Distributed propulsion aircraft power yaw control method based on deep reinforcement learning

Abstract

The invention belongs to the technical field of flight control, and discloses a distributed propulsion aircraft power yaw control method based on deep reinforcement learning. The power yaw control method comprises the steps that the distributed propellers are installed; establishing a six-degree-of-freedom nonlinear dynamic model; pneumatic data and pressure distribution under the differential power are obtained; establishing a distributed propulsion aircraft power yaw control framework; and establishing a distributed propulsion aircraft power yaw control law. According to the power yaw control method, an expected index is introduced into a reward function of an algorithm, a power yaw coupling control model is obtained in an offline training and online use mode, the flight-thrust-control coupling control problem of the distributed propulsion aircraft is solved, and the power yaw control method has an important significance in guaranteeing the aircraft attitude stability during power yaw of the distributed propulsion aircraft. And the method is of great significance in improving course stability and landing safety.

Description

Technical Field

[0001] The present invention belongs to the technical field of flight control, and particularly relates to a dynamic yaw control method for a distributed propulsion aircraft based on deep reinforcement learning. Background Art

[0002] A distributed propulsion aircraft realizes flight control and power output of the aircraft by arranging a plurality of thrusters on the symmetric plane of the aircraft. Each thruster has a certain degree of independence and autonomy, and the attitude control and heading control of the aircraft can be achieved by adjusting the magnitude of the thrust of each thruster and the deflection angle of the rudder surface. Compared with a conventional layout aircraft, a distributed propulsion aircraft has significant advantages in coping with crosswinds and gusts. Especially when landing under strong crosswind conditions, the yaw moment provided only by the rudder is not sufficient to correct the heading. The distributed propulsion aircraft can enhance the heading stability of the aircraft through power differential and improve the landing safety.

[0003] The existing dynamic yaw control technologies for distributed propulsion aircraft mainly utilize modern control theory to design dynamic yaw control laws for aircraft. For example, the invention names disclosed in the Chinese patent literature database are a distributed electric propulsion aircraft and its control method (CN114476093A), and a cooperative control method applicable to multiple thrusters of a distributed electric propulsion aircraft (CN114137839A), as well as the dynamic yaw nonlinear dynamic inversion control of a distributed electric propulsion aircraft published in the Journal of Aerospace Power (https: / / doi.org / 10.13224 / j.cnki.jasp.20220222) [J / OL] (You Shun, Kou Peng, Yao Xuanyu, Wang Jing, Liang Deliang, Liang Zhe).

[0004] The above methods can calculate the required power of each thruster according to the heading command and altitude command, study the thrust power distribution strategy, and achieve stable heading flight in the case of gusts. However, dynamic yaw will change the flow field on the upper wing surfaces of both wings. The air flow velocity on the upper wing surface of the wing with a slower engine speed is relatively slow, and the air flow velocity on the upper wing surface of the wing with a faster engine speed is relatively fast. The side with a faster air flow velocity provides a greater lift, which will generate an additional rolling moment, resulting in an unexpected roll angle of the aircraft. Especially during the landing phase in the case of crosswind, the accompanying roll angle may cause serious flight accidents such as wing touchdown and landing rollover. In addition, when performing flight missions such as transporting personnel and surveying, the accompanying roll of the fuselage will affect the riding experience and surveying accuracy. At this time, it is necessary to rely on aileron differential to compensate for the rolling moment generated by dynamic yaw. Under different flight speeds, angles of attack, tail fin deflection angles, and dynamic differential modes, the required aileron differential compensation deflection angles are also different, and there are complex coupling relationships hidden therein. Relying on traditional iterative fitting and other methods will consume huge amounts of manpower and financial resources, and there are also problems such as low grid accuracy.

[0005] At present, there is an urgent need to develop a dynamic yaw control method for distributed propulsion aircraft based on deep reinforcement learning. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a dynamic yaw control method for distributed propulsion aircraft based on deep reinforcement learning.

