Self-adaptive agile control method for high-speed maneuvering flight-oriented intelligent unmanned aerial vehicle with body

Abstract

The invention relates to a self-adaptive agile control method for a high-speed maneuvering flight-oriented intelligent unmanned aerial vehicle with a body, and the method comprises the steps: building a four-rotor unmanned aerial vehicle kinetic model, and outputting the coupling characteristics of aerodynamic force and torque and the prediction information of a future motion state; constructing a reinforcement learning strategy of the Lyapunov stability constraint, and obtaining a compensation control strategy according to output information and tracking errors of the kinetic model; constructing a hierarchical neural network architecture, and generating a motion control instruction of the quad-rotor unmanned aerial vehicle based on a compensation control strategy and latest state information fed back by a sensor; non-linear interference in flight is estimated through a non-linear interference observer, the non-linear interference is fed back to an adaptive control law in the hierarchical neural network, and control input of the hierarchical neural network is optimized in a prediction window through a rolling horizon optimization algorithm; and generating a motion control instruction for intercepting the target unmanned aerial vehicle by using the compensated and optimized adaptive control law and the control input, and sending the motion control instruction to an execution mechanism of the four-rotor unmanned aerial vehicle.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) control technology, and relates to an adaptive agile control method for an embodied intelligent UAV oriented towards high-speed maneuvering flight. Background Technology

[0002] With the rapid development of drone technology and its expanding application scope, security issues such as illegal drones intruding into sensitive areas are becoming increasingly prominent. In counter-drone missions, quadcopter drones, with their agile maneuverability, have become important equipment for intercepting illegal drones. However, under high-speed and high-maneuver conditions, quadcopter drones face many severe challenges.

[0003] In high-speed maneuvering scenarios, the dynamic characteristics of quadrotor UAVs exhibit significant strong-coupled nonlinear features. Simultaneously, control delays and complex environmental disturbances, such as airflow variations and electromagnetic interference, severely restrict their control performance. Traditional control methods struggle to address these complexities, resulting in the inability to adjust the quadrotor UAV's attitude and trajectory in a timely and precise manner when facing high-speed evasive maneuvers by target UAVs, ultimately leading to interception mission failure.

[0004] Furthermore, existing quadcopter drones have significant shortcomings in structural design and flight control systems. Conventional structural designs cannot meet the high maneuverability requirements of high-speed, high-maneuverability operations, and limitations in computing power, power consumption, size, and weight of flight control systems severely restrict their application in complex scenarios. For example, traditional quadcopter drones are prone to deformation due to insufficient structural strength during high-speed flight, affecting flight stability; the slow computing speed of flight control systems makes it impossible to process large amounts of sensor data and complex control algorithms in a timely manner, resulting in control delays. Therefore, an innovative solution is urgently needed to overcome the challenge of adaptive and agile control of interceptor drones under high-speed, high-maneuverability conditions. Summary of the Invention

[0005] The technical problem solved by this invention is to overcome the shortcomings of the prior art and propose an adaptive and agile control method for embodied intelligent unmanned aerial vehicles (UAVs) for high-speed maneuvering flight. This method addresses multiple problems faced by quadrotor UAVs when performing high-speed, high-maneuvering flight to intercept target UAVs, such as strongly coupled nonlinear dynamics, control delay, and environmental interference, thereby achieving stable and agile control of quadrotor UAVs.

[0006] The solution to the technical problem of this invention is: to provide an adaptive agile control method for an embodied intelligent unmanned aerial vehicle (UAV) oriented towards high-speed maneuvering flight, comprising the following steps:

[0007] A strongly coupled nonlinear dynamic model of a quadrotor UAV is established. The real-time state of the quadrotor UAV and environmental disturbances obtained through sensor fusion are used as the inputs of the dynamic model. The outputs are the coupling characteristics of aerodynamic forces and torques, as well as the prediction information of future motion states.

