Aircraft aerodynamic parameter real-time identification method based on deep learning

Abstract

The invention discloses an aircraft aerodynamic parameter real-time identification method based on deep learning, and belongs to the technical field of aircraft parameter identification. Firstly, aerodynamic parameter perturbation related items are attributed to lumped interference suffered by an aircraft, and for the problem that the lumped interference in an actual system is unknown, an interference observer is used for estimating the lumped interference to obtain an estimated value of the lumped interference. And then perturbation values of all pneumatic parameters implied in lumped interference are separated by means of deep learning, so that an online identification function of the pneumatic parameters is realized. According to the method, the real-time identification problem under the condition that the aerodynamic parameters have perturbation is effectively solved, and the robustness and the accuracy of real-time identification are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of flight vehicle parameter identification, and particularly relates to a method for real-time identification of flight vehicle aerodynamic parameters based on deep learning. Background Art

[0002] A flight vehicle is an apparatus that flies within the atmosphere or in outer space (space), and can be divided into several categories such as aircraft, spacecraft, rockets, and missiles. In recent years, it has been widely used in military and civilian fields. During the research process related to flight vehicles, parameter identification is an important means for establishing the mathematical model of flight vehicles and designing flight control laws, and it plays a crucial role in aspects such as verifying the results of wind tunnel tests, flight vehicle modeling, flight control system design and improvement, and evaluating flight qualities.

[0003] With the rapid development of advanced flight vehicle technologies such as hypersonic flight vehicles, large flexible solar flight vehicles, highly maneuverable unmanned flight vehicles, and variable configuration flight vehicles, the uncertainty of flight vehicle models caused by multiple complexities such as flight environments, structural characteristics, and mission profiles has become increasingly prominent. It is difficult to obtain accurate aerodynamic coefficients through traditional wind tunnel experiments and CFD methods, and there is an urgent need to conduct research on online parameter identification technologies.

[0004] Currently, parameter identification can be divided into two categories: offline identification and online identification. The most studied is offline parameter identification after the flight. At this time, the algorithm can process the entire data set and does not need to consider the real-time performance of the algorithm. The commonly used methods mainly include the least squares method, maximum likelihood method, extended Kalman filter method, and various improved algorithms. The purpose of online identification is to obtain real-time model parameters, providing a model basis for online updates of various advanced flight control systems, real-time flight stability assessment and envelope expansion (or boundary control), and fault detection and system reconstruction. Commonly used methods include the recursive least squares method, iterative extended Kalman filter method, and frequency domain regression analysis method. With the development of machine learning technologies, intelligent methods such as support vector machines and neural networks have also been applied to online parameter identification. However, the offline identification method has limited effects on the above-mentioned advanced flight vehicles and cannot achieve real-time identification functions. In terms of real-time identification, the online identification algorithm based on the recursive idea is very sensitive to the initial value; the frequency domain regression analysis method has relatively poor adaptability to time-varying or non-linear scenarios; intelligent methods basically do not consider parameter perturbations and external interference situations, resulting in poor anti-interference ability.

[0005] Therefore, how to perform real-time identification in the presence of perturbations in flight vehicle aerodynamic parameters is a challenging problem faced by current real-time identification of flight vehicle aerodynamic parameters. Summary of the Invention

[0006] In order to make up for the deficiencies of existing online identification methods, such as sensitivity to initial values, poor non-linear adaptation ability, and poor anti-interference ability, and to solve the problem of real-time identification under the condition of aerodynamic parameter perturbations, combined with the latest research results of the new generation of artificial intelligence, and relying on the powerful feature extraction ability of deep learning methods, a real-time identification method for aircraft aerodynamic parameters based on deep learning is proposed. The specific steps are as follows:

[0007] Step 1: Establish a non-linear kinematic model and a dynamic model for the aircraft's motion about the center of mass.

[0008] The non-linear kinematic model for the aircraft's motion about the center of mass is as follows:

[0009]

[0010] The dynamic model is as follows:

[0011]

[0012] Among them, α, β, γ V represent the angle of attack, sideslip angle, and bank angle of the aircraft respectively; ω x , ω y , ω z represent the roll rate, yaw rate, and pitch rate of the aircraft respectively; m represents the mass of the aircraft; g represents the acceleration due to gravity; V represents the flight speed; θ represents the flight path angle of the aircraft; L and Z represent the lift and side force acting on the aircraft respectively; M x , M y , M z represent the rolling moment, yawing moment, and pitching moment acting on the aircraft respectively; I x , I y , I z represent the moments of inertia of the aircraft about the x-axis, y-axis, and z-axis of the body coordinate system respectively.

[0013] Step 2: Establish an aircraft aerodynamic parameter model and its perturbation model.

[0014] Aircraft aerodynamic parameters can be divided into aerodynamic force coefficients and aerodynamic moment coefficients;

[0015] The aerodynamic force coefficients are expressed as follows:

[0016]

[0017] Among them, C L , C D and C Z are the lift coefficient, drag coefficient, and side force coefficient respectively; is the lift coefficient caused by the angle of attack, is the lift coefficient caused by the control surface; is the drag coefficient caused by the angle of attack, is the drag coefficient caused by the control surface, is the side force coefficient caused by the sideslip angle, is the side force coefficient caused by the control surface, δ a 、δ e 、δ r are the right elevator aileron deflection angle, the left elevator aileron deflection angle, and the rudder deflection angle, respectively.

