WELDING RESEARCH

Primary Circuit Dynamic Resistance

Monitoring and its Application to Quality
Estimation during Resistance Spot Welding
Weld strength and nugget diameter were estimated by using primary circuit
process parameter as a real-time in-process monitoring system

BY Y. CHO and S. RHEE

ABSTRACT. Resistance spot welding is

one of the most widely used processes in
sheet metal fabrication. Improvements in
quality assurance are important to increase welding productivity. Until now,
quality estimation was performed using
dynamic resistance at the secondary circuit, which is easily measured. However,
this method has many problems when applied to in-process real-time systems. In
this study, weld quality was ensured
through the use of a new dynamic resistance monitoring method. This method
used instantaneous current and voltage
values measured at the primary circuit. A
weld quality estimation system is suggested using these variables. For quality
estimation, ten factors relating to quality
were extracted from the primary circuit
dynamic resistance, measured in the
timer. The relationship between these
factors and weld quality was determined
through a regression analysis. Four regression models were developed using
the results of the analysis in order to estimate weld strength and nugget diameter.
Neural network models were also suggested using these factors. The results of
the neural network were compared with
those of the regression models. This was
designed so weld quality estimation could
be obtained upon completion of welding.
Moreover, the proposed method using
the primary circuit dynamic resistance
has an advantage over the conventional
one, such as obtaining information on
quality without the use of extra devices.
Y. CHO is a Research Associate, Department of
Mechanical Engineering, The University of
Michigan, Ann Arbor, Mich. S. RHEE is an Associate Professor, Dept. of Mechanical Engineering, Hanyang University, Seoul, Korea.
This paper was presented at the AWS 81st Annual Convention, April 2527, 2000, Chicago,
Ill.

104-S JUNE 2002

Introduction
For several decades, resistance spot
welding has been an important process in
sheet metal fabrication. The automotive
industry, for example, prefers spot welding for its simple and cheap operation.
Recently, however, attempts are being
made to reduce the number of spot welds
to increase productivity. In order to minimize the number of spot welds and satisfy essential factors such as strength,
weld quality must be obtained. Traditionally, to check weld quality, destructive
and nondestructive tests were used on
randomly sampled workpieces at the production site. These processes, however,
can only be examined off-line, making it
impossible to receive pertinent information regarding the weld quality during the
process. Weld quality estimation must be
done in real time to monitor and repair
weld defects as they occur.
As resistance spot welding has electromechanical elements, various weld quality control techniques have been proposed based upon the following process
parameters: dynamic resistance, welding
current, voltage, and electrode displacement (Refs. 1, 2). In general, electricitybased systems, which maintain the volt-

KEY WORDS
Resistance Spot Welding
Inductive Noise
Primary Circuit
Dynamic Resistance
Regression Analysis
Correlation
Neural Network
Quality Estimation

age or the current at a constant preset

level (Refs. 3, 4), have been used to control those parameters automatically. Although the general voltage supply is reliable, severe voltage fluctuations may
occur in some areas and, in such situations, automatic voltage regulators have
been employed. The idea to maintain
constant power or current for a set period
of time has also been employed in cases
where the impedance changes during
welding. In addition, a combined current
and voltage regulator, which adjusts the
phase-shift control automatically, was introduced. Investigations based upon the
thermal expansion have been undertaken
in attempts to control the quality as well.
A resistance spot welding machine has
moving parts such as electrodes or arms.
Moreover, a direct relationship between
the rate of electrode separation and the
properties of the weld exists. Johnson, et
al. (Ref. 2), developed a spot-weld correction system based on sensing the weld
thermal expansion in which correction
was achieved by automatic adjustment of
weld load. A real-time adaptive spot
welding control system based upon a
measurement of electrode displacement
was also operational on the factory floor
throughout the recent study (Ref. 5). Resistance correction technique forms another basis for an automatic quality control system in spot welding (Ref. 6). In the
early 1970s, commercial units of these
quality control systems were developed
and attached to welding machines in the
auto body industry (Ref. 4).
Measurement of dynamic resistance
has been one of the most effective techniques of quality monitoring and estimation during the past several decades. Some
of the earliest and simplest techniques
were to monitor the voltage and current at
the secondary circuit. The electric parameters, however, vary frequently during
welding cycles due to resistance change.

WELDING RESEARCH

Fig. 1 Schematic diagram of resistance spot welding monitoring system.

