You are on page 1of 10

Proceedings of the ASME 2015 Conference on Smart Materials, Adaptive Structures and Intelligent Systems

SMASIS2015
September 21-23, 2015, Colorado Springs, Colorado, USA

SMASIS2015-9054

STRUCTURAL ASSESSMENT AND DAMAGE IDENTIFICATION ALGORITHMS


USING BINARY DATA

Hadi Salehi1 Saptarshi Das2 Shantanu Chakrabartty2 Subir Biswas2 Rigoberto Burgueño1

1 Dept. of Civil and Environmental Engineering; 2 Dept. of Electrical and Computer Engineering

Michigan State University, East Lansing, Michigan, USA

ABSTRACT the monitored structure and analysis of the measurements for


One of the challenges in structural health monitoring (SHM) structural condition assessment. Several studies have been
is the power required for sensors to collect and communicate carried out to develop appropriate algorithms to extract critical
data. Self-powered sensors are able to harvest power by from structural information from measured data [6-8]. Many SHM
their environment, i.e., strain and vibration of the host structure. algorithms are based on natural frequency analyses that assess
However, the harvested power with current technology is still damage from changes in the natural frequencies of the structure.
limited and improving the system’s efficiency requires reducing For this purpose, traditional SHM methods such as wavelet
the power budget. A way to minimize the communication power transform or vibration control techniques deal with continuous
demand is to transmit the minimum amount of information, data that are physically based [9-12]. In other words, these
namely one bit. The binary signal can be generated at a sensor methods are based on the availability of continuous time-history
node according to a local rule based on physical measurements, data from sensors distributed over the host structure.
but interpretation at the global level requires dealing with The energy required for measuring and sensing data is a
discrete binary (1 or 0) data, which implies system information crucial challenge in SHM. The required energy budget becomes
with reduced resolution. This study presents an investigation on even more important when it is used for communication between
approaches for the interpretation of such kind of binary data for different sensor nodes. Several researchers have studied the use
use in structural assessment and damage identification. Pattern of smart materials and wirelessly networked sensors for
recognition (PR) methods based on image data analysis were embedded monitoring in SHM [13-15]. A limitation is that the
adapted for the study. The methods used were classifiers based operating life of battery-powered sensors is limited by their
on deviation of patterns with respect to each other, two- battery life. On the other hand, replacing the batteries of
dimensional principal component analysis, and two-dimensional structure-embedded sensors is major issue [16,17]. Embedded
linear discriminant analysis. The PR methods and the sensors in passive battery-less sensing system do not need
performance of the interpretation algorithms were evaluated by energy, although the disadvantage of these sensors is their
using virtual data from finite element simulations on an limited communication range [18,19]. Self-powered sensors
aluminum plate. The ability for each of the PR methods to have recently become a reality by overcoming the gap between
identify service demands, load variations and localized material the achievable scavenged energy and the energy required for
degradation was evaluated. Results indicate that PR methods can sensing, computing and communication [20,21]. Self-powered
be used as damage identification algorithms for binary data sets sensors are thus able to directly harvest the required
in novel wireless self-powered sensor networks. computational, storage and transmission power from the signal
being sensed. Chakrabartty et al. [22-24] introduced a self-
INTRODUCTION powering SHM device that exploits the properties of ceramic and
Structural health monitoring (SHM) and damage detection polymer based piezoelectric materials to generate large open-
has become an important issue in many fields such as civil, source voltage, and demonstrated that self-powered sensors are
mechanical and aerospace engineering. Although SHM is a able to use their environments to generate the demand power.
broad research area, SHM methods that can detect and localize Communication, processing, and buffering overheads of the
of damages have been the subject of much study [1-5]. SHM can large number of bits within data packets may lead to energy-
be attained by collecting data from a sensor network installed on inefficiency for event sensing with conventional communication

