applied

sciences
Article
Guided Wave-Based Monitoring of Evolution of
Fatigue Damage in Glass Fiber/Epoxy Composites
Gang Yan * , Xiang Lu and Jianfei Tang
State Key Laboratory of Mechanics and Control of Mechanical Structures, College of Aerospace Engineering,
Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China; chenmo@nuaa.edu.cn (X.L.);
tangjf@nuaa.edu.cn (J.T.)
* Correspondence: yangang@nuaa.edu.cn




Received: 28 February 2019; Accepted: 26 March 2019; Published: 3 April 2019


Featured Application: The technique presented in this paper has potential application in
structural health monitoring of composite structures.

Abstract: This paper presents an experimental study on detecting and monitoring of evolution of
fatigue damage in composites under cyclic loads by using guided waves. Composite specimens
fabricated by glass fiber/epoxy laminates and surface mounted with piezoelectric wafers are fatigued
under tension–tension loads. A laser extensometer is used to obtain the degradation of longitudinal
stiffness of the specimens under fatigue states to reflect the accumulation of internal fatigue damage.
Meanwhile, at different fatigue cycles, one wafer acts as actuator to excite diagnostic guided waves,
and the other acts as sensor to receive corresponding response waves. These guided wave signals are
then processed by wavelet packet transform to extract characteristic features of energies in multiple
frequency bands. A statistical multivariate outlier analysis is then performed to determine the
existence of fatigue damage and to characterize their evolution using Mahalanobis squared distance.
Experimental results have demonstrated the potential applicability and effectiveness of guided waves
for continuous monitoring of fatigue damage in composite structures.

Keywords: structural health monitoring; fatigue damage in composites; guided waves; wavelet
packet transform; statistical multivariate outlier analysis

1. Introduction
Nowadays, composites are widely used in engineering structures due to their advantages of
light weight, high specific stiffness and strength. However, in the long run, fatigue damage initiated
with matrix cracks may inevitably accumulate in composite structures under cyclic loads. If this kind
of fatigue damage cannot be detected in time, it may continue to grow with new emerging damage
patterns such as delamination and fiber breakages, leading to catastrophic failure of the structures [1,2].
Thus it is very important and crucial to develop structural health monitoring (SHM) techniques to
detect the fatigue damage inside composite structures and, if possible, to continuously monitor its
evolution [3–11].
During the past two decades, a lot of techniques and methods as well as sensor networks have been
proposed to detect and identify damage in composites [12–17]. Among them, the guided wave-based
approach is considered as one of the most promising techniques. Using advanced signal processing
techniques, damage could be detected and identified by extracting characteristic information contained
in transient guided waves. A wide range of theoretical and experimental studies have been performed
to demonstrate the effectiveness of damage detection and identification using guided waves [18–27].

Appl. Sci. 2019, 9, 1394; doi:10.3390/app9071394 www.mdpi.com/journal/applsci

Appl. Sci. 2019, 9, 1394 2 of 17

However, most of these methods focus on impact damage which is localized damage phenomenon
instead of extended damage such as those caused by cyclic fatigue loads [28,29].
Usually, in the initial stage, fatigue damage in composites mainly contains accumulated micro
matrix cracks, resulting in a gradual stiffness degradation [1,2]. Thus, for monitoring fatigue damage
in composites, it is a straightforward way to monitor the degradation of stiffness to reflect the internal
damage. For example, Seale et al. studied the change of velocities and amplitudes of guided waves
in thermal–mechanical aged composites. They found that the velocities of symmetric S0 Lamb
waves reduced with the reduction of the in-plane stiffness, demonstrating that the accumulated
fatigue damage in composites can be characterized by measuring the velocities of the waves [28].
Rheinfurth et al. employed noncontact technique to excite and receive guided waves in composite
under different fatigue loading conditions, and studied the relationship between the change of wave
velocities of antisymmetric A0 Lamb waves and the stiffness degradation [29]. Further, to quantitatively
indicate the internal fatigue damage, studies have been performed to identify the elastic properties
of composites under different fatigue states by using guided waves and try to relate the reduction of
elastic properties with internal matrix cracks with micromechanics models. For example, Marzani et al.
used semianalytical finite element (SAFE) method to model the dispersion curves of guided waves in
an anisotropic composite plate, and combined the experimentally obtained dispersion curves from
time–frequency analysis to identify the elastic parameters [30]. Zhao et al. used a laser–ultrasonic
system to measure the phase velocity of ultrasonic guided waves in composites, and then employed
genetic algorithm combined with the measured phase velocity to invert the elastic modulus [31].
Tao et al. selected phase velocity of the S0 mode Lamb wave in the low frequency region to characterize
the fatigue damage in a composite laminate under cyclic loads in different fatigue stages. They also
used the phase velocity and damage model to predict the evolution of fatigue damage [32].
In contrast to the aforementioned velocity-based characterization methods, researchers also used
signal features to characterize fatigue damage in composites with guided waves [33–36]. By extracting
the damage-sensitive feature from the received guided waves, the fatigue damage can be classified and
identified. For example, Peng et al. used guided waves for in situ fatigue life prognosis for composite
laminates. They employed selected features such as normalized amplitude, correlation coefficient and
cross-correlation to characterize the fatigue damage and established a regression model to relate them
with stiffness degradation [33]. Wilson and Chang found that time-of-flight (TOF), amplitude, and
power spectral density (PSD) of the guided waves were sensitive features to fatigue induced matrix
cracks and delamination. They defined damage indices (DIs) with these features to characterize the
growth of matrix cracks and selected the PSD as the most sensitive feature to correlate with the matrix
cracks [34]. In addition to linear waves, nonlinear guided waves and their corresponding features
were also employed for fatigue damage detection [35,36]. For example, Li et al. investigated the
second harmonic generation of guided wave in composite laminates, and the correlation between the
acoustic nonlinearity and the degradation of composite laminate was studied, demonstrating that
the extracted nonlinear feature can characterize damage in an early stage [36]. However, most of
these works extracted only a univariate feature in the time or frequency domain, while deterministic
approaches were employed for characterizing fatigue damage without consideration of uncertainties
from environmental effects and measurement noise.
This paper presents experimental studies on detecting and monitoring of evolution of fatigue
damage in glass fiber/epoxy (GF/EP) composites under cyclic loads by using guided waves. Following
feature extraction of signal energies in multiple frequency bands by wavelet packet transform (WPT),
a statistical multivariate outlier analysis is performed to determine whether fatigue damage exist in
the specimens and to characterize the evolution of fatigue damage using Mahalanobis squared distance
(MSD). The rest of this paper is structured as follows. Section 2 describes the experimental fatigue
tests of the composite specimens and the monitoring process. Sections 3 and 4 present the WPT-based
signal processing and feature extraction, and statistical multivariate outlier analysis, respectively.
Appl. Sci.
Appl. Sci. 2019,
2019, 9,
9, 1394
x FOR PEER REVIEW 33 of
of 17
17

