Intelligent long-endurance composite wing unmanned aerial vehicle system and control method thereof

Abstract

The application discloses an intelligent long-haul composite wing unmanned aerial vehicle system and a control method thereof, collects flight state information, constructs a multi-time scale disturbance state observer, compensates position and speed deviation caused by a steady wind field by using a GPS low-frequency outer ring, estimates lumped disturbance caused by a gust and model uncertainty by using an IMU high-frequency inner ring, calculates a normalized transition factor according to a flight mode, convexly combines a vertical take-off and landing and horizontal flight state space matrix, forms a continuous linear variable parameter model, adaptively adjusts a state weighting matrix according to disturbance high and low frequency energy, and smoothes and adjusts a virtual control input penalty weight in a control weighting matrix according to instantaneous energy consumption; a Riccati equation is solved to generate a virtual control instruction, a constraint quadratic programming is established by taking an instruction tracking error and actuator power consumption as a target, and a rotor speed and rudder deflection angle control instruction is output, so that transition stability, anti-disturbance performance and endurance economy are considered.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicles (UAVs), and in particular relates to an intelligent long-endurance compound wing UAV system and its control method. Background Technology

[0002] In recent years, unmanned aerial vehicle (UAV) technology has been widely applied in various fields such as aerial surveying, material inspection, and disaster relief. Compound-wing UAVs combine the vertical takeoff and landing capabilities of multi-rotor aircraft with the efficient horizontal cruising capabilities of fixed-wing aircraft, offering advantages such as low takeoff and landing site requirements, long flight range, and high cruising speed, making them an important development direction for long-endurance UAVs. However, long-endurance compound-wing UAVs have complex flight profiles, typically involving vertical takeoff and landing, mode transitions, and fixed-wing level flight. When performing long-distance missions, they are also affected by complex meteorological environments such as high-altitude steady-state wind fields and sudden gusts. Furthermore, compound-wing UAVs are equipped with multiple redundant heterogeneous actuation mechanisms, resulting in significant differences in aerodynamic characteristics under different flight modes, and strong nonlinear coupling and model uncertainties within the system. Therefore, maintaining the stability and reliability of the aircraft under multi-modal flight switching and external disturbances is a crucial problem that urgently needs to be solved in the field of UAV control.

[0003] While existing UAV control schemes can maintain basic aircraft operation, they often struggle to effectively distinguish and compensate for slowly changing steady-state low-frequency wind fields and sudden high-frequency gusts when dealing with complex external disturbances. This results in insufficient suppression of position deviations and attitude oscillations during flight. During the transition from vertical takeoff and landing to horizontal flight, existing methods also lack mechanisms for recognizing and smoothly adjusting the continuous evolution of flight states, easily leading to abrupt control changes or even flight instability. Furthermore, long-endurance missions place higher demands on overall energy efficiency, while existing control and allocation strategies primarily focus on trajectory tracking accuracy, failing to adequately consider the power consumption of redundant actuators in high-frequency disturbance rejection adjustments, and lacking online optimization mechanisms that combine disturbance characteristics and real-time energy consumption levels. Therefore, a comprehensive control method is urgently needed that can suppress and compensate for environmental disturbances across multiple time scales, ensure smooth modal transitions, and balance energy optimization with the requirements of long-endurance flight. Summary of the Invention

[0004] According to one aspect of the present invention, a control method for an intelligent long-endurance compound-wing unmanned aerial vehicle (UAV) is proposed, which is used to comprehensively suppress and compensate for environmental disturbances at multiple time scales, ensure smooth transition of modes throughout the entire process, optimize energy consumption online, and achieve long-endurance deployment, including the following steps: The flight status information of the compound-wing UAV is obtained, a multi-timescale disturbance state observer is constructed, and the low-frequency position and velocity deviation caused by the steady-state wind field is estimated online using the outer ring low-frequency channel of the global positioning system information. Based on the compensated state, the lumped disturbance caused by gusts and model uncertainties is estimated using the inner ring high-frequency channel of the inertial measurement unit information, and the estimated values ​​of the aircraft state and lumped disturbance are obtained. Based on the current flight mode, a normalized transition factor representing the transition degree between vertical takeoff and landing and horizontal flight is calculated. Based on the normalized transition factor, the state space matrices of vertical takeoff and landing and horizontal flight are convexly combined to obtain a continuously changing linear variable parameter state space model. A state weighting matrix is ​​constructed, the lumped disturbance estimate is decomposed into high and low frequency components, and the state weighting matrix is ​​adjusted based on the high and low frequency components. The instantaneous energy consumption index is calculated based on the actuator command of the previous cycle, and the weights corresponding to the virtual control input in the control weighting matrix are adjusted based on the instantaneous energy consumption index. Based on the state space model, state weighting matrix, and control weighting matrix, the algebraic Riccati equation is solved to obtain the state feedback gain matrix. The virtual control command is calculated by combining the estimated value. A control allocation objective function is established for the redundant actuator, which includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption prediction values ​​based on local quadratic approximation. The control commands for the rotational speed of each rotor motor and the deflection angle of the servo motor are obtained by solving the constrained quadratic programming problem.

[0005] According to another aspect of the present invention, an intelligent long-endurance compound wing unmanned aerial vehicle system is proposed, comprising the following modules: The module is used to acquire flight status information of the compound-wing UAV, construct a multi-timescale disturbance state observer, use the outer ring low-frequency channel of the global positioning system information to estimate and compensate for the low-frequency position and velocity deviation caused by the steady-state wind field, and use the inner ring high-frequency channel of the inertial measurement unit information as a reference to estimate the lumped disturbance caused by gusts and model uncertainties, so as to obtain the estimated values ​​of the aircraft state and lumped disturbance. The processing module is used to calculate a normalized transition factor representing the transition degree between vertical takeoff and landing and horizontal flight based on the current flight mode, perform a convex combination of the state space matrices of vertical takeoff and landing and horizontal flight based on the normalized transition factor to obtain a continuously changing linear variable parameter state space model, construct a state weighting matrix, decompose the lumped disturbance estimate into high and low frequency components, adjust the state weighting matrix based on the high and low frequency components, calculate the instantaneous energy consumption index according to the actuator command of the previous cycle, and adjust the weights in the control weighting matrix corresponding to the virtual control input based on the instantaneous energy consumption index. The calculation module is used to solve the algebraic Riccati equation based on the state space model, state weighting matrix and control weighting matrix to obtain the state feedback gain matrix, calculate the virtual control command in combination with the estimated value, establish a control allocation objective function for the redundant actuator that includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption prediction values ​​based on local quadratic approximation, and obtain the control command for the rotational speed of each rotor motor and the deflection angle of the servo motor by solving the constrained quadratic programming problem.

