Evaluation Of Partial Discharge Severity Of Power Equipment Based On Meta Learning

Partial discharge is an important cause and external manifestation of insulation deterioration of power equipment.Detection and analysis of partial discharge signal is of great significance to prevent latent failure of power equipment and ensure the safe operation of equipment.Aiming at the partial discharge problem of power equipment,the research contents of this paper are as follows:(1)Based on pulse current method to set up the experimental conditions of partial discharge detection platform,the design of the corona discharge,air gap discharge,creeping discharge and suspended discharge four typical insulation defect models,a collection of four types of typical defects under the step voltage of partial discharge signals,and each type of partial discharge signals according to the pressure of time be divided into four stages of development,corresponding four severity labels.The detection data of onsite live switchgear are collected,and the severity of partial discharge of switchgear is divided according to the disintegration test results.(2)A denoising method of partial discharge signal based on improved variational mode decomposition(VMD)and wavelet transform is proposed.Firstly,on the basis of VMD theory,the preset parameters of VMD are optimized by combining frequency domain analysis and particle swarm optimization(PSO).Then,the optimized VMD is used to decompose the denoised discharge signal into different modal components,and the effective components are screened according to kurtosis.Finally,hard threshold wavelet transform is used to remove residual noise for each effective component,and then the denoised signal is obtained by superposition of each effective component.The simulation results show that the denoising effect of the proposed method is better than that of the traditional method,which provides a favorable condition for the subsequent evaluation of pd severity.(3)In this paper,a partial discharge severity assessment method based on modelagnostic meta-learning(MAML)and convolutional neural network long and short-term memory network(CNN-LSTM)is proposed.Firstly,the severity assessment problems of partial discharge signals under different defect types and measurement methods are divided into different tasks,and a set of initialization parameters with quick adaptability to various tasks are obtained by using MAML algorithm training.Then,the CNN-LSTM network parameters suitable for a specific task are obtained by fine-tuning the initialization parameters with a small number of specific task samples.Finally,the accuracy of the model is verified by test set samples.The results show that this method is suitable for pd severity assessment of different defect types and measurement methods,and has a high accuracy for pd severity assessment of unknown defect types.