Structures 56 (2023) 104980

Contents lists available at ScienceDirect

Structures
journal homepage: www.elsevier.com/locate/structures

Damage detection in power transmission towers using machine

learning algorithms
Mehdi Kouchaki a, Mojtaba Salkhordeh a, Mohammadreza Mashayekhi a, *, Masoud Mirtaheri a,
Hessam Amanollah b
a
Department of Civil and Environmental Engineering, K. N. Toosi University of Technology, Tehran, Iran
b
School of Civil Engineering, College of Engineering, University of Tehran, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: The purpose of this study is to utilize machine learning techniques to detect any damages that occurred in power
Power transmission tower transmission towers. In the first step, various machine learning algorithms were employed for a structural model
Damage detection with a small number of members so that the most efficient algorithm was selected based on the accuracy cri­
Machine learning
terion. Then, the selected method was evaluated using acceleration responses obtained from larger power
Structural health monitoring
transmission towers in various states. It is worth mentioning that the parameters of the learning algorithms have
Classification
been optimized by the Bayesian Optimization (BO) algorithm. Responses of three case studies, including 25-, 36-,
and 160-member structures, under environmental vibrations, were polluted to 10% noise to simulate the field
condition. Studies conducted on selecting the best classifier in detecting damages in these structures indicate that
the Support Vector Machine (SVM) algorithm, with an approximate average of 96%, has the highest accuracy
among different utilized algorithms. Moreover, the Linear Discriminant Analysis (LDA) algorithm, with an
approximate average of 94%, is the second most accurate algorithm, and its computational cost is lower than the
SVM algorithm. The results stemming from investigations of case studies show that the chosen features and
method seem appropriate for identifying power transmission towers’ damages. Finally, a study was conducted to
investigate the impact of the number of input records on accuracy.

1. Introduction safety were mainly through visual inspections. Whereas these methods
remain necessary, they have considerable limitations that make the
Throughout the world, numerous power transmission towers have diagnosis of different types and extent of damage difficult [4]. The
been erected. These towers making the connection of enormous regions procedure of detecting structural damage is called Structural Health
feasible and have profoundly aided the growth of countries and societies Monitoring (SHM) [2,5]. In SHM, the damage identification pursues
[1]. These structures are not only vulnerable to effects stemming from three principal aims: (1) assessing damage existence, (2) determining
wind loads, but also the global experiences have indicated their damage intensity, and (3) detecting the location of damage. Data-driven
vulnerability subjected to earthquakes and typhoon [2]. Given the and model-based methods are two general methods in SHM [6]. In the
enormous expense of repairing power transmission lines, the devasta­ model-based methods, a system identification procedure should be
tions caused by different occurrences seem irreparable [3]. Notwith­ carried out to detect the dynamic properties of the structure. Afterward,
standing many studies that have investigated the vibration by assessing the changes in these dynamic properties, structural health
specifications of power transmission towers under seismic ground mo­ status will be determined [7,8]. However, data-driven techniques
tion and wind load, the evaluation of the health situation of these employ statistical learning algorithms using vibration responses ob­
structures has not attracted much attention [1].It is crucial to constantly tained from the structure. These approaches apply learning algorithms
monitor the status of power transmission lines to preclude such net­ to build a classification or regression learner to predict structural dam­
works from sudden damages. age [9,10].
In the past, operational methods for the evaluation of structural In earthquake engineering, machine learning approaches have been

* Corresponding author.
E-mail address: [email protected] (M. Mashayekhi).

https://doi.org/10.1016/j.istruc.2023.104980
Received 22 April 2023; Received in revised form 26 July 2023; Accepted 26 July 2023
Available online 1 August 2023
2352-0124/© 2023 Institution of Structural Engineers. Published by Elsevier Ltd. All rights reserved.
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 1. (a) The shape of the 25-member structure, (b) connection simulation.

significantly developed, particularly in system identification, damage transmission tower structures using the processing and analysis of its
detection, and risk assessment [11]. Sajedi and Liang [6] proposed a vibration signals. Their results indicated that the proposed method could
damage detection method based on SVM to detect damages in buildings; efficiently identify the local damage in power transmission structures.
moreover, they validated the proposed method with a three-dimensional According to past studies, the rapid assessment of damage condition
reinforced concrete construction. Kourehli [12] suggested a least in power transmission towers requires more attention from researchers.
squares SVM-based damage prediction approach for SHM whenever In other words, limited studies assessed the sufficiency of damage in­
only limited modal data is accessible. The stiffness and mass matrices dicators and learning algorithms for detecting structural damage in this
were compacted using the iterated improved reduction system method. type of structure.
The findings showed that the proposed approach could determine the This paper proposes a framework for damage detection in power
severity of structural damage with promising accuracy. The hyper­ transmission towers under environmental vibrations using machine
parameters of SVMs in a damage detection issue were optimized using learning algorithms. First, three case studies, including 25-, 36-, and
an acquisition function-based BO method proposed by Agrawal and 160-member structures, and their modeling procedure are explained.
Chakraborty [13]. The proposed method was examined using the ASCE Afterward, the damage indicators derived from the vibration data are
benchmark structure [14]. Mariniello et al. [15] investigated the ability presented. Then, various algorithms used to achieve the best classifi­
of the decision tree algorithm in identifying the severity and location of cation learner are described. The next part presents more information on
damages in structures. Three different case studies, consisting of nu­ data generation, noise condition, damage patterns, data labeling, and
merical simulations and experimentally collected data, were used to training procedure. Thereafter, hyperparameters of the machine
verify the suggested damage detection approach. learning algorithms are optimized using the BO algorithm. Eventually,
There are only a small number of studies that have explored the the results are discussed, and the accuracy of the damage detection al­
implementation of damage detection approaches in power transmission gorithm is provided.
towers. According to Xu [16], power transmission towers are vulnerable
to abrupt damages and instabilities that can cause transmission towers 2. Case studies
to collapse entirely. Huang et al. [17] evaluated the influence of foun­
dation settlement on the changes in the natural frequencies of these As mentioned, this study was conducted with the aim of damage
types of structures. A vibration-based system was constructed for identification in power transmission towers using machine learning al­
transmission towers to monitor the settlement. By performing modal gorithms; consequently, three case studies, including 25-, 36-, and 160-
analysis on the least square complex frequency domain, it is possible to member structures, were investigated. These structures are explained in
determine the tower’s modal parameters. The outcomes demonstrate what follows.
that by identifying the modal parameters of the tower, the system is able
to precisely find the settlement information of the structure. Lam and 2.1. 25-member structure
Yang [18] presented an approach for damage detection in power
transmission towers utilizing the Bayesian probability method and The first benchmark structure used here to validate the proposed
updating finite element models. A two-step procedure was provided by method was a 25-member structure with a height of 508 cm(Fig. 1-a).
Qu et al. [19] to detect damage in the vertical components of power This structure has already been studied by Wu and Chow [22], Rajeev
transmission towers. In this approach, damage indicators were wave­ and Krishnamoorthy [23], Lee et al. [24], Li et al. [25], Sadollah et al.
let packet energy and the structure’s modal strain energy. Valeti and [26], and other researchers. The density and the modulus of elasticity of
Pakzad [20] presented an automatic detection method to identify the material are 2768 kg/m3 and 6890000 N/cm2, respectively. All the
corrosion in power transmission towers utilizing image processing, and elements were modeled by “elasticBeamColumn” element in the Open­
their approach was mainly successful in identifying corrosion. Karami- sees platform. The “zeroLength” element were used to simulate the
Mohammadi et al. [2] proposed a unique approach based on modal complexity of truss joints (Fig. 1-b). The rotational stiffness of these
shape curvature along with continuous wavelet transform to detect “zeroLength” elements were defined as 9.0 × 105 N-m/rad. The move­
possible damage locations in power transmission towers; moreover, they ment of all shared nodes shown in Fig. 1b were fixed together using
verified this numerical model with an actual tower modal test. The “equalDOF” command [27].
proposed method in their study was based on the processing of the
structural responses under environmental vibration. Xingjie et al. [21]
proposed a damage detection algorithm based on stochastic subspace for

