Save
Money with Better Tech
Improvement on Level 2 Model Force Prediction
Level 2 force model is
one of the most
significant section of
the Level 2 system. See
Significance of Accurate
Level 2 Force Prediction.
Potential Problems You May
Check
(You may
refer to
Mill Level 2 Model Basics
for the terms and symbols
used here)
Many of
the following Level 2 model
problems seem to be ridiculous, but
before you laugh, check your
own system. You cannot
guarantee your system
doesn't have any of those
issues. At least, one of the
popular Level 2 systems in
the market has ALL those
problems!
-
Two of the frequently used
learning fits assume that
draft has no effect on
rolling force, torque and
power, so 10% draft and 40%
draft require the same
force, torque and energy.
This, of course, is wrong,
but check your log to see
how often your system were
using the FIT2 and FIT3B
(refer to
Mill Level 2 Model Basics
for the terms used).
-
Two of the same frequently
used learning fits assume
that rolling speed has no
effect on rolling force,
torque and power, so to roll
1m/s and 10m/s has the same
force, torque and power.
This, of course, is wrong,
too, but check your log to
see how often your system
were using the FIT2 and
FIT3A.
-
How could this happen?
Because, though the cases
mentioned here are bad,
other cases may be worse.
The Level 2 model usually
tests all the learning fits
and selects the one with the
lowest error. The learning
fit FIT4 doesn't have the
above-mentioned problem, but
it may have still lower
accuracy than FIT2, FIT3A
and FIT3B in certain
situation; that's why the
system sometimes picks the
other one (FIT2, FIT3A or
FIT3B) instead of FIT4. Our
research indicates that FIT4
does have weaknesses in
certain situation and we can
fix the problem.
-
Sometimes with a draft of
10%, for example, in a pass,
you should use e.g. 15% or
even 20% as the draft to
calculate force, torque and
power. Otherwise your
results will be inaccurate
even if all others of your
models are perfectly
correct. Interestingly, this
often happens in finishing
passes and may cause your
shape problems.
-
Level 2 model force
prediction formula may not
apply to a draft below 10%.
To make things worse, the
draft about 10% is often
used in the finishing pass
and this may lead to product
shape problem.
-
Some passes in your
two-piece rolling may, just
like one of our clients,
usually have a force error
of 10% to 30%, and often of
-20% to 60%.
-
Though you don't want to,
can you guarantee you are
not, performing your rolling
in the two-phase region? If
you roll unexpectedly in the
two-phase region, the force
error would be big
and your product shape could
be poor (especially if it is
thin).
Check
whether you have set a
learning parameter (such as
the coefficient for strain
or strain rate) to zero when
you don't use it for
learning. If you have, you
are making the same mistakes
as the problems No. 1 and 2 described
above. The error from the
problem No. 3
is introduced from the
dependence of the learning
parameters, which is a
common weakness of the
adaptive learning. The
problem No.
4 above is due to the
retained strain from
previous pass, and the
problem No. 5 is caused by
valid range problem of
certain flow stress formula.
The problems No. 6 and 7 indicate that
metallurgical issues often
cause significant error to
the Level 2 model.
Force
Model Improvement
Due to
the lack of understanding of the nature
for roll separating force,
especially in the areas such as flow stress and retained strain, many vendors
of Level 2 system overly simplified process and ignored the metallurgical
effects on roll separating force. Sometimes Level 2 administrators may also
have difficulty in applying metallurgical principles in operating the Level 2
system. The Level 2
model may be operated with inaccurate input data
(e.g. high-temperature
properties of specific heat,
E-Modulus, thermal expansion
factor, etc. in dependence
on the temperature), or with improper learning
(feedback) operation.
Consequently, the Level 2 system may be instable in the
force prediction accuracy, with very high error for certain products, such as
the hard and thin ones.
Flow Stress error
One of the toughest areas in the force prediction is the flow
stress. Flow stress is tightly related to the metallurgical process such as
recrystallization and correspondingly the retained strain. If a simple equation is
used to describe flow stress, errors may occur in large strain and small strain passes. More complicated problems exist in,
that in today?s rolling practice a great number of passes are rolled below the recrystallization temperature (technically it?s a cold rolling). Due to the
incomplete recrystallization, up to 80% (Tamura, Ouchi, et al.), or even more,
of the retained strain from previous pass can be added to the current pass.
In this aspect, we have
models to calculate the
retained strain, and we have
experience in integrating the
result into your existing
Level 2 without major change
of your Level 2 source code.
Many Level 2 system use
adaptive learning. However,
due to various issues, the
adaptive learning has
certain weaknesses and they
limit the accuracy of the
Level 2 force model. Through
our consulting work for the
steel mill Level 2 models,
we have summarized a number
of weaknesses of the
adaptive learning. We can
overcome those weaknesses,
by make a very slight system
change, usually only by
changing several lines of
the source code, or by
simply modifying your input
data without any source code
change, to greatly increase
your model accuracy. We have
developed a procedure called
the Guided Two-Parameter Learning
(FIT2G), a
special type of the adaptive
learning, to overcome the
shortcoming of the general
adaptive learning. To
achieve this, we have
developed over 6,000 sets of
the flow stress coefficients
as the start points of the
adaptive learning.
See
Modeling Issues in Level 2
to understand the problems
described above.
Simply letting Level 2 calculate flow stress coefficients is not
an optimal
way to perform learning. Certain guidelines may be
of
great help for better system learning. In this area
our experience on flow
stress modeling can be of great value
- the experience to generate over 2000
flow stress models (see
www.Meta4-0.com/flowstress,
or from North America,
www.flowstress.com).
If you
are using the neural
network, etc. to do the
system learning, we could
provide you with an expert
system to guide your
learning. If you by chance
have an expert system, we
could easily improve your
expert system. Those would,
too, significantly increase
your model accuracy. If you
want to make sure your
neural network never fails
(a black-box neural network
could fail at any time) and
would be continuously
improved, we could improve
your learning logic by
applying our hybrid solution
(a seamless combination of
the empirical model and
neural network).
Some
Level 2 model may divide the
rolling temperature range
into the high, medium and
low temperature regions. How
to optimize the dividing
points for the three
regions, for each
steel grade, and based on the
flow stress tendency and
metallurgical feature, is
also critical for the high
quality model
prediction. During our
consulting work we have
solved the problem and are
ready to apply the results
to you. You don't have to
reinvent the wheel to do
your own research.
Other parameter errors
Depending on the design of your Level 2 system, following areas may also be the
potential sources of error:
-
Roll flattening calculation. Roll flattening should be calculated. Some systems
may overly simplify the calculation process; others may even ignore the roll
flattening for hot rolling.
-
Shape factor. Shape factor covers all the contributions to the roll separating
force beyond the mean flow stress and the contact area. It is affected by
both roll gap geometry and friction. We have studied over a dozen of different formulas for
shape factor, and are capable of improving it in case your Level 2 model has
problem in this aspect.
-
Temperature during rolling, interpass cooling and controlled cooling.
Many Level 2 systems only use the primary models,
which consider a
heat transfer
coefficient as a
constant, to perform
temperature learning.
We can apply our
heat transfer models
to improve your
temperature learning
logics, in which a
heat transfer
coefficient is
described as a
function of various
rolling/cooling
parameters. This
would greatly improve
your
temperature
prediction accuracy.
See
Temperature Model
Improvement.
What we can do for you
We have made a great
progress on the rolling process modeling
since 20 years throughout
our work
in various countries (Germany, USA, China, etc.). During the work we
have identified a long
list of potential sources of force errors and have established the
solution in each aspect.
We can help you, by making a minimal change of your
level 2 system, or by simply modifying your Level 2 input data, to
achieve much
higher model accuracy. For reversing/steckle mill, your Level 2
is expected to have an error of
about 5% for force prediction; for continuous
strip rolling, the accuracy should be still higher
because the rolling
condition is more consistent
than in reversing/steckle
mill.
After the force improvement the system accuracy should be stable for all
products. Our current
result for a reversing/steckle
plate mill, with test data
for previously
troubled grades (hard and
thin ones with shape
problems), is averagely 3.4%
(average of the absolute
force errors)! See the
paper we published, some
jointly with the client.
For this purpose we may work in following areas:
-
To examine your Level 2 system to check potential design problems.
-
To analyze your history
data by accessing your
log files or database to
identify weaknesses. We have a list of programs used to
analyze your data.
-
To establish stable,
optimized learning for your Level
2 model. If you analyze
your log files/tables
for your flow stress
coefficients, you would
be surprised by their
large scattering range.
Certain guidelines
should be applied to
achieve proper
learning.
-
To help you with other model problems besides the force prediction. Dozens of
rolling process models we have created in the past
20 years can be used.
-
If you plan to upgrade your current Level 2 system or design a new
one, one or more of our consultants can join your team. We have excellent steel
mill modeling and software engineering expertise (10 years of software
development experience
in steel mills).