RESISTANCE SPOT WELDING PROCESS IDENTIFICATION
AND INITIALIZATION BASED ON SELF-ORGANIZING MAPS
Heli Junno, Perttu Laurinen, Eija Haapalainen, Lauri Tuovinen, Juha Röning
ISG, Department of Electrical and Information Engineering,PO BOX 4500, 90014 University of Oulu, Finland
Dietmar Zettel, Daniel Sampaio, Norbert Link, Michael Peschl
Fachhochschule Karlsruhe, Institut für Innovation und Transfer, Moltkestr. 30, 76133 Karlsruhe, Germany
Keywords: Resistance spot welding, Se
lf-organizing maps, Process identification, Initialization parameters
Abstract: Resistance spot welding is used to join two or m
ore metal objects together, and the technique is in
widespread use in, for example, the automotive and electrical industries. This paper discusses both the
identification of different spot welding processes and the process initialization parameters leading to high-
quality welding joints. In this research, self-organizing maps (SOMs) were used, and optimal features for
the training parameters were sought. According to the results, processes can be classified by specific
features. When introducing new data to trained SOMs, the welding operator can visually identify similar
processes. After process identification, the most similar process is retrieved and a self-organizing map is
trained for this specific process. The initialization parameters leading to successful welds in that process can
thus be identified, which means that the manufacturers can use them to initialize their welding machines.
1 INTRODUCTION
Spot welding is used to join metal objects. It is
widely used, for example, more than 100 million
spots are produced daily in the European vehicle
industry (TWI). This study explains how SOMs
have been used to identify processes and to find the
initialization parameters leading to good results.
In this paper, the aim is to compare the
characte
ristics of a sample measured from a new
process to information gathered from existing
processes, to find a similar process and then to apply
the process parameters leading to high quality joints.
With this approach, the set-up time of new processes
can be significantly reduced.
The research in the field has concentrated on
esti
mating the quality of welding by using neural
networks and regression analysis. The studies have
utilized different features extracted from data. In
many studies the variation of resistance over time
has been used. Neural network and regression
models have been generated based on the dynamic
resistance pattern by, for example, (Aravinthan,
2001) and (Cho, 2002). Studies using other variables
include approaches involving neural networks with
tip force, the number of weld cycles, the weld
current and the upslope current (Ivezic, 1999).
In this paper, the term ‘process’ is used
diffe
rently compared to the previous studies on
process control of spot welding. In our study, the
welding machines, the material and the thickness of
the material can vary in different processes, but the
changes in current, electrode force or electrode wear
are thought to be internal to the process. In other
studies the term ‘process’ has been used to refer to
the internal changes, including differences in
electrode wear (Mintz, 1995). In other application
areas, such as the copper flash smelting process,
SOMs are used in process control but in these
studies, too, the emphasis has been on the internal
variations of processes (Vermasvuori, 2002).
2 DATA DESCRIPTION AND PRE-
PROCESSING
The data used in this study comprise measurements
of welding tests done at Stanzbiegetechnik (SBT).
The data set contained 5 test series (1107 welding
296
Junno H., Laurinen P., Haapalainen E., Tuovinen L., Röning J., Zettel D., Sampaio D., Link N. and Peschl M. (2004).
RESISTANCE SPOT WELDING PROCESS IDENTIFICATION AND INITIALIZATION BASED ON SELF-ORGANIZING MAPS.
In Proceedings of the First International Conference on Informatics in Control, Automation and Robotics, pages 296-299
DOI: 10.5220/0001128802960299
Copyright
c
SciTePress
a)
b)
c)
d)
Figure 1: a) metal objects are joined using resistance spot welding, b) the welded part, c) the welding joint is torn apart in
a destructive test by a quality assurance device, d) tensile strength is shown on the screen of the quality assurance device.
experiments). The materials can be seen in Table 1.
The experiments were done by welding two metal
objects together and after that tearing the objects
apart in a destructive test (Figure 1). Each of the
observations contains measurements of current and
voltage signals recorded during and the tensile
strength of the spot measured after the welding.
The resistance curve, derived from the voltage
and current signals, contains the necessary
information for comparing the processes. Since it
was not feasible to train a SOM with all the data
points of a signal curve, suitable features were
extracted. Every resistance signal was divided into
ten parts of equal length, and their averages were
chosen as features. However, the whole feature set
was not used, but the number of features was
reduced to avoid cross-correlation and to eliminate
overlap. Furthermore, the data were divided into
training and test data sets, which consisted of 80 and
20 percent of the data, respectively.
