| 摘要: | 異常偵測(Anomaly Detection)是指問題在發生的當下或提前被找出來，是個在現實中資料分析裡常見的問題，像是信用詐欺與醫療問題，而在製造業的製程中，維修人員常會遇到機器故障、零件損壞、耗材斷裂等情況，造成瑕疵發生或製程中斷，因此對於機具及耗材的維護保養或更換，都不希望在問題發生後才去解決。異常偵測常見的資料中，都是僅有極少量的異常資料跟大量正常資料，因此會難以分辨出資料異常的特性，在對資料抽樣的時候我們常見的作法有調整資料抽取的數量、特徵欄位的選取(Feature Selection)、跟計算資料間的距離相似度。在過去的研究中，以距離為算法的抽樣方法都沒有考量到資料在時間序列上的不同時間長度的相似度，因此在本研究中，我們將針對時間序列的資料提出以DTW為計算不同時間長度為基礎的抽樣方法。
關於衡量時間序列的相似度，我們採用的方法是動態時間校正(DTW)，其中兩筆時間序列的DTW距離愈小，代表兩者之間愈相似。相較於歐氏距離，DTW可以用來計算不同的時間長度，因此在我們已知異常的情況下，可以根據相對應的表現對樣本的長度放寬或限縮。在本篇論文的實驗中，我們便首先定義了發生異常前的一段時間作為原本的異常資料，並用DTW將跟其最為相似的資料作為樣本，這樣可以使我們考量到的不僅有發生異常前的時間點，還有可能找到隱藏在不同時間長度中的異常片段。
論文中使用的資料是來自一間半導體公司其中的一個製程資料，由於耗材在製程中損壞的時間難以預估，使得異常偵測的準確率相對較低，本研究希望以我們提出的抽樣方法提升類似此情況的時間序列異常偵測。本研究會將提出的抽樣方法、隨機位置抽樣的過抽樣方法與以歐基里德距離抽樣的方法，在LSTM跟SVC兩個分類模型中進行比較。最後從我們的實驗結果可以觀察到，我們提出的抽樣方法在模型的表現都比其他兩個方法優秀。
;Anomaly detection implies that the problem is found at the moment of occurrence or in advance. It is a common problem in data analysis in reality, such as credit fraud and medical problems. In the manufacturing process, maintenance personnel often meet the event of machine failure, damaged parts, broken consumables, etc., resulting in defects or interruption of the process. We are not willing the maintenance or replacement of equipment and consumables will be found out after the problem occurs. Among the common data for abnormal error detection, there are only a very small amount of abnormal data and a large amount of normal data, which causes it is difficult to distinguish the characteristics of abnormal data. Sampling is a common method to solve the problem by adjust the number and characteristics of data extraction, feature selection, and distance similarity. In past researches, the sampling method based on the distance between sequences did not consider the similarity of sequences at different time lengths in time series. Therefore, in this study, we propose to use DTW as a calculation method in time series sampling method.
For measuring the similarity of time series, the method we use is Dynamic Time Warping (DTW). The smaller the DTW distance between two time series, the more similar they are. Compared with Euclidean distance, DTW can be used to calculate different time lengths, so in the case of known anomalies, the length of the sample can be relaxed or limited according to the corresponding performance. In the experiment of this paper, we first define the period before the abnormality occurs as the original abnormal data, and use DTW to take the most similar data as samples, so that we can consider not only data before the abnormality occurs, but also possible to find anomalous fragments hidden in different lengths of time.
The data used in this paper is from one of the process data of a semiconductor company. Since it is difficult to predict the time when consumables are damaged in the process, the recall rate of abnormal detection is relatively low. We hope to improve the performance by the sampling method we propose in the study while comparing the proposed sampling method, oversampling with random location, and sampling with Euclidean distance, in two classification models, LSTM and SVC. Finally, it can be observed from our experimental results that our proposed sampling method performs better than the other two methods in the model. |
