On Analyzing And Diagnosing Transformer Partial Discharge Based On Deep Learning

In the normal operation of power system,the stability of power transformer plays a very important role.Once the power transformer fails,it will cause immeasurable impact.One of the main causes of power transformer faults is partial discharge in power transformer.Therefore,the analysis of partial discharge of power transformer is an important way to detect power transformer faults and maintain the security and stability of power grid.Based on the analysis of the partial discharge signal of power transformer,this paper studies it from two aspects.First,considering that there are many useless noise interference in the discharge signal,if the signal analysis is carried out directly,the effect is very poor.Firstly,the discharge signal is denoised to ensure the signal stability;secondly,considering the limitations of traditional methods,this paper uses in-depth learning.A neural network suitable for partial discharge signal recognition is constructed.In the analysis of transformer partial discharge signal,the primar y problem is to eliminate the interference in the discharge signal and ensure the stability of the discharge signal.There are many reasons for interference in discharge environment,the most common ones are narrowband interference and white noise interfer ence.Aiming at this kind of interference,this paper presents a threshold de-noising method of stationary wavelet trace based on traditional wavelet trace method,which has good de-noising effect.This method can translate the signal,eliminate the transl ation dependence of the wavelet base,improve the particularity near the singular point and reduce the influence of random oscillation.By choosing appropriate threshold,wavelet basis and decomposition layer,the optimal value of signal denoising is obtained,and then a stationary wavelet trace threshold denoising method suitable for partial discharge signal research is constructed.Combining the simulation signal of MATLAB with the measured data of some substation,the method has a good de-noising effect.In this paper,after eliminating the useless noise interference in the discharge signal,the pattern recognition of the discharge signal is carried out.In traditional methods,feature extraction is the key step of fault diagnosis,and feature extraction mostly depends on personal experience and has some limitations.To overcome this limitation,a partial discharge pattern recognition method based on convolution neural network is proposed,which inputs the time domain waveform into the convolution neural network model.By extracting features through convolution layer,the artificial operation of feature extraction and selection is avoided,and the intelligence of recognition is enhanced.At the same time,the two characteristics of local perception and weight sharing accelerate the training speed of samples and better adapt to today’s big data era.Based on the Alexnet model,this paper trains and recognizes the discharge image.By adjusting the convolution layer and full connection layer of the convolution neural network,the network is optimized,which greatly reduces the number of neurons and the amount of calculation,reduces the training time,and the recognition effect is better than the traditional method.