A novel few-shot learning based feature relation model for robotic welding states monitoring

Research question

How can welding states such as penetration, burn-through and misalignment be recognized when new labeled examples are scarce?

Method

Hybrid supervised pre-training uses unlabeled pool images together with labels for common states. Cross-attention and Brownian distance covariance refine feature relations before comparing a small support set with incoming images.

Main findings

After the reported pre-training, the robotic GTAW five-state experiment achieved 96.5% classification accuracy with 15 labeled support examples per class. The result demonstrates a way to reduce new-state labeling by reusing learned pool features.