Advanced signal processing techniques for online power transformer insulation diagnosis

Abstract

Power transformer is one of the most important assets in an electric utility. However, a large number of existing power transformers worldwide have already approached or even exceeded their designed lifetimes. Any failure of a transformer can be disastrous. Therefore, the conditions of transformers need to be continuously monitored and evaluated. Since a transformer’s condition is largely dependent on its insulation system, a number of diagnostic methods have been developed for assessing transformer insulation conditions over the past decades. Among these methods, partial discharge (PD) measurement is widely adopted due to its capability of providing continuously online monitoring and diagnosis of a transformer without disturbing its normal operation. PD is a rather complicated phenomenon and stochastic in nature. Properly performing online PD measurements of a transformer, effectively analysing the measured PD signals, and subsequently making an informed condition assessment on a transformer's insulation system are still challenging. This thesis is aimed at developing advanced signal processing techniques for online PD monitoring and diagnosis of power transformer insulation systems. PD signals acquired at substation environments are always coupled with extensive noise, which exhibits different distribution properties. Therefore, PD signal de-noising is an essential process for accurately extracting PD signals from the acquired noise-corrupted signals before further analysis. In this thesis, advanced signal processing techniques, such as wavelet transform (WT), empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), blind equalization (BE) and pre-whitening, have been investigated for removing discrete spectral interference (DSI) that exhibits sinusoidal behaviour at various frequencies. Mathematical morphology (MM) has also been investigated for suppressing white noise by adaptively selecting threshold values. To remove stochastic noise, fractal dimension and entropy analyses have been investigated. Based on these techniques, several adaptive PD signal de-noising methods have been proposed in this thesis for removing different types of noise. A number of case studies using PD signals acquired from both laboratory experiments and online PD measurements of field transformers are presented. These case studies demonstrate advantages of the proposed PD signal de-noising methods over conventional methods in PD signal de-noising. In this thesis, phase-resolved pulse sequence (PRPS) diagrams and kurtograms have also been proposed for consistent representations of PD patterns after the noise has been removed. Case studies have been provided to prove a PRPS diagram’s capability for accurately and consistently representing PD patterns. This representation can minimize influences of different types of PD sensors and measurement systems on PD pattern construction. Results are presented to demonstrate that kurtograms can be used to represent PD patterns even in the presence of extensive white noise in PD signals. In practice, it is not uncommon that more than one PD source can co-exist in a transformer and discharge simultaneously. Therefore, this thesis is also aimed at developing an effective method for separating multiple PD sources in a PD signal. Though in the literature, a time-frequency (TF) map has been developed and applied to separate multiple PD sources, its performance can sometimes be compromised due to the lack of an accurate representation of individual PD pulses. Therefore, a TF sparsity map has been proposed in this thesis. The proposed TF sparsity map is based on decomposing individual PD pulses into time and frequency domains at multi-resolutions. A MM-based signal decomposition method has been developed for consistent signal decomposition without being affected by the selected functions as in the conventional WT-based decomposition. After sparsity values are calculated from the decomposed signals in time and frequency domains, sparsity trends are determined to provide a unique representation of PD sources. By taking roughness values of the trends, an accurate separation of multiple PD sources can be obtained on a TF sparsity map. To verify the proposed PD source separation method, a test cell has been designed and manufactured to configure three PD source models simultaneously for simulating multiple PD sources. The PD source models adopted in this thesis include internal discharge, corona, surface discharge, discharge due to floating particles and discharge in transformer oil. The proposed TF sparsity map has also been verified by signals acquired from field transformers. After PD source separation, representative features can be extracted from PD signals for PD source classification. In this thesis, an investigation on the features extracted from conventional phase resolved PD (PRPD) diagrams as well as decomposed signals of DWT or EMD has been performed. A comparison on these features has also been made by evaluating classification performances of a support vector machine (SVM)-based algorithm. Results show that features extracted from decomposed signals are more effective in representing PD patterns for PD source classification when compared with those from PRPD diagrams.