DIAGNOSIS OF PARTIAL DISCHARGE SIGNALS USING NEURAL NETWORKS AND MINIMUM DISTANCE CLASSIFICATION

This paper describes two different methods to classify partial discharge (PD) phenomena by an automated personal computer-aided system. The first is concerned with common minimum distance classification using statistical data on pulse quantities such as apparent charge, energy and phase. Applying the correct algorithms and features, such a system is able to discriminate between unknown defects using conventional discharge patterns. The classification with neural networks discussed here offers the new possibility of classifying the shape of the PD pulses without using statistical tools for data reduction. Examples of diagnostic decisions are shown using a gas insulated switchgear system with several artificially introduced defects. The reliability of a diagnosis must be considered as a main precondition for those who are involved in detection and evaluation of partial discharges. This paper gives new information to estimate the reliability of such a diagnosis for both ways of time-resolved detection evaluated by neural networks and classic phase-resolved PD evaluation. A two step strategy of time-resolved preclassification and automated phase-resolved evaluation is introduced in this report. No measuring evidence to prove the described methods is presented in this paper, but to clarify the descriptions, a few examples are shown.