GUIDE: A knowledge-unified framework for assembly anomaly diagnosis via graph-enhanced retrieval-augmented generation

Assembly anomaly diagnosis in complex manufacturing systems still relies heavily on manual inspection and experience-driven practices, leading to delayed responses, fragmented knowledge use, and inconsistent diagnostic outcomes. Although Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) can improve access to industrial knowledge, similarity-based retrieval often fails to preserve the diagnostic sequence required for root-cause tracing, verification planning, and action recommendation. To address this limitation, this study proposes the Graph-Unified Intelligent Diagnosis Engine (GUIDE), a graph-enhanced RAG framework for assembly anomaly diagnosis. GUIDE organizes heterogeneous assembly records into a diagnosis-oriented knowledge graph (KG) and retrieves topology-constrained diagnostic evidence chains that link observed anomaly phenomena to candidate root causes, verification operations, and corrective measures. The retrieval process combines community-level evidence localization, ontology-constrained causal path traversal, candidate evidence ranking, and structured diagnostic context construction for LLM-based reasoning. Before returning diagnostic recommendations, GUIDE verifies generated diagnostic claims against retrieved evidence and graph-defined relation constraints to reduce unsupported causal explanations and weakly grounded corrective recommendations. Experiments on 286 real-world aero-engine assembly fault records show that GUIDE outperforms parametric, text-based, graph-based, and a root-cause-oriented KG baseline. GUIDE achieved a faithfulness score of 0.913 and an answer relevance score of 0.910, while reducing noise sensitivity to 0.155. These results indicate that topology-constrained graph-enhanced retrieval with claim-level verification can improve the reliability, stability, and interpretability of intelligent diagnosis in complex assembly environments.