Degree Name

Doctor of Philosophy


Department of Mechanical Engineering


Strip thickness, a key quality variable in rolling process, must be controlled to within a narrow tolerance to meet the increasingly commercial demand for more accurate and consistent dimensional requirements. Therefore, a great variety of automatic gauge control systems have been introduced to rolling process since the mid-fifties, among which, the gaugemeter control plays an essential and critical role, since all other gauge controls are normally applied in conjunction with the gaugemeter control to obtain accurate quality.

The gaugemeter control has been successfully applied in industry to satisfy the thickness requirements to a certain extent, particularly when it is implemented in combination with other controls. However, the competitive markets request so ever-increasingly stringent thickness quality that the conventional gaugemeter control can not catch up. This is due to the fact that the conventional gaugemeter control is based on linear models whereas rolling is so complex that it can not be described as a linear process with a satisfied accuracy.

This study proposes an intelligent control by using neural networks and fuzzy systems to improve the thickness control performance. Back-Propagation (BP) neural networks are investigated and applied to model the non-linear relationship between roll gap, rolling force and exit thickness, which replaces the linear gaugemeter equation employed in the gaugemeter control. Meanwhile, a newly adaptive fuzzy controller is developed to produce control signals. This fuzzy controller dynamically approximates the relationship between the exit thickness error and the required roll gap adjustment control signal by adapting itself based on the evaluation of its previous control performance, rather than requiring accurate values of mill elastic modulus and material plastic modulus.

A computer simulation based on a dynamic mathematical model of rolling process verifies the feasibility of this neurofiizzy control scheme. The simulation also confirms that an application of neural networks and fuzzy systems to a rolling process is feasible in extended ways by investigating an integral neural control and an integral fuzzy control.

The neurofiizzy control is developed on an experimental rolling mill to embody a real-time application, which is fully implemented in C/C++ Language. To provide an intuitive input and output interface for this system, an interactive graphical user interface is developed.

Experiments are conducted to test the neurofiizzy control and to compare its performance with the gaugemeter control. Experiments with the gaugemeter control confirm that the control performance is improved as the compensation coefficient C increases, but over-compensation will occur when C = 1.0, which might result in an unstable control. Experiments with the neurofiizzy control indicate that this control scheme improves strip quality significantly in terms of exit thickness standard deviation compared with the gaugemeter control. From the experimental results with different values to adaptation coefficient for the adaptive fuzzy controller, the following conclusions are obtained.

• The neural model possesses the non-linear capability to estimate exit thickness more accurately.

• The adaptive fuzzy controller successfully determines the relationship between the exit thickness error and required roll gap adjustment by adapting itself to produce an accurate control signal.

The above two mechanisms are beneficial to the improvement of thickness control. Therefore, combining them, the neurofuzzy control is an improved control scheme which is demonstrated both by theory and experiments.