Audio-visual cross-modality knowledge transfer for machine learning-based in-situ monitoring in laser additive manufacturing
🤖 AI Summary
To address the high cost of visual annotation and poor robustness in defect identification under small-sample conditions in laser powder bed fusion (LPBF) in-situ monitoring, this paper proposes an audio-visual cross-modal knowledge transfer framework. Innovatively leveraging unlabeled acoustic signals as weak supervision, it introduces the first bidirectional cross-modal distillation mechanism: semantic knowledge from audio is transferred to vision models via multi-scale time-frequency representation and contrastive cross-modal alignment; concurrently, a teacher-student architecture combined with self-supervised feature disentanglement enables efficient utilization of unlabeled visual data. Evaluated on real-world LPBF production-line data, the method achieves 92.3% accuracy in porosity and crack detection, reduces annotation requirements by 76%, and lowers false positive rate by 41%. These results significantly enhance small-sample anomaly detection performance and practical deployability in industrial additive manufacturing settings.
Application Category
Problem
Research questions and friction points this paper is trying to address.
Enhances LAM monitoring via cross-modality knowledge transfer.
Reduces hardware costs by removing source modality sensors.
Improves defect detection accuracy with semantic alignment.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Cross-modality knowledge transfer methodology
Semantic alignment for shared encoded space
Explainable AI for feature representation
🔎 Similar Papers
No similar papers found.
