基于改进 DDPG 的组合动力飞行器上升段轨迹优化

A deep reinforcement learning-based trajectory optimization method is investigated to address the challenges posed by complex propulsion-mode transitions and strongly coupled process constraints in the ascent trajectory optimization of RBCC-powered aerospace vehicles. A longitudinal dynamical and aerodynamic model covering the ejector, ramjet, scramjet, and rocket modes is constructed, and a continuous action space centered on angle-of-attack rate is formulated within a Markov decision process framework. A reward structure is designed to balance feasibility constraints, flight-load constraints, and terminal mission requirements. A guidance-channel compensation mechanism based on a time-sequenced action library is further introduced to steer value estimation during policy updates, thereby enhancing the convergence efficiency and training stability of the deep deterministic policy gradient algorithm under multimodal propulsion conditions. Simulation analyses indicate that the guidance mechanism can improve convergence behavior, reduce peak process loads, and yield smoother control inputs to a certain extent. The results demonstrate the feasibility and potential of incorporating a guidance channel into reinforcement learning-based trajectory optimization, providing an extensible pathway toward intelligent trajectory planning for complex combined-cycle aerospace vehicles.