Gear Relationship
The X-axis represents time in seconds, and Y-axis shows the gear change. The color
indicates the variation in oxygen sensor voltage with the gear change. The orange
section of the graph represents low oxygen sensor voltage: < 0.45 V (lean air/fuel
mixture). The blue area shows the oxygen sensor voltage greater than 0.45 V (rich
air/fuel mixture), where 0.45 is the balanced point with the air/fuel mixture being 1
(Fig. 6).
It is seen that the average oxygen sensor voltage is low (the color orange, < 0.45 V)
throughout the journey. The low sensor voltage indicates the air/fuel mixture is lean,
Fig. 6 The change in gear with time throughout the journey

and more oxygen is in the exhaust. During this phase, the vehicle NOx emission drops,
and there is a gradual decrease in carbon monoxide and hydrocarbon emissions from
their peak values. The blue phase (> 0.45 V) indicates the air/fuel mixture is rich, and
less oxygen is available in the exhaust. During this phase, the vehicle NOx emission
drops drastically. CO and HC emissions are high initially, as shown in Fig. 4, and
with the increase in mixture richness, the CO and HC emission decreases too. The
graph’s colorless sections indicate a balanced air/fuel mixture with oxygen sensor
voltage being around 0.45 V.
It is seen that the oxygen sensor voltage tends to increase when there is a sudden
rise in gear when the vehicle starts accelerating, indicating a rich air/fuel mixture (as
the color of the graph turns blue) and reduction in NOx emissions after exceeding
the balanced point (according to the Fig. 4). During this phase, the HC emissions
increase up to 900 ppm and CO production goes near 100 ppm. It is also seen that
during vehicle idling conditions, the oxygen sensor voltage increases and produces
more HC (Fig. 4). Additionally, on average, the third gear of the car, where the
vehicle speed is around 40 km/hr, shows higher HC production (as we get to see
more color blue throughout the graph). In contrast, the vehicle keeps a lower air/fuel
mixture throughout the journey, excluding gear three, as the color orange dominates
the entire graph. Majority of the time, the vehicle produces more elevated CO and
HC emissions, and low NOx emissions. The car rarely achieves an oxygen sensor
voltage around 0.45 V, where the air/fuel mixture is balanced and produces high NOx
emissions (Fig. 7).
It is seen that the oxygen sensor voltage has majorly concentrated in two places—
between 0.1 to 0.2 V and 0.7 to 0.8 V. The 0.1–0.2 V range indicates the vehicle had
a lean air/fuel mixture, and the 0.7–0.8 V range indicates rich air to fuel mixture. The
Fig. 7 The causal relationship between the oxygen sensor voltage and the change in gear. The
X-axis represents the change in oxygen sensor voltage, which varies from 0 to 1 V. The Y-axis
shows the gear change from 0th gear to 7th gear
car achieved a balanced point of 0.45 V rarely. The data shows that the vehicle had
low NOx emissions, but the CO and HC emissions were high throughout the journey.
Furthermore, the data shows higher data point concentration around the 0.1–0.2 V
range, indicating low NOx emissions (1000 ppm) and moderate HC emissions (700
ppm). Gear 0–1 shows lower data point concentrations of a rich mixture, indicating
low emissions of CO.
3.3 Vehicle Speed and Oxygen Sensor Relationship
The X-axis shows the change in time (s), and the Y-axis shows the change in vehicle
speed (km/h). The color of the graph indicates the change in oxygen sensor voltage
with vehicle speed. Orange sections of the graph show low oxygen sensor voltage
(< 0.45 V), where 0.45 V being the median, and the blue section shows high oxygen
sensor voltage (> 0.45 V) (Fig. 8).
and more oxygen is in the exhaust. During this phase, the vehicle NOx emission drops,
and there is a gradual decrease in carbon monoxide and hydrocarbon emissions from
their peak values. The blue phase (> 0.45 V) indicates the air/fuel mixture is rich, and
less oxygen is available in the exhaust. During this phase, the vehicle NOx emission
drops drastically. CO and HC emissions are high initially, as shown in Fig. 4, and
with the increase in mixture richness, the CO and HC emission decreases too. The
graph’s colorless sections indicate a balanced air/fuel mixture with oxygen sensor
voltage being around 0.45 V (Fig. 9).
Fig. 8 The relationship between the change in vehicle speed and oxygen sensor voltage with respect
to time
Fig. 9 The causal relationship between the change in vehicle speed and oxygen sensor voltage
with respect to time. The X-axis represents the change is vehicle speed and the Y-axis shows the
change in oxygen sensor voltage from 0 to 1 V
Additionally, it is seen that, with a sudden rise in acceleration/vehicle speed, the

air/fuel mixture gets rich (as the color of the graph shifts to blue), indicating higher
HC production (Fig. 4). According to the data, we see that above 100 km/h, the
probability of HC emission is high, and at a range of 20 km/h–60 km/h, the vehicle
purely emits CO. The vehicle rarely reaches its balance point, where NOx production
peaks (Fig. 4).
The oxygen sensor voltage appears to have concentrated in two areas: between
0.1 and 0.2 V and 0.7 and 0.8 V. The range of 0.1–0.2 V suggests a lean air/fuel
combination, whereas the range of 0.7–0.8 V indicates a rich air/fuel mixture. Rarely
did the vehicle reach a balanced point of 0.45 V. The car had low NOx emissions, but
significant CO and HC emissions throughout the drive, according to the statistics.
Additionally, the data indicates greater data point concentrations in the 0.1–0.2 V
range, indicating low NOx (1000 ppm) and moderate HC emissions (700 ppm).
From 0 to 30 km/h, vehicle speed results in minimal CO emissions but a greater
concentration of HC.
3.4 Engine RPM and Oxygen Sensor Relationship
The X-axis represents time in seconds, and the Y-axis shows the engine RPM vari-
ation. The color indicates the variation in oxygen sensor voltage with the change in
engine RPM. The orange section of the graph represents low oxygen sensor voltage:
< 0.45 V (lean air/fuel mixture). The blue area shows the oxygen sensor voltage
Fig. 10 The change in RPM with time throughout the journey
greater than 0.45 V (rich air/fuel mixture), where 0.45 is the balanced point with the
air/fuel mixture being 1 (Fig. 10).
and more oxygen is in the exhaust (as the color of the graph is orange). During this
phase, the vehicle NOx emission drops (Fig. 4), and there is a gradual decrease in
carbon monoxide and hydrocarbon emissions from their peak values. The blue phase
(> 0.45 V) indicates the air/fuel mixture is rich, and less oxygen is available in the
exhaust. During this phase, the vehicle NOx emission drops drastically. CO and HC
emissions are high initially, as shown in Fig. 4, and with the increase in mixture
richness, the CO and HC emission decreases too. The colorless sections of the graph
indicate a balanced air/fuel mixture with oxygen sensor voltage being around 0.45 V.
Additionally, it is seen that above 2500 RPM (vehicle speed 40 km/h), the vehicle
tends to produce more HC from the exhaust, as the graph gets bluer above this data
point. The sudden increase in acceleration tends to make higher HC at the exhaust as
well. On average, RPM below 2500 tends to produce more CO and HC and reduces
NOx emissions. The oxygen sensor voltage being 0.45 V, where the air/fuel mixture
is balanced, is rarely seen throughout the journey.
4 Conclusion
4.1 Vehicle Idling
The data analysis shows us that during the vehicle’s idling, the system provides a
high oxygen sensor voltage, indicating a rich air/fuel mixture at the exhaust. During
this phase, the car produces Hydrocarbons, whereas the NOx and Carbon monoxide
emissions reduce. The hydrocarbon production increases to approx. 900 ppm in this
phase, as shown in Fig. 4. In contrast, the carbon monoxide emission tends to drop
toward 100 ppm and NOx toward 1000 ppm.
To reduce any emissions during the idling period, the driver should switch off his
engine for efficiency. Switching off the engine during an idling state not only reduces
emissions but also saves fuel and money.
4.2 Acceleration
The vehicle’s oxygen sensor voltage shoots above 0.45 V during a sudden rise in
acceleration, indicating a rich air/fuel mixture at the exhaust. During this phase,
the car produces Hydrocarbons, whereas the NOx and Carbon monoxide emissions
reduce. The hydrocarbon production increases to 900 ppm in this phase, as shown in
Fig. 4. In contrast, the carbon monoxide emission tends to drop toward 100 ppm and
NOx toward 1000 ppm. On average, the driving pattern has not surged the acceleration
multiple times and has maintained an oxygen sensor voltage below 0.45 V, indicating
a lean air/fuel mixture at the exhaust. The Carbon monoxide and HC production peaks
during this phase, reaching up to 40,000 ppm and 1400 ppm, respectively. The NOx
production is low, around 1900 ppm. The vehicle rarely achieves a stable air/fuel
ratio throughout the journey.
Hence, the driver should avoid sudden acceleration of the vehicle, as suddenly
accelerating the vehicle causes higher fuel consumption and increases HC emissions
at the exhaust.
4.3 Vehicle Speed
It is seen that vehicle speed above 42 km/h (2500RPM) produces a rich air/fuel
mixture at the exhaust. During this phase, the car produces Hydrocarbons, whereas
the NOx and Carbon monoxide emissions reduce. The hydrocarbon production
increases to 900 ppm in this phase, as shown in Fig. 4. In contrast, the carbon
monoxide emission tends to drop toward 100 ppm and NOx toward 1000 ppm. We
see that RPM speed above 2500 directly relates to the air/fuel mixture as per the data
shown above. At speeds less than 40 km/h, the vehicle produces low oxygen sensor
voltage, indicating a low air/fuel mixture at the exhaust. The Carbon monoxide and
HC production peaks during this phase, reaching up to 40,000 ppm and 1400 ppm,
respectively. The NOx production is low, around 1900 ppm. The vehicle rarely
achieves a stable air/fuel ratio throughout the journey.
Hence, the driver should maintain speeds below 42 km/h at gear values between
0–2 for better fuel efficiency, as speeds above 42 km tend to cause a rich air to fuel
mixture.
4.4 Emissions
The vehicle fluctuates from a lean air–fuel mixture and a rich air–fuel mixture
throughout the journey, indicating high Carbon Monoxide and Hydrocarbon emis-
sions. The vehicle rarely maintains a balanced point of a lambda value around 1, indi-
cating low NOx emissions (fluctuating around 3000 ppm) frequency at the exhaust.
The use of transition metal oxides as described in [6] or customized sensors such as
[7] alongwith accelerometers as described in [8] could facilitate continuous moni-
toring of oxides of nitrogen and sulfur and correlate with acceleration enabling better
understanding and possibility of regulation.
Finally, it can be concluded that Volkswagen Vento, TSI Highline Plus, petrol
engine had low NOx emissions throughout the journey. The vehicle emits higher
Hydrocarbon emissions than Carbon Monoxide. Furthermore, the efficiency of the
vehicle depends on the catalytic converter that determines the concentration of the
emission particles released into the atmosphere.
References
1. “ECU” is a Three letter answer for all the innovative features in your car: know how the story
unfolded. Embitel (2017)
2. Etschberger K, Controller area network, basics protocols, chips and applications. ISBN 3-00-
007376-0 (www.ixxat.com)
3. Lawrenz W, CAN systems engineering, from theory to practical applications. ISBN 0-387-
94939-9
4. SAE International (2018) J1939: serial control and communications heavy duty vehicle network
5. Understanding CAN DBC, telematics & fleet management. Influx Big Data Solutions
6. Nag P, Sadani K (2015) Factors governing gas sensing characteristics of some transition metal
oxides. Int J Control Theory Appl 8(3):1225–1233
7. Sadani K, Nag P (2016) Encapsulated piezoresistor cantilevers as affinity sensors: a review. Int
J Control Theory Appl 9(39):205–215
8. Sadani K, Prabhakar DA, Nag P (2018) A cardio pulmonary resuscitation device for stretchers.
Int J Eng Technol 7(2.21):62–5
Hand Gesture-Controlled Wheeled
Mobile Robot for Prospective Application
as Smart Wheelchairs
Leon Muli Suryavanshi, Ananth Jnana Chandraraj, Kshetrimayum Lochan,

and Pooja Nag
Abstract The wheelchair is used to assist in movement and heavily relied on by

people who have significant difficulty in walking. Manual wheelchairs can be used by
people with good upper limb strength and dexterity, however people suffering from
any kind of locomotive disabilities in the arms would require assistance. This article
reports the proof of concept of a smart electric wheelchair that can be controlled
by utilizing hand gestures. A prototype with wheeled mobile robot was devel-
oped and involved the use of data fusion from sensors mounted on human hand
to create a scheme for motor control of the wheels. Hand gesture dependent data
from accelerometers and gyroscope were inputs to control the wheelchair in the
required direction. The prototype of the wheelchair, PID based controller and a hand
mounted sensor unit (HMSU) was able to function reliably and efficiently on urban
terrain.
Keywords Electric wheelchair (ECW) · Gesture-controlled/control · IMU (inertial

measurement unit) sensor · PID · Filter · Hand mounted sensor unit (HMSU)
1 Introduction
A large number of people around the world suffer from locomotive disabilities [1].
This may be permanent in some cases or a temporary injury. Hence, wheelchairs
have become extremely important for such individuals to move around. Patients who
cannot operate manual wheelchairs effectively would always require additional help.
This limits the freedom of the patients, as they have to constantly ask for assistance
to move around. Conventional electric wheelchairs (ECWs), are operated with a
joystick input. This is suitable for the majority of people that need ECWs, however
Leon Muli Suryavanshi, Ananth Jnana Chandraraj—Authors have contributed equally to the work
L. M. Suryavanshi · A. J. Chandraraj · K. Lochan · P. Nag (B)

Department of Mechatronics Engineering, Manipal Institute of Technology, Manipal,
Karnataka 576104, India
e-mail: pooja.nag@manipal.edu
https://doi.org/10.1007/978-981-19-6913-3_29
438 L. M. Suryavanshi et al.
a small percentage may have additional disabilities that make it difficult to operate
a joystick. One such disability is Symbrachydactyly [2], in which an individual has
missing/underdeveloped/webbed fingers and/or toes.
The purpose of this work was therefore aimed to assist such individuals that have
finger-related disabilities in addition to locomotive disabilities, by providing an alter-
native control scheme to a joystick, at a minimum cost. This could be used to easily
and intuitively maneuver the ECW. Multiple iterations of alternative wheelchair
control system have been conceptualized and realized with a portion of them relying
on IMU sensors in different capacities. This project demonstrates the prospects of
utilizing IMU sensors as the main element to not only derive control signals, but also
provide stabilizing data based on the wheelchair’s angular orientation, to be used in
a PID control algorithm. Furthermore, the use of two IMU sensors enables proper
functioning of the control system on sloped surfaces.
1.1 Related Work
Over the years, several types of smart electric wheelchairs have been reported, which
utilize different strategies to sense the direction command from the patient, and
accordingly maneuver the wheelchair. This includes, mounting of sensors on the
patient hand or head to sense certain gestures or movement [3, 4], muscle activity
monitoring through electromyogram (EMG) [5], tracking of eye movement [6], facial
expressions [7], a combination of few of the above [8] or voice recognition [9].
Khan [3], proposed a flex sensor-based control of wheelchair. The flex sensors were
attached to a glove and finger movement controlled the motors. Four fingers were
used to derive five different outputs corresponding to forward, left, right, reverse and
stop. However, flex sensors do not produce consistent change in resistance over long
periods of usage and easily get damaged. Machangpa and Chingthamb [4], designed
a wheelchair that could be controlled by head movement. Gyroscope, accelerometer
and ultrasonic sensors were mounted on a hat and used to sense the head movement.
The data from the sensors were passed through a novel algorithm for processing.
This enabled each user to have a custom threshold for their head movements. The
ultrasonic sensor was used to prevent accidental collisions. The roll and pitch angles
were taken into consideration in processing when determining the direction to move.
Rakasena and Herdiman [5], designed a wheelchair controlled by muscle activity in
the arm and utilized surface electromyography sensors (SEMGs) to provide signals
based on the user’s muscle activity. This design was mainly aimed at people that have
some sort of muscle dystrophy disabilities. An electrode was placed on the wrist
flexor muscle and wrist extensor muscle. Direction of movement was controlled by
performing different kinds of grip movement i.e. spherical, cylindrical, lateral, hook
and tip grasp. These gave different output potential readings which could be mapped
to different motions of the wheelchair. However, a limitation of such a method is
that EMG sensors require the user’s skin to be cleaned of any oils produced by
the body in order to get reliable readings from the sensor. Rajesh and Mantur [6],
Hand Gesture-Controlled Wheeled Mobile Robot for Prospective … 439
designed an eye tracking based wheelchair control system with an attached camera
mounted on the user’s head. The camera kept track of the position of the pupil and the
wheelchair moved in the direction the user looked. In such a system, the field of vision
is hindered and the user had to wait two seconds for any change in direction or stop.
Rabhi et al. [7], proposed a wheel chair controlled using facial expressions. Yi et al.
[8], proposed a combination or a hybrid of EMG signals and head gestures which
worked as the input to the wheelchair. Rabhi et al. [10], designed an Electric Wheel
Chair controlled by a touchscreen device such as a tablet. Although non-contact
approaches such as these [6–8, 10–12] are frequently used for sensory feedback,
they run a risk of faulty alarms or actuation as well. The advantage of the system was
that it required less effort when compared to a traditional joystick powered wheel
chair and also gave a much quicker speed response. Choudhari et al. [13], proposed
a unique method to control wheelchairs—an electrooculography-based system to
control the motion of the wheel chair (forward, left, right turn and stop) through
number of eye blinks. While some methods are promising, most of them suffer
from issues of user discomfort, reliability, incapability of traversing over slopes or
a complicated approach which thereby increases cost. This work reports of a simple
approach of using accelerometer and gyroscope sensor data fusion and an efficient
control scheme to develop a robust prototype, which can be easily scaled up to build
an efficient and affordable smart electric wheelchair.
2 Methodology
2.1 The Hand Mounted Sensor Unit (HMSU)
The proposed system comprised of two 6-axis inertial measurement unit (IMU)
MPU-6050 sensors [2], consisting of an accelerometer and gyroscope. These sensors
were connected to Arduino Nanos [14] for processing. The first IMU was mounted
on the user’s hand and detected the orientation of their hand (Fig. 1b), based on which
the motor was actuated via a PWM signal. This acted as the hand mounted sensor
unit (HMSU). A minimum threshold was set for the angles required to cause the
motor to be actuated (Table 1). The HMSU communicated to the actuating Arduino
wirelessly via nrfl01 + modules (Fig. 1a).
2.2 Motor Actuator
A second Arduino was mounted on the wheelchair and connected to two motors
via a motor driver (Fig. 2). Optical speed encoders that were attached to the motors
and the second IMU sensor were both used to send feedback signals. This feedback
was passed through a PID system which was used to maintain a constant speed of
Fig. 1 a Hand mounted sensor unit circuit connection; b type of gestures used to control wheelchair
(with the tip of fingers being the nose of the roll axis and base of the wrist being tail end of the roll
axis)
Table 1 Mapping for motor