effectiveness and applicability of the proposed method. Finally, conclusions are summarized in
Experimental
Section 6. results are given in Section 5 to verify the effectiveness and applicability of the proposed
method. Finally, conclusions are summarized in Section 6.
2. Experimental Test
2. Experimental Test
To perform experimental study, two GF/EP laminate with [06/906]s and [04/908]s layups were
To perform
fabricated, experimental
respectively. study,unidirectional
The G15000 two GF/EP laminate prepregs with
were[06purchased
/906 ]s andfrom[04 /90 8 ]s layups
Weihai were
Guangwei
fabricated, respectively. The G15000 unidirectional prepregs were purchased
Composites Co. Ltd., with thickness of 0.12 mm and resin content of 30–40%. After stacking the layers from Weihai Guangwei
Composites
with designed Co. Ltd., the
layups, with thickness
laminates areofmanufactured
0.12 mm and through resin content of 30–40%.
hot-press After stacking
with temperature of 120the
°C
layers with designed layups, the laminates are manufactured through hot-press
for 90 min. The cured laminates have a thickness of 2.6 mm, and they were then cut into 200 mm × 30 with temperature
of 120 ◦ C for 90 min. The cured laminates have a thickness of 2.6 mm, and they were then cut into
mm specimens for tensile and fatigue tests. The average maximum tensile load for three specimens
200
of [0mm × 30 mm specimens for tensile and fatigue tests. The average maximum tensile load for three
6/906]s layup is 22.15 kN, corresponding to an average tensile strength of 283.97 MPa, and for
specimens of [06 /90
three specimens ]s4/90
of 6[0 layup is 22.15 kN, corresponding to an average tensile strength of 283.97 MPa,
8]s the layup is 16.19 kN, corresponding to an average tensile strength of
and forMPa.
207.56 three The
specimens
other twoof [0specimens
4 /908 ]s the from
layup[0 is6/90
16.19 kN, corresponding to an average tensile strength
6]s and [04/908]s layups are labeled as S_A and S_B
of 207.56 MPa. The other two specimens from
for fatigue test, respectively. As illustrated in Figure 1a, 6the[0 6 /90 ]s and
two [0 4 /90
ends of8 ]the
s layups are labeled
specimens as S_A
are reinforced
and S_B for fatigue test, respectively. As illustrated in Figure 1a, the two
for gripping, and on the surface of the specimens, two P-51 piezoelectric wafers purchased from Wuxiends of the specimens are
reinforced
Haiying Co. forwere
gripping,
mounted and on by the surface
epoxy withofathe specimens,
distance of 80 two
mm.P-51
Thepiezoelectric
diameter and wafers purchased
thickness of the
from
wafers are 10 mm and 1 mm, respectively. In the following experiments, one wafer acts asthickness
Wuxi Haiying Co. were mounted by epoxy with a distance of 80 mm. The diameter and actuator
of
to the wafers
excite are 10 guided
diagnostic mm andwaves 1 mm,intorespectively.
the specimens In theandfollowing
the otherexperiments,
acts as sensorone towafer actsthe
receive as
actuator to excite
response waves. diagnostic guided waves into the specimens and the other acts as sensor to receive
the response waves.

(a) (b)
Figure
Figure 1. Experimental setup:
1. Experimental setup: (a)
(a) glass
glass fiber/epoxy GF/EP specimen
fiber/epoxy GF/EP specimen and
and (b)
(b) fatigue
fatigue test
test and
and
monitoring systems.
monitoring systems.

AA test
test setup
setup consisting
consisting of
of aa MTS
MTS 810.25
810.25 fatigue
fatigue test
test system
system andand aa guided
guided waves
waves monitoring
monitoring
system is established. The monitoring system contains a KH-7600 wideband
system is established. The monitoring system contains a KH-7600 wideband amplifier produced amplifier produced by
by Krohn-Hite Corporation, a PXI-5441 arbitrary function generator, a PXI-5105
Krohn-Hite Corporation, a PXI-5441 arbitrary function generator, a PXI-5105 digitizer, and an digitizer, and an
embedded
embedded controller,
controller, which
which were
were all
all produced
produced by by National
National Instrument
Instrument Corporation.
Corporation. InIn addition,
addition, aa
noncontact
noncontact laser extensometer LE-05 produced by Electronic Instruments Research is also used
laser extensometer LE-05 produced by Electronic Instruments Research is also used to to
provide reference stiffness information to the guided wave-based technique.
provide reference stiffness information to the guided wave-based technique. The overall view of theThe overall view of
the
testtest setup
setup is shown
is shown in Figure
in Figure 1b. 1b.
For For
the the fatigue
fatigue test,test, tension–tension
tension–tension cyclic
cyclic loadsloads
are are applied
applied to
to the
the
specimens. The maximum tensile loads applied to specimens S_A and S_B are 11 kN (equals to
specimens. The maximum tensile loads applied to specimens S_A and S_B are 11 kN (equals to
nominal stress of 141.03 MPa) and 8 kN (equals to nominal stress of 102.56 MPa), respectively;
nominal stress of 141.03 MPa) and 8 kN (equals to nominal stress of 102.56 MPa), respectively; the the stress
ratios
stress and
ratiosloading frequencies
and loading of bothofcyclic
frequencies both loads
cyclicare
loads0.1 are
and0.11 Hz,
andrespectively. In order
1 Hz, respectively. Intoorder
followto
the evolution of fatigue damage, the longitudinal stiffness of the specimens is first
follow the evolution of fatigue damage, the longitudinal stiffness of the specimens is first obtained obtained before
before fatigue tests (zero fatigue cycle) and then at periodically paused intervals of each 1000 cycles.
 The narrow frequency band can help to reduce distortion of the response wave signals due to
dispersion effects. The response wave signals before fatigue test (zero fatigue cycle) is first acquired
as the baseline signals and then measured with an interval of 1000 cycles when the fatigue tests are
periodically paused. Figures 4a and 5a show a comparison of six sets of typical response wave signals
under the fatigue state after 1000, 2000, 3000, 4000, and 5000 fatigue cycles, as well as the baseline
Appl. Sci. 2019, 9, 1394 4 of 17
wave signals obtained from the two specimens, respectively. The later acquired wave signals are not
plotted together due to difficulty for clearly discerning each other. Combined with the stiffness
fatigue tests (zero
information fatigue
provided bycycle)
Tableand 1 andthen at periodically
Figure 2, it can bepaused
seen fromintervals of eachthat
the figures 1000with
cycles.
theAincrease
set of
small loadscycles,
of fatigue were applied
fatigue to the specimens
damage accumulateand inthethe
corresponding displacements
specimens, leading were measured
to the change of responseby
the
wavelaser extensometer.
signals, especiallyWith the information
the reduction of loadsAlso,
of amplitudes. and asdisplacements, longitudinal
illustrated in the locally zoomed stiffness
signals
can be calculated.
in Figures 4b andTable5b, at1anlists the stage
early longitudinal stiffness
of the fatigue datathe
tests, forwave
both signals
specimens within
change 10,000and
rapidly, fatigue
then
cycles, while Figure 2 shows the variations of longitudinal stiffness with increase
change slowly. This trend is similar to the degradation of longitudinal stiffness of the composite of fatigue cycles.
Itspecimens,
can been observed
indicating from
theFigure 2 that,ofwith
evolution the increase
fatigue damageofinfatigue cycles, the
composites can longitudinal stiffness
be characterized by
of the two specimens
analyzing the guideddegrades;
wave signals.at the end of 10,000 fatigue cycles, the stiffness of specimens S_A and
S_B reduces by approximately 11.2% and 8.6%, respectively. The stiffness reduction reflects the fatigue
Table 1. Longitudinal
damage accumulation stiffness data
in the composite for GF/EP[24,25].
specimens specimens under fatigue cycles (Unit: GPa).

