Multi-sensor deep learning framework for detection and severity estimation of nozzle clogging in pellet-based 3D printing

Additive manufacturing, particularly fused deposition modelling (FDM), is widely adopted for its ability to fabricate customised, complex geometries at low cost. However, FDM processes continue to face frequent defects and part failures, with reported failure rates of up to 20%. Nozzle clogging is among the most common failure modes, leading to weaker parts, dimensional inaccuracies, and print failure. Previous studies have primarily focused on clog detection in filament-based FDM, often simulating clogging via abrupt nozzle temperature changes or complete blockages, which do not accurately reflect realistic clogging conditions. Pellet-based FDM has recently emerged as a cost-effective alternative with broader material versatility, yet nozzle clogging remains unexplored. Unlike filament systems, pellet extrusion involves complex thermomechanical interactions in screw-driven flow. This limits the direct applicability of the existing clogging models developed for filament-based systems. The present study demonstrates a multisensory data-driven deep learning framework for detection and estimation of nozzle clogging in pellet-based FDM. Sensors are integrated to record feeder motor current, screw speed, nozzle vibration, and temperature. Clogging is systematically simulated using nozzles from 1.0 to 0.2 mm. A stacked long short-term memory (LSTM) network is trained for clogging severity classification. Two model configurations are evaluated: Model A (all nozzle diameters) and Model B (excluding one class—0.3 mm, 0.5 mm, or 0.8 mm) to test generalisation. All models achieved high accuracy (~ 92–94%) and reliably mapped unseen classes to adjacent known ones. A probabilistic diameter estimation framework was introduced based using class probability distributions. The error analysis showed that extended training improved the estimation precision and reduced bias. In addition, a regression-based model was also developed for direct diameter prediction (R² ≈ 0.9 on validation data) but showed weaker generalisation than the LSTM classifier. The developed LSTM framework demonstrates robust clogging detection, accurate severity quantification, and generalisation to unseen conditions. It can be integrated into a supervisory control system, where clogging predictions serve as feedback to initiate corrective actions during printing. Although the present work focuses on recycled ABS under fixed process parameters, this approach can be extended to other materials through transfer learning. Overall, the work highlights the potential of LSTM networks for time-series data driven process monitoring and defect detection. This provides an essential initial step toward AI-driven anomaly detection and closed-loop control strategies.