Binary Particle Swarm Optimization Selected Time-Frequency Features for Partial Discharge Signal Classification

In this study, the development of binary particle swarm optimization (BPSO) feature selection algorithm is proposed for partial discharge (PD) signal classification based on quantified time-frequency features. Firstly, adaptive optimal kernel (AOK) time-frequency representation technique is employed to obtain a high quality of time-frequency distribution of partial discharge ultra-high-frequency (UHF) signals with reasonable resolutions in both time and frequency domains. Then, a non-negative matrix factorization (NMF)-based matrix decomposition (MD) method is applied to obtain a series of base vectors in frequency domain and location vectors in time domain which are further used to extract statistical features to construct an adequate feature space representing the time-frequency information. Finally, the developed BPSO feature selection algorithm is adopted to improve PD classification performance. A fuzzy k-nearest neighbor (FkNN) classifier is responsible for the classification task and used as the fitness evaluator of BPSO. Using a UHF detector, 600 PD signals sampled from four categories of artificial defect models in the laboratory are adopted for testing. Performances of various feature sets, including all the statistical features, artificially combined features with different dimensions and BPSO selected features, are compared. Results demonstrate that the proposed feature extraction and selection algorithm can provide an effective tool for partial discharge signal classification, and it is easy to extend to other image or matrix applications. Copyright (C) 2012 Praise Worthy Prize S.r.l. - All rights reserved.