Fatigue Cycles 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10,000
Table 1. Longitudinal stiffness data for GF/EP specimens under fatigue cycles (Unit: GPa).
Specimen S_A 25.30 23.70 23.32 22.90 22.67 22.58 22.55 22.49 22.51 22.50 22.46
Specimen S_B 18.14
Fatigue Cycles 0 17.60
1000 17.19
2000 16.88
3000 16.67
4000 16.65
5000 16.63
6000 700016.62 8000
16.61 9000
16.60 10,000
16.58
Specimen S_A 25.30 23.70 23.32 22.90 22.67 22.58 22.55 22.49 22.51 22.50 22.46
Specimen S_B 18.14 17.60 17.19 16.88 16.67 16.65 16.63 16.62 16.61 16.60 16.58

1 Test data
Fitted curve

0.98
Normalized stiffness (A.U.)

0.96

0.94

0.92

0.9

0.88
Appl. Sci. 2019, 9, x FOR PEER REVIEW
0 2000 4000 6000 8000 10,000 5 of 17
Fatigue cycles
(a)

1 Test data
Fitted curve
0.99
Normalized stiffness (A.U.)

0.98

0.97

0.96

0.95

0.94

0.93

0.92

0.91
0 2000 4000 6000 8000 10,000
Fatigue cycles
(b)
Figure 2. Longitudinal
Figure 2. Longitudinalstiffness
stiffnessdegradation
degradationwith fatigue
with cycles:
fatigue (a) specimen
cycles: S_A and
(a) specimen S_A(b)
andspecimen S_B.
(b) specimen
S_B.
Meanwhile, for monitoring the fatigue damage by using guided waves, the PXI-5441 arbitrary
function generator sends a diagnostic wave signal which is amplified by KH-7600 amplifier and drives
the actuator to generate guided waves, the response wave signals are then sensed by the sensor and
acquired by the PXI-5105 digitizer, whose sampling rate is set at 10 MHz. In this study, a modulated
five-cycle signal with center frequency of 250 kHz is employed as the diagnostic wave. Figure 3a,b
 Appl. Sci. 2019, 9, x FOR PEER REVIEW 5 of 17

Appl. Sci. 2019, 9, 1394 1 Test data 5 of 17

Fitted curve
0.99

Normalized stiffness (A.U.)

shows the time trace and the 0.98frequency spectrum of the diagnostic excitation, respectively. The narrow

frequency band can help to0.97 reduce distortion of the response wave signals due to dispersion effects.
The response wave signals before
0.96 fatigue test (zero fatigue cycle) is first acquired as the baseline signals
and then measured with an 0.95
interval of 1000 cycles when the fatigue tests are periodically paused.
Figures 4a and 5a show a comparison of six sets of typical response wave signals under the fatigue
0.94
state after 1000, 2000, 3000, 4000, and 5000 fatigue cycles, as well as the baseline wave signals obtained
0.93
from the two specimens, respectively. The later acquired wave signals are not plotted together due
0.92
to difficulty for clearly discerning each other. Combined with the stiffness information provided by
Table 1 and Figure 2, it can0.91
be seen from the figures that with the increase of fatigue cycles, fatigue
0 2000 4000 6000 8000 10,000
damage accumulate in the specimens, leading to the cycles
Fatigue change of response wave signals, especially the
reduction of amplitudes. Also, as illustrated in the(b) locally zoomed signals in Figures 4b and 5b, at an
early stage of the fatigue tests, the wave signals change rapidly, and then change slowly. This trend is
Figure 2. Longitudinal stiffness degradation with fatigue cycles: (a) specimen S_A and (b) specimen
similar to the degradation of longitudinal stiffness of the composite specimens, indicating the evolution
S_B.
of fatigue damage in composites can be characterized by analyzing the guided wave signals.

(a)

(b)
Figure Diagnosticexcitation
Figure 3.3. Diagnostic excitation with
with center
center frequency
frequency of kHz:
of 250 250 (a)
kHz:time(a) time
trace andtrace and (b)
(b) frequency
frequency
spectrum.spectrum.
 Appl. Sci. 2019, 9, x FOR PEER REVIEW 6 of 17
Appl.Sci.
Appl. Sci.2019,
2019,9,9,1394
x FOR PEER REVIEW 6 6ofof1717

(a)
(a)

(b)
(b)
Figure 4. Response wave signals after different fatigue cycles for specimen S_A: (a) six signals and (b)
zoomed
Figure signals.
Figure4.4.Response
Responsewave
wavesignals
signalsafter
afterdifferent
differentfatigue
fatiguecycles
cyclesfor
forspecimen
specimenS_A:
S_A:(a)
(a)six
sixsignals
signalsand
and(b)
(b)
zoomed signals.
zoomed signals.

