Effect of training methods on the accuracy of PCA-KNN partial discharge classification model

Abstract

The aim of this paper is to describe the effect of training methods on the accuracy of PCA-KNN partial discharge (PD) classification model. This model used principal component analysis (PCA) combined with k-nearest neighbor (KNN) model, so called, PCA-KNN PD classification model for PD pattern classification. PD phenomena, corona at high voltage side in air (CHV), corona at low voltage side in air (CLV), surface discharge (SF), and internal discharge (IN) were experimented in the shielding room. Electromagnetic wave due to PD phenomena was detected using a log-periodic antenna and recorded employing a spectrum analyzer. 80 PD experiments in total were performed. The original independent variables for the classification model, skewness and kurtosis of each period of the captured signals, were calculated. To study the effect of training methods: two patterns for data training, odd/even and block training methods were investigated. In case of the block training method, the effect of training data number can be examined as well. Besides, noise signals were generated with the computer program and trained into the PD classification models. The peak of noise signal was set up at 30% of the peak value of the PD signal. These noise signals were added with the PD signals to generated a mixed noise — PD signal. Then, the mixed noise — PD signals were used to evaluate the performance of the PCA-KNN PD classification model. It was found that the block data training method provided the higher accuracy PD classification compared with the odd/event data training method. The block training method with 80% training data/20% testing data gave the highest accuracy (95% correction) for PD classification without noise signal. However, this training technique provided the lowest accuracy (56.25% correction) for PD classification with the mixed noise-PD signals.