Prediction of Railway Vehicles’ Dynamic Behavior with Machine Learning Algorithms

Abstract

The dynamic performance of railway vehicles needs to be carefully monitored to ensure their safe operation. Presently a number of systems such as the Vehicle Track Interaction Monitor and the Instrumented Revenue Vehicles, utilize a number of on-board inertial sensors to obtain near-real time information on the dynamic performance of railway vehicles. These systems provide rich data sets that give an indication of the underlying track condition and the corresponding dynamic response. This paper outlines the use of Machine learning to develop dynamic behavior predictive models for railway vehicles from measured data. This study worked on the development of 2 types of predictive models, viz. regression and classification model. The regression model predicted the time series dynamic response amplitude and the classification model classified the track sections based on the response distribution over it. Train speed and parameters estimated from the unsprung mass were used as predictors in the model. After the trial of a number of predictive models the Ensemble Tree Bagger method was found to have highest overall prediction accuracy. These predictive models can be utilized as a decision making tool to determine safe operational limits and prioritize maintenance interventions.