Structures: Hossein Mahmoudi, Maryam Bitaraf, Mojtaba Salkhordeh, Siavash Soroushian

Structures 47 (2023) 482–499
Contents lists available at ScienceDirect
Structures
journal homepage: www.elsevier.com/locate/structures
A rapid machine learning-based damage detection algorithm for identifying

the extent of damage in concrete shear-wall buildings
Hossein Mahmoudi a , Maryam Bitaraf a , Mojtaba Salkhordeh b , Siavash Soroushian b ,∗
a
Department of Civil and Environmental Engineering, University of Tehran, Tehran, Iran
b
Department of Civil and Environmental Engineering, K. N. Toosi University of Technology, Tehran, Iran
ARTICLE INFO ABSTRACT
Keywords: This paper presents a rapid machine learning-based damage detection framework for identifying the damage
Concrete shear wall extent of concrete shear wall buildings. For this purpose, a parametric study was carried out to determine
Machine learning the most efficient machine learning algorithm in classifying the damage states of the building. According to
Damage detection
this parametric study, the K-Nearest Neighbor (KNN) learner was selected as the reference prediction model
Ground motion
because of the higher accuracy achieved by this algorithm. Bayesian Optimization (BO) algorithm was used
Cumulative absolute velocity
to tune the hyperparameters affecting the accuracy of the model. The most efficient attributes were selected
from the set of damage indicators through the BO algorithm to train the model. Three different benchmark
buildings, including 7-,9-, and 13-story concrete shear wall buildings, were used to evaluate the robustness of
the proposed framework. A suite of 111 pair motions, originally developed for the SAC project, were employed
to create a generalized dataset. These motions were uniformly scaled from 0.05 g to 1.5 g to expand the
intensity range of the events. All the acceleration signals were polluted to 10% noise using white Gaussian
signals to simulate the field condition. Results reveal the efficiency of the proposed framework in identifying
the extent of damage in concrete shear wall elements of the building. In addition, a parametric study was
conducted to illustrate the reliability of two commonly used features, called Cumulative Absolute Velocity
(CAV) and the energy ratio between the acceleration response and the input excitation, in determining the
damage states of the shear walls under seismic motions.
1. Introduction element analysis to update the structural model [6]. The performance
of the model-based approaches directly depends on the accuracy of the
Assessment of structural safety is essential for post-earthquake information about the physical properties of the under-study structure.
restoration. Generally, to evaluate the post-earthquake vitality of the In addition, updating the finite element model based on the physical
exposed structures, a complete visual inspection is required [1]. Co- properties of the structure is computationally expensive for large-scale
ordination and implementation of the manual visual inspection needs structures, and rapid condition monitoring could be challenging in
several dedicated teams and monetary resources. In this regard, consid- this condition [7,8]. On the other hand, data-driven methods apply
erable efforts have been carried out to automate the visual inspection statistical learning algorithms to the vibration data captured from
process, e.g., image-based visual inspection [2]. However, such an the structure. This method uses learning algorithms to construct a
engineering visual inspection is only able to detect the visible defects
classification or regression learner for predicting structural damage [9,
that occurred in the structures [3]. This means that some serious
10]. Tsou and Shen [11] proposed the use of Neural Networks (NNs)
invisible damages may be left latent during the visual inspection. The
for predicting the severity and location of the structural damage. They
process of identifying and tracking the structural damage is known
used the variations in the modal properties of the structure as the
as the Structural Health Monitoring (SHM) [4,5]. In SHM, damage
damage feature to identify the damage. Worden et al. [12] proposed
detection is related to the methods developed for identifying the
probable existence, severity, and location of the structural damage. outlier analysis to compute the deviation of structural response from
Model-based and data-driven methods are two of the most commonly its normal condition for detecting the damage by solving a novelty
used strategies proposed for damage detection. Model-based methods detection problem. Zhao et al. [13] explored the efficiency of counter-
generally involve a system identification algorithm paired with a finite propagation NNs to locate the structural damage and identify the
∗ Corresponding author.
E-mail address: [email protected] (S. Soroushian).
https://doi.org/10.1016/j.istruc.2022.11.041
Received 5 October 2022; Received in revised form 9 November 2022; Accepted 11 November 2022
Available online 28 November 2022
2352-0124/© 2022 Institution of Structural Engineers. Published by Elsevier Ltd. All rights reserved. e-tarjome.com
H. Mahmoudi et al. Structures 47 (2023) 482–499
Fig. 1. Failure mechanisms of the concrete shear walls during the 2010 Chile earthquake.
support movement of beams in their axial direction. Yam et al. [14] linear and nonlinear response of the structure. These damage indicators
presented a method based on capturing the variations in the energy of were employed to identify the extent of damage based on the rec-
the structural vibrations using wavelet transform and Artificial Neural ommendations of ASCE-41 [33] for braced-frame buildings. Selecting
Networks (ANNs). Zhang et al. [15] proposed the regression tree and efficient features depends on the structural properties and the type of
random forest to evaluate the post-earthquake structural safety. Sajedi application that is expected from the method [34,35].
and Liang [16] proposed a near real-time damage detection framework Previous earthquakes show that the damage intensity is significantly
based on a fully convolutional encoder–decoder NN. Salkhordeh et al. varied in buildings [36]. The extent of damage could be either vis-
[17] proposed a decision-tree based algorithm for identifying the struc- ible due to the large deformations in different elements or hidden
tural damage in braced-frame buildings. Neves et al. [18] presented a because of the slight deflection in that elements. For example, primitive
dynamic decision-making algorithm to conduct the structural health evaluation of a building subjected to the 1994 Northridge earthquake
monitoring in bridges. They used the Bayesian theorem to update demonstrated that the structural elements had not suffered extensive
the prior possibilities of the structural condition based on the new damage. However, by eliminating the nonstructural components, the
information provided by the system. Different optimization algorithms site engineers reported that the dynamic performance of the building
can be used to tune the hyperparameters of the machine learning is highly reduced due to the failure of structural elements [37]. Re-
algorithms [19–22]. In general, the application of signal processing inforced concrete shear walls are widely used to provide the lateral
methods along with machine learning algorithms plays an imperative load resistance of buildings under earthquake motions because of their
role in implementing these approaches [1,1,23,24]. high stiffness and strength. Experimental investigations illustrated that
Feature extraction and compressing the size of damage index vec- the seismic performance of shear walls is different regarding the wall
tor are two important objectives of implementing a classification al- characteristics. In general, short wall response is dominated by shear
gorithm. The main object of reducing the size of the dataset is to and slender wall response governed by flexural behavior [38,39]. Fig. 1
ameliorate the learning performance and rate of the learning problem. shows failure mechanisms of concrete shear walls during the 2010 Chile
Principal Component analysis (PCA) [25], Locally Linear Embedding earthquake.
(LLE) [26], and Independent Component analysis (ICA) [27] are well- As shown in this figure, extensive crushing of the concrete followed
known tools in data size compression. Several studies were carried out by local buckling of the rebars led to the complete collapse of the
to find low-dimensional features that could be used as the damage walls. More than 100 concrete shear wall buildings were severely
index in the SHM. This low-dimensional features aim to avoid fitting damaged during the 2010 Chile earthquake (Mw = 8.8) [40]. In the
sophisticated statistical models to the response time series. Reed and light of these experiences, it can be concluded that implementing a
Kassawara [28] proposed two damage indicators based on Cumulative monitoring scheme in these buildings is essential not only for mitigating
Absolute Velocity (𝐶𝐴𝑉 ) and response spectrum acceleration to deter- the lifetime damages but also for early identifying the post-earthquake
mine the existence of damage in the structures. Cabanas et al. [29] damages. According to the previous studies, there are two main re-
proposed the modified 𝐶𝐴𝑉 and Arias intensity (𝐼𝐴 ) as two damage search gaps in (1) rapid estimation of the local damage exposed to the
indexes that have an appropriate correlation with the damage. Ban- shear-walls of concrete buildings and (2) investigating the efficiency of
dara et al. [30] presented a damage index based on the Principal different damage features derived from the ground motion records and
Components (PCs) of the Frequency Response Functions (FRF) of the the structural response in identifying the local damage of the structure
structure. Muin and Mosalam [31] investigated the performance of caused by ground motions. In addition, in machine-learning-based
𝐶𝐴𝑉 for localizing the structural damage in structural systems. Sajedi structural health monitoring, a finite number of studies investigated
and Liang [32] presented two features based on the energy ratio of the the most powerful learning algorithm as well as damage indices for
structural response for detecting the damage. Salkhordeh et al. [17] detecting the structural damage.
showed that the features based on the energy ratio of the structural This paper presents a machine-learning-based damage detection
response are not efficient as a dependent attribute to classify the struc- framework for identifying the extent of damage exposed to the concrete
tural damage after the earthquake. They combined an energy-based shear-wall buildings during the earthquake. For this purpose, a series
feature with two additional attributes, including drift ratio extracted of nonlinear response history analyses are implemented on three 3D
from acceleration response and the correlation coefficient between the (three-dimensional) buildings using a suite of 111 ground motions that
483
Table 1
Material properties of steel rebars.
Element Yield strength (𝑘𝑠𝑖) Modulus of elasticity (𝑘𝑠𝑖) Strain hardening ratio
Steel rebar grade 𝐴𝐼𝐼𝐼 58 29000 0.01
Table 2
Material properties of concrete.
Confinement type Expected compressive Strain at Crushing strength Strain at
strength (𝑘𝑠𝑖) maximum strength crushing strength
Confined concrete 5.55 0.003 1.11 0.03
Unconfined concrete 4.26 0.003 0.85 0.01
Table 3
Results of mesh sensitivity analysis.
Model Mesh size 𝐼𝐷𝑅𝑚𝑎𝑥 Analysis time Error
1 0.5 m × 0.5 m 0.0173 5h , 10min –
2 1.0 m × 1.0 m 0.0177 1h , 17min 2.3%
3 1.5 m × 1.5 m 0.0213 58min 23.1%
4 2.0 m × 2.0 m 0.0228 34min 31.8%
originally developed in the SAC project [41]. To increase the rate

of training and eliminate unnecessary features, Bayesian Optimiza-
tion (BO) is applied. In addition, hyperparameters of the classification
learner are optimized by using the BO algorithm. This paper sheds
significant light on rapid damage estimation of concrete shear-wall
buildings using machine-learning techniques. The present paper is or-
ganized in such a way that first, the benchmark buildings and their
modeling procedure are described. Then, the damage indicators derived
from the vibration data recorded during the earthquake motions are
introduced. Next, different classification algorithms (i.e. decision trees,
support vector machine, k-nearest neighbor, linear discriminant anal-
ysis, and Naive Bayes) used to achieve the best classification learner
are explained. In subsequent, the implementation procedure of the
proposed framework is presented. Finally, the results obtained for
determining the extent of damage in benchmark structures (concrete
shear wall buildings) are discussed.
2. Benchmark buildings
This study aims to predict the structural damage in concrete shear-

wall buildings after a damaging earthquake. To investigate the effi-
ciency of the proposed method, three existing 7-, 9-, and 13-story
concrete shear wall buildings were used as the benchmark models.
These buildings were designed in high seismic zone according to the
corresponding modern code. The modeling procedure and the struc-
tural details of these benchmark buildings are explained in the next
subsections.
2.1. Seven-story building
A 3D seven-story building is selected to assess the performance of

the proposed method in identifying the structural damage of concrete
shear wall buildings. This building includes two shear walls continued
through all the stories. Fig. 2 shows a general view of the building as Fig. 2. General view of the 7-story building.
well as its plan view. It is worthwhile to mention that the first three
natural periods of the structure were respectively equal to 1.27 s, 0.66 s,
and 0.45 s. 2.3. Thirteen-story building
2.2. Nine-story building