In this work, the quality criterion was given as
tensile strength, and the distribution of quality
varied. For example, the quality limit for SBT 1 was
160 N, while for SBT 3 it was 550N.
Table 1: The materials used
Test series Base material Contact material
SBT 1, SBT 2 0.18mm Stainless steel Silver nickel
0.5x0.9mm
2
SBT 3, SBT 4 (Time
variation)
0.18mm Nickel-Beryllium
Stainless steel
Silver nickel
0.5x0.9mm
2
SBT 5 0.18mm Nickel-Beryllium Silver nickel
0.7x1.5x2.3mm
3
3 METHOD
The self-organizing map is a method that visualizes
high-dimensional data in a two-dimensional space.
This is done by keeping the topologic and metric
relations of the two-dimensional space as close as
possible to the relations of the initial space.
The SOM is usually formed of neurons on a
regular low-dimensional grid. The neurons are
model vectors m
i
=[m
i1
, m
i2
, … , m
in
], where n is the
dimension of the input space. The training is done
iteratively by choosing a data sample x and finding
the closest model vector m
c
(best-matching unit).
When the best-matching unit is found, it and its
closest neighbors are updated with the equation
m
i
(t+1) = m
i
(t) + α(t)h
ci
(t)(x(t) - m
i
(t)),
where α(t) is the learning rate factor and h
ci
(t) is the
neighborhood kernel centered on the winner unit c.
In this study, the SOM Toolbox, a function package
for Matlab implementing the Self-Organizing Map
algorithm, was used (HUT). For more information
on SOMs, (Kohonen) is recommended.
4 RESULTS
The study was divided into two phases: process
identification and search for initialization
parameters. The features selected were generally the
same in both cases, but the effect of welding time
was ignored when searching for the initialization
parameters, because it was the same inside process.
A strategy for deploying the results is presented
in Figure 2 and a case study of the implementation is
presented in the following chapters.
1) SOM
containing all
the processes
2) Process
identification
4) Return
parameters
leading to good
quality
3) SOM formed
for the states of
an individual
process
Figure 2: Steps for deploying the results.
4.1 Process identification
Figure 3a) shows a trained SOM. The division into 5
regions can be seen from the U-matrix. Now, the
division given by the SOM is researched in more
detail, because it is not yet certain that the different
regions in the map contain information from
different processes.
In Figure 3b), the spots of the 5 test series are
labeled with the numbers 1-5. These labels are
assigned to the map elements representing the curves
belonging to the corresponding cluster. From Figure
3b), it can be seen that all the regions visible in the
U-matrix contain only different processes. However,
the division in the lower part of the map only points
out the differences inside the processes 1 and 2,
which are thought to be similar processes and it is
therefore not considered as an important division.
The identification of different processes with this
RESISTANCE SPOT WELDING PROCESS IDENTIFICATION AND INITIALIZATION BASED ON
SELF-ORGANIZING MAPS
297
a) b) c)
Figure 3: a) The SOM for all the test series from SBT when the means with the least cross-correlations of ten equally long
parts and welding time were used as features. The following abbreviation is used: r = resistance. The numbers 1,…,10 refer
to the means of the respective tenths of the signal. b) The labels for the map shown in a). c) The labels for the testing set.
method seems to be straightforward, but to be more
confident, the test set can be introduced into the
map. The Figure 3c) shows the corresponding labels.
All the welding experiments of the test set are
located in the correct regions. This allows the
operator to identify manually the most similar
process from a database of existing processes on the
basis of welding experiments conducted on the
process of the user.
4.2 Search for initialization
parameters
After identifying the most similar process, the search
for the suitable initialization parameters for that
particular process can be started. In this paper, the
results for process number 3 are presented in greater
detail. The actual SOM trained with the data from
the SBT process number 3 is not shown, but Figure
4 shows the labels related to the map. In the Figure
4a) the quality values of the welding spots shown as
labeled, b) and c) the pre-set values used by the
welding machine.
A welding spot is classified as successful if its
tensile strength is more than 550N. From Figure 4a)
it can be seen that, in the middle right part of the
map, all the welds are of high quality. The low
quality welds are shown as grey areas, while the
high-quality welds are shown as white. The area that
consists only high-quality welds is marked by a
black line
1
. The figures b) and c) show that the
parameters of the welding machine in that area are:
current=3.7 and force=15.8. In fact, all the spots
welded by a combination of those values are there
2
.