[0008] A reinforcement learning strategy with Lyapunov stability constraints is constructed. Based on the output information of the dynamic model and the tracking error, a compensation control strategy is obtained for fine-tuning the control quantity in complex coupling and interference environments. The tracking error is the relative position deviation between the quadcopter UAV and the target UAV.

[0009] A hierarchical neural network architecture is constructed to generate motion control commands for a quadcopter UAV based on the compensation control strategy and the latest state information from sensor feedback. The hierarchical neural network architecture includes a dynamic feature decoupling layer and a hybrid control generation layer. The dynamic feature decoupling layer is used to separate the coupling features of aerodynamic forces and torques. The hybrid control generation layer integrates robust baseline control and the compensation control strategy to provide an adaptive control law and outputs motion control commands based on real-time state information.

[0010] The nonlinear disturbances encountered during flight are estimated by a nonlinear disturbance observer and fed back to the adaptive control law for compensation. The control input of the hierarchical neural network is optimized within the prediction window by a rolling time-domain optimization algorithm to minimize prediction error and control energy consumption.

[0011] The quadcopter drone generates motion control commands to intercept the target drone using the compensated and optimized adaptive control law and control input, and sends them to the quadcopter drone's actuators.

[0012] Furthermore, the strongly coupled nonlinear dynamic model of the quadcopter UAV is established as follows:

[0013]

[0014] Where p is the position vector, v is the velocity vector, R is the rotation matrix, ω is the angular velocity vector, m is the mass, F is the total thrust, g is the gravitational acceleration, I is the moment of inertia matrix, and T is the torque vector. The antisymmetric matrix representing the angular velocity vector ω.

[0015] Furthermore, in the reinforcement learning policy constrained by Lyapunov stability, a hybrid reward function is designed as follows:

[0016]

[0017] Where e is the tracking error vector, u c The control input vector, obtained from the output information of the dynamic model, is used for energy assessment. Let H be the derivative of the Lyapunov function, H be the tracking error weight, D be the control energy consumption weight, and α be the stability penalty coefficient.

[0018] Furthermore, the dynamic feature decoupling layer has the following network structure:

[0019] Spatiotemporal attention mechanism: 4-head attention, time window size 10, spatial attention weights Where Q is the state feature vector at the current moment, K is the state feature vector within the historical time window, V is the control output feature corresponding to the state feature vector within the historical time window, and d k =16 is the vector dimension of K;

[0020] GRU unit: input dimension 20, hidden layer dimension 64, adopts a 2-layer structure, uses the ReLU activation function, and has a dropout rate of 0.1.

[0021] Furthermore, the hybrid control generation layer has the following network structure:

[0022] Robust baseline control employs a sliding mode control law: u base =-k s sgn(s)-k e e;

[0023] Among them, u base For robust baseline control output, x represents the state tracking error, which includes position error, velocity error, and angle error; λ is the sliding surface coefficient; sgn(s) is the sign function; e is the tracking error vector; k s and k e All are constant coefficients;

[0024] Linear observer for estimating linear disturbances Where y is the actual output state vector. The predicted state vector output by the linear observer. C is the identity matrix, A and B are the coefficient matrices of the linear disturbance observer, and L is the pole placement gain.

[0025] Control weight adjustment: w = σ(κ||d||);

[0026] Where σ is the Sigmoid function and κ is a constant coefficient; For total interference, This is the interference estimate output by the nonlinear interference observer;

[0027] Final control output: u = wu base +(1-w)u comp ;

[0028] Among them, u comp This is the compensation output of the compensation control strategy.

[0029] Furthermore, the estimation of nonlinear disturbances experienced during flight via a nonlinear disturbance observer includes:

[0030] The dynamic equation of the nonlinear disturbance observer is:

[0031] in, A is the interference estimate output by the nonlinear interference observer. d B d C is the coefficient matrix of the nonlinear disturbance observer. d A is the bias matrix; d =-5I,B d =10I,C d =0, where I is the identity matrix.

