INTEGRATION OF MULTILAYER PERCEPTRON NETWORKS AND LINEAR DYNAMIC-MODELS - A HAMMERSTEIN MODELING APPROACH

Recently, neural network dynamic modeling has drawn a great deal of attention not only from academia but also from industry. It has been shown that neural networks can learn to mimic the behavior of a complex chemical process. However, a neural network dynamic model might be difficult to develop when training data do not contain sufficient nonlinear information about the process. In addition, owing to the capacity of a modern control system, a great deal of steady-state data is stored. Most modeling techniques to date do not take into account the fact that the transient data may not contain sufficient nonlinear information and that the nonlinear steady-state data are abundant and easily accessible. In order to fully utilize the abundant steady-state information, this paper presents a neural network Hammerstein (NNH) modeling approach. In the proposed approach, a static neural network is integrated with a linear dynamic model in series. A constrained gradient-based algorithm is derived for training the linear dynamic model in order to retain the nonlinearity established by the static neural network. To demonstrate the proposed modeling approach, a complex polymerization process is used as an example.