NEURAL NETWORKS AS A TOOL FOR RECOGNITION OF PARTIAL DISCHARGES

In this paper the application of three different neural networks (NN) to recognize partial discharge (PD) sources is studied. Results of PD measurements on simple two-electrode models as well as on models of artificial defects in industrial objects are presented. The PD were measured using conventional discharge detection and PD patterns were processed by previously developed statistical tools. Satisfactory results in the past have shown that using mathematical descriptors, the properties of the phase-position distributions can be analyzed. Therefore these descriptors were used as input patterns for back-propagation network, Kohonen self-organizing map and learning vector quantization network. All three NN, as used in this work, recognize fairly well the PD patterns of those insulation defects for which they were trained. On the other hand, the NN might misclassify those PD patterns for which they were not trained. The classifications of PD patterns by NN can be influenced also by the structure of the particular NN, the value of convergence criterion, and the number of learning cycles.