【作者】 刘浩；

【导师】 涂海燕；

【作者基本信息】 四川大学 ， 控制科学与工程， 2023， 硕士

【摘要】 四旋翼编队作为一种多智能体系统,具有结构简单、成本低、可操作性强等明显优势。此外,四旋翼编队可通过协同控制和分布式控制实现群体行为,可在一定范围内完成空中勘测,灾后搜救以及军事作战等任务,具有广阔的应用前景。因此,研究以四旋翼编队为主的多智体控制算法和轨迹规划方法,对于推动无人机技术的发展以及提高多智能体系统的智能化水平具有重要的理论和实践意义。针对四旋翼编队控制问题,本文以编队中的四旋翼为研究对象,建立了四旋翼平移动力学模型和旋转运动学模型,设计了一种基于RBF(Radial Basis Function)神经网络的自适应预定义时间滑模控制算法;针对四旋翼编队轨迹规划问题,为了实现四旋翼编队自主避障的功能,本文设计了基于虚拟力的改进人工势场法和基于Sum_Tree的改进DDPG(Deep Deterministic Policy Gradient)强化学习算法两种轨迹规划算法。最后经过仿真实验验证了算法的有效性和优越性。本文的主要工作如下所述:1)基于RBF神经网络的自适应预定义时间滑模控制算法:首先,本文根据牛顿第二定律以及Newton-Euler方程对编队中的四旋翼建立了平移动力学模型和旋转运动学模型;然后,本文设计了基于RBF神经网络的预定义时间滑模控制算法,该算法不仅可以在预先设定好的时间范围内利用RBF神经网络估计出四旋翼上所搭载未知重物的质量,从而消除该重物对四旋翼控制效果的影响,而且可以在估计出未知重物的基础之上使得四旋翼的位置环和姿态环的误差在预先设定好的时间范围内稳定;最后,本文不仅在理论上证明了该算法的稳定性,而且设计了仿真实验验证了该算法的快速性和稳定性。2)基于虚拟力的改进人工势场法:首先,为了能够为四旋翼编队规划出期望形状的轨迹,本文设计了一种使得规划轨迹能够渐近收敛于期望轨迹的虚拟力;然后,为了避免在障碍物附近所使用的人工势场法陷入局部最优,本文设计了一种干扰力,从而避免产生局部最优;最后,本文设计仿真实验来验证算法的有效性和先进性。3)基于Sum_Tree的改进DDPG强化学习算法:首先,为了克服传统DDPG强化学习算法训练速度慢以及训练结果不稳定的缺陷,本文为传统DDPG强化学习算法引入了基于人类思想的经验收集策略,基于Sum_Tree结构的经验池以及基于OU(Ornstein-Uhlenbeck)随机噪声的动作选择策略;然后,为了使得四旋翼编队更有效的避障,本文设计了相应的状态空间,动作空间以及奖励函数;最后,本文设计了仿真实验来训练改进DDPG强化学习算法,从而验证该算法训练过程的快速性和训练结果的稳定性,然后又设计了多种复杂环境来验证其训练结果的通用性和有效性。更多还原

【Abstract】 Quadrotor formations are a multi-intelligence system with simple structure,low cost and high operability.They can be used for aerial survey,disaster rescue,military operations,and other tasks.Therefore,it is of great theoretical and practical significance to study the control algorithm and trajectory planning algorithm of quadrotor formation to promote the intelligence level of multi-intelligent systems.In the aspect of quadrotor formation control,this paper establishes a translational and rotational kinematic model for each quadrotor in the formation,and designs an adaptive predefined time sliding control algorithm based on RBF neural network to control the model,which is used to verify the control effect of the algorithm for the quadrotor formation.In the aspect of quadrotor formation trajectory planning,this paper designs the improved artificial potential field method based on virtual force and the improved DDPG reinforcement learning algorithm based on Sum＿Tree.This paper also designs the simulation experiments to test the obstacle avoidance effect of these two trajectory planning algorithms.The main work of this paper has three points as follows:1)The adaptive predefined time sliding mode control algorithm based on RBF neural network: Firstly,this paper establishes translational kinematic model and rotational kinematic model for the quadrotor within the formation using Newton’s second law and Newton-Euler equation respectively.Then,this paper designs a simulation to verify the stability of the adaptive predefined time sliding mode control algorithm based on RBF neural network.This algorithm can not only reduce the impact on the quadrotor control by adaptively estimating the mass of the weight carried by the quadrotor,but also make the quadrotor attitude stable within the predefined time range.Finally,this paper designs the simulation experiments to verify the effectiveness of the algorithm.(2)The improved artificial potential field method based on virtual force: First,this paper designs an improved artificial potential field method based on virtual force for trajectory planning of quadrotor formations.The algorithm not only can use the characteristics of the artificial potential field method to plan a path to avoid obstacles,but also can plan a path that converges asymptotically to the desired trajectory in the absence of obstacles.Then,this paper introduces a disturbance force on the basis of the above algorithm to solve the defect that the artificial potential field method falls into local optimum in the process of obstacle avoidance.Finally,this paper designs simulation experiments to verify the effectiveness of the algorithm.(3)The improved DDPG reinforcement learning algorithm based on Sum＿Tree:First,this paper adds an experience collection strategy based on human thoughts,an experience pool based on Sum＿Tree structure and an action selection based on OU random noise to the traditional DDPG reinforcement learning algorithm in order to overcome the shortcomings about slow training speed and unstable training results of traditional DDPG reinforcement learning algorithm.Finally,this paper designs simulation experiments to train the improved DDPG reinforcement learning algorithm,so as to compare with the traditional DDPG reinforcement learning algorithm to verify the speed and the stability of the training results,and also designs various other complex environments to test the generality of the training results.更多还原