Advanced VTSDREF for vehicle-turnout system dynamic reliability analysis: Integration of hybrid deep learning and adaptive probability density evolution method

Turnouts are critical in railway systems, but evaluating the dynamic reliability of the vehicle-turnout system has significant challenges. Current methods have high costs for dynamic response calculations, accuracy issues with surrogate models, and substantial computational expenses for solving probability density evolution equations. To tackle these issues, a new framework, vehicle-turnout system dynamic reliability evaluation framework (VTSDREF), is proposed. A mechanistic model considering multiple random parameters of the vehicle-turnout system is established, yielding random responses. To reduce computational costs, Bayesian optimization is used to optimize the hyperparameters of a temporal convolutional neural network, resulting in a hybrid deep learning surrogate model. Additionally, an adaptive probability density evolution method (APDEM) is introduced, significantly improving computational efficiency. The proposed surrogate model and APDEM outperform traditional methods, with computational times reduced to 25.111 % and 0.098 % respectively. The computational time of the VTSDREF is only 21.361 % of traditional methods, significantly enhancing the efficiency of reliability evaluation. These research findings hold considerable practical value for the maintenance and repair of vehicle-turnout system.