Research On Intelligent Method Of Transformer Fault Diagnosis Based On Dissolved Gas-in-oil Analysis

Using the data of the dissolved gases composition and content can effectively detect the faults of the power transformer.The dissolved gas-in-oil analysis technology has matured after years of research and application,and it is still the first choice for transformer fault diagnosis.The development trend of dissolved gas-in-oil analysis technology is to combine with advanced technologies such as artificial intelligence to get rid of manual dependence and improve the accuracy of fault diagnosis.Based on the existing research results,this paper proposes an intelligent method for transformer fault diagnosis.The main content includes four aspects: correlation analysis between transformer fault and dissolved gases composition and content,intelligent diagnosis method based on single time section data,intelligent diagnosis method considering new fault types and intelligent diagnosis method considering time series correlation characteristics.The main work and innovative results are as follows:(1)The gas production principle and composition of the dissolved gases in the transformer oil are analyzed.The correlation characteristics between the transformer fault type and the dissolved gases composition and content are explored.The typical methods and their shortcomings of using the dissolved gas-in-oil analysis technology to determine whether there are faults inside the transformer and the specific fault types are introduced.(2)A Stacked Contractive Auto-Encoder Net for transformer fault diagnosis is proposed.This model can effectively extract the implicit correspondences between the state of the transformer and the dissolved gases compositons and content.When using the single-time section data of dissolved gases for testing,the diagnostic accuracy rate reaches 93.54%.(3)A two-stage transformer fault diagnosis model combining Stacked Contractive AutoEncoder Net and Isolated Forest algorithm is proposed.This model can overcome the shortcoming of the single-stage model in identifying new fault type data by bringing the step of identifying abnormal data forward.when compared with the thresholded Softmax classifier model,the two-stage model has better performance in identifying new fault types.(4)A transformer fault diagnosis model considering the timing correlation characteristics as well as combining Stacked Contractive Auto-Encoder Net and Gated Recurrent Unit is proposed.This model has the advantages of Stacked Auto-Encoder,Contractive Auto-Encoder and Gated Recurrent Unit.It can mine deep features such as the time series correlation characteristics of dissolved gases data,and can find anomalies based on the trend of the sequence data when the transformer has a potential failure trend.When using the time series data of dissolved gases for testing,the diagnostic accuracy rate reaches 94.31%.