A Fault Diagnosis-Oriented Data-Driven Embedding Fine-Tuning and Domain-Aware Hybrid Retrieval Method for Intelligent Operation and Maintenance

Xueyi Li; Gang Li; Jiannan Dong; Yun Kong; Wenyang Hu; Tianyang Wang

doi:10.37965/jdmd.2026.1505

Authors

Xueyi Li 1 College of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040, China 2 Department of Mechanical Engineering, Tsinghua University, Beijing 100084, China https://orcid.org/0000-0003-0335-0594
Gang Li 1 College of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040, China
Jiannan Dong 1 College of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040, China
Yun Kong 3 School of Mechanical Engineering, Beijing Institute of Technology, Beijing 100081, China
Wenyang Hu 2 Department of Mechanical Engineering, Tsinghua University, Beijing 100084, China
Tianyang Wang 2 Department of Mechanical Engineering, Tsinghua University, Beijing 100084, China

DOI:

https://doi.org/10.37965/jdmd.2026.1505

Keywords:

fault diagnosis; intelligent maintenance; hybrid retrieval; domain knowledge; fine-tuning

Abstract

To improve the support of engineering knowledge resources for fault diagnosis and intelligent maintenance, this paper proposes a fault diagnosis-oriented data-driven embedding fine-tuning and domain-aware hybrid retrieval method. The proposed framework is intended for industrial fault diagnosis and intelligent operation and maintenance, and is instantiated and validated in wind turbine operation and maintenance. Domain-specific training data are constructed to fine-tune the embedding model, enhancing its representation of fault concepts, failure mechanisms, and maintenance knowledge. Query expansion, dense retrieval, sparse retrieval, and domain term matching are then integrated with weighted fusion and cross-encoder re-ranking to obtain relevant and low-redundancy diagnostic contexts. Compared with the general retrieval-augmented generation (RAG) baseline, the proposed method increases Recall@5 by 57.33 percentage points, achieves an MRR of 98%, increases the key point coverage rate by 4.8 percentage points, reaches a factual support rate of 97.83%, and reduces the hallucination rate to 2.17%. These results demonstrate its effectiveness in improving retrieval accuracy and generation reliability for fault diagnosis and maintenance decision-making in the wind turbine operation and maintenance scenario.