Partial Discharge Data Augmentation and Pattern Recognition for Unbalanced and Small Sample Scenarios

Abstract

To improve the accuracy of partial discharge pattern recognition under unbalanced and small sample conditions, a method of partial discharge data augmentation and pattern recognition using the generative adversarial network embedded deep auto-encoder (DAE-GAN) is proposed. First, deep Auto-encoder (DAE) is embedded into a Generative Adversarial Network (GAN), and DAE is used to guide the generation process to improve the authenticity of generated samples. Second, complement samples of PD samples are added to the training process of the Generative Adversarial Network to solve the problem of small PD samples. Finally, extended equalization training samples are used to fine-tune the discriminator of the model to realize PD pattern recognition. DAE-GAN is used for data augmentation and pattern recognition of partial discharge experimental signals. The results show that, compared with other algorithms, the authenticity and probability distribution fitting accuracy of partial discharge samples generated by DAE-GAN are higher and the accuracy of partial discharge recognition is improved by 8.24% after data augmentation.