Partial discharge pattern recognition using variable predictive model‐based class discrimination with kernel partial least squares regression

The pattern recognition of partial discharge (PD) is critical to evaluate the insulation condition and locate the defect of a power transformer. The existing pattern recognition methods fail to make use of the inter‐relations of the extracted features of PD signals. In fact, the inter‐relations can show distinct dissimilarities between different classes of the signals. To overcome the defect of existing pattern recognition methods, the variable predictive model‐based class discrimination (VPMCD), a new pattern recognition method, is introduced for the pattern recognition of PD in this study. However, the original VPMCD lacks general expression ability of the inter‐relations of extracted features and inherits the shortcomings of least squares (LS) regression. To overcome the above defects, an improved VPMCD based on kernel partial LS (KPLS) regression, which is called as KPLS‐VPMCD, is proposed in this study. Experiments and analyses are implemented using both UCI datasets and the extracted features of PD signals. The experiments show that the performance of the proposed KPLS‐VPMCD is better than those of the existing methods such as VPMCD, back propagation neural networks, and support vector machines. The conclusion is that KPLS‐VPMCD is an efficient supervised learning algorithm with consistency and good performance for PD pattern recognition.