Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

J Intell Manuf

DOI 10.1007/s10845-014-0891-x

Statistical modeling and optimization of resistance spot welding


process parameters using neural networks and multi-objective
genetic algorithm
Hamed Pashazadeh · Yousof Gheisari ·
Mohsen Hamedi

Received: 7 January 2014 / Accepted: 24 February 2014


© Springer Science+Business Media New York 2014

Abstract Resistance spot welding (RSW) is a highly used Introduction


joining procedure in automotive industry. In RSW, after
a number of welds the welding electrode starts to wear Because of the continuous developing of automotive and
and its diameter changes. This causes the weld nugget aerospace industry, nowadays the application of spot welding
diameter abnormal variations and consequently reduces the in these fields is more highlighted than the past times. Resis-
weld strength. Therefore the tip of the electrode should be tance spot welding is the most simple and applicable method
dressed in RSW. Selecting the optimum time for the weld- of spot welding which is implemented on thin sheets. The
ing electrode tip dressing operations is very important. In weld strength in RSW is directly affected by the diameter
this research three welding parameters including the weld- and height of the weld nugget which are in turn influenced
ing time, the welding current, and the welding pressure by many process parameters such as the applied electrical
were identified as the main effective parameters on the weld current, the welding gun force, the welding time (including
nugget dimensions including the weld nugget diameter and the squeeze time, the welding time, the holding time, and
height using full factorial design of experiments. Then using the separation and preparation time for the next weld), the
hybrid combination of the artificial neural networks and welding machine setting, and the electrode and the workpiece
multi-objective genetic algorithm, the optimized values of material properties. In RSW, after some welds the welding
the aforementioned parameters were specified. Finally exper- electrode starts to wear and its diameter changes. This causes
iments were fulfilled to estimate the admissible number of the weld nugget diameter abnormal variations and conse-
the weld spots which should be done before the electrode tip quently reduces the weld strength. Therefore the tip of the
dressing operation. electrode should be dressed in RSW. Selecting the optimum
time for the welding electrode tip dressing operations is very
Keywords Resistance spot welding · Electrode tip important. If the dressing operation is done before the opti-
dressing · Design of experiments · Artificial neural mum time, the production costs will rise due to the stop time
networks · Multi-objective genetic algorithm of the production line and earlier replacement of the elec-
trode, and if this operation is done after the optimum time,
the weld strength will reduce due to the wear of the electrode
H. Pashazadeh (B)
Young Researchers and Elites Club, East Azarbaijan Science tip.
and Research Branch, Islamic Azad University, Tabriz, Iran Hamedi et al. (2007) optimized three important process
e-mail: hamed_pashazadeh@yahoo.com; parameters in the spot welding of the body components,
hamedpashazadeh@ut.ac.ir
namely welding current, welding time, and gun force. They
Y. Gheisari studied the effects of these parameters on deformation of the
Bushehr Branch, Department of Mathematics, sub-assemblies experimentally and utilized neural networks
Islamic Azad University, Bushehr, Iran and multi-objective genetic algorithms to select the optimum
values of welding parameters that yield the least values of
M. Hamedi
School of Mechanical Engineering, College of Engineering, dimensional deviations in the sub-assemblies. Eisazadeh et
University of Tehran, Tehran, Iran al. (2010) presented a mechanical–electrical–thermal cou-

