This study investigates a multi-sensor approach to estimate the surface thermal fields of Fused Filament Fabrication (FFF), an additive manufacturing process, at high spatial (4 μm) and temporal (0.1 msec) resolutions. The current state-of-the-art methods for quantifying the thermal, mechanical, and chemical state of the FFF processes are categorized into two main groups: numerical simulations and in-process thermal monitoring. Although numerical simulations can precisely estimate the process state for simple components, they are computationally demanding, become less accurate when complexity increases, and are therefore not suitable for real-time monitoring. In-process thermal monitoring, on the other hand, is becoming vital to track the process’ state, and offers a means to discern the underlying phenomena. While the current thermal sensors offer sub-mm scale spatial resolutions, their temporal resolutions (at best 1 msec) do not allow tracking of the thermal gradients necessary to determine the build geometry, microstructure, and the properties of the components. They are also limited by the cost and size, and need custom fixtures for mounting to a printer. The study proposes an approach based on pairing a primary high-tech thermal sensor with a secondary small footprint vibration sensor using a machine learning model to reconstruct the thermal fields at high time-resolutions from vibration signals. The key steps of this novel ML pipeline consisted of a singular value decomposition (SVD) of the ground truth of the thermal images and a time–frequency transformation with fast Fourier of the vibration signal. The overall model was able to make inferences and predict with 91.36 % accuracy the thermal state of the process using only vibration data as input. In addition, the ML pipeline filtered part of the experimental error contained in the ground truth data.
Read full abstract