Professional Documents
Culture Documents
A Hybrid Learning Strategy For Structural Damage Detection - Flavio Barbosa e Alexandre Cury
A Hybrid Learning Strategy For Structural Damage Detection - Flavio Barbosa e Alexandre Cury
Abstract
Over the past decades, several methods for structural health monitoring have been developed and employed in various
practical applications. Some of these techniques aimed to use raw dynamic measurements to detect damage or struc-
tural changes. Desirably, structural health monitoring systems should rely on computational tools capable of evaluating
the information acquired from the structure continuously, in real time. However, most damage detection techniques fail
to identify novelties automatically (e.g. damage, abnormal behaviors, and among others), rendering human decisions nec-
essary. Recent studies have shown that the use of statistical parameters extracted directly from raw time domain data,
such as acceleration measurements, could provide more sensitive responses to damage with less computational effort.
In addition, machine learning techniques have never been more in trend than nowadays. In this context, this article pro-
poses an original approach based on the combination of statistical indicators—to characterize acceleration measure-
ments in the time domain—and computational intelligence techniques to detect damage. The methodology consists in
the combined use of supervised (artificial neural networks) and unsupervised (k-means clustering) learning classification
methods for the construction of a hybrid classifier. The objective is to detect not only structural states already known
but also dynamic behaviors that have not been identified yet, that is, novelties. The main purpose is to allow a real-time
structural integrity monitoring, providing responses in an automatic and continuous way while the structure is under
operation. The robustness of the proposed approach is evaluated using data obtained from numerical simulations and
experimental tests performed in laboratory and in situ. Results achieved so far attest a promising performance of the
hybrid classifier.
Keywords
Structural health monitoring, damage detection, supervised learning, unsupervised learning, feature extraction
engineering are of great interest to researchers in this be more effective compared to using acceleration
area due to the possible social and economic conse- responses. Moreover, the misclassification rates
quences caused by anomalies in structures. It is known increased as the damage severity decreased. Even
that if damage is not detected in time, the structural though their approach appeared to be promising, the
system can suffer serious consequences related to its number of false alarms prevented the method from
safety.8–10 being generalized for all practical damage detection
Currently, continuous monitoring systems are being applications. In another experimental study, Hakim
used mostly in large structures. As examples, one can et al.22 used the I-beam structures to predict the severity
cite the Millau Viaduct in France,11 the Tsing Ma and location of several damage scenarios. Accelerations
bridge in China,12 the Burj Khalifa skyscraper in along the beam were measured under a white noise
Dubai,13 as well as the Golden Gate Bridge in the excitation. Based on their results, the proposed metho-
United States, the Minami Bisan-Seto Bridge in Japan, dology could be generalized for other applications after
the Grande Belt Bridge in Denmark, and among others verification on large-scale structures. In their work,
mentioned in the work of Xu and Xia.14 Structural nat- Cury and Crémona23 performed a comparative study
ural aging and degradation, along with the lack of to evaluate feedforward ANNs, support vector
proper maintenance strategies, are major reasons for machines (SVMs), and Bayesian decision trees (BDTs).
the installation of monitoring programs. The authors used concepts of symbolic data analysis
Many studies have been carried out over the last (SDA) to condense acceleration data of a steel railway
decades aiming to develop robust and reliable methods bridge in France. The bridge was monitored to assess
focusing on the diagnosis of structural failure. Most the efficiency of a structural strengthening procedure.
damage detection methods are based on determining The authors concluded that the need of large training
modal parameters through a modal identification pro- data sets to train the supervised learning methods was a
cedure.15,16 However, the identification of modal prop- relevant limitation. Meruane and Mahu24 attempted to
erties is a sort of filtering process, which can lead to a apply ANN to detect, locate, and quantify structural
loss of information (when compared to the use of raw damage. Two experimental cases were used to validate
dynamic measurements), hiding small structural their approach: an 8-degree-of-freedom (8-DOF) mass–
changes that could be related to damage.17 Another spring system and a beam with multiple damage scenar-
major drawback is that methods based on modal para- ios. Although the results obtained were reasonable,
meters consider that the structure remains in the linear there was no verification on large-scale structures.
