Knowledge extraction and retrieval-augmented generation for intelligent maintenance of wind power equipment based on graph attention networks

The adoption of deep learning in intelligent operation and maintenance of wind turbines has improved fault detection and classification accuracy. However, most existing studies focus primarily on the fault identification stage, while the crucial task of generating intelligent and personalized maintenance recommendations after a fault occurs has been largely overlooked, reducing the practical guidance value of fault diagnosis systems in real-world industrial scenarios. Retrieval-augmented generation (RAG) technology enhances the knowledge representation capability of large language models by incorporating external documents, but faces insufficient context modeling in complex semantic environments. To address these challenges, this paper proposes a graph attention network (GAT)-enhanced knowledge extraction and retrieval-augmented generation method, termed GAT-RAG, which constructs a unified knowledge graph integrating textual, visual, and structured data, and leverages graph neural networks to model semantic dependencies and contextual relationships between entities, enabling structure-aware retrieval within the RAG framework to preserve multi-hop semantic connections and improve the logical coherence of generated outputs. Experimental results on wind turbine datasets show that GAT-RAG outperforms the baseline RAG model, achieving relative improvements of 27.9% in BLEU and 25.3% in ROUGE-L scores, confirming the effectiveness of GAT-based structural modeling in enhancing both the quality and interpretability of generated maintenance recommendations, and providing a promising direction for integrating intelligent fault diagnosis with context-aware decision support in industrial operation and maintenance systems.