This paper addresses the current challenges of complex substation operating environments, diverse risk factors, and relatively slow response times in traditional operation and maintenance models. Based on intelligent algorithms, it researches new methods for enhancing the operational safety and efficiency of IoT-enabled substations, establishing a comprehensive technical system integrating real-time perception, associative reasoning, and dynamic decision-making. At the methodological level, a "cloud-edge-device" collaborative framework is constructed: at the edge, an improved YOLOv8-n network is used, incorporating CBAM attention mechanism and Ghost convolution. Through channel pruning, knowledge extraction, and INT8 quantization, an inference latency of 87.3% mAP@0.5 and 78 ms is achieved. The platform layer constructs a heterogeneous graph neural network, defining four types of nodes (personnel, equipment, region, and task) and four types of edge relationships (operation, location, topology, and assignment). It employs a Heterogeneous Graph Transformer fused with a temporal GRU, achieving a risk node classification accuracy of 92.5% and an F1-score of 0.901. The application layer utilizes a near-end policy optimization algorithm, conducting Sim-to-Real training in the Unity3D digital twin environment to achieve dynamic optimization of inspection paths. Experimental results show that object detection is improved by 6.1 percentage points compared to the baseline YOLOv8-n, with the model compressed to 0.9MB; risk inference is improved by 7.3% compared to GraphSAGE, with a high-risk pattern recognition rate of 96.5%; reinforcement learning, while ensuring zero safety violations, reduces inspection time by 34.3% compared to human experience, increases coverage to 99.1%, and achieves a dynamic obstacle avoidance success rate of 96.8%.
Huang et al. (Thu,) studied this question.