鑑於半導體產業在數據搜集上的成本降低,擁有大量的即時製程資訊,如何運用收集機台上各項參數的時間序列資料,以進行錯誤偵測與分類、虛擬測量,乃至批次控制,為構建現今智慧製造系統的基石之一。本研究著眼於製造系統中的分析錯誤偵測與分類資料,萃取與機台狀態特性相關的時間特徵,以非監督式學習將原始多維度時間序列簡化成單一虛擬時間訊號,此時間卷積特徵經指數加權移動平均後作為機台單一指標,並藉其偵測與預測機台設備的異常失效。藉由觀察機台指標與實際製程資訊,本研究發現機台發生錯誤前,該指標均有明顯的下降或上升的趨勢。從論文中的蝕刻系統個案來看,以廣義變異數與馬氏距離作為卷積計算出來的時間特徵,皆出現與機台錯誤發生時機有關的模式。此時間卷積特徵能夠以非監督式學習的方式進行特徵萃取,同時保留晶圓製程的時間資訊,提高特徵的代表性。綜合來說,依時間卷積特徵所建立的健康指標,在其解釋上、資料不平衡的處理上都跟過去的資料分析方法有所不同,實務上,透過建立單一指標的健康指數,能夠有效幫助工程師在時間視窗上進行機台是否錯誤的判讀。
Given the cost reduction of data collection in the semiconductor industry and the large amount of real-time process information available, utilizing the temporal data of machine sensors for fault detection and classification, virtual metrology, and run-to-run control is more and more critical and challenging. In this thesis, the fault detection and classification data are used to extract the temporal convolutional features out of the machine sensor readings. By applying the exponentially weighted moving average schema to the temporal convolutional features, the machine condition is monitored in an unsupervised manner and the failure evaluation is performed. Through evaluating the estimated machine condition with actual process information, our study shows that there is a significant upward or downward trend before the machine failures. In the case study of an etching process, we observe the associated patterns between the faults and the temporal features convoluted by the kernels of generalized variance and Mahalanobis distance. Due to its unsupervised characteristic, the issue resulted modeling bias from data imbalance can be avoided. In summary, the health indicators built upon the temporal convolutional features have enabled analyzing the machine condition with respect to the temporal windows in a more comprehensive perspective.