1 December 2021 in Aerospace Science and Technology
1 March 2021 in Acta Astronautica
25 August 2026 in SSRN Electronic Journal
Addressing the challenges of insufficient trajectory reconstruction accuracy for non-cooperative space targets under fragmented observation conditions, as well as the lack of feedback from evolving observations to intercept planning in conventional open-loop frameworks, this paper proposes a two-stage "far-range navigation + close-range guidance" active defense framework. At the trajectory reconstruction level, an improved Gramian Angular Field (GAF) algorithm is employed to transform time-series trajectories into nine-dimensional asymmetric multi-channel images, with a MinMaxPredictor module inferring global orbit extremum parameters from local statistical features. Under the partial observable trajectory reconstruction strategy, a Temporal Feature-Guided Masked Autoencoder (TFG-MAE) achieves semantic-level reconstruction from sparse visible image patches to complete GAF images through multi-scale feature-guided decoding. At the control level, a Progressive Perception Receding-Horizon (PPRH) strategy is proposed throughout the entire pursuit. At each control cycle, TFG-MAE ingests accumulated observational data to generate rolling target trajectory predictions, with KRR correction applied. Sequential Quadratic Programming (SQP) then solves the fuel--time optimal transfer from the current spacecraft state to the predicted intercept point. A genetic algorithm (GA) pre-searches the optimal intercept timing before each SQP solution. Receding-horizon replanning is triggered when the L2 deviation between successive TFG-MAE predictions exceeds a predefined threshold. The MAE-based trajectory predictor exhibits inherent robustness to observation noise: reconstruction MAE remains nearly constant (≈0.05) across noise levels from σ = 0 to 0.010. This inherent denoising capability eliminates the need for a dedicated noise-sensitive terminal guidance law. Combined with KRR correction using time-decay weights, prediction error remains bounded and unbiased. Experimental results show that PPRH achieves superior comprehensive performance over six classical closed-loop guidance laws---PD, ZEM/ZEV, OSG, APN, Bang-Bang, and Pursuit---across all noise levels: 100% mission success rate with the shortest intercept time and near-optimal fuel consumption. These findings support the effectiveness of the "online evolution of prediction model + receding-horizon replanning" technical approach.