2
M. Kouchaki et al. Structures 56 (2023) 104980

responses of the structure are influenced by such changes in dynamic

properties. As a result, the data directly obtained from structures must
be processed to extract features that are deeply connected with damages.
Seven potential features that were taken from the structure’s accelera­
tion response are explained in this paper.

3.1. Maximum peak to peak acceleration

The maximum peak-to-peak acceleration is the maximum accelera­

tion change in the dynamic response history, and some research has
shown that it has a strong correlation to the damage exposed to the
structures [33].

3.2. Mean period

The mean period is a term that was initially utilized to examine the
Fig. 2. The shape of the 36-member structure. ground motion records. It generates a weighted mean period between
0.25 Hz and 20 Hz, utilizing Fourier amplitude values within a corre­
sponding frequency range [33].
∑ 2
FA .(1/fi )
TM = i ∑ i 2 for 0.25 Hz ≤ fi ≤ 20 Hz (1)
i FAi

where FAi is the Fourier amplitude, and fi represents the frequency of

interest.

3.3. Mean frequency

The value of the mean period obtained from Equation 1 is used to

compute the mean frequency, which is defined as the inverse of it [33].

3.4. Cumulative absolute velocity

The value of this vibration parameter depends on the energy of the

output signal of the structure. Cumulative absolute velocity is the sum of
the absolute acceleration values throughout the duration of the vibra­
tion response. The area under the absolute values in the accelerogram is
represented by this parameter, and cumulative absolute velocity ex­
Fig. 3. The shape of the 160-member structure. hibits a strong correlation with the possibility of structural damage [34].
∫t
2.2. 36-member structure CAV = |ẍ(t)|dt (2)
0

As shown in Fig. 2, the second benchmark structure was a three-

where ẍ(t) is the acceleration response of the structure, and t is the total
dimensional truss with 36 members. This structure has already been
duration of the recorded response.
studied by Nobahari et al. [28] and other researchers. All members of
the structure are made of tubular steel with an outer diameter of 17.8 cm
and a wall thickness of 0.89 cm; furthermore, the height of the structure 3.5. Arias intensity
is 27.42 m. The material’s modulus of elasticity and density are
2100000 kg/cm2 and 7850 kg/m3, respectively. The Arias intensity is defined as a quantitative measure describing
the shaking intensity of earthquake records. Several studies showed that
this attribute is highly correlated to damage subjected to structures
2.3. 160-member structure
during ground motions. This feature, a vibration parameter, is
commonly used to explain an earthquake’s vibration energy during its
The third benchmark structure was a tower with 160 members
entire length [33,35].
(Fig. 3). This tower has already been studied by Groenwold and Stander
[29], Zhuang and Chang [30], Kaveh et al. [31], and other researchers. ∫∞
π
The modulus of elasticity and density of the used material are 2047000 IA = . ẍ(t)2 dt (3)
2g
kg/cm2 and 7850 kg/m3, respectively. 0
It should be noted that all introduced structures have been modeled
in the OpenSees software, and all the elements of the structures have where g represents the gravity acceleration.
been modeled using Elastic Uniaxial Material.
3.6. Destructive potential factor
3. Feature extraction
Destructive potential factor is a derivation of Arias intensity con­
The presence of damage in a structural component leads to a change taining the variable v0 to represent the frequency content in addition to
in the dynamic properties of the entire structure [32]. Therefore, the vibration energy [33].

3
M. Kouchaki et al. Structures 56 (2023) 104980

∫
π ∞
ẍ(t)2 dt Table 1
(4)
2g 0
PD = Damage patterns and and their corresponding labels in the 25-member trans­
v20
mission tower.

where v0 indicates the number of zero crossings per a predefined time Damage Explanation Label
pattern
threshold value.
I Undamaged structure 0
P1 Elements numbers 22 and 24 have been considered 1
3.7. Dominant frequency
damaged
P2 Elements numbers 20 and 21 have been considered 2
Another selected feature is the dominant frequency of dynamic damaged
response of the structure, which is obtained using the concept of the fast P3 Elements numbers 3, 4, and 5 have been considered 3
Fourier transform. Readers are encouraged to review Karami Moham­ damaged
P4 Elements numbers 1, 2, and 9 have been considered 4
madi et al. [2] to obtain more information in this regard. damaged
P5 Node number 6 has been considered damaged 5
4. Machine learning methods