1 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


protocols. In this regard, the concept of pulse switching for static damage detection in an aircraft wing based on principal
event sensing in sensor networks was introduced by Biswas et al. component analysis and a simple pattern recognition method.
[25-29]. This technology consists of an energy-efficient pulse In contrast with traditional SHM techniques, this study deals
switching for ultra-weight wireless network applications, which with the development of global interpretation algorithms for
has been demonstrated as an effective means for minimizing the damage identification using discrete binary data that is not
energy communication demand in a sensor network. For this continuously available in time (Fig. 1). The type of integrated
purpose, event localization can be combined with the pulse SHM sensor considered for this study is based on the self-
switching protocol. Events can be localized by a receiver by powered sensors. The proposed damage identification
observing the temporal position of a received pulse with respect methodology is based purely on PR methods. Different PR
to a reference frame. However, pulse based communication methods based on image data analysis were examined for this
decreases the available information to a binary format and this purpose and used for data interpretation and pattern
leads to received information to be discrete and time-based identification. The methodology was evaluated using the
binary at the SHM processor. vibration response of an aluminum plate structure simulated by
Pulse switching is an energy-light data networking paradigm the finite element method.
in which the location and occurrence of a binary sensing event
can be communicated across small-to-medium scale networks of
wireless sensing devices [26]. The primary advantage of pulse
switching is the considerable energy savings in terms of
transmission energy cost when compared to other traditional
approaches such as packet switching. The energy savings arise
from the fact that pulse switching uses a minimal number of
pulses to represent the event location and forwarding
information and provides reliable event delivery characteristics Fig. 1. Discrete binary data vs. time
with no additional overhead such as preamble information. The
latter is often the largest contributor to transmission energy cost MODULATION AND PROBLEM STATEMENT
in packet switching and might turn out to be much larger than the The aim of this study is to develop global interpretation
actual information, which is binary in this case and hence only a algorithms based on discrete, time-based asynchronous binary
single bit in length. Pulse switching trades off delivery delay to data generated by self-powered sensors. For this purpose, binary
transmit the information at very low energies. It events are generated following rules based on measurements at
represents the information to be communicated as a specific a local level. Accordingly, the generated binary events were used
temporal pulse pattern with respect to a synchronized frame to interpret and investigate structural behavior at a global level.
structure thus encoding the location and forwarding information. It has been initially assumed that the system is able to operate
The higher delivery delay can be accommodated in several with full data availability, and the constraints of the
applications that do not need real-time event delivery and can communication power budget for the embedded sensors and the
accept delivery delays within a range, but where energy savings communication time delay were disregarded. In this way, the
are crucial. An energy aware version of the pulse switching proposed method can be examined for simulated binary data. A
approach has been developed [30] and shown to be especially simple finite element (FE) model of a simply supported
useful in networks of devices powered by erratic harvested aluminum plate was used to simulate its vibration response.
energy sources. The latter protocol introduces efficient power Therefore, the input to the assessment algorithm is the structural
management principles to modulate the operation of pulse response according to an FE simulation. It was assumed that the
switching in accordance with the energy availability in the sensor self-powered sensors were powered directly through strain-
devices for optimal operation in energy harvested devices. variations in the substrate and had the ability to measure strain
The interpretation of discrete binary data over a domain and vibration. Once the amplitude of the input signal (strain or
resembles a pattern recognition (PR) problem. Therefore, binary vibration response of the plate) exceeds predefined threshold,
data generated through a pulse communication protocol can be events are flagged and binary events are created.
evaluated through PR methods. In other words, binary data can
be interpreted by PR methods based on image data analysis Pattern Recognition Methods Explored
concepts. PR methods have achieved extensive attention for To reduce the effect of modelling errors, noise and
SHM applications in order to detect changes in a structure they environmental variations on measured structural responses, the
have been identified as a promising set of methods related to PR paradigm can be incorporated into a damage detection
system identification for structural damage evaluation [31,32]. approach. For this purpose, PR methods such as a proposed
Kiremidjian et al. [33] demonstrated that a damage detection method based on deviation of patterns, a two-dimensional
algorithm based on a pattern classification framework can detect principal component analysis (2DPCA), and a two-dimensional
structural changes caused by damage in a bridge structure. linear discriminant analysis (2DLDA), together with use of the
Trendafilova et al. [34] developed a viable methodology for nearest neighbor principle were used for pattern extraction and
classification. The PR methods were considered for this study