[0006] This invention improves the anti-disturbance capability and flight control accuracy of UAVs in complex weather conditions by constructing a multi-timescale disturbance state observer. It utilizes an outer-loop low-frequency channel to estimate and compensate for slow-varying position and velocity deviations caused by steady-state wind fields, and an inner-loop high-frequency channel to estimate lumped disturbances caused by gusts and model uncertainties in real time. By normalizing the transition factor and performing a convex combination of the state space matrices of the vertical takeoff and landing mode and the horizontal flight mode, a continuously changing linear variable parameter model is formed, enabling smooth switching between different flight modes and reducing state abrupt changes and mechanical shocks. Furthermore, the state weighting matrix is ​​adaptively adjusted according to the proportion of high- and low-frequency disturbance energy, dynamically emphasizing attitude stability and position holding in the control strategy. Simultaneously, the penalty weight of the upper-level virtual control input is smoothly increased by combining instantaneous energy consumption indicators, and the optimal control command is solved through a control allocation objective function that includes virtual command tracking error and actuator power consumption prediction values. While ensuring the coordinated operation of redundant actuators, unnecessary power consumption is reduced, improving the long-endurance economic flight performance of the compound-wing UAV. Attached Figure Description

[0007] Figure 1 A flowchart of the first embodiment; Figure 2 This is a graph showing the normalized transition factor as a function of airspeed. Figure 3 This is a time-domain comparison curve of the normalized energy of high and low frequency disturbances. Detailed Implementation

[0008] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0009] In the first embodiment, the present invention proposes a control method for an intelligent long-endurance compound-wing unmanned aerial vehicle, such as... Figure 1 It includes the following steps: S1. Obtain flight state information of the compound-wing UAV, construct a multi-timescale disturbance state observer, use the outer ring low-frequency channel of the global positioning system information to estimate and compensate for the low-frequency position and velocity deviation caused by the steady-state wind field online, and use the inner ring high-frequency channel of the inertial measurement unit information as the benchmark to estimate the lumped disturbance caused by gusts and model uncertainties, and obtain the estimated values ​​of the aircraft state and lumped disturbance.

[0010] The robot obtains 3D position and velocity data from the Global Positioning System (GPS) and 3D acceleration and angular velocity data from the Inertial Measurement Unit (IMU) by subscribing to sensor topics through the robot's operating system nodes. Two independent threads are allocated on the flight control motherboard. The low-frequency thread operates at 10 Hz and uses the extended Kalman filter function from the open-source filtering library Filterpy to construct the outer loop observer, converting latitude, longitude, and altitude coordinates to the local Northeast Geodetic Coordinate System (NGCS) as the measurement input. Using the kinematic equations with wind field and wind speed terms as the state prediction model, the Kalman gain is calculated, and the state covariance matrices of position and velocity are updated to separate the low-frequency steady-state wind field disturbance component. The high-frequency thread operates at 200 Hz, using the compensated state output by the low-frequency thread as the reference state point. The inner loop disturbance observer is constructed based on Newton-Euler six-degree-of-freedom dynamics. The translational channel uses Newton's second law to establish the relationship between the resultant force and linear acceleration of the system, while the rotational channel uses the angular momentum theorem to establish the relationship between the resultant torque and angular acceleration of the system. An explicit Euler integral algorithm is used to discretize and solve the nonlinear state differential equations, thereby calculating in real time the lumped disturbance forces and moments caused by gusts and aerodynamic parameter perturbations. The inner and outer loop data are then combined to output a fused aircraft full-state estimation vector and a lumped disturbance estimation vector. Thus, the three-axis force disturbances are estimated using translational dynamics equations, and the three-axis moment disturbances are estimated using rotational dynamics equations, avoiding the lack of a physical equation source for force disturbance estimation due to reliance solely on the angular momentum theorem.

[0011] In an optional embodiment, the steps of acquiring the flight state information of the composite-wing UAV, constructing a multi-timescale disturbance state observer, using the outer ring low-frequency channel of GPS information to online estimate and compensate for the low-frequency position and velocity deviations caused by the steady-state wind field, and using the compensated state as a reference to estimate the lumped disturbances caused by gusts and model uncertainties using the inner ring high-frequency channel of inertial measurement unit information, to obtain the estimated values ​​of the aircraft state and the lumped disturbances, include: Establish a discrete dynamic state-space observation model with the incremental dimension of the system's expanded state variables; In each control cycle, the actuator control command and actuation mechanism action deviation of the previous cycle are extracted, and the position and velocity compensation amount generated by the outer loop channel and the three-dimensional angular rate and linear acceleration data obtained by the inertial measurement unit are used as the input source of the observation model. Using a preset discrete control step size, the extended state observation algorithm is used to perform iterative integral calculations on unmodeled and high-frequency gust disturbances, outputting the updated aircraft state estimation matrix in real time, and extracting the state variables of the extended dimension as the lumped disturbance estimate for the current period.

[0012] The core of constructing a multi-timescale disturbance state observer lies in establishing an 18-dimensional discrete dynamic state space observation model that includes extended states. The first 12 dimensions are the basic state variables of the aircraft, including three-dimensional position, three-dimensional velocity, three-dimensional attitude angle, and three-dimensional angular rate. The last 6 incremental dimensions correspond to the lumped disturbances of the three-axis forces and three-axis moments caused by unmodeled factors and gusts. In each discrete control cycle, the preset discrete control step size is preferably within the range of 0.005s to 0.02s, exemplarily taken as 0.01s. The actuator control commands of the previous cycle are extracted, such as the rotational speed commands of the four rotors, with a speed range of 0 to 6000 RPM; and the servo deflection angle commands of the ailerons, horizontal stabilizer, and vertical stabilizer, with a deflection range of -30° to +30°, as well as the actual action deviation of the actuation mechanism. The low-frequency position and velocity compensation values ​​generated by the outer-loop Global Positioning System (GPS) channel (typically updated at 10Hz) are combined with the three-dimensional angular rate (range, for example ±300° / s) and three-dimensional linear acceleration (range, for example ±4g) acquired by the inner-loop high-frequency inertial measurement unit (INS) at an update frequency of 100Hz or higher, as the full-state multi-source inputs of this 18-dimensional observation model. An extended state observation algorithm with explicit Euler method is used to iteratively integrate the multi-source inputs. The observer bandwidth parameter needs to be set, preferably in the range of 15 to 40 rad / s, exemplarily taken as 25 rad / s. The prediction deviation is calculated through the state error feedback term, and the 18×1-dimensional aircraft state estimation matrix is ​​updated in real time. The state variables corresponding to the last six extended dimensions of this state estimation matrix are extracted and used as the estimated values ​​of the lumped disturbance of the three-axis forces and three-axis moments for the current control cycle. This implementation mechanism leverages the timescale differences between different sensors and the inner and outer loop frequencies. Based on low-frequency correction of the accumulated spatial position deviation caused by the steady-state wind field, it uses high-frequency broadband detection of high-frequency discrete gusts and nonlinear dynamic coupling and unmodeled errors caused by center-of-gravity shift, ensuring the timeliness and effectiveness of feedforward compensation in subsequent control modules. Specifically, the 18-dimensional state estimation matrix is ​​not directly multiplied by the feedback gain matrix as a whole when solving the subsequent control law. Instead, it is divided into a first 12-dimensional basic state estimate and a last 6-dimensional disturbance estimate. The first 12 dimensions are used for state feedback, and the last 6 dimensions are used for disturbance feedforward compensation.