Therefore, the monitored power input

may not be appropriate, and the true energy state of welding is not represented.
Dynamic resistance, on the other hand,
has relatively accurate information on energy generation during nugget formation
(Ref. 7). In earlier studies (Ref. 8), the potential of dynamic resistance as a parameter indicative of weld quality was explored.
The values of dynamic resistance as a
nominal condition of test welding were
sampled and stored in an electronic memory. These values were compared with values sampled at the production welds. The
values in agreement with those for the test
weld, within an acceptance band, might
then be considered as good as the initial
test weld. Another quality evaluation technique was proposed with microprocessorbased resistance monitoring system as an
in-process scheme (Ref. 9). The dynamic
resistance characteristic of an accepted
weld was computed as a reference in the
monitoring system. The next procedure
was very similar in comparison to the first
example except not only the absolute values were compared to the reference but
also the slopes of the dynamic resistance
characteristics.
In order to acquire the resistance signal, the dynamic resistance was calculated from the welding current, and the
voltage was measured at the secondary
circuit using an oscilloscope in the early
stages (Refs. 1012). The physical meaning of resistance variation, however,
could not be clearly explained in these
studies. Dynamic resistance was calculated more efficiently by applying the root
mean square (rms) value of the measured
signal to the analog circuit using the Halleffect sensor and voltage-measuring devices (Ref. 13). In this study, the effect of
the dynamic resistance on the nugget formation and ensuing dynamic resistance
pattern were considered. The study was
useful in explaining the variations of the
dynamic resistance according to the
nugget formation, such as the collapse of

the contact resis- Fig. 2 Simplified welding machine circuitry including primary circuit montance, the increas- itoring units.
ing temperature
of the faying surface, and the melting and plastic deforresistance was used, which proved effimation of the nugget. Kaiser, et al. (Ref.
cient for in-process application. By com14), and Thornton, et al. (Ref. 15), atparing the rms dynamic resistance actempted to examine the nugget formation
quired from the traditional method of the
according to changes in contact resissecondary circuit monitoring system with
tance. Kaiser, et al., explained the effect
the proposed dynamic resistance based
of the dynamic resistance and initial conon the primary circuit monitoring system,
tact resistance on the lobe curve, while
it became evident there was a direct relaThornton, et al., looked into the contact
tionship between the primary and secresistance of the aluminum alloy and the
ondary circuit dynamic resistance. Furensuing dynamic resistance variance.
thermore, the primary circuit dynamic
Research based on dynamic factors
resistance was used to estimate the weld
continues to be carried out regarding
quality. Tensile shear strength and nugget
weld quality estimation (Refs. 1620). In
diameter were estimated through the rea weld quality estimation system using
gression analysis and neural network.
multiple linear regression analysis (Ref.
Throughout the research, a system was
16), weld quality was examined by definproposed in which the weld quality could
ing the relationship between factors, such
be estimated in real-time upon compleas electrode displacement, dynamic resistion of the weld.
tance, and weld quality. Livshits (Ref. 18)
Process Parameter Monitoring
suggested a system that could more commonly be used. In his research, the dyInductive Noise and Dynamic
namic resistance based on the current
Resistance Monitoring
density of the faying surface was used to
assure weld quality. Research was also
As the resistance spot welding is
performed on weld quality estimation
processed, the electrical resistance variausing an intelligent algorithm such as the
tion of the welds, due to the Joule heat
neural network on resistance spot weldgenerated by the weld current, becomes
ing. Brown, et al. (Ref. 19), used the norone of the most important factors in
malized dynamic resistance, welding curnugget formation. Unlike the initial conrent, and electrode diameter to predict
tact resistance of the faying surface, the
the nugget diameter, which was closely
dynamic resistance includes information
related to weld strength. Dilthey, et al.
on nugget formation as welding occurs
(Ref. 20), however, examined the shear
(Refs. 13, 15, 21, 22). Generally, the restrength using welding current and voltsistance can be obtained by the voltage diage through a similar method.
vided by the current. In resistance spot
Although such studies based on rewelding circuitry, however, many probsults obtained from the secondary circuit
lems occur in the measurement of resiscan be applied to real-time weld quality
tance due to the inductive reactance eleestimation, in-process usage has several
ments of the electrical circuit. A simple
limitations. These limitations include the
means to acquire dynamic resistance, uninstallation location of the voltage meainfluenced by inductive noise, is to use the
suring device and increased costs due to
rms value. When the rms voltage deadditional measuring devices. Therefore,
tected across the electrode, which does
in this study, the primary circuit dynamic

WELDING JOURNAL 105-S

WELDING RESEARCH

Fig. 3 Primary circuit dynamic resistance pattern and quality estimation

factors.

Fig. 4 Linear relationship plot between primary and secondary dynamic

resistance.

