부산대학교 도서관

The self-explosion of insulators in transmission lines can impact the steadiness and promptness of power transmission. This paper introduces an algorithm for detecting aerial insulator defects based on a lightweight version of YOLOv5s as a response. The Backbone and Neck of YOLOv5s are lightened using GhostNet, which reduces the model size and the number of parameters of the algorithm to meet the requirements of the algorithm's lightness and real-time performance. Subsequently, the K-means++ clustering algorithm was used on a self-assembled aerial insulator dataset, resulting in the best initial anchor box sizes. The algorithm's convergence speed was increased, and this step improved its detection accuracy. Finally, a streamlined and effective ECA (Efficient Channel Attention) attention mechanism was implemented to enhance the algorithm's feature map data, decreasing the model size and parameters while increasing the detection accuracy. The model size and number of parameters of the improved algorithm decreased by 29.8% and 31.0%, respectively, compared to YOLOv5s. The mAP was 87.3%, 2.4% higher than YOLOv5s, making the algorithm lighter and more accurate.