The third case study model was an existing thirteen-story concrete
The second case study model was an existing nine-story concrete shear wall building. Fig. 4 shows a general view of this model as well
shear wall building. As shown in Fig. 3, this building had a core as as its plan view. According to this figure, this building has six concrete
well as a single shear wall that continued from the underground level shear walls continued through all the stories of the structure. The first
to the roof level. The remaining walls shown in Fig. 3(b) only belonged three natural periods of the building were equal to 1.99 s, 1.84 s, and
to the underground floors. It is notable that the first three fundamental 0.76 s. The modeling details of the benchmark buildings are explained
periods of the structure were respectively 1.63 s, 0.83 s, and 0.54 s. in the following section.
484
Fig. 3. General view of the 9-story building.
2.4. Numerical modeling methodology
All the steel rebars are modeled using the Giuffre-Menegotto-Menegotto-

Pinto uniaxial strain-hardening material model defined as 𝑆𝑡𝑒𝑒𝑙02
is 𝑂𝑝𝑒𝑛𝑠𝑒𝑒𝑠 platform. The yield stress and modulus of elasticity for
𝐴𝐼𝐼𝐼 steel bars were 58.0 ksi and 29000 ksi, respectively. In addition, Fig. 4. General view of the 13-story building.
the strain hardening ratio of steel material was designated as 0.01.
The recommended values of 0.15, 0.925, and 12 were respectively
considered for the model parameters 𝑐𝑅2, 𝑐𝑅1 and 𝑅0 to control the
𝑓𝑝𝑐𝑢 = 0.2𝑓𝑝𝑐 and the concrete strain at crushing strength was 𝑒𝑝𝑠𝑈 =
shape of hysteresis curves of the rebars. These parameters control the
0.01. The parameters 𝑓𝑝 𝑐 and 𝑒𝑝𝑠𝑐0 for confined concrete were calculated
transition of the hysteresis curve from elastic region to plastic branch.
using the recommendation of Saatcioglu and Razvi [44], depending on
The isotropic hardening parameters 𝑎1, 𝑎2, 𝑎3, and 𝑎4 defined as the
the area, spacing, and configuration of the transverse reinforcement.
default values of 1.0, 0, 1.0 and 0, respectively. Table 1 summarizes
the material properties of the steel rebars considered for reinforced For the confined concrete, the expected compressive strength 𝑓 𝑝𝑐 was
concrete elements. Cross-sectional properties of concrete beams and 5.55 ksi. Similar to the unconfined concrete, the crushing strength of
columns were defined using the 𝐹 𝑖𝑏𝑒𝑟𝑆𝑒𝑐𝑡𝑖𝑜𝑛 command in 𝑂𝑝𝑒𝑛𝑠𝑒𝑒𝑠. confined concrete was also considered equal to 0.2𝑓𝑝𝑐 . The crushing
The material model of both confined and unconfined concrete was strain of the confined concrete was considered similar to the com-
defined using the model presented by Yassin [42] which is designated pression strain 𝜖20 of the Saatcioglu and Razvi model [44]. The ratio
in 𝑂𝑝𝑒𝑛𝑆𝑒𝑒𝑠 as 𝐶𝑜𝑛𝑐𝑟𝑒𝑡𝑒02. For unconfined concrete, the expected com- between the unloading slope and the initial slope was selected to be the
pressive strength 𝑓𝑝𝑐 was 4.26 ksi, the concrete strain 𝑒𝑝𝑠𝑐0 at maximum default value of 0.1. The tensile strength of concrete 𝑓𝑡 was considered
√
strength was 0.003 (which represents an initial concrete modulus (𝐸0 ) as 0.33 𝑓𝑐′ [45] and the softening stiffness of concrete in tension 𝐸𝑡𝑠
equal to that defined in ACI 318 [43]), the concrete crushing strength was defined as 0.05𝐸0 according to the recommendation of Yassin [42].
485
Fig. 6. Inter-story drift ratio of the structure for different mesh sizes.
Bayesian optimization [48]. Doing so leads to increasing the learning

