應用類神經網路於IC封裝業資料探勘

Abstract

資料探勘運用演算法從資料庫中挖掘未知、有用的資訊，其主要的技術包含關聯法則、分類、預測、群集分析與探勘各種複雜的資料（如序列、空間、時間序列、文字資料探勘等）。類神經網路經常被運用於資料探勘之分類、預測與分群上，本研究發展一ART-Counterpropagation 類神經網路作為分類、預測的技術，其演算法基於ART原理與CPN的學習演算法則建構類神經網路，ART-CPN具有可塑性與穩定性類神經網路學習模式，此演算法有較佳的學習效率及良好的分類與預測的準確性。本研究分別運用ART-CPN 與SOM 類神經網路於IC封裝製造業資料探勘的品質預測與機台分群二種問題上，以ART-CPN 建構IC導線架蝕刻過程中定位孔尺寸的預測模式，此演算法有較佳的學習效率與預測準確性；適當機器分群可改善IC封裝瓶頸機台的效率， 整合SOM與K-mean 作為 Wire Bond 機台分群演算法，以機台記錄資料庫中各項產出記錄，將機器作適當的分群改善瓶頸機台生產之效率，此方法可使每一機群群內變異較少，得到一良好的分群結果。

Data mining involves the application of various algorithms for extracting implicit, previously unknown and potentially useful information from databases. Major techniques of data mining include: association rules, classification and prediction, clustering, and mining complex types of data (sequential, spatial, time-series, text, etc.). The neural networks are applied to solve the classification, forecasting and clustering problems in data mining. This study uses an ART-Counterpropagation neural network (ART-CPN) for solving classification and forecasting problems. The network is based on the ART concept and the CPN learning algorithm for constructing the neural network. ART-CPN involves real-time learning and is capable of developing a more stable and plastic classification and prediction models of input patterns by self-organization. The learning algorithm reveals better learning efficiency and good classification and prediction performance. ART-CPN and SOM neural networks are applied to solve the forecasting and clustering problems for dada mining in the IC packaging. ART-CPN is used to predict the dimension of the pilot hole of lead frame in the etching process. The algorithm reveals better learning efficiency and prediction performance. Integration of SOM network and K-means algorithm clusters the Wire Bond machines into groups given the machine performance record. The adaptive machine groups will improve the bottleneck efficiency. The result of clustering method is more homogeneous for each machine group.