1. Dual-Channel and Bidirectional Neural Network for Hypersonic Glide Vehicle Trajectory Prediction
- Author
-
Yangfan Xie, Hongbo Chen, Xuebin Zhuang, and Zepu Xi
- Subjects
Hypersonic speed ,trajectory prediction ,General Computer Science ,Artificial neural network ,Computer science ,General Engineering ,deep learning ,Kalman filter ,Hypersonic glide vehicle ,TK1-9971 ,bidirectional gated recurrent unit ,Nonlinear system ,Recurrent neural network ,Robustness (computer science) ,Control theory ,Trajectory ,General Materials Science ,time series forecasting ,Electrical engineering. Electronics. Nuclear engineering ,Time series - Abstract
Different from traditional aircraft, hypersonic glide vehicles (HGVs) possess stronger maneuverability and a higher flight speed (generally higher than 5 Mach), making trajectory prediction very complicated. Several works have been conducted in this field, which usually analyze the motion characteristics of the HGV first and then use a Kalman filter to track and predict the trajectory. In this way, the accuracy of prediction depends on how to model the control parameters of the target vehicle. The core idea of this paper mainly concerns treating the HGV trajectory prediction as a multivariate time series forecasting problem since HGV trajectories are special multivariate time series with nonperiodic temporal patterns. Moreover, capturing the hidden dependencies between time steps and different time series ensures the accuracy and robustness of predictions. Recently, recurrent neural networks have been widely used in predicting data with temporal patterns between several time steps; however, they fail in capturing nonperiodic temporal patterns. Therefore, we propose a brand-new model named the dual-channel and bidirectional neural network (DCBNN) to intelligently predict the trajectory of hypersonic vehicles in undetectable areas, especially for those with complex maneuver models. DCBNN is constructed with a nonlinear component and a linear component to collect both nonlinear and linear features from input data to improve robustness. Moreover, a dual-branch architecture is utilized in the nonlinear component to capture the complex and mixed dependencies between long-term and short-term patterns. Experiments reveal that the proposed method is effective and intelligent.
- Published
- 2021