[0018] The aerodynamic moment coefficient is expressed as:

[0019]

[0020] where C mx ,C my and C mz represent the rolling moment coefficient, the yawing moment coefficient, and the pitching moment coefficient, respectively; is the rolling moment coefficient caused by the sideslip angle, is the rolling moment coefficient caused by the control surface, is the rolling moment coefficient caused by the roll angular rate, is the rolling moment coefficient caused by the yaw angular rate; is the yawing moment coefficient caused by the sideslip angle, is the yawing moment coefficient caused by the control surface, is the yawing moment coefficient caused by the roll angular rate, is the yawing moment coefficient caused by the yaw angular rate; is the pitching moment coefficient caused by the angle of attack, is the pitching moment coefficient caused by the control surface, is the pitching moment coefficient caused by the pitching angular rate.

[0021] The aerodynamic parameters considering parameter perturbations are expressed as:

[0022]

[0023] where C′ L ,C′ D ,C′ Z ,C′ mx ,C′ my ,C′ mz are the actual aerodynamic coefficients and aerodynamic moment coefficients, C L ,C D ,C Z ,C mx ,C my ,C mz are the nominal values, calculated from Equations (3) and (4), ΔCL , ΔC D , ΔC Z , ΔC mx , ΔC my , ΔC mz is the aerodynamic parameter perturbation caused by modeling error and aerodynamic data deviation, which is characterized by multiplying the corresponding nominal value by a percentage. The calculation formula is as follows:

[0024]

[0025] Among them, the percentage of aerodynamic parameter perturbation Δ i (i = C L , C D , C Z , C mx , C my , C mz ) can be modeled by a first-order Gaussian - Markov process:

[0026]

[0027] Among them, and are both Gaussian white noises. ξ i is the integral of the Gaussian white noise .

[0028] Step 3: Perform an equivalent transformation on the kinematic model and dynamic model of the aircraft's motion about the center of mass to obtain an aircraft model for aerodynamic parameter identification.

[0029] Define the attitude angle vector Ω = [α, β, γ V T , define the angular rate vector ω = [ω x , ω y , ω z T , define the control input u = [δ a , δ e , δ r T , then rewrite equations (1) and (2) into the following affine - nonlinear form:

[0030]

[0031] In equation (8), is the kinematic equation of rotation about the center of mass, which describes the change of attitude angle with angular rate as the control input (actually a virtual control quantity); is the dynamic equation of rotation about the center of mass, which describes the change of angular rate with the deflection of the air rudder as the control input. f s is the dynamic system matrix describing the kinematics of the attitude loop, g​​​s The input matrix for describing the static coupling of the attitude loop The lumped disturbance for describing the kinematics of the attitude loop, mainly including aerodynamic parameter perturbations; f f The dynamic system matrix for describing the dynamics of the angular rate loop, g f The input matrix for describing the static coupling of the angular rate loop The lumped disturbance for describing the dynamics of the angular rate loop

[0032] Through this step, the nominal part and the perturbation part of the aerodynamic parameters are separated. The aerodynamic parameters to be identified are the percentage of aerodynamic parameter perturbations included in the vectors and

[0033] Step 4: Using the aircraft model for aerodynamic parameter identification established in Step 3, design a flight control system based on a disturbance observer

[0034] First, a tracking differentiator is designed to obtain smoother command signals and the differential signals of the command signals. The specific structural forms of the tracking differentiators for the attitude loop and the angular rate loop are as follows

[0035]

[0036] where, Ω d = [α d β d γ Vd T and ω d = [ω xd ω yd ω zd T are the reference input signals, Ω c = [α c β c γ Vc T and ω c = [ω xc ω yc ω zc T are the smoothed command signals, r is the acceleration factor, which determines the speed of Ω c tracking Ω d and ω c tracking ω d , h is the integration step size; k is the current time; fhan(x1, x2, r, h0) is the optimal control synthesis function, and its specific algorithm is as follows

[0037]

[0038] ​​​​​Among them, h0 is a new variable independent of the integration step h, called the fast factor, where h0 > h; d, a0, x2, x1, y, a1, a2, s y , a, s a are all internal parameters of the algorithm; sign() is the sign function.

[0039] Next, a model-assisted extended state observer is designed to observe the lumped disturbances of the attitude loop and the angular rate loop and The specific construction form is:

[0040]

[0041] Among them, and are the estimated values of Ω and ω respectively, and are the estimated values of and respectively, which are used for disturbance compensation and controller design, f s +g s ω and f f +g f u are the known parts of the model, K s1 , K s2 , K f1 and K f2 are the observer gain matrices. Assuming that the observer bandwidths of the attitude loop and the angular rate loop are w s0 and w f0 respectively, then K s1 = 2w s0 , K f1 = 2w f0 ,

[0042] Finally, an anti-disturbance control law based on the disturbance observer is designed:

[0043]

[0044] Among them, ω d is the virtual control quantity, which can be regarded as the desired command of the angular rate loop, is the estimated value of the lumped disturbance of the attitude loop, Ω c is the attitude angle command smoothed by the tracking differentiator, is its differential signal, K s = diag{k s1 ,k s2 ,k s3}, k si > 0 (i = 1, 2, 3) represents the control gain of the attitude loop, e s = Ω - Ωc is the attitude loop tracking error; u is the control input, is the estimated value of the lumped disturbance of the angular rate loop, ω c is the attitude angle command after being smoothed by the tracking differentiator, is its differential signal, K f = diag{k f1 , k f2 , k f3}, k fi > 0 (i = 1, 2, 3) represents the control gain of the angular rate loop, e f = ω - ω c is the angular rate loop tracking error.

