Under earthquake actions, bridge structures may suffer from various forms of damage, which threaten the overall safety of the structure. Traditional finite element methods have high computational costs in nonlinear time-history analysis. In order to rapidly and accurately assess the seismic performance of bridges, a method combined with cluster and multi-level feature fusion deep learning algorithms with low computational cost and high computational accuracy is proposed. To address the challenge of small-sample data, the DTW-Kmedoids time-series clustering framework is proposed to cluster ground motion records and generate the representative small-sample training set for DL model training. The proposed Multi-level feature fusion GRU model with strong generalization capability and high robustness was trained in small-sample scenarios and is capable of accurately and efficiently predicting the nonlinear response of bridge structures under seismic actions. The effectiveness of the proposed method is verified by comparing the computational results with the traditional finite element model. This study provides a novel and efficient solution for seismic response prediction in the small-sample data scenario of bridge engineering.
 

Bridge engineering;Multi-level feature fusion mechanism;Clustering strategy;Small-sample data;Seismic response prediction