Investigation of long short-term memory networks for real-time process monitoring in fused deposition modeling

To improve additive manufacturing (AM), from being limited to creating prototypes to fabricating functional parts with low failure rates and high precision, several challenges have to be addressed. In-situ AM process monitoring plays a central role in the quality assurance of fabricated parts during the actual build job. This paper studies the feasibility of utilizing long short-term memory (LSTM) networks in the real-time monitoring of fused deposition modeling (FDM) processes for detecting pre-specified types of anomalies. Several hybrid LSTM models are developed and validated using the time-series sensor data collected during the operation of a Delta 3D printer. In particular, two distinct approaches are tailored to pre-process the sensor data before being fed into the LSTM classifiers: (i) a handcrafted feature extraction method that leverages statistical abstractions of the data; and (ii) an intelligent feature extraction method through the so-called anomaly images, adopted from the human activity recognition literature. The results show that all image-based LSTM models are more reliable and robust against noise than the models trained using handcrafted feature extraction. The highest mean accuracy of the LSTM models is 99.85%, and the required time for a prediction task is sub-millisecond, making the model feasible for real-time process monitoring and anomaly detection. It is also shown that the proposed system substantially outperforms the traditional machine learning alternatives.