[0013] S2, calculate the normalized transition factor representing the transition degree between vertical takeoff and landing and horizontal flight based on the current flight mode, perform a convex combination of the state space matrices of vertical takeoff and landing and horizontal flight based on the normalized transition factor to obtain a continuously changing linear variable parameter state space model, construct a state weighting matrix, decompose the lumped disturbance estimate into high and low frequency components, and adjust the state weighting matrix based on the high and low frequency components, calculate the instantaneous energy consumption index based on the actuator command of the previous cycle, and adjust the weights in the control weighting matrix corresponding to the virtual control input based on the instantaneous energy consumption index.

[0014] The current feedback angle of the rotor tilt servo motor in the flight controller is read. If the tilt angle feedback is used to represent the transition state, the tilt angle is first divided by 90 degrees to obtain the horizontal mode proportion factor η, and η is limited to the range of 0 to 1. Then, the normalized transition factor is defined as λ = 1 - η, so that λ always represents the vertical take-off and landing mode weight, and 1 - λ always represents the horizontal flight mode weight. The system state matrix and input matrix of the hovering multirotor mode and the fixed-wing level flight mode are pre-stored in memory. In this embodiment, the state space matrix includes the state matrix A and the control input matrix B. The convex combination processing is applied synchronously to the state matrix and control input matrix under the vertical take-off and landing mode and the level flight mode to ensure the integrity of the system matrix dimension required for the subsequent solution of the Riccati equation. The linear algebra module operation function is used to multiply the hovering matrix by the normalized transition factor λ, multiply the level flight matrix by 1 - λ, and then add the two matrices to generate the linear parameter change state space matrix at the current moment. Therefore, regardless of whether tilt angle feedback or airspeed feedback is used, the physical meaning of the normalized transition factor remains the same: the larger λ is, the closer it is to the vertical takeoff and landing mode, and the smaller λ is, the closer it is to the horizontal flight mode.

[0015] Using the Butterworth digital filter design function in the signal processing module, a cross-filter bank consisting of a high-pass filter and a low-pass filter is constructed. The cutoff frequency of the filter bank is a preset frequency, preferably ranging from 0.5Hz to 2.0Hz, and 1.0Hz is used as an example. The lumped disturbance estimation vector is input into the filter bank to obtain high- and low-frequency time series. The high-frequency disturbance energy and low-frequency disturbance energy are calculated by obtaining the square integral value or root mean square energy index within a preset sliding time window, preferably 1.0s to 5.0s. Before constructing the state weighting matrix and the control weighting matrix, the state deviation and virtual control input are dimensionlessly normalized according to preset position range, velocity range, attitude angle range, angular rate range, and rated range of virtual control force and virtual control torque, so that the diagonal elements of the weighting matrix act on the normalized variables. When the high-frequency disturbance energy is greater than the low-frequency disturbance energy, the elements in the diagonal state weighting matrix corresponding to the attitude angle and angular velocity are magnified by two times; conversely, the elements corresponding to the position and linear velocity are magnified by two times. The rotor motor speed command from the previous control cycle recorded by the flight control system is extracted, or the pulse width modulation command is converted into an equivalent speed command through the electronic speed controller calibration model. The control surface deflection command is also extracted and substituted into the pre-identified cubic power fitting polynomial of the motor and the control surface aerodynamic drag power model to obtain the instantaneous power prediction value of the entire aircraft. When this power value exceeds the set cruise economic power limit, the excess difference is calculated and input into a discrete first-order inertial low-pass filter to filter out high-frequency noise. The smoothed scalar output of the filter is extracted as the penalty coefficient increment. A matrix diagonal element update function is used to add the smoothed scalar as the control weight increment to the diagonal elements in the control weighting matrix R corresponding to the three-dimensional virtual control force and three-dimensional virtual control torque, thereby increasing the penalty weight of the upper-level virtual control input channel without changing the elements in the state weighting matrix Q. The energy consumption suppression of the actual rotor motor speed command and servo deflection angle command is achieved by the actuator power consumption prediction term in the subsequent control allocation objective function. The adjustment of the state weighting matrix Q is used to change the convergence focus of the state error, and the adjustment of the control weighting matrix R is used to suppress high-energy-consuming actuator actions. Both affect the state term and control input term in the LQR cost function, respectively. It should be noted that the control input term in the control weighting matrix R does not directly correspond to the actual rotor motor speed command and servo deflection angle command, but rather to the upper-level virtual control input channel after equivalent mapping by the control efficiency matrix. The actual rotor motor speed command and servo deflection angle command are still obtained by subsequent control allocation quadratic programming under actuator physical constraints. Furthermore, each diagonal element in the control weighting matrix R has a preset upper limit and a basic lower limit. When the instantaneous energy consumption index falls below the economic safety threshold, the corresponding diagonal element gradually falls back to the basic control weight value according to a preset recovery coefficient to avoid long-term cumulative increases that could lead to an overly conservative control response.

[0016] In an optional embodiment, the step of calculating a normalized transition factor representing the transition degree between vertical takeoff and landing (VTOL) and horizontal flight based on the current flight mode, and performing a convex combination of the state space matrices of VTOL and horizontal flight based on the normalized transition factor to obtain a continuously changing linear variable parameter state space model includes: Read the current actual airspeed value of the drone as measured by the airspeed sensor; The difference between the preset maximum horizontal reference airspeed and the current actual airspeed value is divided by the maximum horizontal reference airspeed to obtain the reference ratio, and the reference ratio is limited to a closed interval of 0 to 1 to generate a normalized transition factor. The normalized transition factor is used as the weight for vertical takeoff and landing mode to extract the vertical takeoff and landing state space model matrix, which is then multiplied by the weight to obtain the first sub-matrix. The normalized transition factor is subtracted from 1 as the weight for horizontal flight mode to extract the horizontal flight state space model matrix, which is then multiplied by the weight to obtain the second sub-matrix. The first submatrix and the second submatrix are combined by matrix addition to generate a linear variable parameter state-space model that adjusts in real time with airspeed.

[0017] The steps to obtain a continuously varying linear variable parameter state-space model include reading the current UAV's true airspeed value measured in real time by an onboard pitot tube or differential pressure airspeed sensor via a microcontroller bus. The measurement range is typically set to 0 to 40 m / s. A dimensionless normalized transition factor representing the degree of transition between rotor and fixed-wing hybrid power operation is calculated. A preset maximum horizontal reference airspeed parameter is set by looking up a table or through pre-simulated aerodynamics. This parameter represents the boundary airspeed at which the compound-wing UAV relies on the fixed wing surface to generate lift to maintain level flight; the preferred range is 20 to 28 m / s, with 22 m / s used as an example. The difference between this reference airspeed and the current true airspeed is calculated using a formula. This difference is divided by the reference airspeed to obtain a reference percentage, which is then constrained within a closed interval [0,1] using limiting logic: when the true airspeed is equal to or less than 0 m / s, the transition factor output is 1; when the true airspeed is greater than or equal to 22 m / s, the transition factor output is 0, thus generating a normalized transition factor representing the smooth transition characteristics. The normalized transition factor in this paper uniformly represents the weight of the vertical takeoff and landing mode, and therefore its value decreases as the degree of horizontal flight increases. The curve of the normalized transition factor as a function of airspeed is shown below. Figure 2 As shown.

