Partial Discharge Pattern Recognition Of Oil-Immersed Transformer Based On Residual Network And Generative Adversarial Network

As an important equipment in power system,the safe and stable operation of power transformer is directly related to the normal operation of power grid and the production and life of the people.The insulation system structure of oil-immersed power transformer is complicated,because of the increase of service life,it is easy to cause insulation fault,and then cause partial discharge(PD)inside the transformer.The analysis and diagnosis of partial discharge signal can quickly realize the diagnosis of transformer fault type.At present,the research on transformer PD pattern recognition mainly has the following technical difficulties:first,PD sample data is insufficient and sample imbalance among different defect categories;Second,the traditional PD mode construction and feature extraction is difficult,there are artificial experience and subjectivity,so that the fault recognition rate is low.In this paper,residual network(ResNet)and generative adversarial network(GAN)are used to solve the problem of PD pattern recognition in oilimmersed transformers.(1)Build an oil-immersed transformer PD simulation experiment platform.Complete the the phase resolved partial discharge(PRPD)spectrum sample acquisition.The generation mechanism of four typical insulation defects of transformers,including air gap discharge,suspension discharge,surface discharge and corona discharge,is analyzed,and PRPD spectra of four types of insulation defects were obtained by detecting PD signals with high frequency current method,and the spectral characteristics under different defects were analyzed.The initial sample database is constructed by preprocessing the original sample spectrum.(2)The original sample data is expanded based on the improved GAN.The spectrogram generation capability of the improved Wasserstein GAN(WGAN)was evaluated comprehensively from the aspects of loss change curve during model training,and quantitative and qualitative comparison analysis of samples generated by different network models.The results showed that the improved WGAN model could effectively improve the sample diversity and richness,and improve the generalization ability and recognition effect of the classifier under the premise of ensuring the similar distribution of sample data features.(3)The pattern recognition of PD is realized based on squeeze-and-excitation ResNet18(SE-ResNet18).Based on the attention mechanism SE module,ResNet is improved.In order to verify the actual effect of data enhancement,SE-ResNet18 was used as the network model to compare and analyze the classification effect of different data enhancement methods,and the applicability of enhanced sample data to different classifiers was discussed.The recognition performance of SE-ResNet18 is analyzed from the aspects of spectrum size and network training process,and the recognition rate of classical convolutional neural network(CNN)model and machine learning model is compared with it,which verifies the superior performance of SE-ResNet18 in the field of transformer PD pattern recognition in many aspects.The proposed method can effectively balance and expand PD samples,improve the recognition accuracy of classifier to a certain extent,and has a good application prospect in partial discharge diagnosis engineering.