Dynamic Modeling and Adaptive

Control of the Gas Metal Arc

Welding Process
Control of the welding process is a very important step in welding automation.
Jae-Bok Song Since the welding process is complex and highly nonlinear, it is very difficult to
Assistant Professor, accurately model the process for real-time control. In this research, a discrete-
Department of Mechanical time transferfunction matrix modelfor gas metal arc welding process is proposed.
Engineering, This empirical model takes the common dynamics for each output and inherent
Korea University,
Seoul, Korea
process and measurement delays into account. Although this linearized model
is valid only around the operating point of interest, the adaptation mechanism
employed in the control system render this model useful over a wide operating
David E. Hardt range. Since welding is inherently a nonlinear and multi-input, multi-output
Professor,
process, a multivariable adaptive control system is used for high performance.
Laboratory for Manufacturing and
Productivity, The process outputs considered are weld bead width and depth, and the process
Massachusetts Institute of inputs are chosen as the travel speed of the torch and the heat input. A one-
Technology, step-ahead {or deadbeat) adaptive control algorithm combined with a recursive
Cambridge, MA 02139
least-squares methods for on-line parameter estimation is implemented in order
to achieve the desired weld bead geometries. Control weighting factors are used
to maintain the stability and reduce excessive control effort. Some guidelines for
the control design are also suggested. Command following and disturbance
rejection properties of the adaptive control system for both SISO and MIMO
cases are investigated by simulation and experiment. Although a truly independent
control of the outputs is difficult to implement because of a strong output coupling
inherent in the process, a control system for simultaneous control of bead width
and depth was successfully implemented.

1 Introduction
Welding processes have been developed into an auto-
mated operation over the past decades. In the beginning
stages of automation, the welding machines were success-
fully mechanized. Control systems for the welding torch
motion (e.g., joint tracking and robotic welding) are now
commercially available. However, control of the welding
process itself is also very important so that full automation
of the process and, possibly, unmanned operation may be
achieved. Unfortunately, process control systems have not
been fully developed for many reasons such as complexity
of the welding process and lack of reliable sensors; therefore,
much research is underway in order to solve these problems.
Gas Metal Arc Welding (GMAW) is a complex, multi-
Contributed by the Dynamic Systems and Control Division for publication in
the JOURNAL OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL. Manuscript
received by the DSCD August 26, 1993; revised manuscript received July 15,
1993. Associate Technical Editor: J. Stein.
energy domain process that is essential to many types of
manufacturing. Weld quality features such as final metal-
lurgy and joint mechanics are typically not measurable on-
line for control; thus, some indirect way of controlling the
weld quality is necessary. A comprehensive approach to in-
process control of welding includes both geometric features
of the bead (such as the cross section features width, depth
and height) and thermal characteristics (such as the heat-
affected zone width and cooling rate). The definitions of
these features are illustrated in Fig. 1. Geometric features
can be related to the basic mechanics of the joint, while the
thermal features can indicate the final microstructure of the
material. A general multi-input, multi-output approach to
control is illustrated in Fig. 2.
Nonlinearity of welding has been confirmed by several
process modeling investigations. For Gas Tungsten Arc
Welding (GTAW) process, Hardt et al. (1985) developed a

Journal of Dynamic Systems, Measurement, and Control SEPTEMBER 1994, Vol. 116/405