[0007] The dynamic yaw control method for distributed propulsion aircraft based on deep reinforcement learning of the present invention includes the following steps:

[0008] S10. Install distributed thrusters;

[0009] The fuselage of the distributed propulsion aircraft is provided with symmetric left and right wings, ailerons and a V-shaped tail; distributed thrusters are installed on the upper wing surfaces of both wings, and the distributed thrusters include a number of thrusters symmetrically distributed left and right;

[0010] S20. Establish a six-degree-of-freedom nonlinear dynamics model;

[0011] Establish a six-degree-of-freedom nonlinear dynamics model according to the physical parameter relationship of the distributed thrusters and the six-component aerodynamic data obtained from wind tunnel tests;

[0012] S30. Obtain aerodynamic data and pressure distribution under differential power;

[0013] According to the differential mode of the thrusters on both wings of the pre-set aircraft, conduct wind tunnel tests or numerical simulations to obtain the aerodynamic data and pressure distribution under the differential mode, and obtain the roll moment curve and yaw moment curve under thruster differential control;

[0014] S40. Establish a dynamic yaw control framework for distributed propulsion aircraft;

[0015] Build a dynamic differential control neural network, combine the basic control law, and obtain a dynamic yaw control framework for distributed propulsion aircraft based on deep reinforcement learning;

[0016] S50. Establish a dynamic yaw control law for distributed propulsion aircraft;

[0017] Establish a reward function, provide training parameters, and use python to train the neural network to obtain a dynamic yaw control law for distributed propulsion aircraft based on deep reinforcement learning.

[0018] Furthermore, the thruster is a ducted fan.

[0019] Furthermore, the physical parameter relationship of the distributed thrusters is as follows:

[0020]

[0021] Among them, V is the flight speed, β is the sideslip angle, α is the angle of attack, p is the roll rate, q is the pitch rate, r is the yaw rate, φ is the roll angle, θ is the pitch angle, and ψ is the yaw angle; is the rate of change of speed, is the sideslip angular velocity, is the rate of change of the angle of attack, is the rate of change of the roll rate, is the rate of change of the pitch rate, is the rate of change of the yaw rate, is the rate of change of the roll angle, is the rate of change of the yaw angle; m is the mass, Q is the dynamic pressure, S is the wing area, b is the wingspan, c is the wing chord, and g is the acceleration due to gravity; I x is the moment of inertia about the x-axis, I y is the moment of inertia about the y-axis, I z is the moment of inertia about the z-axis, I xz is the product of inertia; The six-component aerodynamic data are the lift coefficient C L , the residual thrust coefficient C R , the residual thrust coefficient C R is the difference between the drag coefficient and the thrust coefficient, the lateral force coefficient C Y , the pitch moment coefficient C m , the roll moment coefficient C l and the yaw moment coefficient C n .

[0022] Furthermore, the six-degree-of-freedom nonlinear dynamic model is as follows:

[0023]

[0024] Among them, N T,i represents the horizontal distance between the i-th thruster and the longitudinal axis of the aircraft, represents the residual thrust coefficient of the i-th thruster; δ a is the aileron deflection angle, δ e is the elevator deflection angle, δ r is the rudder deflection angle; is the lift coefficient at zero angle of attack, is the rate of change of the lift coefficient with respect to the angle of attack, is the rate of change of the lift coefficient with respect to the elevator deflection angle; is the residual thrust coefficient at zero angle of attack, is the rate of change of the residual thrust coefficient with respect to the angle of attack, is the rate of change of the residual thrust coefficient with respect to the elevator deflection angle, is the rate of change of the residual thrust coefficient with respect to the square of the elevator deflection angle; is the rate of change of the lateral force coefficient with respect to the sideslip angle, is the rate of change of the lateral force coefficient with respect to the aileron deflection angle, is the rate of change of the lateral force coefficient with respect to the rudder deflection angle; is the rate of change of the rolling moment coefficient with respect to the roll angular velocity, is the rate of change of the rolling moment coefficient with respect to the yaw angular velocity, is the rate of change of the rolling moment coefficient with respect to the aileron deflection angle, is the rate of change of the rolling moment coefficient with respect to the rudder deflection angle, is the rate of change of the rolling moment coefficient with respect to the sideslip angle; V 0 is the initial velocity of the aircraft performing a pitch maneuver, is the rate of change of the pitch moment coefficient with respect to the velocity, is the rate of change of the pitch moment coefficient with respect to the angle of attack, is the rate of change of the pitch moment coefficient with respect to the pitch angular velocity, is the rate of change of the pitch moment coefficient with respect to the elevator deflection angle; is the rate of change of the yaw moment coefficient with respect to the sideslip angle, is the rate of change of the yaw moment coefficient with respect to the aileron deflection angle, is the rate of change of the yaw moment coefficient with respect to the rudder deflection angle, is the rate of change of the yaw moment coefficient with respect to the roll angular velocity, is the rate of change of the yaw moment coefficient with respect to the yaw angular velocity.