elastic domain after the occurrence of damage, which is In summary, even though some ANN-based
not always the case. approaches have proven their efficiency in particular
Thus, strategies that consider both acceleration mea- cases, all of them require the analyst/engineer/stake-
surements using statistical analysis and computational holder to have prior knowledge about the structure’s
intelligence to assess the structure’s dynamic behavior dynamic behaviors—or even its damage history—to
have been taken as a promising field of research in train the classifier. Hence, this aspect represents a
recent years.18,19 The direct use of acceleration mea- major shortcoming in practical applications since it is
surements may ease the problem of damage detection not always possible to recognize different structural
as the procedure becomes more straightforward since conditions beforehand.
the need for a modal identification process is ruled out. Conversely, methods based exclusively on unsuper-
Moreover, there is no need to know the excitation vised classification do not need any prior data labeling.
source to assess the structure’s behavior since it is pos- However, depending on the quality of the input data
sible to deal with the output measurements directly. (too noisy, hazy, missing, badly sampled, etc.), this
Many damage detection methods are based exclu- advantage could turn into a major disadvantage, thus
sively on supervised learning techniques, such as artifi- leading to inaccurate results. Silva et al.25 proposed
cial neural networks (ANNs), for example. Mehrjoo two fuzzy clustering approaches applied to data from a
et al.20 proposed an ANN-based technique applied to a benchmark structure in Los Alamos National
numerical truss bridge in which modal properties (i.e. Laboratory, USA. The results demonstrated that both
natural frequencies and mode shapes) and acceleration fuzzy clustering algorithms were effective. Nevertheless,
responses were used as damage sensitive features. the damage severity was not accurately assessed, and
Although the authors recommended such an approach both approaches generated a significant number of
for online and real-time damage detection applications, false-negative damage indications. Yu and Zhu26
its testing and validation on real structures were not reported a clustering technique applied to acceleration
presented. Lee and Kim21 also verified a similar measurements from a three-story structure tested in
approach both analytically and experimentally. The laboratory. Their approach was successful for non-
ANN’s performance with strain signals was found to linear damage detection. Yet, they suggested further
Nunes et al. 3
verifications with additional experimental studies, espe- Table 1. Statistical indicators extracted from a signal xi.
cially on large-scale structures. Studies conducted by
Alves et al.1,8 used a methodology based on k-means Peak Mean
xpeak = maxjxi j P
n
cluster analysis coupled with SDA. Several applications x = 1n xi
i=1
considering numerical simulations and experimental Mean square Root s
mean square
data were explored. Fair results were achieved when ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
P
n
Pn
modal parameters were used as damage sensitive fea- xq = 1n ðxi Þ2 1
rms = n ðxi Þ 2
i=1
i=1
tures. However, the use of raw dynamic measurements
Variance Standard deviation ffi
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(e.g. accelerations) led to a significant number of false- P
n
s2 = 1n ðxi xÞ2 Pn
negative alarms. Santos et al.27 implemented a similar s = 1n ðxi xÞ2
i=1
methodology. They applied a clustering technique to a i=1
8
>
> uðT6 , G1 Þ\uðT6 , G2 Þ uðT6 , G1 Þ\dt1 ) T6 2 C 1
<
uðT6 , G1 Þ\uðT6 , G2 Þ uðT6 , G1 Þ.dt1 ) T6 2 C 3
If
>
> uðT6 , G1 Þ.uðT6 , G2 Þ uðT6 , G2 Þ\dt2 ) T6 2 C 2
:
uðT6 , G1 Þ.uðT6 , G2 Þ uðT6 , G2 Þ.dt2 ) T6 2 C 3
Definition of the feature vectors and the ANN’s where Tn Iq is the qth statistical indicator of the nth
inputs dynamic test. Thus, the number of features nf is equal
to n.