Copyright © 1994 by ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab
 W h
L
:
"" J I Weld Bead developed by Vroman and Brandt (1976). A video line-scan
Heat Affected r ~ H >L-
Zone
\ _ Height camera was used to measure the pool width and PI controller
was designed heuristically since no dynamic modeling of
Depth the welding process was attempted. Nomura et al. (1980)
Fi. I *
Pene V
' /
developed a control system for a full penetration submerged
arc welding. They measured the light emission from the
back side of the weld and succeeded in maintaining a con-
Fig. 1 Definitions of bead width and depth
stant back bead width with a integral controller by modulat-
ing the current. Richardson et al. (1982) developed a control
. Desired system for regulating weld pool width of a gas tungsten arc
Attributes
welding using current as a control input. A vision based
Multi-
f Welding GMAW system was used for width measurement. A proportional
1—> variable
Controller
Equipment Process controller was implemented, but no dynamic modeling of
the process was attempted. The performance of the above
Bead Depth
simple control systems was not very good because of sensor
Depth Measurement
Estimator limitations and their inability to cope with nonlinearity and
time-varying nature of the welding process.
Bead Width Image Measurement One of the first adaptive control applications to welding
Analysis
Bead Height was demonstrated by Dornfeld et al. (1982). A Model Refer-
ence Adaptive Control (MRAC) scheme was used to regulate
HA2 HAZand Measurement
Cooling Rate the back bead temperature of GMAW using travel speed as
Cpooling Rate Estimator an input. Recently, Suzuki et al. (1991) developed adaptive
control systems for the full penetration GTAW problem
Fig. 2 Schematic for multivariate control of GMAW outlined in Hardt et al. (1985). In this case, both MRAC
and Self-Tuning Control (STC) were shown to be very effec-
tive in responding to both uncertain nominal parameters
simple non-stationary first-order model relating input weld- and severe parameter disturbances. Doumanidis and Hardt
ing current to output bead width. It was shown both analyti- (1991) addressed the MIMO case of thermal control by using
cally and experimentally that the process parameters were a deadbeat adaptive algorithm with parameter adjustment
dependent both on the operating point of the process and on based on a projection algorithm for parameter identification.
the boundary conditions (e.g., preheat temperature, material This performed well for this rather slow system, but it re-
thickness). In a later investigation concentrating on thermal quired the addition of an external PI loop to reduce the
properties, Doumanidis and Hardt (1989) developed a trans- control effort required. Finally, Hale and Hardt (1990b) dealt
fer function matrix model for GMAW that related wire with the nonlinearities of GMAW geometry control problem
feedrate and torch travel speed to heat affected zone width
using a scheduled gain approach. While this was adequate for
and cooling rate. Again the process was found to be nonlinear
controlled situations, it quickly showed limited performance
and operating point dependent. Finally, most relevant to this
when pushed off nominal conditions, and had poor distur-
paper, Hardt and Hale (1990a) developed a transfer function
bance rejection properties.
matrix model for GMAW relating wire feedrate and travel
speed to bead geometries, bead width, and height. Recently, Andersen et al. (1990) and Smartt et al. (1992)
Recently, some researchers (Andersen et al., 1990) mod- applied artificial neural networks (ANN) to control of arc
eled the welding process based on the artificial neural net- welding processes. Although the ANN has the ability to
work (ANN). If enough training data are given, the ANN handle the multivariable and nonlinear nature of the welding
model can capture the process characteristics including non- process, this scheme requires large I/O data over a wide
linearities and parameter couplings over large operating operating range to tune the controller in advance. Also,
ranges. In addition, the ANN model can be refined with welding control systems using fuzzy linguistic control were
addition of new training data. However, the ANN model developed by Langari and Tomizuka (1988) and Boo and
requires a large amount of I/O data to accurately train the Cho (1991). Both neutral network and fuzzy control ap-
network and cannot describe the physics of the process. proaches assume that arc welding processes are so complex
Even though much research has been done for the model- and nonlinear that they cannot be adequately modeled by
ing and analysis of welding processes, research on real-time conventional modeling techniques.
welding control is a relatively recent development, mainly In this research, the transfer function matrix model for
because of lack of reliable sensors for feedback control. GMAW process developed by Hale and Hardt (1990a) is
Most of the early research was based on single-inputs, single- further extended to include more accurate dynamic charac-
output nonadaptive control schemes. Recently, with the de- teristics. This empirical model takes common dynamics for
velopment of robust control algorithm, computers and sen- each output and inherent time delays of the process into
sors, more sophisticated control schemes (e.g., adaptive con- account. Open-loop tests with a pseudo-random binary se-
trol, artificial neural network, fuzzy control, etc.) have been quence input are conducted to evaluate the proposed model.
employed to improve performance of the welding control In order to consider nonlinear and multivariable nature of the
system. process, an adaptive control system with on-line parameter
One of the first weld pool width control systems was estimation is developed and tested for both SISO and MIMO

406 /Vol. 116, SEPTEMBER 1994 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab

 Pool Image
Welding .-
Depth
Table
Temperatures
Fig. 3 Schematic for welding setup and measurement devices
cases. Some guidelines for designing the adaptive control
system are also suggested.
2 Process Modeling for Control
Among many possible outputs shown in Fig. 2, bead width
and depth, are chosen as process outputs in this research.
Because the process inputs must be able to regulate the
chosen outputs, wire feedrate (f) and travel speed of the
torch (v) (i.e., torch velocity) are selected as inputs. Even
though these two inputs have similar effects on the outputs
and cause strong coupling of I/O pairs, they are considered
the best choices under the current configuration of GMAW
setup. Note that the wire feedrate regulates the welding
current, and thus the heat input into the weld.
In order to investigate the dynamics of the process, a
series of open-loop welding experiments were performed
using the welding setup in Fig. 3. The welding table with
the workpiece clamped to it was set up to move in the
direction opposite to the welding direction, while the torch
and measurement devices remain stationary. The experi-
ments were conducted based on a bead-on-plate welding.
The primary metals were low carbon steels which are 6.35
mm ('/j in) thick. The top surface of the work was ground
to remove oxides, while the back side was cleaned chemi-
cally in order to improve temperature measurement. An Ar
+ 2 percent oxygen gas mixture was used as a shielding gas
for this experiment.
The weld bead depth, which is the key geometric attribute
of a major class of welds, is very difficult to directly measure,
but a robust method to estimate the depth using temperature
measurement was developed (Song and Hardt, 1992). The
estimator employs a 3-D analytical heat conduction relation-
ship with a Gaussian distributed heat source. The algorithm is
designed such that the temperature measurement, nonlinear
parameter identification based on a combined Gauss-Newton
and Steepest Descent method, and estimation of the depth
are performed in real-time. Destructive measurement of the
weld profile indicates that the method can provide depth
estimates of acceptable accuracy and speed for real-time
control. In addition to the depth estimator, a weld pool width
estimator using a CCD camera and image analysis was also
developed for real-time control (Song, 1992).
Extensive open-loop step tests have shown that the
GMAW process can be adequately modeled by the second-
order dynamics (Hale and Hardt, 1990a). Since the process
Journal of Dynamic Systems, Measurement, and Control
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab
is nonlinear as shown above, a locally linearized process
model around the operating point is obtained. Based on
these observations, the following discrete-time second-order
transfer function matrix (TFM) model has been developed:
>i(z)\
J2U).
(bn + *i2Z )z » (bl3 + bHz'l)z~ai2-i
\+auZX + a\2z2 \+anzx + a\2Z2 u\{z)
(b2l + &22Z )z 21 (b23 + b24z~i\„-rfi
l
)z-"22 U2(z)J
1 + a2\Z~l +a22z2 1 + a2\z x
+ a22z .
(1)
where yi(z) and y2(z) are the width and depth, u\(z) and u2(z)
are the wire feedrate (/) and inverse travel speed or inverse
velocity (v_1), and dy is the time delay between the ith output
and j'th input. Note that the inverse travel speed is used
instead of the travel speed in Eq. (1). This is because the
leading coefficients b 13 and b23 corresponding to u2(z) will
have positive sign with the inverse travel speed as input;
both width and depth increase as the inverse travel speed
increases (i.e., the travel speed decreases). This convention
will turn out to be very useful in controller design later.
It is noted that each output has been modeled to have
common denominator dynamics; two process inputs have
similar effects on each output from a physical point of view
because both inputs are regulating the outputs through the
amount of the heat input to the weldment. This common
dynamics also simplifies the subsequent analysis and design;
if each I/O pair had different second-order dynamics, the
resulting transfer function matrix would include a fourth-
order polynomial. This idea of common dynamics was exam-
ined in a series of open-loop welding experiments using a
pseudo-random binary sequence (PRBS) input. The PRBS's
for the feedrate and travel speed were designed such that
the minimum pulse duration is chosen as twice the sampling
time so as to put more energy on the low frequency range
and the levels of pulses are chosen as a compromise between
the linearity consideration and a high signal-to-noise ratio.
These signals combined with a recursive least-squares
method were used to identify the process parameters for
both SISO and MIMO cases. With 20 cm/s of nominal
feedrate and 7 mm/s of nominal travel speed, the following
two TFM models were identified:
y2(z)J
(0.13-0.10z -1l\,-2
)z (2.24-1.85z_1)z""2
l
\-\.\2z- +Q36z- z
l-1.15z _ 1 + 0.38z"i! «i(z)
1 1
(0.03-O.Olz-V (0.82-0.38Z-V «2(z).
1 - 1.05z~' + 0.35z"2 1 + 0.98z~' + 0.29z~
(2)
vi(z)]
V2(Z)
SEPTEMBER 1994, Vol. 116/407
 14 '—°—~~- i
12
I (a)
I 10 f= 30 cm/s j
5 f= 18cm/s i
8
_ 14
13
I
JZ 12
~—~S j f = 24 cm/s |
I 11 f=l8cm/s ! . i .
Time (sec)
Fig. 4 Width responses to step changes in feedrate (a) v = 9 mm/s,
(b) v = 7 mm/s, and (c) v = 5 mm/s
(0.18 - 0.09z-V 2 (2.48 - 1.29z~h,-2
l
)z
1 - l.lOz + 0.40z~ 1 - l.lOz + 0.40z-2 «i(z)"
-1 2 -1
(0.035 - 0.01z_1)z (0.80 - 0.44z_1h,-l
)z u2(z).
1 - 0.99z + 0.32z~ 1 - 0.99z_I + 0.32z"
_1 2 tion. This feature is suitable for control of the welding pro-
where each element of the TFM was identified separately
from each I/O pair in Eq. (2), while a whole TFM was
identified in MEMO fashion with the assumption of the com-
mon dynamics in Eq. (3). It is shown in Eq. (2) that the
two inputs produce similar denominator coefficients (i.e.,
dynamics) for each output, which justifies the use of the
common dynamics of each output.
Although the dynamics are assumed to be common to
each output, the gains (b's) are modeled to be different for
each I/O pair in Eq. (1). Note that comparison of Eqs. (2)
and (3) indicates that the common dynamics does not affect
the gains greatly. Another important point is that each numer-
ator polynomial should have at least two coefficients in order
to compensate for a fractional time delay which cannot be
represented by z~dv, where dy is an integer.
The nonlinear feature of the process is demonstrated in
the following open-loop step tests. Figure 4 plots the width
responses to a step change in wire feedrate for three travel
speeds; 9, 7, and 5 mm/s. The sampling period was chosen
as 0.2 s. The feedrate change for each case was chosen such
that it produced a reasonably large change in width without
causing saturation. It is noted that the response with a travel
speed of 9 mm/s shows a much faster rise time than that
with a travel speed of 5 mm/s. Another point observed in
Fig. 4 is that the steady-state gains are quite different for
three cases. A feedrate increase of about 12 cm/s produces
a width increase of 3 mm with a travel speed of 9 mm/s,
while a feedrate increase of only 4 cm/s results in a width
increase of about 5 mm with a travel speed of 5 mm/sec.
These two observations above, i.e., different dynamic char-
acteristics and steady-state gains for different travel speeds,
indicate that the process under consideration is highly non-
linear.
3 Adaptive Control Algorithm
As shown in Fig. 4, nonlinearities observed in the process
significantly change the process parameters (a's and b's)
408/Vol. 116, SEPTEMBER 1994
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab
depending on the operating points; therefore, the process
parameters found in Eq. (3) cannot be used for different
operating points. These factors necessitate the use of adapta-
tion (e.g., on-line parameter estimation) in the controller
design.
(b)
Only a few multivariable adaptive control algorithms have
been developed, though many adaptive control schemes exist
for single-input, single-output cases (e.g., Astrom and Wit-
tenmark, 1980 Narendra and Lin, 1980, etc.). One of multi-
