Level 2 Model Improvement
Case
Study:
Oregon Steel
Improvements were made to Evraz Oregon Steel
(EOS) steckel mill Level 2 force model. Learning logic and metallurgical effects were identified as the primary sources of errors. Limitation of the adaptive learning was discussed. Concept of guided two-parameter learning was proposed to resolve the issues and over 6000 sets of the flow stress coefficients were designed. In addition, solutions for the problems with resuming passes and the passes with large or small strain were also proposed. Even with troubled grades, the trials still indicated a high accuracy with an average absolute error of 3.4%. The solutions were designed so it would require the least amount of changes to the existing system.
Following sections are
included in this
writing:
-
Introduction
-
Data Collection and Data Analysis
-
First Improvement
-
Testing Results for the First Improvement
-
Second Improvement
-
Summary/References
1.
Introduction
EOS Level 2 Models and Metallurgical Requirements
Evraz Oregon Steel Mills (EOS) is the largest steel
processing company in the Western United States and produces over 1.8
million tons of plate, coiled plate, welded and seamless pipe,
structural tubing, rail and wire rod/bar. Its facilities are
in three locations: Portland, Oregon; Pueblo, Colorado; and Camrose, Alberta, Canada. EOS plate mill has
an equipment
investment of about 800 million dollars and annual sales of about 800
million dollars, too.
The improvement work for the EOS Level 2
models was initially to fix some logical problems with the learning function. In
designing flow stress coefficients to establish a more robust learning process,
the development was faced with very strong challenges from metallurgical issues
involved in the low-temperature rolling. Therefore, the work to resolve metallurgical
issues was incorporated into the learning and modeling solutions. This writing
provides some keynote summaries for the recent Level 2 model Improvements [1].
EOS plate mill strongly
relies on rich functions of Level 2 management, high accuracy of Level 2
models, and great potentials of the metallurgical process in improving rolled
steel properties. Its automation systems have assumed most
functions such as production scheduling (by Level 3), draft schedule generation
(by Level 2 models) and production execution (by Level 2 management system). For
good product property, EOS pushes very aggressively in pursuing good product
properties through controlled rolling/cooling and use of microalloys. This,
however, has placed strong challenges to the Level 2 models, which were initially
designed based on mechanical principles without metallurgical consideration.
Accurate force prediction is one of the
most critical functions of the Level 2 models. Fewer force errors directly leads
to better product quality and higher mill utilization. For example, an overly
conservative force prediction error of 10% results in lower mill utilization to
avoid potential equipment damage. Many product defects can be attributed to the
Level 2 model errors [2]. For this reason, the model improvements were focused
on the force prediction.
EOS Level 2 Force Model
EOS Level 2 system was initially supplied
by Tippins in 1997, but improved significantly by EOS. One of the primary
parameters the Level 2 models take into account in generating draft schedules is
the roll separating force. Roll separating force is usually modeled as the
product of mean flow stress, projective contact area and shape factor (Q-factor)
[4]. Among various factors affecting the roll separating force, the flow stress
is the one that bears the effect of material, strain, strain rate and
temperature.
In the EOS Level 2, the following formula
is used for the flow stress modeling:
(1)
The four parameters C1, C2,
C3 and C4 represent the coefficients of material,
temperature (T in K), strain (e ) and strain rate (u in /s), respectively.
To increase accuracy of the rolling force
predictions across various products, the flow stress model maintains separate
sets of flow stress coefficients for each model grade. A model grade is created
based on the steel grade (chemical composition), product (type and dimension)
and production practice (e.g., regular or controlled rolling). For each model
grade, there are three sets of coefficients that are automatically adjusted by
the long-term learning function to cover the three ranges (either thickness or
temperature) expected during rolling.
The EOS Level 2 uses adaptive learning.
The learning includes the short-term or pass-to-pass learning to shift the
values upwards or downwards based on the error in the previous pass, and the
long-term learning to recalculate and adjust all the coefficients (such as the
flow stress coefficients and the heat transfer coefficient) after a qualified
piece is rolled. The long term learning of the EOS Level 2 initially used four
fitting mechanisms: the Two-Parameter Learning FIT2 (using C1 and C2
only), the Three-Parameter Learning FIT3A (using C1, C2
and C3) and FIT3B (using C1, C2, and C4)
and the Four-Parameter Learning FIT4 (with all four coefficients).
Project Outline
In the second half of 2006, the Level 2
models frequently encountered product shape defects. Analysis indicated that the
Level 2 models had significant force errors, sometimes up to 40%, in the related
passes. Because the draft schedules generated by the Level 2 models were
primarily based on predicted forces, unreasonable draft schedules were created
which led to bad finish shape. Further analysis also identified the issue of the
fluctuation of the flow stress coefficients (C3 and C4)
within a large range. This fluctuation was considered to be due to a logical
design problem. For example, if the coefficient C3 is not used for
learning, the existing system set C3 to zero. This ignored the effect
of the strain (and thus the draft) to the force, and relocated the strain
contribution to the strain rate coefficient C4, etc. and caused the C4
to fluctuate. Similarly, if the FIT3A is used, the existing system set C4=0;
this pushed all the effect of the strain rate (and thus the rolling speed) to C3
and made C3 scatter in a large range. A solution was proposed to
replace the zeros with average values (C3m or C4m), if a
coefficient is not used for learning.
The improvement of the EOS force model
started in November 2006. The early stage of the work was filled with confusion
because the data were highly scattered (e.g. C1, C2, C3
and C4) and usually much higher than the theoretical values (e.g. C3),
even for the Four-Parameter learning. Later, the limitation of the adaptive
learning was identified and the issues of the metallurgical effects in the Level
2 models were disclosed. The solution was proposed and approved to use
well-designed fixed values for C3 and C4 to solve the
learning problems and to include metallurgical effects in the coefficients
(especially C3). The learning would then be conducted only with C1
and C2 (so-called Guided Two-Parameter Learning, or FIT2G). In March
2007, the newly designed coefficients for several most troubled grades were
incorporated into the Level 2 system for testing. With the encouraging results,
the management approved the use of all the designed coefficients in the Level 2
models. Further improvement was then conducted with newly identified issues such
as defining the valid ranges for the formulas in each of the three temperature
regions, considerations for controlled rolling hold temperatures and the
possibility of entry into the two-phase region.
<To
Be Continued>
<Previous>
|
<Next>
|
Part1
|
Part2
|
Part3
|
Part4
|
Part5
|
Part6
|
Part7
|
Part8
|
Part9
|
|