摘要
Real-time inspection in rubber-hose manufacturing demands not only rapid defect localization but also actionable root-cause explanations—capabilities that conventional vision pipelines rarely deliver in a deployable manner. This study presents an end-to-end framework integrating a dual-branch YOLO network, a manually curated hose-defect knowledge graph (KG), and a retrieval-augmented large language model (RAG-LLM). The YOLO backbone performs image-level screening and defect localization in a single forward pass, achieving 86.31% mAP on edge hardware while sustaining production-grade throughput. When a defect event is triggered, a dual-channel retriever gathers k-hop KG subgraphs and dense-vector passages from SOPs and maintenance logs. The evidence is linearized into a schema-constrained prompt, enabling the LLM to generate an operator-oriented diagnostic report with a causal narrative, targeted corrective actions, evidence references, and an evidence-grounded confidence score. For safety–critical deployment, the RAG-LLM is governed by evidence-only constraints and post-generation consistency checks; low-evidence or low-confidence cases trigger abstention and escalation to manual inspection, ensuring accountable human-in-the-loop decision support. Operator feedback is logged and incorporated through routine knowledge maintenance, refining defect–cause–action relations in the KG with provenance preserved for auditability. Beyond offline testing, a long-term shop-floor study under factory-defined UCL monitoring confirms stable alert behavior, bounded false alarms, and manageable workload while maintaining timely, factually grounded explanations. Deployment results also indicate economic impact: removing one inspector per line saves roughly USD 50,000 annually, and a 12% reduction in false negatives lowers scrap and warranty costs, shortening payback to under 12 months.