Parameter correction method and system of fixed-wing unmanned aerial vehicle semi-physical simulation system

Abstract

The invention belongs to the technical field of unmanned aerial vehicle simulation, and particularly relates to a parameter correction method and system for a fixed-wing unmanned aerial vehicle semi-physical simulation system, and the method comprises the steps: calculating the state information of an aircraft through a kinematics model and a dynamics model, and transmitting the state information to an unmanned aerial vehicle guidance control system, the unmanned aerial vehicle guidance control system outputs a rudder instruction according to the received aircraft state information to control the aircraft; a deep learning network model is constructed, the input of the deep learning network model is the state information of the aircraft, and the output of the deep learning network model is the correction of the aerodynamic coefficient; and updating the kinematics model and the dynamics model by adopting the corrected aerodynamic coefficient, recalculating the state information of the aircraft, and transmitting the state information to the unmanned aerial vehicle guidance control system to control the aircraft. According to the method, the modeling precision of the 6DOF model is improved, so that a semi-physical simulation system can better simulate a real flight condition.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) simulation, and particularly to a method and system for parameter correction of a fixed-wing UAV hardware-in-the-loop simulation system. Background Art

[0002] With the rapid development of aerospace technology, the design and research of aircraft have become increasingly complex. During the development process of an aircraft, it is necessary to fully verify and optimize the performance of the aircraft and the stability and reliability of the control system before actual flight tests. However, directly conducting full-scale flight tests is costly, risky, and difficult to modify and adjust once problems occur. Traditional pure digital simulation methods have certain limitations in verifying the performance of aircraft and control systems and cannot fully and realistically reflect various situations in actual flight. To more effectively design, develop, and test aircraft, the hardware-in-the-loop simulation technology has emerged.

[0003] The hardware-in-the-loop simulation system mainly includes a simulator (real-time simulation system), a guidance and control system, and a ground operation station. The real-time simulation system includes hardware and software. The hardware is an industrial control computer, and the software includes a real-time operating system, a UAV kinematics and dynamics model, and a serial communication module. Traditional hardware-in-the-loop simulation systems only use aerodynamic data for kinematics and dynamics modeling and do not use actual flight data to correct the model.

[0004] Since the data used for kinematics and dynamics modeling of UAVs comes from software calculations or wind tunnel experiments, there is a gap between the data and the actual force conditions during the UAV's movement process. The forces acting on the UAV during flight change with the UAV's motion state (speed, altitude, etc.) and environmental temperature changes, which is a continuous state. However, during calculations or wind tunnel experiments, only some discrete points can be collected, resulting in deviations. Therefore, the established kinematics and dynamics model (6DOF model) deviates from the actual situation, affecting the accuracy of the UAV hardware-in-the-loop simulation system. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a method and system for parameter correction of a fixed-wing UAV hardware-in-the-loop simulation system in view of the deficiencies of the prior art, improve the modeling accuracy of the 6DOF model, and thus enable the hardware-in-the-loop simulation system to better simulate the actual flight situation.

[0006] Technical Solution: The method for parameter correction of the fixed-wing UAV hardware-in-the-loop simulation system according to the present invention includes the following steps: S1: Build a hardware-in-the-loop simulation system, including a real-time simulation system, an unmanned aerial vehicle (UAV) guidance and control system, a UAV ground operation station, and a serial communication module. The real-time simulation system conducts data interaction with the UAV guidance and control system through the serial communication module, and the UAV guidance and control system conducts data interaction with the UAV ground operation station through a data link. The real-time simulation system includes a real-time operating system and real-time simulation software. The real-time operating system is installed on a hardware platform, and the real-time simulation software includes a hardware driver module and an aircraft simulation model. The aircraft simulation model includes an aerodynamic model, a dynamics model, a kinematics model, and a sensor model. The dynamics model and the kinematics model are used to calculate the state information of the aircraft, and after packing the state information according to the data protocol of real sensors, send it to the UAV guidance and control system through the serial communication module. The UAV guidance and control system outputs a rudder command according to the received aircraft state information to control the aircraft. S2: Construct a deep learning network model. The input of the deep learning network model is the state information of the aircraft, and the output of the deep learning network model is the correction amount of the aerodynamic coefficient. Use the gradient descent method to train the network model so that the output of the network model approaches the difference between the real aerodynamic coefficient and the original aerodynamic coefficient. S3: Update the kinematics model and the dynamics model with the corrected aerodynamic coefficient, and use the updated kinematics model and dynamics model to calculate the state information of the aircraft and transmit it to the UAV guidance and control system to control the aircraft.

