Variational Inference for Laser Disturbance Detection in Powder Bed Fusion

Anonymous

Keywords: Variational Inference, Machine Learning, Additive Manufacturing, 3D Printing, Dynamics

TL;DR: In this study we use variational inference to learn the unknown dynamics of a laser melting process then quantify the model’s ability to detect anomalous melting.

Abstract: In this study we use variational inference to learn a dynamics model from a high-speed video stream of a laser melting process. We compare two deep generative sequence models and evaluate them on video prediction and anomaly detection tasks. We find that the latent representation provides sufficient robustness to detect anomalies to high levels of performance (AUROC=0.9999). The method is generally applicable to high dimensional time-series modelling and distils the temporal data-stream to a single metric.