Cite this article: Huang, Y., He, J., Zhu, Z. Rapid Assessment of Seismic Risk for Railway Bridge Based on Machine Learning. International Journal of Structural Stability and Dynamics 23 (2023). https://doi.org/10.1142/S0219455424500561
Rapid Assessment of Seismic Risk for Railway Bridge Based on Machine Learning
Yong HUANG, Jing HE, Zhihui ZHU*
Corresponding author: Zhihui Zhu
Yong Huang, Institute of Engineering Mechanics, China Earthquake Administration, Key Laboratory of Earthquake Engineering and Engineering Vibration, No. 29 Xuefu Road, Harbin, Heilongjiang, People’s Republic of China, 150080, huangyong@iem.ac.cn
Jing He, Institute of Engineering Mechanics, China Earthquake Administration, Key Laboratory of Earthquake Engineering and Engineering Vibration, No. 29 Xuefu Road, Harbin, Heilongjiang, People’s Republic of China, 150080, Hejing@163.com.
Zhihuii Zhu, School of Civil Engineering, Central South University; National Engineering Research Center of High-speed Railway Construction Technology, Changsha, Hunan, People’s Republic of China, 410075, ruizhi@iem.ac.cn
Abstract:
When an earthquake occurs, railway bridges will suffer from different degrees of seismic damage, and it is necessary to assess the seismic risk of bridges. Unfortunately, the majority of studies were done on highway bridges without taking into account railway bridge characteristics; hence they are not applicable to railway bridges. Furthermore, current research methods for risk assessment cannot be performed quickly, and suffer from the problems of subjective personal experience, complicated calculations, and time consuming. This paper we use machine learning for earthquake damage prediction and empirical vulnerability curves to represent risk assessment results, creating a rapid risk assessment procedure. We gathered and tallied seismic damage data from 335 railway bridges that were damaged in the Tangshan and Menyuan earthquakes, found six variables that had a substantial impact on seismic risk outcomes, and categorized the damage levels into five categories. It is essentially a multi-classification and prediction problem. In order to solve this problem, four algorithms were tested: Random Forest (RF), Back Propagation Artificial Neural Network (BP-ANN), PSO-Support Vector Machine (PSO-SVM), and K Nearest Neighbor (KNN). It was found that RF is the most effective method, with an accuracy rate of up to 93.31% for the training set and 89.39% for the test set. Then this study describes the new procedure in detail for rapidly assessing seismic risk to 269 bridges chosen at random from the sample pool. Firstly, the seismic damage data of bridges are collated, then the seismic damage rating is predicted using RF, and finally the empirical vulnerability curve is drawn using a two-parameter normal distribution function for the purpose of seismic damage risk assessment. The study's findings can be used as a guide for choosing a machine learning approach and its inputs to build a rapid assessment model for railway bridges.
Keywords: machine learning; railway bridges; seismic risk; empirical vulnerability.
Table1. The defines of the degree of earthquake damage
Degree of seismic damage |
Destruction phenomenon |
Collapsed |
The bridge cannot be used. |
Severely Damaged |
The main load-bearing structure is severely damaged and needs significant repair or reconstruction. |
Moderately Damaged |
The main load-bearing structures suffer damage or local damage. |
Slightly Damaged |
Non-load-bearing structures suffer damage. |
Intact |
No seismic damage. |
(a)Sleeper crack (b) Bearing damaged (c) Pier cracked
(d) Girder large horizontal displacement (e) Pier fracture
Fig.1. Typical railway bridge seismic damage
Fig.2. Data distribution of each feature category in the database
Fig.3. Input variables and output variable values in the model
Fig.4. Bagging training process
Fig.5 Accuracy of RF
Fig.6. BP artificial neural network
Fig.7 Accuracy of BP-ANN
Fig.8. PSO-SVM prediction flow chart
Fig.9 Fitness of PSO-SVM
Fig.10 Accuracy of PSO-SVM
Fig.11. Flowchart of KNN
Fig.12 Accuracy of KNN
Table2. Comparison of the accuracy of each model
method |
Training accuracy (%) |
Testing accuracy (%) |
RF |
93.9086 |
89.3939 |
BP-ANN |
77.3234 |
80.3030 |
PSO-SVM |
91.0781 |
83.3333 |
KNN |
86.1940 |
86.5672 |
(a) RF (b) BP-ANN
(c) PSO-SVM (d) KNN
Fig.13. Confusion matrix of test data in each model
Fig.14. Significance of characteristic
Fig.15. Process of seismic risk assessment
Table3. Correlation between intensity and PGA
Intensity |
Ⅵ |
Ⅶ |
Ⅷ |
Ⅸ |
Ⅹ |
Ⅺ |
PGA(m/s2) |
0.63 |
1.25 |
2.50 |
5.00 |
10.00 |
20.00 |
PGA(g) |
0.064 |
0.128 |
0.255 |
0.510 |
1.020 |
2.041 |
Table4. Seismic predicated matrix of railway bridges %
Intensity DS |
Ⅵ |
Ⅶ |
Ⅷ |
Ⅸ |
Ⅹ |
Ⅺ |
DS1 |
100.00 |
90.77 |
71.88 |
28.21 |
0 |
60.00 |
DS2 |
0 |
3.08 |
3.12 |
7.69 |
60.00 |
0 |
DS3 |
0 |
4.62 |
25.00 |
28.21 |
40.00 |
40.00 |
DS4 |
0 |
1.54 |
0 |
23.08 |
0 |
0 |
DS5 |
0 |
0 |
0 |
12.82 |
0 |
0 |
Table5. Exceedance probability F of bridges with different earthquake damage grades
DS PGA(g) |
DS1 |
DS2 |
DS3 |
DS4 |
DS5 |
|
0.064 |
1.00 |
0 |
0 |
0 |
0 |
|
0.128 |
1.00 |
0.09 |
0.06 |
0.02 |
0 |
|
0.255 |
1.00 |
0.28 |
0.25 |
0 |
0 |
|
0.510 |
1.00 |
0.72 |
0.64 |
0.36 |
0.13 |
|
1.020 |
1.00 |
1.00 |
0.40 |
0 |
0 |
|
2.041 |
1.00 |
0.40 |
0.40 |
0 |
0 |
Fig.16. Vulnerability curve based on a two-parameter log-normal distribution function
Supported by:
This study was supported by Open Foundation of National Engineering Research Center of High-speed Railway Construction Technology; National Natural Science Foundation of China (Grant number: 52078498), Natural Science Foundation of Hunan Province of China (Grant number: 2022JJ30745), the Fundamental Research Funds for the Central Universities of Central South University (Grant number: 2020zzts149).