Research On Insulation Aging Feature Extraction And Condition Diagnosis Of Traction Transformer Based On Big Data Analytics

In recent years,the high speed and heavy haul railways of China have been rapidly developed,in which the high-speed railway plays a crucial role in the development and construction of the One Belt and One Road of our country.Traction transformer is the core part of the traction power supply system.Apart from routine check,inspection and preventive tests,a large numbers of online monitoring systems have been installed and put into operation to master its service condition in real time.How to use the large amount of data generated in the monitoring process and combine with the relevant test and big data technology to extract effective features that can reflect the insulation aging and fault condition of the transformer,thus to prognosis its aging and health condition,is of great theoretical significance and practical value in engineering application.This paper firstly conducted the accererated aging test of the insulation paper and aquisted its mircrocosim images of different aging period and introduced the texture analysis technology to construct the texture feature of the insulation paper.Correlation analysis and orincipal component analysis were used to extract characteristics from those constructed texture features and the multivariate linear regression analysis were used to establish the relationship between texture features and the degree of polymerization(DP).Verification results using supervised and semi-supervised learnig indicated that the texture features based on optical images can effectively characterize the aging codnition of the transformer insulation paper.The texture analysis can be an effective method for non-intrusive and non-destructive detection.This study provides technical support for realizing image big data based transformer condition assessment in the future.Thereafter,a multi-layer Bayesian network model for assessing the aging and health condition of the traction transformer was constructed according to kind of condition monitoring information and statistical data and based on Bayesian net and information fusion theorem.The statistical results of condition monitoring data are proposed to determine the prior probability of each index(variable)in data layer and the principal component analysis method was proposed to determined the condition probability table and joint probability of the net to complementary with experts' experience.At the time of utilizing various condition detection/monitoring data,this model not only realized health condition assessment of the traction transformer,but also put forward the prediction method for its apparent age calculation based on probabilistic health index.Meanwhile,the big data cleaning technology was adopted to preprocess the condition monitoring data of the traction transformer.The data was then used for conducting fault diagnosis.During implementing fault diagnosis using conventional methods,it was found that disequilibrium of different classes of sample data and large quantitative difference between different demensions in one sample are primary reasons causing misdiagnosis.Therefore,on the one hand,the synthetic minority sampling sampling technology(SMOTE)was applied to balance the data distribution of different classes of samples,and the arctangent transformation was applied for preprocessing the input data of fault diagnosis network.On the other hand,the self-adaptive evolutionary extreme learning machine(SaE-ELM)was adopted for fault diagnosis utilizing its strong fitting ability for any nonlinear function.Results shown that the proposed method can improve the data structure of the data effectively and also eliminate the over fitting problems in the diagnosis process,thus improves the accuracy of fault diagnosis fundamentally.At last,in the view of the shortcomings of conventional IEEE/IEC model,knetic model and Monte Carlo simulation based statistical methods rely mainly on the oil temperature and the value of DP in transformer remaining useful life calculation,and also their defects in onestep ahead or even multi-step ahead estimation,this paper proposed to use nonlinear state equation to express the dynamic process of transformer insulation aging,and to use different condition monitoring data with combination with IEEE/IEC model and kinetic model to formulate the measurement equation,thus constructed the state-space model for RUL estimation.The particle filtering was used to solve the proposed equation,while field measured data and experimental data were utilized to verify its feasibility and effectiveness.