Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Computers and Structures 158 (2015) 211–224

Contents lists available at ScienceDirect

Computers and Structures


journal homepage: www.elsevier.com/locate/compstruc

Adaptive reference updating for vibration-based structural health


monitoring under varying environmental conditions
Seung-Seop Jin a, Soojin Cho b, Hyung-Jo Jung a,⇑
a
Department of Civil and Environmental Engineering, KAIST, 291 Daehak-ro, Yuseong-gu, Daejeon 305-701, Republic of Korea
b
School of Urban and Environmental Engineering, Ulsan National Institute of Science and Technology (UNIST), Ulsan 689-798, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, a novel adaptive principal component analysis is proposed for identifying structural
Received 22 March 2014 damage under varying environmental conditions (i.e., both stationary and non-stationary conditions).
Accepted 8 June 2015 The proposed method is based on an adaptive reference scheme in multivariate statistical analysis.
Available online 29 June 2015
One crucial advantage of the proposed method is the ability of generative learning to capture the intrinsic
behavior of the normal condition. The proposed method is numerically tested with noise in measurement
Keywords: under environmental variations. Its performance is compared with conventional methods (i.e., the fixed
Vibration-based structural health
reference scheme). The numerical study demonstrates the superior performance of the proposed method
monitoring
Adaptive reference updating
for structural damage detection.
Environmental variation Ó 2015 Elsevier Ltd. All rights reserved.
Principal component analysis

1. Introduction damage-sensitive features [7,8]. Since the model-based method is


computationally expensive due to an iterative analysis of a com-
The performance of civil engineering structures (e.g., bridges, puter simulation model, the model-free method is well-suited for
buildings, etc.) under operational and environmental conditions the continuous and real-time applications of the early damage
may decrease over time because of the deterioration of structural detection.
materials and extreme events such as earthquakes. Conventionally, However, it is well recognized that the vibration-based
visual inspection has been widely conducted to detect damage and damage-sensitive features (e.g., natural frequency) depend not
evaluate the condition of existing structures. However, it is known only on damage but also on environmental conditions (e.g., tem-
to be subjective and inefficient for large and complex structures. perature) [9–14]. If the change of the environmental condition is
To address these limitations of visual inspection, vibration-based not taken into account, false-positive or false-negative damage
damage detection has been extensively studied over the last several diagnosis could occur, and consequently vibration-based damage
decades, because it is nondestructive and cost-effective to imple- detection may become unreliable. To address the environmental
ment measurement systems. effect, a data-normalization method is required. Data-normalization
The vibration-based damage detection approach can be distin- methods can be categorized as input-output data-normalization
guished into two categories according to whether or not computer and output-only data-normalization. The input-output data-
simulation models (e.g., finite element models) are used [1]: normalization method constructs a mathematical model to map the
model-based and model-free methods. The model-based method environmental variables (i.e., input) on vibration-based damage-
is typically an inverse problem to quantify the damage by updating sensitive features (i.e., output) [9,10,13]. On the other hand, the
a computer simulation model through comparing model prediction output-only data-normalization method requires only the vibration-
with direct measurements [2,3]. On the other hand, the model-free based damage-sensitive features. Once the vibration-based damage-
method is attractive in the sense that it requires only sensitive features are obtained from the normal condition, a
vibration-based damage-sensitive features as training data to multivariate statistical analysis is performed to account for the
define the reference model (i.e., baseline) [4–6]. In the method, pattern of the feature variation due to the environmental effect
the natural frequency, mode-shape, mode-shape curvature and [4–6,11,12].
modal flexibility are commonly used as the vibration-based In the case of large and complex structures such as bridges and
buildings, the output-only data-normalization method based on
multivariate statistical analysis would be more appropriate from
⇑ Corresponding author. the practical point of view, because it is quite challenging to fully
E-mail address: hjung@kaist.ac.kr (H.-J. Jung). represent the input-output relationship due to the difficulty in

http://dx.doi.org/10.1016/j.compstruc.2015.06.001
0045-7949/Ó 2015 Elsevier Ltd. All rights reserved.
212 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

Fig. 1. Procedure of PCA for data-normalization.

