工業大數據之異常偵測分析挑戰：以紙業為例

Abstract

異常偵測應用包括防治金融詐欺、改善工業設備營運等。以紙業為例，造紙流程涉及上千個感測器，形成一個高維度且巨量的資料集。德國PTS指出每次斷紙事件將造成至少6,000歐元(相當於210,000新臺幣)損失，斷紙頻率約每日發生6至9次。此外，斷紙造成的能耗虛耗約占總生產量的2%至7%，不僅是經濟上的虧損還會浪費大量資源。本研究探討異常偵測技術與其應用在斷紙分析之發展狀況，我們以非線性支持向量機分類方法(Non-linear SVM, N-SVM)分析國外斷紙資料集，並從分析結果論述工業數據之異常偵測挑戰、異常因素探索以及資料品質的重要性。

Anomaly detection technology has been widely applied to varied areas, e.g., fraud detection for credit cards, fault detection in safety critical systems, and so on. In paper industry, the paper-making process involves in thousands of sensors, which forms high-dimensional large amounts of data. In analyzing such big data, the state-of-the-art methods probably suffer from computation and distortion problems. According to PTS, a paper break costs around 210,000 NT$, and it occurred 6 to 9 times per day. In addition, paper breaks cause 2–7% of the total production loss. In this paper, we survey the taxonomy of anomaly detection methods and their applications in analyzing paper breaks. Moreover, we adopt the non-linear SVM method (N-SVM) to analyze the paper breaks dataset. Finally, we discuss the findings and illustrate the importance of anomaly exploration and data quality.