Machine Learning for RSO Maneuver Classification and Orbital Pattern Prediction
Abstract
Various Space Situational Awareness (SSA) platforms and capabilities have demonstrated the ability to observe and detect space object / satellite maneuvers as well as predict the resultant orbits after those maneuvers. Upon detection of a maneuver, analysts must quickly assess the potential threat a satellite poses to other orbital assets even though maneuvers may take from hours to weeks to complete. The first step is determining the type of maneuver (e.g., station keeping, Hohmann transfer) and whether that maneuver is benign or a potential threat. These error-prone, nuanced analyses are often accomplished manually with uncertain data, resulting in lower confidence assessments. In this paper we present Probabilistic Assessment of Space Threats (PAST) which employs machine learning tech- niques to assist space domain awareness analysts quickly and accurately. PAST quantitatively categorizes the type of orbital maneuver that a satellite executed and then uses that historical orbit data to determine whether the maneuver is nominal or anomalous. These capabilities provide space domain awareness analysts with quantitative assessments of potentially threatening behaviors in space. PAST addresses three operational challenges that confound rapid and accurate assessment of observed resident space object maneuvers for space situational awareness: (a) producing a quantitative confidence metric for maneuver classifications, (b) performing maneuver classifications with uncertain and/or incomplete data, and (c) assessing whether maneuvers are nominal or anomalous. We tested PAST on the data from the four-week space domain awareness exercise Global Sentinel 2017 (GS17) and proved that PAST pro- vides unique and novel capabilities using Machine Learning and aids in much quicker decision-making after an orbital maneuver is detected.
Operational Challenge 1: Assessing Maneuver Types with Quantitative Confidence Space domain analysts must ultimately assess the output of an analytic tool, such as PAST, and initiate a set of actions in response to its predictions. Machine learning classifiers, however, are black boxes that usually produce one or more "categories" that fit the data, leaving the human decision makers without an understanding of the likelihood or confidences associated with the classifications. Our solution is to use Neural Network Machine Learning techniques that can produce Bolztmann Distribution posteriors (probability distributions across the various maneuver types). The Boltzmann distribution shows the analyst the machine learner's confidence levels for each maneuver type. Separate trained models for each mission-class of satellite e.g., (Iridium, GPS) provide more accurate classifications. Our classification tool exploits an ML-based algorithm to "learn" how to compute a posterior distribution over the class types (i.e., a finite categorical space) given measurements of changes in orbital characteristics. Hence, the output of classification is not a declaration of the form "Maneuver is of Class A," but rather is a set of probabilities over all classes. Operational Challenge 2: Uncertain and/or incomplete maneuver data Real-world orbit data may be based on low-fidelity observations (e.g., some optical telescopes) resulting in missing or low-certainty maneuver data as well as false maneuver alerts. Most state of the practice machine learners will fail with missing or highly inaccurate feature data. PAST's neural network can marginalize the missing or uncertain features to produce posterior distributions with the available feature data, producing lower-confidence (but viable) classifications. Lower-confidence distributions can be generated with early Two Line Element (TLE) data post-maneuver and refined later with more accurate TLEs or state vectors. The machine learner can detect when the input features do not adequately match any trained data, resulting in a "bad solution" classification. This, in addition to very low confidence classifications can help analysts quickly identify potential "false alarm" maneuver alerts. Operational Challenge 3: Assessing Nominal vs Anomalous Maneuvers Real-world satellites maneuver often due to natural effects (e.g., atmospheric drag, gravitational perturbations) as well as mission-specific behaviors (e.g., intentional new orbit). Analysts need to quickly assess a maneuver in light of historical behavior patterns. PAST models each satellite class as its own periodic stochastic process to learn normalcy models for its orbits. PAST can determine if a maneuver is typical/benign, where atypical maneuvers (based on historical data) can indicate threatening behavior by a satellite. The pattern of life model can also assist analysts by predicting the next expected maneuver for each satellite. Experiment Our machine learner was trained using 2017 TLE data from space-track.org with maneuver labels provided by a subject matter expert for Geostationary (GEO) satellites. Note that creation of training data sets for other orbit regimes is in work. We then conducted a 4-week experiment of the PAST prototype using 2016 data replayed from the Global Sentinel (GS) exercise scenario. The unclassified GS Space Surveillance Network (SSN) observation data was ingested by LM iSpace, which processed the observations and detected events such as conjunctions, decays, and maneuvers. iSpace issued a Maneuver Alert to PAST for each maneuver event, which in turn predicted the probability distribution across each maneuver type and presented this distribution to the human analyst. Results During the GS scenario experiment, the PAST IRAD received over 300 maneuvers of GEO satellites to classify from the LM iSpace system in real-time. iSpace SMEs analyzed the results of PAST's classifier and assessed at least90% accuracy. Of the remaining 10%, the SMEs could not score the classifier results due to the nature of the maneuver; for example, the maneuver may have been falsely reported due to poor observations quality or other anomalies. In a real world situation, the maneuvers that could not be classified would be further analyzed and additional sensors would likely be tasked to collect more data to resolve the uncertainty.- Publication:
-
Advanced Maui Optical and Space Surveillance Technologies Conference
- Pub Date:
- September 2019
- Bibcode:
- 2019amos.confE..54D
- Keywords:
-
- maneuver classification;
- threat assessment;
- machine learning;
- patterns of life