Fig. 5 The neural network architecture.

not include much inductance (Ref. 12), is

divided by the rms current of the secondary circuit, resistance can be obtained
with no consideration to the phase shift
due to the inductive noise. Another way
of effectively eliminating the inductive
noise from the dynamic resistance is to
use the voltage and current when the
welding current is at the peak value of
each cycle. This method is based on the
principle that only the pure resistance element, not the inductive reactance element, influences the impedance of the alternating current circuit when the rate of
change of current equals zero (Refs. 1, 9,
15, 23). The greatest advantage of this
method is that the pure resistance variation at the secondary circuit can be monitored at the primary circuit.
Figure 1 is a schematic diagram of the
resistance spot welding system generally
used in auto body manufacturing. The
welding current that is controlled in the
timer of the welding machine is passed
through the transformer TR to the welding gun by the primary power cable. As
the secondary circuit dynamic resistance
monitoring system of this study, welding

106-S JUNE 2002

voltage and current are measured

in terms of rms
values. By attaching the clips to
each electrode
tip, the instantaneous
welding
voltage drop was
measured. The
Hall-effect sensor
located on the
lower arm of the
welding machine
monitored the current flowing in the secondary circuit. By using an analog-to-digital converter (ADC), the analog voltage
and the current were monitored simultaneously. In the earlier study (Ref. 1),
when a 60-Hz alternating current was
used, a sampling rate of more than 1 kHz
should be used. However, in order to effectively reflect the sharp voltage waveform produced by the electrical triggering of the silicon controlled rectifier
(SCR) to the resistance value, a sampling
rate of 6 kHz was used in this study. This
allowed the calculation of the dynamic
resistance with the 50 numbers of data
per half cycle. By using the digitized voltage and current, each rms value of a half
cycle was calculated. As another approach, this research introduced a
process to detect the dynamic resistance
from the timer as a primary circuit monitoring system, using a microprocessor.
In the primary circuit, the welding voltage and current could be detected without extra measuring devices; thus, this
method could be used effectively as an
in-process system. This method will be
addressed in the following sections.

Analysis of Welding Machine Circuit

The resistance spot welding machine

consists of a relatively simple electric circuit that includes a transformer. The simplified equivalent circuit is shown in Fig.
2. In the figure, the primary circuit elements are shown as subscript p and the
secondary circuit elements as s. In the primary circuit, Rp stands for the total resistance value of the primary circuit. Xp
stands for the primary leakage reactance.
In the secondary circuit, Rs stands for the
total resistance value of the secondary
circuit, which includes the welding gun
and the electrode. Xs stands for the secondary leakage reactance. RL stands for
the resistance change across the electrodes. When the secondary circuit elements are shifted to the primary circuit,
the resistance of the circuit, including
shifted secondary circuit elements, which
are shown as the reflected impedance described with the transformer ratio a, can
be written as an equivalent resistance R
and an equivalent inductive reactance X
in Equations 1 and 2, which corresponds
with the primary circuit current.

R = R p + a 2 Rs + RL
X = X p + a2 X s

(1)
(2)

By using the inductive elimination

method, as stated in the preceding section, the inductive reactance X can be removed to calculate the dynamic resistance from the measured current and
voltage of the primary circuit. Also, if assumed the transformer will operate ide-

WELDING RESEARCH
ally and the resistance change due to the
temperature change of the entire welding
machine circuit can be ignored, the relationship between the resistance across
the electrodes RL and the measured dynamic resistance Rmeas from the voltage
Vp and the current Ip in the primary circuit can be shown in Equation 3 below.

V
p
Rmeas =
= R p + a 2 Rs + RL
Ip

)
(3)

Where Ip is the peak current value of the

instantaneous current from the primary
circuit and Vp is the voltage at that
moment.
Primary Circuit Dynamic Resistance
Monitoring and Quality
Estimation Factors

The primary current made by the phase

control was measured using the current
transformer (CT) from the inside of the
timer. This primary current waveform was
appropriately amplified using the gain selector. The signal at this stage includes
high-frequency noise from the SCR current control unit and amplifier. By using
the integrator, which has a suitably selected condenser, the high-frequency
noise signal can be converted into a noisefree welding current waveform. The welding voltage was detected at the primary circuit and was attenuated using the
measuring transformer. The monitoring
trigger pulse was created when the rate of
change of current reached zero, triggering
the sample and holding circuit, which were
connected by the detected welding current
and the attenuated primary voltage. The
instantaneous welding current and voltage
were converted into digital values by the
analog-to-digital converter. These values
were used to calculate the dynamic resistance by dividing the current from the voltage of half cycle. The measured dynamic
resistance Rmeas includes the resistance of
the primary circuit resistance Rp and secondary circuit resistance Rs and RL.
Therefore, to measure the resistance of
the weld, the Rp and Rs must be removed
effectively from the calculated resistance
value. Fortunately, in the general resistance spot welding machine, the square of
the transformation ratio a2, shown in
Equation 3, has an order of hundreds or
thousands, so the Rp included in the calculated dynamic resistance Rmeas is a relatively small value. In other words, the Rp
can be ignored. With the proper measurement, the resistance of the secondary circuit Rs also can be obtained. This method
converts the primary circuit dynamic resistance into the dynamic resistance change
across the electrodes.