2 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


were chosen due to type of data. In other words, binary data to a number of classes [35,37]. Letting 𝐴 denote an 𝑚 × 𝑛
generated from self-powered sensors can be interpreted through image, and 𝑋 denote an n-dimensional column vector, the main
PR methods based on image data analysis techniques. There are idea is to project image 𝐴 , onto 𝑋 by the transformation
several techniques for PR based on image data analysis concepts, presented in (1). Considering that there are 𝐿 known pattern
but only some of them were considered in this paper. In this classes in the training set, and M denotes the size of the training
study, a proposed method based on deviation of patterns was set, the 2DLDA method introduces an image within-class matrix
developed according to the image data analysis concept, while 𝑆𝑊 𝐼𝑀 and an image between-class scatter matrix 𝑆𝐵 𝐼𝑀 that can
2DPCA and 2DLDA are well-known methods for classification be defined as follows [35,37]:
and data representation in PR and image processing.
𝑆𝑊 𝐼𝑀 = ∑𝐿𝑖=1 ∑𝐴𝑘 ∈𝑇𝑖(𝐴𝑘 − 𝐴̅𝑖 )𝑇 (𝐴𝑘 − 𝐴̅𝑖 ) (4)
Method based on Deviation of Patterns. This method
is based on the deviation of patterns with respect to each other. 𝑆𝐵 𝐼𝑀 = ∑𝐿𝑖=1 𝑁𝑖 (𝐴̅𝑖 − 𝐴̅) 𝑇 (𝐴̅𝑖 − 𝐴̅) (5)
In this way, system is able to recognize the change in response
1
of structure through different shapes (patterns). For this purpose, where, 𝐴̅𝑖 = ∑𝐴𝑘 ∈𝑇𝑖 𝐴𝑘 is the mean of the data matrices in
the deformed shape of the plate (pattern) in normal conditions 𝑁𝑖
1
can be memorized. Accordingly, by applying load with different class 𝑇𝑖 , and 𝐴̅ = ∑𝐿𝑖=1 ∑𝐴𝑘 ∈𝑇𝑖 𝐴𝑘 is the global mean matrix.
𝑁
harmonic amplitude or due to a change in material properties, a Fisher optimal projection axes 𝑋1 , … , 𝑋𝑑 are the orthonormal
classifier can recognize deviations from the memorized patterns eigenvectors of 𝑆𝑊 −1 𝑆𝐵 corresponding to the first 𝑑 largest
and identify new common patterns. eigenvalues. Accordingly, these optimal projection vectors and
the nearest neighbor classifier were used for feature extraction
Two-Dimensional Principal Component Analysis. and classification.
Two-dimensional principal component analysis (2DPCA) is an
image feature extraction and data representation technique used Methodology for Data Interpretation based on Pattern
in the areas of pattern recognition and computer vision [35,36]. Recognition
According to image processing methods, each image can be The aim of this study is to develop a computational
shown as a matrix. Letting 𝑋 denote an n-dimensional unitary framework for SHM that relies entirely on discrete binary data.
column vector, the main idea is to project matrix (image) 𝐴, onto Finite element (FE) simulation of an aluminum plate subjected
𝑋 by the following linear transformation [35,36]: to cyclic loading was conducted using the program ABAQUS
[38]. In this regard, the distribution of sensors was considered as
𝑌 = 𝐴𝑋 (1) an image matrix. The main idea is that binary events create and
each sensor flags an event, once the strain or vibration response
where 𝑌is the projected feature vector of matrix 𝐴. The 2DPCA of the plate exceeds predefined threshold. A view of the FE
method introduces the image covariance (scatter) matrix (𝐺𝑡 ), model along with distribution of model output locations (i.e.,
which can be directly determined from the training datasets as simulated sensors) is shown in Fig. 2.
follows:
1
𝑆𝑡 = ∑𝑁 ̅𝑇
𝑖=1[𝐴(𝑖) − 𝐴] [𝐴(𝑖) − 𝐴]
̅ (2)
𝑁

where 𝐴(𝑖) (𝑖 = 1,2, … , 𝑁) is an 𝑚 × 𝑛 matrix and denotes


1
the ith training image, and 𝐴̅ = ∑𝑁 𝐴(𝑖) denotes the average
𝑁 𝑖=1
image of all training data. The optimal projection axes 𝑋1 , … , 𝑋𝑑
are the orthonormal eigenvectors of the image covariance matrix
( 𝑆𝑡 ) corresponding to the first 𝑑 largest eigenvalues.
Accordingly these optimal projection vectors 𝑋1 , … , 𝑋𝑑 were
used for feature extraction. For a given image sample 𝐴, the
principal components vectors 𝑌1 , … , 𝑌𝑑 are defined as follows:

𝑌𝑘 = 𝐴𝑋𝑘 , 𝑘 = 1, … 𝑑 (3)

A nearest neighbor classifier was used for classification after Fig. 2. Illustration of the arrangement of self-powered
the feature extraction step. For this purpose, a nearest neighbor sensors in the FE model
classifier is found on the basis of Euclidean distance.
The FE model for the aluminum plate is shown in Fig. 3
Two-Dimensional Linear Discriminant Analysis. together with representative transverse displacement responses
Two-dimensional linear discriminant analysis (2DLDA) aims to and maximum absolute strain values of strain at time t = 5 s. The
extract features that well discriminate a set of data that belongs schematic simulation of proposed methodology used for data

3 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


interpretation technique and event prediction algorithm based on time spacing of the communication pulse. At the sensor nodes it
pattern recognition methods is illustrated in Fig. 4. is implemented by definiting different local event generation
rules. For this purpose, three thresholds (R1, R2, R3) were
defined. A schematic of the multiple binary concept and the
definition of local event generation rules is presented in Fig. 5.
It is worth mentioning that as a pilot test the considered
thresholds were simply based on displacement or strain
responses from the numerical simulation.

