Real-time defect detection in fused filament fabrication using computer vision and deep learning

Purpose The study aims to develop an intelligent and real-time defect detection framework for Fused Filament Fabrication (FFF) to address limitations in process reliability and quality control. It focuses on establishing a multi-model deep-learning system capable of identifying nozzle, first layer and object-level defects through synchronized visual monitoring and inference. Design/methodology/approach A modular You Only Look Once (YOLO)v8-based architecture comprising five coordinated models was implemented: keypoint detection, instance segmentation and three stage-specific classification networks. The system follows an edge-cloud deployment logic, where a Raspberry Pi camera acquires and preprocesses images, while a GPU host executes inference and graphical user interface visualization. Real-time performance was evaluated across multiple print conditions and results were benchmarked against recent literature. Findings The integrated system achieved 90–99% accuracy (mAP50 ∼ 0.98), demonstrating high detection precision across all print stages. Comparative analysis with the state-of-the-art studies confirmed comprehensive process-wide coverage, hardware efficiency and modularity, enabling scalable deployment in smart manufacturing settings. Originality/value Unlike existing single-stage or single-model frameworks, the proposed approach introduces a context-aware, multi-model pipeline that integrates deep learning with edge-cloud collaboration. The framework represents a deployable and cost-effective solution for intelligent quality monitoring in additive manufacturing and serves as a basis for future closed-loop adaptive control in self-correcting 3D printing systems.