整合田口方法與類神經網路探討CAE的深度學習

Deep Learning Study of CAE with the Integration of Taguchi Method and Neural Network

林韻孜(Yun-Zi Lin)

指導教授 ： 鍾文仁

中原大學/工學院/機械工程研究所/碩士(2017年)

https://doi.org/10.6840/cycu201700874

摘要

近年來，塑膠射出成型是傳統工業中一項非常重要的技術，但很多工廠在模具射出成型時，通常靠老師傅的經驗來解決模具與成型上的問題，對於製程的缺陷安排卻無一套有效的管理方法，造成品質控管困難且整體成本往上提升。針對此問題一般會使用電腦輔助工程(Computer-aided Engineering, CAE)之技術來解決缺陷問題，此方法可以在模具製造前預先排除成型缺陷，但是在現場人員實際成型時，依然無法百分之百預測成型結果，需要靠現場人員排除缺陷。
在射出成型結果之中，不論是翹曲還是收縮都是與控制因子交錯複雜的關係，無法以一般線性規則去進行分類，然而倒傳遞類神經網路(Back Propagation Neural Network, BPNN)在解決非線性問題時有著非常出色的辨識能力，經由多組訓練資料訓練過後，能夠精準的預測未知結果。本研究以田口直交法(Taguchi method)結合倒傳遞類神經網路，建構一個可預測CAE分析結果的計算系統，並以深度學習的角度探討多層式架構對於預測精準度的影響，研究結果顯示，由田口直交表選出的訓練資料用於單一輸出的網路訓練時，可以得到高準確度結果，但是用於多輸出訓練時無法收斂，在訓練設定部分，最佳類神經設定參數為多層隱藏層，在實驗過程中會因為增加隱藏層數目，而訓練計算時間加快且準確率提高。

關鍵字

電腦輔助工程 ； 倒傳遞類神經網路 ； 田口直交法 ； 深度學習

並列摘要

Plastic injection molding is a very important technology in the traditional industry. Usually solving the mold and molding problems depends on the experienced employee and there is not any effective management for the defects of scheduling, resulting in quality control difficulties and the high cost. For this issue, computer-aided engineering (CAE) technology is used to solve the defects. This method can eliminate the molding defects before mold manufacturing. However, it is still impossible to predict the actual molding process therefore experienced personnel are needed to solve the defects.
In the result of injection molding, both warpage and shrinking are complex with control factors, and cannot be classified by general linear rules. However, Back Propagation Neural Network (BPNN) has a very outstanding ability to identify non-linear problem. After multiple sets of training, it can accurately predict the unknown results. This study is based on the Taguchi Method and Back Propagation Neural Network to construct a computational system in order to predict the CAE analysis results. This paper discussed the effect of the multi-layer architecture on the prediction accuracy by Deep Learning. The results show that the training data selected by the Taguchi Orthogonal Array can get high accuracy results for single output network but cannot be used for multi-output training. Thus, the best parameters is multi-layer hidden layer. In the experiment because the hidden layers increases, training time can be improved and the accuracy rate is upgraded.

並列關鍵字

Computer-aided Engineering ； Back Propagation Neural Network ； Taguchi Orthogonal Method ； Deep Learning

參考文獻

[1] Marins, N. H., Mello, F. B. de, Silva, R. M. e, Ogliari, F. A., 2016, “Statistical Approach to Analyze the Warpage, Shrinkage and Mechanical Strength of Injection Molded Parts,” Polymer Processing, 3, pp.376-384.

[3] Castro, C. E., Rios, M. C., Castro, J. M., Lilly, B., 2007, “Multiple Criteria Optimization With Variability Considerations in Injection Molding,” Polymer Engineering and Science, 47(4), pp.400-409.

[5] Yarlagadda, P. K. D.V., and Khong, C. A. T., 2001, “Development of a Hybrid Neural Network System for Prediction of Process Parameters in Injection Molding,” Journal of Materials Processing Technology, 118(1-3), pp.109-115.

[6] Shen, C., Wang, L., and Li, Q., 2007, “Optimization of Injection Molding Process Parameters using Combination of Artificial Neural Network and Genetic Algorithm Method,” Journal of Materials Processing Technology, 183(2-3), pp.412-418.

[8] Rosaa, J. L., Robina, A., Silvab, M. B., Baldana, C. A., and Peresd, M. P., 2009, “Electrodeposition of Copper on Titanium Wires: Taguchi Experimental Design Approach” Journal of Materials Processing Technology, 209(3), pp.1181-1188.