Video Segmentation of Wire + Arc Additive Manufacturing (WAAM) Using Visual Large Model

Area of Science:

• Manufacturing Engineering
• Computer Vision
• Materials Science

Background:

• In-process monitoring videos are crucial for quality assurance in wire + arc additive manufacturing (WAAM) and automated welding.
• Accurate segmentation of video data is needed for feedback control, but is challenging due to fluctuating arc brightness and limitations of conventional methods.
• Deep learning for WAAM video segmentation is hindered by the high cost and time required for dataset creation.

Purpose of the Study:

• To develop a semi-automatic annotation tool for WAAM and welding videos to overcome dataset creation challenges.
• To enable rapid quantitative analysis of WAAM and welding videos with minimal user intervention.
• To demonstrate the effectiveness of the tool in practical applications.

Main Methods:

• Development of a semi-automatic annotation tool integrating the foundation model SAM and the video object tracking model XMem.
• Utilizing the tool for significantly faster video frame annotation compared to manual methods.
• Demonstrating the tool's application in closed-loop control, droplet transfer analysis, and dataset assembly.

Main Results:

• The developed tool accelerates video annotation hundreds of times faster than traditional manual methods.
• The tool facilitates rapid quantitative analysis of WAAM and welding videos.
• Successful demonstration of the tool's utility in three distinct use cases.

Conclusions:

• Large computer vision models offer a viable solution to the challenges of WAAM video segmentation.
• The developed semi-automatic annotation tool significantly reduces the time and cost associated with creating annotated datasets.
• This approach provides a pathway for broader adoption of deep learning in WAAM and welding process monitoring.