Trajectory Estimation And Prediction For Nearspace Hypersonic Targets Via Recurrent Neural Networks

Nearspace hypersonic flight vehicles have become a new threat to national defense due to its high speed,large maneuverability and global arrival characteristics.Since several types of nearspace hypersonic weapons have been released and equipped in recent years,the defense of hypersonic targets is becoming more and more crutial.Since the hypersonic target has a non-inertial trajectory form and large-scale maneuvers,it is necessary to estimate and predict its trajectories with high accuracy for its interception.Most of the existing trajectory estimation methods are based on adaptive filtering or multi-model methods with classic target maneuver models,which are hard to cope with the complex maneuvers of the hypersonic target.Meanwhile,the trajectory prediction methods based on parameter identification and model extrapolation cannot deal with the strategic changes of target maneuvers.Therefore,considering the complex strategic maneuvers of the nearspace hypersonic target,trajectory estimation and prediction problems are investigated in this thesis.Recurrent neural networks are deeply combined with the nonlinear filtering methods to recognize and predict the target's motion behaviors,by which the estimation and long-term prediction of the target's trajectories can be realized.The main work are summarized as follows.First,the dynamic model of the nearspace hypersonic target is given.Two typical motion behavior models are established to describe the quasi-equilibrium and skip gliding trajectories of the target respectively by analyzing the target's motion characteristics,in which several recognizable motion behavior parameters are constructed.And then,a motion behavior model set is designed by analyzing the target's maneuverability,which can capture all motion behaviors of the target.Second,considering the complex strategic maneuvers of the nearspace hypersonic target,a structure design method of motion behavior recognition networks is proposed based on attention LSTM.The networks are trained and optimized by the available trajectory data of the target.By fully mining prior information from the existing trajectories through data-driven,the hidden laws of the target's motion is found.In this way,the target motion behaviors are recognized,which can deal with the unknown maneuvers.Then,an online trajectory estimation algorithm is proposed based the designed motion behav-ior recognition networks,several model selection strategies are designed to select models to be used for trajectories estimation according to motion behavior recognition results,which the trajectory estimation is achieved.Third,considering the strategic changes of the target' s maneuvers,a structure design method of motion behavior prediction networks is proposed based on encodingdecoding framework,by which the long-term motion behavior sequence of the target can be generated.Then,a trajectory prediction algorithm is proposed based on the motion behavior prediction networks,which can autonomously switch to the prediction model sequence provided by the motion behavior prediction results.In this way,the long-term trajectory prediction of the nearspace hypersonic target with complex strategic maneuvers is achieved.Forth,considering nonlinear filtering problems with large-scale model uncertainty,a learnable extended Kalman filtering(L-EKF)method is proposed by embedding recurrent neural networks into the extended Kalman filter.Then,aiming to the real-time requirement in practical trajectory estimation application,a trajectory estimation algorithm is designed based on L-EKF,in which the complex strategic maneuvers is treated as the large-scale uncertainty of the maneuvering model.Finally,the proposed trajectory estimating and predicting methods are applied to the nearspace hypersonic target with complex maneuvers in several typical scenarios.The estimation and prediction performance are compared and analyzed.Simulation experiments show that the proposed methods have higher accuracy and better dynamic performance than the existing methods,especially when the target performs complex strategic maneuvers.