Direction of motor actuation Tilt angle of user’s hand
actuation and control
Roll Pitch
Forward ≤ 30° ≥ 30°
≥ −30° ≥ 30°
Left ≤ −30° ≤ 30°
≤ −30° ≥ -30°
Right ≥ 30° ≤ 30°
≥ 30° ≥ −30°
Back ≤ 30° ≤ −30°
≥ −30° ≤ −30°
the wheelchair in order to enable it to traverse surfaces with varying gradients and
maintain a straight path when required, since a differential drive system was used.
The IMU sensor connected to the actuating Arduino had two purposes. One was to
detect changes in yaw angle. A rotation toward the right was taken as positive change,
while a rotation toward the left was taken as negative change. If the wheelchair was
headed on a straight path, then the yaw angle would be approximately zero. This
information was used to determine the orientation of the wheelchair and fed into the
control loop to ensure stable orientation when in motion.
The second purpose was to determine the roll and pitch angle of the wheelchair.
This is because the orientation of the HMSU’s IMU sensor is affected by two entities,
Fig. 2 Actuating Arduino circuit connection
one being the user’s hand and the other being the wheelchair itself. If we take the
example of a scenario where the wheelchair is traversing an inclined surface, say
a 15° slope, then the hand mounted IMU sensor would not be able to distinguish
between the cause of the change in its orientation and its reading would therefore be
erroneous (it would have an offset of 15°).
Hence, in order to correct any offset created by the incline of the terrain being
traversed, the orientation of the wheelchair in terms of pitch and roll was always calcu-
lated by the actuating Arduino and sent periodically to the hand-mounted Arduino
wirelessly. The change in yaw angle as well as the speed detected by the encoders
is fed to the PID algorithm in a feedback loop. The feedback loop consisted of two
PID algorithms, with one responding to the speed and the other responding to the
change in yaw angle.
2.3 PID Control Loop
Assuming a situation where the wheelchair is meant to move forward or back, in order
to ensure that it moves in a straight path, the wheels must rotate at the same RPM.
However, observing the speed is not enough, since each wheel might experience
different levels of sudden disturbances, caused by either varying surface friction or
obstacles.
Hence the requirement for detecting the yaw angle.
The PID control loop (Fig. 3) was programmed, and the PID gains were found
mainly by trial and error. The method was implemented with the assumption that
the motor driver would output consistent current that was proportional to the PWM
signal being written to it.
Fig. 3 PID control loop for actuating motors. Speed PID Sample Time (Ts1) = 0.25 s, Yaw PID
Sample Time (Ts2) = Program execution time of Arduino microcontroller
The control scheme was applied in the prototype (Figs. 4 and 5) and the system
response was tested.
Fig. 4 Prototype setup top view

Fig. 5 Prototype setup bottom view
2.4 Calculating Angles Form IMU Sensor
The MPU-6050 contains a gyroscope and accelerometer which individually produce

their own readings, a sensor fusion algorithm was used to get the highest accuracy
and speed response possible with the hardware. The first step involved calculation
of the roll and pitch angles from each sensor in the module. The sensitivity and
subsequently the range of the accelerometer and gyroscope sensors are selected by
writing to the control register. The highest sensitivity was used in this case, with the
accelerometer having a max sensitivity of 16,384 LSB/g (least significant bits per g)
and range of ± 2 g. While the gyroscope had a max sensitivity of 131 LSB/°/s (least
significant bits per degree per second) and a range of ± 250°/s.
2.4.1 Euler Angle Calculations from Accelerometer Data
The accelerometer readings were used to calculate Euler angles [15], using the
concept of rotation matrices. However, since the accelerometer cannot detect yaw
angles, any orders of the rotation matrix which contain the yaw angle (RXZY , RYZX ,
RZXY , RZYX ) have to be left out. This leaves only two orders out of the original six
orders. These two orders are RXYZ and RYXZ .
The accelerometer outputs the gravity vector in terms of g-force, therefore by
working backward, it is possible to find the rotation done about a given base vector.
This base vector was taken to be the gravity vector when the sensor was at rest and
upright. Therefore, the output given can be assumed to be:
⎛ ⎞ ⎛ ⎞
GV X 0
G V = ⎝ G V Y ⎠ = R⎝ 0 ⎠ (1)
GV Z 1
where
GV The gravitational vector

GVX The X axis g-force
GVY The Y axis g-force
GVZ The Z axis g-force
R Rotation matrix (RXYZ and RYXZ ).
The pitch ‘θ ’ can hence be deduced:
⎛ ⎞ ⎛ ⎞
GV X − sin(θ )
G V = ⎝ G V Y ⎠ = ⎝ sin(ϕ) cos(θ ) ⎠ (2)
GV Z cos(ϕ) cos(θ )
⎛ ⎞
−G
(θ ) X Y Z = tan−1 ⎝ / ⎠
VX
(3)
G V Y + G 2V Z
2
Similarly, the roll ‘ϕ’ can be found:

⎛ ⎞ ⎛ ⎞
GV X − sin(θ ) cos(ϕ)
GV = ⎝ GVY ⎠ = ⎝ sin(ϕ) ⎠ (4)
GV Z cos(ϕ) cos(θ )
⎛ ⎞
−G
(ϕ)Y X Z = tan−1 ⎝ / ⎠
V Y
(5)
GV X + GV Z
2 2
2.4.2 Gyroscope Data and Sensor Fusion
The gyroscope produces the angular velocity experienced via the Coriolis effect
about the three axes. Therefore, the roll, pitch and yaw can be found by discrete
Fig. 6 Complementary filter flow chart. The complementary filter (Fig. 6) is capable of producing
roll and pitch values with a suitable amount of accuracy and speed
integration of these values:
Σ
n
(ωi × Δt) (6)
i=1
The gyroscope is less prone to noise but suffers from gyroscope drift, which
causes the angles to drift overtime from the true value, hence the use of a sensor
fusion algorithm.
The fusion algorithm involves passing the calculated gyroscope angles through
a discrete high pass filter, while the Euler angles deduced form accelerometer are
passed through a discrete low pass filter. This is because the gyroscope is not affected
by any linear movement/jerks and only responds rotational motion and hence has
barely any high frequency noise. The accelerometer on the other hand has a lot of
high frequency noise since it responds to any kind of motion. The filtered values are
then added to give the final roll and pitch. The resulting equation is a complementary
filter as shown below in the case of pitch ‘θ ’:
θ out[n] = β(θ out[n − 1] + ω[n] × Δt) + (1 − β)(θ accel[n]) (7)
Figure 7a shows the change in pitch angle with time as the IMU sensor was rotated
arbitrarily. The blue line depicts the output given by the complementary filter. As
can be seen, the blue line hugs red line but does ignore most of the noise. However, it
should be noted that that the complementary filter can only produce reliable readings
in the range of ± 90°. This is because of gimbal lock which occurs when calcu-
lating Euler angles with a limited number of rotation orders. This means that the
accelerometer sensor can comprehend angles on the range of ± 90°, beyond which
the values wrap around back to 0°.
Figure 7b shows the PWM signal being written by the PID controller based on
the yaw angle given as input to the PID, with a setpoint of zero degrees. In the given
instance shown here, the model was manually rotated toward the right by 40°. This
caused the PID controller to slow down the left wheel by decreasing the PWM value
(shown by Green line) and keep the right wheel moving faster (Red line) so as to
turn the model back toward the left till the setpoint of zero degrees was reached. As
the setpoint was being approached (shown by Blue line), the left wheel started to
increase speed. The system exhibited slight overshoot in PWM signal being written
to the motors, but the was able to maintain the system close to the setpoint. Once
very near/at the setpoint, the left motor resumes to maintain the same speed as right
motor. The data shown above (Fig. 7b) was taken from the prototype wheelchair in
motion.
Fig. 7 a Filtered pitch angle vs unfiltered pitch reading against time (red graph: raw accelerometer
reading, blue graph: filtered sensor fusion reading); b variation in PWM output and wheelchair yaw
angle with time (red: PWM signal written to right motor, green: PWM written to left motor, blue:
yaw angle of wheelchair)
4 Conclusion
This paper puts forward the concept of using IMU sensors to create a control scheme
that uses hand gestures. The MPU-6050 is to be mounted on the user’s hand and
produces input signals, an Arduino Nano generates control signals based on the
input signal using sensor fusion and Euler angles, which are sent wirelessly to the
wheelchair’s actuation system. A PID control loop actuates the wheelchair motors,
considering the speed of the wheelchair (using encoders), the orientation of the
wheelchair (yaw angle). The PID ensures that the speed of the wheelchair remains
relatively constant and that over all operational stability is maintained. The MPU-
6050 attached to the wheelchair corrects offsets that arise in the HMSU due to inclined
surfaces.
The prototype wheelchair operates as expected and responds to the hand mounted
sensor unit with minimum delay and minor errors. The current system was developed
at minimum cost which would make it more easily available to the general public
and sets a good foundation for the potential improvement, via implementation of
additional sensors, better hardware and further consideration of ergonomics.
References
1. WHO Webpage. https://www.who.int/publications/i/item/9789241547482

2. Woodside JC, Light TR (2016) Symbrachydactyly—diagnosis, function, and treatment. J Hand
Surg 41(1):135–143. https://doi.org/10.1016/j.jhsa.2015.06.114
3. Khan HA (2018) The economical design of a hand-gesture and Bluetooth controlled wheel-
chair by integrating indigenous components: mobility aid for the disabled. IOP conference
series: materials science and engineering, vol 473, Foundation University Islamabad, South
Korea
4. Machangpa JW, Chingthamb TS (2018) Head gesture controlled wheelchair for quadriplegic
patients. In: Procedia computer science, international conference on computational intelligence
and data science, vol 132, pp 342–351
5. Rakasena EPG, Herdiman L (2019) Electric wheelchair with forward-reverse control using
electromyography (EMG) control of arm muscle. In: Journal of physics: conference series,
international conference on applied science and technology (iCAST on Engineering Science)
24–25 Oct 2019, vol 1450, Bali, Indonesia
6. Rajesh A, Mantur M (2017) Eyeball gesture controlled automatic wheelchair using deep
learning. In: 2017 IEEE region 10 humanitarian technology conference (R10-HTC), Dhaka,
pp 387–391. https://doi.org/10.1109/R10-HTC.2017.8288981
7. Rabhi Y, Mrabet M, Fnaiech F (2018) A facial expression controlled wheelchair for people
with disabilities. Comput Methods Programs Biomed 165:89–105
8. Yi Z, Xiaolin F, Yuan L (2015) Intelligent wheelchair system based on sEMG and head gesture.
J China Univ Posts Telecommun 22(2):74–80, 95
9. Joseph C, Aswin S, Sanjeev Prasad J (2019) Voice and gesture controlled wheelchair.
In: Proceedings of the third international conference on computing methodologies and
communication (ICCMC 2019)
10. Rabhi Y, Mrabet M, Fnaiech F, Gorce P, Miri I, Dziri C (2018) Intelligent touchscreen joystick
for controlling electric wheelchair. IRBM 39(3):180–193
11. Nag P, Mathew A, Sadani K (2017) A non-contact flow monitoring system for rotameters.
ICAR
12. Nag P, Shinde S, Sadani K (2018) Developing a smart navigator for surveillance in unmanned
zones. Ind Interact Innov Sci Eng Technol, 255–263
13. Choudhari AM, Porwal P, Jonnalagedda V, Mériaudeau F (2019) An electrooculography based
human machine interface for wheelchair control. Bio Cybern Biomed Eng 39(3):673–685
14. Arduino webpage, Arduino Nano Every. https://store-usa.arduino.cc/products/arduino-nano-
every-with-headers?queryID=undefined
15. Ramessur SA, Oree V (2019) Hand gesture controller for robotic-wheelchair using microelec-
tromechanical sensor ADXL 345. Smart Sustain Eng Next Gener Appl 561:3–10
Manual Dexterity Assessment Using
a Nine-Hole Pegboard Test
K. Aneesha Acharya and Amartya Choudhary
Abstract Manual dexterity is the ability to use the hands in a coordinated way
to grasp and manipulate objects and to demonstrate slight precise movement. This
research aims to develop an outcome measurement unit (pegboard) that tells about
timing information of the hand activity and the accountability of the pegs movement.
This paper seeks to create an automated pegboard with the help of photo-interrupter
sensors, Infrared sensors, microcontroller, and adequate circuitry to reduce human
error in time calculation. This pegboard test consists of two methods to assess hand–
eye coordination: (1) Insertion of the peg and (2) Removal of the pegs. ATmega 2560
microcontroller helps in interfacing the hardware model with the software. Photo-
interrupter helps in sensing removal and insertion of the pegs inside the peg holes. An
inbuilt timer calculates the activity period in milliseconds. Test sessions showed that
the average time taken to complete the test rises as the age increases, indicating the
loss of hand–eye coordination and increment in underlying neural-related conditions
as the age increases.
Keywords Dexterity · Pegboard · Photo-interrupter sensor · Infrared sensor ·

Microcontroller
1 Introduction
Manual dexterity is the ability to make coordinated hand and finger movements to
grasp and manipulate objects [1]. It includes a combination of muscular-skeletal and
neurological functions to execute small precise actions. A well-developed manual
dexterity over time means that a person can adequately assess, plan, and execute a
task within a normal time range. The test used in this paper is the modified Nine-Hole
Peg Test (NHPT), which is also considered a gold standard in manual dexterity tests
[2]. NHPT is used widely in stroke, multiple sclerosis research. A manual pegboard
K. Aneesha Acharya (B) · A. Choudhary

Manipal Academy of Higher Education, Manipal 576104, India
e-mail: ak.acharya@manipal.edu
https://doi.org/10.1007/978-981-19-6913-3_30
450 K. Aneesha Acharya and A. Choudhary
Fig. 1 Traditional Nine-Hole Peg Test (NHPT) set-up
test kit consists of a pegboard with a defined number of holes, pegs, and a stopwatch,
as shown in Fig. 1.
Rivera et al. [3] describe how the automatic pegboard replaces the traditional
pegboard device using sensors and embedded electronics. Using a customized elec-
tronic pegboard, the authors studied the different intraindividual variability in fine
motor skills of varying age groups. Children of various age groups performed the
test, and the result emphasizes that the children with lower dysfunction show less
intraindividual variability. Jobbágy et al. [4] introduced a smart 9-hole peg tester
model, which comprises a standard 9-hole pegboard but with Light Emitted Diodes
(LEDs) next to each hole. In this model, one can freely choose the order of the peg
placement and removal. This smart 9-hole peg tester (s-9-HPT) analyzes a healthy
person’s finger dexterity and a patient suffering from a stroke. Wang et al. [5] focus
on manual dexterity as a critical aspect of motor function across the different age
span. The test was conducted on 4319 subjects using a manual 9-hole pegboard. The
dominant hand completed the test more quickly as compared to the non-dominant
hand. The presence of impairments in dexterity across different age span is possible
to study using standards of the 9-HPT.
Research related to fine motor assessment in multiple sclerosis (MS) and
Parkinson’s disease population also use the pegboard test [6, 7]. With the upcoming
technologies and advancements in the automation, the shortcomings of the manual
pegboard test can be rectified [8]. The automation of the pegboard test certainly will
be cost-friendly and lead to a more accurate result [9]. This work aims to develop
an outcome measurement unit (pegboard) that tells about timing information of the
hand activity and the accountability of the pegs movement.
This paper is structured as follows: Sect. 2 contains methodology, which explains
experimental procedure and algorithm developed in pegboard test set-up develop-
ment. Section 3, results are obtained with our method, followed by a discussion in
Sect. 4, and the concluding remarks in Sect. 5.
Manual Dexterity Assessment Using a Nine-Hole Pegboard Test 451
2 Methodology
The conventional pegboard test had a few shortcomings, such as the clinician need
to operate a stopwatch during the trial, keeping an eye on peg movement, and contin-
uous observation required. Stopwatch handling is a tedious and time-consuming job.
The main idea behind automated pegboard development is to eliminate all human
errors which occur during the test process and make a more efficient assessment
tool. The equipment used here are the photo-interrupters (MOC 7811), LED lights,
microcontroller (Arduino Mega 2560), resistors (10 kΩ and 100 Ω), breadboards,
single strand wires, jumper wire, and a laptop. Figure 2 shows the block diagram
representation of the pegboard test set-up.
A photo-interrupter detects the insertion and removal of pegs. Photo-interrupter
makes the system more efficient and less error-prone. It detects the presence of any
object once it has been inserted right between its opening; hence will help to detect
the pegs inserted. The next part consists of selecting the LEDs and resistor as there
is always a chance to blow the LEDs up if the resistor is not connected. LEDs guide
the patient about the peg placement. It is a 9-hole peg test, and hence we require
nine sensors and nine LEDs attached to each sensor. The next component in this
project is the infrared (IR) sensor which counts the number of hand movements.
Hand movement count ensures that one peg is lifted in each movement and the
participant obeys all the rules related to peg displacement. LCDs or the screen itself
used to display the result message, and lastly, to transfer the data to the portable
devices to track the progress, an HC-05 Bluetooth module is required.
Fig. 2 Block diagram representation of the pegboard test set-up

