ABSTRACT Defect detection in fused filament fabrication (FFF) is essential to ensuring part reliability, as printing faults such as layer shift, under‐extrusion, and poor adhesion can compromise structural integrity and increase material waste. However, existing monitoring approaches often lack the spatial resolution or robustness needed for real‐time, layer‐wise assessment. This study addresses this gap by investigating whether high‐resolution, full‐reference feature comparison can reliably detect catastrophic defects during printing. We propose a real‐time monitoring framework that captures 720 × 1,280 images of each printed layer using an onboard monocular camera and compares them with either simulated reference layers—generated from G‐code using a calibrated Blender environment—or reference images from a defect‐free print. Multi‐resolution embeddings extracted from three intermediate layers of a modified ResNet50 are compared using L2 feature‐space distance to localise anomalous regions. Across 24 experimental prints containing four major defect classes, the system achieved 100% detection of catastrophic defects with zero false positives under the tested conditions, while processing each layer in <2 s on a standard CPU. Limitations include camera blind spots, lighting sensitivity during early layers, and appearance mismatches in simulated references due to unmodelled thermal effects. Future work will incorporate thermo‐mechanical simulation, multi‐camera configurations, and real‐time corrective feedback. The proposed method provides a practical and efficient foundation for industrial FFF quality assurance and automated, in‐situ defect detection.
Ebrahimian et al. (Thu,) studied this question.