Improving recognition accuracy of partial discharge patterns by image-oriented feature extraction and selection technique

This paper aims to improve recognition accuracies of single sources of partial discharge (PD) by employing image-oriented feature extraction and selection algorithms. 419 diversified samples of PD data acquired from typically artificial defect models, where the defect size, applied voltage and insulation aging are taken into account, are employed for algorithms validation. In the proposed method, the PD gray images are decomposed into various vectors by two-dimensional principal component analysis (2DPCA) on horizontal and vertical directions, respectively, in which 9 representative parameters are extracted from each image vector. Based on the 419 PD samples, the proposed image features are confirmed to achieve better PD recognition performances than typical image compression methods relevant to 2DPCA. To further improve the recognition performance and reduce the feature dimension, the feature selection technique based on non-dominated sorting genetic algorithm II (NSGA-II) is adopted to search the optimal 2DPCA features. By using the proposed image-oriented feature extraction and selection algorithm, fuzzy k-nearest neighbor classifier (FkNNC), back-propagation neural network (BPNN) and support vector machine (SVM) with 16-dimensional 2DPCA features can now achieve 96.4%, 94.4% and 97.5% accuracies, respectively. The increment of average recognition accuracies of 5%-7% are obtained by different classifiers compared with phase-resolved partial discharge (PRPD) statistical features in previous studies. The investigations on the recognition results relevant to image distortion, cavity discharge change with time and PD contaminated by random noises with different level confirm the robustness of the 16-dimensional 2DPCA features, which signify a great potential in applying the proposed method to field PD pattern recognition in the future.