Transformer Fault Diagnosis Based On Machine Learning

As a part of the power system,the fault of transformer will lead to the change of the operation state of the power system.The fault of transformer will not only damage and destroy itself,but also affect the power system.It is very necessary to find and deal with the transformer fault in time for the long-term operation of the transformer,and the analysis method based on dissolved gas analysis is widely used in the transformer fault diagnosis.According to the different methods combined with dissolved gas analysis,there are different kinds of transformer fault diagnosis methods with different precision in the transformer fault diagnosis.Dissolved gas analysis is a method to analyze transformer fault according to the difference of concentration of dissolved gas produced in transformer under different operation conditions.In this paper,the problem of transformer fault diagnosis based on machine learning is studied.Aiming at various problems in the process of power transformer fault diagnosis,the corresponding relationship between dissolved gas in power transformer oil and fault type is analyzed.Therefore,BP neural network is applied to power transformer fault diagnosis in this paper.In view of the problem that BP neural network is easy to fall into local optimum and slow convergence speed,a CSO-BP transformer fault diagnosis model is proposed.The weight between input layer,hidden layer and output layer of BP neural network is optimized by CSO,which solves the problem that BP neural network is easy to fall into local optimum and slow convergence speed,and improves the accuracy of fault diagnosis by 6.2%.At present,with the development of deep learning,facing the large sample data BP network can not extract data sample features,simulation run time is long.Deep learning has the advantages of large sample data feature extraction ability and short running time.In this paper,we normalize the data and apply deep belief As a model of transformer fault diagnosis,using the ability of the network to extract the features of large sample data,unlabeled samples can better train the network model,improve the accuracy of power transformer fault diagnosis,and improve the accuracy by 10.6% and 4.4% respectively compared with BP and CSO-BP models,so it is more reliable to identify fault types.In order to solve the problem of low accuracy of power transformer fault diagnosis caused by the small number of transformer faults and the small number of samples in some areas,this paper proposes four transfer learning and neural network models(Tr-BP、Tr-CSO-BP、Tr-DBN、Tr-CSO-DBN)for transformer fault diagnosis.The traditional neural network BP neural network and deep belief are used for the large sample data of power transformer DGA Neural network is trained to retain the weights of the input layer and hidden layer and transfer them to the small sample data diagnosis model.Instead of the input layer and hidden layer in the small sample data diagnosis model,the output layer of the small sample data diagnosis model is established,and the weights and thresholds of the output layer are optimized by the cross algorithm(CSO).The experimental results show that migration learning can effectively solve the problems of low accuracy and poor convergence of small sample data of power transformer DGA,with the accuracy increased by 12.3%,9.6%,9.7% and 11.2% respectively,which verifies the feasibility and effectiveness of the proposed model.