Transfer Learning for Hypersonic Vehicle Trajectory Prediction

Hypersonic glide vehicles (HGVs) introduce challenges in terms of predicting their future flight behavior because they fly at high speeds and maneuver during flight. Furthermore, there are limited examples of actual HGV flights, which impedes the development of prediction methods that model HGV trajectories. In this paper, we investigate transfer learning for HGV trajectory prediction to evaluate how stochastic grammars trained on a limited number of HGV trajectories perform on new, unseen HGV trajectories. Our analysis includes two datasets containing HGV trajectories that exhibit different maneuvers. One dataset contains trajectories that exhibit vertical maneuvers, which are behaviors related to changes in altitude. The second dataset exhibits both horizontal and vertical maneuvers, where horizontal maneuvers refer to changes in crossrange and downrange. The vertical dataset is also significantly smaller than the second dataset (i.e., the dataset that exhibits horizontal and vertical maneuvers). We develop a prediction method using the smaller, less complicated dataset to model HGV trajectories. Specifically, we use an unsupervised machine learning method based on stochastic grammars. Then, we demonstrate that the learned model can be used to predict HGV behavior and engageability for trajectories from the larger, more complicated dataset. Our results show that transfer learning improves prediction performance (even on unseen trajectory maneuvers) and can be used in limited data scenarios.