Deep learning for image-based defect detection in additive manufacturing

Abstract

Additive manufacturing, or 3D printing, has revolutionized industrial production and manufacturing by making it possible to produce more advanced and tailor-made components. But the constant push for a first-pass quality always comes with significant challenges in the industry because defects always creep in some shape or form. This results in wasting materials and raising production costs. Traditional fault detection depends on manual examination, which can be time-consuming and subject to human error. In this project, we attempt to go the route of deep-learning model training for the real-time defect detection in 3D printing, to detect defects like Stringing, Spaghetti, and Warping. In this project, the performance of YOLOv8, YOLOv11, and YOLO-NAS object detection models will be evaluated for detection as well as the classification task. These models were trained and tested on the developed dataset, and metrics of precision, recall, and mean average precision (mAP) were used to evaluate performance. The findings reveal that although all models show promise for real-time defect detection, differences in accuracy and computational efficiency underscore the trade-offs inherent in various architectures. These findings continue to advance the state of the art in automated quality assurance for additive manufacturing, providing enhanced diagnostics and defect mitigation for industrial applications.