Data mining on a transformer partial discharge data using the self-organizing map

Although experts all over the world have investigated methods for partial discharge (PD) detection and classification for over 50 years, until now there is still no universal method available for this purpose. Even in the future, it would not be possible to define a universal method for unambiguous classification and localization of PD sources in complex insulating systems (e.g. transformers or generators) due to the unlimited variations of PD source type and its location. This paper deals with PD signals obtained by remote radiometric measurements performed on a power transformer. Extensive PD measurements were performed in the substation. Since fingerprints of the remote radiometric measurements are not available, the formation of clusters with similar features obtained from captured PD data is crucial. A data mining technique known as self-organizing map (SOM) for the analysis and interpretation of captured PD data is used. Since each signal contains a large number of samples, before applying the SOM, the dimensionality reduction based on principal component analysis is performed. As a result of a data mining process a clear separation of PDs emanating from a transformer and discharges emanating from its surrounding is achieved