On-line sensing and estimation of laser surface modification by computer vision

Laser surface modification with alloying or remelting often yields unstable results. To achieve on-line control, process sensing and estimation are the key technologies. This paper introduces molten pool imaging as laser processing feedback. Well-contrasted molten pool images are acquired in experiments by eliminating the strong light from spatter and plasma. An algorithm for real-time image processing is developed that shows robust and high-speed performance. Analytical models of laser surface modification are studied using moving heat source models. The simplified analytical model shows a roughly linear relationship between laser power and depth of modified surface (molten depth). Based on key feature analysis in analytical models, an on-line estimation model of molten depth is built using a neural network, which applies a time- series of widths of the molten pools as an input vector. By controlling laser power, the neural network model is trained for different heat inputs in a transient process. The testing results by another group of experiments show that the on-line estimation model can predict the depth of the modified surface accurately.