123
J Intell Manuf

pled model in a finite element analysis to study the phenom- gap between the joint parts and the torch angle have an
enon of nugget formation and the effects of process para- important effect on the welded joint strength value and the
meters on this phenomenon. Hashemi et al. (2012) proposed optimum quantity of the shielding gas and the pulse fre-
a 2D FEM simulation of RSW to study the nugget devel- quencies existing in the tensile strength of welded joints.
opment in the different phases of the process. Ganjigatti et Karimzadeh et al. (2006) reported an exploration on the
al. (2008) attempted to determine input–output relationships effect of epitaxial growth on microstructure of Ti–6Al–4V
of the MIG welding process by using regression analysis alloy weldment which was examined by artificial neural net-
based on the data collected as per full-factorial design of works. The artificial neural network was developed to pre-
experiments. They investigated the effects of the welding dict grain size of fusion zone (FZ) at different currents and
parameters and their interaction terms on different responses welding speeds. The results showed that a coarse primary
considering both linear as well as nonlinear regression analy- phase develops in the fusion zone as a result of epitax-
ses. Darwish and Al-Dekhial (1999) reported experimental ial nucleation on coarsened grains near the heat affected
investigations on the spot welding of commercial aluminum zone (HAZ) which grows competitively into the molten
sheets. They carried out experiments to study the influence weld pool. Podrzaj et al. (2004) proposed a linear vector
of spot welding parameters (welding current, welding time, quantization (LVQ) neural network system to achieve the
electrode force and sheet thickness) on the strength of spot expulsion detection goal in RSW. The network was ana-
welded aluminum sheets. They established mathematical lyzed with different sensor combinations and different mate-
models correlating process parameters and their interactions rials. Their observations showed that the LVQ neural net-
with the response parameters on the basis of response surface work is able to detect the expulsion in different materials.
methodology (RSM). Zhang et al. (2000) employed a statis- Their experiment also pointed to the welding force signal
tical model to explore expulsion considering the influence as the most important indicator of the expulsion occurrence.
of random factors. Their models were built based on exper- Yildiz (2008, 2009a,b,c,d,e, 2012a,b, 2013a,b,c,d,e,f) in var-
imental data and one steel and two aluminum alloys were ious studies proposed new approaches of optimization meth-
used as examples. They presented expulsion probabilities ods to solve manufacturing problems. Yildiz (2013a) pro-
as a function of electrode force, welding current, and time. posed a hybrid optimization technique based on differen-
Li et al. (2001) investigated the relationship between weld tial evolution algorithm for structural design optimization
quality and various process conditions, including abnor- problems. His results showed that the proposed approach
mal process conditions using a newly developed two-stage, gives better solutions compared to genetic algorithm, particle
sliding-level experiment. From the analysis, they found that swarm, immune algorithm, artificial bee colony algorithm.
abnormal process conditions, such as axial misalignment, Yildiz (2012a) presented a hybrid optimization method based
angular misalignment, poor fitup, edge weld, and electrode on differential evolution algorithm. The hybrid approach
wear, significantly affect weld size and thus cause large vari- (DERE) was used to select optimal machining parameters
ation in the weld quality. Although they may help increase in turning operations. The DERE outperformed all the com-
the weld size in some cases, abnormal process conditions pared algorithms in solving the turning optimization prob-
generally lead to a less robust process. They suggested that lems. Yildiz (2013b) developed a novel hybrid optimization
high current and large electrodes should be used for reduc- method (HRABC) based on artificial bee colony algorithm
ing the weld quality variation. Li et al. (2000) established and Taguchi method. The results demonstrated the superi-
an ANN model to calculate the weld quality in RSW. They ority of the HRABC over the other techniques like differ-
used the fixture force, the dynamic resistance and the elec- ential evolution algorithm, harmony search algorithm, parti-
trode displacement as the input parameters in their net- cle swarm optimization algorithm, artificial immune algo-
work. Meran (2006) studied the use of stochastic search rithm, ant colony algorithm, hybrid robust genetic algo-
process that is the basis of genetic algorithms (GAs), in rithm, scatter search algorithm, genetic algorithm in terms of
developing estimation of the welding parameters for the convergence speed and efficiency. Yildiz (2009a) proposed
joined brass plates. They utilized Genetic Algorithm Weld- an approach based on two-stages which are (1) Taguchi’s
ing Current Estimation Model and Genetic Algorithm Weld- robust design approach to find appropriate interval levels
ing Velocity Estimation Model to estimate the welding cur- of design parameters (2) Immune algorithm to generate
rent and velocity according to the welding environment for optimal solutions using refined intervals from the previous
the brass material. Canyurt (2005) extended GA approach stage, and he showed that the proposed approach not only
to the estimation of mechanical properties of the joining can find optimal but also can obtain both better and more
of brass materials. The effects of five welding design para- robust results than the existing algorithm. Yildiz (2009b,c)
meters on the strength value using the Genetic Algorithm presented the first application of immune algorithm to the
Welding Strength Estimation Model (GAWSEM) have been optimization of machining parameters in the literature. The
examined. Their results indicated that the changes of the hybrid approach was applied to a case study for milling oper-