long-term monitoring. If the insufficient or invalid data set is used


for the fixed reference model, the performance of damage detec-
tion would be quite disastrous [14]. However, the issue related to
obtaining a valid set of the normal condition has not been widely
investigated yet, although the performance of output-only
data-normalization is strongly affected by the description of the
normal condition.
In this paper, an adaptive reference approach is proposed to
describe the normal condition properly. It enables to capture the
intrinsic system behavior by continuously updating the reference
model before detecting damage. The feasibility of the adaptive ref-
erence approach as an alternative of the conventional fixed refer-
ence approach in damage detection is investigated. To evaluate
Fig. 2. Adaptation to construct new data window in APCA.
the performance of the proposed method to the various situations,
two main factors are considered: (1) dynamic environmental vari-
ations (i.e., stationary and non-stationary temperature history) and
determining proper environmental variables and measurement
(2) intrinsic system behavior on environmental variations (i.e.,
locations in advance. In the output-only data-normalization
quasi-linear and bi-linear). In this regard, a series of the numerical
method, it is generally required to decompose a covariance matrix
studies with different damage scenarios are carried out as follows:
of the vibration-based damage-sensitive features over the normal
(1) quasi-linear behavior under stationary; (2) quasi-linear behav-
condition for establishing the multivariate statistical analysis.
ior under non-stationary; and (3) bi-linear behavior under
Therefore, it is very important to properly describe the normal
non-stationary. The rest of the paper is organized as follows.
condition.
Section 2 describes both conventional and adaptive reference
Previous studies have focused on normalization techniques (i.e.,
schemes in multivariate statistical analysis. Section 3 presents
multivariate statistical models) such as principal component (PCA)
the preliminary works related to the numerical study. Section 4
[11] and factor analysis [12]. These works used a fixed reference
shows the numerical study to evaluate the performances of both
data set from the normal condition to construct a multivariate sta-
methods under various system conditions with noise. In the end
tistical model. However, it was reported that the fixed reference
of Section 4, the computational efforts of both methods are com-
model was not appropriate to describe the normal condition if a
pared. The paper concludes with some discussion in Section 5.
system showed bi-linear behavior which was not stationary and
time-invariant [9,10]. If vibration-based damage-sensitive features
and environmental variations show bi-linear relationship around 2. Multivariate statistical analysis for data-normalization
0 °C and the vibration-based damage-sensitive features above
0 °C are used as the normal condition, the vibration-based 2.1. Principal component analysis (PCA)
damage-sensitive features below 0 °C may be misinterpreted as
the damage or the false-negative alarm (i.e., the invalid data set). Principal component analysis (PCA) is a multivariate statistical
Generally in real applications, there is no prior information on analysis method, also known as proper orthogonal decompo-
the system behavior at the initial phase, and the effects of environ- sition. In the method, a set of possibly correlated variables are
mental and operational conditions can be identified through converted into a set of linearly uncorrelated variables. For
S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224 213

Fig. 3. Comparison between CPCA and APCA.

Fig. 4. Numerical model.

data-normalization, PCA considers the environmental effect as hid- us denote by the 2-dimentional vector xk a set of the features iden-
den variables. The fundamental idea is that a variation of the struc- tified at time tk (k = 1, . . ., N) with the number of samples (N). It is
tural response due to the change of environmental conditions can assumed that environmental variations are mainly responsible
be accounted by a multivariate statistical analysis. For example, let for the dispersion of the features. In the training data set from
214 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

Table 1 PC 1) is expressed to the highest variation of the features, so it is


Material and geometrical properties. associated with the main environmental effects. On the other hand,
Properties Values the PC 2 expresses the effect of secondary factors which are uncor-
Mass density 7827 kg m3 related to the PC1. The procedure of PCA based on the fixed refer-
Poisson’s ratio 0.3 ence model is graphically illustrated in Fig. 1 and summarized as
Young’s modulus 200 GPa follows:
Total length 20 m (1 m@20EA)

Step 1: The PCs are extracted by singular value decomposition


of a covariance matrix of the features from the normal condi-
tion (Fig. 1(a)).
30
Step 2: The original features are projected into the major PC
Temperature (T)

(i.e., dimension reduction, Fig. 1(b)) and re-mapped to the orig-


20
inal spaces (i.e., dimension reconstruction, Fig. 1(c)).
10 Step 3: This projection and re-mapping using certain PCs leads
0 to the loss of information (i.e., E). The novelty is defined using
the Euclidean or Mahalanobis distances based on the loss of
-10
information (Fig. 1(d)).
100 200 300 400 500 600 700 800 Step 4: Statistical analysis is performed to construct a threshold
Sample (N) value for outlier statistics.
Step 5: Once a new feature is obtained, projection and remap-
Fig. 5. Time history of temperature.
ping of the new feature is performed by the PCs of the normal
condition. The novelty index of new data is used to detect the
damage.
Temperature (T)

35 The basic concept of PCA for data-normalization is to remove


30 the variance due to the environmental effect under the normal
25 condition. By removing this effect, the loss of information via pro-
20 jection and re-mapping is used to construct threshold values for
50 100 150 200 250 300 350 400 novelty detection. In the absence of damage, the outlier statistics
Sample (N) value of new data should remain at the same level within reference
data (i.e., below threshold). Damage should lead to departure of the
(a) Stationary environment variation (1~400th samples) feature from the level of reference data, and outlier statistics
Temperature (T)

30 increases significantly as shown in Fig. 1(e).


20
10
0
-10
2.2. Adaptive principal component analysis (APCA)
100 200 300 400 500 600 700 800
Sample (N)
Although many successful applications of PCA can be found in
(b) Non-stationary environment variation (1~889th samples)
various fields, it is obvious that the success of PCA is contingent
Fig. 6. Dynamic environmental variations. on how much accurately a normal model is described. In the case
of a large and complex structural system, however, it may not be
easy to describe the normal condition because the system exposes
large variations of ambient temperatures and operational loads
the normal condition, the major variability may be expressed by (i.e., live loads). In addition, abundant data sets to establish a PC
the combination of environmental effects. Therefore, the features model at the initial phase may not be available from the normal
from the normal condition vary mainly along the directions of condition, so that it would be insufficient to describe the normal
the principal components (PCs) with the highest variations. The condition. Furthermore, using the invalid data for a reference PC
two PCs are generated as shown in Fig. 1(a). The major PC (i.e., model may cause a more serious problem. If there is the invalid

1.06
1.25
1.05
Elastic Modulus (E)
Elastic Modulus (E)

1.2
1.04

1.03 1.15

1.02 1.1

1.01
1.05
1
1

-10 0 10 20 30 -10 0 10 20 30
Temperature (T) Temperature (T)
(a) Quai-linear behavior (b) Bi-linear behavior

Fig. 7. Temperature dependence for Young’s modulus.


S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224 215

Fig. 8. Flowchart of simulation for feature extraction.

