Knowledge-Informed Wheel Wear Prediction Method for High-Speed Train Using Multisource Signal Data

Wear prediction for train wheels is essential for evaluating the health status of wheel-rail systems. Existing prediction approaches mainly focus on the physics-based approach or data-driven approach, which either involve complex mechanisms or lack interpretability. A data-driven wear prediction method regarding domain knowledge and multisource signals is developed herein to improve the difficulties in the two approaches. The presented method involves three modules. First, axle box acceleration (ABA) data are investigated via spectral analysis, and domain knowledge associated with wheel wear degradation is concluded. Then, data fusion and feature extraction are performed to modify the vertical ABA signals and extract effective features. Next, a supervised regression model is built to predict wheel tread wear using the extracted feature and wear data. While the model is established, on-board monitoring for wheel tread wear can be realized by inputting the measured ABA signals. The performance of our method is evaluated and tested on real-world data from three service lines. Experimental results show that the developed method performs satisfactorily in terms of mean absolute percentage error, root mean square error, and R2, registering average values of 0.0939, 0.0224, and 0.9457, respectively.