Gcs-Unet: A lightweight attention network for coaxial melt pool monitoring in laser powder bed fusion

In laser powder bed fusion (L-PBF), coaxial melt pool monitoring methods based on spontaneous radiation often miss low-radiation regions such as the trailing edge, resulting in incomplete information. Additionally, traditional image processing techniques like threshold segmentation lack robustness under complex backgrounds caused by auxiliary lighting, limiting their effectiveness for real-time applications. To address these challenges, a new coaxial melt pool monitoring system was developed, providing clearer and more comprehensive images that capture both geometry and texture. Building on this foundation, an attention-enhanced deep learning network, Gcs-Unet, was proposed to enable robust semantic segmentation under complex conditions. The proposed model achieved an inference time of 6.75 ms while maintaining high performance (99.5 % accuracy, 87.6 % Dice, 86.2 % mIoU) and reducing parameters by 42.91 %, meeting real-time deployment requirements. Furthermore, it was found that scanning speed significantly influences melt pool behavior, with a 33 % speed increase resulting in a 37.45 % rise in the variation of the high-temperature zone's center. These results provide strong support for process optimization and melt pool analysis in L-PBF.