Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Measurement 205 (2022) 112195

Contents lists available at ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Machine learning-aided scenario-based seismic drift measurement for RC


moment frames using visual features of surface damage
Mohammadjavad Hamidia a, *, Sina Mansourdehghan b, Amir Hossein Asjodi b, Kiarash
M. Dolatshahi b, c, *
a
Faculty of Civil, Water, and Environmental Engineering, Shahid Beheshti University, Tehran, Iran
b
Department of Civil Engineering, Sharif University of Technology, Tehran, Iran
c
Department of Civil and Environmental Engineering, University of California, Los Angeles, USA

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

Keywords: This paper presents a novel computer vision-based methodology for assessment of the seismic damage in rein­
Surface crack pattern forced concrete moment frames using visual characteristics of surface damage following an earthquake. An
Post-earthquake damage assessment extensive collected database comprising 974 images associated with 256 cyclic-loaded damaged beam-column
Seismic loss measurement
joints, providing a set of cracking and crushing progression with increasing the evolution of damage level, is
Structural health monitoring
collected and used for the development and validation of the methodology. Employing image processing tech­
Image processing
Reinforced concrete moment frame niques, the characteristics of the surface damage, including the cracking length and crushing areas, are measured
and used in a scenario-based assessment for the seismic peak drift prediction. Based on the availability of the
structural information, four scenarios are proposed using input parameters among cracking length, crushing
areas, concrete compressive force, and the aspect ratio of the joint. The machine learning regression method is
employed for developing nonlinear regression models for each scenario. The proposed models measure the
seismic peak drift ratio during the earthquake excitation based on the visual damage features at the external
surface of the components. Finally, the seismic peak drift ratio obtained by the proposed methodology of this
paper can be used as an input engineering demand parameter in the existing seismic loss measurement frame­
works. An example specimen at various drift ratios is also presented as a case study to evaluate the predicted
versus actual experimental drift ratio.

1. Introduction structural components. However, the final decision is highly affected by


the experience and judgment of the engineer, which may vary in
After an extreme event such as an earthquake, visual inspection is a different situations. Furthermore, visual assessment is a time-intensive
crucial task aiming to evaluate the earthquake-induced structural process also prone to observational errors.
damage to the building components. Risk assessment, loss measurement, In recent years, computer vision-based methods have been widely
and damage state (DS) identification are some of the prominent out­ used as a robust alternative for visual inspection to perform post-disaster
comes of visual assessment [1–3]. For instance, HAZUS [1] and FEMA P- loss measurements at the structure and community level. In particular,
58 [3], the guidelines for seismic performance assessment of buildings, image processing and machine learning are among the state-of-the-art
employ visual inspection to determine the structural damage states damage assessment methods attracting many researchers. For
based on the qualitative description of cracking and crushing, such as instance, one of the image-based features of surface cracking, named the
the length and width of crack patterns and the spatial distribution of fractal dimension, has been recently used by Hamidia and Gangizadeh
crushed areas. Another guideline for evaluating earthquake-damaged for loss measurement [5] and residual stiffness estimation [6] of seis­
concrete and masonry wall buildings, FEMA 307 [4], determines the mically damaged nonductile reinforced concrete (RC) moment frames.
post-earthquake damage states according to the cracking length and In another study, Madani and Dolatshahi [7] collected 236 images
crushing areas. Therefore, using the qualitative description of surface corresponding to 72 damaged RC shear walls and used the symbolic
damage, a trained inspector is able to decide on the damage states of the regression method to quantify the stiffness and strength deterioration of

* Corresponding authors.
E-mail addresses: m_hamidia@sbu.ac.ir (M. Hamidia), dolatshahi@sharif.edu (K.M. Dolatshahi).

https://doi.org/10.1016/j.measurement.2022.112195
Received 14 August 2022; Received in revised form 17 October 2022; Accepted 8 November 2022
Available online 13 November 2022
0263-2241/© 2022 Elsevier Ltd. All rights reserved.
M. Hamidia et al. Measurement 205 (2022) 112195

Fig. 1. Seismic loss assessment framework flowchart.

damaged RC shear walls based on the fractal dimension of surface crack structural integrity of the earthquake-damaged components. For
patterns. The generalized fractal dimension of surface cracking was also instance, Hamidia et al. [32] employed machine learning regression
used by Hamidia et al. [8] to predict the maximum drift ratio of RC models for the loss estimation of nonductile beam-column subassemblies
moment frames. The predictive equations are presented in four scenarios based on information on cracking length and crushing density. Another
based on the availability of information, such as the shape of crack research by Paal et al. [33] used the visual features of surface cracking
patterns, mechanical properties of beam-column subassemblies, and and spalling, as well as structural information, to evaluate the residual
external loading conditions. drift ratio and damage state of the RC column. Lattanzi et al. [34] also
Although available literature [9–12] has shown that the fractal predicted the maximum drift ratio of RC columns based on the param­
dimension is a substantial damage indicator, interpreting the fractal eterized image information of 360 images associated with four RC
dimension with engineering judgment is still challenging. Also, bridge columns. In a more recent study, Mansourdehghan et al. [13]
employing fractal theory for a large number of crack patterns in an used six supervised learning techniques to classify the performance level
immense earthquake-affected region is time- and labor-consuming. of damaged RC shear walls based on the cracking length and crushing
Instead, recent development in image-based filtering has introduced areas. The research results demonstrated that the nonlinear tree-based
the cracking length and crushing area as the alternative for fractal models, such as Decision Tree, Random Forest, and Boosting methods,
dimension. For instance, Mansourdehghan et al. [13] used the normal­ outperform others in identifying post-earthquake performance levels.
ized cracking length and crushing areas as predictor variables to esti­ The performance-based probabilistic seismic loss measurement
mate the drift ratio, performance level, and strength deterioration of studies [1–3] have been commonly conducted on hazard analysis,
damaged RC shear walls. Moreover, Asjodi et al. [14] investigated the response estimation, and earthquake consequences (economic losses,
2D spatial distribution of cracking length and crushing areas to provide a casualties, and business interruption). At a glance, such frameworks use
visual benchmark for estimating the drift ratio of damaged RC shear a structural response as an Engineering Demand Parameter (EDPs) to a
walls. Another scenario-based research by Asjodi and Dolatshahi [15] seismic event with a given intensity measure. Then, the probability of
correlated the peak drift ratio of unreinforced masonry walls to the meeting or exceeding various damage states is quantified through
cracking length and crushing entropy. The study also distinguished fragility analysis for structural and non-structural components. The total
various types of cracking, including the cracks appearing within and at loss can be estimated at the final stage, considering the likelihood of
the interface of masonry blocks. Eventually, the predictive equations for damage and repair costs.
drift ratio based on the length of joint and block cracks, crushing en­ This study proposes a computer-vision-based framework capable of
tropy, aspect ratio, and prism strength are provided. estimating the total loss of RC buildings in the aftermath of an earth­
Machine learning is another branch of the computer vision-based quake. In particular, the visual damage features of cracking and crushing
method capable of recognizing and quantifying structural damage. are used in this paper to estimate the drift ratio of damaged beam-
Several studies have been conducted on detecting structural cracking at column joints (BCJ). The drift ratio is then used as an EDP to be incor­
the exposed face of damaged components. For instance, Convolutional porated into existing fragility curves. The critical difference dis­
Neural Networks (CNNs) are potent machine learning models used to tinguishing this study from previous literature is using tangible damage
classify or detect [16–18] specific objects in an input image. The ar­ features for damage assessment of structural components. Therefore,
chitecture of CNNs relies on a fully connected network automatically proposing robust damage features for visual inspection of damaged
extracting the features through a convolutional-kernel-based operation. beam-column subassemblies is the knowledge gap investigated in this
CNN has also been employed in some studies to predict the structural paper.
response [19,20], although more researchers used CNNs to detect and
measure the crack patterns [21,22]. It should be noted that the efficacy 2. Scope and objectives
of CNNs strongly depends on the network architecture and training data
size. Instead, employing image processing method is another alternative Considering the above-mentioned literature, both the visual features
to facilitating the quantification of crack patterns [23–26]. For instance, of the damage and fractal quantities can be used for quantifying the
Asjodi et al. [27] designed a crack detection and measurement algo­ damage. However, using the less complicated visual features, such as
rithm, the Arc length method, capable of quantifying the prominent cracking length and crushing density, is more applicable since it can be
characteristics of crack patterns, such as crack width, crack length, and easily measured through in-site measurements or image processing
pattern inclination. The Arc length method uses Support Vector Ma­ techniques. Moreover, crack width, which is conventionally considered
chines (SVMs) to detect the crack patterns and employs a filter-based the prominent indicator of damage to RC elements, is not a reliable
algorithm to measure the crack parameters. Further studies are avail­ parameter because several studies have shown that the residual crack
able in the literature [28–31] to detect surface damage, including crack width after a seismic event could significantly differ from the peak crack
pattern, crush area and spalling, at the exposed face of the concrete width [7,9]. Furthermore, there is a knowledge gap in seismic loss
components. measurement using crack patterns of RCMFs. Therefore, this paper aims
Machine learning models have also been carried out to monitor the to find robust visual features of the occurred damage in RCMFs to

2
M. Hamidia et al. Measurement 205 (2022) 112195

Fig. 2. Schematic flowchart for proposed damage state identification methodology.

AvJ s
f'
P Ag fc' fcc'

fc'

FyB
FCyB
FyB
P A f'
P vJAggs fcc'
A

PBAg fc'
B
B

AvJ s
ACvJ s
'
B
'
A B s
'vJ Fyv
FyB
B

Fyv
AvB s
ACCvB s
AvB s
B
fc'
C

F
AyB
vC s
A'yC
F vC s
AvC
B s

F
P yC
Ag fc'
B

FyB
A
FvB s Fyv
yB

FyB
'
B
ABvJ s
ABvC s

ABvB s
'
B
F'yC
B

Fig. 3. Relative frequency histogram of the beam-column subassemblies properties.

evaluate the seismic loss measurement based on the inference from an stages to quantify the total loss for a given damage level. Fig. 1 shows the
extensively collected databank. flowchart of the seismic loss measurement frameworks. Based on this
At a glance, loss estimation frameworks [1,3] include four general figure, the first step of loss measurement is hazard analysis, in which an

3
M. Hamidia et al. Measurement 205 (2022) 112195

the guidelines using the qualitative and quantitative description of


surface damage, such as the width of cracks, extent of the crushed areas,
and reinforcement yielding and buckling. Finally, seismic loss analysis
provides the conditional probability of exceedance of the decision var­
iable (i.e., financial loss) for a given DS.
Equation (1) shows the fundamental relation for computing total loss
through a triple integration. The equation implies the total probability
theorem in which the output of each four stages of Fig. 1 is simulta­
neously integrated.
∫ ∫ ∫
P(DV) = P(DV|DM)dP(DM|EDP)dP(EDP|IM)dλ(IM) (1)
im dm edp
Fig. 4. Beam-column-joint region for (a) exterior specimen [35], (b) interior
specimen [36], and (c) knee specimen [37]. In Equation (1), P(EDP|IM) is the probabilistic distribution of EDP, P
(DM|EDP) is the probabilistic relation of EDP and DM, and P(DV|DM) is
the conditional probability of exceedance of the decision variable for a
given DS.
The primary purpose of this paper is to use the post-earthquake
damage features as the EDP for the seismic loss measurement using
Equation (1). To this end, image-based damage features are employed to
predict the drift ratio of BCJs. The estimated peak drift ratio is then
served as an EDP to the fragility curves to compute the conditional
probability of P(DM|EDP) in Equation (1). Accordingly, the loss curve is
the ultimate output of the proposed loss measurement frameworks,
which enables deciding on the required post-earthquake repair action.
The schematic flowchart of the proposed damage assessment
framework, including corresponding sections in the paper, is presented
in Fig. 2. Based on this figure, the central core of this framework is
Fig. 5. The variation of the proposed parameters (a) crack length index (Xl) and machine-learning-based models that use visual damage features to
(b) crushing areal density index (Xc) versus drift ratio for five different speci­
predict the maximum drift ratio that a beam-column joint (BCJ) has
mens in four drift steps.
experienced. These models are explained in section 5. Section 6 in­
troduces the predictive models considering four scenarios based on data
availability. The seismic maximum drift ratio is the ultimate output of
these developed predictive models, which can be used as the input EDP
for the seismic loss estimation through fragility curves developed by
relevant loss measurement frameworks.

