Partial discharge pattern classification for an oil-pressboard interface

This paper compares the ability of a Single Neural Network (SNN) and an Ensemble Neural Network (ENN) in classifying and discriminating oil-pressboard interface partial discharge (PD) degredation. Discharges were sustained for 15 hours and PD patterns captured, evaluated and correlated with the tracking damage on the pressboard surface. For the same experimental arrangement two samples were stressed, one at 18.5kV (rms) and the other at 30kV (rms). Training data for both the SNN and the ENN comprised statistical parameters obtained from the Φ-q-n discharge patterns. Data sets for application were were split into periods of the first 9 hours and last 6 hours, as these time periods appeared to show most variabilty and stabilty of the statisical paremeters respectively. The results show that both the ENN and the SNN are able to discriminate the Φ-q-n patterns over these periods. It is also shown that the ENN always provides a higher recognition rate for unseen trained data while the SNN actually appears to show a higher ability to discriminate the patterns.