應用類神經網路建立偵測自我相關製程平均值偏移之管制法

A Neural Network-based Control Procedure for Monitoring the Mean Shifts of Autocorrelated Process

李秋蓁(Chiu-Chen Li)

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(2006年)

https://doi.org/10.6838/YZU.2006.00079

摘要

統計製程管制圖是在製程數據滿足常態分配且獨立性的假設下發展。然而，在實際的工業製程中已採用自動化之生產及檢驗方式，導致製程數據間具有高度之自我相關性，在這種情況下傳統之製程管制法並不適用，將造成錯誤警報增加。

在過去的研究中，大多數的學者所提出類神經網路之方法只能處理單一自我相關係數下之製程，然而，在實際製程中，自我相關係數將隨時間或批量之差異而改變。因此，就必須有多個類神經網路去監控不同的自我相關係數下之製程數據，才能達到全面性效益。

本研究以多個不同自我相關係數所獲得之數據，以類神經網路為基礎，建立一個管制程序去偵測平均值偏移。在製程偏移之偵測，以平均連串長度來評估類神經網路的偵測能力。

本研究的主要貢獻有兩點。第一、不論製程處於何種程度之自我相關係數下，都能具有良好的偵測能力。第二、本研究所提出之類神經網路管制法對於偵測平均值的偏移比EWMA管制法及X管制法好。

關鍵字

自我相關製程、統計製程管制、類神經網路、平均連串長度

並列摘要

In statistical process control (SPC), a state of statistical control is developed with a process which is generated independently and distributed normally. In fact, the manufacturing processes can not be satisfied with the assumption of uncorrelated or independent observations. This is due to automatic production and inspection procedures. When this assumption is violated, the control charts do not perform as well. Specifically, the control charts will give much higher false alarms when the data is correlated.

In the past researches, the neural network approach could only tackle single correlation coefficient processes. However, in the practical process the autocorrelation coefficient will be influenced by the difference of time and batch. Hence a design of neural network is based on a function of multiple correlation coefficients instead of single correlation coefficient on traditional control charts. They can detect the process mean shift quickly and correctly.

In this research, we develop a control procedure based on neural network to monitor the process mean shifts. We use the back-propagation neural network to solve these problems. The performance of neural networks are evaluated by the average run lengths (ARL) using simulation.

In this study, there are two contributions. Firstly, it can provide good detecting ability for processes under any level of autocorrelation coefficient. Second, neural network controls the method to detect process of mean shift which is superior to EWMA and X charts.

並列關鍵字

autocorrelation、statistical process control、neural networks、average run lengths

參考文獻

萬維君，「應用類神經網路於製程平均值變化之偵測及參數之估計」，元智大學工業工程研究所碩士論文，2001。

吳聰宏，「類神經網路應用在品質管制中相關性製程數據之管制」，元智大學工業工程研究所碩士論文，1994。

鄭春生、鄭靜彥，「以類神經網路辨識製程個別值數據之平均值、變異數及相關性的變化」，品質學報，6，29-43 (1999)。

Alwan, L. C., and Radson, D., “ Time-series investigation of subsample mean charts,” IIE Transactions, 24, 66-80 (1992).

Alwan, L. C. and Roberts, H. V. “ Time-series modeling for statistical process control,” Journal of Business and Economic Statistics, 6, 87-95 (1988).