2.1 Algorithm of Automated Pegboard Test
The flowchart shown in Fig. 3 gives a basic idea of how the logic works in an
automated pegboard test set-up. IR sensor connected to digital pin 6 of the Arduino
Mega 2560 board. Nine photo-interrupter sensors have been initialized with a specific
Personal Identification Number (PIN). Only one peg is supposed to be picked up
during the peg test, and the IR sensor is positioned properly near the peg holder.
When the participant’s hand approaches the peg holder to collect one peg, the IR
sensor detects this movement. IR sensor triggers the timer, and once the peg has
been inserted in the pegboard, it increments the counter. IR sensor keeps track of
hand movement during the test. After hand holding the peg from the peg holder,
the next step is to detect correct peg placement. The output of the photo-interrupter
circuit is observed for the proper placement of peg placement. If the output given by
this photo-interrupter sensor is high, the second LED lights up, indicating the hole
in which the peg is supposed to be kept.
In between this process, if the subject cannot keep a single peg into the indicated
sensor within 250 s, a “time out” message will be displayed on the screen. This is
particularly required when the clinician examines the patient population and monitors
the progress in multiple sessions. In healthy adults, pegboard tests can complete
within 50 s and for the patient population, time given is five times more (250 s).
The “time out” feature helps in identifying the intentional delay from the patient
population. In this automated pegboard test, guided LEDs have been used to indicate
the position where the peg is to be inserted. In case the subject keeps the peg in
a sensor that he was not supposed to keep, the subsequent LED will not light up.
This indicates that previous peg movement was wrong and participant can correct it
immediately. Hence, LED arrangements guide the participant during the peg insertion
and removal process.
Once all the nine pegs have been inserted, and the hand movement sensed by the
IR sensor count is 9, a message to remove the peg is displayed on the screen. Once
the message to remove has appeared, the subject is supposed to remove the pegs out
from the sensors in reverse order. The code first checks whether the peg has been
taken out of the sensor or not by checking the output of the photo-interrupter sensor.
During removal of the peg, a low signal has been detected. The next thing that’s
detected is the hand movement of the subject. Once both the sensor’s output and
hand movement have been detected, the counter corresponding to hand movement
is incremented. If the time taken by the subject to remove the peg from the sensor
is more than 250 s, the “time out” message will be displayed, and the test will be
terminated. Once all the pegs from the nine sensors have been removed, the hand
movement count and the total time taken to perform the test will be displayed on the
screen.
Fig. 3 A flowchart
explaining the programming
of the set-up
3 Results
The initial part of the research work was to measure the peg insertion time. In this
part, time was measured only till the last peg was inserted without using the third
person’s interference with the help of LED lights. These LED arrangements guide the
subject to place the pegs in an orderly fashion. Figure 4 shows the subject performing
the test by inserting the pegs into the sensor.
Once the insertion part is done, it is time to test the removal part and cross verify
the results with the manual stopwatch. In the removal part, another important thing to
be added was the glowing of LED’s backward, with the difference being the fact that
as soon as the peg was removed, subsequent LED will light up contrary to the last
part in which the LED used to light up once the peg was inserted. A specific pattern
is followed, and this is ensured using the infrared (IR) sensor. The circuit has been
coded so that unless the IR sensor has not detected one hand movement, the sensor
will not detect the peg, making sure that the subject follows the rule. Once insertion
and removal parts were done, the total time and hand movement were displayed on
the screen, as shown in Fig. 5. HC-05 Bluetooth module is used, which transmits the
data on the screen to mobile phone using an application called Bluetooth terminal
(Fig. 6).
Overall, five subjects of different age groups were tested, and the results are
displayed as shown in Table 1. It can also be inferred from Table 1 that the time
taken to complete the test increases as the age increases. In all the cases, the right
hand is the dominant hand, and hence the test was done using the same as well.
Thus, Table 1 shows the results obtained when each of the five subjects was tested
for their reflexes associated with their right hand. The subjects were of varying with
same hand dominance (right-hand user). This observation leads to believe that the
average reaction time tends to gradually increase in the age range of 20–30 years,
whereas the rise is exponential when seen in the range of 40–80 years. Figure 7 was
the graphical analysis of the subjects when they were tested for the first time.
The hand–eye coordination of all the subjects is improved by practicing pegboard
tests at regular intervals. Normative data from the study [10] also confirm the same.
Fig. 4 Subject inserting the pegs

Fig. 5 Output of the pegboard test experiment in the Arduino IDE environment
A separate pegboard test session was conducted after the two weeks of training, and
the results are shown in Table 2.
As shown in Table 2 and Fig. 8, the reaction time improved after repeatedly
practicing the same test over a significant time. Hence it can be inferred that the
same test was used to detect and improvise the patient’s condition.
4 Discussions
The advantages provided by the automated pegboard to the traditional pegboard are
as follows:
• Presents the result with greater accuracy by using internal timer and appropriate
sensors.
• It allows registering the time for each one of the pegs.
• It allows knowing if two or more pegs move at the same time.
• It allows comparing the time taken to complete the test using the right hand and
the left hand with greater accuracy.
• It allows determining the number of pegs moved successfully during the test.
Fig. 6 Transmission of data

using HC-05 Bluetooth
module
Table 1 Age of the subject with the time taken for the test
S. No. Age Time taken (s)
Person 1 09 25
Person 2 22 29
Person 3 45 35
Person 4 52 36
Person 5 80 60
• It gives a visual indication if the peg has been inserted properly or not. When the
peg is inserted correctly, the LED goes off, and when the peg is removed, the LED
is on.
The pegboard was constructed with the help of breadboards, single strand wire,
and jumper wire. Also, the insertion and removal sequence of the pegs were success-
fully achieved. The timer was incorporated into the system to calculate the whole
Fig. 7 Graph representing the time taken for the pegboard test with age
Table 2 Improved performance during pegboard test

S. No. Age Time taken (s)
Person 1 09 22
Person 2 22 26
Person 3 45 34
Person 4 52 34
Person 5 80 56
Fig. 8 Rehabilitation graph showing improvement in the test performance

process’s completion time using the millis () function. All the functions of the
pegboard were demonstrated using the flowcharts, block diagrams, and serial monitor
and serial plotter outputs. In future work, portability of the device will be increased
by developing pegboard in a 3D printer with proper Printer Circuit Board (PCB).
5 Conclusion
Conducting the automated pegboard test gives vital information related to the indi-
vidual participant’s hand–eye coordination and enhances the fine motor skills of the
subject. One advantage of this test over the traditional tests is that this test can be
conducted without any external supervision and can give accurate results compared
to the conventional tests. The test process can be incorporated by including the diver-
sified demographical nature of the participants in the future study. Due to the present
pandemic situation, we have considered convenience sampling rather than random
selection from the population. These samples individually might not be representa-
tive of the general population. The prototype will be transferred on the PCB board
considering the design specifications and portable in future work.
References
1. Kellor M, Frost J, Silberberg N, Iversen I, Cummings R (1971) Hand strength and dexterity.
Am J Occup Therapy Off Publ Am Occup Therapy Assoc 25(2):77–83 PMID: 5551515
2. Johansson GM, Häger CK (2019) A modified standardized Nine-Hole Peg test for valid and
reliable kinematic assessment of dexterity post-stroke. J NeuroEng Rehabil 16(8). https://doi.
org/10.1186/s12984-019-0479-y
3. Rivera D, García A, Ortega JE, Alarcos B, van der Meulen K, Velasco JR, Del Barrio C
(2019) Intraindividual variability measurement of fine manual motor skills in children using an
electronic pegboard: cohort study. JMIR Mhealth Uhealth 7(8). https://doi.org/10.2196/12434
4. Jobbágy Á, Marik AR, Fazekas G (2018) Quantification of the upper extremity motor functions
of stroke patients using a smart Nine-Hole Peg tester. J Healthc Eng. https://doi.org/10.1155/
2018/7425858
5. Wang YC, Bohannon RW, Kapellusch J, Garg A, Gershon RC (2015) Dexterity as measured
with the 9-Hole Peg Test (9-HPT) across the age span. J Hand Ther 28(1):53–59. https://doi.
org/10.1016/j.jht.2014.09.002
6. Feys P, Lamers I, Francis G, Benedict R, Phillips G, LaRocca N, Hudson LD, Rudick R (2017)
Multiple sclerosis outcome assessments consortium. The Nine-Hole Peg Test as a manual
dexterity performance measure for multiple sclerosis. Multiple Scler 23(5):711–720. https://
doi.org/10.1177/1352458517690824
7. Earhart GM, Cavanaugh JT, Ellis T, Ford MP, Foreman KB, Dibble L (2011) The 9-hole PEG
test of upper extremity function: average values, test-retest reliability, and factors contributing
to performance in people with Parkinson disease. J Neurol Phys Ther 35(4):157–163. https://
doi.org/10.1097/NPT.0b013e318235da08
8. Bowler M, Amirabdollahian F, Dautenhahn K (2011) Using an embedded reality approach to
improve test reliability for NHPT tasks. IEEE Int Conf Rehab Robot. https://doi.org/10.1109/
ICORR.2011.5975343
9. Lemmens RJ, Janssen-Potten YJ, Timmermans AA, Smeets RJ, Seelen HA (2015) Recognizing
complex upper extremity activities using body worn sensors. PLoS One 10(3). https://doi.org/
10.1371/journal.pone.0118642
10. Mathiowetz V, Weber K, Kashman N, Volland G (1985) Adult norms for the Nine-Hole Peg
test of finger dexterity. Occup Therapy J Res 5(1):24–38. https://doi.org/10.1177/153944928
500500102
Implementation of Indoor Navigation
Control for Two-Wheeled Self-balancing
Robot
B. Vignesh, Deepa Jose, and P. Nirmal Kumar
Abstract Two-wheeled self-balancing robots (TWSBR) overcomes the limitations

of conventional three or four-wheeled robots by offering a zero-turning radius with
two motors. The usage of conventional dc motors with encoders in these two-wheeled
self-balancing robots requires a complex design and tuning of cascaded controllers.
In this paper, we implement a model that reduces the complexity of the design, and
we have implemented a simple low-cost indoor navigation technique along with
it, to know about the robustness of the design. The proposed system uses stepper
motors in place of dc motors such that a single controller can be used that reduces
the complexity of the design. We have implemented laser-based navigational control
along with obstacle avoidance control on this robot for the indoor environments to
show the stability of the robot during the motion for reaching the destination goal
along with the detection and avoidance of the obstacles in the path of the robot.
Keywords ROS · Lidar · Two-wheeled self-balancing robot · Obstacle

avoidance · Navigation
1 Introduction
In today’s scenario robotics are developed in different structures that have their
own advantages when compared with other robotic structures. The conventional
four-wheeled robots which are used for various applications in indoor and outdoor
environments have their own limitations such as requirement of high number of
wheels and motors which increases the energy consumption of the robot and also
B. Vignesh · P. Nirmal Kumar

Electronics and Communication Engineering, College of Engineering, Anna University, Guindy,
Chennai 600025, India
e-mail: nirmal@annauniv.edu
D. Jose (B)
Department of Electronics and Communication, KCG College of Technology, Chennai, Tamil
Nadu 600097, India
e-mail: deepa.ece@kcgcollege.com
https://doi.org/10.1007/978-981-19-6913-3_31
462 B. Vignesh et al.
these robots require additional spaces on the path for making some rotations or
when changing its directions which makes it difficult to control over narrow paths.
Such limitations were overcome by the two-wheeled self-balancing robots (TWSBR)
which requires only two motors for locomotion and hence the power requirement is
less and also it requires zero turning radius which is the most advantage of the robot
making it suitable for narrow spaces [1].
Many papers were proposed with different types of cascaded controllers for better
positional and directional stability. Cascaded controller like PD-PI was proposed [2]
where the PD controller is maintaining the balance and the PI controller is tuned for
the position control of the TWSBR and compared with the adaptive fuzzy controller.
A two-level adaptive controller along with PD controller [3] was designed to tackle
the instability of the TWSBR which provided a better stability, but the implementation
is complex due to multiple controllers being cascaded. As given in [4] the sliding
mode-based controllers offer a robust control of the TWSBR over the disturbance and
noise. Cascaded PID controllers [5] were designed for TWSBR with conventional
DC-geared motors. In the above-discussed papers have commonly used conventional
DC-geared motors with encoders to identify the position of the shaft and to rotate
it precisely. And also, the cascaded form of controllers discussed above are quite
difficult to tune for a good stability.
Some of the indoor navigational methodologies were studied to find a way for
low-cost robust navigational technique. Ultrasonic based obstacle avoidance along
with Fuzzy controller [6] were adopted for collision-free movement. A 3D SLAM
was developed by using a rotating 2D laser scanner [7] that was aimed to provide a 3D
map with 2D lidar sensor with a complex algorithm. Autonomous mobile robots were
designed in the ROS environment [8] for establishing a mapping and path planning
for indoor navigation. ROS is nowadays widely used in various robotic applications
because of its dedicated platform for robotics.
In this paper, we proposed a TWSBR model with single PID controller by using
stepper motors in place of DC geared motors along with the integration of the 2D
Lidar and ultrasonic sensors for a low-cost collision free navigation of the robot in
the indoor environments.
2 Proposed System Design
The proposed system has three sections where each section does different operations.
The overall block diagram of the proposed system is given in Fig. 1.
Those three sections are (i) Balancing controller design (ii) obstacle avoidance
controller (iii) navigational controller.
A. Balancing Controller Design
It is the important design of the robot which ensures the stability of the robot at
all circumstances that contains PID controller, IMU sensor, motor control, As said
Implementation of Indoor Navigation Control for Two-Wheeled … 463
Fig. 1 Overall block diagram of the TWSBR model
earlier, since we use Stepper motors in TWSBR a single PID controller is sufficient
for maintaining the stability of the robot.
The robot’s upright position is maintained by computing the error from the actual
setpoint to the current angle of the robot. The current angle of the robot is measured
by the MPU 6050 Inertial Measurement Unit (IMU) sensor. The values out from
the sensor are not accurate; some offsets are there due to environmental noises.
These will affect the stability of the system, Sun et al. [8] has used a kalman filter
that gives more accurate sensor values at a cost higher computation. Since we are
implementing in a low-cost hardware like an arduino, implementation of kalman
filter increases the time lag which causes instability of the system. Hence, we have
designed a complementary filter for the filtering purposes. The block diagram of the
complementary filter is shown in Fig. 2.
The complementary filter is simple to design and requires lower computations
time. The equation corresponding to the complementary filter is shown in Eq. (1),
where α is the constant calculated from the sampling frequency of the code.
Tilt_angle = α ∗ (acc_Angle) + (1 − α) ∗ (gyro_angle) (1)
Fig. 2 Complementary filter

The error calculated is given as input to the PID controller. The PID controllers
are linear aThe PID controller requires the value of three gains proportional gain
(P), integral gain (I), derivative gain (D). Each gain has certain significance in some
parameters like the peak overshoot, settling time, rise time, etc. By tuning the PID
gain values by trial-and-error method offers a better controller for the robot. The
equation for the PID controller is given in the Eq. (2).

derr
pid_out = K p ∗ err + K i err ∗ dt + ∗ Kd (2)
dt
where K p —proportional gain, K i —integral gain, K d —differential gain.

The output of the PID controllers is sent as pulse width modulation (PWM) pulses
to the stepper motor drivers. The advantage of the stepper motors is the rotation
angle of the shaft can be controlled by the number of pulses given to the motor
driver and the speed of the motor is controlled by adjusting the frequency of the
pulses. We use NEMA 17 stepper motors which is a low-cost stepper motor that
offers a torque of 4.44 kg cm. Since it is a two-wheeled robot, I have implemented a
differential drive system for the robot which makes turning or rotation easier. Simple
driver circuitry and elimination of the encoder makes the design simpler and more
effective while compared to the cascaded controllers. Compared to other controllers
like fuzzy or LQR which might be more robust compared to PID are difficult in real-
time implementation. The PID controllers are simple and its robustness is sufficient
for the robot to maintain its stability during directional movements [9, 10].
To execute the directional commands unlike the normal robots the direction of
the TWSBR is altered by changing the setpoint of the robot. Making the setpoint to
positive makes the robot to move forward and setpoint to negative makes the robot
to move backward. The directional commands are given from the other two sections
based on the obstacles around the robot and optimal path chosen by the algorithm.
The flowchart for the balancing algorithm is shown in Fig. 3.
B. Navigational Controller
For adding an indoor navigation technique, the methods discussed earlier uses
sensors like lidar, RGBD camera, odometry sensors, etc., by fusing different sensor
values together a sophisticated navigation can be implemented but this comes with
higher hardware and computational costs. In this paper, I have used 2D lidar sensors
for creating a map of the indoor environment and path planner creates a path based
on the goal, and the directional commands are sent to the microcontroller of the
balancing controller unit. And also, the model receives the directional commands
from the obstacle avoidance section where both of these commands are fused in such
a way that the model is able to locate itself in an environment and navigate to the
respective goal without colliding with the dynamic obstacles present in the particular
chosen path. This way of integrating the data from the lidar and ultrasonic sensors
provides a better navigation with a low-cost investment.
The slamtec 2D lidar is used which detects the distance of the obstacles by the
laser triangulation method. The lidar sensor covers a distance of 8 m in radius with
Fig. 3 Flowchart of the balancing algorithm
a 360 angular rate. The sampling frequency of the lidar sensor is 5.5 Hz that gives
approximately 4000 readings in a minute. The accuracy of the distance is high when
compared with ultrasonic sensors with resolution of 0.5°.
This navigation controller is developed in the Robotic Operating System (ROS)
which is a dedicated middleware for the robotic applications. In the ROS we develop
the code in blocks where each block of the code is called a node. One node for getting
the data from the lidar, another node for creating the map from obtained data from
the lidar and the other node for identifying the respective optimal path for the given
destination and sends the commands to the balancing controller unit.
Raspberry pi 4 is used for implementing this technique in ROS. At first every node
was developed in the ROS environment of pi and simulated in the available sample
home environments of the ROS successfully. But on real-time execution with real-
time data from the lidar sensor since the GPU of the Raspberry pi is not sufficient
for running all nodes along with the application softwares like Rviz and Gmapping
simultaneously the outcomes on movement of the robot were not accurate. So, I
have used the ROS toolbox of matlab software in remote pc for creating the map and
running the path planner in it for finding the optimal path and directional commands
are sent from the pc to the balancing controller unit.
C. A* Algorithm
The robot at first is made to move around the given indoor environment such that
a map is generated from the lidar data. Then the local plannar uses A* algorithm to
find the optimal route for the selected destination goal in the created map. The A*
algorithm assign cost values to different motions in the adjacent cells and choses the
cells with lower cost and repeat the algorithm.
After generating the path, the directional commands are sent to the Balancing
Controller Unit (BCU) via bluetooth for producing the respective movements.
D. Obstacle Avoidance Unit
The lidar sensor is capable of detecting the obstacles only to its range of vicinity.
The obstacles that are located on the floor which are very much smaller are not
detected by the lidar which might result in collision of the robot with the obstacle.
To overcome this issue, we added a section called obstacle avoidance unit.
Here, the ultrasonic sensors are used for obstacle detection which offers better
obstacle detection at lower price. The developed model senses the nearby obstacles
which are in the range of 50 cm that could collapse the stability of the robot. When
the obstacle is detected the location of the obstacles is first determined by measuring
the distance of the object from various sensors placed at different angles and then the
direction of the movement is chosen. Whenever an object is detected the robot holds
the directional commands sent by the navigational unit and waits for some seconds
to find whether the object is moved from the path, if not moved then it avoids that
particular obstacle by moving around it where now the directional commands are sent
by obstacle avoidance unit and after overcoming the obstacle once again it resumes
the navigational unit and receives the directional commands from the navigational
unit for reaching the destination.
We have used six ultrasonic sensors which are placed at different angles and
different heights such that it covers 150° of area in front of the robot. Three sensors
are placed at the lower part of the robot in a tilted manner so that they are capable
of detecting obstacles on the floor which have a height above 1 cm. the sensors are
tilted by an angle of 50°. And the other three sensors are placed in the top below the
lidar sensor where one in the middle is facing straight forward and the other two are
facing towards the diagonal direction of the robot. This kind of placement of sensors
offers an optimum obstacle avoidance for the TWSBR model.
3 Structure of the TWSBR Model
The structure of the robot looks similar to the arrangement of the stacks, where each
stack contains certain components of the robot. The structure of the robot is built
over acrylic material which provides low weight structure. The stepper motors are
attached under the base layer of acrylic sheet and above the base layer contains the
obstacle avoidance unit placed on it that is interfaced with the ultrasonic sensors.
The complete structure of the developed TWSBR model is shown in Fig. 4.
The first layer consists of the balancing controller unit that comprises the MPU
6050 sensor and motor driver and microcontroller for maintaining the stability of the
TWSBR model. The second layer sheet consists of the battery pack that powers the
entire setup. A Lithium-ion battery of 12v 1800 maH battery is used which offers
higher current density at lower weight. The third layer houses the raspberry pi 4 board
which is going to fetch the data from the lidar sensor and send it to the computer via
Fig. 4 Complete structure

