Autonomous orbital maintenance using a supervised-learning-based target point approach

Autonomous orbital maintenance is a fundamental component of spacecraft autonomy and has become an active area of research. This study investigates the feasibility of implementing an autonomous onboard Target Point Approach (TPA) for the stationkeeping of periodic orbits, enabled by supervised learning. A stochastic optimization framework based on the TPA is first employed to generate optimal stationkeeping parameters from a range of initial state deviations. Based on these solutions, a large balanced dataset is constructed and used to train supervised learning models, including a multilayer perceptron (MLP) classifier to distinguish feasible from infeasible initial deviations, and MLP regressors to predict optimal stationkeeping parameters directly from initial deviations. The trained models are then integrated into an onboard TPA-based stationkeeping framework and evaluated through large-scale simulations involving 100,000 initial state deviations for a candidate Near Rectilinear Halo Orbit (NRHO). The simulation results demonstrate the effectiveness and robustness of the proposed approach. Furthermore, regularities observed from the large-scale stationkeeping analysis are identified and analysed, providing insight into the structure of the stationkeeping solution space and the learning-enabled decision process. • Supervised Learning enables autonomous stationkeeping using Target Point Approach. • Lightweight neural networks assess feasibility and predict stationkeeping parameters. • Data processing based on distributions improves stationkeeping parameter prediction. • Large-scale simulations show robust long-term autonomous stationkeeping performance.