Phillip Schmedeman, US Space Command; David Beskow, USMA; Daniel Herber, Colorado State University
https://doi.org/10.64861/WYTC9204
Keywords: satellite maneuver detection, propulsion classification, multimodal sensor fusion, passive RF (TDOA/FDOA), geosynchronous orbit, collection requirements, satellite characterization, machine learning
Abstract:
Space Domain Awareness depends on detecting orbital maneuvers promptly and identifying the propulsion capabilities of non-cooperative satellites. Physics-based detection is reactive, identifying a maneuver only after the orbit drifts far enough from prediction to be distinguishable. That can take hours to days, compressing the time available for conjunction assessment. Benchmarking of the commercially provided maneuver-detection services covering this study’s geosynchronous fleet shows that none delivers a high detection rate, a low false-alarm rate, and broad coverage at once. The service with competitive recall reports only after the maneuver, and the prompt services cover few maneuvers. The second gap is propulsion system characterization, because hardware specifications for non-cooperative satellites are often unknown or deliberately obscured. The goal of this research is to close both gaps with supervised machine learning over fused sensor observations, and to convert the results into collection-requirements guidance for Space Situational Awareness (SSA) sensor networks._x000D_
_x000D_
This research develops supervised machine-learning models that integrate electro-optical (EO) and passive radio-frequency (RF) observations of geosynchronous satellites, trained and evaluated against 13,312 operator-confirmed maneuver events spanning five propulsion technologies over nine months. A long short-term memory (LSTM) ensemble detector operates on residual features referenced to precision orbit determination (i.e., state vectors). Passive-RF observations take the form of time-difference-of-arrival (TDOA) and frequency-difference-of-arrival (FDOA) measurements. These are evaluated at signature level under a pre-registered, hash-pinned protocol that compares two-line element (TLE) and precision orbit-determination residual references on identical observations. Per-satellite propulsion classifiers are evaluated under a frozen one-shot protocol across a three-rung ladder of label quality: operator truth, commercial maneuver reports, and the deployed maneuver detector’s own alerts, then checked against an independent commercial reference. Protocols were registered and hashed before any answer existed, models and thresholds were frozen, sealed sets were scored exactly once, and every claim is bounded by the population it was measured on._x000D_
_x000D_
Both methods achieve operational quality and are deployed in support of US Space Command. The detector has run as a live service since July 2026, publishing machine-confirmed maneuver records to the Unified Data Library. Evaluated once on a sealed, satellite-disjoint test set, it recovers 0.669 of priority maneuvers at a 5 percent false-alarm budget, at a median time-to-detection of 6.6 hours. At a 10 percent budget, it recovers 0.869, at a median of 4.5 hours. Priority maneuvers are operator-confirmed events with a commanded velocity change of at least 0.5 meters per second. That recall is comparable to the best commercial baseline (0.676), a recovery service that reports a median of 9.4 hours after maneuver start. Detection quality depends strongly on the residual reference. On identical observations, residuals referenced to precision orbit determination reach 0.963 PR-AUC in development, against 0.598 for residuals referenced to TLE sets. The same dependence governs the passive-RF modality. Maneuver-window TDOA separability reaches an AUC of 0.925 (95 percent confidence interval 0.839 to 0.978), a paired improvement of 0.144 over the TLE reference. This establishes the signature-level basis for combining EO with passive RF. A replication under the same frozen protocol extends the signature to two independent commercial passive-RF networks, at a union AUC of 0.815. The propulsion classifier reaches 0.933 PR-AUC for chemical propulsion, 0.874 for electric, and 0.739 accuracy for propulsion scheme, when using features derived from the deployed maneuver detector. Observation-span analysis shows 81 to 85 percent of satellites reach a stable classification within three months of routine observation, where stable means settled rather than verified correct. These results translate into collection-requirements trades across observation cadence, modality mix, and observation span for Space Domain Awareness sensor-network design and procurement._x000D_
Date of Conference: September 15-18, 2026
Track: Space Domain Awareness