Vision-Based Weld Pool Surface Geometry Measurement System for Pulsed-GMAW Process

Measurement of the three-dimensional weld pool surface geometry in GMAW process provides an essential approach to understand and future control of the complex welding process. Precisely measured weld pool surface geometry during the process also provides critical boundary conditions to evaluate and validate numerical welding process models.

This study demonstrates a vision-based weld pool surface geometry measurement system for pulsed-GMAW process. A 375mW 4-by-4 grid laser pattern was projected onto the liquid weld pool surface during the welding process. The laser reflected from the weld pool surface was intercepted by an imaging plane and was captured. Three-dimensional surface geometry is successfully reconstructed using an iteration-based reconstruction algorithm based on captured pattern. By processing sequential images obtained in the experiment, the system is capable of capturing the dynamics of weld pool surface.

Reconstruction was achieved by exploiting law of reflection and the correlation established between projected pattern and captured reflected pattern. Correspondence was established between captured grid intersection points and the projected interception points. Corresponding surface normal was calculated. A virtual surface is computed using assigned surface normal field and resultant reflection pattern from the computed surface was calculated and compared with the acquired pattern. The assigned surface normal was adjusted accordingly based on the displacement, a new virtual surface was computed until the calculated and captured patterns converge. The reconstruction accuracy of the proposed weld pool surface geometry measurement system is affected by the weld pool fluctuation and image capturing frame rate.

Experiments were conducted in bead-on-plate experiment using pulsed GMAW process on 1/8 inch mild steel sheet. Pulse frequency was 8.7 Hz, average current 106 amps, peak current 220 amps and base current 50 amps. Reconstruction algorithm error was evaluated using simulation data and the result demonstrated that the error was within 10%.