Abstract
This study addresses online threat intention recognition for space non-cooperative targets approaching a formation comprising a chief spacecraft and multiple deputy spacecraft. With only line-of-sight (LOS) angles available at long range, intention-dependent LOS differences are jointly constrained by relative geometry and observation errors, yielding weak evidence and hindering early recognition. We propose a method based on robust evidence recursion. Local LOS-motion features are used to construct window-level composite log-evidence. Building on the sequential decision principle of the sequential probability ratio test (SPRT), exponential smoothing mitigates measurement fluctuations, an evidence-bounding mapping limits anomalous local evidence, and recursive updating with exponential forgetting reduces the persistent influence of unstable early evidence. Across 600 independent test scenarios, the method achieved 98.57% accuracy, with a 0.37% false alarm rate, 1.98% miss rate, and 14.09 s mean remaining time margin. Relative to baseline SPRT, accuracy increased by 2.93 percentage points and the mean remaining time margin by 3.88 s. In separate single-factor tests, accuracy remained above 97% for noise standard deviations ≤0.20 mrad, angular resolutions ≤0.50 mrad, and effective lateral separations ≥1.0 km. The results demonstrate that reliable and timely online recognition can be achieved using LOS angles alone.