Deep learning with mixup augmentation for improved pore detection during additive manufacturing.

Area of Science:

• Materials Science
• Manufacturing Engineering
• Data Science

Background:

• Additive manufacturing (AM) processes like laser powder bed fusion (LPBF) are prone to defects such as keyhole pores, which compromise material quality.
• Predicting these defects is challenging due to imbalanced datasets where pore events are rare compared to non-pore events.
• Existing machine learning (ML) approaches struggle with sparse and imbalanced data common in AM monitoring.

Purpose of the Study:

• To develop an ML approach to address data imbalance in AM defect prediction.
• To investigate the effectiveness of data augmentation, specifically Mixup, for improving pore and non-pore prediction.
• To enhance the accuracy of ML models for identifying keyhole pores in LPBF.

Main Methods:

• Collected and registered X-ray radiography data of keyhole pores with acoustic emission measurements from an LPBF experiment.
• Utilized Mixup, a weak-supervised learning data augmentation technique, to address dataset imbalance.
• Trained Convolutional Neural Networks (CNNs) on both original and augmented datasets to predict pores.

Main Results:

• CNN models trained on augmented data achieved a 99% accuracy on test datasets, a significant improvement over models trained on original data (95% accuracy).
• Data augmentation with Mixup demonstrably enhanced the prediction performance for both pores and non-pores.
• Performance improvements were consistent across five different experimental trials, validating the robustness of the approach.

Conclusions:

• Data augmentation, particularly Mixup, is an effective strategy for overcoming data imbalance in AM defect prediction.
• The proposed ML approach significantly enhances the accuracy and reliability of identifying critical process defects like keyhole pores.
• Further research can explore optimal Mixup parameters to maximize CNN performance in AM monitoring contexts.