應用時間序列相似度量測方法於異常偵測與分類

Abstract

本論文主要探討如何在滿足正常狀態之工作範圍的條件下, 偵測出異常的時間序列資料並將其分類,且該方法與系統對時間序列位移量與雜訊大小的敏感性越小越好。 相較於傳統的工作範圍偵錯法只能偵測出超出工作範圍的異常, 以及主成份分析法容易損失資料資訊, 造成微小異常的不易偵測與分類, 本論文以直接量測時間序列的相似度來改進這些缺失。 本研究以半導體製程中, 質流控制器的五種異常狀態為例, 希望在滿足正常工作範圍的條件下, 能將異常訊號偵測出來, 並可將正常與異常共六種類別分類。 首先, 使用歐基里德距離量測法與動態時間扭曲法, 測量出時間序列間的相似度。 接著, 利用多重維度度量法的二維知覺圖與 $F$-試驗值初步評估系統成效。 之後, 再使用 $k$-最鄰近點分類法來進一步驗證偵錯與分類結果, 並比較不同相似度量測方法的系統成效差異。 同樣的系統除了可應用在本論文所舉的半導體製程例子之外, 亦可應用於與時間序列有關的議題上, 例如金融與醫學等領域。

The purpose of this thesis is to detect and classify abnormal time-series data under normal working range. In addition, the less sensitive the offset and noise of time-series data is, the better the system is. The traditional working-range-detection method is just used to detect faults over normal working range, and using principal components analysis is not easy to detect and classify little abnormality because it loses information of data. The thesis improves these shortcomings by measuring the similarity of time-series data directly. This research takes five faults of mass flow controller in semiconductor manufacturing process for example. It is expected that faults can be detected and classified under normal working range. First, we measure similarity between different time-series data with Euclidean distance metric and dynamic time warping. Second, we make use of two-dimensional perceptual map of multidimensional scaling and F-test to evaluate the preliminary performance of the system. And then, we validate the results of fault detection and classification by using k-nearest neighbors classification, and comparing the performance of the system using different similarity measurement. The same system can be applied not only to semiconductor manufacturing process in this thesis, but to the issues about time-series data in financial and medical domain.