Civil Engineering Reference
In-Depth Information
30
Bead Width
Penetration
Reinforcement Height
Cross Section Area
25
20
Error Standard
Deviations [%]
15
10
5
0
2x18
3x12
4x9
2x24
3x16
4x12
2x30
3x20
4x15
2x36
3x24
4x18
Configurations: Hidden layers x Nodes
FIGURE 7.7
Standard deviations of mapping errors (%) for various network configurations, shown for all four DWPs.
as it requires chemical etching, which results in a relatively blurred boundary between the bead and the
surrounding base metal. This difference is reflected in the consistently lower accuracy of the penetration
modeling, compared with the width modeling.
The two neural network schemes presented by Zeng and Andersen apply to static parameter settings
only. Therefore, dynamic control of the welding process is left to the various controllers of the welding
equipment, such as the automatic voltage controller (AVC), power source (usually constant current for
GTAW), etc. Thus, it is the responsibility of these controllers to maintain the parameters, selected by the
neural network, constant in the presence of noise or process perturbations.
Real-Time Monitoring of Weld Joint Penetration
A rather elaborate implementation of monitoring weld joint penetration using machine vision and a
high-shutter-speed camera assisted with pulsed laser illumination is presented by Kovacevic, et al. [35,
36, 37, 38, 39]. In their research, they employ an ANN to emulate the human operator's ability to extract
weld penetration information from the geometrical appearance of the weld pool [36]. Conceptually,
Kovecevic believes that the human welder relies on the geometrical appearance parameters, which include
size and shape information, of the weld pool to represent weld penetration. Kovacevic's model uses an
empirically derived relationship between the sag of the weld near the rear of the pool and the root face
bead width [15, 29, 30]. Thus, their ANN weld penetration estimator is trained with the length and the
included rear angles of the weld pool to represent the pool's size and shape, respectively.
The objective of Kovacevic's experiments was to monitor, in real-time, the full penetration state of the
weld pool which is described by the root face bead (backside) width. Extensive neural network config-
urations were tested in order to examine the possibility of using different pool parameters as inputs to
represent the weld penetration. Each of the networks, however, contained only one output node which
was responsible for calculating the backside bead width.
They employed a commercial neural network software package to facilitate the training of the various
neural network configurations. An extended delta-bar-delta (EDD) learning algorithm was used to
overcome the slow convergence rate associated with the back propagation learning algorithm. Only one
hidden layer was used in which the number of nodes,
n
2
, was calculated by the following equation:
N
10
n
1
N
--------------------------
-----------------------------
n
2
(7.8)
(
n
3
)
5
n
1
(
n
3
)
