PNN AND ITS ADAPTIVE VERSION - AN INGENIOUS APPROACH TO PD PATTERN CLASSIFICATION COMPARED WITH BPA NETWORK

Balasundaram Karthikeyan – Srinivasan Gopal – Srinivasan Venkatesh –Subramanian Saravanan

   The reliability of insulation systems is a major requirement of any power apparatus. The incidence of minor flaws and irregularities such as voids, surface imperfections etc, in insulation systems is however inevitable and leads to partial discharges (PD). Classification of PD patterns plays an important role during manufacturing and on-site assessment of power apparatus. The innovative trend of using artificial neural network towards classification of PD patterns is perceptible. A novel method for the classification of PD patterns using the original probabilistic neural network (PNN) and its variation has been proposed and implemented in this work. The classification of single-type insulation defects such as voids, surface discharges and corona has been considered primarily. The efficacy and merits of PNN and its adaptive version over that of the back propagation algorithm based feed forward neural network has been established through exhaustive comparisons on the performance of the neural networks in PD pattern classification task.

Keywords: partial discharge, probabilistic neural network, smoothing parameter, Bayesian strategy, back propagation algorithm, adaptive probabilistic neural network

[full-paper]