Deep learning object detection for optical monitoring of spatters in L-PBF

• Published in: Journal of materials processing technology 2023-10, Vol.319, p.118063, Article 118063
• Main Authors: Chebil, G., Bettebghor, D., Renollet, Y., Lapouge, P., Davoine, C., Thomas, M., Favier, V., Schneider, M.
• Format: Article
• Language: English
• Subjects: Additive manufacturing; Deep learning; Engineering Sciences; Image analysis; Physics; Process monitoring; Spatter
• ISSN: 0924-0136; 1873-4774
• DOI: 10.1016/j.jmatprotec.2023.118063

Laser-powder bed fusion process (L-PBF) is one of the most widespread additive manufacturing processes for metallic materials, where powder layers are successively melted with a laser beam to generate a 3D object. During L-PBF, the rapid extraction of thermal energy involves complex hydrodynamic behaviors inducing the ejection of metal particles (called “spatters”) from the laser-matter interaction. These spatters fall down on the powder bed and are known to create various defects, which might have a detrimental effect on surface roughness and mechanical properties. For the sake of understanding the complex spatter phenomenon and its consequences on the build quality, a novel global system called SP(AM)² (SPatter Analysis Method for Additive Manufacturing) was developed. It provided a new rapid and automated control method for additive manufacturing using deep learning, by combining in situ experimental study on an innovative instrumented L-PBF setup and camera-based monitoring. In this work, a traditional detection method was used to help in the complex labeling task required to train deep learning models to detect particles. A near real time deep learning architecture (YOLOv4) has been employed and tweaked to the peculiarities of spatter detection. Completed with a tracking stage, SP(AM)² made it possible to successfully study pollution rates in L-PBF process by determining the main characteristics of spatters (number, size, speed, ejection angle, landing area on the powder bed). These quantitative and qualitative information can be used to optimize L-PBF parameters and scan strategies to minimize the defects related to spatters. •Monitoring and tracking of spatters in AM is critical for quality.•A state-of-the-art deep learning technique is proposed to detect spatters.•Our new approach combines traditional CV (labelling) and YOLO for detection.•Our model is iteratively trained and outperforms existing techniques.•The system provides quantitative and qualitative data on spatter ejections.