隨著科技迅速發展，現今大多工廠漸漸往工業4.0邁進，而穩定高效的生產效率、良好的產品品質與高穩定度的設備機具，皆是製造業在全球市場中脫穎而出的重要關鍵。過去在設備維護方法中使用被動維護的方式，此種方式不僅會造成產能下降更會導致機台壽命減少，以致於成本增加，因此近年來預測性維護成為主流方法，在設備上安裝感測器，並運用物聯網、雲端運算、數據分析等方式挖掘資料中有意義的資訊加以利用，在故障發生前提早發現異常並迅速維修，以改善過去被動維護方式的種種缺點。 本研究使用A公司所提供之塗佈機運轉數據，使用長短期記憶網路（Long Short-Term Memory, LSTM）及單類支持向量機（One-Class Support Vector Machines, OCSVM），兩種方法混和建立半監督式學習模型，欲達成預測性維護方法中的預診斷與健康管理（Prognostics and Health Management, PHM）。實驗結果評價指標準確率達99.27%、召回率達100%，以及F2-Score達81.83%，且預測模型相較於實際的張力異常紀錄能提前20秒檢測到異常徵兆，以利機台檢驗負責人能於機台故障前做出決策確保產線順暢，並降低因機台異常導致產品良率下降及機台壽命減短等問題。 ;With the rapid development of technology, most factories nowadays are gradually moving towards Industry 4.0, and stable and efficient production efficiency, good product quality and high stability of equipment and machinery are all important keys for the manufacturing industry to stand out in the global market. In the past, passive maintenance was used in equipment maintenance methods, which not only caused a decrease in production capacity but also led to a reduction in machine life, resulting in increased costs. In order to improve the shortcomings of the past reactive maintenance methods, we use the information provided by Company A to provide the sensors. This study, a semi-supervised learning model was built using a mixture of Long Short-Term Memory (LSTM) and One-Class Support Vector Machines (OCSVM) using the data provided by Company A for the operation of the coating machine, in order to achieve Prognostics and Health Management. The experimental results were evaluated as accuracy 99.27%, recall 100%, and F2-Score 81.83%. The prediction model can detect abnormal symptoms 20 seconds earlier than the actual tension abnormality records, so that the machine inspection personnel can make decisions before the machine failure to ensure smooth production line and reduce the problems of decreased product yield and shortened machine life caused by machine abnormality.
