Pattern Recognition of Partial Discharge in Power Transformer Based on InfoGAN and CNN

As an important equipment in the power system, it has a significant meaning in scientifically diagnosing the insulation state of oil-immersed power transformers. At present, the pattern recognition of partial discharge (PD) in a transformer has the problem of the insufficient generalization ability of the classifier due to scarcity and imbalance of samples, resulting in low recognition accuracy. To solve this problem, this paper proposes a pattern recognition of the PD method based on information maximizing generative adversarial nets (InfoGAN) and convolutional neural networks. In this method, phase-resolved partial discharge (PRPD) maps, constructed from pulse current waveforms, are chosen as the training samples. First, the InfoGAN is trained to generate new samples which expanded the original sample database, then various classifiers are trained by using the expanded sample database to realize the pattern recognition. Results of the test show that the proposed method can generate new highly similar samples more stable than other data enhancement methods, and effectively enrich the data diversity. In addition, the classifier trained by the expanded sample database has better generalization ability and is applicable to different classifiers, while residual network 18 has the highest recognition rate of 99.0%. This method can effectively balance and expand PRPD samples, and improve the recognition accuracy of the classifier to a certain extent. It has a good application prospect in PD diagnosis engineering.