Embedded process monitoring of stringing defects in fused filament fabrication for intelligent manufacturing

Stringing defect in fused filament fabrication (FFF) is a critical process monitoring challenge for intelligent manufacturing, which adversely affects dimensional accuracy and surface quality of printed components. This work presents an embedded, vision-based defect detection system for real-time classification of stringing defects in FFF processes. A defect-specific image dataset comprising 1379 annotated images was generated using a design of experiments approach to systematically induce stringing under controlled variations of key process parameters. Visual data were gathered using a dual-camera setup integrated with a Raspberry Pi 4 that enables in-situ image capture and on-device inference. A lightweight convolutional neural network was custom-designed to achieve high classification accuracy while maintaining low computational complexity for embedded deployment. The proposed system achieved a validation accuracy of 93.83%, with class-wise precision, recall, and F1 scores of 0.94, which demonstrates strong discrimination between stringing and non-defective prints. The trained model was quantized and deployed using TensorFlow Lite, achieving stable real-time performance with low latency on embedded hardware. The results demonstrate the feasibility of cost-effective, on-device vision-based quality monitoring for desktop-scale FFF systems, providing a foundation for early-stage defect detection in additive manufacturing.