Characteristic classification and extraction of robotic multi-layer multi-pass MAG welding pool—An extended UNet network implementation based on transfer learning

Research question

How can changing pool contours be segmented and categorized as welding pass and torch posture vary in multi-layer MAG welding?

Method

A labeled multi-pass MAG pool-image dataset trains a transfer-learning UNet with a ResNet101 backbone. Geometric features are then used to formulate seven pool categories spanning triangular, trapezoidal and parallelogram-like forms.

Main findings

The custom pool dataset supported comparisons of segmentation networks and validation of seven geometric pool categories. The method turns pixel segmentation into interpretable contour and shape descriptors for subsequent process monitoring.