CHEN Xiao ¹, XU Shu ¹, ZHANG Chengwei ¹, XU Haiyuan ¹, MIN Jianliang ²
(1. Shenzhen Power Supply Bureau Co., LTD., China Southern Power Grid, Shenzhen 518000, Guangdong, China; 2. Key Laboratory of Intelligent Transportation System, School of Intelligent Systems Engineering, Sun Yat-sen University, Guangzhou 510006, Guangdong, China)

Abstract: The traditional inspection of ceramic insulator surface defects on power transmission towers has various problems, such as low detection efficiency, poor accuracy, high leakage rate and so on. A UAV intelligent inspection method, based on improved U-net network, was proposed to detect and identify the surface defects of ceramic insulators on power transmission towers. In this method, the null convolution matrix was used to optimize the expansion coefficient of the convolutional layer in the U-net network, which helped to increase the convolutional kernel perception field to enhance the integrity of detail defect information. Secondly, a full-scale skip connection model was employed to fuse high-level feature information with low-dimensional feature information, thus increasing the detection accuracy of the ceramic insulator surface defects. Experimental results indicated that the UAV intelligent inspection method has an identification accuracy (Accuracy) of 97.6%, an average accuracy (mPA) of 95.28% and an average cross-merge ratio (mIOU) of 91.56%. Comparatively, this method for surface defects of ceramic insulators achieved an increment in accuracy by 7.8%.
Key words: ceramic insulator; surface defect; improved U-net network; empty convolution matrix; full-scale skip connection model