LAUSR

Health condition monitoring of insulators is one of the key factors to ensure a stable power supply. This article proposes an automated insulator surface condition analysis approach based on vision measurement and the convolutional neural network (CNN). Specifically, we consider the condition analysis of insulators as an image multiclassification problem. An insulator image is first fed into a feature extractor, implemented by leveraging CNN, to extract features automatically. Then, a softmax classifier is utilized to map the feature representation into a probability vector to obtain the analysis results. Considering that neither each layer nor each location of features has the same impact on classification performance, we further focus on feature relationships and present a novel attention architectural unit, called the joint fully convolutional and multiscale spatial pooling attention (JFCMSPA) module. This new attention module not only considers the friendly fusion of different pooled information but also introduces the local similarity property of images into visual attention for the first time. These new insights allow it to efficiently learn feature importance. Finally, experimental results demonstrate the effectiveness of the proposed method and the value of its application in the inspection of power equipment.