Train Slide Prediction and Risk Assessment Using Vehicle-Signal Data: A Data-Model Fusion Method

Abstract Insufficient wheel–rail adhesion during braking poses a significant threat to operational safety and accelerates component wear. Although anti-slide valves modulate braking force to mitigate sliding, the resulting speed fluctuations can inadvertently trigger emergency braking commands from the signaling system, potentially exacerbating the slide. Existing detection methods are largely reactive, relying on post-event identification rather than predictive foresight. To bridge this gap, an onboard module for predictive risk warning is proposed. Utilizing real-world vehicle and signaling data, a cross-attention CNN–GRU model is proposed for the accurate prediction of whole-train wheel slide probability. In parallel, a train dynamics model projects the braking trajectory to assess the displacement deviation relative to target positions and safety margins. By integrating the data-driven slide probability with the model-based displacement deviation, a dynamic risk matrix is constructed to assess different levels of train slide risk. With an optimal prediction horizon of 2.5 s and established alarm thresholds, the proposed module provides actionable inputs for the train operation control system. Experimental results demonstrate that the proposed method effectively predicts both wheel sliding and the associated overrun risk. This enables the signaling system to proactively intervene in sliding control, preemptively mitigating sliding occurrences and enhancing overall operational safety.