Table 1 Correlation Coefficients, Root Mean Square Errors, and Maximum Errors between
Primary and Secondary Dynamic Resistance

R
0.9207

6.5 kA
Erms

Emax

8 kA
Erms

Emax

9.5 kA
Erms

Emax

11 kA
Erms

Emax

2.0223 7.6392 0.9519 1.5340 4.1564 0.9625 1.5967 4.4573 0.9682 2.1217 6.0091

Table 2 Correlation Coefficients between Independent Variables Extracted from the Primary
Circuit Dynamic Resistance and Dependent Variables of Weld Strength YS and Nugget
Diameter YD
YS

YD

X1

X2

X3

X4

X5

X6

X7

X8

X9

X10

X1 0.957 0.978 1.000

X2
0.864 0.895 0.892 1.000
X3
0.894 0.910 0.891 0.972 1.000
X4
0.589 0.626 0.636 0.884 0.817 1.000
X5
0.524 0.488 0.458 0.639 0.689 0.535 1.000
X6
0.374 0.347 0.326 0.359 0.401 0.229 0.363 1.000
X7 0.906 0.937 0.924 0.950 0.947 0.799 0.509 0.357 1.000
X8 0.551 0.592 0.609 0.816 0.740 0.942 0.353 0.152 0.771 1.000
X9
0.255 0.274 0.219 0.526 0.536 0.636 0.624 0.114 0.428 0.577 1.000
X10 0.576 0.616 0.616 0.850 0.790 0.961 0.534 0.211 0.780 0.898 0.654 1.000

The primary circuit dynamic resistance is generally in a formation, as

shown in Fig. 3. After monitoring, ten factors were extracted and used in the weld
quality estimation based on the dynamic
resistance pattern. First, the location of
the beta peak X1, which is related to the
initial nugget formation, and the rising
speed of the dynamic resistance after the
alpha peak X2, which is related to the
heating rate of the faying surface, were
selected as the geometric extraction factors. Based upon similar dynamic resistance analysis (Refs. 13, 14), beta peak is
a balance point between a resistance increase, resulting from increasing temperature, and a resistance drop due to
molten nugget growth and mechanical
collapse. Therefore, the beta peak tends
to be reached earlier in high heating rate
(i.e., current level) and later in low heating rate. To make the estimation factor
reflect the variations in resistance value,
the difference between the maximum and
minimum values of the dynamic resistance X3 and the maximum value of the
dynamic resistance variation in each cycle
X4 were selected. At low currents, the
heating rate is relatively slow, resulting in
small changes in resistance. At expulsion
level currents, however, large variations
of resistance can be observed, such as
from a sharp drop due to the reduction in
material thickness and an increase in effective contact area. Also, the absolute
values of the monitored dynamic resistance, such as the maximum value of the
dynamic resistance X5, the initial resistance X6, the final dynamic resistance
X7, and the average value of the total dynamic resistance X8 were selected as estimation factors. The standard deviation

of the dynamic resistance X9 and the

standard deviation of the dynamic resistance variation per cycle X10 were used
to comprehend the effect of the trends in
resistance variation.

Experiment
Welding was performed on a 0.7-mmthick sheet of uncoated low-carbon steel.
A resistance spot welding machine using
single-phase, 60-Hz alternating current
with an attached pneumatic cylinder was
used. A 16-mm-diameter, dome-type electrode with a 6-mm-diameter, flat tip end
made with copper alloy of RWMA class II
was used. Welding was performed on a
specimen prepared according to an AWS
standard (Ref. 24) while varying the current under ten cycles of welding time and
2.45kN of electrode force. The welding
current was increased at 1.5-kA intervals
starting from 6.5 kA up to 11 kA. Welds
were tested in tensile-shear to determine
the ultimate strength and measured
nugget diameter. These readings were
used as the criterion for weld quality.

