Real-time flight trajectory optimization for TF/TA using an enhanced RBF-LSTM network with attention mechanisms

In this paper, we present a real-time three-dimensional flight trajectory optimization method for fixed-wing unmanned aerial vehicles (UAVs) to achieve terrain-following-terrain-avoidance (TF-TA) capabilities in mountainous flight scenarios. This approach employs an innovative dual-layer structure that combines discrete trajectory optimization with an enhanced radial basis function-long short-term memory (RBF-LSTM) network for real-time trajectory planning. The designed network is obtained by introducing a multi-head attention mechanism into the classical LSTM network and utilizing the pre-planned trajectories from the RBF network as the initial input sequence for the LSTM network. At the upper layer, the method generates an optimal trajectory dataset for fixed-wing UAVs during specific tasks, encompassing the state and control of the trajectory. In the lower online planning layer, the pre-generated trajectory dataset is utilized to train the enhanced RBF-LSTM network, ensuring that the resulting network can accurately represent the mapping relationship between the state and control within the optimal trajectory. This enables its application in the optimal real-time feedback control of the vehicle system. The reliability of the proposed real-time flight trajectory planning approach is validated through Monte Carlo (MC) experiments. Furthermore, the optimality and real-time performance of the designed dual-layer framework are verified through comprehensive simulation studies. Finally, an explanation regarding the generalization ability of the proposed network is provided.