[0045] Step 5: Establish an aerodynamic parameter identification network based on the LSTM neural network, and train and test this aerodynamic parameter identification network with the simulation data of the flight control system based on the disturbance observer;

[0046] The structure of the aerodynamic parameter identification network is designed as: 4 LSTM layers and 1 fully connected layer. Each LSTM layer has 100 LSTM neurons, and the fully connected layer has 6 nodes, corresponding to the network output dimension.

[0047] The training process is as follows:

[0048] Step 501: Obtain training sample data;

[0049] Randomly give a desired command signal to the flight control system, and at the same time randomly set the reference value of the aerodynamic parameter perturbation percentage. Use the flight controller based on the disturbance observer to track the desired attitude loop command Ω c , record the dynamic pressure Q, angle of attack α, Mach number Ma, velocity V, roll angular rate ω x , yaw angular rate ω y , pitch angular rate ω z , the estimated value of the lumped disturbance of the attitude loop the estimated value of the lumped disturbance of the angular rate loop and use the above data as the network input; use the aerodynamic parameter perturbation percentage as the network output to obtain the training sample data.

[0050] Step 502: Preprocess the training sample data;

[0051] Normalize the network input and output data of the training samples to between [-1, 1]. The data preprocessing formula is:

[0052]

[0053] Among them, x' is the data after normalization, x is the original data, and μ and σ are the mean and standard deviation of the original data, respectively.

[0054] Step 503: Set the training parameters, and based on these training parameters, use the Adam optimizer to iteratively train the aerodynamic parameter identification network and update the network parameters.

[0055] The training parameters include the number of training epochs, batch training size, learning rate, etc.

[0056] The Adam optimizer updates the parameters of the aerodynamic parameter identification network according to the oscillation situation of the historical gradients and the true historical gradients after filtering the oscillations. The loss function is selected as the mean squared error function, and its expression is:

[0057]

[0058] Step 504: When the number of iterations reaches the set number of training epochs or the loss function reaches the minimum value, test the aerodynamic parameter identification network using the test data set. If the average value of the loss function is less than the expected value, the trained aerodynamic parameter identification network is obtained; otherwise, readjust the training plan for retraining until the effect is good on the test set.

[0059] Step 6: Add the trained aerodynamic parameter identification network to the online loop, receive the aircraft state data and the observation data of the disturbance observer in real time, and output the identification value of the aerodynamic parameter perturbation percentage in real time to achieve the function of intelligent real-time identification of aerodynamic parameters based on deep learning.

[0060] So far, the extended state observer and the deep learning aerodynamic parameter identification network together constitute the intelligent real-time identification framework of aerodynamic parameters based on deep learning.

[0061] The advantages of the present invention are as follows:

[0062] (1) In the present invention, the nominal part and the perturbation part of the aerodynamic parameters are separated, the perturbation part of the aerodynamic parameters is regarded as lumped disturbance, and then the extended state observer is used to observe the lumped disturbance to extract the rough information of the aerodynamic parameter perturbation, making it easy for the parameter identification network to extract the aerodynamic parameter perturbation characteristics.

[0063] (2) Based on the aircraft state information and the lumped disturbance observation information, the present invention uses the aerodynamic parameter identification network based on deep learning to accurately extract the perturbation percentage of each aerodynamic parameter hidden in the lumped disturbance, improving the accuracy, real-time performance, and robustness of aerodynamic parameter identification. Description of the Drawings

[0064] Figure 1 It is the step flow chart of a method for real-time identification of aircraft aerodynamic parameters based on deep learning according to the present invention;

[0065] Figure 2 This is the system structure diagram of a method for real-time identification of aircraft aerodynamic parameters based on deep learning according to the present invention;

[0066] Figure 3 This is the structure diagram of a flight control system based on a disturbance observer according to the present invention;

[0067] Figure 4 This is the network structure diagram of the aerodynamic parameter identification network according to the present invention;

[0068] Figure 5 This is a schematic diagram of a method for generating training and test samples of the aerodynamic parameter identification network according to the present invention;

[0069] Figure 6 This is the training structure diagram of the aerodynamic parameter identification network according to the present invention;

[0070] Figure 7 This is the training curve diagram of the aerodynamic parameter identification network according to the present invention;

[0071] Figure 8 This is the curve diagram of the identification effect of the aerodynamic parameter identification network on the perturbation of aerodynamic parameters in the present invention. Detailed implementation manner

