Industrial Camera-Based Powder Bed Defect Detection in Laser Powder Bed Fusion: Dataset Construction and Model Evaluation

Laser Powder Bed Fusion (LPBF) is a crucial additive manufacturing technique that builds complex geometries by selectively melting metal powders. However, it faces challenges from defects such as ditches, stacking, insufficient spreading, and craters, which can affect the mechanical properties and quality of the final product. To enhance manufacturing quality and stability, we developed a specialized dataset for detecting these defects, comprising 420 images of ditches, 800 of insufficient spreading, 657 of craters, and 600 of stacking. We evaluated 16 state-of-the-art deep learning models on this dataset and implemented data augmentation techniques, including rotation, scaling, and noise addition, to increase the dataset size fivefold to 12,385 images. Our results showed that data augmentation significantly improved model performance. The YOLOv10 series excelled in post-processing speed at 0.7 milliseconds, while YOLOv8-n had strong inference capabilities at 12.4 milliseconds. YOLOv9-t achieved the fastest preprocessing at 0.3 milliseconds, and YOLOv9-e attained the highest mAP@0.50 score, with precision and recall rates of 0.965 and 0.975, respectively. This study provides a high-quality dataset for powder bed defect detection and validates the efficacy of advanced deep learning models in this field.