Fig 1.
Device for measuring weld surface parameters, which consists of a laser profile sensor, electric slide-in Z-axis, manual slide in Y-axis, and computer.
Fig 2.
Flowchart of the laser profile sensor-based welding seam surface profile parameter detection system.
Fig 3.
Schematic diagram of parameter definition of butt weld 4.
(a) Definition of weld reinforcement, width, and undercut. (b) Definition of weld misalignment.
Fig 4.
Schematic diagrams of welding seam parameter measurement index under different welding seam shapes.
(a) Normal weld. (b) Welds with the wrong sides. (c) Welds with single undercut and misalignment. (d) Welds with double undercut and misalignment.
Fig 5.
Network structure diagram of EDE-net.
Fig 6.
Deconvolution up-sampling mechanism diagram.
Fig 7.
Theoretical output results of EDE-net branch 1.
Fig 8.
Theoretical output results of EDE-net branch 2.
Fig 9.
Simulation of normal weld curve.
Table 1.
Simulation feature point position of normal weld contour curve.
Table 2.
Location of the feature point in the simulation of the undercut weld profile curve with the defect parameter.
Fig 10.
Simulation of weld profile curve with single defect parameter undercut Setun_cut.
Fig 11.
Simulation of weld profile curve with single defect parameter misalignment Setalign.
Table 3.
Tested CNN network information.
Fig 12.
Acquisition of contour images to conduct simulations to generate contour images.
(a) No defect parameter welds profile data set D3. (b) Weld profile data set D3 for parameter simulation of defects with a single misalignment error. (c) Weld profile data set D5 for parameter simulation of single undercut defect. (d) Weld profile data set D5 for parameter simulation of undercut and misalignment defect parameters. (The undercut is on the misalignment side.) (e) Weld profile data set D5 for parameter simulation of undercut and misalignment defect parameters. (The undercut is not on the misalignment side.) (f) Weld profile data set D7 with parameter simulation of undercut defects on both sides. (g) Weld profile data set D7 for simulation of the misalignment and undercut parameters on both sides (h) Actual collection data set D3. (j) Actual collection data set D5. (j) Actual collection data set D7.
Table 4.
Training hyper-parameter information.
Fig 13.
Loss function trend chart of different CNN backbones based on the encoding–decoding depth feature point extraction network.
(a) Comparison of transfer learning and non-transfer learning. (b) Comparison of loss functions of different network backbones. (c) Final convergence of the network.
Table 5.
AP results of feature points extracted from different CNNs.
Table 6.
Feature point extraction error information.