[0025] Furthermore, the input layer of the dynamic differential control neural network includes the angle of attack α, the roll angle φ, the sideslip angle β, the aileron differential deflection angle △δ a and the V-tail deflection angle δ V ; the neural network transfers each parameter of the input layer to the hidden layer, and after calculation, the output layer Q π (s, a 1 )、Q π (s, a 2 )……Q π (s, a n ) is obtained, and the maximum Q value is selected, that is, the obtained optimal differential mode A = a i is the selected dynamic differential control strategy;

[0026] where Q π is the Q value obtained by executing the strategy a in the state s, s is the state, a is the strategy, and the subscripts 1, 2, …… n represent different dynamic differential control strategies.

[0027] Furthermore, the basic control law includes:

[0028] Velocity control law:

[0029]

[0030] Among them, e V = V r -V is the speed tracking error, and V r is the flight speed reference signal; k V is the speed proportional control parameter, and k Vi is the speed integral control parameter; T is the ducted fan power, and T p is the trim power of the ducted fan under cruise conditions;

[0031] Pitch angle control law:

[0032] δ e = k θ e θ - k q q(4)

[0033] Among them, e θ = θ r -θ is the pitch angle tracking error, and θ r is the pitch angle reference signal; q is the pitch angle rate; k θ is the pitch angle control parameter, and k q is the pitch angle rate control parameter;

[0034] Roll angle control law:

[0035] δ a = k φ e φ - k p p(5)

[0036] Among them, e φ = φ r -φ is the roll angle tracking error, and φ r is the roll angle reference signal; p is the roll angle rate; k φ is the roll angle control parameter, and k p is the roll angle rate control parameter;

[0037] To avoid the roll angle φ from being too large, a roll angle protection system is designed:

[0038]

[0039] Among them, δ ac is the aileron deflection angle after amplitude limiting;

[0040] Yaw angle control law:

[0041]

[0042] Among them, e β = β r-β is the sideslip angle tracking error, β r is the sideslip angle reference signal; r is the yaw rate; k β is the sideslip angle proportional control parameter, k βi is the sideslip angle integral control parameter, k r is the yaw angular velocity control parameter.

[0043] Furthermore, the reward function is:

[0044] R = -|e θ |-|e φ | (8)

[0045] where R is the reward value.

[0046] Furthermore, the dynamic yaw control law of the distributed propulsion aircraft: First, the basic control law calculates the ducted fan power T, elevator deflection angle δ e , aileron deflection angle δ a , rudder deflection angle δ r , and then the dynamic differential control neural network calculates the optimal power distribution strategy according to the input layer parameters.

[0047] Furthermore, the training parameters include the attenuation coefficient, learning law, number of neural network layers and nodes, experience pool size, random sampling sample size, discount factor, time step, and target network update speed.

[0048] The dynamic yaw control method of the distributed propulsion aircraft based on deep reinforcement learning of the present invention has the following characteristics:

[0049] (1) It can calculate the power distribution strategy of the distributed thrusters according to the wind field information and stabilize the flight heading;

[0050] (2) It can give the aileron differential strategy according to the power distribution strategy of the distributed thrusters, flight speed, angle of attack and other information, compensate for the additional roll caused by the dynamic differential, and stabilize the attitude of the distributed thrusters;

[0051] (3) It adopts the off-line training and on-line use method, completes the work that takes a long time and requires a large amount of computing power in the ground preparation work, directly writes the trained policy network into the flight control board, and the existing commercially available flight control board can meet the computing power required during the flight of the distributed thrusters, which has engineering practical value.

