Computer-Aided Civil and Infrastructure Engineering 00 (2012) 1–12

Signal Pattern Recognition for Damage Diagnosis

in Structures
Long Qiao
Department of Engineering Technology, Missouri Western State University, Saint Joseph, MO, USA

&

Asad Esmaeily∗ & Hani G. Melhem

Department of Civil Engineering, Kansas State University, Manhattan, KS, USA

Abstract: A signal-based pattern-recognition approach infrastructures aims at monitoring the performance of

is used for structural damage diagnosis with a single or a system to enhance its safety and reduce its life-cycle
limited number of input/output signals. The approach cost by detection of possible damage in an early stage.
is based on extraction of the features of the structural This includes various infrastructures from pavements
response that present a unique pattern for each spe- to bridge decks (Lajnef et al., 2011; Cusson et al., 2011;
cific damage case. In this study, frequency-based fea- Adewuyi and Wu, 2011). Structural health monitoring
tures and time–frequency-based features were extracted (SHM) monitors the performance of a structural system
from measured vibration signals by Fast Fourier Trans- with an identical goal. In this field, nondestructive
form (FFT) and continuous wavelet transform (CWT) damage detection (NDD) techniques are of special
to form one-dimensional or two-dimensional patterns, interest in monitoring structures for possible damage.
respectively. Three pattern-matching algorithms includ- Basically, NDD techniques can be classified into either
ing correlation, least-square distance, and Cosh spec- local or global methods. Most currently used methods
tral distance were investigated for pattern matching. To such as ultrasonic, eddy-current, and thermal methods
demonstrate the validity of the approach, numerical and are visual or localized experimental methods that detect
experimental studies were conducted on a simple three- damage on or near the surface of the structure. Limita-
story steel building. Results showed that features of the tions of local methods are the need to know the vicinity
signal for different damage scenarios could be uniquely of the damage and accessibility of the portion of the
identified by these transformations, and suitable corre- structure being inspected. Chang and Liu (2003) pro-
lation algorithms could perform pattern matching that vided detailed information about “local” methods. The
identified both damage location and damage severity. need for global damage detection techniques has led to
Meanwhile, statistical issues for more complex structures development of vibration-based detection methods that
as well as the choice of wavelet functions are discussed. rely on the change of vibration characteristics and sig-
nals as indication of damage. Over the last two decades,
extensive research has been conducted on this detection
1 INTRODUCTION
approach, leading to various experimental techniques,
methodologies, and signal processing algorithms.
Civil structures are susceptible to damage over their
Doebling et al. (1996) and Sohn et al. (2003) presented
service life due to aging, environmental factors, fa-
comprehensive literature reviews of vibration-based
tigue, and excessive load. Health monitoring of civil
damage detection and health monitoring methods for
∗ To whom correspondence should be addressed. E-mail: asad@ structural and mechanical systems. These methods can
k-state.edu. be classified into parametric- or signal-based categories.


C 2012 Computer-Aided Civil and Infrastructure Engineering.

DOI: 10.1111/j.1467-8667.2012.00766.x
2 Qiao, Esmaeily & Melhem

Parametric-based methods use changes in measured frequency-domain methods employ Wigner–Ville