Machine learning-based algorithms have been extensively utilized to non-linear problems. The SVM classifier’s primary objective is to iden­
identify vulnerabilities; in addition, numerous machine learning algo­ tify a decision boundary that is capable of classifying desperate classes.
rithms exist as prediction approaches. In the current research, five ma­ In other words, taking into account the data that has been collected, the
chine learning algorithms have been used to identify desperate states. SVM generates a hyperplane that divides two different classes of data
samples.
4.1. Decision tree
5. Methodology
The decision tree is defined as a hierarchical classifier [36]. This
The steps required for implementing the proposed damage detection
algorithm’s output is basic and simple to comprehend [37,38]. Using
training data, it generates a number of classification branches to a framework are mentioned in the following sections.
particular pattern [39]. The decision tree’s components include nodes
and arcs. The terminal and non-terminal nodes represent pattern classes 5.1. Data generation
and features, respectively; moreover, the arcs indicate feature values
[40,41]. Learning a decision tree entails learning the succession of if/ This study used 1000 environmental vibration records with a dura­
else questions which will lead to the correct answer as fast as possible tion of 10 s and a time step of 0.01 s to obtain a generalized data set. In
[42]. OpenSees software, the structures were modeled, and the mentioned
environmental vibrations were applied to them. Finally, the acceleration
responses of one node from each height level were recorded.
4.2. K-Nearest Neighbor (KNN)

The KNN algorithm is a learning approach that makes judgments 5.2. Considering additional noise in structural response
about data classes by comparing new examples of data to previously
known samples using a distance measure [43]. One of the simplest-to- In order to simulate the fieldconditions, a maximum of 10% noise
use methods in machine learning is the KNN algorithm, which belongs was added to the output signals. For this purpose, a fixed coefficient
within the category of supervised type learning [44]. The KNN method called signal-to-noise ratio (SNR) which is the ratio of signal variance to
predicts the target output of objects based on the outcomes of the closest noise variance is defined as follows.
samples or a group of the closest objects in a training set’s feature space σ 2signal
[45]. SNR = (5)
σ2noise

4.3. Naïve Bayes method (NBM) A Gaussian white noise with a mean value of 0 and a variance equal
σ2signal
to SNR
is added to the output signals of the structures.
The Bayes theorem is the basis of the NBM, and the assumption made
by this approach is that the features do not depend on one another [46]. 5.3. Feature extraction and data labeling
The possibility that a point corresponds to any of the listed classes is
assessed by naïve Bayes classifiers. In this method, a point is classified In the next step, the features are extracted from the acceleration
into the class with the greatest possibility [43]. responses. These features should be labeled in such a way that different
damage conditions can be determined. In this regard, different damages
4.4. Linear Discriminant analysis (LDA) have been applied to structures, and scenarios consisting of a set of
damaged states and an undamaged state have been considered so that
One of the most significant and frequently utilized supervised algo­ the results of this method can be examined for damage detection in
rithms for extracting features or reducing dimensions is LDA [47]. MATLAB software. In order to simulate damage, the stiffness of the
Despite being largely used to divide patterns into two groups, LDA is damaged member was reduced to 10% of the stiffness of that member in
able to be utilized to classify a variety of patterns [48]. The LDA method the undamaged state.
utilizes class labels to transfer the data into a new space with a lower In the 25-member structure, different states and their corresponding
dimension [49,50]. labels are shown in Table 1. Moreover, the locations of the damaged
elements are illustrated in Fig. 4. In order to validate the proposed
4.5. Support Vector Machines (SVM) method in identifying multiple damages in the structure, six damage
scenarios consisting of an undamaged state and damaged states, given in
SVM is considered a supervised learning model that has been Table 2, have been investigated for damage detection.
established in the methodology of statistical learning theory [51]. This In the 36-member structure, ten damaged states and related labels
approach possesses a variety of advantages in solving large-scale and are written in Table 3; furthermore, the locations of the damaged

4
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 4. Damage patterns in the 25-member structure: (a) P1, (b) P2, (c) P3, (d) P4, and (e) P5.

Table 2
Considered scenarios for the 25-member transmission structure.
Scenario number Damage patterns

1 I and P1
2 I, P1, and P2
3 I, P3, and P4
4 I, P1, P2, P3, and P4
5 I, and P5
6 I, P1, P2, P3, P4, and P5

Table 3
Damage patterns and and their corresponding labels in the 36-member trans­
mission tower.
Damage Explanation Label
pattern

I Undamaged structure 0
P1 Elements numbers 30, 32, 34, and 36 have been 1
considered damaged
P2 Elements numbers 18, 20, 22, and 24 have been 2
considered damaged
P3 Elements numbers 6, 8, 10, and 12 have been considered 3
damaged
P4 Elements numbers 6, 8, 10,12, 18, 20, 22, and 24 have 4
been considered damaged
P5 Elements numbers 29, 31, 33, and 35 have been 5
considered damaged
P6 Elements numbers 17, 19, 21, and 23 have been 6 Fig. 5. Damage patterns in the 36-member structure: (a) P1, (b) P2, (c) P3, (d)
considered damaged P4, (e) P5, (f) P6, (g) P7, (h) P8, (i) P9, and (j) P10.
P7 Elements numbers 5, 9, and 11 have been considered 7
damaged
P8 Elements numbers 25, 26, 27, and 28 have been 8
Table 4
considered damaged
P9 Node number 5 has been considered damaged 9 Considered scenarios for the 36-member transmission structure.
P10 Node number 10 has been considered damaged 10 Scenario number Damage patterns

1 I, P1, and P2
2 I, P3, and P4
elements are shown in Fig. 5.
3 I, P1, P2, P3, and P4
The damage cases include eight scenarios consisting of an undam­ 4 I, P5, P6, and P7
aged state along with different damaged conditions are given in Table 4. 5 I, P7, and P8
In the end, ten damage patterns have been considered for the 160- 6 I, P1, P2, P3, P4, P5, P6, P7, and P8
member structure, shown in Table 5. furthermore, the locations of the 7 I, P9, and P10
8 I, P1, P2, P3, P4, P5, P6, P7, P8, P9, and P10
damaged elements are shown in Fig. 6.