of the TWSBR model
the ROS master URI protocol for further mapping and path planning. The Rplidar is
mounted to the fourth layer which is the top most layer of the robot structure.
The balancing algorithm was first experimented. Instead of modelling the controller
in MATLAB and tuning the gain values, I have directly implemented on the hardware.
First, the proportional gain is set such that the robot is able to balance with oscillatory
movements; then secondly, the differential gain is adjusted to damp out the oscillatory
movements, and lastly, the integral gain is set to nullify the offset in the position of the
robot. The responses of the pid controller for disturbances and directional commands
are plotted as a graph is shown below in Fig. 5a, b.
From the above responses, it may be seen that output of the pid controller is near
to zero when the robot is at still position. And pid is positive for forward direction
and pid is negative for backward.
Fig. 5 a Pid_output
response on disturbances. b
Pid_output response on
direction movements
After computing the balancing algorithm, the obstacle controller was tested the
ultrasonic sensors were able to detect the obstacles and sent the right directional
commands to the balancing controller unit when an obstacle is detected which is
shown in the Fig. 6a–c are given.
At last, the navigation controller was tested for mapping and path planning. As
said earlier the developed code was tested first with the simulational environments
available in ROS environment is shown in Figs. 7 and 8
Fig. 9 shows the map created for simulation environment and Fig. 10 shows the
execution of the path planning and navigation in the simulation window on the created
map.
After successful simulation, the controller was implemented in raspberry pi and
MATLAB. The robot was able to create the map successfully and planned an optimal
path for the particular destination which is shown in Figs. 11 and 12.
After checking the functionalities of each controller unit everything was inte-
grated together for a complete operation. As discussed in the abstract we are able to
implement a TWSBR model with a single PID controller and also, we have success-
fully implemented a low-cost navigation system for the robot to navigate in the given
indoor environment along with the obstacle avoidance.
5 Conclusion
The proposed paper has discussed the implementation of the lidar and ultrasonic
sensor-based navigation for a two wheeled self-balancing robot. Instead of using
conventional DC motor with encoders we have used stepper motors which also
produced a better stability like the dc motors which reduced the cost of the hardware
and also the need of tuning multiple controllers are eliminated where a single PID
controller is sufficient for maintaining the stability of the robot. The navigation using
ROS environment by the data from the lidar for map creation and fusing the values of
the ultrasonic sensors for a collision free navigation in the determined path defined
by the A* star algorithm were successfully integrated together providing a better
navigation technique for the robot with low-cost hardware.
6 Future Work
The proposed TWSBR model is capable of maintaining the stability in the event of
navigation and collision avoidance. The balancing algorithm has to modified such
that the robot moves with same stability on the slopes. At present, the robot is able
to stand still on a slope surface of about 20° but finds it difficult to move from the
flat surface to the sloped surface which must be eliminated by the modification in
the algorithm. Similarly, I have an idea on implementing swarm robotic concept
to this structure such that multiple TWSBR robots can be coordinated for carrying
Fig. 6 a–c Directional

commands sent to the BCU
based on the position of the
obstacle detected
Fig. 6 (continued)
Fig. 7 Home environment for simulation
the load of different weights by varying the orientation and number of the robots
involved in the load carrying task which ensures the efficient utilization of the robots
on dynamically changing load weights.
Fig. 8 Initial mapping
Fig. 9 Complete map of the environment

Fig. 10 Simulation of A* path planning in map
Fig. 11 Real-time binary occupancy grid map

Fig. 12 Real-time path planning and navigation
References
1. Zimit A, Yap HJ, Hamza M, Siradjuddin I, Hendrik B, Herawan T (2018) Modelling and
experimental analysis two-wheeled self balance robot using PID controller. https://doi.org/10.
1007/978-3-319-95165-2_48
2. Iwendi C, Alqarni MA, Anajemba JH, Alfakeeh AS, Zhang Z, Bashir AK (2019) Robust
navigational control of a two-wheeled self-balancing robot in a sensed environment. IEEE
Access 7:82337–82348. https://doi.org/10.1109/ACCESS.2019.2923916
3. Wasif A, Raza D, Rasheed W, Farooq Z, Ali SQ (2013) Design and implementation of a two
wheel self balancing robot with a two level adaptive control. In: Eighth international conference
on digital information management (ICDIM 2013), Islamabad, Pakistan, pp 187–193. https://
doi.org/10.1109/ICDIM.2013.6694021
4. Abeygunawardhana PKW, Defoort M, Murakami T (2010) Self-sustaining control of two-wheel
mobile manipulator using sliding mode control. In: The 11th IEEE international workshop on
advanced motion control, Nagaoka, Japan
5. Pratama D, Ardilla F, Binugroho EH, Pramadihanto D (2015) Tilt set-point correction system
for balancing robot using PID controller. In: 2015 international conference on control, elec-
tronics, renewable energy and communications (ICCEREC), Bandung, Indonesia, pp 129–135.
https://doi.org/10.1109/ICCEREC.2015.7337031
6. Ruan X, Li W (2014) Ultrasonic sensor based two-wheeled self-balancing robot obstacle avoid-
ance control system. In: 2014 IEEE international conference on mechatronics and automation,
Tianjin, China, pp 896–900. https://doi.org/10.1109/ICMA.2014.6885816
7. Fang Z, Zhao S, Wen S (2017) A real-time and low-cost 3D SLAM system based on a continu-
ously rotating 2D laser scanner. In: 2017 IEEE 7th annual international conference on CYBER
technology in automation, control, and intelligent systems (CYBER), Honolulu, HI, USA, pp
454–459. https://doi.org/10.1109/CYBER.2017.8446162
8. Sun F, Yu Z, Yang H (2014) A design for two-wheeled self-balancing robot based on Kalman
filter and LQR. In: 2014 international conference on mechatronics and control (ICMC), Jinzhou,
China, pp 612–616. https://doi.org/10.1109/ICMC.2014.7231628
9. Natarajan E, Hong LW, Ramasamy M, Hou CC, Sengottuvelu R (2018) Design and development
of a robot gripper for food industries using Coanda effect. In: 2018 IEEE 4th international
symposium in robotics and manufacturing automation, ROMA, https://doi.org/10.1109/ROM
A46407.2018.8986699
10. Someshwaran M, Jose D, Paul JP (2020) Autonomous unmanned ground vehicle for
enhancement of defence strategies. Lect Notes Netw Syst 89:873–880
Application of NIR Spectroscopy
with Chemometrics for Discrimination
of Indian Black Pepper Berries
Arnab Giri, Dilip Sing, Sudarshana Ghosh Dastidar, Pallab Kanti Halder,
Nanaocha Sharma, Pulok K. Mukherjee, and Rajib Bandyopadhyay
Abstract The objective of this study was to discriminate the samples with different
piperine content using near-infrared reflectance (NIR) spectroscopy. Twenty black
pepper samples were collected from different parts of West Bengal, India. Clus-
tering analysis was carried out using principal component analysis (PCA) and linear
discriminant analysis (LDA) on the collected samples. The cluster separability crite-
rion was used for quantifying the separability. The separation index was obtained as
3368.4, which establishes the utility of the NIR instrument for discrimination of the
black pepper samples based on their piperine content.
Keywords NIR spectroscopy · Medicinal plant · Piperine · Black pepper ·

Principal component analysis (PCA) · Linear discriminant analysis (LDA) ·
Separability index
1 Introduction
Piper nigrum L. (family—Piperaceae), commonly known as black pepper, is found

in India, China, Brazil, and Southeast Asia. Being considered as the “King of Spices,”
it is one of the earliest spices used for seasoning to food and to enhance the flavor
of other spices [1]. Black pepper is a natural antioxidant and is extensively used in
anti-inflammatory, anti-cancer, anti-tumor, anti-periodic, anti-bacterial, anti-fungal
analgesic, and anti-pyretic treatment in traditional medicine like, Ayurveda, Unani,
A. Giri · D. Sing (B) · S. G. Dastidar · R. Bandyopadhyay

Department of Instrumentation and Electronics Engineering, Jadavpur University, Salt Lake
Campus, Kolkata, India
e-mail: dilipsing.ju@gmail.com
P. K. Halder · P. K. Mukherjee
School of Natural Product Studies, Jadavpur University, Kolkata, India
N. Sharma · P. K. Mukherjee
Department of Biotechnology, Institute of Bioresources and Sustainable Development,
Government of India, Imphal, India
https://doi.org/10.1007/978-981-19-6913-3_32
476 A. Giri et al.
and traditional Chinese medicine. It also contains vitamin A, C, E, K, niacin, β-

carotene, and cholesterol-reducing properties [2].
Piperine is the major alkaloid present in black pepper, responsible for its pungent
smell, taste, and medicinal properties. The concentration of piperine in black pepper
varies from 2 to 9% depending on climate, place of origin, and growing conditions,
which regulates the degree of therapeutic effectiveness [3]. Due to this varying quality
and its vast commercial importance, quality control and market grading of black
pepper are needed by estimating piperine, the marker molecule concentration.
Traditional quality assessment of black pepper involves visual and olfactory segre-
gation, which does not consider the piperine concentration. Modern physiochemical
studies involve estimation of piperine molecule concentration using analytical tech-
niques like chromatography, mass spectrometry. However, these analytical processes
are limited in applications as they are time-consuming, expensive, complicated,
labor-extensive, and involve samples [4, 5].
NIR spectroscopy has proved to be an effective, low-cost, non-invasive, and accu-
rate analytical method for estimating chemical constituents. The basic molecular
bonds of X–H (X:C, O, N, and S), which are the primary basic components of
any organic molecule, mainly reflect different features of the NIR spectrum. Near-
infrared reflectance spectroscopy (NIR) is an accurate and non-invasive analytical
method, which is used in diverse industrial applications to study the characteristics of
different chemical constituents. This is particularly useful in study of medicinal plants
[6–8]. The technique offers several advantages over other techniques like localized
surface plasmon resonance (LSPR) [9, 10], evanescent wave-based measurements
[11], and surface-enhanced Raman measurements involving the synthesis and use of
customized nanoparticle systems [12, 13]. NIR techniques offer higher sensitivity
and require minimum sample preparation procedure.
In this paper, the feasibility of NIR spectroscopy was studied for the quantifica-
tion of black pepper. NIR spectroscopy has been used for cluster analysis to find
the discrimination of black pepper samples with different piperine content in them.
The NIR spectral data of black pepper samples were subjected to principal compo-
nent analysis (PCA) and linear discriminant analysis (LDA). In addition, the cluster
separability criterion has been used for quantitative measurement, and the separation
index was calculated.
2.1 Sample Collection
Specimens of P. nigrum L. were collected from the local markets and plantations of
Kolkata and nearby districts of West Bengal, India. Twenty samples were considered
for experimentation.
Application of NIR Spectroscopy with Chemometrics … 477
2.2 Chemical Analysis
The RP-HPLC analysis of the black pepper berries was carried out using piperine
as the marker compound. The RP-HPLC method was validated based on calibration
curve with high correlation coefficient and low LOD, LOQ values. The solvent system
was optimized (methanol: water 1% glacial acetic acid—60:40) with retention time
for piperine at 5.5 min. The percentage content of piperine in the different samples
varied from 0.53% w/w to 6.5% w/w. The details of experimental procedure and
result can be found in [14].
2.3 Spectra Acquisition
NIR spectra of the black pepper samples were collected using the DWARF-Star NIR
spectrometer from StellarNet Inc. USA. It contains an InGaAs detector array of 256
diodes. The wavelength range is 900–1700 nm. The spectrometer is accommodated
with a bi-directional fiber-optic bundle installed at the high-intensity contact probe,
tungsten halogen lamp with 12 V/20 W bulb as the light source and a sample holder.
The block diagram and the experimental setup are shown in Figs. 1 and 2, respectively.
For each sample, 5 replicates of spectra were collected and stored for analysis.
Fig. 1 Block diagram Tungsten

Lamp Source
Sample in Bandpass Focusing Detector

sample Filter Lens
holder
478 A. Giri et al.
Fig. 2 Experimental setup
2.4 Software
SpectraWiz, Spectrometer OS v5.3, 2013 software was used for collecting the NIR
spectra. PCA and LDA techniques were implemented in Matlab V10.0 (Mathworks
Co., USA).
3 Data Analysis
3.1 Principal Component Analysis (PCA)
Principal component analysis (PCA) is a multivariate method to analyze a data table

whose datasets are related by quantitative dependent variables [15]. The primary
objective of the technique in this experiment was to obtain the important statistical
measures to represent as principle components and display the similarities as spot
maps on orthogonal axes.
3.2 Linear Discriminant Analysis (LDA)
Linear discriminant analysis (LDA) is a very common technique for dimension-

ality reduction problems as a preprocessing step for machine learning and pattern
classification applications [11]. For reduction features, LDA technique uses a lesser
dimension hyper-plane and in a direction that maximizes the separation between the
samples.
3.3 Separability Index (SI)
Separability is the quantitative measure to segregate the fairly different set of points
in a dataset. The separability measure is obtained by the ratio of the trace of the
“within class scatter matrix” (SW) to that of the “between class scatter matrix” (SB)
[14].
4.1 Wavelength Versus Absorbance Plot
The wavelength versus absorbance plot of the spectral data of black pepper samples
is shown in Fig. 1. The X axis of each of the spectral profile represents the wavelength
(in nm) in the range of 900–1700 nm, and the Y axis represents the absorbance.
4.2 Principal Component Analysis
PCA analysis was applied on the black pepper spectral data. The PCA plot for the
black pepper samples by considering the first two principal components is shown in
Fig. 3. For each sample, five replicates were available, and thus, there are 60 points
on the plot. The information content in the first two axes of PCA is PC1-89.31%
and PC2-6.46%. The PCA plot revealed that the samples of same variety have been
observed to appear as clusters, and no overlaps were observed among them (Fig. 4).
4.3 Discrimination Using Linear Discriminant Analysis
The spectral data were subjected to LDA analysis. A cluster separation criterion has
been used in our study for quantitative measurement, and the separation index was
obtained as 3368.4. Thus, the discrimination obtained is better than PCA. The LDA
plot is shown in Fig. 3 (Fig. 5).
480 A. Giri et al.
Fig. 3 Wavelength versus absorbance plot
PCA SCORE PLOT OF BLACK PEPPER SAMPLES
0.8 S1
S2
S3
0.6 S4
S5
S6
PC2 Explained: 22.622426 perc
S7
0.4
S8
S9
S10
0.2 S11
S12
S13
0 S14
S15
S16
-0.2 S17
S18
S19
S20
-0.4
-0.6
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
PC1 Explained: 74.144083 perc
Fig. 4 PCA plot of black pepper samples

LDA plot of Black pepper samples

0.1
S1
S2
S3
0.08 S4
S5
S6
S7
0.06
S8
S9
S10
0.04 S11
S12
S13
LDA2
0.02 S14
S15
S16
S17
0 S18
S19
S20
-0.02
-0.04
-0.06
-0.12 -0.1 -0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06
LDA1
Fig. 5 LDA plot of black pepper samples
5 Conclusion
The results highlight that different samples of black pepper with variation in piperine
content in them could be distinguished visually by PCA and LDA analysis. The
separability index is obtained as 3368.4 which shows a very good discrimination
using this technique. The results demonstrate that NIR spectroscopy is capable of
discrimination between different black pepper samples without any prior chemical
analysis.
Acknowledgements This project was funded by the National Medicinal Plant Board, Ministry of
AYUSH, Government of India [grant numbers: Z.18017/187/CSS/R&D/WB-02/2017/-18-NMPB-
IV A, 2017]. The authors express their gratitude toward the Institute of Bioresources and Sustainable
Development, an autonomous institute under Departmentof Biotechnology (DBT), Government of
India, Imphal, India for necessary help and support through IBSD-JU joint collaboration.
References
1. Black Pepper Plant, Editors of Encyclopaedia Britannica