123
J Intell Manuf

ations to show its effectiveness in machining operations. In Table 1 The chemical composition of used steel sheets
another work Yildiz (2009d) developed a novel hybrid opti- C Mn P S Si Al N Fe
mization method based on particle swarm optimization algo-
rithm and receptor editing property of immune system. He 0.08 0.32 0.009 0.018 0.024 0.046 0.007 Bal.
also developed a hybrid Taguchi–Harmony search algorithm
for solving engineering optimization problems. Yildiz (2008,
2013c) proposed a new optimization method (HRDE) based
on Taguchi’s method and differential evolution algorithm.
He used Taguchi’s method to find appropriate interval levels
of design variables. In another work Yildiz (2009e) devel-
oped an effective hybrid immune-hill climbing optimization
approach for solving design and manufacturing optimization
problems in industry. He focused on a hybrid method com-
bining immune algorithm with a hill climbing local search
algorithm for solving complex real-world optimization prob-
lems. This research is the first application of immune algo-
rithm to the optimization of machining parameters in turning
and also shape design optimization problems in the literature.
Among many studies, Yildiz (2013d,e,f) investigated on the
development of different evolution algorithms for the selec-
tion of optimal machining parameters in milling operations. Fig. 1 The welding gun pressure measurement
Recently A. R. Yildiz proposed a structural design optimiza-
tion of vehicle components using Cuckoo search algorithm
(Yildiz 2012b). welding current, the electrode force and the welding time
In this paper, the effect of the welding parameters namely were calibrated. The welding time is normally measured
the welding time, the welding pressure and the welding in units called cycles (one sixtieth of a second on 60 Hz
current on the diameter and the height of the nugget have power system) or in milliseconds (one thousandth of a sec-
been investigated using full factorial design of experiments ond). Figure 1 illustrates the welding gun pressure measure-
methodology. After proper design of experiments the opti- ment. The sheets’ thickness was 0.7 mm. The best minimum
mum values of the three welding parameters are specified value of the weld nugget diameter is 3.5 mm as reported
using hybrid combination of the artificial neural networks and in the KES standard.1 Figure 2 shows a view of the back
the multi-objective genetic algorithm in order to achieve the of the automotive body sheet with the location of the spot
best nugget sizes. Finally the number of spots which should welds.
be welded before the electrode tip dressing operation was
calculated. The proposed approach can be utilized in man-
ufacturing sub-assemblies that can fit and match better with Experimental investigation
adjacent parts in the automotive body. It enhances quality
of the joint and will result in improving overall quality the In Figs. 3 and 4 for applied different welding currents and
welded sheets. times the measured nugget diameter and height are presented,
respectively. As seen, with increasing electrical current the
nugget diameter increases and also increasing of the welding
Experimental setup time leads to a relative increase in the nugget diameter. There
is not an obvious correlation between the weld nugget height
The materials used in the present work are AISI 1008 com- and the welding time, however increasing the welding cur-
mercially steel sheets. The chemical composition of used rent causes the reduction in the weld nugget height. In Fig. 5
steel sheets is given in Table 1. The specimens were cut the measured nugget diameter for different applied welding
parallel to the rolling direction of the sheets. The surfaces pressures and currents with a constant welding time of six
of the specimens were cleaned using acetone, then washed cycles is illustrated. With increasing the welding gun pres-
and dried before welding. A small circle was marked on sure to 3.6 kgf/cm2 the nugget diameter decreases, but after
the center of the upper face of the specimens in order to that with increasing the pressure to 4.4 kgf/cm2 the nugget
help the accuracy of alignment of the electrode tip on the diameter increases.
center of the overlap area during welding, using a specially
designed fixture. Before starting the experimental work, the 1 KES C-G006 P.14 STANDARD.

123
J Intell Manuf

Fig. 2 The back of the automotive body sheet and the location of the spot welds

4.5 4.5
5 kA 5 kA
nugget diameter (mm)

nugget diameter (mm)


4 5.5 kA 5.5 kA
3.5
6 kA 6 kA
3
3.5 6.5 kA 6.5 kA
2.5
3 2
1.5
2.5 1
0.5
2
2 4 6 8 10 12 14 0
2 2.5 3 3.5 4 4.5 5
welding time (cycles)
welding pressure (kgf/cm^2)
Fig. 3 The measured nugget diameter at different welding time and
currents (the welding pressure was 3 kgf/cm2 ) Fig. 5 The measured nugget diameter at diameter at different welding
pressure and currents (the welding time was six cycles)

0.8
5 kA
0.7
5.5 kA
nugget height (mm)

0.6
causes and the chance causes influence the process outputs.
6 kA
Since the effects of the chance causes on the outputs are
0.5 6.5 kA
more dispensable, the process is more controllable. With
0.4
the DOE methodology we can evaluate the effects of the
0.3 assignable causes on the outputs and explore the control-
0.2 lability of the process. In this research a full factorial 2k
0.1 design of experiments was designed, that k is 3 here and
0 two main levels of −1 and +1 (Montgomery 2000). Thus, a
2 4 6 8 10 12 14
total of 16 combinations of the input parameters were con-
welding time (cycles)
sidered, according to the full factorial design of experiments
Fig. 4 The measured nugget height at diameter at different welding (DOE). Table 2 shows the experimental data collected as per
time and currents (the welding pressure was 3.6 kgf/cm2 ) the above DOE.
Figure 6 shows the Pareto diagram which is used to eval-
uate the effects of factors on the outputs. Using this diagram
Full factorial design of experiments the main factors of the process are characterized. The Pareto
diagram of the nugget diameter (see Fig. 6a) indicates that
Three main welding parameters namely, the welding pres- all the parameters and their combinations have direct influ-
sure, the welding time and the welding current were cho- ence on the nugget diameter, except factor A (the welding
sen as the input parameters which have the most important pressure). The welding current has the maximum effect. In
effects on the weld nugget diameter and height which were the case of the nugget height (see Fig. 6b), all factors and
considered as the outputs. In every process the assignable their combinations are effective. Similar to the nugget diam-

