On multivariable adaptive control of a class of nonlinear systems using neural networks

Multilayer neural networks are used in a nonlinear adaptive control problem. The plant under consideration is an unknown feedback-linearizable continuous-time system, represented by an input-output multivariable model. A linearizing feedback control is derived in terms of some unknown nonlinear functions. To model these nonlinear functions and generate the feedback control there are used layered neural networks. Based on the errors between the plant outputs and the model outputs, the weights of the neural network are updated on-line according to a gradient learning rule. A local convergence result is provided which says that if the initial parameters errors are small enough, then the tracking errors will converge to a bounded area. Computer simulations are included to demonstrate the performance of the controller for a continuous fermentation process. Copyright (C) 1998 IFAC.