In-line laser weld surface inspection and abnormality detection via a human-knowledge integrated hybrid framework using 3D point cloud measurements

Abstract The current research contributes toward developing a Machine Learning (ML) framework for surface quality monitoring and abnormality detection in the laser welding process using 3-dimensional (3D) point cloud data. In this research, an enhanced segmentation framework for large aspect ratio images based on You Only Look Once (YOLO)–based segmentation model serves as the foundation network to automatically identify and isolate weld path regions from surface images in a single forward pass. The model, fine-tuned using a custom-annotated dataset of weld path images, achieved a high Intersection-over-Union (IoU) score of 0.93 and an F1-score of 0.96, demonstrating its strong localization and detection performance under varying surface conditions. The extracted regions are subsequently mapped to their corresponding 3D point cloud representations, which are analyzed through machine learning and mathematical surface-fitting methods. A reference plane is established based on weld surface morphology to quantify geometric deviations and identify abnormality regions such as concavities, convexities, and discontinuities along the weld path. The proposed abnormality regions were validated against leak-test data containing large and medium surface leaks. The comparison results confirm that this method can detect large leaks with an accuracy of approximately $$95\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>95</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> . This study underscores the potential of integrating 3D point cloud analytics with machine learning for reliable, automated quality inspection in laser welding applications.