(a)
Figure (a)
5. Cont.
Appl.
Appl.Sci. 2019,9,9,1394
Sci.2019, x FOR PEER REVIEW 7 7ofof1717

Amplitude (V)

(b)
Figure5.5.Response
Figure Responsewave
wavesignals
signalsafter
afterdifferent
differentfatigue
fatiguecycles
cyclesfor
forspecimen
specimenS_B:
S_B:(a)
(a)six
sixsignals
signalsand
and(b)
(b)
zoomed
zoomedsignals.
signals.

3. Signal Processing and Feature Extraction

3. Signal Processing and Feature Extraction
For characterizing the accumulated internal fatigue damage, the first step is to extract
For characterizing the accumulated internal fatigue damage, the first step is to extract damage-
damage-sensitive features from the sensed guided wave signals to indicate the presence and evolution
sensitive features from the sensed guided wave signals to indicate the presence and evolution of
of damage. The damage feature is a characteristic parameter or a set of parameters that could be
damage. The damage feature is a characteristic parameter or a set of parameters that could be
obtained by signal processing techniques in time, frequency or time–frequency domains. Compared to
obtained by signal processing techniques in time, frequency or time–frequency domains. Compared
traditional Fourier transform (FT)-based signal processing techniques, wavelet analysis is a widely
to traditional Fourier transform (FT)-based signal processing techniques, wavelet analysis is a widely
used signal processing and feature extraction technique in guided wave-based SHM with its powerful
used signal processing and feature extraction technique in guided wave-based SHM with its
capabilities for multiresolution analysis and time–frequency analysis [18,19,37,38].
powerful capabilities for multiresolution analysis and time–frequency analysis [18,19,37,38].
Within the framework of wavelet analysis, a signal is decomposed into an approximation and
Within the framework of wavelet analysis, a signal is decomposed into an approximation and a
a detail. The approximation is then itself split into a second-level approximation and detail, and
detail. The approximation is then itself split into a second-level approximation and detail, and the
the process is repeated [39]. For the sake of improvement, WPT decomposes not only the wavelet
process is repeated [39]. For the sake of improvement, WPT decomposes not only the wavelet
approximations at each level, but also the wavelet details. Therefore, WPT offers a richer analysis
approximations at each level, but also the wavelet details. Therefore, WPT offers a richer analysis on
on the original signal and a complete binary decomposition tree can be produced. In this study,
the original signal and a complete binary decomposition tree can be produced. In this study, the
the guided wave signals obtained after different fatigue cycles are first decomposed by WPT into
guided wave signals obtained after different fatigue cycles are first decomposed by WPT into
component signals to extract characteristic features, and then further analyzed to indicate the existence
component signals to extract characteristic features, and then further analyzed to indicate the
and evolution of internal fatigue damage.
existence and evolution of internal fatigue damage.
AAwavelet
waveletpacket
packet(WP)
(WP)isisa alinear
linearcombination
combinationofofwavelet waveletfunctions.
functions.ItItcan
canbe
bepresented
presentedasasthe
the
function ψij,k (i t):
function  j , k (t ) :
ψij,k (t) = 2 j/2 ψi (2 j t − k) (1)
i j/2 i j
 (t )  2  (2 t -k ) (1)
where i, j, and k are the modulation, the scale,j , kand the translation parameter, respectively. The wavelet
i is obtained from a recursive filtering process as
ψwhere i, j, and k are the modulation, the scale, and the translation parameter, respectively. The
wavelet  i is obtained from a recursive filtering process as
√ ∞
ψ2i (t) = 2 ∑ h(k )ψi (2t − k ) (2)
 2i (t ) k=−2∞ h( k ) i (2t -k ) (2)
k  
√ ∞
(t) = 2 ∑ g(k)ψi (2t − k)
ψ 2i +1
(3)
 (t )  2  g ( k ) (2t -k )
2 i 1 k =− ∞ i
(3)
k 
where the discrete filters h(k) and g(k ) are quadrature mirror filters associated with the scaling function
and the mother
where wavelet
the discrete h(k ) and
function,
filters g (k ) are quadrature mirror filters associated with the scaling
respectively.
function and the mother wavelet function, respectively.
For a guided wave signal S(t), according to Equation (1), at each level j, the WPT produces 2j
component signals S ij (t ) that are related to 2j sets of coefficients cij , k :
Appl. Sci. 2019, 9, 1394 8 of 17

For a guided wave signal S(t), according to Equation (1), at each level j, the WPT produces 2j
component signals Sij (t) that are related to 2j sets of coefficients cij,k :

2j
S(t) = ∑ Sij (t) (4)
i =1

∞
Sij (t) = ∑ cij,k ψij,k (t) (5)
k =−∞

Each WP coefficient cij,k is the inner product of the signal with the wavelet as
Z ∞
cij,k = S(t)ψij,k (t)dt (6)
−∞

After WPT decomposition, the WP component energy related to a single component signal is
equal to Z ∞
Eij = Sij (t)2 dt (7)
−∞

which stands for energy stored in the component signal Sij (t).
Based on the WP component signals, an energy vector containing WP component energies in
multiple frequency bands is defined by Equation (8) in this study as a multivariate signal feature:
h j j
iT
E= E1j E2j · · · E2j −1 E2j (8)

It is used to characterize the occurrence and evolution of fatigue damage in the composite
specimens. For real applications, the energies in some WP components are very small, thus the energy
vector defined in Equation (8) can be further simplified by excluding these negligible components.

4. Statistical Multivariate Outlier Analysis

In this study, the experimental tests are under controlled laboratory condition, and the guided
waves signals are not affected by various uncertainties. However, in real structural applications
with operational environmental effects, the guided wave signals will being inevitably influenced by
environmental conditions such as temperature and loads, and contaminated by measurement noise.
Thus a statistical analysis is needed to consider these various uncertainties. Since outlier analysis
have been demonstrated as a robust unsupervised learning pattern recognition tool for damage
detection [40–44], it is employed in this study to analyze the WPT features obtained from the response
guided wave signals under different fatigue states.
In statistics, an outlier is an observation that is significantly different from the rest of the population
and the outlier is believed to be generated by an alternative mechanism, for example, damage in
structures [40,41]. Statistical outlier analysis can be considered as a special class of pattern recognition,
which classifies the abnormal data or state from the normal ones. To determine whether an observed
data is an outlier, discordancy tests are usually performed. In the case of multivariate data, the
discordancy test is based on deviation statistics and given by

