A Dual Data-Model-Driven Approach for Rail Wear Prediction in Heavy-Haul Railways: Integrating Vehicle-Track Dynamics and Bayesian Residual Correction

With the increase in axle load and traffic density of heavy-haul railways, intense cyclic wheel-rail contact forces accelerate rail wear, especially in small-radius curves where lateral forces induce complex flange-rail contact. Accurate rail wear prediction is critical for optimizing maintenance schedules and ensuring operational safety. Traditional rail wear prediction relies on vehicle-track dynamics simulations or finite element methods, but these suffer from high computational costs and insufficient accuracy in heavy-haul scenarios due to simplified dynamics modeling. To address this, this paper proposes a data-model dual-driven approach integrating vehicle-track dynamics and data-driven residual correction. First, field-measured wheel-rail force data are collected to calibrate a vehicle-track coupling dynamics model: sensitive parameters are identified via the Morris method, and a genetic algorithm is used to minimize the deviation between simulated and measured wheel-rail forces, enhancing the fidelity of dynamics-based wear simulation. Second, the calibrated dynamics model is combined with the Archard wear model to generate simulated wear values. Finally, Bayesian regression predicts the residual between actual and simulated wear using gross tonnage and simulated values, with the final wear prediction obtained by fusing simulated values and residuals. Experimental results on the Shenshuo heavy-haul railway show that parameter calibration reduces wheel-rail force errors by an average of 69.5%, and the proposed approach lowers wear prediction relative error from 15% to 4.3%. This framework provides a reliable tool for rail maintenance decision-making, demonstrating significant engineering value for heavyhaul railway safety and efficiency.