基於人工智慧之鋁合金板材殘留應力消除預測研究;Artificial Intelligence Based Prediction of Residual Stress Relief in Aluminum Alloy Plates

NCU Institutional Repository > 工學院 > 機械工程研究所 > 博碩士論文 > Item 987654321/98758

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98758

題名:	基於人工智慧之鋁合金板材殘留應力消除預測研究;Artificial Intelligence Based Prediction of Residual Stress Relief in Aluminum Alloy Plates
作者:	高靖桐;Kao, Ching-Tung
貢獻者:	機械工程學系
關鍵詞:	7055鋁合金板材;殘留應力;有限元素法;機器學習;深度學習;粒子群演算法;7055 aluminum alloy plates;Residual stress;Finite element method;Machine learning;Deep neural network;Particle swarm optimization
日期:	2025-08-25
上傳時間:	2025-10-17 13:16:16 (UTC+8)
出版者:	國立中央大學
摘要:	鋁合金板因其高強度重量比與良好加工性，廣泛應用於航太與結構工程領域。然而，為提升其機械性質所採用的淬火製程，常導致顯著的殘留應力，特別是在厚板區域，進而引發變形、裂紋與尺寸不穩定等問題。為解決此類問題，已有多種應力消除技術被提出，其中以機械拉伸法因其操作簡便且具產業實用性，成為最有效的方法之一。儘管有限元素法可用於模擬預測應力變化，但其高計算成本與耗時限制了於製造現場之即時應用。本研究結合有限元素模擬與機器學習模型，探討7055鋁合金厚板之殘留應力消除行為。有限元素模擬架構係基於並延伸自本研究室先前之研究，模擬水淬與拉伸過程，板厚範圍涵蓋10 mm至250 mm，拉伸應變比則介於1.0%至4.5%。為提升預測效率，本研究使用逾九百萬筆有限元素模擬資料，訓練與測試兩種機器學習方法，包括以隨機森林演算法（Random Forest）與極限梯度提升演算法（XGBoost）進行殘留應力迴歸預測，並以結合粒子群最佳化演算法（PSO）之多層感知器（MLP）分類模型，判別適當之拉伸應變比。研究結果顯示，板材厚度對殘留應力分布具有關鍵影響，其中中等厚度板件產生最高應力值。拉伸處理選擇應變比在2.0%至2.5%之範圍內對應力消除效果尤為顯著。在機器學習模型的殘留應力預測方面，隨機森林模型於殘留應力預測中表現最佳，其決定係數（R²）達0.9995，均方根誤差（RMSE）僅1.3252 MPa，而XGBoost模型則展現較佳之泛化能力，其R²為0.9783、RMSE為4.6333 MPa。分類模型以PSO優化後之MLP分類器在拉伸應變比分類中達到83.06%(F1-score)之準確率，優於基準模型。綜上所述，有限元素模擬與機器學習之整合，為鋁合金板材製程中之殘留應力分析與應力消除製程優化提供一套具體且高效之解決方案。 ;Thick aluminum alloy plates are widely used in aerospace and structural applications due to their high strength-to-weight ratio and machinability. However, the quenching process used to enhance their mechanical properties introduces significant residual stresses, particularly in thick sections. It can lead to distortion, cracking, and dimensional instability. To address these issues, various stress relief methods have been investigated, with tensile stretching emerging as one of the most effective techniques due to its simplicity and industrial applicability. Despite the availability of finite element method (FEM) simulations for predicting stress evolution, their high computational cost limits real-time application in manufacturing environments. This study investigates the residual stress relief behavior of thick 7055 aluminum alloy plates through FEM simulations and machine learning (ML) models. The FEM framework, based on and extended from a previous study, simulated water quenching and tensile stretching processes with varying plate thicknesses ranging from 10 mm to 250 mm and stretching ratios from 1.0% to 4.5%. To enhance prediction efficiency, over 9 million FEM-derived data points were used to train and validate two ML approaches, including two regression models (Random Forest and XGBoost) to predict remaining residual stresses, and a classification model (MLP with PSO optimization) to identify appropriate stretching ratios. The results confirmed that thickness plays a dominant role in the residual stress distribution, with intermediate-thickness plates showing the highest stress magnitude. Tensile stretching proved highly effective in reducing residual stresses, particularly with a stretching deformation between 2.0% and 2.5% strain. The Random Forest model achieved the highest predictive accuracy with an R² of 0.9995 and RMSE of 1.3252 MPa, while the XGBoost model showed stronger generalization with an R² of 0.9783 and RMSE of 8.7259 MPa. The optimized MLP classifier reached an F1-score of 83.06%, outperforming baseline models in classifying applied stretching ratios. Overall, the integration of FEM and ML offers a practical and efficient approach to residual stress analysis and stress relieving process optimization in aluminum plate manufacturing.
顯示於類別:	[機械工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	6	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....