Fig. 3. FE analysis model: (a) Plate and loading, (b)


Transverse displacements (t = 5 s), (c) Maximum principal
strains (t =5 s)

Fig. 5. Distribution of self-powered sensors on a plate and


the concept of multiple binary events

Finite Element Analysis based on the Variance in


Amplitude of Harmonic Load. The task was to investigate
the effect of a change in loading conditions on the response
patterns. For this purpose, once the load was applied the
deformed shape of the plate (pattern) was memorized and the
system could recognize the change in response of structure
through different shapes (patterns). Three different arbitrary
displacement values were considered as thresholds to evaluate
the plate’s behavior due to changes in loading. The finite element
(FE) simulations were prepared according to different loading
Fig. 4. Framework of proposed methodology for data magnitudes. Dynamic analyses were performed for 5 sec and the
interpretation measurement parameter were displacement normal to the plate
surface. A uniformly distributed load (Fig. 3) with harmonic
Multiple Binary Event Concept used for Pattern amplitudes (Fig. 6) was considered. The resulting transverse
Identification. Simple pilot-type local rules for binary event displacement at the FE output locations (i.e., simulated sensor
generation were defined in terms of transverse displacements nodes) were provided as input to the proposed model to
and maximum principal strains at the sensor nodes and recognize the most common patterns. Additional FE analyses
introduced to the PR system. Plate response for normal operating were conducted for a distributed load with harmonic variation
conditions need to be classified to identify changes. along with random noise (Fig. 6).
Furthermore, in order to investigate the structure’s behavior due
to changes in loading condition or material degradation, it is Finite Element Analysis based on the Localized
necessary to obtain different levels of response. For this purpose, Decrease in Material Properties. Damage was simulated in
the concept of multiple binary events was introduced and used in the FE model by gradually reducing the stiffness (elastic
this study. This concept can be implemented in the substrate modulus) with time. Dynamic analyses were performed for 5
communication technique through variations in frequency or seconds and strains were the measured criteria. Maximum

4 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


principal strains were considered as thresholds. Accordingly, considered to classify patterns from the training data for the PR
strains from the virtual sensor nodes were introduced to the PR method based on two-dimensional principal component analysis
algorithm. The response of a plate with uniform stiffness was and two-dimensional linear discriminant analysis.
first memorized. Once the stiffness was locally decreased the PR Harmonic amplitudes with different noise levels (Fig. 6)
algorithm was used to successfully identify new patterns. The FE were considered for several FE analysis and obtained results
model for the plate with the region of reduced stiffness and the were used to evaluate the algorithms. In this way, the proposed
maximum absolute values of strain based on a 40% reduction in algorithms were assessed for the effects of signal noise. It was
stiffness for the 5th sec of analysis are presented in Fig. 7. determined that the system has acceptable performance with
respect to variations due to noise. However, performance of the
system can be affected by higher levels of noise.

Results based on the Change in Loading Condition


This section presents results based on the variance in loading
condition.

Results based on the Deviation of Patterns. For the


proposed pattern recognition methodology, the distribution of
virtual self-powered sensors was considered as an image matrix.
Finite element analyses were carried out for 5 seconds with a
time step of 0.01 s; therefore, the dataset size was 500. Three
arbitrary displacement values corresponding to the structural
Fig. 6. Harmonic amplitudes used for dynamic loading response of the plate were assumed as a threshold in order to
recognize the normal plate response. In this way, 4 regions
related to the different levels of displacement responses of sensor
nodes were established (Regions 1 to 4). It should be noted that
regions 1 and 4 represented the structural response
corresponding to the least and maximum displacement,
respectively. For the proposed classifier and using multiple
binary event generation concept (Fig. 5), the proposed method
could identify 3 common patterns according to Load #1.
Common patterns recognized by the system are shown in Fig. 8.
The different colors represent the plate’s response based on the
measurement/output at the sensor nodes.

