Spatio-temporal deep learning model-aided seismic fragility analysis of long-span bridges considering spatial variability of ground motions

Structural fragility analysis plays a critical role in the assessment of infrastructures.This process commonly requires large numbers of nonlinear time history analyses via extremely timeconsuming finite element simulations.Although booming deep learning (DL) techniques have been served as surrogate models to lessen computing burden, the previous DL surrogate models neglect the spatial variability of ground motions in response prediction for long-span bridges.Against this background, this paper proposes a spatio-temporal DL model for structural dynamic analysis of long-span bridges considering spatial variability of ground motions.In this model, convolution operations that can capture spatial features are integrated with gated recurrent unit (GRU) network with extraction capability of temporal features.Therefore, the spatial variability of ground motion is extracted through convolution, and the inherent temporal correlations within structural response is established via GRU.For validating the proposed spatio-temporal model, the prediction performance of this model was systematically evaluated through a case study involving a long-span cable-stayed bridge.Meanwhile, the traditional time series prediction model was used for comparison.Subsequently, the fragility analysis of bridge component was conducted, which further demonstrated the application feasibility of the proposed model for structural seismic fragility assessment.The results indicated that the proposed model could efficiently and accurately predicting the structural seismic responses and conducting fragility analysis considering spatial variability of ground motions.