Neural Network–Assisted Initial Orbit Determination Method for Libration Point Orbits

Abstract

Classic initial orbit determination (IOD) methods are mainly based on the hypothesis of unperturbed Keplerian motion and are inapplicable for Libration point orbits (LPOs). To this end, this paper proposes a neural network–assisted method for accurate and efficient IOD of LPOs. The proposed method consists of a long short-term memory (LSTM) neural network–based initial guess generator (IGG) combined with a second-order optimal corrector. First, the LSTM neural network is developed to compensate for the residuals between the two-body IOD solution and the true solution to generate a more accurate initial guess. Two sample forms, one in the rotating frame and the other in the inertial frame, are investigated and compared to select the better form to train the LSTM neural network. Then, a second-order optimal corrector is proposed based on the second-order state transition tensor and a two-step procedure, which takes the initial guess from the LSTM neural network and iteratively obtains a more accurate IOD solution. Finally, the proposed method is applied to solve the IOD problem in two Earth–Moon LPO scenarios: a 4:1 synodic resonant Near-rectilinear Halo orbit (NRHO) and a transfer trajectory from a lunar orbit to the 4:1 NRHO. Numerical simulations show that the estimated errors are reduced by 99% using the proposed LSTM-based IGG. Moreover, for a comparable level of accuracy, the second-order optimal corrector converges one time faster than the first-order corrector.