Separation of partial discharges from pulse-shaped noise signals with the help of neural networks

Abstract

The paper introduces a method to separate partial discharges (PDs) from pulse-shaped noise signals using a neural network. After a short introduction to the problems of PD measurements on-site, the structure of neural networks and their ability for pattern recognition is presented. The adaptive resonance theory (ART) architectures, which are suitable for PD measurement, and especially the fast simulating algorithm ART 2-A, are explained. To ensure the suitability of the chosen network for PD measurement, the electrical noises and PD signals measured on a distribution transformer as well as on a high voltage transformer are classified. Furthermore, it is shown that the same algorithm with changed parameters can make a contribution to PD localisation in a transformer. This takes place with the help of calibration pulses, which are injected in different points of a transformer coil. It is shown that the ART 2-A network is able to classify these pulses in accordance with their origin for the distribution transformer. The paper ends with an examination of the signals measured on a power transformer under high voltage on-site.