次 级 通 道 模 型 先 验 驱 动 的 直 升 机 结 构 强 化 学 习振 动 主 动 控 制 方 法 研 究

High vibration has always been one of the major challenges hindering the development of helicopters. Active vibration control technology has become a preferred method for reducing helicopter vibration. With the development of artificial intelligence, reinforcement learning has provided new ideas and solutions for active vibration control. To solve the active vibration control problem under sinusoidal disturbances, this paper proposes a reinforcement learning method driven by a secondary model. The fundamental concept is to design the state and reward functions through prior knowledge to effectively improve training stability and convergence speed, while the controller is trained without online interaction with the real system. The method uses the identified secondary model to complete offline simulation training and obtain a controller that can be deployed directly. The secondary model is identified by the least squares method. The identified secondary model is used to reconstruct the disturbance signal as the environment state, thereby enhancing the ability of reinforcement learning to handle partial observability problems. In addition, the error signal is used to construct the reward function, and alignment with the secondary inverse model is carried out to achieve phase compensation for the delayed observation of rewards. The deep deterministic policy gradient algorithm is used to optimize the network controller to achieve robust control in the presence of noisy observations. Through the above strategies, the proposed method demonstrates excellent training convergence and control performance across multiple typical sinusoidal disturbance frequencies, and the effectiveness, robustness, and applicability of the controller have been verified through numerical simulations and experimental studies.