A flying wing aircraft attitude control method based on neural network incremental dynamic inversion

Abstract

The present invention discloses a flying wing aircraft attitude control method based on neural network incremental dynamic inversion, the steps include: establishing a nonlinear model of the flying wing aircraft attitude angle and angular velocity loop based on the principle of time scale separation; designing the attitude angle loop control law based on the dynamic inversion method, and designing the angular velocity loop control law based on the incremental dynamic inversion method; introducing a single hidden layer neural network to online compensate for the inverse error caused by sensor delay, external interference and inaccurate modeling, so as to ensure that the aircraft attitude angle can achieve the expected dynamic performance. The present invention utilizes the characteristics of the neural network to fit unknown dynamics online, solves the inverse error caused by sensor delay in the traditional inverse control method, further improves the robustness of the controller under the premise of ensuring that the controller is insensitive to the dynamics of the flying wing aircraft and external interference, and improves the dynamic performance and steady-state performance of the flight attitude.

Description

Technical Field

[0001] The present invention belongs to the field of advanced aviation control technology, and in particular relates to a flying wing aircraft attitude control method based on neural network incremental dynamic inversion. Background Art

[0002] Flying wing aircraft have become the main configuration of high-performance military UAVs on the battlefield today due to their excellent penetration capability and battlefield survivability. However, since the flying wing layout eliminates the horizontal tail and vertical tail, the longitudinal and lateral stability of the aircraft is reduced, the efficiency of the control surface operation is also greatly reduced, and the control surfaces are highly coupled. These defects make it difficult for traditional aircraft control methods to meet the design requirements of flying wing aircraft. Usually, the design of the flight control law is to linearize the aircraft state differential equation with a small perturbation at a certain working point to obtain the time-invariant system matrix of the current working state, and then use linear control theory to design the controller structure and parameters. This design method is difficult to ensure the control effect while taking into account multiple state quantities when facing highly coupled nonlinear flying wing layout aircraft.

[0003] Therefore, in the face of more complex nonlinear and highly coupled control objects, it is necessary to develop more advanced flight control methods. After experiments and tests on advanced foreign aircraft, the dynamic inversion method is considered to be an effective nonlinear control method. This method obtains a pseudo-linear system by inverting the original system for control design. It is usually combined with the time-scale separation method to divide the aircraft state quantity into multiple groups to meet the dynamic inversion method's requirement that the number of control quantities and state quantities is equal. However, the direct application of the dynamic inversion method of state feedback requires accurate system information, and in actual engineering, the system model is often accompanied by many uncertainties caused by interference and errors.

[0004] Based on the dynamic inversion method, the incremental dynamic inversion method is proposed to reduce the dependence on accurate system information. Incremental dynamic inversion replaces the precise transfer function of the system required by direct dynamic inversion by first-order differential feedback of the state quantity, which greatly improves the robustness of the controller while simplifying the calculation of the control law. Therefore, this method is very suitable for control objects with uncertainty. In the control design of flying wing aircraft, the angular rate loop is very susceptible to the fluctuation of aerodynamic parameters and external interference, so it is very suitable to use the incremental dynamic inversion method to improve the reliability of the controller. However, in actual flight, the angular rate differential feedback required by the incremental dynamic inversion method is often accompanied by the delay error of the filter. These errors will affect the quality of aircraft attitude control after entering the controller.

[0005] In summary, the prior art lacks an effective solution to the filter error in incremental dynamic inverse control, and cannot meet the precise attitude control requirements of flying wing aircraft. The present invention is based on the inverse control principle of time scale separation, regards the feedback delay error as a kind of inverse error, and uses the characteristics of neural networks that can fit unknown nonlinear functions online to compensate for pseudo-control quantities with errors. The attitude tracking performance of flying wing aircraft is improved by reasonably selecting controller and neural network parameters. The simulation results show that the control method has good control quality for the attitude of flying wing aircraft with floating aerodynamic parameters, external interference and filter delay. Summary of the invention

[0006] In order to solve the problems in the prior art, the present invention provides a flying wing aircraft attitude control method based on neural network incremental dynamic inversion. The method realizes rapid and reliable control of the attitude angle of a flying wing aircraft with floating aerodynamic parameters, external interference and filter delay.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A flying wing aircraft attitude control method based on neural network incremental dynamic inversion comprises the following steps:

[0009] S1 establishes the nonlinear model of the attitude loop and angular rate loop of the flying wing aircraft according to the principle of time scale separation, and obtains the mathematical relationship between the aircraft control surface, angular rate state quantity, and attitude angle state quantity;

[0010] S2 establishes an attitude loop model according to the attitude angle state quantity, solves the attitude loop control law based on the dynamic inverse method, and obtains a stable attitude angle loop controller;

[0011] S3 establishes an angular rate loop model according to the angular rate state quantity, selects the second-order reference model as the angular rate loop command model, solves the angular rate loop control law based on the incremental dynamic inverse method, introduces a single hidden layer neural network to perform online compensation for the inverse error of the incremental dynamic inverse controller, and constructs the adaptive learning law of the network based on the Lyapunov stability principle to obtain the angular rate loop controller.

[0012] Furthermore, the nonlinear models of the attitude loop and angular velocity loop of the flying wing aircraft in step S1 are specifically:

[0013] a. Nonlinear differential equation of angular rate state:

[0014]

[0015] In the above formula, p, q and r represent the roll, pitch and yaw angular rates of the aircraft respectively, J is the inertia matrix of the aircraft, and L A 、M A and NA It represents the rolling, pitching and yaw aerodynamic moments generated by the fuselage and control surfaces in the body coordinate system, M T It is represented by the engine thrust torque, and the specific expression of the torque is:

[0016]

[0017]

[0018]

[0019] M T =z T T max u t

[0020] In the above formula, u a ,u e and u r represents the control input of the aircraft's total aileron, total elevator and total rudder, u t represents the throttle opening of the aircraft, Q is the dynamic pressure of the aircraft, S w is the wing reference area, b is the wing span, c is the wing span A is the average aerodynamic chord length, z T Indicates the coordinate of the thrust point on the z-axis of the aircraft body, T max Indicates the maximum thrust of the engine; represents the aerodynamic moment derivative related to the rolling moment of the aircraft, represents the aerodynamic moment derivative related to the pitch moment of the aircraft, represents the aerodynamic moment derivative related to the aircraft yaw moment, and β represents the aircraft sideslip angle;

[0021] b. Nonlinear differential equation of attitude angle state quantity:

[0022]

[0023] In the above formula, μ, α and β represent the roll angle, angle of attack and sideslip angle of the aircraft respectively, γ and χ represent the track tilt angle and heading tilt angle respectively, T BV Represents the transformation matrix from the velocity coordinate system to the body coordinate system.

[0024] Furthermore, the step S2 of determining the relationship between the attitude angle reference signal and the angular rate reference signal based on the dynamic inverse method comprises the following steps:

[0025] Rewrite the aircraft attitude loop model into affine form:

[0026]

[0027] In the above formula, x2=[μ α β] T is the state quantity of the attitude loop, x1=[pqr] T is the state quantity of the angular rate loop, f2 is the attitude angle loop system transfer matrix, and g2 is the attitude angle loop control input matrix;

[0028] According to the dynamic inverse control theory, when the number of state quantities is equal to the number of control quantities and the input control matrix g2 is reversible, the dynamic inverse control law of the attitude angle loop is:

[0029]

[0030] In the above formula is the angular rate loop command signal calculated by the attitude loop controller, v2 = [v μ v α v β ] T It is the pseudo control input of the attitude loop, which is calculated by the linear PID controller. By properly selecting the parameters of the linear PID controller, the attitude angle state can be controlled through the pseudo control input.

[0031] Furthermore, the step S3 of determining the relationship between the angular rate reference signal and the control input based on the incremental dynamic inverse method comprises the following steps:

[0032] Rewrite the vehicle angular rate loop model into affine form:

[0033]

[0034] In the above formula represents the control input that can produce the ideal angular rate, f1 is the angular rate loop system transfer matrix, and g1 is the control input matrix;

[0035] According to the dynamic inverse control theory, when the number of state quantities is equal to the number of control quantities and the input control matrix g1 is reversible, the dynamic inverse control law of the angular rate loop is:

[0036]

[0037] In the above formula, v1=[v p v q v r ] T Represents the pseudo control input of the angular rate loop, which is calculated by the linear PD controller; Since the transfer matrix f1 of the angular rate loop system contains a large number of aerodynamic derivatives that are easily affected by changes in the external environment, dynamic inverse control depends on accurate transfer function matrices and control input matrices. Incremental dynamic inversion methods are usually used to reduce the control law's dependence on f1;