[0072] To facilitate the understanding and implementation of the present invention by those of ordinary skill in the art, the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0073] The present invention proposes a method for real-time identification of aircraft aerodynamic parameters based on deep learning. First, the terms related to the perturbation of aerodynamic parameters are reduced to the lumped disturbance received by the aircraft through model equivalent transformation. Aiming at the problem that the lumped disturbance in the actual system is unknown, a disturbance observer is used to estimate the lumped disturbance to obtain an estimated value of the lumped disturbance. Then, by virtue of the powerful feature extraction ability of deep learning, the perturbation values of each aerodynamic parameter hidden in the lumped disturbance are separated, and thus the online identification function of aerodynamic parameters is realized. The present invention makes up for the deficiencies of the existing online identification methods, such as sensitivity to initial values, poor non-linear adaptation ability, and poor anti-interference ability, and effectively solves the problem of real-time identification in the case of aerodynamic parameter perturbation.

[0074] The implementation steps of the method for real-time identification of aircraft aerodynamic parameters based on deep learning are as Figure 1 shown, and the specific steps are as follows:

[0075] Step 1: Establish a non-linear kinematic model and a dynamic model of the aircraft's motion around the center of mass.

[0076] When establishing the flight vehicle motion model, on the one hand, it is necessary to correctly reflect the kinematic and dynamic characteristics of the hypersonic flight vehicle to ensure the accuracy of system modeling; on the other hand, it must be conducive to the design and implementation of the control system and the parameter identification system. This requires that in the process of model analysis and derivation, the main contradictions of the problem must be grasped according to the needs of the control and identification tasks, and the secondary factors should be ignored. The flight vehicle in the present invention is an in-atmosphere hypersonic glide vehicle, and the following assumptions are made: (1) The flight vehicle is an ideal rigid body, and the elastic degrees of freedom of the fuselage and the control surfaces are not considered; (2) The engine of the flight vehicle is shut down, and it relies only on the air rudder to provide the control moment to achieve power-off flight; (3) Due to the high speed of the flight vehicle, it is assumed that the influence of the wind speed on its speed can be ignored.

[0077] Based on the above assumptions, the non-linear kinematic model and dynamic model of the flight vehicle moving around the center of mass are as follows:

[0078]

[0079] Among them, α, β, γ V respectively represent the angle of attack, sideslip angle and bank angle of the flight vehicle; ω x , ω y , ω z respectively represent the roll rate, yaw rate and pitch rate of the flight vehicle; m represents the mass of the flight vehicle; g represents the acceleration due to gravity; V represents the flight speed; θ represents the flight path angle of the flight vehicle; L, Z respectively represent the lift and side force received by the flight vehicle; M x , M y , M z respectively represent the roll moment, yaw moment and pitch moment received by the flight vehicle; I x , I y , I z respectively represent the moments of inertia of the flight vehicle about the x-axis, y-axis and z-axis of the body coordinate system.

[0080] Step 2: Establish the aerodynamic parameter model of the flight vehicle and its perturbation model.

[0081] Regarding the aerodynamic force and aerodynamic moment appearing in equations (1) and (2), their models are established as follows:

[0082]

[0083] Among them, D is the drag force received by the flight vehicle; is the dynamic pressure, ρ is the air density at the altitude where the flight vehicle is located; S is the reference area; b is the wingspan; c is the mean aerodynamic chord; C L , C D , C Z are the lift coefficient, drag coefficient and side force coefficient respectively; C mx , C my , Cmz The roll moment coefficient, yaw moment coefficient, and roll moment coefficient respectively.

[0084] The aerodynamic coefficients in Equation (3) can be modeled as:

[0085]

[0086] Where, is the lift coefficient caused by the angle of attack, is the lift coefficient caused by the control surface; is the drag coefficient caused by the angle of attack, is the drag coefficient caused by the control surface, is the side force coefficient caused by the sideslip angle, is the side force coefficient caused by the control surface, δ a , δ e , δ r The right elevator aileron deflection angle, left elevator aileron deflection angle, and rudder deflection angle respectively.

[0087] The aerodynamic moment coefficients in Equation (4) can be expressed as:

[0088]

[0089] Where, is the roll moment coefficient caused by the sideslip angle, is the roll moment coefficient caused by the control surface, is the roll moment coefficient caused by the roll angular rate, is the roll moment coefficient caused by the yaw angular rate; is the yaw moment coefficient caused by the sideslip angle, is the yaw moment coefficient caused by the control surface, is the yaw moment coefficient caused by the roll angular rate, is the yaw moment coefficient caused by the yaw angular rate; is the pitch moment coefficient caused by the angle of attack, is the pitch moment coefficient caused by the control surface, is the pitch moment coefficient caused by the pitch angular rate.

