Transformer Fault Diagnosis Based On One-dimensional Convolution Variational Autoencoder Network

With the gradual progress of smart grid construction,using artificial intelligence methods such as neural network to mine and analyze power equipment data has become the development trend of intelligent operation and maintenance of power equipment.As an important power equipment in the power system,it is of great significance to master its operating state to ensure the safe and reliable operation of the power grid.According to the different characteristics of dissolved gas in oil of oil-immersed transformer under different operating conditions,DGA(dissolved gas analysis,DGA has long been an effective means to judge its operating conditions.Based on the analysis of the existing transformer fault diagnosis methods,this paper takes the data of dissolved gas in oil as the characteristic quantity,uses artificial intelligence technology such as deep neural network as the method,and tries to combine emerging digital technology with condition monitoring to study the expansion of DGA data and the estimation of power transformer operation state.In order to solve the problems of data leakage and uneven sample distribution in the actual collection of DGA data,a data expansion method based on one-dimensional convolutional encoder is proposed to enhance the training data set.Replace the fully connected layer of the traditional self-encoder with one-dimensional convolution layer to obtain better feature learning ability;Abandon the pooling layer in the traditional convolution structure to better adapt to the characteristics of DGA data with few features.The practical application proves that this method is feasible,and the new samples generated by data expansion can effectively improve the accuracy of transformer fault diagnosis,and the generated samples are close to the actual samples.Considering the characteristic of few characteristic values of transformer gas dissolution data,a transformer fault diagnosis method based on improved graph convolution network is proposed.The traditional pool layer is replaced by convolution operation,and the dimension reduction can be realized by adjusting convolution parameters,while the fitting ability of the network can be enhanced.Finally,the multiclassification activation function is added to the top layer to realize the state estimation of the transformer.The example analysis shows that this method can effectively identify the fault type and has high diagnostic accuracy,which can better adapt to the data characteristics of dissolved gas in transformer.The performance of the data enhancement method and transformer fault diagnosis method proposed in this paper is tested by engineering examples,and compared with the current mainstream deep neural network model.The results show that the method proposed in this paper has better performance than its similar models and can better meet the engineering needs.Figure [17] Table [20] Reference [70]...