MBSNet: A deep learning model for multibody dynamics simulation and its application to a vehicle-track system

Abstract

In multibody dynamics simulation (MBS) analysis, researchers usually face three challenges: high modeling difficulty, large calculation amount, and restricted solver. Therefore, it is of interest to explore a new technology that can solve or avoid one or more of these challenges. This article first explores how to use deep learning (DL) to realize dynamics simulation of a multibody system and contributes to the modeling and analysis of MBS in two meaningful aspects: (1) Based on the 3D convolutional neural network (3DCNN), long short-term memory(LSTM), and fully connected network (FCNN), we develop a DL network called MBSNet, which considers the relationship between system variables and disturbance variables of a general multibody system, to realize MBS analysis. (2) Based on the short-term MBS results of a vehicle-track vertically coupled dynamics model, we train MBSNet and apply it to long-term system dynamic predictions. The comparison between the MBSNet result and the MBS result shows that MBSNet has high robustness in the face of different track irregularities and can accurately and quickly achieve long-term predictions of low-frequency dynamic responses. Finally, the evidence from the vehicle-track case shows that MBSNet has the potential to be used in MBS analysis, but it faces some challenges, including inaccurate prediction of high-frequency impact components caused by the coupled motion of bodies in a multi-degree-of-freedom system, improper training strategy, insufficient working conditions and samples considered during training, etc.