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

Zhenan Zhou,Chuyan Zhang,Min Xie,Bin Cao

DOI: https://doi.org/10.1109/tdei.2024.3355031

IF: 2.509

2024-01-01

IEEE Transactions on Dielectrics and Electrical Insulation

Abstract: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 data sets are collected based on the composite insulator surface images, and three convolutional neural networks (CNN) models VGG, GoogleNet and ResNet are established. However, the small number of images in the data set has made it difficult to achieve accurate classification, so the authors also establish three generative adversarial network (GAN) models - original GAN, DCGAN, and WGAN - to fit the existing image data and generate virtual images of the insulator surface to expand the data set. Using transfer learning, the authors obtained pre-trained models and trained them on the expanded data set. By comparing the test results of different models, the authors 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 data set. The results of this study suggest that these models have potential for engineering applications.

engineering, electrical & electronic,physics, applied