A Continual Learning Approach for Adaptive Error Detection in 3D Printing

Despite its wide use, material extrusion additive manufacturing still presents accuracy and speed challenges, which translate into a range of different defects, and ultimately scrap parts. To address this, data-driven approaches based on machine learning for error detection and process control have been proposed. These methods rely on large amounts of sensor and vision data during training to achieve high robust-ness and generalisation. While this may be effective for a given process configuration, methods that are adaptable to new conditions are more desirable, given the need for high customisation of products, batch-of-one demands and often complex part geometries. In these conditions, large amounts of data is less likely to be available. Instead, it is necessary to be able to build robustness over time, without having to retrain a model from scratch with an ever growing data set. In this paper, a continual learning approach to 3D printing temperature error detection based on Elastic Weight Consolidation is proposed, where a model learns to detect errors in an incremental manner as new temperature conditions are presented. Preliminary experimental results show the ability to learn and adapt to temperature changing conditions in a resource efficient manner, while retaining knowledge of previously known conditions.