You are on page 1of 8

Structures 36 (2022) 927–934

Contents lists available at ScienceDirect

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

Machine learning applications for assessment of dynamic progressive


collapse of steel moment frames
Yan Fei Zhu a, *, Yao Yao a, b, *, Ying Huang a, Chang Hong Chen b, Hui Yun Zhang a,
Zhaohui Huang a
a
School of Civil Engineering, Xi’an University of Architecture and Technology, Xi’an 710055, China
b
School of Mechanics and Civil Engineering, Northwestern Polytechnical University, Xi’an 710129, China

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

Keywords: A highly efficient implementation of machine learning (ML) framework is developed for assessing the dynamic
Progressive collapse increase factor (DIF) used in nonlinear static analysis (pushdown). Analysis datasets with a total of 3992 samples
Dynamic increase factor consisting of training (70%) and testing (30%) are generated by the Monte Carlo simulation that carried out by
Dynamic response
Sap2000 program application programming interface (API) link with Python in PyCharm software. The gener­
Steel
ated datasets are evaluated by the correlation matrix and relative feature importance of features that consist of
Machine learning
damping ratio (Zeta) range, period of vertical vibration (Tv), ratio of duration of a column removal to Tv (Tau),
ratio of factored moment to yielding moment (MMR) and ratio of factored rotation to yielding rotation (TTR).
The current study incorporates implementation of K-nearest neighbors algorithm (KNN), extreme gradient
boosting (XGBoost), back-propagation neural network (BPNN), and one-dimensional convolutional neural
network (1DCNN). The results confirm the ability of the developed framework to efficiently implement
regression analysis for the DIF.

1. Introduction [9–13] and the robustness of steel modular buildings [14–16]. The
alternate load paths (ALP) method has been widely applied to the
Progressive collapse is defined as the spread of an initial local failure assessment of progressive collapse under column removal scenarios.
to a chain failure reaction that leads to disproportionate partial or total Other than static analysis, more attention has been paid to the study of
collapse of a structure [1,2]. The initial event could be terrorist attacks, dynamic progressive collapse of steel structures. Dynamic experiments
vehicular collisions, gas explosions, earthquakes, fires, or human errors and finite element (FE) analysis of dynamic progressive collapse in steel
in design and construction [3,4]. Extreme progressive collapse events frames are widely performed [17–20].
indicate the importance of resistance against progressive collapse [4], Although nonlinear dynamic analysis with good accuracy is one of
which consisting of the Ronan Point in London caused by gas explosion, the best approaches for progressive collapse assessment, the analysis
the Alfred P. Murrah Federal Building in Oklahoma (1995) caused by processing is complex and time consuming. Nonlinear static analysis
terrorist truck bombing, the twin towers of the World Trade Center in (pushdown) incorporating dynamic increase factor (DIF) can be utilized
New York (2001) caused by terrorist attack and fire, the Rana Plaza in to assess structural responses. Thus, different DIF formulas are proposed
Savar (2013) caused by overload, and the Plasco Building in Tehran by researchers, and various of influence parameters are considered in
(2017) caused by fire. those proposed formulas [4], such as ductility demand of plastic rotation
Steel structures are commonly adopted for mitigation of progressive divided by yield rotation [21–23], damping ratio [24], post-elastic
collapse, which attract the attention of structural engineers and re­ stiffness ratio [21,25], applied loads (energy demand) and energy ca­
searchers. Extensive research has been carried out to discuss and pacity [26,27]. However, for multi-factor evaluation problem, machine
investigate progressive collapse of steel structures, involving the per­ learning (ML) is an efficient and convenient implementation tool.
formance of beam-column joint [5–8], the failure mode of steel frames Due to the rapid development of machine learning (ML), structural

* Corresponding authors at: School of Civil Engineering, Xi’an University of Architecture and Technology, Xi’an 710055, China.
E-mail addresses: zhuyanfei@xauat.edu.cn (Y.F. Zhu), yaoy@nwpu.edu.cn (Y. Yao), changhong.chen@nwpu.edu.cn (C.H. Chen), leslie_zhanghuiyun@xauat.edu.
cn (H.Y. Zhang).

https://doi.org/10.1016/j.istruc.2021.12.067
Received 17 July 2021; Received in revised form 17 December 2021; Accepted 23 December 2021
Available online 30 December 2021
2352-0124/© 2021 Institution of Structural Engineers. Published by Elsevier Ltd. All rights reserved.
Y.F. Zhu et al. Structures 36 (2022) 927–934

Fig. 1. Dynamic progressive collapse analysis using machine learning.

performance and condition can be predicted and assessed by the ML increase factor (DIF) used in nonlinear static analysis (pushdown),
models [28]. The machine learning applications in building structural which can be utilized to efficiently assess structural maximum responses
design and performance assessment consist of estimating structural under a sudden column removal scenario. Different influence parame­
response [29–32], calculating the ultimate bearing capacity of compo­ ters are incorporated in the obtained DIF based on the proposed machine
nents [33], recognizing patterns in structural health monitoring data learning framework, which consist of damping ratio, period of vertical
[34,35] and predicting the variability of material properties [36–38]. vibration, duration of a column removal, applied loads and ductility
Although research on construction problems has been carried out in the demand. The current study presents implementation of K-nearest
past decade by the machine learning models, the ML has rarely been neighbors algorithm (KNN), extreme gradient boosting (XGBoost), Back-
applied to design against progressive collapse. Fu [39] assessed the fire propagation neural network (BPNN), and one-dimensional convolu­
induced progressive collapse of steel frames using machine learning tional neural network (1DCNN).
models, which consisting of Decision Tree, KNN and Neural Network.
Esfandiari and Urgessa [40] presented progressive collapse analysis 2. Data preparation for assessing the DIF
based on an efficient structural optimization technique, called DMPSO.
Therefore, it is essential to develop a machine learning framework for 2.1. Data generation
the assessment of dynamic progressive collapse.
The current study focuses on developing a highly efficient machine The whole analysis process of dynamic progressive collapse based on
learning (ML) framework, and provides a new analytical idea and machine learning method is shown in Fig. 1. The process consists of data
analytical tool for the rapid implementation of DIF evaluation. The preparation, building machine learning models along with hyper­
framework enables the achievement of a regression analysis for dynamic parameter optimization and performance evaluation, and applications

Fig. 2. Illustration of the steps to obtain the DIF.

928
Y.F. Zhu et al. Structures 36 (2022) 927–934

Table 1
Statistics of the datasets.
Attribute Abbreviation Unit Mean Std Min Max

Damping ratio Zeta — 0.030 0.012 0.010 0.050


The period of Tv s 0.659 0.086 0.481 0.944
vertical vibration
Ratio of duration of Tau — 0.154 0.083 0.010 0.300
a column removal
to Tv
Ratio of factored MMR — 1.008 0.053 0.745 1.060
moment to
yielding moment
Ratio of factored TTR — 1.756 0.638 0.745 2.999
rotation to
yielding rotation
Dynamic increase DIF — 1.288 0.100 1.142 1.789
factor

Fig. 4. Correlation matrix of features in datasets presented in Table 1.

Fig. 3. The DIF histogram display.

