Track Residual Deformation Prediction of Railway Bridges after Earthquakes Using Multi-Scale CNN-Attention

This study proposes an intelligent prediction method for post-earthquake track residual deformation based on a Multi-scale CNN-Attention architecture. To effectively capture the cross-scale characteristics of non-stationary seismic signals, heterogeneous convolution kernels are designed to extract both amplitude evolution and time-frequency features across multiple frequency bands. An additive attention mechanism is introduced to dynamically quantify the contributions of different high-dimensional features and identify the dominant seismic attributes driving track deformation, thereby enhancing the model’s ability to learn system-level coupling effects. A deep learning framework combining serial and parallel pathways is constructed to establish an end-to-end prediction pipeline from ground motion acceleration time histories to the residual deformation of key track nodes. A five-span simply supported railway bridge is used as a case study, for which 1[Formula: see text]800 synthetic ground motion samples are generated to form a training database. The proposed model achieves a coefficient of determination [Formula: see text] of 0.9885, a root mean square error (RMSE) of only 4.5[Formula: see text]mm, and prediction stability within [Formula: see text] — demonstrating significant improvements over traditional models. This work not only provides an efficient tool for assessment of post-earthquake track deformation, but also offers a generalizable paradigm for response prediction of complex coupled systems.