Running safety assessment of trains crossing bridges using hybrid machine-learning based vehicle-structure algorithm

Dynamic vehicle–structure interaction analyses for train running safety assessment require accurate modelling of wheel–rail contact behaviour, as contact forces strongly influence the dynamic response of both the vehicle and the structure. Realistic contact formulations involve highly nonlinear algorithms capable of capturing the complex spatial and temporal behaviour of wheel–rail interaction, although at significant computational cost and implementation complexity. To improve the efficiency of these analyses, this paper presents a hybrid machine-learning-based vehicle–structure interaction framework employing deep neural networks for the prediction of wheel–rail contact parameters within a three-dimensional nonlinear dynamic model. In contrast to conventional feedforward approaches, the proposed methodology investigates convolutional and long short-term memory neural network architectures to capture the sequential and nonlinear characteristics of the contact behaviour during train crossing of railway bridges. The models are trained using numerically generated datasets obtained from a nonlinear three-dimensional vehicle–structure interaction algorithm considering multiple track irregularity profiles and train speeds representative of operational and running safety conditions. The developed networks are designed to learn the complex spatial-temporal behaviour of the coupled dynamic system while maintaining reduced model complexity and improved computational efficiency. Numerical results demonstrate that the proposed CNN–LSTM framework is capable of accurately reproducing the nonlinear wheel–rail contact behaviour and contact parameters throughout the dynamic interaction process. The study highlights the potential of advanced deep-learning architectures for efficient running safety assessment of trains crossing bridges using hybrid vehicle–structure interaction methodologies.