3. Database collection

The database of this study contains 974 images of surface crack


patterns associated with 256 damaged beam-column subassemblies at
various drift ratios. As summarized in Table A1, the BCJs have been
tested under the quasi-static cyclic loading in 56 experimental programs
during 1970–2021. The ranges and relative frequencies for important
Fig. 6. The variation of the proposed parameters (a) length index (Xl), and (b) properties of the specimens shown in Fig. 3 cover a broad range of
crushing areal density index (Xc) versus drift ratio for all the specimens. essential properties. For instance, the figure shows that most of the crack
patterns belong to the drift ratios below 5.0 %, and a few tests have
proceeded to drift ratios larger than 5.0 %. Fig. 3 also shows that most
Table 1 tested specimens had a concrete compressive strength in the range of 20
Input variables for each scenario. to 50 MPa. Noteworthy to mention that the database of this study has
Scenario number Input variables been adopted from [8]. Detailed information on the characteristics of
Scenario I Xl , Xc , Rj
the specimens is summarized in Appendix A.
Scenario II Hcolumn Lbeam
Xl , Xc , Rj, ,
dcolumn dbeam 4. Selection of input parameters
Scenario III Hcolumn Lbeam ’
Xl , Xc , Rj, , , f
dcolumn dbeam c
Scenario IV Hcolumn Lbeam FyB FyC Fs yB Fs yC Four scenarios are considered to develop the predictive models for
Xl , Xc , Rj,
estimating the peak drift ratio based on the selected input parameters.
, , , ’ , ’ , ’ , (Av /s)J
dcolumn dbeam f’c fc fc fc
The input parameters basically contain two groups. The first group in­
cludes damage-induced features, which are indications created after the
Intensity Measure (IM), primarily spectral acceleration or peak ground occurrence of damage. The second group contains features such as
acceleration, is estimated at the site of the structure. Afterward, struc­ design, geometry, and material properties.
tural analysis at different hazard levels is carried out to obtain specific
Engineering Demand Parameters (EDPs). The third loss measurement
stage investigates the probabilistic relation between EDP and Damage 4.1. Damage-induced features
measures (DM) through fragility models. The DM includes various
structural damage states (DSs) describing the severity of damage to the This study uses two visual features developed in the previous work
structural components. It should be noted that the DSs are determined in on RCSWs [13] to show the extent of damage in the reinforced concrete

4
M. Hamidia et al. Measurement 205 (2022) 112195

Xl

Transvers Joint
Xc

column
normalized Fystirrup
stirrup
Fybeam

columnReinforcement

normalized Fycolumn

normalized Fybeam

beamReinforcement
normalized
Av/S

FycolumnStirrups
/d

axial load ratio


beam

stirrup

column
Fybeam

/d
f'c

Fy
L

column
beam

beam
Rj

Fy

+
-

H
Fig. 7. Absolute values of the correlation coefficient of each variable and the drift ratio.

Fig. 9. Test-train accuracy regarding scenario I including the most items of


input information (a) R2 of the model (b) mean squared of error for every
utilized ML model.

measurements. Thus, for the practical case of a post-earthquake damage


assessment, an inspector can measure the cracking length and crushing
areas and use the proposed models to predict the peak-experienced drift
ratio of a damaged beam-column joint. As an example, extracted pa­
rameters are visualized in Fig. 5 for five specimens from the database,
including KJ3, KJ6, KJ8 [38], JC2 [35] and S1BC1 [39]. Fig. 5 shows the
variation of Xl and Xc versus the increment of the drift ratio. In fact, the
Fig. 8. Predicted values and actual values of the dataset in scenario I (a) pre­ increasing trend in these parameters is representative of the progression
diction vs observation plot of GB model (b) distribution of actual data against of the damage.
predicted values by GB model. (c) prediction vs observation plot of RF model Fig. 6 (a) and (b) show the variation of drift ratio versus the Xl and Xc
(d) distribution of actual data against predicted values by RF model. parameters extracted for all the 974 images. Similarly, these figures also
show the increasing trend between extracted parameters and the drift
BCJs. In particular, the parameter Xl indicates the cracking intensity ratio, indicating the amount of damage in this concept. Also, as observed
index and is defined as the cumulative cracking length in the beam- in Fig. 5 (b) and Fig. 6 (b), the crushing areal density index (Xc) is
column-joint region divided by the perimeter of the joint. The second negligible for drift ratios smaller than 1.5 %.
parameter Xc, which is named crushing areal density index, shows the
percentage of the extent of crushing and is defined as the ratio of the
cumulative crushing area in the beam-column-joint region to the entire 4.2. Beam-column properties
area of the region. The beam-column-joint region includes the beam-
column intersection area along with a square area from the top and Potential complementary features for predicting the seismic peak
bottom column with the dimension of column width and a square area drift ratio comprise mechanical properties (e.g., the concrete compres­
from the right and left beam with the beam width dimension. This region sive strength (f ’c ), beam and column longitudinal and transverse rebar
is depicted in Fig. 4 for different BCJ arrangements. strength (Fyb ,Fyc ,Fyv ) , design parameters (e.g., transverse reinforcement
In this study, the Arc Length method developed by Asjodi et al. [27] ratio in joint (Av /s), column axial load ratio (P/Ag fc’ ), beam bottom and
is used for automated quantification of the cracking length in the top reinforcement ratio (ρb , ρ’b ), column reinforcement ratio (ρc )), and
mentioned zone. The crushing areal density is also measured using the geometry properties (e.g., the aspect ratio of the joint (Rj), the ratio of
proposed filtering of Asjodi et al. [14]. It should be noted that the pro­ the beam length to beam depth (Lbeam/dbeam), and the column height to
posed features in this paper can be extracted using in-situ column depth (Hcolumn/dcolumn)). Moreover, two dimensionless

5
M. Hamidia et al. Measurement 205 (2022) 112195

Fig. 10. Predicted values and actual values of the dataset in scenario II (a) Fig. 12. Predicted values and actual values of the dataset in scenario III (a)
prediction vs observation plot of GB model (b) distribution of actual data prediction vs observation plot of GB model (b) distribution of actual data
against predicted values by GB model. (c) prediction vs observation plot of RF against predicted values by GB model. (c) prediction vs observation plot of RF
model (d) distribution of actual data against predicted values by RF model. model (d) distribution of actual data against predicted values by RF model.

Fig. 11. Test-train accuracy regarding scenario II including the most items of Fig. 13. Test-train accuracy regarding scenario III including the most items of
input information (a) R2 of the model (b) mean squared of error for every input information (a) R2 of the model (b) mean squared of error for every
utilized ML model. utilized ML model.

parameters are also considered, which are obtained by dividing the and ElasticNet regression methods are fitted to the database. These
strength of rebars by the concrete compression strength parametric regression methods work based on assuming a linear rela­
(Fyb /f ’c , Fyc /f ’c , Fyv /f ’c ). tionship between the predictors and the target, which is the drift ratio.
The output model of the linear method is shown in Equation (2):
5. Machine learning models ∑
n
y = θ0 + θ i Xi (2)
The purpose of this study is to develop predictive models for esti­
i=1

mating the seismic peak drift ratio experienced by BCJs damaged in an


where Xi is the ith predictor variable, θ0 is the intercept of the equation,
earthquake. The drift ratio is a significant index indicating the perfor­
θi is the coefficient of each predictor variable, and n is the number of
mance level of the BCJs and the major ingredient of loss measurement
predictor variables. In each method, the training procedure includes
frameworks, as stated in Section 2. Furthermore, ML methods are
minimizing a specific objective function considering the training data­
designed for establishing meaningful relations between various features
set. The objective functions associated with linear and ElasticNet models
when the relations are complicated, such as pattern recognition and
are presented in Equations (3) and (4), respectively.
where the amount of data is beyond human capability [40]. The present
study uses ML methods to investigate the relationship between struc­ (
∑ N
[

n
]2 )
tural properties, visual features, and the observed peak drift ratio. θ = argmin Yj − θ 0 − θi Xi (3)
Linear models: First, the ordinary least squared linear regression j=1 i=1

6
M. Hamidia et al. Measurement 205 (2022) 112195

of a single model. Random Forest (RF) and Gradient Boost (GB) are
ensemble learning algorithms that make predicting by using the results
of a group of multiple decision trees.
Artificial Neural Networks (ANN) model: ANNs are deep net­
works, including several layers, formingthe architecture of the models.
ANN model combines input variables through nodes of each layer and
establishes nonlinear relations between the features and target using the
activation function in each layer.
The open-source scikit-learn package [43] in Python is utilized to
develop ML models, including linear regression, ElasticNet, RF, GB, and
SVR models. The ANN regression model is also created using Keras an
open-source library in Python. The hyper-parameters used for these
models are calculated using the GridSearchCV function in the scikit-
learn module. GridSearchCV is an exhaustive search tool over speci­
fied hyperparameters. The evaluation criteria for comparing various
permutations are the k-fold cross-validation carried out over the training
dataset. Further explanation about hyper-parameter selection and
model design is provided in Appendix B. aforementioned models are
fitted to the database of beam-column joints so that 80 percent of the
database is randomly considered as the training data, and 20 percent is
selected as the testing dataset. Afterward, the testing dataset determines
the accuracy of the models. In addition to the contribution of each set of
data in predicting the target, the performance of each regression model
is also interpreted to introduce the best model. The functionality of each
regression method is highly dependent on the correlations of the fea­
Fig. 14. Predicted values and actual values of the dataset in scenario IV (a)
tures, the type of data, and the kind of problem.
prediction vs observation plot of GB model (b) distribution of actual data
against predicted values by GB model. (c) prediction vs observation plot of RF
model (d) distribution of actual data against predicted values by RF model. 6. Predictive models