Fig. 7. Evaluating effects from localized stiffness reduction:


(a) Maximum absolute strains (t = 5 s) based on 40%
stiffness reduction, (b) FE plate model with reduced
stiffness element

RESULTS AND DISCUSSIONS


The proposed damage detection technique based on
different pattern recognition (PR) methods was examined. The
effect of change in loading condition on structural assessment
and event prediction was investigated as well as change in Fig. 8. Common patterns identified by the system based on
material properties of the plate. As previously mentioned, the load#1
input to the computational model (dataset) was the structural
response according to the FE analysis. A custom program in the The definition of multiple thresholds (shown in Fig. 5) can
Matlab [39] platform was used for implementing the different be used to realize and interpret the different regions in Fig. 8. In
pattern recognitions techniques. Furthermore, k-means this way, inactive sensor (white) denote that the displacement
clustering along with the nearest neighbor classifier were response of that region of the plate does not exceed the

5 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


predefined threshold R1; while a yellow region means that the
displacement response of the plate exceeds threshold R2 but it is
less than threshold R3. Table 1 presents the number of
occurrences of each pattern within the entire dataset.

Table 1. Number of occurrences of patterns based on


load#1.
Number of occurences
Percentage
of patterns
Class (Pattern) 1 133 27%
Class (Pattern) 2 57 11%
Class (Pattern) 3 278 56%

Results obtained based on distributed load with harmonic


amplitude and random noise (Load #2) is shown in Fig. 6. The
proposed system could identify 21 patterns (18 new patterns with
respect to the 3 memorized patterns based on Load #1). The Fig. 10. Frequency of new recognized patterns for Load #2
distribution of 10 patterns (from the set of 21 new patterns) with
respect to the time at which they were identified is presented in
Fig. 9. The frequency histogram of these patterns is illustrated in
Fig. 10. It should be noted that 15 of the 21 new patterns with a
low number of occurrences were disregarded by the system and
only 3 patterns were considered by the system as new memorized
patterns. Accordingly, these new identified patterns were used
for comparison with new input patterns.

Fig. 11. Response of sensor nodes based on discrete


binary events (for 3 consecutive time steps & the envelope
of these steps according to load#2)

Results based on the 2DPCA Method. Feature


Fig. 9. Distribution of new recognized patterns with respect extraction was performed according to the 2DPCA method for
to time according to Load #2 determining feature matrices according for Load #1.
Accordingly, k-means clustering was used to identify the initial
Spatial-temporal integration of binary events based on the patterns required for classification. For this purpose, by means
discrete data was determined for Load #2. Results are provided of the 2DPCA method and assuming predefined thresholds
for 3 consecutive time steps of the dynamic analysis along with (displacement), three different classes were identified from the
an envelope of these steps in Fig. 11. This figure presents the possible patterns (Table 1). Once Load #2 was applied, new
creation of binary events with respect to the virtual sensor inputs were provided to the proposed algorithm and a nearest
locations for different analysis time steps. In other words, the neighbor criterion was used for classification. One of the
figure shows that variations in the response of sensor nodes due advantages of this method is that it can successfully overcome
to load variation can be recognized. the complexity of the existing noise of the system. It means that
the patterns were classified regardless of noise due to different
loading condition. Table 2 present the number of patterns
classified for each of classes according to Load #2.

6 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


Table 2. Number of pattern occurrences based in 2DPCA stiffness reduction leads to change in patterns, and this variation
Method for Load #2. is at the same location as the element with reduced stiffness in
Number of pattern the FE model.
Percentage
occurences An attempt was also made to link the binary events to
Class (Pattern) 1 150 34 % physical structural parameters. A preliminary result of this effort
Class (Pattern) 2 34 8% is shown in Fig. 14, which qualitatively shows the behavior of
Class (Pattern) 3 258 58 % the plate as determined from the interpolation of discrete binary
events. In this way, physical behavior of the plate based on
Results based on the 2DLDA Method. Feature binary data could be examined.
extraction was performed according to the 2DLDA method for
determining the Fisher feature matrices for Load #1. As for the
2DPCA method, k-means clustering was used to identify the
initial patterns (Table 1). Accordingly, a nearest neighbor
criterion was used to recognize and classify patterns once Load
#2 was applied. The number of patterns classified for each of the
three memorized classes is presented in Table 3.