rate of the classifier and reducing the computational cost. The following
subsections are going to explain these damage indexes.
3.1. Arias intensity (𝐼𝐴 )
Fig. 5. Shell element for modeling the concrete shear walls [46]. Cabanas et al. [29] proved that the Arias intensity of the ground
motions is related to the severity of damage in earthquake-exposed
buildings. The Arias intensity of a signal can be defined as follow:
Table 2 shows the material properties of both confined and unconfined 𝜋
𝑡
concrete used for all the case studies. All the beam and column 𝐴𝐼 = ̈ 2 𝑑𝑡
(𝑢(𝑡)) (1)
2𝑔 ∫0
elements were modeled using the displacement-based beam–column
where 𝑢(𝑡)
̈ is the earthquake acceleration at time step 𝑡. Since the Arias
element. The shear wall elements were assigned to the structure using
intensity measure does not include any information about the structural
the 𝑆ℎ𝑒𝑙𝑙𝑁𝐿𝐷𝐾𝐺𝑄 element. Material properties of each layer and
properties, this index is not accurate in identifying the local damage in
corresponding geometrical characteristics of the wall can be assigned
structural elements. For example, different ground motions may have
to this element as shown in Fig. 5(a). This element is developed to take
similar input energies, but impose different damage severities to the
into account the geometric nonlinearity of large deformations using the
building because of differences in the frequency content of the records.
updated Lagrangian method. The 𝑆ℎ𝑒𝑙𝑙𝑁𝐿𝐷𝐾𝐺𝑄 element simplifies
the nonlinear behavior of 3D concrete shear walls by discretizing
3.2. Cumulative Absolute Velocity (CAV)
different elements of the wall into separate fully-bonded layers along
the thickness direction. 𝐿𝑎𝑦𝑒𝑟𝑒𝑑𝑆ℎ𝑒𝑙𝑙 command was used to create the
As mentioned before, Reed and Kassawara [28] illustrated that
fiber-based multi-layer section of the shear walls. As shown in Fig. 5(b),
the CAV is correlated with the extent of damage that imposed to the
the rebars are defined as spread orthotropic layers considering the
structure during an earthquake. The CAV of the structural response can
direction and location of them. The axial strain and curvature of the
be written as below:
middle layer is firstly calculated. Then, the strain of other layers is 𝑇
calculated by applying the plane-section assumption [46,47]. 𝐶𝐴𝑉 = |𝑎(𝑡)| 𝑑𝑡 (2)
Remarkably, all the slabs were defined as rigid diaphragms in both ∫0
horizontal directions. Since the size of meshes considered for the shear where 𝑇 is the duration of the structural response, 𝑎 is the acceleration
wall elements highly affects the computational time, a mesh sensitivity response of the building’s floor. Unlike the Arias intensity, the CAV
analysis was performed to achieve an optimum model with acceptable criterion includes the nonlinearity condition of the structure. However,
accuracy and analysis time. Fig. 6 shows the effect of meshing size on the value of CAV highly depends on the duration of the ground motion
the inter-story drift ratio of the building. In addition, Table 3 shows record. For instance, a ground motion with a lower value of PGA and
the results of mesh sensitivity analysis for 7-story building. This Table a larger duration may results in a larger value of CAV compared to a
summarizes the computational time and maximum inter-story drift record with a larger PGA and lower duration.
ratio of the building under the 𝐶𝑜𝑦𝑜𝑡𝑒𝐿𝑎𝑘𝑒 earthquake record. As can be
seen, the size of mesh was selected as 1.0 m×1.0 m because of promising 3.3. Modified Cumulative Absolute Velocity (MCAV)
accuracy and computational time compared to the 0.5 m×0.5 m meshing
size. According to the study implemented by Salkhordeh et al. [17], the
CAV may not be efficient in detecting the extent of damage because
3. Feature extraction of its dependence on the earthquake duration. Therefore, this study
uses the Modified Cumulative Absolute Velocity (MCAV) as one of the
Vibration-based damage detection methods highly depends on the damage features. The MCAV attribute is written as below:
damage index used for revealing the extent of the damage. This paper ∑
𝑁 𝑖
uses seven different damage features and investigates the efficiency 𝑀𝐶𝐴𝑉 = (𝐻(𝑃 𝐺𝐴𝑖 − 𝑐) |𝑎(𝑡)|𝑑𝑡) (3)
∫𝑖−1
of these attributes in diagnosing the extent of damage in concrete 𝑖=1
shear wall buildings. This research reduced the number of damage where 𝑁 is the number of one-second time intervals of the ground
features by selecting efficient attributes from the set of features through motion, 𝐻(𝑥) is the Heaviside step function defined as 0 for 𝑥 ≺ 0 and
486
1 for 𝑥 ≻ 0, 𝑃 𝐺𝐴𝑖 is the peak ground acceleration for 𝑖th time interval, Notably, this method is different with the conventional linear filtering
and 𝑐 is a 𝑃 𝐺𝐴 threshold to neglect the summation for values smaller approaches because of implementing a nonlinear shrinkage on coeffi-
than this threshold value [49]. cients in the transform domain. Moreover, the conventional de-noising
methods assume the low signal-to-noise ratio for the signal, which
3.4. Spectral acceleration may apply additional bias to the filtered signal. In addition, unlike the
parametric methods that need an initial approximation of the model
It seems necessary to define a damage indicator that includes both parameters, wavelet de-noising approaches are non-parametric, and
characteristics of the structure and input ground motions. To do so, the the correction is related to specific properties of the signal. Hence,
spectral acceleration of the earthquake records at the first fundamental wavelet de-noising approach is applied in this study to prepare the
period of the building is defined as a damage feature. This attribute noisy signal for integration. Fundamental characteristics required to
enables the framework to learn the extent of damage based on the implement the method were defined same as the recommendations
intensity of ground motion and also the dynamic properties of the of Ansari et al. [56],Soroushian et al. [57], which evaluate the ef-
building. ficiency of wavelet multi-resolution analysis in correcting the noisy
signals. This method includes two phases based on wavelet de-noising
3.5. Energy ratio to modify the acceleration time histories.
In the first phase, the signal decomposes into 𝑁 levels, in which the
Damage severity in a building has a direct relation to the nonlin- 𝑁 value is selected to achieve the frequency range of interest for the
𝑁th detail of the signal. Next, two modifiers are subjected to the detail
earity that occurred in the structural elements [9]. This nonlinearity is
values of the signal if its amplitude violate a pre-defined threshold
estimable from the acceleration response of the structure. To capture
level [51–54] The modification functions depend on the single positive
the nonlinearity in the structure, a relative energy ratio is defined
parameter 𝜂, called the threshold as:
below:
𝑇 ⎧𝐷 − 𝜂, 𝐷⪰𝜂
∫0 |𝑢(𝑡)|
̈ 𝑑𝑡 ⎪
𝑅𝐸 = (4) 𝑆𝑜𝑓 𝑡 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑𝑖𝑛𝑔 ∶ 𝑇𝑠 (𝐷, 𝜂) = ⎨0, |𝐷| ≺ 𝜂 (9)
𝑇 | |
∫0 |𝑢̈ 𝑔 (𝑡)| 𝑑𝑡 ⎪
| | ⎩𝐷 + 𝜂, 𝐷 ⪯ −𝜂
where, 𝑢(𝑡)
̈ and 𝑢̈ 𝑔 (𝑡) are the recorded acceleration from the floor {
0, |𝐷| ≺ 𝜂
and ground level of the structure, respectively. Nonlinearity in the 𝐻𝑎𝑟𝑑 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑𝑖𝑛𝑔 ∶ 𝑇𝑠 (𝐷, 𝜂) = (10)
structural elements leads to the change in the relative ratio of 𝐶𝐴𝑉 𝐷, |𝐷| ⪰ 𝜂
in different Degree of Freedoms (DOFs) of the building. Generally, thresholds can be classified into two categories: (1) ‘‘single
thresholds’’ and (2) ‘‘level-dependent thresholds’’. In the first method,
3.6. Drift a single threshold determined according to the properties of the signal
is imposed to the wavelet coefficients. In the second approach, the
It is well-understood that the severity of damage exposed to the threshold values are determined regarding the characteristics of the
building under the earthquake motions has a direct relation with signal at different levels. After the thresholding step, the corrected
the inter-story drift [50]. Ideally, the story drift time–history can be signal was reconstructed according to its approximation and detail
computed by implementing two consecutive integration on the rel- values.
ative acceleration response of the structure. However, there is an The second phase has the same steps as the first one, and only the
unwanted noise in the acceleration response histories captured by the following differences should be considered.
accelerometers. This makes the obtained drift history unreliable. There-
• Modification factors are used for the velocity signal obtained
fore, signals acquired by sensors should be filtered from noise by using
by applying multiple integration to the acceleration response
the process known as ‘‘de-noising’’. The present study implements the
achieved in the first phase.
non-parametric de-noising approach known as ‘‘wavelet shrinkage de-
• The velocity signal should be parsed to 𝑀 levels (where 𝑀 ⪰
noising’’. This method was originally developed by Donoho et al. [51–
(𝑁 + 2)).
54]. The wavelet function 𝜓(𝑡) ∈ 𝐿2 (ℜ) can be written as follow [55]: • After the thresholding step, the modified velocity signal is re-
( )
𝑡−𝑏 constructed by applying a summation on only the details of the
𝜓𝑎,𝑏 (𝑡) = |𝑎|−0.5 𝜓 (5)
𝑎 velocity signal.
where 𝐿2 (ℜ) represent the suite of integrable functions, 𝑎 and 𝑏 are
It should be mentioned that the first phase is useful for modifying the
respectively the scale and translation factors, |𝑎|−0.5 is defined to nor-
high-frequency noises, and the second phase is suitable for correcting
malize the value of ‖ ‖
‖𝜓𝑎,𝑏 ‖, and 𝜓 is the wavelet function. The wavelet the low-frequency noises of the signal. In this study, according to the
function has a zero mean value, and must meet the admissibility study of Soroushian et al. [57], hard-thresholding is used together with
condition as below: the SureShrink approach to modify the acceleration signals. Notably,
∞ 2
|𝜓(𝜔)|
̂ SureShrink is an efficient approach for computing the level-dependent
𝐶𝜓 = 𝑑𝜔 < ∞ (6)
∫−∞ |𝜔| thresholds. In addition, to prevent the phase distortion that is usual
in conventional methods of de-noising, the symlet wavelet is used
where 𝜔 refer to the frequency, and 𝜓(𝜔)
̂ = ∫ℜ 𝜓(𝑥) exp (−𝑖𝑥𝜔)𝑑𝑥 is the
because of its approximately linear phase response [56]. For example,
Fourier transform of the function 𝜓. The inner product of the wavelet
Fig. 7 shows the result of using ‘‘wavelet de-noising’’ for correcting
families 𝜓𝑎,𝑏 (𝑡) to the signal 𝑥(𝑡) is defined as Continuous Wavelet
the drift response of the first floor of 9-building under the 𝐶𝑜𝑦𝑜𝑡𝑒𝐿𝑎𝑘𝑒
Transform (CWT):
earthquake (26.86s duration) in presence of 10% noise. This Figure
∞
1 𝑡−𝑏 also illustrates a drift response that is directly calculated from the
𝐹𝑊 (𝑎, 𝑏) =≺ 𝑥(𝑡), 𝜓𝑎,𝑏 (𝑡) ≻= 𝑥(𝑡) √ 𝜓 ∗ ( )𝑑𝑡 (7)
∫−∞ 𝑎 𝑎 acceleration response of the building polluted with 5% noise. It is
clear that the drift response obtained through direct integrating on
where 𝜓 ∗ indicates the complex conjugate of 𝜓. This equation converts
the acceleration history is far from the real response even when the
to the Discrete Wavelet Transform (DWT) where parameters 𝑎 and 𝑏
acceleration response of the building contains 5% noise. However,
are defined as discrete values. To reconstruct the signal 𝑥(𝑡) from its
the corrected drift history is well-correlated with the response of the
wavelet transform, Calderon’s identity can be written as below:
building. It should be noted that the ‘‘wavelet de-noising’’ approach
∞ ∞
1 𝑡 − 𝑏 𝑑𝑎 𝑑𝑏 may not be efficient for certain response because of different frequency
𝑥(𝑡) = 𝐹 (𝑎, 𝑏)𝜓( ) √ (8)
𝐶𝜓 ∫−∞ ∫−∞ 𝑊 𝑎 𝑎2 𝑏 content of the ground motions.
487
Fig. 7. Efficiency of the ‘‘wavelet de-noising’’ in correcting the drift response.
3.7. Correlation coefficient
Fig. 8. The effect of nonlinearity on the correlation coefficient damage indicator.

THe correlation coefficient between the nonlinear acceleration re-
sponse of the building (real structure) and the acceleration response
obtained from the corresponding linear system was used as a damage
indicator. A linear and time-invariant transfer function 𝐻(𝜔) was de- these responses becomes lower (close to zero for severe damage). The
fined as the proportion of the response signal 𝑦 and input excitation 𝑥 correlation of two signals depends on the intensity of the relationship
as below: between the relative movements of these two signals [59]. Here, the
𝑆𝑦𝑦 (𝜔) correlation coefficient between the nonlinear response of the building
𝐻(𝜔) = (11) (real building) and corresponding linear system was written as follow:
𝑆𝑥𝑦 (𝜔)
𝐸[(𝑋 − 𝑚𝑋 )(𝑌 − 𝑚𝑌 )]
where 𝑆𝑥𝑦 represent the cross power spectral density between the exci- 𝐶𝑋,𝑌 = 𝜌𝑥,𝑦 = (12)
𝜎𝑋 𝜎𝑌
tation and response signals, and 𝑆𝑦𝑦 is the power spectral density of the
acceleration response of the structure. The equivalent linear response where, 𝑋 was the acceleration response recorded from the structure,
of the building, 𝑦(𝑡), is estimable by calculating the inverse Fourier and 𝑌 was the acceleration response of the equivalent linear sys-
transform of 𝐻(𝜔)𝑋(𝜔), where 𝑋(𝜔) is the Fourier spectrum of the tem. Moreover, 𝑚𝑋 and 𝑚𝑌 were the average values of the acceler-
input excitation. The transfer function of a structure can be obtained ation response for the real building and its equivalent linear system,
by: (1) using sensory data acquired from the minor ground motions respectively.
recorded during the lifetime of the structure, or (2) by conducting an
ambient vibration-based experiment [58]. In the present research, the 4. Machine learning algorithms
acceleration response of the building under a minor earthquake is used
to calculate the transfer function of the structure. Machine learning algorithms are extensively progressed in earth-
In general, by increasing the severity of damage in the building, the quake engineering, especially for damage detection, system identi-
acceleration response deviates more from the response of the equivalent fication, and risk assessment applications. This paper implements a
linear system because of nonlinearity in the structural elements. Fig. 8 parametric study to determine the best classification learner among the
illustrates this matter for two examples, one for the response of the Support Vector Machine (SVM), K-Nearest Neighbors (KNN), decision
structure under the scaled 𝐶𝑜𝑦𝑜𝑡𝑒𝐿𝑎𝑘𝑒 ground motion with 𝑃 𝐺𝐴 = tree, linear discriminant, and Naive Bayes algorithms. The next sub-
0.1 g and another for 𝑃 𝐺𝐴 = 1.2 g. It is obvious that whatever the sections are going to present a brief description of the mathematical
ground motion intensity increases, the correlation coefficient between foundations of these algorithms.
488
Table 4
Distance measures proposed for W-KNN.
Distance measure Formula
√∑
𝑛
Euclidean 𝐷𝐸𝑢𝑐𝑙𝑖𝑑𝑒𝑎𝑛 = 𝑖=1
(𝑎𝑖 − 𝑏𝑖 )2
∑𝑛
𝐷𝑀𝑖𝑛𝑘𝑜𝑤𝑠𝑘𝑖 = ( 𝑖=1 ||𝑎𝑖 − 𝑏𝑖 || ) 𝑐
𝑐 1
Minkowski
∑𝑛 |
Manhattan 𝐷𝑀𝑎𝑛ℎ𝑎𝑡𝑡𝑎𝑛 = 𝑖=1 |𝑎𝑖 − 𝑏𝑖 ||
∑𝑛
𝐷𝐶ℎ𝑒𝑏𝑦𝑠ℎ𝑒𝑣 = lim𝑐→∞ ( 𝑖=1 ||𝑎𝑖 − 𝑏𝑖 || ) 𝑐
𝑐 1
Chebyshev
√
Mahalanobis 𝐷𝑀𝑎ℎ𝑎𝑙𝑎𝑛𝑜𝑏𝑖𝑠 = (𝑎⃗ − 𝑏)⃗ 𝑇 𝑆(𝑎⃗ − 𝑏)⃗
⃗ 𝑏⃗
𝑎.
Cosine 𝐷𝐶𝑜𝑠𝑖𝑛𝑒 =
‖𝑎‖‖
⃗ 𝑏‖⃗
of occurrence (𝑃𝑖 ), 𝐵𝑖 is defined as follow:

𝑃
𝐵𝑖 = ∑𝑚 𝑖 (16)
𝑖=1 𝑃𝑖
In order to quantify the misclassification cost, 𝛾 is defined. The misclas-
sification penalty will increase for positive class by increasing the value
Fig. 9. Response spectra for the set of earthquake motions. of hyperparameter 𝛾. This parameter is used to establish a uniform
distribution for different labels [60].
4.2. K-Nearest Neighbors (KNN)
KNN is a non-parametric algorithm developed for the classification

and regression [61,62]. The non-parametric term means that the algo-
rithm does not consider any pre-defined assumption on the underlying
data. This algorithm consists of two steps: (1) finding a group of KNN
for a given dataset using different distance metrics, and (2) labeling the
query based on the dominant class in KNN. Supposing the training set
{ }𝑛
(𝑥𝑖 , 𝑦𝑖 ) 𝑖=1 in which, 𝑥𝑖 is a 𝑧-dimensional vector and 𝑦𝑖 is the corre-
sponding label, for a query 𝑥𝑗 , the algorithm determines its unknown
𝑦𝑗 as follows:
• Compute the distance between 𝑥𝑗 and each 𝑥𝑖 in the dataset

• Arrange the calculated distances in descending order
• Select K samples that have the nearest distance to 𝑥𝑗
• The label of 𝑥𝑗 is assigned based on the majority classes of the
KNNs
The Weighted-KNN (W-KNN) is an extension of KNN, which assigns a

Fig. 10. Effect of 10% noise on the structural response. weight to the neighbors based on their distance. The value of weight
would be increased by decreasing the distance of the neighbors to the
test query 𝑥𝑗 . There are various types of distance and weight mea-
4.1. Support Vector Machine (SVM) sures proposed for W-KNN such as Euclidean, City block, Chebyshev,
Jaccard, etc. In addition, the squared inverse and inverse rules are
the main weighting measures presented for KNN. Table 4 summarizes
SVM is a non-probabilistic binary classification learner. Typically, the formulation of some important distance measures defined for two
SVMs enlarge the features space by applying effective calculations hypothetical vectors 𝑎⃗ and 𝑏⃗ where 𝑆 is the covariance matrix of these
through different nonlinear kernel functions. This statistical learner is vectors.
typically developed for binary classification at which each observation
is labeled as 𝑦𝑖 ∈ {−1, 1}. The optimization problem in SVM can be 4.3. Decision tree
written as below:
A decision tree classifier is defined as a recursive partitioning of
1 ∑ ∑
𝑚𝑖𝑛 ‖𝜔‖2 + 𝛾𝜆 𝐵𝑖 𝜉𝑖 + 𝜆 𝐵𝑖 𝜉𝑖 (13) the paradigm space to map the observations into the target values.
𝜔,𝛽 2
𝑦 ∈𝐶𝑖
+ 𝑦 ∈𝐶 −
𝑖 The learning process of decision trees is similar to the structure of
a flowchart in which the uppermost node (root node) implements a
𝑠𝑢𝑏𝑗𝑒𝑐𝑡𝑒𝑑 𝑡𝑜 ∶ (14)
[ ] test on the features. The outcome of each node is partitioned through
𝑇
𝜉𝑖 ≥ 0, ; 𝑦𝑖 ℎ(𝑥𝑖 ) 𝜔 + 𝛽 ≥ 1 − 𝜉𝑖 ∀𝑖; (15) branches and the corresponding label is determined using leaf nodes.
In general, decision trees are categorized as (1) classification trees,
where 𝜔 is the set of weights, 𝛽 is the bias parameter, and 𝜉 is the
and (2) regression trees. In the former one, the set of features are
slack variable utilized to model a soft-margin classifier. Slack variables labeled as the finite number of classes they belong to it. However, the
are useful to construct a generalized decision boundary by permitting samples are labeled as arbitrary real numbers in the regression trees.
some error near the boundaries. 𝜆 is the box constraint used to make Detailed information on the regression trees is neglected because the
a balance between the generalization and training accuracy. To train a present study is focused on the classification learners. The main issue
classifier that performs better for observations with higher probability in providing a classification tree is selecting the most effective feature
489
Table 5 finally, the NBC uses the Maximum A Posteriori (MAP) decision rule [68]
Different splitting criteria proposed for decision tree algorithms.
to determine the class for each sample. Therefore, the NBC classifier is
Splitting criteria Formula a function that assigns a class label 𝑦̂ = 𝐶𝑘 to a set of 𝑘 observations as
∑𝑘
Entropy 𝐸(𝑋) = − 𝑗=1 𝑝𝑗 log2 (𝑝𝑗 ) define log2 (0) = 0 below:
∑𝑘 ∑𝑘 ∏𝑛
Gini index 𝐺(𝑋) = 𝑗=1 𝑝𝑗 (1 − 𝑝𝑗 ) = 1 − 𝑗=1 𝑝2𝑗 𝑃 (𝑥𝑖 |𝑐𝑘 )𝑃 (𝑐𝑘 )
[∑
𝑘 |
]
| 2 𝑎𝑟𝑔𝑚𝑎𝑥𝑘∈(1,…,𝐾) = 𝑖=1 (21)
Towing rule
𝑃 𝑃
𝛷(𝑋) = 𝐿4 𝑅 𝑗=1 ||𝑃 (𝑐𝑗 |𝐴𝐿 ) − 𝑃 (𝑐𝑗 |𝐴𝑅 )||
𝑃 (𝑥)
∑𝑘 despite the simplified assumptions that laid at the foundation of the
Deviance 𝐷(𝑥𝑖 ) = −2 𝑗 𝑥𝑖𝑗 log 𝑝𝑗
NBC, several studies are reported surprising effectiveness of this algo-
rithm even with strong dependencies [67].
for partitioning the samples at each node. Several splitting criteria 4.5. Linear Discriminant Analysis (LDA)
were introduced during development of decision trees, e.g. towing rule,
deviance, Gini index, and entropy [63]. Table 5 summarizes some
The discriminant analysis is a linear combination of the features to
common attribute selection criteria proposed for different decision tree
reduce the size of feature’s space and scale the features according to
algorithms. In a binary classification tree, at a given node 𝐴, there
their importance [69]. Many feature extraction techniques apply linear
are 𝑁-point, 𝐾-class dataset. This node splits dataset to 𝐴𝑅 and 𝐴𝐿
transformations on the original vector of patterns to achieve new vec-
(right and left nodes) with a proportion of 𝑃𝑅 and 𝑃𝐿 for the right
tors with lower dimension. The main objective of Linear Discriminant
and left branches,
{ respectively.
} The outcome of each node labeled by 𝐾
class labels, 𝑐𝑖 , … , 𝑐𝑘 . The probability of the sample 𝑋 is calculated Analysis (LDA) is to search a vector from the vectors space that presents
∑ better separation between different classes of data. The separability of
as 𝑝(𝑋 = 𝑐𝑗 ) = 𝑝𝑗 , in which 𝑗 = 1, 2, … , 𝑘 and 𝑘𝑗=1 𝑝𝑗 = 1. In
this Table, the minimum value of the entropy is 0 and occurs when each vector can be examined by projecting the original set of data onto
one of 𝑝𝑗 = 1 (𝑗 = 1, 2, … , 𝑘). This means that all other 𝑝𝑗 are these vectors. When the transformed data illustrate significant overlap,
equal to 0 [64] The minimum value of Gini index is equal to 0. This the LDA algorithm seeks a better separation by using a transformation
minimum value will occur if all samples in the node be of a single rule know as Fisher ratio [70]. The fisher ratio for a binary problem is
category [65]. Notably, the towing rule is efficient for a dataset labeled defined as below:
by large number of various classes [66]. In the deviance measure, (𝜇1 − 𝜇2 )2 𝑆𝐵
𝑥𝑖𝑗 (𝑖 = 1, … , 𝑛 𝑎𝑛𝑑 𝑗 = 1, … , 𝑘) is the 𝑖th observation related to the Kth- 𝐹𝑅 = = (22)
𝜎12 + 𝜎22 𝑆𝑊
component from the random vector 𝑋. Generally, the splitting criteria
designate the type of splitting that results in the maximum reduction where 𝜇1 and 𝜇2 are the mean or center points of the first and second
in the a parameter called ‘‘impurity measure’’. The impurity measures classes, respectively. It is obvious that 𝜇1 − 𝜇2 represents the distance
represent the homogeneity of training examples in the child node after between the centroids of two classes. 𝜎1 and 𝜎2 are also the variance
each partitioning [66]. of the first and second classes, respectively. Thus, the LDA tries to
Generally, the training process in decision tree learners is continued maximize the distance between two classes by maximizing the Fisher
until all the samples at the subsets have a unique label. In this condi- ratio. In other words, it maximizes the scatter between two classes
tion, it is being said that the tree is a ‘‘complex tree’’ and is overfitted (𝑆𝐵 ) while making two classes as condense as possible by minimizing
on the dataset. The overfitting problem reduces the generalization within class scatter (𝑆𝑊 ). To meet this goal, a transformation vector 𝑤
of the predictive model, and the accuracy of an overfitted model is is applied as below:
low for new data. The pruning is defined as the process of removing
𝑤𝑇 𝑆𝐵 𝑤
the branches and leaves of the tree that decreases the test accuracy 𝐹𝑅 (𝑤) = (23)
of the model. Two general approaches are recommended for pruning 𝑤𝑇 𝑆 𝑊 𝑤
process, (1) pre-pruning, and (2) post-pruning technique. The readers finally, the algorithm is intended to find a transformation vector 𝑤 that
are encouraged to review [66] to achieve more information on the maximizes the Fisher ratio. By differentiating 𝐹𝑅 (𝑤) in terms of 𝑤, the
pruning process. following eigen value problem will obtain:
−1
4.4. Naive Bayes 𝑆𝑤 𝑆𝐵 𝑤 = 𝜆𝑤 𝑤ℎ𝑒𝑟𝑒 𝜆 = 𝐹𝑅 (𝑤) = 𝑠𝑐𝑎𝑙𝑎𝑟 (24)
solving the eigen value problem yields:

