High speed video-based melt pool surveillance in laser spot welding

Laser welding processes are today increasingly observed with cameras. To this end, systems are employed which mostly observe the laser-induced process radiation. Algorithms to evaluate such acquired images are under development, but they only indirectly allow to draw conclusions about the current state of the melt pool, thus leading to corresponding uncertainty. To observe the melt pool directly, we previously tested an approach for monitoring it coaxially with additional illumination. The thus gained information is much more meaningful as it relies on the measurement of geometrical information. These directly measured data are much better suited to derive conclusions about the actual process state.Furthermore, applicable real time approaches with throughput rates of 5000 frames per second and beyond are yet difficult to realize especially depending on the type of information to be monitored. We therefore developed a framework which complements the required throughput rates with a high degree of fault tolerance for determining the desired melt pool parameters such as position, size, size evolution, and a sputter count. The underlying principle is a robust contour point detection method, combined with a smart outlier regression and an algorithm for fitting a circle model to the melt pool. Laboratory results are presented confirming the proposed procedure.Laser welding processes are today increasingly observed with cameras. To this end, systems are employed which mostly observe the laser-induced process radiation. Algorithms to evaluate such acquired images are under development, but they only indirectly allow to draw conclusions about the current state of the melt pool, thus leading to corresponding uncertainty. To observe the melt pool directly, we previously tested an approach for monitoring it coaxially with additional illumination. The thus gained information is much more meaningful as it relies on the measurement of geometrical information. These directly measured data are much better suited to derive conclusions about the actual process state.Furthermore, applicable real time approaches with throughput rates of 5000 frames per second and beyond are yet difficult to realize especially depending on the type of information to be monitored. We therefore developed a framework which complements the required throughput rates with a high degree of fault tol...