Discrete transfer function modeling of non-linear systems using neural networks

ANNs (Artificial Neural Networks) have been used successfully in various system identification applications. However, in general, the ANN-based identification is primarily nonparametric in nature, as the system information is hidden within the neural network architecture, which is not transparent enough to be modeled explicitly. In this paper, a series parallel neural network architecture has been proposed for parametric system identification, whose weights are optimized using adaptive learning through training data, collected from unknown nonlinear system. The optimized weights of the proposed ANN structure are then directly utilized to model the unknown system, in terms of an equivalent discrete transfer function model. The estimated model using proposed technique can be effectively utilized to analyze, test and control the unknown system dynamics. Simulation examples with measurement noise are used to test the effectiveness of proposed identification technique.