結合物理引導機器學習進行機械手臂金屬研磨表面粗糙度預測

Abstract

本論文針對協作型機械手臂於金屬研磨作業之表面粗糙度預測問題，提出結合物理建模與機器學習之預測與補償架構。考慮到協作型機器人回饋性能不足，常導致表面品質不穩定，尤其在無法即時控制研磨力與軌跡重疊區域的情境下，易產生紋理不均與品質偏差，提出了軌跡優化與殘差迭代補償來進行改善。現有文獻多聚焦於傳統加工機或單一品質指標（如 Ra 或 MRR）建模，對於協作型機械手臂尚缺乏能同時兼顧表面粗糙度預測、物理一致性與參數補償的完整方法。 本研究首先以達明協作型機械手臂（TM5-700）與研磨平台，建立金屬片表面多條軌跡的均勻去除策略，並提出結合截面效率因子與增益的材料去除率（MRR）混合模型。進一步將此物理建模所得之 MRR 嵌入機器學習回歸模型，並提出物理引導機器學習（Physics-Informed Machine Learning, PIML）架構：一方面將 MRR 作為額外特徵提升模型可解釋性，另一方面設計物理一致性損失函數（Penalty），於多宇宙優化（MVO）超參數尋優過程中，同時最小化預測誤差與物理趨勢違反率，在提高物理可解釋性的情況下也能提升模型的穩健性。 於多種研磨條件下進行系統性實驗與交叉驗證，結果顯示：PIML 架構可在預測精度（RMSE/ MAPE）與外插穩健性間取得更佳平衡，明顯降低模型違反物理趨勢的比率，並透過 SHAP 可解釋性分析驗證模型對物理主控因子（如 MRR）之合理性。比較不同模型（決策樹、SVR、XGBoost）結果，證實所提出之整合式PIML 方法，具備兼顧高精度、物理一致性與可應用於補償的優勢。

This study presents a hybrid prediction and compensation framework that integrates physical modeling with machine learning to address surface roughness prediction in automated metal grinding using collaborative robots. As collaborative arms increasingly replace manual labor in hazardous and repetitive finishing tasks, challenges such as unstable surface quality—especially when real-time force feedback and trajectory overlap control are limited—remain unsolved by most existing methods, which focus on conventional machining or single-quality metrics. This research uses a Techman TM5-700 collaborative robot and a custom grinding platform to develop a uniform multi-path removal strategy for metal plate surfaces. A hybrid material removal rate (MRR) model is established, combining theoretical and experimental calibration. The resulting dimensionless MRR feature is incorporated into machine learning regressors, and a physics-informed loss function (Penalty) is introduced to guide training via Multi-Verse Optimizer (MVO) hyperparameter tuning, enhancing both interpretability and model robustness. Experimental results across various grinding conditions show that the proposed PIML approach achieves a superior balance between predictive accuracy and physical consistency, significantly reducing physical trend violations. SHAP-based analysis confirms that dominant physical factors, such as MRR, are captured correctly. Overall, the integrated PIML method provides a robust solution for automated surface quality prediction and compensation in collaborative robotic metal grinding.