[0038] The affine form of the angular rate loop differential equation is expanded into a first-order Taylor series form. When the controller sampling time interval is extremely small, the state increment Δx is ignored, and the high-order terms in the expansion are ignored to obtain its simplified form:

[0039]

[0040] In the above formula, u is the control input at the current moment, and u0 is the control input at the previous moment. The angular rate differential at the previous moment is shifted and used as the control quantity to replace the angular rate differential at the current moment to obtain the incremental dynamic inverse control law:

[0041]

[0042] In the above formula, v=[v p v q v r ] T represents the pseudo control input of the angular rate loop, which is calculated by the PD controller, It is measured using a second-order low-pass filter, and its transfer function is in the form of:

[0043]

[0044] In the above formula, s represents the independent variable of the transfer function after Laplace transformation in the frequency domain, ω n is the natural angular frequency of the filter, ξ n is the filter damping ratio;

[0045] The control input u0 at the previous moment is measured using the integral form of a second-order filter, and its transfer function is specifically expressed as follows:

[0046]

[0047] According to the angular velocity derivative measured at the previous moment And the control variable u0 at the previous moment, combined with the control input matrix g1, the complete acceleration loop incremental dynamic inverse control law is obtained:

[0048]

[0049] u des =Δu des +u0

[0050] In the above formula, Q is the dynamic pressure of the aircraft, I xx is the moment of inertia of the aircraft around the x-axis of the body, I yy is the moment of inertia of the aircraft around the y-axis of the body, I zzis the moment of inertia of the aircraft around the z-axis of the body. By properly selecting the parameters of the linear PD controller, the angular rate state can be controlled through the pseudo-control input.

[0051] Furthermore, the use of a single hidden layer neural network to compensate for the pseudo control amount in the incremental dynamic inverse law in step S3 includes the following steps:

[0052] The second-order reference model is appropriately chosen for the angular rate loop controller to limit the input rate of the desired command:

[0053]

[0054] In the above formula, ω m is the natural angular frequency of the reference model, ξ m is the reference model damping ratio, x des is the ideal angular rate command output by the attitude loop, x c is the reference angular rate command output by the second-order reference model. According to the reference angular rate command output by the second-order reference model and the actual angular rate measured by the sensor, the output of the PD linear controller is obtained:

[0055]

[0056] In the above formula, K p , K d is the controller parameter to be designed, and the pseudo control quantity of the inverse controller is set to track the second-order differential reference instruction of the state quantity. And add neural network for compensation, then the pseudo control signal entering the controller is:

[0057]

[0058] In the above formula, v ad is the neural network compensation result, Δ is the inverse error caused by external interference and sensor delay. The error characteristic equation of the angular rate loop controller is obtained by shifting the above equation:

[0059]

[0060] In the above formula is the error vector, assuming that the neural network can compensate for the inverse error in real time, that is, v ad =Δ, and the control parameter K of the linear controller p and K d If A is a Hurwitz matrix, the tracking error can be guaranteed to converge to 0;

[0061] Assume that the number of neural network inputs is n, the number of intermediate layers is m, and the number of outputs is l. Based on the single hidden layer neural network structure with bias, write the network output:

[0062] v nn =W T [b w ; f(V T [b v ;x1;…;x n ])]

[0063] In the above formula, x1,…,x n is the neural network input vector, b v and b w is the input layer bias and output layer bias, usually set to 1, f(·) is the hidden layer sigmoid activation function, and its expression is f(x) = 1 / (1+e -x ), V is the input layer weight matrix of (n+1)×m, and W is the output layer weight matrix of (m+1)×l;

[0064] The neural network input layer is represented as X = [b v x1 … x n ] T , then the output of the intermediate layer activation function is Z = V T X=[z1 … z m ] T , the output of the middle layer of the network is expressed as f(Z) = [b w f(z1) … f(z m )] T , then the network output layer is v nn =W T f(Z)=[v1 … v l ] T , the specific form of the online learning algorithm of the two matrices is as follows:

[0065]

[0066]

[0067] In the above formula, k V and k W represents the network learning error adjustment factor, τ V and τ W Represents the network learning rate matrix, satisfying τ V >0 and τ W >0, Represents the derivative matrix of the intermediate layer activation function f(Z) to the intermediate layer input Z.