Based on the availability of information about the damaged struc­


ture, four different scenarios are considered. The number of accessible
features increases from the first to the last scenario. The first scenario
only uses the visual features of the damaged component. However, in
the last scenario, it is assumed that all the features of the damaged
specimen, including design, mechanical and geometric properties, as
well as visual features, are available. Table 1 lists the selected input
parameters for each scenario. It is noted that the parameters of Table 1
are introduced in section 4.2.
The absolute value of the correlation coefficient (R) of all the
candidate parameters and the target parameter (i.e., the drift ratio) is
plotted in Fig. 7. This figure gives a deep insight into how the variation
of the various variables is related to the variation of the drift ratio. The
visual features of the damage are the most significant parameters with
Fig. 15. Test-train accuracy regarding scenario IV including the most items of the highest correlation with the drift ratio. Furthermore, among the
input information (a) R2 of the model (b) mean squared of error for every various properties of the beam-column specimen, normalized strength of
utilized ML model. reinforcement, transverse reinforcement ratio, and the beam and col­
umn length to depth ratio show a better correlation with the drift ratio.
In contrast, reinforcement ratios and rebar strengths show less
where Yj indicates the jth observation, N shows the number of data rows correlation.
in the database. The following sub-sections describe four scenarios considered for
(
∑N
[(

n
)2 ]
∑n ∑n ( )
) developing the predictive models. The basis of selecting the input vari­
θ = argmin Yj − θ 0 − θi Xij + L1 |θ i | + L2 θ i
2
ables and the hyperparameters is a matter of trial and error. However,
1 1
j=1 i=1 the reported correlation coefficients in Fig. 7 give significant insight into
(4) selecting or replacing variables.

where L1 and L2 are the regularization parameters. The case of L1 and L2 6.1. Scenario I
equal to zero is equivalent to an ordinary linear regression model [41].
Furthermore, four nonlinear models are also used in this research, The first scenario considers that the only available information about
including: the BCJ is a specimen image. Accordingly, the only available data are the
Support Vector Regression (SVR) model: This method seeks to visual feature of surface damage. These visual features are the two
match the optimal hyperplane inside a boundary value [42], where each previously introduced indices (Xl and Xc) as well as the aspect ratio of the
data point of the training dataset is mapped in the feature space of the joint (Rj). Fig. 8 shows the prediction versus observation plot and the
model then the selected kernel of the model transforms data points into distribution of the predicted and observed drift ratio for the best pre­
the required form. dictive models, namely Random Forest (RF) and Gradient Boost (GB)
Ensemble learning models: These models assemble a group of models. Based on the results, the coefficient of determination (R2) is
multiple ML models as estimators in order to enhance the performance reported as 0.77 and 0.76 for GB and RF models, respectively.

7
M. Hamidia et al. Measurement 205 (2022) 112195

Fig. 16. The performance of the various regression models in different scenarios.

compression strength, which can be measured by field tests as well as the


visual damage indices and dimensional information, are fed to the ma­
chine learning models. According to Fig. 12, the accuracy of predictive
models is enhanced by adding strength information to the inputs.
The MSE and R-squared regarding the test and train dataset for all
models in the third scenario are presented in Fig. 13. the RF model is the
most accurate predictor in the third scenario.

6.4. Scenario IV

The last scenario develops predictive models using all structural in­
formation of the specimens. The ultimate combinations of the available
information in this scenario are selected based on the reported corre­
Fig. 17. Relative importance of each variable.
lations in Fig. 14, so that the yield stress of longitudinal and transverse
rebars are divided by the compressive strength of concrete in order to
Moreover, the accuracy and mean-squared errors (MSE) for all the make it normalized and dimensionless features. Fig. 14 shows the pre­
machine learning methods for both test and train datasets are presented diction versus observation plot and the distribution of the predicted and
in Fig. 9. In the case of linear models, using ElasticNet model has slightly observed drift ratio for the best predictive models, which are RF and GB.
enhanced the test accuracy. However, the training accuracy decreased The obtained R2 in this scenario equals 0.80 and 0.83 for RF and GB
because of the influence of ElasticNet algorithm to make a bias-variance models, respectively. It can be concluded that further information about
trade-off to increase the predicting accuracy facing new data. It can be the material and reinforcement ratio does not considerably enhance the
observed that nonlinear models are markedly more accurate. SVR, ANN, predictive models, so the visual features and concrete strength used in
and ensemble learning models have considerably decreased the MSE. the third scenario provide sufficient information for drift ratio
prediction.
6.2. Scenario II Moreover, the accuracy and MSE for all the machine learning
methods for both test and train datasets are presented in Fig. 15. Similar
The second scenario uses information on the length of the column to other scenarios, nonlinear models have made a breakthrough in
and beam in addition to the cracking and crushing indices and the joint predicting the target value.
aspect ratio. Dimensional data of the BCJs has been normalized by
dividing the length of the column and beam by the depth of the column
6.5. Comparison of predictive models and scenarios
and beam, respectively, to produce dimensionless features. Fig. 10
shows that the R2 of the model is enhanced to 0.79 and 0.8 for GB and RF
In this subsection, the results of all models and scenarios are
models, respectively.
compared and interpreted. For the sake of a more thoughtful compari­
Fig. 11 also presents the MSE and R-squared accuracy for all models
son, a new scenario is also defined as Scenario 0, in which all the
in this scenario. The trend is similar to what was observed in Fig. 9. In
selected variables except the visual features (Xl and Xc) are included, and
this scenario RF model is the most accurate model.
the results of this scenario are brought along with other scenarios.
Fig. 16 (a) indicates that linear models, including ordinary least squared
6.3. Scenario III regression and ElasticNet models, have failed to properly fit the data­
base, which in turn implies the complexity of the database. By com­
The third scenario incorporates the compressive strength of concrete parison, SVR has shown better performance; Particularly, by adding
(f’c) into the training data. Thus, in the third scenario, concrete dimensional information to the input parameters, SVR prediction results

8
M. Hamidia et al. Measurement 205 (2022) 112195

Fig. 18. Summary plot of SHAP values calculated for the test dataset.

Fig. 19. Importance factor of each input.

have improved significantly. Finally, ensemble learning models, RF and Finally, Fig. 17 shows the relative importance corresponding to the
GB models, have perfectly fitted to the database in all scenarios last scenario, which includes all properties. This parameter has been
compared to other models because of the flexibility of tree-based models extracted from GB and RF models, using the reduction in the impurity in
in fitting to complicated datasets. Moreover, it can be seen from Fig. 16 each node [44]. As it is shown, crushing and cracking indices are the
(b) that in the third and fourth scenarios, although important informa­ most significant properties. Other features, such as length, aspect ratio,
tion has been added to the models, in the presence of visual features, and yielding stress of reinforcements, have a similar contribution in
further information does not considerably enhance the models. More­ predicting the drift ratio.
over, the gray line in Fig. 16 (a) showing the accuracy of non-visual
feature scenario indicates that visual features play a crucial role in 7. Sensitivity analysis
predicting the drift ratio, and ML are incapable of making an accurate
prediction without visual features. In other words, the proposed visual Machine learning models are powerful tools capable of establishing
features are strong representatives of the imposed damage to the spec­ complicated relations between input and output variables. However, the
imen. Therefore, considering the cost and time required for achieving lack of interpretability concerning the input variables of the model is
the required information on the mechanical and design properties of a one of the drawbacks of these models. In the previous section, the
BCJ along with the sufficiency of the visual features, the second scenario relative importance of the tree-based models is provided based on the
is proposed as the selected scenario for the drift prediction of BCJs. calculated impurity during the training procedure. Still, this method

9
M. Hamidia et al. Measurement 205 (2022) 112195

Table 2
Example application of the proposed methodology for specimen EH60 [47].
Crack patterns of the specimen Drift ratio measured from the experiment Input features Predicted drift ratio

Xl Xc Rj Lbeam/dbeam Hcolumn/dcolumn

1.00 % 0.45 0.00 1.1 8.7 6.2 1.01 %

2.00 % 2.81 0.00 1.1 8.7 6.2 2.11 %

3.00 % 3.21 0.01 1.1 8.7 6.2 2.63 %

4.00 % 3.25 2.65 1.1 8.7 6.2 4.04 %

Fig. 20. Predicted versus actual drift ratio for the specimen tested by Alaee [47] using (a) RF model and (b) GB model, in all scenarios.

10
M. Hamidia et al. Measurement 205 (2022) 112195

8. Example application

In this section, a sample example of the proposed methodology using


the RF model of the second scenario is presented. The required visual
features of the BCJ specimen EH60 tested by Alaee [47] have been
extracted and presented in Table 2. This table demonstrates the crack
pattern of the damaged specimen along with the corresponding extrac­
ted features and the predicted value for drift ratio. It is worth noting that
the specimen EH60 was selected from the testing dataset, and this
specimen was not used during the training procedure.
In addition to the presented predicted value in Table 2, the required
input data in each scenario are also fed to the selected models (RF and
GB), and the predicted versus actual drift ratio is shown in Fig. 20. Ac­
cording to this figure, all scenarios can adequately predict the drift ratio.
Based on the provisions of FEMA P-58 elaborated in Section 2, the
fragility curves showing the probability of exceedance from each dam­
age state for the specimen EH60 [47] are depicted in Fig. 21. The pre­
dicted values of Table 2 are used as the EDP to identify the DS in each
drift ratio, denoted with black points in the figure. Presented results of
Table 2 are based on second scenario and the RF predictive model. The
input parameters are also listed in Table 2.
Fig. 21. DS fragility curves [3] for specimen EH60 [47].
In order to show the application of the presented framework for all
types of beam-column joints, Table 3 presents two examples for each
cannot be implemented on non-tree-based models, such as the ANN interior, exterior, and knee type. It is worth noting that all of these
model, because the term impurity is not defined for these methods. To specimens have been selected from the testing dataset.
solve this issue, the SHapely Additive exPlanation (SHAP) method, one
of the state-of-the-art techniques in machine learning expandability [45] 9. Summary and conclusions
is employed to interpret the sensitivity of a predictive model concerning
the input features. This approach is based on cooperative game theory An innovative computer vision-based methodology aimed at
and is used to increase the transparency and interpretability of ML obtaining the seismic loss estimation for reinforced concrete moment
models. In fact, the SHAP value quantifies the contribution of each frames in compliance with existing loss estimation frameworks is
feature to the prediction made by the model. As Equation (5) shows, the developed in this paper. The methodology takes advantage of an
output of the model is defined as a linear addition of input features. extensive collected database of 265 cyclic-loaded beam-column-joint