Results and Discussion

Relationship between Primary and
Secondary Circuit Dynamic Resistance

In order to investigate the relationship

between the dynamic resistance of the
primary and secondary circuits, the experimental results were analyzed statistically. First, the welding was carried out by
selecting four levels of current. The experiment used the one-way factorial design with nine replicates under each condition. The selection order of the

WELDING JOURNAL 107-S

WELDING RESEARCH
A

in Fig. 4. Figure 4
shows the relationship between
the two dynamic
resistances, according to the
changes in current. This figure
shows a small
change in the resistance value of
about 2340 in
the low current.
On the other
hand, 1846
can be observed in
the high current.
The distribution
of the resistance
values at 8 kA was
gathered around
Fig. 6 Strength estimation results of the models. A Linear model; B
the central line
nonlinear model; C neural network model.
without any abnormal protruexperimental schedule was done ransions, showing the lowest rms error. The
domly to reduce, as much as possible, the
distribution also coincided relatively well
error generated by noise such as elecwith the central line under most conditrode wear or measurement settings. To
tions, excluding a few anomalies.
quantitatively check the relationship beThrough the result described herein, it
tween the secondary and primary circuit
can be seen the suggested primary circuit
dynamic resistance, a correlation analysis
dynamic resistance can be used in the
was carried out. The results, shown in
same way as the secondary circuit dyTable 1, are the correlation coefficient of
namic resistance.
the two dynamic resistance values and the
Quality Estimation by
corresponding rms error Erms and maximum error Emax. Although the correlaRegression Analysis
tion coefficient somewhat increased as
To estimate weld quality for resistance
the current increased, the rms error
spot welding, regression models were forshowed a minimum value at the current
mulated through multiple linear and nonset point of 8 kA. This is not because the
linear regression analyses. The tensile
two data coincide better as the current inshear strength and the nugget diameter
creased but because the sudden decrease
were selected as dependent variables,
in resistance in the expulsion of the high
and ten estimation factors were used to
current caused a gap in the resistance
determine the regression equation. Bevalue, which is relatively larger than that
fore performing the linear regression
of the other conditions. This can be seen
C

108-S JUNE 2002

analysis, the correlation between dependent variables, such as weld strength YS

and nugget diameter YD, and the estimation factors X1 to X10, the independent variables, were observed. The results are shown in Table 2.
Compared to the relationship between the dependent variables and welding condition (i.e., current set point),
which has the correlation coefficients of
0.884 for weld strength and 0.909 for
nugget diameter, it can be seen the location of the beta peak X1, the difference
between the maximum and minimum
value of the dynamic resistance X3, and
the final dynamic resistance X7 were in
relative good linearity with strength YS
and nugget diameter YD. However, the
correlation coefficient between X2 and
X3 exceeded 0.97, showing the two variables were very closely related in terms of
linearity. Therefore, when these variables
are used simultaneously in the regression
model, a multicolinearity effect can be
observed. Thus, there was a possibility
the error rate of the regression models,
which included all these variables, would
increase. In order to overcome such problems, the independent variables were inputted in order from those having the
highest explainability regarding weld
quality, and a regression analysis using
the stepwise method to determine the regression model was performed. The variables were applied to the regression
model in stages, based on the partial correlation and significance probability values of the regression coefficient, until the
significance level exceeds 0.10, at which
point they were eliminated. Tables 3 and
4 show the variables entered at each
stage, the significance probability values
of the coefficients, and the coefficients of
determination R2 according to the model
in each stage.
An observation of Model I, which is a

WELDING RESEARCH
linear regression model for strength estimation, shows X1, X5, X4, and X3 are entered into the regression equation in that
order. In the case of X5 in Model I-4, the
variable is excluded in Model I-5 as it is
not statistically significant. Though X5
has been eliminated, the determination
coefficient maintains a value of 0.941. In
Model II, which is a linear model for
nugget diameter estimation, X1, X7, X3,
X4, and X2 are selected into the estimation model at each stage, and then X3 is
excluded as stated above. The final linear
regression equations of Model I and
Model II, based on the above results, are
shown in Equations 4 and 5.
YS = 3.4530.0973 X10.0108 X4
+ 0.0218 X3
(4)
YD = 7.5960.277 X10.0802 X7
+ 0.0641 X4 + 0.0901 X2
(5)
The regression coefficients, which have
been standardized to observe the effect of
each factor on the strength and diameter,
are shown in Table 5. According to this
table, the location of beta peak X1 has the
greatest effect on both strength and diameter estimation, followed by the difference between the maximum and minimum value of the dynamic resistance X3
in strength estimation and the rising
speed of the dynamic resistance X2 in diameter estimation.
Nonlinear regression models using
the above variables are shown in Equations 6 and 7. The nonlinear regression
models are formulated by analyzing the
factors used in the linear model, with logarithm, and then inverse-transformed.
The factors, which determine the regression model, are selected with the same
method used in the linear regression
analysis. The results are shown in Tables
6 and 7. In Model III, which is a model for
the estimation of strength, X2, X4, X7,
and X5, are entered into the nonlinear regression model in that order. X2, X4, X3,
and X5 are used in Model IV, which estimates nugget diameter. As in the linear
regression model, the determination coefficient increases as the factors are entered. The effects of each factor on the
nonlinear model are shown in Table 8.
YS = 2.6143 X20.119 X40.0461
X70.166 X50.160