system or finite element (FE) modal parameters such as distribution and wavelet analysis to extract the signal
frequencies, damping parameters, stiffness parameters, features. Examples of this category are spectro-
and mode shapes as a sign of damage in a structure. gram, continuous wavelet transform (CWT) coeffi-
Parametric-based methods have been applied success- cients, wavelet packet energies, and wavelet entropy
fully to identify the dynamic properties of linearized and (Staszewski et al., 1997; Hera and Hou, 2004; Melhem
time-invariant equivalent structural systems (Kosmatka and Kim, 2003; Sun and Chang, 2002; Ren and Sun,
and Ricles, 1999; Ren and Roeck, 2002; Shi et al., 2000; 2008). A comparison between parametric-based meth-
Kim et al., 2003; Moaveni et al., 2009; Soyoz and Feng, ods and signal-based methods for damage detection in
2009; Jafarkhani and Masri, 2011). Various algorithms bridges can be found in Cruz and Salgado (2009).
such as Neural Networks, Wavelets, and Chaos The- As an enhancement for feature extraction, selec-
ory have been successfully used for damage detection, tion, and analysis, statistical pattern-recognition tech-
intelligent transportation systems, and smart structures niques are deeply integrated into signal-based damage
(Adeli and Jiang, 2009; Adeli and Kim, 2009). Wavelet- detection. Staszewski (2000) and Farrar et al. (2001)
and Hilbert-based approaches have been developed as presented the detailed descriptions of feature extrac-
enhanced techniques for parametric identification of tion, selection, and analysis based on pattern recogni-
nonlinear and time-variant systems (Staszewski, 1998; tion. Some cases of successful application of the pro-
Kijewski and Kareem, 2003; Yang et al., 2004; Huang cedure for damage detection can be found in Sohn
et al., 2005; Hou et al., 2006; Chen et al., 2006; Yan and et al. (2000), Sohn and Farrar (2001), Trendafilova
Miyamoto, 2006; Umesha et al., 2009). “These methods, (2001), Qiao et al. (2009), Fang et al. (2005), Jiang
however, depend strongly on the accuracy of the mea- et al. (2007), Jiang and Adeli (2005), and Adeli and
sured data. They cannot provide the required accuracy Jiang (2006). Compared with parametric-based meth-
and reliability needed for complex system identifica- ods, signal-based methods are particularly effective for
tions of real life structures due to complicated nonlinear large-scale structures due to their complicated nonlin-
nature of behavior of civil structures, and incomplete, ear behavior and the incomplete, incoherent, and noise-
incoherent, and noise-contaminated measurement of contaminated measurements of structural response un-
structural response under extreme loadings” (Adeli and der extreme loadings (Adeli and Jiang, 2006).
Jiang, 2006). In this study, a signal-based pattern-extraction and
Signal-based (or nonparametric-based) methods recognition method, using a number of signal transfor-
examine changes in the features derived directly mation and pattern-matching algorithms, is investigated
from measured time histories or their corresponding for damage detection. The vibration signals of a struc-
spectra through proper signal processing methods and ture excited by a dynamic excitation such as an impulse
algorithms to detect damage. These features may not load were decomposed by Fast Fourier Transform
represent any explicit physical-dynamic parameters. (FFT) or CWT for feature extraction. Two types of
Based on different signal processing techniques for pattern formed by normalized FFT magnitudes or CWT
feature extraction, these methods are classified into coefficients of the signal were used in this phase of the
time-domain methods, frequency-domain methods, and study. Three statistical algorithms, correlation, least-
time–frequency (or time-scale)-domain methods. Time- square distance, and Cosh spectral distance, were also
domain methods use linear and nonlinear functions investigated to perform pattern recognition separately.
of time histories to extract the signal features. Exam- A damage pattern database was developed analytically
ples of this category are Auto-Regressive (AR) model, by simulating various damage scenarios. Damage
Auto-Regressive with eXogenous inputs (ARX) model, location and level were identified simultaneously by
Auto-Regressive Moving Average (ARMA) model, best matching the unknown damage feature with that of
and Extended Kalman Filter (EKF) model (Sohn et al., known ones in the database. To show the applicability
2000; Sohn and Farrar, 2001; Nair et al., 2006; Yan of the method, numerical and experimental case studies
et al., 2004; Chen and Liu, 2010; Gul and Catbas, 2011). were conducted on a three-story steel structure. At the
Frequency-domain methods use Fourier analysis and first phase of the numerical study, a two-dimensional
cepstrum (the inverse Fourier transform of the loga- (2-D), three-story steel structure model was numerically
rithm of the Fourier spectra magnitude squared) analy- simulated and the method was applied to detect rep-
sis to extract features in a given time window. Examples resentative damage cases. The results encouraged the
of this category are frequency response functions authors to expand the study to a real three-story steel
(FRFs), frequency spectra, cross power spectra, power structure, for which a detailed finite element model was
spectra, and power spectral density (Tang et al., 1991; developed and tuned against the physical structure. The
Kamarthi and Pittner, 1997; Lee and Kim, 2007). Time– detailed finite element model of the structure using the
 Damage diagnosis in structures 3

