Greg Martin, Pacific Defense Solutions – a Centauri Company; Charles J. “Jack” Wetterer, Pacific Defense Solutions, A Centauri Company; Jenna Lau, Pacific Defense Solutions, A Centauri Company; Jeremy Case, Pacific Defense Solutions; Nathan Toner, Pacific Defense Solutions, a Centauri Company; C. Channing Chow, Pacific Defense Solutions, a Centauri Company; Phan Dao, AFRL
Keywords: machine-learning, cislunar, orbit determination
Abstract:
As a means of aiding initial orbit determination for objects in cislunar space, we train various machine-learning algorithms to predict membership in one of 31 elementary periodic orbit families (e.g. HN2 for northern halo orbit about L2 libration point ) using observational features, including uncertainties, generated from the synodic positions and velocities. Additionally, various methods are investigated to allow for “null” membership where the observation is assessed to not belong to any of the periodic orbit families. The machine-learning algorithms investigated include k-nearest neighbor, support vector machines, decision tree ensembles, random forests, and deep neural networks. The features used include combinations of right ascension, declination, range, their associated rates, and visual magnitude. Individual and integrated performance results are reported and, as expected, vary depending on the features used and uncertainty in the features. Performance accuracy of up to 98% is possible.
Note: Pending public release approval by AFRL
Date of Conference: September 15-18, 2020
Track: Cislunar SSA