Identification of solidification cracking using multiple sensors and deep learning in laser overlap welded Al 6000 alloy

Abstract

Solidification cracking, one of the most critical weld defects in laser welding of Al 6000 alloys, occurs at the final stage of solidification owing to shrinkage of the weld metal and deteriorates the joint strength and integrity. The filler metal can control the chemical composition of the weld metal, which mitigates solidification cracking. However, the chemical composition is difficult to control in autogenous laser welding. Temporal and spatial laser beam modulations have been introduced to control solidification cracking in autogenous laser welding because weld morphology is one of the factors that influences the initiation and propagation of solidification cracking. Solidification cracks generate thermal discontinuities and visual flaws on the bead surface. In this study, a high-speed infrared camera and a coaxial charge-coupled device camera with an auxiliary illumination laser (808 nm) were employed to identify solidification cracking during laser welding. Deep learning models, developed using two sensor images of a solidified bead, provided location-wise crack formation information. The multisensor-based convolutional neural network models achieved an impressive accuracy of 99.31% in predicting the crack locations. Thus, applying deep learning models expands the capability of predicting solidification cracking, including previously undetectable internal cracks.