Table 3. Number of pattern occurrences based on 2DLDA


Method for Load #2.
Number of pattern
Percentage
occurences
Class (Pattern) 1 120 27 %
Class (Pattern) 2 37 8%
Class (Pattern) 3 283 64 %

Results based on the change in Material Properties


This section provides results for the case of a localized Fig. 12. Distribution of new recognized patterns with
decreased in material properties. respect to the time based on a localized 40% reduction in
stiffness
Results based on the Deviation of Patterns.
Distributed damage was introduced to the proposed system by
reducing the elastic modulus to a region of the plate in the FE
model. Three strain levels were assumed as thresholds. The
distributed load with harmonic amplitude (Fig. 6) was applied
and common patterns were identified as the stiffness of the
critical region was gradually gradually reduced. However,
results are provided for 40% of stiffness reduction. Table 4
presents the number of occurrences of each pattern within the
entire 500 data for a stiffness reduction of 40%. The distribution
of new patterns with respect to time is shown in Fig. 12.

Table 4. Number of pattern occurrences for a constant


stiffness.
Number of pattern
Percentage
occurences
Class (Pattern) 1 283 65 %
Class (Pattern) 2 106 24 %
Class (Pattern) 3 23 5%
Class (Pattern) 4 22 5%

Spatial-temporal integration of the binary events based on


the discrete data was determined for the case of 40% stiffness
reduction. Results are provided in Fig. 13 for three consecutive Fig. 13. Response of sensor nodes based on discrete
binary events for 3 consecutive time steps and the
time steps of the dynamic analysis along with an envelope of
envelope of these steps based on a localized 40% reduction
these steps. It can be seen that the effect of material degradation in stiffness
on response of the sensor nodes is noticeable. In other words,

7 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


Fig. 14. Response of sensor nodes based on the interpolation of discrete binary events for 3 consecutive time steps & the
envelope of these steps based on a localized 40% reduction stiffness

Connection between Identified Patterns and


Structural Parameters of the Plate. The aim of this section
is to present a preliminary effort to establish a connection
between recognized patterns and structural properties. For this
purpose, the plate’s flexural rigidity was considered and assessed
through the localized change in elastic modulus. In this regard,
three different displacement values were assumed as thresholds
for investigating the behavior of the sensor nodes with respect to
their neighbors. Patterns based on the responses at the virtual
sensor nodes with a constant stiffness were determined due to
Load #1. Displacements at the sensor nodes were provided as
input to the pattern recognition algorithm. Accordingly, the
elastic modulus for one section of the plate was decreased from
20% up to 90% through the FE simulations.
The frequency of recognized patterns with respect to the
stiffness reduction are illustrated in Fig. 15 and Fig. 16 for
modulus reductions of 30% and 80%, respectively The Fig. 15. Frequency of patterns with respect to a localized
frequency of new recognized patterns due to an 80% decrease in stiffness reduction (up to 30% decrease)
Young modulus is presented in Fig. 17. As observed from results,
Pattern 2 is completely insensintive to the stiffness reduction.
However, Patterns 1 and 3 depend on the stiffness reduction
since the frequency of these patterns almost tends to zero with
increased stiffness reduction. Accordinlgy, new patterns were
recognized by the PR algorithm as the localized stiffness
decreases.

8 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


better a performance because this method deals directly with
discrimination between classes, whereas the 2DPCA deals with
the data in its entirety for the principal components analysis.
Furthermore, obtained results show that PR methods can be used
as damage identification algorithms based on binary data sets.