123
J Intell Manuf

Table 2 Experimental data collected as per full-factorial design of tors. The termSE Coef. indicates the standard error for the
experiments estimated coefficient, which measures the precision of the
Run order Welding Welding Welding Nugget Nugget estimation. The T-values are determined as the ratio of cor-
pressure current time diameter height responding value under coefficient and standard error. The p
value is the minimum value for a preset level of significance
1 3 5 6 2.68 0.44
at which the hypothesis of equal means for a given factor
2 3 5 6 2.53 0.47
can be rejected. All of the terms except the welding pressure
3 4.4 5 12 3.79 0.51
(A) have significant contributions on the nugget height and
4 4.4 6.5 6 3.53 0.38
diameter, as their p values are found to be less than the sig-
5 4.4 6.5 12 3.89 0.33
nificance level α = 0.05. The regression equation is written
6 3 6.5 12 4.2 0.37
to establish an input–output relationship in processes. This
7 4.4 5 6 3.45 0.7
equation is derived from the regression analysis, as only the
8 3 5 12 3.57 0.5 effective terms are considered in the equation. The following
9 4.4 5 12 3.7 0.51 response equation is obtained for the nugget diameter:
10 3 6.5 6 3.91 0.35
11 3 5 12 3.43 0.52 Nugget diameter = 3.586 + 0.2725 (current) + 0.210 (time)
12 4.4 6.5 12 3.79 0.36 − 0.2125 (pressure ∗ current) − 0.0525 (pressure ∗ time)
13 3 6.5 6 3.98 0.37
− 0.0988 (current ∗ time)
14 3 6.5 12 4 0.39
15 4.4 5 6 3.36 0.66 + 0.0863 (pressure ∗ current ∗ time)
16 4.4 6.5 6 3.57 0.37
Table 4 shows the analysis of variance which were derived
from the full factorial DOE. The different terms used in the
eter, the welding current is the most important parameter for ANalysis Of VAriance (ANOVA) are defined as follows. The
the weld nugget height. The normal probability plot of the termDF represents the degrees of freedom, i.e., the number
outputs is presented in Fig. 7. In these plots as the factor of terms that will contribute to the error in prediction. The
is further from the normal line, it would be more effective. termSS indicates the sum of squares for each term repre-
As seen, all parameters and their combinations are effec- senting the variability in the data contributed by that term.
tive except the welding pressure for the nugget diameter. In The sum of squares is divided by the DF to calculate the
Table 3 the regression analysis which was carried out from mean square (MS). The F-value is used to test the hypoth-
the full factorial DOE, is shown. In this Table, the termEf- esis, which is defined as the ratio of adjusted mean square
fect indicates the effect of different terms on the response value to the residual error. Normally, the analysis of variance
(expressed in the coded form). The termCoef.stands for the is used to identify the main factors and the analysis of regres-
coefficients of the coded form of the regression equation. It sion to establish a quantitative model in order to contribute
represents the relationship between the response and the fac- the effective factors to the responses. As seen from Table 4a,

Fig. 6 Pareto chart of the


factors

123
J Intell Manuf

Fig. 7 Normal probability plots


of factors

Table 3 Analysis of regression:


(a) nugget diameter (b) nugget Term Effect Coef SE Coef T P
height
(a) Estimated effects and coefficients for nugget diameter (coded units)
Constant 3.5863 0.02118 169.35 0.000
Pressure 0.0975 0.0487 0.02118 2.30 0.050
Current 0.5450 0.2725 0.02118 12.87 0.000
Time 0.4200 0.2100 0.02118 9.92 0.000
Pressure ∗ current −0.4250 −0.2125 0.02118 −10.03 0.000
Pressure ∗ time −0.1050 −0.0525 0.02118 −2.43 0.038
current ∗ time −0.1975 −0.0988 −0.02118 −4.66 −0.002
Pressure ∗ current ∗ time 0.1725 0.0863 0.02118 4.07 0.004
(b) Estimated effects and coefficients for nugget height (coded units)
Constant 0.4518 0.00428 105.46 0.000
Pressure 0.0512 0.0256 0.00428 5.98 0.000
Current −0.1737 −0.0868 0.00428 −20.28 0.000
Time −0.0312 −0.0156 0.00428 −3.65 0.007
Pressure ∗ current −0.0612 −0.0306 0.00428 −7.15 0.000
S = 0.0847054, R-Sq = Pressure ∗ time −0.0687 −0.0343 0.00428 −8.02 0.000
98.11 %, R-Sq (adj) = 96.45 % Current ∗ time 0.262 −0.0131 0.00428 3.06 0.016
S = 0.0171391, R-Sq =
98.71 %, R-Sq (adj) = 97.58 % Pressure ∗ current ∗ time 0.0437 0.0218 0.00428 5.11 0.001