M specimens, which includes 974 images from quasi-static cyclic loading.
yi = φi0 + φi xj (5) Images of the damaged specimens contain crack patterns and crushed
j=0
areas at various segments of beam-column subassemblies. The extent of
surface damage, including the cracking length and the crushing area, is
where yi isthei th output of the model, φixj isthe SHAP value related to jth
extracted as indices of surface damage used in a scenario-based assess­
feature in ith prediction where the number of features is equal to M.
ment to predict the seismic peak drift ratio of the damaged beam-column
Fig. 18 shows the calculated SHAP values for the ANN model in the
subassemblies. Four scenarios are proposed considering the available
fourth scenario (including the most features). Each point on the plot in
input parameters among visual indices of surface damage (i.e., cracking
Fig. 18 is a SHAP value for the input variables for a specimen. The y-axis
length and crushing areas), mechanical properties (e.g., concrete
indicates the input variables in order of importance from top to bottom.
compressive strength), geometric properties (e.g., aspect ratio and
Each dot is colored by the value of the input variable, from low to high
length of components), and design properties of the specimen (e.g.,
according to the color map. As it is depicted, the higher the Xc and Xl
rebar percentage). Using this set of input parameters, six ML-based
values, the higher their SHAP value and the greater their impact on the
models for linear (e.g., elastic net and ordinary linear) and nonlinear
output. The longitudinal and transverse reinforcement strength of col­
(e.g., support vector machine, ANN, random forest, and gradient boost)
umns and beams are the next significant inputs.
regression are trained to predict the seismic peak drift ratio. The
As the calculated SHAP value of each prediction indicates the
robustness of the methodology is verified using a testing dataset which is
contribution of each input variable in predicting the output, the average
basically 20 % of the total images of the database which have not been
absolute SHAP value represents the importance factor of input variables
used for the training of the ML-based models. The seismic peak drift
[45,46]. These values are listed in Fig. 19 for each machine-learning
ratio obtained by the proposed methodology can finally be incorporated
algorithm. Note that the summation of the presented value in Fig. 19
into the existing fragility curves developed by relevant seismic loss
is normalized to unity. Due to variations in different algorithms, the
estimation frameworks to measure the loss following an earthquake. The
importance factor of each feature varies in different models. However, in
application of the proposed methodology is finally shown for two sets of
almost all algorithms, Xl and Xc indicating the propagation of damage
crack patterns with various in-plane configurations. Based on the results,
have the most significant contribution to the functionality of the models,
it can be concluded that:
which is similar to the relative importance result of the tree-based
models presented in Fig. 17. Moreover, as Fig. 19 shows, more vari­
• For all considered scenarios, the GB and RF models yield the highest
ables contribute to the prediction procedure than linear models, which
accuracy among all considered models. The ANN models possess the
incorporate a few variables in prediction. In particular, linear models
3rd accuracy ranking for all considered scenarios.
(linear regression and ElasticNet regression), because of the linear as­
• The predictive equation based on the visual indices of damage as the
sumptions and also the high correlation between Xl and Xc use one of
only input variable, namely scenario I, has the lowest accuracy
them (Xl) as the main visual feature.
among other predictive equations. The coefficient of determination
for testing dataset R2 = 0.77 and 0.76 for GB model and RF models,

11
M. Hamidia et al. Measurement 205 (2022) 112195

Table 3
Predicted drift ratio using the presented framework for all three types of BCJs.
Crack patterns of the specimen Drift ratio measured from the experiment Input features Predicted drift ratio

Xl Xc Rj Lbeam/dbeam Hcolumn/dcolumn

0.25 % 0.20 0.00 1.5 6.7 8.8 0.35 %

1.00 % 0.60 0.00 1.1 8.7 6.2 1.09 %

1.75 % 2.07 0.01 0.8 8.8 3.6 1.56 %

2.00 % 1.58 0.01 0.7 9.2 7.1 1.80 %

2.88 % 2.76 0.86 1.3 7.5 8.8 2.96 %

3.20 % 2.04 0.27 1 10 10 3.08 %

respectively. The mean-squared error MSE = 0.55 and 0.59 for GB and 0.49 for GB and RF models, respectively. The corresponding
model and RF models, respectively. values for Xl are equal to 0.21 in both GB and RF models.
• The coefficient of determination in the second scenario is enhanced
by 4 % for RF model compared to the first scenario. The mean- In conclusion, the second scenario, which includes surface damage
squared error also decreased by 0.13 in this scenario compared to features and geometric parameters as input parameters, provides a
the first scenario. In this scenario, the geometric dimension data is sufficient level of accuracy for measuring the seismic peak drift ratio
added to the input variables of Scenario I to provide more data from and, consequently, for measuring the loss in damaged RCMFs. The
the damaged specimen. proposed practical methodology can be used for rapid post-earthquake
• Interestingly, the addition of the compressive strength of the con­ damage assessment and loss estimation by measuring crack length and
crete and other structural properties of the RC moment frames to the crushing area in the field. The methodology can also be used in an
input data, namely scenario III and scenario IV, leads to a slight automated procedure using image processing techniques after collecting
improvement in the accuracy. images by the reconnaissance teams. Moreover, the effect of features
• The sensitivity analysis through SHAP values of scenario IV reveals reviling more details about the damage such as crack width, rebar
that the cracking length index and crushing areal density index yielding status, and rebar exposure, can be further explored in future
impose the highest impact on the drift ratios predicted by ML-based studies to improve predictive models.
models. The mean normalized SHAP values for Xc is equal to 0.37

12
M. Hamidia et al. Measurement 205 (2022) 112195

Table A1
Properties of beam-column joint specimens of the database.
Researcher Specimen ID fc, (MPa) P FyB (MPa) FyC (MPa) Fyv (MPa) ρB (%)

ρB (%) ρC (%) TRiJ1? (Y/N) Drift (%)
Ag fc,

Megget and Unit 1 28 0.00 286 305 317 1.3 1.3 1.2 Y 0.5,1.0
Park [48] Unit 2 27 0.00 290 305 317 1.3 1.3 1.2 Y 1.0,1.9
Unit 3 36 0.00 286 305 250 1.3 1.3 1.2 Y 2.3
Renton [49] Unit1 26 0.16 297 308 318 1.7 1.7 1.8 Y 1.0,3.8
Unit2 38 0.16 291 312 309 1.7 1.7 1.8 Y 0.7,3.6,3.8,5.3
Unit3 24 0.16 289 312 309 1.7 1.7 1.8 Y 0.9,1.1,2.8,4.7
Unit4 29 0.16 301 313 311 1.7 1.7 1.8 Y 1.9,2.0
Smith [50] Unit4 20 0.05 296 274 334 1.3 1.3 1.2 Y 0.3,1.0,2.0,4.0
Unit5 20 0.05 301 274 334 1.3 1.3 1.2 Y 0.3,1.0,2.0,4.0
Unit6 18 0.06 289 297 334 1.3 1.3 1.2 Y 0.3,1.0,2.0,4.0
Park and Unit1 32 0.25 405 405 289 0.2 0.2 1.3 Y 2.8,4.5
Thompson [51]
Uzumeri and Unit 6 37 0.10 352 340 347 1.3 0.9 2.7 Y 1.5,3.5
Seckin [52] Unit 7 31 0.12 352 340 347 1.3 0.9 2.7 Y 4
Unit 8 26 0.14 352 389 347 1.7 1.3 2.7 Y 3.6
Megget [53] UnitA 22 0.07 374 365 401 1.8 1.1 2.4 Y 2.2,3.3
Keong [54] Unit11 44 0.18 328 439 328 0.2 0.2 1.5 Y 0.9,3.0,5.4
Unit12 41 0.20 328 431 328 0.2 0.2 1.3 Y 3.0,4.6
Unit13 48 0.17 328 439 328 0.2 0.2 1.5 Y 3.0,5.0
Birss [55] Unit B1 28 0.05 288 427 398 1.1 1.1 2.6 Y 0.9,1.8,3.3,3.4
Unit B2 32 0.44 288 427 398 1.1 1.1 2.6 Y 0.4,0.6,1.5,2.8
Beckingsal [56] Unit B11 35 0.05 298 422 331 1.1 0.6 2.2 Y 0.4,1.3,2.5,3.5
Unit B12 35 0.05 298 422 331 0.9 0.9 2.2 Y 1.2,2.4,3.7
Unit B13 29 0.50 298 398 331 0.9 0.9 2.2 Y 0.8,1.7,2.5,3.3
Milburn [57] Unit1 41 0.10 315 473 321 1.5 1.5 2.2 Y 2.5,4.2,4.6
Unit2 47 0.10 307 473 321 1.8 1.8 2.2 Y 3.9,5.8
Unit3 38 0.10 303 485 321 1.7 1.7 0.8 Y 2.9,5.4,5.8
Unit4 39 0.10 305 485 321 2.3 2.3 0.8 Y 1.7,3.5,5.0,5.4
Ehsani and Unit1 33 0.06 410 489 336 0.9 0.8 1.8 Y 2.5
Wight [58] Unit4 45 0.06 410 489 336 1.0 0.8 2.1 Y 4.6,4.8
Philleo and LIJ3 33 0.27 400 400 400 1.2 1.2 2.7 Y 1.9,2.8
Abrams [59] LIJ4 33 0.03 420 420 420 0.6 0.6 1.0 Y 0.6,1.0,1.7,2.2
Ruitong and Unit1 39 0.01 294 498 344 1.0 0.4 1.3 Y 2.4,4.2
Park [60] Unit2 30 0.02 307 476 324 1.2 0.6 2.0 Y 3.0
Unit3 31 0.01 294 514 323 1.0 0.4 1.0 Y 2.2,3.9
Unit4 36 0.02 307 487 311 1.2 0.6 1.6 Y 2.9
Pessiki et al. [61] Unit 1 28 0.40 483 456 420 1.2 0.6 2.0 N 5
Unit 2 28 0.40 483 456 420 1.2 0.6 2.0 N 2
Unit 3 26 0.39 483 486 420 1.2 0.6 2.0 N 2
Unit 4 26 0.39 483 518 420 1.2 0.6 2.0 N 2
Unit 5 28 0.30 531 427 420 1.2 0.6 2.0 Y 2.0,4.0,5.0
Unit 6 24 0.34 531 507 420 1.2 0.6 2.0 Y 2.0,5.0
Unit 7 24 0.32 480 461 420 0.7 0.5 2.0 N 2.2,4.0,5.0
Unit 8 24 0.32 480 461 420 0.7 0.5 2.0 N 2.0,3.0,5.0
Unit 9 28 0.08 422 461 420 0.7 0.5 2.0 N 2.0,3.0,4.0
Unit 10 24 0.34 453 476 420 0.7 0.5 2.0 N 2.0,5.0
Alameddine [62] LLS 55 0.04 457 468 445 1.3 1.3 2.8 Y 3.2,4.8
Mazzoni et al. Unit 2H 21 0.00 414 414 414 1.1 1.1 3.7 Y 6.3
[51] Unit 4H 21 0.00 414 414 414 1.1 1.1 3.7 Y 6.3
Unit R 21 0.00 414 414 414 1.1 1.1 3.7 Y 6.3
Beres et al. [52] I-11 24 0.39 414 414 414 0.5 0.5 2.0 N 0.9,1.2,1.6
E-01 24 0.11 414 414 414 0.5 0.5 2.0 N 0.9,1.2,1.6
Xin [53] Unit1 28 0.08 453 472 350 0.7 0.7 1.4 Y 1.7,3.4,5.0
Unit2 30 0.07 445 453 350 0.7 0.4 0.9 Y 0.4,1.2,2.3,3.5,4.6
Unit3 32 0.07 445 470 348 0.7 0.7 1.6 Y 0.5,1.3,2.7,4.0,5.4
Unit4 37 0.06 473 454 348 0.6 0.4 1.3 Y 0.4,1.1,2.3,3.4,4.6
Unit5 61 0.04 492 470 345 0.8 0.8 1.6 Y 0.6,1.6,3.2,4.7,6.3
Unit6 59 0.04 478 470 345 1.0 0.6 1.6 Y 0.5,1.3,2.6,4.0,5.3
Hoffschild [54] RCBC1 26 0.47 566 566 566 1.0 0.5 1.3 N 5.3
RCBC2 26 0.47 566 566 566 1.0 0.5 1.3 N 5.3
Higazy [55] SA1 28 0.15 400 400 400 0.6 0.6 1.3 Y 0.5,1.4,2.5
SA2 27 0.15 400 400 400 0.6 0.6 1.3 Y 0.4,1.4,2.3
SA3 28 0.15 400 400 400 0.6 0.6 1.3 Y 0.7,1.3,2.4
SB1 29 0.15 400 400 400 1.1 1.1 1.9 Y 0.7,1.5,2.5
SB2 29 0.15 400 400 400 1.1 1.1 1.9 Y 0.6,1.4,2.2
SB3 28 0.15 400 400 400 1.1 1.1 1.9 N 0.7,1.5,2.3
SC1 42 0.15 400 400 400 1.3 1.3 2.2 Y 0.8,1.8,2.4
SC2 41 0.15 400 400 400 1.3 1.3 2.2 Y 0.8,1.9,2.5
SC3 43 0.15 400 400 400 1.3 1.3 2.2 Y 0.8,1.8,2.5
Cote and KJ4 36 0.00 448 448 444 0.4 0.2 1.0 Y 4
Wallace [56]
Hakuta [57] O1 34 0.00 325 325 339 1.4 0.7 2.0 N 0.4,0.9,1.2,2.3
R1 50 0.00 302 462 316 0.7 0.5 1.0 Y 0.3,0.4,0.8,1.6,2.4,3.2
(continued on next page)

