Abstract
隨著電腦整合製造的發展，自動化檢驗已經成為一種趨勢。為了配合自動化檢驗需時甚少的特性，必須建立一套統計製程管制的專家系統，輔助操作員在製造的過程能迅速做決定。本文的目的是建構一套能應用於線上及時檢驗的統計製程管制的專家系統。此一系統的主要功能包括製程能力分析，選擇合適的管制圖，以及偵測管制圖上超出管制的狀況。為了要辨認超出管制的情況，本文引用 Western Electric提出的補充判斷規則。然而，這些規則有時候無法正確的區分出來是屬於哪一種的異常型態，因此考慮以類神經網路來取代。利用類神經網路彈性較大以及運算快速的特性，來辨認管制圖上的異常型態。本文利用逆向傳遞學習法則來訓練類神經網路。而實驗的結果顯示，類神經網路偵測管制圖上的異常型態之準確性高於 Western Electric的補充規則。因此意欲取代Western Electric提出的補充規則來判斷管制圖上異常型態，類神經網路是很好的選擇。此外，類神經網路對於變異不太敏感，使得它更適於實際應用在製程當中。Automatic inspection is a trend which leads short time ofinspection when the movement is toward computer-integratedmanufacturing. In order to match the need of automaticinspection, it is necessary to do the decision very fast. SPCexpert system is built to play the role of an intelligientcounselor to operator in a manufacturing process so the time ofdecision making will be decreased. The objective of thisresearch is to construct an expert system for SPC which can baapplied for on-line real-time inspection. Major tasks involvedin this system include the process capability analysis, theselection of an appropriate control chart, and the detection ofout-of-control situation in control charts. To detect the out-of-control patterns, the Western Electric supplementarry runsrules are adopted in this research. However, since these rulessometimes cannot correctly identify the type of unnaturalpattern, neural network is considered to replace it. Neuralnetwork approach is adopted here to utilized its advantages offlexibility and high-speed computation for identifyingunnatural pattern on control chart. The proposed neural networkis trained by using back-propagation learning rule. Theexperimental results show the neural network is a good tool toreplace Western Electric supplementary runs rules in detectingunnatural pattern of control chart. The accuracy of unnaturalnetwork detecting unnatural pattern is better than WesternElectric supplementary runs rules. The another advantage isthat the neural network is not sensitive to random noise, so itis more suitable to be applied in manufacturing process.