Transcript R (LIB) - Ee Oulu Fi

Product design model for
Impact Toughness
Estimation in Steel Plate
Manufacturing
Satu Tamminen
ISG - Data Mining Group,
Department of Electrical and
Information Engineering,
University of Oulu, Finland
TOC
Introduction
 Data
 Model
 Conclusions

IJCNN 2008
Introduction



Impact toughness (notch toughness) describes
how well does the steel resist fracturing at
predefined temperature, when hard impact
suddenly hits the object.
The property is crucial for steel products that are
used in cold and harsh environments e.g. ships,
derricks and bridges.
The harder the steel the lower the impact
toughness.
IJCNN 2008





In a room temperature the steel can perform well in the
impact toughness test, but when the temperature falls,
the performance weakens.
In this research, the most demanding steel qualities are
tested as low as in -100˚C.
Transition behavior (ductile-to-brittle transition in certain
temperature) is typical for ferritic steel qualities.
Factors that rise transition temperature have a negative
effect on impact toughness.
The complicated interactions between these factors
bring challenge to the modelling (a harmful elements can
produce a desirable effect together with another
component).
IJCNN 2008


Impact toughness is defined by Charpy-V impact
toughness test (CVT).
The test piece is broken with a pendulum and
the energy absorbed in fracturing is measured.
IJCNN 2008


The test is performed for three different samples from
every steel plate, and it will be accepted if the average
of the measurements is higher than the requirement. In
addition, only one of the measurements is allowed to
be not more than 30% under the requirement.
The average of the measurements does not serve well
as the target of the model. Instead, the target is
1 3 1 
LIB   log 10   2 
 3 i 1 mi 
IJCNN 2008
100 J
The difference between the average of the
measurements (left) and the lib transformation
(right).
IJCNN 2008
Data




The data was collected at Ruukki's steel plate mill in
Raahe, Finland during 2002-2007 and it consists of
information about over 200 000 low-alloy steel plates
and over 70 variables.
After careful pre-processing, the final data included
202 667 observations and 42 variables.
The variables included information about the shape and
position of the test bar, the test temperature, chemical
composition and process parameters.
63% of plates rejected in CVT were rejected because of
one too-low measurement.
IJCNN 2008
Model





The model is developed for product design group, who
plans the chemical composition, possible treatments
during melting and some production requirements for
heating, working and thermomechanical treatments.
The model predicts the rejection probability in CVT.
The model will guide designers in producing desired
properties in the product at lower cost.
The working allowance that keeps the product within
tolerance can be decreased with the model.
Ruukki competes with high quality, short delivery time,
and a large product range.
IJCNN 2008






Target values for chemical composition and for the rest
of the process variables were used.
MLP (multilayer perceptron) networks were trained with
MATLAB R2007a.
Half of the data was used for training and one quarter for
validating and one quarter for model selection.
The independence of the data sets was verified by not
allowing plates from the same melting to belong to
different sets.
The network for LIB-model had two hidden layers with 39
and 5 neurons.
For comparison a network for mean of the three
measurements was trained as well (AVG-model).
IJCNN 2008
Over 96%
of plates
Scatter plot between the predicted LIB and the model error.
IJCNN 2008
The performance of the LIB-model and AVG-model.
MSE (LIB)
R (LIB)
R (AVG)
Train
0.0487
0.8747
0.9056
Test
0.0502
0.8676
0.9016
The rejection probability can be calculated with cumulative
Normal distribution function
 L  ˆ i 
Pi  

 ˆ 
where L is the rejection limit, ̂ i is the estimated LIB
and ̂ is the calculated sample deviation.
IJCNN 2008
ROC-analysis of the results show that the LIB transformation discriminate
better the rejected plates (LIB-model solid, AVG-model dashed).
Rejection level LIB model
AVG model
20J
0.9860
0.9586
27J
0.9850
0.9644
40J
0.9452
0.9192
100J
0.9110
0.9049
120J
0.9125
0.9104
Area under ROC curve
IJCNN 2008
Conclusions




Analysis showed that most of the rejections could have
been recognized with the model, and thus, it is expected
that the number of rejections will be reduced when the
model is entered into the product design.
The LIB-transformation improves the prediction result.
At the moment, the model is in test use at the product
design department in Rautaruukki, Raahe.
The assumption of independence between model error
and the variables is not valid, and the results can be
improved further with the use of a variance model
(heteroscedastic regression).
IJCNN 2008
[email protected]
[email protected]
http://www.ee.oulu.fi/research/isg/
IJCNN 2008