深度學習物件偵測模型已廣泛應用於無人機與邊緣人工智慧系統，其中 YOLO 系列因具備良好的即時偵測能力而受到廣泛採用。然而，多數模型高度依賴深度學習框架，導致其內部推論流程不易分析、控制邏輯不易驗證，亦增加嵌入式部署與系統工程化之困難。為解決此問題，本論文以 YOLO26 物件偵測模型為研究對象，導入 Model-Integrated Architecture and Technology (MIAT) 方法論進行逆向工程與系統重構。首先透過 IDEF0 建立功能模型，再利用 GRAFCET 建立控制流程模型，並依據 Controller 與 DataPath 分離概念重構系統架構，完成具備獨立推論能力之 YOLO26-MIAT-C 系統實作。為驗證重建系統之正確性與可行性，分別進行 Tensor Validation、偵測結果驗證、COCO Subset 驗證以及 NVIDIA Jetson Orin NX 嵌入式部署測試。實驗結果顯示，YOLO26-MIAT-C 所產生之中間特徵張量與原始 YOLO26 模型具有高度一致性，並可穩定完成物件偵測推論工作。研究成果驗證 MIAT 方法論應用於深度學習推論系統逆向工程與重構之可行性，並建立由神經網路模型至系統工程架構之形式化轉換流程，可作為未來深度學習模型系統工程化發展之參考。;Deep learning-based object detection models have been widely adopted in unmanned aerial vehicles (UAVs) and edge artificial intelligence applications. Among these， the YOLO family has attracted considerable attention due to its real-time object detection capability. However， most deep learning models are highly dependent on framework-specific implementations， making their inference processes difficult to analyze， their control logic difficult to verify， and their deployment in embedded systems challenging. To address these issues， this thesis applies the Model-Integrated Architecture and Technology (MIAT) methodology to reverse engineer and reconstruct the YOLO26 object detection model. An IDEF0 functional model and a GRAFCET control model are established to formally represent the system architecture and execution flow. Based on the Controller–DataPath separation concept， the system is reconstructed and implemented as a standalone C-based inference system named YOLO26-MIAT-C. To verify the correctness and feasibility of the reconstructed system， tensor validation， detection result validation， COCO subset validation， and embedded deployment experiments on the NVIDIA Jetson Orin NX platform are conducted. Experimental results demonstrate a high degree of consistency between the intermediate feature tensors generated by YOLO26-MIAT-C and those of the original YOLO26 model. The reconstructed system also maintains stable object detection functionality. The proposed approach demonstrates the feasibility of applying the MIAT methodology to the reconstruction of deep learning inference systems and establishes a formal transformation workflow from neural network models to system engineering architectures， providing a foundation for future system-oriented artificial intelligence research.
