• CN: 11-2187/TH
  • ISSN: 0577-6686

Journal of Mechanical Engineering ›› 2026, Vol. 62 ›› Issue (16): 369-382.doi: 10.3901/JME.260719

Previous Articles    

Wheel Profile Wear Evolution Prediction Method Based on GM(1,1)-ISSA-SVM

WEN Bingguang, LI Wenjie, TAO Gongquan, WEN Zefeng   

  1. State Key Laboratory of Rail Transit Vehicle System, Southwest Jiaotong University, Chengdu 610031
  • Received:2025-08-15 Revised:2026-03-02 Published:2026-10-14

Abstract: To address the challenge that traditional machine learning methods face in accurately predicting wheel profile wear under small-sample conditions, a wheel profile wear prediction method (GM(1,1)-ISSA-SVM) is proposed. This method integrates a non-equidistant grey prediction model GM(1,1) and a support vector machine(SVM) optimized by an improved sparrow search algorithm(ISSA). First, an enhanced GM(1,1) model with optimized background values is used to generate an initial prediction of the wear sequence at wheel profile points. Next, an improved sparrow search algorithm is proposed to automatically optimize the penalty coefficient and kernel parameter of the SVM model. The optimized SVM is then applied to perform residual correction on the initially predicted wear sequence. The corrected results are further integrated with a NURBS curve fitting approach, which uses an optimal weighting factor combination, to obtain the final prediction of wheel profile wear. Finally, the effectiveness and superiority of the proposed model are validated using wheel profile wear tracking data collected from a specific metro line. Experimental results show that the proposed ISSA significantly enhances the global optimization capability of the original SSA. With this improvement, the developed GM(1,1)-ISSA-SVM model achieves high-precision fitting and prediction on small-sample wheel wear datasets. Compared with the standalone GM(1,1) model, the LSTM model, and the GM(1,1)-SVM combined model, the GM(1,1)-ISSA-SVM model demonstrates superior predictive performance across all evaluation metrics. The mean absolute and relative errors between the predicted and measured wheel profile data are only 0.005 mm and 0.5%, respectively. Furthermore, the normal error between the predicted and measured wheel profiles does not exceed 0.05 mm, and the wheel-rail contact characteristics closely match those derived from the measured profile. These results could help facilitate the transition of wheel maintenance strategies from scheduled maintenance to condition-based maintenance.

Key words: wheel wear prediction, data driven, GM (1,1) model, sparrow search algorithm, support vector machine

CLC Number: