Deep Learning-Based Multi-Sensor Fusion for Process Monitoring: Application to Fused Deposition Modeling

Abstract

In the realm of additive manufacturing, process monitoring is typically realized through multi-sensor data fusion (MSDF) employing either classical methods like Kalman filtering or methods powered by artificial intelligence. In the latter approach, standard machine learning-based methods that involve handcrafted feature extraction and signal processing face challenges in generalization due to missing signal information and require domain expertise in data and feature selection. This study investigates MSDF in process monitoring from a signal-to-image encoding perspective within deep learning (DL), where intelligent fusion models are developed in different fusion levels, namely data, feature, and decision levels. Various signal imaging encoders, namely Gramian angular field, Markov transition field, and recurrence plots, are adopted and tested for fusion at these three levels. The fusion algorithms are implemented through three different DL-based classifiers, spanning different capacities and architectures recently established in this domain. The developed fusion frameworks are applied to the problem of process monitoring and anomaly detection in fused deposition modeling, utilizing a sensor dataset collected from a Delta 3D printer. Overall, the results indicate that the highest accuracies (up to 99.6%) can be achieved when employing feature-level fusion through a hybrid convolutional and recurrent deep model trained using recurrence plot anomaly images. Conversely, all data-level fusion models offer lower computational time at the cost of a slightly decreased accuracy. Considering the models’ response to various malfunctioning or glitching scenarios, once again, the feature-level fusion demonstrates outstanding stability and robustness, effectively attenuating considerable corruptions in the input signals without requiring model adjustments.