Recognition of Distribution Power Line Insulator Surface Defects Based on Convolutional Neural Networks

Power distribution mainly relies on overhead lines which are composed of towers, transmission lines, insulators, and other components. Condition monitoring of insulators is one of the key factors to ensure stable power transmission. Failure to deal with insulator defects (e.g., contamination, missing-cap) may result in interruption of power transmission. In this paper, a convolutional neural network-based method for surface defect identification of insulators is presented. Specifically, we propose to use the wide residual network (ResNet) whose feature map in each stage is wider than vanilla ResNet to improve the defect recognition performance of vanilla ResNet. Considering that not each layer of feature maps has the same impact on the classification performance, we further propose attention mechanism-based wide ResNet to learn the importance of feature maps. Evaluation results on our collected insulator surface defect dataset demonstrate the effectiveness of our proposed method.