Power outages often happen as a result of electrical insulation breakdown in power equipment. Partial discharge is a key indicator of the occurrence of insulation deterioration. As global efforts are focused on creating a smart grid, online condition monitoring of power equipment is an area of interest to both utilities and researchers. Identifying the type of partial discharge can also determine the nature of the repair required. However, one of the challenges with online condition monitoring is background noise, making partial discharge difficult to detect.This thesis presents the investigation and development of efficient digital signal processing techniques to de-noise the signal and extract robust features for the classification of a variety of partial discharge types. It also analyses existing methods of de-noising partial discharge and develops a new adaptive thresholding algorithm that uses the arithmetic and geometric means in the discrete wavelet domain. To complement this, a Fourier Transform based signal boosting technique is also proposed, to de-noise partial discharge. Results demonstrate that the new algorithms outperform existing techniques. At a feature level, this thesis proposes two novel spectral features called Octave Frequency Moment Coefficients (OFMC) and Octave Frequency Cepstral Coefficients (OFCC) as the front-end to a classifier. In addition, a wavelet transform based feature called Time-Frequency Domain Coefficients (TFDC) is also introduced. These three spectral features provide a greater robustness and better classification accuracy compared to the well-known higher order statistical features.This thesis also introduces the use of the sparse representation classifier and compares its classification performance of partial discharge against existing classifiers such as the probabilistic neural network and support vector machine. Additionally, the use of Pronys method (pole-zero model) is introduced, to estimate the system transfer function such that the impulse response of the estimated transfer function approximately matches the original partial discharge waveform. Step-response features were extracted from the estimated system transfer function to discriminate between various types of partial discharge. Pronys method is also compared with an all-pole model to fit a digital transfer function to a partial discharge waveform. The all-pole model based step response features provided a better classification accuracy compared to the pole-zero model based features. Overall, the results presented in this thesis indicate that the novel techniques developed provide a robust solution that can be considered in online condition monitoring systems for implementation in the smart grid.
展开▼