 n¯o T ¯
Dξ = ( xξ − x ) [Σ]−1 ({xξ } − {x}) (9)

where Dξ is the Mahalanobis squared distance (MSD), xξ is the measured data corresponding to the
¯
potential outlier, and x and Σ are the mean vector and covariance matrix of the samples, respectively.
Superscript T indicates transpose. The latter two values may be calculated with or without the potential
outlier depending upon whether inclusive or exclusive measures are preferred [40]. In this study,
Appl. Sci. 2019, 9, 1394 9 of 17

the statistics is exclusively computed from the signal features, i.e., WP energy vectors defined in
Appl. Sci. 2019, 9, x FOR PEER REVIEW 9 of 17
Equation (8) in the pristine state, without including potential outliers, i.e., WP energy vectors in the
fatigue
and eachstates.copy is subsequently corrupted with amplitude fluctuation and measurement noise as
described into
In order consider(10).
Equation the various uncertainties on guided wave signals and construct a suitable
mean vector and a covariance matrix, the guided c
wave signals e
in pristine state are copied 500 times and
S
each copy is subsequently corrupted with amplitude fluctuation ( t )  (1  e )  S ( t )  and measurement noise as described (10)
inwhere
Equation S e ((10).
t ) is the experimentally obtained guided wave signals and S c (t ) is the corrupted wave
S c ( t ) = (1 + e ) × S e ( t ) + σ (10)
signals, e is an random number to reflect the amplitude fluctuation caused by environmental
where Se (t) isand
conditions,  is the measurement
the experimentally obtained guided noise whichwave is usually
signals Sc (t) is the
andmodeled as Gaussian
corruptedwhite noise. In
wave signals,
this
e is an study,
randomwe assume
number that ethe
to reflect comes
amplitudefrom fluctuation
a uniform caused distribution [emin, emax], and
by environmental the level
conditions, andof
σ measurement
is the measurement noise isnoise
described
whichbyissignal-to-noise
usually modeled ratioas(SNR).
Gaussian Thenwhite
WPTnoise.
is applied to these
In this study,copies
we
to obtain
assume thatthe signal from
e comes features. For determining
a uniform distribution whether
[emin , eanmax observation of signal
], and the level features or its
of measurement MSDisis
noise
an outlier,
described bya signal-to-noise
proper threshold value
ratio (SNR). Dth Then
should WPT be calculated.
is applied to If the
theseMSD is less
copies obtainDthe
to than th , it can be
signal
features.
considered For thatdetermining
the changewhether an observation
of response guided waves of signal features orby
is introduced its environmental
MSD is an outlier, a properor
conditions
threshold
measurement valuenoise;Dth should be calculated.
otherwise, it is because If the MSD isdamage
of fatigue less than Dth , it canand
occurrence be accumulation.
considered that the
change of response guided waves is introduced by environmental conditions or measurement noise;
5. Experimental
otherwise, Results
it is because of fatigue damage occurrence and accumulation.
With the aforementioned
5. Experimental Results WPT-based signal processing and multivariate feature extraction
method, the wave signals obtained under different fatigue states are analyzed. The mother wavelet
With the
function of aforementioned
Daubechies wavelet WPT-based
‘db1’ signal
is used processing and multivariate
to decompose each of the feature
wave extraction
signals method,
into sub-
the wave signals obtained under different fatigue states are analyzed. The
components, and a 4-level decomposition is performed with 2 = 16 components in total. Figure 6
4 mother wavelet function
ofshows
Daubechies
the WPwavelet ‘db1’energies
component is used toof decompose
the baseline each waveofsignals
the wave signals intoS_A
for specimens sub-components,
and S_B at zero
and a 4-level decomposition is performed with 2 4 = 16 components in total. Figure 6 shows the WP
cycle, i.e., without fatigue damage. It can be seen that the WP component energies mainly concentrate
component
on the firstenergies
several of the baselineFor
components. wave
bothsignals for specimens
specimens, the sumsS_A and S_Bin
of energies atthe
zero cycle,
first fivei.e., without
components
fatigue damage. It can be seen that the WP component energies mainly concentrate
exceed 98% of the total energies of the wave signals, thus the first five energies form the energy on the first several
vector
components. For both specimens, the sums of energies in the first five components
as the multivariate signal features for characterizing the fatigue damage in this study. Figure 7 shows exceed 98% of the
total energies of the wave signals, thus the first five energies form the energy vector
the evolution of the first five WP component energies of the wave signals with the increase of fatigue as the multivariate
signal
cycles.features
It can for
be characterizing the fatigue
seen that, in general, withdamage in this of
the increase study. Figure
fatigue 7 shows
cycles, the evolution
the energies of the
for each WP
first five WP component energies of the wave signals with the increase of fatigue cycles.
component decay quickly at the initial stage and then slow down. This trend is very similar to that of It can be seen
that,
the in general,
decay with the increase of fatigue cycles, the energies for each WP component decay quickly
of stiffness.
at the initial stage and then slow down. This trend is very similar to that of the decay of stiffness.
Energy (J)

(a)
Figure 6. Cont.
 Appl. Sci. 2019, 9, x FOR PEER REVIEW 10 of 17
Appl.
Appl.Sci. 2019,
Sci. 9, 9,
2019, 1394
x FOR PEER REVIEW 1010
ofof
1717

180

180
160

160
140

140
120

120
100

100
80

80
60

60
40

40
20

20
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
0 Component
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
(b)
Component

Figure 6. Wavelet packet (WP) component energies (b) for the baseline wave signal: (a) specimen S_A
and
Figure(b)6.specimen
Figure Wavelet
6. WaveletS_B.
packet (WP)
packet component
(WP) energies
component for the
energies for baseline wavewave
the baseline signal: (a) specimen
signal: S_A and
(a) specimen S_A
(b) specimen
and S_B. S_B.
(b) specimen
(J) (J)
Energy
Energy

(a)
(a)
(J) (J)
Energy
Energy

(b)
Figure7.7.Variations
Figure VariationsofofWP
WPcomponent
componentenergies (b)
energiesofofwave
wavesignals
signalswith
withfatigue
fatiguecycles:
cycles:(a)
(a)specimen
specimenS_A
S_A
and
and (b)
(b) specimen
specimen S_B.
S_B.
Figure 7. Variations of WP component energies of wave signals with fatigue cycles: (a) specimen S_A
and (b) specimen S_B.
Appl. Sci. 2019, 9, 1394 11 of 17

