Cross-Process Weld Penetration Prediction via Unsupervised Domain Adaptation for Laser and TIG Welding

Forum topic · 小凯 · 2026-06-26

Summary

A research paper (arXiv:2606.19223) by Sen Li, Haichao Cui, and Chendong Shao proposes an unsupervised domain adaptation (UDA) framework combined with a gradual source domain expansion (GSDE) strategy for weld penetration state classification in laser and TIG welding. Supervised deep learning models typically degrade under domain shift when transferred between welding processes with distinct physical mechanisms, such as from arc-dominated TIG welding to keyhole-based laser welding. The proposed method achieves 90.65% on TIGFH and 90.72% on LSPS in same-process settings, exceeding a supervised baseline by 35.83% and 38.87%. In cross-process transfer, it reaches 80.48% (TIG-to-laser) and 81.13% (laser-to-TIG), improvements of 43.39% and 43.40% over the baseline. UMAP visualization confirms the model learns domain-invariant features while preserving discriminative class boundaries, reducing relabeling costs for new welding processes and improving generalization of intelligent weld monitoring across systems.

Overview

Field: Computer Vision
Authors: Sen Li, Haichao Cui, Chendong Shao
Published: 2026-06-25
arXiv: 2606.19223

Key Points

Supervised deep learning has been widely used for weld penetration state classification; however, its performance often degrades significantly under domain shift, such as when transferring models between welding processes with distinct physical mechanisms — for instance, from arc-dominated tungsten inert gas (TIG) welding to keyhole-based laser welding.

To overcome this limitation, the authors propose an unsupervised domain adaptation (UDA) framework integrated with a gradual source domain expansion (GSDE) strategy.

Results

Evaluated on dedicated TIG and laser welding datasets, the approach achieves high accuracy in both same-process and cross-process transfer tasks:

Same-process transfer: average accuracies of 90.65% on TIGFH and 90.72% on LSPS, surpassing a supervised baseline by 35.83% and 38.87%, respectively.
Cross-process transfer: 80.48% for TIG-to-laser and 81.13% for laser-to-TIG, improving over the baseline by 43.39% and 43.40%.

Analysis

UMAP visualization verifies that the model learns domain-invariant features while maintaining discriminative class boundaries. This approach significantly reduces relabeling costs for new welding processes and enhances the generality of intelligent monitoring across different welding systems.

---

*Auto-collected on 2026-06-26.*