[0090] Due to the technical limitations of the wind tunnel experiment department, there is a certain degree of uncertainty in the aerodynamic data obtained from the wind tunnel experiments. At the same time, considering that the aerodynamic parameters of hypersonic vehicles have large perturbations during large-scale maneuvering flights, in order to make the vehicle model more realistic, the influence of aerodynamic parameter perturbations is further considered. The actual aerodynamic parameters can be modeled as follows:

[0091]

[0092] Among them, C′ L , C′ D , C′ Z , C′ mx , C′ my , C′ mz are the actual aerodynamic force coefficients and aerodynamic moment coefficients, C L , C D , C Z , C mx , C my , C mz are the nominal values, calculated from Equations (3) and (4), ΔC L , ΔC D , ΔC Z , ΔC mx , ΔC my , ΔC mz are the aerodynamic parameter perturbations caused by modeling errors and aerodynamic data deviations, characterized by multiplying the corresponding nominal values by percentages, and the calculation formula is as follows:

[0093]

[0094] Among them, the aerodynamic parameter perturbation percentage Δ i (i = C L , C D , C Z , C mx , C my , C mz ) can be modeled by a first-order Gaussian - Markov process:

[0095]

[0096] Among them, and are both Gaussian white noises, ξ i is the integral of the Gaussian white noise .

[0097] Step 3: Perform an equivalent transformation on the kinematic model and dynamic model of the aircraft's motion around the center of mass, separate the terms related to aerodynamic parameter perturbations as lumped disturbance terms from the nominal model, and obtain an aircraft model for aerodynamic parameter identification.

[0098] Define the attitude angle vector Ω = [α, β, γ V T , define the angular rate vector ω = [ω x , ω y , ω z T , define the control input u = [δ a , δ e , δ r ​​​T , the system of equations (1) and (2) can be rewritten in the following affine non-linear form with strict feedback:

[0099]

[0100] where,

[0101]

[0102]

[0103] In equation (10), is the kinematic equation of rotation about the center of mass, which describes the change of the attitude angle under the control input of the angular rate (actually the virtual control quantity); is the dynamic equation of rotation about the center of mass, which describes the change of the angular rate under the control input of the air rudder deflection. f s is the dynamic system matrix describing the kinematics of the attitude loop, g s is the input matrix describing the static coupling of the attitude loop, is the lumped disturbance describing the kinematics of the attitude loop, mainly including the perturbation of aerodynamic parameters, f f is the dynamic system matrix describing the dynamics of the angular rate loop, g f is the input matrix describing the static coupling of the angular rate loop, is the lumped disturbance describing the dynamics of the angular rate loop.

[0104] Through this step, the nominal part and the perturbation part of the aerodynamic parameters are separated. The perturbation terms of the aerodynamic parameters to be identified are included in the vectors and , and the aircraft model established for aerodynamic parameter identification is as shown in Figure 2 .

[0105] Step 4: For the aircraft model established for aerodynamic parameter identification in Step 3, design a flight control system based on a disturbance observer.

[0106] As shown in Figure 3 , the present invention designs the control law by dividing the loop using the backstepping method and the dynamic inversion idea. The control system mainly consists of three parts: a tracking differentiator, an extended state observer, and a controller. Among them, the tracking differentiator is mainly used to obtain a smoother command signal and the differential signal of the command signal; the extended state observer is mainly used to estimate the lumped disturbances of the attitude loop and the angular rate loop; the controller is used to calculate the control signal.

[0107] First, a tracking differentiator is designed. The specific structural forms of the tracking differentiators for the attitude loop and the angular rate loop are:

[0108]

[0109] wherein, Ω d = [α d β d γ Vd T and ω d = [ω xd ω yd ω zd T is the reference input signal, Ω c = [α c β c γ Vc T and ω c = [ω xc ω yc ω zc T is the smoothed command signal, r is the acceleration factor, which determines the speed of Ω c tracking Ω d and ω c tracking ω d . h is the integration step; k is the current time; fhan(x1, x2, r, h0) is the fastest control synthesis function, and its specific algorithm is as follows:

[0110]

[0111] wherein, h0 is a new variable independent of the integration step h, called the fast factor, and generally h0 > h; d, a0, x2, x1, y, a1, a2, s y , a, s a are all internal parameters of the algorithm; sign() is the sign function.

[0112] Next, a model-assisted extended state observer is designed to observe the lumped disturbances of the attitude loop and the angular rate loop The specific construction form is:

[0113]

[0114] wherein, and are the estimated values of Ω and ω respectively, and are the estimated values of and respectively, which are used for disturbance compensation and controller design, f s + g s ω and f f + g f u are the known parts of the model, K s1 , K s2 , K f1 ​​​​and K f2 is the observer gain matrix. Assume that the observer bandwidths of the attitude loop and the angular rate loop are w s0 and w f0 respectively. Then, we have K s1 = 2w s0 , K f1 = 2w f0 ,

[0115] Finally, design the anti-jamming control law based on the disturbance observer.

[0116] Define the state error variables:

[0117]

[0118] where, e s = [e α e β e μ T , e f = [e p e q e r T represent the three-channel tracking errors of the attitude loop and the angular rate loop respectively.

[0119] First, consider the attitude loop system of the aircraft

[0120]

[0121] The control task is to design a control system to ensure that the attitude angle Ω can track the given command signal Ω c quickly and stably in the presence of uncertainties, and give the command signal for the angular rate loop.

[0122] Differentiate the attitude loop tracking error e s , and we can get

[0123]

[0124] Therefore, the attitude loop control law can be designed as

[0125]

[0126] where, K s = diag{k s1 , k s2 , k s3}, k si > 0 (i = 1, 2, 3) is the attitude loop control gain matrix.

