Fault-tolerance analysis of neural network for high voltage transmission line fault diagnosis

Abstract

On the basis of theoretical analysis and testing of the fault-tolerance property for two kinds of associative memory (AM) neural networks (NN) (multiple-layered feedforward NN-FNN and feedback Hopfield NN), this paper proposes to adopt the Hopfield NN with high fault-tolerance property to solve the problem of transmission line fault diagnosis. It builds an AM NN model structure to realize transmission line fault diagnosis, presents a generalized converse learning algorithm by improving a projection based fake converse learning algorithm and the principle of dividing original samples into modules, and jointly uses them to train NN; thus original samples can be fully memorized, so that the NN's storage capacity and fault-tolerance ability can be increased greatly. Results show that Hopfield AM NN possesses high fault-tolerance ability for disturbed real-time input information sequences, and its fault-tolerance property is obviously better than FNN's, this also sufficiently demonstrates the practical application superiority of Hopfield AM NN in power system fault diagnosis.