In this article, two strategies are proposed to define the
For this case, the ANN input data matrix is assem-
feature vectors. The first one is based on a sensor-wise
bled as follows: the first row of the matrix represents
arrangement, as shown in equation (3)
the first test (T1 ) and the first 10 columns are the 10 sta-
tistical indicators (Iq , q = 1, . . . , 10) extracted from sen-
S1 I1 , S1 I2 , . . . , S1 I10 , S2 I1 , S2 I2 , . . . , S2 I10 , . . . , Sp I1 , . . . , Sp I10
sor 1 (S1 ) of the first test (T1 ). The next 10 columns are
ð3Þ the 10 indicators extracted from sensor 2 (S2 ) of the first
test (T1 ) and so on, until reaching the last 10 indicators
where Sp Iq is the qth statistical indicator of the pth sen-
extracted from sensor Sp of test (T1 ) (see equation (6))
sor. Thus, the number of features nf is equal to p q.
For this case, the ANN input data matrix is orga- 2 3
T1 I1S1 . . . T1 I2S1 . . . T1 I10S1 . . . T1 I1Sp . . . T1 I10Sp
nized as follows: the first row of the matrix represents 6 T2 I1S1 . . . T2 I2S1 . . . T2 I10S1 . . . T2 I1Sp . . . T1 I10Sp 7
6 7
sensor 1 (S1 ) and the first 10 columns are the 10 statisti- 6 .. .. .. .. .. .. .. .. .. 7
4 . . . . . . . . . 5
cal indicators (Iq , q = 1, . . . , 10) extracted from sensor
Tn I1S1 . . . Tn I2S1 . . . Tn I10S1 . . . Tn I1Sp . . . T1 I10Sp
S1 of the first test (T1 ). The next 10 columns are the 10
indicators extracted from the signal of this same sensor ð6Þ
(S1 ), but now from the second test (T2 ) and so on, until
reaching the last 10 indicators extracted from sensor S1
of test Tn (see equation (4)) Creating the hybrid classifier
2 3 The first phase of the methodology consists in the fol-
S1 I1T1 S1 I2T1 ... S1 I10T1 ... S1 I1Tn ... S1 I10Tn lowing steps (also depicted in Figure 4):
6 S2 I1T1 S2 I2T1 ... S2 I10T1 ... S2 I1Tn ... S2 I10Tn 7
6 7
6 .. .. .. .. .. .. .. .. 7
4 . . . . . . . . 5
1. Create the ANN and set its hyperparameters (e.g.
Sp I1T1 Sp I2T1 ... Sp I10T1 ... Sp I1Tn ... Sp I10Tn
number of neurons in the hidden layer, training
ð4Þ method, and error function).
Nunes et al. 7
Figure 4. Creation of the proposed damage detection model (first phase) considering two initial structural states.
2. Pass the input data matrices along with arbitrary expresses a dynamic behavior completely different
labels to the ANN. from the first one and does not influence the classifi-
3. Train, validate, and test the ANN. cation of the proposed strategy. For instance, an
4. Set the k-means clustering parameters (e.g. desired input data matrix formed by ‘‘ones’’ (or ‘‘twos,’’
number of clusters, distance metric, and initializa- ‘‘tens,’’ etc.) could represent such a fictitious input
tion algorithm). data set. Similarly, arbitrary labels are created. Then,
5. Cluster the ANN’s outputs, yielding s groups after a new actual behavior is identified, the fictitious
(C 1 , C 2 , . . . , C s ), where each one represents a dif- state is discarded and only the real structural condi-
ferent structural behavior. tions are used. Thus, the proposed methodology can
6. Calculate the threshold distances (dt1 , dt2 , . . . , dts ) be seamlessly applied to a continuous SHM program
related to each cluster as mentioned in section where just a single structural state needs to be known
‘‘Determination of threshold distances for clusters’’ a priori.