[0127] Next, consider the angular rate loop system of the aircraft​​

[0128]

[0129] The control task is to design a control system to ensure that the attitude angle ω can quickly and stably track the given command signal ω in the presence of uncertainties c , and give the control signal of the rudder surface.

[0130] Deriving the derivative of the angular rate loop tracking error e f , we can obtain

[0131]

[0132] Therefore, the angular rate loop control law can be designed as

[0133]

[0134] where, K f = diag{k f1 , k f2 , k f3}, k fi > 0 (i = 1, 2, 3) is the angular rate loop control gain matrix.

[0135] Step 5: Based on the long short-term memory (LSTM) neural network, design an aerodynamic parameter identification network, including the design of network input and output and the design of network structure.

[0136] First, design the input and output of the aerodynamic parameter identification network. First, analyze specifically what physical quantities affect the change of aerodynamic parameters, and then select the key influencing factors among them as the input of the aerodynamic parameter identification network. In the present invention, the input of the aerodynamic parameter identification network is designed as: dynamic pressure Q, angle of attack α, Mach number Ma, velocity V, roll angular rate ω x , yaw angular rate ω y , pitch angular rate ω z , attitude loop lumped disturbance estimation value angular rate loop lumped disturbance estimation value The input totals 13 dimensions. The output of the aerodynamic parameter identification network is designed as: percentage perturbation of aerodynamic parameters The output totals 6 dimensions.

[0137] Then, design the structure of the aerodynamic parameter identification network. The structure of the aerodynamic parameter identification network designed in the present invention is as Figure 4 shown, specifically including: 4 LSTM layers and 1 fully connected layer. Each LSTM layer has 100 LSTM neurons, and the fully connected layer has 6 nodes, corresponding to the network output dimension.

[0138] Step 6: Conduct a large number of simulations based on the flight control system designed in Step 4, and use the simulation data as samples to train and test the aerodynamic parameter identification network, and finally obtain the aerodynamic parameter identification network.

[0139] First, construct a sample generation environment and simulate to generate a large number of training data. As Figure 5 shown, based on the designed flight control system, randomly give the desired command signal within the feasible range, and at the same time randomly set the reference value of the aerodynamic parameter perturbation percentage within the controllable range. Use the flight controller based on the disturbance observer to track the desired attitude loop command Ω c , record the dynamic pressure Q, angle of attack α, Mach number Ma, speed V, roll rate ω x , yaw rate ω y , pitch rate ω z , the total disturbance estimation value of the attitude loop the total disturbance estimation value of the angular rate loop and aerodynamic

[0140] parameter perturbation percentage as training samples.

[0141] Then, use the training samples to train the aerodynamic parameter identification network. The training structure of the aerodynamic parameter identification network is as Figure 6 shown. Before training the aerodynamic parameter identification network, data preprocessing, training parameter setting, training algorithm selection, and loss function setting are required.

[0142] The role of data preprocessing is to normalize the input and output data of the aerodynamic parameter identification network to between [-1, 1], so as to avoid the dominant role of the characteristics of some data caused by different orders of magnitude of the data. The data preprocessing formula is:

[0143]

[0144] where x′ is the normalized data, x is the original data, and μ and σ are the mean and standard deviation of the original data respectively.

[0145] Training parameter setting mainly includes the setting of hyperparameters such as the number of training epochs, batch training size, learning rate, etc., so that the network is trained according to the above settings during training. In the present invention, the settings of each hyperparameter are shown in Table 1.

[0146] Table 1 Hyperparameter settings of the aerodynamic parameter identification network

[0147]

[0148] The training algorithm selects the Adam optimizer, which can update the parameters of the aerodynamic parameter identification network according to the oscillation of historical gradients and the true historical gradients after filtering the oscillations.

[0149] The loss function selects the mean squared error function, and its expression is:

[0150]

[0151] Under the above settings, the aerodynamic parameter identification network can perform sufficient training iterations on the training set and be tested on the test set. If the average value of the loss function is less than the expected value, the trained aerodynamic parameter identification network can be directly obtained; if the effect of the network on the test set does not meet the accuracy requirements, the training scheme is readjusted for retraining until the effect on the test set is good.

[0152] After the above steps, an aerodynamic parameter identification network that can accurately identify the perturbation values of aerodynamic parameters is finally obtained.

[0153] Step 7: Add the trained aerodynamic parameter identification network to the online loop, receive the aircraft state data and the observation data of the disturbance observer in real time, and output the identification value of the aerodynamic parameter perturbation percentage in real time, so as to realize the function of intelligent real-time identification of aerodynamic parameters based on deep learning.

[0154] In the present invention, the training loss change curve of the aerodynamic parameter identification network is as Figure 7 shown. It can be seen that under the current design scheme, the training loss of the aerodynamic parameter identification network can converge rapidly.

[0155] In the present invention, the identification effect of the aerodynamic parameter identification network on the perturbation values of aerodynamic parameters is as Figure 8 shown. It can be seen that the aerodynamic parameter identification network can quickly and accurately identify the perturbation percentage of aerodynamic parameters, and the RMSE (root mean square error) of the identification result is within the range of 0.0495, verifying the identification accuracy of the parameter identification network. At the same time, the single operation time of the aerodynamic parameter identification network added to the online loop is 12 ms, which is less than the control period of the system, 20 ms, verifying the real-time performance of the parameter identification network.

