Construction method of knowledge enhanced industrial-LLM for intelligent manufacturing operation and maintenance

In response to the limitations of general retrieval-augmented generation (RAG) methods in handling term ambiguity and multi-hop causal reasoning in industrial scenarios, this paper proposes a knowledge-enhanced industrial large language model construction method for intelligent manufacturing operation and maintenance. It integrates knowledge graph entity alignment with RAG to enable dynamic knowledge recall, and employs low-rank adaptation for efficient fine-tuning using multi-source heterogeneous operation and maintenance data. The method also incorporates a constraint decoding mechanism based on real-time knowledge graph verification to suppress hallucinations and ensure output interpretability. Experimental results demonstrate significant improvements in fault diagnosis accuracy and maintenance plan generation quality compared to baseline models.