REFERENCES

[1] Chang, Ch., Chen, L., 2004. “Damage Detection of a


Rectangular Plate by Spatial Wavelet based Approach”. Applied
Acoustics, 65, pp. 819-832.
[2] Cornwell, P., Dobeling, S. W., and Farrar, C. R., 1999.
“Application of the Strain Energy Damage Detection Method to
Plate Like Structures”. Journal of sound and Vibration, 224(2),
pp. 359-374.
Fig. 16. Frequency of patterns with respect to a localized [3] Worden, K., Manson, G., and Allman, D., 2003.
stiffness reduction (up to 80% decrease) “Experimental Validation of a Structural Health Monitoring
Methodology: Part I. Novelty detection on a laboratory
structure”. Journal of sound and Vibration, 259(2), pp. 323-343.
[4] Manson, G., Worden, K., and Allman, D., 2003.
“Experimental Validation of a Structural Health Monitoring
Methodology: Part II. Novelty detection on a gnat aircraft”.
Journal of Sound and Vibration, 259(2), pp. 345-363.
[5] Manson, G., Worden, K., and Allman, D., 2003.
“Experimental Validation of a Structural Health Monitoring
Methodology: Part III. Damage location on an aircraft wing”.
Journal of Sound and Vibration, 259(2), pp. 365-385.
[6] Abdel, M. M., Rock, G., and Peeters, B., 1999.
“Parametrization of Damage in Reinforced Concrete Structures
using Model Updating”. Journal of Sound and Vibration, 228(4),
pp. 717-730.
[7] Jain, A. K., Duin, R. P., and Mao, J., 2000. “Statistical Pattern
Recognition: A Review”. IEEE Trans. Pattern Analysis and
Machine Intelligence, 22(1), pp. 4-3.
Fig. 17. Frequency of new recognized patterns with respect
to a localized stiffness reduction (up to 80% decrease) [8] Ho, T., Baird, H. S., 1997. “Large-scale Simulation Studies
in Image Pattern Recognition”. IEEE Trans. Pattern Analysis
and Machine Intelligence, 19(10), pp. 1067-1079.
CONCLUSIONS [9] Worden, K., Manson, G., and Denoux, T., 2009. “An
This study investigated damage identification algorithms Evidence-based Approach to Damage Location on an Aircraft
using discrete binary data. The proposed damage identification Wing”. Mechanical Systems and Signal Processing, 23, pp.
methodology was based on pattern recognition (PR) methods. 1792-1804.
PR methods based on image data analysis techniques were [10] Shen, T., Wan, F., Song, B., and Wu, Y., 2011. “Damage
adapted for this study and used for data interpretation and pattern Location and Identification of the Wing Structure with
identification. Finite element (FE) simulations were used to Probabilistic Neural Networks”. In Proceedings of Prognostics
generate virtual experimental data in order to evaluate the & System Health Management Conference, Shenzen, China, pp.
proposed method. Simple pilot-type local rules for binary event 118-123.
generation were defined in terms of displacements and strains at [11] Li, Y. Y., Cheng, L., Yam, L. H., and Wong, W. O., 2002.
the virtual sensor nodes in the FE models, and their response was “Identification of Damage Locations for Plate-like Structures
provided as input to the PR algorithms. Results indicate that using Damage Sensitive Indices: Strain Modal Approach”.
proposed PR method based on deviation of patterns can Computer and Structures, 80, pp. 1881-1894.
successfully recognize patterns due to variance in the amplitude [12] Fan, W., Qiao, P., 2012. “A Strain Energy-based Damage
of load and on material stiffness. Moreover, 2DPCA and 2DLDA Severity Correction Factor Method for Damage Identification in
methods were found to be effective techniques for the PR Plate-type Structures”. Mechanical Systems and Signal
problem. According to the obtained the 2DLDA method had Processing, 28, pp. 660-678.

