Int J Adv Manuf Technol (2009) 44:512519

DOI 10.1007/s00170-008-1847-0

ORIGINAL ARTICLE

Weld formation control for arc welding robot

Shen Hongyuan & Huang Xixia & Lin Tao &
Chen Shanben

Received: 11 June 2008 / Accepted: 10 November 2008 / Published online: 2 December 2008
# Springer-Verlag London Limited 2008

Abstract The control of weld formation in gas tungsten arc

welding is required for teach and playback robot to suit
the gap variation in the multi-pass welding process. This
paper presents a robotic system based on the real-time
vision measurement. A close-loop control system is
developed for weld formation control in backing welding.
A novel prediction model of reinforcement is investigated
as the feedback element. At last, the feasibility of proposed
system was validated in the experiments.
Keywords Weld formation . Reinforcement .
Prediction model . Welding robot . Vision

1 Introduction
At present, welding robots have been extensively applied in
manufacturing. They can work all day and all night to
reduce the labor. But welding robots primarily work in
teach and playback mode in which mode they have no
enough flexibility, especially during welding process.
Because welding usually encounters many variables such
as the errors of pre-machining, fitting of work-piece, and
in-process thermal distortions, which will result in changes
of the gap size, the welding procedures must be adjusted
according to these changes. But robots in teach and
playback mode have no such functions and usually weld a
weldment with lots of flaws, even burning through or
incomplete penetration. So, it is necessary to develop an

S. Hongyuan (*) : H. Xixia : L. Tao : C. Shanben

Institute of Welding Technology, Shanghai Jiao Tong University,
Shanghai 200240, China
e-mail: [email protected]

automatic control system for those welding robots working

in teach and playback mode.
Generally, the sound welding procedures of the productions
were obtained after lots of experiments. For example, a
seasoned welder can work well for all kinds of work-pieces.
So, the welding procedures are not the key problem in
automatic welding. The most important is to exactly check
the changes of welding conditions and to research the control
rules of the welding procedures and these changes.
In this paper, a visual sensor device is used to check
these changes. Because the visual sensing has many
advantages, for example, it does not make contact with
the weld pool, and the information is very abundant, and
the visual sensing has become the focus in the region of
studying on sensing technology of the welding robot. Most
welding-related applications are used in seam tracking and
checking the weld pool [18]. Meta is a leading manufacturer
of laser vision system for welding applications worldwide
[8], and its vision systems have been extensively applied in
the manufacture.
A number of significant achievements have been made
in the research field of autonomous welding robot by means
of visual sensing [912]. In some studies, the camera was
directly used to view the weld pool and its vicinity to obtain
control information such as the size, the position of the
weld pool, and the width of the gap [1314]. Zhou et al. [4]
presented a system based on the real-time vision measurement
and control of the upper surface weld pool size. Zhao [14] and
Wang [15] viewed the topface and underside geometrical
properties of the weld pool to construct a prediction model to
control the underside width of the pool at next step. Then,
Zhang [16] and Fan [17] added the measurement of the seam
gap to improve the prediction model. Chen [18] proposed an
intelligent technique for fulfilling quality control of bead-onplate welding.

Int J Adv Manuf Technol (2009) 44:512519

In fact, most of the work-pieces are welded with welding

backing, and the underside of weld pool is not visible.
During continuous gas tungsten arc welding (GTAW),
nothing but the topface width of the pool and the seam
gap can be measured [19]. So, the research on the underside
width of the pool as measurement signal is difficult to put
into practice.
In our research, the visual sensor viewed the weld pool
in topside front of it through the optical filters. The
computer extracted the topside width of weld pool and
seam gap. Then, a reinforcement prediction model was built
for a close-loop control system to adjust the welding current
and wire feed rate. At last, this system was validated in the
experiments, and a favorite result was obtained.

513

welding robot is a six-axis industrial robot. It works in

teach and playback mode in Shanghai Spaceflight Precision
Machinery Research Institute, China. The photo of the
welding robot is shown in Fig. 2. The visual sensor is fixed
on the end joint of the robot and moves with the robot. The
CCD camera receives the weld pool image through doublelayer filter system after two reflections. Then, the computer
captures the images by the image capturing board, extracts
the width of weld pool and seam gap in real time using the
image processing algorithms, and calculates the adjustment
of the welding current and wire feed rate. Finally, the
calculated analogue control voltages are sent to the power
source and the wire feeder controller.

