• CN: 11-2187/TH
  • ISSN: 0577-6686

Journal of Mechanical Engineering ›› 2026, Vol. 62 ›› Issue (16): 355-368.doi: 10.3901/JME.260720

Previous Articles    

Large Models with Embedded Expert Knowledge for Highly Generalized Interpretable High-speed Train Bogie Axle-box Bearing Fault Diagnosis

GUO Zhibin1, XIE Jingsong1, WANG Tiantian1,2, PAN Tongyang1, NIU Buzhao1,3   

  1. 1. School of Traffic & Transportation Engineering, Central South University, Changsha 410075;
    2. College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410082;
    3. National Engineering Research Centre, CRRC Qingdao Sifang Co., Ltd., Qingdao 266111
  • Received:2025-08-11 Revised:2026-03-06 Published:2026-10-14

Abstract: Axlebox bearings constitute a core component of high-speed train running gear. The high-load, high-speed characteristics render failures prone to triggering safety risks. Existing deep learning-based fault diagnosis models exhibit two major limitations:insufficient generalization capability, making them difficult to adapt to the variable operating conditions of high-speed trains; and a lack of interpretability in diagnostic results, failing to meet the credibility requirements of intelligent maintenance operations. Inspired by large models’ strengths in semantic understanding and task generalization, a highly generalisable, explainable fault diagnosis large model incorporating expert knowledge is proposed. This model achieves dual outputs of diagnostic results and supporting rationale through multimodal fusion and integrated reasoning of monitoring data, task texts, and expert knowledge. Its core strengths lie in:① efficiently embedding expert knowledge via retrieval-enhanced generation to bridge the knowledge gap in data-driven models; ② enhancing cross-scenario generalization through multimodal feature alignment; ③ integrating multi-model diagnostic rationale to perform diagnostic attribution via a reasoning large model. Experimental results demonstrate that the proposed model achieves an average diagnostic accuracy of 89.8% in cross-operating-condition generalization tasks and 87.1% in cross-data generalization tasks. Both figures significantly outperform traditional domain adaptation models, which generates diagnostics aligned with professional logic, effectively enhancing the reliability and explainability of intelligent operations and maintenance.

Key words: fault diagnosis, expert knowledge, prognostics and health management, intelligent maintenance, large model

CLC Number: