微射出壓縮成型模內訊號感測與特徵提取於微透鏡陣列光學品質預測之研究

Micro-Injection Compression Molding in Mold Sensing and Feature Extraction for Optical Quality Prediction of Microlens Array

周海棠(Hai-Tang Chou)

指導教授 ： 楊申語

國立臺灣大學/工學院/機械工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202203494

摘要

微透鏡陣列廣泛應用於光場相機、頭戴式顯示器等3D成像設備，隨著虛擬實境市場備受關注微透鏡陣列需求也隨之提升，射出壓縮成型具有成品精度高且殘留應力低等優勢，因此透過射出壓縮成型可達到大量生產高品質微透鏡陣列目的，然而射出成型製程在生產過程中由於無法得知熔膠在模穴內的行為，導致當產品品質發生變異時無法即時反應，需透過線下儀器量測才可檢測產品品質，如此需花費大量檢測時間及人力成本。因此本研究利用射出壓縮成型製做微透鏡陣列並於模內設置壓力及溫度感測器來獲取模內成型資訊，透過提取模內訊號與成品光學品質有關的特徵作為品質指標，並將品質指標作為訓練資料建立神經網路模型，透過神經網路模型來達到微透鏡陣列光學品質預測目的。
本研究先以Moldex3D模擬軟體進行射出壓縮成型參數分析，藉由敏感因子找出對成品光學品質敏感之參數，並透過田口法及變異數分析找出最具影響力之參數，模擬結果顯示熔膠溫度及壓縮間距為影響成品光學品質之重要參數。
實際成型根據模擬分析得出的重要參數進行全因子實驗，透過全因子實驗分析品質指標與成品光學品質相關性，結果顯示本研究所設計的五個品質指標皆與成品光學品質呈現0.8以上的強相關性。接著透過全因子實驗所獲得的250組資料建立神經網路模型，首先針對離群值移除範圍進行分析，結果表明在z值設置為2時模型有較佳的驗證及測試準確度，分別為88.2 %及86.4 %，接著進行超參數優化，優化結果顯示在初始學習率為0.4、學習率衰減因子0.8、隱藏層節點數為50、批次大小為10、訓練週期數量為700及激活函數為Sigmoid時有最佳的模型效能，其模型驗證準確度達97.2%，最後根據測試資料進行模型測試，結果顯示超參數優化後模型測試準確度達到95.5 %，成功透過神經網路模型達到光學品質預測目標。

關鍵字

微透鏡陣列 ； 射出壓縮成型 ； 模內訊號感測 ； 品質指標 ； 神經網路模型

並列摘要

Micro-lens arrays are widely used in 3D imaging equipment such as light field cameras and head-mounted displays. Injection compression molding has the advantages of superior part accuracy and low residual stress. Therefore, injection compression molding can mass produce high-quality micro-lens arrays. However, it is hard to determine the melt quality in the cavity during the injection molding process. Thus unable to respond immediately when the product quality changes. Part quality can only be measured by offline instruments, which takes much testing time and labor costs. Therefore, this study used injection compression molding to fabricate a micro-lens array, and in-mold pressure and temperature sensors were set up to obtain molding information. Extract pressure and temperature curve features strongly correlate with optical quality and define them as quality indices, which were input into the neural network model for learning and prediction—using the neural network model to predict the micro-lens array optical quality.
In this study, Moldex3D simulation software was used to analyze the parameters of injection compression molding. The parameters sensitive to the optical quality of the product were impressed by the sensitivity factor, and the most influential parameters were found by the Taguchi method and analysis of variance. The simulation results show that the melt temperature and the compression gap are important parameters affecting the optical quality of the part.
For the molding experiment, the parameters setting of the full-factor experiment were according to the critical parameters obtained by the simulation analysis. The full-factor experiment analyzed the correlation between the quality indices and the optical quality of the part. The results show that the five quality indices designed in this study have a strong correlation of more than 0.8 with the optical quality of the part. Thus, a neural network model is established through the 250 sets of data obtained from the full-factor experiment. The outlier removal range is analyzed, and the results show that the model has better validation and test accuracy when the z value is set to 2, which are 88.2% and 86.4%, respectively. The hyperparameter optimization results show that the optimal hyperparameter combination is an initial learning rate of 0.4, a learning rate decay factor of 0.8, the number of hidden layer nodes is 50, the batch size is 10, the number of training epochs is 700, and the activation function is sigmoid. The accuracy of model validation reaches 97.2% after hyperparameter optimization, and finally, the model test is carried out according to the test data. The results show that the model test accuracy reaches 95.5% after hyperparameter optimization, and the optical quality prediction target is successfully achieved through the neural network model.

並列關鍵字

Micro-lens array ； Injection compression molding ； In-mold sensing ； Quality indices ； Neural network model

參考文獻

[1] J. Duparré and F. Wippermann, "Micro-optical artificial compound eyes," Bioinspiration biomimetics, vol. 1, no. 1, p. R1, 2006.

[2] J. Yu, "A light-field journey to virtual reality," IEEE MultiMedia, vol. 24, no. 2, pp. 104-112, 2017.

[3] W. Yuan, L.-H. Li, W.-B. Lee, and C.-Y. Chan, "Fabrication of microlens array and its application: a review," Chinese Journal of Mechanical Engineering, vol. 31, no. 1, pp. 1-9, 2018.

[4] T. E. Bishop and P. Favaro, "The light field camera: Extended depth of field, aliasing, and superresolution," IEEE transactions on pattern analysis and machine intelligence, vol. 34, no. 5, pp. 972-986, 2011.

[5] R. Ng, M. Levoy, M. Brédif, G. Duval, M. Horowitz, and P. Hanrahan, "Light field photography with a hand-held plenoptic camera," Stanford University, 2005.