2. Goswami T, Meghwal M (2012) Chemical composition, nutritional, medicinal and functional
properties of black pepper: a review. J Nutr Food Sci 01(S1). https://doi.org/10.4172/scientifi
creports.172
482 A. Giri et al.
3. Stojanović-Radić Z et al (2019) Piperine-a major principle of black pepper: a review of its

bioactivity and studies. Appl Sci 9(20):1–29. https://doi.org/10.3390/app9204270
4. Mukherjee PK, Bahadur S, Chaudhary SK, Kar A, Mukherjee K (2015) Quality related
safety issue-evidence-based validation of herbal medicine farm to pharma. In: Evidence-based
validation of herbal medicine, Elsevier Inc., pp 1–28
5. Mukherjee PK (2019) Chapter 17—Plant metabolomics and quality evaluation of herbal drugs.
In: Mukherjee PK (ed) Quality control and evaluation of herbal drugs. Elsevier, pp 629–653
6. Sing D et al (2020) “Prediction of andrographolide content in Andrographis paniculata Using
NIR Spectroscopy. IEEE Appl Signal Process Conf (ASPCON) 2020:335–338. https://doi.org/
10.1109/ASPCON49795.2020.9276668
7. Sing D et al (2021) Estimation of andrographolides and gradation of Andrographis paniculata
leaves using near infrared spectroscopy together with support vector machine. Front Pharmacol
12(May):1–8. https://doi.org/10.3389/fphar.2021.629833
8. Qu JH et al (2015) Applications of near-infrared spectroscopy in food safety evaluation and
control: a review of recent research advances. Crit Rev Food Sci Nutr 55(13):1939–1954.
https://doi.org/10.1080/10408398.2013.871693
9. Sadani K, Nag P, Mukherji S (2019) LSPR based optical fiber sensor with chitosan capped gold
nanoparticles on BSA for trace detection of Hg (II) in water, soil and food samples. Biosens
Bioelectron 134:90–96
10. Nag P, Sadani K, Mukherji S, Mukherji S (2020) Beta-lactam antibiotics induced bacteriolysis
on LSPR sensors for assessment of antimicrobial resistance and quantification of antibiotics.
Sensors Actuators B Chem 311:127945
11. Nag P, Sadani K, Mohapatra S, Mukherji S, Mukherji S (2021) Evanescent wave optical fiber
sensors using enzymatic hydrolysis on nanostructured polyaniline for detection of β-lactam
antibiotics in food and environment. Anal Chem 93(4):2299–2308
12. Nag P et al (2021) Polyphenol stabilized copper nanoparticle formulations for rapid disinfection
of bacteria and virus on diverse surfaces. Nanotechnology
13. Sadani K et al (2020) A point of use sensor assay for detecting purely viral versus viral-bacterial
samples. Sensors Actuators B Chem 322:128562
14. Sing D, et al (2021) Rapid estimation of piperine in black pepper: exploration of Raman
spectroscopy. Phytochem Anal
15. López del Val JA, Alonso Pérez de Agreda JP (1993) Principal components analysis. Aten
Primaria 12(6):333–338. https://doi.org/10.5455/ijlr.20170415115235
16. Tharwat A, Gaber T, Ibrahim A, Hassanien AE (2017) Linear discriminant analysis: a detailed
tutorial. AI Commun 30(2):169–190. https://doi.org/10.3233/AIC-170729
A Comparative Study Between Partial
Least Squares and Principal Component
Regression for Nondestructive
Quantification of Piperine Contents
in Black Pepper by Raman Spectroscopy
Dilip Sing, Sudarshana Ghosh Dastidar, Wasim Akram, Sourav Guchhait,

Shibu Narayan Jana, Subhadip Banerjee, Pulok Kumar Mukherjee,
and Rajib Bandyopadhyay
Abstract The aim of this work was to compare principal component regression
(PCR) and partial least squares (PLS) regression methods while estimating the
piperine content in black pepper using Raman spectroscopy. The calibration and
prediction models of the regression analysis on Raman spectra were developed using
PCR and PLS algorithm. The efficiency of the developed models was evaluated by
means of root mean square error of calibration (RMSEC), root mean square error
of prediction (RMSEP), and correlation coefficient (R2 ). For PCR algorithm, these
parameters were obtained as 0.1, 0.1, and 0.97, respectively; and for PLS regression,
the parameters were found as 0.05, 0.08, and 0.99, respectively. The results revealed
that Raman spectroscopy with PCR and PLS algorithm could be used for determining
the concentration of piperine in black pepper with an accuracy of 92.35% and 94.74%
respectively.
Keywords Raman spectroscopy · Partial least squares (PLS) · Principal

component regression (PCR) · Medicinal plant · Black pepper · Piperine
D. Sing (B) · S. G. Dastidar · W. Akram · S. Guchhait · R. Bandyopadhyay

Department of Instrumentation and Electronics Engineering, Jadavpur University, Salt Lake
Campus, Kolkata 700106, India
e-mail: dilipsing.ju@gmail.com
S. N. Jana · S. Banerjee · P. K. Mukherjee
School of Natural Product Studies, Jadavpur University, Kolkata 700032, India
P. K. Mukherjee
Institute of Bioresources and Sustainable Development, An Autonomous Institute under
Department of Biotechnology, Government of India, Imphal 795004, India
https://doi.org/10.1007/978-981-19-6913-3_33
484 D. Sing et al.
1 Introduction
The multivariate calibration methods are utilized abundantly for extracting relevant
information from spectral data for predicting concentrations of analytes in complex
samples [1–5]. There are several regression models that have been used for the devel-
opment of calibration models. In this regard, PCR and PLS are two of the commonly
used regression analysis techniques. PCR takes care of the collinearity problem with
a few factors. However, the collinearity problem could also be handled by PLS algo-
rithm utilizing fewer latent variables than with PCR. Several reports reveal that PLS
attains the minimum mean square error (MSE) with a lesser latent variable than
PCR. Therefore, partial lest square regression provides better optimization of latent
variables with less time as compared to PCR [6].
Here in this experiment, we have used Raman spectroscopy with black pepper
samples. In black pepper, piperine is the most important marker molecule. It has
lots of different pharmacological features like anti-inflammatory [7], anticancer [7],
antibacterial, antioxidant [8], antidepressant [9], analgesic, and anti-pyretic [10]. For
quality assessment of black pepper, the prediction of piperine content in black pepper
is extremely important. The conventional methods for quantifying piperine involve
high-end chromatographic and mass spectroscopic technologies. These techniques
in combination with chemometrics help in assessment of the secondary metabolites
present in plants [11, 12]. These methods are laboratory-based, which need tedious
processing of samples causing sample destruction. Other detection techniques have
been evolving that includes more advantages, such as in situ estimation without
prerequisite of chemicals, portability, and low cost such as [13–15]. An alternative
low-cost and effective analytical technique based on Raman spectroscopy shows
enormous application in numerous fields and products like in essential oil [16],
material science [17], geology [18], and medicine [19].
In our previous study, we explored the efficiency of Raman spectroscopy for
estimation of piperine in black pepper [20]. In this paper, we present the comparison
of the performances of PCR and PLS algorithms for the data of black pepper with
Raman shift in the range of 1097–1627 cm−1 .
2.1 Black Pepper Samples and Chemical Analysis
Twelve Piper nigrum samples were collected from different geographical locations of
India. For chemical analysis, methanol extracts were prepared and then the analyses
were carried out using the RP-HPLC system (Waters, USA). It is equipped with
a LC-30AD pump, ultraviolet/visible detector (three Lines degasser) with volume
of four hundred micro-litter and a rheodyne 7725i injector with twenty micro-litter
loop. The details of experimental procedure and result can be found in [20].
A Comparative Study Between Partial Least Squares and Principal … 485
2.2 Spectra Acquisition
The spectrum of Raman shift was collected using an IDRaman micro-Microscope

from Ocean Optics, Inc., Florida, USA. It was equipped with the laser source (785 nm)
and has an output power of 100 mW (40 mW at sample). A thermo-electric cooler
was used for minimizing the effect due to variation in temperature. The groove
frequency of the grating was 1200 lines/mm, and the spectral response range was
200–3200 cm−1 . Resolution was maintained at 8 cm−1 . The average of sixteen was
taken for each sample, and the exposure time was set at 1 s. Before acquiring each
spectral response, a clean and dry sample container is used for taking the background
spectrum.
2.3 Principal Component Regression
Principal component regression (PCR) is a data analysis method based on linear

regression. In PCR, the spectral points primarily break up into suitable components
(latent variables). The important relevant information of the spectral data is contained
in the initial few latent variables [21].
2.4 Partial Least Squares Regression
Partial least squares (PLS) regression is a mathematical model like principal compo-
nents regression [20, 22]. In PLS, training data with maximum variance is projected
to a new space to create the calibration model.
2.5 Software
ToupView Program and the OceanView Program software (version 7, Ocean Optics,
Inc., Florida, USA) were used for controlling the instrument. The implementation of
the PCR and PLS regression technique was done in the R Studio.
3 Results and Discussions
The raw data collected from the Raman spectrometer was used to PLS and PCR
regression technique for building the calibration model. The detailed experimental
workflow can be found in Fig. 2 in our paper [1] for training and testing data set
486 D. Sing et al.
with chemical analysis. The whole data was divided in two parts. Here, the number
of data for calibration and prediction is 9 and 3, respectively. In this process, the
number of components was decided by computing the lowest root mean square
error cross-validation (RMSECV). The optimal model was chosen with the overall
lowest RMSECV and highest prediction accuracy. The number of components versus
RMSECV plot for raw spectral data using PCR and PLS algorithm is shown in Figs. 1
The calibration model for black pepper samples was built using 80% of the data
and 20% of the data was used for testing. PCR and PLS regression algorithms were
used and the content of piperine in black pepper was predicted. The results are shown
in Table 1. FourPCR and PLS components were chosen as optimal. The average
accuracy of prediction for PCR and PLS regression was found to be 92.35% and
94.73%, respectively. Also, the root mean square error of prediction (RMSEP) of
three samples for PCR and PLS regression was found to be 0.1 and 0.08, respectively.
The correlation coefficient (R2 ) for test samples for PCR and PLS regression was
found to be 0.97 and 0.99, respectively. This revealed good correlation between
reference concentration and predicted concentration of piperine in black pepper. The
maximum and minimum standard deviations of the predicted piperine content were
± 0.012 and ± 0.167, respectively, for PCR and ± 0.006 and ± 0.138, respectively,
for PLS regression. Also, the average RPD (residual prediction deviation) value of the
test data set for PCR and PLS regression was obtained as 7.01 and 8.92, respectively.
Fig. 1 Number of components versus RMSECV using PCR algorithm

Fig. 2 Number of components versus RMSECV using PLS algorithm
Table 1 Result of PCR and PLSR models of piperine in black pepper

Model LV Calibration Prediction
RMSEC RC 2 RPDc RMSEP Rp 2 RPDc
PCR 4 0.1 0.96 5.12 0.1 0.97 7.01
PLS 4 0.05 0.99 10.43 0.08 0.99 8.92
4 Conclusion
A comparative study has been conducted using principal component regression and
partial least squares regression algorithms on the spectral data from a Raman spec-
trometer for predicting estimating piperine content in black pepper seed. The efficacy
of Raman spectrometer for quality estimation of black pepper has been illustrated.
The accuracy of prediction of piperine in black pepper samples using PCR and PLS
was obtained as 92.35% and 94.73%, respectively. The correlation coefficient (R2 )
for test samples for PCR and PLS regression was found to be 0.97 and 0.99, respec-
tively. Also, the root mean square error of prediction (RMSEP) for PCR and PLS
regression was found to be 0.1 and 0.08, respectively. The average RPD value of the
test data for PCR and PLS algorithm was obtained as 7.01 and 8.92, respectively. The
results reveal the efficacy of both the regression analysis methods for the estimation
of piperine in the black pepper using Raman spectral data.
488 D. Sing et al.
Acknowledgements The present study was sponsored by the National Medicinal Plant Board,
Ministry of AYUSH, Government of India. The authors are grateful to the Institute of Biore-
sources and Sustainable Development, an autonomous institute under Department of Biotechnology
(DBT), Government of India, Imphal, India, for necessary help and support through IBSD-JU joint
collaboration.
References
1. Sing D et al (2021) Estimation of andrographolides and gradation of Andrographis paniculata

leaves using near infrared spectroscopy together with support vector machine. Front Pharmacol
12(May):1–8. https://doi.org/10.3389/fphar.2021.629833
2. Sing D et al (2020) Prediction of andrographolide content in Andrographis paniculata using
NIR spectroscopy. IEEE Appl Signal Process Conf (ASPCON) 2020:335–338. https://doi.org/
10.1109/ASPCON49795.2020.9276668
3. Balabin RM, Safieva RZ, Lomakina EI (2007) Comparison of linear and nonlinear calibration
models based on near infrared (NIR) spectroscopy data for gasoline properties prediction.
Chemom Intell Lab Syst 88(2):183–188. https://doi.org/10.1016/j.chemolab.2007.04.006
4. Hemmateenejad B, Akhond M, Samari F (2007) A comparative study between PCR and PLS in
simultaneous spectrophotometric determination of diphenylamine, aniline, and phenol: effect
of wavelength selection. Spectrochim Acta Part A Mol Biomol Spectrosc 67(3–4):958–965.
https://doi.org/10.1016/j.saa.2006.09.014
5. Ghasemi J, Niazi A (2001) Simultaneous determination of cobalt and nickel. Comparison of
prediction ability of PCR and PLS using original, first and second derivative spectra. Microchem
J 68(1):1–11. https://doi.org/10.1016/S0026-265X(00)00159-4
6. Yeniay O, Göktas A (2002) A comparison of partial least squares regression with other
prediction methods. Hacettepe J Math Stat 31:99–111
7. Tasleem F, Azhar I, Ali SN, Perveen S, Mahmood ZA (2014) Analgesic and anti-inflammatory
activities of Piper nigrum L. Asian Pac J Trop Med 7(S1):S461–S468. https://doi.org/10.1016/
S1995-7645(14)60275-3
8. Zarai Z, Boujelbene E, Ben Salem N, Gargouri Y, Sayari A (2013) Antioxidant and antimicro-
bial activities of various solvent extracts, piperine and piperic acid from Piper nigrum. LWT
Food Sci Technol 50(2):634–641. https://doi.org/10.1016/j.lwt.2012.07.036
9. Li S, Wang C, Li W, Koike K, Nikaido T, Wang MW (2007) Antidepressant-like effects of
piperine and its derivative, antiepilepsirine. J Asian Nat Prod Res 9(5):421–430. https://doi.
org/10.1080/10286020500384302
10. Parmar VS et al (1997) Phytochemistry of the genus Piper. Phytochemistry 46(4):597–673.
https://doi.org/10.1016/S0031-9422(97)00328-2
11. Mukherjee P et al (2016) Metabolomics of medicinal plants—a versatile tool for standardization
of herbal products and quality evaluation of ayurvedic formulations. Curr Sci 111:1624. https://
doi.org/10.18520/cs/v111/i10/1624-1630
12. Mukherjee PK (2019) Chapter 17—Plant metabolomics and quality evaluation of herbal drugs.
In: Mukherjee PK (ed) Quality control and evaluation of herbal drugs. Elsevier, pp 629–653
13. Nag P, Sadani K, Mohapatra S, Mukherji S, Mukherji S (2021) Evanescent wave optical fiber
sensors using enzymatic hydrolysis on nanostructured polyaniline for detection of β-lactam
antibiotics in food and environment. Anal Chem 93(4):2299–2308
14. Nag P, Sadani K, Mukherji S, Mukherji S (2020) Beta-lactam antibiotics induced bacteriolysis
on LSPR sensors for assessment of antimicrobial resistance and quantification of antibiotics.
Sensors Actuators B Chem. 311:127945
15. Sadani K et al (2020) A point of use sensor assay for detecting purely viral versus viral-bacterial
samples. Sensors Actuators B Chem 322:128562
16. Hanif MA et al (2017) Evaluation of the effects of Zinc on the chemical composition and
biological activity of basil essential oil by using Raman spectroscopy. Ind Crops Prod 96:91–
101. https://doi.org/10.1016/j.indcrop.2016.10.058
17. Zhang X, Tan QH, Bin Wu J, Shi W, Tan PH (2016) Review on the Raman spectroscopy of
different types of layered materials. Nanoscale 8(12), pp 6435–6450. https://doi.org/10.1039/
c5nr07205k
18. Jehlička J, Edwards HGM (2008) Raman spectroscopy as a tool for the non-destructive identi-
fication of organic minerals in the geological record. Organic Geochem 39(4):371–386. https://
doi.org/10.1016/j.orggeochem.2008.01.005
19. Nabiev I, Chourpa I, Manfait M (1994) Applications of Raman and surface-enhanced Raman
scattering spectroscopy in medicine. J Raman Spectrosc 25(1):13–23. https://doi.org/10.1002/
jrs.1250250104.
20. Sing D et al (2021) Rapid estimation of piperine in black pepper: exploration of Raman
spectroscopy. Phytochem Anal
21. Lee KM, Herrman TJ (2016) Determination and prediction of fumonisin contamination in maize
by surface-enhanced Raman spectroscopy (SERS). Food Bioprocess Technol 9(4):588–603.
https://doi.org/10.1007/s11947-015-1654-1
22. SNJ, Dilip Sing PKM, Mallik R, Dastidar SG, Bandyopadhyay R, Banerjee S (2020) Prediction
of andrographolide content in Andrographis paniculata using NIR spectroscopy. IEEE Appl
Signal Process Conf 335–338
Power Quality Data Mining Using
Hybrid Feature Extraction Technique
Vidhya Sivaramakrishnan, Balaji Mahadevan, and Kamaraj Vijayarajan
Abstract This paper proposes an approach for identifying the nature of power
quality disturbances using hybrid feature extraction technique combining S trans-
form and Hilbert transform. Using the combined features obtained, the classification
is performed using Extreme Learning Machine (ELM). The effectiveness of the
proposed approach is tested using wide spectrum of power quality disturbances.
The comparison with existing methods indicates that the proposed hybrid signal
processing approach for feature extraction results in improved classification accu-
racy. Sensitivity of the proposed approach is examined for signals with noise and the
results are presented.
Keywords Power quality disturbance · S transform · Hilbert transform · Extreme