The Naive Bayes Classifier (NBC) is a supervised classification al-
gorithm which founded based on the class conditional independence −1
𝑤 = arg max 𝐹𝑅 (𝑤) = 𝑆𝑤 (𝜇1 − 𝜇2 ) (25)
𝑤
assumption [67]. For an input vector, 𝑋 = (𝑥1 , 𝑥2 , … , 𝑥𝑛 ), the con-
ditional probability of 𝑘th possible outcome (𝐶1 , 𝐶2 , … , 𝐶𝑘 ) can be after projecting the feature’s space onto the projection vector, a Gaus-
defined as below: sian distribution can be fitted on each class to conduct the classification
𝑃 (𝑋|𝐶𝑘 )𝑃 (𝐶𝑘 ) problem.
𝑃 (𝐶𝑘 |𝑋) = (17)
𝑃 (𝑋)
5. Implementation
the denominator 𝑃 (𝑋) is as follow:
∑
𝑘 ∏
𝑛
So far, the mathematical foundation of the proposed framework is
𝑃 (𝑋) = ⟮𝑃 (𝐶𝑗 ) 𝑃 (𝑥𝑖 |𝐶𝑗 )⟯ (18)
𝑗=1 𝑖=1
discussed. For better understanding, this section presents an overall
view of the proposed approach.
the following equation can be written by assuming that all factors in 𝑋
are conditionally independent:
5.1. Input excitations
∏
𝑛
𝑃 (𝑋|𝐶𝑘 ) = 𝑃 (𝑥𝑖 |𝐶𝑘 ) (19)
𝑖=1
In order to provide a generalized dataset, a suite of 111 pair mo-
tions (Table A.1), originally developed for the SAC project [41], are
the following equation is derived by substituting Eq. (19) and (18) into
employed. These motions are uniformly scaled to 16 different Peak
Eq. (17):
∏𝑛 Ground Acceleration values (PGAs), ranging from 0.05 g to 1.5 g, to
𝑃 (𝑥𝑖 |𝐶𝑘 )𝑃 (𝐶𝑘 ) generate a wide range of severity for the input excitations. Notably, the
𝑃 (𝐶𝑘 |𝑋) = ∑𝑘 𝑖=1 ∏𝑛 (20)
𝑗=1 ⟮𝑃 (𝐶𝑗 ) 𝑖=1 𝑃 (𝑥𝑖 |𝐶𝑗 )⟯
SAC motions include a set of far-field, near-field, and simulated records
490
representing a wide range of frequency contents. Therefore, the set of

ground motions provide a generalized set of structural response for
the classification problem. Fig. 9 illustrates the PGA-normalized elastic
acceleration response spectra with 5% damping ratio for the suite of
ground motions.
5.2. Effect of noise
It is well understood that the sensory data are accompanied by

