Application of DOE To CAE
Application of DOE To CAE
Application of DOE To CAE
Engineering
1 Introduction
The recent advances in Digital Engineering have been contributing the reduc-
tion of lead-time of the product development in business process. Especially
in the product design and the production process design, the simulation ex-
periment by Computer-Aided Engineering (CAE) has been developing instead
of the actual experiment and it is surely one of the powerful supports to the
concurrent engineering. In the actual CAE working, however, some issues on
the simulation results appear. How to cope with them often depends on some
experimental or institutional skill of the CAE engineers. We cannot deny that
it can be an obstacle to progress of the reduction of lead-time.
A proper introduction of the Design of Experiment (DOE) to a series of
working in CAE is needed in order to improve the working efficiency in CAE. It
need scarcely be said that DOE has been applied to industry such as Taguchi
Method. But almost of them has been done to actual experiments. We can
get some controversial point issues in the case of application to simulation
experiments.
154 Ken Nishina and Mutsumi Yoshino
Fig. 2. Design parameters (punch depth (d), the punch width (b) and the shape of
the chamfer of the die (w, h)) and effect (height) (presented by N. Taomoto)
Fig. 3. Comparison of the results of the actual experiment and of the simulation
experiment (presented by N. Taomoto)
The parameters of the simulation model which may be used for the calibra-
tion are referred to as “calibration factor”. In general the materials property
index such as Young’s modulus, Poisson’s ratio, density and yielding point are
designed as the calibration factors, which are the variable factors. In addition,
whether or not to set the constraint conditions (translational constraint and
Application of DOE to Computer-Aided Engineering 157
rotational constraint) are designed as the calibration factor, also. These are
the attribute factors. In the example presented in Sec. 2, the calibration is
carried out by removing the constraint conditions of the friction between the
die and the material.
In the example in Sec. 2, the design parameters are the shape of the cham-
fer of the die and the punch depth. Then the punch depth was selected as
a factor in the actual experiment. Such a factor is referred to as “signal fac-
tor”, which should be selected out of the design parameters from the viewpoint
of being easy in carrying out the actual experiment. The number of the sig-
nal factor should be small because of decreasing the number of the actual
experiment.
At first, the calibration factors are selected. As mentioned earlier, the con-
straint conditions (translational constraint and rotational constraint) are de-
signed as the attribute calibration factors. These factors are assigned to the
two levels, that is, “free” and “constraint”. The materials property index such
as Young’s modulus, Poisson’s ratio, density and yielding point are designed
as the variable calibration factors. These factors are assigned to the three lev-
els. The theoretical value of the material property is assigned to the second
level. The first and third levels are assigned to the theoretical value ± about
10%, respectively.
As mentioned earlier, a signal factor is selected out of the design parame-
ters from the viewpoint of being easy in carrying out the actual experiment.
The signal factor is assigned to more than or equal to three levels. If the num-
ber of the level is three, then the level expected to be the optimal condition
should be assigned to the second level.
Fig. 4 shows an example of the design of experiment for the calibration
in order to identify some active calibration factors. As shown in Fig. 4, the
orthogonal array is useful (Fig. 4 shows L18 as an example). The calibration
factors are assigned to the inner factors and the signal factor (G) is assigned
to the outer factor.
The actual experiment is carried out only with the signal factor and at least
two repetitions. As mentioned in Sec. 2, in the actual experiment the other
design parameters are assigned as a constant level considering the facility of
the actual experiment.
If the number of the calibration factors is too large, then the experimental
design correspondent to the inner orthogonal array shown in Fig. 4 should
be divided into two orthogonal arrays. One is designed for the attribute cal-
ibration factors. The other is for the variable calibration factors. A two-level
orthogonal array is applied to the former case, and a mixed-level orthogonal
array is done to the latter case.
158 Ken Nishina and Mutsumi Yoshino
Let xjr be a result of the actual experiment at Gj of the signal factor and at
the rth repetition (see Fig. 4); and cij a numerical solution by the simulation
experiment at the ith experiment of the inner orthogonal array at Gj . Then
yij denotes the difference between the actual experiment and the simulation
experiment as follows:
R
xjr
r=1
yij = cij − x̄j· , x̄j· = (R: the number of the repetition).
R
The difference yij means “the nominal (zero)–the best” character, which
comes from the Taguchi method. So we propose applying the Taguchi’s two
step method, given for example by Taguchi (1993).
The first step is that the condition such that the variance (see Equa-
tion (3.1)) is minimized is determined. The analysis is referred to as the cali-
bration for stabilization. Then the analytic character should be as follows:
J
Si = log (yij − ȳi· ) ,
2
(3.1)
j=1
J
yij
j=1
where ȳi· = (J: the number of level of the signal factor).
J
The next step is that the bias from the nominal value (zero) is calibrated.
The analysis is referred to as the calibration for trueness. Then the analytic
Application of DOE to Computer-Aided Engineering 159
character should be ȳi· . The conditions such that ȳi· is equal to zero may be
countless; however, the priority of determining the calibrated condition should
be given to the determination at the first step.
Let c0j a numerical solution by the simulation experiment with the calibrated
condition at Gj . Then the ratio of the following two statistics, F , should be
given as the evaluation index for the calibration:
&
(xjr − c0j )2 (xjr − x̄j· )2
j,r j,r
F = . (3.2)
JR J(R − 1)
The numerator and the denominator of Equation (3.2) mean the luck of
calibration and the experimental error of the actual experiment, respectively.
The statistic F can be applied in the framework of the statistical test.
3.3 Discussion
5 Conclusive remarks
We have discussed the practical issues and given a proposal on the optimiza-
tion problem using the simulation experiments by the CAE. As mentioned ear-
lier, the content is an outcome of the WG1 of the research group “Simulation
and QC”. The WG1 is now in action and our mission is unchanging.
We are posing some remaining issues as follows:
1. Verification of the proposed method for the calibration through some case
studies
2. Calibration in the case of the multi-characteristic
Application of DOE to Computer-Aided Engineering 161
Acknowledgement. The authors would like to thank the members of the WG1 of the
research group “Simulation and SQC” of the Japanese Society for Quality Control
for their useful comments.
References
1. Taguchi, G. (1993). TAGUCHI on Robust Technology Development, ASME Press.
2. Yoshino, M. and Nishina, K. (2006) SQC and Digital Engineering —
Technological Trends in Design Parameter Optimization and Current Issues —,
Grammar of Technology Development, 129–145.
3. http://member.nifty.ne.jp/QYR03001/calc/calctop.htm
4. Yamada, S. (2004) Selection of active factors by stepwise regression in the data
analysis of supersaturated design, Quality Engineering, 16, 501–513.