[0018] Based on this, a linear variable-parameter state-space model adapted to the continuous time-varying characteristics within the entire envelope domain of the transition corridor is constructed. Let A1 and B1 be the pure vertical takeoff and landing (VTOL) state-space model matrices obtained offline using a system identification method, and A2 and B2 be the pure horizontal flight state-space model matrices. Both the VTOL and horizontal flight state-space model matrices are established based on the same state variable ordering, the same coordinate system, and the same six-dimensional virtual control input definition. Using the normalized transition factor as the convex combination weight scalar of the VTOL mode, scalar multiplication is performed on all elements in matrices A1 and B1 to obtain the first sub-matrix. Using the difference between 1 and the normalized transition factor as the horizontal flight mode weight scalar, matrix A2 and B2 are multiplied and scaled to obtain the second sub-matrix. The first and second sub-matrices, having the same dimension, are then merged using standard matrix addition to generate the fused state-space system matrix A and control input matrix B. The model generated by the aforementioned convex combination algorithm realizes the mapping of rotor aerodynamic evolution to fixed wing aerodynamics, reducing lift abrupt changes and pitch attitude step oscillations that are easily caused by control law switching. When using tilt servo angle instead of airspeed as the transition state input, the tilt angle should first be converted into the proportion of horizontal modes, and then its complement should be taken as the normalized transition factor to keep the direction of matrix convex combination consistent with the airspeed implementation method.

[0019] In an optional embodiment, constructing the state weighting matrix, decomposing the lumped disturbance estimate into high- and low-frequency components, and adjusting the state weighting matrix based on the high- and low-frequency components, includes: Construct a low-pass filter with a preset cutoff frequency, input the lumped disturbance estimate into the filter to extract the gradually changing signal at the output as the low-frequency disturbance component, and extract the signal exceeding the low-pass frequency band as the high-frequency disturbance component by subtracting the original lumped disturbance estimate from the low-frequency disturbance component. Within a preset sliding time window, the high-frequency disturbance component and the low-frequency disturbance component are respectively normalized to the same dimension, and then square integral operation is performed to calculate the high-frequency disturbance energy value and the low-frequency disturbance energy value. By comparing the high-frequency disturbance energy value with the low-frequency disturbance energy value, if the high-frequency disturbance energy value is greater than the low-frequency disturbance energy value, a weight adjustment command is triggered to increase the weight value of the diagonal elements in the state weighting matrix corresponding to the attitude angle and angular rate; otherwise, the weight value of the diagonal elements in the state weighting matrix corresponding to the position and velocity is increased.

[0020] The process of decomposing the lumped disturbance estimate into high- and low-frequency components is achieved through a combination of software-defined digital filters and energy spectrum analysis. A second-order Butterworth or first-order RC low-pass filter is constructed in the digital signal processing module, with a filter cutoff frequency set. The preferred range for this frequency is 0.5 to 2.0 Hz, exemplarily set to 1.0 Hz. This threshold corresponds to steady-state winds and slowly accumulating model errors in space meteorology. The six-DOF lumped disturbance estimate output from the observer is input to this filter in real time for filtering. The continuously changing low-frequency signal at the output is extracted as the low-frequency disturbance component. Using the principle of linear signal subtraction, the original lumped disturbance estimate sequence at the filter inlet is subtracted from the synchronously output low-frequency disturbance component, separating and extracting the changing signal exceeding the 1.0 Hz low-frequency threshold, which represents the high-frequency disturbance component representing local gust wind shear and blade aerodynamic flutter. A fixed memory buffer is allocated to establish a preset sliding time window, with the preferred span of this time window set to 1.0 to 5.0 s, for example, 2.0 s as the statistical sample interval. Since the separated force and torque disturbance components belong to different dimensions, namely Newton and Newton-meter, it is necessary to first perform division operations on the disturbance components of each channel using the system's rated maximum disturbance tolerance to complete the dimensionless normalization process. Within this sliding window period, the normalized high-frequency and low-frequency time-domain data are squared one by one, and definite integrals or discrete summation operations are performed to calculate and output the high-frequency disturbance energy value and the low-frequency disturbance energy value within this time scale. Energy threshold comparisons are performed. If the system determines that the high-frequency disturbance energy value is greater than the low-frequency disturbance energy value, indicating that the high-frequency discrete gust energy dominates the current disturbed state of the aircraft, an event-driven weight adjustment command is triggered. This involves increasing the weight values ​​of the diagonal elements in the LQR controller's system state weighting matrix corresponding to the three-dimensional Euler attitude angle and three-dimensional body axis angular rate by a multiplication factor, such as increasing it by 1.5 to 3.0 times. This sacrifices position control tolerance for attitude convergence in response to high-frequency wind field excitation. Conversely, when the assessment determines that the low-frequency disturbance energy is dominant, the weight values ​​of the diagonal elements in the matrix corresponding to the three-dimensional Earth inertial coordinate system position and linear velocity are increased, such as to 2.0 times the initial value. This strengthens the stiffness of the outer loop control loop, resisting position drift caused by steady-state crosswinds or headwinds and maintaining three-dimensional trackpoint tracking capability. The time-domain comparison curves of normalized energy for high and low frequency disturbances are shown below. Figure 3 As shown.

[0021] In an optional embodiment, the step of calculating the instantaneous energy consumption index based on the actuator instruction of the previous cycle, and adjusting the weights corresponding to the virtual control input in the control weighting matrix based on the instantaneous energy consumption index, includes: Read the rotor motor speed commands and servo deflection angle commands from the previous control cycle, calculate the instantaneous power consumption prediction value of each actuator using the preset actuator power mapping model, and sum the instantaneous power consumption prediction values ​​of all actuators to obtain the instantaneous energy consumption index for evaluating the current total energy consumption level of the aircraft. Determine whether the instantaneous energy consumption index is greater than a preset long-endurance economic safety threshold; When the instantaneous energy consumption index is greater than the economic safety threshold, the overshoot deviation between the instantaneous energy consumption index and the economic safety threshold is calculated. The overshoot deviation is multiplied by a preset cost penalty coefficient that has a dimension conversion function, and smoothed by a preset low-pass filter to obtain a continuously changing control increment. The control increment is then superimposed on the diagonal elements of the control weighting matrix corresponding to the three-dimensional virtual control force and the three-dimensional virtual control torque. The energy consumption constraints of the actual rotor motor speed command and servo deflection angle command are handled by the actuator power consumption prediction term in the subsequent control allocation objective function, so as to increase the penalty weight of the upper controller on the virtual control input and its corresponding energy consumption overshoot.