the p value of the welding pressure (namely factor A) in – When the welding time increases from 6 to 12 cycles,
the analysis of variance is 0.05, which validates the previous the increase of the nugget diameter at the pressure of
results obtained from the Pareto chart, the normal probability 3 kgf/cm2 is more than that at the pressure of 4.4 kgf/cm2 ,
plot and the regression analysis. However, the combinations however both lines of variations are conduced to a same
of this factor (A) with the other factors are effective, so it can value.
be noted that the welding pressure has an important effect on – When the welding time rises from 6 to 12 cycles, the
the weld nugget diameter with a high probability factor. The increase in the nugget diameter at the electrical cur-
interaction plots are presented in Fig. 8. These plots repre- rent of 5 kA is more than that at the current of 6.5 kA.
sent the interaction effect of the inputs on the responses. The Furthermore the measured nugget diameters at the cur-
following highlights can be derived from Fig. 8: rent of 6.5 kA are higher than that at the current of
5 kA.
– When the applied current rises from 5 to 6.5 kA, the – When the welding current increases from 5 to 6.5 kA, the
increase of the nugget diameter at the pressure of 3 kgf/cm2 decrease in the nugget height at the pressure of 3 kgf/cm2
is more than that at the pressure of 4.4 kgf/cm2 . is higher than that at the pressure of 4.4 kgf/cm2 .

123
J Intell Manuf

Table 4 Analysis of variance: (a) nugget diameter (b) nugget height


Source DF SS MS F P

(a) Analysis of variance for nugget diameter


Pressure 1 0.0380 0.0380 5.30 0.050
Current 1 1.1881 1.1881 165.59 0.000
Time 1 0.7056 0.7056 98.34 0.000 Fig. 9 The modeled network
Pressure ∗ current 1 0.7225 0.7225 100.70 0.000
Pressure ∗ time 1 0.0441 0.0441 6.15 0.038
Neural network modeling
Current ∗ time 1 0.1560 0.1560 21.75 0.002
Pressure ∗ current ∗ time 1 0.1190 0.1190 16.59 0.004
ANNs have proven quite reliable when it comes to approx-
Error 8 0.0574 0.0071 imating complex mappings with multiple outputs. They are
Total 15 3.0307 also quite flexible in terms of number of output parame-
(b) Analysis of variance for nugget height ters and, when trained properly, demonstrate a generalization
Pressure 1 0.1050 0.1050 35.77 0.000 capability which is superior to their rivals. An ANN basically
Current 1 0.1207 0.1207 411.09 0.000 utilizes the principal of empirical risk minimization to min-
Time 1 0.0039 0.0039 13.30 0.007 imize the functional risk, defined on the basis of its training
Pressure ∗ current 1 0.0150 0.0150 51.09 0.000 data (Bishop 1996). The function is defined in terms of the
Pressure ∗ time 1 0.0189 0.0189 64.36 0.000 loss function and is commonly given as
Current ∗ time 1 0.0027 0.0027 9.38 0.016
1
l
Pressure ∗ current ∗ time 1 0.0076 0.0076 26.06 0.001
Remp (w) = [yi − f (xi , w)]2
Error 8 0.0023 0.0002 l
i=1
Total 15 0.1818
where yi is the desired output of the network for a given
S = 0.0847054, R-Sq = 98.11 %, R-Sq (adj) = 96.45 %
S = 0.0171391, R-Sq = 98.71 %, R-Sq (adj) = 97.58 % input vector xi and f (xi , w) is the actual output produced by
the network. An error function, the quadratic loss function,
is defined as the squared distance between the actual output
– When the welding time increases from 6 to 12 cycles, the and the predicted response.
nugget height at the pressure of 3 kgf/cm2 increases, while In the current study the process to find a decent topology
at the same variation of the welding time and the pressure for the required network started with a 3 × 4 × 2 architecture
of 4.4 kgf/cm2 the nugget height decreases. while the middle layer is a hidden layer with four neurons.
– When the welding time increases from 6 to 12 cycles, at The number of neurons in the hidden layer was gradually
the welding current of 5 kA the nugget height decreases, increased, and the MSE of the neural network was observed
while at the welding current of 6.5 kA the nugget height in each configuration. The final neural network, a 3 × 12 × 2
does not change considerably. feed forward multi-layer-perception (MLP) architecture with
tansig transfer functions for all hidden neurons and pureline

Fig. 8 Interaction plot: a nugget diameter, b nugget height

123
J Intell Manuf

Table 5 Data implemented in the neural network modeling


Row Kind of Pressure Current Time Diameter Height
data (kgf/cm2 ) (kA) (cycles) (mm) (mm)

