LLM-based Causal Analysis for Manufacturing Defects Through Knowledge Graph Construction in Cigarette Quality Control

Cigarette appearance defects can adversely affect brand reputation, yet current quality control systems lack causal analysis capabilities. This study presents a framework combining large language models with knowledge graphs to extend defect detection toward causal diagnosis. We fine-tuned the DeepSeek-R1-Distill-Llama-8B model using an optimized 4-bit QLoRA strategy on domain-specific corpora. An ablation study on an NVIDIA A100 identified QLoRA as a suitable training strategy, balancing memory efficiency (6.53[Formula: see text]GB VRAM) with computational speed (1.91[Formula: see text]s/step) compared to full fine-tuning and standard LoRA. Using a Top-M constraint with [Formula: see text], the model achieved a precision of 1.00 in causal extraction. A human-in-the-loop mechanism was employed to validate 11,545 relationships, yielding an expert-weighted score of 3.85/5.00. The resulting knowledge graph organizes causes into five hierarchical levels, facilitating systematic troubleshooting in manufacturing environments.