Machine learning enhanced radio location of partial discharge

  • Ephraim Tersoo IORKYASE

Student thesis: Doctoral Thesis


Partial Discharge (PD) is a well-known indicator of plant failure in electricity facilities. A considerable proportion of assets including transformers, switch gears, and power lines are susceptible to PD due to incipient weakness of their dielectric components. These discharges may cause further degradation of the insulation, which in turn may lead to subsequent catastrophic failure. The damage that results from PD activity is worth millions of pounds and endangers the lives of personnel. PD emits electrical pulses in the form of Radio Frequency (RF) signals which propagate as a travelling wave in the vicinity of the discharge site and can be detected using dedicated sensors. This has motivated the use of an enhanced radio-based technique to detect its occurrence at early stage. Early detection of PD helps utility operators to initiate an emergency maintenance outside the scheduled times when it is most cost-effective and before the equipment loses performance or suffers catastrophic failure, hence improving asset management. Therefore this thesis presents an investigation of an enhanced machine learning approach to continuous PD localisation using a network of radio sensors. The approach being investigated relies on location dependent parameters which will be extracted from PD measurements. This thesis demonstrates RF-based fingerprinting technique for locating PD sources using Received Signal Strength (RSS). Furthermore, Signal Strength Ratios (SSR) between pairs of sensor nodes are used as robust fingerprints given that the energy emitted by each PD event may be different due to progressive nature of PD severity as deterioration continues and the fact that different types of PD occur in nature. Sophisticated machine learning techniques are investigated and used to develop PD localisation models. This work also investigates the plausibility of using other PD received signal parameters for locating PD sources. It has been found that the statistical characterisation of the received RF signals produces manifold PD features beside RSS. The developed localisation approach based on the analysis of these statistical features assumes that PDs generate unique RF spatial patterns due to the complexities and non-linearities of RF propagation. This approach exploits two distinct frequency bands which hold different PD information. PD location features are extracted from the main PD signal and the two sub-band signals. These features are then used to infer PD location. Moreover, due to the increased dimensionality of data that may result from PD feature generation, feature selection algorithm; Correlation Based Feature Selection (CFS) is employed for feature selection and dimensionality reduction. The use of statistical PD features improves localisation accuracy. This study further presents a novel method for RF-based PD localisation. The technique uses Wavelet Packet Transform (WPT) and machine learning ensemble methods to locate PDs. More specifically, the received signals are decomposed by the Maximal Overlap Discrete Wavelet Packet Transform (MODWPT) version of wavelet packet and analysed in order to identify localised PD signal patterns. The Regression Tree algorithm, Bootstrap Aggregating method and Regression Random Forest (RRF) are used to develop PD localisation models based on the wavelet PD features. The proposed PD localisation scheme has been found to successfully locate PD with negligible error. Additionally, the principle of the developed PD localisation system has been validated using a separate test dataset. This approach is based on purely practical reasons, given the enormity of separate experiments to be carried out. The data required is collated over an extended time period. The results of the investigation presented in this thesis show that an auto
Date of Award20 Sep 2019
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorRobert Atkinson (Supervisor) & John Soraghan (Supervisor)

Cite this