1 Training 3 5 6 2.6 0.45


2 Test 3 5 8 3.1 0.58
3 Training 3 5 10 3.3 0.59
4 Training 3 5 12 3.5 0.51
5 Training 3 5.5 6 3.4 0.42
6 Training 3 5.5 8 3.65 0.39
7 Training 3 5.5 10 3.7 0.38
8 Training 3 5.5 12 3.8 0.37
9 Validation 3 6 6 3.85 0.37
10 Training 3 6 8 3.75 0.43
11 Training 3 6 10 3.9 0.33
12 Training 3 6 12 4.15 0.49 Fig. 10 Performance of the modeled network
13 Training 3 6.5 6 3.95 0.36
14 Training 3 6.5 8 4.02 0.34
15 Training 3 6.5 10 4.05 0.39
functions for the two output neurons, was found to produce
16 Training 3 6.5 12 4.1 0.38
the most accurate results and to have the best generaliza-
17 Training 3.6 5 6 2.85 0.56
tion ability, as shown in Fig. 9. The combination of tansig
18 Test 3.6 5 8 3.3 0.61
and pureline for transfer functions was achieved at by trial
19 Training 3.6 5 10 3.9 0.64
and error. A radial-basis function network performed equally
20 Training 3.6 5 12 4 0.67
well but was outperformed by the MLP from a generaliza-
21 Training 3.6 5.5 6 2.9 0.51
tion viewpoint. The architecture should also comply with
22 Training 3.6 5.5 8 3.6 0.59
the practical recommendation that prescribes the minimum
23 Training 3.6 5.5 10 3.75 0.52
number of training data by the equation (Rogers and Vemuri
24 Training 3.6 5.5 12 4.15 0.63
1994)):
25 Validation 3.6 6 6 3 0.43
26 Test 3.6 6 8 3.5 0.39
27 Training 3.6 6 10 3.65 0.30 N (I + T + 1)
28 Training 3.6 6 12 3.85 0.40
C =1+
T
29 Training 3.6 6.5 6 3.4 0.37
30 Training 3.6 6.5 8 3.55 0.32
31 Training 3.6 6.5 10 3.78 0.39 in which C is the minimum number of training sets, N is the
32 Training 3.6 6.5 12 3.95 0.26 number of hidden neurons, I is the number of neurons in the
33 Validation 4.4 5 6 3.4 0.68
input layer, and T is the number of neurons in the output
34 Training 4.4 5 8 3.45 0.68
layer. In this case N = 12, I = 3, and T = 2 which gives 37
35 Training 4.4 5 10 3.7 0.66
as the minimum number of training sets required to train the
36 Validation 4.4 5 12 3.75 0.51
network. These data are presented in Table 5. Of these 48 sets
of experimental data, 38 were used as training data and five
37 Training 4.4 5.5 6 3.42 0.44
as test queries, and the other five as validation data as repre-
38 Training 4.4 5.5 8 3.5 0.29
sented in Table 5. The validation data is utilized here in order
39 Training 4.4 5.5 10 3.63 0.53
to prevent the network to be overtrained. The network was
40 Validation 4.4 5.5 12 4 0.62
trained using back propagation with a Levenberg–Marquardt
41 Training 4.4 6 6 2.84 0.40
optimization algorithm. In Fig. 10 the performance of the net-
42 Training 4.4 6 8 3.9 0.41
work is illustrated. As seen, after 19 epochs performance of
43 Test 4.4 6 10 4.03 0.31
0.0017 was achieved and the training error became relatively
44 Test 4.4 6 12 3.95 0.40
constant. The MSE errors of 37 training data were 0.0839
45 Training 4.4 6.5 6 3.55 0.37
and 0.0275 for the nugget diameter and height, respectively,
46 Training 4.4 6.5 8 3.9 0.32
which are acceptable. Table 6 shows the error of the five test
47 Training 4.4 6.5 10 3.98 0.28
data. Consequently it can be noted that the modeled network
48 Training 4.4 6.5 12 3.85 0.34
with 12 neurons in the hidden layer was desired for this study.

123
J Intell Manuf

Table 6 The error of the test data


Row Data 1 Data 2 Data 3 Data 4 Data 5

Measured nugget diameter 3.1 3.3 3.55 3.5 3.95


Network output of nugget 3.04 3.31 3.45 3.42 4.01
diameter
Error −0.06 0.01 −0.1 −0.08 0.06
Measured nugget diameter 0.58 0.61 0.37 0.39 0.4
Network output of nugget 0.5 0.61 0.26 0.41 0.37
diameter
Error −0.08 0 −0.11 0.02 −0.03

Multi-objective genetic algorithm optimization

Several variations of this algorithm have been reported in the


literature, of which the non-dominated sorting genetic algo- Fig. 11 Optimal Pareto points
rithm II (NSGA II) by Deb et al. (2000) has been reported to
be the most successful in a vast range of engineering appli- Table 7 Comparison of the experimental and GA results
cations. A brief pseudo-code for the NSGA II is presented as
Pressure Current Time Diameter Height
follows (Hamedi et al. 2007): (kgf/cm2 ) (kA) (cycles) (mm) (mm)

