Noise reduction and source recognition of partial discharge signals in gas-insulated substation

Abstract

To minimise the failure risk of GIS due to partial discharge (PD), PD diagnosis system is required to detect and identify the harmful PD reliably and efficiently. In this thesis, two major issues for achieving successful PD diagnosis, namely: influence of noises and extraction of effective features from measured data, are addressed. A wavelet-packet-based method is developed to effectively reduce the noises. The method identifies the most significant frequency bands for noise suppression from the wavelet-packet domain using a novel variance-based criterion. The method leads to successful de-noising for signals having various noise levels. Subsequently, PD features best describing each PD source are extracted from the de-noised signals using either the method of independent component analysis or wavelet packet transform. A neural network is then employed using features extracted from either method to identify the PD sources. The relative merits of both methods of feature extraction are discussed.