(steps 1–4). The second phase of the methodology consists in the
following steps (also shown in Figure 5):
After step 6, the damage detection model is created
and ready to classify new dynamic tests as either known 1. Input a new feature vector T (i.e. dynamic test) to
or new structural behaviors. Moreover, steps 3 and 5 the ANN trained during Phase 1.
are critical to this strategy. In fact, if the ANN yields 2. Evaluate the distance between the ANN’s output
bad classification results, this will directly influence the (O) and the centroids (G1 , G2 , . . . , Gs ) of each clus-
k-means algorithm outputs, thus consequently affecting ter, denoted by u(O, Gi ), with i = 1,..., s, according
the threshold distances, which are pivotal for the sec- to step 5 of section ‘‘Determination of threshold
ond phase of the proposed approach. distances for clusters.’’
Although the first three steps of Phase 1 are classi- 3. Verify whether the new data represent a known
fied as supervised since they use ANN, they do not structural behavior or not according to step 6 of
necessarily require prior knowledge of several different section ‘‘Determination of threshold distances for
structural states for their application. In fact, knowing clusters,’’ yielding two outcomes:
several—or even just two—structural conditions
(a) Known behavior(s): the new test is classified
beforehand are a rare situation when it comes to SHM
according to one of the previous known
applications. Thus, to circumvent this limitation, a sim-
behavior(s).
ple yet efficient approach can be used to initialize the
(b) Unknown behavior(s) (i.e. novelty): the new
proposed methodology. By ensuring that a set of
test is addressed to a novel cluster.
dynamic tests are related to a single structural
condition—which could be achieved by acquiring sev- 4. Repeat the entire procedure from Phase 1, but now
eral tests during a short controlled period of time—a considering s + 1 initial structural conditions, that
second data set could be artificially created so that it is, clusters.
8 Structural Health Monitoring 00(0)
Figure 5. Classification of new dynamic tests (second phase) considering two initial structural states.
In summary, the proposed strategy allows perform- genetic algorithms, grid search, Bayesian optimization,
ing a real-time dynamic monitoring program where regression analyses, and among many others. However,
new tests are continuously categorized by the hybrid the use of such techniques implicates on the definition
classifier into clusters representing known or unknown of new sets of parameters related to them. To avoid this
behavior. situation, one proposes a sensitivity analysis over the
most relevant parameters to select the best configura-
tions. Similar analyses were performed in Ranjbar and
Applications Saffar,45 Zhang and Wallace,46 Fock,47 and Cao et al.48
In this section, the performance of the hybrid classifier For the k-means algorithm, the distance metric and
is evaluated to detect changes in the dynamic responses initialization techniques were defined as ‘‘City Block’’
of three different structures. All algorithms and imple- and ‘‘plus,’’ respectively, as explained in section
mentations were developed using toolboxes and inter- ‘‘k-means clustering.’’
nal functions available in MATLAB. Initially, the
proposed methodology is applied to a numerical model
of a simply supported beam. Then, data collected from
Numerical analysis
a four-story frame tested in laboratory at the University A numerical example using a finite element model
of British Columbia (UBC) in Canada, and from a rail- (FEM) of a simply supported beam is initially chosen
way viaduct, located in France, are used to assess the to verify the efficiency of the proposed approach. The
hybrid classifier’s performance. FEM consists of 200 Euler–Bernoulli elements, where
In all applications, the ANN model implemented is each element has two nodes with 2 degrees of freedom
a feedforward MLP with a single hidden layer. The (vertical translation and rotation) per node. The
learning algorithm used is the Levenberg–Marquardt mechanical and geometrical properties are follows: the
and the mean square error is used as a cost function.44 beam is 6 m long; Young’s modulus (E) is equal to
The activation functions used in the hidden and output 210 GPa; the cross-sectional area and moment of iner-
layers were the hyperbolic tangent sigmoid and the soft- tia are 2:813103 m2 and 1:8453108 m4 , respectively;
max functions, respectively.41 The number of neurons and the mass density is 7850 kg m3 . A random force
used in the hidden layer of the network influences the excitation F(t) with different amplitudes and frequen-
ANN’s output, which consequently influences the final cies was applied at 0.69 m from the right support. To
response of the hybrid classifier. reproduce an actual instrumentation with acceler-
At this point, it is noteworthy that there are a con- ometers, vibration signals are simulated as vertical
siderable number of methods in the SHM literature to accelerations at 10 equidistant measurement points
optimally select the ANN’s hyperparameters, such as (‘‘sensors’’) collected during 100 s with a sampling
Nunes et al. 9
frequency of 100 Hz. The beam model is represented in columns are the statistical indicators extracted from
Figure 6. each sensor for all 10 tests performed (10 indicators
Three structural conditions are simulated: D0— 3 10 tests = 100 indicators per sensor).
