| Abstract Scope |
Fused filament fabrication machines are prone to mistakes during printing, which significantly impacts the functionality and reliability of printed parts. To address these challenges for mission critical parts, we propose a novel solution using a deep learning architecture for defect detection. To achieve high-accuracy classification of print anomalies, we train our model on a large thermal imaging dataset, extracting features across multiple error classes, including temperature variations associated with various defects. We begin by developing a deep learning model, such as VGG-16, as a strong baseline for image classification. We then integrate YOLOv5 for real-time localization of multiple defect types during printing, with a focus on the nozzle, the primary heat source. Finally, we incorporate Grad-CAM visualization to highlight the regions in thermal images that most influence the model’s decisions. This combined approach improves classification and localization accuracy, enabling intervention by identifying exactly where and why failures occur. |