Research On Autonomous Flight Control Of Rotary-Wing Unmanned Aerial Vehicle Based On Deep Spatiotemporal Features

The rotary-wing UAV can realize autonomous hovering,vertical take-off and landing,coordinated steering,formation operations and other flight tasks in a small space due to its simple structure,strong adaptability,and high flexibility.So it has been widely used in remote sensing,aerial photography,environmental monitoring,mapping,and other fields,as well as surveillance and reconnaissance,ground strike and electronic countermeasures,and other tasks.With the increasing demands for aerial operations,the requirements of the rotary-wing UAV are gradually developing towards the direction of complex maneuvering and flocking coordination,which pose great challenges to the existing autonomous flight control methods.Such challenges mainly lie in the following aspects:(1)When the rotary-wing UAV is flying in complex environments such as dense buildings and vast jungles,there may be unmodeled errors that cannot be observed and described due to the dynamic and changeable aerodynamic characteristics,resulting in the inability to achieve fast and smooth attitude and position changes;(2)When the rotary-wing UAV performs a fast maneuvering flight mission,it may reach the physical limitations due to the drastic change of its own state,causing the UAV to deviate from the equilibrium position and lose control;(3)When the rotary-wing UAV performs a remote networked flight mission,the flight data may be lost or delayed due to the instability of the network,which affects the performance of the UAV flight control;(4)When the rotary-wing UAVs perform a flocking flight mission,the information exchange between UAVs may be blocked due to the instability of data transmission,and it is difficult to achieve uniform stable UAV flocking control.In response to the above challenges,this thesis conducts research on autonomous flight control of rotary-wing UAVs based on deep spatiotemporal features.The main contents include:(1)Considering the analysis of aerodynamic characteristics of the rotary-wing UAVs,a dynamic modeling based on deep spatiotemporal feature representation is proposed.Firstly,in order to achieve the characterization of unmodeled dynamics,a spectral normalization constrained adaptive neural network is designed to extract deep spatiotemporal features,which represent the hidden states and intrinsic laws of the system.Secondly,in order to meet the requirements of a fast and dynamic online update of the model,a multi-model adaptive structure is adopted.And the spectral normalization constraints are introduced into the training process to ensure the Lipschitz stability of the network.Finally,the unmodeled dynamics based on deep spatiotemporal feature representation are combined with the traditional UAV dynamic model to achieve a complete characterization of the time-varying aerodynamic characteristics of the rotarywing UAV.This thesis verifies the effectiveness of the proposed modeling method on a real flight dataset.The experimental results show that the proposed method has the characteristics of high modeling accuracy,short training time,and fast model response speed.(2)Considering the complex maneuvering flight problem of the rotary-wing UAV,an aerobatic control method based on deep imitation learning is proposed.Firstly,the motion patterns with the same statistical characteristics are mined from the demonstration action sequences by convolutional neural network with the local perception and global integration capabilities.And the spatiotemporal feature vectors that characterize the UAV flight action primitives are generated.Secondly,a deep spatiotemporal feature representation network is designed to map the UAV flight sequence into the state-action space spanned by spatiotemporal feature vectors,so as to accurately describe any maneuvering flight action.Finally,in the spatiotemporal state-action space,the complex maneuvering flight control based on imitation learning is realized by aligning the target state features with their own state features.This thesis verifies the effectiveness of the proposed method on a real aerobatic dataset provided by Stanford University.Experimental results show that the proposed method can perform arbitrary aerobatic maneuvers by observing a limited set of expert demonstrations.(3)Considering the instability of communication in remote flight missions of the rotary-wing UAV,a predictive control method is proposed based on time sequence reconstruction.Firstly,a deep convolution unit for multivariate time sequence prediction is constructed,which uses the dynamic changes between adjacent action states to simulate the non-stationary and approximate stationary characteristics of the UAV during flight through a cascaded self-updating storage mechanism.Secondly,the relevant features with dynamic consistency are mined from the UAV state-action sequence samples by stacking multiple convolutional units,to realize the potential representation of the high-order non-stationary flight process.Finally,by using Lyapunov’s theorem,the stability of the system under the proposed method is proved.In this thesis,the trajectory tracking control performance of the UAV is analyzed under the network control condition with random packet loss.The experimental results show that the proposed method can achieve high precision prediction compensation for lost data packets,thereby ensuring better trajectory tracking performance.(4)Considering the problem of flocking flight control of the rotary-wing UAVs,a flocking control method based on deep spatiotemporal graph features is proposed.Firstly,the UAVs are represented as the nodes of the spatiotemporal graph,and the information interaction between UAVs is used as the edge weight of the graph to construct a spatiotemporal graph network with dynamic flocking representation capability.Secondly,in the time dimension,a deep time sequence learning module is used to aggregate the temporal trends of the historical observation sequences of each node.While in the spatial dimension,a deep graph convolution module is used to aggregate the spatial dependencies between adjacent nodes.Finally,the temporal and spatial features are fused,and the flocking data packets are predicted and reconstructed by the historical states and the spatio-temporal dependencies between adjacent nodes,to realize the flocking consistency control in the case of data packet loss.In this thesis,the proposed flocking control method is experimentally verified on a flocking formation composed of five Crazyflie UAVs.The results show that the proposed method can still ensure good flocking control in a communication environment with a random packet loss rate of 50%.Based on the above research,this thesis systematically studies the autonomous flight control problem of the rotary-wing UAV based on deep spatiotemporal features,which promotes the development of the UAV aerial operation capabilities and provides reference for the research of other types of robots.