Advanced approach for automatic PRPD pattern recognition in monitoring of HV assets

This paper deals with progress in automatic approach for phase resolved partial discharge (PRPD) pattern recognition to be applied in the monitoring of HV equipment. To define classes of defects generating partial discharges (PD), a combination of scalar and vector features is proposed. The feature generation, extraction, selection and classification based on the comprehensive analysis of PRPD patterns are performed. The most important features are indicated by PD human experts and automatic knowledge extraction methods are applied. In particular, statistics, object and shape analysis, as well as fractals and wavelet analysis are taken into account. For the feature selection and reduction, different linear and non-linear methods are investigated by using supervised cluster analysis. This permits to choose the most relevant features for classification. Several classification schemes including statistical model, linear and nonlinear classifiers are evaluated in order to distinguish PRPD patterns in the feature space. The pattern recognition approach is verified on PRPD patterns derived from laboratory and field PD measurements performed on HV XLPE cables and rotating machines. The correlation with PD human experts' interpretation of such patterns was also evaluated.