Fault diagnosis method for arrester in infrared images based on improved U-Net

Abstract

Conducting target detection on arresters in practical settings is challenging due to complex backgrounds and occlusion by surrounding objects. This study proposes a method based on improved U-Net for segmenting arresters in infrared images, and achieves fault diagnosis by obtaining surface temperature of the arrester. Firstly, traditional standard convolution in the U-Net backbone is replaced with depthwise separable convolution, reducing computational complexity to about 11% of the original. Secondly, ECA is introduced in the skip connection section to enable cross-channel information exchange, improving segmentation accuracy by 3.21%. Thirdly, to reduce potential errors during the conversion of grayscale images into temperature maps, the data file storage format defined in DL/T 664 is adopted, ensuring the accuracy of the temperature matrix obtained for the arrester. Finally, three fault diagnosis methods based on the surface temperature distribution curve of the arrester were proposed, and the effectiveness of the methods was verified through experiments.