unwanted noise. The effect of noisy structural response on the learning
process was considered by adding white Gaussian noise to the acceler-
ation signals acquired from each floor of the buildings. The coefficient
of signal-to-noise ratio is defined to measure the level of noise that is
added to the signals. This coefficient is written as below:
2
𝜎𝑠𝑖𝑔𝑛𝑎𝑙
𝑆𝑁𝑅 = (26)
2
𝜎𝑛𝑜𝑖𝑠𝑒
2
where 𝜎𝑠𝑖𝑔𝑛𝑎𝑙 2
and 𝜎𝑛𝑜𝑖𝑠𝑒 are the variance of the signal and noise, respec-
tively. Random noise is generated through the Gaussian process with
𝑠𝑖𝑔𝑛𝑎𝑙
𝜎2
a variance of 𝑆𝑁𝑅 and added to the clean signals acquired from the Fig. 11. Bayesian optimization algorithm.
response history analysis to simulate the field condition. In this study,
a maximum level of 10% noise is considered as the worst case. Fig. 10
shows the effect of 10% on the original acceleration response of the optimization problem. The value of 𝐶𝑉𝑒𝑟𝑟𝑜𝑟 is calculated through a
structure. Gaussian process as follow:
∑
𝐶𝑉𝑒𝑟𝑟𝑜𝑟 ∼ 𝐺𝑃 (𝑀, +𝜎 2 𝐼) (28)
5.3. Feature extraction and labeling
where 𝑀 is the average value of the Gaussian process with an initial
∑
After simulating the field condition by adding natural noise to the value of 0. is the covariance matrix calculated from the squared
structural response, a finite number of features should be extracted exponential kernel function [72]. 𝐼 is an identity matrix with com-
from these raw acceleration histories to feed the classification learn- patible dimensionality. It is assumed that a Gaussian noise with a
ers. In this regard, all the acceleration records were de-noised using variance of 𝜎 2 is added to the observations. To evaluate and update the
the wavelet de-noising approach explained in Section 3.6. Then, all hyperparameters, BO maximizes the following acquisition function:
the damage features were extracted from the acceleration signals as 𝐸𝐼(𝑥) = 𝐸[𝑚𝑎𝑥(0, 𝜇(𝑧𝑏𝑒𝑠𝑡 ) − 𝐶𝑉𝑒𝑟𝑟𝑜𝑟 )] (29)
discussed in Section 3. The performance level of the shear wall is de-
termined based on the recommendation of ASCE 41-17 [71] (Table A.2) where 𝑧𝑏𝑒𝑠𝑡 is the location of the minimum posterior’s mean, and 𝜇(𝑧𝑏𝑒𝑠𝑡 )
to assign the proper label for each sample. According to ASCE 41-17, is the minimum value of the posterior’s average. BO may stop after
the performance level of a shear wall with specific design properties satisfying each of the following assumptions: (1) a threshold value for
and construction type is determined based on the amount of plastic the number of iterations, (2) a threshold for the learning time, and (3) a
rotation that occurred in the body of the wall. Readers are encouraged specific stopping criterion that applied to the algorithm. In the present
to review Table A.2 for detailed information on the way of determining research, the stopping criterion was a threshold value for the number
the performance metrics. After determining the extent of damage of the of iterations. This value was chosen based on the complexity of the
shear walls, the 𝐼𝑂, 𝐿𝑆, and 𝐶𝑃 performance levels were labeled as 0, under-study problem. Readers are referred to study [48,72] for detailed
1, and 2, respectively. information on employing the BO algorithm for hyperparameter selec-
tion in the classification learners. Fig. 11 illustrates the BO flowchart
for calculating the optimized hyperparameters.
5.4. Hyperparameter optimization
5.5. Training procedure
An important problem in training a classifier is selecting optimum
hyperparameters to achieve the minimum cost for predictive model. To A parametric study is implemented to determine the most reliable
prevent the over-fitting of the classification learner, the 𝐾-fold cross- classification learner for mapping the feature’s space onto the damage
validation approach is employed. This method randomly divides the scenarios of the shear wall buildings. BO algorithm is used to tune
dataset into 𝐾 equal parts in which (𝐾 − 1) partitions are utilized for the hyperparameters of each model. Besides, K-fold cross-validation
training process, and the rest is considered to investigate the validity technique is considered to prevent the over-fitting problem during the
of the training algorithm. This process is iterated until all partitions are training process. For this purpose, the training set is partitioned into 10
considered at least one time as the validation part. Finally, the average equal subset. The feature’s space is fed to the classification learners in
loss for these 𝑘 trained models is reported as the cross-validation error. the following form:
[ ]
𝑘 1 2 6 7
𝑂𝑖,𝑗−𝑥 = 𝑓𝑖,𝑗−𝑥 , 𝑓𝑖,𝑗−𝑥 , … , 𝑓𝑖,𝑗−𝑥 , 𝑓𝑖,𝑛−𝑥 𝑎𝑛𝑑 (30)
∑
𝐾
[ ]
𝐶𝑉𝑒𝑟𝑟𝑜𝑟 = 𝑒𝑘 (27) 𝑘 1 2 6 7
𝑖=1
𝑂𝑖,𝑗−𝑦 = 𝑓𝑖,𝑗−𝑦 , 𝑓𝑖,𝑗−𝑦 , … , 𝑓𝑖,𝑗−𝑦 , 𝑓𝑖,𝑛−𝑦 (31)
Where 𝐶𝑉𝑒𝑟𝑟𝑜𝑟 is the cross-validation error, and 𝑒𝑘 represent the value where 𝑓 represents the feature, 𝑖 is the observation number, 𝑗 =
of error obtained for each fold. Bayesian Optimization (BO) algorithm {1, 2, … , 𝑛} is the story number, and 𝑘 is the number of features.
conducts the classifier to achieve a model with minimum 𝐶𝑉𝑒𝑟𝑟𝑜𝑟 . To As mentioned before, the correlation feature is only determined for
this end, the 𝐶𝑉𝑒𝑟𝑟𝑜𝑟 is considered as an objective function for the 7
the roof level. Therefore, 𝑓𝑖,𝑛−𝑥 7
and 𝑓𝑖,𝑛−𝑦 are the correlation features
491
Fig. 12. Outline of the proposed framework.
Table 6
Comparison between the performance of different models.
Learning algorithm 7-story building 9-story building 13-story building Average
Decision tree 93% 83.5% 94.9% 90.5%
Linear discriminant 93.4% 85.6% 92.9% 90.6
Naive Bayes 87.4% 83.7% 90.6% 87.2%
SVM 94.6% 89.3% 95.4% 93.1%
KNN 95.7% 89.7% 95.6% 93.6%
obtained for the roof level of each building in 𝑥 and 𝑦 directions, in detecting the performance level of various stories of the buildings.
receptively. Notably, the dataset is randomly shuffled, and 80% of the According to this table, the KNN algorithm performs better than the
examples are employed for training process. Besides, the remaining bin other classification learners in terms of cross-validation accuracy. Of
is utilized to test the accuracy of the trained classifier. Fig. 12 shows course, the SVM algorithm demonstrates comparable accuracy to the
the general view of the proposed algorithm. KNN, but the optimization process of SVM shows more computational
A parametric study was performed to obtain a sufficient number of cost. On the other hand, the Naive Bayes algorithm, with an average
features that should be used for achieving a predictive model with max- cross-validation accuracy of 87.2%, was the worst case in learning the
imum accuracy and minimum computational cost. According to this damage detection problem. As a result, the KNN algorithm is proposed
investigation, it is concluded that 4 of the proposed features improve to identify the extent of damage in concrete shear wall buildings. In the
the accuracy of the classification learner. Therefore, in addition to following subsections, the results of the KNN algorithm in detecting
tuning the classifier hyperparameters, another problem that should be the performance level of each case study model are explained. The
solved during the BO algorithm is selecting four of the more significant confusion matrix calculated for the test set, the Receiver Operating
features among the seven attributes proposed in this study. Characteristic (ROC) curve, and the cross-validation accuracy of each
classifier are presented to evaluate the performance of the predictive
6. Results models in a perceptible manner. A confusion matrix represents the
robustness of the model in terms of the true predicted instances against
This section first describes the results of different classifiers in the false classified samples. In addition, a ROC curve presents the
detecting the structural damage for the mentioned case study buildings. True Positive Rate (TPR) versus the False Positive Rate (FPR) along
Then, by selecting the best classifier, the results of this model are the training process for different thresholds. Readers can refer to Fan
explained in detail to assess the efficiency of the presented approach et al. [73] to obtain more information on the ROC curve characteristics.
in estimating the damage level of concrete shear wall buildings under Whatever the ROC curve is closer to the upper-left corner of the plot,
earthquake motions. the prediction model presents more accuracy and generalization. The
area under the curve (AUC) is defined to measure the closeness of the
6.1. Performance of different classifiers ROC curve to the upper-left corner. It is obvious that an AUC closer to
1 indicates a more reliable prediction model with a higher estimation
As mentioned before, a parametric study is carried out to determine accuracy. This index is more applicable when the dataset is polluted
the best learner for solving the under-study classification problem. For to noise. On average, the number of optimized neighbors is selected
this purpose, the score of different models was considered as the perfor- as 11 for all stories of the structure. Moreover, 𝑀𝑎ℎ𝑎𝑙𝑎𝑛𝑜𝑏𝑖𝑠 distance
mance metric. Table 6 summarizes the average accuracy of each model and 𝑠𝑞𝑢𝑎𝑟𝑒𝑑 𝑖𝑛𝑣𝑒𝑟𝑠𝑒 rule were respectively selected as the most efficient
492
Fig. 13. The ROC curve and confusion matrices for damage extent detection of the 7-story building.
distance metric and weighting measure during the BO algorithm. It to the true predicted borders. Thus, the false predicted cases do not
should be mentioned that the 𝐷𝑟𝑖𝑓 𝑡, 𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛, and 𝑀𝐶𝐴𝑉 attributes significantly influence the decision-making process after the event.
were chosen as the optimized features among the seven pre-defined Moreover, expert engineers can make a reliable decision for the whole
damage indicators. Fig. 13 shows the confusion matrices for different structure by considering the performance levels achieved for each story
floors of the seven-story building. A close look to this figure reveals of the structure.
that the proposed method provides promising accuracy in estimating Fig. 15 shows the confusion matrices for shear walls in the 𝑌
the local damage exposed to the shear walls of the building. A mean direction of the building. The average accuracy of the classifiers was
accuracy of 95.6% is achieved for estimating the performance level of 89.4% for identifying the extent of damage in the 𝑌 direction. It is
each floor. The ROC curve that is presented in Fig. 13(d) belongs to obvious that the average accuracy of the prediction models is less than
the story level with minimum cross-validation accuracy (second story). the seven-story for both directions. This problem is due to the complex
The minimum AUC among different performance levels was 0.96, behavior of the nine-story building compared to the seven-story case
which represents the effectiveness of the selected features in estimating because the intense vertical irregularity exists in the architecture of
the damage state of the floors. The prediction model illustrates the the nine-story building. Fig. 15(d) illustrates the ROC curve of the
minimum true positive rate of 73.0% for the last floor. Notably, the prediction model for the second story of the building, which presents
false negative rate in this story can be assigned to the proximity of those the minimum prediction accuracy among all the stories of the structure.
portions of samples to the true predicted boundaries, and also the low A closer look at this figure reveals that all the AUC values are greater
number of samples exceed the LS level in the last floor of the building. than 0.94, which demonstrates the promising accuracy of the models
Hence, the false estimated cases do not remarkably influence decision- in identifying the severity of damage for each floors of the building.
making metrics for the confronting actions after the earthquake. In These figures indicate that the proposed method capable of estimating
addition, since the prediction models determine the performance level the damage state of the concrete shear walls of the building with an
of the building for each story, a reliable decision can be made for the average accuracy of 89.4%. This means that the prediction models
whole building. may represent an average false prediction rate of 11.6% for marginal
Fig. 14 illustrates the confusion matrices for shear walls in 𝑋 cases. However, these false prediction cases have similar damage in-
direction of the nine-story building. On average, the accuracy of the dicators to the proximate performance level. Therefore, these false
prediction models was equal to 91.5% for identifying the extent of classified scenarios generally do not impose significant errors on the
damage in 𝑋 direction. Fig. 14(d) shows the ROC curve of the clas- decision-making metrics.
sification learner for the story with minimum prediction accuracy Fig. 16 reveals the confusion matrices for shear walls in 𝑋 direction
(second story). According to this figure, The AUC values are close to of the thirteen-story building. An average accuracy of 91.4% was ob-
1.0, which indicates the generalization and robustness of the proposed tained for detecting the extent of damage in 𝑋-direction shear walls. As
classification process. The minimum true positive rate for different an example, Fig. 16(d) illustrates the ROC curve of the prediction model
performance levels through all the stories was 75.7% for the shear for the third story (story with minimum cross-validation accuracy).
walls in 𝑋 direction. Accordingly, the false negative rate in that story This figure indicates that the AUC values are close to 1.0 for all the
was 24.3%. As the features presented in this study are directly related performance levels. It can be concluded that the prediction models
to the structural damage of the building, this rate of false predicted presents promising performance in estimating the level of damage. The
cases can be specified to the vicinity of these portions of instances minimum true positive rate belongs to the CP state of the last story
493
Fig. 14. The ROC curve and confusion matrices for damage detection model of the 9-story building (Shear walls in 𝑋 direction).
Fig. 15. The ROC curve and confusion matrices for damage detection model of the 9-story building (Shear walls in 𝑌 direction).
which was equal to 68.8%. It is obvious from Fig. 16(c) that the pre- It should be mentioned that expert engineers can also make their
dictive model underestimates the extent of damage in 31.2% of cases. decision considering the damage state obtained for adjacent floors of
However, approximately in 29.5% of cases, the false predicted cases the building.
do not significantly affect the decision-making process because these Similarly, Fig. 17 summarizes the confusion matrices for the shear
false predicted samples have marginal damage indicators between two walls in 𝑌 -direction of the thirteen-story building. The overall view
adjacent classes. The lowest level of performance for the CP damage of this figure reveals the efficiency of the proposed framework in
state was related to the low number of samples that exceeded the LS estimating the structural damage subjected to the concrete shear wall
level in the last story of the building. This behavior was predictable buildings. Similar to the 𝑋-direction, an average accuracy of 91.4% was
taking into account the fundamental concepts of dynamics of structures. achieved for prediction models of the shear walls in the 𝑌 direction.
494
Fig. 16. The ROC curve and confusion matrices for damage detection model of the 13-story building (Shear walls in 𝑋 direction).
Fig. 17. The ROC curve and confusion matrices for damage detection model of the 13-story building (Shear walls in 𝑌 direction)
Fig. 18. Comparison between the 𝐶𝐴𝑉 obtained for two different ground motions.
495
Fig. 19. Comparison between the 𝑅𝐸 obtained for two different ground motions.
extent in each event, the damage extents obtained from two records
have interfered. The same behavior was observed for the energy ratio
feature (Fig. 19). To illustrate the total efficiency of these two features
in detecting the severity of damage at concrete shear wall buildings, a
𝐾𝑁𝑁 classifier was trained, and an example confusion matrix of this
process is shown in Fig. 20. The total accuracy of 69% achieved for the
prediction model indicates the insufficiency of these damage indicators
for identifying the local damage subjected to the concrete shear walls.
According to this study, it can be concluded that the extent of damage
has a direct relation to the lateral behavior of the building system, and
the damage indicators proposed for a specified building may not be
effective for another structural system.
Fig. 20. Predictive model based on 𝐶𝐴𝑉 and 𝑅𝐸 attributes.
7. Conclusion
According to Fig. 17(d), the ROC curves of the story with minimum This paper presented a rapid algorithm for identifying the severity
cross-validation accuracy are close to the upper-left corner of the of local damage in the concrete shear wall buildings. A total number
diagram and their AUC is close to 1. Therefore, the prediction models
of 1884 nonlinear response history analyses were conducted for each
perform acceptably even for the worst prediction model. Fig. 17(a)
building using the SAC motions. A suite of damage indicators was ex-
indicates that the prediction model of this story underestimates the LS
tracted from the acceleration signals to construct the prediction models.
state in 19.7% and overestimates them in 7.5% of cases. In addition,
A parametric study was carried out to determine the most efficient
this predictive model underestimates the CP state in 26.4% of cases. Of
learner for classifying the damage states of the buildings. The KNN
course, approximately 22.9% of these underestimated cases belong to
classifier was selected to construct the predictive models because of its
those samples that have similar damage indicators to the LS state. In
fact, these scenarios do not impose much error on the decision-making maximum accuracy compared to the other algorithms. The Bayesian
metrics as they are in the margins of the acceptance criterion defined optimization algorithm implemented to tune the hyperparameters of
in 𝐴𝑆𝐶𝐸 − 41. the classification learners. The main conclusions derived from the study
are as below:
6.2. Discussion
• The proposed features were efficient in detecting the probable
extent of local damage subjected to the concrete shear walls.
This research investigated the efficiency of two commonly used
• The parametric study showed that the 𝐾𝑁𝑁 classifier is the most
damage indicators proposed for identifying the extent of damage in
reliable algorithm for identifying the structural damage at the
concrete shear walls buildings. Many studies showed the efficiency
concrete shear wall buildings.
of energy-based damage indicators, such as output energy and the
• Bayesian optimization algorithm prevented the possible errors
relative energy between the output and input signals of the structure,
