WT-Fault-LLM:Knowledge Graph Enhanced Large Model for Wind-turbine Faults

This is the authors' abstract. We don't add key points for this paper.

This study introduces an innovative framework integrating Large Language Models (LLMs) with knowledge graphs to significantly enhance wind turbine fault diagnosis, specifically addressing critical LLM limitations in industrial settings, such as knowledge inaccuracies and deficiencies in symbolic reasoning. While LLMs exhibit proficiency in general domains, their application in specialized vertical fields remains constrained. Knowledge graphs offer a solution by providing structured, interpretable reasoning through networks of triplets. The proposed architecture employs a three-layer "knowledge storage-retrieval-generation reasoning" structure to integrate heterogeneous data sources—including wind turbine manuals and maintenance records—into a domain-specific knowledge graph that explicitly models fault entities, their attributes, and causal relationships. A novel dual-channel semantic retrieval mechanism is implemented, concurrently leveraging graph-based reasoning paths within the knowledge graph and semantic similarity matching via a vector database, thereby substantially improving adaptability to complex operational scenarios. Utilizing local deployment with the DeepSeek-R1 model and Retrieval-Augmented Generation (RAG) techniques, the framework transforms both structured and unstructured data into contextual constraints through dynamic prompt engineering, resulting in markedly improved fault diagnosis accuracy. Case studies establish the system’s capacity for sophisticated multi-hop causal reasoning, such as identifying that "gearbox noise" can originate from underlying causes like lubricant contamination. Future research will prioritize the integration of multi-modal data, including vibration spectrums, to accelerate the transformation of industrial operations from experience-based practices to knowledge-driven methodologies. This work provides a valuable technical paradigm for the reliable application of LLMs in vertical industrial domains.