IMM-AIS-UKF: A GAN-Enhanced Dynamics-Constrained Trajectory Predictor With DDPG Method for UAV-USV Interaction

In order to solve the problem that the real-time position of USVs in heterogeneous collaborative systems cannot be used as an effective position for tasks such as tracking and positioning, this paper proposes an Interactive Multi-Model Adaptive Incremental-Step Unscented Kalman Filter (IMM-AIS-UKF) method based on deep deterministic policy gradient (DDPG) to predict the position of USVs after a certain period of time. First, in order to solve the problem that traditional methods are prone to deviate from the true trajectory during the prediction of large-scale circular arc trajectories, this paper proposes an AIS-UKF method that considers angular acceleration and linear acceleration, so that the predicted trajectory is more biased towards the inside of the arc and is suitable for the kinematic laws of most heterogeneous subsystems. At the same time, DDPG training is used to obtain the most appropriate number of incremental steps required for different states. Then, in order to avoid DDPG from repeatedly selecting the same wrong action during training and testing, an Ornstein-Uhlenbeck (OU) noise based on a real physical model is proposed as the strategy selection noise of the DDPG model to increase the richness of action selection. In addition, a trusted interactive multi-model method based on generative adversarial network (GAN) enhancement is proposed, which enables the method to adapt to more complex surface environment and rely on generative artificial intelligence instead of traditional expert experience to generate process noise covariance matrix of each model. The feasibility and innovation of the proposed algorithm are verified by practical training data acquisition and real-time prediction experiments.