Level 2 Model Improvement
Case Study: Oregon Steel
<Continue>
4.
Testing Results for the
First Improvement
In early March 2007, the newly designed
coefficients for the four most troubled model grades (for hard and thin
products), as listed in Table 4, were entered into the Level 2 system for
a trial. The model grade 05010531CN1 caused a redrafting (redistribution of
draft) due to poor force predictions, so for this grade the system was switched
back to use the original coefficients. In this very early stage, coefficients C1
and C2 had not yet been fully tested against the large amount of
historical data. This might have been the reason for the difference between the
new and old models behavior that led to the redrafting. The idea mentioned in
section 3, that C1 and C2 should also be carefully
designed and fully tested against historical data, came from this redrafting
case. The satisfactory results from the tests provided encouragement for the
completion of the remaining grades, with over 6000 sets of flow stress
coefficients being designed.
Table 4:
Model grades for testing
Model Grade |
Chemical composition (Aimed) |
Description |
05012506CN1 |
C 0.08, Mn 1.1, Si 0.25, Al 0.03 |
Thickness below 0.2" (5mm), coil, 1
stage |
05010531CN1 |
C 0.14, Mn 0.8, Si 0.25, Al 0.03 |
Thickness below 0.2" (5mm), coil, 1
stage |
05010002SN1 |
C 0.14, Mn 0.8, Si 0.25, Al 0.3 |
Thickness below 0.2" (5mm), steckel,
1 stage |
04010531CN1 |
C 0.14, Mn 0.8, Si 0.25, Al 0.03 |
Thickness between 0.2" (5mm) and
0.3?? (7.6mm), coil, 1 stage |
For grade 05010002SN1, 13 of the 15 passes
were within an error of 5%. The maximum errors with the old model were 12-14%.
With the new model, the maximum errors were about 10%, which were encountered in
the penultimate pass.
Fig. 2:
Slab NT2254A5 of Grade 05012506CN1
The predicted values for Grade 05012506CN1
from the old and the new models were compared with the measurements, as showed
in the Fig. 2. The old model led to an error over 20%. The new model
showed improvements, but still, with fairly high errors slightly over 10%. The
fact that the new model still had certain error in those two model grades could
be explained.
-
No X-Ray correction was observed that
would indicate an inaccurate gage-meter from the AGC.
-
It is believed that there was a phase
transformation in the last pass, so the much softer ferrite was generated. In
the measured data (Fig. 2), from the penultimate pass to the last pass,
the flow stress went lower with decreasing temperature. Further data processing
to draw the flow stress curves vs. temperature at a constant strain (0.3) and
constant strain rate (10/s) showed the same trend [8]. The logical explanation
is that a newer, softer material (ferrite) was generated through metallurgical
transformation (in austenite). In this case the flow stress model failed because
more than one materials (phases) were involved.
-
The range of the third temperature region
was too wide and thus there were too many passes in the region. Although passes
are weighted more heavily towards the last pass, this negatively affected the
learning regression.
If those two problems had not been
significant, a much higher accuracy would have resulted, as in the case of the
model grade 04010531CN1, showed in Fig. 3. With the new model, the errors
for all passes were near or below 5%. With the old model, however, there were
still 40% passes with errors beyond 5%. The better results obtained with this
grade over the previous example may also be attributed to the larger final
thickness of the product (grade 04010531CN1, "04" series, with its name starting
with "04") than the other two grades ("05" series). It is expected that for the
thick grades ("03", "02" and "01" series), most passes would enjoy a small force
error below 5%, while a small portion of passes with 5-10% error.
Fig. 3:
Grade 04010531CN1, Slab NT2291A4
<To
Be Continued>
<Previous>
|
<Next>
|
Part1
|
Part2
|
Part3
|
Part4
|
Part5
|
Part6
|
Part7
|
Part8
|
Part9
|
|