[0068] In order to ensure the compensation effect of the compensator when the inverse error is large, a high gain robust term at the output is introduced:

[0069] v r =-kr ||T x ||ζ

[0070] In the above formula, ||T x || is the total weight matrix T at the previous moment x = the matrix norm of diag(V,W), k r is the robust term coefficient greater than 0, ζ=e T PB is the error dynamic vector, P is the Lyapnov equation A T The positive solution of P+PA+Q=0 is as follows if Q=2I:

[0071]

[0072] In summary, the output of the neural network adaptive compensator is:

[0073] v ad =v nn +v r

[0074] Choose the neural network input as During the calculation process of the aircraft attitude control, the neural network weight matrix can be continuously adjusted to compensate in real time for the inverse error caused by sensor delay.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] The present invention utilizes a flying wing aircraft attitude control method based on neural network incremental dynamic inversion, which does not rely on accurate aerodynamic parameter measurement, but uses the incremental form of dynamic inversion to simplify the controller calculation complexity while reducing the controller's sensitivity to external changes. In addition, for unforeseen inverse errors such as sensor delay, a neural network compensator is introduced to compensate for the pseudo-control input of the incremental dynamic inversion, further improving the robustness of the flying wing aircraft attitude loop controller. Since the incremental form of the inverse control method and the single hidden layer neural network compensation do not require the introduction of additional hardware, nor do they need to increase the system sampling frequency, the improved adaptive compensation method will not increase the development cost, nor will it add additional computational burden to the flight controller hardware. On the whole, the flying wing aircraft attitude control method based on neural network incremental dynamic inversion proposed in this patent improves the reliability of the controller while ensuring the aircraft attitude stability and transient performance, and is easy to expand to other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a schematic diagram of the structural principle of the flying wing aircraft attitude control method based on neural network incremental dynamic inversion of the present invention;

[0078] Figure 2Schematic diagram of a single hidden layer neural network structure in an embodiment of the present invention;

[0079] Figure 3 is a schematic diagram of the structure of a second-order reference model in an embodiment of the present invention;

[0080] Figure 4 is a time domain response diagram of the roll angle in an embodiment of the present invention;

[0081] Figure 5 is a time domain response diagram of the angle of attack in an embodiment of the present invention;

[0082] Figure 6 is a time domain response diagram of the sideslip angle in an embodiment of the present invention. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The embodiments are only used to explain the present invention and are not used to limit the present invention.

[0084] The present invention provides a flying wing aircraft attitude control method based on neural network incremental dynamic inversion. The complete control scheme is as shown in the attached figure. Figure 1 As shown, it includes a command filter, an attitude angle loop controller and an angular rate loop control law. The input values ​​of the entire aircraft attitude control system include a roll angle command, an angle of attack command and a sideslip angle command, and the output values ​​include the aircraft attitude angle and attitude angular rate information.

[0085] Taking a certain type of flying wing aircraft as an example, the goal of the present invention is to ensure its stability, good adaptability and control quality when the aircraft is in the normal cruising flight phase and there are aerodynamic parameter fluctuations and sensor delays. According to the principle of time scale separation, the state quantity of the aircraft inner loop is divided into two groups of subsystems based on different time scales. The attitude angle loop that changes slowly over time adopts dynamic inverse control, and the angular rate loop that changes quickly over time and is easily affected by aerodynamic parameter fluctuations adopts incremental dynamic inverse control, and uses neural networks to compensate for sensor delay errors online. The input instructions of the aircraft attitude control system include the reference roll angle μ c , reference angle of attack α c and the reference sideslip angle β c The output is roll angle μ, angle of attack α, sideslip angle β, roll angle rate p, pitch angle rate q and roll angle rate r. The attitude controller includes attitude angle dynamic inverse controller and angular rate increment dynamic inverse controller. The input value of attitude angle dynamic inverse control law is the reference roll angle μ of the aircraft after instruction filtering. c , reference angle of attack α c and the reference sideslip angle β c , the output is the roll angle rate command p des , pitch angle rate command αdes and the yaw rate command r des The angular rate incremental dynamic inverse controller includes a second-order reference model, an incremental dynamic inverse control law, and a neural network adaptive compensator. The input is the angular rate command The output is the total aileron control amount Total elevator control amount and total rudder control