[0052] The dynamic yaw control method for a distributed propulsion aircraft based on deep reinforcement learning of the present invention introduces an expected index into the reward function of the algorithm, and obtains a dynamic yaw coupling control model through offline training and online use, solving the problem of flight-thrust-control coupling control of a distributed propulsion aircraft, and is of great significance for ensuring the aircraft attitude stability during the dynamic yaw of a distributed propulsion aircraft, improving the course stability and landing safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a schematic structural diagram of a distributed power aircraft;

[0054] Figure 2a is the influence curve of the power differential of the distributed power aircraft on the yaw moment;

[0055] Figure 2b is the influence curve of the power differential of the distributed power aircraft on the roll moment;

[0056] Figure 3 is the neural network diagram of power differential control;

[0057] Figure 4 is the flowchart of the dynamic yaw control method for a distributed propulsion aircraft based on deep reinforcement learning of the present invention.

[0058] In the figure, 1. Thruster. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] The present invention will be described in detail below with reference to the drawings and embodiments.

[0060] Embodiment: The distributed power aircraft of this embodiment is shown in Figure 1 , Figure 1 in which, the label 1 represents a thruster, and the label 2 represents a pressure sensor. The training parameters of this embodiment are shown in Table 1.

[0061] Table 1

[0062]

[0063] The dynamic yaw control method for a distributed propulsion aircraft based on deep reinforcement learning of this embodiment includes the following steps:

[0064] S10. Install distributed thrusters;

[0065] The fuselage of the distributed propulsion aircraft is provided with symmetric wings, ailerons and a V-shaped tail on the left and right; distributed thrusters are installed on the upper wing surfaces of both wings, and the distributed thrusters include a plurality of thrusters symmetrically distributed on the left and right;

[0066] S20. Establish a six-degree-of-freedom nonlinear dynamics model;

[0067] Establish a six - degree - of - freedom non - linear dynamic model based on the physical parameter relationships of distributed thrusters and six - component aerodynamic data obtained from wind tunnel tests;

[0068] S30. Obtain aerodynamic data and pressure distribution under differential power;

[0069] According to the differential mode of the thrusters on both sides of the preset aircraft, conduct wind tunnel tests or numerical simulations to obtain aerodynamic data and pressure distribution under the differential mode, and obtain the roll moment curve and yaw moment curve under the differential control of the thrusters as shown in Figure 2a 、 Figure 2b ;

[0070] S40. Establish a dynamic yaw control framework for distributed - propulsion aircraft;

[0071] Build a dynamic differential control neural network as shown in Figure 3 , and combine it with the basic control law to obtain a dynamic yaw control framework for distributed - propulsion aircraft based on deep reinforcement learning;

[0072] S50. Establish a dynamic yaw control law for distributed - propulsion aircraft;

[0073] Establish a reward function, provide training parameters, and use Python to train the neural network to obtain a dynamic yaw control law for distributed - propulsion aircraft based on deep reinforcement learning as shown in Figure 4 ;

[0074] Furthermore, the thruster is a ducted fan.

[0075] Furthermore, the physical parameter relationships of the distributed thrusters are as follows:

[0076]

[0077] where V is the flight speed, β is the sideslip angle, α is the angle of attack, p is the roll rate, q is the pitch rate, r is the yaw rate, φ is the roll angle, θ is the pitch angle, ψ is the yaw angle; is the rate of change of speed, is the sideslip angular velocity, is the rate of change of the angle of attack, is the rate of change of the roll rate, is the rate of change of the pitch rate, is the rate of change of the yaw rate, is the rate of change of the roll angle, is the rate of change of the yaw angle; m is the mass, Q is the dynamic pressure, S is the wing area, b is the wingspan, c is the wing chord, g is the acceleration due to gravity; I x is the moment of inertia about the x - axis, I yis the moment of inertia about the y-axis, I z is the moment of inertia about the z-axis, I xz is the product of inertia; the six-component aerodynamic data are the lift coefficient C L , the residual thrust coefficient C R , the residual thrust coefficient C R is the difference between the drag coefficient and the thrust coefficient, the lateral force coefficient C Y , the pitch moment coefficient C m , the roll moment coefficient C l and the yaw moment coefficient C n .

