用於精密加工之邊緣運算系統

Abstract

隨著工業4.0與永續製造的發展，精密加工設備逐步深度整合感測技術以實現智能化。然而，切削過程中同時存在高頻動態行為（如切削顫振）與低頻準靜態效應（如熱變形），兩者在物理機制、時間尺度與運算需求上差異顯著，使傳統集中式或單一邊緣運算架構難以同時兼顧即時反應能力與模型精度。 為此，本研究提出一套應用於精密加工之異質邊緣運算系統，整合多模態感測介面、物理與資料驅動並行之混合建模方法，並依據不同任務特性設計運算分流策略，以全面提升加工穩定性與幾何精度。針對切削加工監測需求，本研究開發多項專用感測技術以實現多物理量量測。在切削力監測方面，自主研發整合壓電元件之「智慧虎鉗」，可直接量測動態切削力，使其能廣泛應用於各類加工場景；同時，設計無線刀尖溫度量測模組，以獲取刀尖熱源資訊，彌補傳統僅依賴主軸溫度量測時，於乾式切削條件下對切削熱估測不足之問題。於人機互動與聲學感測層面，為解決工業高噪音環境與邊緣端算力受限之矛盾，本研究結合MEMS振動感測器與赫姆霍茲共振器（Helmholtz Resonator）陣列，利用其物理共振特性於類比端進行頻譜分解與特徵擷取，使訊號於數位化前即完成頻帶濾波，有效取代高耗能之數位傅立葉轉換（FFT），實現物理層即時運算。 在建模與演算法層面，本研究採取物理機制或資料驅動的建模方法。於加工穩定性預測方面，首先建立納入刀具偏擺效應之半經驗切削力模型，於離線階段生成貼近實際加工行為之穩定界限圖（SLD）；再依據物理模型所呈現之特徵規律，推導低計算複雜度之穩定性指標（Stability Index, SI），並部署於線上邊緣平台以實現顫振即時判定。針對熱誤差補償，本研究依據濕式與乾式切削之熱傳特性，分別提出對應之建模策略。在濕式切削條件下，由於主軸與刀尖熱源受冷卻液影響顯著，研究重點聚焦於主軸溫度量測點，並系統性比較多種統計學習模型；同時透過 PFI-RFE 與 SFS 特徵選取方法，篩選具高泛化能力之溫度特徵組合。於乾式切削情境中，考量刀尖與工件間劇烈溫升所造成之高溫熱源轉移效應，本研究進一步整合無線刀尖溫度資訊，經交叉驗證後選定 Ridge 迴歸演算法，以建立具強健性的熱誤差補償模型，有效降低預測偏差。此外，針對 MEMS 感測器先天頻寬不足所導致之高頻資訊缺失問題，本研究利用深度神經網路（DNN）建立非線性映射關係，於數位端重建高頻頻譜，驗證資料驅動方法補償物理感測限制之可行性。 在系統部署層面，本研究依任務特性實現異質運算分流：將需毫秒級反應之顫振監測部署於輕量化邊緣平台，而計算量較大且時間變化較緩之熱誤差補償則由 PC 平台執行，並透過 VMX 通訊中介層與 CNC 控制器整合，形成閉迴路即時補償機制。實驗結果顯示，本系統可於毫秒等級內完成顫振辨識，並於乾式切削條件下降低 80% 以上之熱誤差。研究結果證實，所提出之異質邊緣運算架構能有效處理具多時間尺度與多物理特性之加工訊號，在即時性與運算深度之間取得良好平衡，為精密加工智慧監控提供一套高度整合且具實用價值之系統性解決方案。

With the advancement of Industry 4.0 and sustainable manufacturing, precision machining equipment increasingly integrates sensing technologies to achieve intelligence. However, the simultaneous presence of high-frequency dynamic behaviors (e.g., cutting chatter) and low-frequency quasi-static effects (e.g., thermal deformation) during the cutting process reveals significant disparities in physical mechanisms, time scales, and computational requirements. These differences make traditional centralized or single-tier edge computing architectures insufficient for balancing real-time responsiveness with model accuracy. To address these challenges, this research proposes a heterogeneous edge computing system for precision machining. The system integrates multi-modal sensing interfaces, a hybrid modeling approach combining physical and data-driven methods, and a task-specific computation offloading strategy to enhance machining stability and geometric accuracy. To facilitate comprehensive process monitoring, a suite of specialized sensing interfaces is developed to capture multi-physical phenomena. In cutting force monitoring, a “Smart Vise” integrating piezoelectric elements is developed to directly measure dynamic cutting forces, ensuring broad applicability across diverse machining scenarios. Simultaneously, a wireless tool-tip temperature measurement module is designed to obtain heat source information. This module compensates for the limitations of traditional spindle temperature measurement, which often underestimates cutting heat under dry cutting conditions. Regarding human-machine interaction and acoustic sensing, to resolve the conflict between high-noise industrial environments and limited edge computing power, MEMS vibration sensors are combined with Helmholtz Resonator arrays. By leveraging physical resonance for spectral decomposition and feature extraction in the analog domain, frequency band filtering is completed prior to digitization. This effectively replaces power-hungry Digital Fourier Transforms (FFT) and achieves real-time computation at the physical layer. At the modeling and algorithmic level, this research adopts both physical-based and data-driven methodologies. For machining stability prediction, a semi-empirical cutting force model incorporating tool runout effects is established to generate Stability Lobe Diagrams (SLD) that reflect actual machining behavior during the offline phase. Based on the feature patterns identified by the physical model, a low-complexity Stability Index (SI) is derived and deployed on the online edge platform for real-time chatter detection. Regarding thermal error compensation, distinct modeling strategies are proposed based on the thermal transfer characteristics of wet and dry cutting. In wet cutting, where the spindle and tool-tip heat sources are significantly affected by coolant, the research focuses on spindle temperature measurement points and systematically compares various statistical learning models. Feature selection methods such as PFI-RFE and SFS are utilized to identify temperature feature combinations with high generalization capabilities. In dry cutting scenarios, considering the intense heat source transfer caused by the temperature rise between the tool tip and the workpiece, wireless tool-tip temperature data is integrated. After cross-validation, the Ridge regression algorithm is selected to establish a robust thermal error compensation model, effectively reducing prediction bias. Furthermore, to address the lack of high-frequency information caused by the inherent bandwidth limitations of MEMS sensors, a Deep Neural Network (DNN) is employed to establish non-linear mapping relationships for high-frequency spectrum reconstruction in the digital domain, validating the feasibility of using data-driven methods to compensate for physical sensing limitations. Regarding system deployment, heterogeneous computation offloading is implemented based on task characteristics: chatter monitoring, which requires millisecond-level response, is deployed on a lightweight edge platform; meanwhile, thermal error compensation, characterized by larger computational loads and slower temporal changes, is executed on a PC platform. Integration with the CNC controller is achieved via the VMX communication middleware, forming a closed-loop real-time compensation mechanism. Experimental results indicate that the system identifies chatter within milliseconds and reduces thermal errors by over 80% under dry cutting conditions. The findings confirm that the proposed heterogeneous edge computing architecture effectively processes machining signals with multiple time scales and multi-physical characteristics, achieving an optimal balance between real-time performance and computational depth. This provides a highly integrated and practical systemic solution for the intelligent monitoring of precision machining.