Research On Feature Extraction And Pattern Recognition Of GIS Partial Discharges UHF Signal

The advantages of gas insulated substation(GIS)are good insulation,high safety and stability and small space occupied.GIS is widely used in the control and protection of high-voltage electrical equipment in power system.However,in the complex operation environment of high temperature and high pressure of power system or in the process of GIS manufacturing,assembly and transportation,there are some inevitable security risks,such as metal burr,dust particles and loose installation parts.These potential danger may lead to partial discharge(PD)phenomenon.Partial discharge will slowly erode the surrounding insulating medium.Eventually it may lead to the breakdown of the whole insulation and seriously endangering the safe operation of electrical equipment.Therefore,in the insulation state diagnosis of GIS equipment,it is necessary to detect the internal partial discharge in GIS and recognize its defect types,which is of great significance for the safe and stable operation of the power system.In this thesis,the research status of PD feature extraction and pattern recognition is analyzed,and the existing problems are summarized.The research is carried out from the aspects of GIS PD experiment,data preprocessing,feature extraction and pattern recognition.The main research achievements and innovation points are as follows:1.On the 252 k V GIS partial discharge experiment platform,this thesis set up 4 kinds of typical defects of partial discharge model,namely protrusion discharge,free moving particle discharge,surface discharge and floating electrode discharge.This thesis researches the properties of various types of partial discharge defects,and collects partial discharge ultra high frequency(UHF)signals and phase resolved partial discharge(PRPD)spectra data.After data pre-processing,this thesis builds the GIS partial discharge experiment data set.Based on the PD time analysis mode and phase analysis mode,the statistical characteristic parameters are extracted respectively.The defect types are recognized and classified by binary tree multi-classification support vector machine,and the recognition accuracy are83.25% and 86.75% respectively.2.Aiming at the problems of high feature dimension and redundant information in the feature extraction methods of GIS PD statistical parameters,this thesis proposes a feature extraction and pattern recognition method of GIS PD based on multi-scale fractal dimension.In this method,the energy distribution spectra of UHF PD signals are extracted by continuous wavelet transform,and then the multi-scale fractal dimensions are extracted as the feature attribute.The feature vector dimensions are reduced by linear discriminant analysis.Finally,the defect types are classified and recognized by support vector machine.The recognition accuracy of this method can reach 96.5%,which is significantly better than that of statistical parameter feature extraction method.In this thesis,14 other recognition models are built according to the common framework of PD pattern recognition,and the compatibility of PD feature extraction method and pattern recognition method is studied.Moreover,the experimental results of each model are analyzed and compared to verify the superiority of the proposed method.3.It is found in the research process that the selection of the feature parameters of partial discharge is subjective and limited,and it is easy to lose part of the feature information of the sample.Moreover,the feature extraction process requires more manual participation and the degree of automation is low.To solve these problem,deep learning method is introduced into PD pattern recognition in this thesis.Deep learning method can automatically extract the sample features and complete the recognition and classification,avoiding the artificial operation of feature selection and extraction.In this thesis,convolutional neural network model,deep residual network model and lightweight convolutional neural network model are designed and built for the PD PRPD spectrum data set.Conditional generative adversarial network is used to expand PRPD spectrum data set,which effectively solved the problem that the number of experimental samples could not meet the amount of deep learning data.The experimental results show that the deep learning method can efficiently and accurately complete the task of partial discharge pattern recognition.The highest recognition accuracy of the deep residual network model can reach98.75%,and the recognition accuracy of the lightweight convolutional neural network is s91.87% with short training time.The deep residual network model is applied to the GIS insulation diagnosis process,which has 95% recognition accuracy and has good application and promotion value.