- 學位論文
矽氣凝膠產品製程參數最佳化
Process Parameter Optimization of Silica Aerogel Products
摘要
本研究使用田口實驗結合灰關聯分析多重品質特性，來處理二氧化矽氣凝膠之製程參數最佳化問題，經由實際實驗來了解影響氣凝膠的製程關鍵因素，進而找出影響製程的因子並進行實驗的探討。經專家建議，選擇密度、比表面積、孔體積為品質特性，並考慮特性間的權重問題，以田口實驗結合灰關聯進行多重品質特性之參數最佳化組合。 本研究採用直交表以三水準為主要的實驗配置架構，經由文獻探討所得三個實驗因子，為水矽比、PH值、改性劑比例，以直交表 L9(34)進行實驗，結合灰關聯分析法，並採用客觀權重-熵權重來進行權重配比，以灰關聯度進行資料分析，最後再進行確認實驗，以驗證本方法之有效性。為發掘最佳的參數水準，本研究利用田口實驗結果，嘗試以倒傳遞類神經網路建立製程參數(網路輸入)與品質特性(網路輸出)之關係，試圖找出田口實驗參數外的更加參數組合。 研究結果顯示，採用田口方法結合灰關聯分析法，所得到的最佳灰關聯度之實驗條件，可同時滿足多重品質特性。而進一步以田口回應圖分析所求得之製程參數，經確認實驗，無論是單一品質特性之SN比或是總和之灰關聯度，其結果均落在95%信賴區間，亦可同時滿足單一品質特性及多重品質特性。上述兩種方法所得參數皆可應用於實際氣凝膠之生產。最後採用倒傳遞類神經網路來推論品質是否有落於其他區間，但由於訓練樣本不足，推論結果不佳，若後續有相關研究有要採用倒傳遞類神經網路應加大訓練樣本，以利樣本訓練。
並列摘要
The research combines the Taguchi experiment and grey relational analysis (GRA) to deal with the problem of optimizing the parameters of the SiO2 aerogel. In addition, this paper tries to understand the main process factors influencing the quality characteristics of the aerogel through the experiments. After discussing with the domain experts, we chose three quality characteristics and three experimental factors with three levels and adopted L9 (34) orthogonal array to proceed with the experiment. The three quality characteristics are density, surface area, and pore size. To obtain the weight of the quality characteristics, the entropy measurement was applied. Based on the results of GRA with entropy weight and confirmation run, the optimal process parameters of silicon aerogel were obtained. To find the optimum level of process parameters, the research applies back-propagation neural network to build the relationship between the process parameters and the quality characteristics. The experiment results show that the experiment conditions with the best grey relation grade together with the parameters obtained by the method combining the Taguchi experiment and GRA can simultaneously improve the multiple quality characteristics. Both of the parameters can be used in the production of the aerogel in practice. Because of the insufficient of the training samples, the deduction of the back-propagation neural network doesn’t work well. To get the better results of the neural network, more samples should be required in future research.