Research On Rural Network Transformer Fault Diagnosis Method Based On Fuzzy Information Entropy And GCS-EN-SCNs

As the hub power equipment of agricultural distribution network,the safe and stable operation of power transformer directly affects the reliability and economy of rural residents’life and agricultural production.Therefore,it is of great significance to further research on transformer online monitoring technology to improve the stability of agricultural distribution network and ensure the development of rural economy.At present,with the development of computer technology and the continuous improvement of Dissolved Gases Analysis(DGA)database,the data-driven transformer intelligent diagnosis model has become the current research hotspot.However,DGA data has uncertainty that was affected by transformer capacity and fault location,which directly affects the diagnostic performance of the model.In addition,the data imbalance caused by the small probability events of transformer fault leads to the normal class and ignores the feature extraction and learning of the fault data,which seriously affects the improvement effect of the model on transformer reliability in practical application.In view of this,this paper further studies the transformer fault diagnosis method from the perspective of feature selection and algorithm improvement to reduce the impact of DGA data uncertainty and imbalance on the fault diagnosis model.Firstly,this paper constructed a feature search model based on fuzzy information entropy theory and binary multi-objective particle swarm optimization algorithm(B-CMOPSO),and it was realized that feature subsets searching and evaluation in the original DGA feature space.Then,the optimal feature subsets were selected according to the classification effects of 4 common classifiers.Secondly,in order to improve the generalization ability of the diagnosis model,this paper innovatively applied Stochastic Configuration Networks(SCN_S)in the field of transformer fault diagnosis,and the integrated cost-sensitive mechanism based on class effective sample number into the solution process of the model,so as to improve the accuracy of the model in the diagnosis of a few classes of faults.Finally,the proposed model based on the preferred feature subset in this paper was compared with other common models and improved methods by simulation and field example analysis to verify the effectiveness of the model.The main conclusions of this paper are as follows:(1)The feature correlation function and redundancy function based on fuzzy information entropy could reasonably evaluate the feature information of DGA data with continuity and uncertainty.Compared with Relief F and MIC,the means of Acc of ELM,SVM,Ada Boost.M1and BPNN based on the preferred feature subset in this paper improved 12.35%,10.84%,8.12%and 6.26%on average,respectively;(2)This paper constructed a transformer hybrid feature selection model based on fuzzy information entropy theory,which could effectively reduce the impact of DGA data uncertainty on the fault diagnosis model and improve its generalization.Simulation results showed that accuracy means of ELM,SVM,Ada Boost.M1 and BPNN based on the preferred feature subset in this paper respectively increase by 17.65%,18.57%,18.21%,and 16.43%on average,compared with the input methods of Doernenburg ratio,Rogers ratio,and Ew-DGA.(3)This paper innovatively integrated SCNs,with good learning ability and generalization characteristics,with the cost-sensitive mechanism based on the class effective samples number,which could effectively improve the fault diagnosis accuracy of minority class.Simulation results showed that compared with Adaboost.M1,and ELM,BPNN,SCNs based on undersampling and oversampling balanced data methods,the model achieved the best diagnostic effect in each fault type,with the mean Avg-Acc and G-mean of 0.9260 and 0.9254and on average improvement of 0.1012 and 0.1056,respectively.In addition,the model analyzed the field instance data and compared it with the IEC three-ratio method and Duval triangle method in the national standard.The results showed that the model has the optimal diagnostic accuracy,and accurately diagnosed 12 samples out of 15 samples,which reflected the study has a certain engineering application value.