A new image-oriented feature extraction method for partial discharges

Partial discharge (PD) measurement and recognition is of great importance to assess the health condition of power transformers. However, the variation of defect size, applied voltage as well as insulation aging gives rise to the dispersion and crossover of PD features, which would influence PD recognition reliability of transformers. An image-oriented two-directional modified fuzzy-weighted two-dimensional linear discriminant analysis (TD-MFW-2DLDA) for feature extraction of PD gray images, aiming at solving the above problem, is proposed in this paper. Two classification models including fuzzy C-means (FCM) clustering and support vector machine with genetic algorithm (GA-SVM) are designed for features evaluation and PD pattern recognition. 419 PD samples of four typically artificial defect models are measured in laboratory, in which the multi-factors of defect size, applied voltage and insulation aging are taken into account, and further adopted for algorithms testing. It is shown that TD-MFW-2DLDA achieves optimal successful FCM clustering rate of about 93% and GA-SVM recognition accuracy of about 96%, which are better than that of typical PD features influenced by the multi-factors. Additionally, FCM clustering validity measures provide better compactness within class and separability between classes of TD-MFW-2DLDA which is suitable for on-site PD diagnostic applications.