for detecting the structural damage in buildings [29,31,32]. However, related to the improper architecture of the classification learner.
based on an investigation conducted in the present study, it is shown • According to the outcome of the Bayesian optimization, it was
that these features are not capable of classifying the severity of local concluded that the 𝐷𝑟𝑖𝑓 𝑡, 𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛, and 𝑀𝐶𝐴𝑉 are the most
damage subjected to the concrete shear wall frames under ground mo- powerful indicators in classifying the damage states of the con-
tions. Fig. 18 illustrates the CAV diagrams for two records (𝐶𝑜𝑦𝑜𝑡𝑒𝐿𝑎𝑘𝑒 crete shear wall buildings.
and 𝑀𝑜𝑟𝑔𝑎𝑛𝐻𝑖𝑙𝑙) with different durations, where both motions scaled • Results showed that vertical irregularity could decrease the ac-
similarly. This figure indicates that the maximum value of CAV is curacy of the prediction models because of the more complex
dependent on the ground motion duration. In this figure, the CAV dynamic behavior of such buildings.
response space was partitioned into three windows of 𝑊 1, 𝑊 2, and 𝑊 3 • The task of damage identification was successfully implemented
based on the performance level of the building under the 𝐶𝑜𝑦𝑜𝑡𝑒𝐿𝑎𝑘𝑒 by achieving an average accuracy of 95.6%, 90.4%, and 91.4%
earthquake. Despite the good agreement between the 𝐶𝐴𝑉 and damage for 7-, 9-, and 13-story buildings, respectively.
496
Table A.1
SAC ground motions.
ID Event name Duration (s) PGA (g) ID Event name Duration (s) PGA (g)
1 CoyoteLake 26.86 0.59 56 LongBeach 39.1 0.174
2 ImperialValley 39.1 0.143 57 MorganHill 60 0.663
3 Kernmodified 78.62 0.159 58 WestWashington 80 0.384
4 Landersm 80 0.338 59 WestWashington 66.7 0.389
5 MorganHill 60 0.546 60 NorthPalm 60 0.588
6 Parkfield 43.94 0.781 61 PugetSound 81.84 0.752
7 Parkfield 26.16 0.79 62 PugetSound 74.1 0.369
8 NorthPalm 60 0.518 63 EasternWa 60 0.575
9 SanFernando 79.48 0.253 64 Llolleom 100 0.698
10 Whittier 40 0.769 65 Vinadel Mar 100 0.542
11 ImperialValley 53.48 0.676 66 Mendocino 60 0.756
12 ImperialValley 39.4 0.488 67 Erzincan 20.78 0.605
13 ImperialValley 39.1 0.302 68 Olympiamod 80 0.896
14 Landersmod 80 0.426 69 Seattlemod 81.84 1.756
15 Landersmod 80 0.52 70 Valpariso 100 1.637
16 LomaPrieta 40 0.97 71 Valpariso 100 1.271
17 Northridge 60 0.678 72 DeepInterplate 80 0.797
18 Northridge 14.96 0.58 73 Miyagioki 80 0.784
19 Northridge 60 0.817 74 ShallowInter 80 0.563
20 NorthPalm 60 1.019 75 ShallowInter 80 0.75
21 Kobemodi 60 1.283 76 Tabasmod 50 0.978
22 lomaPrieta 25 0.473 77 LomaPrieta 25 0.718
23 Northridge 14.96 0.944 78 LomaPrieta 40 0.686
24 Northridge 60 1.33 79 CMendocino 60 0.655
25 Tabasmod 50 0.992 80 Erzincan 20.78 0.457
26 ElysianPark 30 1.296 81 Landersmod 49.3 0.799
27 ElysianPark 30 0.782 82 Nothridge 14.98 0.89
28 ElysianPark 30 1.101 83 Nothridge 60.02 0.732
29 PalosVerdes 60 0.776 84 Kobemodi 60 1.088
30 PalosVerdes 60 0.625 85 Kobemodi 40.1 0.786
31 hangingwall 30 0.124 86 Elysianmod 40 0.892
32 footwall 30 0.144 87 Elysianmod 40 1.803
33 NewHampshire 19.24 0.576 88 Elysianmod 40 1.013
34 Nahannimod 20.34 0.088 89 Elysianmod 40 0.922
35 Nahannimod 18.76 0.074 90 Elysianmod 40 1.162
36 Nahannimod 19.02 0.136 91 Palosmod 80 0.974
37 Saguenay 17.74 0.29 94 Palosmod 80 0.968
38 Saguenay 29.58 0.524 95 PalosVerdes 80 0.874
41 footwall 30 0.316 98 Northridge,CA 30 0.516
42 footwall 30 0.335 99 Northridge,CA 20 0.482
43 footwall 30 0.309 100 Duzce,Turkey 55.9 0.822
44 Nahannimod 20.34 0.252 101 HectorMine,CA 45.32 0.337
45 Nahannimod 18.76 0.211 102 ImperialValley 99.92 0.351
46 Nahannimod 19.02 0.389 103 ImperialValley 39.04 0.38
47 Saguenay 17.74 0.783 104 Kobe,Japan 40.96 0.509
48 Saguenay 29.58 1.504 105 Kobe,Japan 40.96 0.243
49 Saguenay 39.06 0.651 106 Kocaeli,Turkey 27.2 0.358
50 Saguenay 33.26 0.781 107 Kocaeli,Turkey 30 0.219
51 Landers,CA 27.98 0.417 108 Chi-Chi,Taiwan 61.78 0.44
52 LomaPrieta,CA 39.96 0.541 109 Chi-Chi,Taiwan 53.66 0.512
53 LomaPrieta,CA 39.96 0.555 110 SanFernando 28 0.21
54 Manjil,Iran 46 0.538 111 Friuli,Italy 36.36 0.351
55 SuperHills,CA 40 0.358
Table A.2
Modeling parameters and numerical acceptance criteria for nonlinear procedures—reinforced concrete structural walls.
Conditions Acceptable plastic hinge rotation (Radians)
Performance level
IO LS CP
(𝐴𝑠 − 𝐴′𝑠 )𝑓𝑦𝐸 + 𝑃 𝑉
′ √ Confined Boundary
𝑡𝑤 𝐼𝑤 𝑓𝑐𝐸 ′
𝑡𝑤 𝐼𝑤 𝑓𝑐𝐸
≤ 0.1 ≤4 Yes 0.005 0.015 0.020
≤ 0.1 ≥6 Yes 0.004 0.010 0.015
≥ 0.25 ≤4 Yes 0.003 0.009 0.012
≥ 0.25 ≥6 Yes 0.0015 0.005 0.010
≤ 0.1 ≤4 No 0.002 0.008 0.015
≤ 0.1 ≥6 No 0.002 0.006 0.010
≥ 0.25 ≤4 No 0.001 0.003 0.005
≥ 0.25 ≥6 No 0.001 0.002 0.004
497
• This research illustrates that the efficiency of damage indicators [18] Neves AC, Leander J, González I, Karoumi R. An approach to decision-making
depends on the lateral behavior of the building. For example, analysis for implementation of structural health monitoring in bridges. Struct
Control Health Monit 2019;26(6):e2352. http://dx.doi.org/10.1002/stc.2352.
several studies show that damage indexes like output energy of
[19] Salkhordeh M, Alishahiha F, Mirtaheri M, Soroushian S. A rapid neural
the structure and relative energy between the output and input network-based demand estimation for generic buildings considering the effect
signals are efficient in determining the damage level in structures. of soft/weak story. Struct Infrastruct Eng 2022;1–20.
However, this study reveals that these features are not suitable [20] Wang L, Liu Y, Li M. Time-dependent reliability-based optimization for structural-
for achieving a reliable local-damage detection framework in topological configuration design under convex-bounded uncertain modeling.
Reliab Eng Syst Saf 2022;221:108361.
reinforced concrete shear wall buildings. [21] Wang L, Liu Y, Liu D, Wu Z. A novel dynamic reliability-based topology
• The proposed method did not consider the uncertainties related to optimization (DRBTO) framework for continuum structures via interval-process
the ground motion directionality, material properties, etc. These collocation and the first-passage theories. Comput Methods Appl Mech Engrg
uncertainties can be applied to the framework to analyze the 2021;386:114107.
[22] Salkhordeh M, Mirtaheri M. A swift neural network-based algorithm for demand
reliability of the proposed technique [21,74].
estimation in concrete moment-resisting buildings. Numer Methods Civ Eng
2021;6(3):28–36.
Declaration of competing interest [23] Liu Y, Wang L, Gu K, Li M. Artificial neural network (ANN)-Bayesian prob-
ability framework (BPF) based method of dynamic force reconstruction under
multi-source uncertainties. Knowl-Based Syst 2022;237:107796.
The authors declare that they have no known competing finan-
[24] Ghiasi R, Torkzadeh P, Noori M. A machine-learning approach for structural
cial interests or personal relationships that could have appeared to damage detection using least square support vector machine based on a new
influence the work reported in this paper. combinational kernel function. Struct Health Monit 2016;15(3):302–16. http:
//dx.doi.org/10.1177/1475921716639587.
[25] Ni Y, Zhou X, Ko J. Experimental investigation of seismic damage identification
Appendix
using PCA-compressed frequency response functions and neural networks. J
Sound Vib 2006;290(1–2):242–63. http://dx.doi.org/10.1016/j.jsv.2005.03.016.
See Tables A.1 and A.2 [26] De Ridder D, Duin RP. Locally linear embedding for classification. Tech. rep.
PH-2002-01, Pattern Recognition Group, Dept. of Imaging Science & Technology,
Delft University of Technology, Delft, the Netherlands; 2002, p. 1–12.
References
[27] Hyvärinen A, Oja E. Independent component analysis: Algorithms and appli-
cations. Neural Netw 2000;13(4–5):411–30. http://dx.doi.org/10.1016/S0893-
[1] Azimi M, Pekcan G. Structural health monitoring using extremely compressed 6080(00)00026-5.
data through deep learning. Comput-Aided Civ Infrastruct Eng 2019. http://dx. [28] Reed JW, Kassawara RP. A criterion for determining exceedance of the operating
doi.org/10.1111/mice.12517. basis earthquake. Nucl Eng Des 1990;123(2–3):387–96. http://dx.doi.org/10.
[2] Cha Y-J, Choi W, Büyüköztürk O. Deep learning-based crack damage detec- 1016/0029-5493(90)90259-Z.
tion using convolutional neural networks. Comput-Aided Civ Infrastruct Eng [29] Cabanas L, Benito B, Herráiz M. An approach to the measurement of the
2017;32(5):361–78. http://dx.doi.org/10.1111/mice.12263. potential structural damage of earthquake ground motions. Earthq Eng Struct
[3] Rehman SKU, Ibrahim Z, Memon SA, Jameel M. Nondestructive test methods for Dyn 1997;26(1):79–92. http://dx.doi.org/10.1002/(SICI)1096-9845(199701)26:
concrete bridges: A review. Constr Build Mater 2016;107:58–86. http://dx.doi. 1<79::AID-EQE624>3.0.CO;2-Y.
org/10.1016/j.conbuildmat.2015.12.011. [30] Bandara RP, Chan TH, Thambiratnam DP. Structural damage detection method
[4] Mirtaheri M, Salkhordeh M, Mohammadgholiha M. A system identification- using frequency response functions. Struct Health Monit 2014;13(4):418–29.
based damage-detection method for gravity dams. Shock Vib 2021;2021. http: http://dx.doi.org/10.1177/1475921714522847.
//dx.doi.org/10.1155/2021/6653254. [31] Muin S, Mosalam KM. Cumulative absolute velocity as a local damage indicator
[5] Chang PC, Flatau A, Liu S. Health monitoring of civil infrastructure. Struct Health of instrumented structures. Earthq Spectra 2017;33(2):641–64. http://dx.doi.org/
Monit 2003;2(3):257–67. http://dx.doi.org/10.1177/1475921703036169. 10.1193/090416EQS142M.
[6] Karami-Mohammadi R, Mirtaheri M, Salkhordeh M, Hariri-Ardebili MA. Vibration [32] Sajedi SO, Liang X. A data-driven framework for near real-time and robust dam-
anatomy and damage detection in power transmission towers with limited age diagnosis of building structures. Struct Control Health Monit 2019;e2488.
sensors. Sensors 2020;20(6):1731. http://dx.doi.org/10.1002/stc.2488.
[7] Arici Y, Mosalam KM. Modal identification of bridge systems using state-space [33] American Society of Civil Engineers. Seismic evaluation and retrofit of existing
methods. Struct Control Health Monitor: Official J Int Assoc Struct Control buildings. American Society of Civil Engineers; 2017.
Monitor Eur Assoc Control Struct 2005;12(3–4):381–404. http://dx.doi.org/10. [34] Avci O, Abdeljaber O, Kiranyaz S, Hussein M, Gabbouj M, Inman DJ. A review
1002/stc.76. of vibration-based damage detection in civil structures: From traditional methods
[8] Jafarkhani R, Masri SF. Finite element model updating using evolutionary strat- to machine learning and deep learning applications. Mech Syst Signal Process
egy for damage detection. Comput-Aided Civ Infrastruct Eng 2011;26(3):207–24. 2021;147:107077. http://dx.doi.org/10.1016/j.ymssp.2020.107077.
http://dx.doi.org/10.1111/j.1467-8667.2010.00687.x. [35] Lin Y-z, Nie Z-h, Ma H-w. Structural damage detection with automatic
[9] Farrar CR, Worden K. Structural health monitoring: A machine learning feature-extraction through deep learning. Comput-Aided Civ Infrastruct Eng
perspective. John Wiley & Sons; 2012. 2017;32(12):1025–46. http://dx.doi.org/10.1111/mice.12313.
[10] Ying Y, Garrett Jr JH, Oppenheim IJ, Soibelman L, Harley JB, Shi J, [36] Zhao B, Taucer F, Rossetto T. Field investigation on the performance of building
et al. Toward data-driven structural health monitoring: Application of ma- structures during the 12 May 2008 wenchuan earthquake in China. Eng Struct
chine learning and signal processing to damage detection. J Comput Civ Eng 2009;31(8):1707–23. http://dx.doi.org/10.1016/j.engstruct.2009.02.039.
2013;27(6):667–80. http://dx.doi.org/10.1061/(ASCE)CP.1943-5487.0000258. [37] Watanabe E, Sugiura K, Nagata K, Kitane Y. Performances and damages
[11] Tsou P, Shen M-HH. Structural damage detection and identification using neural to steel structures during the 1995 Hyogoken-Nanbu earthquake. Eng Struct
networks. AIAA J 1994;32(1):176–83. http://dx.doi.org/10.2514/3.11964. 1998;20(4–6):282–90. http://dx.doi.org/10.1016/S0141-0296(97)00029-1.
[12] Worden K, Manson G, Fieller NR. Damage detection using outlier analysis. J [38] Massone L, Orakcal K, Wallace J. Shear-flexure interaction for structural walls.
Sound Vib 2000;229(3):647–67. http://dx.doi.org/10.1006/jsvi.1999.2514. Spec Publ 2006;236:127–50.
[13] Zhao J, Ivan JN, DeWolf JT. Structural damage detection using artificial neural [39] Orakcal K, Wallace JW. Flexural modeling of reinforced concrete walls-
networks. J Infrastruct Syst 1998;4(3):93–101. http://dx.doi.org/10.1061/(ASCE) experimental verification. ACI Mater J 2006;103(2):196.
1076-0342(1998)4:3(93). [40] Sherstobitoff J, Cajiao P, Adebar P. Repair of an 18-story shear wall building
[14] Yam L, Yan Y, Jiang J. Vibration-based damage detection for composite struc- damaged in the 2010 Chile earthquake. Earthq Spectra 2012;28(1_suppl1):335–
tures using wavelet transform and neural network identification. Compos Struct 48. http://dx.doi.org/10.1193/1.4000028.
2003;60(4):403–12. http://dx.doi.org/10.1016/S0263-8223(03)00023-0. [41] Somerville PG. Development of ground motion time histories for phase 2 of the
[15] Zhang Y, Burton HV, Sun H, Shokrabadi M. A machine learning framework FEMA/SAC steel project. SAC Joint Venture; 1997.
for assessing post-earthquake structural safety. Struct Saf 2018;72:1–16. http: [42] Yassin MHM. Nonlinear analysis of prestressed concrete structures under
//dx.doi.org/10.1016/j.strusafe.2017.12.001. monotonic and cyclic loads [Ph.D. thesis], University of California, Berkeley;
[16] Sajedi SO, Liang X. Vibration-based semantic damage segmentation for 1994.
large-scale structural health monitoring. Comput-Aided Civ Infrastruct Eng [43] Committee A. Building code requirements for structural concrete (ACI 318-14)
2020;35(6):579–96. http://dx.doi.org/10.1111/mice.12523. and commentary. American Concrete Institute; 2014.
[17] Salkhordeh M, Mirtaheri M, Soroushian S. A decision-tree-based algorithm for [44] Saatcioglu M, Razvi SR. Strength and ductility of confined concrete. J Struct Eng
identifying the extent of structural damage in braced-frame buildings. Struct 1992;118(6):1590–607. http://dx.doi.org/10.1061/(ASCE)0733-9445(1992)118:
Control Health Monit 2021;28(11):e2825. http://dx.doi.org/10.1002/stc.2825. 6(1590).
498
[45] Cando M, Hube M, Parra P, Arteta C. Effect of stiffness on the seismic [61] Fix E, Hodges JL. Discriminatory analysis. Nonparametric discrimination: Con-
performance of code-conforming reinforced concrete shear wall buildings. Eng sistency properties. Int Stat Rev/Rev Int Stat 1989;57(3):238–47. http://dx.doi.
Struct 2020;219:110724. http://dx.doi.org/10.1016/j.engstruct.2020.110724. org/10.2307/1403797.
[46] Xie L, Lu X, Lu X, Huang Y, Ye L. Multi-layer shell element for shear walls [62] Altman NS. An introduction to kernel and nearest-neighbor nonparametric re-
in OpenSees. In: Computing in civil and building engineering (2014). 2014, p. gression. Amer Statist 1992;46(3):175–85. http://dx.doi.org/10.1080/00031305.
1190–7. 1992.10475879.
[47] Lu X, Xie L, Guan H, Huang Y, Lu X. A shear wall element for nonlinear [63] Al-Hegami AS. Classical and incremental classification in data mining process.
seismic analysis of super-tall buildings using OpenSees. Finite Elem Anal Des Int J Comput Sci Netw Secur 2007;7(12):179–87.
2015;98:14–25. http://dx.doi.org/10.1016/j.finel.2015.01.006. [64] Quinlan JR. Induction of decision trees. Mach Learn 1986;1(1):81–106. http:
[48] Snoek J, Larochelle H, Adams RP. Practical bayesian optimization of machine //dx.doi.org/10.1023/A%3A1022643204877.
learning algorithms. In: Advances in neural information processing systems. [65] Quinlan JR. C4. 5: Programs for machine learning. Elsevier; 2014.
2012, p. 2951–9. [66] Esposito F, Malerba D, Semeraro G, Kay J. A comparative analysis of methods for
[49] Campbell K, Bozorgnia Y. Use of cumulative absolute velocity (CAV) in damage pruning decision trees. IEEE Trans Pattern Anal Mach Intell 1997;19(5):476–91.
assessment. In: Proc. of 15th world conf. in earthquake eng. 2012, p. 1–10. [67] Feng X, Li S, Yuan C, Zeng P, Sun Y. Prediction of slope stability using naive
[50] ATC58. Seismic performance assessment of buildings, Vol. 1 (Methodology). Bayes classifier. KSCE J Civ Eng 2018;22(3):941–50. http://dx.doi.org/10.1007/
California: Redwood City; 2012a. s12205-018-1337-3.
[51] Donoho DL. Progress in wavelet analysis and WVD: A ten minute tour. Progress [68] Prémaud A, Le Meur Y, Debord J, Szelag J-C, Rousseau A, Hoizey G, et al. Max-
Wavelet Anal Appl 1993;109–28. imum a posteriori Bayesian estimation of mycophenolic acid pharmacokinetics
[52] Donoho DL, Johnstone JM. Ideal spatial adaptation by wavelet shrinkage. in renal transplant recipients at different postgrafting periods. Ther Drug Monit
Biometrika 1994;81(3):425–55. 2005;27(3):354–61.
[53] Donoho DL. De-noising by soft-thresholding. IEEE Trans Inform Theory [69] Mbo’o CP, Hameyer K. Fault diagnosis of bearing damage by means of the linear
1995;41(3):613–27. discriminant analysis of stator current features from the frequency selection. IEEE
[54] Donoho DL, Johnstone IM, Kerkyacharian G, Picard D. Wavelet shrinkage: Trans Ind Appl 2016;52(5):3861–8.
Asymptopia? J R Stat Soc Ser B Stat Methodol 1995;57(2):301–37. http://dx. [70] Ali L, Zhu C, Zhang Z, Liu Y. Automated detection of Parkinson’s disease based
doi.org/10.1111/j.2517-6161.1995.tb02032.x. on multiple types of sustained phonations using linear discriminant analysis
[55] Daubechies I. Ten lectures on wavelets. SIAM; 1992. and genetically optimized neural network. IEEE J Transl Eng Health Med
[56] Ansari A, Noorzad A, Zare M. Application of wavelet multi-resolution analysis 2019;7:1–10.
for correction of seismic acceleration records. J Geophys Eng 2007;4(4):362–77. [71] ASCE. Seismic evaluation and retrofit of existing buildings (ASCE/SEI 41-17).
http://dx.doi.org/10.1088/1742-2132/4/4/002. American Society of Civil Engineers; 2017.
[57] Soroushian S, Maragakis EM, Ansari A. Estimation of vertical floor displacement [72] Rasmussen CE. Gaussian processes in machine learning. In: Summer school on
using a wavelet de-noising method. J Earthq Eng 2016;20(2):279–97. http: machine Learning. Springer; 2003, p. 63–71.
//dx.doi.org/10.1080/13632469.2015.1085466. [73] Fan J, Upadhye S, Worster A. Understanding receiver operating characteristic
[58] Brownjohn JM. Ambient vibration studies for system identification of tall (ROC) curves. Canad J Emerg Med 2006;8(1):19–20. http://dx.doi.org/10.1017/
buildings. Earthq Eng Struct Dyn 2003;32(1):71–95. http://dx.doi.org/10.1002/ s1481803500013336.
eqe.215. [74] Mendler A, Döhler M, Ventura CE. A reliability-based approach to determine the
[59] Newland DE. An introduction to random vibrations, spectral & wavelet analysis. minimum detectable damage for statistical damage detection. Mech Syst Signal
Courier Corporation; 2012. Process 2021;154:107561.
[60] Suthaharan S. Support vector machine. In: Machine learning models and
algorithms for big data classification. Springer; 2016, p. 207–35.
499