13
M. Hamidia et al. Measurement 205 (2022) 112195

Table A1 (continued )
Researcher Specimen ID fc, (MPa) P FyB (MPa) FyC (MPa) Fyv (MPa) ρB (%)

ρB (%) ρC (%) TRiJ1? (Y/N) Drift (%)
Ag fc,

R2 54 0.00 302 462 316 0.7 0.5 1.0 N 0.2,0.3,0.6,1.3,2.0,2.6


R3 39 0.00 308 308 302 1.2 0.6 1.0 N 0.2,0.2,0.5,1.0,1.6,2.1
O4 41 0.00 308 321 378 1.4 1.4 1.3 N 0.4,0.5,1.1,2.2,3.3,4.4
O5 27 0.00 306 321 398 1.2 1.2 1.3 N 0.4,0.6,1.2,2.4,3.6,4.9
O6 31 0.00 308 308 398 1.0 0.7 0.9 Y 0.3,0.4,0.8,1.6,2.4,3.3,4.9
O7 27 0.00 308 308 398 1.0 0.7 0.9 Y 0.3,0.4,0.9,1.9,2.8,3.8
Clyde et al. [58] Test#2 46 0.10 454 470 428 2.0 2.0 2.2 N 0.2,0.5,0.9,1.9,2.4
Test#4 41 0.25 454 470 428 2.0 2.0 2.2 N 0.3,0.5,0.7,1.4,1.8
Test#5 37 0.25 454 470 428 2.0 2.0 2.2 N 0.2,0.5,0.9,1.7,2.4
Test#6 40 0.10 454 470 428 2.0 2.0 2.2 N 0.2,0.6,0.9,1.9,2.6

Hamili [59] C4ALN1CY 44 0.05 500 500 500 1.7 1.7 3.6 Y 1.4,3.8,4.6
C4ALN7CY 50 0.04 500 500 500 1.7 1.7 3.6 Y 1.4,3.8,4.6
C6LN1CY 45 0.05 500 500 500 1.7 1.7 3.6 Y 1.4,3.8,4.6
C6LN7CY 47 0.05 500 500 500 1.7 1.7 3.6 Y 1.4,3.8,4.6
C6LNP4 46 0.05 500 500 500 1.7 1.7 3.6 Y 1.4,3.8,4.6
C6PLNP4 41 0.05 500 500 500 1.7 1.7 3.6 Y 1.4,3.8,4.6
C3LN7CY 43 0.05 500 500 500 1.7 1.7 3.6 Y 3.8,4.6,5.3
C3XLN7CY 44 0.05 500 500 500 1.7 1.7 3.6 Y 3.8,4.6,5.3
Walker [60] PEER-14 32 0.10 463 425 662 0.8 0.4 1.3 N 0.5,0.7,1.0,1.5,2.0,3.0
CD15-14 30 0.10 463 425 662 0.8 0.4 1.3 N 1.5
CD30-14 43 0.10 463 425 662 0.8 0.4 1.3 N 3
PEER-22 38 0.10 527 538 662 1.1 0.8 2.8 N 0.2,0.5,0.7,1.0,1.5
CD30-22 38 0.10 516 510 662 1.1 0.8 2.8 N 3.0,5.0
Chutarat [63] Unit I 28 0.00 488 488 362 1.3 1.3 2.8 Y 2.1,5.7
Unit A 33 0.00 488 488 362 0.5 0.5 1.5 Y 1.8,4.8
Pantelides Test#1 33 0.10 459 470 427 1.6 1.6 2.5 N 0.8,1.5,2.0,3.0
et al. [62] Test#2 30 0.25 459 470 427 1.6 1.6 2.5 N 0.5,1.5,2.0,3.0
Test#3 34 0.10 459 470 427 1.6 1.6 2.5 N 0.5,1.5,2.0,5.0
Test#4 32 0.25 459 470 427 1.6 1.6 2.5 N 0.5,1.5,2.0,5.0,5.0
Test#5 32 0.10 459 470 427 1.6 1.6 2.5 N 1.5,2.0,2.0,5.0
Test#6 31 0.25 459 470 427 1.6 1.6 2.5 N 0.5,1.5,2.0,5.0
Alire [63] PEER-0850 34 0.10 462 462 462 0.4 0.4 0.6 N 0.2,0.5,0.7,1.0,1.5,2.0
PEER-0995 65 0.10 462 462 462 0.9 0.6 1.2 N 0.2,0.5,0.7,1.0,1.5,2.0,3.0
PEER-1595 65 0.10 862 462 662 0.8 0.6 3.5 N 0.2,0.5,0.7,1.0,1.5,2.0,3.0
PEER-4150 34 0.10 462 462 462 1.9 1.9 3.5 N 0.2,0.5,0.7,1.0,1.5,2.0
Wong [63] BS-L 31 0.15 520 520 500 0.8 0.8 2.2 N 0.2,0.4,1.0,1.4,1.9
BS-OL 31 0.15 520 520 500 0.8 0.8 2.2 N 0.2,0.4,1.0,1.4,1.9,2.8,3.8
BS-U 37 0.15 520 520 500 0.8 0.8 2.2 N 0.2,0.4,1.0,1.3,1.9,2.8
BS-LL 37 0.15 520 520 500 0.8 0.8 2.2 N 0.2,0.4,1.0,2.0,4.0
BS-L-LS 32 0.15 520 520 500 0.8 0.8 2.2 N 0.2,0.4,1.0,2.0,4.0
BS-L-V2T10 33 0.15 520 520 500 0.8 0.8 2.9 N 0.2,0.4,1.0,2.0,4.0
BS-L-V4T10 28 0.15 520 520 500 0.8 0.8 3.6 N 0.2,0.5,1.0,2.0,4.0
BS-L-H1T10 33 0.15 520 520 500 0.8 0.8 2.2 Y 0.2,0.4,1.0,2.0,4.0
BS-L-H2T10 42 0.15 520 520 500 0.8 0.8 2.2 Y 0.2,0.4,1.0,2.0,4.0
BS-L-300 34 0.15 520 520 500 1.2 1.2 2.2 N 0.2,0.4,1.0,2.0,4.0
BS-L-600 36 0.15 520 520 500 0.6 0.6 2.2 N 0.2,0.4,1.0,2.0,4.0
JA-NN-03 45 0.03 520 520 500 0.6 0.6 2.2 N 0.2,0.4,1.0,1.9,2.8
JA-NN-15 46 0.15 520 520 500 0.6 0.6 2.2 N 0.2,0.4,1.0,1.9,2.8,3.8,4.8
JA-NY-03 41 0.03 520 520 500 0.6 0.6 2.2 Y 0.2,0.4,1.0,1.9,2.8,3.8,4.8
JA-NY-15 37 0.15 520 520 500 0.6 0.6 2.2 Y 0.2,0.4,1.0,1.9,2.8,3.8
JB-NN-03 43 0.03 520 520 500 0.8 0.8 2.2 N 0.2,0.4,1.0,1.9,2.8,3.8
JB-NY-03 34 0.03 520 520 500 0.8 0.8 2.2 Y 0.2,0.4,1.0,1.9,2.8,3.8,4.8,5.7
Hertanto [64] TDD-1 21 0.07 326 326 424 0.5 0.2 0.9 N 0.2,0.5,1.0,1.5,2.0,2.5
TSD 21 0.07 326 326 424 1.4 0.7 0.9 Y 0.5,1.0,1.5,2.0,2.5,3.0,4.5
TDD-2 23 0.06 354 354 408 0.7 0.5 0.9 N 0.5,1.0,1.5,2.0,2.5,3.0,3.5
Liu [65] NZ-7 24 0.07 307 307 384 0.6 0.6 1.8 Y 0.2,0.5,1.0,2.0,4.0,5.0
Shiohara and C1 28 0.08 456 357 326 1.2 1.2 2.4 Y 4
Kusuhara [66] A1 28 0.08 456 357 326 1.2 1.2 2.4 Y 4
A2 28 0.08 456 357 326 1.2 1.2 2.4 Y 4
A3 28 0.08 456 357 326 1.2 1.2 2.4 Y 4
B1 28 0.08 456 357 326 1.5 1.5 2.2 Y 4
B2 28 0.08 456 357 326 1.5 1.5 2.2 Y 4

