A Data-Centric Understanding of Satellite Maneuver Anomaly Classification within Space Domain Awareness

In 2023, a record-breaking 2664 objects were launched into space. As space traffic increases, situational awareness in the space domain becomes more important. Space systems must identify abnormal, or anomalous, satellite behavior to determine whether the behavior is of concern. This capability requires vast amounts of data from different sources and modalities. These data types can include Electro-Optical and Infrared imagery, astrometric and photometric features, and other radio-frequency features. A Space Domain Awareness (SDA) system should take advantage of any data it receives, but also be able to operate reliably with sparse amounts of data. Artificial Intelligence and Machine Learning (AI/ML) is a suitable approach to understand multi-source and multi-modal data but must be implemented in a careful manner. Important works like Google’s Data Cascades and the Berryville Institute of Machine Learning’s Architecture Risk Analysis of Machine Learning Systems highlight the most common flaws in AI/ML systems such as training a model on data that does not properly reflect the model’s real-world environment or overfitting a model so that it memorizes that data rather than learn from it. AI/ML systems for SDA must avoid mistakes such as these due to the highly critical nature of SDA. Therefore, machine learning models should be carefully evaluated and provide users with sufficient information.  A data-centric satellite maneuver anomaly classification system will determine how uncertain the machine learning model is in its prediction based on the input data. Our proposed project will establish the feasibility of such a system. The system processes input data from multiple sources and modalities to provide a user with an abnormal satellite maneuver alert. This alert will provide an explanation of how the input data and trained machine learning model relates to the detected anomaly. This proof-of-concept will result in an SDA Anomaly Classification Framework.