A Machine Learning Approach for Real-Time Wheel-Rail Interface Friction Estimation

Abstract

Predicting friction at the wheel-rail interface is a key problem in the rail industry. Current forecasts give regional-level predictions, however, it is well known that friction conditions can change dramatically over a few hundred meters. In this study, we aimed to produce a proof-of-concept friction prediction tool which could be used on trains to give an indication of the limiting friction present at a precise location. To this end, field data including temperature, humidity, friction, and images were collected. These were used to fit a statistical model including effects of local environmental conditions, surroundings, and railhead state. The model predicted the friction well with an R2 of 0.97, falling to 0.96 for naive models in cross validation. With images and environmental data collected on a train, a real-time friction measurement would be possible.