Abstract
We examine a fault-diagnosis method with an artificial neural network which recognizes partial discharges or corona occurring before full breakdowns of power cables in our research. We use a feedforward type of a neural network with three layers, i. e., input, hidden and output layers. The input, hidden and output layers include 50, 4_??_16 and 2 neurons, respectively. Connection weights of the neural network are self-organized through back-propagation learning with normalized data of corona and noise waveforms in real power cables. It is shown that this system makes it possible to recognize partial discharges automatically with the correct rate of 95% at maximum. Also examined is the relationship between feature-extracting neurons self-organized in the hidden layer and the patterns of the input waveforms. Moreover, we discuss influence of temporal shifts in the input waveforms and methods to be tolerant to the influence. This research shows prospects of diagnosing breakdowns of power cables with neurocomputing.
