INEEL/CON-02-00596

PREPRINT

Intelligent Control of Modular Robotic Welding

Cell

H. B. Smartt
K. L. Kenney
C. R. Tolle

April 15, 2002 – April 19, 2002

6th International Conference on Trends in

Welding Research

This is a preprint of a paper intended for publication in a

journal or proceedings. Since changes may be made
before publication, this preprint should not be cited or
reproduced without permission of the author.
This document was prepared as a account of work
sponsored by an agency of the United States Government.
Neither the United States Government nor any agency
thereof, or any of their employees, makes any warranty,
expressed or implied, or assumes any legal liability or
responsibility for any third party's use, or the results of
such use, of any information, apparatus, product or
process disclosed in this report, or represents that its
use by such third party would not infringe privately
owned rights. The views expressed in this paper are
not necessarily those of the U.S. Government or the
sponsoring agency.
6th International Conference on Trends in Welding Research INEEL/CON-2002-00596
15-19 April 2002
Pine Mountain, GA USA
ASM International

Intelligent Control of Modular Robotic Welding Cell

Herschel B. Smartt, Kevin L. Kenney, Charles R. Tolle

Idaho National Engineering and Environmental Laboratory
P.O. Box 1625
Idaho Falls, ID 83415-2210

Abstract Introduction
Although robotic machines are routinely used for welding, James Albus [1991] at NIST has defined machine
such machines do not normally incorporate intelligent intelligence as “the ability of a system to act appropriately in
capabilities. We are studying the general problem of an uncertain environment, where appropriate action is that
formulating usable levels of intelligence into welding which increases the probability of success, and success is the
machines. From our perspective, an intelligent machine achievement of behavioral subgoals that supports the system’s
should: incorporate knowledge of the welding process, know ultimate goal.” Following Albus’ intent, we can say that
if the process is operating correctly, know if the weld it is intelligent machines are those that either know or can learn
making is good or bad, have the ability to learn from its everything they need to know to perform a process or task.
experience to perform welds, and be able to optimize its own Such machines may be able to perform a process or task
performance. To this end, we are researching machine autonomously (without operator intervention) or semi-
architecture, methods of knowledge representation, decision autonomously (with operator intervention).
making and conflict resolution algorithms, methods of
learning and optimization, human/machine interfaces, and In this paper, we present a modular design of a machine
various sensors. using distributed knowledge represented as local rules of
behavior. In the design there is no central knowledge of how
This paper presents work on the machine architecture and
to perform a weld. Our approach is inspired by biological
the human/machine interface specifically for a robotic, gas
models of social insects. Related work has been conducted,
metal arc welding cell. Although the machine control problem
for example, by Dorigo and Colorni [1996] using ant-based
is normally approached from the perspective of having a
local behavior of multiple agents to solve the Travelling
central body of control in the machine, we present a design
Salesman Problem and other classical hard problems. Schatz
using distributed agents. This new design is loosely based on
et al. [1999] formulated a model for route learning in ants.
biological models of social insects. For example, in an ant
Lambrinos et al. [2000] used a similar model for navigation of
colony each ant functions according to local rules of behavior
a mobile robot. Overgaard, Petersen, and Perram [1995, 1996]
[Hölldobler and Wilson, 1990, see chapters 8 and 9]. There is
used local agent control of dynamic motion and path planning
no “king” or “queen”, although the latter name has been given
in multiple link robot arms.
to the reproducing ant. Following a similar approach, we
present a modular machine architecture in which each Consider an intelligent machine in which various machine
machine element has local rules of behavior but no single functions are carried out in a distributed manner. A schematic
element understands how to make a weld. A prime goal of of such a machine for arc spot welding is shown in Figure 1.
this work is to develop an architecture for an intelligent In addition to the machine hardware required (most of which
machine that will support a modular, plug and play standard. is not shown) there are several “agents”, see Figure 2. These
A secondary goal of this work is to formulate a agents have local control of various machine functions and are
human/machine interface that treats the human as an active able to communicate with each other and with an operator
agent in the modular structure. agent. The operator agent may be a human or may be an
interface to a human (or even an interface to another machine).
(Although it would be possible to focus on autonomous
machines, we chose not to do so; our machines interact with
humans who have supervisory control authority.) The various

