In-situ monitoring and online prediction of keyhole depth in laser welding by coaxial imaging

Abstract

A comprehensive understanding of welding penetration and the role of process parameters is crucial for ensuring high-quality joints in laser welding. In-situ process monitoring can aid in detection of defects, reducing material usage and time-consuming inspection operations. In this study, we propose a novel approach for online prediction of keyhole depth in laser welding operations. Using in-process images captured with a coaxial camera and active illumination, our software employs pre-Trained CNNs from the EfficientNet and DenseNet families to extract features. These features serve as input for data-efficient regression models, trained to predict the keyhole depth. The results have shown that both methods yield percentage errors of approximately 3%. Ultimately, this methodology facilitates real-Time analysis of welding operations.