learning machine
1 Introduction
The distortion in the current and voltage signals lead to failure and malfunctioning of
several sensitive loads. These distortions are mainly attributed to the switching load
capacitors, power electronic equipments and natural phenomena like lightning [1,
2]. The recent advancements in classification algorithms along with efficient feature
extraction techniques have encouraged the researchers to take up studies in power
quality data analysis and classification. Numerous literatures have been reported
on improving the efficiency of classification using new variants in soft computing
V. Sivaramakrishnan (B)
New Prince Bhavani College of Engineering and Technology, Chennai, India
e-mail: vidhya.s@newprincehribhavani.com
B. Mahadevan · K. Vijayarajan
Sri Sivasubramaniya Nadar College of Engineering, Chennai, India
e-mail: balajim@ssn.edu.in
K. Vijayarajan
e-mail: kamarajv@ssn.edu.in
https://doi.org/10.1007/978-981-19-6913-3_34
492 V. Sivaramakrishnan et al.
and efficient signal processing techniques. The major challenge lies in selecting
appropriate features that form the basis of classifying power quality disturbance.
The time frequency algorithms have been extensively applied to extract the features
and thereby recognize the type of disturbance.
In [3], the authors have applied wavelet technique for extracting the features for
classifying power quality disturbance. The performance of the classifier with different
wavelet family has been tested under normal and noisy conditions. An S transform
based feature extraction technique has been proposed in [4]. The statistical features
extracted by applying S transform form the input to fuzzy expert system based clas-
sifier. The main advantage of S transform is that it has superior characteristics when
the signals are prone to noise. Owing to this merit, it has widely applied for detection
and classification of various time series events. The classification of events using S
transform and TT transform has been discussed in [4]. The patterns generated as the
result of applying transformation forms the basis for classification. Feature extrac-
tion using Hilbert transform has been discussed in [5]. The results of this method are
compared with other feature extraction methods to highlight its merits.
The authors have employed Teager energy operator and Hilbert transform for
detecting voltage flicker events [6]. For efficient classification of power quality
signals, the authors have integrated wavelet and Hilbert transform [7]. In this method,
the features are extracted by applying Hilbert transform on all the decomposed coef-
ficients obtained through wavelet transform. The results indicate that the proposed
method has enhanced accuracy of classification.
From the literature, it is evident that integrating two transformation techniques
results for feature extraction results in improved classification accuracy. However,
the classification with random noise signals remains a challenge and there is a need
to explore efficient feature extraction and classification techniques. In this trend, the
concept of integrating two transformation techniques for feature extraction generates
interest. This paper focuses on a hybrid feature extraction technique using S transform
and Hilbert transform. To highlight the merits, the performance of the hybrid feature
extraction technique is compared with other feature extraction methods. The feature
classification is done using extreme learning machine.
2 S Transform
The S transform [8] is a feature extraction technique that provides the required data
for analyzing the disturbances [8, 9]. This method incorporates the features of short-
term Fourier transform and the wavelet transform. The S transform is derived from
wavelet transform by changing the phase of the window function or mother wavelet.
The S transform
of the signal, y(t), is the product of the signal and a phase correction
function e− j2 f t [8]. The S transform of y(t) is defined as
Power Quality Data Mining Using Hybrid Feature Extraction Technique 493
∞
S(τ, f ) = y(t)ga f (τ − t) e−i2f t dt (1)
−∞
where ga f (τ ) is the Gaussian modulation function defined as

|f| − τ2
ga f = √ e 2σ 2 (2)
2π
and
1
σ = (3)
|f|
The expression becomes

∞ (τ − t)2 f 2/
|f| − 2
S(τ, f ) = y(t) √ e e−i2 ft
dt (4)
2π
−∞
The discrete Fourier transform of the time series y(t) is expressed as
N −1
n − (2i )
nk
H = y(kT )e N
(5)
NT K =0
/
By substituting f → n N T and τ → j T .
the discrete S transform is obtained as below
N −1
n m+n
S j T, = H G(m, n)ei2mπ j/N (6)
NT m=0
NT
where
G(m, n) = e−2π m /n and j, m, n = 0, 1,…, N − 1.
2 2 2
The discrete inverse of the S transform can be obtained as

⎡ ⎤
N −1 N −1
⎣ n ⎦ei2nk
y(kT ) = S( j T, (7)
n=0 j=0
N T
The rows and columns of the matrix obtained as output of S transform correspond
to frequency and time, respectively. The features extracted from time, frequency,
amplitude and from high frequency and low frequency areas characterize the power
quality events.
Figure 1 illustrates the signal without disturbance together with the characteristics
obtained from S transform. The characteristics in Fig. 2 depict these plots for sag
disturbance.
Fig. 1 Signal without disturbance and its S transform characteristics
Fig. 2 Sag disturbance and its S transform characteristics

Even though the study of literature show that the S transform exhibits supe-
rior characteristics, the performance under noisy conditions remains a challenge.
Hence, this work explores the combination of S transform with Hilbert transform for
improved classification accuracy.
3 Hilbert Transform
The Hilbert transform results in a signal whose amplitude is the same as that of
the original signal and the phase component lags the original by 90°. The result of
Hilbert transform is the convolution of original signal y(t) with function πt1 [5, 10].
Hilbert transform can be expressed as
1
H (y(t)) = ẑ(t) = ∗ y(t) (8)
πt
1
H z (t) = y(t) = ∗ ẑ (t) (9)
πt
As the integrals defining the convolution do not converge. HT can be expressed

as below:
∞ ∞
1 y(τ ) 1 y(t − τ )
H (y(t)) = P V dτ = P V dτ (10)
π (t − τ ) π τ
−∞ −∞
where PV represents the Cauchy’s principal value of the singular integral.

The discrete Hilbert transform is given by
∞
1 1 y(λ)
y H (t) = y(t) ∗ = dλ (11)
πt π t −λ
−∞
The discrete Hilbert transform output is 90° phase shifted and the complete signal
is expressed as
yc [k] = y[k] + j y H [k] (12)
The envelope of the original signal is then defined as

/
|y A [k]| = y 2 [k] + y H2 [k] (13)
The envelope of Hilbert transforms aids in discriminating different PQ distur-

bances.
Figure 3 shows the Hilbert transform envelope for normal signal, sag signal and
sag signal with noise.
In this work, the features extracted from S transform [11] and Hilbert trans-
form are used for identifying power quality disturbances namely PQD1-sag, PQD2-
Harmonic, PQD3-Interruption, PQD4-Swell, PQD5-Swell with Harmonic, PQD6-
Sag with Harmonic, PQD7-Flicker, PQD8-Oscillatory Transient. The 9 features
extracted from Hilbert envelope given in Table 1 together with 20 features extracted
Fig. 3 a Signal without disturbance. b Output of Hilbert transform for signal without disturbance.
c Sag disturbance. d Output of Hilbert transform for sag signal. e Output of Hilbert transform for
normal signal and sag signal with noise
Table 1 Features extracted

F1—mean of the Hilbert array
from Hilbert transform
F2—standard deviation of the Hilbert array
F3—maximum value of the Hilbert array
F4—energy of the Hilbert array
F5—kurtosis of the Hilbert array
F6—skewness of the Hilbert array
F7—median of the Hilbert array
F8—range of the Hilbert array
F9—minimum value of the Hilbert array
from S transform [12] are used for identifying the nature of disturbance. These
features have been obtained by programming in MATLAB/Simulink environment.
4 Extreme Learning Machine Classifier
Extreme Learning Machine (ELM) classifier consists of input layer, a single hidden
layer and an output layer. In ELM, input weights are selected randomly and the
output weights are calculated analytically [13]. The various activation functions for
the hidden layer used in ELM include Sigmoid, Sine, Gaussian and hard limiting
functions. Linear Activation function is applied for the output neurons. The merits
of ELM include faster learning speed and enhanced generalization performance
[11, 12, 14, 15].
The output Ok of the ELM network composed of H hidden neurons is described
as
H
Ok = wk j G j (V , b, X i ), k = 1, 2, . . . , C (14)
j=1
where V represents (H × n) input weight, b represents (H × 1) bias values for each

hidden neuron, W represents (C × H) output weight, C is multi-class classification
representing distinct classes with N observations (X i , T i , i = 1, 2, …, n), Gj (.) is the
output of the jth hidden neuron and Gj (.) is the activation function.
The output of the jth hidden neuron Gj (.) with sigmoidal activation function is
described as
N

G j (V , b, X i ) = tanh b j + v jk xik (15)
k=1
Equation (14) can be expressed in matrix form as

O = W Oh (16)
where Oh is an H × n matrix which is given by

⎛ ⎞
G 1 (V , b, X 1 ) G 1 (V , b, X 2 ) . . . G 1 (V, b, X N )
⎜ .. .. .. ⎟
Oh = ⎝ . . . ⎠
G H (V , b, X 1 ) G H (V, b, X 2 ) . . . G H (V , b, X N )
The target tik is given as
tik = 1 if ci = k
= −1 otherwise k = 1, 2, . . . C (17)
where ci is the class label for X i

In ELM, the output weights are calculated by presuming the network output Y
equal to the coded class label t,
W = O Oh+ (18)
where Oh+ is the Moore–Penrose pseudo inverse of matrix Oh

The estimated class label is computed as

C i = arg max Oik (19)

k=1,2...c
5 Extreme Learning Machine Classifier
The evaluation of performance of the proposed technique is done by taking into

consideration different types of power quality events. For each disturbance, around
125 samples are generated using Matlab. In addition 30, 40 and 50 dB noise is added
to the original signal to analyze the performance for signals with noise. For these
signals, the features are obtained and the training and testing data set is generated. The
classification accuracy obtained based on these data sets are illustrated in Tables 2
and 3.
From Tables 2 and 3, it is evident that the combined feature extraction method has
better classification accuracy in comparison with S transform based feature extraction
method. For the signals with noise, the combined feature extraction method performs
better. The S transform technique performs better for sag and swells signals with
30 dB noise. The analysis of the proposed method with Support Vector Machine
(SVM) classifier is presented in Tables 4 and 5, respectively. The result indicates
Table 2 Training accuracy for the combined feature extraction method

Signal Without noise 50 dB noise 40 dB noise 30 dB noise
S transform Hybrid S transform Hybrid S transform Hybrid S transform Hybrid
features features features features
PQD1 99 100 99 100 100 100 96 100
PQD2 100 100 99 100 100 100 99 100
PQD3 100 100 100 100 100 100 100 100
PQD4 100 100 100 100 100 100 90 100
PQD5 100 100 99 100 99 100 99 100
PQD6 98 100 99 100 100 100 95 100
PQD7 100 100 99 100 100 100 81 100
PQD8 100 100 100 100 99 99.15 100 100
Table 3 Testing accuracy for the combined feature extraction method

Signal Without noise 50 dB noise 40 dB noise 30 dB noise
S transform Hybrid S transform Hybrid S transform Hybrid S transform Hybrid
features features features features features features features features
PQD1 95 100 93 100 80 100 93 90
PQD2 100 100 100 100 100 100 77 100
PQD3 100 100 100 100 100 100 98 100
PQD4 100 100 100 100 100 100 97 94.11
PQD5 100 100 100 100 100 98.14 100 100
PQD6 100 100 100 98.14 100 100 100 100
PQD7 100 100 100 100 95 100 95 100
PQD8 100 100 98 100 90 100 98 100
that the classification accuracy of both ELM and SVM classifiers has improved with
hybrid feature selection.
Table 4 Training accuracy with different classifiers

Feature selection ELM (%) SVM (%)
S transform 98 97.5
Hybrid features 99.83 99.42
Table 5 Testing Accuracy with different classifiers

Feature selection ELM (%) SVM (%)
S transform 94 94
Hybrid features 99.05 98.5
6 Performance of Classifiers with Real Time and Dataset
The performance of the proposed hybrid feature extraction technique is evaluated

with the power quality event data set available in national database repository of
power system events (https://pqmon.epri.com/). From the database data set with
respect to events sag, oscillatory transient and interruption are taken and given as
input to the proposed algorithm employing feature extraction and ELM classifier.
Figure 4 illustrates the waveform of the events obtained from the data set. The
accuracy of classification results are illustrated in Table 6 and the results signify the
accuracy of hybrid feature extraction in classification of power quality events.
Site0010 Site0006
Event 2787 - 2005-07-27 21:55:35.6620 Event 3557 - 2006-02-12 20:29:36.9340
12000
10000
10000
8000
8000
6000 6000
4000 4000
INST AN VOLTS
INST AN VOLTS
2000 2000
0 0
-2000 -2000
-4000
-4000
-6000
-6000
-8000
-8000
-10000
-10000
-12000
0 200 400 600 800 1000 1200 1400 0 100 200 300 400 500 600 700 800 900
Time (s) Time (s)
Electrotek Concepts® PQDiffractor® Electrotek Concepts® PQDiffractor®
(a) (b)
Site0010
Event 2857 - 2006-05-19 19:44:57.5420
3000
2500
2000
1500
1000
INST AN VOLTS
500
0
-500
-1000
-1500
-2000
-2500
-3000
0 200 400 600 800 1000 1200
Time (s)
Electrotek Concepts® PQDiffractor®
(c)
Fig. 4 Waveforms obtained from the national database repository of power system events. a Sag
waveform. b Oscillatory transient. c Interruption
Table 6 Classification with respect to real-time data

Events Total Number of correct classification
Sag 9 8
Oscillatory transient 2 1
Interruption 1 0
7 Conclusion
This work investigates the performance of hybrid feature extraction method

combining S transform and Hilbert transform for classifying single and combined
power quality disturbances. ELM is used as the classifier for classifying the distur-
bances. The analysis illustrates that the hybrid feature extraction technique improves
the classification accuracy. The proposed approach performs better for noisy signals
which have been a major concern in the area of power quality disturbance classifica-
tion. The hybrid feature extraction technique improves the classification accuracy of
both ELM and support vector machine classifier to indicate that this technique can
satisfactorily classify power quality disturbances. The performance of the proposed
feature extraction techniques is tested on real-time power quality events obtained
from national database repository of power system events and the classification accu-
racy from these data sets highlight the efficacy of the proposed feature extraction
technique. Future work will focus on optimization of features using evolutionary
computation algorithm.
References
1. Ahsan MK, Pan T, Li Z (2018) A three decades of marvellous significant review of power
quality events regarding detection & classification. J Power Energy Eng 6:1–37
2. Moravej Z, Banihashemi SA, Velayati MH (2009) Power quality events classification and
recognition using a novel support vector algorithm. J Energy Conv Manag 50:3071–3077
3. Uyar M, Yildirim S, Gencoglu MT (2008) An effective wavelet-based feature extraction method
for classification of power quality disturbance signals. J Electr Power Syst Res 78:1747–1755
4. Behera HS, Dash PK, Biswal B, Power quality time series data mining using S-transform and
fuzzy expert system. J Appl Soft Comput 10:945–955
5. Suja S, Jerome J (2010) Pattern recognition of power signal disturbances using S Transform
and TT transform. J Electr Power Energy Syst 32:37–53
6. Jayasree T, Devaraj D, Sukanesh R (2010) Power quality disturbance classification using Hilbert
transform and RBF networks. J Neurocomput 73:1451–1456
7. Abdel-Galil TK, El-Saadany EF, Salama MMA (2004) Online tracking of voltage flicker
utilizing energy operator and Hilbert transform. IEEE Trans Power Delivery 19(2):861–867
8. Sahani M, Mishra S, Ipsita A, Upadhyay B (2016) Detection and classification of power
quality event using hybrid wavelet-Hilbert transform and extreme learning machine. In: 2016
international conference on circuit, power and computing technologies, pp 1–6
9. Stockwell RG, Mansinha L, Lowe RP (1996) Localization of the complex spectrum: the S-
transform. IEEE Trans Signal Process 44:998–1001
10. Rodríguez JA, Aguado F, Martín JJ, López F, Mu˜noz JE, Ruiz (2012) Rule-based classification
of power quality disturbances using S-transform. J Electr Power Syst Res 86:113–121
11. Vidhya S, Kamaraj V (2017) Particle swarm optimized extreme learning machine for feature
classification in power quality data mining. Automatika 58(4):487–494
12. Huang G-B, Zhu Q-Y, Siew C-K (2006) Extreme learning machine: theory and applications. J
Eurocomput 70(1–3):489–501
13. Abd-el-Maleka MB, Abdelsalam AK, Hassana OE (2018) Novel approach using Hilbert trans-
form for multiple broken rotor bars fault location detection for three phase induction motor.
ISA Trans 80:439–457
14. Ahila R, Sadasivam V, Manimala K (2015) An integrated PSO for parameter determination and
feature selection of ELM and its application in classification of power system disturbances. J
Appl Soft Comput 32:23–37
15. Mishra M (2019) Power quality disturbance detection and classification using signal processing
and soft computing techniques: a comprehensive review. Int Trans Electr Energy Syst 1–42
Low-Cost, IOT-Based Child Safety
Monitoring Robot with User-Friendly
Mobile App
Kalyan Kasturi, Rajani Dharanikota, Khaleelu Rehman,

Senthilkumar Meyyappan, and Akhil Kommineni
Abstract Current population trend in India is moving from single-income to dual-

income families, and there is an increasing need for parents to remotely monitor the
safety of their children, while they are busy at work. In this research work, we propose
a low-cost child safety monitoring robot based on IOT technology that can perform
real-time live video streaming of the children and send notifications about child
safety to the parents. Using this low-cost, IOT-based child safety monitoring robot,
parents can monitor their child’s activities in real time through live video feed by
controlling the robot’s movement in real time from anywhere using a user-friendly
mobile app. In addition to this, the parents will get alerted automatically through
notifications in case of any gas leakage or fire accident. The main advantage of the
proposed system is instead of using mobile network (GSM) we are using Webhooks
for sending alerts in case of gas leakage or fire accident which reduces the cost
of hardware and works perfectly when connected to the home network. The entire
IOT-based child safety monitoring system developed in this research paper makes of
low-cost, common electronic components and hence provides a cost-effective IOT
system for child safety monitoring.
Keywords NodeMCU · Arduino · Real-time database · Webhooks · Wireless

camera · Child safety and child monitoring
1 Introduction
The Internet of Things (IOT) technology consists of a network of connected objects

and modernizes the human society in several aspects [1]. The Internet of Things
has become a popular technology bandwagon with multitudes of applications in
wide-ranging scientific fields such as smart cities, healthcare systems, automobile
systems and financial managements systems. The IOT technology is also used in the
development of smart houses [2]. Most of the connected objects in the IOT network
K. Kasturi (B) · R. Dharanikota · K. Rehman · S. Meyyappan · A. Kommineni

