Case-based Reasoning System For Transformer Fault Diagnosis Based On Deepand Wide Neural Network

Transformer is the pivotal device for power transmission and conversion in the power supply system.Its reliable operation is essential for the safe and economic operation of the entire power system.Therefore,it is necessary to pre-diagnose the transformer.This paper proposes a case-based reasoning system for transformer fault diagnosis based on deep and wide neural network.Firstly,the dissolved gas method in transformer oil is introduced.Some common causes of transformer fault are analyzed,and the relationship between features and fault types is presented.Then the fault diagnosis method based on case-based reasoning is proposed.The case retrieval is realized by K-Nearest Neighbor.The cross-validation method is used to compare the influence of K on the diagnosis results,and an optimal K value is determined.The case-based reasoning system is compared with the support vector machine through simulation experiments,and the classification performance of the two is compared.Then,an improvement is made to the case representation of the case-based reasoning system,and the sample features of each transformer are represented as embedded vectors.Then two improved deep-wide neural network algorithms,which named Neural Network &the Generalized Linear Parallel Model(DWM)and Neural Network & Factorization Machine Model(DNN-FM),are proposed to train and optimize the parameters in the embedded vector.In the training process of the two models,take the influence of feature combination on training results into account: 1)Added cross-features of discrete features in a wide model with memorization and used a deep neural network with generalization to capture high-order interactions between continuous features.Parameters of the above two models were optimized by joint training.The DWM model is constructed;2)Used a deep neural network with generalization to capture high-order interactions of continuous features and used a factorization machine with memorization to realize low-order interactions of discrete features,the two part above shares the same input.The DNN-FM model is constructed.When training models,the sample is divided into the training set and the test set.The transformer fault type is regarded as the label.The model is trained in combination with the label of the training set to realize supervised model training method.The two models proposed in this paper are compared with the shallow BP neural network by simulation experiments,and the classification performance of the three models is compared.After the case is embedded in the vector,it is written into the database,and the KD tree is constructed to realize the case retrieval link.The similarity calculation is performed by the Euclidean distance method.The CBR based on the two algorithms proposed is used to diagnose the transformer,and the results are compared with the results obtained by the support vector machine to analyze the advantages and disadvantages of various methods.Finally,this paper builds a B/S framework.In this framework,the core of the system function implementation is centralized on the server,and the system interacts with the database through the server,so that the staff can operate the system anytime and anywhere through the browser.