Level 2 Model Improvement
Case Study: Oregon Steel
<Continue>
3. First Improvement
Improvement of the existing Level 2 force
model was primarily based on the following principles:
The logical learning error must be
corrected, so the average values C3m and C4m should be
used instead of zeros in the FIT2, FIT3A and FIT3B;
Limitation of the adaptive learning in
FIT4 should be removed, especially the use of the widely scattered values caused
by the dependence of C3-C4;
Metallurgical principles should be
considered; so the effect of the retained strain should be reflected;
There should not be significant change to
the existing source code.
Data analysis showed earlier indicated
that the metallurgical process had great effects on the flow stress coefficients
and was one of the primary reasons to cause C3 to scatter in a wide
range and to be larger than the theoretical value. Even for the FIT4, due to the
use of the widely scattered data caused by the dependence of C3 and C4,
the learning quality is still limited. In the log files there were many
occasions where the FIT4 had poorer quality than other fits (FIT2, FIT3A or
FIT3B), even though those fits were defective with the use of 0 instead of a
medium value of C3 or/and C4.
There are various solutions possible. In
developing a new Level 2 system, adding logic to improve C3-C4
dependence issue for FIT4, improving logic for FIT3A and FIT3B, and predicting
retained strain, etc. would be the right solution. However, for existing system,
too many modifications would introduce potential bugs and would be time
consuming. For this reason, a simple but effective solution was accepted.
In this solution, carefully designed
average values for C3 and C4, covering all the
above-mentioned issues including the medium values and the dependence of C3
and C4 as well as the retained strain effects, etc., were generated.
Because the retained strain is affected by material composition, rolling
temperature and inter-pass time, the coefficients C3 and C4
were specifically designed for each model grades in every temperature regions.
As a result of implementing these coefficients in the learning process, the
designed C3 and C4 remained constant and only C1
and C2 were adapted. Manipulating the retained strain into the flow
stress coefficients meant that there was no need to change the flow stress
formula and therefore, the change to the source code would be minimal. This
learning procedure can be referred to as the Guided Two-Parameter Learning
(FIT2G).
The concept to manipulate the retained
strain into the flow stress coefficient C3 was also observed in many
publications such as [7] mentioned above. Those authors usually argued that the
contributions from the strain and the strain rate were much less than those from
the material and the temperature. Even though, using a single pair of C3
and C4 for all grades still seems too rough. In the EOS Level 2
models project, over 6000 sets of C3 and C4 were actually
used! In the current production practice, EOS has over 2000 model grades. As
mentioned earlier in the Introduction section, a model grade is created based on
the steel grade, product and production practice, etc. For each model grade,
there are three sets of coefficients, for the three temperature regions.
Designing over 6000 sets of C3
and C4 needs a good understanding of the flow stress model. In this
specific case, the contribution of the metallurgical parameters to the C3
and C4 added much more difficulty to the work. In the first step, a
basic range of the C3 and C4 was created as a guideline
for the coefficient design. The basic range of the C3 and C4
was determined based on various considerations such as
the initial values of the coefficients C3
and C4 in typical hot rolling [5],
the pattern of the initial values (C3
and C4) in different temperature regions, and
the retained strains in different
temperature regions.
The pattern of the initial values was
established through studies of published flow stress models in various sources
[5].
In EOS case, if the inter-pass time is
30-40 seconds, the percentage of the strain retained from a pass to the next is
estimated in Table 3. For further details, different steel grade, etc.
may have different ratio of the retained strain. At this plate mill, the
inter-pass time may be longer than some strip mills, so the ratio of the
retained strain here should be slightly lower than that in many strip mills with
the similar rolling temperatures.
Table 3:
Ratio of Retained Strain in various Temperatures
T (Deg C) |
1000 |
900 |
850 |
800 |
750 |
T (Deg F) |
1830 |
1650 |
1560 |
1470 |
1380 |
Ratio of Retained Strain |
0% |
15% |
21% |
33% |
42% |
Application of the newly designed C3
and C4 sets in the production system also involved another issue. If
C1 and C2 corresponding to the new C3 and C4
were not carefully handled, there would be problem for the first piece (at least
the first pass) after the new coefficients are introduced into the Level 2
system. Therefore, C1 and C2 were also carefully designed
and the entire set of coefficients C1, C2, C3
and C4 were tested against the 1.5 million records of the history
data. This would guarantee the smooth transition from the old coefficients to
the new ones.
<To
Be Continued>
<Previous>
|
<Next>
|
Part1
|
Part2
|
Part3
|
Part4
|
Part5
|
Part6
|
Part7
|
Part8
|
Part9
|
|