healthy beam (undamaged); D1—damage level #1 As already mentioned, sensitivity analyses regarding
(characterized by a reduction of 20% in Young’s mod- the number of neurons in ANN’s hidden layer are per-
ulus at the midspan), illustrated by the light gray region formed. The number of neurons tested is 25, 35, and 50.
in Figure 6; and D2—damage level #2 (characterized To define this set of values, many others were tested.
by a reduction of 10% in Young’s modulus at the The tests started with the number of neurons equal to
quarter length), illustrated by the dark gray region in 10 and were varied on a step of 5 until reaching 100
Figure 6, plus the second structural configuration neurons. Of all values tested, this set presented the most
(damage level #1). Moreover, the simulated signals consistent results.
are contaminated with three levels of noise: no noise In the following simulations, the neural networks
(N0), 5% (N1), and 10% (N2), evaluated by means of used within the hybrid method are executed and
equation (7) trained, considering a 10-fold cross-validation method,
thus generating 10 hybrid classifiers. The k-fold cross-
xi, noise = xi + hnoise :sxi :V ;N ð0, 1Þ ð7Þ validation aims to ensure the ability to generalize com-
putational intelligence models when subjected to data
where xi, noise and xi are the vectors containing accelera-
not yet seen.49
tion measurements with and without noise measured
To simulate a real-case monitoring scenario, the ini-
by sensor i, respectively, hnoise is the level of noise, sxi is
tial input data considered a fictitious damage level
the standard deviation of xi , and V ;N (0, 1), in which
(matrix of ‘‘ones’’) and the undamaged case (D0). Thus,
V is a vector with random values extracted from a stan-
the ANN’s input matrix has dimensions [20 3 100],
dard normal distribution.
where rows from 1 to 10 represent the tests with a ficti-
In this study, 10 different dynamic tests are simu-
tious level of damage and rows from 11 to 20 corre-
lated for each level of damage and noise, yielding 90
spond to undamaged tests.
tests (10 vibration tests 3 three structural conditions 3
Since it is a supervised learning method, ANN also
three noise levels). For each test, 10,000 points are mea-
need labels (damage classes) corresponding to each
sured per sensor. For these measurements, 10 statistical
observation in the input matrix. Thus, an arbitrary tar-
indicators are calculated, as presented in Table 1.
get matrix [20 3 2] is generated, where the rows indi-
cate the class of the test using the following binary
Results. For all simulations regarding this application, coding: [1 0]—fictitious damage and [0 1]—undamaged.
the best results are obtained when data are organized as As Phase 1 begins, the ANNs are trained and unseen
‘‘sensors’’ (see equations (3) and (4)), where nine input subsets of both training (training and validation) and
[10 3 100] matrices are formed. Each matrix represents testing are inputted to the networks. Then, the k-means
a structural configuration, where rows correspond to method is fed with the ANN’s outputs, yielding two
the number of sensors used in the tests, and the 100 clusters related to each structural state, that is, one
10 Structural Health Monitoring 00(0)
Figure 8. (a) Structural system, (b) support system, and (c) positioning of masses (Dyke et al.50 and Dyke et al.51).