of the machine learning model. Based on the basic principle of structural


dynamics and researches of dynamic progressive collapse as described in
Introduction section, inputs consists of damping ratio (Zeta), period of
vertical vibration (Tv), ratio of duration of a column removal to Tv (Tau,
Fig. 5. Relative feature importance of features in datasets presented in Table 1.
τ = Δt/Tv ), ratio of factored moment to yielding moment (MMR,
MMR = Ml /My ), and ratio of factored rotation to yielding rotation (TTR,
TTR = θl /θy ). The damping ratio (Zeta) mainly affects the maximum support reaction method, and given by:
dynamic response of a steel structure. The vertical elastic stiffness of a FSR,NLP
DIF = (1)
steel frame structure can be presented by the first-mode vertical natural FSR,NL
vibration period (Tv). Different failure modes of a column caused by
where FSR,NLP is the support reaction obtained by Step3 when the
different initial events are presented by ratio of duration of a column
structural responses from the pushdown analysis best match those from
removal to Tv (Tau). The factored moment (Ml )and rotation (θl ) are
the nonlinear dynamic analysis in Step2as recommended by Zhu, et al.
indexes of external load energy demand, and the yielding moment (My )
[27], uNLP = uND,max ; FSR,NL is obtained by Step1 under the applied load
and rotation (θy ) are indexes of energy capacity. The output is dynamic
combination of 1.2DL + 0.5LL, where DL is dead load and LL is live load
increase factor (DIF). Typically, labeled dataset are split into training
as described in DoD [1] and GSA [2].
(70%) and testing (30%) datasets by the function of train_test_split im­
ported from model_selection in Sklearn module. The testing subset is to
develop the machine learning model. The testing subset is for evaluating 2.2. Data evaluation
performance of the proposed model. In the current study, Monte Carlo
simulation is automatically carried out using the SAP2000 program The statistics of the datasets with a total of 3992 samples are listed in
application programming interface (API) incorporated with the Python. Table 1, which calculated from the above analysis process. The DIF
The trained machine learning model can be then used repeatedly and histogram are described and shown in Fig. 3. It is worth reminding that
quickly in other assessment of structural progressive collapse. the range of values of DIF is 1.142 to 1.789. The maximum value of DIF
The major steps to investigate the output parameter DIF are illus­ (DIFmax = 1.789) obtained by the steps as shown in Fig. 2. is close to, but
trated in Fig. 2. To incorporate the material yielding on the structural less than, the constant value (DIF = 2.0) used in the traditional analysis
dynamic response, plastic hinges with default hinge properties obtained and design. Also, the maximum and minimum values are close to the
from ASCE4-17 [41] are utilized in SAP2000, as shown on the left side of value recommended in guidelines [1,2] and in references [22,24,25]. It
Fig. 2. For a given column removal scenario, the DIF is calculated by is noted that Zeta, Tau, MMR, TTR, DIF are dimensionless. The unit of Tv
is second (s). The value of each feature parameter is variable over a

929
Y.F. Zhu et al. Structures 36 (2022) 927–934

Fig. 6. Visualization of the XGBoost model architecture.

certain range so as to generate large volumes of analysis data. The value Hyperparameter optimization of the KNN is performed by the grid
of Zeta ranges from 0.010 to 0.050, Tv ranges from 0.481 to 0.944, and search method in the current study. The grid search provided by Grid­
Tau ranges from 0.01 to 0.3. The variable factored moment and rotation SearchCV exhaustively generates candidates from a grid of parameter
are obtained by adjusting the sections of members and the applied load. values specified with the param_grid parameter.
The correlation matrix and relative feature importance of features in
datasets presented in Table 1 are shown in Fig. 4 and Fig. 5 respectively.
The output variable DIF is highly correlated with the MMR (-0.799) and 3.2. Extreme gradient boosting (XGBoost)
TTR (-0.714). The Zeta and Tau show little effect on the DIF, and the
corresponding correlation coefficients are − 0.168 and − 0.135, respec­ Extreme gradient boosting (XGBoost) library supported with
tively. Moreover, the influence of structural vertical stiffness on the DIF “xgboost” package in python provides an efficient and effective imple­
should not be neglected. mentation of the gradient boosting framework. Just like any other ma­
chine learning models, the XGBoost model can be built-up and
3. Machine learning algorithms implemented by the following steps: 1) Prepare data; 2) Develop a
XGBoost model imported from “bayesian-optimization” package; 3)
3.1. K-nearest Neighbors algorithm (KNN) Predict the input values; 4) Evaluate with the prediction performance
metrics; 5) Apply the XGBoost model to structural design against pro­
Supervised neighbors-based learning consists of classification for gressive collapse. The critical parameters of the model algorithm are
data with discrete labels, and regression for data with continuous labels. max_depth, learning_rate, n_estimators, gamma, min_child_weight, max_del­
The nearest neighbor method is to find a predefined number of training ta_step, subsample, colsample_bytree, reg_alpha, reg_lambda. Default set­
data closest in distance to the new point and predict the label. Kneigh­ tings are recommended for other parameters, such as booster, base_score,
borsRegressor class of Sklearn Neighbors library is imported in the KNN random_state and so on. Hyperparameter optimization of the XGBoost is
model to predict DIF. The DIF output of the KNN is based on the mean or performed by the bayesian optimization framework and grid search
median of the k-nearest neighbors in the feature space. The critical pa­ method. The bayesian optimization is based on bayesian inference and
rameters of this class are n_neighbors, weights and p. Default values are gaussian process, and the maximum value of an unknown function can
used for other parameters, such as algorithm, leaf_size and metric. be found. The XGBoost model architecture for the first tree is visualized
using a convenient function plot_tree imported from XGBoost in the