commercially available finite element software, AN- detection of damage in terms of level and location, but
SYS, simulated the structural dynamic response excited fails if damage time is a factor in the algorithm. This in-
by an impulsive load, without damage, as well as under cludes various structural control systems, or methods in
different damage scenarios and, the recorded response which time is implemented in the algorithm to distin-
was processed using MATLAB. The normalized signal guish concurrent damages at different locations.
features from this detailed model, generated for the The detailed descriptions about mother wavelet and
base (healthy) structure, as well as various damage CWT can be found in Melhem and Kim (2003). In this
cases were collected in a database. The normalized study, the Daubechies 6 wavelet was used as the mother
signal features of the real structure under the same type wavelet. Meanwhile, other types of mother wavelets
of excitation for an unknown damage case, was then were also investigated.
compared against this database, using three different
pattern matching methods separately, to detect the
most probable damage case. 1.3 Pattern-recognition techniques
A pattern can be a set of features recorded as discrete
1.1 Fourier transform values forming a vector or matrix. The purpose of pat-
tern recognition is to implement the algorithms that op-
This is a frequency-based transform widely used in anal- erate on the extracted features and qualify the damage
ysis of linear systems. It decomposes a signal into sine state of the structure. In this study, three algorithms
waves of different frequencies, which sum to the orig- were used to perform pattern matching of the extracted
inal waveform, distinguishing different frequency sine features against the database to identify the damage lo-
waves and their respective amplitudes. cation and level (severity).
FFT is an efficient algorithm for calculating discrete Correlation analysis, as the first method calculates the
Fourier transform and its inverse by reducing the num- correlation value Ci j of two patterns (Posenato et al.,
ber of computations needed for N points from 2N2 to 2008). A correlation value of 1 indicates that the two
2N log2 N. FFT is of great importance to digital signal patterns are identical, a correlation value of −1 means
processing. It has been widely used to extract the fre- that they are diametrically opposite, and a correlation
quency response of structures and has successfully been value of 0 means that they are completely different. A
applied for fault detection in beam and rotating machin- closer value to 1 shows a closer match between the two
ery. However, it should be noted that Fourier transform patterns.
is not capable of preserving the information on time do- n
main. If there is a local oscillation representing a partic- k=1 (Si (k) − Si )(S j (k) − S j )
Ci j =   (1)
ular frequency in the signal, its location on the time do- n
(S i (k) − S i ) 2 n
(S j (k) − S j ) 2
k=1 k=1
main will be lost. Note that although a structure might
have been pushed into its nonlinear range of response The second method was the Least Square Distance
when damaged, the response due to the excitation used (LSD), which has been widely applied for system mod-
for damage detection will be linear, even if the dynamic eling and identification, speech recognition, and finger-
properties of the structure have been affected by dam- print identification. It is defined as
age. So FFT can theoretically be applied for damage de-  n  12
tection as outlined in this method. 
di j = (Si (k) − S j (k))2 (2)
k=1
1.2 Wavelet transform The least value shows a closer match and vice versa.
This is computationally similar to the FFT. However, The third method was the Cosh Spectral Distance
unlike the sine waves used in the FFT, the wavelet trans- (CSD), which gives an indication about the global
form decomposes a signal into a set of orthogonal ba- difference between two patterns (Trendafilova, 2001;
sic functions, also called mother wavelets. The mother Owen, 2003; Haritos and Owen, 2004). It is defined as
wavelets are typically chosen to have compact sup- n 
1  Si (k) Si (k) S j (k)
ports in both time and frequency domains, so that they Coij = − log +
have local time–frequency properties. This addresses 2n S j (k) S j (k) Si (k)
k=1
the aforesaid deficiency mentioned for FFT. In other 
S j (k)
words, the information on time and frequency will be − log −2 (3)
Si (k)
preserved, depending on the scale-time range used in
wavelet transformation, while the information on time where n is the number of vector points in the pattern;
is lost using FFT. FFT may serve as a suitable tool for Si (k) and S j (k) are the vector values of the patterns i
4 Qiao, Esmaeily & Melhem

Fig. 1. Flowchart of pattern recognition.

and j at point k; and Si and S j are the average values of 2 PRELIMINARY NUMERICAL
the patterns i and j, respectively. INVESTIGATIONS
If i is the unknown damage feature pattern, and j
is a known feature pattern in the database, then the A 2-D steel structure as shown in Figure 2 was nu-
highest correlation coefficient, the lowest LSD coeffi- merically simulated to compare the performance of
cient, and the lowest CSD coefficient indicate the most the aforesaid algorithms and illustrate the applicability
similar pattern in the database, which shows the un- of the proposed damage detection procedure. The
known case. Figure 1 shows the process of pattern- material had a mass density of 7.85 g/cm3 , modulus
recognition method for damage detection in this study. of elasticity E = 2 × 105 MPa, and Poisson ration 0.3.
It mainly includes five operation stages: numerical sim- The area of each floor cross-section was 258 cm2 , with
ulation of the dynamic response of the structure under a moment of inertia of 555 cm4 ; column cross-sectional
different known damage scenarios, signal processing area was 8.06 cm2 , with a moment of inertia of 0.27 cm4 .
and feature extraction and normalization, damage pat- This structure was modeled by ANSYS. The element
tern database construction, signal acquisition on a struc- type for floors and columns was 2-D elastic beam
ture with an unknown damage, and pattern matching to (beam3). The floors were rigid compared to columns.
find the most probable damage case from the database The ratio of unit nodal rotation moment of the floor
which indicates the damage location and severity. For to that of column was more than 1 × 103 . All of the
continuous structural monitoring, it is necessary to up- connections were assumed to be fixed. Therefore there
date the numerical model once damage has been found were a total of three noticeable horizontal degrees
to accurately represent the physical condition of the of freedom (DOFs) in the numerical structure. The
structure. damage was simulated by using the baseline model with
 Damage diagnosis in structures 5