[0032] Furthermore, the optimization of the control input of the hierarchical neural network within the prediction window using a rolling temporal optimization algorithm includes:

[0033] Set the prediction window N=20 for the rolling time-domain optimization algorithm;

[0034] The optimization objective of the rolling time-domain optimization algorithm is:

[0035]

[0036] The constraint condition is: u c,k ∈[-10,10]N·m;

[0037] The interior point method is chosen as the solution method.

[0038] Among them, e k Let u be the error vector between the current state information fed back by the sensor and the predicted state at time k. c,k The optimized hierarchical neural network control input vector is defined by Q', where Q' is the prediction error weight and D is the control energy consumption weight.

[0039] Furthermore, the quadcopter drone's fuselage is made of carbon fiber composite material, and the arm structure adopts a streamlined variable cross-section arm, which is narrow at the front end and wide at the rear end to reduce air resistance; the battery is placed at the bottom center of the quadcopter drone's fuselage, and the motor is installed at the end of the arm, with the motor installation angle adjusted.

[0040] Furthermore, redundant design and fault-tolerant control of the quadcopter UAV's power system were implemented.

[0041] The Smith prediction compensation algorithm for brushless motors eliminates power lag. When a single rotor fails, a three-rotor fault-tolerant mode is adopted: the speed of the remaining motors is adjusted to the saturation zone, the yaw differential torque is used to compensate for the unbalanced torque, and the sliding mode control attitude reconstruction algorithm is combined to improve the success rate of recovery after loss of control.

[0042] The advantages of this invention compared to the prior art are:

[0043] (1) This invention combines Lyapunov stability theory with reinforcement learning to construct an end-to-end adaptive control method. Precise dynamic modeling and stability constraints ensure the reliable operation of the system in complex flight environments, avoiding runaway phenomena caused by complex dynamic characteristics and environmental interference, and enhancing the success rate of interception missions. It effectively solves the control problem of quadcopter UAVs intercepting target UAVs under high-speed and high-maneuver conditions, and improves the stability and control performance of the system.

[0044] (2) This invention achieves decoupling and adaptive control of aerodynamic / torque coupling characteristics through the design of a hierarchical neural network architecture, which improves the system’s adaptability to different operating conditions and can maintain good control performance under various flight conditions.

[0045] (3) This invention uses a nonlinear interference observer to optimize the control output of the neural network and uses a rolling time-domain optimization algorithm to optimize the future state prediction information within the prediction window, thereby achieving control input compensation for the neural network. This enables the system to complete real-time trajectory reconstruction and parameter self-tuning at a millisecond-level response speed, ensuring the real-time performance and accuracy of interception in high-speed maneuvering scenarios and effectively coping with environmental interference and uncertainty.

[0046] (4) The embodiments of the present invention improve the maneuverability of the UAV through hardware innovation, namely the improvement of the structure of the quadcopter UAV and the upgrade of the flight control, while taking into account the improvement of computing power and the reduction of power consumption, size and weight, providing strong support for the practical application of the system. Attached Figure Description

[0047] Figure 1 This is a flowchart of the adaptive agile control method for an embodied intelligent unmanned aerial vehicle (UAV) designed for high-speed maneuvering flight, according to the present invention. Detailed Implementation

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0049] like Figure 1 As shown, the adaptive agile control method for embodied intelligent unmanned aerial vehicles (UAVs) oriented towards high-speed maneuvering flight proposed in this invention includes the following steps:

[0050] S1. Establish a strongly coupled nonlinear dynamic model of the quadrotor UAV. The real-time state of the quadrotor UAV and environmental disturbances obtained through sensor fusion are used as the input of the dynamic model. The output is the coupling characteristics of aerodynamic force and torque, as well as the prediction information of future motion state.

[0051] S2. Construct a reinforcement learning strategy with Lyapunov stability constraints. Based on the output information of the dynamic model and the tracking error, obtain a compensation control strategy for fine-tuning the control quantity in complex coupling and interference environments. The tracking error is the relative position deviation between the quadcopter UAV and the target UAV.

