The use of neural networks for discrimination of partial discharges in transformer oil

Abstract

The detection of partial discharges in laboratory samples of transformer oil has been performed using both electrical and ultrasonic techniques. The oil was subjected to a pin-plane electric field in a purpose built cell. The oil was contaminated using particles of a given size with only one size being used during any one experiment. Electrical signals produced by discharges were recorded using a current measuring resistor. Acoustic signals were taken by fixing a commercial acoustic emission probe to the exterior of the cell and recording acoustic signals given by the discharge. The electric and acoustic signals were analysed and yielded little to allow for discrimination of the particle type. Analysis of the signals in the frequency domain yielded similar negative results. A simple 3-layer back propagation neural network was used to try and discriminate between the particle types by using their frequency spectrum. The use of the frequency domain allows the network to be trained without having to accurately position the discharge signals in the data window. With acoustic signals, the training data was recognised to 100% accuracy, with test data then being recognised with over 95% accuracy. Electrical signals were less amenable to discrimination, with 95% of the training data being recognised correctly and over 90% of the test data. Reasons for the poorer recognised performance with electrical signals may rest with electromagnetic interference and stray capacitance.