Chun et al. JCT 11–1 31 0.05 468 431 396 0.9 0.9 2.2 Y 4
[67] JMT 11-2B 42 0.05 468 431 396 0.9 0.9 2.2 Y 4
Supaviriyakit J3A 29 0.13 480 480 299 1.3 1.3 2.9 Y 0.3,0.8,1.0,1.8,3.0,5.0
Pimanmas [68] J3B 24 0.13 480 480 299 1.3 1.3 2.9 Y 0.3,0.8,1.0,1.8,3.0,5.0
Ishikawa [69] MJIS 86 0.15 626 626 742 1.0 1.0 2.7 Y 2.0,3.0,5.0
MJ5H 86 0.15 626 626 742 1.0 1.0 2.7 Y 2.0,3.0,5.0
MJ3H 86 0.15 626 626 742 1.0 1.0 2.7 Y 2.0,5.0
Au [70] RCB1 32 0.12 329 316 418 0.6 0.6 2.0 Y 0.5,0.7,1.0,1.5,2.0,2.5,3.5,4.5
Yuen [71] NS-O1 17 0.27 362 362 400 0.3 0.3 0.9 N 1.0,2.0,3.0,3.3
Ibrahim [72] SS-5 35 0.08 431 461 484 1.0 1.0 1.8 Y 1.2,2.4,3.6,4.8,5.4
SS-1 36 0.08 495 493 484 1.0 1.0 1.8 Y 1.5,2.4,3.6,4.8,5.4
(continued on next page)

14
M. Hamidia et al. Measurement 205 (2022) 112195

Table A1 (continued )
SS-2 37 0.08 495 493 484 1.0 1.0 1.8 Y 1.5,3.6,4.8
SS-3 35 0.09 495 493 484 1.0 1.0 1.8 Y 1.5,2.4,3.6,4.8
SS-4A 34 0.09 495 493 484 1.0 1.0 1.8 Y 1.5,2.4,3.6,4.8
SS-6 38 0.08 466 493 484 1.0 1.0 1.8 Y 1.2,2.4,3.6,4.8
SS-7 37 0.08 431 493 484 1.0 1.0 1.8 Y 1.2,2.4,3.6,4.8,5.4
SS-4B 35 0.09 495 493 484 1.0 1.0 1.8 Y 1.6,4.1
C-2 36 0.08 495 493 484 1.0 1.0 1.8 Y 1.5,2.4,3.6,4.8,5.4
C-1 35 0.08 495 493 484 1.0 1.0 1.8 N 0.6,1.5,3.3,4.8
Lai [73] NS-1 60 0.00 510 512 358 0.7 0.7 1.4 Y 1.0,2.0,3.0,4.0,4.2
AS-1 60 0.30 510 512 358 0.7 0.7 1.4 Y 1.0,2.0,3.0,4.0,4.4
NS-2 60 0.00 508 512 358 0.7 0.4 0.9 Y 1.0,2.0,3.0,4.0,4.2
AS-2 60 0.30 508 512 358 0.7 0.4 0.9 Y 1.0,2.0,3.0,4.0,4.3
NS-3 60 0.00 508 512 356 0.7 0.7 1.6 Y 1.0,2.0,3.0,4.0,4.2
AS-3 60 0.30 508 512 356 0.7 0.7 1.6 Y 4.1
NS-4 60 0.00 510 507 356 0.6 0.4 1.3 Y 1.0,2.0,3.0,4.0,4.1
AS-4 60 0.30 510 507 356 0.6 0.4 1.3 Y 1.0,2.0,3.0,4.0,4.2
Decker [74] 8–3 54 0.10 476 482 479 1.0 1.0 3.5 Y 5.5
8–4 56 0.10 476 482 479 1.0 1.0 3.5 Y 4.5
8–5 55 0.10 478 482 479 1.5 1.5 3.5 Y 5
Ashtiani BCJ 1 124 0.07 312 306 371 1.1 1.1 2.5 Y 1.0,1.5,2.0,2.5,3.5,4.5
[75] BCJ 2 113 0.01 312 306 371 1.1 1.1 2.5 Y 1.0,1.5,2.0,2.5,3.5,4.5
BCJ 3 101 0.08 312 306 371 1.1 1.1 2.5 Y 1.0,1.5,2.0,2.5,3.5,4.5
BCJ 4 122 0.07 530 537 371 0.9 0.9 1.4 Y 1.0,1.5,2.0,2.5,3.5,4.5
BCJ 5 83 0.10 312 306 371 1.1 1.1 2.5 Y 1.0,1.5,2.0,2.5,3.5,4.5
BCJ 6 48 0.08 312 306 371 1.1 1.1 2.5 Y 1.0,1.5,2.0,2.5,3.5,4.5
BCJ 7 46 0.08 312 306 371 1.1 1.1 2.5 Y 1.0,1.5,2.0,2.5,3.5,4.5
Jemaa [76] JA-2 31 0.07 554 554 533 0.8 0.8 1.8 N 0.2,0.5,0.7,1.0,1.5,2.0
JA-3 32 0.11 554 554 533 0.8 0.8 1.8 N 0.2,0.5,1.0,1.5,2.0,3.0,4.0,5.0
JB-1 31 0.07 554 554 616 0.8 0.8 2.4 N 0.2,0.5,0.7,1.0,1.5,2.0,3.0,4.0
JB-2 31 0.12 554 554 616 0.8 0.8 2.4 N 0.2,0.5,0.7,1.0,1.5,2.0,3.0,5.0
JC-1 27 0.08 554 554 616 0.8 0.8 2.4 N 0.2,0.3,0.5,0.7,1.0,1.5,2.0,3.0,4.0
JC-2 33 0.11 554 554 616 0.8 0.8 2.4 N 0.2,0.5,0.7,1.0,1.5,2.0,3.0,4.0,5.0
Shafaei et al. C-1 21 0.15 460 460 350 1.1 1.1 2.0 Y 1.0,2.2,3.5
[77] C-2 23 0.14 460 460 350 1.1 1.1 2.0 N 1.0,1.4,1.7,2.2,2.7,4.5
C-3 22 0.14 460 460 350 1.1 1.1 2.0 N 1.0,1.4,1.7,2.2,2.7,3.5,4.5
SC2-1 23 0.14 460 460 350 1.1 1.1 2.0 N 1.4,1.7,2.2,2.7,3.5
SC2-2 25 0.13 460 460 350 1.1 1.1 2.0 N 1.0,2.2,2.7,4.5
SC2-3 27 0.12 460 460 350 1.1 1.1 2.0 N 1.0,1.4,2.2,2.7,3.5,4.5
SC3-1 21 0.15 460 460 350 1.1 1.1 2.0 N 1.0,1.7,2.2,2.7,3.5,4.5

Meas et al. LS 43 0.15 520 520 520 1.3 1.3 2.0 Y 1.0,3.0
[78] LSU 43 0.15 520 520 520 1.3 1.3 2.0 Y 1.5
VR3 35 0.30 490 490 370 0.9 0.9 2.4 Y 1.0,3.0
VR3U 35 0.30 490 490 370 0.9 0.9 2.4 Y 1.5,3.0
VR4 35 0.30 490 490 370 0.9 0.9 2.4 Y 1.0,3.0
VR4UU 35 0.30 490 490 370 0.9 0.9 2.4 Y 1.5,3.0
Mobin et al. JNC 30 0.15 420 420 390 0.8 0.8 3.4 Y 1.0,2.0,3.5,4.8
[79] JSCC-1 29 0.00 420 420 390 0.8 0.8 3.4 Y 1.0,2.0,3.5,3.9
JSCC-2 29 0.08 420 420 390 0.8 0.8 3.4 Y 1.0,2.0,3.5,4.2
JSCC-3 29 0.15 420 420 390 0.8 0.8 3.4 Y 1.0,2.0,3.5,4.1
Alaee [33] IN80 80 0.00 564 535 563 0.6 0.3 2.0 Y 1.0,2.0,3.0,4.0
IH80 80 0.00 712 592 563 0.6 0.3 2.6 Y 1.0,2.0,3.0,4.0
IH80A 80 0.30 712 592 563 0.6 0.3 2.6 Y 1.0,2.0,3.0,4.0
IN100 100 0.00 564 535 563 0.6 0.3 2.0 Y 1.0,2.0,3.0,4.0
IH100 100 0.00 712 592 563 0.6 0.3 2.6 Y 1.0,2.0,3.0,4.0
IH60 60 0.00 708 592 464 0.5 0.5 2.6 Y 1.0,2.0,3.0,4.0
IH60A 60 0.30 708 592 464 0.5 0.5 2.6 Y 1.0,2.0,3.0,4.0
EN80 80 0.00 564 558 563 0.6 0.3 2.0 Y 1.0,2.0,3.0,4.0
EH80 80 0.00 712 638 563 0.6 0.3 2.6 Y 1.0,2.0,3.0,4.0
EH80A 80 0.30 712 638 563 0.6 0.3 2.6 Y 1.0,2.0,3.0,4.0
EH60 60 0.00 708 638 464 0.4 0.4 2.6 Y 1.0,2.0,3.0,4.0
EH60A 60 0.30 708 638 464 0.4 0.4 2.6 Y 1.0,2.0,3.0,4.0
Ridwan [80] BCJ-CS-ATest1 34 0.08 512 512 581 1.0 1.0 2.7 Y 1.0,2.0,3.0
BCJ-CS-ATest2 31 0.08 512 512 581 1.0 1.0 2.7 Y 3
BCJ-SS-4 46 0.08 512 512 581 1.0 1.0 2.7 Y 2
BCJ-SS-8 25 0.08 512 512 581 1.0 1.0 2.7 Y 1.0,2.0,3.0
Behnam [81] S1-BC1 36 0.13 522 522 485 0.9 0.9 1.5 Y 1.0,2.0,3.0,4.0,5.0
S2-BC1.5 36 0.13 522 522 485 0.9 0.9 1.9 Y 0.5,1.0,2.0,3.0,4.0,5.0
Zhang [82] KJ0 38 0.00 520 520 500 1.0 1.0 2.0 N 0.2,0.7,1.0,2.0,4.0
KJ1 36 0.00 520 520 500 1.0 1.0 2.0 Y 1.2,2.0,2.8
KJ2 29 0.00 520 520 500 1.0 1.0 2.0 Y 0.8,1.0,1.6,2.4,3.2
KJ3 32 0.00 520 520 500 1.0 1.0 2.0 Y 0.2,0.8,1.6,2.8,3.6
KJ4 35 0.00 520 520 500 1.0 1.0 2.0 Y 1.2,1.6,2.0,2.8
KJ5 31 0.00 520 520 500 0.6 0.6 2.0 Y 0.2,0.8,1.2,2.0,2.4,3.2,3.6
KJ6 31 0.00 520 520 500 1.0 1.0 2.0 Y 0.2,0.8,1.2,2.0,2.8,3.2
KJ8 32 0.00 520 520 500 1.0 1.0 1.3 Y 0.8,1.6,2.4,2.8
KJ9 32 0.00 520 520 500 0.5 0.5 2.0 Y 0.8,1.2,2.8,3.2,3.6
(BD700)
(continued on next page)

