| While
analytical model components
describe the part of the process
behavior that can be expressed with
equations from physics, Neural Networks
model effects that cannot efficiently
be modeled analytically.
For rolling
mills, such effects are:
 material behavior (e.g. modification
of heat capacity or flow stress due to
chemical analysis),
process specialties (e.g. adaptation
to heat transfer and friction
variations)
Thus, in general, Neural Networks adapt
the boundary conditions and material
parameters for the solution of the
model equations. They are the key to
fit the generally valid analytical
models to the needs of the automation
of a specific mill in order to achieve
tight product tolerances for all
materials and operation modes used at
that mill.
Siemens piloted the application of
Neural Networks in the early 90’s. In a
rolling force modeling application
Neural Networks increased modeling
accuracy by 30% as compared to
inheritance tables.
Since then, Siemens was able to gather
wide experience, exploiting the
advantages of Neural Networks for
process automation in over 100
applications at over 45 rolling mills
worldwide.
More than a decade of research and
development in this field has resulted
in specific neural network methods that
guarantee robust on-line adaptation to
process variations, precise modeling of
material behavior even for rare steel
grades, safe operation even if "extrapolating"
to new materials and reliable
implementations.
|
