Partial Discharge Identification Method Based On Deep Forest And Siamese Network

Insulation performance of power equipment is related to the reliability and safety of power supply system.Partial discharge under high voltage environment is an important cause of insulation deterioration and breakdown.The correct identification of discharge type is of great significance for defects and fault diagnosis of power equipment.Based on the integrated transformer test platform with built-in insulation defect model,the partial discharge data are generated,and the classification of discharge signals is studied by using depth model.In this paper,a sample library generation method based on time-frequency analysis of pulse segment was designed.Firstly,the Otsu adaptive double threshold method was used to determine the position of the center point and the boundary point.Secondly,the power frequency signals collected by the exper iment are intercepted into pulse waveform data of specified length.Then,an appropriate time-frequency analysis method was selected based on Renyi entropy to t ransform the pulse waveform into a time-spectrum image.Finally,the coordinate axis was removed and the grayscale was processed,and the time-frequency grayscale image sample library was generated as the input of classifier.A deep forest based partial discharge type recognition method was proposed.Firstly,following the random forest self-sampling method,the pre-training of deep forest was carried out and the image features were extracted independently.Secondly,the differences between deep forest,convolutional network and artificial feature extraction methods were analyzed,and the distribution of feature extracted by each algorithm on the two-dimensional plane was observed.Then,each classifier model was trained and the sample size of the training set was changed.Finally,the recognition accuracy of each model and its changing trend were compared.The experimental results show that the autonomous feature extraction can make the distribution of the sample points cluster according to the categori es,and the recognition accuracy of the deep forest is less affected by the change of the sample number.Aiming at the problem of low accuracy of data recognition,this paper improve d the method of type recognition based on siamese network,and designed an integrated siamese network architecture.Firstly,the improved reconstruction method of sliding window data selection was used to reconstruct the time-frequency grayscale image sample library and increased the sample size.Secondly,the branch network structure and the combination of boundary value and threshold value were designed to determine the overall structure of the siamese network.Finally,the identification accuracy of siamese networks before and after integration was compared,and the convolutional neural network and deep forest were compared.The advantages and disadvantages of each algorithm were comprehensively analyzed by introducing the calculation cost.Experimental results show that this method can achieve high accuracy of partial discharge classification.