[0086] According to the principle of time scale separation and engineering practice, the state variables are divided into two groups, attitude angle and angular rate, with the attitude loop control of flying wing aircraft as the requirement. The control laws are designed for each group respectively. The state variables are divided into: x1 = [pqr] T The angular rate state quantity, which changes the fastest over time, will directly affect the three angular rate differentials of the aircraft during the flight mission. Then we can think of the change in angular velocity; x2 = [μ α β] T is the attitude angle state quantity. This group of state quantities changes relatively slowly over time. When the aircraft control input changes, the angular rate x1 changes immediately and enters a stable state. After x1 changes, the attitude angle x2 changes gradually, thereby enabling the aircraft to achieve the ideal flight attitude desired by the pilot or controller.

[0087] The attitude angle command [μ c α c β c ] T Based on the dynamic inversion method, three angular rate instructions [p des q des r des ] T The attitude angle state equation of the flying wing aircraft is:

[0088]

[0089] In the above formula, γ and χ represent the track tilt angle and heading tilt angle respectively, T BV Represents the transformation matrix from the velocity coordinate system to the body coordinate system, and its matrix form is:

[0090]

[0091] T BV =T VB T

[0092] Arrange the attitude angle state equation into affine form:

[0093]

[0094] Where f2 is the transfer function of the attitude angle loop system, and g2 is the attitude angle loop control input function;

[0095] The above formula can be organized into a dynamic inverse control law:

[0096]

[0097] In the above formula is the angular rate loop command signal calculated by the attitude angle dynamic inverse controller, v2 = [v μ v α v β ] T It is the pseudo control input of the attitude loop and is calculated by the linear PID controller.

[0098] The angular velocity state equation of a flying wing aircraft is:

[0099]

[0100] In the above formula, p, q and r represent the roll, pitch and yaw angular rates of the aircraft respectively, J is the inertia matrix of the aircraft, and L A 、M A and N A It represents the rolling, pitching and yaw aerodynamic moments generated by the fuselage and control surfaces in the body coordinate system, M T It is represented by the engine thrust torque, and the specific expression of the torque is:

[0101]

[0102]

[0103]

[0104] M T =z T T max u t

[0105] In the above formula, u a ,u e and u r represents the control input of the aircraft's total aileron, total elevator and total rudder, u t represents the throttle opening of the aircraft, Q is the dynamic pressure of the aircraft, S w is the wing reference area, b is the wing span, c is the wing span A is the average aerodynamic chord length, z T Indicates the coordinate of the thrust point on the z-axis of the aircraft body, T max Indicates the maximum thrust of the engine; represents the aerodynamic moment derivative related to the rolling moment of the aircraft, represents the aerodynamic moment derivative related to the pitch moment of the aircraft, represents the aerodynamic moment derivative related to the aircraft yaw moment, and β represents the aircraft sideslip angle.

[0106] The vehicle angular rate state equation is organized into an affine form:

[0107]

[0108] In the above formula represents the control input that can produce the ideal angular rate, f1 is the angular rate loop system transfer matrix, and g1 is the control input matrix;

[0109] Expand the above formula into a first-order Taylor series form, ignore the higher-order terms and Δx, and ignore the higher-order terms in the expansion to obtain its simplified form:

[0110]

[0111] In the above formula, u is the control input at the current moment, and u0 is the control input at the previous moment. The angular rate differential at the previous moment is shifted and used as the control quantity to replace the angular rate differential at the current moment to obtain the incremental dynamic inverse control law:

[0112]

[0113] In the above formula, v=[v p v q v r ] T Represents the pseudo control input of the angular rate loop, which is calculated by the linear PD controller. The state quantity differential The second-order low-pass filter sZ(s) is used to remove noise and measure the control quantity u0 at the previous moment. The control quantity u0 is obtained through the integral form Z(s) of the filter. The specific transfer function is as follows:

[0114]

[0115]

[0116] In the above formula, s represents the independent variable of the transfer function after Laplace transformation in the frequency domain, ω n is the natural angular frequency of the filter, ξ n is the filter damping ratio.