Table 2 aggressive environmental variations (i.e., non-stationary) or com-


Damage scenarios.
plex intrinsic behavior corresponding to the environmental varia-
Scenario Process Intrinsic Damage tions such as the bi-linear behavior between temperature and
condition system natural frequencies [9,10], because a reference PC model from
Location and severity Occurrence
behavior
insufficient normal data set is no longer representative of the cur-
S#1a Stationary Quasi- No damage for false-alarm – rent behavior of a structural system. In such a situation, it is desir-
linear testing
S#1b 10% Reduction in stiffness 300th
able that the reference PC model is updated based on a new
of 10th and 11st elements observation to adapt the current behavior of the structural system.
S#2a Non- Quasi- No damage for false-alarm –
In order to settle the limitation of CPCA, adaptive PCA (APCA)
stationary linear testing has been developed mainly in the field of process monitoring
S#2b 10% Reduction in stiffness 700th [15–18]. The main difference between APCA and CPCA is to update
of 10th and 11st elements the reference PC model which represents the current system
S#3a Non- Bi-linear No damage for false-alarm – behavior by using the adaptive reference data set. APCA can be dis-
stationary testing tinguished into the recursive PCA (RPCA) and the moving window
S#3b 10% Reduction in stiffness 700th
PCA (MWPCA). RPCA is a recursive technique to update the refer-
of 10th and 11st elements
ence model for every new sample [15]. RPCA includes older sam-
ples that become increasingly unrepresentative of the current
data from a damaged state or outlier in the training data set, it may system behavior. For addressing this drawback, a forgetting factor
diagnose the damaged state as an undamaged state. is introduced to reduce the effect of older samples on the current
The conventional PCA (CPCA) based on the fixed reference system behavior. It is not easy to select this factor without a priori
would lead to numerous false-alarms under the dynamic knowledge. MWPCA has been developed to overcome this

0.99 0.995
0.95 0.99
0.9
0.75
0.95
0.5 0.9
Probability
Probability

0.25 0.75

0.1 0.5
0.05
0.25
0.1
0.01 0.05
0.005 0.01
0.005
1 2 3 4 5 6
1 2 3 4 5 6
-3
Residual x 10 Residual -3
x 10
(a) Extreme value distribution (b) Normal distribution

0.995
0.995 0.99
0.99
0.95
Probability

Probability

0.9
0.95
0.75
0.9
0.5
0.75
0.25
0.5 0.1
0.25 0.05
0.1 0.01
1 2 3 4 5 6 1 2 3 4 5 6
-3
Residual x 10 Residual -3
x 10
(c) Exponential distribution (d) Rayleigh distribution
Fig. 9. Probability plots of residuals from initial data set.
216 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

W/O noise With noise [SNR: 60dB]


Freq. #1 (Hz)

Freq. #1 (Hz)
2.87 2.87
2.86 2.86
2.85 2.85
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
Sample (N) Sample (N)
Freq. #2 (Hz)

Freq. #2 (Hz)
11.5 11.5
11.45 11.45
11.4 11.4

50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
Sample (N) Sample (N)
Freq. #3 (Hz)

Freq. #3 (Hz)
25.85 25.85
25.8 25.8
25.75 25.75
25.7 25.7
25.65 25.65
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
Sample (N) Sample (N)
Freq. #4 (Hz)

Freq. #4 (Hz)
45.9 45.9
45.8 45.8
45.7 45.7
45.6 45.6
45.5 45.5
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
Sample (N) Sample (N)
Freq. #5 (Hz)

Freq. #5 (Hz)

71.8 71.8
71.6 71.6
71.4 71.4
71.2 71.2
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
Sample (N) Sample (N)

Fig. 10. Simulated first 5 natural frequencies (S#1a).

limitation [16]. In the application of civil structural systems, the new observation is not an outlier, the oldest sample is dis-
Posenato et al. [4–6] proposed the damage detection method using carded from the old moving window and then new sample is
MWPCA and robust regression analysis (RRA) for continuous struc- added to new moving window. It updates the reference PC model
tural monitoring. It transforms measurement time series (i.e., by calculating the current covariance matrix from a new moving
quasi-static data) into eigenvector time series by the moving win- window as shown in Fig. 2.
dow. When the damage occurs, the mean and covariance matrix The proposed method based on APCA is as follows:
are changed, so that the values of eigenvalues and eigenvectors
are also changed. As the moving window is sliding to all the mea- Step 1: Sample the training data set (Xk = [x1 x2, . . ., xk]) to con-
surements, the damage can be identified by change of eigenvectors struct the initial PC model and compute the principle compo-
of the moving window. They used MWPCA to extract nent (Pk).
damage-sensitive features for damage detection. Step 2: Compute the residuals (Ek = Xk–XkPkPTk) between the
One challenge in implementing MWPCA is to select the optimal original data (Xk) and the reconstructed data (XkPkPTk) and esti-
length of the moving window. It should be determined according mate the threshold value (Ta,k) with the significant level a
to the change rate of the normal process which may not be known under a goodness-of-fit test for the probability distribution for
in advance. Instead of the fixed length of the moving window, He residuals. Superscript ‘‘T ’’ denotes the transpose.
and Yang [17] proposed the variable MWPCA method which Step 3: Collect the new data sample (xk + 1) and calculate the
changes the length of the moving window adaptively in accor- residual (Ek+1 = xk+1–xk+1PkPTk).
dance with the rate of the system change. Once the optimal length Step 4: If Ta,k < Ek+1, the PC model updating is stopped and an
of the moving window is adaptively selected, the additional chal- abnormal alarm is triggered. If Ta,k > Ek+1, update the PC model
lenging that there is not sufficient observed data may arise at the based on the new data set (Xk+1 = [x2, x3, . . ., xk+1]) and compute
initial phase of monitoring. One approach to address this issue is the principle component (Pk+1), update threshold value (Ta,k+1).
to combine MWPCA with RPCA [18]. RPCA is used at the initial Step 5: Go to Step 3 for evaluation of the new data (Xk+2).
phase of the monitoring since this method does not require a speci-
fic number of observed data (i.e., length of moving window). Once 2.3. Comparison between CPCA and APCA
enough observations have been obtained to fill the moving win-
dow of MWPCA, MWPCA is used. CPCA is reasonable to describe the process under the assump-
The proposed adaptive reference updating is based on APCA tion that a system is stationary and time-invariant, because a PC
(i.e., MWPCA) for both accurate description of the normal model is fixed. On the other hand, APCA has the adaptation ability
condition and damage detection. Contrary to Posenato’s study, of the intrinsic change in system behavior by updating the refer-
the proposed method is able to adaptively select and reject a ence PC model. Thanks to its adaptive selection for PC model
new measurement and update the reference model for data- update, APCA is applicable to both stationary/non-stationary sys-
normalization and damage detection. It uses the concept of the tems and time-variant/invariant systems. However, the computa-
moving window which slides along the sequential data stream. If tional effort of APCA is relatively heavy than that of CPCA, since
S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224 217

