以類神經網路、田口方法結合反應曲面法建立製程參數優化之預測模型
—	以射出成型機為例

塑膠射出成型，已經是個歷史悠久的製程方式，其製程優勢在於能夠製造出較為複雜的形體、結構且精密度高、生產快速，搭配一模多穴的設計產能更能獲得提升，但隨著行業發展得越久，機器與技術的水平不斷上升，製造成本不斷下降，但利潤也是為了競爭越來越低，因此如何以最快的速度最低的成本，來找出最佳參數，製造出品質最好的產品，是此行業提升競爭力的一大重點。目前塑膠射出成形參數設定，往往需要依賴工程師之經驗與實驗設計，但影響產品的參數眾多，各參數之間具有高度非線性關係，要找出適合之製程參數組合，須花費大量時間、人力及成本，因此本研究提出塑膠射出成形製程參數最佳化系統，利用式田口方法進行實驗與資料分析，找出引響產品品質特性的重要參數，在使用反映曲面法來找出此參數的最佳值。最後使用類神經網路，來建立品質特性的預測模型，減少公司進行不必要的實驗。具結果顯示，產品的品質特性，由原本的翹曲0.26mm 降低到0.1133mm，改善幅度達到56%，製程時間降低1.5秒。而建置的預測模型精準度約為16%，未來隨著訓練資料增加，精準度會再持續提升。

關鍵字

塑膠射出成型；田口方法；反應曲面法；類神經網路

並列摘要

Plastic injection molding is a long-established manufacturing method. Its manufacturing advantages are that it can produce more complex shapes, structures, high precision, and rapid production. At present, the plastic injection molding parameter setting often depends on the experience and experimental design of the engineer, but there are many parameters that affect the product, and each parameter has a highly nonlinear relationship. Finding a suitable process parameters combination requires a lot of time, labor, and Cost, so this study proposes a plastic injection molding process parameter optimization system, using the Taguchi method for experiment and data analysis, to find out the important parameters that induce product quality characteristics, and use the reflection surface method to find the best parameter value. Finally, a neural-like network is used to establish a predictive model of quality characteristics, reducing unnecessary experiments by companies. The results show that the quality characteristics of the product are reduced from the original warpage of 0.26mm to 0.1133mm, the improvement is up to 56%, and the process time is reduced by 1.5 seconds. The accuracy of the built prediction model is about 16%. In the future, as the training data increases, the accuracy will continue to increase.