An Adaptive Control Algorithm for Multiple-Input Multiple-Output Systems Using Neural Networks

Many industrial processes have multiple inputs and outputs. Such systems are known by the acronym MIMO. In order to properly control such processes some linear techniques like multivariable control, robust control, etc., have been used [1]-[3]. These techniques require a known model of the controlled process. However, many industrial processes are poorly known and nonlinear. Therefore, a technique to model and control nonlinear and partly known systems is needed. It is known that artificial neural networks, particularly Multi-Layer Perceptrons, partly fulfill such requirements. In this paper the gradient descent optimization rule is combined with a trained neural network model for the computation of the control vector [5],[7]. The computed control vector would drive the nonlinear MIMO system outputs to the desired operating point. For this purpose, a quadratic cost function is defined. The computation of a new control vector requires the previous computation of the gradient of the process. Since direct information of the gradient of the process is not available, an accurate approximation of the systems gradient is computed from the neural network model. The general performance of the controller is illustrated using simulations. A simple mathematical model of coupled tanks systems is used for this purpose [10]. Key-Words: Neural networks, control, stability, modeling, optimization.