[0007] Further improve the above technical solution. The dynamics model and the kinematics model calculate the state information of the aircraft through the centroid dynamics equation, the centroid kinematics equation, the rotational dynamics equation, the rotational kinematics equation, the force and moment equation acting on the aircraft, and the longitude and latitude calculation equation. The state information includes position (x, y, z), attitude (θ, ψ, φ), velocity (u, v, w), angular velocity (p, q, r), and acceleration (ax, ay, az).

[0008] Further, the deep learning network model adopts a fully connected neural network, including a first channel channel1 and a second channel channel2. Among them, the first channel channel1 consists of 1 input layer and 6 hidden layers, and the hidden layer is a 48-dimensional state quantity; the second channel channel2 consists of 1 input layer and 3 hidden layers, and the hidden layer is a 16-dimensional state quantity.

[0009] Further, the input of the deep learning network model is the flight state information of the drone, including position (x, y, z), attitude (θ, ψ, φ), velocity (u, v, w), angular velocity (p, q, r), and acceleration (ax, ay, az). Among them, θ, ψ, and φ are the pitch angle, yaw angle, and roll angle of the x, y, and z axes in the body coordinate system; u, v, and w represent the velocity components of the x, y, and z axes in the body coordinate system, p, q, and r are the angular velocity components of the x, y, and z axes in the body coordinate system, and ax, ay, and az are the acceleration components of the x, y, and z axes in the body coordinate system. The output of the deep learning network model is the correction amount of the aerodynamic coefficients, and the aerodynamic parameters include lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient, and roll moment coefficient.

[0010] Further, for different aerodynamic coefficients, the vector dimensions of the input of the deep learning network model are different: When correcting the lift coefficient, the input of the first channel is 12-dimensional state information, which are x, y, θ, ψ, φ, u, v, p, q, r, ax, and ay respectively, and the input of the second channel is 3-dimensional state information, which are z, w, and az respectively; When correcting the drag coefficient, the input of the first channel is 12-dimensional state information, which are x, z, θ, ψ, φ, u, w, p, q, r, ax, and az respectively, and the input of the second channel is 3-dimensional state information, which are y, v, and ay respectively. When correcting the side force coefficient, the input of the first channel is 12-dimensional state information, which are y, z, θ, ψ, φ, v, w, p, q, r, ay, and az respectively, and the input of the second channel is 3-dimensional state information, which are x, u, and ax respectively; When correcting the pitch moment coefficient, the input of the first channel is 13-dimensional state information, which are x, y, z, ψ, φ, u, v, w, p, r, ax, ay, and az respectively, and the input of the second channel is 2-dimensional state information, which are θ and q respectively; When correcting the roll moment coefficient and yaw moment coefficient, the input of the first channel is 11-dimensional state information, which are x, y, z, θ, u, v, w, q, ax, ay, and az respectively, and the input of the second channel is 4-dimensional state information, which are φ, ψ, p, and r respectively.

[0011] Further, the corrected aerodynamic coefficient consists of the following two parts. The first part is the theoretical calculation value or wind tunnel experiment value; the second part is the output value of the deep learning network model.

[0012] Further, the deep learning network model uses the state quantity of the previous moment of the aircraft as the input, and outputs the updated correction coefficient after iterative calculation; the kinematic and dynamic models calculated using the updated correction coefficient are in a dynamic update process.

