Transformer Fault Diagnosis By Combining Semi-supervised Generative Adversarial And Deep Stacked Auto-encoder Networks

The power transformer is the most vital devices in the power grid,which is the root of assurance of security and stabilization of electricity in the supply.Currently,due to the big data characteristics of electricity transformers,such as mass data,fast growth rate and low-value density,The historical manual labeling techniques are no more applicable.Therefore,it becomes an urgent problem to investigate the ways to have a sufficient and equitable distribution of large datasets generated with small labeled samples.To address the issue above,this paper researches the fault diagnosis model of petroleum chromatography data in semi-supervised state which is based on semisupervised generative adversarial network algorithm.The paper introduces the semi-supervised generative adversarial network-depth stacked self-coding transformer trouble diagnosis model with only a limited quantity of marked samples and a great quantity of unmarked samples for transformer oil chromatography detection.The technique produced enough and balanced sample datasets by both the generator and classifier training.It is demonstrated that the technique can achieve excellent diagnostic results on the oil chromatography of data set used in this paper.Compared with the traditional data enhancement algorithm,It can increase the classifier’s diagnostic power and address the issue of low diagnostic accuracy of few samples.By combining convolution neural network and semi-supervised generation of confrontation network,this paper improves the network structure of semi-supervised generation of confrontation network.This method can improve the diagnosis performance of classifier model and has better robustness.In this paper,EM distance and spectral normalization methods were used to improve the network stability of Semi-supervised generative adversarial networks.This method can better measure the original data distribution and generation data distribution,achieved a more balanced state,acquired the most optimum solution for the model,as well as settled the unstable question of model training.