Dynamic inertia identification of combined spacecraft using a sequence-aware AESP-LSTM with excitation-embedded features

Accurate identification of the inertia tensor of combined spacecraft after capturing a non-cooperative target is crucial for stable attitude takeover in on-orbit servicing. Conventional estimators often degrade under abrupt inertia variations and measurement noise. This paper presents a compact inertia identification approach, termed Adaptive Entropy-Structured Pruned LSTM (AESP-LSTM), designed for real-time onboard execution under limited computational resources. In the offline stage, an over-parameterized model is trained on diverse post-capture scenarios and pruned to suppress redundant neurons that amplify noise, while a Dynamic Inertia Transfer Ratio (DITR) provides excitation-aware prior features. The strategy reduces model size by over 29%, lowering storage to 719 KB. In the online stage, the pruned network runs on a processor with millisecond-level inference 200–300 ms and a rolling adaptation scheme to preserve accuracy under time-varying inertia. Simulations in disturbance-rich orbital environments demonstrate superior accuracy, robustness, and efficiency compared with classical RLS, RPBSID, and FNN estimators. Notably, more than 90% of the predictions generated by this method fall within a 5% relative error margin, the proposed AESP-LSTM framework thus ensures high-accuracy, lightweight, and real-time inertia identification, offering a practical solution for reliable attitude takeover in future on-orbit servicing missions.