[0013] The UAV hardware-in-the-loop simulation system for implementing the above method includes: A real-time simulation system, which includes a real-time operating system and real-time simulation software. The real-time operating system is installed on a hardware platform, and the real-time simulation software includes a hardware driver module and an aircraft simulation model. The aircraft simulation model includes an aerodynamic model, a dynamics model, a kinematics model, and a sensor model. The aerodynamic data model is used to provide basic aerodynamic parameters. The sensor model is used to generate data that conforms to the output format of a real sensor. The dynamics model and the kinematics model are used to calculate the state information of the aircraft. A deep learning network model. The input of the deep learning network model is the state information of the aircraft, including position (x, y, z), attitude (θ, ψ, φ), velocity (u, v, w), angular velocity (p, q, r), and acceleration (ax, ay, az). The output of the deep learning network model is the correction amount of the aerodynamic coefficients. The aerodynamic parameters include lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient, and roll moment coefficient. A UAV guidance and control system, which is used to obtain the state information of the aircraft calculated by the dynamics model and the kinematics model using the corrected aerodynamic coefficients for aircraft control.

[0014] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows: The present invention proposes a method for correcting aerodynamic parameters based on deep learning, and designs a method for processing flight data and training a network using flight data to correct the lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient, and roll moment coefficient of the UAV respectively, improving the modeling accuracy of the kinematics and dynamics models, and further enabling the hardware-in-the-loop simulation system to better simulate real flight conditions, helping to design and optimize the control system of the aircraft, and being able to quickly test and evaluate different design schemes and control strategies. Description of the Drawings

[0015] Figure 1 is a schematic structural diagram of the hardware-in-the-loop simulation system in the present invention; Figure 2 is a network schematic diagram of the deep learning model in the present invention. Detailed Embodiments

[0016] The technical solution of the present invention will be described in detail below with reference to the drawings, but the protection scope of the present invention is not limited to the described embodiments.

[0017] Embodiment 1: The present invention proposes a parameter correction method based on deep learning. By using a deep learning network and data from actual flight records, the lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient, and roll moment coefficient of the unmanned aerial vehicle are corrected respectively, improving the modeling accuracy of the 6DOF model, and further enabling the hardware-in-the-loop simulation system to better simulate the real flight situation.

[0018] I. Building a hardware-in-the-loop simulation system As Figure 1 shown, the hardware-in-the-loop simulation system mainly includes a real-time simulation system, an unmanned aerial vehicle guidance and control system, and an unmanned aerial vehicle ground operation station. Its system composition includes: a real-time simulation system, an unmanned aerial vehicle guidance and control system, an unmanned aerial vehicle ground operation station, and a serial communication module. Among them, the real-time simulation system includes hardware and software. The hardware is an industrial control computer, and the software is a real-time operating system and real-time simulation software. The real-time simulation software includes an aerodynamic model, a dynamics model, a kinematics model, and a sensor model. In each module, the accuracy of the establishment of the unmanned aerial vehicle dynamics and kinematics models directly determines the accuracy of the hardware-in-the-loop simulation system.

[0019] First, establish the kinematics model and dynamics model: (1) Aircraft center-of-mass dynamics equation Use the following formula to calculate the projection of the aircraft speed in the body axis system , , .

[0020] Among them, is the projection of the aircraft speed in the body coordinate system; is the projection of the aircraft angular velocity in the body coordinate system; is the projection of the resultant external force acting on the aircraft in the velocity coordinate system; is the mass of the aircraft.

[0021] (2) Aircraft center-of-mass kinematics equation According to the transformation matrix from the body coordinate system to the ground coordinate system, calculate the speed of the aircraft in the ground coordinate system , , and the position x, y, z: Among them, is the roll angle, pitch angle, and yaw angle representing the axis in the body coordinate system.

[0022] (3) Aircraft rotational dynamics equation Among them, are the moments of inertia about the axes of the aircraft body coordinate system; is the projection of the resultant external moment of the aircraft on the body coordinate system.

