Civil Engineering Reference
In-Depth Information
Weld
Profile
Weld Edge
Location
Undercut
Location
Edge
Edge
Undercut
Undercut
Cro
Crowwn
Crown
Location
Undercut
Undercut
Undercut
Location
Edge
Edge
Weld Edge
Location
75 Nodes
FIGURE 7.14
The neural network system used to locate the crown, undercuts, and weld boundaries from weld
profile data.
Cross Section No. 154
File "06180717.W91"
0.24
N
Neural Network
0.23
Previous Method
0.22
N
N
N
0.21
N
0.2
N
0.19
0.18
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
Distance Across Section, inches
FIGURE 7.15
A scanned weld profile, with the weld crown, undercuts, and parent metal boundaries indicated by
the proposed neural network and the previously used system.
on
Fig. 7.15
with the arrows marked ''
N
''. For comparison, the previously used data analysis algorithms
located the same features as shown with the arrows pointing from below the profile trace. In this case
the previous method failed in correctly identifying the right weld boundary, while the neural network
succeeded. In most cases the previous system performed satisfactorily, while occasionally it gave incorrect
results. The errors usually occurred in determining the weld boundaries, and these errors could be
significant. For comparison, the neural network performed somewhat better, on the average, and the
worst case errors were always much smaller than those of the previous system.
A total of 60 laser bead profiles were used to test and train the neural network used in the weld profile
analysis. Of the 60 profiles taken, 30 were used for training and 30 were used for testing the trained
network. Six of the testing profiles were sampled for error analysis, and the resulting errors were typically
less than 1%.
