Robust pattern recognition for measurement of three dimensional weld pool surface in GTAW

Abstract

Ability to observe and measure weld pool surfaces in real-time is the foundation for next generation intelligent welding that partially mimics welders' sensory capability, i.e., acquiring weld status information from the observation of weld pool. To image and measure the mirror-like weld pool surface under the strong arc radiation, a structured light laser pattern, dot matrix, has been projected onto the weld pool surface and its specular reflection is intercepted and imaged by an imaging plane placed with a distance from the arc. The reflection pattern, deformed by the specular liquid weld pool, contains three-dimensional (3D) geometry information of the weld pool. In this paper, a robust recognition procedure has been proposed to identify the reflection pattern in real-time. In particular, an adaptive thresholding algorithm is proposed to distinguish the laser dots in the reflection pattern. Then a reflection pattern recognition algorithm is proposed to determine the row and column number for each reflected laser dot such that reflection pattern can be matched with the projection pattern. The identified reflection pattern can be used to reconstruct the 3D weld pool surface. Experiments with different welding conditions have been conducted to verify the real-time performance of the proposed procedure, including the effectiveness, robustness and time complexity. It has been found that the procedure is capable of identifying the reflection pattern from a captured image in less than 19 ms. The real-time performance meets the most widely used arc welding process for precise welding-gas tungsten arc welding (GTAW) that typically requires to be controlled a few times per second.