Discharge severity evaluation of porcelain insulators based on UV images and YOLOv3

The safety operation of the power system is seriously threatened by the discharge and even flashover accidents of the porcelain insulator in wet and contamination conditions. Therefore, the research on the pollution discharge experiments of insulators were conducted in the artificial climate chamber to solve the problem above. Ultraviolet videos of different discharge states that based on the relative UV count and spot area parameters defined and their relationship between the leakage current were obtained. Then the ultraviolet image database was established and their discrimination of the four discharge severities were completed. Simultaneously, A deep learning platform was built based on YOLOv3 and then the classification of ultraviolet images was accomplished. The effects of learning rate and activation function on recognition accuracy and training errors were analysed. The network evaluation parameter and optimization methods are calculated and proposed. The recognition accuracy on the test set reaches 94.5%. This work can provide advice for engineering application.