Time-frequency features extraction and classification of partial discharge UHF signals

Partial discharge (PD) measurement can be of practical value for condition monitoring and diagnosis of power equipment. In the current work, ultra-high-frequency (UHF) signals are measured and used to represent each PD source. A new group of time-frequency features is proposed for partial discharge classification. First of all, adaptive optimal kernel (AOK) time-frequency representation is employed to acquire the joint time-frequency information of partial discharge UHF signals, which are characterized by AOK amplitude (AOKA) matrices. Then, A new group of features are extracted from AOKA based time-frequency matrices by non-negative matrix factorization aided principal component analysis (NMF-PCA) which is developed to solve the difficulties of PCA for feature extraction of AOKA matrices due to the high dimensionality. Finally, all the extracted features are used as input vectors of fuzzy k-nearest neighbor (FkNN) classifier to obtain the PD recognition results. 600 partial discharge UHF signals sampled from four typical artificial defect models in laboratory are adopted for algorithms testing. It is shown that the maximum classification accuracy of 94.33% is obtained, which proves the effectiveness of the proposed time-frequency features. Besides, the classification performance of the NMF-PCA features is superior to that of two-dimensional NMF (2DNMF) features. The obtained results in this work provide a solid basis for the data mining technique that can be used for PD pattern recognition based on UHF detection arrangements.