類神經網路在品質管制上之應用：以倒傳遞網路偵測製程個別數據之平均 值及變異數的變化

The Application of Artificial Neural Networks for Quality Control : Detecting Changes in Mean and Variability of the Individual Measurements by Back-Propagation Neural Network

曾慶安(Ching-An Tzeng)

指導教授 ： 鄭春生

元智大學/工程學院/工業工程與管理學系/碩士(1993年)

摘要

在生產過程中由於受到可歸屬原因之影響，產品品質特性數據可能產生平
均值及/或變異數之改變。這些變化都代表著製程是處於管制外(異常)之
狀態。在統計製程管制中，當管制對象為個別值時，通常是以個別值及移
動全距管制圖來監視製程之變化。過去之研究結果說明此兩種管制程序並
無法區分製程之異常是來自於平均值改變、變異數改變或兩者同時產生變
化。本研究是以類神經網路(artificial neural networks)為基礎發展一
辨認製程異常之程序。我們所考慮之製程異常類型有(1) 製程平均值變
化 (向上或向下移動)；(2)變異數變化；(3)製程平均值和變異數同時變
化之情形。類神經網路之成效是以蒙地卡羅模擬法產生數據來評估，並與
其它管制法比較。本研究所採用之評估比較指標為平均連串長度和正確辨
認率。本論文所提出之異常類型辨認系統較傳統之管制圖及已發表過之類
神經網路辨認系統，除了有更快的偵測速度及較高之正確辨認率外，更提
昇了系統整體的穩定性及即時、連線的可行性。

關鍵字

統計製程管制 ； 類神經網路

並列摘要

The combined individual measurements and moving range charts
are often used in statistical process control to monitor a
series of individual measurements coming from a manufacturing
process. Previous research has shown that these charts are not
effective in identifying the process change caused by a shift
in mean or in variability or both. In this research, a set of
pattern recognizeres based on artificial neural network
techniques have been developed as an alternative means to
traditional methods. The out-of-control conditions considered
in this research includes the following: (1) a shift in the
process mean; (2) a change in the process variability; (3) a
shift in the mean as well as the variability. The performance
of the proposed neural networks has been evaluated based on the
average run length (ARL) and the recognition accuracy. A Monte
Carlo simulation has shown that the proposed neural network
approach out performs other traditional methods in terms of
recognition accuracy.

並列關鍵字

Statistical Process Control ； Artificial Neural Networks