the iterative updating of the reference PC model is required to to simulate the non-stationary process (Fig. 6(b)). As shown in
describe the current normal condition as shown in Fig. 3. The com- Fig. 6(b), the temperature history has the non-stationary property
putational issue is comprehensively discussed in the Section 4.4. due to a seasonal variation (i.e., from summer to winter). It is found
that there is a drift of the mean value after the 300th sample, and
3. Preliminary works for numerical studies for data- the variance of the temperatures after the 680th sample has the
normalization relatively large compared to that of the first 200 samples.
As the intrinsic system behavior on environmental variations,
To evaluate the performance of the proposed method for dam- the temperature dependences to Young’s modulus are assumed
age detection under varying environmental conditions, a numeri- to be quasi-linear and bi-linear as shown in Fig. 7. It is assumed
cal model is constructed by using the Bernoulli–Euler beam that there is no temperature gradient, so that the temperature is
theory. This model is a simply supported beam with 20 elements distributed uniformly.
as shown in Fig. 4. Its material and geometrical properties are tab- The natural frequencies of the first 5 modes are considered as
ulated in Table 1. the feature for damage detection, since the natural frequency is
The time history of temperature is shown in Fig. 5. To obtain the more stable and accurate to estimate than mode-shapes and sensi-
realistic time history of temperature, the temperature record of tive to both damage and environmental conditions (i.e., tempera-
Seoul in South Korea from June 18th in 2012 (i.e., summer) to ture). To simulate the environmental effect, it is assumed that
January 24th in 2013 (i.e., winter) is used [19]. Total 889 samples Young’s modulus is dependent on the temperature [9–12]. The fea-
are obtained under the sampling rate of 6 h. Temperature ranges tures are computed as shown in Fig. 8. To account for the presence
from 16 to 35.5 °C. of the random or secondary effects (i.e., operational effect), the
For the numerical study, the following two main factors are simulated features (f i ) are perturbed by adding Gaussian noise
considered: (1) dynamic environmental variations; and (2) intrin- with the level b as shown in the Eq. (1). The signal-to-noise ratios
sic system behavior on environmental variations. For the dynamic (SNRs) are adjusted to 100 dB (b ¼ 0, no noise), 80 dB (b ¼ 0:0001,
environmental variations, the temperature history in Fig. 5 is used low noise) and 60 dB (b ¼ 0:001, high noise).
to describe the stationary and non-stationary environment. The _

first 400 samples are used to simulate the stationary condition f i ¼ f i þ b  RMSðf i Þ  Nð0; 1Þ ð1Þ
(Fig. 6(a)), while the whole samples (i.e., 889 samples) are used

Conventional PCA Adaptive PCA


0.3

0.25
0.25
Euclidean distance (E)

Euclidean distance (E)

0.2
0.2

0.15 0.15

0.1 0.1

0.05 0.05

50 100 150 200 250 300 350 400 50 100 150 200 250 300 350
Sample (N) Sample (N)
(a) False alarm test of CPCA (S#1a) (b) False alarm test of APCA (S#1a)
Conventional PCA Adaptive PCA
0.7
0.7

0.6
0.6
Euclidean distance (E)

Euclidean distance (E)

0.5
0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350
Sample (N) Sample (N)
(c) Damage detection test of CPCA (S#1b) (d) Damage detection test of APCA (S#1b)
(Damage at 300th sample) (Damage at 300th sample)

Fig. 11. Results of numerical study #1 with SNR of 60 dB.


218 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

Freq. #1 (Hz)

Freq. #1 (Hz)
W/O noise With noise [SNR: 60dB]
2.9 2.9

2.85 2.85
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #2 (Hz)

Freq. #2 (Hz)
11.8 11.8

11.6 11.6

11.4 11.4
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #3 (Hz)

Freq. #3 (Hz)
26.4 26.4
26.2 26.2
26 26
25.8 25.8
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #4 (Hz)

Freq. #4 (Hz)
47 47
46.5 46.5
46 46

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #5 (Hz)

Freq. #5 (Hz)

73.5 73.5
73 73
72.5 72.5
72 72
71.5 71.5
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)

Fig. 12. Simulated first 5 natural frequencies (S#2a).

