In this study, two problems that can occur in designing a CNN(Convolutional Neural Network)?based partial discharge pattern classifier are presented and a solution to each is proposed.
It is difficult to distinguish between noise and discharge through a CNN-based pattern classifier because information about noise in the image of partial discharge data processed through the conventional preprocessing method is more prominent than information about the discharge signal. We analyzed that the cause of this reason lies in the conventional partial discharge data preprocessing method that stores the number of occurrences of the signal as the pixel value of the image. In order to highlight the information on the discharge signal on the image, we propose a new preprocessing method for partial discharge data that stores the average value of the signal size in the pixel value of the image by applying the orthogonal projection technique.
Through GIS simulation, partial discharge data for 5 cases of free conductor discharge, insulator discharge, corona discharge, floating discharge, and noise that can occur inside and outside of GIS are acquired. Then design three of CNN-based partial discharge pattern classifiers using images obtained by applying the conventional pre-processing method PRPS (Phase Resolved Pulse Sequence), PRPD (Phase Resolved Partial Discharge), and proposed pre-processing methods, respectively. Finally through a comparative analysis of each model''s feature map and results, the superiority of the proposed preprocessing method is proved.
In order to solve the problem of generalization performance degradation and unstable model design that can occur by using partial discharge data significantly less than the amount of data normally used to learn the CNN structure, we propose a method to design a model by adding fake partial discharge data generated from DCGAN(Deep Convolutional Generative Adversarial Network) to the training data.
The real image data used to generate the fake data is the image data obtained through the PRPS method, Lastly through a comparative analysis of the result of a model designed only with real PRPS image data and the result of a model designed by adding fake data to the real training data, the potential of proposed method for improving the problem of data shortage is confirmed.