[0022] The instantaneous energy consumption of the entire aircraft is quantitatively calculated based on the actuator commands from the previous control cycle, relying on the actuator power mapping model obtained by fitting ground test data. At the beginning of each control step, the actuator input command data generated by the flight controller in the previous cycle is read. Specifically, this includes the electronic speed controller speed commands for each brushless rotor motor of the compound wing in the current operating state, such as speed values ​​from 1500 to 5500 RPM, and the deflection angle commands for each aerodynamic servo of the wing and tail, such as the range from -25° to +25°. These commands are substituted into their respective independently constructed nonlinear actuator power mapping models. The rotor power consumption can be mapped by the cubic aerodynamic drag polynomial of the rotational speed variable, while the servo power consumption is related to the square of the change in deflection angle when executing commands and the nonlinear frictional resistance to overcome the deflection hinge torque. The instantaneous power consumption prediction value required by each motor and servo actuator at this time is calculated separately. The power prediction values ​​of these multiple independent branches are arithmetically summed, and the output represents the instantaneous energy consumption index for evaluating the actual total discharge load of the aircraft's current battery terminal, in watts (W).

[0023] After the calculation is completed, the built-in power consumption monitoring process will verify and determine whether the total instantaneous energy consumption index exceeds the long-endurance economic safety threshold preset by the battery's maximum continuous discharge capacity and expected flight time. For example, for a 15kg-class flight platform, this safe power threshold can be preferably set to 1200W. When the current energy consumption level is detected to be greater than this economic safety threshold, the instantaneous energy consumption index is subtracted from the economic safety threshold to calculate the power overshoot deviation. Since the power data alone is incompatible with the control matrix in terms of dimensions and orders of magnitude, the overshoot deviation needs to be multiplied by a preset penalty coefficient that has the function of dimension unification conversion and proportional adjustment. The preferred order of magnitude range for this coefficient is 0.01 to 0.1. Example values To reduce high-frequency power fluctuations in the control signal, the converted product is input to a first-order low-pass smoothing filter, with a preferred time constant of 0.5s, to obtain a continuously and slowly varying control increment over time. This control increment is then scalarized and superimposed in real-time onto the diagonal elements of the control weight matrix corresponding to the upper-level virtual control input in the quadratic cost function. As energy consumption exceeds the threshold, the value of the diagonal elements expands, causing the calculated controller to impose a higher penalty on action energy loss, suppressing excessive amplitude or high-frequency changes in upper-level virtual control commands. Furthermore, the subsequent control allocation objective function suppresses high-frequency reciprocating corrections or large-scale deflection behaviors of the actuators. By tolerating minor flight attitude fluctuations, the control platform's current output is controlled, improving overall mission range economy and meeting long-endurance requirements.

[0024] S3. Based on the state space model, state weighting matrix, and control weighting matrix, solve the algebraic Riccati equation to obtain the state feedback gain matrix. Combine the estimated value to calculate the virtual control command. Establish a control allocation objective function for the redundant actuator, which includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption prediction values ​​based on local quadratic approximation. Obtain the control commands for the rotational speed of each rotor motor and the deflection angle of the servo motor by solving the constrained quadratic programming problem.

[0025] Before solving the algebraic Riccati equation, the state weighting matrix Q is set as a positive semi-definite diagonal matrix, and the control weighting matrix R is set as a positive definite diagonal matrix. A preset lower limit ensures that all diagonal elements of R are always greater than zero. Simultaneously, the state-space model satisfies the stability condition to guarantee the existence of a stable solution for feedback control in the Riccati equation. When the linear parameter-changing state-space matrix is ​​a continuous-time model matrix, it is combined with the state weighting matrix Q and the control weighting matrix R to solve the continuous algebraic Riccati equation. When the state-space matrix has been discretized into a discrete-time model matrix according to the control step size, a discrete algebraic Riccati equation is constructed to solve the state feedback gain matrix. The corresponding Riccati equation solver is called according to the time form of the state-space model. When the model is in continuous-time form, the continuous algebraic Riccati equation solver is called to obtain a symmetric positive semi-definite stable solution; when the model is in discrete-time form, the discrete algebraic Riccati equation solver is called to obtain the corresponding matrix solution, and the state feedback gain matrix is ​​calculated accordingly. Before calculating the virtual control commands, the first 12 dimensions of the basic state estimates are extracted from the 18-dimensional aircraft state estimation matrix by selecting matrix Sx=[I12 0], forming a basic state deviation vector with the same dimensions as the state feedback gain matrix. Then, the last 6 dimensions of the extended disturbance estimates are extracted by selecting matrix Sd=[0 I6], serving as the disturbance feedforward compensation for three-dimensional forces and three-dimensional moments. The deviation vector between the set reference trajectory and the first 12-dimensional basic state estimates is multiplied by this feedback gain matrix, and combined with the last 6-dimensional lumped disturbance estimate vector for feedforward compensation, outputting a virtual control command vector containing the desired three-dimensional forces and three-dimensional moments. By separating the basic state feedback and the extended disturbance feedforward, the matrix dimensions of the 18-dimensional observation model output and the 12-dimensional LQR feedback gain matrix are kept consistent.

[0026] For an actuator system with multiple rotors, elevators, ailerons, and rudders, a cost function is defined that includes command tracking error and power consumption optimization. The first part calculates the sum of squares of the differences between the virtual control command and the virtual control output predicted by the actuator control increment, based on the control performance matrix obtained by local linearization at the actuator operating point of the previous cycle. The second part uses a Taylor expansion function to perform a second-order Taylor polynomial expansion of the nonlinear actuator aerodynamic power consumption function at the actuator operating point of the previous cycle or the current predicted nominal operating point, extracting the Hessian matrix and gradient vector from the expansion to construct a locally quadratic form power consumption prediction term. To ensure that the quadratic programming objective function satisfies the convex optimization solution conditions, the Hessian matrix is ​​first symmetricized, then subjected to positive semi-definite projection or diagonal regularization. When the minimum eigenvalue is less than 0, the Hessian matrix is ​​corrected to... Where V is the eigenvector matrix, Λ is the eigenvalue diagonal matrix, and ε is the non-negative small regularization coefficient, thus obtaining the positive semi-definite local power consumption quadratic form matrix. The weighted summation of the two parts is transformed into a standard convex quadratic programming cost function. The upper and lower speed limits of each rotor motor and the maximum mechanical deflection angle limits of each control surface are extracted to construct a linear inequality constraint matrix and boundary vector. For the servo motor actuation angular velocity limit, it is first discretized into the current cycle's dynamic boundary by combining the control step size Δt and the actual servo motor deflection angle δ(k-1) of the previous cycle, i.e. , The dynamic upper and lower bounds are then incorporated into the linear boundary constraints of the quadratic programming. The optimization solution function provided in the open-source operator splitting quadratic programming solver library OSQP is called to perform iterative calculations within the preset real-time solution time limit and maximum iteration count constraints. When the convergence tolerance is met, the optimal control quantity that minimizes the total cost function is output. When the convergence tolerance is not met but a feasible solution exists, the current feasible approximate optimal solution is output. If necessary, the feasible solution from the previous cycle or the projected safety control quantity is used as a degraded output. This optimal solution is inversely mapped into a pulse width modulation signal and then distributed to each rotor brushless motor ESC and servo drive module for execution via the controller area network bus. The real-time solution time limit can be preset according to the processor computing power, the number of actuator variables, the number of constraints, and the upper limit of the iteration count. For example, the number of actuator control allocation variables can be limited to the total number of rotor motor and servo channels, the maximum number of iterations can be set to a fixed upper limit, and the optimal solution from the previous cycle can be used for hot start-up to ensure that the solution is completed within the online control cycle.

