Feature extraction and recognition of UHF partial discharge signals in GIS based on dual-tree complex wavelet transform

Ultra-high-frequency (UHF) partial discharge (PD) detection is an effective means for evaluation of dielectric condition of gas insulated switchgear (GIS). Time-resolved data pattern is utilized to recognize insulation defect because of its attractive advantages that there exists some direct relationship between the shape of PD signal and the physics of the defect. Therefore, the main task is feature extraction of time-resolved data. However, it is hard to extract features in time domain because there is no obvious difference in types of signals. UHF PD signals are non-stationary, so feature extraction in frequency domain is not suitable as well. Feature extraction in time-frequency domain based on discrete wavelet transform (DWT) is a solution for this problem, but there is a weak point which hampers the application of this method, that is, lack of shift invariance, which means that small shifts in the input signal can cause major variations in the distribution of energy between DWT coefficients at different scales. In this paper, a new method of feature extraction of UHF PD signal on the basis of dual-tree complex wavelet transform (CWT) which can reduce shift sensitive is proposed. UHF PD signals are acquired in lab and decomposed by a dual-tree CWT. Then the feature space is constructed through the energy on all frequency bands and modulus maxima on all scales. At last feature space is inputted into a radial base function (RBF) neural network classifier. The results show the superiority of this method over traditional ones. Copyright (C) 2009 John Wiley & Sons, Ltd.