variable adaptive control algorithms developed for discrete-
time system (Goodwin et al., 1980 and Goodwin and Sin,
1984) is a multivariable one-step-ahead adaptive control
algorithm (also referred to deadbeat adaptive algorithm in
some literature). This algorithm is applicable to discrete-
time multi-input, multi-output (MEVIO) deterministic linear
systems. It has a simple structure that is similar to a minimum
variance control scheme, and it guarantees global and asymp-
totical stability of an output tracking system with bounded
input sequences. Since it is based on a transfer function
model, it requires only information on process inputs and
outputs for feedback without an explicit state space descrip-
cess because some of the state variables in the state-space
(3) representation are not physical quantities and, thus, can only
be observed at best. A brief introduction of this algorithm
is presented below.
3.1 Generalized One-Step-Ahead Control. A gen-
eral MBVIO linear discrete-time system with r inputs and r
outputs can be described by the following matrix polynomial
representation
A ^ 1 ) y(k) = B(q-'Mk) (4)
where q~l is a backward shift operator, and y(k) and u(k)
are output and control vectors, respectively. In general, there
exist time delays between I/O pairs (i.e., z~d'J in Eq. (1)).
Since each delay is generally different from each other, a
new output-dependent delay is defined as follows:
dj = min dtj i = 1, •••, r (5)
lS/'Sr
If we define D(q) = diag(#di ••• qdr] as the delay matrix,
then A(#-1) and B(q~x) of Eq. (4) can be represented as
A(q ! ) = I + Ai q * + A2Q 2
+ ••• + A.„q " (6)
l l 1 2 m
B(q~ ) = D' (q) (B 0 + Btf" + B2q~ + - + Bmq~ )
= V-\q)B\q-x) (7)
Note that Bo is the matrix consisting of the leading coeffi-
cients. With this output-dependent delay, the future output
vector y{k) can be expressed as
y ( ^ ) = { y 1 ( ^ + J I ) - v r ( f c + rfr)}r (8)
Then, the predictor form can be given by
y(k) = a(q-l)y(k) + - i iL
^(q- )u(k) (9)
where a^" 1 ) = <*o + «i <fl + a2<7-2 + •••
P(<f 1 ) = Po + 3i<7-1 + P 2 < T 2 + -
po + s-Vte -1 ) (io)
This predictor form enables the future outputs to be ex-
Transactions of the ASME
 pressed in terms of the outputs and inputs up to time k
(i.e., current time). This predictor form will be used in the
subsequent control algorithm.
Consider the following cost function
J{k) = {y(*) y*(*)}r{y(*)-y*(*)} •
+ {u(k) - u(k l)}TR{u(k)-u(k-l)} (11)
where R is a control weighting matrix which is positive
definite; therefore, a diagonal matrix (i.e., R = diag[fb •••,
rrJ) can be selected for simplicity. The first term in Eq. (11)
represents the cost due to tracking errors, while the second
term represents the cost due to large control signals. With
the first term alone, the control system can achieve the
desired output in just one step, but it may produce an exces-
sively large control signal and sometimes cause instability.
In addition, in the subsequent closed-loop control system
with the first term alone, the closed-loop poles tend to cancel
process zeros and thus it cannot be used for nonminimum-
phase process. The control weighting matrix R in the second
term has been introduced to overcome these problems
(Goodwin and Sin, 1984). The importance of the control
weights will be shown later in experimental results. Note
that the difference between u(k) and u{k - 1) appears in the
second term so that a zero steady-state tracking error for a
step command or a step disturbance can be achieved (i.e.,
digital integrator). Minimizing the cost function with respect
to the control u(k) yields the following generalized one-
step-ahead control law:
uW = [pg + R ] - I O 0 { y * W oi{q-l)y(k)
-p'iq-^uik- 1)} + R u ( * - 1 ) } (12)
where y*(k) is the desired future output vector, which is
assumed to be known in advance and bounded for all time.
On the other hand, the closed-loop control system can be
derived from the cost function as follows:
[B'te-VO-? -1 )
B ' ^ P o ' R B ' ^ - V A O T 1 ) ] ^ ) = B'(q-l)y*(k)
(13)
-1
where B'(^ ) is defined in Eq. (7).
3.2 Parameter Estimation. As mentioned earlier, the
parameters of the welding process are changing depending
on both the operating conditions and the past thermal history.
Thus, the parameters cannot be determined accurately before
the welding process and on-line parameter estimation is
required to achieve consistent control. Among numerous on-
line parameter estimation methods, a recursive least-squares
(RLS) algorithm was selected. Because the control algorithm
is based on the predictor form, the parameter estimation
vector 0(£) and regression vector $(k) are also expressed in
this form.
The parameter estimate at time k is the sum of the estimate
at time k - 1 and the term that is proportional to the prediction
error (i.e., actual output-estimated output) as follows:
6(/t) = 9 ( i t - 1)
P ( j t - 1)<K£-1)
e(k) (14)
\(k) + <$>(k- l)TP(k- l)<|)(it-l)
Journal of Dynamic Systems, Measurement, and Control
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab
where the prediction error e(k) = y(k) - §T(k - l)Q(k - 1)
and \(k) is the forgetting factor. The covariance vector P(k)
is updated by
P(*)
P(/t - l)<|>(jt - l)V(k - l)P(Jfc - 1)
P(k- 1)
k(k) \(k) + <$>\k- l ) P ( / t - 1 )<|>(]fc-1)
(15)
Note that the forgetting factor k(k) is used in the RLS method
above. With this forgetting factor, the recent data is given
more weight than the old data. Therefore, the RLS method
work more effectively for the system with time-varying pa-
rameters by discarding the old data. The forgetting factor
can be set to a constant number (typically, 0.98) or to a
variable number determined by the prediction error each
time (Fortescue et al., 1981). It was shown that both worked
well in actual experiments.
When actually implemented, an ordinary RLS method
has some drawbacks. One of these problems is related to
persistent excitation of the input signals. In closed-loop con-
trol, variations of the process inputs and outputs' become
small when the controller becomes well tuned and no
changes in the reference command or disturbance exist. In
this case, parameter update is more likely to be affected by
the measurement errors than by the parameter estimation
errors. In order to prevent this problem, parameter estimation
is switched off when prediction errors become very small.
Another practical aspect is coping with potentially large
measurement errors. The welding process is inherently noisy
and the measurement systems are not perfect, so large mea-
surement errors are likely to occur. Large measurement error
can be detected by monitoring the prediction error, even
though the large prediction error can also be caused by
parameter changes of the process. If I/O data are suspected
of large measurement errors, then small weights are given
to these data in order to reduce the adverse effects of the
large measurement errors on the parameter estimation; it is
called weighted RLS algorithm.
On the other hand, the data collected from the experiments
are usually contaminated by high frequency noise. Low pass
filtering of the data prior to parameter estimation improves
performance since the welding process is inherently slow
and thus the low frequency portion is more important. The
same low pass filters must be used for both input and output
signals.
3.3 Multivariable One-Step-Ahead Adaptive Con-
trol. The multivariable generalized one-step-ahead control
law combined with the recursive least-squares parameter
estimation scheme constitutes the multivariable generalized
one-step-ahead adaptive control algorithm. The convergence
and stability properties were proved by Goodwin et al.
(1980).
Like other adaptive control algorithms, this algorithm can
be implemented in either an indirect or direct form. In the
indirect approach, the controller parameters are computed
based on the process parameters that are identified first by
a parameter estimator. In the direct approach, however, the
controller parameters are estimated directly by reparameter-
izing the process model in advance. The direct approach is
SEPTEMBER 1994, Vol. 116/409
 "*| Supervisory Loop H "