[0117] In order to limit the input rate of the desired command, a second-order reference command model is added before the controller; then a neural network adaptive compensator is introduced to compensate for the inverse error contained in the pseudo-control input, and a neural network compensation structure is established for the pseudo-control quantity of each state quantity of the angular velocity loop, as shown in the attached figure. Figure 2 As shown, the number of input layers, intermediate layers, and output layers of each single hidden layer neural network is selected to be 6, 7, and 1 respectively, and the input is where x c , and From the second-order reference model, x and from the angular velocity loop sensor and the flushing filter with time delay, ||T x ||The matrix norm of the total weight matrix from the previous moment, the compensation output of each neural network is:

[0118] v nn =W T [b w ; f(V T [b v ;x])]

[0119] In order to ensure the compensation effect of the compensator when the inverse error is large, a high gain robust term at the output is introduced:

[0120] v r =-k r ||T||ζ

[0121] Where k r >0 is the robust term coefficient, ζ=e T PB is the error dynamic vector, P is the Lyapnov equation A T The positive solution of P+PA+Q=0 is as follows if Q=2I:

[0122]

[0123] The neural network weight matrices V and W are learned online according to the following equations:

[0124]

[0125]

[0126] in:

[0127] X=[b v x1 … x6] T ,Z=V T X

[0128] f(Z)=[b w f(z1) … f(z7)]T

[0129]

[0130] The output of the neural network adaptive compensator is:

[0131] v ad =v nn +v r

[0132] Then the complete form of the pseudo-control input of the angular rate loop is:

[0133]

[0134] When the neural network adaptive module is able to compensate the inverse error online, the angular rate loop tracking error will converge to 0 in a very short time.

[0135] In the actual flight process, the pilot gives the angle of attack command α through the longitudinal control stick. c , give the side slip angle command β through the pedal c (usually kept at 0), and the roll angle instruction μ is given through the lateral joystick c ,In numerical simulation research, usually only the command generator can solve the required ,attitude according to the prescribed flight trajectory. The command model includes the ,command filter and the second-order reference model.

[0136] As attached Figure 3 As shown, the command filter and the second-order reference model are both selected as follows:

[0137]

[0138] Take ω m =2.5rad / s,ζ m =0.8.

[0139] For the second-order angular rate filter required by the angular rate controller, choose the following form:

[0140]

[0141] Take ω n =25rad / s,ζ n =0.8.

[0142] The main parameters of the flying wing aircraft are shown in Table 1:

[0143] Table 1 Main parameters of flying wing aircraft

[0144]

[0145]

[0146] Maintain the aircraft airspeed V = 0.6Ma and the throttle opening u t = 0.3, ignoring the slower changing outer loop state quantity, that is, γ = χ = 0. To keep the aircraft balanced, set β c = 0. Initialize all attitude angles and angular velocities to 0. The aerodynamic coefficient deviation is 20%, and the filter delay link is 0.1s.

[0147] Select the linear controller parameters as shown in Table 2:

[0148] Table 2 Linear controller parameters

[0149]

[0150] The parameters of the neural network compensator are selected as shown in Table 3:

[0151] Table 3 Neural network compensation structure parameters

[0152]

[0153] At the beginning of the simulation, μ des = -5° and α des =5°, μ is given at 10s des =5° and α des =-5°, μ is given at 20s des =0 and α des =0, to stabilize the aircraft, the simulation process maintains β des = 0, the attitude loop is simulated using the incremental dynamic inversion method without and with online neural network compensation, the simulation time is 30s, and the results are as follows Figure 4-6 shown.

[0154] It can be seen from the simulation results that the incremental dynamic inversion method can well decouple the control of the complex nonlinear flying wing aircraft model. On this basis, the neural network compensator further improves the dynamic performance of the controller: it eliminates the 6.7% overshoot of the roll angle and pitch angle of the original method, reduces the overshoot of the yaw angle by about 5%, and reduces the adjustment time of the roll angle and pitch angle by 3.5s (Δ=±2%). Each attitude angle can quickly track the reference instruction, and avoids oversaturation of the control input under the adjustment of the reference model.

[0155] The flying wing aircraft attitude control method of the present invention does not rely on the solution or stability analysis of the nonlinear system, but only needs to discuss the inverse transformation of the system. The incremental dynamic inverse method is simple to design and simple to calculate. The neural network adaptive compensator can compensate for the inverse error not considered in the control system. The overall method has good command tracking performance and strong robustness. In addition to the flying wing aircraft in the embodiment, the present invention can also be applied to linear or nonlinear systems in other fields, and has strong adaptability.

