Optimizing power system edge computing with a high-performance and light-weight YOLO-based substation equipment defect detection network
Year: 2025
Type: article
Abstract: • Optimize YOLOv5n: Improve its multi-object detection by refining loss function, backbone, and neck networks. • Innovate modules: Devise Sigma-CIoU loss function, C3-DCNv2 module, and integrate CAM module into the neck network. • Develop YOLO-SS-tiny: A high-performance lightweight model tailored for substation equipment defect detection. • Validate experimentally: Using a custom dataset, YOLO-SS-tiny reduces total loss by 24.34 % and boosts mAP by 6.5 %. In the current fast-paced development of the power industry, traditional inspection is insufficient to meet the requirements of substation operation and maintenance and safety control. The unattended automatic operation and management mode can achieve remote monitoring functions, but a large amount of equipment data obtained during the inspection process still needs to be transmitted back to the monitoring main station for further analysis and judgment by the operation and inspection personnel. YOLOv5 is a universal object detection model, but there is still great room for improvement in multi-object detection problems such as unbalanced sample quality, mutual object occlusion, and poor distinguishability between the object and background. Therefore, this paper has made targeted improvements to the YOLOv5n loss function, backbone network and neck network, designed a new loss function Sigma-Complete Intersection over Union (CIoU) and C3-DCNv2 modules and integrated the Context Augmentation Module (CAM) into the model neck network. YOLO-Substation-tiny (YOLO-SS-tiny) high-performance lightweight model is proposed in this paper and constructs a defect detection dataset for substation equipment for multiple comparative experiments. The results show that the YOLO-SS-tiny reduces the total loss by 24.34 %, achieves mean average precision (mAP) of 69.1 %, which is 6.5 % higher than that of the original model, and reaches a detection speed of 274.2 FPS, based on increasing the parameter number by one-third of the original model. Compared with multiple advanced detection models, YOLO-SS-tiny has the best comprehensive performance and can better meet the requirements of the intelligent inspection of the distribution network in the new stage.
Authors: Qian Wang, Rui Liu, Sichen Qin, Jiawei Pu, Rong Shi, Yulu Wang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,920
Funders: National Natural Science Foundation of China, China Postdoctoral Science Foundation, Education Department of Shaanxi Province, Key Research and Development Projects of Shaanxi Province, Science and Technology Support Plan for Youth Innovation of Colleges and Universities of Shandong Province of China
OpenAlex ID: W4414653169