David Witman, Slingshot Aerospace; Timothy Olson, Slingshot Aerospace; Clarice Reid, Slingshot Aerospace; Belinda Marchand, Slingshot Aerospace; Adrian Thompson, Slingshot Aerospace
https://doi.org/10.64861/OTZY4616
Keywords: Anomaly detection, Time Series, LEO, GEO, Optical Systems, Modeling, Data Analysis, Machine Learning
Abstract:
Real-time monitoring, characterization, and anomaly detection of objects in Low-Earth Orbit depend on diverse datastreams with irregular update rates. Considering all available information is critical for operators building an understanding of space operations to better inform downstream decisions.
We propose a Self-Similarity Anomaly Detection framework, Marvin, that partitions an object’s observation history into representations of its behavior over defined time intervals, or temporal-actors, and compares those intervals to identify departures from the object’s own historical patterns. This formulation enables behavioral changes to be localized relative to the object’s own historical pattern of life, rather than requiring comparison with a peer population.
In this paper, we present an overview of two methodologies that use different constructions of temporal-actors: Masked Subspace Scorer (MSS) and Action-Free Inverse Reinforcement Learning (AFIRL) self-similarity. We evaluate both algorithms using two synthetic scenarios representing astrometric and photometric deviations from nominal behavior.
Finally, we evaluate the operational applicability using observations associated with the June 2025 MethaneSAT failure.
Date of Conference: September 15-18, 2026
Track: Space Domain Awareness