(6)

YD = 30.9384 X20.347 X40.0888

(7)
X30.161 X50.778

Table 3 Significance Probability Values of Regression Coefficients for the Stepwise

Regression Analysis and Its Results for Model I

Constant
Model I-1
Model I-2
Model I-3
Model I-4
Model I-5

Significance probability values of entered variables

X1
X5
X4
X3

4.9E-102
2.5E-20
5.02E-19
1.63E-21
9E-58

9.79E-44
8.48E-40
7.83E-37
3.15E-13
5.65E-16

2.52E-03
3.83E-04
7.25E-01

1.95E-04
2.55E-07

0.916
0.926
0.930
0.941
0.941

Table 4 Significance Probability Values of Regression Coefficients for the Stepwise

Regression Analysis and Its Results for Model II

Constant
Model II-1
Model II-2
Model II-3
Model II-4
Model II-5
Model II-6

Significance probability values of entered variables

X1
X7
X3
X4

2.79E-86
2.94E-45
3.00E-19
6.04E-25
1.96E-23
6.92E-25

5.96E-55
3.63E-22
9.96E-22
8.28E-17
1.12E-10
1.7E-10

1.42E-04
8.60E-02
3.92E-05
1.44E-05
9.62E-07

9.70E-02
9.61E-05
2.11E-01

1.21E-07
4.44E-07
6.48E-09

X2

R2

1.92E-02
1.19E-05

0.957
0.964
0.964
0.976
0.978
0.977

Table 5 Standardized Regression Coefficients of the Final Linear Models

Model I-5
Model II-6

X1

X2

X3

X4

X5

X6

X7

X8

X9

X10

0.672
0.478

0.439

0.491

0.239
0.354

0.361

Table 6 Significance Probability Values of Regression Coefficients for the Stepwise

Regression Analysis and Its Results for Model III

Constant
Model III-1
Model III-2
Model III-3
Model III-4

5.54E-86
1.02E-89
2.51E-10
1.18E-03

Significance probability values of entered variables

In X2
In X4
In X7
In X5
1.33E-42
3.55E-40
4.38E-21
2.62E-19

agation method, was used to estimate

weld quality. The formation of the neural
network was determined through the previously mentioned regression analysis results. The X1, X2, X3, X4, X5, and X7 factors applied to linear and nonlinear
models were selected as input variables
for the neural network quality estimation
models such as Model V for strength estimation and Model VI for nugget diameter estimation. Two hidden layers with six
nodes were used to determine the output
values of weld strength and nugget diameter, respectively. This neural network architecture is shown in Fig. 5.
Performances and Evaluation
of the Models

Quality Estimation by Neural Network

Following the regression analysis, the

multilayer artificial neural network, which
was learned through the error back-prop-

4.18E-02
4.26E-05
1.32E-05

R2

Two methods were used to evaluate

the performance of each model, and each
estimation result was represented using
the coefficient of correlation R and root

1.22E-08
3.10E-09
5.57E-10

1.19E-02
2.31E-03

3.17E-02

R2
0.910
0.941
0.946
0.949

mean square error. In the first method,

the original data used in formulating each
regression model and neural network
were used in a model to estimate the performance of the suggested model. In the
second method, multiple regression with
a validation test (Ref. 16) was introduced.
First, a single data point was removed
from the data set of N data points, and the
remaining set of N-1 data points was used
to make an estimation model. The removed single data point was then applied
to this model to obtain a validation result.
This procedure was repeated N-times for
each piece of N data in the data set, obtaining N validation results. Tables 9 and
10 show the results of strength and diameter estimation through such methods.
As expected, the results from the data
used in formulating the model provided
more accurate results, but the results of
the validation test also showed good estimation performance with only a small dif-

WELDING JOURNAL 109-S

WELDING RESEARCH
A

and predicted values are presented

in Table 10, showing values higher
than 0.98. It can
also be seen in the
nugget diameter
estimation model
that the neural
network demonstrated the best
performance.

