About this Abstract |
| Meeting |
11th Conference on Trends in Welding Research + Additive Manufacturing (TWR+AM)
|
| Symposium
|
11th Conference on Trends in Welding Research + Additive Manufacturing (TWR+AM)
|
| Presentation Title |
Deep Learning Based System for Real-time Arc Welding Monitoring |
| Author(s) |
Molan Zhang, Jian Chen, Peijun Hou, Zhili Feng |
| On-Site Speaker (Planned) |
Molan Zhang |
| Abstract Scope |
Real-time weld monitoring from optical images is challenged by limited data, low resolution, and varying conditions. Authors proposed a deep learning-based system that simultaneously segments and extracts geometric parameters for melt pool and others, to provide quantitative feedback. An enhanced DDRN-Res23 network was developed for small datasets. To improve boundary precision and address class imbalance, a composite loss function combining boundary and focal loss was employed, along with extensive data augmentation and dynamic learning rates. This model achieved mean IoU of 0.96 for segmentation. An automated analysis module extracts melt pool geometry and wire tip offset relative to the melt pool centroid. Visual outputs include annotated images with overlays of elements boundaries, tip markers, and quantitative data. This paper showed that accurate real-time weld monitoring is achievable with limited data through improved network design and loss functions, providing a foundation for closed-loop control in automated welding. |
| Proceedings Inclusion? |
Undecided |