ELECTRICAL TREE TESTS - PROBABILISTIC INFERENCE AND INSULATING MATERIAL EVALUATION

In this paper the application of neural network (NN) to the probabilis tic inference of partial discharge (PD) phenomena generated from elect rical tree growth is presented. On the basis of experimental results o f measurements of trees occurring in a needle-plane arrangement, stoch astic quantities are derived, which are relevant to PD pulse amplitude and phase. The NN trained by these quantities shows the feasibility o f evaluations that connect tree-growth stage, i.e, the amount of damag e produced by the tree, with a reduced set of these quantities. This s et is, in turn, obtained applying a NN operating for data compression. In this framework, the NN can also recognize a material, among those used for training, associating to it the specific tree-growth feature.