Partial discharge pattern recognition method based on variable predictive model-based class discriminate and partial least squares regression

Both the feature extraction method and pattern recognition method are of great importance to assess the health condition of a power transformer. Since the partial discharge (PD) signals of the transformer are non-stationary and non-linear, and the existing pattern recognition methods fail to capitalise on the inter-relations between the extracted features of the signals, a novel pattern recognition method, namely variable predictive model based class discrimination (VPMCD) is introduced for the PD pattern recognition in this research. However, the parameters of VPMCD are estimated by using least squares (LS) regression which is sensitive to multiple correlations between independent variables. Fortunately, partial LS (PLS) regression is usable even if features are highly correlated or the number of trained samples is very small. It is novelly adopted to overcome the defects of LS regression. Therefore, an automatic PD source classifier based on PLS and VPMCD, i.e. PLS-VPMCD, is put forward in this study. PD signal features from either pulse shape characterisations or phase-resolved PD are extracted. Then, the features are used as the input vectors of PLS-VPMCD classifier. PD signals sampled from four artificial defect models are adopted for the algorithms testing. Compared with the original VPMCD and back propagation recognition methods, PLS-VPMCD has much higher recognition accuracy.