A Convolutional Neural Networks Model Based on Wavelet Packet Transform for Enhanced Earthquake Early Warning

Abstract

The earthquake early warning (EEW) system is critical in mitigating the effects of seismic hazards by providing valuable response time. Damage to buildings can be determined based on peak ground acceleration (PGA). If PGA can be predicted in advance, the seismic risk of a building can be estimated. A novel method for predicting PGA using wavelet packet transform (WPT) and convolutional neural networks (CNN) is proposed. Early arrival waves generated by earthquakes are decomposed using WPT to produce a matrix of wavelet coefficients. This matrix serves as the input to the CNN model to predict the PGA. To achieve the best prediction performance, different setups were investigated, including the use of Daubechies Wavelet 4 (db4) and four-level decomposition. The results indicate that this configuration yields better prediction accuracy. P-waves of three-second duration are commonly used for EEW. Compared to the prediction results of a CNN model used to validate the method, the proposed method has a lower average error and better use of early arrival waves with shorter duration. Overall, the proposed method demonstrates the potential of WPT and CNN in EEW systems. The proposed approach can utilize shorter early arrival waves to predict PGA and gain more reaction time to get rid of seismic hazards. To estimate structural damage using the predicted PGA under an impending earthquake, a method is proposed to quickly determine structural damage based on the predicted PGA and fragility curves. It provides a method to estimate the potential structural damage caused by an impending earthquake.