To cite this article: Dongcheol Kim & Sehun Rhee (2003) Optimization of GMA welding process
using the dual response approach , International Journal of Production Research, 41:18,
4505-4515
To link to this article: http://dx.doi.org/10.1080/0020754031000/595800
1. Introduction
Establishing a welding process model in reference to weld quality is important as
the model identies the eect of the welding process parameters on weld quality and
also determines which welding process parameters will generate high quality welds.
Arc welding processes, however, are generally nonlinear and complicated, and thus
make it very dicult to obtain a welding process model based on physical principles.
In order to overcome such diculties, experimental data are used to derive the
welding process model, and a method of optimizing the welding process is suggested
based on the welding process model and the optimization algorithm (Tarng et al.
1999, Kim and Rhee 2002).
To date, welding process optimization has been carried out on the assumption
that the induced model is accurate and there is no variance or disturbance in working
environment. This has led to various cases where the desired weld quality is not
generated because of variances or disturbances actually extant in the welding environment. Several studies have tried to overcome this problem by adopting the
Taguchi method to determine specic welding conditions that would generate
good quality welds without having disturbance as a major inuence factor
(McConnell and McPherson 1997, Tarng and Yang 1998). The Taguchi method
takes into consideration the variance as well as the mean value of the output variable
4506
D. Kim et al.
to determine the optimal input variables, and this method is applied to various
processes (Phadke 1989). The interaction between the variables, however, is not
adequately considered in the Taguchi method, and it eliminates the possibility of
conducting sequential experiments. Therefore, a new method combining the Taguchi
method with the response surface method is suggested to overcome such deciencies.
Vining and Myers (1990) adopted the dual response approach developed by
Myers and Carter (1973) to achieve the goal set out by the philosophy of the
Taguchi method. Vining and Myers employed Lagrangian multipliers to solve the
dual response optimization problems, while Dell Castillo and Montgomery (1993)
and Lin and Tu (1995) used nonlinear programming algorithms.
In this study, the dual response approach is adopted to optimize the output
variables aecting the weld quality in the arc welding process as follows. First, the
experimentation is conducted in the region of interest according to the experimental
design. Second, the regression analysis is used to obtain the models of the standard
deviation as well as the mean value of the output variable based on the results of the
experiment. Last, the optimization algorithm is used to determine the optimal
parameters of the welding process. The algorithms used in previous studies tend
to converge on the local optimal point according to the initial point of the search.
To solve this problem, a genetic algorithm is adopted in this study wherein the
objective function is not required to be dierential and has a high tendency to
converge on the global optimal point.
2.
4507
A regression model of the mean value, in general, is induced in the response surface
method, but a regression model of the standard deviation is also required in the dual
response approach to reect the Taguchi theory. When ^ and ^ are the regression
models of the mean and standard deviation respectively, the three basic problems
examined in the Taguchi method are as follows (Vining and Myers 1990).
Case 1. Target value is best, in which the goal is to minimize the variance while
keeping the mean at a specied target value 0 . Mathematically this can be formulated as
minimize ^
subject to ^ 0 .
Case 2. The larger, the better, in which the goal is to maximize the mean while
holding the variance at a xed value 0 . Mathematically this can be formulated as
maximize ^
subject to ^ 0 .
Case 3. The smaller, the better, in which the goal is to minimize the mean while
holding the variance at a xed value 0 . Mathematically this can be formulated as
minimize ^
subject to ^ 0 .
In order to solve the optimization problems above, Vining and Myers (1990) used
Lagrangian multipliers, while Dell Castillo and Montgomery (1993) showed that the
dual response surface optimization could be solved using nonlinear programming
techniques, such as a generalized reduced gradient (GRG) algorithm. There are,
however, two deciencies in the optimization method suggested by Vining and
Myers. First, because the optimization method Vining and Myers proposed is
based on the quadratic form of the model, a full second-order model is required
regardless of the signicance levels of the model ttings. Second, because ^ and ^
only reect the approximate value of the actual response, either ^ or ^ is forced to a
xed value in the optimization problem, and the preferred values may be excluded.