3 Control method
2 System description
The welding robot system developed in this study consists
of a teach and playback robot, a visual sensor composed of
a charge coupled device (CCD) camera and the optical
filters, a welding power source, a control computer, and an
interface box. According to the spectral character of the
aluminum alloy welding, a wideband filter and two dimmer
glasses are used for the sensor. The transmission of the
wideband is 590 to 710 nm. The attenuation of the two
dimmer glasses is 99% and 70%, respectively. Figure 1
shows a schematic diagram of the real-time weld pool
control system of welding robot with computer vision. The

Fig. 1 The schematic diagram

of the real-time weld pool
control system of welding robot
with computer vision

It is most important for welding process to keep the

stability of welding. So, the parameters should be adjusted
as little as possible during welding process. Reference [20]
has proposed a method of welding process control. The
welding parameters should be adjusted in sectional type
with the different seam gap, as shown in Table 1. The
adjusted rule of wire feed rate is not only simply based on
the groove geometry proposed in Ref. [20] but is also
variable with the time. So, an advanced control method of
wire feed rate will be proposed in this paper based on Ref.
[20].
Multi-pass welding is a kind of general welding methods
for medium plate in manufacturing. Usually, the first pass,

514

Int J Adv Manuf Technol (2009) 44:512519

4.1 Prediction model principle

The prediction model is based on the volume conservation
of fed wire volume and filled wire volume. In order to build
the mode easily, some supposed conditions were provided:
1. All the metal volume fed by wire feeder fill the seam.
2. The work-piece is close-fitted with the welding
backing.
3. The root reinforcement is the same size as the groove in
the welding backing, and its value is constant during
stable welding process.
Therefore, in some period, the relationship of the metal
volume can be expressed as:

t2
t1

Z
Sr t
Vw t dt

t2

t1


Sw Vf t  Sj t Vw t dt

I t2  I0  hI
Fig. 2 The welding robot system

named backing welding, is the most prone to defects, which

affects the joint quality. There are two focuses on backing
welding: one is welding penetration; the other is uniform
welding reinforcement. In order to solve the penetration,
different controllers are chosen according to different gap
value (shown in Table 1). Welding current fuzzy controller
has been introduced in Ref. [20]. Wire feed rate controller
will be improved in this paper, which will solve the uniform
welding reinforcement problem. Figure 3 is the block
diagram of the weld process controller.

4 Prediction model of reinforcement

As it is quite difficult to accurately measure reinforcement
on line, a prediction model is constructed to estimate the
reinforcement value in real time instead of the real value,
which outputs the feedback value in Fig. 3.

Where Sr(t) is real sectional area of welded joint at the time

of t as shown in Fig. 4, Sw is sectional area of the wire, Vf is
wire feed rate at the time of t, Sj(t) is sectional area of
welded joint when the reinforcement is zero at the time of t
as shown in Fig. 5, Vw(t) is welding speed at the time of
t, I(t) is welding current at the time of t, I0 is original
welding current, and I is impact factor of welding current
to the weld pool, which depends on welding materials and
welding procedure.
As arc force enlarges with welding current increasing,
arc crater size also becomes bigger, which will result in
higher reinforcement, so welding current is considered as
an impact variable. But, so far, there are few researches
on the relationship between welding parameters and arc
crater size. Since welding current is variable in a small
range in this research, it is supposed that their relationship
is linear.
Then, Eq. 1 can be expressed in computer as follows:
n2
X
in1

Sr i
Vw i
$t

n2
X



Sw Vf i  Sj iVw i
$t
in1

I n2  I0  hI
Table 1 The control rules for different gaps [20]
Gap (mm)

Controlled variables

Controller

00.3
0.31.2
1.22
>2

vf
vf and I

Without control
PID for vf
PID for vf and fuzzy for I
Without control

2
Where t is the cycle of capturing sample points and also
the control cycle.
In the background of this paper, welding speed is
assumed as constant, so Vw(i) should be Vw. In order to
play down the complexity of solving equation and improve

