Prediction of wheel-rail contact forces using simple onboard monitoring system and machine learning

Abstract

For safe railway operation, periodic measuring of vehicle dynamics (wheel-rail-contact forces) is important, especially for tilting trains since they run faster through curves than normal traffic. So far, these forces are determined in test runs once a year using instrumented wheelsets. To get information more regularly and more economically, a simple onboard monitoring system for daily use on a commercial train has been developed. This system is predicting the forces relevant to assess running safety of tilting trains, so it is optimised for curves with lateral forces close to the critical values. Vertical forces are predicted by metering the primary spring deflection, which is already a proven method. The ambitious part research is focussing on is the prediction of the lateral forces on the whole wheelset and on the guiding wheel. This is obtained by transferring lateral accelerations using machine learning to manage even non-linear effects of the train’s undercarriage. Finally, the used Random Forest regressor thereby shows a good accuracy of the predicted forces compared to the original forces of the instrumented wheelset with correlations of around 95% for the relevant tilting train track sections.