Neural networks as a tool for recognition of partial discharges

Three neural networks (NNs) were used to classify partial discharge (PD) patterns: 1. Back-propagation (BP) network with hyperbolic tangent transfer function and momentum term. 2. Kohonen self-organizing map (KOH) with conscience mechanism. 3. Extended version of learning vector quantization (LVQ) network with conscience mechanism. The results showed that all three types of NNs classified correctly test finger prints of trained PD patterns. Testing of the unknown and impossible finger prints resulted in a number of misclassifications. The important factor for obtaining satisfactory classifications was the choice of the correct number of neurons, learning cycles, and values of learning coefficients. Teaching times ranged from tens of seconds in the case of the BP network to some minutes in the case of the KOH and LVQ networks