Neural networks as a tool for recognition of partial discharges

The application of three different neural networks (NNs) to the recognition of partial discharge (PD) 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 PDs are measured using conventional discharge detection, and PD patterns are processed by previously developed statistical tools. Mathematical descriptors are used as input patterns for a backpropagation network, Kohonen self-organizing map, and learning vector quantization network. All three NNs recognize fairly well the PD patterns of those insulation defects for which they were trained. On the other hand, the NNs could misclassify those PD patterns for which they were not trained. The classification of PD patterns by NNs can be influenced also by the structure of the particular NN, the value of the convergence criterion, and the number of learning cycles