Int J Adv Manuf Technol (2009) 44:512519

515

Fig. 3 Block diagram of the

weld process controller

calculation speed, the notching curve of welded seam

surface is supposed as triangle shape, instead of arc shape.
In Fig. 6a, the blue line is the real cure, and the red line is
supposed. So, Sr(i) can be presented as:

According to Eqs. 25,

n2 
X
1
in1

1
Sr i
Wp i
Hr i
2

Where Wp(i) is the width of weld pool at the time of i,

and Hr i is real reinforcement after welding at the time
of i.
Sw and Sj(i) are presented respectively as:

Sw p
f2 4



Wp i
Hr i
Vw
$t

n2

X
in1




p
f2
Vf i=4  H
g i tga
H22 Sl
Vw

$t I n2  I0  hI

6
As welding process is dynamic with the flowing metal in
weld pool, the reinforcement value Hr t is unknown when
the current liquid metal is freezing. In fact, in the current
time, the liquid metal at the past time of m is just freezing.
The reinforcement should be expressed as Hr T (shown in
Fig. 6b). The relationship between T and t is expressed as:

Where is the diameter of the wire.

T tm
Sj i H
gi tga

H22

Sl

m 1; 2; 3:::

Where Sl is the sectional area of root reinforcement.

Usually, the arc stirs the weld pool fully, so the

frozen reinforcement is supposed as the average rein-

Fig. 4 Sketch of Sr(t)

Fig. 5 Sketch of Sj(t)

516

Int J Adv Manuf Technol (2009) 44:512519

Fig. 8 The work-pieces with two segments of 2-mm gap

Fig. 6 Schematic plan of groove in butt plate; a the cross section, b

lengthwise direction

forcement of the whole liquid metal under the arc.

Then,
t 
X
1
iT



Wp i
Hr T
Vw
$t

t 
X
1
iT



p
f2
Vf i  H
g i tga
H22 Sl
Vw

Equation 8 expresses that the predicted reinforcement is

not current according to the current welding parameters, but
the one in the past time of m, which does not break the
definition of real-time predicting, reversely, correctly
accords with the characteristic of liquid metal fluid during
welding process. The reinforcement value at the past time
of m is affected by the seam gap, wire feed rate, and
welding current from current time to the past time of
(m1). So, the freezing prediction reinforcement can be
expressed as:

$t I t  I0  hI

t

1
P

H r T

itm




 p  f2  Vf i  H  gi tga  H22 Sl  Vw  $t I t  I0  hI
t

1
P
itm

 Wp i  Vw  $t

Where Hr(T) is the prediction reinforcement.

Table 2 Aluminum alloy welding parameters

Fig. 7 Experiment data for constructing model

Parameter

Value

Process
Material

Square-wave AC GTAW at 60 Hz
6-mm-thick LD10 aluminum alloy
plate with a Y-groove of 80
2 mm
1Cr18Ni9Ti
240 A
24.7 mm s1
2.67 mm s1
Ar

Thickness of root face

Backing bar
Welding current
Wire feed rate
Welding speed
Shielding gas

Int J Adv Manuf Technol (2009) 44:512519

517

n  h
i @H j
@Jmin X
r

2 Hr j  Hr j 
0
@h
@hI
I
j1

11

According to Eqs. 10 and 11,

Fig. 9 Welding results for the work-pieces with two segments of 1.5-mm
gaps with constant welding procedure

@Hr j
ij
jm
@hI
P
1

4.2 Prediction model realization

ij

In Eq. 9, there are two unknown parameters, I and m,

which can be worked out with least-squares procedure, as
shown in the following:
Jmin

n h
X

Hr j  Hr j

jm
P

i2

I i  I0 
$t


2
Wp i
Vw
$t

hI

12

Then,

10

j1

9
>
14
p
f2
Vf iH
gitga
H22 Sl
Vw 
I jmI0  >
=
ij
I jmI0 


H




2
r
jm
P
jm
>
P
j1 >
>
>
1
;
: 12
Wp i
Vw 

Wp i
Vw 

2
ij
ij
hI
92
8
>
>
=
<
n
P
8
>
>
n <
P

jm
P

j1 >
:

