A Hybrid Neural-Decoupling Pole Placement CONTROLLER AND ITS APPLICATION

DggA Hybrid Neural-Decoupling Pole Placement Controller And ITS APPLICATION Abstract A hybrid control architecture is proposed integrating recurrent dynamic neural networks into the pole placement context. The neural network topology involves a modi¯ed recurrent Elman network to capture the dynamics of the plant to be controlled, being the learning phase implemented on-line using a truncated backpropagation through time algorithm. At each time step the neural model, modelling a general non-linear state space system, is linearized to produce a discrete linear time varying state space model. Once the neural model is linearised some well-established standard linear control strategies can be applied. In this work the design of a decoupling pole placement controller is considered at each instant, which combined with the on-line learning of the network results in a self-tuning adaptive control scheme. Experimental results collected from a laboratory three tank system con¯rm the viability and e®ectiveness of the proposed methodology. Keywords: Hybrid methods, recurrent neural networks, pole placement, decoupling, multivariable adaptive control. 1 Introduction The evolution of automatic control in the last ten years has been characterised by a certain antagonism between two schools: the one based on the analytical-algebraic approach and the other based on information processing tools issued from arti¯cial intelligence. Both have worked to develop control systems for complex, non-linear, hardly modelled processes. The analytical-algebraic school, using rigorous methods of linear and non-linear systems, built a coherent body of knowledge but still fails to solve problems where it is not possible to obtain su±ciently precise processes and disturbances models. The other school, based on neural networks and fuzzy systems, developed a high number of methods and control architectures that solve e±ciently some di±cult problems, but the resulting body of knowledge lacks coherence, systematisation and generality.