Determining the structure of nonlinear models

In order to perform systems analysis or synthesis, it is compulsory to deduce a model of the process. Artificial Neural Networks (ANN) have shown their suitability to identify nonlinear dynamic processes without modelling them theoretically. Since no modelling is performed, the important issue for the Neural Network approach is to determine the required time delays. In this paper, different methods are presented that make it possible to reach this goal. First, some pruning methods are presented to detect non-required input neurons belonging to certain time delays. In order to avoid the high computational efforts of these methods, a new approach is presented which is based on the estimation of the gradient vector of the system nonlinearity. All methods are applied to a continuous-stirred tank reactor.