In many plants of the steel-, paper-, printing-, plastic- or textile industry the first section is an unwinder. But during storage and transport of the rolls, the rolls are often non-circular. If such a roll is unwinded, the radius is variable around the circumference. Therefore a changing web velocity is caused depending on the variable radius and big oscillations of the web forces are the consequence, if the winder speed is constant.

To solve the problem, it is necessary to use new methods to compensate the disturbances without additional mechanical systems. To do this, there are to solve two fundamental problems. The first is to get the unknown, nonlinear changes of the radius and the second is the design of a fast speed control of the winder which is able to compensate these changes.

Therefore a neural network is proposed to identify online the unknown, nonlinear variable radius with respect to the angle position continuously during running. The advantage of neural networks is that they do not need any information about the nonlinearities, they are able to handle unknown nonlinearities. So the neural network is like a "learning" observer.

To compensate the disturbances in the web force, the output of the neural network is fed into the speed control of the winder to change the motor speed of the winder in this way that the web velocity is nearly constant. This reduces the oscillations of the web force in a big range. But the requirement is a fast speed control of the winder. A conventional PI-controller is not able to fulfill that conditions. Therefore an adaptive state space control of the motor speed was designed.

The advantage of this method is, that no additional mechanical equipment is necessary. The identification of the unknown eccentricity of the roll and the compensation are done only in the digital control system.