Comparative study of feature extraction methods applied to partial discharge signals

This paper aims to provide comparative study of different feature extraction methods applied to partial discharge signals. It has been clearly shown in research literature that partial discharge signals generated from different PD sources can have different shapes. It is highly beneficial to separate and classify such different signals as different signal shapes can indicate different underlying physical mechanisms and states. Time and frequency localisation characteristics, discrete wavelet composition (DWT) and principal component analysis (PCA) are popular feature extraction techniques that have been widely applied to signal analysis. While these techniques have their own advantages and disadvantages, their efficiency can be domain dependent. In this paper, we applied all three techniques to PD pulse analysis in order to identify the most suitable method for the purpose of PD pulse separation and anomaly detection. The comparisons are based on case studies using real data collected on site. We will show that PCA outperforms the other two methods in terms of finding efficient features for pulse separation purpose.