Cite this article: Wang Y, Zhao Q, Qian K, Wang Z, et al. Cumulative absolute velocity prediction for earthquake early warning with deep learning, Computer‐Aided Civil and Infrastructure Engineering,(2023). https://doi.org/10.1111/mice.13065

Cumulative absolute velocity prediction for earthquake early warning with deep learning

Yanwei Wang, Qingxu Zhao, Kai Qian, Zifa Wang*,

Zhenzhong Cao, Jianming Wang

Corresponding author: Zifa Wang

Yanwei Wang, Guangxi Key Laboratory of Geomechan-ics and Geotechnical Engineering, Guilin University of Technology, Guilin,541004, China, wywiem@163.com

Qingxu Zhao, Guangxi Key Laboratory of Geomechan-ics and Geotechnical Engineering, Guilin University of Technology, Guilin,541004, China, 742884711@qq.com

Kai Qian, Guangxi Key Laboratory of Geomechan-ics and Geotechnical Engineering, Guilin University of Technology, Guilin,541004, China, qiankai@glut.edu.cn

Zifa Wang, 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; CEAKJ ADPRHexa Inc., Shaoguan, Guangdong 512000, China, zifa@iem.ac.cn

Zhenzhong Cao, Guangxi Key Laboratory of Geomechan-ics and Geotechnical Engineering, Guilin University of Technology, Guilin,541004, China, iemczz@163.com

Jianming Wang, CEAKJ ADPRHexa Inc., Shaoguan, Guangdong 512000, China, jwang780@gmail.com

Abstract:

Rapid and accurate estimate of earthquake damage is a key component in a successful earthquake early warning (EEW) system. The Cumulative Absolute Velocity (CAV) is an important and widely used parameter to measure ground motion intensity, but it cannot be correctly estimated via the traditional approach with the limited information available in typical EEW systems. Therefore, current EEW systems cannot effectively use CAV to predict earthquake damage. Herein, a CAV prediction model (DLcav) based on Convolutional Neural Networks was proposed for EEW systems. DLcav is an end-to-end solution to continuously predict CAV using arriving seismic waves of increasing length and supplemented with additional auxiliary information. The effectiveness of DLcav to predict CAV was tested based on Japanese ground motion records, and the generalization ability of DLcav was assessed using the ground motion records from Chile. The results demonstrate that DLcav can rapidly predict CAV with good accuracy which will help better estimate earthquake damage in EEW systems.

Keywords: Cumulative absolute velocity; Ground motion prediction; Earthquake early warning; Deep learning

Fig.1 CAV changes with increasing duration for a typical ground motion record

Table 1 Adjustments of DLcav architecture

Fig.2 Losses of different architectures with initial 3 s waves of the validation dataset

Fig.3 Architecture and hyperparameters of DLcav

Fig.4 Distribution of the selected Japan accelerograms. (a) Distribution of accelerograms with M_w and hypocentral distance. (b) Number of accelerograms for different M_w bins. (c) Number of accelerograms for each hypocentral distance bin. (d) Number of accelerograms for different Vs30 bins. (e) Number of accelerograms for different CAV bins

Fig.5 The distribution of CAV predicted by DLcav with initial 3 s waves of the test data set. Each black dot represents one prediction. The black solid line is the 1:1 line showing perfect agreement between the predicted and the observed. The two blue dashed lines represent the range of ± 1 standard deviation. r is the correlation coefficient. pre is the predicted CAV (CAV_pre), and obs is the observed CAV (CAV_obs)

Fig.6 Distributions of DLcav prediction error with magnitude (a) - (d), hypocentral distance (e) - (h) and Vs30 (i) - (l) with initial 3 s waves of the test dataset. The black dot represents the error of predicted CAV, and σ is the standard deviation of the error while μ is the mean error, and the two blue lines represent the range of ± σ

Fig.7 The distribution of CAV predicted by DLcav with initial 4 - 6 s waves of the test dataset. Each ‘○’ represents one prediction. The black solid line is the 1:1 line showing perfect agreement between the predicted and the observed. The two blue dashed lines represent the range of ± 1 standard deviation. r is the correlation coefficient, pre is the predicted CAV (CAV_pre), and obs is the observed CAV (CAV_obs)

Fig.8 Correlation coefficient (a), standard deviation of errors (b), and mean of the errors (c) of CAV predicted by DLcav with initial 3 - 6 s waves in the test dataset

Fig.9 Distribution of the selected Chile accelerograms. (a) Distribution of accelerograms with M_w and hypocentral distance. (b) Number of accelerograms for different M_w bins. (c) Number of accelerograms for different hypocentral distance bins. (d) Number of accelerograms for different Vs30 bins. (e) Number of accelerograms for different CAV bins. (The dashed line represents the boundary of the training dataset; ‘×’ marks the outside records)

Fig.10 The distribution of CAV predicted by DLcav with initial 3 - 6 s waves of generalization dataset. (‘○’ are for records in the range of the training database; ‘×’ are outside records not in the boundary of the training dataset)

Fig.11 Correlation coefficient (a), the standard deviation of the error (b) and mean of error (c) of CAV predicted by DLcav with initial 3 - 6 s waves in the generalization dataset

Fig.12 Confusion matrices of CAV threshold for generalization dataset

Fig.13 MCC of DLcav with initial 3 - 6 s waves of generalization dataset

Supported by：

This work was supported by the National Natural Science Foundation of China (Grant No. 51968016, No. 51978634 and No. 51968015) and the Natural Science Foundation of Guangxi (No.2021GXNSFFA196001).