利用機器學習結合蟻群演算法最佳化聚醯胺六之碳纖維複合材料射出成型縫合線強度;Optimization of weld line strength for injection molding process parameters of polyamide 6 and carbon fiber composites using machine learning and ant colony algorithm

NCU Institutional Repository > 工學院 > 機械工程研究所 > 博碩士論文 > Item 987654321/98734

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98734

题名:	利用機器學習結合蟻群演算法最佳化聚醯胺六之碳纖維複合材料射出成型縫合線強度;Optimization of weld line strength for injection molding process parameters of polyamide 6 and carbon fiber composites using machine learning and ant colony algorithm
作者:	傅家芸;Fu, Chia-Yun
贡献者:	機械工程學系
关键词:	特徵工程;機器學習;蟻群演算法;feature engineering;machine learning;ant colony algorithm
日期:	2025-08-05
上传时间:	2025-10-17 13:11:43 (UTC+8)
出版者:	國立中央大學
摘要:	本研究旨在最佳化含短碳纖維含量之聚醯胺6射出成型製程參數，以提升縫合線強度，並將機器學習技術與蟻群演算法相結合，探索製程參數最佳化的能力。研究過程首先基於Box-Behnken設計實驗規劃表與實驗數據建立數據資料，隨後採用數據前處理和特徵工程方法，提高模型輸入數據品質。接著，分別建立支持向量迴歸(SVR)模型和類神經網路模型，並利用網格搜索交叉驗證、Keras Tuner和田口法，對模型超參數進行最佳化。實驗結果表明，經過超參數最佳化，SVR模型在訓練集上的R2為0.83、RMSE為2.954的，而在測試集上的R2為0.789、RMSE為2.735；相較之下，Keras Tuner最佳化的類神經網路模型性能更為出色，訓練集的R2為0.914、RMSE為2.306，測試集的R2為0.864、RMSE為2.890，展現了更佳的擬合能力和強健性。隨後，本研究將包括Box-Behnken模型在內的四種機器學習模型納入蟻群演算法框架，以尋求製程參數的最佳組合。研究表明，不同模型對應的最佳參數存在一定差異。其中，Keras Tuner最佳化模型的預測誤差最小，僅為0.49%相對於BBD的改善率達到4.65%；而SVR模型和Box-Behnken模型的改善率則相對較低，源於其最佳參數組合導致產品收縮率增加或殘留應力提高所致。本研究成功利用有限的實驗數據，建立了描述製程參數與縫合線強度的關係之優秀機器學習模型，並將其與蟻群演算法相結合，獲得了最佳化的製程參數組合，從而提升產品性能。這為技術人員提供了全面的製程最佳化方法，展現了機器學習與啟發式演算法相互融合在智慧製造領域的廣闊應用前景。 ;This study aimed to optimize the injection molding process parameters of polyamide 6 reinforced with short carbon fiber to enhance the tensile strength of weld line, and explore the capability of integrating machine learning with ant colony optimization (ACO) for parameter optimization. The dataset was established based on the experimental plan and results of Box-Behnken design (BBD), followed by data preprocessing and feature engineering to improve the model’s input quality. Support vector regression (SVR) and neural network models were constructed, with optimized hyperparameters using the grid search cross-validation, Keras Tuner, and the Taguchi method. After hyperparameter optimization, the Keras Tuner optimized ANN model achieved R2 = 0.914 and RMSE = 2.306 on the training dataset, as well as R2 = 0.864 and RMSE = 2.890 on the testing dataset, demonstrating superior fitting and robustness. Four machine learning models, including BBD, were incorporated into the ACO framework for optimal parameter combinations. The Keras Tuner optimized model exhibited the smallest prediction error (0.49%) and BBD highest improvement rate (4.65%). In contrast, the SVR and BBD models had lower improvement rates due to optimal parameters causing increased product shrinkage or residual stress. This study successfully established good machine learning models from limited data to describe the relationship between process parameters and the weld-line strength, and then integrated them with ACO to obtain optimized parameter combinations, enhancing product performance. It provides a comprehensive optimization approach and showcases the broad application prospects of machine learning with heuristic algorithms in smart manufacturing.
显示于类别:	[Graduate Institute of Mechanical Engineering] Electronic Thesis & Dissertation

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	23	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....