[0156] The content not described in detail in the specification of the present invention belongs to the prior art well-known to those skilled in the art.

Claims

1. A real-time identification method for aircraft aerodynamic parameters based on deep learning, characterized in that, Specifically, it includes the following steps: Step 1: Establish the non - linear kinematic model and dynamic model of the aircraft's motion around the center of mass, rewrite them into the non - ray form, and obtain the aircraft model for aerodynamic parameter identification; Define the attitude angle vector Ω = [α, β, γ V T , define the angular rate vector ω = [ω x , ω y , ω z T , define the control input u = [δ a , δ e , δ r T , rewrite the non - linear kinematic model and dynamic model into the following affine non - linear form:​​​ Among them, α, β, γ V respectively represent the angle of attack, sideslip angle, and bank angle of the aircraft; ω x , ω y , ω z respectively represent the roll rate, yaw rate, and pitch rate of the aircraft; δ a , δ e , δ r are the right elevator aileron deflection angle, left elevator aileron deflection angle, and rudder deflection angle respectively; is the kinematic equation of rotation about the center of mass, describing the change of the attitude angle under the control input of the angular rate; is the dynamic equation of rotation about the center of mass, describing the change of the angular rate under the control input of the air rudder deflection; f s is the dynamic system matrix describing the kinematics of the attitude loop, g s is the input matrix describing the static coupling of the attitude loop, is the lumped disturbance describing the kinematics of the attitude loop, including aerodynamic parameter perturbations; f f is the dynamic system matrix describing the dynamics of the angular rate loop, g f is the input matrix describing the static coupling of the angular rate loop, is the lumped disturbance describing the dynamics of the angular rate loop; Step 2: Establish the aircraft aerodynamic parameter model and its perturbation model, and separate the nominal part and the perturbation part of the aerodynamic parameters; The aircraft aerodynamic parameters are divided into aerodynamic force coefficients and aerodynamic moment coefficients; The aerodynamic force coefficients are expressed as follows: Among them, C L , C D and C Z are the lift coefficient, the drag coefficient and the side force coefficient respectively; is the lift coefficient caused by the angle of attack, is the lift coefficient caused by the control surface; is the drag coefficient caused by the angle of attack, is the drag coefficient caused by the control surface, is the side force coefficient caused by the sideslip angle, is the side force coefficient caused by the control surface; The aerodynamic moment coefficients are expressed as: Among them, C mx , C my and C mz respectively represent the rolling moment coefficient, the yaw moment coefficient, and the pitch moment coefficient; is the rolling moment coefficient caused by the sideslip angle, is the rolling moment coefficient caused by the control surface, is the rolling moment coefficient caused by the roll angular rate, is the rolling moment coefficient caused by the yaw angular rate; is the yaw moment coefficient caused by the sideslip angle, is the yaw moment coefficient caused by the control surface, is the yaw moment coefficient caused by the roll angular rate, is the yaw moment coefficient caused by the yaw angular rate; is the pitch moment coefficient caused by the angle of attack, is the pitch moment coefficient caused by the control surface, is the pitch moment coefficient caused by the pitch angular rate; The aerodynamic parameters considering parameter perturbation are expressed as: Among them, C′ L ,C′ D ,C′ Z ,C′ mx ,C′ my ,C′ mz are the actual aerodynamic force coefficients and aerodynamic moment coefficients, C L ,C D ,C Z ,C mx ,C my ,C mz are the nominal values, calculated from Equations (2) and (3), ΔC L ,ΔC D ,ΔC Z ,ΔC mx ,ΔC my ,ΔC mz are the aerodynamic parameter perturbations caused by modeling errors and aerodynamic data deviations, characterized by multiplying the corresponding nominal values by percentages, and the calculation formula is as follows: Among them, the perturbation percentage Δ of the aerodynamic parameters i (i = C L , C D , C Z , C mx , C my , C mz ) is modeled by a first-order Gaussian-Markov process: Among them, and are both Gaussian white noises, and ξ i is the integral of Gaussian white noise . Step 3: Based on the aircraft model for aerodynamic parameter identification, separate the nominal part and the perturbation part of the aerodynamic parameters. The aerodynamic parameters to be identified are included in the vectors and ; Step 4: Use the aircraft model for aerodynamic parameter identification to design a flight control system based on a disturbance observer; First, design a tracking differentiator to obtain the command signal and the differential signal of the command signal; The specific structural forms of the tracking differentiators for the attitude loop and the angular rate loop are: where, Ω d = [α d β d γ Vd T and ω d = [ω xd ω yd ω zd T is the reference input signal, Ω c = [α c β c γ Vc T and ω c = [ω xc ω yc ω zc T is the smoothed command signal, r is the acceleration factor, which determines the speed of Ω c tracking Ω d and ω c tracking ω d , h is the integration step size; k is the current time; fhan(x1, x2, r, h0) is the fastest control synthesis function;​​​​ Next, a design model-assisted extended state observer is used to observe the lumped disturbances of the attitude loop and the angular rate loop and The specific construction form is as follows: Among them, and are the estimated values of Ω and ω respectively, and are respectively and 's estimated values, which are used for interference compensation and controller design. f s +g s ω and f f +g f u are the known parts of the model. K s1 , K s2 , K f1 and K f2 are the observer gain matrices; it is assumed that the observer bandwidths of the attitude loop and the angular rate loop are w s0 and w f0 respectively, then K s1 = 2w s0 , K f1 = 2w f0 , Finally, design an anti - interference control law based on a disturbance observer: where, ω d is a virtual control variable and is the desired command of the angular rate loop, is the estimated value of the total disturbance of the attitude loop, Ω c is the attitude angle command smoothed by the tracking differentiator, is its differential signal, K s = diag{k s1 , k s2 , k s3}, k si > 0 (i = 1, 2, 3) represents the control gain of the attitude loop, e s = Ω - Ω c is the attitude loop tracking error; u is the control input, is the estimated value of the total disturbance of the angular rate loop, ω c is the attitude angle command smoothed by the tracking differentiator, is its differential signal, K f = diag{k f1 , k f2 , k f3}, k fi > 0 (i = 1, 2, 3) represents the control gain of the angular rate loop, e f = ω - ω c is the angular rate loop tracking error; Step 5: Establish an aerodynamic parameter identification network based on LSTM, and train and test this aerodynamic parameter identification network through the simulation data of the flight control system based on a disturbance observer; The training process is as follows: Step 501: Obtain the training sample data and perform pre - processing; Randomly give the desired command signal to the flight control system, and simultaneously randomly set the perturbation percentage reference value of the aerodynamic parameters. Use the flight controller based on the disturbance observer to track the desired attitude loop command Ω c , record the dynamic pressure Q, angle of attack α, Mach number Ma, velocity V, roll angle rate ω during the simulation x , yaw angle rate ω y , pitch angle rate ω z , the lumped disturbance estimation value of the attitude loop , the lumped disturbance estimation value of the angular rate loop And use the above data as the network input; use the perturbation percentage of the aerodynamic parameters as the network output to obtain the training sample data; Step 502: Set the training parameters, and based on these training parameters, perform iterative training on the aerodynamic parameter identification network through the Adam optimizer and update the network parameters; The Adam optimizer updates the parameters of the aerodynamic parameter identification network according to the oscillation situation of the historical gradients and the true historical gradients after filtering the oscillations. The loss function is selected as the mean squared error function, and its expression is: Step 503: When the number of iterations reaches the set number of training epochs or the loss function reaches the minimum value, test the aerodynamic parameter identification network through the test data set. If the average value of the loss function is less than the expected value, obtain the trained aerodynamic parameter identification network; otherwise, adjust the network structure and training parameters and retrain until the requirements are met on the test set; Step 6: Add the trained aerodynamic parameter identification network to the online loop, receive the aircraft state data and the observation data of the disturbance observer in real - time, and output the identification value of the percentage of aerodynamic parameter perturbation in real - time to realize the function of intelligent real - time identification of aerodynamic parameters based on deep learning.