4. Numerical study Section 4.4. The simulated natural frequencies under the two
SNRs of 100 dB (i.e., no noise) and 60 dB without damage are plot-
A total of three scenarios are constructed based on the two main ted in Fig. 10. It is clearly observed from the figure that the natural
factors as follows: (1) quasi-linear behavior under the stationary frequencies of SNR of 60 dB are more dispersed due to the noise
environment (S#1); (2) quasi-linear behavior under the effect.
non-stationary environment (S#2); and (3) bi-linear behavior under Fig. 11 shows the results of both the false-alarm and damage
the non-stationary environment (S#3). As tabulated in Table 2, each detection tests. The Euclidean distances from CPCA and APCA are
scenario consists of the false alarm and damage detection tests to plotted with the threshold value (black dotted line). The fixed refer-
compare the performance of the proposed APCA-based damage ence of CPCA is described by the green line. The sample beyond the
detection method and that of the CPCA-based method. threshold value is marked with a red circle. Theses notations are also
For brevity and comparison, the first 200 samples (i.e., the 1st to applied to the following figures such as Figs. 11, 13, 15 and 16. As
200th samples) are used as the fixed reference data of CPCA and shown in Fig. 11, both CPCA-based and APCA-based damage detec-
the length of the moving window in ACPA. For all the scenarios, tion methods are successful to interpret the healthy condition and
the number of principal components is chosen by the cumulative identify the damage after the 300th sample. It figures out that both
percent variance (CPV) method [20]. The predetermined value is methods have the similar performance in this case.
95% of CPV. The Euclidean distance is used as a novelty index for
further numerical studies, since the Mahalanobis distance may 4.2. Scenario #2: quasi-linear behavior under the non-stationary
not be appropriate under a non-stationary condition due to the environment
time-variant covariance matrix of residuals (i.e., CPCA). For the
estimation of the threshold value, the Kolmogorov–Smirnov test In this scenario case, both methods are investigated under the
(K–S test) is performed to estimate the probability distribution dynamic aggressive environmental variations (i.e., non-stationary
function (PDF) of residuals. Four PDFs are considered as (1) condition). The whole samples (i.e., a total of 889) are used to sim-
Extreme value, (2) Normal, (3) Exponential, and (4) Rayleigh distri- ulate the non-stationary process as shown in Fig. 12. Similarly to
butions. The one of the probability plots are shown in Fig. 9. It is Section 4.1, the results of the numerical study from SNR of 60 dB
observed that Rayleigh distribution properly describes the popula- are illustrated.
tion of the residuals in all the scenario cases. The threshold value is As shown in Fig. 13(b) and (d), it is clearly observed that the
estimated by the one-sided 99.7% (3r) confidence level. threshold value is adaptively constructed in the APCA-based dam-
age detection method. The APCA-based method updates the refer-
4.1. Scenario #1: quasi-linear behavior under the stationary ence condition by adapting or rejecting the new observation, once
environment the non-stationary behavior in temperature history appears. The
CPCA-based damage detection method diagnoses the false-alarms
This damage scenario is the simplest one among all the scenar- due to the non-stationary condition (i.e., relatively larger variation)
ios. For brevity, the results of the numerical study from SNR of after the 700th sample, while the APCA-based method results in
60 dB are illustrated. The other results are tabulated in much less false-alarms.
S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224 219

Conventional PCA Adaptive PCA


0.7
0.35
0.6
0.3
Euclidean distance (E)

Euclidean distance (E)


0.5
0.25

0.4
0.2

0.3
0.15

0.2
0.1

0.1 0.05

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
(a) False alarm test of CPCA (S#2a) (b) False alarm test of APCA (S#2a)

Conventional PCA Adaptive PCA

1 1

0.9 0.9
Euclidean distance (E)

0.8
Euclidean distance (E)
0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
(c) Damage detection test of CPCA (S#2b) (d) Damage detection test of APCA (S#2b)
(Damage at 700th sample) (Damage at 700th sample)

Fig. 13. Results of numerical study #2 with SNR of 60 dB.

The CPCA-based damage detection method may identify the For the false-alarm test, the CPCA-based damage detection
damage by the relatively large deviation of the Euclidean distance method generates many false-alarms in both different SNRs. It is
at the 700th sample. However, there are false-alarms around the very unreliable to interpret the healthy condition. In contrast to
680th sample before the damage occurs at the 700th sample. the CPCA-based method, the APCA-based method properly inter-
These false-alarms result in unnecessary expenditure and the cred- prets the normal condition by reflecting the natural change of
ibility of the damage detection method is getting lacked. On the the intrinsic system behavior under the non-stationary
other hand, the APCA-based damage detection method successfully environment.
detects the damage with much less false-alarms. Despite no noise effect (i.e., 100 dB of SNR), the CPCA-based
damage detection method misinterprets the damage (i.e.,
4.3. Scenario #3: bi-linear behavior under the non-stationary false-alarms) as similar to Fig. 13(c). In real applications, these
environment false-alarms may be misinterpreted as the damage, since they
are misinterpreted continuously after the 680th sample. In the
According to the bi-linear behavior as shown in Fig. 7(b), there SNR of 60 dB, the CPCA-based method does not identify the dam-
are two slopes in the relation between temperature and Young’s age by masking in the false-alarms. On the other hand, the
modulus. The simulated frequencies above 0 °C (i.e., the first 200 APCA-based damage detection method is successful to identify
samples) are used for the description of the normal condition at the damage without false-alarms regardless of the SNRs.
the initial phase, so that the initial samples represent the normal
condition above 0 °C partially and both methods are lacked to rep-
resent the behavior below 0 °C. Furthermore, non-stationary envi- 4.4. Summary and discussion
ronment makes more challenging to interpret the healthy
condition and identify the damage. Fig. 14 shows the simulated To consider the randomness of noise, sensitivity with respect to
natural frequencies. It is observed that the natural frequencies sig- the noise ratio is analyzed. Under the three different SNRs (i.e., 60,
nificantly increase from the 680th to 870th samples, once the 80, 100 dB), each scenario is repeated five times. Then, the false
intrinsic behavior of the system changes. alarm rate is computed as a measure of the performance.
220 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