Controller
Parameters Parameter
Desired Estimator
Outputs

Multivariate
Adaptive Welding
Wire Feedrate Process
Controller Width
Travel Speed
7 Depth

Fig. 5 Schematic for multlvarlable adaptive control (direct approach)

generally better than the indirect one because the controller

parameters that determine the control actions are obtained
directly. The experimental comparison will be shown later.
Figure 5 shows the block diagram for the direct approach,
where the parameter estimator identifies the controller pa-
rameters from the I/O data every sampling period. A super-
visory loop primarily monitors the parameter estimator, im-
proves its performance and prevents its malfunction; some
of the supervisory functions were described in the previous
section.
3.4 Guidelines for Control Design. Since the closed-
loop poles tend to cancel the process zeros in the one-step-
ahead (OSA) control scheme without control weights, the
locations of the transmission zeros are very important to the
stability of the control system in that case. This is no longer
true in the generalized OSA control law with control weights
where the closed-loop poles can be chosen arbitrarily to
some extent, but a priori information on the transmission
zeros is still important to the controller design because the
values of the control weighting factors may be determined
by the transmission zeros. In general, the following guide-
lines can be used in determining control weighting factors.
(1) no or small control weights for well-damped zeros
(i.e., zeros well inside the unit circle),
(2) relatively large control weights for poorly-damped
zeros (i.e., zeros inside but near the unit circle),
(3) large control weights for unstable zeros (i.e., zeros
outside the unit circle or nonminimum phase system).
Another important point in determining control weights is
their effect on the closed-loop poles. As shown in the left
hand side of Eq. (13), the closed-loop poles depend on the
choice of the control weighting matrix R; therefore, the
closed-loop stability should be the first consideration when
selecting R.
Since the system under consideration is square (i.e., the
same number of inputs and outputs), the transmission zeros
and closed-loop poles can be computed by taking the deter-
minant of each side of Eq. (13); the transmission zeros are
found as the roots of detB'(<jf') = 0 and the closed-loop poles
are computed as the roots of det[B'(q~l) + ••• A(q )] = 0.
However, for the complex system, selecting R analytically
is often difficult because of computational complexity.
Therefore, proper values of the control weighting factors are
often found by simulation or experiment by trial and error.
4 Control Experiments
In order to evaluate the performance of the adaptive con-
trol system, a series of closed-loop control experiments were
10 15 20 25
Time (sec)
Fig. 6 Depth responses of OSA adaptive controller for (a) no control
weight, and (h) control weight = 0.5 (input; feedrate, output; depth, T, =
0.5 s)
conducted. The same measurement systems and GMAW
setup in Fig. 3 were used. The sampling rate was chosen as
2 Hz because excessively fast sampling rates lead to time
delays of large samples, which is undesirable from a control
standpoint; in the welding process under consideration, time
delays (mainly due to measurements) exist between the pro-
cess inputs and outputs (i.e., 0.9 - 1.0 s for width output
and 0.4 - 0.5 s for depth output).
Although the final objective of this work is to develop and
implement a multivariable control system for the welding
process, a single-input, single-output (SISO) control system
is initially investigated for the following two reasons: (1)
the control of a single output is of practical interest since
control of the width or depth alone is sufficient to achieve
a quality weld in some welding situations, and (2) the SISO
case provides a valuable insight into the MIMO case.
4.1 SISO Cases. There are four combinations of the
input-output pairs: feedrate-width, feedrate-depth, travel
speed-width, and travel speed-depth. Although all four
combinations are equally important since each output must
be controllable from each input in a multivariable control
system, only a few results are shown here.
Choice of Control Weighting Factors. The importance
of the control weighting factor to the generalized one-step-
ahead (OSA) adaptive control was mentioned previously
and is experimentally verified in Fig. 6. Figure 6(a) plots
the depth responses when the process is subjected to step
changes in reference command without a control weight.
The feedrate was used as the control input, while the travel
speed was kept constant. When no control weight was given,
the response shows a large overshoot and a undershoot im-
mediately after step changes in a reference command. In the
absence of control weight, the OSA control algorithm tries
to achieve the desired output in a single step, which leads
to a large variation in the control signal and an actuator
saturation. In an attempt to reduce this large control effort,
the control weighting factor of 0.5 was introduced as shown

410/Vol. 116, SEPTEMBER 1994 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab


10 15 20
Time {sec)
Fig. 7 Comparison of OSA adaptive controller to nonadaptive control-
lers (a) OSA adaptive controller, (b) PI controller, and (c) nonadaptive
OSA controller (input; feedrate, output; width, travel speed = 7 mm/s,
T, = 0.5 s)
in Fig. 6(b); inputs and outputs are normalized so that they
are of about the same magnitude (typically, =1). The over-
shoot observed in Fig. 6(a) is significantly reduced and the
output maintains good command-following. Additionally,
the control signal shows a reasonable variation.
Command Following. The command following per-
formance by the adaptive controller is compared to two
nonadaptive controllers in Fig. 7 for feedrate-width pair.
Figure 1(b) shows the response of the digital PI controller
in the following form
K,
u(z) = KP + - e(z) (16)
l-z~
where Kp and Kj are the proportional and integral gains,
respectively, and e(z) is the error signal defined as y*(z) -
y(z). It is observed that the width response cannot follow
the step change in reference command quickly enough,
though the steady-state performance is relatively good. This
slow response is because the PI controller is operated based
on the error signal between the reference command and the
measured output. Since there are time delays of two samples
(delay of 1 s) and a computational delay of one sample in
the width control, the control action is always delayed by
at least three samples. In Fig. 1(b), a time delay of five or
six samples (2.5 to 3.0 s) is observed in the step change.
This delay can be slightly reduced by adjusting the gains
(e.g., increasing the proportional gain Kp), but at the expense
of a possible overshoot. When compared to the PI controller,
the OSA adaptive controller shows much faster time re-
sponse (Fig. 7(a)). This is because the OSA adaptive con-
troller is based not on the error signal but on the predictor
form which accounts for the time delays including a compu-
tational delay. In the width control case involving three
samples of time delay, for example, the control signal at
time kTs is commanded to track the reference input at time
(k + 3)7;, thus leading to good tracking for the step change.
Figure 7(c) shows the performance of the nonadaptive
generalized OSA controller. The parameters used for the
controller were obtained from the operating condition corres-
Journal of Dynamic Systems, Measurement, and Control
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab
16
e
E 14 r
2 12 «r-« rLXA-^T*"
10.
Fig. 8 Disturbance rejection of OSA adaptive control (input; travel
speed, output; width, disturbance; feedrate)
ponding to the portions before t = 2 s and after t = 22 s in
Fig. 7(c). As a result, tracking performance for the portion
between t = 2 and t = 22 s was very poor. This indicates
that the performance of the OSA controller depends strongly
upon the accuracy of the parameters; thus, it cannot be
effectively used in a system with time-varying parameters
unless some adaptation mechanism (e.g., on-line parameter
estimation scheme) is also used.
Disturbance Rejection. In arc welding, various forms of
disturbances may exist (e.g., thickness change, change of
material properties, parameter drifts of the welding ma-
chines, etc.). The disturbances need to be rejected in most
cases since they adversely affect the performance of control
systems. Figure 8 shows the disturbance rejection perform-
ance of the OSA adaptive controller for travel speed-width
pair. Step changes in the feedrate were used as the distur-
bance. At the time t = 5 s, the feedrate decreases from 22
to 18 cm/s, thus resulting in a quick decrease in the width
output. The control parameters then adapt to the new parame-
ters; the controller decreases the travel speed to compensate
for the decreased width. A similar trend is observed in the
disturbance introduced at time t =25 s.
4.2 MIMO Cases. The same control experiments as
in the SISO case were extended to the two-input, two-output
case. In the actual implementation with the sampling period
of 0.5 s, the width (yt) and depth (y2) outputs have 2 units
and 1 unit of time delay (i.e., dt = 2 and d2 = 1 in Eq. (3)),
respectively; the time delays are identified from the open-
loop step tests. Therefore, the predictor forms can be ob-
tained by successive substitutions as follows:
yi(k + 2) = anyi(k - 1) + oiUyi(k - 2)
+ 0 n m ( * ) + Pi2»i(* - 1) + Pi3«i(* - 2)
+ Pl4«l(* - 3) + Pl5"2(£) + Pl6"2(£ - 1)
+ Pl7«2(*-2) + Pl8«2(*-3) (17)
and
y2(k + 1) = a2iy2(k - 1) + 0L22y2(k - 2)
+ P21«lW + 022"l(£ - 1) + P23"l(£ - 2)
+ P24"2(fc) + ?>25»2(k " 1) + $26U2(k - 2) (18)
where y\(z) and y2(z) are the width and depth, u\(z) and u2(z)
are the wire feedrate (f) and inverse travel speed or inverse
velocity (v_1). The coefficients a's and P's are obtained by
SEPTEMBER 1994, Vol. 116/411
 3 5 10 15 20 25 30 35
f 30

& 25
5 10 15 20 25 30 35
I 20 Time (sec)
CD
^
0)
"- 15 Fig. 10 Output responses of Indirect OSA adaptive controller

-0.1
0.4 ; „/A/\
^^Vv

Fig. 9 Output responses and control inputs of multivariate direct OSA

adaptive controller
0.2

0.0
w
1-0-3
^r jsT \

the recursive least-squares method for both direct and indi-

0.5
AA 9.0

„ , 0 <j*r
J\
0.3 \
5.0
rect approaches; in indirect approach, the process parameters
are obtained first (i.e., the parameters of Eq. (1)), and then
'•

the coefficients are computed. From the above predictor
forms, the control law can be defined by Eq. (12) with
"3n3i5~ riO" y?(k + :
3o = ,R = , J*(k) =
321 324- .0 r2- •y$(k +
anvi(fc- 1) + otnyiik - 2)
«(«"')y(i) = »
u-i\yi(k - 1) + a22y2(k - 2).
and p ' ^ M A : - 1)
3l2«l(& - 1) + - + Pi 8 w 2 (fc-
132 iu\{k- - 1 ) + - + $26U2(k - 2 ) J '
where r\ and r2 are the control weighting factors for the
feedrate and the inverse travel speed, respectively.
The control weighting factors are also important as in the
SISO case. With no control weights, either unstable or highly
oscillatory behavior was observed. In addition to the proper
magnitudes of the control weights, the proper balance be-
tween two control weights are important as well in the
MIMO case. Since the control weight tends to decrease the
control signal, excessively large control weight reduces the
corresponding control signal more than the other, as a result,
the whole process is dominated by the other control signal,
which is undesirable. In most cases, it is desirable to achieve
equal participation from both control inputs, so that a good
control performance may be accomplished. Therefore, an
appropriate choice of the control weighting factors is an
important step in the design of the controller. The proper
level and balance can be found either by simulation or by
experiment as indicated in Section 3.4.
Figure 9 shows the output responses and control inputs
of direct adaptive controller. With proper control weights,
both control signals show reasonable variations. Both output
responses are shown to follow the reference commands rea-
sonably well even with the step changes in the commands.
!\S/
12 16 20 24 28 32 12 16 20 24 28 32
Tims Time
Fig. 11 Parameter estimates for OSA adaptive controller (dotted lines
indicate changes in reference commands)
Figure 10 shows the output responses of an indirect adaptive
controller. The indirect controller also produces a reasonably
good performance, but die steady-state responses are some-
what oscillatory when compared to those of the direct con-
troller.
Figure 11 plots some of the parameter estimates that are
identified by the RLS method. The parameters around time
t = 22 s are plotted in order to emphasize the parameter
changes in a step change of a reference command. It is
observed that all parameters change significantly after the
set point change. The control signals after t = 22 s are
computed based on these changed control parameters and
thus improves performance of the control system.
In the MIMO case, the control range is also an important
issue (Hale and Hardt, 1990a, Doumanidis and Hardt, 1990,
and Hardt, 1992). In gas metal arc welding process, two
outputs (i.e., width and depth) are highly coupled partly
because two inputs (i.e., feedrate and travel speed) have
similar effects on each output. Therefore, there is a limit to
the range which combinations of the two control inputs
can achieve. For example, an increase in die width while
decreasing or maintaining a constant depth cannot be at-
tained with any combination of the control inputs. No control
algorithm can overcome this inherent problem since the
limitations are due to the welding process and machines
used. Therefore, the desired bead geometries must be in die
achievable control range. The adaptive controller cannot
provide good command following performance when de-
sired bead geometries derive greatly from die achievable
control range.
This output coupling may make the disturbance rejection

