Optimization of energy-absorbing parameters for train crash energy management using a variable-fidelity neural network surrogate model

Abstract

Train crash energy management is a design technique used to improve the safety performance of trains in collision accidents. It reduces the impact force during a collision and protects the safety of passengers through the appropriate configuration of energy-absorbing parameters. In order to balance the computational efficiency of existing one-dimensional (1D) models and the simulation accuracy of three-dimensional (3D) models, this paper established a variable-fidelity surrogate model (VFSM) based on a neural network to optimize the energy-absorbing parameters of the train. A Long Short-Term Memory network with Motion Constraints (LSTM-MC) was constructed to predict the collision curve of the train. Then, the pre-training results of the 1D low-fidelity model (LFM) were transferred to the 3D high-fidelity model (HFM) using transfer learning. Two verified cases show that the proposed VFSM achieves high prediction accuracy for the displacement, velocity, and acceleration curves of the train, with the peak acceleration error within 10 %. Finally, the energy-absorbing parameters were optimized with the goal of minimizing the acceleration. The optimization results show that there is no significant surge in the impact force of the train, and the peak acceleration is about 3.4 g.