1
 Gas
Agent WIRE FEEDER

Timer
Operator
Agent
Wire Feeder Agent 0.6-1.6+ mm diameter Agent
steel, aluminum
GUI
Power Agent
Supply Agent Physics
Ar Agent
Ar + O2
POWER Ar + CO2
SUPPLY Robot
Agent

50-500 mm/s 1/100 ohm

- + Sensor
Agent
12-25 mm Transfer
CV 20-30 volts 1-1000 s -1 Mode Detector
CC 50-500 amperes 1/10 ohm Agent

1-100+ mm

5-50 mm/s X-Axis

Agent

Figure 1. Gas Metal arc spot welding machine with various agents including power supply, electrode wire
feeder, robot, sensor, metal transfer mode detector, process physics, operator, and user interface.

agents incorporate knowledge of how to perform their local is modular in that individual machine elements or functions
tasks, although there is no single agent that has knowledge of may be removed or replaced with other elements or functions.
the entire process. The agents communicate via a bus using a The total machine configuration is defined in an external data
protocol similar to TCP/IP and a vocabulary incorporating file that is downloaded to the machine agents at run time. This
both generic and process specific words. Each agent is allows the machine configuration to be changed without
dedicated to a specific machine hardware element or function. modifying the source code.
The individual agents incorporate specific knowledge
To formulate such a machine, we need a variety of
expressed as fuzzy logic rules of behavior. The total machine
methods. In addition to distributed learning and control, we

tasks, dead man

state
methods analog

com dead
port I/O
I man
state digital
Ethernet/TCP-IP
high-level
messages
serial port

mid-level
motion
control

Figure 2. Agent block diagram. I designates a command interpreter.

2
also chose to have our machines learn rules of behavior. This materials (e.g., advanced high strength steels) the heat input
is distinct from learning control trajectories, a method must be limited to minimize the metallurgical degradation
frequently employed for machine learning. Our rules are associated with the high peak temperatures and slow cooling
embodied using a variant of fuzzy logic [Johnson and Smartt, rates experienced during welding. The relationship between
1995] that allows the system to learn by back propagation heat input and the welding process variables is given by
[Rumelhart, 1986]. [Rosenthal, 1946]
Welding Application
H = EItη (1)
Consider a specific welding control problem. We desire to
fabricate a spot welded steel structure using gas metal arc where H is heat input, E is arc voltage, I is arc current, t is arc-
welding (GMAW), Figure 1. Thus, steel sheet will be welded on time and η is heat transfer efficiency from the process to
to an underlying structure by means of weld nuggets deposited the base metal. Arc voltage and time are independent
into circular holes in the sheet. variables selected by the weld engineer. Arc current, on the
In this situation, the weld torch may be moved to a suitable other hand, is a dependent variable that is function of the
position over a weld site, using motion control as discussed independent variables of wire feed speed, electrode diameter
earlier. The welding power supply contactor is activated, the and electrode stick-out in. This relationship is given by
power supply voltage is set, the shielding gas is turned on, and [Smartt & Einerson, 1993]
the electrode wire is fed downward. This will result in
ignition of an arc with corresponding heat and mass transfer to
the weldment. After a suitable time, the power supply
I = K 0 + K1S + K 2 E + K 3 (CT ) (2)
contactor is deactivated and the electrode wire feed is stopped.
A short time later the shielding gas is turned off. Although where S is wire feed speed, E is arc voltage, CT is contact tip
this is perhaps the simplest arc welding example we can to workpiece distance and K0, K1, K2 and K3 are constants.
consider, there are still important control decisions that ensure For constant-voltage gas metal arc welding at a given CT, arc
that the weld will meet its acceptance requirements current is directly controlled by the wire feed speed.
To obtain a good weld in this example, the current must be Second, the weld nugget volume (or mass input to the
high enough but not too high and the weld time (the time the weld) is an important characteristic of a good weld. The
arc is on) must be equal to or greater than some critical relationship between mass input (G) and the weld variables is
minimum. This will ensure that adequate heat and mass have given by
been transferred to the weldment. It is also necessary for the
voltage to be above some minimum (to reduce spatter) and
below some maximum (to avoid melt through and burn back). πd 2
G=( ) St (3)
Weld procedures are normally developed by a weld 4
engineer based on the requirements identified for a “good”
weld. A good weld in this spot weld application is one that is where d is the electrode diameter, S is the wire feed speed and
strong enough, does not excessively over or under fill the hole, t is the arc-on time. For a given electrode, the values for WFS
has minimal spatter, and does not contain gross defects such as and time are selected to provide adequate fill.
cracks or porosity that could lead to failure. To be strong
enough, the weld bead must adequately penetrate the lower A third consideration for weld procedure development is
structure (but not excessively melt through that structure) arc length. If the arc length is too short the electrode will stub
while fusing into the upper sheet. For most applications, the into the weld pool, producing spatter which is undesirable for
cross-sectional area of the weld bead in the plane of the aesthetic reasons as well as for the potential to interfere with
interface between the upper sheet and lower structure needs to mating or adjacent parts. If the arc length is too long the
be equal to or greater than some critical amount. The electrode will burn back into the contact tube, the arc transfers
independent weld variables that are specified in the weld to the contact tube, and the process becomes uncontrollable.
procedure include arc voltage, wire feed speed and weld time. Arc length is also a function of the weld variables and is given
The appropriate settings for these variables are based on by [Reutzel, 1996]
several key characteristics of the weld process.
C1
First, the heat transferred to the base metal directly from L = C0 E + + C2 I (4)
the arc and molten metal droplets is often a key consideration
I
in procedure development. The weld heat input must be where L is arc length, E is arc voltage, I is arc current and C0,
sufficient to provide the penetration and weld bead interface C1 and C2 are constants.
area required for joint strength. However, when joining some