In the second phase of the preliminary numerical

study, time–frequency-based features were extracted
by CWT. The acceleration signal was decomposed by
CWT and the extracted features were time-scale-based
CWT coefficients. The CWT is the inner product or
cross correlation of the signal f (t) with the scaled and
time-shifted wavelet ψa,b(t). Variable a determines the
amount of time scaling or dilation, it is referred to as
the scale or dilation variable. The value of “scale a”
is proportional to the reciprocal of the frequency. The
smaller the value of a, the more the bandpass shifts to
a higher frequency, implying that the CWT at small
scales contains information about signal f (t) at the
higher end of its frequency spectrum. The variable b
represents time shift or translation and “time b” is the
moment of the wavelet along the time axis. The CWT
coefficients show the similarity between the signal and
the scaled and shifted wavelet. The coefficients can
be plotted in 2-D contour with time on the horizontal
axis, scale on the vertical axis, and values given by
gray-scale colors. Figures 3a–3d show an example
of the CWT coefficients contours of acceleration
signals of the structure under the selected damage
cases.
Fig. 2. Two-dimensional three-story numerical structure. Lighter shading in the contour indicates a higher
wavelet coefficient value. Comparison of the four
various dynamic properties, that is, flexural stiffness, of figures shows that the time–frequency-based CWT
the damaged components. Various damage cases were coefficients were sensitive to different damage cases,
introduced by symmetrically reducing the column stiff- forming a 2-D pattern that presents a unique condition
ness at different stories to preserve the symmetry of the for a given damage case. Each coefficient vector in a
structure. For instance, damage case 0-40-20 refers to pattern was also normalized with respect to the square
a case where stiffness of the columns at the second and root of the sum of squares of the corresponding pattern.
third stories was reduced by 40% and 20%, respectively.
Transient dynamic analysis was performed by AN-
2.1 Damage pattern database construction
SYS to generate the dynamic response of the healthy
structure, and the response under different damage sce- As mentioned earlier, different damage levels and loca-
narios. The excitation was an impulse force of 0.2 kN tions were numerically simulated by changing the model
with 0.02 second duration acting at the very top cor- properties of the structure, that is, flexural stiffness, of
ner of the model (point A, Figure 2), and the numerical the damaged components. For demonstration, the dam-
acceleration response was recorded at the opposite top age level was set on a scale of 0% to 60% with incre-
corner (point B, Figure 2) for 2 seconds at a sampling ments of 20% at different locations. A total of 64 sets of
frequency of 250 Hz. damage cases, as shown in Table 2, including the base-
In a preliminary effort, frequency-based features line condition, were selected to represent the possible
were extracted by FFT. The frequencies and magni- structural damage conditions (level and location) for
tudes corresponding to the three peaks in each of the the sample structure.
FFT spectrums are listed in Table 1. The FFT magni- All of the 64 sets of simulated acceleration responses
tude vectors in frequency domain were selected as the were transformed by FFT and CWT into FFT magni-
sensitive features which also preserved the information tude vectors and CWT coefficient vectors, respectively.
of frequency shifting, forming a 1-D pattern, present- The resulting 64 sets of normalized FFT magnitude vec-
ing a unique damage condition. To eliminate the effects tors and 64 sets of CWT coefficient matrices form the
of possible variation of the other factors such as pulse representative damage feature patterns in the database.
intensity, each magnitude vector in a pattern was nor- It should be noted that theoretically, a much larger
malized with respect to the square root of the sum of number of damage cases could have been generated
squares of the corresponding pattern. by combining various levels and locations. However,
6 Qiao, Esmaeily & Melhem

Table 1
Peak values on the FFT spectrums

Peak 1 Peak 2 Peak 3

Damage case Frequency Magnitude Frequency Magnitude Frequency Magnitude

0-0-0 1.996 1,911.9 5.489 3,220.8 7.984 1,351.7
20-40-60 1.497 1,858 3.992 2,468.5 5.988 709.95
60-20-40 1.497 1,376.3 3.992 3,601.2 6.487 1,181.9
60-60-60 1.497 882.4 3.493 2,429.4 4.990 1,366.1

Fig. 3. CWT contours for (a) damage case 0-0-0 (baseline condition), (b) damage case 40-60-60, (c) damage case 60-40-60,
and (d) damage case 60-60-60.
 Damage diagnosis in structures 7

Table 2
Damage cases in database

Case no. Damage case Case no. Damage case Case no. Damage case Case no. Damage case Case no. Damage case
1 0-0-0 14 0-60-20 27 20-40-40 40 40-20-60 53 60-20-0
2 0-0-20 15 0-60-40 28 20-40-60 41 40-40-0 54 60-20-20
3 0-0-40 16 0-60-60 29 20-60-0 42 40-40-20 55 60-20-40
4 0-0-60 17 20-0-0 30 20-60-20 43 40-40-40 56 60-20-60
5 0-20-0 18 20-0-20 31 20-60-40 44 40-40-60 57 60-40-0
6 0-20-20 19 20-0-40 32 20-60-60 45 40-60-0 58 60-40-20
7 0-20-40 20 20-0-60 33 40-0-0 46 40-60-20 59 60-40-40
8 0-20-60 21 20-20-0 34 40-0-20 47 40-60-40 60 60-40-60
9 0-40-0 22 20-20-20 35 40-0-40 48 40-60-60 61 60-60-0
10 0-40-20 23 20-20-40 36 40-0-60 49 60-0-0 62 60-60-20
11 0-40-40 24 20-20-60 37 40-20-0 50 60-0-20 63 60-60-40
12 0-40-60 25 20-40-0 38 40-20-20 51 60-0-40 64 60-60-60
13 0-60-0 26 20-40-20 39 40-20-40 52 60-0-60