GA result 4.02 6.14 12.00 4.68 0.43


Step 1. Initialize a random population of potential solutions
Experiment 4 6.1 12 4.62 0.46
(chromosomes). test
Step 2. Evaluate individual objective functions for each Error 0.02 0.04 0.0 0.06 −0.03
potential solution.
Step 3. Perform the non-dominated sorting; determine the
number of dominations for each potential solution,
rank it accordingly, and assign to it a fitness value the Pareto point will be more feasible. So our best point is
inversely proportional to its rank. the point with the maximum nugget diameter as shown in
Step 4. Perform the fitness-proportionate selection and add Fig. 11. In Table 7 the results of the MOGA and experimen-
selected individuals to the parent population; re-rank tally measured data with the input welding parameters of the
the new (doubled) population. selected optimal Pareto point are compared. The errors in
Step 5. Form the mating pool from the best (lowest ranking) this table could be divided into two main errors, namely the
individuals in the doubled population while apply- experimental tests errors and the neural network model error.
ing similarity restrictions (e.g. crowding distance).
Step 6. Apply fitness-proportionate selection together with
other genetic operators to the mating pool and pro- Specification of the optimum electrode tip dressing time
duce the next generation.
Step 7. Check the termination criteria; if not met, repeat the In RSW after a number of welds the electrode starts to wear
above steps. due to the electrical current passing through it and the pres-
sure applying on it. The wear of the electrode causes the
Using the neural network of section “Experimental inves- nugget dimensions especially its diameter to digress the stan-
tigation” as function evaluator, the multi-objective genetic dards. So it is necessary to dress the electrode tip after a
optimizer is employed with a population size of 30, a (two- specified time. The aim of this section is to specify the opti-
point) crossover rate of 0.8, a (uniform) mutation probability mum number of spot welds which should be joined before
of 0.01, and the maximum number of generations of 100. the tip dressing operation. The welding parameters were set
Values of the 40 Pareto optimal points are depicted graphi- to be optimum status achieved through MOGA optimization
cally in Fig. 11. Considering the sheet thickness of 0.7 mm methodology and the nugget diameter at different number of
the best values of the nugget height are between 25 and 70 % spot welds was measured. The results are shown in Table 8.
of the thickness, which results in the optimum range of the A line was fitted through the data as illustrated in Fig. 12
nugget height between 0.18 and 0.49 mm (see footnote 1). with an error of ±12. It can be seen that the nugget diameter
On the other hands with the nugget diameter getting larger, deceases linearly with the number of the spot welds. The line

123
J Intell Manuf

Table 8 The measured nugget diameter after a specified number of sizes. Finally the number of spots which should be welded
spot welds before the electrode tip dressing operation was calculated.
No. of spot welds Measured nugget diameter The proposed approach can be utilized in manufacturing sub-
assemblies that can fit and match better with adjacent parts
1 4.62
in the automotive body. It enhances quality of the joint and
90 4.49
will result in improving overall quality the welded sheets.
180 4.35 The findings of this study are summarized as follows:
270 4.14
360 3.94
– With increasing electrical current the nugget diameter
450 3.81
increases and also increasing the welding time leads to
540 3.64
a relative increase in the nugget diameter. There is not an
obvious correlation between the weld nugget height and
the welding time, however increasing the welding current
causes the decrease in the weld nugget height.
– The welding current is the most important parameter
affecting the nugget dimensions.
– Hybrid combination of the artificial neural networks
(ANNs) and Multi Objective Genetic Algorithm (MOGA),
successfully found the optimized values of the welding
parameters.
– It is found that, the range of 591 up to 615 spot welds
is suitable for the tip dressing operation in the studied
welding operations.

References
Fig. 12 The fitted line of data represented in Table 8
Bishop, C. M. (1996). Neural network for pattern recognition. Oxford:
Oxford University Press.
Canyurt, O. E. (2005). Estimation of welded joint strength using genetic
equation is: algorithm approach. International Journal of Mechanical Sciences,
47, 1249–1261.
y = (−0.0019) x + (4.647) Darwish, S. M., & Al-Dekhial, S. D. (1999). Statistical models for spot
welding of commercial aluminium sheets. International Journal of
where, y denotes the nugget diameter and x the number of Machine Tools and Manufacture, 39, 1589–1610.
spot welds. Solving this equation for the desired nugget diam- Deb, K., Agrawal, S., Pratap, A., & Meyarivan, T. (2000). A fast elitist
non-dominated sorting genetic algorithm for multi objective opti-
eter of y = 3.5 mm, x = 603.68 is gained. It means that the
mization: NSGA-II. In M. Schoenauer, K. Deb, G. Rudolph, X. Yao,
range of 591 up to 615 spot welds is suitable for the tip E. Lutton, J. J. Merelo, & H. P. Schwefel (Eds.), Parallel problem
dressing operation. On the other hands, before this range all solving from nature (PPSN VI). Lecture notes in computer science
welded spots have the nugget diameter of 3.5 mm or near (Vol. 1917). Berlin: Springer.
Eisazadeh, H., Hamedi, M., & Halvaee, A. (2010). New parametric
that, but after this range the electrode tip starts to wear and
study of nugget size in resistance spot welding process using finite
so performing the dressing operation is necessary. element method. Materials and Design, 31, 149–157.
Ganjigatti, J. P., Pratihar, D. K., & RoyChoudhury, A. (2008). Mod-
eling of the MIG welding process using statistical approaches. The
International Journal of Advanced Manufacturing Technology, 35,
Conclusions 1166–1190.
Hamedi, M., Shariatpanahi, M., & Mansourzadeh, A. (2007). Optimiz-
In this paper, the effect of the welding parameters namely the ing spot welding parameters in a sheet metal assembly by neural
networks and genetic algorithm. Proceedings of the Institution of
welding time, the welding pressure and the welding current Mechanical Engineers, Part B: Journal of Engineering Manufac-
on the diameter and the height of the nugget have been inves- ture, 221, 1175–1184.
tigated using full factorial design of experiments methodol- Hashemi, R., Pashazadeh, H., & Hamedi, M. (2012). An incrementally
ogy. After proper design of experiments the optimum values coupled thermo-electro-mechanical model for resistance spot weld-
ing. Materials and Manufacturing Processes, 27, 1442–1449.
of the three welding parameters are specified using hybrid Karimzadeh, F., Ebnonnasir, A., & Foroughi, A. (2006). Artificial neural
combination of the artificial neural networks and the multi- network modeling for evaluating of epitaxial growth of Ti6Al4V
objective genetic algorithm in order to achieve the best nugget weldment. Materials Science and Engineering A, 432, 184–190.

