Attention mechanism based neural networks for structural post-earthquake damage state prediction and rapid fragility analysis
Youjun Chen, Zeyang Sun, Ruiyang Zhang, Liuzhen Yao, Gang Wu
Abstract
This paper is devoted to the research on applying the deep learning method to nonlinear structural post-disaster damage state assessment. Transformer and Informer networks with a classification network customized according to the adopted damage assessment framework are proposed for data-driven structural seismic response and damage state modeling. Compared with recurrent neural network and convolution neural network, the networks in this paper can predict the elastoplastic response of nonlinear structures more effectively. In addition, this paper presents a method for rapid structural fragility analysis, which can consider multiple damage assessment indexes at the same time. The performance of the proposed approach is successfully demonstrated through two examples, including a numerical analysis validation and a field sensing validation. The results show that the Transformer network used in this paper is a reliable and computationally efficient approach for predicting the structural seismic response and damage category, and appears great potential in structural health monitoring and rapid assessment on post-disaster structural resilience.