Since the structure’s establishment, several experi- ambient excitation of the tested scenarios (more or less
ments were carried out, including ambient vibration wind and traffic).
tests, impact tests, and shaker tests. Only ambient
vibration tests are used to assess the ability of the pro-
posed approach to continuous monitoring applications. Results. For each one of the five damage levels, a
Therefore, excitations were due to wind, pedestrians, dynamic test consisted of 60,000 acceleration samples
and traffic in the vicinities of the structure. Further measured by each one of the 15 sensors, yielding
details about the experiment setup and data can be [60,000 3 15] data matrices per test. These matrices are
found in Dyke et al.50 and Dyke et al.51 divided into six parts of dimensions [10,000 3 15],
Fifteen accelerometers (FBA and EPI sensors, 0– yielding six matrices per level of damage, as if each part
50 Hz frequency range, sensitivity of 5 V/g) were placed was a test performed separately. This division is neces-
throughout the frame, three at each floor including the sary to guarantee a larger number of training and test-
ground level. Such transducers were placed so as they ing samples for the hybrid classifier. Then, the 10
could measure motions in all directions and torsional statistical indicators are extracted.
modes. Anti-aliasing filter cut-off of 50 Hz was used, For all simulations concerning this application, the
and the data were sampled at 200 Hz. best results are obtained when data are organized by
The structure was subjected to different scenarios of ‘‘tests’’ (see equations (5) and (6)), where five input
damage (1–5), where the braces were removed and matrices of dimension [6 3 150] are assembled, with each
placed gradually as shown in Figure 9. In order of matrix representing a structural configuration. The rows
acquisition, levels 1, 5, 4, 3, and 2 simulate gradual correspond to the number of tests performed by each
damage to the support system. These levels are level of damage (six, in this case) and the columns are the
described in Table 2. statistical indicators extracted from the signals of the 15
Figure 10 depicts the acceleration histories measured sensors (10 indicators 3 15 sensors = 150 indicators).
by the 15 sensors. Magenta vertical lines indicate the Since this is an experimental application, it has
limits between cases, that is, simulated damage. It is uncertainties inherent to the data acquisition process.
worth mentioning that the wide variation in terms of Thus, in addition to sensitivity analyses regarding the
response amplitudes is mainly due to different levels of number of neurons, another sensitivity study is carried
12 Structural Health Monitoring 00(0)
Figure 9. Levels of damage imposed on the structure (Dyke et al.50 and Dyke et al.51).
Table 2. Cases of damage imposed to the structure out considering the probability rate of a new test to
(Dyke et al.50 and Dyke et al.51). belong to existing clusters (step 4—section
‘‘Determination of threshold distances for clusters’’).
Case Configuration To this end, three different rates of the CDF are evalu-
ated: 97.5%, 90%, and 85%. The number of neurons
1 Fully braced configuration.
2 All east side braces removed. tested in the hidden layer is 15, 25, and 55.
3 Removed braces on all floors in one bay on southeast Initial simulations are performed training the neural
corner. network using a fictitious damage state (matrix of
4 Removed braces on first and fourth floors in one bay ‘‘ones’’) and damage level #1 data. The ANN’s input
on southeast corner. matrix has dimensions [6 3 150], where rows from 1 to
5 Removed braces on first floor in one bay on southeast
corner. 3 represent the tests from the fictitious damage state,
while rows from 4 to 6 correspond to tests pertaining
Figure 13. Experimental setup of the bridge:52 (a) view of the bridge during the strengthening procedure, (b) strengthening system,
and (c) localization of the sensors.
limits (see step 6, section ‘‘Determination of thresh- vertical accelerometers and two horizontal acceler-
old distances for clusters’’). ometers (longitudinal and transversal) located under
By reducing the CDF rate to 90%, the number of the bridge deck acquired at a sampling rate of 4096 Hz,
correct classifications remains the same for known respecting the Nyquist frequency.
damage states but increases for new damage levels. Two measurement campaigns were carried out: the
With the rate reduced to 85%, the performance of first one on 24 June 2003, before reinforcement, when
the hybrid method in detecting known behaviors 15 dynamic tests were registered; and the second one,
begins to drop while correct classification percen- which took place on 26 June 2003, after the reinforce-
tages for new structural behaviors remain the same. ment was completed, when 13 tests were carried out.
