A Study on LLM-Driven Knowledge Graph Construction and Production Anomaly Inference for Semantic Standardization of Multi-Factory Data in Automotive Manufacturing

Source abstract

Multi-plant automotive-component data contain inconsistent field names, device codes, quality codes, and unstructured production logs. This study presents an LLM-driven method for manufacturing semantic standardization and knowledge-graph reasoning. Equipment signals, production beats, quality inspections, maintenance work orders, VIN/OCR results, shutdown events, and process files are integrated to construct cross-factory semantic mapping tasks. Sentence-BERT and GPT-4o-mini perform field alignment and context correction; BERT-CRF supports entity extraction, while GraphSAGE and R-GCN infer typed anomaly relations. KG-RAG and an LLM Agent are further used for quality tracing and case retrieval. Experiments cover 4 plants, 9 production lines, 156 field categories, 780,000 production records, and 120,000 maintenance and quality texts. Under leave-one-factory-out evaluation, Sentence-BERT plus GPT-4o-mini achieves 94.1% field-matching accuracy and 94.6% precision. The 94.1% accuracy is 21.3 percentage points above rule-based mapping at 72.8%, corresponding to a 29.3% relative improvement. GraphSAGE-R-GCN obtains Hits@10 of 0.876 on the mixed split and 0.842 under factory holdout. The complete KG-RAG-Agent system reaches retrieval P@20 of 0.842, R@20 of 0.816, and 92.1% answer accuracy, while reducing mean anomaly-tracing time from 2.8 h to 54 min. These results show that semantic constraints and typed graph reasoning support cross-factory anomaly transfer and evidence-grounded knowledge reuse.