Table 3 matching to identify the damage case even when the

Test cases signal was highly contaminated with noise, and struc-
Multiple ture had a damping property slightly different from the
Single Multiple damage Highest damping ratio used in the database. As a sample of the
damage damage locations and damage damage cases, Figure 4 shows the correlation pattern-
location location severities severity recognition results for the test damage case 58-38-19
(G1) (G2) (G3) (G4) (environmental condition: damping & noise), by using
FFT and CWT pattern, respectively. The highest corre-
0-0-19 0-38-38 19-38-58 0-58-58
lation value was achieved for pattern with damage con-
0-19-0 38-0-38 19-58-38 58-0-58
19-0-0 38-38-0 38-19-58 58-58-0 dition of 60-40-20 in each pattern database, correctly
0-0-58 38-38-38 38-58-19 58-58-58 detecting the closest damage case in the database.
0-58-0 58-19-38 It should be noted that the process will be more
58-0-0 58-38-19 probabilistic and will require a statistical component
for more complicated structures, and the probability of
a certain damage scenario will be the outcome of the
algorithm.
considering the limitations of the simple test structure,
and as a preliminary step in exploring the approach,
the damage cases generated and stored in the database
were limited to 64. 3 EXPERIMENTAL TESTS AND VERIFICATION

3.1 Descriptions of test structure, impulse applicator,

2.2 Case studies and pattern matching and signal acquisition
Twenty damage cases listed in Table 3, slightly different Successful numerical implementation of the proposed
from identical cases in the database were analytically damage detection procedure encouraged the authors
simulated, and the corresponding dynamic response to experimentally validate it. As shown in Figure 5,
under the impulse excitation was numerically gener- the test structure was 91.4 cm tall and consisted of 3
ated. Gaussian white noise was added to the generated floors and 30 columns. Each floor was supported on 10
acceleration signals of the test cases to simulate the columns. The floors were steel plates with dimensions
condition of signal contaminated with noise. The signal- of 38.1 cm × 25.4 cm × 2.54 cm and the columns were
to-noise ratio (SNR) was 5 dB. The damping ratio steel flat bars with dimensions of 41.91 cm × 1.91 cm ×
(ζ ) was 2% when generating the structure dynamic 0.32 cm.
response with damping. To make the rigid connection between the steel slab
The pattern-recognition results for all of the test cases and the steel flat column, four pieces of steel angles
by using the three different matching algorithms show (0.64 cm × 3.18 cm × 3.18 cm; length: 25.4 cm) were
that correlation algorithm could best perform pattern welded on the two faces and on the short edges of the
8 Qiao, Esmaeily & Melhem

Fig. 4. Correlation matching for damage case 58-38-19, FFT and CWT pattern matching.

Fig. 5. Test structure.

floor plates; and two pieces of steel angles (0.64 cm × traveling on a circular path to its lowest position, where
3.18 cm × 3.18 cm; length: 25.4 cm) were welded on the it hit the third floor slab and then bounced off the struc-
top face and on the short edges of the foundation slab. A ture to create an impulsive force on the structure. The
total of 14 pieces of steel angle were used. The columns impact was mostly elastic; however, since the response
were connected to the angles’ vertical legs using four was normalized, the impulse magnitude did not affect
bolts. To prevent rotation and drift, the foundation slab the recognition process as long as it did not push the
was fixed to the laboratory strong floor. structure into nonlinear response range.
To apply a consistent impulse force on the structure, a MicroStrain, Inc.’s (Williston, VT, USA) G-Link was
steel ball with a diameter of 4.45 cm was used. The steel installed on the top of the third floor to sample and store
ball was magnetically adhered to the top of a frame. It the acceleration response. The sampled data were sent
was tied by a 52.07 cm chain to this frame so that when to the computer by a USB base station through a wire-
the magnet was turned off, the ball dropped 52.07 cm less connection. Agile-LinkTM software (MicroStrain,
 Damage diagnosis in structures 9