Conclusions
In order to develop a real-time
quality estimation
system for inprocess resistance
spot welding, the
Fig. 7 Nugget diameter estimation results of the models. A Linear welding variables
model; B nonlinear model; C neural network model.
were monitored at
the primary circuit
of the welding machine and the dyference from the previous results.
namic resistance obtained was used to esIn strength estimation, the neural nettimate weld strength. In order to
work model showed the best perforeffectively eliminate the unwanted inducmance followed by the nonlinear regrestive noise, the primary current and voltsion model and the linear regression
age were measured when the rate of
model. These results can be seen in Fig.
change of current was zero. These were
6. The estimation performance was espeused to calculate the primary circuit dycially good in the neural network model.
namic resistance using the microprocesEvaluating the nugget diameter estimasor of the welding machine timer. The
tion, the linear model showed similar, but
welding machine circuit with a transforslightly better, results than the nonlinear
mer was also analyzed, and the primary
model. Figures 7A and B shows the recircuit dynamic resistance was converted
sults for nugget diameter estimation for
into a dynamic resistance of the welds. To
linear and nonlinear regression models.
consider the relationship between dyAlthough the distribution range of the esnamic resistance and secondary circuit
timated result near a nugget diameter of
dynamic resistance, a correlation analysis
45 mm is up to 1.3 mm, overall estimawas carried out. It was shown the correlation results show good correspondence
tion coefficient was more than 0.92 and
with measured nugget diameter. The cormaximum error was only 7.6392 .
relation coefficients between measured
Therefore, the dynamic resistance pat-

110-S JUNE 2002

tern, which contains information about

the nugget formation mechanism of the
welds, could also be effectively obtained
in the welding machine timer. We have
also observed the effect of dynamic resistance on weld quality through the regression analysis of the factors extracted from
the dynamic resistance pattern. In the linear regression model, weld quality estimation was influenced by beta peak location X1, the rising speed of dynamic
resistance X2, difference between the
minimum and maximum values of dynamic resistance X3, the maximum value
of dynamic resistance variation per cycle
X4, and the final dynamic resistance X7
while it was influenced by the rising speed
of dynamic resistance X2, difference between the minimum and maximum values
of dynamic resistance X3, the maximum
value of dynamic resistance variation per
cycle X4, the maximum value of dynamic
resistance X5, and the final dynamic resistance X7 in the nonlinear model.
As another method of evaluating weld
quality, a neural network with two hidden
layers was used. The factors determined
in the regression analysis were used as the
input while the strength and nugget diameters were selected as the output. According to the performance results of
quality estimation through such models,
the weld strength estimation showed a
maximum error of 0.2061 kN, a correlation coefficient exceeding 0.96, and an
rms error under 0.0836 kN. The neural
network showed the most accurate results in nugget diameter estimation as
well as in the strength estimation. The
performance of the nugget diameter estimation showed a maximum error of
0.66093 mm, correlation coefficient over
0.98, and an rms error under 0.2317 mm.
According to the results described in this
paper, not only can weld quality estimation be performed without the attach-

WELDING RESEARCH
ment of additional monitoring devices in
the secondary circuit of the welding machine but also real-time quality estimation is made possible.

Table 7 Significance Probability Values of Regression Coefficients for the Stepwise

Regression Analysis and Its Results for Model IV

Constant

Acknowledgment

This work was supported by a grant

from the Brain Korea 21 Project and Critical Technology 21 Project of the Ministry
of Science and Technology, Korea.
Reference
1. Gedeon, S. A., Sorensen, C. D., Ulrich,
K. T., and Eagar, T. W. 1987. Measurement of
dynamic electrical and mechanical properties
of resistance spot welds. Welding Journal
66(12): 378-s to 385-s.
2. Johnson, K. I., and Needham, J. C. 1972.
New design of resistance spot welding machine
for quality control. Welding Journal 51(3): 122s to 131-s.
3. Wood, R. T., Bauer, L. W., Bedard, J. F.,
Bernstein, B. M., Czechowski, J., Dandrea, M.
M., and Hogle, R. A. 1985. A closed-loop control systems for three-phase resistance spot
welding. Welding Journal 64(12): 26 to 30.
4. Johnson, K. I. 1973. Quality control resistance welding quality-control techiniques.
Metal Construction and British Welding Journal
5(5): 176181.
5. Haefner, K., Carey, B., Bernstein, B.,
Overton, K., and Dandrea, M. 1991. Realtime adaptive spot welding control. Transactions of the ASME, Journal of Dynamic Systems,
Measurement, and Control 113(3): 104112.
6. Towey, M., and Andrews, S. R. 1968. Instantaneous resistance during spot weld formation as a parameter for an automatic control system. Welding and Metal Fabrication
36(10): 383392.
7. Tsai, C. L., Dai, W. L., Dickinson, D. W.,
and Papritan, J. C. 1991. Analysis and development of a real-time control methodology in
resistance spot welding. Welding Journal
70(12): 339-s to 351-s.
8. Andrews, D. R., and Bhattacharya, S.
1973. Quality control resistance-weld monitoring for production. Metal Construction and
British Welding Journal 5(5): 172175.
9. Patange, S. R., Anjaneyulu, T., and
Reddy, G. P. 1985. Microprocessor-based resistance welding monitor. Welding Journal
64(12): 3338.
10. Studer, F. J. 1939. Contact resistance in
spot welding. Welding Journal 18(10): 374-s to
380-s.
11. Tylecote, R. F. 1941. Spot welding part
II: contact resistance. Welding Journal 20(12):
591-s to 602-s.
12. Roberts, W. L. 1951. Resistance variations during spot welding. Welding Journal
30(11): 1004-s to 1019-s.
13. Dickinson, D. W., Franklin, J. E., and
Stanya, A. 1980. Characterization of spot welding behavior by dynamic electrical parameter

