A Data-Driven Wheel Wear Prediction Model for Rail Train Based on LM-OMP-NARXNN

Abstract The wheel wear situation on the railway vehicles will affect the train running stability and riding comfort. Thus, the prediction model of wheel tread wear is critical for anticipating the wheelset state information and formulating the reprofiling strategy. However, for the wheel wear analysis, the physical simulation models based on vehicle track system dynamics are time consuming and do not have universal adaptability. Moreover, it underutilized the large amount of raw data accumulated by the wheelset detection system in the long-term service of the vehicle. This article presents a data-driven method for precisely predicting wheel wear in future. This method includes nonlinear autoregressive models with exogenous inputs neural networks (NARXNNs), Levenberg Marquardt (LM), and orthogonal matching pursuit (OMP) algorithm, i.e., LM-OMP-NARXNN, and LM-OMP is used to ascertain the network weight and nodes of the prediction model structure. Datasets of the case study are derived from a motor station for three consecutive years. The experiment results demonstrate that the proposed method leads to a more compact model with the reduced size. Besides, it has higher accuracy in the prediction of wheelset tread wear status in the short term when compared with other prediction models and other training algorithms used in NARXNN.