[0078] Furthermore, the six-degree-of-freedom nonlinear dynamic model is as follows:

[0079]

[0080] where, N T,i represents the horizontal distance between the i-th thruster and the longitudinal axis of the aircraft, represents the residual thrust coefficient of the i-th thruster; δ a is the aileron deflection angle, δ e is the elevator deflection angle, δ r is the rudder deflection angle; is the lift coefficient at zero angle of attack, is the rate of change of the lift coefficient with respect to the angle of attack, is the rate of change of the lift coefficient with respect to the elevator deflection angle; is the residual thrust coefficient at zero angle of attack, is the rate of change of the residual thrust coefficient with respect to the angle of attack, is the rate of change of the residual thrust coefficient with respect to the elevator deflection angle, is the rate of change of the residual thrust coefficient with respect to the square of the elevator deflection angle; is the rate of change of the lateral force coefficient with respect to the sideslip angle, is the rate of change of the lateral force coefficient with respect to the aileron deflection angle, is the rate of change of the lateral force coefficient with respect to the rudder deflection angle; is the rate of change of the roll moment coefficient with respect to the roll angular rate, is the rate of change of the roll moment coefficient with respect to the yaw angular rate, is the rate of change of the roll moment coefficient with respect to the aileron deflection angle, is the rate of change of the roll moment coefficient with respect to the rudder deflection angle, is the rate of change of the roll moment coefficient with respect to the sideslip angle; V 0 is the initial velocity of the aircraft during pitch maneuver, is the rate of change of the pitch moment coefficient with respect to the velocity, is the rate of change of the pitch moment coefficient with respect to the angle of attack, is the rate of change of the pitching moment coefficient with respect to the pitching angular velocity, is the rate of change of the pitching moment coefficient with respect to the elevator deflection angle; is the rate of change of the yaw moment coefficient with respect to the sideslip angle, is the rate of change of the yaw moment coefficient with respect to the aileron deflection angle, is the rate of change of the yaw moment coefficient with respect to the rudder deflection angle, is the rate of change of the yaw moment coefficient with respect to the roll angular velocity, is the rate of change of the yaw moment coefficient with respect to the yaw angular velocity.

[0081] Furthermore, the input layer of the dynamic differential control neural network includes the angle of attack α, the roll angle φ, the sideslip angle β, the aileron differential deflection angle △δ a and the V-tail deflection angle δ V ; the neural network transmits each parameter of the input layer to the hidden layer, and after calculation, the output layer Q π (s, a 1 )、Q π (s, a 2 )……Q π (s, a n ) is obtained, and the maximum Q value is selected, that is, the obtained optimal differential mode A = a i is the selected dynamic differential control strategy;

[0082] where Q π is the Q value obtained by executing the strategy a in the state s, s is the state, a is the strategy, and the subscripts 1, 2, …… n represent different dynamic differential control strategies.

[0083] Furthermore, the basic control law includes:

[0084] Speed control law:

[0085]

[0086] where e V = V r - V is the speed tracking error, V r is the flight speed reference signal; k V is the speed proportional control parameter, k Vi is the speed integral control parameter; T is the ducted fan power, T p is the trim power of the ducted fan under cruise conditions;

[0087] Pitch angle control law:

[0088] δ e = k θ e θ - k qq(4)

[0089] where e θ = θ r -θ is the pitch angle tracking error, and θ r is the pitch angle reference signal; q is the pitch rate; k θ is the pitch angle control parameter, and k q is the pitch rate control parameter;

[0090] Roll angle control law:

[0091] δ a = k φ e φ - k p p(5)

[0092] where e φ = φ r -φ is the roll angle tracking error, and φ r is the roll angle reference signal; p is the roll rate; k φ is the roll angle control parameter, and k p is the roll rate control parameter;

[0093] To avoid the roll angle φ from being too large, a roll angle protection system is designed:

[0094]

[0095] where δ ac is the aileron deflection angle after limiting;

[0096] Yaw angle control law:

[0097]

[0098] where e β = β r -β is the sideslip angle tracking error, and β r is the sideslip angle reference signal; r is the yaw rate; k β is the sideslip angle proportional control parameter, and k βi is the sideslip angle integral control parameter, and k r is the yaw angular velocity control parameter.

