A deep learning-based surrogate model for dynamic interaction assessment of high-speed overhead conductor rail system

Overhead conductor rail (OCR) is a critical power-supplying structure for trains in railway tunnels. As operating speed increases, assessing the dynamic interaction of the pantograph-OCR system (POCR) becomes increasingly crucial, which is widely analysed using the finite element method. However, this approach has high computational costs when applied to large-scale cases. To tackle this issue, a surrogate model that simultaneously predicts multiple indicators for evaluating the dynamic performance is developed using deep learning in this paper. Firstly, a mathematical model simulating the dynamic behaviour of the POCR is proposed and validated against measurement data. Five input OCR structural parameters are extracted, based on which three output indicators are calculated by the mathematical model. Next, a sampling strategy is employed to establish a parameter variable space for 30,000 cases. The numerical model is used to generate the 20,000 cases for setting up a database. Thirdly, a hybrid network architecture combining convolutional neural networks (CNN) and long short-term memory (LSTM) network is proposed to construct the surrogate model and simulate the remaining 10,000 cases, with optimal hyperparameters determined through an optimisation strategy. The results indicate that the maximum relative errors in three output indicators between numerical simulation and surrogate model are 4.17 %, 6.73 %, and 4.75 %, respectively. The sensitivity analysis is performed to reveal the effect of structural parameters on the dynamic performance of the POCR, and span length is the most influential factor.