Civil Engineering Reference
In-Depth Information
TORCH
SCANNER
TACK
WELD
INDISTINCT
REGION
NEURAL NETW
NEURAL NETWORK
ORK
LASER/
VIDEO
SCANNER
POSITION
POSITION
CONTR
CONTROLLER
OLLER
SEAM
FIGURE 7.13
The seam tracking system.
The objective of the seam tracker is to monitor the location of the unwelded seam with respect to the
moving welding torch, and to adjust the location of the torch with respect to the seam in real-time. In
addition to seam tracking applications, the profile scanning system can be utilized to evaluate the quality
of the VPPA weld. The quality of a VPPAW root pass can, to a large extent, be determined from the shape
of its surface profile. Therefore, means for automatically detecting abnormal shapes, excessive undercuts,
asymmetry of the weld profile, etc., are important.
Referring to
Fig. 7.13
,
the laser scanner scans the seam in front of the moving torch and locates the
seam with respect to the torch. Although the laser signal produces an adequate indication of this location
most of the time, there are occasions when this method may become unreliable or fail. One situation is
where isolated tack welding points along the welding seam cover the seam line, resulting in a laser signal
that confuses the tracking system. Another cause for tracking errors occurs when the junction between
the joined parts (e.g., for butt welds) is faint, or blends in with the parent metal surfaces, so that the
seam apparently disappears over a section of the joint. Previously used data analysis algorithms for
processing the output of the laser scanner sometimes became unreliable when the signal indicating the
seam location became degraded in the manner discussed. The ability of the neural network to ignore
minor disturbances made it an ideal candidate for this purpose.
A weld bead profile is obtained from the laser scanner where it is available as a list of coordinates.
Typically, the list contains on the order of 80 coordinates, obtained for a fixed cross section of the weld
profile. Each coordinate consists of the location of the measured point, as measured transversely across
the bead width (
y
-axis value) and the height of the bead at that location (
z
-axis value). To locate, for
example, the crown, undercuts and parent metal boundaries of the bead, the entire 80 data point heights
(
z
-values) are fed into the network as distinctive inputs (refer to
Fig. 7.14
).
Based on this given profile
the network determines the locations of the desired bead profile parameters and presents these values as
five separate outputs. Given the locations of the crown, undercuts, and bead boundaries, the crown height
and undercut levels are easily determined by looking up the
z
-axis values at these locations. Deriving
other properties of the bead, such as symmetry of the undercuts, both in terms of location and magnitude,
is straightforward once the above information has been obtained.
Referring to
Fig. 7.14
,
the height values of the digitized profiles were entered into the 80 inputs of the
neural network, which was trained to locate the five features of the profile and present these locations
at the outputs.
Figure 7.15
illustrates a typical cross-sectional weld profile. The objective of the neural
network was to find the locations of the weld crown, the left and right undercut, and the left and right
weld boundaries. To the human observer, these features are relatively clear, with the possible exception
of the weld boundaries. The locations of these features, as determined by the neural network, are indicated
