Compensation of Periodic Disturbances in Continuous Processing Plants by Means of a Neural Controller

Abstract

In this paper, a new approach for the compensation of unknown periodic disturbances by means of a neural network is presented. The neural controller supports the conventional controller by suppressing periodic disturbances. This is done by online learning in order to adapt to different operating conditions and to time varying unknown disturbances. The neural network learns an optimal compensation signal, such that the effect of the disturbance becomes zero in the considered output signal. With this method, there is no need to redesign existing control loops. Exemplified by the compensation of eccentricities of the unwinder of a continuous processing plant, the neural controller is explained and simulation results are shown. An extension to the basic method is to consider an additional input dimension in the neural network, which represents the current operating point. The information about the optimal compensation signal of a specific operating point is stored in the network weights of a multidimensional Radial Basis Function Network. For pre-trained operation ranges, this guarantees an optimal compensation result even if the operating point changes. The main benefit of the presented method in industrial applications is the capability to augment the production speed and to improve the product quality, by reducing tension force oscillations caused by eccentricities of rollers or unwinders.