Small-Sample Visual Monitoring and Defect Analysis in FDM Additive Manufacturing Processes

Abstract

Addressing the issues in Fused Deposition Modeling (FDM) additive manufacturing, such as dimensional deviations, surface roughness, and structural defects caused by unstable material extrusion and temperature non-uniformity, and the difficulty of real-time monitoring and closed-loop control with traditional offline detection, this paper proposes and develops an online visual monitoring and defect recognition system for small-sample data. A lightweight meta-learning network architecture, DynaFDM-Net, is proposed. Its three-level linkage mechanism of "feature enhancement - rapid adaptation - cross-domain fusion" effectively addresses the challenge of high-precision defect recognition under small-sample conditions, enabling real-time identification of typical surface defects like layer lines, stringing/oozing, and warping deformation. Experimental results show that DynaFDM-Net achieves an overall recognition accuracy of 97.2% across three defect categories, with good performance in precision, recall, and F1-score for each defect category. Furthermore, combined with measurements from a surface roughness tester and a 3D measuring instrument, the system reveals the formation mechanisms and influence patterns of key process parameters, such as printing speed and layer thickness, on different types of defects. This research provides an effective technical solution for enhancing real-time quality monitoring and process optimization in FDM additive manufacturing processes.