As aforementioned, to construct a suitable mean vector and a covariance matrix for outlier analysis,
the guided wave signals in pristine state are copied 500 times and each copy is subsequently corrupted
with amplitude fluctuation and measurement noise, as described in Equation (10). Two contamination
levels are considered: (I) Level 1: the fluctuation factor e comes from uniform distribution [−0.05 0.05]
and the SNR of measurement noise is 20 dB. (II) Level 2: The fluctuation factor e comes from
uniform distribution [−0.1 0.1] and the SNR of measurement noise is 10 dB. Figure 8a,b illustrates the
contaminated baseline signals with different levels for specimen S_A. Similar contaminated signals
can be obtained through Equation (10) for baseline signal for specimen S_B. Then for each set of
contaminated 500 baseline copies, mean vector and covariance matrix are constructed and baseline
MSDs are calculated according to Equation (9). To set an appropriate threshold value, statistical
analysis is performed and lognormal distribution is used to fit the probability density function (PDF)
for baseline MSDs. As illustrated in Figure 9a, for wave signals from specimen S_A with Level 1
contamination, threshold is determined by taking the value corresponding to 99.9% confidence of the
fitted lognormal distribution, which is 19.3. To test whether the wave signals after a certain fatigue
cycles are outlier or not, the wave signals are copied 100 times and the same contamination levels for
the baseline signals are applied. The corresponding MSDs for different sets of contaminated wave
signals can be also calculated by Equation (9) and compared with the threshold value. If the MSDs are
less than the threshold value, it can be considered that the change of guided wave features compared to
the baseline set is caused by environmental conditions or measurement noise; otherwise it is caused by
damage in the structure, i.e., fatigue damage in our case. It can be seen from Figure 9b, the calculated
MSDs after different fatigue cycles all exceed the threshold value, and with the increase of fatigue
cycles, the MSDs have a general increased trend. The means of the MSDs increase almost monotonically
with the fatigue cycles. This may because only matrix cracks increasingly accumulate in specimen
S_A before 10,000 fatigue cycles. It demonstrates that MSD can be used as a measure to integrate
the multivariate signal features to detect and characterize fatigue damage in composites. For wave
signals from specimen S_A with Level 2 contamination, Figure 10a shows the histogram of baseline
MSDs and the lognormal fit, the threshold value with 99.9% confidence is 19.2, and it can be seen from
Figure 10b that with the increase of contamination level, unlike the results shown in Figure 9b, most of
the MSDs for wave signals right after 1000 fatigue cycles are below the threshold value, indicating that
at this stage fatigue damage cannot be discerned. However, the MSDs from the later wave signals are
all above the threshold value and have a general trend of increase with the increase of fatigue cycles.
A similar processing procedure is applied to wave signals for specimen S_B, and the corresponding
results are illustrated in Figures 11 and 12. It can be seen that, for both contamination levels, the MSDs
of wave signals obtained after different fatigue cycles all exceed the threshold of the baseline MSDs,
and have general trends of increase with the increase of fatigue cycles until last several thousand cycles.
This may because matrix cracks accumulation in specimen S_B reaches saturation after ~5000 fatigue
cycles, and then new damage mechanism such as delamination occurs. Although further study is
needed to disclose the interaction between guided waves and damage phenomena, it still indicates
the effectiveness of MSD in integrating multiple signal features and characterizing fatigue damage
in composites.
 Appl.
Appl. Sci.
Sci. 2019,
2019, 9, 9, x FOR PEER REVIEW
1394 1212
ofof
1717

(a)

(b)
Figure8. 8. Contaminated
Figure Contaminated baseline
baseline signals
signals with
with different
different levels
levels for
for specimen
specimen S_A:
S_A:(a)
(a)Level
Level1 1
contaminationand
contamination and (b)
(b) Level
Level 2 contamination.
2 contamination.
 Appl.
Appl. Sci.
Sci.
Appl. Sci. 2019, 9,xxFOR
9, 9,
2019,
2019, 1394 FORPEER
PEERREVIEW
REVIEW 13of
13 of17
17

(a)
(a)

(b)
(b)
Figure
Figure
Figure 9.9.9.Outlier
Outlieranalysis
analysisresults
analysis results for
resultsfor specimen
specimen
for S_A
S_A
specimen with
with
S_A Level
Level
with 11contamination:
contamination:
Level (a)probability
(a)
1 contamination: probability density
density
(a) probability
function
function (PDF)
(PDF) of
of baseline
baseline Mahalanobis
Mahalanobis squared
squared distance
distance (MSD)
(MSD) and
and (b)
(b) MSDs
MSDs for
for
density function (PDF) of baseline Mahalanobis squared distance (MSD) and (b) MSDs for different different
different fatigue
fatigue
states.
states.states.
fatigue

(a)
(a)
Figure 10. Cont.
 Appl. Sci. 2019, 9, x FOR PEER REVIEW 14 of 17
Appl.
Appl.Sci. 2019,9,9,1394
Sci.2019, x FOR PEER REVIEW 14of
14 of17
17

(b)
(b)
Figure
Figure 10. Outlier
10.Outlier
Figure10. analysisresults
Outlieranalysis
analysis resultsfor
results forspecimen
for specimenS_A
specimen S_Awith
S_A withLevel
with Level22contamination:
Level contamination:(a)
contamination: (a)PDF
(a) PDFof
PDF ofofbaseline
baseline
baseline
MSD
MSD
MSDandand (b)
and(b) MSDs
(b)MSDs fordifferent
MSDsfor
for differentfatigue
different fatiguestates.
fatigue states.
states.

(a)
(a)
MSD
MSD

(b)
(b)
Figure 11. Outlier analysis results for specimen S_B with Level 1 contamination: (a) PDF of baseline
Figure 11.Outlier
Figure11. Outlieranalysis
analysisresults
resultsfor
forspecimen
specimenS_B
S_Bwith
withLevel
Level11contamination:
contamination:(a)
(a)PDF
PDFofofbaseline
baseline
MSD and (b) MSDs for different fatigue states.
MSD
MSDandand(b)
(b)MSDs
MSDsforfordifferent
differentfatigue
fatiguestates.
states.
Appl.
Appl.Sci. 2019,9,9,1394
Sci.2019, x FOR PEER REVIEW 15
15ofof17
17

Probability

(a)

(b)
Figure 12.Outlier
Figure12. Outlieranalysis
analysisresults
resultsfor
forspecimen
specimenS_B
S_Bwith
withLevel
Level22contamination:
contamination:(a)
(a)PDF
PDFofofbaseline
baseline
MSD and (b) MSDs for different fatigue states.
MSD and (b) MSDs for different fatigue states.