[0023] (4) Aircraft rotational kinematic equations To describe the flight attitude of the aircraft in the atmosphere, the equations for the rotation of the aircraft as a rigid body about its center of mass need to be established in the ground coordinate system The three angles describing its attitude are the pitch angle , the yaw angle , and the roll angle . The angular velocities are the pitch angular velocity , the yaw angular velocity , and the roll angular velocity .

[0024] (5) Force and moment equations acting on the aircraft The forces acting on the aircraft body are divided into aerodynamic forces, thrust, and gravity: The projections of the aerodynamic forces onto the three axes of the aircraft body coordinate system: Among them, is the projection of the aerodynamic force in the aircraft body coordinate system; are the drag, lift, and side force acting on the aircraft respectively; are the angle of attack and sideslip angle of the aircraft.

[0025] The projections of the gravity onto the three axes of the aircraft body coordinate system: Among them, is the projection of the gravity in the aircraft body coordinate system; is the gravity of the aircraft.

[0026] Define three aerodynamic forces in the velocity coordinate system, and the calculation formulas are: Among them, is the dynamic pressure; is the wing area; are the three aerodynamic force coefficients.

[0027] Define three aerodynamic moments in the body coordinate system, and the calculation formulas are Among them, is the mean aerodynamic chord; is the wingspan; are the three aerodynamic moment coefficients.

[0028] The aerodynamic forces and moments acting on the aircraft are calculated according to the above formulas, and the force and moment coefficients are obtained by two-dimensional or three-dimensional interpolation of the original aerodynamic data.

[0029] The aerodynamic force coefficient is composed of the aerodynamic force coefficient generated by the wing and the aerodynamic force coefficient of the control surface: In the formula, the subscript represents the wing, represents the elevator, represents the aileron, represents the rudder.

[0030] The aerodynamic moment coefficient is composed of the moment generated by the wing, the moment coefficient generated by the control surface and the damping moment coefficient: (6) Longitude and latitude calculation equation Since the Earth is an oblate spheroid of revolution about two axes, the distance from any point on the Earth's surface to the center of the Earth is related to the geocentric latitude of that point.

[0031] The formula for calculating the geocentric latitude is as follows: The formula for calculating the longitude is as follows: The real-time altitude on the flight path is . Among them, is the distance from any point on the flight path to the center of the Earth, is the geocentric radius vector of the real-time flight path point.

[0032] (7) Calculation equations for the aircraft speed, angle of attack and sideslip angle The formula for calculating the aircraft speed is: The formulas for calculating the aircraft angle of attack and sideslip angle are: (8) The state information of the aircraft such as position, attitude, velocity, acceleration, and acceleration calculated according to the equations provided in (1) to (7) is packed according to the data protocol of the real sensors used in the flight control system, and finally the packed data is sent to the unmanned aircraft guidance and control system through the serial port, and the rudder command sent by the unmanned aircraft guidance and control system is received through the serial port, and the force and moment equations acting on the aircraft in (5) are input for calculation to complete the closed-loop control.

[0033] II. Optimization and correction of aerodynamic parameters Whether the aerodynamic parameters obtained by CFD calculation or the aerodynamic parameters obtained by wind tunnel experiment, there are certain errors compared with the real values, which will cause the established kinematic and dynamic models to be unable to accurately simulate the real flight situation of the aircraft. Therefore, it is necessary to use the actual flight data to optimize and correct the aerodynamic parameters. Since the correction methods of the lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient, and roll moment coefficient are the same, the correction of the lift coefficient is selected for detailed description, and the correction methods of other coefficients are the same.

[0034] (1) Lift coefficient calculation: Combine equations (1), (2), (3), (4), (5), (6), (7), (8), (9), and combine the state data (position, attitude, velocity, angular velocity, acceleration) of the unmanned aircraft to calculate the real lift coefficient .

