High‐speed railway track geometry robust maintenance with Bayesian deep learning‐enhanced physical‐informed neural networks

Millimeter‐level active control of track geometry irregularity (TGI) is crucial for intelligent maintenance of high‐speed railways. To address the efficiency limitations of existing time‐domain characteristics‐dominated TGI maintenance scheme design methods and vehicle–track coupled dynamics evaluation models, this paper proposes a new Bayesian deep learning‐enhanced physical‐informed neural network for TGI robust maintenance (PINN‐RM) to overcome technical barriers in frequency‐domain energy regulation and active optimization of driving performance. First, a Bayesian deep learning‐enhanced multilayer perceptron architecture is developed, which converts traditional constrained optimization into neural network parameter optimization and effectively solves the computational complexity and robustness challenges in high‐dimensional decision spaces. Furthermore, a dense scale constraint method and loss weighting method based on attention transition mechanism are proposed to significantly enhance dynamic interaction capabilities and convergence efficiency. Case studies demonstrate that compared with the traditional design method under equivalent constraints, the proposed approach can maintain the similarity of repeated test results over 0.98 and improve the computational efficiency by more than seven times. Compared with traditional design modes, the optimized schemes can reduce carbody acceleration by 10%–25% and improve the Sperling index by 8.8%, confirming its superior capability in dual active control of static track geometry smoothness and dynamic driving performance.