融合大型語言模型與多源製造資料之精密加工 監控、品質預警與產線決策支援

金屬加工業面臨少量多樣、交期縮短與缺工，急需將分散的設備、生產與品質資料即時轉為決策 資訊。本文以六條自動化產線、12 台 CNC 與 6 台機械手臂之精密加工廠為場域，建構以大型語言模型為 核心之生產戰情室，整合機聯網、製造執行系統與巡檢量測資料，經六個月量產驗證。結果顯示，兩台 執行同一程式的配對機台，主軸負載相差 23.5%（正常配對相差不到 1%），而此時品檢尺寸全數合格、 Cpk 最低仍達 2.46；系統可據此在品檢察覺異常之前，指出高風險的機台與程式組合。系統並能產生管制 圖與製程能力指標、回溯換刀等事件歸因；亦可由歷史報工反推機台與人員能力，支援兼顧交期、產能、 品質與人力之派工決策。

Under high-mix low-volume production, shortened delivery schedules, and labor shortages, turning dispersed equipment, production, and quality data into timely, actionable decisions is a critical challenge for the metalworking industry. This article develops a production control room centered on a large language model (LLM) for a precision machining factory with six automated lines, 12 CNC machines, and six industrial robots, integrating machine-network, manufacturing execution system (MES), and in-process inspection data, and validated over six months of production. Two paired machines running the same program showed a 23.5% difference in spindle load (normal pairs differ by less than 1%) while every inspected dimension was within specification and the lowest Cpk was 2.46; the system can therefore flag high-risk machine and program combinations before inspection detects any anomaly. It also generates control charts and process capability indices with event-based root-cause tracing, and infers actual machine and operator capability from historical records to support dispatching decisions that balance delivery, capacity, quality, and workforce.