Lin and Tu (1995) proposed the following optimization method using the mean
square error (MSE) criterion to overcome the deciencies of the Vining and Myers
method.
minimize MSE ^ 0 2 ^ 2 .
The MSE method, nevertheless, does not take into consideration the acceptable
deviation of the target value, and thus the optimal solution, which minimizes the
MSE, may exceed the acceptable limits of the target value. In this study, an optimization problem considering the acceptable deviation of the target value is suggested
in order to overcome such deciencies.
minimize MSE ^ 0 2 ^ 2
such that j^ 0 j
where is the acceptable deviation from target.
The genetic algorithm is used in this study to determine the optimal solution for
the above optimization problem. The genetic algorithm is an optimization algorithm,
which has been developed on the basis of the mechanics of natural selection and
4508
D. Kim et al.
Experimental procedure
In this study, the dual response approach is adopted to optimize the penetration
in the GMA welding process. The base metal to be welded was mild steel with a
thickness of 5.8 mm and the joint type of I-groove. The electrode wire used was AWS
classication ER 70S-6 with a diameter of 1.2 mm. The shielding gas used in this
experiment was 100% CO2 with a ow rate of 20 litre/min. The welding power
source was operated in a constant voltage mode. In this study, the variation of the
root opening was used as a disturbance with its range established at 0.4 1.2 mm.
The weld geometry variables were measured after the experiment.
4.
4509
Whether a complete joint penetration weld, or a partial joint penetration weld are
made is determined on the design of welded connections (AWS 1994). In this study,
we make a partial joint penetration weld. The weld bead geometry has a direct
correlation with the weld quality and the penetration has the greatest eect on
weld quality. In addition, changes in the working environment cause a variance in
the penetration, undermining the quality of the weld. Therefore, it is necessary to
minimize the variance in penetration, and the dual response approach is used in this
study to determine the welding conditions under which the penetration is least
aected by the disturbance.
A set of experiments was conducted according to the experimental design. Models
for the mean value and standard deviation of the output variable were then induced,
and the optimal welding conditions were determined based on these models. The
input variables consisted of the wire-feed rate (X1 ) and welding speed (X2 ), while
the output variables consisted of the mean value and standard deviation of the penetration (P). The experimental method used for the inner array, comprising two input
variables, was the 23 factorial design. The level of each input variable used in the
experiment is shown in table 1. The sampling data acquired by considering the disturbance in each input variable of the inner array were obtained by altering the root
opening, which easily occurs during the welding process. The range of variance was
established at 0.4 1.2 mm, and the penetration was measured at ve dierent sites.
Table 2 shows the nine pairs of welding conditions obtained according to the
level of each input variable and the results of the experiments conducted under each
welding condition. The two input variables x1, x2 in table 2 have the following
relationship with the variables X1, X2.
xk
Factor
X1
X2
Xk Xk, 0
,
Xk, 1 Xk, 0
1k2
Factor name
1 Level
0 Level
1 Level
60
6
75
8
90
10
Table 1.
Natural units
Coded units
Penetration
Expt.
no.
X1
X2
x1
x2
y1
y2
y3
y4
y5
1
2
3
4
5
6
7
8
9
60
75
90
60
75
90
60
75
90
6
6
6
8
8
8
10
10
10
1
0
1
1
0
1
1
0
1
1
1
1
0
0
0
1
1
1
2.7
3.0
3.6
2.4
2.5
2.9
1.9
1.8
2.6
2.9
3.2
3.8
2.6
2.7
3.3
1.9
2.4
2.7
3.1
3.3
4.2
2.6
3.0
3.6
2.2
2.7
3.2
3.1
3.6
4.3
2.7
3.5
3.7
2.2
2.9
3.5
3.4
4.0
4.5
2.8
3.8
4.1
2.5
3.2
3.9
3.04
3.42
4.08
2.62
3.10
3.52
2.14
2.60
3.18
0.2608
0.3899
0.3710
0.1483
0.5431
0.4494
0.2510
0.5339
0.5450
Table 2.