Since the experimental tests were recorded during high-
Those three remarks highlight the importance of speed trains (TGV) crossings, ambient vibrations are
properly defining the CDF rate for each application. not controlled (output-only accelerations). In total, 280
signals were recorded (28 tests 3 10 accelerometers per
test). For each signal, 10 statistical indicators are
Experimental test II: railway bridge calculated.
The structure analyzed in this section is an embedded
steel bridge (Figure 13(a)) located at a high-speed track Results. For all simulations concerning this application,
in southeastern France, kilometer point 075 + 317, the best results are obtained when data are organized
between the cities of Sens and Soucy in the county of by ‘‘tests’’ (see equations (5) and (6)), where two input
Yonne. A dynamic monitoring campaign was per- matrices are assembled. For the structural state before
formed to analyze the influence of a strengthening pro- reinforcement, the matrix has dimensions [15 3 100].
cedure, as Figure 13(b) shows. For the structural state after reinforcement, the matrix’s
The strengthening procedure consisted in tightening dimensions are [13 3 100]. The rows correspond to
special bearings in an effort to shift the first natural fre- each test, whereas the columns are related to the num-
quency from the excitation frequency due to the trains’ ber of statistical indicators per sensor (10 indicators
crossings.52 The data comprised measurements of eight 3 10 sensors = 100 indicators).
Nunes et al. 15
17. Alves V, Cury A and Cremona C. On the use of symbolic 34. Avci O, Abdeljaber O, Kiranyaz S, et al. Structural health
vibration data for robust structural health monitoring. monitoring with self-organizing maps and artificial neural
Struct Build 2016; 169: 715–723. networks. Topic Modal Anal Test 2020; 8: 237–246.
18. Finotti R, Cury A and Barbosa F. An SHM approach 35. Prı́ncipe J, Euliano N and Lefebvre W. Neural and adap-
using machine learning and statistical indicators extracted tive systems: fundamentals through simulations. London:
from raw dynamic measurements. Latin Am J Solid John Wiley & Sons, 2000.
Struct 2019; 16(2): e165. 36. Jain A, Murty M and Flynn P. Data clustering: a review.
19. Santos A, Figueiredo E, Silva M, et al. Machine learning ACM Computing Surveys 1999; 31(3): 264–323.
algorithms for damage detection: kernel-based 37. Torres A, Alves V, Cury A, et al. Advanced statistical tech-
approaches. J Sound Vib 2016; 363: 584–599. niques applied to raw data for structural damage detection.
20. Mehrjoo M, Khaji N, Moharrami H, et al. Damage Experiment Vib Anal Civil Struct 2017; 5: 94–103.
detection of truss bridge joints using Artificial Neural 38. Bock H and Diday E. Analysis of symbolic data: explora-
Networks. Expert Syst Appl 2008; 35(3): 1122–1131. tory methods for extracting statistical information from
21. Lee J and Kim S. Structural damage detection in the fre- complex data. Berlin: Springer, 2000.
quency domain using neural networks. J Intell Mater Syst 39. Park S, Lee J, Yun C, et al. Electro-mechanical
Struct 2007; 18(8): 785–792. impedance-based wireless structural health monitoring
22. Hakim S, Razak H and Ravanfar S. Fault diagnosis on using PCA-data compression and k-means clustering
beam-like structures from modal parameters using artifi- algorithms. J Intel Mater Syst Struct 2008; 19(4):
cial neural networks. Measurement 2015; 76: 45–61. 509–520.
23. Cury A and Crémona C. Pattern recognition of struc- 40. Singh A, Yadav A and Rana A. K-means with three dif-
tural behaviors based on learning algorithms and sym- ferent distance metrics. Int J Comput Appl 2013; 67(10):
bolic data concepts. Struct Control Health Monit 2012; 13–17.
19(2): 161–186. 41. K-means. Mathworks, 2020, https://www.mathworks.