ECE Department, Nalla Malla Reddy Engineering College, Hyderabad 500088, India
e-mail: kalyankasturi.ece@nmrec.edu.in
https://doi.org/10.1007/978-981-19-6913-3_35
504 K. Kasturi et al.
are based on some kind of microcontroller device and used to collect information
and perform monitoring in houses or offices [3].
There has been research work conducted in the area of security and surveillance
based on IOT technology. The general idea of research work in this field is to employ
a digital camera which is connected to the Internet. The digital camera connected
to the Internet or called as a web camera serves as a connected object which can be
used to perform some form of surveillance in some particular location of interest on
a particular subject or person of interest. The web camera connects to the Internet
which serves as the backbone network of the IOT system. Zafar et al. [4] investi-
gated about building a real-time monitoring system with the help of IOT devices
and cloud technology in their research work using Arduino and cloud service. The
research work performed by Deshmukh et al. [5] discussed the use of Internet-based
camera or webcam technology for the purpose of intelligent surveillance. Ramakr-
ishna and Swathi [6] discussed using IOT technology for developing a smart security
surveillance system.
In addition to the surveillance research, another related area of interest in home
security or safety is the detection of gas leakage system. In the country of India,
in which millions of households totally and completely depend on the LPG gas for
cooking their food materials on a daily basis, the gas leakage detection is a topic
of significant interest. Gupta [7] discussed developing a gas leakage detection and
alerting system. Sharma [8] developed a microcontroller-based LPG gas leakage
detector which sends alert with the help of GSM module in case of any gas leakage.
In this research paper, we develop a low-cost IOT-based child safety monitoring
robot that performs video surveillance of children present in the house and generates
live video feed which can be monitored by their parents who are present at office or
some other location outside their home. We also developed a complete IOT-based
child monitoring as well as safety system, by performing gas leakage detection
and flame detection. The low-cost IOT-based child safety monitoring robot can be
very conveniently monitored by the parents using a simple mobile app, which was
developed as part of the proposed research work.
2 Model of IOT-Based Child Safety Monitoring Robot
This research work is aimed at developing a low-cost, IOT-based robot which can be
controlled remotely from anywhere by using an android application to help parents
to monitor their child’s activities. Nowadays, mostly CC cameras are being used for
this monitoring purpose. But children will not stay at a single location or room always
in the house. Fixing cameras at each and every location in the house is practically not
possible, and also, it is a costly process, and maintaining all those cameras is a tough
task for the parents. To overcome all these problems, in this research work, we propose
low-cost, IOT-based robot which is capable of monitoring the child’s activities and
automatically alerting the parents in case of any danger in the house, especially near
the surroundings of the child. This low-cost, IOT-based robot can be used by the
Low-Cost, IOT-Based Child Safety Monitoring Robot … 505
parent to monitor their child’s activities whenever they want from anywhere like
office, workplace, etc., using an android application which will be user-friendly and
can be used by anyone with basic knowledge of using smartphones.
Currently, for alerting the user in case of fire detection and gas leakage detection,
GSM-based systems are being widely used. GSM-based alerting systems consist of
SIM card which is used for sending alerts in case of any accident. The main purpose
of low-cost, IOT-based robot developed in this research work is for indoor use. GSM-
based alerting systems may not work properly due to various reasons like low signal
strength. So, in case of any danger in the house if there is no proper signal strength
then there may be a delay in sending the alert to the parent. This delay may result
in a great danger as we know fire and LPG leak can cause huge danger in confined
places like houses if no proper action is taken as soon as possible.
To overcome all these problems, we are using Webhooks in place of GSM-based
alerting system. Webhooks are automated messages sent from apps when something
happens. They have a message—or payload—and are sent to a unique URL—essen-
tially the app’s phone number or address. Webhooks are almost always faster than
polling and require less work on your end. They are much like SMS notifications.
So, when something like gas leakage occurs, the sensor will send the alert to the
NodeMCU. As we know NodeMCU is an IOT development board which is capable
of connecting to the Internet and capable of sending and receiving the data using
HTTP requests. In this research paper, we are making use of cloud computing to
connect both the android application and robot virtually through Internet. So, this
process makes the possibility of controlling the robot from anywhere with the help of
Internet. A block diagram of the low-cost, IOT-based child safety monitoring system
is as shown in Fig. 1.
Initially, both the robot and the user’s smartphone in which the controller appli-
cation is installed should be connected to the Internet. Once the user initiates a
command, then that particular instruction will be first sent to the real-time database.
Real-time database works in such a way that if any changes occur in the existing
data present in the database, then the new data will be updated in the configured
application or device. Here, the real-time database is connected to the NodeMCU.
So, as soon as the command is initiated by the user, that particular command will be
sent in real time to the NodeMCU which is one of the main components of the child
safety monitoring robot.
From the NodeMCU, the data will be transferred to Arduino UNO via serial
communication. Then, based on the instruction received, Arduino UNO sends the
control signals to the L298N motor driver module which controls the movement of
DC motors. The above-mentioned process completes within seconds which makes
it possible to control the robot’s movement in real time.
Wireless camera will be mounted on the dashboard of the robot. The camera will
be configured with home WIFI network. So, the user can watch the live video feed in
order to monitor the child’s activities as well as observing the surroundings in order
to control the robot’s movement.
Fig. 1 Block diagram representation of the proposed low-cost, IOT-based child safety monitoring
robotic system
For the detection of fire and gas leakage, two different sensors are being used:
MQ5 sensor for gas leakage detection and IR flame sensor for fire detection. Both the
IR flame sensor and MQ5 sensor will be directly connected to the NodeMCU. In case
of gas leakage or fire detection, sensors will send the data to NodeMCU. NodeMCU is
preconfigured in such a way that in case of fire or gas leakage detection a particular set
of commands must be executed. The preconfigured logic consists of a code to send a
GET request to the preconfigured Webhook URL. GET is one of the HTTP methods
used to communicate with the servers. GET is used to retrieve and request data
from a specified resource in a server. GET is one of the most popular HTTP request
techniques. In simple words, the GET method is used to retrieve whatever information
is identified by the request-URL. Once the logic is triggered, a GET request will be
initiated to the Webhook URL. A Webhook URL works in such a way that if it
is triggered then an alert will be automatically sent to preconfigured email address
or phone number with the preconfigured alert message like “FIRE DETECTED”
or “GAS DETECTED”. Using Webhooks instead of GSM-based alerting system
reduces the cost as well as size and weight of the hardware.
3 Components and Technologies Employed in Child Safety

Monitoring Robotic System
In this research work, we are making use of NodeMCU [9] for connecting Arduino
UNO with the real-time database because Arduino UNO [10] itself does not have
any inbuilt hardware to connect to the Internet. NodeMCU is a low-cost open-source
IOT platform. It initially included firmware which runs on the ESP8266 Wi-Fi SOC
from Espressif Systems and hardware which was based on the ESP-12 module.
We are making use of Arduino UNO for sending the control signals to motor
driver module in order to control the movement of DC motors. L298N driver module
[11] is employed to control the movement of robot in different directions based on the
instructions received from the Arduino UNO microcontroller. MQ5 gas sensor and
IR flame sensor are employed for detecting gas leakage and fire accident. An android
application which is used to control the movement of the robot is developed using
MIT app inventor. MIT app inventor is an intuitive, visual programming environment
that allows everyone to build fully functional apps for smartphones. We used Google
Firebase, a real-time database, to transfer the instructions from android application
to the robot in real time.
4 Demonstration Kit
The entire hardware components for the child safety monitoring robotic system were
connected together to develop the demonstration kit. Figure 2 illustrates the side view
of the demonstration kit showing the important hardware parts.
We can observe in Fig. 2 that the Arduino UNO board which is the heart of the
child safety monitoring robotic system can be located at the center of the kit. Wireless
camera which presents the live video feed for child monitoring is fixed at the front
side of the robot’s chassis for the purpose of monitoring as well as guiding the user
in order to move the robot and can be seen on the far right-hand side of the picture.
A top view of the demonstration kit is shown in Fig. 3. The motor drivers which
control the DC motors used to move the child safety monitoring robot in various
directions are connected to the kit, as shown in the picture. Remaining hardware
components such as MQ5 gas sensor and IR flame sensor are placed on the chassis
as shown in Fig. 3. Power supply for all the hardware components is given through
onboard rechargeable battery. The entire cost of all components used in the child
monitoring system is less than Rs. 5000, and hence, this child monitoring system
can be built with very less cost.
Fig. 2 Side view of the demonstration kit of the low-cost, IOT-based child safety monitoring robot
5 Experimental Results
If a parent who is at some location other than home such as workplace or office wants
to check the status of their child who was alone in the house, then the parent can start
the child safety monitoring by launching the android application and corresponding
application for wireless camera which were already installed in his/her smartphone.
The android mobile application is developed in a user-friendly approach so that by
clicking the left/right/front/back arrows on the mobile app the parent can move the
IOT-based child safety monitoring robot as per their requirements and the parent can
watch the child’s activities in real time.
The working of the low-cost, IOT-based child safety monitoring robot was
successfully verified by moving the demonstration kit presented in Figs. 2 and 3
in various directions and collecting the live video feed of the child present in the
home in real time. In Fig. 4, the live video feed of child collected in real time by
the low-cost, IOT-based child safety monitoring robot is presented. In Fig. 4, we can
observe the child who is seated, and in addition, we can note the time and date stamp
at the top, right-hand corner of the picture.
In case of fire accident or gas leakage detection near the child, an alert notification
will be automatically sent to the parent through SMS or email based on their prefer-
ence. Figure 5 shown below illustrates the notification which will be sent to the parent
in case of fire accident. We can clearly observe that the fire detection message was
sent by seeing “What: fire_detected” header displayed in the picture. From Fig. 5,
we can also observe that the fire detection message was sent using Webhooks. In
Fig. 3 Top view of the

demonstration kit of the
low-cost, IOT-based child
safety monitoring robot
Fig. 4 Live video feed of

child with time and date
stamp of the low-cost,
IOT-based child safety
monitoring robot
Fig. 5 Fire detection message using Webhooks delivered by the demonstration kit of the low-cost,
IOT-based child safety monitoring robot
addition, we can observe the time and date stamp of fire detection notification in
Fig. 5.
6 Conclusion
Currently, IP cameras are available for child monitoring purpose. It is not possible
to fix cameras at each and every location in the house. Installing cameras at each and
every location in the house is costly process, and also, it is not possible to monitor
all the cameras. So, by using this low-cost IOT-based child safety monitoring robot,
parents can monitor their children’s activities from their workplace, etc., and will
get alerted automatically in case of gas leakage or fire accident. Once the robot is
connected to the Internet, it can be controlled from any location such as office, work-
place and almost anywhere in the world. Parent can control the robot’s movements
by operating the left/right/front/back arrows on the mobile app and can move the
robot in all directions. The parent can watch the live video feed through the onboard
IP camera. Also, the child monitoring system can be built with very less cost. This
low-cost child monitoring system will be very useful for the parents whose children
will reach home from schools before their parents arrive from their workplace.
References
1. Atzori L, Lera A, Morabito G (2010) The Internet of Things: a survey. Comput Netw
54(15):2787–2805
2. Kasturi KS, Reddy PV, Rao AN, Vinod S (2016) A review of architecture and applications for
Internet of Things. Adv Nat Appl Sci 10(9):261–266
3. Kasturi K, Sri Mourya K, Mahendra I, Maheshwari V (2017) Gesture recognition based device
control using MEMS accelerometer. Int J Eng Sci Adv Comput Bio-Technol 8(4):256–264
4. Zafar S, Miraj G, Baloch R, Murtaza D, Arshad K (2018) An IoT based real-time environmental
monitoring system using Arduino and cloud service. Eng Technol Appl Sci Res 8(4):3238–3242
5. Deshmukh A, Wadaskar H, Zade L, Dhakate N, Karmore P (2013) Webcam based intelligent
surveillance system. Int J Eng Stud 2(8):2278–4721
6. Ramakrishna U, Swathi N (2016) Design and implementation of an IoT based smart security
surveillance system. Int J Sci Eng Technol Res 5(4):697–702
7. Gupta A (2017) Economical and optimal gas leakage detection and alert system. Int J Sci Res
Publ 7(11):260–263
8. Microcontroller Based LPG Gas Leakage Detector using GSM Module: Engineers Garage.
https://www.engineersgarage.com/contributions/microcontroller-based-lpg-gas-leakage-det
ector-using-gsm-module/. Last accessed 03 Feb 2021
9. NODEMCU—A perfect board for IOT. https://www.circuito.io/blog/nodemcu-esp8266/. Last
accessed 16 May 2021
10. Arduino UNO. https://www.arduino.cc/en/Main/arduinoBoardUno&gt. Last accessed 16 May
2021
11. L298N Driver Module. https://components101.com/modules/l293n-motor-driver-module/.
Last accessed 17 May 2021
Secure Image Classification Using Deep
Learning
K. Gururaj, Alaka Ananth, and Sachin S. Bhat
Abstract Machine learning and security are the buzzwords these days. Just like
other fields, privacy concern is a major issue in machine learning systems as well.
Current privacy techniques focus on allowing multiple input parties to collaboratively
train machine learning models without releasing their private data in its original form.
One of the most sensitive data in this regard is medical images. Usage of such data
for collectively training models might be against the policies of hospitals, which
assure patients that their information would be kept confidential. In such a scenario,
privacy preserving machine learning poses several advantages over the conventional
methods. In this paper, we have implemented a secure machine learning model based
on the multi-party protocol described in SecureML (Mohassel and Zhang in 2017
IEEE symposium on security and privacy. IEEE, pp 19–38, 2017, [1]), on the medical
dataset of X-ray images for pneumonia. The performance of these privacy preserving
techniques against conventional machine learning algorithms is evaluated.
Keywords Image classification · Neural network · Linear regression · X-ray

images
K. Gururaj · A. Ananth
NMAM Institute of Technology, Nitte, India
S. S. Bhat (B)
Shri Madhwa Vadiraja Institute of Technology and Management, Bantakal, Udupi, India
e-mail: sachinbhat88@gmail.com
https://doi.org/10.1007/978-981-19-6913-3_36
514 K. Gururaj et al.
1 Introduction
Machine learning (ML) algorithms are widely popular in today’s applications. They
are commonly used to produce predicting models for use in pharmacy, accounting,
advisory systems, and vulnerability detection. Machine learning aims to transform
industries through its novel and impactful applications. Machine learning involves
manipulating huge quantities of training data. Consequently, large amount of data
collection leads to the problem of its privacy. With rising concerns of digital breaches
and security concerns, ensuring privacy of data in machine learning is an important
concern. For instance, machine learning models that predict human diseases using
images require sensitive medical data. While these algorithms indeed benefit humans,
they require access to data that are usually private to hospitals. In such scenarios,
hospitals, legal rules, and restrictions may prohibit them from outsourcing their
medical data. With the standardization of newer and better machine learning tech-
niques, there is a need to integrate data privacy with machine learning arises, which
led to the concept called privacy preserving machine learning. Privacy preserving
ML through secured multi-party computation (MPC) is a propitious solution that
enables distinct entities to train different models on their combined data without
revealing any information beyond the outcome. The leakage of data used for training
could be potentially leaked by the models, and this has been studied extensively [2–
4]. However, the assertions given by privacy preserving machine learning provide a
first line of defense that can be supported by integrating with other privacy-related
methods such as differential privacy [5, 6].
Recently, there have been data collection schemes that need privacy as a necessary
feature owing to the sensitivity of the data. The most common setting for privacy
preserving machine learning is one where clients and servers interact in the following
way: Multiple clients encrypt their data and send it to the server, which performs
training on the combined data to classify new data samples. This being said, conven-
tional training algorithms do not incorporate privacy inherently. Therefore, privacy
preserving machine learning solutions aim at optimizing this gap [7–10]. In the last
few years, privacy (especially differential privacy) has acquired popularity as a formal
and quantifiable measure of privacy risk in data publishing [11, 12].
The most common scenario considered in these privacy preserving techniques is
the server-aided model. Here, owner (otherwise known as clients) sends encrypted
data to multiple servers which perform the training procedure on combining similarly
sent data. There are also techniques where the server has a pre-trained model and
uses it to classify new data samples.
In this paper, we provide an implementation of a secure deep learning model using
PySyft [13] that uses convolutional neural networks to recognize pneumonia in chest
X-ray images [14]. We provide evaluation of both these models, and present graphs
of variation of the model accuracy against different parameters.
Remaining part of the paper is organized as follows. Section 2 provides an
overview of the concepts as well as the preliminaries need to understand the rest
of the sections. We give a theoretical background to secure machine learning in
Secure Image Classification Using Deep Learning 515
Sect. 3. Section 4 describes our work and implementation of two secure models.
Analyzed results are presented in Sect. 5. We conclude in Sect. 6.
2 Background
The problem of secure computation [16] has been well studied in cryptography. In
the setting of multi-party computation, two or more parties engage in a protocol.
The parties each have their inputs, and agree on a common functionality that accepts
these inputs and produces an output. For example, in the widely popular problem of
multi-party private set intersection, the parties’ inputs are sets of elements, and the
functionality computes the intersection of these sets. The goal of any Secure MPC
protocol is to achieve two main concerns: the protocol should output the correct value
according to the defined functionality, and each party should learn no additional
information than what can be inferred from its own input and the obtained output.
The main focus in this paper is on the semi-honest model, also otherwise known
as the honest-but-curious model. Here, the parties are assumed to follow the speci-
fied Secure MPC protocol, but they may try to learn additional information from the
protocol transcript. Semi-honest security in the context of machine learning captures
a realistic scenario, since for currently used popular machine learning libraries, mali-
ciously modifying the code is prevented through software attestations during deploy-
ment. Therefore, all that the parties can hope to obtain is the information received
during the protocol, which mainly constitutes the transcript.
Oblivious transfer (OT) is a commonly used cryptographic functionality which
is widely used in MPC protocols. Here, a sender S has two strings s0 and s1 , and
the receiver has a choice bit b. The OT protocol allows the receiver to obtain sb
without gaining any information about s1−b . Moreover, the sender does not gain any
information about b. Multiple oblivious transfers can be executed efficiently with a
reduced cost using the OT extension technique [17]. OT protocols can be applied
to realize any secure functionality using Yao’s garbled circuit approach [18]. This
approach provides a generic mechanism for constructing a semi-honest secure two-
party protocol to compute a functionality f , using its Boolean circuit representation.
In a garbled circuit protocol, one party (a circuit-generator) generates an encrypted
version of the Boolean circuit for f . The other party (the circuit-evaluator) obliviously
computes the output of the circuit without learning any intermediate values.
3 Privacy Preserving Machine Learning with SecureML
In 2017, Mohassel and Zhang proposed SecureML [1], a privacy preserving machine
learning technique that designs efficient protocols for various types of regressions
and artificial neural networks. They consider the two-server model, where parties
distribute their sensitive data among two servers S 1 and S 2 . It is assumed that the two
servers do not collude, and thus, train models jointly using two-party MPC protocols.
3.1 Server Architecture
There are n parties P1 , P2 … Pn , who wish to collectively train their data. This data
may be sensitive, and the parties (also referred to as clients in context of a client–
server architecture) do not wish to share their private data information to other parties.
The work in [1] does not make any assumption about the distribution of client data
and can also be output shares of a previous MPC protocol’s output.
For reasons of efficiency, the servers outsource their data to remote servers, who
are more computationally capable than clients. This scenario has been widely used in
machine learning-based applications today, where clients take the role of lightweight
mobile devices who outsource computation to servers.
An advantage of a server-aided approach is that clients can send their data (as
secret shares) to servers in a setup phase that is independent of the training phase.
In our implementation described in Sect. 4, we focus on the two-server model,
where the servers are assumed to not collude (i.e., there is no common semi-honest
adversary A that simultaneously corrupts both the servers).
3.2 Secret Sharing
In the SecureML protocol, the two servers exchange information between them
as secret shares. Because of the assumption that the two servers never collude, it
is impossible for one server to obtain information about the data with the other
server and any of the clients. Three different secret sharing techniques are used:
Boolean, Additive, and Yao shares. A finite field is used to generate additive shares. To
additively secret share a field element a ∈ Z 2 l, a server chooses a random a0 ∈ Z 2 l, and
sends a1 = a a0 mod 2l . The shares can be added and multiplied using corresponding
operations over the finite field, as described in [1]. Similar to arithmetic sharing,
Boolean shares can be obtained through similar operations in Z 2 . In this case, addition
of shares corresponds to bit-wise XOR, and multiplication of shares corresponds to
bit-wise AND. Finally, Yao shares are generated based on the encrypted values of
wires present in the garbled circuit [18].
The advantage of having three different shares is that certain type of shares is
used to carry out specific operations very efficiently. For instance, if two arithmetic
values need to be multiplied, doing so with Boolean shares is very inefficient. With
arithmetic sharing, a multiplication of the shares efficiently results in shares of the
multiplication of the arithmetic values in the field.
3.3 Definition of Security
As mentioned in Sect. 2, we assume that the adversary A is semi-honest, meaning that