3
 An examination of Equations 1-4 shows heat input, mass droplets transferred from the electrode to the work piece. The
input and arc lengths are all functions of the independent weld metal transfer mode is known to affect weld deposition rate,
variables and are thus all inter-related. A change in wire feed penetration, and spatter. Metal transfer mode is an important
speed affects heat input, masses input and arc length. A factor in out-of-position welding, and its effect on weld
change in weld time affects both heat and mass input, and penetration makes it important for thin section welding where
changes in arc voltage affect both heat input and arc length. burn-through is a concern. Three distinct metal transfer modes
This means that independent adjustment of heat input, mass have been identified for gas metal arc welding: spray, globular
input or arc length requires combined changes of wire feed and short-circuiting (also referred to as short-arc) transfer.
speed and weld time in some unique ratio. For example, Spray transfer mode is characterized by consistency, good
increasing the mass input heat input can be accomplished by penetration, and a low degree of spatter. For applications
increasing wire feed speed or time (it is impractical to change where maximum penetration and/or minimal spatter are
electrode size). However, either action also increases the heat desired, spray transfer is often preferred. Short-arc is often
input. Consequently, if wire feed speed is increased than the preferred for welding thin sections or for out-of position
weld time must be decreased, or visa versa. Thus it is difficult welding due to the lower current levels (lower heat input)
to control heat and mass inputs to the weld independently. offered by the process. In addition, spray transfer may not be
However, such independence can be obtained by solving achievable with higher concentrations of CO2 in the shielding
Equations 1-4 appropriately. First, Equation 2 must be gas. Since metal transfer mode is an important feature, a weld
defined explicitly. An empirically derived relationship was mode identifier was considered an integral component of an
obtained assuming a constant contact tip workpiece distance intelligent welding system.
(CT). This expression for arc current (I) is given by
This task of identifying the metal transfer mode is
dedicated to one of the independent agents of the intelligent
I = 0.97S + 3.25V + 5.84 (5) weld system. This weld mode analysis agent performs a fuzzy
logic based analysis of the acquired current and voltage
Equations 1, 3 and 5 may be solved iteratively to obtain signals [Smartt, et.al. 2001], and is referred to as a Fuzzy
values for wire feed speed (S) and weld time for a given heat Logic Weld Mode Identifier (FLWMI) agent. The FLWMI
input, mass input and voltage. agent performs a fuzzy logic comparison of the average
current (Iavg), the minimum value of current (Imin), the standard
This procedure is the basis for the methodology employed deviation of current (Isd), and the standard deviation of voltage
by the intelligent robotic weld system for providing the human (Vsd) and, based on this comparison, identifies the metal
operator with advanced tools for developing the weld transfer mode as spray, globular, short-circuiting, or shorting
procedure. After all, since the values of wire feed speed, streaming. The Fuzzy Logic Rules from which the diagnostic
voltage and weld time are all based on the requirements of algorithm is based are as follows:
heat input, mass input and arc length, why not present the
operator with controls to select the latter parameters directly. • If Isd is low, then mode is spray
As such, the virtual weld panel of the robotic welding cell • If Isd is high AND Vsd is high AND Imin is high, then mode
contains controls for heat input and mass input, and, since arc is globular
length varies linearly with arc voltage, a control for arc • If Isd is high AND Vsd is high AND Imin is low AND Iavg is
voltage is also provided. Using this methodology, the low, then mode is short-circuiting
operator can dial in a heat-input value, a fill value and a • If Isd is high AND Vsd is high AND Imin is low AND Iavg is
voltage value directly. high, then mode is shorting streaming.
It is interesting how the system adjusts the wire feed speed As shown in Figure 1, the robotic gas metal arc welding cell is
and weld time with independent changes of heat input, mass composed of the following independent agents:
input or voltage. For example, an increase in heat input is
accomplished by decreasing the wire feed speed while • Analog Input Agent
increasing the weld time to maintain a constant mass input. • Power Supply Agent
An increase of the mass input is accomplished by increasing • Wire Feeder Agent
the wire feed speed while decreasing the weld time to • Timer Agent
maintain a constant heat input with additional fill. When • Gas Agent
voltage is increased, wire feed speed increases and time • Robot Agent
decreases so that both heat input and fill remain unchanged as • GUI Agent
the voltage is increased.
• FLWMI (Fuzzy Logic Weld Mode Identifier) Agent
In addition to the weld process variables, another important
process characteristic is metal transfer mode. The Each agent is assigned a specific, independent task,
characteristics of the constant-voltage GMAW process are requiring the agents to work together to accomplish the global
best described in terms of the size and frequency of metal objective of making a spot weld. In the process of performing