[0035] (2) Correct the aerodynamic parameters in equation (10): Transform the above formula into: Among them, is the real lift coefficient, is the original aerodynamic data, obtained from equation (10). Since the real aerodynamic parameters are continuous variables that change with the state of the unmanned aircraft (position, velocity, etc.), and the output of the deep learning network also changes according to the change of the input state quantity of the network, the present invention proposes to use the deep learning network model DPnet to identify the real-time correction amount of the aerodynamic parameters of the aircraft. Since the original aerodynamic data obtained by CFD calculation or wind tunnel experiment generally does not differ greatly from the real value, that is, it is generally near the optimal value, so there is no need to worry about the problem of the network falling into a local optimal value during the training process. Therefore, a fully connected deep network is selected, and its network schematic diagram is as Figure 2As shown, it consists of a network with two channels: channel1 and channel2. Among them, channel1 consists of 1 input layer and 6 hidden layers, and the hidden layer has a state quantity of 48; channel2 consists of 1 input layer and 3 hidden layers, and the hidden layer has a state quantity of 16 dimensions.

[0036] The input vector X of the entire network is a state quantity of 15 dimensions: position (x, y, z), attitude (θ, ψ, φ), velocity (u, v, w), angular velocity (p, q, r), acceleration (ax, ay, az). Assume that the coordinate system uses north-east-up, that is, x is to the right, y is forward, and z is upward. Then the state quantity in the z-axis direction is closely related to the lift, so there is no need to use too many abstraction layers (hidden layers) between the input and output. To accelerate the training of the network, channel2 is used. The connection between other state quantities and the lift is an indirect influence with a low degree of correlation, so channel1 with more abstraction layers is used. Therefore, when training the lift coefficient deviation, the input vector x1 of channel1 is 12 dimensions (x, y, θ, ψ, φ, u, v, p, q, r, ax, ay), and the input vector x2 of channel2 is 3 dimensions (z, w, az). The output layer is 1 dimension , and the target value is , the activation function uses Relu, and the loss function is .

[0037] Using the gradient descent method and using flight data, the network parameters are trained to obtain the trained parameter-corrected deep network model DPnet.

[0038] Furthermore, the corrected lift coefficient is obtained: (3) Using the same method, the corrected drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient, and roll moment coefficient are obtained respectively.

[0039] The overall structure of the deep learning network model remains unchanged, but the input vector needs to be adjusted accordingly based on the degree of association between the state variables and the coefficients to be corrected. When training the correction amount of the drag coefficient, the input vector x1 of the first channel of the deep network is 12-dimensional (x, z, θ, ψ, φ, u, w, p, q, r, ax, az), and the input vector x2 of the second channel is 3-dimensional (y, v, ay). When training the correction amount of the side force coefficient, the input vector x1 of the first channel of the deep network is 12-dimensional (y, z, θ, ψ, φ, v, w, p, q, r, ay, az), and the input vector x2 of the second channel is 3-dimensional (x, u, ax). When training the correction amount of the pitching moment coefficient, the input vector x1 of the first channel of the deep network is 13-dimensional (x, y, z, ψ, φ, u, v, w, p, r, ax, ay, az), and the input vector x2 of the second channel is 2-dimensional (θ, q). When training the correction amount of the rolling moment coefficient, the input vector x1 of the first channel of the deep network is 11-dimensional (x, y, z, θ, u, v, w, q, ax, ay, az), and the input vector x2 of the second channel is 4-dimensional (φ, ψ, p, r). When training the correction amount of the yawing moment coefficient, the input vector x1 of the first channel of the deep network is 11-dimensional (x, y, z, θ, u, v, w, q, ax, ay, az), and the input vector x2 of the second channel is 4-dimensional (φ, ψ, p, r).

[0040] Replace the lift coefficient, drag coefficient, side force coefficient, pitching moment coefficient, yawing moment coefficient, and rolling moment coefficient in Equations (1) and (3) with the corrected coefficients, such as the lift coefficient in the corrected Equation (19). Re-establish the kinematic and dynamic models according to steps (1)-(8) in building the hardware-in-the-loop simulation system, and replace the models in the original hardware-in-the-loop simulation system, so as to optimize the simulation accuracy of the hardware-in-the-loop simulation system.

