Comparative Analysis of Machine Learning-Based Surrogate Modeling Approaches for Multi-Body Dynamic Simulation in Railway Digital Twin Platform

Abstract

Machine learning (ML)-based surrogate models offer a promising alternative for Multibody Dynamics (MBD) Simulation of railway vehicle-track dynamics systems. A well-built ML model can accurately and quickly predict the dynamic responses to various track irregularities, significantly reducing computation time and unifying integration interfaces of different vehicle-track systems. However, training effective surrogate models is a complex process, influenced by the specific needs of the analysis. Different algorithms and training sets might be required for different surrogate models, making it essential to research the impact factors for building these models. This paper presents a comparative analysis of ML-based surrogate modeling approaches tailored for MBD Model within a railway digital twin platform. The primary focus is evaluating the performance of various ML-based surrogate modeling approaches, as well as the influence of neural networks, training parameters, and data sources. By leveraging extensive simulation and measurement data, we assess the ability of these surrogate models to predict key performance indicators under varying operational conditions. This analysis provides valuable insights for railway engineers in selecting appropriate surrogate modeling approaches, ultimately contributing to the advancement of predictive maintenance and optimization in railway operations.