In this paper, we present two on-line adaptive control algorithms for non-linear plants using neural networks. The architecture used is based on the concept of specialized learning, which was first proposed by Psaltis et al. and suffers from two main problems, namely lack of knowledge of the plant Jacobian and slow training speed if the standard backpropagation algorithm is used. Specialized learning has been tested successfully by several researchers using the sign of the plant Jacobian, chosen on the basis of prior qualitative knowledge of the plant. It has also been proposed to calculate the plant Jacobian through a pretrained model. However, if off-line training of the model is not possible and qualitative knowledge of the plant Jacobian is not available, then specialised learning may not be feasible for an on-line neurocontroller. We propose that it is possible to estimate the plant Jacobian through the model on-line for a slowly varying plant. Some analysis for calculating the plant Jacobian on-line is discussed. For rapidly varying plant, we propose using the recursive least-squares (RLS) training algorithm instead of the standard back-propagation (BP) algorithm in the specialised learning architecture. We show that RLS training is faster than the BP algorithm. The proposed fast algorithm is then tested on simulated non-linear plants as well as in a real-time application to a coupled-tanks test rig.
1998 John Wiley & Sons, Ltd.
1. Introduction
The application of artificial neural networks (ANNs) in control has been an area of active research since the late 1980s, owing to the difficulty of controlling unknown non-linear systems by conventional adaptive control techniques. ANNs, which show several desirable properties for non-linear systems are used to overcome this difficulty. Currently, there are various ANN modelling and control architectures proposed in the literature. Among these structures are: forward and inverse plant modelling and control, direct inverse control, internal model control, feedback error learning and specialized learning, sometimes also known as indirect inverse modelling. These structures are briefly described in References 2 and 3.
ΒΉhis paper was recommended for publication by editor M. J. Grimble