Research on L-PBF Defect Detection Based on VAE-GAN Data Augmentation with Multi-Head Attention Mechanism

Abstract Additive Manufacturing (AM), particularly Laser Powder Bed Fusion (L-PBF), faces critical challenges in defect detection due to the scarcity of high-quality training data and severe class imbalance, which significantly degrade the accuracy of deep learning models. To address these issues, this study proposes a novel data augmentation framework combining geometric transformations with an enhanced Variational Autoencoder-Generative Adversarial Network (VAE-GAN). Traditional augmentation techniques (rotation, scaling, flipping) are first applied to alleviate sample imbalance, followed by the improved VAE-GAN to synthesize high-fidelity defect images, thereby enriching dataset diversity. Experimental results on an L-PBF defect dataset demonstrate significant improvements in detection performance: This study conducted a comparative experiment evaluating the performance of YOLOv4, YOLOv7, YOLOv8, SSD, and Faster R-CNN on defect detection tasks before and after data augmentation. The results demonstrated significant mAP improvements across all models, with YOLOv4 achieving the most substantial enhancement (+20.03%, from 65.18% to 85.21%) despite its lower baseline performance. Faster R-CNN attained the highest post-augmentation mAP (87.75%), representing the best overall performance. YOLOv8 exhibited an optimal balance between real-time processing and accuracy (67.07%→84.66%, +17.59%), approaching Faster R-CNN's performance level. While SSD showed the smallest improvement (+12.34%), it maintained a relatively high baseline mAP (73.99%). These results validate the effectiveness of the proposed method in overcoming data scarcity and improving defect detection accuracy in AM.