應用均勻實驗設計法優化倒傳遞網路之超參數以預測射出成型成品之比容;Optimizing BPNN hyperparameters via uniform design to predict specific volume of injection molded part

NCUIR > Engineering College > Graduate Institute of Mechanical Engineering > Electronic Thesis & Dissertation > Item 987654321/90206

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/90206

Title:	應用均勻實驗設計法優化倒傳遞網路之超參數以預測射出成型成品之比容;Optimizing BPNN hyperparameters via uniform design to predict specific volume of injection molded part
Authors:	鄧修奇;Teng, Justin H.C.
Contributors:	機械工程學系
Keywords:	倒傳遞神經網路;實驗設計法;超參數;均勻實驗法;Back propagation neural network (BPNN);Design of experiment (DOE);Hyperparameters;Uniform design (UD)
Date:	2022-09-07
Issue Date:	2022-10-04 12:16:29 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究利用模內感測器收集並記錄射出成型之溫度、壓力值，並利用壓力、溫度與比容之關聯性以公式量化成品上三點位置比容的不均勻性。利用 python 程式語言撰寫 Tensorflow 框架之倒傳遞神經網路模型，將射出成型製程參數與相應比容數據分成訓練組與驗證組進行訓練，實際以驗證組預測後得到 2.8802%的誤差。為了要降低預測誤差提升準確度，利用實驗設計法探討超參數之學習率、迭代次數、批量與預測誤差之關係，分別使用田口方法規劃 L25(53)實驗表與均勻實驗法規劃 U20(203)實驗表，兩種方法經過迴歸望小分析後皆得到最佳參數組為學習率 0.01、迭代次數 4800、批量 29 之組合，以此組合實際訓練模型過後得到 1.1015%之預測誤差，相較初始設定之預測誤差 2.8802%優化了 61.76%，顯示實驗設計法於超參數優化之可行性。 ;In this research, in-mold sensors were used to record pressure and temperature during the injection molding process. A relationship between pressure, temperature and specific volume was established to quantify the nonuniformity of three spots on the molded part. The measured data was then divided into training set and validation set for a back propagation neural network model based on TensorFlow framework to make prediction on specific volume of the part, with a prediction error of 2.8802% on validation set. In order to improve prediction accuracy, applying design of experiment methods to make observation on the relationship among hyperparameters of learning rate, epoch, batch size and prediction error. The aforementioned design of experiment methods were a L25(53 ) table of Taguchi method and a U20(203 ) table of uniform design, respectively. The two methods obtained the same outcome of best parameters set with learning rate 0.01, epoch 4800, and batch size 29 from smaller-the-better regression analyzation. The best parameters set leads to a lower prediction error of 1.1015%. The improvement of 61.76% compared to initial hyperparameters setting shows that tuning hyperparameters via design of experiment methods is feasible.
Appears in Collections:	[Graduate Institute of Mechanical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	51	View/Open

社群 sharing

Loading...