Deep Learning for Spacecraft Anomaly Detection

The research problem is to develop an on-board technological approach that will enable spacecraft to autonomously detect anomalies, identify the anomalies, self-diagnose the root causes of the anomaly, and self-prescribe remedies to fix the identified faults that are causing the anomaly. We propose a purely data-driven deep learning approach that can autonomously detect, identify, diagnose and remedy anomalies. This approach allows for an initial "operator-in-the-loop" supervised approach to guiding every stage of this process. Part of the problem is to identify open-source or commercially available datasets that contain historical data with anomalous patterns that we can use to train this deep learning model. Given the unavoidable paucity of labelled anomalous data, we propose to synthesize training datasets using available historical data.