OM
O.T. Maddock
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Dual-colour fused deposition modelling using a rotational nozzle can produce continuous colour gradients from a single dual-colour filament, but the colour deposited at any point is a direct function of the nozzle's rotational angle, so small orientation errors translate into visible, uncorrectable colour misregistration. This thesis investigates the extent to which computer vision and machine learning can detect filament colour orientation in real time and use it to drive closed-loop correction of nozzle rotation. A nozzle-mounted camera feeds a convolutional neural network that predicts orientation directly from the deposited filament's colour distribution; the estimate drives corrective G-code between toolpath segments. Under in-distribution conditions, the final model (Dataset v3) detected orientation accurately (mean absolute error 1.73°, 90th-percentile error 3.61°, 100% direction accuracy), but this did not transfer to unseen geometries (MAE 11.09°, P90 19.97°). Mechanistic analysis showed that the model infers orientation from nozzle-body reflectance and surrounding geometry, rather than solely from the colour boundary it was designed around. This behaviour propagated to closed-loop performance: large resolvable errors were corrected, confirming the control architecture is sound, but errors near the filament's resting state were imperceptible, and sustained non-equilibrium orientations drove the filament into states absent from training, causing the loop to act on confidently incorrect estimates. Closed-loop performance is therefore governed not by the control loop but by the reliability of perception under untrained conditions.
...
Dual-colour fused deposition modelling using a rotational nozzle can produce continuous colour gradients from a single dual-colour filament, but the colour deposited at any point is a direct function of the nozzle's rotational angle, so small orientation errors translate into visible, uncorrectable colour misregistration. This thesis investigates the extent to which computer vision and machine learning can detect filament colour orientation in real time and use it to drive closed-loop correction of nozzle rotation. A nozzle-mounted camera feeds a convolutional neural network that predicts orientation directly from the deposited filament's colour distribution; the estimate drives corrective G-code between toolpath segments. Under in-distribution conditions, the final model (Dataset v3) detected orientation accurately (mean absolute error 1.73°, 90th-percentile error 3.61°, 100% direction accuracy), but this did not transfer to unseen geometries (MAE 11.09°, P90 19.97°). Mechanistic analysis showed that the model infers orientation from nozzle-body reflectance and surrounding geometry, rather than solely from the colour boundary it was designed around. This behaviour propagated to closed-loop performance: large resolvable errors were corrected, confirming the control architecture is sound, but errors near the filament's resting state were imperceptible, and sustained non-equilibrium orientations drove the filament into states absent from training, causing the loop to act on confidently incorrect estimates. Closed-loop performance is therefore governed not by the control loop but by the reliability of perception under untrained conditions.