Physics-based data-driven interpretation and prediction of rolling contact fatigue damage on high-speed train wheels

Abstract Rolling contact fatigue (RCF) of high-speed train wheels is detrimental to reliability and entails huge maintenance costs. Accurate prediction of RCF damage is important for train operators, however, physics-based methods are not suitable for practical applications. Based on accumulated RCF data, this paper develops a physics-based data-driven method for the interpretation and prediction of RCF damage risk. First, based on RCF mechanism, it is found that many relevant parameters can be used for RCF prediction. Then, eight predictors are extracted from RCF-related data through dimensionality reduction methods, including linear discriminant analysis and kernel Fisher discriminant analysis. It shows that greater wheel out-of-roundness or smaller wheel diameters leads to higher RCF risks; previously defective wheels are more likely to reencounter RCF damage; the RCF risk varies in different seasons; the effect of wear on RCF damage is significantly nonlinear. Afterwards, a predictive model is established based on logistic regression. Validation results yield an average accuracy of 75% and also demonstrate its easy interpretability, fast computational speed and stable performance. Further, test result shows an accuracy of 85% to distinguish high-risk wheelsets prior to RCF inspections. The proposed method can be applied to the risk-based maintenance of high-speed train wheels.