4
a spot weld, the operator specifies the maximum allowable − Send message to Timer to End Weld (goto 11)
weld time, the mass input, the heat input and the arc voltage 9. AI Agent
for the spot weld. The Physics Agent, which is embedded in − Receives message to “End Weld”
the GUI Agent, then iteratively solves Equations 1, 3, and 5 − Read waveforms
for the appropriate wire feed speed setting and the weld time − Compute statistics
setting, as discussed in detail above. The Wire Feeder, Power − Send waveforms and statistics to GUI Agent
Supply and Timer agents then retrieve their appropriate − Send waveforms to FLWMI Agent (goto 13)
settings from the GUI Agent. Before the weld process can 10. Gas Agent
proceed, the operator must Arm each agent. If an agent is not
− Receives message to “End Weld”
armed, it will not perform any task beyond initialization.
− Turn gas off
Once armed, the agents are able to perform their tasks
− Send message to Robot Agent to End Weld (goto 14)
associated with the spot welding process. The weld process is
11. Timer Agent
initiated by the operator and the weld operation proceeds in
the sequence that follows. − Receives message to “End Weld”
− Request time from WF Agent
1. Operator presses “Weld” button − Compares desired weld time to actual weld time,
− Message is sent to Power Supply Agent to “Weld” compute time offset for next weld
2. Power Supply Agent 12. FLWMI Agent
− Receives message to “Weld” − Receives message to “End Weld”
− Check status of Deadman Switch − Receives waveform data
− If Deadman is open − Executes analysis algorithm
− Then abort sequence and notify operator − Sends weld mode to GUI Agent
− If Deadman is closed 13. CRS Agent
− Then − Receives message to “End Weld”
− Enable power supply contactor − Move torch to new position
− Send message to AI Agent to “Weld” (goto 3)
− Send message to Gas Agent to “Weld” (goto 4) During the welding sequence, each agent independently
− Send message to WF Agent to “Weld” (goto 5) monitors the status of the Deadman Switch. If the switch
3. AI Agent opens up before the Timer Agent initiates the “End Weld”
− Receives message to “Weld” sequence, then each agent terminates the weld sequence.
− Start data acquisition
4. GAS Agent
Conclusion
− Receives message to “Weld”
An approach to design of an intelligent machine has been
− Turn gas on presented based on distributed intelligence. Local agents are
5. WF Agent used to control individual machine functions and to process
− Receives message to “Weld” information needed by the machine functions. Examples of
− Enable wire feeder how this approach may be used to build a specific machine are
− Send message to Timer Agent to Weld (goto 6) presented for an arc spot welding application. A possible
6. Timer Agent agent internal structure is presented that provides for local
− Receives message to “Weld” rules of behavior and safety considerations. Additional
− Reset Timer research on this approach is presented in [Smartt, et. al.,
− Monitor weld time. When weld time expires, send 2000].
message to Wire Feeder Agent to “End Weld” (goto
7) Acknowledgement
7. Wire Feeder Agent
− Receives message to “End Weld” This work is supported by the Basic Energy Sciences, Office
− Disable wire feeder of Science, U.S. Department of Energy, under DOE Idaho
− Send message to Power Supply Agent to “End Weld” Operations Office Contract DE-AC07-99ID13727.
(goto 8) References
− Send message to AI Agent to “End Weld”, also send
time to AI Agent (goto 9) Albus, James. S., “Outline for a Theory of Intelligence,” IEEE
8. Power Supply Agent Transactions on Systems, Man, and Cybernetics, vol 21, no. 3,
− Receives message to “End Weld” pp. 473-509, May/June (1991).
− Disable Contactor
− Send message to Gas Agent to End Weld (goto 10)