15
M. Hamidia et al. Measurement 205 (2022) 112195

Table A1 (continued )
KJ9 33 0.00 520 520 500 1.0 1.0 2.0 Y 0.2,0.8,2.0,2.8,3.6
(CW600)
Lee and A5 70 0.05 690 690 790 1.0 1.0 2.4 Y 4
Chang [83] B3 78 0.05 690 690 790 1.4 1.4 2.4 Y 4
B5 64 0.05 690 690 790 1.4 1.4 2.4 Y 4
C3 65 0.05 690 690 790 1.8 1.8 2.4 Y 4
C5 68 0.05 690 690 790 1.8 1.8 2.4 Y 4
Mogili [84] KJ1 37 0.02 556 556 526 1.0 1.0 2.0 Y 1.2,2.0,3.2,4.0
KJ2 38 0.02 556 556 526 0.7 1.0 2.0 Y 1.2,2.0,3.2,4.0
KJ3 34 0.03 556 556 526 1.0 0.7 2.0 Y 1.2,2.0,3.2,4.0
KJ4 27 0.03 558 552 526 0.6 0.6 3.3 Y 1.2,2.0,3.2,4.0,5.0
KJ5 28 0.03 550 552 526 0.4 0.6 3.3 Y 1.2,2.0,3.2,4.0,5.0
KJ6 30 0.03 550 552 526 0.6 0.4 3.3 Y 1.2,2.0,3.2,4.0,5.0
KJ7 29 0.02 556 556 526 1.0 1.0 1.5 Y 1.2,2.0,3.2,4.0,5.0
KJ8 35 0.01 558 556 526 0.7 1.0 1.5 Y 1.2,2.0,3.2,4.0,5.0

KJ9 32 0.02 558 556 526 1.0 0.7 1.5 Y 1.2,2.0,3.2,4.0,5.0


KJ10 33 0.02 556 556 556 1.0 1.0 2.0 Y 1.2,2.0,3.2,4.0,5.0
KJ11 33 0.02 556 556 526 1.0 1.0 2.0 Y 1.2,2.0,3.2,4.0,5.0
Wentong [85] C-2 29 0.30 540 540 379 1.3 0.8 2.0 Y 0.7,1.0,1.3,1.7,3.5,4.5,5.2
Dehkordi et al.[86] CNS-70 70 0.04 605 439 425 1.0 1.0 1.4 Y 2.2,3.5,4.5,5.3
RCHS-70 70 0.04 605 602 425 1.0 1.0 1.0 Y 2.2,3.5,4.5,5.3,5.4
NS-30 30 0.08 439 449 425 0.9 0.9 1.9 Y 4.8
NS-70 70 0.08 439 449 425 0.9 0.9 1.9 Y 4.8
RHS-30 30 0.08 439 604 425 0.6 0.6 1.4 Y 5.3
RHS-70 70 0.08 439 604 425 0.6 0.6 1.4 Y 5.3
Adibi et al. [87] JRD1 23 0.15 460 460 350 1.1 0.8 2.0 Y 1.0,1.7,3.5
JRD2 23 0.15 460 460 350 1.1 0.8 2.0 N 1.0,1.7,3.5
JRD3 25 0.15 460 460 350 1.1 0.8 2.0 N 1.0,1.7,3.5
Shen et al. J1 44 0.20 443 460 450 0.5 0.5 1.8 Y 0.7,2.1,3.5,4.9
[88] J2 46 0.20 443 460 450 0.5 0.5 1.8 Y 0.7,2.1,3.5,4.9,5.6
J3 47 0.20 443 460 450 0.5 0.5 1.8 Y 0.7,2.1,3.5,4.9,5.6
Abuzeid et al. [89] CB01 34 0.00 400 400 400 0.9 0.7 1.5 N 2.8
CB02 34 0.00 400 400 400 0.9 0.7 1.5 Y 2.8
CB03 55 0.00 400 400 400 0.9 0.7 1.5 N 2.8
CB04 74 0.00 400 400 400 0.9 0.7 1.5 N 3.7
1
TRiJ: Transverse Reinforcement exists in the Joint?

10. Model availability Appendix B

The ultimate trained ML-based models and the extracted visual


features of this study are provided on GitHub. Model design

CRediT authorship contribution statement The design of the ML models and hyperparameters selection of this
study has been carried out via GridSearchCv tools in the Sklearn library
Mohammadjavad Hamidia: Conceptualization, Methodology, Su­ [43]. Hyperparameters are elements that are not set and improved
pervision, Writing – review & editing, Project administration, Re­ through a learning procedure. A typical example of hyperparameters is
sources. Sina Mansourdehghan: Software, Data curation, Writing – the number of estimators in the random forest model. It is recommended
original draft, Formal analysis, Visualization. Amir Hossein Asjodi: to search the hyper-parameter space for the best cross-validation score
Project administration, Conceptualization, Writing – review & editing, [43]. To this end, a wide range of possible hyperparameters, which are
Investigation. M. Kiarash Dolatshahi: Conceptualization, Validation, selected based on designer experience and several trial-and-error pro­
Visualization, Writing – review & editing, Writing – original draft. cess results, is fed to a GridSearchCv model, and the model will be
trained and validated with every permutation of entry parameters and
Declaration of Competing Interest the mean accuracy calculated from 10-fold cross-validation over
training dataset is reported as the score of that set. As an example, for the
The authors declare that they have no known competing financial random forest regression model, the following ranges are considered as
interests or personal relationships that could have appeared to influence entry values to the grid search method:
the work reported in this paper. n_estimators (the number of trees in the forest) = [60, 70, 80, 90,
100, 120, 140, 150, 160, 180, 200, 250].
Data availability
• min_sample_split (The minimum number of samples required to split
Data will be made available on request. an internal node) = [1–5]
• min_sample_leaf (The minimum number of samples required to be at
Acknowledgement a leaf node; otherwise, the split will not happen.) = [1,3,4,5]
• max_depth (The maximum depth of the tree. If “None”, then nodes
The authors acknowledge Amirhossein Ganjizadeh for reading the are expanded until all leaves are pure or until all leaves contain less
manuscript and making comments. than min_samples_split samples.) = [None, 2, 3, 5]

Appendix A These possible numbers lead to 960 different sets of hyper-


parameters. For each case, a model is created and validated then all
See Table A1. 960 models are sorted based on their cross-validation scores.