6. Conclusions
6. Conclusions
This paper presents an experimental study on detecting and monitoring of evolution of fatigue
This paper presents an experimental study on detecting and monitoring of evolution of fatigue
damage in composites under cyclic loads by using guided waves. Piezoelectric wafers are used to excite
damage in composites under cyclic loads by using guided waves. Piezoelectric wafers are used to
and receive guided wave signals to monitor the evolution of fatigue damage after different fatigue
excite and receive guided wave signals to monitor the evolution of fatigue damage after different
cycles. The guided wave signals are processed by WPT to extract features based on WP component
fatigue cycles. The guided wave signals are processed by WPT to extract features based on WP
energies to characterize the internal fatigue damage accumulation which is reflected by measured
component energies to characterize the internal fatigue damage accumulation which is reflected by
stiffness degradation. A statistical multivariate outlier analysis is then performed with consideration
measured stiffness degradation. A statistical multivariate outlier analysis is then performed with
of uncertainties from amplitude fluctuations caused by environmental effects and measurement noise.
consideration of uncertainties from amplitude fluctuations caused by environmental effects and
It is used to determine the existence of fatigue damage in composite specimens and to characterize
measurement noise. It is used to determine the existence of fatigue damage in composite specimens
their evolution using MSD. Experimental results have demonstrated the effectiveness and potential
and to characterize their evolution using MSD. Experimental results have demonstrated the
application of the guided wave-based monitoring technique: the extracted WP component energies of
effectiveness and potential application of the guided wave-based monitoring technique: the extracted
the guided wave signals can be used as effective signal features to characterize the internal fatigue
WP component energies of the guided wave signals can be used as effective signal features to
damage in composites, and the statistical outlier analysis is an effective tool to detect the damage with
characterize the internal fatigue damage in composites, and the statistical outlier analysis is an
consideration of uncertainties, while the merged MSDs of the signal features have a similar trend to
effective tool to detect the damage with consideration of uncertainties, while the merged MSDs of the
evolution of fatigue damage in composites.
signal features have a similar trend to evolution of fatigue damage in composites.
Author Contributions: Conceptualization, G.Y.; Methodology, G.Y., X.L., and J.T.; Investigation, G.Y., X.L., and
Author
J.T.; DataContributions: Conceptualization,
Curation, X.L. and G.Y.; Methodology,
J.T.; Writing—Original G.Y., X.L.,
Draft preparation, G.Y.;and J.T.; Investigation,
Writing—Review and G.Y., X.L.,G.Y.,
Editing, and
J.T.; and
X.L., DataJ.T.;
Curation, X.L. and
Supervision, J.T.; Writing—Original Draft preparation, G.Y.; Writing—Review and Editing, G.Y.,
G.Y.
X.L., and J.T.; Supervision, G.Y.
Appl. Sci. 2019, 9, 1394 16 of 17

Funding: This research is supported by the National Nature Science Foundation (Grant No. 11602104) and the
Fundamental Research Funds for the Central Universities (Grant No. NS2016011).
Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Talreja, R. Multi-scale modelling in damage mechanics of composite materials. J. Mater. Sci. 2006, 41,
6800–6812. [CrossRef]
2. Reifsnider, K.L. Fatigue of Composite Materials; Elsevier: New York, NY, USA, 2012.
3. Reis, P.; Ferreira, J.; Richardson, M. Fatigue damage characterization by NDT in polypropylene/glass fibre
composites. Appl. Compos. Mater. 2011, 18, 409–419. [CrossRef]
4. Shin, C.S.; Chiang, C.C. Fatigue damage monitoring in polymetric composites using multiple fiber bragg
gratings. Int. J. Fatigue 2006, 28, 1315–1321. [CrossRef]
5. Gao, L.; Thostenson, E.T.; Zhang, Z.G.; Chou, T.W. Sensing of damage mechanisms in fiber-reinforced
composites under cyclic loading using carbon nanotubes. Adv. Sci. 2009, 19, 123–130. [CrossRef]
6. De Baere, I.; van Paepegem, W.; Degrieck, J. Electrical resistance measurement for in situ monitoring of
fatigue of carbon fabric composites. Int. J. Fatigue 2010, 32, 197–207. [CrossRef]
7. Mouritz, A.P.; Townsend, C.; Shah Khan, M.Z. Non-destructive detection of fatigue damage in thick
composites by pulse-echo ultrasonics. Compos. Sci. Technol. 2000, 60, 23–32. [CrossRef]
8. Toubal, L.; Karama, M.; Lorrain, B. Damage evolution and infrared thermography in woven composite
laminates under fatigue loading. Int. J. Fatigue 2006, 28, 1867–1872. [CrossRef]
9. Bourchak, M.; Farrow, I.R.; Bond, I.P.; Rowland, C.W.; Menan, F. Acoustic emission energy as a fatigue
damage parameter for CFRP composites. Int. J. Fatigue 2007, 29, 457–470. [CrossRef]
10. Jespersen, K.M.; Zangenberg, J.; Lowe, T.; Withers, P.J.; Mikkelsen, L.P. Fatigue damage assessment
of uni-directional non-crimp fabric reinforced polyester composite using X-ray computed tomography.
Compos. Sci. Technol. 2016, 136, 94–103. [CrossRef]
11. Ling, Y.; Mahadevan, S. Integration of structural health monitoring and fatigue damage prognosis.
Mech. Syst. Signal Process. 2012, 28, 89–104. [CrossRef]
12. Staszewski, W.J.; Mahzan, S.; Traynor, R. Health monitoring of aerospace composite structures—Active and
passive approach. Compos. Sci. Technol. 2009, 69, 1678–1685. [CrossRef]
13. Diamanti, K.; Soutis, C. Structural health monitoring techniques for aircraft composite structures.
Prog. Aerosp. Sci. 2010, 46, 342–352. [CrossRef]
14. Kinet, D.; Megret, P.; Goossen, K.W.; Qiu, L.; Heider, D.; Caucheteur, C. Fiber bragg grating sensors toward
structural health monitoring in composite materials: Challenges and solutions. Sensors 2014, 14, 7394–7419.
[CrossRef] [PubMed]
15. Giurgiutiu, V.; Santoni-Bottai, G. Structural health monitoring of composite structures with
piezoelectric-wafer active sensors. AIAA J. 2011, 49, 565–581. [CrossRef]
16. Bosse, S.; Lechleiter, A. A hybrid approach for structural monitoring with self-organizing multi-agent
systems and inverse numerical methods in material-embedded sensor networks. Mechatronics 2016, 34,
12–37. [CrossRef]
17. Neuschwander, K.; Moll, J.; Memmolo, V.; Schmit, M.; Bucker, M. Simultaneous load and structural
monitoring of a carbon fiber rudder stock: Experimental results from a quasi-static tensile test. J. Intell. Mater.
Syst. Struct. 2019, 30, 272–282. [CrossRef]
18. Lemistre, M.; Balageas, D. Structural health monitoring system based on diffracted Lamb waves analysis by
multiresolution processing. Smart Mater. Struct. 2001, 10, 504–511. [CrossRef]
19. Seth, S.K.; Spearing, S.M.; Constantinos, S. Damage detection in composite materials using Lamb wave
methods. Smart Mater. Struct. 2002, 11, 269–278.
20. Basri, R.; Chiu, W.K. Numerical analysis on the interaction of guided lamb waves with a local elastic stiffness
reduction in quasi-isotropic composite plate structures. Compos. Struct. 2004, 66, 87–99. [CrossRef]
21. Su, Z.; Ye, L.; Lu, Y. Guided lamb waves for identification of damage in composite structures: A review.
J. Sound Vib. 2006, 295, 753–780. [CrossRef]
22. Ng, C.T.; Veidt, M. A lamb-wave-based technique for damage detection in composite laminates. Smart Mater.
Struct. 2009, 18, 074006. [CrossRef]
Appl. Sci. 2019, 9, 1394 17 of 17

