Partial Discharge Data Enhancement and Pattern Recognition Method Based on a CAE-ACGAN and ResNet

In order to address the issue of the accuracy of partial discharge pattern recognition being constrained by unbalanced samples and the deep structure of the deep learning network, a method for partial discharge data enhancement and pattern recognition based on a convolutional autoencoder auxiliary classifier generative adversarial network (CAE-ACGAN) and a residual network (ResNet) is proposed. The initial step involves the preprocessing of the acquired partial discharge signals, with the phase resolved partial discharge (PRPD) spectra subsequently employed as the training samples. Secondly, a CAE-ACGAN is constructed. The model combines the advantages of a deep convolutional self-coding structure and a generative adversarial paradigm to generate high-quality phase resolved partial discharge spectrograms. Subsequently, a ResNet is employed as the classifier for partial discharge pattern recognition, utilising the CAE-ACGAN-enhanced partial discharge dataset for network training to achieve accurate recognition of partial discharge signals. The experimental findings demonstrate that the SSIM and PSNR indexes of the CAE-ACGAN model utilised in this study are 0.92 and 45.88 dB, respectively. The partial discharge pattern method employing the CAE-ACGAN and ResNet exhibits superiority in identifying partial discharges, attaining an identification accuracy of 98%, which is 7.25% higher than the pre-enhancement level.