Classification Method of Composite Insulator Surface Image Based on GAN and CNN

Composite insulators have been widely used in power transmission lines due to their excellent insulation performance. However, it is impossible to completely avoid flashovers in power transmission lines, so it is crucial to ensure the safety and reliability of these lines. To address this issue, the datasets are collected based on the composite insulator surface images, and three convolutional neural network (CNN) models, VGG, GoogleNet, and ResNet, are established. However, the small number of images in the dataset has made it difficult to achieve accurate classification, so we also establish three generative adversarial network (GAN) models—original GAN, deep convolutional GAN (DCGAN), and Wasserstein GAN (WGAN)—to fit the existing image data and generate virtual images of the insulator surface to expand the dataset. Using transfer learning, we obtained pretrained models and trained them on the expanded dataset. By comparing the test results of different models, we found the best classification model with an accuracy of 97.5%. This shows that the CNN models can effectively classify the surface images of composite insulators, especially after using the virtual images generated by GAN to expand the dataset. The results of this study suggest that these models have potential for engineering applications.