5
M. Kouchaki et al. Structures 56 (2023) 104980

Table 5 5.5. Training process

Damage patterns and and their corresponding labels in the 160-member trans­
mission tower. In this section, machine learning methods were used along with the
Damage Explanation Label BO algorithm utilized to tune the hyperparameters. In addition, the K-
pattern fold method was used for cross-validation, which is a method to evaluate
I Undamaged structure 0 the machine learning model and test its efficiency. It should be noted
P1 Elements numbers 1, 2, 3, and 4 have been considered 1 that the value of K is considered equal to 10 in the current research. The
damaged space of feature is fed to the learners in the following form:
P2 Elements numbers 13 and 15 have been considered 2
[ ]
damaged Oi = (F 11i , ⋯, F1in ),(F12i , ⋯, F2in ), ⋯, (F1mi , ⋯, Fmi
n
) (9)
P3 Elements numbers 49 and 50 have been considered 3
damaged
P4 Elements numbers 5, 6, 7,8, 9, and 10 have been 4
where F represents the feature described in section 3, n and i are the
considered damaged number of features and the observation number, respectively; moreover,
P5 Elements numbers 17, 18, 19, and 20 have been 5 m represents the node number from which the acceleration responses
considered damaged were recorded. Fig. 8 shows the framework of this method.
P6 Elements numbers 29, 30, 33, and 34 have been 6
considered damaged
P7 Elements numbers 81, 85, and 89 have been considered 7 6. Results
damaged
P8 Elements numbers 99, 100, and 104 have been considered 8 In this section, the results of different classifiers in identifying
damaged
damages of the 25-member structure are presented in section 6.1.
P9 Node number 10 has been considered damaged 9
P10 Node numbers 16 and 28 have been considered damaged 10 Following the selection of the best classifiers, the results of damage
identification for all structures are shown. In section 6.2, a study was
conducted to examine and compare the results using two algorithms
Eight different scenarios consisting of an undamaged state and with the highest accuracy among other algorithms. Results achieved
distinct damaged conditions, written in Table 6, have been investigated from evaluating the effect of the number of input environmental vi­
to validate the reliability of the proposed method. bration records on the prediction accuracy are presented in section 6.3.

5.4. Hyperparameter optimization

6.1. Performance of different classifiers in the 25-member structure
Classification accuracy is strongly influenced by the selection of
As mentioned, five different machine learning algorithms were used
hyper-parameters. The K-fold cross-validation has been utilized in this
for the 25-member structure. The accuracy values using different algo­
paper in order to obtain a generalized classifier. In this approach, the
rithms are given in Table 7.
entire data set is randomly divided into K parts. One of the parts is used
It is obvious in Table 7 that among the five algorithms, SVM and LDA
for validation, and another K-1 parts are used for training. This pro­
have the highest performances in terms of accuracy. Given the high
cedure is repeated K times, and all parts are used for training and vali­
accuracy of the mentioned algorithms, the SVM algorithm has been used
dation. The procedure is performed repeatedly for each subset, and the
to detect damages in the other structures. In addition, the results ob­
cross-validation loss is calculated here as the average of all losses.
tained from the LDA algorithm in those structures were compared to the
∑
K SVM results.
CVloss = lk (6) In order to express the result in a manner that can be easily com­
prehended, the confusion matrix and the Receiver Operating Charac­
k=1

where CVloss and lk represent the cross-validation loss and loss of each teristic (ROC) curve, as well as the validation accuracy, are shown for
fold, respectively. To construct an appropriate classifier with a mini­ various states. A confusion matrix is a special table layout where each
mum cross-validation loss, the BO method was used. The value of CVloss row of the matrix represents true classes while each column represents
was deemed as an objective function that will be calculated utilizing the predicted classes. A ROC curve is drawn on a unit square. The y and ×
Gaussian method, which is expressed in equation 7. axes represent the true positive rate (TPR) and the false positive rate
∑ (FPR), respectively. It should be noted that ROC curves pass the upper
CVloss GP(M, +σ 2 I) (7) right corner and the lower left corner of the plot [52]. Whatever the ROC
curve is closer to the upper-left corner of the chart, the model of pre­
where the covariance matrix and the mean value of the Gaussian method diction has more accuracy. More information about ROC curves is
are denoted by Σ and M, respectively. Additionally, it is presumed that available in the study by Fan et al. [53]. The area under this curve (AUC)
data with Gaussian noise are inserted with a variance of σ 2 , and the is defined as a measure of accuracy; moreover, when AUC is closer to 1,
identity matrix has been shown by I. In order to assess the subsequent it indicates a classification with a more accurate prediction. The results
sample point, BO maximizes an acquisition function described as: of investigations using the SVM algorithm are presented in sections 6.1.1
to 6.1.3.
EI(x) = E[max(0, μ(xbest ) − CVloss )] (8)

where xbest denotes the position of the minimum posterior mean, and 6.1.1. 25-member structure results using SVM
μ(xbest ) denotes the posterior mean’s minimum values. Figs. 9 to 14 show the confusion matrices corresponding to scenarios
The Bayesian condition is satisfied whenever it has attained the considered in the 25-member structure. The accuracy values of the
specified time limitation, the maximum number of iterations, and the prediction model in the first to sixth scenarios were 97.8%, 96.7%,
stopping condition which is given to the algorithm. The stopping crite­ 99.4%, 91.5%, 93.2%, and 90.3%, respectively. Moreover, these figures
rion in this study is reaching the maximum number of iterations, which illustrate the ROC curves for this learner. According to these figures, the
is selected following the amount of complexity of the selected classifier. algorithm is able to estimate the damage with promising accuracy.
The flowchart of selecting optimum hyper-parameters using the BO al­ Furthermore, the AUC values are larger than 0.97, which demonstrates
gorithm is shown in Fig. 7. the generalization and strength of the suggested classification process.
As shown in Fig. 9-b, 10-b, and 11-b, the ROC curves are close to the top-
left corner, which means that the proposed algorithm is flexible in

6
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 6. Damage patterns in the 160-member structure: (a) P1, (b) P2, (c) P3, (d) P4, (e) P5, (f) P6, (g) P7, (h) P8, (i) P9, and (j) P10.