16
M. Hamidia et al. Measurement 205 (2022) 112195

GridSearchCv assigns an attribute called best_params_ to the selected [13] S. Mansourdehghan, K.M. Dolatshahi, A.H. Asjodi, Data-driven damage assessment
of reinforced concrete shear walls using visual features of damage, J. Build. Eng. 53
model, which contains the best hyperparameters. In the following, the
(2022), 104509, https://doi.org/10.1016/j.jobe.2022.104509.
best_params_ attribute associated with the third scenario is listed: [14] A.H. Asjodi, K.M. Dolatshahi, A. Ebrahimkhanlou, Spatial analysis of damage
evolution in cyclic-loaded reinforced concrete shear walls, J. Build. Eng.. (2022),
• RF model: 104032, https://doi.org/10.1016/j.jobe.2022.104032.
[15] A.H. Asjodi, K.M. Dolatshahi, Peak drift ratio estimation for unreinforced masonry
1. n_estimators = 80 walls using visual features of damage, Bull. Earthq. Eng. (2022), https://doi.org/
2. min_sample_split = 2 10.1007/s10518-022-01523-8.
3. min_sample_leaf = 1 [16] S. Dorafshan, R.J. Thomas, M. Maguire, Comparison of deep convolutional neural
networks and edge detectors for image-based crack detection in concrete, Constr
4. max_depth = None Build Mater. 186 (2018) 1031–1045, https://doi.org/10.1016/j.
• GB model: conbuildmat.2018.08.011.
1. n_estimators = 90 [17] A.M.A. Talab, Z. Huang, F. Xi, L. Haiming, Detection crack in image using Otsu
method and multiple filtering in image processing techniques, Optik (Stuttg). 127
2. min_sample_split = 4 (2016) 1030–1033, https://doi.org/10.1016/j.ijleo.2015.09.147.
3. min_sample_leaf = 1 [18] Y.J. Cha, W. Choi, O. Büyüköztürk, Deep learning-based crack damage detection
4. max_depth = 5 using convolutional neural networks, computer-aided civil and infrastructure,
Engineering. 32 (2017) 361–378, https://doi.org/10.1111/mice.12263.
5. subsample (The fraction of samples to be used for fitting the indi­ [19] R.-T. Wu, M.R. Jahanshahi, Deep convolutional neural network for structural
vidual base learners.) = 1 dynamic response estimation and system identification, J. Eng. Mech. 145 (2019)
6. learning_rate (scales the contribution of each tree) = 1 4018125, https://doi.org/10.1061/(ASCE)EM.1943-7889.0001556.
[20] B.K. Oh, Y. Park, H.S. Park, Seismic response prediction method for building
• SVR model:
structures using convolutional neural network, Struct. Control Health Monit. 27
1. Kernel (Specifies the kernel type to be used in the algorithm) = “rbf” (2020), https://doi.org/10.1002/stc.2519.
2. Gamma (Kernel coefficient) = 0.1 [21] X. He, T. Wang, K. Wu, H. Liu, Automatic defects detection and classification of low
3. C (Regularization parameter. There is always a trade-off between C carbon steel WAAM products using improved remanence/magneto-optical imaging
and cost-sensitive convolutional neural network, Measurement (Lond). 173 (2021),
and gamma setting the regularization and curvature of “rbf” kernel https://doi.org/10.1016/j.measurement.2020.108633.
respectively.) = 10 [22] J. Xing, M. Jia, A convolutional neural network-based method for workpiece
• ElasticNet model: surface defect detection, Measurement (Lond). 176 (2021), https://doi.org/
10.1016/j.measurement.2021.109185.
1. alpha (Constant that multiplies the penalty terms in objective func­ [23] J. Valença, D. Dias-Da-Costa, E. Júlio, H. Araújo, H. Costa, Automatic crack
tion) = 0.1 monitoring using photogrammetry and image processing, Measurement (Lond). 46
2. l1_ratio (The ElasticNet mixing parameter) = 0.01 (2013), https://doi.org/10.1016/j.measurement.2012.07.019.
[24] C. Bernstone, A. Heyden, Image analysis for monitoring of crack growth in
hydropower concrete structures, Measurement (Lond). 42 (2009), https://doi.org/
In the case of the ANN regression model, the GrideSearchCv does not 10.1016/j.measurement.2009.01.007.
provide optimization tools. Therefore, the parameters of the ANN ar­ [25] A. Reddy, V. Indragandhi, L. Ravi, V. Subramaniyaswamy, Detection of Cracks and
damage in wind turbine blades using artificial intelligence-based image analytics,
chitecture, including the number of layers, and the number of nodes in Measurement (Lond). 147 (2019), https://doi.org/10.1016/j.
each layer, should be selected through a for-loop trial and error pro­ measurement.2019.07.051.
cedure. The selected attributes for the ANN regression are listed below: [26] T. Ni, R. Zhou, C. Gu, Y. Yang, Measurement of concrete crack feature with android
smartphone APP based on digital image processing techniques, Measurement
(Lond). 150 (2020), https://doi.org/10.1016/j.measurement.2019.107093.
1. number of layers = 2 dense layers [27] A.H. Asjodi, M.J. Daeizadeh, M. Hamidia, K.M. Dolatshahi, Arc Length method for
2. number of nodes = 256 (first layer) and 128 (second layer) extracting crack pattern characteristics, Struct Control Health Monit. 28 (2021)
3. activation function = “relu” 1–14, https://doi.org/10.1002/stc.2653.
[28] X. Tan, Y. Bao, Measuring crack width using a distributed fiber optic sensor based
on optical frequency domain reflectometry, Measurement (Lond). 172 (2021),
References https://doi.org/10.1016/j.measurement.2020.108945.
[29] J. Chen, F. Xiong, Y. Zhu, H. Yan, A crack detection method for underwater
concrete structures using sensing-heating system with porous casing, Measurement
[1] M.-H.L.E.M. FEMA, Earthquake Model HAZUS® MH MR4 Technical Manual,
(Lond). 168 (2021), https://doi.org/10.1016/j.measurement.2020.108332.
National Institute of Building Sciences, Washington, DC. (2019).
[30] Q. Feng, J. Cui, Q. Wang, S. Fan, Q. Kong, A feasibility study on real-time
[2] G.G. Deierlein, H. Krawinkler, C.A. Cornell, A framework for performance-based
evaluation of concrete surface crack repairing using embedded piezoceramic
earthquake engineering G.G., in: 2003 Pacific Conference on Earthquake
transducers, Measurement (Lond). 122 (2018), https://doi.org/10.1016/j.
Engineering, 2003.
measurement.2017.09.015.
[3] FEMA P-58, Seismic Performance Assessment of Buildings - methodology, Fema P-
[31] Z. Ismail, Application of residuals from regression of experimental mode shapes to
58-1. 1 (2012).
locate multiple crack damage in a simply supported reinforced concrete beam,
[4] D.C. Washington, FEMA 307 EVALUATION OF EARTHQUAKE DAMAGED
Measurement (Lond). 45 (2012), https://doi.org/10.1016/j.
CONCRETE AND MASONRY WALL BUILDINGS Technical Resources The
measurement.2012.03.006.
Partnership for Response and Recovery, 1998.
[32] M. Hamidia, S. Mansourdehghan, A.H. Asjodi, K.M. Dolatshahi, Machine learning-
[5] M. Hamidia, A. Ganjizadeh, Post-earthquake damage evaluation of non-ductile RC
based seismic damage assessment of non-ductile RC beam-column joints using
moment frames using surface crack patterns, Struct Control Health Monit (2022)
visual damage indices of surface crack patterns, Structures 45 (2022) 2038–2050,
e3024.
https://doi.org/10.1016/j.istruc.2022.09.010.
[6] M. Hamidia, A. Ganjizadeh, Computer vision-based automated stiffness loss
[33] S.G. Paal, J.-S. Jeon, I. Brilakis, R. DesRoches, Automated damage index estimation
estimation for seismically damaged non-ductile reinforced concrete moment
of reinforced concrete columns for post-earthquake evaluations, J. Struct. Eng. 141
frames, Bull. Earthq. Eng. (2022), https://doi.org/10.1007/s10518-022-01408-w.
(2015) 4014228, https://doi.org/10.1061/(ASCE)ST.1943-541X.0001200.
[7] H.M. Madani, K.M. Dolatshahi, Strength and stiffness estimation of damaged
[34] D. Lattanzi, G.R. Miller, M.O. Eberhard, O.S. Haraldsson, Bridge column maximum
reinforced concrete shear walls using crack patterns, Struct Control Health Monit.
drift estimation via computer vision, J. Comput. Civil Eng. 30 (2016) 4015051,
27 (2020) 1–18, https://doi.org/10.1002/stc.2494.
https://doi.org/10.1061/(ASCE)CP.1943-5487.0000527.
[8] M. Hamidia, A. Ganjizadeh, K.M. Dolatshahi, Peak drift ratio estimation for RC
[35] Y. Jemaa, Seismic behaviour of deficient exterior RC beam-column joints, UK:
moment frames using multifractal dimensions of surface crack patterns, Eng Struct.
University of Sheffield, 2013. PhD Thesis.
255 (2022), 113893, https://doi.org/10.1016/j.engstruct.2022.113893.
[36] M.S. Ashtiani, Seismic performance of high-strength self-compacting concrete in
[9] A. Farhidzadeh, E. Dehghan-Niri, A. Moustafa, S. Salamone, A. Whittaker, Damage
reinforced concrete structures, University of Canterbury, New Zealand, 2013. PhD
assessment of reinforced concrete structures using fractal analysis of residual crack
Thesis.
patterns, Exp Mech. 53 (2013) 1607–1619, https://doi.org/10.1007/s11340-013-
[37] N. Zhang, Seismic performance and shear strength of reinforced concrete beam-
9769-7.
column knee joints, University of Hong Kong, China, 2017. PhD Thesis.
[10] A. Rezaie, A.J.P. Mauron, K. Beyer, Sensitivity analysis of fractal dimensions of
[38] S. Mogili, Seismic Behaviour and Strength of Reinforced Concrete Beam-Column
crack maps on concrete and masonry walls, Autom Constr. 117 (2020), 103258,
Knee Joints, University of Hong Kong, China, 2019. PhD Thesis.
https://doi.org/10.1016/j.autcon.2020.103258.
[39] H. Behnam, Seismic performance and failure mechanisms of reinforced concrete
[11] J. Carrillo, D. Dominguez, N. Prado, Seismic damage index based on fractal
wide beam-column connections, Hong Kong University of Science and Technology,
dimension of cracking on thin reinforced concrete walls, ACI Struct J. 114 (2017).
2017. Ph.D. Dissertation,.
[12] M. Hamidia, M. Afzali, S. Jamshidian, Post-earthquake residual stiffness estimation
[40] M. Mohri, A. Rostamizadeh, A. Talwalkar, Foundations of machine learning, MIT
using crack maps of reinforced concrete columns, Struct. Concr. (2022). Submitted
press, 2018.
for publication.

17
M. Hamidia et al. Measurement 205 (2022) 112195

[41] C. de Mol, E. de Vito, L. Rosasco, Elastic-net regularization in learning theory, [53] L.M. Megget, Cyclic behaviour of exterior reinforced concrete beam-column joints,
J Complex. 25 (2009), https://doi.org/10.1016/j.jco.2009.01.002. Bull. New Zealand Soc. Earthq. Eng. 7 (1974) 27–47.
[42] H. Drucker, C.J.C. Surges, L. Kaufman, A. Smola, V. Vapnik, Support vector [54] Y.S. Keong, Prestressed Concrete Beam-column Joints, University of Canterbury,
regression machines, in: Adv Neural Inf Process Syst, 1997. New Zealand, 1978. Master’s Thesis.
[43] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, [55] B.R. Birss, The elastic behaviour of earthquake resistant reinforced concrete
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, interior beam-column joints, University of Canterbury, New Zealand, 1978.
D. Cournapeau, M. Brucher, M. Perrot, É. Duchesnay, Scikit-learn: Machine Master’s Thesis.
learning in Python, Journal of Machine Learning Research. 12 (2011). [56] C.W. Beckingsale, Post elastic behaviour of reinforced concrete beam-column
[44] A. Geron, Hands on Machine Learning with Scikit Learn, Keras, 2019. joints, University of Canterbury, New Zealand, 1980. PhD Thesis.
[45] S.M. Lundberg, S.I. Lee, A unified approach to interpreting model predictions, in: [57] J.R. Milburn, Behaviour of reinforced concrete beam-column joints designed to
Adv Neural Inf Process Syst, 2017. NZS 3101. Master’s Thesis, University of Canterbury, New Zealand, 1982.
[46] S. Mangalathu, S.H. Hwang, J.S. Jeon, Failure mode and effects analysis of RC [58] M.R. Ehsani, J.K. Wight, Behavior of external reinforced concrete beam to column
members based on machine-learning-based SHapley Additive exPlanations (SHAP) connections subjected to earthquake type loading, Department of Civil Engineering
approach, Eng Struct. 219 (2020), https://doi.org/10.1016/j. the University of Michigan, Michigan, United States, 1982. Report UMEE 825R;.
engstruct.2020.110927. [59] P.R. Philleo, D.P. Abrams, Scale relationships of concrete beam-column joints,
[47] P. Alaee, Experimental and analytical investigations on seismic behavior of National Science Foundation Research University of Colorado, Boulder, United
reinforced concrete frames and members using high-strength materials. PhD States, 1984.
Thesis. Singapore: University of Nanyang;, 2016. Doi: 10.32657/10356/69894. [60] D. Ruitong, R. Park, A comparison of the behaviour of reinforced concrete beam-
[48] L.M. Megget, R. Park, Reinforced concrete exterior beam-column joints under column joints designed for ductility and limited ductility, Bull. New Zealand Soc.
seismic loading, NZ Eng. 26 (1971) 341–353. Earthq. Eng. 21 (1988) 255–278, https://doi.org/10.5459/200Bbnzsee.21.4.255-
[49] G.W. Renton, The behaviour of reinforced concrete beam-column joints under 278.
cyclic loading, University of Canterbury, New Zealand, 1972. Master’s Thesis. [61] S.P. Pessiki, C.H. Conley, P. Gergely, R.N. White, Seismic behavior of lightly-
[50] B.J. Smith, Exterior Reinforced Concrete Joints with Low Axial Load Under Seismic reinforced concrete column and beam-column joint details. Technical Report
Loading, University of Canterbury, New Zealand, 1972. Master’s Thesis. NCEER-90-0014;, 1990.
[51] R. Park, K.J. Thompson, Behaviour of a prestressed concrete interior beam-column [62] F.F. Alameddine, Seismic design recommendations for high-strength concrete
assembly under cyclic loading, Bull. New Zealand Soc. Earthq. Eng. 6 (1973) beam-to-column connections, University of Arizona, United States, 1990. PhD
158–169, https://doi.org/10.5459/bnzsee.6.4.158-169. Thesis.
[52] S.M. Uzumeri, M. Seckin, Behaviour of reinforced concrete beam-column joints [63] N. Chutarat, Control of plastic mechanism in corner reinforced concrete beam-
subjected to slow load reversals. Publication No. 74-05. Department of Civil column joints using headed reinforcement bars, University of Syracuse, United
Engineering, University of Toronto, (1974). States, 2001. PhD Thesis.

18

You might also like