Fig. 7. Visualization of the BPNN model architecture.

930
Y.F. Zhu et al. Structures 36 (2022) 927–934

Fig. 8. Visualization of the 1DCNN model architecture.

current study, as shown in Fig. 6. The features of MMR and Tv are layer, Conv1D, MaxPooling1D, Flatten and Dense [45]. In the con­
contained in the first tree diagram. Other features of Zeta, Tau and TTR volutional layer, the multiplication is performed between input data and
are represented in other tree diagrams. weights (filter or kernel). The max pooling layer is used to select the
maximum element from the region of feature map and reduce the size of
3.3. Back-propagation neural network (BPNN) input data. Once the pooled featured map is obtained, flatten layer
transforms entire pooled feature map matrix (two-dimensional array)
A neural network consists of multiple interconnected and combined into a one-dimensional array. After flattening, two fully connected
neurons. The neurons interact with each other via weighted connections layers (Dense) are added to the 1DCNN model. The dense layer involves
and are logically arranged into two or more layers. The neural network input layer and output layer. Hyperparameter optimization of the
model adopted in the current study utilizes the back-propagation 1DCNN model is similarly performed by the grid search method. Opti­
learning algorithm, namely back-propagation neural network (BPNN) mization results are shown in Fig. 9. The red results shown at the top
[42]. The adopted BPNN model architecture after hyperparameter position in the radar plot are the optimized analysis results. The possible
optimization is visualized and shown in Fig. 7 consisting of 1 input layer, values for val_filter are set to be 2, 4, 8, 16, 32, and output_dim are set to
5 hidden layers and 1 output layer. The “keras” package in python be 4, 8, 16, 32, 64. Thus, there are 5 × 5 = 25 combinations of the two
provides an efficient and effective implementation of the BPNN frame­ hyperparameters (val_filter and output_dim) as shown in Fig. 9 (a). The
work. A convenient tool of NNSVG is utilized to design and visualize possible values for batch_size are set to be 8, 16, 32, 64, 128, and epochs
architecture of neural network. Hyperparameter optimization of the are set to be 300, 600, 1200, 1800. Thus, there are 4 × 5 = 20 combi­
BPNN model is performed by the grid search method. The critical pa­ nations of the two hyperparameters (batch_size and epochs) as shown in
rameters of the BPNN model are number of layers, number of neurons in Fig. 9 (b). The keras package in python provides seven options for
each layer of the network, optimizer, epochs and batch_size. Default set­ optimizer, namely: SGD, RMSprop, Adagrad, Adadelta, Adam, Adamax,
tings are recommended for other parameters, such as activation, init_­ Nadam, as shown in Fig. 9 (c). Activation. In the current study, the ef­
mode, learn_rate and so on. fects of nine different activation functions on the analysis results were
analyzed, namely softmax, softplus, softsign, relu, tanh, sigmoid, hard_sig­
moid, linear, as shown in Fig. 9 (d). The critical parameters of the CNN
3.4. One-dimensional convolutional neural network (1DCNN) model are filters = 16, output_dim = 32, batch_size = 8, epochs = 1200,
optimizer = adamax, activation = relu, kernel_size = 2. It is worth
Deep learning has developed as an efficient and high-performance reminding that the output_dim = 32 is for the first fully connected layer
machine learning method to solve a variety of complex problems, as shown in Fig. 8. Default settings are recommended for other param­
especially for image processing. Convolutional neural network (CNN) is eters, such as padding = ’same’, loss = mean squared error (MSE), met­
a deep neural network approach, which has been widely applied in rics = mean absolute error (MAE), etc.
various scenarios [43], such as face recognition, activity recognition etc.
In a classical convolutional neural network (CNN), one or more con­ 4. Results and discussions
volutional layers are followed by a pooling layer and a fully connected
layer [43]. Basically, these layers can be repeated as often as desired Scatter plots of the correlation between the target DIF and the pre­
[44]. In the present study, the 1DCNN model architecture is converted to dicted values obtained by KNN, XGBoost, BPNN and 1DCNN are illus­
dot format and visualized using a convenient function plot_model from trated in Figs. 10-13, respectively. The performance of the proposed
Keras, as shown in Fig. 8. The 1DCNN model is convolved with the Input

