A new model structure selection method for non-linear systems in neural modelling

A new model structure selection method is proposed for neural network modelling of non-linear systems. When neural networks are employed to model a non-linear system for which no a priori knowledge is available, a problem which arises is how to determine the model structure in terms of the system order and the time delay. The model structure considered in this paper is the NARX model. The new method proposed utilizes linearisation techniques to evaluate the differentiates of the non-linear process with respect to different terms in the NARX model. Consequently, information for selection of a NARX model structure can be drawn from identification of linearised models. Neural modelling of a numerical example is investigated to demonstrate the application procedure and effectiveness of the method.