Methods of image recognition of overhead power line insulators and ice types based on deep weakly‐supervised and transfer learning

Abstract

Abstract Insulator icing endangers the safe operation of overhead power lines. The current measures for image monitoring of icing overhead power lines rely on manual observation. There is an urgent need for image intelligent analysis methods with strong generalisation, accuracy, and efficiency to improve power grid decision‐making. Based on the icing monitoring data from China Southern Power Grid, this paper first proposes rules for insulator image data cleaning, classification, and annotation, establishing a dataset of insulator icing monitoring images during 2014–2018. Then, based on Yolo (You only look once) v5, a weakly supervised and phased transfer learning method is proposed to recognise insulators and ice types such as snow, rime, mixed rime, glaze, and the normal. This method reduces the interference of icing background, uneven illumination, and camera occlusion through multidimensional feature fusion. The precision, recall, and mean average precision of the recognition can reach 86.6%, 91.3%, and 90.1%, respectively, and the recognition speed is as fast as 8 ms/image. Using image pseudo‐labelling based on weakly supervised learning enables the intellectualisation of image annotation, which significantly changes the inefficiency of manual labelling.