基於作動深度學習於物理氣相沉積設備故障預測與健康管理

Abstract

隨者科技不斷演進與發展，有許多人工智慧與資料科學相關的新技術出現用來輔助解決現實生活中的問題。目前世界上產值相當高的製造業中因為需求不斷的提升，許多企業都將資料科學相關的技術用於提升產能以及降低成本。而為了要兼顧成本以及產能的考量，設備的故障預測以及健康管理是很重要的一個方法，透過監控機台狀況來使機台能在良好的情況下運作，在必要時才進行保養及更換。因此近幾年不僅學界以及業界都投入許多資源研究相關的議題。本研究針對物理氣相沉積設備，建立基於資料導向的健康評估方式及異常偵測模型。本研究與台灣頂尖面板製造公司合作，使用較有解釋性的特徵來建構機台健康指標，並透過實證資料驗證本研究所提出之方法。除此之外，也提出基於深度學習相關的技術萃取特徵並使用異常偵測的模型及驗證於公開資料集上。本研究貢獻在於根據不同使用情境建立不同監控機台狀況的方法，並透過健康指標來提早預警機台故障，同時權衡保養成本及突發錯誤導致的產能損失。

With the continuous evolution and development of technology, many technologies related to artificial intelligence and data science have emerged to assist in solving real-world problems. Due to the continuous increase in demand in the manufacturing industry which has a high value in the world now, many companies use data science to increase production capacity and reduce costs. Prognostic and Health Management is an important method that considers cost reduction and production capacity at the same time. The machine can operate in a good condition, and only carry out maintenance and replacement when necessary by monitoring the machine. Therefore, not only academia but also the industries have devoted a lot of resources to research related issues. In this study, a data-driven health assessment method and anomaly detection model was established for physical vapor deposition equipment. This study cooperated with a leading panel manufacturing company in Taiwan. Use explanatory features to construct health indicators, and validate the proposed method through empirical data. In addition, anomaly detection models based on the deep learning features are proposed and validated by public datasets. The contribution of this study is to establish different monitoring methods for the machine according to different scenarios and to use health indicators to waring machine failures in advance. Taking into consideration maintenance costs and capacity losses caused by sudden errors at the same time.