Study On The Evaluation Method Of Oil Paper Insulation Aging Stage Based On Local Discharge Rractal Characteristics And Convolution Neural Network

With the gradual implementation and vigorous promotion of the smart grid development plan in China,the life cycle management of power equipment as one of the specific contents has been paid more and more attention.In order to achieve life cycle management,condition monitoring and fault diagnosis are important technical support,especially insulation aging condition evaluation and prediction technology.In this paper,the oil immersed power transformer,which is widely used in power system,is taken as the research object.Based on the measured partial discharge signals of 7 aging stages,the evaluation method of oil paper insulation aging stage is studied in depth.Firstly,the thermal aging experiment and partial discharge experiment of oil paper insulation of transformer are carried out in high voltage laboratory,and the typical air gap discharge model is made.The samples with different aging degree are sampled to collect the original partial discharge signal.Adaptive spectrum subtraction is used to denoise the collected noisy PD signal.In view of the fact that the characteristics of PD signal are non-stationary and non-linear,this paper divides the PD signal into frames,introduces the spectral order reduction based on the traditional spectral subtraction,links it with each frame signal,adjusts the coefficients in the definition,and achieves better denoising effect.In the simulation,the adaptive spectrum subtraction method,db8 wavelet method and EMD threshold method are used to denoise and compare the results.Finally,the adaptive spectrum subtraction method is used to denoise the PD signal.Secondly,according to the PRPD model,two-dimensional spectrum,three-dimensional spectrum and gray-scale image with different aging degree are constructed,statistical characteristic parameters are extracted,and the statistical characteristic parameters are reduced by the principal component factor method.After dimension reduction,8-dimensional characteristic quantity is obtained;based on gray-scale image,box dimension and gap degree are extracted as fractal characteristic parameters,and the box dimension value is obtained after the scale-free area is determined by the Trifold method,The change of box dimension of samples with different aging degree was studied.In the process of research,it is found that the box dimension values in the early stage of aging can be effectively distinguished.In the later stage of aging,the box dimension values of samples are the same.Therefore,the porosity is introduced to distinguish the fractal characteristic parameters in the laterstage of aging and increase the recognition accuracy.Finally,the convolution neural network and support vector machine are combined to replace the full connection layer of convolution neural network with support vector machine.The key factors affecting the recognition results are identified one by one by using the control variable method.The CNN-SVM model is constructed by selecting the optimal parameters according to the recognition results.Based on the combination of statistical feature parameters,statistical feature parameters and fractal feature parameters,the recognition results are compared with the recognition results of traditional BP neural network algorithm,support vector machine algorithm and two improved recognition algorithms to determine the effectiveness of CNN-SVM algorithm and identify and classify the remaining sample data.The recognition accuracy is 97.14%,achieving the expected effect Fruit.There are 36 papers,16 tables and 101 references in the paper.