[0041] The formulas of the corrected kinematic and dynamic models, and the aerodynamic parameters (such as the lift coefficient ) consist of two parts. One part is the theoretical calculation value or the wind tunnel experiment value, and the other part is the estimated value DPnet of the deep network. The value of DPnet can be regarded as a function of the real-time state data (position, attitude, etc.) of the aircraft. In the calculation process of this real-time state data, the data of the previous step is used, and the calculation process is continuously iterated. Update Equations (5) and (8) again, and the updated kinematic and dynamic models.

[0042] According to the characteristics of flight data and aerodynamic parameters, namely, the aerodynamic parameters change in real time according to the aircraft state and the initial value of the aerodynamic parameter calculation is near the optimal value, and the network training will not fall into local extrema, the present invention proposes a parameter correction method based on deep learning, designs a deep learning network suitable for aerodynamic parameter correction, uses the deep learning network to identify the real-time correction amount of the aircraft aerodynamic parameters, and designs a flight data processing method and a method of using flight data for network training, respectively corrects the lift coefficient, drag coefficient, side force coefficient, pitching moment coefficient, yaw moment coefficient, and rolling moment coefficient of the aircraft, improves the modeling accuracy of the kinematic and dynamic models, and further enables the hardware-in-the-loop simulation system to better simulate the real flight situation, helps to design and optimize the control system of the aircraft, and can quickly test and evaluate different design schemes and control strategies.

[0043] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as a limitation of the present invention itself. Various changes can be made in its form and details without departing from the spirit and scope of the present invention defined by the appended claims.

Claims

1. A parameter correction method for a fixed-wing UAV semi-physical simulation system, characterized in that: The steps include: S1: Building a semi-physical simulation system, including a real-time simulation system, a UAV guidance and control system, a UAV ground operation station, and a serial communication module. The real-time simulation system exchanges data with the UAV guidance and control system through the serial communication module, and the UAV guidance and control system exchanges data with the UAV ground operation station through a data link; the real-time simulation system includes a real-time operating system and real-time simulation software, the real-time operating system is mounted on a hardware platform, and the real-time simulation software includes a hardware driver module and an aircraft simulation model; the aircraft simulation model includes an aerodynamic model, a dynamic model, a kinematic model, and a sensor model, the dynamic model and the kinematic model are used to calculate the state information of the aircraft, and the state information is packaged according to the data protocol of the real sensor and sent to the UAV guidance and control system through the serial communication module, and the UAV guidance and control system outputs a rudder command according to the received aircraft state information to control the aircraft; S2: constructing a deep learning network model, wherein the input of the deep learning network model is the state information of the aircraft, and the output of the deep learning network model is the correction amount of the aerodynamic coefficient. The network model is trained by the gradient descent method so that the output of the network model is close to the difference between the real aerodynamic coefficient and the original aerodynamic coefficient; S3: Using the corrected aerodynamic coefficients to update the kinematic model and the dynamic model, using the updated kinematic model and the dynamic model to calculate the state information of the aircraft and transmit it to the UAV guidance and control system to control the aircraft.

2. The parameter correction method of the fixed-wing UAV semi-physical simulation system according to claim 1 is characterized in that: The dynamic model and kinematic model calculate the state information of the aircraft through the center of mass dynamic equation, the center of mass kinematic equation, the rotational dynamic equation, the rotational kinematic equation, the force and torque equation acting on the aircraft, and the longitude and latitude solution equation. The state information includes position (x, y, z), attitude (θ, ψ, φ), speed (u, v, w), angular velocity (p, q, r), and acceleration (ax, ay, az).

3. The parameter correction method of the fixed-wing UAV semi-physical simulation system according to claim 2 is characterized in that: The deep learning network model adopts a fully connected neural network, including a first channel channel1 and a second channel channel2, wherein the first channel channel1 consists of 1 input layer and 6 hidden layers, and the hidden layer is a 48-dimensional state quantity; the second channel channel2 consists of 1 input layer and 3 hidden layers, and the hidden layer is a 16-dimensional state quantity.