[0052] S3. Construct a hierarchical neural network architecture to generate motion control commands for the quadcopter UAV based on the compensation control strategy and the latest state information from the sensor feedback. The hierarchical neural network architecture includes a dynamic feature decoupling layer and a hybrid control generation layer. The dynamic feature decoupling layer is used to separate the coupling features of aerodynamic forces and torques. The hybrid control generation layer integrates robust baseline control and the compensation control strategy to provide an adaptive control law and output motion control commands based on real-time state information.

[0053] S4. Estimate the nonlinear disturbances encountered during flight using a nonlinear disturbance observer, feed them back to the adaptive control law for compensation, and optimize the control input of the hierarchical neural network within the prediction window using a rolling time-domain optimization algorithm to minimize prediction error and control energy consumption.

[0054] S5. Using the compensated and optimized adaptive control law and control input, generate motion control commands for the quadcopter drone to intercept the target drone, and send them to the quadcopter drone's actuators.

[0055] Example 1

[0056] The adaptive agile control method for embodied intelligent unmanned aerial vehicles in this embodiment includes the following steps:

[0057] S1. Establish a strongly coupled nonlinear dynamic model of the quadrotor UAV. The real-time state of the quadrotor UAV and environmental disturbances obtained through sensor fusion are used as the inputs to the dynamic model. The outputs are the coupling characteristics of aerodynamic forces and torques, as well as the prediction information of future motion states.

[0058] The real-time status of the quadcopter UAV includes position, attitude, and velocity. Environmental disturbances include influencing factors such as wind and airflow. The strongly coupled nonlinear dynamic model of the quadcopter UAV is established as follows:

[0059]

[0060] Where p is the position vector, v is the velocity vector, R is the rotation matrix, ω is the angular velocity vector, m is the mass, F is the total thrust, g is the gravitational acceleration, I is the moment of inertia matrix, and T is the torque vector. The antisymmetric matrix representing the angular velocity vector ω.

[0061] S2. Construct a reinforcement learning strategy with Lyapunov stability constraints. Based on the output information of the dynamic model and the tracking error, obtain a compensation control strategy for fine-tuning the control quantity in complex coupling and interference environments. The tracking error is the relative position deviation between the quadcopter UAV and the target UAV.

[0062] The convergence condition of the Lyapunov function is used as a hard constraint for policy updates to ensure the stability of the quadcopter UAV during autonomous interception. By designing a suitable Lyapunov function and incorporating its derivative as a penalty term into the reward function, the reinforcement learning algorithm can optimize the control policy while maintaining system stability. A hybrid reward function is designed in the reinforcement learning policy, including tracking error, control energy consumption, and a Lyapunov derivative penalty term. Tracking error measures the relative positional deviation between the quadcopter UAV and the target UAV, control energy consumption evaluates the energy consumption of the control input, and the Lyapunov derivative penalty term ensures system stability. By appropriately setting the weights of each component, the reinforcement learning algorithm can achieve high-maneuverability control while ensuring system stability.

[0063] The hybrid reward function takes the following form:

[0064]

[0065] Where e is the tracking error vector, u c The control input vector, obtained from the output information of the dynamic model, is used for energy assessment. Let H be the derivative of the Lyapunov function, H be the tracking error weight, D be the control energy consumption weight, and α be the stability penalty coefficient. The hybrid reward function balances tracking accuracy, control energy consumption, and stability.

[0066] S3. Construct a hierarchical neural network architecture and generate motion control commands for the quadcopter UAV based on the compensation control strategy and the latest state information fed back by the sensors.

[0067] The input to the hierarchical neural network is the compensation control strategy obtained in step S2 and the latest state information directly fed back from the sensors. The hierarchical neural network includes a dynamic feature decoupling layer and a hybrid control generation layer. The dynamic feature decoupling layer is used to separate the coupling features of aerodynamic forces and torques. The network structure is as follows:

[0068] Spatiotemporal attention mechanism: 4-head attention, time window size 10, spatial attention weights Where Q is the state feature vector at the current moment, K is the state feature vector within the historical time window, V is the control output feature corresponding to the state feature vector within the historical time window, and d k =16 is the dimension of the K vector;

[0069] GRU unit: input dimension 20, hidden layer dimension 64, 2-layer structure, using ReLU activation function, dropout rate 0.1;

[0070] Function: Separates the aerodynamic coupling feature F from the torque coupling feature T, and outputs the decoupled feature vector.

