Feature importance analysis of low adhesion risk model to find key parameters influencing friction

Low wheel-rail adhesion presents a major safety and economic challenge for the rail industry by causing station overruns and delays. The impact of this (with estimates of costs up to £355 m annually) presents a research opportunity with the potential for significant safety improvements and cost savings. Poor adhesion can be mitigated somewhat using weather forecasting, although this is limited to regional-level resolution and results in reduced averaged speeds. Recent studies have investigated rail cleaning to remove railhead contamination prior to wheel contact, which increases friction and braking performance. Determining when to activate systems for removing contamination on the rail surface is an area of active research with systems combining sensor and camera data with machine learning models recently demonstrating accurate prediction of Coefficient of Friction (CoF) values. While these systems are effective in initial testing, they display a common problem for safety-critical ML systems in a lack of model interpretability due to the “black box” nature of the neural network models used. This study seeks to mitigate poor ML interpretability through the use of Feature Importance (FI) algorithms, to better understand the causes of low wheel-rail friction prediction in ML models. A dataset correlating rail friction measurements against sensor data and rail images is used to train an ensemble of ML models, which are then used alongside a testing dataset to evaluate an array of FI algorithms (SHapley Additive exPlanations, Local Interpretable Model-agnostic Explanations and Permutation Importance). A feature importance fusion algorithm is then applied to assist in interpreting these results. Across a range of feature importance techniques, it is consistently found that atmospheric data such as air temperature and humidity is the strongest predictor of low friction conditions. Tests of similar friction prediction models using only atmospheric data with no photographic images show only a small loss in accuracy. Edge detection is found to be the most useful feature extracted from images, while single-pixel features such as brightness and colour are found to be the least important. This provides insight into best practices for feature selection when designing ML models for the rail industry in future.