Learning-assisted adaptive filter for tracking hypersonic glide vehicles

Accurate tracking of hypersonic glide vehicles (HGVs) remains challenging because their aerodynamic maneuver accelerations typically exhibit pronounced temporal variations, while the parameters of conventional maneuver models are generally unavailable in advance. This paper develops a learning-assisted adaptive filtering framework in which the drag, longitudinal-lift, and lateral-lift accelerations are explicitly represented by three-axis Singer stochastic processes with parameters adapted online. In the offline stage, a sliding-window maximum-likelihood (ML) estimator is first applied to representative HGV trajectory samples to identify locally optimal maneuver-model parameters. These estimates are subsequently used as supervised labels to train a radial basis function neural network (RBFNN) using measurement-innovation features. During online tracking, the trained RBFNN efficiently maps recent innovation-driven statistics to the Singer time constants and process-noise intensities, which are then incorporated into an extended Kalman filter for recursive state estimation. The proposed scheme therefore approximates likelihood-based parameter adaptation through a feedforward learning model, thereby avoiding repeated online optimization while preserving the recursive structure and interpretability of stochastic filtering. Numerical simulations involving quasi-equilibrium glide and skip-glide trajectories show that the proposed method achieves tracking accuracy close to that of the sliding-window ML adaptive filter, while substantially reducing the computational burden. The results further indicate that the proposed method consistently improves estimation accuracy and robustness compared with fixed-parameter and interacting-multiple-model filters, particularly under substantially time-varying maneuver conditions.