the parties corrupted by A follow the specified protocol without deviating from it.
However, any information received during the protocol is analyzed by the corrupted
parties to obtain additional information. Although the two servers do not collude, an
adversary can corrupt a server and a subset of the clients. The security requirement
states that the adversary should learn only the inputs of the parties it has corrupted
and should obtain no information about the other parties’ data than what is implied
by the output.
This definition is attributed to the real-world and ideal-world models defined in
multi-party computation [16]. In an ideal world, a trusted third party exists that cannot
be corrupted and does not reveal information to any of the parties. In such a case, the
parties send their inputs to this trusted third party over a private channel. The third
party locally computes the desired functionality, and outputs them to each party. In
this world, any additional information obtained by the clients is a function of its own
input and the local output. In the real world, where parties can be corrupted, a secure
MPC protocol should be designed so that the corrupted parties learn no more than
what can be obtained from their collective inputs and the obtained protocol output.
This protocol, therefore, would “simulate” the ideal world, and the MPC protocol is
said to securely evaluate the desired functionality.
In the context of privacy preserving machine learning, the functionality to be
evaluated should train the required model in our implementation. Therefore, the
execution of the secure MPC protocol should not leak any information to the adver-
sary, regarding the model parameters and the input dataset of honest clients used for
training.
4 Implementation of the Two-Server Model Over the Chest

X-ray Dataset
We provide implementation of secure machine learning models in the semi-honest

two-party setting. This model is a protocol for deep convolutional neural networks
implemented using the PySyft library [13] on Pneumonia X-ray images [14].
We highlight the differences in performance of privacy preserving algorithms
against ideal (non-secure) machine learning algorithms in Sect. 5. In the remainder
of this section, we discuss on the implementation architecture and the code
optimizations used to make the protocols more efficient.
We implemented privacy preserving machine learning model on chest X-ray
images [14] to recognize pneumonia. This dataset, being a collection of medical
images, has more features and pixel-information in each image. Using a simple tech-
nique such as linear regression would perform poorly on this data. We, therefore,
decided to use a more complicated machine learning technique using neural networks.
More specifically, we use convolutional neural networks (CNNs) to train over this
data. The advantage is that these type of neural networks are great at capturing
localized information in images, which greatly improves accuracy.
However, the implementation for a neural network is highly complicated, with
multiple back-propagated derivatives at each node. Hence, a native C++ implemen-
tation posed many difficulties. We decided to use a pre-built Python library to achieve
this task. We use PyTorch (a popular machine learning library for Python) as well
as the PySyft [13] to implement a convolutional neural network. PySyft is a library
for secure, private machine learning. It provides efficient implementations of many
privacy preserving machine learning models. The setting used in our implementation
is again the two-server model, where the servers do not collude.
The general flow of the PyTorch script is as follows. Two servers are present
that hold information corresponding to a model that has been trained on a specific
dataset. To begin with, the server characterizes and trains a model with this private
training data (securely, using a privacy preserving machine learning technique). At
this point, the server connects with a client C, who holds its own portion of the
information related to the dataset, and wishes to obtain the server’s model and make
a few predictions.
The server encrypts its trained model (which is a neural network in this case). The
client encodes its information which it requires to query on the server’s model. At this
point, both the server and client use the encrypted information present with them (the
trained model and the encoded query, respectively). Finally, the server performs the
query computation securely and this result is sent back to the client in an encrypted
manner. This query is sent back in such a way that the server infers nothing about
the client’s query data—neither the sources of information nor the final prediction
results. The computation is secure in the honest-but-curious adversary model which
is standard in many MPC frameworks as well.
The structure of our implementation is as follows. We first split the dataset into
three sets, the training, validation, and test sets. We also initialize two workers (corre-
sponding to two servers training the data). This is followed by a data share function
that sends data to these workers thereby splitting the dataset into nearly equal sets.
A traditional convolutional neural network is used to train the medical image data.
Two 2D convolutional layers (of sizes 3 × 20 × 5 and 20 × 50 × 5) are implemented
in the neural network, followed by two fully connected (or linear) layers. We also
used two rectified linear units (ReLUs) as the activation functions along with two
layers for max-pooling. In PyTorch, the most optimal way to implement a network
is to write a separate class. The class contains an initialization function as well as
forward function with relevant dimensions. The hyperparameters used to train the
model are given in Table 1.
The training of the dataset is then launched. Now, the server securely sends the
model to the workers holding the data. Because the model contains sensitive infor-
mation, the server does not wish to disclose its weights. So, we use secret sharing
(described in Sect. 3.2) to share the model among the workers (who are assumed to
not collude) without revealing any information. This is followed by a testing function
that performs an encrypted evaluation. The model weights, data inputs, prediction,
Table 1 Hyperparameters
Hyperparameters Selection
fine-tuned
Input 224 × 224
Activation function ReLu
Optimizer SGD
Error rate Categorical cross entropy
Learning rate 0.0005–0.01
Kernel size 3×3
Momentum 0.9
Epochs 100
Batch size 32–512
Pooling Max-pooling
Padding Zero padding
Searching algorithm Randomized search
and the target used for scoring are all encrypted, thereby revealing no additional
information to the adversary.
5 Results
For the secure convolutional neural network, we used the PySyft [13] library for the
popular machine learning framework PyTorch. We applied different techniques such
as max-pooling, SGD optimizers, and ReLU layers to find the optimal hyperparame-
ters for the network. This secure model was trained on the medical dataset containing
X-ray images to detect Pneumonia. Since we use a convolutional neural network,
we decided to move from a simpler dataset such as MNIST, to a more complicated
medical image-based dataset [14]. This allows our model to capture localized infor-
mation more effectively and identify complex patterns that simple linear regression
could not. The experiments were carried out on an Intel×86 processor running on
Linux environment.
We implemented a convolutional neural network to securely train a model for the
X-ray medical datasetto detect pneumonia [15]. Figure 1 shows a few images from
this dataset.
As described in Sect. 4, the network architecture for the CNN consists of two 2D
convolutional layers (of sizes 3 × 20 × 5 and 20 × 50 × 5), two linear layers, and
two max-pooling layers. The ReLU activation function was used in PyTorch, along
with the SGD optimizer.
Fig. 1 X-ray images [14] to detect pneumonia
We obtained several graphs in 3 to compare the behavior of accuracy of the trained

model against different parameters.
Figure 2a shows the variation of accuracy against the number of epochs. The
accuracy slightly increases in the beginning as the number of epochs’ increases, but
it later converges to a constant value. Figure 2b shows the variation of accuracy against
the batch size used for training. The optimal batch size to obtain highest accuracy
was found to be 128. Figure 2c shows the variation of accuracy with learning rate.
As the learning rate increases, the accuracy of our secure neural network decreases.
This can be attributed to an increase in aggressiveness of the optimization equation.
With a higher learning rate, the model overshoots the global minimum and tends to
oscillate about that point. This leads to a decrease in overall accuracy. We found that
a learning rate of 0.0005 was optimal to obtain maximum accuracy.
6 Conclusion
In this work, we implemented practical algorithms for privacy preserving machine

learning techniques. Specifically, we used the convolutional neural network to
securely classify X-ray images to detect pneumonia. Using the PySyft library [13],
we analyzed the practicality of using complex machine learning algorithms in the
context of secure multi-party protocols. The performance of this privacy preserving
technique was evaluated against conventional machine learning algorithms.
Fig. 2 a Accuracy versus epochs, b accuracy versus batch size, c accuracy versus learning rate, d
total execution time versus iterations, e setup time versus iterations, f online time versus iterations
Fig. 2 (continued)
References
1. Mohassel P, Zhang Y (2017) Secureml: a system for scalable privacy-preserving machine

learning. In: 2017 IEEE symposium on security and privacy. IEEE, pp 19–38
2. Tramèr F, Juels A, Reiter MK, Ristenpart T (2016) Stealing machine learning models via
prediction Apis. In: Proceedings of the 25th USENIX conference on security symposium,
SEC’16, (USA). USENIX Association, pp 601–618
3. Song T, Ristenpart T, Shmatikov U (2017) Machine learning models that remember too much.
In: Proceedings of the 2017 ACM SIGSAC conference on computer and communications
security, CCS’17, (New York, NY, USA). Association for Computing Machinery, pp 587–601
4. Shokri R, Stronati M, Song C, Shmatikov V (2017) Membership inference attacks against
machine learning models. In: 2017 IEEE symposium on security and privacy (SP), pp 3–18
5. Abadi M, Chu A, Goodfellow I, McMahan HB, Mironov I, Talwar K, Zhang L (2016) Deep
learning with differential privacy. In: Proceedings of the 2016 ACM SIGSAC conference on
computer and communications security, CCS ’16, Association for Computing Machinery, New
York, NY, USA, pp 308–318
6. McMahan HB, Ramage D, Talwar K, Zhang L (2017) Learning differentially private language
models without losing accuracy. ArXiv, vol. abs/1710.06963
7. Mohassel P, Rosulek M, Zhang Y (2015) Fast and secure three-party computation: the garbled
circuit approach. In: Proceedings of the 22nd ACM SIGSAC conference on computer and
communications security, CCS ’15, Association for Computing Machinery, New York, NY,
USA, pp 591–602
8. Furukawa J, Lindell Y, Nof A, Weinstein O (2017) High-throughput secure three-party compu-
tation for malicious adversaries and an honest majority. In: Coron J-S, Nielsen JB (eds)
Advances in cryptology—EUROCRYPT 2017. Springer International Publishing, Cham, pp
225–255
9. Araki T, Furukawa J, Lindell Y, Nof A, Ohara K (2016) High-throughput semi-honest secure
three-party computation with an honest majority. In: Proceedings of the 2016 ACM SIGSAC
conference on computer and communications security, CCS ’16, Association for Computing
Machinery, New York, NY, USA, pp 805–817
10. Mohassel P, Rindal P (2018) Aby3: a mixed protocol framework for machine learning. In:
Proceedings of the 2018 ACM SIGSAC conference on computer and communications security,
CCS ’18, Association for Computing Machinery, New York, NY, USA, pp 35–52
11. Dwork C, Nissim K (2004) Privacy-preserving datamining on vertically partitioned databases.
In: Franklin M (ed) Advances in cryptology—CRYPTO 2004. Springer, Berlin, Heidelberg,
pp 528–544
12. Dwork C (2006) Differential privacy in automata, languages and programming. Springer,
Berlin, Heidelberg, pp 1–12
13. OpenMinded, “PySyft.” www.openmined.org
14. Mooney P, Chest X-Ray images. www.kaggle.com/paultimothymooney/chest-xray-pneumonia
15. Bishop CM (2006) Pattern recognition and machine learning, information science and statistics.
Springer, Berlin, Heidelberg
16. Oded G (2009) Foundations of cryptography: volume 2, basic applications, 1st edn. Cambridge
University Press, USA
17. Ishai Y, Kilian J, Nissim K, Petrank E (2003) Extending oblivious transfers efficiently. In: Boneh
D (ed) Advances in cryptology—CRYPTO 2003. Springer, Berlin, Heidelberg, pp 145–161
18. Yao C (1986) How to generate and exchange secrets. In: Proceedings of the 27th annual
symposium on foundations of computer science, SFCS ’86, IEEE Computer Society, USA, pp
162–167

Smart Sensors Measurement and Instrumentation: Shreesha Chokkadi Rajib Bandyopadhyay Editors

Uploaded by

Document Information

Original Title

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

Smart Sensors Measurement and Instrumentation: Shreesha Chokkadi Rajib Bandyopadhyay Editors

Uploaded by

Copyright:

Available Formats

Lecture Notes in Electrical Engineering 957

Smart Sensors Measurement

ISSN 1876-1100 ISSN 1876-1119 (electronic)

Kolkata, India Prof. Rajib Bandyopadhyay

Manipal, India Prof. Shreesha Chokkadi

Control Instrumentation System (CISCON) is the annual conference event organized

Muffler Transmission Loss Optimization for a Vehicle Using

System Identification, Stability Analysis and PID Controller

Automation of Weight-Based Sorting System Using Programmable

A Comparative Study Between Partial Least Squares

About the Editors

Prof. Rajib Bandyopadhyay completed bachelors, masters and Ph.D. in Electronics

Yashwant Adhav Cummins College of Engineering for Women, Pune, India

Wasim Akram Department of Instrumentation and Electronics Engineering,

Ananth Jnana Chandraraj Department of Mechatronics Engineering, Manipal

Nikhil Kadam Department of Electrical Engineering, College of Engineering Pune,

Madhuri More Department of Renewable Energy Engineering, CTAE, MPUAT,

K. Rajesh Deparment of EEE, RGM College of Engineering and Technology,

S. Sibi Selvan Department of Electronics and Communication Engineering, St.

M. Vishnu Deepika Department of Electronics and Communication Engineering,

Riziyamaalisa Gavit and Kiran Wani

Abstract Automotive reactive mufflers attenuate sound using interference which

Keywords Transmission loss · Muffler · Genetic algorithm

1.1 The Targets in the Project to Be Achieved Are as Follows

• Optimize the multichamber muffler lengths and diameters.

2.1 Acoustic Filter Design Parameters

1. Adequate insertion loss (attenuation)

2.2 Acoustic Filter Performance Parameters

Transmission loss (TL)

TL = Lw1 − Lw2 (dB) (1)

2.3 Calculation of Noise Transmission Loss by Impedance

Fig. 1 Diagram of transmission loss measurement setup using impedance tube

2.4 Genetic Algorithm (GA)

Fig. 2 Workflow of research thesis

4 Simulation and Experimentation

Table 1 Original dimensions

Computer-aided engineering simulation is done in COMSOL Multiphysics software

Fig. 3 Three-dimensional model of muffler

Fig. 4 Meshed model in COMSOL

Fig. 5 Variable used for calculation

5.1 Experimental Testing

Fig. 6 Actual experimental setup

6 Result and Discussion

6.1 Transmission Loss Experimental Results for Base

6.2 Transmission Loss CAE Simulation Results for Base

Table 2 Transmission loss

Fig. 7 Transmission loss

6.3 Modified Geometry

Effect of varying lengths of components

Fig. 8 Transmission loss

Table 3 Effect of length variation of components on TL

Table 4 Effect of diameter variation of components on TL

Effect of varying diameters of components

6.4 Curve Fitting to Obtain Polynomial Equation

Fig. 9 Combined graph of

Coefficient values are different for all six conditions.

Fig. 10 Polynomial curve for input diameter

Fig. 11 Optimized inlet length

Fig. 12 Optimized inlet diameter

Table 5 Optimization result table

Fig. 13 Transmission loss of optimized muffler in COMSOL

Nikhil Kadam and Archana Thosar

N. Kadam (B) · A. Thosar