Abstract: Objective To address the critical challenge in marine propulsion shafting fault diagnosis, where fault types are identifiable but specific faulty components remain difficult to locate, this study proposes a novel intelligent diagnostic framework that integrates data-driven analytics with knowledge-based reasoning. Methods The proposed framework, termed ShaftAgent, is established upon a three-layer architecture. First, the Mechanism Modeling Layer extracts high-dimensional features from component-level vibration signals and auxiliary systems. Second, the Interpretable Analysis Layer employs XGBoost for robust fault classification and introduces a component-level SHAP attribution aggregation method to achieve automated fault localization. Finally, the Knowledge-Enhanced Reasoning Layer constructs a hierarchical “Equipment-Phenomenon-Mechanism-Fault” knowledge graph. By leveraging multi-stage prompt engineering, this layer drives a Large Language Model (LLM) to generate comprehensive diagnostic reports, while a consistency verification mechanism ensures that all outputs strictly adhere to physical laws. Results Experimental validation demonstrates that ShaftAgent achieves a fault classification accuracy of 96.8% and a component localization accuracy of 94.2%. Furthermore, diagnostic reports generated by the framework received an average expert score of 4.70. Ablation studies confirm the indispensable contribution of each functional module, and representative case analyses illustrate the end-to-end diagnostic workflow from raw multi-source vibration signals to actionable maintenance recommendations.Conclusion ShaftAgent effectively overcomes the limitations of traditional diagnostic methods regarding insufficient localization precision and poor interpretability. The findings validate the feasibility of employing knowledge-graph-constrained LLMs for industrial fault diagnosis, offering a transformative technical paradigm for the intelligent operation and maintenance of marine propulsion systems.