Model IV-1
Model IV-2
Model IV-3
Model IV-4

Significance probability values of entered variables

In X2
In X4
In X3
In X5

3.98E-42
1.71E-48
8.36E-40
3.08E-08

6.92E-47
5.44E-49
1.41E-17
6.49E-19

2.41E-13
2.77E-15
3.15E-13

4.52E-04
1.57E-06

R2

2.37E-06

0.930
0.965
0.971
0.978

Table 8 Standardized Regression Coefficients of the Final Nonlinear Models

Model III-4
Model IV-4

In X1

In X2

In X3

In X4

In X5

0.962
0.867

0.380

0.443
0.264

0.076
0.115

In X6 In X7

In X8 In X9 In X10

0.320

Table 9 Strength Estimation Results as Correlation Coefficients, Root Mean Square Errors,
and Maximum Errors for the Original and Validation Data Set

Linear
Erms

Emax

Nonlinear
Erms
Emax

Original data set

0.9702 0.0751 0.1719 0.9726 0.0721
Validation data set 0.9652 0.0836 0.1967 0.9684 0.0774

Neural
Erms

Emax

0.1838 0.9959 0.0284 0.1002

0.2061 0.9943 0.0290 0.1053

Table 10 Nugget Diameter Estimation Results as Correlation Coefficients, Root Mean Square
Errors, and Maximum Errors for the Original and Validation Data Set

Linear
Erms

Emax

Nonlinear
Erms
Emax

Original data set

0.9889 0.1871 0.6013 0.9877 0.1949
Validation data set 0.9853 0.2040 0.6609 0.9816 0.2317

monitoring. Welding Journal 59(6): 170-s to

176-s.
14. Kaiser, J. G., Dunn, G. J., and Eager, T.
W. 1982. The effect of electrical resistance on
nugget formation during spot welding. Welding
Journal 61(6): 167-s to 174-s.
15. Thornton, P. H., Krause, A. R., and
Davies, R. G. 1996. Contact resistances in spot
welding. Welding Journal 75(12): 402-s to 412-s.
16. Hao, M., Osman, K. A., Boomer, D. R.,
and Newton, C. J. 1996. Developments in characterization of resistance spot welding of aluminum. Welding Journal 75(1): 1s-s to 8-s.
17. Chakalev, A. A., and Vishnyakov, I. V.
1994. Controlling the properties of welds in resistance spot welding. Welding International
8(10): 810813.
18. Livshits, A. G. 1997. Universal quality
assurance method for resistance spot welding
based on dynamic resistance. Welding Journal
76(9): 383s-s to 390-s.
19. Brown, J. D., Rodd, M. G., and
Williams, N. T. 1998. Application of artificial
intelligence techniques to resistance spot
welding. Ironmaking and Steelmaking 25(3):
199204.
20. Dilthey, U., and Dickersbach, J. 1999.
Application of neural networks for quality

Neural
Erms

Emax

0.5790 0.9989 0.0571 0.1767

0.6378 0.9975 0.0583 0.1790

evaluation for resistance spot welds. ISIJ International 39(10): 10611066.

21. Cho, Y., Rhee, S., Shin, H. I., and Bae,
K. M. 1999. Characterization of primary dynamic resistance in resistance spot welding.
Journal of the Korean Welding Society 17(2):
159165.
22. Cho, Y. 2000. A study of dynamic resistance monitoring and intelligent quality estimation for the manufacturing process automation during resistance spot welding,
Ph. D. dissertation. Seoul, Korea, Hanyang
University.
23. Savage, W. F., Nippes, E. F., and Wassell, F. A. 1978. Dynamic contact resistance of
series spot welds. Welding Journal 57(2): 43-s to
50-s.
24. American Welding Society. 2000. Recommended Practice for Resistance Welding.
C1.1M/C1.1.
25. Kimchi, M. 1984. Spot weld properties
when welding with expulsion comparative
study. Welding Journal 63(2): 58-s to 63-s.

WELDING JOURNAL 111-S