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

Bogie axle-box bearings are among the most critical core components in the running gear of high-speed trains. Due to the high load-bearing capacity and high operating speed, any damage to them can severely affect the stability and safety of high-speed trains. Despite advances in deep learning-based fault diagnosis, existing models lack generalization for variable operating scenarios and offer poor interpretability. Inspired by the excellent performance of large models such as Deepseek and ChatGPT in semantic understanding, reasoning ability, and task generalization, this paper proposes an expert knowledge-embedded, highly generalizable, and interpretable large model for such fault diagnosis. A knowledge vector database based on expert knowledge texts is constructed using retrieval-augmented generation (RAG) technology. The multimodal large fault diagnosis model that integrates monitoring data, task texts, and expert knowledge vectors is constructed to output vectorized diagnostic results and textual information of diagnostic criteria. The results demonstrate that the proposed large model for axle-box bearing fault diagnosis can effectively fuse monitoring data and textual information, and exhibits excellent performance in cross-operating condition and cross-data generalization diagnosis tasks. Additionally, the large model can effectively embed expert knowledge and output diagnostic bases, thereby enhancing the reliability of operation and maintenance results.