Classification of EMI discharge sources using time-frequency features and multi class support vector machine

Imene Mitiche, Gordon Morison, Alan Nesbitt, Michael Hughes-Narborough, Brian G. Stewart, Philip Boreham

Research output: Contribution to journalArticle

3 Citations (Scopus)
9 Downloads (Pure)

Abstract

This paper introduces the first application of feature extraction and machine learning to Electromagnetic Interference (EMI) signals for discharge sources classification in high voltage power generating plants. This work presents an investigation on signals that represent different discharge sources, which are measured using EMI techniques from operating electrical machines within power plant. The analysis involves Time-Frequency image calculation of EMI signals using General Linear Chirplet Analysis (GLCT) which reveals both time and frequency varying characteristics. Histograms of uniform Local Binary Patterns (LBP) are implemented as a feature reduction and extraction technique for the classification of discharge sources using Multi-Class Support Vector Machine (MCSVM). The novelty that this paper introduces is the combination of GLCT and LBP applications to develop a new feature extraction algorithm applied to EMI signals classification. The proposed algorithm is demonstrated to be successful with excellent classification accuracy being achieved. For the first time, this work transfers expert's knowledge on EMI faults to an intelligent system which could potentially be exploited to develop an automatic condition monitoring system.
Original languageEnglish
Pages (from-to)261-269
Number of pages9
JournalElectric Power Systems Research
Volume163
Issue numberPart A
Early online date5 Jul 2018
DOIs
Publication statusPublished - 31 Oct 2018

Keywords

  • EMI
  • partial discharge
  • GLTC
  • uniform LBP
  • multi-class support vector machine
  • classification accuracy
  • intelligent system
  • expert system

Fingerprint Dive into the research topics of 'Classification of EMI discharge sources using time-frequency features and multi class support vector machine'. Together they form a unique fingerprint.

Cite this