A cross-process welding penetration status prediction algorithm based on unsupervised domain adaptation in laser and TIG welding

Researchers have developed an unsupervised domain adaptation framework that enables deep learning models to predict weld penetration status across different welding processes without extensive relabeling. The approach achieves 80-81% accuracy in cross-process transfer between TIG and laser welding, significantly outperforming supervised baselines and reducing the cost of deploying AI systems to new welding environments.

This research addresses a critical challenge in industrial AI deployment: the brittleness of machine learning models when applied to different operational contexts. The paper demonstrates how unsupervised domain adaptation can bridge the gap between fundamentally different welding processes—arc-based TIG welding and keyhole-based laser welding—which operate through distinct physical mechanisms. Traditional supervised approaches fail dramatically in this scenario because labeled data from one process provides limited signal for another, but the proposed framework learns domain-invariant features that capture generalizable welding physics.

The manufacturing sector has increasingly adopted computer vision for quality control, yet the high cost of relabeling data for each new process or equipment variant has limited widespread adoption. This research directly tackles that bottleneck by reducing relabeling requirements while maintaining strong predictive performance. The 80%+ accuracy in cross-process scenarios represents a meaningful advancement for factories seeking to deploy unified monitoring systems across multiple welding technologies without rebuilding datasets from scratch.

The industrial implications are substantial. Smart manufacturing facilities operating heterogeneous equipment can now implement intelligent penetration monitoring more cost-effectively, potentially reducing defect rates and material waste. Equipment manufacturers could offer pre-trained models that require minimal customization for customer deployments. The gradual source domain expansion strategy also suggests the framework could scale to additional welding processes beyond the two tested.

Future developments should focus on real-time implementation on edge devices and validation across additional welding parameters and environmental conditions. The methodology may prove applicable to other manufacturing domains where process variations create domain shift challenges.

• →Unsupervised domain adaptation achieves 80-81% accuracy transferring weld penetration models between TIG and laser welding without extensive relabeling
• →Same-process performance reaches 90%+ accuracy, substantially surpassing supervised baselines by 35-39 percentage points
• →The framework reduces deployment costs for intelligent monitoring systems across heterogeneous manufacturing equipment
• →Domain-invariant feature learning enables generalization across physically distinct welding processes with different mechanisms
• →Approach has potential applications beyond welding to other manufacturing domains facing similar domain shift challenges