A neural network approach to identification of structural systems

Artificial neural networks are widely used for identification of nonlinear control systems. Two common approaches are multichannel neural networks and recurrent networks. The nonlinear autoregressive moving averages with exogenous input (NARMAX) model is usually used as a general input-output representation. Using the NARMAX model it is not necessary to make any assumptions regarding the structure of identified system except for the maximal values of delays. Another problem considered here is identification of a structural system, i.e., a system consisting of a few interconnected subsystems. First, we assume the internal structure of the identified system to be known and containing interconnected nonlinear static and linear or nonlinear dynamical subsystems. Identification of the system can then be performed using a neural network of a mixed linear-nonlinear perceptron architecture. Next, it is shown that recurrent models should be used to obtain uncorrelated residuals in the case of additive output noise. Finally, we also show how the backpropagation learning algorithm can be specialized to adjust weights of both the linear and nonlinear parts of the network. Some simulation examples show the high effectiveness of the structural approach (in comparison with the technique of NARX modeling).