[0071] The hybrid control generation layer integrates robust baseline control and the compensation control strategy to provide an adaptive control law, outputting motion control commands based on real-time state information; the network structure is as follows:

[0072] Robust baseline control (sliding mode control law): u base =-k s sgn(s)-k e e;

[0073] Among them, u base For robust baseline control output, x represents the state tracking error (including position, velocity, and angle errors), λ represents the sliding surface coefficient (a positive constant that determines the speed at which the system state converges to the sliding surface), sgn(s) is the sign function (used to ensure robustness), e is the tracking error vector, and k s and k e All are constant coefficients. In this embodiment, λ = 5, k s =10,k e =5.

[0074] Linear observer for estimating linear disturbances

[0075] Where y is the actual output state vector of the system (such as position, velocity, etc.). The predicted state vector output by the linear observer. C is the identity matrix, A and B are the coefficient matrices of the linear disturbance observer, and L is the pole placement gain.

[0076] Control weight adjustment: w = σ(κ||d||);

[0077] Where σ is the Sigmoid function and κ is a constant coefficient; in this embodiment, κ = 0.5.

[0078] For total interference, This is the interference estimate output by the nonlinear interference observer;

[0079] Final control output: u = wu bass +(1-w)u comp

[0080] Among them, u comp The compensation output of the compensation control strategy is used to adaptively compensate for coupling, disturbances, and model uncertainties.

[0081] S4. The nonlinear disturbances encountered during flight are estimated by the nonlinear disturbance observer and fed back to the adaptive control law for compensation. The control input of the hierarchical neural network is optimized within the prediction window by the rolling time-domain optimization algorithm to minimize the prediction error and control energy consumption.

[0082] The Nonlinear Disturbance Observer (NDOB) estimates the nonlinear disturbances experienced by the system in real time by observing the system state and output, and feeds them back to the adaptive control law for compensation.

[0083] The dynamic equation of the nonlinear disturbance observer is: in, A is the interference estimate output by the nonlinear interference observer. d B d C is the coefficient matrix of the nonlinear disturbance observer. d This is the bias matrix. In this embodiment, A... d =-5I,B d =10I,C d =0, where I is the identity matrix.

[0084] The performance specifications of the nonlinear interference observer are: interference estimation error ≤ 5%, response time ≤ 2ms.

[0085] The rolling time-domain optimization algorithm optimizes future control inputs based on the current state and the predicted future state, in order to minimize prediction error and control energy consumption.

[0086] Set the prediction window N=20 for the rolling time-domain optimization algorithm, corresponding to a time of 0.2s;

[0087] The optimization objective of the rolling time-domain optimization algorithm is:

[0088]

[0089] The constraint condition is: u c，k ∈[-10, 10]N·m;

[0090] The interior point method is selected as the solution method, and the calculation time is ≤1ms.

[0091] Among them, e k Let u be the error vector between the current state information fed back by the sensor and the predicted state at time k. c,k The optimized hierarchical neural network control input vector is defined by Q', where Q' is the prediction error weight and D is the control energy consumption weight.

[0092] S5. Using the compensated and optimized adaptive control law and control input, generate motion control commands for the quadcopter drone to intercept the target drone, and send them to the quadcopter drone's motors and other actuators.

