Industrial Automatic Control Systems and Controllers Annotation << Back Implementation of a Hybrid Image Analysis System in Remote Computing Systems Borodov E.V., Goryunov A.G. This paper describes the practical implementation of a hybrid anomaly detection system for deployment within industrial automation systems (IAS) at power and nuclear facilities. The focus is on efficient distribution of computational load across specialized units of the Radxa Rock 5B (RK3588) edge platform, enabling real-time operation under resource constraints. The system integrates YOLOv8n (CNN), Topological Data Analysis (TDA), Haar wavelet transform, and Shannon entropy. Each module is optimized for specific hardware: NPU handles YOLO inference via rknn-toolkit2, GPU (Mali-G57) accelerates Lucas-Kanade alignment and wavelet filtering via OpenCL, while CPU computes TDA, entropy, and feature fusion. An asynchronous architecture using queues and multithreading achieves up to 10 FPS at ~12 W power consumption. Frame skipping and ROI size limitation (64×64 pixels) balance performance and accuracy. Integration into industrial networks is implemented via MQTT and OPC UA, supporting HMI visualization, SCADA alarm transmission, and MES reporting. Experiments demonstrate a 64 % reduction in false positives compared to standalone YOLO. The work confirms the feasibility of developing interpretable, energy-efficient, and scalable soft sensors for industrial vision, ready for mass deployment on accessible edge devices. Keywords: Radxa Rock 5B, RK3588, hybrid system, CNN, NPU, GPU, Industrial Automation, OQC, Computer Vision. DOI: 10.25791/asu.4.2026.1651 Pp. 14-24.