23. Polimeno, U.; Meo, M. Detecting barely visible impact damage on aircraft composite structures. Compos.
Struct. 2009, 91, 398–402. [CrossRef]
24. Yan, G.; Zhou, L. Damage detection for composite structure using statistical outlier analysis with temperature
effect. Trans. Nanjing Univ. Aero. Astro. 2011, 28, 231–239.
25. Watkins, R.; Jha, R. A modified time reversal method for Lamb wave based diagnostics of composite
structures. Mech. Syst. Signal Process. 2012, 31, 345–354. [CrossRef]
26. Rogge, M.D.; Leckey, C.A.C. Characterization of impact damage in composite laminates using guided
wavefield imaging and local wavenumber domain analysis. Ultrasonics 2013, 53, 1217–1226. [CrossRef]
27. Capriotti, M.; Kim, H.E.; Lanza di Scalea, F.; Kim, H. Non-destructive inspection of impact damage in
composite aircraft panels by ultrasonic guided waves and statistical processing. Materials 2017, 10, 616.
[CrossRef]
28. Seale, M.D.; Madaras, E.I. Lamb wave evaluation of the effects of thermal-mechanical aging on composite
stiffness. J. Compos. Mater. 2000, 34, 27–38. [CrossRef]
29. Rheinfurth, M.; Kosmann, N.; Sauer, D.; Busse, G.; Schulte, K. Lamb waves for non-contact fatigue state
evaluation of composites under various mechanical loading conditions. Compos. Part A 2012, 43, 1203–1211.
[CrossRef]
30. Marzani, A.; Marchi, L.D. Characterization of the elastic moduli in composite plates via dispersive guided
waves data and genetic algorithms. J. Intell. Mater. Syst. Struct. 2013, 24, 2135–2147. [CrossRef]
31. Zhao, J.L.; Qiu, J.; Ji, H.L. Reconstruction of the nine stiffness coefficients of composites using a laser
generation based imaging method. Compos. Sci. Technol. 2016, 126, 27–34. [CrossRef]
32. Tao, C.C.; Ji, H.L.; Qiu, J.H.; Zhang, C.; Wang, Z.; Yao, W. Characterization of fatigue damages in composite
laminates using Lamb wave velocity and prediction of residual life. Compos. Struct. 2017, 166, 219–228.
[CrossRef]
33. Peng, T.; Liu, Y.; Saxena, A.; Goebel, K. In-situ fatigue life prognosis for composite laminates based on
stiffness degradation. Compos. Struct. 2015, 132, 155–165. [CrossRef]
34. Wilson, C.L.; Chang, F.K. Monitoring fatigue-induced transverse matrix cracks in laminated composites
using built-in acousto-ultrasonic techniques. Struct. Health Monit. 2016, 15, 335–350. [CrossRef]
35. Deng, M.; Pei, J. Assessment of accumulated fatigue damage in solid plates using nonlinear lamb wave
approach. Appl. Phys. Lett. 2007, 90, 121902. [CrossRef]
36. Li, W.; Cho, Y.; Achenbach, J.D. Detection of thermal fatigue in composites by second harmonic lamb waves.
Smart Mater. Struct. 2012, 21, 085019. [CrossRef]
37. Rizzo, P.; Sorrivi, E.; Lanza di Scalea, F.; Viola, E. Wavelet-based outlier analysis for guided wave structural
monitoring: Application to multi-wire strands. J. Sound Vib. 2007, 307, 52–68. [CrossRef]
38. Sohn, H.; Park, G.; Wait, J.R.; Limback, N.P.; Farrar, C.R. Wavelet-based active sensing for delamination
detection in composite structures. Smart Mater. Struct. 2004, 13, 153–160. [CrossRef]
39. Mallat, S. A theory for multiresolution signal decomposition: The wavelet representation. IEEE Trans. Pattern
Anal. Mach. Intell. 1989, 11, 674–693. [CrossRef]
40. Worden, K.; Manson, G.; Fieller, N.R.J. Damage detection using outlier analysis. J. Sound Vib. 2000, 229,
647–667. [CrossRef]
41. Worden, K.; Manson, G. The application of machine learning to structural health monitoring. Philos. Trans.
R. Soc. A 2007, 365, 515–537. [CrossRef] [PubMed]
42. Cheung, A.; Cabrera, C.; Sarabandi, P.; Nair, K.K.; Kiremidjian, A.; Wenzel, H. The application of statistical
pattern recognition methods for damage detection to field data. Smart Mater. Struct. 2008, 17, 065023.
[CrossRef]
43. Park, S.; Inman, D.J.; Yun, C.B. An outlier analysis of MFC-based impedance sensing data for wireless
structural health monitoring of railroad tracks. Eng. Struct. 2008, 30, 2792–2799. [CrossRef]
44. Pavlopoulou, S.; Worden, K.; Soutis, C. Novelty detection and dimension reduction via guided ultrasonic
waves: Damage monitoring of scarf repairs in composite laminates. J. Intell. Mater. Syst. Struct. 2016, 27,
549–566. [CrossRef]

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).