I jmI0 

jm
P
ij

;
12
Wp i
Vw  >

In Eq. 13, m is an unknown parameter. But, it can be

calculated by welding speed, control cycle, and the distance of
the center of weld pool to the critical edge of liquid and solid.
Although it is not sure, it is limited in the range of [928],
during which range the values of m and I should be chosen
when Jmin is minimal. This process was run in Matlab Tools.
The data to be used to build the model are from qualified
weldment the reinforcement of which is in the range of

Fig. 10 Comparison of
prediction reinforcement and
off-line test reinforcement

13

(0.5, 0) mm (this range is a standard of first-order welding

seam). The detailed experiment data is shown in Fig. 7.
Therefore, m and I are worked out.
(

m 16
hI 2:9112

14

518

Int J Adv Manuf Technol (2009) 44:512519

Fig. 13 The curves of the gap, welding current, and wire feed rate for
the testing work-piece
Fig. 11 The weldment with two segments of 2-mm gap without
control: a topface, b underside

Then,
j16
P

Hr T

ij

0:402
Vf i  3:192
gi 12:04I j 16  250
2:9112

j16
P

ij

0:266
Wp i

15

5 Experiment and results

After the prediction model of reinforcement was constructed,
a butt weld work-piece with two segments of 1.5-mm gap has
been designed to validate the prediction model, as shown in
Fig. 8. Table 2 shows the constant welding parameters for
aluminum alloy plates. In order to prevent the work-piece

from being burned through in such a constant procedure, the

gap is preset to be 1.5 mm and not 2 mm. The welding result
is shown in Fig. 9. The reinforcement was predicted with the
proposed model. The reinforcement comparison result of
prediction value and off-line test value is shown in Fig. 10.
The root-mean-square error is 0.265 mm, and the maximum
absolute error is 0.5 mm. As the limit of detection tools, it is
difficult to reduce the error.
In this research, a classical PID terms as shown in Eq. 16
is used in the wire feed rate controller.

$vf k kp $ek ki ek kd $ek  $ek  1

16

Fig. 12 The weldment with two segments of 2-mm gap with control:
a topface, b underside

Where vf(k) is the adjustment of the wire feed rate, e(k)

is the change of the reinforcement, kp is the proportional
gain, ki is the integral gain, and kd is the derivative gain.
The kp, ki, and kd can be ensured according to lots of
experiments, and their values are 4.27, 0.85, and 1.26,
respectively.
In Fig. 3, 0.2 mm is the expectation value of the
reinforcement change in the close-loop control, instead of
zero as a general expectation value. If the expectation value
is zero, the real reinforcement in most seams would be
convex due to the error of the prediction model. The worker

Int J Adv Manuf Technol (2009) 44:512519

must smooth the convex joint before cosmetic welding,

which will cost plenty of time.
At last, the proposed process control shown in Fig. 3 was
validated in the work-pieces with two segments of 2-mm gap
that is of the same shape as Fig. 8. The original welding
procedure shown in Table 2 was used to weld the seam
without process control, and the result is shown in Fig. 11.
When the gap first changes from zero to 2 mm, the heat input
is too large, and there is not enough wire fed, so the seam is
burned through. At the second gap slot, the seam is not
burned through as the gap reduces due to thermal distortion,
but the reinforcement is sunken at least 2 mm, which must
increase the possibility of flaws during cosmetic welding.
Compared with Fig. 11, Fig. 12 shows the results using
the process control as described above. The curves of the
gap, welding current, and wire feed rate during the welding
process are shown in Fig. 13. The wire feed rate is all along
controlled with the change in the seam gap. But, the welding
current is only adjusted when the gap is more than 1.2 mm.

6 Conclusions
This paper has researched a welding process control for
teach and playback robot with computer vision. A novel
reinforcement model was proposed for backing welding
with welding backing. The model was demonstrated to be
feasible. And, it was used in close-loop control as feedback
element. Process control was implemented in the form of a
PID method to adjust the wire feed rate. The experiment
result showed a favorite weldment with the proposed
technique. The whole system researched was successfully
applied in the teach and playback robot welding process to
help the robot adjust the welding current and wire feed rate.
Future work should improve the prediction model of
reinforcement through analyzing more impact factors and
enhance its precision.
Acknowledgments This work is supported by the National Natural
Science Foundation of China under Grant No. 50575144.
The authors wish to thank the anonymous reviewers for their
valuable comments on the earlier draft of this paper.

