基於自編碼器的動態管制界線偵測製程與量測異常

Abstract

隨著數據量的增長以及技術的成熟，現今各個行業都開始仰賴各種人工智慧及資料科學相關的技術來幫助解決各種問題，而在擁有大量資料並且器材昂貴的製造業中，更為準確的量測以及進一步的預測更是需要資料科學的技術輔助，其中異常偵測又為製造業最常探討的議題。異常偵測旨在透過機台數值的分析來得知機台的健康狀況並且進一步進行機台維護策略，藉此取代過往只能透過人工檢查機台的方式來節省時間並減少不必要的成本，但過往異常偵測的議題往往圍繞在機台本身，每當機台數值發生異常便認為機台可能需要進行維護，實際上造成數值異常的原因有另一大主因，便是量測異常，也就是負責偵測數值的感測器本身發生了異常，此時需要維護的不是機台本身而是感測器。因此，本研究以自編碼器為基礎提出了一個能夠線上動態偵測製程與量測異常的PSAD架構，並且透過一系列的實驗設計來嘗試證實架構的有效性及穩固性。本研究貢獻在於能夠偵測到過往較少探討的量測異常情形並且同時對兩種異常進行線上動態偵測，以幫助各種使用到感測器進行數值量測的實際場域做出更好的維護決策。

With the growth of data volume and the maturity of technology, various industries have begun to rely on various artificial intelligence and data science-related technologies to help solve various problems. In the manufacturing industry with a large amount of data and expensive equipment, more accurate measurement and further prediction require technical assistance from data science. Among this, anomaly detection is the most frequently discussed topic in the manufacturing industry. The purpose of anomaly detection is to know the health status of the machine through the analysis of the machine value and further implement the machine maintenance strategy to replace the manual inspection of the machine in the past to save time and reduce unnecessary costs. However, in the past, the issue of anomaly detection often revolved around the machine process itself. Whenever the machine value was abnormal, it was considered that the machine might need to be maintained. In fact, there is another major reason for the anomaly. That is, the sensor responsible for detecting the value itself has an anomaly. At this time, it is not the machine itself but the sensor that needs to be maintained. Therefore, based on the autoencoder, this research proposes a PSAD architecture that can dynamically detect process and sensor anomalies online, and try to verify the effectiveness and robustness of the architecture through a series of experimental designs. The contribution of this research lies in the ability to detect sensor anomalies that have been less discussed in the past and perform online dynamic detection of two anomalies at the same time, so as to help various practical fields that use sensors for numerical measurement to make better maintenance decisions.