應用不同監督式網路於多變量製程偵測與分析

Detection and Analysis of a Multivariate Process by Multiple Supervised Networks

林偉成(Wei-Cheng Lin)

指導教授 ： 陳雲岫

元智大學/工程學院/工業工程與管理學系/碩士(2008年)

https://doi.org/10.6838/YZU.2008.00022

摘要

隨著產品和製程的日漸精密複雜，只監控一個品質特性已不足以幫助工程師有效控制品質，加上品質變數之間存在著一定的相關性，若使用多個單變量管制圖管制，則容易造成誤判的情形，因此近年來許多學者提出了多變量管制圖，但是多變量管制圖還是有幾項缺點，第一為不同偏移量大小需要使用不同管制圖，第二為當點超出管制界線無法得知是由哪的變數造成的，第三為無法得知偏移變數的偏移量。
因此在90年代開始類神經網路成為使用的新方法，不同的網路有不同優缺點，因此本研究使用四種神經網路去作整理比較，分別是倒傳遞網路、機率神經網路、廣義迴歸神經網路和徑向基神經網路，採用兩階段方法，階段一偵測製程是否偏移且偏移變數為哪個變數，階段二量化偏移變數的偏移量。實驗結果顯示倒傳遞網路不論是分類問題還是量化問題，都為四種網路中最為準確和穩定。而訓練時間以機率神經網路和廣義迴歸網路最為快速，徑向基函數最為緩慢，最後並以案例驗證得知類神經比傳統多變量管制圖具有更佳的偵測能力。

關鍵字

多變量管制圖 ； 倒傳遞網路 ； 機率神經網路 ； 廣義迴歸神經網路,徑向基神經網路

並列摘要

With the complexicity of product process, we may not have enough quality control by supervising the single quality characteristic only. Due to the correlation among several characteristics, it may cause miscarriage of justice if we use several univariate control charts to inspect multivariate observations. As the reason, the researchers take multivariate control charts recently. However, there still exists some problems on using multivariate control chart. First of all, different magnitudes of shifts may cause us to use different control charts. Secondly, we cannot identify which variable or group of variables has caused the out-of-control signal. Finally, it’s hard to know the exactly amount of the shift magnitude when data is out of control.
Neural networks have been used to solve the problems since 1980’s. Many different networks are proposed to optimize the system or models. In our research, we consider four different networks, Back Propagation Network (BPN), General Regression Neural Network (GRNN), Radial Basis Function Neural Network (RRFN) and Probabilistic Neural Network (PNN) in detecting the shifts of multivariate inspection data. In the thesis, we use a two-stage model to solve the problem. Stage one is to detect the process whether there are in or out of control signal. Stage two is to quantify the magnitude of shift variables. Study shows that BPN performs the best either in classification and quantification among four networks. As to the training time of the networks, PNN and GRNN perform equally well among four while RBFNN needs the longest training time. We use cases to show networks have better detection ability than conventional multivariate control charts

並列關鍵字

multivariate control chart ； Back Propagation Network ； General Regression Neural Network ； Radial Basis Function Neural Network ； Probabilistic Neural Netwok

參考文獻

1. 阮冰如(2005)，“應用類神經網路與支援向量機於多變量製程

2. 吳重澤(2006)，“利用類神經網路與支援向量機技術偵測多變量

47, 195-205.

weighted moving average control chart,”Technometrics,

34, 46-53 .