Comparison of typical machine learning algorithms for rapid prediction of operational safety indicators for heavy-haul trains

Abstract

With the expansion of heavy-haul train (HHT) configurations, the increase in axle load, and the enhancement of operating speed, the issue of longitudinal impulse during emergency braking has become increasingly prominent. Therefore, how to quickly assess the operational safety of trains is particularly important for guiding train handling and ensuring service safety. To this end, a co-simulation model combining the longitudinal train dynamics (LTD) and the multibody dynamics (MBD) of HHT was developed in this study. Based on field test data, the accuracy of the established co-simulation model was verified. Furthermore, three surrogate models for the dynamic behavior analysis of HHT were constructed using machine learning algorithms, and the accuracy of three machine learning algorithms, Random Forest (RF), Backpropagation Neural Network (BPNN), and Least Squares Support Vector Machine (LSSVM), in predicting the safety indicators of HHT operation under different conditions was compared. The research results demonstrate that the surrogate model built based on the LSSVM has the highest prediction accuracy and can be considered the preferred algorithm for developing surrogate models for the dynamic behavior analysis of HHT. It is particularly noteworthy that, compared to traditional MBD models, the surrogate model achieves a significant improvement in computational speed (approximately 8.0 × 10⁵ to 2.1 × 10⁶ times), while substantially reducing computational costs. This approach provides new insights into the widespread application of machine learning in the field of railway transportation.