9 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo


[13] Liu, S. C., Tomizuka, M., 2003. “Strategic Research for Distributed Computing and Networking, ICDCN, January,
Sensors and Smart Structures”. In Proceedings of the Mumbia, India.
International Conference on Structural Health Monitoring and [29] Huo, Q., Dong, B., and Biswas, S., 2012. “A Cellular Pulse
Intelligent Infrastructure, Tokyo, Japan, Nov 13-15, 1, pp. 113- Switching Architecture for Binary event Sensing”. IEEE Global
117. Communications Conference (GLOBECOM), December,
[14] Lynch, J. P., Loh, K. J., 2006. “A Summary Review of pp.379-384.
Wireless Sensors and Sensor Networks for Structural Health [30] Das, S., Lorenz, S., and Biswas, S., 2015. “Through-
Monitoring”. The Shock and Vibration Digest, 38, pp. 91-128. Substrate Event Reporting using Harvested Energy in
[15] Tanner, N. A., Farrar, C. R., and Sohn, H., 2002. “Structural Ultrasound Sensor Networks”. Accepted for publishing in IEEE
Health Monitoring using Wireless Sensing Systems with Global Communications Conference (GLOBECOM),
Embedded Processing”. In Proceedings of SPIE, 4704, pp. 215- December.
224. [31] Ren, W. X., Lin, Y. Q., and Fang, S. E., 2011. “Structural
[16] Wang, A., Chandrakasan, A. P., 2002. “Energy-efficient Damage Detection based on Stochastic Subspace Identification
DSPs for Wireless Sensor Networks”. IEEE Signal Processing and Statistical Pattern Recognition: I. Theory”. Smart Materials
Magazine, pp. 68-78. and Structures, 20, pp. 1-13.
[17] Paradiso, J.A., Starner, T., 2005. “Energy Scavenging for [32] Gul, M., Catbas, F., 2009. “Statistical Pattern Recognition
Mobile and Wireless Electronics”. IEE Pervasive Computing, for Structural Health Monitoring using Time Series Modeling:
4(1), pp. 18-27. Theory and Experimental Verifications”. Mechanical Systems
[18] Mita, A., Takahira, S., 2004. “Damage Index Sensor for and Signal Processing, 23, pp. 2192-2204.
Smart Structures”. Structural Engineering and Mechanics, 17(3- [33] Cheung, A., Cabrera, C., Sarabandi, P., Nair, K. K., and
4), pp. 331-346. Kiremidjian, A., 2008. “The Application of Statistical Pattern
[19] Farrar, C. R., Park, G., Allen, D. W., and Todd, M. D., 2006. Recognition Methods for Damage Detection to Field Data”.
“Sensor network Paradigms for Structural Health Monitoring”. Smart Materials and Structures, 17, pp. 1-12.
Structural Control and Health monitoring, 13, pp. 210-225. [34] Trendafilova, I., Cartmell, M. P., Ostachowicz, W., 2008.
[20] Park, G., Farrar, C.R., Todd, M.D., Hodgkiss, W., and “Vibration-based Damage Detection in an Aircraft Wing Scaled
Rosing, T., 2007. “Energy Harvesting for Structural Health Model using Principal Component Analysis and Pattern
Monitoring Sensor Networks”. Los Alamos National Laboratory recognition”. Journal of Sound and Vibration., 313, pp. 560-566.
Report, LA-14314-MS. [35] Wang, L., Wang, X., and Feng, J., 2006. “On Image Matrix
[21] Elvin, N. G., Lajnef, N., and Elvin, A., 2006. “Feasibility of based Feature Extraction Algorithms”. IEEE Trans. System.
Structural Monitoring with Vibration Powered Sensors”. Smart Man. And Cybernetics, 36(1), pp. 194-197.
Materials and Structures, 15(4), pp. 977-986. [36] Yang, J., Zhang, D., Frangi, A. F., Yang, J., 2004.
[22] Chakrabartty, S., Lajnef, N., and Elvin, A., 2006. “Toward “Two Dimensional PCA: A New Approach to Appearance-based
Self-powered Sensors and Circuits for Biomechanical Implants”. Face Representation and Recognition”. IEEE Trans. Pattern
VLSI Circuits for Biomedical Applications, eds. Artech House. Analysis and Machine Intelligence, 26(1), pp. 131-137.
[23] Huang, Ch., Lajnef, N., and Chakrabartty, Sh., 2010. [37] Li, M., Yuan, B., 2005. “2DLDA: A Statistical Linear
“Calibration and Characterization of Self-powered Floating- Discriminant Analysis for Image Matrix”. Pattern Recognition
Gate Usage Monitor with Single Electron Per Second Operation Letters, 26, pp. 527-532.
Limit”. IEEE Trans. Circuits and Systems, 57(3), pp. 556-567. [38] Dassault Systèmes Simulia Corp., "Abaqus Analysis User's
[24] Chakrabartty, Sh., Cauwenberghs, G., 2007. “Sub- Manual," Abaqus Documentation, version 6.11, Dassault
microwatt Analog VLSI Trainable Classifier”. IEEE Trans. Systèmes, Providence, RI, 2011.
Journal of Solid-State Circuits, 42(5), pp. 1169-1179. [39] MATLAB, The Math Works, Inc, Natick, MA, 2013.
[25] Huo, Q., Plummer, A., and Biswas, A., 2011.
“Ultra Wide Band Impulse Switching Protocols for Event and
Target Tracking Applications ”. In Proceedings of IEEE
SECON2,11, June, Salt Lake City, Utah.
[26] Huo, Q., Rao, J., and Biswas, S., 2013. “Pulse Switching:
Towards a Packet-less Protocol Paradigm for Event Sensing”.
IEEE Transactions on Mobile Computing (TMC), 12(1), pp. 35-
50.
[27] Huo, Q., Plummer, A., and Biswas, A., 2011. “Pulse
Switching for Static Event Sensing in Sensor Networks”. In
Proceedings of Globecom 2011, IEEE Global Communications
Conference, GLOBECOM, December, Houston, Texas.
[28] Huo, Q., Dong, B., and Biswas, S., 2013. “Cellular Pulse
Switching: An Architecture for event Sensing and Localization
in Sensor Networks”. 14th International Conference on

10 Copyright © 2015 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/conferences/asmep/86532/ on 03/23/2017 Terms of Use: http://www.asme.org/abo

You might also like