Freq. #1 (Hz)

Freq. #1 (Hz)
3 3
W/O noise With noise [SNR: 60dB]
2.9 2.9

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #2 (Hz)

Freq. #2 (Hz)
12 12

11.5 11.5

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #3 (Hz)

Freq. #3 (Hz)
26.8 26.8
26.6 26.6
26.4 26.4
26.2 26.2
26 26
25.8 25.8
25.6 25.6
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #4 (Hz)

Freq. #4 (Hz)
48 48
47 47
46 46
100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
Freq. #5 (Hz)

Freq. #5 (Hz)

74 74

72 72

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)

Fig. 14. Simulated first 5 natural frequencies (S#3a).

Table 3
False-alarm rate in Scenario #1.

Signal-to-noise ratio Type False-alarm rate from the 201st to 400th samples (%)
1 2 3 4 5 Mean Standard deviation
S#1a: False-alarm test
SNR 100 dB (No noise) CPCA 0 0 0 0 0 0 0
APCA 0 0 0 0 0 0 0
SNR 80 dB (Low noise) CPCA 0.5 0 0 1 0 0.3 0.45
APCA 0.5 0 0 0.5 0 0.2 0.27
SNR 60 dB (High noise) CPCA 0.5 0.5 2 1 0.5 0.9 0.65
APCA 0.5 0 2.5 1 0.5 0.9 0.96

False-alarm rate from the 201st300th samples (%)


S#1b: Damage detection test
SNR 100 dB (No noise) CPCA 0 0 0 0 0 0 0
APCA 0 0 0 0 0 0 0
SNR 80 dB (Low noise) CPCA 0.5 0 1 0.5 0 0.4 0.42
APCA 0.33 0 0.67 0.33 0 0.27 0.28
SNR 60 dB (High noise) CPCA 0 0 0.5 0 0.5 0.2 0.27
APCA 0 0 0.5 0 0.5 0.2 0.27

At first, the performance of both methods is evaluated under a small sensitivity to noise under stationary condition, since
stationary dynamic environmental variation with the quasi-linear the higher noise ratio does not lead to an increase of false alarm
behavior (S#1). False-alarm rates of S#1a are calculated from the rate.
201st to 400th sample, while false-alarm rates of S#1b are Under non-stationary dynamic environmental variation with
calculated from the 201st to 300th samples, since both methods the quasi-linear behavior (S#2), false-alarm rates of S#2a are cal-
successfully identify the damage state after the 300th samples. culated from the 201st to 889th sample and those of S#2b are cal-
Both methods are successful to interpret the normal condition culated from the 201st to 700th sample. In the false alarm test
and damage detection. According to Table 3, both methods have (S#2a), the CPCA-based damage detection method generates many
S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224 221

Conventional PCA -3 Adaptive PCA


x 10
0.03
8
Euclidean distance (E) 0.025 7

Euclidean distance (E)


6
0.02
5
0.015
4

0.01 3

2
0.005
1

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
(a) Conventional PCA (SNR = 100 dB) (b) Adaptive PCA (SNR = 100 dB)
Conventional PCA Adaptive PCA
2
0.35
1.8

1.6 0.3
Euclidean distance (E)

Euclidean distance (E)


1.4
0.25
1.2

1 0.2

0.8 0.15
0.6
0.1
0.4

0.2 0.05

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
(c) Conventional PCA (SNR = 60 dB) (d) Adaptive PCA (SNR = 60 dB)

Fig. 15. False-alarm test with different SNRs.

false-alarms around the 700th sample due to non-stationary prop- As found in Tables 3–5, the APCA-based damage detection
erties. It means that the fixed reference PC model is not working method is more robust with respect to the noise effect than the
properly in the non-stationary environment, since their basic CPCA-based one. Considering that the noise effect in real applica-
assumption is that a system is stationary and time-invariant. On tions is inevitable to the measurements, the robustness to the
the other hand, the APCA-based method has still relatively small noise effect is important. Based on the previous results, the perfor-
false-alarms than the CPCA-based one. In the damage detection mances of both methods are summarized in Table 6. It is expected
test (S#2b), both methods can detect the damage under SNR that the APCA-based method can apply to various situations (i.e.,
100 dB (i.e., no noise). However, the CPCA-based damage detection similarly to Scenarios #1–3) successfully by obtaining the valid
method generates more false-alarms before damage as shown in data set of the normal condition adaptively.
Fig. 15, as the noise ratio increases. According to Table 4, the The computational effort is an important concern in algorithms.
CPCA-based method is not reliable under the non-stationary envi- The main computational difference between the CPCA-based and
ronment, while the APCA-based method still works successfully APCA-based methods is the iterative update of the reference
regardless of the noise effect. model. The number of iterative update is depending on the number
In the last case, the performance of both methods is evaluated of initial data (i.e., the length of the moving window in APCA) and
under the non-stationary environment with the bi-linear behavior the total number of data. These are used for the parametric study
(S#3). The samples used for false-alarm rates are similar to S#2. to evaluate the computational effort. The computation of each
The APCA-based damage detection method is successful to inter- method has been performed by MATLAB 2014a with the identical
pret the change of the normal condition under the insufficient computer (IntelÒ Core™ i7-950 processor running with 3.06 GHz
description of normal condition at the initial phase (S#3a), while with Window 7 OS). To estimate the average computational time,
the CPCA-based method leads to disastrous results with numerous a total of 30 simulations are performed for each method. The esti-
false alarms. In the damage detection test (S#3b), the CPCA-based mated computing time from this parametric study is shown in
damage detection method are very disappointing to identify the Fig. 17. As seen from the figure, the computational time of the
damage. On the other hand, the APCA-based method is successful APCA-based method (0.41–1.56 s) is much heavier than those of
to detect the damage with small false-alarms. According to the CPCA-based method (0.0027–0.055 s), since the APCA-based
Table 5, the CPCA-based damage detection method is not reliable method has the recursive framework to update the reference PC
under the non-stationary environment, while the APCA-based model. Considering that APCA in this study is based on singular
method still works successfully regardless of the noise effect. value decomposition, its computational time can be significantly
222 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

