A multimodal TransTCN fusion network embedding prior knowledge for predicting seismic responses in high-speed railways considering pier height

Abstract

Given that the significant impact of stochastic pier height on the seismic response of high-speed railway track-bridge systems (HSRTBS) has been revealed, related seismic design and post-earthquake safety assessments face greater computational demands. To address the inadequacies of traditional methods in meeting these demands—specifically, the heavy computational burden of purely physical methods and the insufficient sample size required for training purely data-driven models—we propose a multimodal deep learning framework called the Prior Domain Knowledge-Transformer Temporal Convolutional Network (PDK-TransTCN) to achieve rapid and accurate prediction of key seismic response indicators of HSRTBS under seismic excitation. In this framework, the integration of prior domain knowledge (PDK) significantly reduces the number of training samples needed to achieve the same accuracy, while the use of Transformers and Temporal Convolutional Networks (TCN) with strong capabilities in fitting nonlinear temporal tasks enhances the model's precision. The model was validated using reliable training samples from shaking table tests and compared with several neural network models that have demonstrated strong performance in seismic damage prediction through ablation experiments. Results showed an 8.0 % improvement in the coefficient of determination (R²) for predicting peak track irregularities (PTIs), requiring only 30 % of the training samples that traditional models need to achieve the same accuracy. These improvements in computational efficiency and predictive performance make the proposed model suitable for real-time post-earthquake assessments, providing critical technical support for quantifying uncertainties in high-speed rail network safety and assessing regional operational risks.