Structures: Hossein Mahmoudi, Maryam Bitaraf, Mojtaba Salkhordeh, Siavash Soroushian

Uploaded by

Copyright:

Available Formats

Structures: Hossein Mahmoudi, Maryam Bitaraf, Mojtaba Salkhordeh, Siavash Soroushian

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Structures: Hossein Mahmoudi, Maryam Bitaraf, Mojtaba Salkhordeh, Siavash Soroushian

Uploaded by

Copyright:

Available Formats

Structures 47 (2023) 482–499

Contents lists available at ScienceDirect

A rapid machine learning-based damage detection algorithm for identifying

ARTICLE INFO ABSTRACT

originally developed in the SAC project [41]. To increase the rate

This study aims to predict the structural damage in concrete shear-

2.1. Seven-story building

A 3D seven-story building is selected to assess the performance of

2.2. Nine-story building

Fig. 3. General view of the 9-story building.

2.4. Numerical modeling methodology

All the steel rebars are modeled using the Giuffre-Menegotto-Menegotto-

Bayesian optimization [48]. Doing so leads to increasing the learning

3.1. Arias intensity (𝐼𝐴 )

Fig. 7. Efficiency of the ‘‘wavelet de-noising’’ in correcting the drift response.

3.7. Correlation coefficient

Fig. 8. The effect of nonlinearity on the correlation coefficient damage indicator.

of occurrence (𝑃𝑖 ), 𝐵𝑖 is defined as follow:

4.2. K-Nearest Neighbors (KNN)

KNN is a non-parametric algorithm developed for the classification

• Compute the distance between 𝑥𝑗 and each 𝑥𝑖 in the dataset

The Weighted-KNN (W-KNN) is an extension of KNN, which assigns a

solving the eigen value problem yields:

representing a wide range of frequency contents. Therefore, the set of

5.2. Effect of noise

It is well understood that the sensory data are accompanied by

Fig. 12. Outline of the proposed framework.

You might also like