5
Dorigo, Marco; Maniezzo, Vittorio; Colorni, Alberto. “Ant
System: Optimization by a Colony of Cooperating Agents,”
IEEE Transctions on Systems, Man, and Cybernetics-PartB:
Cybernetics, v 26, n 1, pp. 29-41, Feb. 1, (1996).
Hölldobler, B. and Wilson, E. O., the ANTS, The Belknap
Press of Harvard University Press, Cambridge, Mass., (1990).
Johnson, J. A., and H. B. Smartt, "Advantages of an
Alternative Form of Fuzzy Logic," IEEE Transactions on
Fuzzy Logic, 3, pp. 149-157, (1995).
Lambrinos, D., Möller, R., Labhart, T., Pfeifer, R., and
Wehner, R., “A mobile robot employing insect strategies for
navigation”, Robotics and Autonomous Systems, 30, pp. 39-
64, (2000).
Overgaard, L., Petersen, H. G., and Perram, J. W., “A General
Algorithm for Dynamic Control of Multilink Robots”, The
International Journal of Robotics Research, Vol. 14, No. 3, pp.
281-294, June (1995).
Overgaard, L., Petersen, H. G., and Perram, J. W., “Reactive
Motion Planning: a Multiagent Approach”, Applied Artificial
Intelligence, 10:35-51, (1996).
Reutzel, E. W., 1739 Blue Course Dr., State College, PA,
Personal communication (1996).
Rosenthal, D., The theory of moving sources of heat and its
application to metal treatments, Trans. ASME, pp. 849-866,
(1946).
Rumelhart, D. E., G. E. Hinton, and R. J. Williams, "Learning
Internal Representations by Error Propagation," Parallel
Distributed Processing: Explorations in the Microstructures
of Cognition, Vol. 1, ed. D. E. Rumelhart and J. L.
McClelland, MIT Press, Cambridge, MA, pp. 318-362 (1986).
Schatz, B., Chameron, S., Beugnon, G., and Collett, T. S.,
“The use of path integration to guide route learning in ants”,
Nature, Vol. 399, No., 24, pp. 769-772, June (1999).
Smartt, H. B., and Einerson, C. J., “A model for heat and mass
input control in GMAW”, Welding Journal 72(5):217-s to
229-s (1993).
Smartt, H. B., Kenney, K. L., Johnson, J. A., Carlson, N. M.,
Clark, D. E., and Reutzel, E. W., “Method and Apparatus for
Assessing Weld Quality”, United States Patent 6,236,017,
May 22, 2001.
Smartt, H. B., Tolle, C. R., and Kenney, K. L., “Complex
Intelligent Machines”, Proceedings, Eighteenth Symposium on
Energy Engineering Sciences, DOE-BES, Argonne National
Laboratory, May 15-16, (2000).