Research On Satellite Attitude Control Algorithm Based On Drl Algorithm

Satellite attitude control is one of the most important aspects in aerospace engineering.It refers to moving the satellite to the target attitude according to the predetermined target attitude,and the control should also meet certain robustness and anti-interference ability.Some control tasks still exist.Various const raints such as status.The satellite itself has nonlinear,parameter coupling and other factors that make modeling difficult.Therefore,in this paper,the deep reinforcement learning algorithm is used as a controller that does not depend on the satellite accurate model.Aiming at this problem,this paper is based on the deep reinforcement learning control idea,taking satellite attitude control as the blueprint,and studying the application of deep reinforcement learning algorithm in the control field.Firstly,in the second chapter,the satellite attitude dynamics and kinematics model based on quaternion are given.Based on this,the kinematics equation based on the error quaternion is derived.It is pointed out that the satellite attitude maneuver control can be abstracted into input.The mapping to the fixed output provides a theoretical basis for solving the problem with a reinforcement learning algorithm.And the established model provides the basic th eory for the simulation experiment and control of Chapter 4 of Chapter 3.Then the paper introduces the model identification network.In order to better apply the deep reinforcement learning to an unknown model and improve the utilization of data in the deep reinforcement learning algorithm,this paper firs tly carries out the nonparametric identification of the model and selects the neural network.In this paper,we compare the effects of BP neural network and RBF neural network fitting,and finally choose the RBF neural network with better generalization ability.This paper introduces the context of reinforcement learning development,and focuses on the algorithm of depth deterministic strategy gradient class.This paper intends to implement attitude control for satellites using the Deep Deterministic Policy Gradient(DDPG)algorithm,and proposes related improvement schemes,combined with the well-known RBF neural network to find the optimal control strategy.In this chapter,the overall framework of the mo delbased DDPG algorithm is given,and the end of training is given,that is,the relevant test set is combined to ensure the convergence of the model.Finally,the control results of DDPG algorithm and traditional PID algorithm are compared.The state constraint of DDPG algorithm is also verified,and the stability analysis and convergence analysis of DDPG algorithm are carried out.In order to make the controller training better,this paper combines the prior knowledge of the existing controlled objects and uses DDPG to optimize the parameters of the PID controller.At the same time,this paper gives the direction that the deep reinforcement learning algorithm should be improved.