Table 4 3.3 Damage pattern database construction

Test cases in experimental study
Various damage cases were introduced by removing
Single Double Triple columns at different locations, which simulated the fail-
location location location ure of one or more columns in the structure. Sixty-four
damage damage damage damage cases including the baseline condition were
0-0-20 0-20-20 20-20-20 designed to represent possible structural damage condi-
20-0-0 20-0-20 20-20-40 tions. In this study, the numerical dynamic responses of
0-20-0 20-20-0 20-40-20 the structure under the 64 damage cases were simulated
0-0-40 40-40-0 20-60-20 by removing corresponding columns from the structure
0-40-0 0-40-40 40-20-20 FE model. The resulting 64 sets of normalized FFT
40-0-0 40-0-40 40-40-20 magnitude vectors and 64 sets of CWT coefficient
0-0-60 20-40-0 40-40-40 matrices formed the damage feature patterns in the
0-60-0 40-20-0 40-60-20
database.
60-0-0 40-0-20
0-20-40
0-40-20
3.4 Case studies and pattern matching
Twenty-eight damage cases were chosen to test as listed
Inc.) was used to communicate with G-Link and con-
in Table 4. Note that as mentioned earlier, each case is
figure data logging. The sampling rate was 2,048 Hz.
shown by the percentage of damage in the first, second,
and third stories. As an example, 20-60-20 denotes a
case where the strength of the columns has dropped
3.2 Numerical model simulation of the structure
20% in the first story, 60% in the second story, and
and its dynamic response
20% in the third story. The acceleration response of
A 3-D FE model of the test structure shown in Figure 5 the structure with each damage case was measured
was constructed by ANSYS. The ANSYS element types after application of the impulsive force on the structure
for floors and columns were shell63 and beam4, respec- by the ball. These acceleration signals were then de-
tively. Transient dynamic analysis was carried out to noised and transformed by FFT and CWT. The three
determine the dynamic response of the structure under pattern-matching algorithms were used for pattern
a time-varying load. The time-step was 0.000488 second recognition. The results show that both FFT and CWT
(1/2,048 second). The FE model for healthy condition transformations could preserve the damage informa-
was tuned against test data to fine-tune the value of tion enough for distinguished patterns in this study,
each parameter in the model. The tuned FE model and the correlation algorithm could perform a better
represented the structure’s baseline (healthy) condition pattern recognition. Figures 6 and 7 show part of these
and was used in setting damage pattern database. results.

Fig. 6. Correlation matching for damage case 0-20-20, FFT and CWT pattern matching.
10 Qiao, Esmaeily & Melhem

Fig. 7. Correlation matching for damage case 20-20-40, FFT and CWT pattern matching.

4 DISCUSSION AND CONCLUSIONS can be automated in terms of detection and continuous

fine-tuning of the model and database. The study can
Experimental tests and case studies further validated be advanced further by using a more complicated
the overall feasibility of the method for damage detec- structure.
tion. Fourier and especially wavelet transforms could It should be noted that in addition to variations in the
well extract and preserve the features of the signal dynamic response of a structure inflicted by damage,
under damage conditions. Since the CWT pattern dynamic properties such as damping ratio, stiffness, and
preserves the frequency and time sensitive features, mass are not deterministic values and are affected by
it results in higher pattern-matching resolution than numerous factors such as temperature and environmen-
FFT pattern. The choice of wavelet functions will tal conditions like humidity, etc. So, even if the model
affect the computing time and detection resolution. should be calibrated against these changes to enhance
In the experimental studies, all the available main the accuracy of the detection process, the values remain
mother wavelets were compared. Haar, Daubechies, probabilistic and there is a certain level of uncertainty
Symlets, and Gaussian wavelets had the best perfor- associated with them. Also, for large-scale structures
mance. Among the three algorithms used, correlation with a large number of details and redundancy, some
was the best at performing pattern matching, even damage patterns may overlap each other. Therefore, in
when the signal was contaminated with noise. How- addition to fine-tuning the model used for generation
ever, the numerical model must be carefully tuned of the damage database, a relatively large number of
to accurately represent the physical conditions of the damage cases are needed to specify the most probable
structure. damage case, once damage in the real structure occurs.
The potential advantages of this approach, as per the In a large structure, although major damage cases
results of this study, may be summarized as follows. The can be detected with a good probability, detection of
method can potentially be used with a single or limited some local damages that may lead to similar patterns
number of input/output signals. It can be used to detect in the dynamic response of the structure have a lower
multiple damage locations and also detect the severity probability of being distinctly detected. This issue
of damage at each damage location. It gives good accu- may be addressed by a multilayer detection process,
racy in the presence of noise. After the case database which needs more damage databases at local levels. The
is generated for the structure, no additional cases are method and algorithm in this study are elementary steps
needed to “train” the system. However, after any with promising results, which can have some merits
physical change including changes due to any possible when it comes to detection of major damages, not de-
damage, the database should be updated based on the tectable by visual inspection, in a large structure with a
new dynamic properties of the structure. Note that the good probability. The same procedure can be used with
method can be implemented in various layers, starting more input–output signals to enhance the resolution/
from global (the whole structure) and ending at a probability of damage detection for minor/local damage
structural member for a detailed detection. The process cases.
 Damage diagnosis in structures 11

