Deep Learning-based Key Equipment Recognition Method for Substation Inspection Robots

When used for real-time detection in substation inspection robots, issues like the complex substation environment often lead to low detection accuracy for key equipment. Networks are also usually too heavy to be deployed on the edge. To solve these problems, this paper proposes the KDI-YOLO algorithm. Based on YOLOv8, it uses the MobileViT network to increase training speed and parameter efficiency, keeping the model lightweight. Then, a multi - attention mechanism detection head with third-generation deformable convolutions is adopted to enhance target localization accuracy. Experiments show that KDI-YOLO achieves a detection accuracy of 90.7% while reducing GFLOPs to 5.2. It meets the requirements for recognizing key substation equipment. Ablation and comparative experiments confirm the effectiveness and superiority of the proposed algorithm.