Conventional PCA Adaptive PCA


1.2 1.2

1 1
Euclidean distance (E)

Euclidean distance (E)


0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2
600 700

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
(a) Conventional PCA (SNR = 100 dB) (b) Adaptive PCA (SNR = 100 dB)
Conventional PCA Adaptive PCA
1.2

2
1
Euclidean distance (E)

Euclidean distance (E)


1.5 0.8

0.6
1

0.4

0.5
0.2

100 200 300 400 500 600 700 800 100 200 300 400 500 600 700 800
Sample (N) Sample (N)
(c) Conventional PCA (SNR = 60dB) (d) Adaptive PCA (SNR = 60 dB)
Fig. 16. Damage detection with different SNRs (Scenario #6).

Table 4
False-alarm rate in Scenario #2.

Signal-to-noise ratio Type False-alarm rate from the 201st to 889th samples (%)
1 2 3 4 5 Mean Standard deviation
S#2a: False-alarm test
SNR 100 dB (No noise) CPCA 14.80 18.43 14.80 15.82 13.64 15.50 1.81
APCA 0.87 0.15 0.29 0.29 0 0.32 0.33
SNR 80 dB (Low noise) CPCA 11.18 14.66 10.89 22.06 19.01 15.56 4.90
APCA 1.45 0.73 0.29 1.16 0.87 1.57 1.68
SNR 60 dB (High noise) CPCA 22.79 28.01 20.46 27.58 24.82 24.73 3.19
APCA 0.73 1.16 1.02 1.16 1.60 1.13 0.32

False-alarm rate from the 201st700th samples (%)


S#2b: Damage detection test
SNR 100 dB (No noise) CPCA 0.58 1.16 0.73 0.58 1.31 0.87 0.34
APCA 0.77 0.51 1.03 0.29 0.52 0.82 0.33
SNR 80 dB (Low noise) CPCA 2.47 1.16 1.74 1.02 2.90 1.86 0.82
APCA 1.29 1.29 1.80 0.77 0.26 1.08 0.59
SNR 60 dB (High noise) CPCA 3.34 2.61 6.68 7.11 4.93 4.93 1.98
APCA 1.03 1.03 2.32 0.52 0.77 1.13 0.70

reduced by using rank-one modification [15–18]. It is also noted those of the acquisition and modal identification (i.e., feature
that the computational time of both methods is much smaller than extraction).
S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224 223

Table 5
False-alarm rate in Scenario #3.

Signal-to-noise ratio Type False-alarm rate from the 201st to 889th samples (%)
1 2 3 4 5 Mean Standard deviation
S#3a: False-alarm test
SNR 100 dB (No noise) CPCA 21.05 19.01 18.43 21.77 19.30 19.91 1.42
APCA 2.18 2.18 2.47 1.45 2.62 2.18 0.45
SNR 80 dB (Low noise) CPCA 19.74 20.03 18.29 20.46 22.93 20.29 1.69
APCA 0.44 1.31 1.02 0.73 1.45 0.99 0.41
SNR 60 dB (High noise) CPCA 34.40 32.66 32.95 28.01 35.56 32.71 2.88
APCA 2.33 1.16 2.18 5.67 0.58 2.38 1.97

False-alarm rate from the 201st to 700th samples (%)


S#3b: Damage detection test
SNR 100 dB (No noise) CPCA 1.31 2.32 2.03 2.47 1.60 1.94 0.49
APCA 0.77 1.8 1.29 1.03 0.77 1.13 0.43
SNR 80 dB (Low noise) CPCA 2.61 2.18 1.88 2.47 3.19 2.47 0.49
APCA 1.55 0.77 1.03 1.29 0.77 1.08 0.34
SNR 60 dB (High noise) CPCA 9.14 9.58 11.76 9.29 8.56 9.67 1.23
APCA 2.84 2.06 3.61 2.06 2.06 2.53 0.69

Table 6
Summary of numerical simulations.

