Deep Learning-based Characterization of Fused Filament Fabrication from Temporal Thermal Data

Abstract

The stability of fused filament fabrication (FFF), crucial for achieving high production efficiency, is considerably affected by the manner in which temperature distributes and propagates throughout the printed region. The application of infrared (IR) thermal cameras for the collection of spatiotemporal thermal data during strand deposition is investigated in this research. A classification framework utilizing the X3D deep learning model was developed to categorize process states as Normal, Transition, or Failure. The X3D model demonstrated high reliability in distinguishing Normal from Abnormal states with an accuracy of approximately 95% and a weighted F1-score of 0.95, although performance in categorizing Transition from Failure states was constrained to around 65%. The effectiveness of the model in discerning stable and unstable process conditions has been affirmed through the findings, resulting in the provision of a practical tool for real-time quality assurance in FFF. In addition, the X3D model showcased exceptional computational efficiency, processing a 1-second IR video clip (32 Hz) in 6 milliseconds while utilizing only 0.67 GB of GPU memory, making this setup suitable for in-process monitoring and control.