基于可学习EKF的高超声速飞行器航迹估计

临近空间高超声速目标因具有高速、大机动、全球到达的特点,已成为国防安全的一类新型威胁.此类目标具有非惯性的航迹形式、并可进行复杂的策略性机动,给其航迹估计带来了新的挑战.为应对目标的机动特性,提升航迹估计性能,将循环神经网络与扩展卡尔曼滤波深度嵌合,提出了基于可学习扩展卡尔曼滤波的航迹估计方法.首先,通过分析目标机动特性,建立了参数化的目标机动模型.然后,考虑目标复杂机动特性对航迹估计造成的影响,将循环神经网络与扩展卡尔曼滤波深度嵌合,提出了可学习扩展卡尔曼滤波方法.通过使用已有航迹数据进行训练,所嵌入的两个循环神经网络,可发现目标机动的隐含规律,并对目标复杂机动所引起参数与模型不确定性的进行在线识别与动态补偿.最后,以某临近空间高超声速目标的航迹估计为例,选取典型机动场景,对所提出方法与EKF、AEKF等传统方法进行了对比分析. 分析结果表明,所提出的可学习扩展卡尔曼滤波方法可有效应对目标复杂的机动,具有比EKF、AEKF方法更高的估计精度和更优的估计动态性能.

Trajectory estimation of a hypersonic flight vehicle via L-EKF

Nearspace hypersonic flight vehicles have become a new threat to national defense due to their characteristics of high speed, large maneuverability, and global arrival. The vehicle has non-inertial trajectory and complex strategic maneuvers, which brings new challenges to its trajectory estimation. In order to deal with vehicle maneuvers and improve the trajectory estimation performance, a trajectory estimating method was proposed based on the learnable extended Kalman filter (L-EKF) by combining the recurrent neural networks (RNNs) and the extended Kalman filter (EKF). First, a parametric characteristic model was established to describe the vehicle maneuvers. Then, the L-EKF was proposed, in which two RNNs were designed and embedded into EKF. By training with available trajectory data of the vehicle, the two embedded RNNs could find the hidden laws of the vehicle maneuvers and dynamically compensate for the parametric and model uncertainties online. Finally, the proposed L-EKF method was compared with EKF and adaptive EKF methods in several typical estimation scenarios of a hypersonic vehicle. Simulation results show that the proposed L-EKF had a higher estimation accuracy and better dynamic performance than EKF and adaptive EKF, especially when the flight vehicle performs unknown complex maneuvers.