4510
D. Kim et al.
In equation (1), Xk,0 is the actual value of the input variable before it is converted
to level 0, and Xk, 1 is the actual value of the input variable before it is converted
to level 1. Therefore, (Xk, 1 Xk,0) refers to the dierence between two levels of
the input variable.
The correlation between the input variables of the welding process and the correlation between the input and output variables of the welding process are shown in
table 3. The correlation coecient between the input variables is 0, denoting that
multi-collinearity does not exist and that both input variables must be included in the
models. Moreover, the mean value of the wire feed rate and the penetration display
a positive correlation while the welding speed and penetration show a negative
correlation in the correlation between the input and output variables.
Regression models are induced through regression analysis based on the
experimental data. The following are the regression models of the mean value
and standard deviation of the penetration obtained by using the least squares
method.
^ 3:0422 0:4967x1 0:4367x2 0:0567x21 0:0033x22
The performance of the models obtained through the regression analysis can be
determined by the coecient of multiple determination, R2 (Myers and Montgomery
1995). The coecient of determination is a numerical expression of how much the
input variables of the models can reect the output variables. The coecients of
determination of regression models (2) and (3) are 0.9939 and 0.9210 respectively.
This shows that the regression model for the mean value has superior prediction
abilities to the regression model of the standard deviation.
Next, a set of input variables, which produces the desired penetration, was determined through optimization with the target value of the penetration determined by
the designer. The weld quality improves as the penetration approaches the objective
value and as the standard deviation of the penetration decreases. In this study, the
optimal solution is determined by using the genetic algorithm suggested in section 2,
with the target value of the penetration set at 4, 3.5, 3 mm, and the acceptable
deviation () at 0.2. First, the constrained problem of the optimization process is
transformed to the unconstrained problem using a penalty method. The following is
the original constrained problem in a minimization form:
minimize MSE ^ 0 2 ^ 2
such that g1 ^ 0 0, g2 ^ 0 0
x1
x2
Table 3.
x1
x2
1.0
0.0
0.7478
0.7049
1.0
0.6575
0.3092
1.0
0.3054
1.0
Correlations between input variables and between input variable and output
variable.
4511
2
X
gi
i1
where is a penalty function and r is a penalty coecient. In this study, gi g2i
and r 1.
Because the genetic algorithm is, in essence, a maximization problem, objective
function J is converted into 1/(J 1) to determine the optimal solution. The search
range of the two input variables is 1 1, and the control parameters of the genetic
algorithm are established at a population size of 30, crossover rate of 0.95,
and mutation rate of 0.01 based on the study conducted by Grefenstette (1986).
In addition, elitism (Goldberg 1989) is employed to force the best string of the
previous generation into the population of the next generation.
Primarily, the optimal conditions for the welding process when the target value of
the penetration is 0 4 are searched. The results of the genetic algorithm executed
until the 20th generation are shown in gure 1. Marker is the average objective
function value of the population, which makes up a generation, while marker g is
the maximum objective function value of the population. As seen in the gure,
the objective function value converges to a smaller value as the number of the generation increases. Figure 2 shows the convergence process of the input variables
indicating the maximum objective function value in each generation. The coded
values of the optimal values determined through the genetic algorithm were
(x1,x2) (1.000,1.000), while the predicted output variables under the optimal
conditions were ^ , ^ 4:0290, 0:3610. In order to verify whether the optimal
solution from the genetic algorithm was found, level curves of the objective function
were produced. The level curves of the objective function are shown in gure 3,
where the solution from the genetic algorithm is veried to be the optimal solution.
