| 摘要: | 隨著工業4.0的蓬勃發展，製造業正積極導入人工智慧（Artificial Intelligence, AI）、物聯網（IoT）、大數據（Big Data）、雲端運算(Cloud Computing)、機器人技術以及自動化等現代技術使工廠轉型為智慧製造。利用這些高科技技術建立智慧工廠，將產品組件和設備上安裝感測器以利收集即時數據與分析，預測設備故障的可能性和時間點，實現預測性維護同時實現更多製造和供應鏈運營的智慧決策。在傳統的預防性維護模式中，工廠通常會根據固定的時間間隔或使用時間來進行設備檢修，但這種方法存在許多缺點，如維護成本高、資源浪費、無法即時調整等。若採用預測性維護則可以避免過度維護所增加的停機時間，且能提前檢測到異常進行維護，提高設備的運行穩定性。智慧製造的關鍵在於穩定且高效的生產效率，有著優質品質和可靠的設備將成為製造業在全球市場中取得成功的重要因素。
本研究以某塗佈製程公司的感測器所收集之數據為研究對象，以預診斷與健康管理 (Prognostics and Health Management, PHM)為架構，進行設備預測模型。透過利用長短期記憶(Long Short-Term Memory, LSTM)自編碼器(AutoEncoder)建立異常檢測分類模型，結果顯示模型的評價指標如準確率達99.96%、召回率達100% 以及 F2-Score 為96.7%，整體成效皆有良好表現，且透過計算健康指標制定機台健康管理規範，提高機台健康透明度。再使用一維卷積神經網路(1D Convolutional Neural Network, 1D CNN) 結合LSTM建立預測主速度模型，透過分析歷史數據找出判斷異常的跡象，並使用不同超參數與滑動窗口來進行效能比較，評估出最佳模型且驗證其效能。本研究以15秒的時間序列資料預測未來第20秒的主速度值，最終實驗結果顯示設備可於14秒前偵測到異常速度下降，模型效能在R2達到96%、MSE為2.2。基於上述結果，可作為設備健康狀況的評估，提前規劃維修保養、保障機台的持續穩定運轉、提高生產效率以及優化整體生產運營，達到預測性維護之目的。
;With the vigorous development of Industry 4.0, the manufacturing industry is actively adopting modern technologies such as the Internet of Things (IoT), Artificial Intelligence (AI), Big Data, Cloud Computing, robotics, and automation to transform factories into smart manufacturing facilities. By integrating automation equipment and combining it with AI, smart factories are established to collect real-time data through sensors installed on product components and equipment, predict the likelihood and timing of equipment failures, achieve predictive maintenance. In traditional preventive maintenance models, factories typically perform equipment maintenance based on fixed time intervals or usage time, which has many disadvantages such as high maintenance costs, resource waste, and the inability to adjust in real time. Adopting predictive maintenance can avoid excessive downtime and detect anomalies for maintenance in advance, thereby improving the stability of equipment operation. The key to smart manufacturing is to achieve high-quality products, and reliable equipment, which will be important factors for success in the global market for the manufacturing industry.
This study focuses on the data collected from the sensors of a coating process company, aiming at prognostics and health management (PHM) for equipment anomaly detection. A predictive anomaly detection classification model is established using Long Short-Term Memory (LSTM) AutoEncoder. The results show that the model achieves important evaluation indicators such as accuracy rate of 99.96%, recall rate of 100%, and F2-Score of 96.7%, demonstrating excellent overall performance. Additionally, the Health Index calculation serves as a machine health management mechanism. Furthermore, a predictive model is established using 1D Convolutional Neural Network (1D CNN) combined with LSTM. By analyzing historical data to identify signs of anomalies and using different hyperparameters and sliding windows for performance comparison, the optimal model is evaluated. This study uses 15 seconds of time series data to predict the main speed value at the 20th second in the future. The final results show that the equipment can detect an abnormal speed drop 14 seconds in advance, with the model′s performance achieving R2of 96% and MSE of 2.2. Based on these results, it can be used to assess equipment health, schedule maintenance in advance, boost production efficiency, and optimize overall operations, achieving predictive maintenance goals. |
