Research And Application Of Partial Discharge Pattern Recognition Technology Of Transformer Based On Deep Learning

As the core equipment of power system,transformer’s running state directly affects the safety of power system.Partial discharge(PD)is the main symptom of transformer insulation defect,accurate pattern recognition of PD signal can provide strong support for transformer fault warning and other maintenance work,which is of great significance to the safe and stable operation of power system.Traditional PD pattern recognition methods mainly have two problems:Firstly,the discharge data is scarce,so it is difficult to support the characteristic research.Secondly,the algorithms based on traditional feature extraction method have high complexity,low generalization and intelligent analysis level in practical application.In view of the limitations of current transformer PD pattern recognition field,the mechanism and characteristics of transformer PD are analyzed and the PD pattern recognition technology based on deep learning platform is deeply studied in this thesis.The main research work is as follows:First,the PD sample library is constructed by building a PD simulation test system.Firstly,the generation mechanism of different PD modes in transformer and five defect physical models including tip discharge,suspension discharge,air gap discharge in oil,surface discharge and oil gap discharge are studied.Secondly,the defect model is implanted into the transformer to build a PD simulation test system,and the transformer PD sample library is constructed.Finally,the discharge characteristics are analyzed and studied.Different from the traditional PD simulation experiment,the PD simulation experiment is carried out under the working state of transformer to obtain the real and effective PD data,and the collected PD data can truly reflect the mode characteristics of PD.Second,a structural similarity based conditional generative adversarial network(SCGAN)is proposed to break through the limitations of sample size and diversity of training sets on classification model performance,which can expand the PRPD spectrum data representing the discharge amplitude-phase information.SCGAN constrains the generated samples to maintain the same structural characteristics as the real samples by adding the structural similarity to the loss function of the generator.Label information is added to improve the model into a supervised conditional generation network to achieve controllable sample categories generated by the model.The feature decoupling effect of the non-linear mapping network is used to solve the problem of misalignment of features between classes that easily occurs in the traditional conditional generative model.Both qualitative and quantitative experiments have verified the validity of the model.Thirdly,a convolution based Vision Transformer(CViT)model applicable to PD pattern recognition is proposed to achieve automated feature extraction and pattern recognition for PRPD mapping.The CViT model has both local and global feature extraction capabilities.Firstly,the structure of multiple multi-channel convolutions is used to extract the local discharge density features of the PRPD data.Secondly,the output feature map is transformed into sequences containing phase information by the longitudinal segmentation module.Finally,the sequences are fed into the Transformer encoder in parallel,and the global structural distribution features are extracted with the help of the multi-head attention mechanism.The pattern recognition results are output by the Multi-layer Perceptron(MLP).The CViT model outperforms other models in the enhanced data sets.Furthermore,a remote fault diagnosis system for transformer is designed,which can provide a theoretical basis for the subsequent implementation of the pattern recognition algorithm research results in this thesis into specific applications.