Li et al.'s deep neural network predicts laser welding with 99.35% accuracy
99.35% penetration state accuracy and 1.79mm depth error from weld pool images
Sen Li et al. from an undisclosed institution (submitted to arXiv) introduced a multi-task spatiotemporal deep neural network for laser penetration welding quality assessment. The model uses top weld pool images captured by a complementary metal-oxide-semiconductor (CMOS) camera and combines them with welding parameters. The architecture leverages convolutional neural networks (CNNs) for spatial feature extraction and state space models (SSMs) for temporal dynamics, enabling efficient processing of spatiotemporal information. The researchers also devised a reliable dataset construction method to boost robustness and generalization.
Validation on test data showed outstanding performance: 99.35% accuracy in classifying penetration state, a mean absolute error of only 1.79mm for penetration depth prediction, and 95.65% accuracy in reconstructing the weld cross-section morphology. This approach enables real-time, non-destructive monitoring of weld quality during the laser welding process, potentially reducing defects and rework in industrial manufacturing. The paper is published in arXiv (cs.CV/2606.26260) and has a DOI linking to Engineering Applications of Artificial Intelligence.
- Model integrates CNN and state space models to process spatiotemporal features from weld pool images
- Achieves 99.35% penetration state classification accuracy and 1.79mm depth prediction error
- Weld cross-section reconstruction accuracy reaches 95.65%, enabling non-destructive quality control
Why It Matters
Enables real-time, in-situ quality control for laser welding, reducing defects and manual inspection costs.