Neural-Enhanced Micro-Kalman Filter Outperforms Classical Methods in Satellite Tracking

A new paper by Moh Kamalul Wafi presents a Neural-enhanced micro-Kalman filter (μKF) designed to improve satellite state estimation. The method builds on a linearized state-space model of orbital dynamics and introduces a lightweight neural scaling mechanism that adapts process and measurement noise covariances in real time. This adaptation is integrated within an information-form μKF framework, preserving the core Bayesian filtering structure while adding flexibility. The approach targets a fundamental challenge in satellite tracking: uncertainty in noise parameters that can degrade conventional Kalman filter performance.

Simulation results using a linear Gaussian satellite tracking model show that the Neural-μKF achieves consistently low mean square estimation errors (MSEE). It performs comparably to classical Kalman filter (KF), extended Kalman filter (EKF), unscented Kalman filter (UKF), and an adaptive Kalman filter—and slightly better for selected states. Critically, it retains the computational efficiency of the information-form formulation. These results demonstrate that integrating lightweight neural covariance adaptation into micro-Kalman filtering provides an effective, flexible framework for autonomous space operations and real-time satellite navigation.