Subsequently, the genetic algorithm was adopted to search for the optimal
conditions when the target value of the penetration is 0 3.5. The coded
values of the input variables determined through the genetic algorithm were
(x1,x2) (0.0987,0.9961), while the predicted responses under the optimal conditions were ^ , ^ 3:4254, 0:4329. The level curves of the objective function were
produced to verify whether the optimal solution had been found. The level curves of
the objective function are shown in gure 4 where A and B are the local extreme
points. Point A is the global optimal point, and thus gure 4 veries that the genetic
algorithm has found the global optimal point. When an optimization algorithm
based on derivatives is applied to the optimization problem, the algorithm has the
defect of converging on local extreme points A and B depending on the initial values
of the input variables.
Finally, the optimal conditions were explored when the target value of the penetration was 0 3.0. The coded values of the optimal values of the input variables
determined through the genetic algorithm were (x1,x2) (1.0000,0.8983), while
the predicted responses under the optimal conditions were ^ , ^ 2:9918, 0:2169.
The optimal conditions determined through this study were compared with the
conditions, which considered only the mean value of the penetration. The objective
function in the optimization method, which only considered the mean penetration,
was established at J ^ 0 2 . The genetic algorithm was adopted as the
4512
D. Kim et al.
1.5
0.5
Figure 1.
10
Generation number
15
20
History of the objective function for the dual response problem with the genetic
algorithm when 0 4.
1
x
1
x
2
0.5
x1,x2
Objective function, J
Average
Best
-0.5
-1
Figure 2.
10
Generation number
15
20
History of x1 and x2 for the dual response problem with the genetic algorithm
when 0 4.
optimization algorithm, and optimization was performed 10 times for each target
value. The minimum objective function values were used in the comparison.
The comparisons of the two objective functions are shown in tables 4 through 6.
Table 4 shows the comparison when the target value of the penetration is 4.0 mm.
The two methods show similar results. It is thought that the region near, or including, the penetration of 4.0 mm is relatively small and the variation of the standard
deviation of the penetration within the region is not large enough to aect the
determination of the optimal solution. Table 5 shows the result of the comparison
when the penetration is 3.5 mm. The two methods show dierent results under this
condition since the variation in the standard deviation within the proximity of the
region where the penetration is 3.5 mm aects the determination of the optimal
solution. Table 6 shows the result of the comparison when the target value of the
4513
5
1.
0.5
4
2
5
1.
0
3
0.
5
1.5
-0.5
5
0.
-0.5
0.
2
1.5
-1
-1
0
x
0.5
0.5
0.
3
1
1.5
5
0.
1. 5
0.5
0.
22
Figure 3.
0.3
2
0.2
0.5
-0.5
0.5
0.2
-0.5
0.
22
0.22
0.3
-1
-1
2
0.2
0. 3
0
x
0.5
3
0.
Figure 4.
Objective function
Optimal solution
^
^
MSE
^ 0 2
^ 0 2 ^ 2
(0.9883, 0.9433)
(1.0000, 0.9824)
3.9975
4.0214
0.3679
0.3623
0.1354
0.1317
Table 4.
Comparison of the optimal solutions for the two dierent objective functions when
0 4.0.
penetration is 3.0 mm. The average penetration of the two optimal solutions is
similar. Nevertheless, the standard deviation of the penetration when the optimal
solution only considers the mean value is twice as high as the optimal solution, which
considers the standard deviation of the penetration. Therefore, the former optimal
solution is more likely to produce weld defects.
4514
D. Kim et al.
Objective function
Optimal solution
^
^
MSE
(0.9941, 0.2102)
(0.0987, 0.9961)
3.5001
3.4254
0.4693
0.4329
0.2202
0.1930
^ 0
^ 0 2 ^ 2
Table 5.
Comparison of the optimal solutions for the two dierent objective functions
when 0 3.5.