With this knowledge, it can be assumed that the
correct parameters to be used with the welding
machine are the current value of 3.7 and the force
value of 15.8. These assumptions can be tested with
the test data. In Figure 4d), the tensile strengths for
the testing data are shown labeled. For every weld
from the testing set, the best-matching unit from
SOM is identified, and the corresponding tensile
strength is used as a label. Because in Figure 4d), the
successful tests are located inside the successful area
that was formed in the training phase, the parameters
can be considered good enough. Therefore, the
parameters can be delivered to the manufacturer,
who can use them to initialize his welding machines.
5 CONCLUSIONS
The study was divided into two different phases:
process identification and search for initialization
parameters. According to the results, the different
processes could be identified on the basis of the
features extracted from the signal curves. Also,
processes close to each other could be differentiated.
Furthermore, after identifying the most similar
process, the initialization parameters leading to high-
quality welds could be found inside that process.
The authors will continue to explore the usage of
more extensive data sets and to address in more
detail the questions that arose during the study.
Answers will be sought to the following questions:
Can the method of setting up the welding parameters
be automated? Can the initialization parameters be
1
There is one node in the upper right part of the
successful area with one unsuccessful test, but a closer
analysis shows it to have been formed when the force had
a value of 14.8.
2
There is also another region where the current value of
3.7 and the force value of 15.8 co-occur. However, in that
ICINCO 2004 - INTELLIGENT CONTROL SYSTEMS AND OPTIMIZATION
298
region the welding help parameter, which does not show
in the maps, was not applied.
a) b) c
)
d)
Figure 4: The labels for the map trained for SBT3, a) tensile strength, b) pre-set current values, c) pre-set force values. d)
Tensile strength labels for the testing set.
delivered to the manufacturer if there are also
unsuccessful welds in the successful area? And is it
possible to differentiate between processes without
taking the effect of time into account?
ACKNOWLEDGEMENTS
We would like to express our gratitude to our
colleagues at Fachochschule Karlsruhe, Institut für
Innovation und Transfer, in Stanzbiegetechnik, in
Harms + Wende GmbH & Co.KG and in Technax
for providing the data set and the expertise needed at
the different steps of the research project.
Furthermore, this study has been carried out with
financial support from the Commission of the
European Communities, specific RTD programme
“Competitive and Sustainable Growth”, G1ST-CT-
2002-50245, “SIOUX” (Intelligent System for
Dynamic Online Quality Control of Spot Welding
Processes for Cross(X)-Sectoral Applications”). It
does not necessarily reflect the views of the
Commission and in no way anticipates the
Commission’s future policy in this area.
REFERENCES
Aravinthan, A., Sivayoganathan, K., Al-Dabass, D.,
Balendran, V., 2001. A neural network system for spot
weld strength prediction, UKSIM2001, Conf. Proc. of
the UK Simulation Society, Pages: 156-160.
Cho, Y., Rhee, S., 2002. Primary Circuit Dynamic
Resistance Monitoring and its Application on Quality
Estimation during Resistance Spot Welding, Welding
Researcher, Pages 104-111.
Fachochschule Karlsruhe,
http://www.fh-karlsruhe.de/, the
homepage of the university, referenced 10.2.2004.
Harms+Wende, Web site of Harms&Wende
http://www.harms-wende.de/. Referenced: 10.2.2004.
HUT, Helsinki University of Technology,
http://www.cis.hut.fi/projects/somtoolbox/, referenced
10.2.2004.
Ivezic, N., Alien, J. D., Jr., Zacharia, T., 1999. Neural
network-based resistance spot welding control and
quality prediction, Intelligent Processing and
Manufacturing of Materials, IPMM '99. Proceedings
of the Second International Conference on, Pages: 989
–994.
Kohonen, T., 1997. Self-organizing maps, Springer-
Verlag, 2
nd
edition.
Mintz, D., Wen, J.T., 1995. Process monitoring and
control for robotic resistive welding, Proceedings of
the 4th IEEE Conference on Control Applications,
Pages: 1126 –1127
Stanzbiegetechnik, SBT,
http://www.stanzbiegetechnik.at,
referenced 10.2.2004.
Technax,
http://www.technaxindustrie.com/, referenced
10.2.2004.
TWI World Centre for Materials Joining Technology,
information available at their homepage:
http://www.twi.co.uk/j32k/protected/band_3/kssaw001
.html
, referenced 10.2.2004.
Vermasvuori, M., Enden, P., Haavisto, S., Jamsa-Jounela,
S.-L., 2002. The use of Kohonen self-organizing maps
in process monitoring, Proceedings of IEEE
International Symposium Intelligent Systems, Pages:
2- 7 vol.3.
RESISTANCE SPOT WELDING PROCESS IDENTIFICATION AND INITIALIZATION BASED ON
SELF-ORGANIZING MAPS
299