Optimization and New Product Development
Metal Pass LLC
www.meta4-0.com (Industry 4.0 Metaverse)
3. A Simple and Efficient Way to Integrate New Models into Level 2
(1) General Learning
We have developed a simple, easy but effective way
to integrate large number of new models into Level 2 System, thus to upgrade
conventional Level 2 System into the New-Generation one! In this way, only a
very few source code modifications is required.
Usually machine learning can be carried out in following ways:
(1) . With chemical composition and temperature
(2) . (1) plus strain
(3) . (1) plus strain rate
(4) . (1) plus strain and strain rate
In real operation usually all 4 cases are considered, and the best one was
selected. Of those selected, there occur often the bigger strain factor plus
small strain rate factor, or the small strain factor plus bigger strain rate
factor. Therefore, both the strain factor and strain rate factor can be any
value from large to small. There would be other problems. Those problems reduced
the accuracy of the model prediction.
(2) Guided Two-Parameter Learning
To solve this problem, we classify the all possible
chemical composition into
material grades/Steel grades (most plants have about 100-200 grades, average
as 150), Thickness (considered as 5 cases), Product types (4 cases), Rolling
stages (3 cases), Slab thickness (3 cases), and 3 temperature zones. Under this
classification, there are 54000-108000 cases (average 81000 cases. Each case has
a data file containing over 100 data and for a production process. Material data
for E modulus and specific heat, etc. are pretty strong temperature dependent;
in particular, flow stress coefficients as material factor, temperature factor,
etc. can be determined through offline design, wile strain factor and strain
rate factor can be determined through machine learning. These learning
procedure is called guided Two-Parameter Learning. So-called guided is because of
the offline design into above mentioned cases, and Two-Parameter is referred to
strain factor and strain rate factor.
In some situation there may not be all cases. For example, for thin product, the
high thickness of the slab may not be used.
In each production only one file is used for learning, and the learnt factor
(strain and strain rate) is updated in the data file (only small portion of the
data is changed). So in the source code, only the source code regarding the
learning need modification; others, such as use cases/user scenarios, do not need
modification. Therefore, source code modification is very limited.
Due to large number of data files, and due to complicated data process, software
is developed to do the offline design. In the design software, metallurgic
models and manufacturing knowledge is integrated, such as those related to
equipment, process, entry material, automation, etc.
This type of software is beyond the reach of pure software engineers. It would
also be hard for non-software engineer to integrate such data and learning in
the large software such as Level 2 system.
<To Be Continued>