7
M. Kouchaki et al. Structures 56 (2023) 104980

Table 6 6.1.2. The most complicated scenario considered for this structure was
Considered scenarios for a 160-member transmission structure. scenario #6
Scenario number Damage patterns Fig. 14 illustrates the confusion matrix and ROC curves for this case.
Although this scenario contains six damage conditions, the proposed
1 I, P1, and P2
2 I, P1, P2, and P3 algorithm is well-trained. The maximum false negative rate of this case
3 I, P4, P5, and P6 was 15.3%, which indicates the capability of the method in identifying
4 I, P1, P2, P3, P4, P5, and P6 the damage pattern where multiple damage cases are combined. 36-
5 I, P7, and P8 member structure results using SVM.
6 I, P1, P2, P3, P4, P5, P6, P7, and P8
7 I, P9, and P10
The ROC curve and the confusion matrices in all scenarios consid­
8 I, P1, P2, P3, P4, P5, P6, P7, P8, P9, and P10 ered for the 36-member structure are shown in Figs. 15 to 22. The ac­
curacy of the test in the trained models in the first to eighth scenarios
was 99.0%, 96.9%, 93.6%, 97.2%, 97.9%,93.3%, 93.3%, and 93.1%,
predicting unseen scenarios in these conditions. respectively. These figures show that the prediction model is trusty in
It is obvious that the more damage scenarios become complicated, estimating the damage. Furthermore, the AUC values for all patterns are
the more reduction in the prediction accuracy is achieved. The close to 1.0, which indicates the high performance of these prediction
maximum false negative rate in scenario #4 was 16.8%, which means models.
that the minimum true positive rate is 83.2% in this condition (Fig. 12- As shown in Fig. 15, the algorithm is well-trained for scenario #1.
a). As shown in Fig. 12-b, all the ROC curves are close to the top-left The minimum true positive rate for this scenario was 98.6%, and all the
corner of the diagram, and minimum AUC value is 0.97, which proves ROC curves are stuck to the top-left corner of the diagram. In scenario
the robustness of the proposed method. #2, the maximum false negative rate was 4.5%, which means that the
Figs. 13 and 14 show the damage cases that consist of joint damage performance of the prediction model is assuring. In addition, the devi­
scenarios. As depicted in Fig. 13, the proposed algorithm is capable of ation of ROC curves from the top-left corner of the diagram is negligible.
identifying damage subjected to the connections of the structure with The prediction model trained for scenario #3 presented lower accuracy
promising accuracy. Fig. 13-b indicates that the ROC curves are close to than scenarios #1 and #2. As can be seen, the maximum false negative
the top-left corner of the diagram, and the AUC values are 0.98, which rate was 15.2% for this case. This means that the algorithm results in
proves the efficiency of the algorithm in predicting these damage 84.8% true predictions for this pattern, which is acceptable. The mini­
scenarios. mum AUC value was 0.98 in this scenario, which demonstrates the
flexibility of the algorithm for new unseen data.
In scenario #4, the maximum false negative rate was 6.5%, and all

Fig. 7. Flowchart of the selection of optimized hyper-parameters using the BO algorithm.

8
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 8. Outline of the proposed framework.

Table 7
The accuracy of different algorithms in the 25-member structure.
Algorithm Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Average

Decision tree 95% 92.5% 97.4% 84.4% 89.8% 83.2% 90.4%

LDA 97.5% 96.6% 98.6% 90.3% 92.8% 88.7% 94.1%
Naïve bayes 95.7% 91% 98.4% 85.2% 89.3% 83.6% 90.5%
SVM 97.8% 96.7% 99.4% 91.5% 93.2% 90.3% 94.8%
KNN 97.4% 94.5% 99.2% 85.4% 92% 85.4% 92.3%

Fig. 9. Result of damage detection in the 25-member structure for scenario number 1: (a) Confusion matrix; (b) ROC curve.

9
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 10. Result of damage detection in the 25-member structure for scenario number 2: (a) Confusion matrix; (b) ROC curve.

Fig. 11. Result of damage detection in the 25-member structure for scenario number 3: (a) Confusion matrix; (b) ROC curve.

Fig. 12. Result of damage detection in the 25-member structure for scenario number 4: (a) Confusion matrix; (b) ROC curve.

10
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 13. Result of damage detection in the 25-member structure for scenario number 5: (a) Confusion matrix; (b) ROC curve.

Fig. 14. Result of damage detection in the 25-member structure for scenario number 6: (a) Confusion matrix; (b) ROC curve.

the ROC curves are near the top-left corner of the diagram. This means 6.1.3. 160-member structure results using SVM
that the algorithm predicts the worst pattern with 93.5% accuracy. The Figs. 23 to 30 illustrate the ROC curves and confusion matrices in all
prediction model performed better in scenario #5 compared to scenario scenarios considered for the 160-member structure. The values of cross-
#4. As shown, the minimum true positive rate was 96.4% in this sce­ validation accuracy of the learner in the first to eighth scenarios were
nario, and all the ROC curves are stuck to the top-left corner of the 97.2%, 97.2%, 99.5%, 97.4%, 91%, 94.5%, 96.5% and 93.4%, respec­
diagram. tively. Furthermore, according to these ROC curves, the AUC values for
As a complicated scenario, the maximum false negative rate of the approximately all patterns are equal to one. This means that the clas­
model was 18.1% for scenario #6 (Fig. 20). The ROC curves of different sifier acted in damage detection accurately. As depicted in Figs. 23 to 25,
damage patterns indicate that the algorithm performed promising in this the prediction model was well-trained for damage scenarios #1, #2, and
scenario. The true positive rate of all patterns was more than 93.0% #3. All the ROC curves are far from line TPR = FPR, which means that
except for damage patterns #2 and #4. the algorithm is flexible for unseen damage patterns.
Scenario #7 is related to the damage in the joints of the structure. In damage pattern #4, despite the complexity of the scenario, the
According to Fig. 21, the minimum true positive rate was 89.7%, and all maximum false negative rate was 5.7%, which proves the efficiency of
the ROC curves presented an AUC value larger than 0.98. This means the proposed algorithm in this case. In addition, all the AUC values were
that the prediction model performed promising in identifying the very close to 1.0. This means that the reliability of the prediction model
damage patterns in this scenario. The most complicated damage sce­ for new data is promising, and the algorithm is flexible to be extended by
nario for 36-member structure was damage scenario #8 (Fig. 22). As new cases.
shown, the prediction model performed promising for this case. All the In damage scenarios #5 and #6, the algorithm provides lower ac­
ROC curves are close to the top-left corner of the diagram, and the curacies compared to scenario #4. The maximum true positive rate was
minimum true positive rate was 84.3% in this condition. 11.4% and 16.5% for damage scenarios #5 and #6, respectively. The
ROC curves did not deviate from the top-left corner of the diagram. As

11
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 15. Result of damage detection in the 36-member structure for scenario number 1: (a) Confusion matrix; (b) ROC curve.

Fig. 16. Result of damage detection in the 36-member structure for scenario number 2: (a) Confusion matrix; (b) ROC curve.

Fig. 17. Result of damage detection in the 36-member structure for scenario number 3: (a) Confusion matrix; (b) ROC curve.

12
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 18. Result of damage detection in the 36-member structure for scenario number 4: (a) Confusion matrix; (b) ROC curve.

Fig. 19. Result of damage detection in the 36-member structure for scenario number 5: (a) Confusion matrix; (b) ROC curve.