[0027] In an optional embodiment, the establishment of a control allocation objective function for the redundant actuator, which includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption predictions based on a local quadratic approximation, and the obtaining of control commands for the rotational speed of each rotor motor and the deflection angle of the servo motor by solving a constrained quadratic programming problem, includes: Based on the aircraft dynamics control performance matrix, a predicted output of virtual control commands is constructed. The difference between the predicted output and the virtual control commands generated by the upper-level state feedback controller is calculated, and a first penalty term representing the tracking accuracy is constructed by performing a quadratic calculation on the difference. Based on the preset actuator power mapping model and the actuator control allocation variable to be solved, a local quadratic approximation expansion is performed at the actuator operating point of the previous cycle to calculate the quadratic approximate prediction value of the current instantaneous power consumption of each actuator. The predicted value is then weighted and summed using weights with dimension conversion coefficients to construct a quadratic second penalty term representing energy consumption. The first penalty term and the second penalty term are added to construct a control allocation quadratic programming optimization objective function with the goal of reducing overall energy consumption and reducing tracking error; Extract the upper and lower limits of the rotational speed of each rotor motor and the upper and lower limits of the deflection angle of each servo motor. Calculate the dynamic deflection boundary corresponding to the servo motor angular velocity based on the control step size and the servo motor deflection angle of the previous cycle, and use it as the constraint boundary of the quadratic programming optimization objective function. The minimum value of the quadratic programming optimization objective function is found in the solution space that satisfies the constraints, and the optimal control allocation variable that minimizes the objective function is output, generating actual control commands sent to each rotor motor and each servo motor.

[0028] For UAVs equipped with redundant actuator mechanisms such as quadrotor systems and multi-channel fixed aerodynamic control surfaces, the control allocation logic needs to be transformed into a quadratic programming extremum problem with feasible region constraints. Before constructing the control allocation objective function, the nonlinear performance mapping between rotor thrust, control surface aerodynamic torque, and actuator control quantity at the actuator operating point of the previous cycle is first-order locally linearized. The variable to be solved is defined as the actuator control increment, and the predicted virtual control output is composed of the virtual control quantity at the previous operating point, the local performance matrix multiplied by the control increment. After obtaining the control increment, it is superimposed with the actuator command of the previous cycle and limited by boundary constraints to generate the current actual control command. The aircraft dynamics control efficiency matrix, obtained by locally linearizing the actuator operating point at the current real airspeed, is retrieved. This matrix encompasses the mathematical transformation relationship between the multi-actuator control increments and the six-degree-of-freedom virtual control increments of the fuselage. This matrix is ​​then multiplied with the current actuator control increment to be solved, and superimposed with the virtual control output corresponding to the previous operating point to construct the predicted output of the virtual control command generated by the actual aircraft. This predicted output matrix vector is then subtracted from the three-dimensional desired control force and three-dimensional desired control torque vectors generated by the upstream state feedback controller (i.e., the ideal virtual control command). The residual tracking difference is then calculated using a standard quadratic matrix inner product with a specific diagonalized weight matrix to construct a first penalty term representing the accuracy of the torque tracking command error. This penalty term is equipped with corresponding tracking weights to ensure that the allocated solution conforms to the flight control's intended intent. The energy dissipation characteristics at the redundant hardware level are also considered.

[0029] Based on the pre-calibrated power mapping function model of each independent actuator, a local second-order Taylor expansion containing a first-order Jacobian matrix vector and a second-order Hessian matrix is ​​applied to the current sequence of actuator control allocation variables to be solved. This is done near the actual steady-state actuator operating point of the previous control cycle, such as a specific motor RPM and yaw angle point. This local quadratic approximation method reduces computational burden while transforming the energy consumption function into a locally approximating quadratic formula, calculating the quadratic approximate total value of the current instantaneous power consumption of each corresponding independent channel actuator. Before forming the quadratic programming objective function, the Hessian matrix obtained from the local second-order Taylor expansion is made convex. Preferably, eigenvalue truncation, positive semi-definite projection, or adding an εI regularization term are used to make the quadratic matrix corresponding to the power consumption approximation term a positive semi-definite matrix. When the power consumption model itself satisfies convexity near the local operating point, its symmetric Hessian matrix can be directly used. Furthermore, a conversion ratio parameter containing a dimensionless conversion coefficient is used to unify the dimensions of the tracking error term and the power consumption term. This parameter is typically set between 0.1 and 1.0. The aforementioned power consumption approximation terms are weighted and summed to construct a quadratic second penalty term representing the reduction of the platform's real-time comprehensive energy consumption. The main body of the first penalty term equation and the main body of the second penalty term equation are algebraically added together to establish a multi-objective comprehensive control allocation quadratic programming optimization objective function that comprehensively minimizes the command torque tracking error and reduces control allocation energy consumption.

[0030] Before solving for the minimum value, the external hardware characteristics are extracted and abstracted into mathematical constraints based on inequalities. The inherent lower and upper bounds of the operating speed of each brushless rotor motor are extracted, for example, limited to the lower limit of idle speed (800 RPM) and the upper limit of full throttle (6500 RPM). The maximum deflection angle limits of the mechanical structures achievable by all wing and tail servo motors are also extracted, for example, limiting the actuation stroke to -25° to +25°. For actuation angular velocity constraints, such as a speed constraint of less than 60° / s, these are not directly used as fixed angle boundaries, but are calculated based on the discrete control step size Δt of the current control cycle to determine the achievable angle range for that cycle. If the control surface angle of the previous cycle was δ(k-1), then the control surface angle δ(k) of the current cycle must satisfy... The intersection of this intersection with the upper and lower limits of the mechanical deflection angle yields dynamic upper and lower bounds that can be directly incorporated into the quadratic programming model. These real hardware constraint boundaries are combined to generate a constraint boundary matrix inequality with a defined boundary space. At the processor level, using solver modules embedded in the microcontroller unit, such as the effective set method, the primal dual interior-point algorithm based on barrier functions, or the operator split quadratic programming algorithm, the minimum value of the quadratic programming objective function is calculated within a preset allowable real-time solution time limit within the feasible mathematical solution space satisfying the constraints. A set of multidimensional control allocation variable vectors that ensures the minimization of the objective cost functional is locked and output. This set is converted into a multi-path pulse width modulation (PWM) signal sequence via the communication layer and generates actual servo control signal commands sent to each underlying rotor motor and multi-channel servo motor. The real-time solution time limit is preferably less than the control cycle and can be achieved by limiting the variable size, setting an upper limit for the number of iterations, using the solution from the previous control cycle as an initial value for hot start, and pre-caching the constraint matrix structure.

