Long and Short Term Maneuver Trajectory Prediction of UCAV Based on Deep Learning

Autonomous air combat technology of unmanned combat air vehicles (UCAVs) is a hot issue that is currently being studied by various countries, and maneuvering trajectory prediction is an important part of autonomous air combat research. To address the difficulty of maintaining high prediction accuracy and short prediction time simultaneously in maneuvering trajectory prediction, this paper proposes a maneuvering trajectory prediction method that is based on a layered strategy, which combines long-term maneuvering unit prediction and short-term maneuvering trajectory prediction. In long-term maneuvering unit prediction, the complex trajectory is divided into 21 types of maneuvering units using the four characteristics of maneuvering trajectories, and a maneuvering unit library is established. On the basis of the deep echo state network(DeepESN), to capture multiscale prediction input parameters, autoencoder (AE) technology is incorporated. In addition, to increase the prediction accuracy, adaptive boosting (Ada) learning technology is utilized to build a strong predictor, and seven prediction networks are compared. The results demonstrate that the proposed method realizes the highest prediction accuracy. The single-step prediction time is about 0.002 s, which meets the time requirement. In short-term maneuvering trajectory prediction, the long and short-term memory (LSTM) network is analyzed, and the gaussian random walk strategy particle swarm optimization (GWSPSO) algorithm is used to update the internal weights and biases of the network to overcome the problems of "gradient disappearance" and "gradient explosion", and a data sharing method is proposed for overcoming the no directionality of optimization algorithms. Compared with four traditional networks, the results demonstrate the method that is proposed in this paper performs better. Compared with the sampling time of 0.3 s, the short-term prediction time of 0.05 s can also meet the requirements. Finally, a long- and short-term layered prediction method is used on a group of complex maneuvering trajectories. The results demonstrate that the prediction accuracy is significantly increased and the real-time requirements are satisfied.