Application Research Of Transformer DGA Fault Diagnosis Based On Deep Learnin

Power transformer fault diagnosis can effectively avoid bad losses in the process of service,and it was an important means to ensure the continuous operation of the power system.Dissolved Gas Analysis(DGA)in oil can flexibly identify the internal faults of transformers,and it was an effective method in the field of transformer fault diagnosis.However,the current research on transformer fault diagnosis based on DGA has the following problems.First,the characteristic expression of transformer fault samples was insufficient,which cannot reflect the relationship between fault and characteristic gas.Second,it was difficult to obtain transformer fault samples,and there were insufficient sample data and the balance of samples among various fault types was difficult to meet the requirements.Third,the accuracy of fault diagnosis can not meet practical engineering requirements.This paper aimed at the above problems from the following three aspects.(1)Determination of transformer DGA data sample characteristics.The ratio between different gases was used to expand the feature vector,and the comprehensive factor was extracted scientifically by the method of factor analysis to fully express the feature.The results of different diagnostic models showed that the expanded feature can further enrich the correlation information between the feature and the fault,and the feature processed by the factor analysis method was used as the input of the diagnostic model.It can improve the accuracy of fault model diagnosis.(2)The expansion of the number of fault samples and the balance of the number of different kinds of faults.By using the the conditional generative adversarial network to expand the number of fault samples and balance the number of different fault types,the similarity of the probability distribution between the generated sample data and the original data was measured by negative logarithmic likelihood function,and a fault diagnosis model based on convolutional neural network was established to analyze the influence of the addition of generated data samples on the diagnosis effect of the fault diagnosis model.Compared with other generation data models,conditional generation adversarial network generation has better data quality and can improve the learning ability of the fault diagnosis model.(3)Transformer fault diagnosis model based on Ada Boost-CNN.Aiming at the low accuracy of transformer fault diagnosis,firstly,three integrated algorithm models were established,and Adaptive Boosting(Ada Boost)algorithm was validated for its superiority in transformer fault diagnosis.Subsequently,by using the Convolutional Neural Network(CNN)as its basic classifier,Ada Boost can weigh each convolutional layer in CNN and adjust the weights according to the errors of the model,while CNN learns new features to improve the accuracy of the diagnosis.By comparing with SVM,Decision Tree,single CNN,Ada Boost-Decision and Ada Boost-SVM,the validity and accuracy of the proposed model in transformer fault diagnosis and its feasibility in engineering application are verified.