State-space nonlinear process modeling: Identification and universality

A practical approach was used to identify nonlinear models from available data. Knowing the true dynamics of a process allows implementation of a better and more efficient control scheme. Since this knowledge is not always available, a valid alternative to obtain a model is to identify the system. For that purpose, the DABNet (Decoupled A-B Net) structure was used that is composed of a decoupled linear dynamic system followed by a nonlinear static map. The linear dynamic,system was initially spanned by a set of discrete Laguerre systems and then cascaded with a single hidden layer Perceptron. A model reduction technique (linear balancing) used on the hidden nodes of the neural network as part of the identification process makes it possible not only to identify the main time constants, but reduce the dimensionality of the Perceptron input. It was proven that the DABNet structure can approximate every nonlinear, causal, discrete time invariant, multiple-input single-output system with fading memory. The final model consists of a linear state-space system whose states, decoupled by input, are mapped by a neural network. Results concerning the application of the methodology to the approximation of a CSTR and a polymer process are presented.