圖形化高斯模型應用於自動化生產資料之關聯性分析

Abstract

自動化生產過程有上百個步驟，每個步驟都包含相當多的測量項目，所以會得到相當龐大的原始數據。這些數據不但有極大量的變數，同時變數之間也存在高度的關聯性，因此會有多餘資訊的產生。依據這些特性我們選擇使用圖形化高斯模型的方法建立模型，提供研究者分析資料中解釋變數對於被解釋變數的影響以及解釋變數之間互相的關聯性。 以晶圓製造為例，論文中將資料作前置處理後，先就一般的方法作討論，並測試其計算的極限和一些影響的因素。接著結合因素分析和多維縮放比例的方法進行變數的分群，藉由分群建立模型的方式來簡化一般方法，並且針對偏差值提出改進。利用簡化方法所建立的模型在分析更多變數量的同時，更能使模型偏差值保持於使用者要求的範圍之內。 同樣的方法亦可應用於與製程類似自動化過程所測量的大量數值資料中，模型建立之後再結合專家系統或貝氏網路，將可以對結果進行預測與診斷之工作。

There are hundreds of steps in the process of automated manufacture operation. Every step contains lots of measurements. As a result a tremendous amount of data is available. These data have a great deal of variables, which are highly correlated. According redundancy exits. In order to provide analysts the influence of predictors upon dependents, and to explain the correlations of variables, we use Graphical Gaussian Models(GGMs) to establish models based on the characteristic of the gathered data. Take manufacture of silicon wafers for example, data will be preprocessed first. Then we will discuss the measured limits and the factors of the general GGMs. Through the combination between factor analysis (FA)and multidimensional scaling (MDS), clusters of variables will be proceeded. According to the clusters, he procedure of modeling will be simplified and an improved method will be introduced to analyze more variables while maintaining the requested deviance. This method also can be applied to the massive data gathered by the similar procedure like automated manufacturing operation. Combining Expert system or Bayesian Network, we can prognosis and diagnosis results after a model is built.