[0031] In a second embodiment, the present invention also proposes an intelligent long-endurance compound wing unmanned aerial vehicle system, comprising the following modules: The module is used to acquire flight status information of the compound-wing UAV, construct a multi-timescale disturbance state observer, use the outer ring low-frequency channel of the global positioning system information to estimate and compensate for the low-frequency position and velocity deviation caused by the steady-state wind field, and use the inner ring high-frequency channel of the inertial measurement unit information as a reference to estimate the lumped disturbance caused by gusts and model uncertainties, so as to obtain the estimated values ​​of the aircraft state and lumped disturbance. The processing module is used to calculate a normalized transition factor representing the transition degree between vertical takeoff and landing and horizontal flight based on the current flight mode, perform a convex combination of the state space matrices of vertical takeoff and landing and horizontal flight based on the normalized transition factor to obtain a continuously changing linear variable parameter state space model, construct a state weighting matrix, decompose the lumped disturbance estimate into high and low frequency components, adjust the state weighting matrix based on the high and low frequency components, calculate the instantaneous energy consumption index according to the actuator command of the previous cycle, and adjust the weights in the control weighting matrix corresponding to the virtual control input based on the instantaneous energy consumption index. The calculation module is used to solve the algebraic Riccati equation based on the state space model, state weighting matrix and control weighting matrix to obtain the state feedback gain matrix, calculate the virtual control command in combination with the estimated value, establish a control allocation objective function for the redundant actuator that includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption prediction values ​​based on local quadratic approximation, and obtain the control command for the rotational speed of each rotor motor and the deflection angle of the servo motor by solving the constrained quadratic programming problem.

[0032] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0033] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A control method for an intelligent long-endurance compound-wing unmanned aerial vehicle, characterized in that, include: The system acquires flight state information of a compound-wing UAV, constructs a multi-timescale disturbance state observer, and uses the outer low-frequency channel of GPS information to estimate and compensate for low-frequency position and velocity deviations caused by steady-state wind fields online. Based on the compensated state, the inner high-frequency channel of inertial measurement unit information is used to estimate the lumped disturbances caused by gusts and model uncertainties, thus obtaining estimates of the aircraft state and lumped disturbances. A normalized transition factor representing the transition between vertical takeoff and landing (VTOL) and horizontal flight is calculated based on the current flight mode. Based on the normalized transition factor, the state space matrices of VTOL and horizontal flight are convexly combined to obtain a continuously changing linear variable-parameter state space model. A state weighting matrix is ​​constructed, and the estimated lumped disturbances are decomposed into high- and low-frequency components. The state weighting matrix is ​​adjusted based on these components. Instantaneous energy consumption indicators are calculated based on the actuator commands of the previous cycle, and the weights corresponding to the virtual control inputs in the control weighting matrix are adjusted based on these indicators. The state feedback gain matrix is ​​obtained by solving the algebraic Riccati equation based on the state space model, state weighting matrix and control weighting matrix. Virtual control commands are calculated by combining the estimated values. A control allocation objective function is established for the redundant actuator, which includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption prediction values ​​based on local quadratic approximation. The control commands for the rotational speed of each rotor motor and the deflection angle of the servo motor are obtained by solving the constrained quadratic programming problem. A state weighting matrix is ​​constructed to decompose the lumped disturbance estimate into high- and low-frequency components. The state weighting matrix is ​​then adjusted based on these components. This process includes: constructing a low-pass filter with a preset cutoff frequency; inputting the lumped disturbance estimate into the filter to extract a gradually changing signal at the output as the low-frequency disturbance component; and subtracting the original lumped disturbance estimate from the low-frequency disturbance component to extract a signal exceeding the low-pass frequency band as the high-frequency disturbance component. Within a preset sliding time window, the high-frequency and low-frequency disturbance components are normalized to the same dimension, and then square integrals are performed to calculate the high-frequency disturbance energy value and the low-frequency disturbance energy value. The high-frequency disturbance energy value and the low-frequency disturbance energy value are compared. If the high-frequency disturbance energy value is greater than the low-frequency disturbance energy value, a weight adjustment command is triggered to increase the weight values ​​of the diagonal elements in the state weighting matrix corresponding to the attitude angle and angular rate; otherwise, the weight values ​​of the diagonal elements in the state weighting matrix corresponding to the position and velocity are increased. The instantaneous energy consumption index is calculated based on the actuator commands from the previous cycle, and the weights corresponding to the virtual control inputs in the control weighting matrix are adjusted based on the instantaneous energy consumption index. This includes: reading the rotor motor speed commands and servo deflection angle commands from the previous control cycle; calculating the predicted instantaneous power consumption of each actuator using a preset actuator power mapping model; summing the predicted instantaneous power consumption of all actuators to obtain the instantaneous energy consumption index for assessing the current total energy consumption level of the aircraft; determining whether the instantaneous energy consumption index is greater than a preset long-endurance economic safety threshold; calculating the overshoot deviation between the instantaneous energy consumption index and the economic safety threshold when the instantaneous energy consumption index is greater than the economic safety threshold; and multiplying the overshoot deviation by a preset cost penalty coefficient with dimensional conversion function, with a value ranging from 0.01 to... The control increment is obtained by smoothing through a preset low-pass filter, and the control increment is superimposed on the diagonal elements in the control weighting matrix corresponding to the three-dimensional virtual control force and the three-dimensional virtual control torque. The energy consumption constraints of the actual rotor motor speed command and servo deflection angle command are processed by the actuator power consumption prediction term in the subsequent control allocation objective function, so as to increase the penalty weight of the upper controller on the virtual control input and its corresponding energy consumption overshoot.

2. The method according to claim 1, characterized in that, To acquire flight state information of a compound-wing UAV, a multi-timescale disturbance state observer is constructed. The outer-loop low-frequency channel using GPS information is used to online estimate and compensate for low-frequency position and velocity deviations caused by steady-state wind fields. Based on the compensated state, the inner-loop high-frequency channel using inertial measurement unit (IMU) information is used to estimate lumped disturbances caused by gusts and model uncertainties, yielding estimates of the aircraft state and lumped disturbances. This process includes: establishing a discrete dynamic state-space observation model with an incremental dimension of system extended state variables; extracting actuator control commands and actuation mechanism action deviations from the previous cycle within each control cycle, combining the position and velocity compensation amounts generated by the outer-loop channel with the three-dimensional angular rate and linear acceleration data acquired by the IMU as the input source for the observation model; using a preset discrete control step size, iterative integration calculations are performed on unmodeled and high-frequency gust disturbances using an extended state observation algorithm, outputting the updated aircraft state estimation matrix in real time, and extracting the extended-dimensional state variables as the lumped disturbance estimate for the current cycle.

3. The method according to claim 2, characterized in that, Based on the current flight mode, a normalized transition factor representing the transition degree between vertical takeoff and landing (VTOL) and horizontal flight is calculated. Based on this normalized transition factor, the state space matrices of VTOL and horizontal flight are convexly combined to obtain a continuously changing linear variable-parameter state space model. This includes: reading the current actual airspeed value of the UAV measured by the airspeed sensor; dividing the difference between the preset maximum horizontal reference airspeed and the current actual airspeed value by the maximum horizontal reference airspeed to obtain a reference ratio, and limiting the reference ratio within a closed interval of 0 to 1 to generate a normalized transition factor; using the normalized transition factor as the weight for VTOL mode, extracting the VTOL state space model matrix and multiplying it by the weight to obtain a first sub-matrix; using 1 minus the normalized transition factor as the weight for horizontal flight mode, extracting the horizontal flight state space model matrix and multiplying it by the weight to obtain a second sub-matrix; and combining the first and second sub-matrices by matrix addition to generate a linear variable-parameter state space model that adjusts in real time with airspeed.