931
Y.F. Zhu et al. Structures 36 (2022) 927–934

Fig. 9. Hyperparameter optimization for the 1DCNN model by the grid search method.

Fig. 10. DIF from KNN predictive model and target. Fig. 11. DIF from XGBoost predictive model and target.

machine learning model is evaluated based on the 10-fold cross vali­ error (RMSE), mean absolute error (MAE) and mean absolute percentage
dation method. Machine learning model performance, such as underfit error (MAPE). The detailed calculation formulas are given by
or overfit, is also diagnosed by learning curves in the current study. In

n
addition, a set of 4 key indicators are used to quantify the prediction (y − ̂y )2
2
R = 1− ∑ (2)
performance of the model as recommended by Nguyen, et al. [38], n
(y − y)2
which consisting of coefficient of determination (R2), root mean square

932
Y.F. Zhu et al. Structures 36 (2022) 927–934

obtained by the function of mean_squared_error, mean_absolute_error and


r2_score imported from the Sklearn library. Here R2 ranges from − 1 to +
1. If the always-positive RMS E, MAE and MAPE are close to zero, it
represents better prediction of the machine learning algorithm. Appar­
ently, R and MAPE are dimensionless, while the unit of RMSE and MAE
is same as the output datasets.
Comparison of the different machine learning methods performance
and computation cost for prediction of the DIF are listed in Table 2. The
KNN model parameters are calibrated as: n_neighbors = 4; Weights=’­
distance’; p = 1. The XGBoost model parameters are calibrated as:
n_estimators = 800; max_depth = 6; learning_rate = 0.1; objective=’reg:
squarederror’. The BPNN model parameters are calibrated as: neurons
= 10; layers_num = 6; activation=’relu’; optimizer=’Adamax’; epochs =
300; batch_size = 8. Finally, the 1DCNN model parameters are calibrated
as val_filter = 16; output_dim = 32; kernel_size = 2; pool_size = 2; pad­
ding=’same’; activation=’relu’; optimizer=’Adamax’; epochs = 1200,
batch_size = 8. Then these parameters are used in training and testing the
proposed machine learning model. According to the results as shown in
Fig. 12. DIF from BPNN predictive model and target. Figs. 10-13 and listed in Table 2, the prediction accuracy of the 1DCNN
model with R2 = 0.959, RMSE = 0.020, MAE = 0.012 is the highest, the
BPNN model in the second place, the XGBoost model in the third place,
and the KNN model in the end for assessment of dynamic progressive
collapse. It is noted that the 1DCNN model can make an excellent rating
estimation because the R2 score is high, and the MSE and RMSE values
are close to zero in despite of the higher MAPE value and computation
cost.
The present study provides accurate assessment on the dynamic
progressive collapse response of steel frames based on machine learning
method. The machine learning model that is trained by the proposed
analysis framework provides a new design idea of the engineering
practice to enhance the robustness of steel building against progress
collapse. However, it has the following limitations: a) it does not account
the 3-dimensional effects; b) it does not incorporate the effect of
connection flexibility on the dynamic behavior of structure.

