| About this Abstract |
| Meeting |
2024 AWS Professional Program
|
| Symposium
|
2024 AWS Professional Program
|
| Presentation Title |
Weld Penetration Prediction using GAN with Inaccurate Labels |
| Author(s) |
Yue Cao, Edison Mucllari, Yuming Zhang, Qiang Ye |
| On-Site Speaker (Planned) |
Yue Cao |
| Abstract Scope |
This work aims to train a generative adversarial network (GAN) for weld penetration prediction with automatically obtained but inaccurate labels. Accurate weld penetration measurement is vital for ensuring high-quality welding. Typically, supervised deep learning models for predicting penetration require large, accurately labeled datasets derived from backside observations of the workpiece, which demand excessive time and resources. A key principle is that the welding current, which influences weld pool formation, should mirror the distribution of weld penetration. Hence, the readily available welding current can serve as inaccurate labels for GAN training. The generator processes topside weld pool images and then predict current values, with the aim of replicating the actual distribution of current. After that, the easily established relationship between current and weld penetration can then be employed to restore the penetration from generated current values. Experimental results verified the effectiveness of the proposed method. |
| Proceedings Inclusion? |
Undecided |