[0099] Furthermore, the reward function is:

[0100] R = -|e θ |-|e φ |(8)

[0101] where R is the reward value.

[0102] Further, the power yaw control law of the distributed propulsion aircraft: First, the basic control law calculates the ducted fan power T, elevator deflection angle δ e , aileron deflection angle δ a , rudder deflection angle δ r . Then, the dynamic differential control neural network calculates the optimal power distribution strategy according to the input layer parameters.

[0103] Further, the training parameters include the attenuation coefficient, learning law, number of neural network layers and nodes, experience pool size, random sampling sample size, discount factor, time step, and target network update speed.

[0104] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. For those familiar with the art, without departing from the principle of the present invention, all the features disclosed in the present invention, or all the steps in the disclosed methods or processes, except for mutually exclusive features and / or steps, can be combined in any way. The present invention is not limited to the specific details and the illustrated examples here.

Claims

1. A distributed propulsion aircraft powered yaw control method based on deep reinforcement learning, characterized in that: The power yaw control method comprises the following steps: S10. Install distributed thrusters; The fuselage of the distributed propulsion aircraft is provided with left-right symmetrical wings, ailerons and V-shaped tail; distributed propellers are installed on the upper wing surfaces of the wings on both sides, and the distributed propellers include a plurality of propellers distributed left-right symmetrically; S20. Establish a six-degree-of-freedom nonlinear dynamic model; A six-degree-of-freedom nonlinear dynamic model is established based on the physical parameter relationship of the distributed thruster and the six-component aerodynamic data obtained from the wind tunnel test; S30. Obtaining aerodynamic data and pressure distribution under differential power; According to the preset differential mode of the wing thrusters on both sides of the aircraft, a wind tunnel test or numerical simulation is performed to obtain aerodynamic data and pressure distribution under the differential mode, and obtain a rolling moment curve and a yaw moment curve under the thruster differential control; S40. Establish a framework for the dynamic yaw control of distributed propulsion aircraft; Build a power differential control neural network and combine it with the basic control law to obtain a distributed propulsion aircraft power yaw control framework based on deep reinforcement learning; S50. Establish the dynamic yaw control law of distributed propulsion aircraft; Establish a reward function, provide training parameters, use Python to train the neural network, and obtain the dynamic yaw control law of distributed propulsion aircraft based on deep reinforcement learning.

2. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 1, characterized in that: The propeller is a ducted fan.

3. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 2, characterized in that: The physical parameter relationship of the distributed thruster is shown in the following formula: Wherein, V is the flight speed, β is the sideslip angle, α is the angle of attack, p is the roll rate, q is the pitch rate, r is the yaw rate, φ is the roll angle, θ is the pitch angle, and ψ is the yaw angle; is the speed change rate, is the sideslip angular velocity, is the rate of change of angle of attack, is the rate of change of the roll angle, is the pitch angle rate of change, is the rate of change of yaw angle, is the rate of change of roll angle, is the rate of change of yaw angle; m is mass, Q is dynamic pressure, S is wing area, b is wing span, c is wing chord length, g is gravitational acceleration; I x is the moment of inertia about the x-axis, I y is the moment of inertia about the y-axis, I z is the moment of inertia about the z-axis, I xz is the inertia product; the six-component aerodynamic data is the lift coefficient C L , residual thrust coefficient C R , residual thrust coefficient C R is the difference between the drag coefficient and the thrust coefficient, and the lateral force coefficient C Y , pitch moment coefficient C m , rolling moment coefficient C l and the yaw moment coefficient C n .

4. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 3, characterized in that: The six-degree-of-freedom nonlinear dynamic model is shown in the following formula: Among them, N T,i represents the horizontal distance between the ith thruster and the longitudinal axis of the aircraft, represents the residual thrust coefficient of the i-th thruster; δ a is the aileron rudder deflection angle, δ e is the elevator deflection angle, δ r is the rudder deflection angle; is the lift coefficient at zero angle of attack, is the rate of change of lift coefficient with respect to angle of attack, is the rate of change of lift coefficient to elevator deflection angle; is the residual thrust coefficient at zero angle of attack, is the rate of change of the residual thrust coefficient with respect to the angle of attack, is the rate of change of the residual thrust coefficient to the elevator deflection angle, is the rate of change of the residual thrust coefficient to the square of the elevator deflection angle; is the rate of change of the lateral force coefficient to the sideslip angle, is the rate of change of the lateral force coefficient to the aileron rudder deflection angle, is the rate of change of the lateral force coefficient to the rudder deflection angle; is the rate of change of the rolling moment coefficient to the rolling angular rate, is the rate of change of the rolling moment coefficient to the yaw rate, is the rate of change of the rolling moment coefficient to the aileron rudder deflection angle, is the rate of change of the rolling moment coefficient to the rudder deflection angle, is the rate of change of the rolling moment coefficient to the sideslip angle; V0 is the initial velocity of the aircraft when doing pitch maneuvers, is the rate of change of the pitch moment coefficient with respect to velocity, is the rate of change of the pitch moment coefficient with respect to the angle of attack, is the rate of change of the pitch moment coefficient to the pitch angular velocity, is the rate of change of the pitch moment coefficient to the elevator deflection angle; is the rate of change of the yaw moment coefficient to the sideslip angle, is the rate of change of the yaw moment coefficient to the aileron rudder deflection angle, is the rate of change of the yaw moment coefficient to the rudder deflection angle, is the rate of change of the yaw moment coefficient to the roll angular rate, is the rate of change of the yaw moment coefficient with respect to the yaw angular rate.

5. The distributed propulsion aircraft powered yaw control method based on deep reinforcement learning according to claim 4, characterized in that: The input layer of the power differential control neural network includes the angle of attack α, the roll angle φ, the sideslip angle β, the aileron differential deflection angle △δ a and the V-tail deflection angle δ V ; The neural network passes the various parameters of the input layer to the hidden layer, and after calculation, the output layer Q is obtained π (s, a1), Q π (s, a2)……Q π (s,a n ), select the maximum Q value, that is The optimal differential mode A = a i is the selected power differential control strategy; Among them, Q π is the Q value obtained by executing strategy a in state s, s is the state, a is the strategy, and subscripts 1, 2, ... n represent different power differential control strategies.

6. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 5, characterized in that: The basic control law includes: Speed ​​control law: Among them, e V =V r -V is the speed tracking error, V r is the flight speed reference signal; k V is the speed proportional control parameter, k Vi is the speed integral control parameter; T is the ducted fan power, T p is the trim power of the ducted fan in cruise mode; Pitch angle control law: δ e =k θ and θ -k q q (4) Among them, e θ =θ r -θ is the pitch angle tracking error, θ r is the pitch angle reference signal; q is the pitch angle rate; k θ is the pitch angle control parameter, k q is the pitch angle rate control parameter; Roll angle control law: d a =k φ e φ -k p p (5) Among them, e φ =φ r -φ is the roll angle tracking error, φ r is the roll angle reference signal; p is the roll angle rate; k φ is the roll angle control parameter, k p is the roll angle rate control parameter; In order to prevent the roll angle φ from being too large, a roll angle protection system is designed: Among them, δ ac is the aileron rudder deflection angle after limiting; Heading angle control law: Among them, e β =β r -β is the sideslip angle tracking error, β r is the sideslip angle reference signal; r is the yaw rate; k β is the sideslip angle proportional control parameter, k βi is the sideslip angle integral control parameter, k r is the yaw rate control parameter.

7. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 6, characterized in that: The reward function is: R=-|e θ |-|and φ | (8) Among them, R is the reward value.

8. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 7, characterized in that: The distributed propulsion aircraft power yaw control law is as follows: First, the basic control law calculates the ducted fan power T and the elevator deflection angle δ e 、Aileron rudder deflection angle δ a 、Rudder deflection angle δ r ,Then the power differential control neural network calculates the optimal power distribution strategy based on the input layer parameters.

9. The method for distributed propulsion aircraft powered yaw control based on deep reinforcement learning according to claim 8, characterized in that: The training parameters include attenuation coefficient, learning law, number of neural network layers and nodes, experience pool size, random sampling sample size, discount factor, time step, and target network update speed.

Images