Identification of partial discharges immersed in noise in large hydro-generators based on improved wavelet selection methods

Abstract

Background noise is a major problem in online Partial Discharge (PD) detection. Particularly in large hidro-generator windings, PD pulses originated in locations far from the PD sensors are strongly attenuated due to the propagation characteristics of the pulses, and arrive to the sensors completely buried in the background noise. Therefore, it is of paramount importance to identify the PD signals whose power is at least around the same as that of noise, thus with a medium to high signal to noise ratio, to extend the PD predictive diagnosis to the innermost bars of the winding. The wavelet shrinkage technique provides the best results in eliminating this type of noise. For this purpose, it is essential to choose the most appropriate wavelet decomposition, defined by the topology of the decomposition tree, its number of levels and the selected wavelet functions. In this paper a new algorithm for the automatic selection of the number of decomposition levels is proposed, and two new methods for the selection of the wavelet decomposition filters applied to PD signals measured from two large hydro generators are advanced. In addition a new methodology to evaluate the performance of denoising methods, which takes into account the average results for all possible relative time shifts of the PD pulses and several noise threshold levels is described. One of the proposed methods presented much better results than do the traditional CBWS, EBWS and SNRBWS methods, with respect to both the denoising performance and the runtime.