Gregory Badura, Georgia Tech Research Institute; Christopher R. Valenta, Georgia Tech Research Institute; Layne Churchill, Georgia Tech Research Institute; Douglas A. Hope, Georgia Tech Research Institute
Keywords: Machine Learning, Light Curves, Unified Data Library, Spin Stability
Abstract:
Providing real-time updates on satellite spin-stability is a critical task for continuously assessing collision risk. Space Domain Awareness (SDA) community analysts often manually inspect light curves of satellites in order to assess whether a satellite is stabilized or tumbling. While human analysts can perform stability classification, the ever-growing volume of data collected and published to repositories such as the Unified Data Library (UDL) necessitates the development of explainable automated Machine Learning (ML) stability assessment tools to assist them. Previous efforts to assess satellite stability using ML solutions assume idealized assumptions about the light curve data quality, such as having evenly time-sampled data, and full-night observations. Because of factors such as limited telescope tasking time and weather fluctuations, data uploaded to the UDL rarely meets these assumptions meaning that a significant amount of useful UDL data is under-utilized.
In order to address this gap, we developed an autoencoder-based anomaly detection algorithm that provides spin-stability assessments on light curves as short as 5 minutes in duration with up to a 91% score in balanced accuracy. The autoencoder was trained to reconstruct noisy light curves with explicit knowledge of the phase angle and visual magnitude uncertainty of each data point. We quantitatively and qualitatively show that the inclusion of this ancillary light curve information provides the potential for the network to overcome environmental noise, weather corruption, and uneven sensor sampling cadences in the process of learning to differentiate typical behaviors of stable light curves (i.e. glints) from those of tumbling light curves (i.e. oscillations , aliasing). The autoencoder’s light curve reconstruction prediction accuracy was found to serve as an “anomaly score” for discriminating unstable light curves from stable light curves.
The autoencoder’s uncertainty-weighted mean reconstruction error was combined with several expert-derived features as a feature embedding for a Balanced Random Forest (RF) classifier. The Balanced RFs that were trained using the autoencoder’s reconstruction error metric outperformed those trained solely using expert level features by 9±3.5% in terms of Matthews Correlation Coefficient (MCC) and 8.5±3.7% in terms of F1-score, with a significance level of p
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications