Research On Partial Discharge Type Recognition Based On Convolutional Neural Network

The phenomenon of partial discharge(PD)generated by gas insulated combined electrical appliance(GIS)equipment is not only a manifestation of the damage of the insulating material,but also the cause of the deepening of the damage degree of insulating material.In order to prevent GIS equipment from malfunctioning and causing losses,it is necessary to monitor the degree of insulation defect degradation of on-site GIS equipment in real time,so as to avoid serious accidents caused by insulation breakdown.Different types of partial discharge caused by different reasons in different parts of electrical equipment are also different,and the influence of different types of discharge on the work of equipment and the damage degree of insulation are also very different.Judging the discharge type of insulation defects of GIS equipment can provide guidance for understanding the degree of equipment damage,and can provide a strong basis for how to repair the system to ensure the safe and stable operation of the power system.In order to solve the problems of difficult data acquisition and unbalanced categories,poor feature selection sensitivity,and low classification and recognition accuracy of existing partial discharge recognition methods.In cooperation with an electric company in Guangzhou,this thesis proposes an improved partial discharge type recognition algorithm based on convolutional neural network.First,four typical partial discharge defect models are designed in the laboratory environment,and partial discharge signals are collected,and the phase resolved partial discharge(PRPD)map image characteristics and PRPD map statistical characteristics are extracted,which laid a foundation for PD type recognition in the following thesis.In this thesis,the methods of data expansion and transfer learning are used to solve the problem of small samples and category imbalance.In order to improve the discharge classification and recognition accuracy of four kinds of partial discharge defects,this thesis is based on the convolutional neural network(CNN)recognition method,and on the basis of the improvement of the transfer learning method,using the D-S evidence theory method,the PRPD map image feature and the map statistical feature are fused and identified.The feature richness is effectively improved,and the algorithm has a significant improvement in recognition accuracy and robustness.In the study of the activation function of convolutional neural network,three common activation functions were compared and the Relu function with better effect was finally selected.In order to verify the superiority of the convolutional neural network designed in this thesis,the traditional BPNN network and SVM network are compared.The results show that the convolutional neural network in this thesis has significantly improved the recognition accuracy of various types of partial discharges.Finally,in order to verify the improved D-S evidence theory multi-feature fusion recognition method,a comprehensive comparison experiment was conducted between this method and the migration learning + CNN method,BPNN network and SVM network.The results show that the recognition rate of multi-feature fusion network reaches a high level for air gap and free metal particle discharge with high similarity and feature overlap,which is much higher than the other three recognition methods.