Real-time railhead friction estimation using machine learning: development of an on-board-train data capture system

Low adhesion between the wheel and rail interface remains a significant challenge for the railway industry, particularly during the autumn season, leading to delays and safety risks such as station overruns and signals passed at danger (SPADs). The impact of low adhesion is estimated to cost the UK railway industry approximately £355 million annually. Current methods for estimating railhead adhesion lack real- me, high-resolution spa al and temporal capability, which is critical for improving safety and operational efficiency. This research introduces a novel real- me railhead friction estimation approach, utilizing an estimation model trained on a variety of environmental and rail-specific data, such as railhead images, friction measurements, air temperature, relative humidity and railhead temperature. To test this model in real-world conditions, a specialized data capture system (camera box) was developed and mounted on rolling stock, capturing relevant data while ensuring accuracy in location and railhead condition. Field tests conducted at the Wensleydale Heritage Railway in the UK demonstrated the feasibility of this system, with consistent and reliable friction estimations. The results indicate that the model can effectively estimate railhead friction levels and identify potential low adhesion hotspots in real me, thus providing valuable insights for mitigating risks, reducing delays, and improving overall safety.