REFERENCES Computer-Aided Civil and Infrastructure Engineering, 20,

408–23.
Adeli, H. & Jiang, X. (2006), Dynamic fuzzy wavelet neural Jafarkhani, R. & Masri, S. F. (2011), Finite element model
network model for structural system identification, Journal updating using evolutionary strategy for damage detec-
of Structural Engineering, ASCE, 132(1), 102–11. tion, Computer-Aided Civil and Infrastructure Engineering,
Adeli, H. & Jiang, X. (2009), Intelligent Infrastructure – Neu- 26(3), 207–24.
ral Networks, Wavelets, and Chaos Theory for Intelligent Jiang, X. & Adeli, H. (2005), Dynamic wavelet neural
Transportation Systems and Smart Structures, CRC Press, network for nonlinear identification of high rise build-
Taylor & Francis, Boca Raton, Florida, 440 p. ings, Computer-Aided Civil and Infrastructure Engineering,
Adeli, H. & Kim, H. (2009), Wavelet-Based Vibration Control 20(5), 316–30.
of Smart Buildings and Bridges, CRC Press, Taylor & Fran- Jiang, X., Mahadevan, S. & Adeli, H. (2007), Bayesian wavelet
cis, Boca Raton, FL, 238 pp. packet denoising for structural system identification, Struc-
Adewuyi, A. P. & Wu, Z. (2011), Vibration-based dam- tural Control and Health Monitoring, 14(2), 333–56.
age localization in flexural structures using normalized Kamarthi, S. & Pittner, S. (1997), Fourier and wavelet trans-
modal macrostrain techniques from limited measurements, form for flank wear estimation: a comparison, Mechanical
Computer-Aided Civil and Infrastructure Engineering, 26 Systems and Signal Processing, 11(6), 791–809.
(3), 154–72. Kijewski, T. & Kareem, A. (2003), Wavelet transforms for
Chang, P. C. & Liu, S. C. (2003), Recent research in nonde- system identification in civil engineering, Computer-Aided
structive evaluation of civil infrastructures, Journal of Ma- Civil and Infrastructure Engineering, 18, 339–55.
terials in Civil Engineering, ASCE, 15(3), 298–304. Kim, J. T., Ryu, Y. S., Cho, H. M. & Stubbs, N. (2003), Dam-
Chen, B. & Liu, W. (2010), Mobile agent computing paradigm age identification in beam-type structures: frequency-based
for building a flexible structural health monitoring sensor method vs mode-shape-based method, Engineering Struc-
network, Computer-Aided Civil and Infrastructure Engi- tures, 25, 57–67.
neering, 25(7), 504–16. Kosmatka, J. B. & Ricles, J. M. (1999), Damage detection in
Chen, S. L., Lai, H. C. & Ho, K. C. (2006), Identification of structures by modal vibration characterization, Journal of
linear time varying systems by Haar wavelet, International Structural Engineering, 125(12), 1384–92.
Journal of Systems Science, 37(9), 619–28. Lajnef, N., Rhimi, M., Chatti, K., Mhamdi, L. & Faridazar,
Cruz, P. J. S. & Salgado, R. (2009), Performance of vibration- F. (2011), Toward an integrated smart sensing system and
based damage detection methods in bridges, Computer- data interpretation techniques for pavement fatigue moni-
Aided Civil and Infrastructure Engineering, 24(1), 62–79. toring, Computer-Aided Civil and Infrastructure Engineer-
Cusson, D., Lounis, Z. & Daigle, L. (2011), Durability mon- ing, 26(7), 513–23.
itoring for improved service life predictions of concrete Lee, J. & Kim, S. (2007), Structural damage detection in the
bridge decks in corrosive environments, Computer-Aided frequency domain using neural networks, Journal of Intelli-
Civil and Infrastructure Engineering, 26(7), 524–41. gent Materials Systems and Structures, 18, 785–92.
Doebling, S. W., Farrar, C. R., Prime, M. B. & Shevitz, D. Melhem, H. & Kim, H. (2003), Damage detection in concrete
W. (1996), Damage Identification and Health Monitoring of by Fourier and wavelet analyses, Journal of Engineering
Structural and Mechanical Systems from Change in Their Mechanics, 129(5), 571–77.
Vibration Characteristics: A Literature Review, Report, LA- Moaveni, B., Conte, J. P. & Hemez, F. M. (2009), Uncertainty
13070-MS, US-900, Los Alamos National Laboratory, NM. and sensitivity analysis of damage identification results ob-
Fang, X., Luo, H. & Tang, J. (2005), Structural damage detec- tained using finite element model updating, Computer-
tion using neural network with learning rate improvement, Aided Civil and Infrastructure Engineering, 24(5), 320–34.
Computers and Structures, 83, 2150–61. Nair, K. K., Kiremidjian, A. S. & Law, K. H. (2006), Time
Farrar, C. R., Doebling, S. W. & Nix, D. (2001), Vibration- series-based damage detection and localization algorithm
based structural damage identification, Philosophical with application to the ASCE benchmark structure, Jour-
Transactions of Royal Society of London Series A: Math- nal of Sound and Vibration, 291, 349–68.
ematical, Physical and Engineering Science, London, 359, Owen, J. S. (2003), Monitoring the degradation of reinforced
131–49. concrete beams under cyclic loading, Key Engineering Ma-
Gul, M. & Catbas, F. (2011), Damage assessment with ambi- terials, 245–246, 307–14.
ent vibration data using a novel time series analysis method- Posenato, D., Lanata, F., Inaudi, D. & Smith, I. F. C. (2008),
ology, Journal of Structural Engineering, 137(12), 1518–26. Model-free data interpretation for continuous monitoring
Haritos, N. & Owen, J. S. (2004), The use of vibration data of complex structures, Advanced Engineering Informatics,
for damage detection in bridges: a comparison of system 22(1), 135–44.
identification and pattern recognition approach, Structural Qiao, L., Esmaeily, A. & Melhem, H. (2009), Structural dam-
Health Monitoring, 3(2), 141–63. age detection using signal pattern-recognition, Key Engi-
Hera, A. & Hou, Z. K. (2004), Application of wavelet ap- neering Materials, 400–402, 465–70.
proach for ASCE structural health monitoring benchmark Ren, W. X. & Roeck, G. D. (2002), Structural damage identifi-
studies, Journal of Engineering Mechanics, 130(1), 96–104. cation using modal data. I: Simulation verification, Journal
Hou, Z., Hera, A. & Shinde, A. (2006), Wavelet-based struc- of Structural Engineering, 128(1), 87–95.
tural health monitoring of earthquake excited structures, Ren, W. X. & Sun, Z. S. (2008), Structural damage identi-
Computer-Aided Civil and Infrastructure Engineering, 21, fication by using wavelet entropy, Engineering Structures,
268–79. 30(10), 2840–49.
Huang, C. S., Huang, S. L., Lin, C. I. & Su, W. C. (2005), Shi, Z. Y., Law, S. S. & Zhang, L. M. (2000), Structural dam-
A wavelet-based approach to identifying structural modal age detection from modal strain energy change, Journal of
parameters from seismic response and free vibration data, Engineering Mechanics, ASCE, 126(12), 1216–23.
12 Qiao, Esmaeily & Melhem

