Control architecture based on a radial basis function network. Application to a fluid level system

Nonlinear radial basis functions (RBF) at single layer hidden units of a neural net have proven to be effective in generating complex nonlinear mapping and at the same time facilitate fast learning. In the present paper off-line Gaussian control and identification of general nonlinear plants are realized. An iterative method to determine the desired controller output is described, and based upon this, a neural controller is adjusted by using the orthogonal least squares method. Neural identification of the plant is necessary to derive the parameter adjustments of the neural controller. The performance of the Gaussian approach has been demonstrated by off-line reference model neural control. Applications both to a general nonlinear plant and to a highly nonlinear fluid level system are detailed. It is shown that training times with orthogonal least squares method are dramatically reduced during control compared to standard backpropagation-of-the-error-through-the-plant technique. Finally, neural control is compared to PI control, showing the neural approach better generalization properties.