References
1. Xu PQ, Xu GX, Tang XH, Yao S (2008) A visual seam tracking
system for robotic arc welding. Int J Adv Manuf Technol 37(12):70
75. doi:10.1007/s00170-007-0939-6

519
2. Moon HS, Ko SH Kim JC (2008) Automatic seam tracking in
pipeline welding with narrow groove. Int J Adv Manuf Technol
April 3, 2008. doi:10.1007/s00170-008-1474-9
3. Jeremy SS, Balfour C (2005) Real-time top-face vision based
control of weld pool size. J Ind Robot 32(4):334341.
doi:10.1108/01439910510600209
4. Zhou L, Lin T, Chen SB (2006) Autonomous acquisition of seam
coordinates for arc welding robot based on visual servo. J Intell
Robot Syst 47:239256. doi:10.1007/s10846-006-9078-9
5. http://www.scout-sensor.com/
6. Lee SK, Chang WS, Yoo WS, Na SJ (2007) A study on a vision
sensor based laser welding system for bellows. J Manuf Syst 19
(4):249255
7. Lee SK, Na SJ (2002) A study on automatic seam tracking in
pulsed laser edge welding by using a vision sensor without an
auxiliary light source. J Manuf Syst 21(4):302315
8. Meta Vision Systems 2008. http://www.meta-mvs.com
9. Luo H, Chen XQ (2005) Laser visual sensing for seam tracking in
robotic arc welding of titanium alloys. Int J Adv Manuf Technol
26(910):10121017. doi:10.1007/s00170-004-2062-2
10. Zhu ZY, Lin T, Piao YJ, Chen SB (2005) Recognition of the initial
position of weld based on the image pattern match technology for
welding robot. Int J Adv Manuf Technol 26(78):784788.
doi:10.1007/s00170-003-2053-8
11. Bae KY, Lee TH, Ahn KC (2002) An optical sensing system for
seam tracking and weld pool control in gas metal arc welding of
steel pipe. J Mater Process Technol 120(2):458466. doi:10.1016/
S0924-0136(01)01216-X
12. Ge JG, Zhu ZQ, He DF, Chen LG (2005) A vision-based
algorithm for seam detection in a PAW process for large-diameter
stainless steel pipes. Int J Adv Manuf Technol 26(10):10061012.
doi:10.1007/s00170-004-2070-2
13. Yamane S, Kaneko Y, Kitahara N, Ohshima K, Yamamoto M
(1993) Neural network and fuzzy control of weld pool with
welding robot. Ind Appl Soc Annu Meet: Conf Rec IEEE 3:2175
2180
14. Zhao DB, Chen SB, Wu L, Dai M, Chen Q (2001) Intelligent
control for the shape of the weld pool in pulsed GTAW with filler
metal. Weld J 80(11):253313
15. Wang JJ, Lin T, Chen SB (2005) Obtaining weld pool vision
information during aluminium alloy TIG welding. Int J Adv
Manuf Technol 26(3):219228. doi:10.1007/s00170-003-1548-7
16. Zhang GJ, Chen SB, Wu L (2005) Intelligent control of pulsed
GTAW with filler metal. Weld J 84(1):917
17. Fan CJ, Chen SB, Lin T (2006) Visual sensing and image
processing in aluminum alloy welding: 2006 International
Conference on Robotic Welding, Intelligence and Automation,
Lecture Note in Control and Information Sciences. Springer,
Germany, 362, 275280
18. Chen SB, Lou YJ, Wu L, Zhao DB (2000) Intelligent methodology
for sensing, modeling and control of pulsed GTAW: part 1beadon-plate welding. Weld J 79(6):151164
19. Shen HY, Lin T, Chen SB (2006) A study on vision-based realtime seam tracking in robotic arc welding: 2006 International
Conference on Robotic Welding, Intelligence and Automation,
Lecture Note in Control and Information Sciences. Springer,
Germany, 362:311318
20. Shen HY, Ma HB, Lin T, Chen SB (2007) Research on weld pool
control of welding robot with computer vision. Ind Rob 34
(6):467475. doi:10.1108/01439910710832066