Dynamic Bayesian Optimization Network-Based Method for Geostationary Satellite Orbital Behavior Recognition

To address the challenges of temporal dependencies, environmental uncertainties, and engineering constraints in identifying the orbital behavior of noncooperative objects in high Earth orbit, we propose a dynamic Bayesian optimization network (DBON) framework. This framework transforms orbital behavior identification into a partially observable stochastic game, using state transition probabilities to model behavioral sequences and integrating a proximal policy optimization algorithm for training. By defining seven behavioral primitives'circumnavigation, flyby, following, hovering, no behavior, rendezvous, and collision' we establish a reverse inference mechanism from observation data to orbital behavior. Experimental results demonstrate that DBON achieves a behavior recognition accuracy of 70.5% under both single and complex behavior space conditions, outperforming traditional algorithms in key metrics, such as behavior recognition Brier score and early warning stability. This research offers an effective method for addressing behavior identification in high Earth orbit games and provides a theoretical basis and engineering support for space security autonomous decision-making systems.