Neural-Enhanced Micro-Kalman Filtering for Satellite Tracking: A Comparative Study
For satellite tracking applications, this work offers an incremental improvement by integrating neural adaptation into an information-form Kalman filter, though performance gains over baselines are marginal.
The paper proposes a Neural-enhanced micro-Kalman filter (μKF) that adapts noise covariances online via a lightweight neural mechanism for satellite tracking. In simulations, it achieves comparable or slightly better mean square estimation errors than the classical Kalman filter while retaining computational advantages.
Satellite state estimation plays a fundamental role in orbital navigation, tracking, and autonomous space operations. Accurate estimation remains challenging due to uncertainties in process and measurement noise, which may degrade the performance of conventional Kalman filtering techniques. This paper presents a Neural-enhanced micro-Kalman filter ($μ$KF) for satellite tracking based on an information-form state estimation framework. Starting from a linearized state-space model of orbital dynamics, a lightweight neural scaling mechanism is introduced to adapt the process and measurement noise covariances online while preserving the underlying Bayesian filtering structure. The proposed estimator is formulated within the information-form $μ$KF framework and evaluated through numerical simulations using a linear Gaussian satellite tracking model. Its performance is compared with the classical Kalman filter (KF), the extended Kalman filter (EKF), the unscented Kalman filter (UKF), and an adaptive Kalman filter under identical operating conditions. Simulation results demonstrate that the proposed Neural-$μ$KF accurately tracks the satellite states with consistently low mean square estimation errors (MSEE). Furthermore, the proposed method achieves estimation performance comparable to, and for selected states slightly better than, the baseline Kalman filter while retaining the computational advantages of the information-form formulation. These results demonstrate that integrating lightweight neural covariance adaptation into the $μ$KF provides an effective and flexible framework for satellite state estimation.