- 學位論文
開發 AI 嵌入式系統於 CNC 控制器之應用
Development of AI Embedded System for CNC Controller Applications
摘要
主流的控制器往往會透過搭配機邊電腦的方式進行製程監控和生產可視化，但隨著人工智慧的興起，如加工參數優化和機台元件預診等等 AI 智慧製造技術也日臻成熟，如何能將這些 AI 技術整合至同一平台中，又同時能夠有效率地收集到高品質的感測器數據進而做為模型訓練和改進的依據，儼然成為一項重要課題。 現今的廠商或是研究大多採用 NI 擷取卡搭配機邊電腦的方式進行訊號擷取，但這樣的方案在連接多感測器時往往會需要處理資料對齊與同步擷取等等問題，同時該方案也具有較高的架構成本與有限的運算效能，較不利於控制器整合以及AI 智慧製造技術的部署。 因此，本研究擬針對 CNC 控制器相關應用建立一套 AI 嵌入式系統，其具有高效率的資料擷取、高整合性與專責 AI 運算的高性能處理單元三項優勢。本系統硬體部分採用研華 SMARC 2.0 模組方案，以 NXP i.MX 8M Plus SoC 為核心。而資料擷取部分則使用實驗室自行開發之 ECB 資料擷取模組，其具有比 NI 擷取卡更低的架構成本。 此外，本研究以.NET Framework 開發專屬集成式人機介面，使用者可在其中完成模組連線、感測器訊號擷取、數據處理以及 AI 推論等功能。該人機介面將以共享記憶體做為 IPC 技術實作，使各行程間可相互通訊和協同運作。同時，本研究亦針對該平台建立多種函式庫以完善資料流機制，並在三軸加工機的激振實驗中取得和 NI 擷取卡相符的峰值頻率特徵分析結果，亦於機台 x 軸熱變位跑合實驗中完成感測器訊號擷取與 AI 即時推論之整合驗證。相較於 CPU，在熱變位預測模型中調用神經網路處理單元（NPU）降低了30.7%的推論時間。
並列摘要
Production monitoring and visualization are usually implemented on mainstream CNC controllers by edge computing devices. However, with the growth of artificial intelligence, AI smart manufacturing techniques such as processing parameter optimization or machine component diagnosis are becoming more mature. There is an issue with how to integrate AI smart manufacturing techniques into a system or platform while efficiently collecting high-quality sensor data for model training and improvement. Machining manufacturers or researchers of today mostly use the NI DAQ system with edge computing devices for sensor signal acquisition. Nevertheless, such solution needs to address some issues such as data alignment and synchronization acquisition when connecting multiple sensors. Additionally, the solution has high setup costs and limited computing performance, making it unsuitable for CNC controller integration and the deployment of AI smart manufacturing techniques. Therefore, this research aims to establish an AI embedded system for CNC controller applications, offering three advantages: high-efficiency data acquisition, high functional integration, and high-performance processing unit for AI inference. The main hardware of this system adopts Advantech SMARC 2.0 module solution, with NXP i.MX 8M Plus SoC as its core. The data acquisition module uses the Edge Computing Box (ECB) module self-developed by the laboratory, which has the lower setup costs compared to the NI DAQ system. In addition, this research develops and designs an integrated human-machine interface (HMI) using .NET Framework, through which users can complete functions such as acquisition module connection, sensor signal acquisition, data processing, and AI inference. The human-machine interface uses shared memory as Inter-Process Communication (IPC) technique, allowing processes to communicate and operate with each other. This research also develops various function libraries for AI embedded system to improve the data flow mechanism. In the peak frequency features comparative analysis of the three-axis vibration experiment, the AI embedded system developed by this research obtains consistent results with NI DAQ system. In the tool machine X-axis thermal expansion running test, the integration verification of sensor signal acquisition and AI real-time inference is completed. Compared with the CPU, applying the neural network processing unit (NPU) in the thermal expansion error prediction model reduces the inference time by 30.7%.