Stacking Deep Scalogram Features With the Enhanced GAN for Defect Recognition in Power Transformers With Highly Imbalanced Partial Discharge Signals

Measuring partial discharge (PD) in a power transformer (PT) is one of the most crucial metrics for determining how well an insulating system is performing. However, in onsite environments, single/multiple PD sources or imbalance conditions may limit the signal’s acquisition ability. Numerous oversampling techniques are used on imbalanced datasets, but these techniques could be more extensive in displaying a non-deterministic correlation between the regional and global distributions. The proposed work uses stacking deep scalogram features with an enhanced generative adversarial network (SDSF-EGAN) for oversampling based on a minority sample global underlying structure. Three specific tactics are offered in our proposed work: The generator’s input random vectors are sampled from a rough estimate of the minority sample distribution to create fake samples more accurately; a residual about minority samples is added to the discriminator to reinforce the loss function’s constraint; and the generated samples are redistributed using a reshaper. At last, a fine-tuned VGG19 model and three different pre-trained DNN models, ResNet101, InceptionV2, and VGG16, are used for feature extraction to attain the varied SDSF map. The proposed results demonstrate that the system gets a realistic identification rate of 99.1% and is resistant to fluctuations in terms of occlusion and noise.