Research On Cooperative Control Of Multiple UAVs Based On Reinforcement Learnin

The application of UAVs has gone deep into all aspects of life at present.Therefore,new challenges have been presented to the performance and control strategy of UAVs,especially in the cruising of power lines.However,a single UAV cannot meet the complex task requirements of changing geographical environment,so the multi-UAV collaborative task planning emerges.And reinforcement learning,with its strong ability of environment perception and decision making,can fit the application scenarios of multi-machine systems.Cooperative inspection of power lines by multi-UAV involves target recognition,autonomous obstacle avoidance,path planning,following cruise and other fields.According to kinematic constraints and preset formation requirements,planning a smooth and feasible route for each UAV to reduce the difficulty of UAV trajectory tracking is an important content of the current research on quadrotor UAV formation.The main content of my thesis can be shown as follows:(1)The UAV dynamics model and mission constraint model were studied.The selfconstruction and coordinate system modeling of UAV are studied,and the path planning model of UAV is established.The transformation relationships between coordinate systems and different coordinates and the dynamics models between UAV formations were established,and some of the models were simplified.The multi-machine system task constraint model was established by combining the multi-travel salesman problem(MTSP)with the artificial potential field method(APF).(2)An improved MADDPG obstacle avoidance algorithm was proposed.The trajectory planning and obstacle avoidance algorithm of UAV based on Multi-Agent Deep Deterministic Policy Gradient(MADDPG)technology were studied.Firstly,the model was introduced by training convolutional neural network.Then,aiming at the practical requirements of cooperative inspection of power lines,a UAV obstacle avoidance algorithm based on MADDPG algorithm was proposed,which enabled the multi-aircraft system to avoid multiple or single obstacles and planed a shortest path to the target point.In order to enhance the UAV path-planning ability,the idea of MTSP and improves the exploration process to improve the UAV path planning ability was introduced.Compared with the traditional algorithm,the accuracy,convergence and antidisturbance ability of environmental factors of the model have been greatly improved.The obstacle avoidance ability of UAV was promoted by combining the improved reward function with APF algorithm.Several scenarios were simulated and designed,including the obstacle avoidance ability and flight path planning ability of UAV formation under single or multiple obstacles.The accuracy of algorithm was verified.(3)The real machine testing based the promoted algorithm was studied.The indoor flight control platform was built based on binocular vision positioning principle,and the simulation was verified on Matlab/Simulink platform.In this thesis,Feisi X150 UAV platform was used,and its structure and binocular vision positioning system were also analyzed.Meanwhile,multiple quadrotor UAVs were used to conduct formation experiments,and codes were deployed on the simulation platform to verify its capability of flight path planning in real scenes.From the experimental results,it can be seen that this method can efficiently and accurately achieve the predicted effect.