4. The method according to claim 1, characterized in that, A control allocation objective function is established for the redundant actuators, including a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption predictions based on local quadratic approximation. The control commands for each rotor motor speed and servo deflection angle are obtained by solving a constrained quadratic programming problem. This includes: constructing a predicted output of the virtual control command based on the aircraft dynamics control efficiency matrix; subtracting the predicted output from the virtual control command generated by the upper-level state feedback controller; and constructing a first penalty term representing the tracking accuracy by performing a quadratic form calculation on the difference; and based on a preset actuator power mapping model and the current actuator control allocation variables to be solved, performing a local quadratic approximation expansion at the actuator operating point of the previous cycle to calculate a quadratic approximation prediction of the current instantaneous power consumption of each actuator. The first penalty term is calculated by weighting the predicted values ​​with dimensional conversion coefficients and constructing a second penalty term representing the energy consumption quadratic form. The first penalty term is added to the second penalty term to construct a control allocation quadratic programming optimization objective function aimed at reducing overall energy consumption and tracking error. The upper and lower limits of the rotational speed of each rotor motor and the upper and lower limits of the deflection angle of each servo motor are extracted. The dynamic deflection boundary corresponding to the servo motor angular velocity is calculated based on the control step size and the servo motor deflection angle of the previous cycle, which serves as the constraint boundary of the quadratic programming optimization objective function. The minimum value of the quadratic programming optimization objective function is solved in the solution space that satisfies the constraint boundary. The optimal control allocation variable that minimizes the objective function is output, and the actual control commands sent to each rotor motor and each servo motor are generated.

5. An intelligent long-endurance compound-wing unmanned aerial vehicle system, applied to the method of claim 1, characterized in that, include: The module is used to acquire flight status information of the compound-wing UAV, construct a multi-timescale disturbance state observer, use the outer ring low-frequency channel of the global positioning system information to estimate and compensate for the low-frequency position and velocity deviation caused by the steady-state wind field, and use the inner ring high-frequency channel of the inertial measurement unit information as a reference to estimate the lumped disturbance caused by gusts and model uncertainties, so as to obtain the estimated values ​​of the aircraft state and lumped disturbance. The processing module is used to calculate the normalized transition factor representing the transition degree between vertical takeoff and landing and horizontal flight based on the current flight mode, perform convex combination of the state space matrices of vertical takeoff and landing and horizontal flight based on the normalized transition factor to obtain a continuously changing linear variable parameter state space model, construct a state weighting matrix, decompose the lumped disturbance estimate into high and low frequency components, adjust the state weighting matrix based on the high and low frequency components, calculate the instantaneous energy consumption index based on the actuator command of the previous cycle, and adjust the weights in the control weighting matrix corresponding to the virtual control input based on the instantaneous energy consumption index. The calculation module is used to solve the algebraic Riccati equation based on the state space model, state weighting matrix, and control weighting matrix to obtain the state feedback gain matrix. It calculates virtual control commands by combining the estimated values, establishes a control allocation objective function for redundant actuators that includes a quadratic term of virtual command tracking error and a weighted sum of actuator power consumption predictions based on local quadratic approximation, and obtains the control commands for the rotational speed of each rotor motor and the deflection angle of the servo motor by solving the constrained quadratic programming problem.

6. The system according to claim 5, characterized in that, The flight state information of the compound-wing UAV is acquired, and a multi-timescale disturbance state observer is constructed. The outer low-frequency channel of the GPS information is used to estimate and compensate for the low-frequency position and velocity deviations caused by the steady-state wind field online. Based on the compensated state, the inner high-frequency channel of the inertial measurement unit information is used to estimate the lumped disturbances caused by gusts and model uncertainties, obtaining estimates of the aircraft state and lumped disturbances, including: Establish a discrete dynamic state-space observation model with the incremental dimension of the system's expanded state variables; In each control cycle, the actuator control command and actuation mechanism action deviation of the previous cycle are extracted, and the position and velocity compensation generated by the outer loop channel and the three-dimensional angular rate and linear acceleration data obtained by the inertial measurement unit are used as the input source of the observation model. Using a preset discrete control step size, the extended state observation algorithm is used to perform iterative integral calculations on unmodeled and high-frequency gust disturbances, outputting the updated aircraft state estimation matrix in real time, and extracting the state variables of the extended dimension as the lumped disturbance estimate for the current period.

7. The system according to claim 6, characterized in that, Based on the current flight mode, a normalized transition factor representing the transition degree between vertical takeoff and landing (VTOL) and horizontal flight is calculated. Then, based on this normalized transition factor, a convex combination of the state space matrices for VTOL and horizontal flight is performed to obtain a continuously varying linear variable-parameter state space model, including: Read the current actual airspeed value of the drone as measured by the airspeed sensor; The difference between the preset maximum horizontal reference airspeed and the current actual airspeed is divided by the maximum horizontal reference airspeed to obtain the reference ratio, and the reference ratio is limited to a closed interval of 0 to 1 to generate a normalized transition factor. The first sub-matrix is ​​obtained by multiplying the vertical takeoff and landing state space model matrix by using the normalized transition factor as the weight of the vertical takeoff and landing mode and subtracting the normalized transition factor from 1 as the weight of the horizontal flight mode and then multiplying it by the weight. The first submatrix and the second submatrix are combined by matrix addition to generate a linear variable parameter state-space model that adjusts in real time with airspeed.

8. The system according to claim 7, characterized in that, Construct a state weighting matrix, decompose the lumped disturbance estimate into high- and low-frequency components, and adjust the state weighting matrix based on the high- and low-frequency components, including: Construct a low-pass filter with a preset cutoff frequency, input the lumped disturbance estimate into the filter to extract the gradually changing signal at the output as the low-frequency disturbance component, and extract the signal exceeding the low-pass frequency band as the high-frequency disturbance component by subtracting the original lumped disturbance estimate from the low-frequency disturbance component. Within a preset sliding time window, the high-frequency disturbance component and the low-frequency disturbance component are normalized to the same dimension, and then square integrals are performed to calculate the high-frequency disturbance energy value and the low-frequency disturbance energy value. By comparing the energy values ​​of high-frequency disturbances with those of low-frequency disturbances, if the energy value of high-frequency disturbances is greater than that of low-frequency disturbances, a weight adjustment command is triggered to increase the weight values ​​of the diagonal elements in the state weighting matrix corresponding to the attitude angle and angular velocity. Conversely, if the energy value of high-frequency disturbances is greater than that of low-frequency disturbances, the weight values ​​of the diagonal elements in the state weighting matrix corresponding to the position and velocity are increased.

Images