Adaptive neural network structures for non-linear process estimation and control

While modern feedback controllers are based on linear models, many processe s of concern to chemical engineers are non-linear. The models for these pro cesses are often too complex for non-linear controller design, and their li nearized counterparts may not adequately represent the process dynamics. Ad aptive radial basis function neural networks have been implemented in proce ss estimators and controllers to approximate non-linear functions that desc ribe the process. Existing algorithms have been proposed which require spec ification of network properties such as dilation of the basis functions. Th is work demonstrates that the function representation may be improved by se lection of an optimal dilation, and presents an algorithm that allows simul taneous adaptation of dilation and node weights. In order to deal with func tions that may have more than one dominant dilation, a multiresolution netw ork adaptation algorithm is proposed. Lyapunov stability is proven for both strategies, and performance is evaluated for the control of an exothermic CSTR. (C) 1999 Elsevier Science Ltd: All rights reserved.