Process monitoring via prediction of laser beam absorptance in partial penetration laser keyhole welding of aluminum alloy

Abstract

Achieving high-quality laser keyhole welding requires precise control of energy absorption, which is inherently linked to the complex shape and dynamics of the keyhole. While absorptance-based monitoring methods have been successfully applied to full penetration welding, their application to partial penetration welding remains a challenge due to the limited visibility of the keyhole interior. In this study, we developed a deep learning-based absorptance monitoring system for partial penetration laser welding, utilizing a combination of object detection and CNN regressors. The system estimates keyhole aperture shape and penetration depth from melt pool images and constructs a two-channel keyhole representation to predict laser absorptance. The proposed method was trained and validated using melt pool image datasets collected from 14 fiber laser welding experiments on aluminum alloy 1050P-H16, covering laser power levels (538-739 W) and scanning speeds (8.2-15.4 mm/s). The deep learning model achieved high accuracy in predicting laser beam absorptance, with an R-squared value of 0.99706, closely matching ray tracing simulations. Additionally, three distinct absorptance patterns were identified, correlating with variations in penetration depth, weld bead shape, and defect formation. This study presents the first application of deep learning for absorptance prediction in partial penetration keyhole welding. The proposed framework enables real-time welding quality assessment and provides insight into hidden keyhole dynamics, addressing a critical gap in laser welding research.