Hydrophobicity Classification of Composite Insulators Based on Light‐Weight Convolutional Neural Networks

To achieve accurate and rapid measurement of the hydrophobicity class (HC) of composite insulators, an intelligent spray image recognition technique based on light-weight convolutional neural networks (CNN) is proposed in this paper. A spray image data set contains clean, contaminated and aged insulators with various illuminations, shooting angles and distances, about 10 400 images of shed surface were collected by spray tests and data augmentation. Five classification models were established by different CNNs, including GoogLeNet, ResNet101, ShuffleNet 0.5×, ShuffleNet 0.25× and MobileNet V2, while the first four of them were pre-trained by ImageNet dataset. These models were trained, validated and tested by spray image data set. Six indexes were designed to evaluate each model and the discriminative regions for classification were visualized by gradient weighted class activation mapping (Grad-CAM) method. The results show that these models can effectively recognize spray images with HC1–HC7 and the light-weight ShuffleNet 0.5× has the best performance, with the classification accuracy of 97.09% for 2022 test images. The Grad-CAM visualizations indicate that the results have high reliability. This study can provide reference for on-line detection and intelligent identification of hydrophobicity levels of composite insulators. © 2022 Institute of Electrical Engineers of Japan. Published by Wiley Periodicals LLC.