123
J Intell Manuf

Li, W., Cheng, S., Hu, S. J., & Ni, J. (2000). On-line quality estimation Yildiz, A. R. (2009e). An effective hybrid immune-hill climbing opti-
in resistance spot welding. Journal of manufacturing Science and mization approach for solving design and manufacturing optimiza-
Engineering, 122, 511–512. tion problems in industry. Journal of Materials Processing Technol-
Li, W., Cheng, S., Hu, S. J., & Shriver, J. (2001). Statistical investiga- ogy, 209(6), 2773–2780.
tion on resistance spot welding quality using a two-state. Sliding- Yildiz, A. R. (2012a). A comparative study of population-based opti-
Level Experiment, Journal of Manufacturing Science and Engineer- mization algorithms for turning operations. Information Sciences,
ing, 123, 513–520. 210, 81–88.
Meran, C. (2006). Prediction of the optimized welding parameters Yildiz, A. R. (2012b). Structural design optimization of vehicle com-
for the joined brass plates using genetic algorithm. Materials and ponents using Cuckoo search algorithm. Materials Testing, 54(3),
Design, 27, 356–363. 185–188.
Montgomery, D. C. (2000). Design and analysis of experiments (5th Yildiz, A. R. (2013a). Comparison of evolutionary-based optimization
ed.). London: Wiley. algorithms for structural design optimization. Engineering Applica-
Podrzaj, P., Polajnar, I., Diaci, J., & Kariz, Z. (2004). Expulsion detec- tions of Artificial In telligence, 26(1), 327–333.
tion system for resistance spot welding based on a neural network. Yildiz, A. R. (2013b). A new hybrid bee colony optimization approach
Measurement Science and Technology, 15, 592–598. for robust optimal design and manufacturing. Applied Soft Comput-
Rogers, R. D., & Vemuri, V. R. (1994). Artificial neural networks fore- ing, 13(5), 2906–2912.
casting time series. Los Alamitos, CA: IEEE Computer Society Yildiz, A. R. (2013c). Hybrid Taguchi-differential evolution algorithm
Press. for optimization of multi-pass turning operations. Applied Soft Com-
Yildiz, A. R. (2008). Hybrid Taguchi–Harmony search algorithm for puting, 13(3), 1433–1439.
solving engineering optimization problems. International Journal Yildiz, A. R. (2013d). Optimization of cutting parameters in multi-
of Industrial Engineering Theory, Applications and Practice, 15(3), pass turning using artificial bee colony-based approach. Information
286–293. Sciences, 220, 399–407.
Yildiz, A. R. (2009a). A new design optimization framework based Yildiz, A. R. (2013e). A new hybrid differential evolution algorithm for
on immune algorithm and Taguchi method. Computers in Industry, the selection of optimal machining parameters in milling operations.
60(8), 613–620. Applied Soft Computing, 13(3), 1561–1566.
Yildiz, A. R. (2009b). A novel hybrid immune algorithm for global Yildiz, A. R. (2013f). Cuckoo search algorithm for the selection of
optimization in design and manufacturing. Robotics and Computer- optimal machining parameters in milling operations. International
Integrated Manufacturing, 25(2), 261–270. Journal of Advanced Manufacturing Technology, 64(1–4), 55–61.
Yildiz, A. R. (2009c). Hybrid immune-simulated annealing algorithm Zhang, H., Hu, S. J., Senkara, J., & Cheng, S. (2000). A statistical
for optimal design and manufacturing. International Journal of analysis of expulsion limits in resistance spot welding. Journal of
Materials and Product Technology, 34(3), 217–226. Manufacturing Science and Engineering, 122, 510–510.
Yildiz, A. R. (2009d). A novel particle swarm optimization approach
for product design and manufacturing. International Journal of
Advanced Manufacturing Technology, 40(5–6), 617–628.

123

You might also like