False-alarm Stationary, quasi-linear Non-stationary, quasi-linear Non-stationary, bi-linear


S#1a S#2a S#3a
SNR 100 dB CPCA No false-alarms False-alarms False-alarms
APCA No false-alarms No false-alarms No false-alarms
SNR 80 dB CPCA No false-alarms False-alarms False-alarms
APCA No false-alarms No false-alarms No false-alarms
SNR 60 dB CPCA No false-alarms False-alarms False-alarms
APCA No false-alarms No false-alarms No false-alarms

Damage detection S#1b S#2b S#3b


SNR 100 dB CPCA Distinct detection Distinct detection Faint detection
APCA Distinct detection Distinct detection Distinct detection
SNR 80 dB CPCA Distinct detection Distinct detection Faint detection
APCA Distinct detection Distinct detection Distinct detection
SNR 60 dB CPCA Distinct detection Faint detection Undetected
APCA Distinct detection Distinct detection Distinct detection

environmental conditions. The proposed method has the ability


of adaptive learning to capture the intrinsic behavior of the normal
condition, while the conventional approach (i.e., the CPCA-based
damage detection method) does not have. Based on the numerical
studies, it is demonstrated that the proposed APCA-based damage
detection method is able to capture the intrinsic characteristics of
the normal condition and detect the damage with much less false
alarms. The false alarm rates of the APCA-based method do not
increase, even though the noise ratio increases. From the perspec-
tive of the computational effort, the APCA-based method is much
heavier than the CPCA-based method due to its iterative update
of the reference model. However, this is negligible in the real appli-
cation, because the computational time of the APCA-based method
is much smaller than the time for feature extraction (i.e., modal
(a) CPCA (b) APCA identification). Therefore, the proposed method can be effectively
used for detecting the structural damage in practice.
Fig. 17. Comparison of computational efforts required by both methods.

Acknowledgements
5. Conclusions
This research was partially supported by a Grant (13SCIPA01)
In this paper, the APCA-based damage detection method is from Smart Civil Infrastructure Research Program funded by
proposed for identifying structural damage under varying Ministry of Land, Infrastructure and Transport (MOLIT) of Korea
224 S.-S. Jin et al. / Computers and Structures 158 (2015) 211–224

government and Korea Agency for Infrastructure Technology [9] Peeters B, De Roeck G. One-year monitoring of the Z24-Bridge: environmental
effects versus damage events. Earthq Eng Struct Dyn 2001;30:149–71.
Advancement (KAIA) and the New & Renewable Energy of the
[10] Moser P, Moaveni B. Environmental effects on the identified natural
Korea Institute of Energy Technology Evaluation and Planning frequencies of the Dowling Hall Footbridge. Mech Syst Signal Process
(KETEP) Grant funded by the Korea government Ministry of 2011;25:2336–57.
Knowledge Economy (No. 20123030020090). [11] Yan AM, Kerschen G, De Boe P, Golinval J-C. Structural damage diagnosis under
varying environmental conditions – Part I: a linear analysis. Mech Syst Signal
Process 2005;19:847–64.
References [12] Deraemaker A, Reynders E, De Roeck G, Kulla J. Vibration-based structural
health monitoring using output-only measurements under changing
[1] ASCE SEI Committee. Structural identification of constructed systems: environment. Mech Syst Signal Process 2008;22:33–56.
approaches, methods, and technologies for effective practice of St-ID. Reston, [13] Sohn H, Dzwonczyk M, Straser EG, Kiremidjian AS, Law KH, Meng T. An
Virginia; 2013. experimental study of temperature effect on modal parameters of the Alamosa
[2] Mottershead JE, Link M, Friswell MI. The sensitivity method in finite element Canyon Bridge. Earthq Eng Struct Dyn 1999;28:879–97.
model updating: a tutorial. Mech Syst Signal Process 2011;25:2275–96. [14] Manson G, Pierce G, Worden K, Monnier T, Guy P, Atherton K. Long-term
[3] Rober-Nicoud Y, Raphael B, Smith IFC. System identification through model stability of normal condition data for novelty detection. Proc SPIE
composition and stochastic search. J Comput Civil Eng 2005;19:239–47. 2000:324–34.
[4] Laory I, Trinh TN, Posenato D, Smith IFC. Combined model-free data- [15] Li W, Yue H, Valle-Cervantes S, Qin SJ. Recursive PCA for adaptive process
interpretation methodologies for damage detection during continuous monitoring. J Process Control 2000;10:471–86.
monitoring of structures. J Comput Civil Eng 2013;27:657–66. [16] Wang X, Kruger U, Irwin GW. Process monitoring approach using fast moving
[5] Posenato D, Lanata F, Inaudi D, Smith IFC. Model-free data interpretation for window PCA. Ind Eng Chem Res 2005;44:5691–702.
continuous monitoring of complex structures. Adv Eng Inform [17] He XB, Yang YP. Variable MWPCA for adaptive process monitoring. Ind Eng
2008;22:135–44. Chem Res 2008;47:419–27.
[6] Posenato D, Kripakaran P, Inaudi D, Smith IFC. Methodologies for model-free [18] Jeng J-C. Adaptive process monitoring using efficient recursive PCA and
data interpretation of civil engineering structures. Comput Struct moving window PCA algorithms. J Taiwan Inst Chem Eng 2010;41:475–81.
2010;88:467–82. [19] National Climate Data Service System, Database for Korean temperature
[7] Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage identification and records. <http://sts.kma.go.kr/jsp/home/contents/statistics/
health monitoring of structural and mechanical systems from changes in their newStatisticsSearch.do?menu=SFC>.
vibration characteristics: a literature review. Technical Report LA-13070-MS, [20] Malinowski FR. Factor analysis in chemistry. New York: Wiley-Interscience;
Los Alamos National Laboratory. New-Mexico; 1996. 1991.
[8] Carden EP, Fanning P. Vibration based condition monitoring: a review. Struct
Health Mon 2004;3:355–77.

You might also like