4. The parameter correction method of the fixed-wing UAV semi-physical simulation system according to claim 3 is characterized in that: The input of the deep learning network model is the flight status information of the UAV, including position (x, y, z), attitude (θ, ψ, φ), speed (u, v, w), angular velocity (p, q, r), acceleration (ax, ay, az), wherein θ, ψ, φ are the pitch angle, yaw angle and roll angle of the x, y, z axes in the body coordinate system; u, v, w represent the velocity components of the x, y, z axes in the body coordinate system, p, q, r are the angular velocity components of the x, y, z axes in the body coordinate system, and ax, ay, az are the acceleration components of the x, y, z axes in the body coordinate system; the output of the deep learning network model is the correction amount of the aerodynamic coefficient, and the aerodynamic parameters include lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient and roll moment coefficient.

5. The parameter correction method of the fixed-wing UAV semi-physical simulation system according to claim 4 is characterized in that: For different aerodynamic coefficients, the vector dimensions of the deep learning network model input are different: When correcting the lift coefficient, the first channel input is 12-dimensional state information, which are x, y, θ, ψ, φ, u, v, p, q, r, ax, ay, and the second channel input is 3-dimensional state information, which are z, w, az; When correcting the drag coefficient, the first channel input is 12-dimensional state information, which are x, z, θ, ψ, φ, u, w, p, q, r, ax, az, and the second channel input is 3-dimensional state information, which are y, v, ay, When correcting the lateral force coefficient, the first channel input is 12-dimensional state information, which are y, z, θ, ψ, , v, w, p, q, r, ay, az, the second channel input is 3D state information, which are x, u, ax respectively; When correcting the pitch moment coefficient, the first channel input is 13-dimensional state information, which are x, y, z, ψ, , u, v, w, p, r, ax, ay, az, the second channel input is 2D state information, which are θ and q respectively; When correcting the rolling moment coefficient and the yaw moment coefficient, the first channel input is 11-dimensional state information, which are x, y, z, θ, u, v, w, q, ax, ay, az, and the second channel input is 4-dimensional state information, which are ,ψ,p,r.

6. The parameter correction method of the fixed-wing UAV semi-physical simulation system according to claim 1 is characterized in that: The corrected aerodynamic coefficient consists of the following two parts: the first part is the theoretical calculation value or wind tunnel experimental value; the second part is the output value of the deep learning network model.

7. The parameter correction method of the fixed-wing UAV semi-physical simulation system according to claim 1 or 6, characterized in that: The deep learning network model uses the state of the aircraft at the previous moment as input, and outputs an updated correction coefficient after iterative calculation; the kinematic and dynamic models calculated using the updated correction coefficient are in a dynamic updating process.

8. A UAV semi-physical simulation system for implementing the method described in claim 1, characterized in that: The system comprises: A real-time simulation system, wherein the real-time simulation system comprises a real-time operating system and real-time simulation software, wherein the real-time operating system is mounted on a hardware platform, and the real-time simulation software comprises a hardware driver module and an aircraft simulation model; the aircraft simulation model comprises an aerodynamic model, a dynamic model, a kinematic model, and a sensor model, wherein the aerodynamic data model is used to provide basic aerodynamic parameters, the sensor model is used to generate data in a format that conforms to a real sensor output, and the dynamic model and the kinematic model are used to calculate aircraft status information; A deep learning network model, wherein the input of the deep learning network model is the state information of the aircraft, including position (x, y, z), attitude (θ, ψ, φ), speed (u, v, w), angular velocity (p, q, r), acceleration (ax, ay, az), and the output of the deep learning network model is the correction of the aerodynamic coefficients, and the aerodynamic parameters include lift coefficient, drag coefficient, side force coefficient, pitch moment coefficient, yaw moment coefficient and roll moment coefficient; The unmanned aerial vehicle guidance and control system is used to obtain the dynamic model and kinematic model and use the corrected aerodynamic coefficients to calculate the state information of the aircraft to control the aircraft.

Images