Autonomous Navigation And Flight Control Of Quadrotor Drone Based On Neural Networks

The potential applications of autonomous Unmanned Aerial vehicles(UAVs)can be extended to many fields,such as industry,agriculture,intelligent transportation and national defense.As these applications involve the interaction of UAVs with the environment,a number of fundamental theoretical and technical issues need to be addressed,among which,autonomous navigation and flight control are the most important and challenging ones.Recently,imitation learning,deep reinforcement learning and neuroadaptive control have gained increasing attention and are deemed useful and valuable methods for autonomous navigation and flight control.Imitate learning algorithms to learn autonomous navigation strategies from automatic driving and image classification datasets,but these strategies are not imitated from drone operators.Deep reinforcement algorithms manipulate drones to interact with environment,learn flight strategies for autonomous navigation,in the absence of accurate states of the environment,regular deep reinforcement learning methods are not so efficient in general.Adaptive control methods have been used for the quadrotor dynamics with parametric uncertainties,which however cannot deal with nonparametric uncertainties and external disturbance as well as actuator failures.NN-based tracking control methods can only ensure semiglobal stability.In this thesis,the focus is on autonomous navigation and flight control of UAVs under modeling uncertainties.The features and contributions of this work can be summarized as follows:(1)The problem of autonomous navigation based on imitation learning is investigated.Firstly,an autonomous navigation algorithm based on imitation learning is proposed,an observation-action dataset is collected when drone operators fly a drone,where the dataset is used to train a convolutional neural network,the trained network can fly a drone like human experts.Secondly,an autonomous navigation algorithm based on position and attitude control strategies is proposed,a yaw angle classification network is designed to classify the yaw angle of the drone,and a lateral offset classification network is constructed to classify the lateral offset of the drone,combining classification results to calculate yaw angle for autonomous navigation.Finally,the effectiveness of the autonomous navigation algorithm based on imitation learning is verified on the road,the effectiveness of the autonomous navigation algorithm based on position and attitude control strategies is verified in the corridor of a building.(2)The problem of autonomous navigation based on deep reinforcement learning is investigated.Firstly,as the state of environment in the real world is difficult to measure accurately,environment observations instead of environment states is used to establish a Markov decision process for autonomous navigation.Secondly,an autonomous navigation algorithm based on deterministic policy gradient and convolutional neural network is proposed.An actor network is designed for observe the environment and generates flight action;a critic network is designed for score the actions of the actor network.The ability of actor network to fly a drone has been continuously improved,the ability of critic network to judge the actor network has also been improved,eventually,the actor network can fly a drone safely.Finally,the effectiveness of autonomous navigation algorithms based on deep reinforcement learning is tested and verified in the corridor of a building.(3)The problem of position and attitude tracking control of quadrotor UAVs under uncertain dynamic and actuation faults is studied.Firstly,a comprehensive dynamic model is established,which is more effective in reflecting the behaviors of the vehicle.Secondly,indirect NN-based adaptive FTC algorithms are developed to ensure stable position and attitude tracking of UAVs,which are featured with simplicity and cost effectiveness.Finally,numerical simulation is conduced to verify and demonstrate the effectiveness of the proposed method.(4)The problem of globally exponentially stable tracking control of self-restructuring nonlinear UVAs systems is studied.Firstly,a class of self-restructuring nonlinear dynamics are established.Secondly,a NN-based adaptive control algorithm is developed to guarantees exponentially globally stable tracking control of self-restructuring nonlinear UVAs systems,and address the issue of reliable in-loop operation of NN approximation-based control unit.Finally,numerical simulation is conduced to verify and demonstrate the effectiveness of the proposed method.