Impact of Dimensionality Reduction Techniques on the Classification of Ceramic Insulators Defects

Abstract

Ultra-high frequency (UHF) based testing of disc ceramic insulators has been predominantly used for the detection and classification of partial discharge (PD) defects. The initiated electromagnetic waves due to PD currents can be captured using UHF antennas. In this paper, three classes of ceramic insulator defects namely corona discharge, cracks on insulator, and voids are classified using machine learning (ML) techniques. The classification accuracies are presented with and without the use of two dimensionality reduction techniques, i.e. principal component analysis (PCA) and recursive feature elimination (RFE). A total of 322 signals were obtained from laboratory tests using a wideband Horn antenna. Then, wavelet decomposition was applied to the obtained signals, and some statistical features, which were fed to the ML algorithms, were obtained at each decomposition level. Four score metrics are used for the classification, namely accuracy, precision, recall, and f1-score. Recall (sensitivity) and f1-score are important metrics when dealing with imbalanced data. It has been shown that although PCA is very efficient in reducing the number of input features, it reduces the classification score metrics. This is attributed to the loss of important information associated with the use of PCA. On the other hand, RFE does not have a large impact on the different score metrics.