An approach to fault diagnosis in dynamic systems using Kohonen neural networks

In this paper, a fault diagnosis system for linear dynamic systems based on Kohonen neural networks is proposed. The technique of pattern recognition is taken into account for the classification of the modes of operation of the system. The pattern is given by the coefficients of the transfer matrix which are estimated by a least squares algorithm; in this way classification can also be achieved under dynamic conditions. The method employs an unsupervised neural network based on competitive learning. An example is proposed to show the effectiveness of this approach.