Influence of modifications in partial discharge patterns on a neural network's recognition ability

For several years the interest in automated analysis of partial discharge (PD) measurement has increased, because manual PD pattern interpretation by human experts is one of the most powerful diagnostic methods for design and improvement of electrical HV equipment. The main aspect is the interpretation of the phase resolved PD patterns. A database of 4000 PD patterns of on-site measurements has now been used for basic investigations with neural networks. In a first step, 500 of them have been pre-selected and classified by a human expert. These files have been classified into 44 classes. For a better understanding of the general recognition abilities of NNs, patterns of on-site measurements have been artificially modified by splitting of PD matrices, using picture smoothing algorithms and zero value adjustment.