田口實驗法可以藉由部分因子實驗與因子回應分析推論出分析特性的較適參數因子組合,但是正交化的實驗因子規劃,所造成的分析誤差是無法避免的。為了改善此誤差,本文提出整合直交表實驗因子正交化特性與類神經網路學習能力的網路建構方法,藉由階段式的訓練程序建構訊噪比推論網路。文中以兩個IC導線架橋帶薄板剪切製程例,說明網路的建構流程。經過驗證實驗的證實,增加少許的追加實驗組後,可以得到較傳統 田口法更為客觀且正確的參數水準組合。
In Taguchi method, a preferred factor solution can be derived by factional factor experiments and factor level response analysis. However, the experimental error caused by factor levels orthogonality in orthogonal arrays is inevitable. In order to improve this, a network approach combines with Taguchi's method by progressive training is proposed, which integrates the experimental orthogonality and the learning ability of neural network to establish an inferring network model. Two cases of IC leadframe dam-bar shearing have carried out to demonstrate the modeling process. By increasing a few additional experiments, an optimal factor level combination can be inferred, which is more objective and accurate than the traditional Taguchi's method does.