Fig. 20. Result of damage detection in the 36-member structure for scenario number 6: (a) Confusion matrix; (b) ROC curve.

shown, the minimum AUC values were 0.97 and 0.99 for scenarios #5 6.2. Results of LDA in comparison to the results using SVM
and #6, respectively.
In damage scenarios #7 and #8, joint damage patterns are also In this part, damage detection in all scenarios is carried out using the
considered. It is obvious that the prediction algorithm performed LDA algorithm, and its results are compared with the results of the SVM
promising in these scenarios because the maximum false negative rate algorithm. The prediction accuracy values of results using SVM and LDA
was 5.4% and 18.2% in damage scenarios #7 and #8, respectively algorithms are given in Figs. 31 to 33.
(Figs. 29 and 30). As can be observed, the majority of damage patterns Fig. 31 demonstrates that the averages of accuracy values utilizing
were predicted with a true positive rate higher than 90.0% in these SVM and LDA algorithms are 94.8% and 94.1%, respectively.
scenarios. In addition, all the ROC curves were far from line TPR = FNR, It is evident in Fig. 32 that the values of average accuracy using SVM
which proves the stability of the models in predicting damage patterns of and LDA algorithms are 95.5% and 93.1%, respectively.
these scenarios. According to Fig. 33, the averages of accuracy values using SVM and
LDA algorithms are 95.8% and 95.1%, respectively.

13
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 21. Result of damage detection in the 36-member structure for scenario number 7: (a) Confusion matrix; (b) ROC curve.

Fig. 22. Result of damage detection in the 36-member structure for scenario number 8: (a) Confusion matrix; (b) ROC curve.

Fig. 23. Result of damage detection in the 160-member structure for scenario number 1: (a) Confusion matrix; (b) ROC curve.

14
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 24. Result of damage detection in the 160-member structure for scenario number 2: (a) Confusion matrix; (b) ROC curve.

Fig. 25. Result of damage detection in the 160-member structure for scenario number 3: (a) Confusion matrix; (b) ROC curve.

Fig. 26. Result of damage detection in the 160-member structure for scenario number 4: (a) Confusion matrix; (b) ROC curve.

15
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 27. Result of damage detection in the 160-member structure for scenario number 5: (a) Confusion matrix; (b) ROC curve.

Fig. 28. Result of damage detection in the 160-member structure for scenario number 6: (a) Confusion matrix; (b) ROC curve.

Fig. 29. Result of damage detection in the 160-member structure for scenario number 7: (a) Confusion matrix; (b) ROC curve.

16
M. Kouchaki et al. Structures 56 (2023) 104980

Fig. 30. Result of damage detection in the 160-member structure for scenario number 8: (a) Confusion matrix; (b) ROC curve.

Fig. 33. Comparison of box plot of validation accuracy using LDA and SVM
Fig. 31. Comparison of box plot of validation accuracy using LDA and SVM algorithms in 160-member structure.
algorithms in the 25-member structure.

Fig. 34. The impact of the number of input records on prediction accuracy in
the 25-member structure.
Fig. 32. Comparison of box plot of validation accuracy using LDA and SVM
algorithms in the 36-member structure.

17
M. Kouchaki et al. Structures 56 (2023) 104980

7. Conclusions

In this research, the detection of damages in power transmission

towers was carried out using different machine learning algorithms. For
this purpose, three structures, including 25-, 36-, and 160-member
structures, in various states were subjected to environmental vibra­
tion, and the acceleration responses of a specific node in each height
level were recorded. Furthermore, a maximum of 10% noise was added
to the acceleration data, and a set of features was extracted. Various
scenarios including different states such as an undamaged state and
damaged conditions were considered. Initially, different machine
learning algorithms, including decision tree, KNN, naïve base, LDA, and
SVM, were first trained for the 25-member structure. It should be noted
that hyperparameters were tuned utilizing the BO algorithm. The algo­
rithms with the best performance were selected and trained for the
Fig. 35. The impact of the number of input records on prediction accuracy in larger structures. Moreover, the influence of the number of input records
the 36-member structure. on the accuracy was assessed. The significant results obtained from this
study are written in what follows.
The results stemming from the parametric study demonstrated that
the SVM algorithm is the most trustworthy classifierfor detecting
structural damage in the investigated power transmission towers. In the
25-, 36-, and 160-member structures, investigating damage scenarios
using the SVM algorithm demonstrated that the averages of accuracy
values were 94.8%, 95.5%, and 95.8%, respectively. Further assessment
revealed that the averages of accuracy values utilizing the LDA classifier
in the 25-, 36-, and 160-member structures, were 94.1%, 93.1%, and
95.1%, respectively, which are close to the results of the SVM algorithm;
moreover, according to the investigations, the LDA algorithm has a
lower computational cost in comparison to SVM algorithm. Moreover,
The features extracted in this study, including maximum peak-to-peak
acceleration, mean period, mean frequency, cumulative absolute ve­
locity, Arias intensity, destructive potential factor, and dominant fre­
quency, have been effective in identifying damages in assessed power
Fig. 36. The impact of the number of input records on prediction accuracy in transmission towers. The reduction in the number of input environ­
the 160-member structure. mental vibration records had an adverse effect on the average accuracy.
Moreover, the analysis of all considered structures revealed that the
According to Figs. 31 to 33, it can be seen that the average accuracy prediction accuracy is greater than 90% when the dataset contains more
using the SVM algorithm is higher than LDA; however, the difference in than 750 records. Bayesian optimization algorithm hindered the prob­
the accuracy using these two algorithms in various scenarios is not able errors pertaining to the inappropriate architecture of the classifi­
significant. Moreover, the computational cost of the SVM optimization cation learner used for considered structures.
process is higher than that of LDA.

