Physics-guided spatiotemporal deep learning for early melt pool instability detection in dual-beam laser welding of steel–titanium joints

The concept of dual-beam laser welding is the new method of welding dissimilar materials like steel and titanium, with a better process stability and quality of the joint, in contrast with single-beam methods. This paper presents a deep learning-based model to classify melt pool instabilities during high-speed thermal imaging of steel-titanium joints of high-speed laser welding. Infrared image sequences with a high frame-rate of changing melt pool dynamics are obtained over a variety of conditions of the process. Spatiotemporal deep learning architecture that is based on convolutional neural networks to extract spatial features and recurrent temporal models learning is utilized to learn discriminative thermal patterns of various instability modes. The proposed model is trained with a categorical cross-entropy criterion and is tested with extensive measures, such as accuracy, F1-score, and receiver operating characteristic (ROC)-area under the curve (AUC). The experimental findings indicate strong classification and all instability classes with the F1-scores of over 88% and the AUC of over 0.90. Absorption-informed feature inclusion gives significant increases in first instability detection and general discriminative ability over baseline feature representations. The results confirm that deep learning using high-speed thermal imaging can be effective to model the high-level spatiotemporal signatures of the melt pool instabilities and this forms a promising direction to real-time monitoring and closed-loop control in advanced laser welding systems.