請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90624
標題: | 射出壓縮成型微透鏡陣列品質提升及即時監測 Study on Quality Improvement and Real-time Monitoring for Injection Compression Molding of Microlens Array |
作者: | 張鈞程 Jun-Cheng Zhang |
指導教授: | 楊申語 Sen-Yeu Yang |
關鍵字: | 微透鏡陣列,射出壓縮成型,頂針式澆口關閉機構,模內訊號感測,品質指標,比容數值,神經網路模型, Micro-lens array,Injection compression molding,Shut-off nozzle,In-mold sensors,Quality indices,PVT,Neural network model, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 高分子材料光學性質優異,然而一般射出成型因模穴冷卻熔膠造成流動阻力隨流長比上升而提高,使模內壓力分佈不均並留下流動殘留應力,導致模穴末端與澆口處尺寸差異性過大、光彈條紋產生,因此本實驗採用射出壓縮成型,並設計一頂針式澆口封閉機構改善壓縮回流所造成近澆口區域的品質差異性,接續於模內加裝感測器,將模內數據建立神經網路模型監測技術,穩定成品的尺寸均勻性。
本實驗使用 Moldex3D 模流分析軟體,確認射出壓縮成型壓縮回流成因與文獻一致,由單因子實驗得出熔膠溫度、模具溫度及壓縮間距為成品尺寸均勻性之敏感因子。實際成型實驗驗證頂針式澆口關閉機構的可行性,成品重量穩定性提升 2.5 倍,光彈條紋不影響微透鏡陣列範圍,並使用成品尺寸均勻性之敏感因子進行三因子三水準之全因子實驗,實驗結果趨勢與 Moldex3D 模流分析結果一致。 最後將全因子實驗模內感測器特徵作為品質指標,並透過 Tait 方程式計算比容數值,經由相關性分析,品質指標與比容差皆為0.75 以上之強相關性。接續通過全因子實驗 216 組資料建立神經網路模型,首先通過 Z-score 離群值分析移除離群值,當 z = 1.5 時模型有較佳的驗證及測試準確度,接著對超參數進行優化,結果顯示優化器為 Adam、神經元節點數量為 128、激活函數為 sigmoid、批次大小為32、訓練週期數量為 500、學習率為 0.01 即時有最佳的模型效能,其模型驗證準確度達 0.963、測試準確度達到 0.939,成功透過神經網路模型達到預測微透鏡尺寸均勻性之目標。 Polymer materials possess excellent optical properties. However, in conventional injection molding, flow resistance increases as the flow length ratio increases due to cooling of the mold cavity. This unevenly distributes pressure within the mold, leading to residual flow stress, which causes significant differences in size between the mold cavity and gate, resulting in the production of photoelastic stripes. To address this issue, this experiment utilizes injection compression molding and designs a gate sealing mechanism to improve the quality disparity near the gate caused by compression and reflow. Additionally, sensors are installed in the mold, and in-mold data is used to establish a neural network model monitoring technology to ensure dimensional uniformity of the finished products. Using Moldex3D mold flow analysis software, this experiment confirms that the cause of compression reflow in injection compression molding aligns with existing literature. Through single-factor experiments, it is determined that melt temperature, mold temperature, and compression distance significantly affect the uniformity of the finished product size. The actual molding experiment verifies the feasibility of the gate closing mechanism, resulting in a 2.5 times increase in the weight stability of the finished product. The presence of photoelastic stripes does not impact the range of the microlens array. Furthermore, a three-factor, three-level full-factor experiment is conducted, focusing on the sensitive factors affecting the size uniformity of the finished product. The results obtained are consistent with the Moldex3D mold flow analysis findings. Subsequently, the in-mold sensor's characteristics from the full-factorial experiment are utilized as the quality index. The specific volume value is calculated using the Tait equation. Through correlation analysis, it is observed that the quality index and specific tolerance are strongly correlated above 0.75. A neural network model is then established using 216 sets of data from the full-factorial experiment. Initially, outliers are removed using Z-score outlier analysis, with a threshold of z = 1.5, resulting in improved verification and test accuracy of the model. Hyperparameter optimization is performed, revealing that the Adam optimizer, 128 neuron nodes, sigmoid activation function, a batch size of 32, 500 training cycles, and a learning rate of 0.01 yield the best model performance. The accuracy of model verification reaches 0.963, and the test accuracy reaches 0.939. As a result, the neural network model successfully achieves the goal of predicting the size uniformity of the microlens. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/90624 |
DOI: | 10.6342/NTU202302616 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 機械工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 8.96 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。