DC 欄位 值 語言 DC.contributor 資訊工程學系 zh_TW DC.creator 蔡祐翔 zh_TW DC.creator Yu-Hsiang Tsai en_US dc.date.accessioned 2020-7-28T07:39:07Z dc.date.available 2020-7-28T07:39:07Z dc.date.issued 2020 dc.identifier.uri http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=107522096 dc.contributor.department 資訊工程學系 zh_TW DC.description 國立中央大學 zh_TW DC.description National Central University en_US dc.description.abstract 晶圓線切割機台用於矽晶圓製造的一環，機台的作用是將晶圓切片。然而，當機台無預警地停機或者切割的線鉅斷裂，整批的晶圓將會成為次等品甚至必須拋棄，這將導致成本增加，而這伴隨著一個大挑戰-資料集不平衡問題。正常與異常資料比例懸殊，為21:1。因此提出了一套異常偵測的策略，由三個部分組成：表徵學習、監督式分類器、警報機制來做出最後決策。k-平均演算法與自編碼器之表徵學習法，僅使用正常資料來學習正常的特徵，這不僅解決資料集不平衡的問題，也讓實驗採用的四個監督式學習分類器:隨機森林、單純貝氏分類器、支援向量機器、極限學習機表現得更好，並且設計一套發警報機制有效減少假警報數量。我們評定這套異常偵測策略的好壞是根據一家合作半導體矽晶圓材料製造公司的真實機台收集之資料，在測試資料集達到偵錯率0.57以及錯報率0.10。此外，這套預測系統已經在產線上實裝和測試，我們也提出了最影響模型好壞的原始工程資料。 zh_TW dc.description.abstract Wafer wire saw machines are used in a link of silicon wafer manufacturing, that saws wafers into individual die. However, while the machine shutdown or the sawing wire broken unexpectedly, that batch of wafers will be secondary products or wasted wafers leading to cost increase. Also, it comes up with a challenging issue - the imbalance dataset. The ratio of normal and abnormal data is 21:1. Therefore, an anomaly detection strategy is proposed, composed of three parts: representation learning methods, supervised classifiers and alarm rules. K-means clustering and autoencoders are the representation learning methods that learn normal features from normal data only, that not merely solves the imbalanced data challenge, but also helps the 4 experimental supervised classifiers: random forest, Naïve Bayes, support vector machine, extreme learning machine perform better, whereas the alarm rules help reduce false alarm. The anomaly detection strategy is evaluated on two machines from a real semiconductor silicon wafer material manufacturing company, where the catching rate is 0.57 and false alarm is 0.10. Moreover, this predictive system has been implemented and tested in production line, and we put forward the considerable engineering profiles that are highly related to the models. en_US DC.subject 異常偵測 zh_TW DC.subject 不平衡資料集 zh_TW DC.subject 機器學習 zh_TW DC.subject 表徵學習 zh_TW DC.subject anomaly detection en_US DC.subject imbalance dataset en_US DC.subject machine learning en_US DC.subject representation learning en_US DC.title 使用表徵學習和機器學習方法於晶圓線切割機台之異常偵測 zh_TW dc.language.iso zh-TW zh-TW DC.title Anomaly detection in wafer wire saw machines using representation learning and machine learning methods en_US DC.type 博碩士論文 zh_TW DC.type thesis en_US DC.publisher National Central University en_US