本文爲了克服傳統的田口方法的缺點,提出基於類神經網路與交叉驗證法的田口方法,其要點有:(1)採用交叉驗證法以克服類神經網路必須區隔實驗數據集爲訓練集、測試集,導致訓練集數據不足的缺點。(2)提出敏感性分析與帶狀主效果圖以改善類神經網路黑箱模型的缺點。(3)採用非線性規劃以設計最佳品質因子的組合。爲了證明本法可行,本研究以二個已發表的田口方法的實例進行實證。研究結果顯示:(1)傳統的田口方法的未區隔資料集爲測試集、訓練集之作法,其誤差被嚴重低估。(2)在同樣使用可正確評估模型誤差的交叉驗證法下,本文提出的基於類神經網路的田口方法遠比傳統的田口方準確。(3)本文提出的「敏感性分析」與「帶狀主效果圖」確實可以表現因子與反應間的關係,改善類神經網路的解釋能力。(4)非線性規劃確實可以設計出最佳品質因子的組合。
To overcome the shortcomings of traditional Taguchi method, this study proposed a Taguchi method based on neural networks and cross validation methodology. Its main ideas included that (1) it used cross-validation methodology to overcome the shortcoming that neural networks need to separate the dataset into training set and testing set, which makes training set insufficient, (2) it used sensitive analysis and the main effect diagram to over come the shortcoming that neural networks are black box model, and (3) it used nonlinear programming to find the optimum combination of factor level. Two examples that had been solved by traditional Taguchi method were examined to verify the proposed method. It was demonstrated that (1) the prediction error is seriously underestimated in traditional Taguchi method, which does not separate the dataset into the testing set and training set, (2) under the same cross validation methodology, the proposed neural network-based Taguchi method is much more accurate than the traditional one, (3) the sensitive analysis and the main effect diagram truly can express the relationship between factors and responses, and improve the explanatory capacities of neural networks, (4) the nonlinear programming truly can find the optimum combination of factor level.