Leveraging Synthetic Data For Defect Detection In Metal Binder Jetting

Metal Binder Jetting is a promising additive manufacturing technology, but powder spreading defects can compromise part quality and cause process interruptions. This study presents a Deep Learning pipeline leveraging synthetic data to detect such defects in real time. For training this pipeline, a total of 1,107 real process images are used. First, a pixel-diffusion UNet2D model generates synthetic images, which are filtered using a VQ-VAE model to ensure realistic representations. To automatically label synthetic data classified as realistic, a Faster R-CNN model is trained, eliminating the need for manual labelling. Finally, a YOLO11 model, trained only on synthetic data set, is tested for in-situ defect detection. It could be shown that the YOLO11 model achieves a mean Intersection over Union value above 0.6, successfully detecting two pre-defined powder spreading defects. This approach significantly reduces the time and cost of data collection while enabling robust, real-time defect detection.