Sohn, H., Czarnecki, J. A. & Farrar, C. R. (2000), Structural Sun, Z. & Chang, C. C. (2002), Structural damage assessment
health monitoring using statistical process control, Journal based on wavelet packet transform, Journal of Structural
of Structural Engineering, 126(11), 1356–63. Engineering, ASCE, 128(10), 1354–61.
Sohn, H. & Farrar, C. R. (2001), Damage diagnosis using time- Tang, H., Cha, J. Z., Wang, Y. & Zhang, C. (1991), The princi-
series analysis of vibration signals, Smart Materials & Struc- pal of cepstrum and its application in quantitative fault di-
tures, 10(3), 446–51. agnostics of gears, in ASME Design Technical Conference,
Sohn, H., Farrar, C. R., Hemez, F. M., Shunk, D. D., Stine- Miami, FL, Vol. 38, 141–44.
mates, D. W. & Nadler, B. R. (2003), A Review of Struc- Trendafilova, I. (2001), Pattern recognition methods for dam-
tural Health Monitoring Literature: 1996–2001, Report, age diagnosis in structure from vibration measurements,
LA-13976-MS, Los Alamos National Laboratory, NM. Key Engineering Materials, 204–205, 85–94.
Soyoz, S. & Feng, M. Q. (2009), Long-term monitoring and Umesha, P. K., Ravichandran, R. & Sivasubramanian, K.
identification of bridge structural parameters, Computer- (2009), Crack detection and quantification in beams using
Aided Civil and Infrastructure Engineering, 24(2), wavelets, Computer-Aided Civil and Infrastructure Engi-
82–92. neering, 24(8), 593–607.
Staszewski, W. J. (1998), Identification of non-linear system Yan, A. M., Boe, P. D. & Golinval, J. C. (2004), Structural
using multi-scale ridges and skeletons of the wavelet trans- damage diagnosis by Kalman model based on stochastic
form, Journal of Sound and Vibration, 214(4), 639–58. subspace identification, Structural Health Monitoring, 3(2),
Staszewski, W. J. (2000), Advanced data pre-processing for 103–19.
damage identification based on pattern recognition, Inter- Yan, B. & Miyamoto, A. (2006), A comparative study
national Journal of Systems Science, 31(11), 1381–96. of modal parameter identification based on wavelet and
Staszewski, W. J., Worden, K. & Tomlinson, G. R. (1997), Hilbert-Huang transform, Computer-Aided Civil and In-
Time-frequency analysis in gearbox fault detection us- frastructure Engineering, 21, 9–23.
ing the Wigner-Ville distribution and pattern recogni- Yang, J. N., Lei, Y., Lin, S. & Huang, N. (2004), Hilbert-
tion, Mechanical Systems and Signal Processing, 11(5), Huang based approach for structural damage detection,
673–92. Journal of Engineering Mechanics, 130(1), 85–95.