折床工具機壓力預測之機器學習應用

Abstract

傳統折床加工對操作人員經驗的依賴影響製程穩定性與加工精度，也限制我國折床工具機在國際市場的競爭力。針對國產折床工具機，現行壓力預測與補償模組大多基於固定公式，對貼膜材料、板厚變異與模具磨耗等非典型情境反應能力有限。雖有限元素分析(Finite Element Analysis, FEA)具物理解釋性，然其建模需要依據力學學理，且運算成本高，難以應用於一般中小型加工廠。考量機器學習廣為現代科技各領域所應用，本研究導入機器學習技術，建構一套具備彈性且成本可控的折床壓力預測模型。本研究初步選定支持向量迴歸(Support Vector Regression, SVR)、梯度提升決策樹(Gradient Boosting Decision Tree, GBDT)、極端梯度提升(Extreme Gradient Boosting, XGBoost)、極端隨機樹(Extremely Randomized Trees, ETR)與隨機森林(Random Forest, RF)共五種監督式學習模型進行資料預處理，經由誤差指標、學習曲線等多面向評估，篩選出SVR、GBDT與XGBoost三種表現較佳的模型，進一步進行超參數優化、正則化與田口試驗設計，最終發現SVR模型於泛化能力與預測穩定性方面表現最佳，特別是在低壓區段與貼膜材料樣本中，相較於內建系統與其他模型皆有較佳準確性。儘管高折彎角度與長材料樣本仍有偏差，經過L18田口表格補充資料後，模型效能明顯提升，驗證資料完整性的重要性。本研究所提出之預測模型，建構成本低、部署彈性高，適用於中小型加工廠多變製程條件，並具備拓展應用於模具補償、刀具壽命監控、數位雙生模擬等智慧製造場景的潛力。透過系統性建模與驗證流程，本研究揭示機器學習方法於折床壓力參數預測的可行性，為傳統加工產業智慧升級提供參考依據。

The reliance of traditional press brake processing on operator experience affects process stability and machining accuracy, and also limits the international competitiveness of domestically produced press brake machine tools. For domestically produced press brake machine tools, the current pressure prediction and compensation modules are mostly based on fixed formulas, and have limited responsiveness to a typical situation such as film-coated materials, sheet thickness variation, and die wear. Although Finite Element Analysis (FEA) has physical interpretability, its modeling requires mechanical theory and has high computational costs, making it difficult to apply to general small and medium-sized processing plants. Considering that machine learning is widely applied in various fields of modern technology, this study attempts to introduce machine learning technology to construct a flexible and cost-controllable pressure prediction model for press brakes. This study initially selects five supervised learning models—Support Vector Regression (SVR), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Extremely Randomized Trees (ETR), and Random Forest (RF)—to perform data preprocessing. Through multi-faceted evaluation such as error metrics and learning curves, SVR, GBDT, and XGBoost, the three models with better performance, are selected for further hyperparameter optimization, regularization, and Taguchi experimental design. It is finally found that the SVR model demonstrates the best performance in terms of generalization ability and prediction stability, especially in low-pressure regions and film-coated material samples, showing better accuracy compared to the built-in system and other models, and presenting good application potential. Although there are still deviations in high bending angles and long material samples, after supplementing the data through the L18 Taguchi design, the model performance improves significantly, verifying the importance of data completeness. The prediction model proposed in this study has low construction cost and high deployment flexibility, suitable for the diverse processing conditions of small and medium-sized processing plants, and has the potential to be extended to smart manufacturing scenarios such as die compensation, tool life monitoring, and digital twin simulation. Through systematic modeling and verification processes, this study reveals the feasibility of using machine learning for press brake pressure parameter prediction and provides a reference for the smart upgrading of traditional manufacturing industries.