n  the work  described  in  this  article, controlling  the  strip
I  thickness of a steel rolling mill provides a real-world applica-
tion  of  nonlinear neural control. Different control structures
based on neural models of the simulated plant are proposed. The
results of  the neural controllers, among  them  internal model
control  and model  predictive control, are compared with the
performance of  a conventional PI controller. Exploiting the ad-
vantage  of  the nonlinear modeling technique,  all  neural ap-
proaches  increase the control precision.  In this  case,  the
combinalion of  a neural model as a feedforward controller with
a feedback controller of  integral type gives the best results.
Artificial Neural Networks 6778
Artificial neural networks have been the focus of a great deal
of attention during the  last decade, due to their capabilities in
solving nonlinear problems by learning. Such networks provide
a paralle. structure with very simple processing elements. This
paradigm has been applied to solve different technological prob-
lems, such as signal classification and, more recently, nonlinear
and adaptive control problems [  1  1.
Controlling the strip thickness of  a cold rolling mill is an
application  of the Daimler-Benz company AEG. The steel rolling
mill has :several  characteristics,  such as varying time delays and
varying  gains, which make it  appropriate to  be  tackled with
neural approaches. This is because neural networks offer some
distinct advantages over conventional approaches, providing a
general framework for modeling and control of  nonlinear sys-
tems. Some work on using neural nets for modeling  a  rolling mill
has alreajy been done (see [2] and [3]).  In particular, both papers
describe the use of  neural network models  to  predict process
parameters for a certain rolling task. By  incorporating a priori
knowledge about  the rolling process into  the training procedure
the authors of  [2] achieved a better generalization  in regions in
which training data was insufficient. The predicted values were
used  to  !.et up the controller parameters of  a conventional con-
trolling xheme, which is a type of  supervisory control. On  the
other hand, the use of  neural models for direct control has no
been  investigated. Here, we focus on  the generation of  neura
models  of  the  plant  and  their online use  in  different contro
schemes. such as model predictive control. Neural net based structures can be  applied even if no exact
mathematical representation  of the process exists; that is, models
can be derived from observations  of the input-output behavior  of
the plant. This leads to a significantly reduced controller design
effort in comparison with design techniques needed for setting
up  the parameters of  conventional controllers. Neural network
based controllers generally depend less on apriori knowledge of
the process to be controlled.
The next section of  this article defines the control problem
and briefly introduces the model used for simulation  of the plant.
The third section describes  the traditional approach  to solve  this
control problem  by using a  linear controller. The fourth  part deals
with modeling aspects using neural networks, and section five
compares control solutions using these neural network models.
Finally, conclusions are drawn including possible directions of 轧钢机英文文献和中文翻译:http://www.youerw.com/fanyi/lunwen_4452.html