[0093] This embodiment also includes hardware optimization design for the quadcopter drone:

[0094] (1) Quadcopter UAV Airframe Structure Optimization: The quadcopter UAV airframe is made of lightweight, high-strength carbon fiber composite material, which has the characteristics of low density and high strength. While ensuring the structural strength of the airframe, it effectively reduces weight and significantly improves the power-to-weight ratio. The arm structure is innovatively designed, adopting a streamlined variable cross-section arm with a narrower front end and a wider rear end, reducing air resistance and improving the stability and maneuverability of the quadcopter UAV during high-speed flight. The airframe layout is optimized by placing the battery at the bottom center of the quadcopter UAV airframe, lowering the center of gravity; the motors are installed at the ends of the arms, and the anti-interference capability and agility are enhanced by optimizing the motor installation angle (such as inward or outward tilt angle).

[0095] (2) Redundancy Design and Fault-Tolerant Control of Power System: The Smith prediction compensation algorithm of brushless motor (BLDC) is used to eliminate power lag and significantly shorten the motor step response time to the millisecond level. When a single rotor fails, a three-rotor fault-tolerant mode is adopted: the speed of the remaining motors is adjusted to the saturation range, and the yaw differential torque is used to compensate for the unbalanced torque. Combined with the sliding mode control attitude reconstruction algorithm, the success rate of recovery after loss of control is improved.

[0096] (3) Flight control system hardware optimization:

[0097] Main processor: ARM Corex-A72 (4 cores, 1.8GHz), responsible for complex algorithm calculations, with a data processing capability of 2 TOPS;

[0098] Low-power coprocessor: TI MSP430F5529 is used to handle sensor data preprocessing and low-power management, with a static power consumption of ≤1μA;

[0099] Functional module integrated circuit board: integrates sensor module, control module and communication module on a single 100mm×80mm circuit board; adopts 4-layer PCB layout to reduce line loss (≤3%) and signal interference;

[0100] Functional module connection relationship: The sensor module communicates with the control module through the I2C / SPI interface (transmission rate 10Mbps), and the communication module uses the CAN bus for data transmission (transmission delay ≤1ms), supporting data interaction with the ground station and the onboard computer;

[0101] Dimensions and weight specifications: Flight control system total weight ≤ 80g, volume ≤ 80cm² 3 It is suitable for installation in high-speed mobile scenarios.

[0102] In summary, this invention solves multiple problems faced by quadcopter UAVs when performing high-speed, high-maneuver flight to intercept target UAVs, such as strongly coupled nonlinear dynamics, control delay, and environmental interference, and achieves stable and agile control of quadcopter UAVs.

[0103] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

[0104] The contents not described in detail in this specification are common knowledge to those skilled in the art.

Claims

1. A body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight, characterized in that, The method comprises the following steps: A strong coupling nonlinear dynamics model of the quadrotor UAV is established, taking the real-time state of the quadrotor UAV obtained through sensor fusion and environmental disturbance as the input of the dynamics model, and outputting the coupling characteristics of aerodynamic force and moment and the prediction information of future motion state; A reinforcement learning strategy with Lyapunov stability constraint is constructed, and a compensation control strategy is obtained according to the output information of the dynamics model and the tracking error, which is the relative position deviation between the quadrotor UAV and the target UAV, to fine-tune the control quantity in a complex coupling and disturbance environment; A hierarchical neural network architecture is constructed to generate the motion control instruction of the quadrotor UAV based on the compensation control strategy and the latest state information fed back by the sensor; the hierarchical neural network architecture comprises a dynamic characteristic decoupling layer and a hybrid control generation layer, the dynamic characteristic decoupling layer is used to separate the coupling characteristics of aerodynamic force and moment, and the hybrid control generation layer fuses the robust baseline control and the compensation control strategy to provide an adaptive control law, which outputs the motion control instruction according to the real-time state information; Nonlinear disturbance observer is used to estimate the nonlinear disturbance in flight, which is fed back to the adaptive control law for compensation, and the control input of the hierarchical neural network is optimized in the prediction window through the receding horizon optimization algorithm to minimize the prediction error and control energy consumption; The adaptive control law and the control input after compensation are used to generate the motion control instruction of the quadrotor UAV to intercept the target UAV, which is sent to the actuator of the quadrotor UAV.

2. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 1, characterized in that, The strong coupling nonlinear dynamics model of the quadrotor UAV is established as follows: where p is the position vector, v is the velocity vector, R is the rotation matrix, ω is the angular velocity vector, m is the mass, F is the total force, g is the gravitational acceleration, I is the moment of inertia matrix, T is the torque vector, denotes the skew-symmetric matrix of the angular velocity vector ω.

3. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 2, characterized in that, In the reinforcement learning strategy with Lyapunov stability constraint, the mixed reward function is designed as follows: where e is the tracking error vector, u c is the control input vector, derived from the output information of the dynamics model, for energy evaluation; is the derivative of the Lyapunov function, H is the tracking error weight, D is the control energy weight, and a is the stability penalty coefficient.

4. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 3, characterized in that, The network structure of the dynamic characteristic decoupling layer is as follows: Space-time attention mechanism: 4 heads, time window size 10, space attention weight where Q is the state feature vector at the current time, K is the state feature vector within the historical time window, V is the control output feature corresponding to the state feature vector within the historical time window, d k = 16 is the vector dimension of K; The GRU unit: input dimension 20, hidden layer dimension 64, 2-layer structure, ReLU activation function, and dropout rate 0.

1.

5. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 4, characterized in that, The network structure of the hybrid control generation layer is as follows: Robust baseline control employs a sliding mode control law: u bass = -k s sgn(s) - k e e; wherein u base is a robust baseline control output, x is a state tracking error, including position error, velocity error and angle error; λ is a sliding mode surface coefficient, sgn(s) is a sign function, e is a tracking error vector, k s and k e are constant coefficients; Linear observer for estimating linear disturbance where y is the actual output state vector, is the predicted state vector output by the linear observer, C is the identity matrix, A, B are the coefficient matrices of the linear disturbance observer, and L is the pole placement gain. Control weight adjustment: w=σ(κ||d||); where σ is a sigmoid function, and κ is a constant coefficient; is the total interference, is the interference estimation value of the nonlinear interference observer output; Final control output: u = wu bass + (1 - w)u comp ; wherein u comp is the compensation output of the compensation control strategy.

6. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 5, characterized in that, The nonlinear disturbance observer used to estimate the nonlinear disturbance in flight comprises: The dynamic equation of the nonlinear disturbance observer is: wherein is the disturbance estimate output by the nonlinear disturbance observer, A d , B d is the coefficient matrix of the nonlinear disturbance observer, C d is the bias matrix; A d = -5I, B d = 10I, C d = 0, I is the identity matrix.

7. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 6, characterized in that, The receding horizon optimization algorithm used to optimize the control input of the hierarchical neural network in the prediction window comprises: The prediction window N of the receding horizon optimization algorithm is set to 20; The optimization objective of the receding horizon optimization algorithm is established as follows: The constraint is: u c，k ∈ [-10, 10] N-m; The solving method is selected as the interior point method; where e k is the error vector between the current state information fed back by the sensor and the predicted future k-time state, u c,k is the optimized hierarchical neural network control input vector, Q' is the prediction error weight, and D is the control energy consumption weight.

8. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 1, characterized in that, The quadrotor UAV body is made of carbon fiber composite material, the arm structure adopts a streamlined variable cross-section arm with a narrow front end and a wide rear end to reduce air resistance; the battery is placed at the bottom center of the quadrotor UAV body, and the motor is installed at the end of the arm, and the motor installation angle is adjusted.

9. The body-aware intelligent UAV adaptive agile control method for high-speed maneuvering flight according to claim 1, characterized in that, The quadrotor UAV power system redundancy design and fault-tolerant control are also performed: The Smith prediction compensation algorithm of the brushless motor is used to eliminate the power lag, when a single rotor fails, a three-rotor fault-tolerant mode is adopted: adjusting the remaining motor speed to the saturation zone, using the yaw differential speed torque to compensate for the unbalanced moment, and combining the sliding mode control attitude reconstruction algorithm to improve the recovery success rate after losing control.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-9. The computer program is executed by the processor to implement the steps of the method according to any one of claims 1-7.

Images