Enhancing partial discharge pattern recognition via WGAN-GP and inception-resnet-v2

In this study, a novel method for identifying local discharge defects in transformers is introduced, leveraging the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and the Inception-ResNet-v2 network to enhance the recognition of partial discharge patterns. Initially, four typical partial discharge (PD) defect models are established, and phase resolved partial discharge (PRPD) spectra are collected to compile a comprehensive dataset. This dataset is then augmented using WGAN-GP, and the efficacy of this augmentation is assessed using the T-distributed Stochastic Neighbor Embedding (t-SNE) algorithm. The augmented training set undergoes subsequent processing by the classification network, wherein PD features are autonomously learned and discharge types are classified by the Inception-ResNet-v2 network. Experimental results indicate that, post-augmentation, the recognition accuracy of the Inception-ResNet-v2 network reaches 99%, marking a 5% improvement over its pre-augmentation performance and significantly surpassing that of conventional neural networks.