[0156] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A flying wing aircraft attitude control method based on neural network incremental dynamic inversion, characterized in that: The following steps are involved: S1 establishes the nonlinear model of the attitude loop and angular rate loop of the flying wing aircraft according to the principle of time scale separation, and obtains the mathematical relationship between the aircraft control surface, angular rate state quantity, and attitude angle state quantity; S2 establishes an attitude loop model according to the attitude angle state quantity, solves the attitude loop control law based on the dynamic inverse method, and obtains a stable attitude angle loop controller; S3 establishes an angular rate loop model according to the angular rate state quantity, selects the second-order reference model as the angular rate loop command model, solves the angular rate loop control law based on the incremental dynamic inverse method, introduces a single hidden layer neural network to perform online compensation for the inverse error of the incremental dynamic inverse controller, and constructs the network's adaptive learning law based on the Lyapunov stability principle to obtain the angular rate loop controller; The nonlinear models of the attitude loop and angular velocity loop of the flying wing aircraft in step S1 are specifically: a. Nonlinear differential equation of angular rate state: In the above formula, p, q and r represent the roll, pitch and yaw angular rates of the aircraft respectively, J is the inertia matrix of the aircraft, and L A 、M A and N A It represents the rolling, pitching and yaw aerodynamic moments generated by the fuselage and control surfaces in the body coordinate system, M T It is represented by the engine thrust torque, and the specific expression of the torque is: M T =z T T max u t In the above formula, u a ,u e and u r represents the control input of the aircraft's total aileron, total elevator and total rudder, u t represents the throttle opening of the aircraft, Q is the dynamic pressure of the aircraft, S w is the wing reference area, b is the wing span, c is the wing span A is the average aerodynamic chord length, z T Indicates the coordinate of the thrust point on the z-axis of the aircraft body, T max Indicates the maximum thrust of the engine; represents the aerodynamic moment derivative related to the rolling moment of the aircraft, represents the aerodynamic moment derivative related to the pitch moment of the aircraft, represents the aerodynamic moment derivative related to the aircraft yaw moment, and β represents the aircraft sideslip angle; b. Nonlinear differential equation of attitude angle state quantity: In the above formula, μ, α and β represent the roll angle, angle of attack and sideslip angle of the aircraft respectively, γ and χ represent the track tilt angle and heading tilt angle respectively, T BV Represents the transformation matrix from the velocity coordinate system to the body coordinate system; The step S2 of determining the relationship between the attitude angle reference signal and the angular rate reference signal based on the dynamic inverse method comprises the following steps: Rewrite the aircraft attitude loop model into affine form: In the above formula, x2=[μ α β] T is the state quantity of the attitude loop, x1=[pqr] T is the state quantity of the angular rate loop, f2 is the attitude angle loop system transfer matrix, and g2 is the attitude angle loop control input matrix; According to the dynamic inverse control theory, when the number of state quantities is equal to the number of control quantities and the input control matrix g2 is reversible, the dynamic inverse control law of the attitude angle loop is: In the above formula is the angular rate loop command signal calculated by the attitude loop controller, v2 = [v μ v α v β ] T It is the pseudo control input of the attitude loop, which is calculated by the linear PID controller. By properly selecting the parameters of the linear PID controller, the attitude angle state quantity can be controlled through the pseudo control input. Step S3 determines the relationship between the angular rate reference signal and the control input based on the incremental dynamic inverse method, including the following steps: Rewrite the vehicle angular rate loop model into affine form: In the above formula represents the control input that can produce the ideal angular rate, f1 is the angular rate loop system transfer matrix, and g1 is the control input matrix; According to the dynamic inverse control theory, when the number of state quantities is equal to the number of control quantities and the input control matrix g1 is reversible, the dynamic inverse control law of the angular rate loop is: In the above formula, v1=[v p v q v r ] T Represents the pseudo control input of the angular rate loop, which is calculated by the linear PD controller; Since the transfer matrix f1 of the angular rate loop system contains a large number of aerodynamic derivatives that are easily affected by changes in the external environment, dynamic inverse control depends on accurate transfer function matrices and control input matrices. Incremental dynamic inversion methods are usually used to reduce the control law's dependence on f1; The affine form of the angular rate loop differential equation is expanded into a first-order Taylor series form. When the controller sampling time interval is extremely small, the state increment