管制圖非隨機性樣式之辨認及參數之估計

Control Chart Patterns Recognition and Parameters Characterization

陳信嘉(Shin-Jia Chen)

指導教授 ： 鄭春生

元智大學/工程學院/工業工程與管理學系/博士(2001年)

摘要

在統計製程管制 (Statistical Process Control, SPC) 中，當製程存在可歸屬原因時，其製程可能會表現出特定的非隨機性樣式的變化。正確辨認率非隨機性樣式，可以縮小診斷製程異常原因之範圍。若能估計參數之變化量，更有助於規劃改善方案。本研究之目的是發展一個辨認系統，可以有效地辨認非隨機性樣式之種類，並估計其參數。本研究將考慮單一樣式出現及多種樣式併發之情形。
在本研究中，我們提出兩種方法來做管制圖非隨機性樣式之辨認及重要參數之估計。第一種方法是以計算累積和平均為基礎之原創方法，若將累積和平均以圖形方式表示，不僅可協助使用者辨認非隨機性樣式，同時也可以由圖形估計樣式之重要參數。
第二種方法是以類神經網路為基之辨認程序。我們的主要貢獻有下列幾點。第一，我們發展啟發式特徵擷取方法。這些特徵不僅可提高應用時之辨認能力，同時也可降低訓練時間及應用時之計算時間。第二，我們以類神經網路發展一個篩選訓練樣本之程序，將不良之樣本篩選後，有助類神經網路之學習並提高類神經網路之一般化能力。
本研究所建立之辨認系統是以辨認率、平均連串長度及平均百分比誤差來評估其效益。經由模擬方法分析，證明本研究所提出之方法可以有效地辨認非隨機性樣式並估計其參數。

關鍵字

非隨機性樣式 ； 圖形法 ； 類神經網路

並列摘要

Control chart patterns recognition is an important aspect of statistical process control (SPC). A control chart may indicate an out-of-control condition when some nonrandom patterns occur. Different nonrandom patterns can be associated with a specific set of assignable causes. Hence, identification of nonrandom patterns can greatly narrow the set of possible causes that must be investigated, and thus the diagnostic search could be reduced in length. The purpose of this research is to develop a pattern recognition system to recognize nonrandom patterns and identify the key parameters.
In this research we propose two pattern recognizers for the analysis of control chart patterns: a graphical approach and a neural network-based pattern recognizer. The proposed graphical approach is based on the concepts of data smoothing and is capable of detecting unnatural patterns as well as describing the key parameters of the specific pattern detected.
The neural network-based pattern recognizer looks for the following nonrandom patterns: trend, cycle, shift and the multiple combinations of these patterns. In addition to the raw data, some important features extracted from process data were used as the inputs of neural network.
The pattern recognizers have been evaluated by estimating the average run length, the rate of correct classification and mean absolute percent errors. The simulation results show that the pattern recognizer can recognize the control chart patterns with correct classification rate of about 90%. The results also show that the pattern recognizers developed in this research can accurately identify the key parameters of the patterns detected.

並列關鍵字

nonrandom patterns ； graphical method ； neural networks

參考文獻

58.謝昆霖，類神經網路在品質管制上之應用：非隨機性變化之偵測，中

Foote, "Spectral analysis in quality control: a control chart

control charts with supplementary run rules," Technomtrics,

detecting and classifying process mean shifts," International

5.Cheng, C. S., "A multi-layer neural network for detecting