Study On Pattern Recognition Of Power Cables Partial Discharge Based On Deep Learning

As the infrastructure equipment of power distribution system and large-scale clean energy,the reliability of the cables affects the stability and security of the power system.Partial discharge is an important characterization of the deterioration of cables insulation.Identifying partial discharge can provide a reliable basis for evaluating the risk of cables insulation defects and realizing real-time cables maintenance.Deep learning is a research hotspot in machine learning.It is widely used in image recognition,speech recognition and other fields because of its characteristics suitable for processing large-scale data.With the advent of the era of electric power big data,deep learning will have broad application prospects in the electric power industry.Therefore,this paper takes deep learning as the core technical means to study the pattern recognition of cables partial discharge.In this paper,experimental models of four types of typical partial discharge defects in cables are designed and built.By using PD data to plot time-domain pulse waveforms and 3D spectrograms of partial discharge phase distribution(PRPD).The characteristics of different types of partial discharge are studied to provide conditions for the identification of cables partial discharge.The traditional shallow neural network classifier relies on manual feature extraction,which is difficult to Process large amounts of partial discharge data.In order to solve this problem,a cables partial discharge pattern recognition method based on automatic feature extraction is proposed.Converting partial discharge time-domain signals to PRPD grayscale images,and the Convolutional Neural Networks(CNN)in deep learning are selected as pattern recognition classifiers in this method.A CNN model was built and the effects of PRPD grayscale image size,network depth,pooling method,activation function,and convolution kernel on the recognition performance of CNN were studied.Compared with the shallow neural network,such as SVM and BPNN,the recognition rate of the CNN model is higher.The number of partial discharge gray-scale images for cables is small,and it is difficult to train deep residual network model based on large-scale data sets.A method of high-precision partial discharge pattern recognition for cables under few samples condition is proposed.The method uses the idea of combining expanded samples with simplified model.In terms of data augmentation,a method for augmenting PRPD grayscale images with Deep Convolutional Generative Adversarial Network(DCGAN)in deep learning is proposed.In terms of model optimization,a structurally optimized residual network model is proposed to match the small-scale partial discharge dataset.The experimental results show that the recognition rate of the optimized residual network reaches 98.5%,and the average iteration time is only 7.3s.The DCGAN model has some problems,such as unstable training and slow convergence speed.Therefore,it is difficult to expand PRPD grayscale images.In order to solve these problems,the DCGAN model is improved from the aspects of algorithm and model structure,and a WDRGAN model is proposed.In terms of algorithm,the Wasserstein distance with gradient penalty optimization is used instead of JS divergence to improve the stability of model training.In terms of model structure,the residual network is used to build the generator of the model,which speed up the convergence speed of the model.Finally,four evaluation metrics are set,such as Inception Score,Fréchet Inception Distance,box dimension,recognition rate,to verify the effectiveness of WDRGAN generated samples.