2. The real-time identification method of aircraft aerodynamic parameters based on deep learning according to claim 1, characterized in that The non - linear kinematic model and dynamic model of the aircraft's motion around the center of mass are respectively: The non - linear kinematic model of the aircraft's motion around the center of mass is as follows: The dynamic model is as follows: Among them, m represents the mass of the aircraft; g represents the acceleration due to gravity; V represents the flight speed; θ represents the ballistic inclination angle of the aircraft; L and Z respectively represent the lift and side force received by the aircraft; M x , M y , M z respectively represent the rolling moment, yaw moment and pitch moment received by the aircraft; I x , I y , I z respectively represent the moments of inertia of the aircraft about the x-axis, y-axis and z-axis of the body coordinate system.

3. The real-time identification method for the aerodynamic parameters of an aircraft based on deep learning according to claim 1, characterized in that, The algorithm of the fastest control synthesis function fhan(x1,x2,r,h0) is as follows: Among them, h0 is a new variable independent of the integration step size h, called the fast factor, where h0 > h; d, a0, x2, x1, y, a1, a2, s y , a, s a are all internal parameters of the algorithm, and sign() is the sign function.

4. The real-time identification method for aircraft aerodynamic parameters based on deep learning according to claim 1, wherein, The structural design of the aerodynamic parameter identification network is: 4 LSTM layers and 1 fully - connected layer. Each LSTM layer has 100 LSTM neurons, and the fully - connected layer has 6 nodes, corresponding to the network output dimension.

5. The real-time identification method of the aerodynamic parameters of an aircraft based on deep learning according to claim 1, characterized in that, The pre - processing of the training sample data refers to: Normalize the network input and output data of the training samples to the range of [-1,1]. The data pre - processing formula is: Where x′ is the normalized data, x is the original data, and μ and σ are the mean and standard deviation of the original data respectively.

6. The real-time identification method for aircraft aerodynamic parameters based on deep learning according to claim 1, characterized in that The training parameters include the number of training epochs, the batch training size, and the learning rate.

Images