Contextual Predictive Model for Early Identification of High-Covariance Conjunctions

Abstract Satellite operators regularly assess conjunction risks and weigh trade-offs between mission operations and collision avoidance actions in the increasingly congested orbital environment. As a conjunction event approaches, operators typically receive a sequence of conjunction data messages (CDMs) from the US Space Force’s 18th and 19th Space Defense Squadrons. The CDMs identify objects involved in the conjunction, report a measure of risk and margins for the event, and estimate uncertainty in the form of state covariance matrices. Operators then use the information in the CDMs, along with supplemental data, to determine whether to take action. The methods presented in this paper aim to facilitate operators’ decision-making by identifying which conjunction events are likely to benefit from additional data acquisition, and which are likely to self-resolve from routine tracking, well in advance of when they must commit to maneuvers. We introduce a machine learning based model developed by Slingshot Aerospace that takes into account the temporal evolution of CDM sequences, and it predicts, in advance, the conjunctions for which the covariance will not decline enough via typical tracking. The model incorporates contextual information about the objects involved in the conjunction, including physical characteristics and pattern of life details, that enable it to outperform baseline covariance-only approaches. Our model has demonstrated reliable performance, identifying high-covariance (i.e. high-uncertainty) conjunctions up to five days prior to the time of closest approach, which equips operators with sufficient advanced notice to request observations and obtain refined state estimates with enough time to incorporate the enhanced information into their collision avoidance maneuver decision.