6.3. The impact of the number of input records on prediction accuracy Declaration of Competing Interest

This section has studied the effect of changing the number of input The authors declare that they have no known competing financial
vibration records using the LDA algorithm on damage detection in the interests or personal relationships that could have appeared to influence
considered structures. The box plots of this subject are illustrated in the work reported in this paper.
Figs. 34 to 36.
As can be seen from Fig. 34, the accuracy has decreased as the References
number of records has reduced. When 500, 750, and 1000 records were
input, the averages of accuracy values in this structure were 87.9%, [1] Yin T, Lam HF, Chow HM, Zhu HP. Dynamic reduction-based structural damage
detection of transmission tower utilizing ambient vibration data. Eng Struct 2009;
91.8%, and 94.1%, respectively. 31:2009–19. https://doi.org/10.1016/j.engstruct.2009.03.004.
According to Fig. 35, the average accuracy in this structure fluctuates [2] Karami-Mohammadi R, Mirtaheri M, Salkhordeh M, Hariri-Ardebili MA. Vibration
between 83.4% and 93.2%. Anatomy and Damage Detection in Power Transmission Towers with Limited
Sensors. Sensors 2020;20:1731. https://doi.org/10.3390/s20061731.
According to Fig. 36, the average accuracy in the structure with 160
[3] Albermani F, Kitipornchai S, Chan RWK. Failure analysis of transmission towers.
members when 500, 750, and 1000 records were used gave rise to ac­ Eng Fail Anal 2009;16:1922–8. https://doi.org/10.1016/j.
curacy values of 88.5%, 92%, and 95.1%, respectively. engfailanal.2008.10.001.
[4] Azimi M, Pekcan G. Structural health monitoring using extremely compressed data
In general, the results demonstrate that the prediction accuracy ­
through deep learning. Comput-Aided Civ Infrastruct Eng 2020;35:597–614.
grows as more records are fed into the algorithm. It is observed from the https://doi.org/10.1111/mice.12517.
investigation of three structures that the utilization of more than 750 [5] Arici Y, Mosalam KM. Modal identification of bridge systems using state-space
records as input leads to obtaining an accuracy of over 90% in the re­ methods. Struct Control Health Monit 2005;12:381–404. https://doi.org/10.1002/
stc.76.
sults. Additionally, it should be noticed that as the quantity of input [6] Sajedi SO, Liang X. A data-driven framework for near real-time and robust damage
records rises, the computational time required for the training process diagnosis of building structures. Struct Control Health Monit 2020;27. https://doi.
increases. org/10.1002/stc.2488.
[7] Basseville M, Mevel L, Goursat M. Statistical model-based damage detection and
localization: subspace-based residuals and damage-to-noise sensitivity ratios.
J Sound Vib 2004;275:769–94. https://doi.org/10.1016/j.jsv.2003.07.016.

18
M. Kouchaki et al. Structures 56 (2023) 104980

[8] Mirtaheri M, Salkhordeh M, Mohammadgholiha M, Lin T-K. A System [32] Curadelli RO, Riera JD, Ambrosini D, Amani MG. Damage detection by means of
Identification-Based Damage-Detection Method for Gravity Dams. Shock Vib 2021; structural damping identification. Eng Struct 2008;30:3497–504. https://doi.org/
2021:1–15. 10.1016/j.engstruct.2008.05.024.
[9] Farrar CR, Worden K. Structural health monitoring: a machine learning [33] Moughty JJ, Casas JR. Performance assessment of vibration parameters as damage
perspective. John Wiley & Sons; 2012. indicators for bridge structures under ambient excitation. Procedia Eng 2017;199:
[10] Ying Y, Garrett JH, Oppenheim IJ, Soibelman L, Harley JB, Shi J, et al. Toward 1970–5. https://doi.org/10.1016/j.proeng.2017.09.306.
data-driven structural health monitoring: application of machine learning and [34] Salkhordeh M, Mirtaheri M, Soroushian S. A decision-tree-based algorithm for
signal processing to damage detection. J Comput Civ Eng 2013;27(6):667–80. identifying the extent of structural damage in braced-frame buildings. Struct
[11] Mahmoudi H, Bitaraf M, Salkhordeh M, Soroushian S. A rapid machine learning- Control Health Monit 2021:28. https://doi.org/10.1002/stc.2825.
based damage detection algorithm for identifying the extent of damage in concrete [35] Kordestani H, Xiang Y, Ye X, Yun C, Shadabfar M. Localization of damaged cable in
shear-wall buildings. Structures 2023;47:482–99. https://doi.org/10.1016/j. a tied-arch bridge using Arias intensity of seismic acceleration response. Struct
istruc.2022.11.041. Control Health Monit 2020:27. https://doi.org/10.1002/stc.2491.
[12] Kourehli SS. LS-SVM regression for structural damage diagnosis using the iterated [36] Zhang Y, Tang Z, Yang R. Data anomaly detection for structural health monitoring
improved reduction system. Int J Struct Stab Dyn 2016;16:1550018. https://doi. by multi-view representation based on local binary patterns. Measurement 2022;
org/10.1142/S0219455415500182. 202:111804. https://doi.org/10.1016/j.measurement.2022.111804.
[13] Agrawal AK, Chakraborty G. On the use of acquisition function-based Bayesian [37] Liu R, Li S, Zhang G, Jin W. Depth detection of void defect in sandwich-structured
optimization method to efficiently tune SVM hyperparameters for structural immersed tunnel using elastic wave and decision tree. Constr Build Mater 2021;
damage detection. Struct Control Health Monit 2021;28. https://doi.org/10.1002/ 305:124756. https://doi.org/10.1016/j.conbuildmat.2021.124756.
stc.2693. [38] Gordan M, Sabbagh-Yazdi S-R, Ismail Z, Ghaedi K, Carroll P, McCrum D, et al.
[14] Johnson EA, Lam HF, Katafygiotis LS, Beck JL. Phase I IASC-ASCE structural health State-of-the-art review on advancements of data mining in structural health
monitoring benchmark problem using simulated data. J Eng Mech 2004;130:3–15. monitoring. Measurement 2022;193:110939. https://doi.org/10.1016/j.
https://doi.org/10.1061/(ASCE)0733-9399(2004)130:1(3). measurement.2022.110939.
[15] Mariniello G, Pastore T, Menna C, Festa P, Asprone D. Structural damage detection [39] Athanasiou A, Ebrahimkhanlou A, Zaborac J, Hrynyk T, Salamone S. A machine
and localization using decision tree ensemble and vibration data. Comput-Aided learning approach based on multifractal features for crack assessment of reinforced
Civ Infrastruct Eng 2021;36:1129–49. https://doi.org/10.1111/mice.12633. concrete shells. Comput-Aided Civ Infrastruct Eng 2020;35:565–78. https://doi.
[16] Xu Y, Lin J, Zhan S, Wang F. Multistage damage detection of a transmission tower: org/10.1111/mice.12509.
Numerical investigation and experimental validation. Struct Control Health Monit [40] Chen D, Montano V, Huo L, Song G. Depth detection of subsurface voids in
2019:e2366. https://doi.org/10.1002/stc.2366. concrete-filled steel tubular (CFST) structure using percussion and decision tree.
[17] Huang X, Zhao Y, Zhao L. A Vibration-based Monitoring System for Transmission Measurement 2020;163:107869. https://doi.org/10.1016/j.
Tower Settlement. 2018 Cond. Monit. Diagn. CMD, Perth, WA: IEEE; 2018. p. 1–4. measurement.2020.107869.
[18] Lam H-F, Yang J. Bayesian structural damage detection of steel towers using [41] Amarnath M, Sugumaran V, Kumar H. Exploiting sound signals for fault diagnosis
measured modal parameters. Earthq Struct 2015;8:935–56. https://doi.org/ of bearings using decision tree. Measurement 2013;46:1250–6. https://doi.org/
10.12989/EAS.2015.8.4.935. 10.1016/j.measurement.2012.11.011.
[19] Qu W, Song W, Xia Y, Xu Y, Qin W, Jiang Z. Two-Step Method for Instability [42] da Silva Coelho J, de Sousa AASR, Machado MR, Dutkiewicz M. Study of Machine
Damage Detection in Tower Body of Transmission Structures. Adv Struct Eng 2013; Learning Techniques for Damage Identification in a Beam. In: Dimitrovová Z,
16:219–32. https://doi.org/10.1260/1369-4332.16.1.219. Biswas P, Gonçalves R, Silva T, editors. Recent Trends Wave Mech. Vib., vol. 125,
[20] Valeti B, Pakzad S. Automated Detection of Corrosion Damage in Power Cham: Springer International Publishing; 2023, p. 817–26. https://doi.org/
Transmission Lattice Towers Using Image Processing. Struct. Congr. 2017, Denver, 10.1007/978-3-031-15758-5_84.
Colorado: American Society of Civil Engineers; 2017, p. 474–82. https://doi.org/ [43] Miller BA, Anton SR. Investigation into machine learning with impedance SHM for
10.1061/9780784480427.040. damage detection and classification within simulated total knee replacements.
[21] Xingjie W, Tingting Y, Meigen C, Chong Z, Wensong Z, Yutong P. Damage J Intell Mater Syst Struct 2022;33:2405–21. https://doi.org/10.1177/
Detection of Transmission Tower Based on Stochastic Subspace and Statistic 1045389X221086668.
Model. IOP Conf Ser Earth Environ Sci 2021;687(1):012086. https://doi.org/ [44] Bansal M, Goyal A, Choudhary A. A comparative analysis of K-Nearest Neighbor,
10.1088/1755-1315/687/1/012086. Genetic, Support Vector Machine, Decision Tree, and Long Short Term Memory
[22] Wu S-J, Chow P-T. Steady-state genetic algorithms for discrete optimization of algorithms in machine learning. Decis Anal J 2022;3:100071. https://doi.org/
trusses. Comput Struct 1995;56:979–91. https://doi.org/10.1016/0045-7949(94) 10.1016/j.dajour.2022.100071.
00551-D. [45] Triguero I, García-Gil D, Maillo J, Luengo J, García S, Herrera F. Transforming big
[23] Rajeev S, Krishnamoorthy CS. Genetic Algorithms-Based Methodologies for Design data into smart data: An insight on the use of the k-nearest neighbors algorithm to
Optimization of Trusses. J Struct Eng 1997;123:350–8. https://doi.org/10.1061/ obtain quality data. WIREs Data Min Knowl Discov 2019:9. https://doi.org/
(ASCE)0733-9445(1997)123:3(350). 10.1002/widm.1289.
[24] Lee KS, Geem ZW, Lee S, Bae K. The harmony search heuristic algorithm for [46] Soltani E, Ahmadi E, Gueniat F, Salami MR. A review of bridge health monitoring
discrete structural optimization. Eng Optim 2005;37:663–84. https://doi.org/ based on machine learning. Proc Inst Civ Eng - Bridge Eng 2022:1–11. https://doi.
10.1080/03052150500211895. org/10.1680/jbren.22.00030.
[25] Li LJ, Huang ZB, Liu F, Wu QH. A heuristic particle swarm optimizer for [47] Ma Y, He Y, Wang L, Zhang J. Probabilistic reconstruction for spatiotemporal
optimization of pin connected structures. Comput Struct 2007;85:340–9. https:// sensor data integrated with Gaussian process regression. Probabilistic Eng Mech
doi.org/10.1016/j.compstruc.2006.11.020. 2022;69:103264. https://doi.org/10.1016/j.probengmech.2022.103264.
[26] Sadollah A, Bahreininejad A, Eskandar H, Hamdi M. Mine blast algorithm for [48] Nguyen A, Long Nguyen C, Gharehbaghi V, Perera R, Brown J, Yu Y, et al.
optimization of truss structures with discrete variables. Comput Struct 2012; A computationally efficient crack detection approach based on deep learning
102–103:49–63. https://doi.org/10.1016/j.compstruc.2012.03.013. assisted by stockwell transform and linear discriminant analysis. Structures 2022;
[27] Zheng W, Yu Y. Bayesian Probabilistic Framework for Damage Identification of 45:1962–70.
Steel Truss Bridges under Joint Uncertainties. Adv Civ Eng 2013;2013:1–13. [49] Al-Ghalib AA. Structural damage detection of old ADA steel truss bridge using
https://doi.org/10.1155/2013/307171. vibration data. Struct Control Health Monit 2022:29. https://doi.org/10.1002/
[28] Nobahari M, Ghasemi MR, Shabakhty N. A fast and robust method for damage stc.3098.
detection of truss structures. Appl Math Model 2019;68:368–82. https://doi.org/ [50] Chen D-W, Miao R, Yang W-Q, Liang Y, Chen H-H, Huang L, et al. A Feature
10.1016/j.apm.2018.11.025. Extraction Method Based on Differential Entropy and Linear Discriminant Analysis
[29] Groenwold AA, Stander N. Optimal discrete sizing of truss structures subject to for Emotion Recognition. Sensors 2019;19(7):1631. https://doi.org/10.3390/
buckling constraints. Struct Optim 1997;14:71–80. https://doi.org/10.1007/ s19071631.
BF01812508. [51] Vapnik VN. An overview of statistical learning theory. IEEE Trans Neural Netw
[30] Juang DS, Chang WT. A revised discrete Lagrangian-based search algorithm for the 1999;10:988–99. https://doi.org/10.1109/72.788640.
optimal design of skeletal structures using available sections. Struct Multidiscip [52] Grzybowski M, Younger JG. Statistical Methodology: III. Receiver Operating
Optim 2006;31:201–10. https://doi.org/10.1007/s00158-005-0571-8. Characteristic (ROC) Curves. Acad Emerg Med 1997;4:818–26. https://doi.org/
[31] Kaveh A, Kalatjari VR, Talebpour MH. Optimal Design of Steel Towers Using a 10.1111/j.1553-2712.1997.tb03793.x.
Multi-Metaheuristic Based Search Method. Period Polytech Civ Eng 2016;60(2): [53] Fan J, Upadhye S, Worster A. Understanding receiver operating characteristic
229–46. (ROC) curves. CJEM 2006;8:19–20. https://doi.org/10.1017/
S1481803500013336.

19