Δx is ignored, and the high-order terms in the expansion are ignored to obtain its simplified form: In the above formula, u is the control input at the current moment, and u0 is the control input at the previous moment. The angular rate differential at the previous moment is shifted and used as the control quantity to replace the angular rate differential at the current moment to obtain the incremental dynamic inverse control law: In the above formula, v=[v p v q v r ] T represents the pseudo control input of the angular rate loop, which is calculated by the PD controller, It is measured using a second-order low-pass filter, and its transfer function is in the form of: In the above formula, s represents the independent variable of the transfer function after Laplace transformation in the frequency domain, ω n is the natural angular frequency of the filter, ξ n is the filter damping ratio; The control input u0 at the previous moment is measured using the integral form of a second-order filter, and its transfer function is specifically expressed as follows: According to the angular velocity derivative measured at the previous moment And the control quantity u0 at the previous moment, combined with the control input matrix g1, the complete acceleration loop incremental dynamic inverse control law is obtained: you des =Δu des +u0 In the above formula, Q is the dynamic pressure of the aircraft, I xx is the moment of inertia of the aircraft around the x-axis of the body, I yy is the moment of inertia of the aircraft around the y-axis of the body, I zz is the moment of inertia of the aircraft around the z-axis of the fuselage. By properly selecting the parameters of the linear PD controller, the angular rate state can be controlled through the pseudo control input. In step S3, a single hidden layer neural network is introduced to perform online compensation for the inverse error of the incremental dynamic inverse controller, which includes the following steps: The second-order reference model is appropriately chosen for the angular rate loop controller to limit the input rate of the desired command: In the above formula, ω m is the natural angular frequency of the reference model, ξ m is the reference model damping ratio, x des is the ideal angular rate command output by the attitude loop, x c is the reference angular rate command output by the second-order reference model. According to the reference angular rate command output by the second-order reference model and the actual angular rate measured by the sensor, the output of the PD linear controller is obtained: In the above formula, K p , K d is the controller parameter to be designed, and the pseudo control quantity of the inverse controller is set to track the second-order differential reference instruction of the state quantity. And add neural network for compensation, then the pseudo control signal entering the controller is: In the above formula, v ad is the neural network compensation result, Δ is the inverse error caused by external interference and sensor delay. The error characteristic equation of the angular rate loop controller is obtained by shifting the above equation: In the above formula is the error vector, assuming that the neural network can compensate for the inverse error in real time, that is, v ad =Δ, and the control parameter K of the linear controller p and K d If A is a Hurwitz matrix, the tracking error can be guaranteed to converge to 0; Assume that the number of neural network inputs is n, the number of intermediate layers is m, and the number of outputs is l. Based on the single hidden layer neural network structure with bias, write the network output: v nn =W T [b w ;f(V T [b v ;x1;...;x n ])] In the above formula, x1,...,x n is the neural network input vector, b v and b w is the input layer bias and output layer bias, usually set to 1, f(·) is the hidden layer sigmoid activation function, and its expression is f(x) = 1 / (1+e -x ), V is the input layer weight matrix of (n+1)×m, and W is the output layer weight matrix of (m+1)×l; The neural network input layer is represented as X = [b v x1…x n ] T , then the output of the intermediate layer activation function is Z = V T X=[z1…z m ] T , the output of the middle layer of the network is expressed as f(Z) = [b w f(z1)…f(z m )] T , then the network output layer is v nn =W T f(Z)=[v1…v l ] T , the specific form of the online learning algorithm of the two matrices is as follows: In the above formula, k V and k W represents the network learning error adjustment factor, τ V and τ W Represents the network learning rate matrix, satisfying τ V >0 and τ W >0, Represents the derivative matrix of the intermediate layer activation function f(Z) to the intermediate layer input Z; In order to ensure the compensation effect of the compensator when the inverse error is large, a high gain robust term at the output is introduced: v r =-k r ||T x ||ζ In the above formula, ||T x || is the total weight matrix T at the previous moment x = the matrix norm of diag(V,W), k r is the robust term coefficient greater than 0, ζ=e T PB is the error dynamic vector, P is the Lyapnov equation A T The positive solution of P+PA+Q=0 is as follows if Q=2I: In summary, the output of the neural network adaptive compensator is: v ad =v nn +v r Choose the neural network input as During the calculation process of the aircraft attitude control, the neural network weight matrix can be continuously adjusted to compensate in real time for the inverse error caused by sensor delay.

Images