Real-Time Segmentation of Overheating Faults in Disconnectors Using YOLOv7

Substation disconnectors are critical components that are highly susceptible to overheating failures, necessitating the development of efficient and accurate monitoring solutions. This article addresses the limitations of current methods, which are constrained by limited datasets and lack real-time detection capabilities, especially when dealing with the challenges posed by infrared images, such as blurry visuals, lack of texture detail, and loss of crucial edge information. Responding to these issues, this article presents the development of a specialized disconnectors’ dataset, and we propose a reliable tool for the real-time monitoring method named thermal disconnectors’ network (TDNet) to detect overheat faults on infrared images of disconnectors based on YOLOv7. We use transposed convolution for its ability to improve the resolution of feature maps. To optimize the model’s performance, we design efficient feature fusion block (EFFB), striking a balance between robust feature extraction and computational load. Eventually, our method excels in real-time segmentation small targets in complex backgrounds with 81 fps and outperforms YOLOACT++ in key metrics, with a 2.44% increase in PA, 2.23% in mean intersection over union (mIoU) and 5.35% in $F1$ -score.