5. Conclusions

Four machine learning models consisting of KNN, XGBoost, BPNN


and 1DCNN are adopted to predict the DIF in the current study. Based on
Fig. 13. DIF from 1DCNN predictive model and target. the analysis results, the following conclusions can be drawn:

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ 1) Machine learning application can provide an efficient and effective


1∑ tool for the assessment of dynamic progressive collapse of steel
n
RMSE = (y − ̂y )2 (3)
n moment frames instead of performing complex numerical simulation
and theoretical analysis.
1∑ 2) The determination coefficient R2 of 1DCNN, BPNN and XGBoost
n
MAE = |y − ̂y| (4)
n model are over 0.9, range from 0.937 to 0.959. The prediction ac­
⃒ ⃒ curacy of the 1DCNN model is the highest, the BPNN model in the
1 ∑⃒⃒y − y ⃒⃒
̂ second place, the XGBoost model in the third place, and the KNN
MAPE = ⃒ y ⃒ × 100 (5)
n model in the end for assessment of dynamic progressive collapse.
where y and ̂y are target and predicted values for the testing datasets, 3) Hyperparameter optimization or tuning of the machine learning
respectively; y represents the mean of the target value; n is the number models can be performed by the grid search method and bayesian
of testing datasets. The key indicators adopted in this work can be easily optimization framework.

Table 2
Comparison of the performance of different machine learning methods for prediction of DIF.
Method Module / Function Hyperparameter Performance indicator Time
R2 RMSE MAE MAPE / (s)
%

KNN Sklearn / KNeighborsRegressor n_neighbors = 4; Weights=’distance’; p = 1 0.859 0.037 0.018 1.378 1


XGBoost XGBoost /XGBRegressor n_estimators = 800; max_depth = 6; learning_rate = 0.1; objective=’reg: 0.937 0.025 0.014 1.065 2
squarederror’
BPNN Keras /Sequential neurons = 10; layers_num = 6; activation=’relu’; optimizer=’Adamax’; epochs = 0.944 0.023 0.014 8.166 76
300; batch_size = 8
1DCNN Keras /Sequential, Conv1D, val_filter = 16; output_dim = 32; kernel_size = 2; pool_size = 2; padding=’same’; 0.959 0.020 0.012 8.245 88
MaxPooling1D activation=’relu’; optimizer=’Adamax’; epochs = 1200, batch_size = 8;

