Civil Engineering Reference
In-Depth Information
Consequently, Messler uses an ANN to capture the dynamics of the RSW process through the training
of actual welding data.
A time-delay neural network was developed by Messler in order to map the percent heat input to
electrode displacement. In addition to percent heat input, the time-dependent electrode displacement
output of the network is fed back as fixed-time delayed inputs. These tapped delay lines are necessary in
order to consider the dynamic electrode displacement response. The result is a single-channel temporal
sequence neural network which can map the percent heat input to electrode displacement.
Four different neural network structures using various time delays were trained using the back prop-
agation learning algorithm in order to maximize performance. Training data sets were obtained under
normal production welding conditions, i.e., no special surface preparation or cleaning was performed
and electrodes were replaced as needed, for various heat inputs. A second training set was evaluated when
a worn electrode was used under similar heat inputs. In both cases the lowest maximum error of 5 %
was obtained using four input nodes. (percent heat input and three time delays), seven hidden layer
nodes, and one output node.
Closed-Loop Control of GTA Weld Bead Geometry
An approach to closed-loop control, using neural networks, has been attempted by a number of research-
ers using various welding processes, such as GTAW, GMAW, RSW, VPPAW, etc., and a wide range of
sensing devices, such as optical, arc, infrared, acoustical, ultrasonic, etc. [5, 6, 7, 11, 25, 27].
Andersen developed a digital welding control system for the GTAW process in which the objective was
to maintain constant weld bead dimensions when a variation in the welding environment was encoun-
tered, such as a variation in the thickness of the plate upon which the weld was being performed [6].
This is a practical consideration when welding items of relatively complex shapes, where the varying
thermal conditions can affect the geometry of the molten pool.
Andersen's control system is most readily explained by reference to
Fig. 7.9
,
which illustrates the main
components. The desired bead width and penetration are specified by the user as W
REF
and P
REF
, respectively.
These parameters, as well as the workpiece thickness,
H
, are fed to a neural network setpoint selector,
which yields the nominal travel speed, current, and arc length (
0
,
and
L
0
, respectively). Arc initiation
and stabilization is controlled in an open-loop fashion by the weld start sequencer. Given the desired
equipment parameters the arc is typically initiated and established at a relatively low current, with the
other equipments parameters set at some nominal values. Once the arc has been established the equipment
I
0
,
H
V
Start
V
0
Neural
Net
Setpoint
Selector
Weld
Start
Sequencer
I
0
I
Start
L
Start
L
0
Weld
Process
W
+
W
REF
+
PI-Ctrl
+
t = T
t = T
P
-
+
P
REF
+
PI-Ctrl
+
t = T
t = T
-
H
W
Model
Neural
Net
Weld
Model
V
P
Model
I
L
FIGURE 7.9
Closed-loop welding control system.
