Estimation of railway vehicle response for track geometry evaluation using branch Fourier neural operator

Abstract: In railway transportation, the evaluation of track geometry is an indispensable requirement to ensure the safety and comfort of railway vehicles. A promising approach is to directly use vehicle dynamic responses to assess the impact of track geometry defects. However, the computational cost of obtaining the dynamic response of the vehicle body using dynamics simulation methods is large. Thus, it is important to obtain the dynamic response of the vehicle-track coupled system efficiently and accurately. In this work, a branch Fourier neural operator (BFNO) model is proposed to obtain the dynamic response of the vehicle-track coupled system. The model takes into account the nonlinear relationship of the vehicle-track coupled system and realizes the fast and accurate estimation of the system dynamic response. The relative loss (rLSE) of BFNO model is 2.04%, which is reduced by 64%, compared with the traditional neural network (CNN-GRU). In the frequency domain, BFNO model achieves the effective estimation of the dynamic response of the system within the primary frequency range. Compared with the existing methods, our proposed model can make predictions at unseen time steps, enabling predictions from low to high time resolutions. Meanwhile, our proposed model is superior to commercial software in terms of efficiency. In the evaluation of track geometry, users can use pre-trained BFNO to obtain the dynamic response with almost no computational cost.