933
Y.F. Zhu et al. Structures 36 (2022) 927–934

4) The performance of the machine learning model requires to be [19] Li H, Cai X, Zhang L, Zhang B, Wang W. Progressive collapse of steel moment-
resisting frame subjected to loss of interior column: Experimental tests. Eng Struct
evaluated by using multiple indicators, such as R2, RMSE, MAE and
2017;150:203–20.
MAPE. [20] Xie F, Gu B, Qian H. Experimental study on the dynamic behavior of steel frames
during progressive collapse. J Constr Steel Res 2021;177:106459.
Declaration of competing interest [21] Tsai M-H, Lin B-H. Dynamic amplification factor for progressive collapse resistance
analysis of an RC building. The Structural Design of Tall and Special Buildings
2009;18(5):539–57.
The authors declare that they have no known competing financial [22] McKay A, Marchand K, Diaz M. Alternate path method in progressive collapse
interests or personal relationships that could have appeared to influence analysis: Variation of dynamic and nonlinear load increase factors. Practice
Periodical on Structural Design and Construction 2012;17(4):152–60.
the work reported in this paper. [23] Qian K, Li B. Research advances in design of structures to resist progressive
collapse. J Perform Constr Facil 2015;29(5):B4014007.
Acknowledgments [24] Mashhadi J, Saffari H. Effects of damping ratio on dynamic increase factor in
progressive collapse. Steel and Composite Structures 2016;22(3):677–90.
[25] Mashhadi J, Saffari H. Modification of dynamic increase factor to assess
The authors would like to acknowledge the financially supported by progressive collapse potential of structures. J Constr Steel Res 2017;138:72–8.
the National Natural Science Foundation of China (51408489, [26] Liu M. A new dynamic increase factor for nonlinear static alternate path analysis of
building frames against progressive collapse. Eng Struct 2013;48:666–73.
51248007, 51308448, 51301136 and 51508464), the Natural Science [27] Zhu YF, Chen CH, Yao Y, Keer LM, Huang Y. Dynamic increase factor for
Foundation of Shaanxi Province (2017JM5007, 2020JM-152). progressive collapse analysis of semi-rigid steel frames. Steel and Composite
Structures 2018;28(2):209–21.
[28] Sun H, Burton HV, Huang H. Machine learning applications for building structural
References
design and performance assessment: state-of-the-art review. Journal of Building
Engineering 2021;33:101816. https://doi.org/10.1016/j.jobe.2020.101816.
[1] DoD, Design of Buildings to Resist Progressive Collapse, Vol. UFC 4-023-03, [29] Zhang Yu, Burton HV, Sun H, Shokrabadi M. A machine learning framework for
Department of Defence, Washington, DC., 2016. assessing post-earthquake structural safety. Struct Saf 2018;72:1–16.
[2] GSA, Alternate Path Analysis and Design Guidelines for Progressive Collapse [30] Kiani J, Camp C, Pezeshk S. On the application of machine learning techniques to
Resistance, General Services Administration, Washington, DC., 2016. derive seismic fragility curves. Comput Struct 2019;218:108–22.
[3] Alshaikh IM, Bakar BA, Alwesabi EA, Akil HM. Experimental investigation of the [31] Sun H, Burton H, Wallace J. Reconstructing seismic response demands across
progressive collapse of reinforced concrete structures: An overview. Structures multiple tall buildings using kernel-based machine learning methods. Structural
2020;25:881–900. Control and Health Monitoring 2019;26(7). https://doi.org/10.1002/stc.
[4] Kiakojouri F, De Biagi V, Chiaia B, Sheidaii MR. Progressive collapse of framed v26.710.1002/stc.2359.
building structures: Current knowledge and future prospects. Eng Struct 2020;206: [32] Kaveh A, Eslamlou AD, Sheikhi R. Seismic performance of steel structures
110061. retrofitted with optimal slack cable collapse prevention system. Journal of Building
[5] Han Q, Li X, Liu M, Spencer BF. Performance analysis and macromodel simulation Engineering 2020;31:101392.
of steel frame structures with beam-column joints using cast steel stiffeners for [33] Kaveh A, Dadras Eslamlou A, Javadi SM, Geran Malek N. Machine learning
progressive collapse prevention. Thin-Walled Structures 2019;140:404–15. regression approaches for predicting the ultimate buckling load of variable-
[6] Qian K, Lan X, Li Z, Li Y, Fu F. Progressive collapse resistance of two-storey seismic stiffness composite cylinders. Acta Mech 2021;232(3):921–31.