Objective function
Optimal solution
^
^
MSE
(0.1359, 0.2532)
(1.0000, 0.9003)
3.0000
2.9927
0.5098
0.2169
0.2599
0.0471
^ 0
^ 0 2 ^ 2
Table 6.
Comparison of the optimal solutions for the two dierent objective functions
when 0 3.0.
The optimization results of the optimization problems, which consider the three
target values of the penetration, show that the optimal solution obtained by applying
the dual response approach has a smaller standard deviation than the optimal
solution obtained by considering only the average penetration. The former solution,
therefore, has a lower probability of producing defective welds. In addition, the
genetic algorithm can overcome the deciency of the derivative-based-optimization
method wherein the algorithm converges to a local extreme point.
5.
Conclusion
Disturbances and changes in the welding working environment lead to possible
deteriorations of weld quality. The dual response approach is adopted in this study
to determine the welding process parameters under which the output variables
associated with weld quality are less aected by disturbances or changes in the
working environment. In addition, it is adopted to determine the welding conditions
with the potential to produce the desired partial penetration. The input variables of
the welding process are the wire feed rate and welding speed, while the output
variable is the penetration. The set of experiments is conducted on the region of
interest according to the experimental design, and regression analysis is employed to
produce the models of the standard deviation as well as the mean of the penetration
based on the experimental results, leading the models to determine the optimal
conditions of the welding process. The results of the optimization show that the
optimal conditions determined through this study have a lower possibility of producing defective welds than the conditions determined through the existing methods.
Acknowledgements
The Brain Korea 21 Project and Critical Technology 21 Project of the Ministry of
Science and Technology, Korea supported this work.
References
AWS, 1994, Structural Welding Code-Steel, 13th Ed. (American Welding Society).
BOX, G. E. P, 1985, Discussion of O-line quality control, parameter design, and the Taguchi
method by R. N. Kackar. Journal of Quality Technology, 17, 189190.
DELL CASTILLO, E. and MONTGOMERY, D., 1993, A nonlinear programming solutions to the
dual response problem. Journal of Quality Technology, 25, 199204.
4515
GOLDBERG, D. E., 1989, Genetic Algorithms in Search, Optimization, and Machine Learning
(Addison-Wesley).
GREFENSTETTE, J. J., 1986, Optimization of control parameters for genetic algorithms. IEEE
Proceedings on Systems, Man and Cybernetics, 16, 122128.
KIM, D. and RHEE, S., 2002, Modeling and optimization of a GMA welding process by genetic
algorithm and response surface methodology. International Journal of Production
Research, 40, 16991711.
LIN, D. K. J. and TU, W., 1995, Dual response surface optimization. Journal of Quality
Technology, 27, 3836.
MCCONNELL, I. A. and MCPHERSON, N. A., 1997, The application of statistical process design
to a FCAW process. Welding Journal, 76(10), 412-s416-s.
MYERS, R. H. and CARTER, W. H., 1973, Response surface techniques for dual response
systems. Technometrics, 15, 301317.
MYERS, R. H. and MONTGOMERY, D. C., 1995, Response Surface Methodology: Process
and Product Optimization using Designed Experiments (Wiley).
PHADKE, M. S., 1989, Quality Engineering Using Robust Design (AT&T Bell Laboratories).
TARNG, Y. S. and YANG, W. H., 1998, Optimization of the weld bead geometry in gas tungsten
arc welding by the Taguchi method. International Journal of Advanced Manufacturing
Technology, 14, 549554.
TARNG, Y. S., TSAI, H. L. and YEH, S. S., 1999, Modeling, optimization and classication of
weld quality in tungsten inert gas welding. International Journal of Machine Tools &
Manufacture, 39, 14271438.
VINING, G. G. and MYERS, R. H., 1990, Combining Taguchi and response surface philosophies: a dual response approach. Journal of Quality Technology, 22, 3845.