412/Vol. 116, SEPTEMBER 1994 Transactions of the ASME

Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab

 -> Disturbed
E 12
§
E
10
4r
10 15
> Disturbed
%r
0 5 10 15 20 25 30 35
Fig. 12 Simulation of disturbance rejection by OSA adaptive controller
(disturbances; changes in steady-state gains)
difficult. If the disturbance affects one output much more
than the other, then it is difficult to reject this disturbance.
For example, a certain disturbance tends to increase the
width with no change in the depth. Then, the control signals
must decrease the width while maintaining the depth to
achieve the desired bead geometries. As mentioned before,
such a combination of the control inputs are hard to find
because of the strong I/O coupling. However, if the distur-
bance affects both outputs and they are in the achievable
control range, then the adaptive controller can reject it by
adapting to new process parameters as in the SISO case.
This can be shown in Fig. 12. In this simulation, the gain
disturbances (b's in Eq. (1)) were injected as step functions
from the nominal values at t = 14 s. These disturbances
make both responses increase from the nominal conditions
significantly at the introduction of the disturbances. It is
noted that the disturbed outputs are in the achievable control
range. The adaptive controller quickly adapts to the new
parameters and commands the proper control inputs to reject
the disturbances.
5 Conclusions
Gas metal arc welding process is highly nonlinear and its
dynamics depends on the operating and boundary conditions.
It is therefore very difficult to obtain the process model
accurately in advance and on-line parameter estimation and
adaptive control algorithm are required. A one-step-ahead
adaptive control strategy combined with a recursive least-
squares methods has been investigated. Command following
and disturbance rejection properties of the adaptive control
system for both SISO and MIMO cases are investigated by
simulations and experiments. The adaptive control system
shows reasonable performance on command following for
both SISO and MIMO cases. This system also shows good
disturbance rejection in the SISO case, but is shown to be
of limited success in the MIMO case due to inherent lack
of output decoupling in the GMAW process. Process modifi-
cation to decouple the process outputs to a great extent is
underway (Masmoudi, 1992).
Even though a truly independent control of the outputs
is difficult to implement due to a strong output coupling
inherently existing in the process, a control system for simul-
taneous control of bead width and depth was successfully
implemented in this paper. More important, a closed-loop
control system to directly regulate the depth of penetration
Journal of Dynamic Systems, Measurement, and Control
Downloaded From: http://dynamicsystems.asmedigitalcollection.asme.org/pdfaccess.ashx?url=/data/journals/jdsmaa/26207/ on 05/18/2017 Terms of Use: http://www.asme.org/ab
was first successfully implemented in this research thanks
to the thermally based depth estimator (Song and Hardt,
1992).
Acknowledgment
This project was supported by the U.S. Department of En-
ergy under contract number DE-FG02-85ER13331.
References
Andersen, K., Cook, G. E., Karsai, G., and Ramaswamy, K., 1990, "Artificial
Neural Networks Applied to Arc Welding Process Modeling and Control," IEEE
Transactions on Industry Applications, Vol. 26, No. 5, pp. 824-830.
Astrom, K. J., and Wittenmark, B., 1984, Computer Controlled Systems,
Prentice-Hall.
Astrom, K. J., and Wittenmark, B., 1989, Adaptive Control, Addison Wesley.
Boo, K. S., and Cho, H. S., 'A Fuzzy Linguistic Control Approach to the
Control of Weld Pool Size in Gas Metal Arc Welding Processes," Welding and
Joining Processes, ASME WAM, PED-Vol. 51, pp. 73-84.
Cook, G. E., Andersen, K., and Barrett, R. J., 1989, "Feedback and Adaptive
Control in Welding," Proceedings of the 2nd International Conference on Trends
in Welding Research, Gatlinburg, TN, ASM International, pp. 891-903.
Dornfeld, D. A., Tomizuka, ML, and Langari, G„ 1982, "Modeling and Adap-
tive Control of Arc Welding Processes," Measurement and Control for Batch
Manufacturing, ASME WAM, pp. 53-64.
Doumanidis, C. C , and Hardt, D. E., 1989, "A Model for In-Process Control
of Thermal Properties during Welding," ASME JOURNAL OF DYNAMIC SYSTEMS,
MEASUREMENT, AND CONTROL, Vol. I l l , No. 1, pp. 40-50.
Doumanidis, C. C , and Hardt, D. E., 1991, "Multivariable Adaptive Control
of Thermal Properties During Welding," ASME JOURNAL OF DYNAMIC SYSTEMS,
MEASUREMENT, AND CONTROL, Vol. 113, No. 1, pp. 82-92.
Fortescue, T. R., Kershenbaum, L. S., and Ydstie, B. E., 1981, "Implementa-
tion of Self-Tuning Regulators with Variable Forgetting Factors," Automatica,
Vol. 17, pp. 831-835.
Goodwin, G. C , and Sin, K. S., 1984, Adaptive Filtering Prediction and
Control, Prentice-Hall.
Goodwin, G. C , Ramadge, P. J., and Caines, P. E., 1980, "Discrete-Time
Multivariable Adaptive Control," IEEE Transactions on Automatic Control,
Vol. AC-25, No. 3, pp. 449-456.
Hale, M. B., and Hardt, D. E., 1990a, "Multivariable Geometry Control of
Welding—Part I: Process Modeling," Automation of Manufacturing Processes,
ASME Winter Annual Meeting, DSC-Vol. 22, pp. 1-9.
Hale, M. B., and Hardt, D. E., 1990b, "Multivariable Geometry Control of
Welding—Part U: Process Control," Automation of Manufacturing Processes,
ASME Winter Annual Meeting, DSC-Vol. 22, pp. 143-151.
Hardt, D. E., Garlow, D. A., and Weinert, J. B„ 1985, "A Model of Full
Penetration Arc-Welding for Control System Design," ASME JOURNAL OF DY-
NAMIC SYSTEMS, MEASUREMENT, AND CONTROL, Vol. 107, No. 1, pp. 40-46.
Hardt, D. E., 1992, "Welding Process Modeling and Redesign for Control,"
Proceedings of Japan/USA Symposium on Flexible Automation Vol. 1, ASME,
pp. 275-281.
Langari, G., and Tomizuka, M., 1988, "Fuzzy Linguistic Control of Arc
Welding," Sensors and Controls for Manufacturing, ASME WAM, pp. 157-162.
Masmoudi, R. A., 1992, "Modeling and Control of Thermal and Geometric
Properties in Arc Welding," Ph.D. thesis, Massachusetts Institute of Technology,
Cambridge, MA.
Nomura, H., Satoh, Yu., Tohno, K., Satoh, Yo., and Kurotori, M., 1980, "Arc
Light Intensity Controls Current in SA Welding System," Welding and Metal
Fabrication, Vol. 48, pp. 457-463.
Richardson, R. W., Gutow, D. A., and Rao, S. H., 1982, "A Vision Based
System for Arc Weld Pool Size Control," Measurement and Control for Batch
Manufacturing, ASME WAM, November, pp. 65-75.
Smartt, H. B., 1992, "Intelligent Control of Arc Welding," Proceedings of
the 3rd International Conference on Trends in Welding Research, Gatlinburg,
TN, ASM International, June.
Song, J.-B., and Hardt, D. E., 1991, "Multivariable Adaptive Control of Bead
Geometry in GMA Welding," Welding and Joining Processes, ASME WAM,
PED-Vol. 51, pp. 123-124.
Song, J.-B., 1992, "Multivariable Adaptive Control in GMA Welding Using
a Thermally Based Depth Estimator," Ph.D. thesis, Massachusetts Institute of
Technology, Cambridge, MA.
Song, J.-B., and Hardt, D. E., 1992, "A Thermally Based Weld Pool Depth
Estimator for Real-Time Control," Proceedings of the 3rd International Confer-
ence on Trends in Welding Research, Gatlinburg, TN, ASM International.
Suzuki, A., Hardt, D. E., and Valavani, L., 1991, "Application of Adaptive
Control Theory to On-Line GTA Weld Geometry Regulation," ASME JOURNAL
OF DYNAMIC SYSTEMS, MEASUREMENT, AND CONTROL, Vol. 113, No. 1, pp.
93-103.
Vroman, A. R., and Brandt, H., 1976, "Feedback Control of GTA Welding
Using Puddle Width Measurement," Welding Journal, Vol. 57, No. 9, pp.
742-749.
SEPTEMBER 1994, Vol. 116/413