configured steel sub-frames using welded connections. J Constr Steel Res 2020; [34] Mangalathu S, Sun H, Nweke CC, Yi Z, Burton HV. Classifying earthquake damage
170:106117. to buildings using machine learning. Earthquake Spectra 2020;36(1):183–208.
[7] Tang H, Deng X, Jia Y, Xiong J, Peng C. Study on the progressive collapse behavior [35] Rafiei MH, Adeli H. A novel machine learning-based algorithm to detect damage in
of fully bolted RCS beam-to-column connections. Eng Struct 2019;199:109618. high-rise building structures. The Structural Design of Tall and Special Buildings
[8] Wang F, Yang J, Pan Z. Progressive collapse behaviour of steel framed 2017;26(18):e1400. https://doi.org/10.1002/tal.v26.1810.1002/tal.1400.
substructures with various beam-column connections. Eng Fail Anal 2020;109: [36] Timur Cihan M. Prediction of Concrete Compressive Strength and Slump by
104399. Machine Learning Methods. Advances in Civil Engineering 2019;2019:1–11.
[9] Chen K, Zhang Y, Tan KH. Behaviour of steel beam-column joints subjected to [37] Liu JC, Zhang ZG. A machine learning approach to predict explosive spalling of
quasi-static and impact loads. J Constr Steel Res 2021;183:106721. heated concrete. Archives of Civil and Mechanical Engineering 2020;20(4):1–25.
[10] Ding Y, Song X, Zhu H-T. Probabilistic progressive collapse analysis of steel frame [38] Nguyen H, Vu T, Vo TP, Thai HT. Efficient machine learning models for prediction
structures against blast loads. Eng Struct 2017;147:679–91. of concrete strengths. Constr Build Mater 2021;266:120950.
[11] Ferraioli M. Evaluation of dynamic increase factor in progressive collapse analysis [39] Fu F. Fire induced progressive collapse potential assessment of steel framed
of steel frame structures considering catenary action. Steel and Composite buildings using machine learning. J Constr Steel Res 2020;166:105918.
Structures 2019;30(3):253–69. [40] Esfandiari MJ, Urgessa GS. Progressive collapse design of reinforced concrete
[12] Kong D-Y, Yang Y, Yang Bo, Zhou X-H. Experimental study on progressive collapse frames using structural optimization and machine learning. Structures 2020;28:
of 3D steel frames under concentrated and uniformly distributed loading 1252–64.
conditions. J Struct Eng 2020;146(4):04020017. https://doi.org/10.1061/(ASCE) [41] American Society of Civil Engineers, Seismic Evaluation and Retrofit of Existing
ST.1943-541X.0002537. Buildings, Vol. ASCE/SEI 41-17, American Society of Civil Engineers, Reston,
[13] Wang H, Yang B, Chen K, Elchalakani M. Parametric analysis and simplified Virginia, 2017.
approach for steel-framed subassemblies with reverse channel connection under [42] Goh ATC. Back-propagation neural networks for modeling complex systems. Artif
falling-debris impact. Eng Struct 2020;225:111263. Intell Eng 1995;9(3):143–51.
[14] Alembagheri M, Sharafi P, Hajirezaei R, Samali B. Collapse capacity of modular [43] Dhillon A, Verma GK. Convolutional neural network: a review of models,
steel buildings subject to module loss scenarios: The role of inter-module methodologies and applications to object detection. Progress in Artificial
connections. Eng Struct 2020;210:110373. Intelligence 2020;9(2):85–112.
[15] Alembagheri M, Sharafi P, Hajirezaei R, Tao Z. Anti-collapse resistance [44] Ragab MG, Abdulkadir SJ, Aziz N, Al-Tashi Q, Alyousifi Y, Alhussian H, et al.
mechanisms in corner-supported modular steel buildings. J Constr Steel Res 2020; A Novel One-Dimensional CNN with Exponential Adaptive Gradients for Air
170:106083. Pollution Index Prediction. Sustainability 2020;12(23):10090. https://doi.org/
[16] Luo FJ, Bai Yu, Hou J, Huang Y. Progressive collapse analysis and structural 10.3390/su122310090.
robustness of steel-framed modular buildings. Eng Fail Anal 2019;104:643–56. [45] Kiranyaz S, Avci O, Abdeljaber O, Ince T, Gabbouj M, Inman DJ. 1D convolutional
[17] Bregoli G, Vasdravellis G, Karavasilis TL, Cotsovos DM. Static and dynamic tests on neural networks and applications: A survey. Mech Syst Sig Process 2021;151:
steel joints equipped with novel structural details for progressive collapse 107398. https://doi.org/10.1016/j.ymssp.2020.107398.
mitigation. Eng Struct 2021;232:111829.
[18] Li G-Q, Li L-L, Jiang B, Lu Y. Experimental study on progressive collapse resistance
of steel frames under a sudden column removal scenario. J Constr Steel Res 2018;
147:1–15.

934

You might also like