Optimization of Cable Fault Recognition Using Fisher’s Ratio Feature Selection

Lau, Theng Sing (2018) Optimization of Cable Fault Recognition Using Fisher’s Ratio Feature Selection. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Lau Theng Sing.pdf Restricted to Registered users only Download (1MB)

Abstract

In urban distribution network, XLPE which is also known as cross-linked polyethylene has been widely used and it is one of the most famous polyethylene polymers since its first introduction to the world. There are several significant factors which contribute to its wide usage in distribution power grids. During power supply in the distribution network, XLPE possessed a high safety and reliability level and not to mention also its excellent insulating performance which contributes to its success. External condition such as environment, corrosion which caused by chemicals, mechanical stress and heat effects are some of the factors which will cause insulation failure in XLPE cables. The insulation failure in power cables can bring a serious economic loss to companies due to its vital role in power transformation and transmission. Therefore it is important to diagnose early stage of insulation failure to reduce the negative impacts. Partial discharge is created when high insulating material breaks down, measured in units of charge known as pico-Coulombs or millivolts. Partial discharge is defined as electrical discharge occurring inside a medium and high voltage insulation. The common terms for partial discharge are called flaws, cracks, voids, and irregularities. These kind of imperfections create voltage stress gradually and after some period it will eventually cause failure of insulation. Insulation failures begin with and are characterized by small but detectable releases of energy which is partial discharge. In this research, PD data which collected previously from real cable joint and artificial defects are created on the cable joint to imitate exact on-site situation rather than done it a noise-free environment and use lab fabricated insulator as material. In all the PD data measured, not all of them are important and significant in classifying different type of PD pattern. Thus, feature extraction namely Fisher’s Ratio Feature Selection is proposed to select significant data only in classifying PD pattern. On the other hand, MATLAB built in function named Relieff is also a feature extraction, which also used to compare the result to Fisher’s Ratio Feature Selection. In terms of intelligent classifiers, Artificial Neural Network and Support Vector Machine have been introduced in this research. Upon completion of this research work, it was found that ANN performs better in classifying PD data as compared to SVM due to its higher capability of noise tolerance and its ability to be insensitive towards small changes in PD data. As for feature selections, both Fisher’s Ratio and ReliefF both possess similar attributes which contributes to its similar performance but in an overall view, Fisher’s Ratio Feature Selection perform better as its accuracy is generally higher than ReliefF. Through this findings, it was also proved that, performing feature selection techniques increase the performance of the intelligent classifier significantly and thus increasing the accuracy in PD classification.