Enhancing defect detection model with frequency dynamic convolution and attention mechanism for LPBF additive manufacturing

Abstract

Laser Powder Bed Fusion (LPBF) often produces defects such as overlap, spatter, and fusion anomalies, which seriously affect part quality. Real-time in-situ defect detection remains difficult due to low contrast, multi-scale defects, and complex backgrounds. This paper proposes an improved YOLOv11-based detection model. We integrate Frequency Dynamic Convolution (FDConv) into the backbone and neck to enhance frequency-adaptive feature extraction. A Parallelized Patch-Aware Attention (PPA) module is added to the detection head for better multi-scale feature representation. A multi-energy condition dataset with 2,557 annotated defects is built for training and validation. Ablation experiments show that the proposed model increases mAP@0.5 by 5.1% compared with the baseline. Comparative results demonstrate that our model achieves 0.849 mAP@0.5, 0.775 recall, and 109 FPS, outperforming mainstream YOLO models while maintaining real-time performance. This model provides an effective solution for industrial LPBF in-situ quality control.