24. Meruane V and Mahu J. Real-time structural damage com/help/stats/kmeans.html?s_tid=doc_ta (accessed 2
assessment using artificial neural networks and antireso- November 2019).
nant frequencies. Shock Vib 2013; 2014: 14. 42. David A and Vassilvitskii S. K-means+ +: the advantages
25. Silva S, Dias M, Lopes V, et al. Structural damage detec- of careful seeding. In: SODA ‘07: proceedings of the eight-
tion by fuzzy clustering. Mech Syst Signal Pr 2008; 22(7): eenth annual ACM-SIAM symposium on discrete algo-
1636–1649. rithms, New Orleans, Louisiana, USA, 7–9 January 2007,
26. Yu L and Zhu J-H. Nonlinear damage detection using pp. 1027–1035.
higher statistical moments of structural responses. Struct 43. Cury A and Crémona C. Assignment of structural beha-
Eng Mech 2015; 54(2): 221–237. viours in long-term monitoring: application to a strength-
27. Santos J, Cremona C, Orcesi D, et al. Early damage ened railway bridge. Struct Health Monit 2012; 11(4):
detection based on pattern recognition and data fusion. J 422–441.
Struct 2017; 143(2): 0401612. 44. Hagan M and Menhaj B. Training feedforward networks
28. Li F, Meng G, Ye L, et al. Wavelet transform-based with the Marquardt algorithm. IEEE T Neural Netw
higher-order statistic for fault diagnosis in rolling element 1994; 5(6): 989–993.
bearings. J Vib Control 2008; 14(11): 1691–1709. 45. Ranjbar M and Saffar MG. A sensitivity analysis on
29. Farrar C and Worden K. Structural health monitoring: a application of artificial neural networks in structural
machine learning perspective. Hoboken, NJ: John Wiley & acoustics. J Robot Mechatron Syst 2016; 1(2): 23–26.
Sons, 2012. 46. Zhang Y and Wallace BC. A sensitivity analysis of (and
30. De La Rosa J, Aguera-Pérez A, Palomares-Salas J, et al. Practitioners’ Guide to) convolutional neural networks
Higher-order statistics: discussion and interpretation. for sentence classification, https://arxiv.org/abs/
Measurement 2013; 46(8): 2816–2827. 1510.03820v4
31. Behnia A, Chai H, Ghasemigol M, et al. Advanced dam- 47. Fock E. Global sensitivity analysis approach for input
age detection technique by integration of unsupervised selection and system identification purposes—a new
clustering into acoustic emission. Eng Fract Mech 2019; framework for feedforward neural networks. IEEE T
210: 212–227. Neural Netw Learn Syst 2014; 25(8): 1484–1495.
32. Haykin S. Neural Networks: a comprehensive foundation. 48. Cao MS, Pan LX, Gao YF, et al. Neural network
Upper Saddle River, NJ: Prentice Hall, 1999. ensemble-based parameter sensitivity analysis in civil
33. Azam S, Rageh A and Linzell D. Damage detection in engineering systems. Neural Comput Appl 2017; 28:
structural systems utilizing artificial neural networks and 1583–1590.
proper orthogonal decomposition. Struct Control Health 49. Kohavi R. A study of cross-validation and bootstrap for
Monit 2018; 26(2): e2288. accuracy estimation and model selection. In: Proceedings
18 Structural Health Monitoring 00(0)
of the 14th international joint conference on artificial intel- 51. Dyke J, Bernal D, Beck J, et al. Experimental phase II of
ligence, Montreal, QC, Canada, 20–25 August 1995, pp. the structural health monitoring benchmark problem. In:
1137–1145. New York: ACM. Proceedings of the 16th ASCE engineering mechanics con-
50. Dyke S, Agrawal A, Caicedo J, et al. NEES: database for ference, Seattle, WA, 16–18 July 2003.
structural control and monitoring benchmark problems, 52. Cury A, Cremona C and Diday E. Application of sym-
2015, https://datacenterhub.org/resources/257 bolic data analysis for structural modification assess-
ment. Eng Struct 2010; 32(3): 762–775.