Reducing Failed Prints in Fused Filament Fabrication using Real-Time Vision-Based Closed-Loop Control

Abstract

Fused filament fabrication can produce useful parts with little tooling and modest equipment costs. However, unnoticed extrusion faults can ruin long builds and waste material, energy, and machine time. This paper examines whether real-time vision can support safe closed-loop control during printing. A secondary-data design was selected because no new physical printer tests were available. Public datasets and experimental studies supplied the evidence. The analysis covers nozzle images, layer images, defect labels, model accuracy, inference speed, and tested control actions. One major dataset contained 946,283 cleaned images from 192 prints. A second dataset contained 776 unique monolayer images and a separate 120-image test set. Published systems reported test performance between 84.3% and 100%. Yet these values came from different tasks and cannot be treated as direct rankings. The evidence supports a two-view control design. A nozzle camera should detect recoverable flow and height errors. A wider camera should detect warping, detachment, layer shift, and uncontrolled extrusion. Temporal voting should confirm each warning before control begins. Recoverable errors should receive small parameter changes. Dangerous or uncertain errors should pause the machine. Firmware acknowledgement should prevent repeated commands and unstable control. An analytical model shows that early detection can protect much of the material assigned to failed builds. Actual savings will depend on recall, detection time, and defect recoverability. The proposed design is practical, evidence-based, and suitable for later printer trials. Physical validation remains necessary before industrial use.