This dissertation is a study of artificial intelligence for diagnosing the condition of
high voltage bushings. The techniques include neural networks, genetic algorithms,
fuzzy set theory, particle swarm optimisation, multi-classifier systems, factor analysis,
principal component analysis, multidimensional scaling, data-fusion techniques,
automatic relevance determination and autoencoders. The classification is done using
Dissolved Gas Analysis (DGA) data based on field experience together with
criteria from IEEEc57.104 and IEC60599. A review of current literature showed
that common methods for the diagnosis of bushings are: partial discharge, DGA,
tan- (dielectric dissipation factor), water content in oil, dielectric strength of oil,
acidity level (neutralisation value), visual analysis of sludge in suspension, colour of
the oil, furanic content, degree of polymerisation (DP), strength of the insulating
paper, interfacial tension or oxygen content tests. All the methods have limitations
in terms of time and accuracy in decision making. The fact that making decisions
using each of these methods individually is highly subjective, also the huge size of
the data base of historical data, as well as the loss of skills due to retirement of
experienced technical staff, highlights the need for an automated diagnosis tool that
integrates information from the many sensors and recalls the historical decisions and
learns from new information. Three classifiers that are compared in this analysis are
radial basis functions (RBF), multiple layer perceptrons (MLP) and support vector
machines (SVM). In this work 60699 bushings were classified based on ten criteria.
Classification was done based on a majority vote. The work proposes the application
of neural networks with particle swarm optimisation (PSO) and genetic algorithms
(GA) to compensate for missing data in classifying high voltage bushings. The work
also proposes the application of fuzzy set theory (FST) to diagnose the condition of
high voltage bushings. The relevance and redundancy detection methods were able
to prune the redundant measured variables and accurately diagnose the condition
of the bushing with fewer variables. Experimental results from bushings that were
evaluated in the field verified the simulations. The results of this work can help to
develop real-time monitoring and decision making tools that combine information
from chemical, electrical and mechanical measurements taken from bushings.
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:wits/oai:wiredspace.wits.ac.za:10539/4970 |
| Date | 20 June 2008 |
| Creators | Dhlamini, Sizwe Magiya |
| Source Sets | South African National ETD Portal |
| Language | English |
| Detected Language | English |
| Type | Thesis |
| Format | 1828934 bytes, application/pdf, application/pdf |
Page generated in 0.002 seconds