Neural networks to improve mathematical model adaptation in the flat steel cold rolling process

Abstract

In the flat steel cold rolling process, real-time controllers get their reference values (setpoints) using a mathematical model. Such a model is carried out at the process optimization level of the plant automation architecture. Since not all variables needed by the model can be effectively measured, and since a very accurate modeling would be unsuitable for real-time application or unachievable at all, the mathematical model must have adaptive capabilities, that is, its key parameters must be continuously adjusted based on real process values. This work proposes the application of Artificial Neural Networks to improve the adaptation of two hardly modeled process variables: the material yield stress and the friction coefficient between the work rolls and the strip. The text describes the theoretical foundations, the development methodology and the preliminary results achieved by implementing the proposed system in a real tandem cold mill.