A Fault diagnosis method for shafting systems integrating physics-driven knowledge and large language model reasoning

Objective To address the challenge in marine propulsion shafting fault diagnosis where fault types can be identified but the faulty equipment is difficult to localize, this study proposes an intelligent diagnostic method that integrates mechanism-based feature modeling with knowledge graph-constrained reasoning.

Methods A three-layer ShaftAgent diagnostic framework is developed. The mechanism modeling layer is used to extract equipment-level vibration features and auxiliary system features. The interpretable analysis layer employs XGBoost for fault classification and introduces an equipment-level SHAP attribution aggregation method to enable automatic localization of faulty components. The knowledge-enhanced reasoning layer is designed to build a hierarchical knowledge graph of “equipment-phenomenon-mechanism-fault”, which, together with multi-stage prompt engineering, guides large language models to generate diagnostic reports. A consistency verification mechanism is further incorporated to ensure that the generated outputs conform to physical laws.

Results Experimental results show that ShaftAgent achieves a fault classification accuracy of 96.8%, an equipment localization accuracy of 94.2%, and an expert-evaluated comprehensive score of 4.70 for diagnostic reports. Ablation experiments validate the effectiveness of each module.

Conclusion The results indicate that ShaftAgent can effectively address the limitations of traditional methods in terms of insufficient equipment-level localization capability and weak interpretability. Moreover, the study verifies the feasibility of applying large language models to industrial fault diagnosis under knowledge graph constraints, providing a new technical pathway for intelligent operation and maintenance of marine propulsion shafting systems.