Aircraft parameter identification method combining gated circulation unit neural network model and Gaussian-Newton optimization algorithm

Abstract

The invention provides an aircraft parameter identification method combining a gating circulation unit neural network model and a Gaussian-Newton optimization algorithm, and the method combines a GRU neural network model and the Gaussian-Newton optimization algorithm, and is used for solving the problem of complex aircraft parameter identification. The GRU neural network model is designed to model aircraft dynamics by establishing a non-linear function mapping relationship directly between measurement inputs and outputs of the aircraft system without any dynamics priori hypothesis. And a Gaussian-Newton optimization algorithm is combined with the trained GRU network model, and an optimal parameter identification result is obtained by minimizing a cost function related to unknown aerodynamic parameters in an iteration mode.

Description

Technical Field

[0001] The present invention relates to the field of time domain aircraft system identification methods, and specifically to an aircraft aerodynamic parameter identification method combining a gated recurrent unit neural network model and a Gauss-Newton optimization algorithm. Background Art

[0002] Aircraft aerodynamic parameter identification describes a process of obtaining the most likely values ​​of unknown aerodynamic parameters that characterize a dynamic aircraft system. It helps to accurately and comprehensively understand the aerodynamic model structure of the target aircraft and is crucial for diagnosing and improving the design of the aircraft. In the past few decades, parameter identification methods based on the maximum likelihood principle (the most commonly used methods: output error method and filter error method) have been widely used to capture aerodynamic information closely related to the dynamic behavior of the aircraft system from flight data through deterministic analysis. However, both the output error method and the filter error method require not only a priori assumptions about the dynamics of the aircraft system, but also multiple numerical integrations of the assumed dynamic model. Therefore, the output error method or the filter error method is not suitable for parameter identification problems that lack a comprehensive understanding of the dynamics of the aircraft system (such as: aircraft parameter identification problems under complex flight conditions (stall lag, high angle of attack flight and large-scale time-varying maneuvers), deformation aircraft parameter identification problems with large-scale deformation of the wing, etc.), and are likely to encounter integral divergence problems.

[0003] Artificial neural networks provide a comprehensive framework for nonlinear function mapping in the input-output subspace and have been proven to be a reliable nonlinear function approximator. In addition, the potential and effectiveness of artificial neural network-based parameter identification methods in solving complex problems related to aircraft parameter identification have been successively demonstrated. Although the current artificial neural network-based identification methods have brought gains to aircraft parameter identification because they do not require dynamic assumptions, most of these methods rely on feedforward neural networks to map the dynamics of the aircraft system. Unlike standard recurrent neural networks, feedforward neural networks cannot handle input sequences of arbitrary time lengths because there are no feedback connections between nodes / units and layers. However, it is extremely challenging to train standard recurrent neural networks with stable convergence while avoiding the occurrence of gradient vanishing or exploding problems.

[0004] The improved recurrent neural network is not only superior to the feedforward neural network in processing long time series data, but also avoids the gradient explosion or vanishing problem that may occur in the standard recurrent neural network. Therefore, it is necessary and valuable to explore the potential and effectiveness of the improved recurrent neural network in the field of aircraft parameter identification. Summary of the invention

[0005] In order to solve the problem that the parameter identification method based on the maximum likelihood principle requires a priori assumptions about the dynamics of the aircraft when processing complex aircraft parameter identification, and to overcome the convergence problem of the existing parameter identification methods based on feedforward neural networks or standard recurrent neural networks when processing long time series flight data, the present invention proposes a parameter identification method based on a gated recurrent unit (GRU) neural network, which combines the GRU neural network model and the Gauss-Newton optimization algorithm to solve the problem of complex aircraft parameter identification. The GRU neural network model is designed to model the aircraft dynamics by directly establishing a nonlinear function mapping relationship between the measured input and output of the aircraft system without any prior assumptions about dynamics. The Gauss-Newton optimization algorithm is combined with the trained GRU network model to obtain the optimal parameter identification result by iteratively minimizing the cost function related to the unknown aerodynamic parameters.

[0006] The technical solution of the present invention is:

[0007] A method for aircraft parameter identification combining a gated recurrent unit neural network model and a Gauss-Newton optimization algorithm comprises the following steps:

[0008] Step 1: Build and train the GRU neural network prediction model;

[0009] Step 1.1: Establish a GRU neural network prediction model; the GRU neural network prediction model includes an input layer, an L-layer GRU layer and an output layer; a Dropout layer is added to the GRU layer;

[0010] The input of the GRU neural network prediction model is X = [X1, X2, ..., X N-1 ], the output is Y=[Y2,Y3,…,Y N ], using the input vector X at time t t To predict the output vector Y at time t+1 t+1 , t=1,2,…,N-1; where N represents the total number of discrete time points contained in the flight data;

[0011] Step 1.2: Train the GRU neural network prediction model established in step 1.1:

[0012] The data collected by the aircraft sensors are collected through flight tests, including the angle of attack, sideslip angle, attitude angle, attitude angular velocity, flight speed and the three-axis acceleration under the body axis system. The angular acceleration of the aircraft in three directions around its body axis system is calculated by numerical differentiation method, and the aerodynamic coefficient and torque coefficient of the aircraft are calculated by Newton's law, so as to obtain the observation matrix Z = [Z2, Z3, ..., Z N ] and the input matrix X = [X1,X2,...,X N-1]; the meaning of the vectors in the observation matrix is ​​the same as that of the vectors in the output matrix;

[0013] Use the input matrix X and the observation matrix Z to train the GRU neural network prediction model;

[0014] Step 2: Combine the trained GRU network with the Gauss-Newton optimization algorithm to realize the aircraft parameter identification function:

[0015] Step 2.1: Select the corresponding aerodynamic model structure according to the actual maneuvering flight conditions of the selected aircraft in the flight test;

[0016] Step 2.2: Set the initial value of the aerodynamic parameter vector Θ to be identified in the aerodynamic model;

[0017] Step 2.3: Using the values ​​of the aircraft motion state variables at the initial moment obtained by the sensor in the flight data, and calculating the corresponding aerodynamic force and moment coefficient values ​​according to the aerodynamic model; combining the values ​​of the aircraft motion state variables at the initial moment obtained by the sensor with the aerodynamic force and moment coefficient values ​​calculated according to the aerodynamic model to obtain the input vector X1 at the initial moment;

[0018] Step 2.4: Input the input vector X1 at the initial moment into the trained GRU neural network prediction model, and solve the aerodynamic force and torque coefficient at each time point in combination with the aerodynamic model to predict the output matrix Y = [Y2, Y3, ..., Y N ]; Calculate the measurement noise covariance matrix R;

[0019] Step 2.5: Minimize the cost function J(Θ) using the standard unconstrained Gauss-Newton algorithm and one-dimensional line search, and iteratively update the aerodynamic parameter vector Θ:

[0020] J(θ)=det(R)

[0021] Where FΔΘ=-G

[0022]

[0023] In the formula, i represents the number of iterations; represents a dimensionless parameter, whose value is determined by the one-dimensional line search algorithm in each iteration; F and G represent the information matrix and the gradient vector respectively; ΔΘ represents the change vector of the aerodynamic parameter to be identified;

[0024] Step 2.6: Return to step 2.3 until the convergence condition is met and the aerodynamic parameter vector is obtained.

[0025] Furthermore, in step 1.1, the Dropout layer is used to quantify the probability of units in each GRU layer being discarded or excluded.

[0026] Furthermore, for longitudinal flight data, the input vector X of the GRU neural network prediction model is t And the output vector Y t+1 Respectively expressed as:

[0027] X t,纵向 =[V(t),α(t),θ(t),q(t),C D (t),C L (t),C m (t)] T

[0028]

[0029] Where V is the flight speed, α is the angle of attack, θ is the pitch angle, q is the pitch angular velocity, and C is D is the drag coefficient, C L is the lift coefficient, C m is the pitching moment coefficient, is the pitch angular acceleration, a x represents the acceleration component of the aircraft in the x direction under the body axis system, a z It represents the acceleration component of the aircraft in the z direction under the body axis system;

[0030] For lateral flight data, input vector X t And the output vector Y t+1 They are respectively expressed as:

[0031] X t,横航向 =[β(t),φ(t),ψ(t),p(t),r(t),C y (t),C l (t),C n (t)] T

[0032]

[0033] Among them, β represents the sideslip angle, φ represents the roll angle, ψ represents the yaw angle, p represents the roll angular velocity, r represents the yaw angular velocity, C y represents the lateral force coefficient, C l Denotes the rolling moment coefficient, C n represents the yaw moment coefficient, is the roll angular acceleration, represents the yaw angular acceleration, a y It represents the acceleration component of the vehicle in the y direction under the body axis system.

[0034] Furthermore, in step 1.2, the root mean square propagation algorithm is used to train the GRU neural network prediction model, and the neural network prediction accuracy is evaluated by the root mean square error and the coefficient of determination.

[0035] Furthermore, in step 2.1, if the aircraft performs a longitudinal maneuver at a small angle of attack, the linear longitudinal aerodynamic model is selected; when the aircraft performs a longitudinal maneuver at a large angle of attack, the nonlinear longitudinal aerodynamic model with time effect is selected; if the aircraft performs a lateral heading maneuver, the lateral heading aerodynamic model is selected.

[0036] Furthermore, in step 2.1, for the longitudinal aerodynamic model, the aerodynamic parameter vector to be identified is

[0037] Θ 纵向 =[C D0 ,C L0 ,C Lα ,C Lq ,C Lδe ,C m0 ,C mα ,C mq ,C mδe ] T

[0038] For the lateral aerodynamic model, the aerodynamic parameter vector to be identified is

[0039] Θ 横航向 =[C y0 ,C yβ ,C yr ,C yp ,C yδr ,C l0 ,C lβ ,C lr ,C lp ,

[0040] C lδa ,C lδr ,C n0 ,C nβ ,C nr ,C np ,C nδr ] T

[0041] In the formula, C ·0 represents the aerodynamic coefficient deviation term, C ·α and C ·β represents the aerodynamic derivative of static stability, C ·p , C ·q and C ·r represents the dynamic aerodynamic derivative term, and represents the static aerodynamic derivative term related to the control surface deflection, and · represents the symbols D, L, m, y, l and n.

[0042] Beneficial Effects

[0043] The present invention provides an aircraft parameter identification method combining a gated recurrent unit neural network model and a Gauss-Newton optimization algorithm. The present invention trains a gated recurrent unit neural network model GRU based on data collected from real flights, and is used to directly establish a nonlinear mapping relationship between the input and output of the aircraft system, effectively avoiding the uncertainty problem of aircraft dynamics modeling. Then, the trained GRU neural network is used to predict the output matrix Y, and a cost function and gradient based on the predicted noise covariance matrix are constructed. The cost function J(Θ) is iteratively optimized through the Gauss-Newton optimization algorithm and one-dimensional line search until the cost function residual meets the iterative convergence condition, thereby completing the complete parameter identification process. The present invention is a parameter identification method combining a gated recurrent unit neural network model and a Gauss-Newton optimization algorithm. By introducing a gated recurrent unit neural network, the inaccurate modeling of the dynamics model and the problem of integral divergence can be effectively avoided, and the parameter identification problem of various aircraft can be solved.

[0044] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0046] Figure 1 :GRU-GN method flow chart;

[0047] Figure 2 : Gated Recurrent Unit Neural Network (GRU) structure;

[0048] in: Figure 2 (a) shows the unfolded structure of the GRU network; Figure 2 (b) represents the detailed structure of a hidden unit in the Lth GRU layer at time t;

[0049] Figure 3 :Application example of GRU-GN method in parameter identification of a small unmanned aerial vehicle;

[0050] in: Figure 3 (a) shows a real picture of a small unmanned aerial vehicle collecting flight data in the air; Figure 3 (b) shows the elevator of the small unmanned aerial vehicle (δ e ) The waveform of the excitation; Figure 3(c) shows the verification of the recognition effect of the GRU-GN method. DETAILED DESCRIPTION

[0051] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0052] like Figure 1 As shown, the present invention is an aircraft parameter identification method combining a gated recurrent unit neural network model and a Gauss-Newton optimization algorithm, including a gated recurrent unit neural network model (GRU) training and a Gauss-Newton parameter identification method combined with the GRU; the specific steps are as follows:

[0053] Step 1: Build and train the GRU neural network prediction model; including the following steps:

[0054] Step 1.1: GRU neural network prediction model building:

[0055] Build as Figure 2 The GRU neural network structure shown. Figure 2 (a) shows the unfolded structure of the GRU neural network, including the input layer, L GRU layers, and the output layer; Figure 2 (b) Represents the detailed structure of a hidden unit in the Lth GRU layer at time t.

[0056] In the present invention, a Dropout layer is added to the GRU layer of the neural network to prevent the network from overfitting. The Dropout layer is used to quantify the probability of units in each GRU layer being discarded or excluded. Figure 2 (b) It can be seen that there are two gates in the hidden unit (i.e., update gate and reset gate ), represents the output vector of the hidden unit in the Lth GRU layer, represents the output vector of the hidden unit in the L-1th GRU layer. At time t, the mathematical form of the hidden unit in the Lth GRU layer is:

[0057]

[0058] In the formula, σ and tanh represent the sigmoid function and the hyperbolic tangent function respectively; and Respectively represent the input weight matrix in the L-th GRU layer; and They represent the cyclic weight matrices in the L-th GRU layer respectively; and They represent the offset vector in the L-th GRU layer respectively.

[0059] The input and output of the GRU network model are represented by the matrix X = [X1, X2, ..., X N-1 ] and the matrix Y = [Y2,Y3,…,Y N ], where N represents the total number of discrete time points contained in the flight data. The function of the GRU network prediction model is: using the input vector X at time t t To predict the output vector Y at time t+1 t+1 (t=1,2,…,N-1).

[0060] For longitudinal flight data, the input vector X t And the output vector Y t+1 Respectively expressed as:

[0061] X t,纵向 =[V(t),α(t),θ(t),q(t),C D (t),C L (t),C m (t)] T

[0062]

[0063] Where V is the flight speed, α is the angle of attack, θ is the pitch angle, q is the pitch angular velocity, and C is D is the drag coefficient, C L is the lift coefficient, C m is the pitching moment coefficient, is the pitch angular acceleration, a x represents the acceleration component of the aircraft in the x direction under the body axis system, a z It represents the acceleration component of the vehicle in the z direction under the body axis system.

[0064] For lateral flight data, input vector X t And the output vector Y t+1 They are respectively expressed as:

[0065] X t,横航向 =[β(t),φ(t),ψ(t),p(t),r(t),C y (t),C l (t),C n (t)] T

[0066]

[0067] Among them, β represents the sideslip angle, φ represents the roll angle, ψ represents the yaw angle, p represents the roll angular velocity, r represents the yaw angular velocity, C y represents the lateral force coefficient, C lDenotes the rolling moment coefficient, C n represents the yaw moment coefficient, is the roll angular acceleration, represents the yaw angular acceleration, a y It represents the acceleration component of the vehicle in the y direction under the body axis system.

[0068] Step 1.2: Before realizing the prediction function of the network model, the network model needs to be trained. Through flight tests, the data collected by the aircraft sensors are collected, including the angle of attack, sideslip angle, attitude angle, attitude angular velocity, flight speed and the three-axis acceleration under the body axis system. The angular acceleration of the aircraft in three directions around its body axis system is calculated by numerical differentiation method, and the aerodynamic coefficient and torque coefficient of the aircraft are calculated by Newton's law, so as to obtain the observation matrix Z = [Z2, Z3, ..., Z N ] and the input matrix X = [X1,X2,...,X N-1 ]; the meaning of the vectors in the observation matrix is ​​the same as that of the vectors in the output matrix.

[0069] The GRU neural network prediction model is trained by using the root mean square propagation (RMSProp) algorithm using the input matrix X and the observation matrix Z. The prediction accuracy of the neural network is measured by the root mean square error (RMSE) and the coefficient of determination (R 2 )Evaluate:

[0070]

[0071] Step 2: Combine the trained GRU network with the Gauss-Newton optimization algorithm to realize the aircraft parameter identification function:

[0072] Step 2.1: Select the corresponding aerodynamic model structure according to the actual maneuvering flight conditions of the selected aircraft in the flight test; for example, if the aircraft performs longitudinal maneuvers at a small angle of attack, select a linear longitudinal aerodynamic model; and when the aircraft performs longitudinal maneuvers at a large angle of attack, select a nonlinear longitudinal aerodynamic model with time effect; if the aircraft performs lateral heading maneuvers, select a lateral heading aerodynamic model.

[0073] In the aerodynamic model, the aerodynamic parameters to be identified are:

[0074] For the longitudinal aerodynamic model, the aerodynamic parameter vector to be identified is

[0075]

[0076] For the lateral aerodynamic model, the aerodynamic parameter vector to be identified is

[0077]

[0078] In the above formula, C·0 represents the aerodynamic coefficient deviation term, C ·α and C ·β represents the aerodynamic derivative of static stability, C ·p , C ·q and C ·r represents the dynamic aerodynamic derivative term, C ·δe , C ·θm and C ·δr represents the static aerodynamic derivative term related to the deflection of the control surface, and represents the symbols D, L, m, y, l and n, which are well-known concepts in the field of flight mechanics.

[0079] Step 2.2: Set the initial value of the aerodynamic parameter vector Θ to be identified in the aerodynamic model.

[0080] Step 2.3: Use the aircraft state values ​​obtained by sensors in the flight data and calculate the corresponding aerodynamic force and torque coefficient values ​​according to the aerodynamic model; combine the aircraft state values ​​obtained by the sensors with the aerodynamic force and torque coefficient values ​​calculated according to the aerodynamic model to obtain the input vector X1 at the initial moment.

[0081] For example, when the aircraft performs longitudinal maneuvers at a small angle of attack, the longitudinal aerodynamic model is used, and the aerodynamic parameter vector to be identified is Θ 纵向 , the input vector of the GRU network model is X t,纵向 , where V, α, θ, q at the initial moment are collected by sensors, and C D ,C L ,C m It is calculated by the longitudinal aerodynamic model.

[0082] If the aircraft performs lateral maneuvers, the lateral aerodynamic model is used, and the aerodynamic parameter vector to be identified is Θ 横航向 , the input vector of the GRU network model is X t,横航向 , where β, φ, ψ, p, r at the initial moment are collected by sensors, and C y ,C l ,C n It is calculated through the lateral aerodynamic model.

[0083] Step 2.4: Input the input vector X1 at the initial moment into the trained GRU neural network prediction model, and combine the aerodynamic model to solve the aerodynamic force and torque coefficient at each time point, and predict the output matrix Y(=[Y2,Y3,...,Y N ]); Calculate the measurement noise covariance matrix R:

[0084]

[0085] Step 2.5: Minimize the cost function J(θ) using the standard unconstrained Gauss-Newton (GN) algorithm and one-dimensional line search. The iterative update of the aerodynamic parameter vector θ to be predicted involves predicting the output vector Y t and the response gradient The calculation formula is as follows:

[0086] J(θ)=det(R)

[0087] Where FΔΘ=-G

[0088]

[0089] In the above formula, i represents the number of iterations; represents a dimensionless parameter, whose value is determined by the one-dimensional line search algorithm in each iteration; F and G represent the information matrix and the gradient vector, respectively; ΔΘ represents the change vector of the aerodynamic parameter to be identified.

[0090] Step 2.6: Return to step 2.3 until the convergence condition is met:

[0091]

[0092] In the above formula, J(Θ)0 represents the initial value of the cost function; J(Θ) represents the cost function value corresponding to the i-th iteration; e0 represents J(Θ) i About J(Θ) i-1 The relative error of .

[0093] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A method for aircraft parameter identification combining a gated recurrent unit neural network model and a Gauss-Newton optimization algorithm, characterized in that: The following steps are involved: Step 1: Build and train the GRU neural network prediction model; Step 1.1: Establish a GRU neural network prediction model; the GRU neural network prediction model includes an input layer, an L-layer GRU layer and an output layer; a Dropout layer is added to the GRU layer; The input of the GRU neural network prediction model is X = [X1, X2, ..., X N-1 ], the output is Y=[Y2,Y3,…,Y N ], using the input vector X at time t t To predict the output vector Y at time t+1 t+1 , t=1,2,…,N-1; where N represents the total number of discrete time points included in the flight data; Step 1.2: Train the GRU neural network prediction model established in step 1.1: The data collected by the aircraft sensors are collected through flight tests, including the angle of attack, sideslip angle, attitude angle, attitude angular velocity, flight speed and the three-axis acceleration under the body axis system. The angular acceleration of the aircraft in three directions around its body axis system is calculated by numerical differentiation method, and the aerodynamic coefficient and torque coefficient of the aircraft are calculated by Newton's law, so as to obtain the observation matrix Z = [Z2, Z3, ..., Z N ] and the input matrix X = [X1,X2,...,X N-1 ]; the meaning of the vectors in the observation matrix is ​​the same as that of the vectors in the output matrix; Use the input matrix X and the observation matrix Z to train the GRU neural network prediction model; Step 2: Combine the trained GRU network with the Gauss-Newton optimization algorithm to realize the aircraft parameter identification function: Step 2.1: Select the corresponding aerodynamic model structure according to the actual maneuvering flight conditions of the selected aircraft in the flight test; Step 2.2: Set the initial value of the aerodynamic parameter vector Θ to be identified in the aerodynamic model; Step 2.3: Using the values ​​of the aircraft motion state variables at the initial moment obtained by the sensor in the flight data, and calculating the corresponding aerodynamic force and moment coefficient values ​​according to the aerodynamic model; combining the values ​​of the aircraft motion state variables at the initial moment obtained by the sensor with the aerodynamic force and moment coefficient values ​​calculated according to the aerodynamic model to obtain the input vector X1 at the initial moment; Step 2.4: Input the input vector X1 at the initial moment into the trained GRU neural network prediction model, and solve the aerodynamic force and torque coefficient at each time point in combination with the aerodynamic model to predict the output matrix Y = [Y2, Y3, ..., Y N ]; Calculate the measurement noise covariance matrix R; Step 2.5: Minimize the cost function J(Θ) using the standard unconstrained Gauss-Newton algorithm and one-dimensional line search, and iteratively update the aerodynamic parameter vector Θ: J(θ)=det(R) Where FΔΘ=-G In the formula, i represents the number of iterations; represents a dimensionless parameter, whose value is determined by the one-dimensional line search algorithm in each iteration; F and G represent the information matrix and the gradient vector respectively; ΔΘ represents the change vector of the aerodynamic parameter to be identified; Step 2.6: Return to step 2.3 until the convergence condition is met and the aerodynamic parameter vector is obtained.

2. The method according to claim 1, characterized in that: In step 1.1, the Dropout layer is used to quantify the probability of units being dropped or excluded in each GRU layer.

3. The method according to claim 1, characterized in that: For longitudinal flight data, the input vector X of the GRU neural network prediction model is t And the output vector Y t+1 Respectively expressed as: X t,纵向 =[V(t),α(t),θ(t),q(t),C D (t),C L (t),C m (t)] T Where V is the flight speed, α is the angle of attack, θ is the pitch angle, q is the pitch angular velocity, and C is D is the drag coefficient, C L is the lift coefficient, C m is the pitching moment coefficient, is the pitch angular acceleration, a x represents the acceleration component of the aircraft in the x direction under the body axis system, a z It represents the acceleration component of the aircraft in the z direction under the body axis system; For lateral flight data, input vector X t And the output vector Y t+1 They are respectively expressed as: X t,横航向 =[β(t),φ(t),ψ(t),p(t),r(t),C y (t),C l (t),C n (t)] T Among them, β represents the sideslip angle, φ represents the roll angle, ψ represents the yaw angle, p represents the roll angular velocity, r represents the yaw angular velocity, C y represents the lateral force coefficient, C l Denotes the rolling moment coefficient, C n represents the yaw moment coefficient, is the roll angular acceleration, represents the yaw angular acceleration, a y It represents the acceleration component of the vehicle in the y direction under the body axis system.

4. The method according to claim 1, characterized in that: In step 1.2, the root mean square propagation algorithm is used to train the GRU neural network prediction model, and the neural network prediction accuracy is evaluated by the root mean square error and the coefficient of determination.

5. The method according to claim 1, characterized in that: In step 2.1, if the aircraft performs longitudinal maneuvers at a small angle of attack, the linear longitudinal aerodynamic model is selected; when the aircraft performs longitudinal maneuvers at a large angle of attack, the nonlinear longitudinal aerodynamic model with time effect is selected; if the aircraft performs lateral heading maneuvers, the lateral heading aerodynamic model is selected.

6. The method according to claim 5, characterized in that: In step 2.1, for the longitudinal aerodynamic model, the aerodynamic parameter vector to be identified is Θ 纵向 =[C D0 ,C L0 ,C Lα ,C Lq ,C Lδe ,C m0 ,C mα ,C mq ,C mδe ] T For the lateral aerodynamic model, the aerodynamic parameter vector to be identified is Θ 横航向 =[C y0 ,C yβ ,C yr ,C yp ,C yδr ,C l0 ,C lβ ,C lr ,C lp , In the formula, C ·0 represents the aerodynamic coefficient deviation term, C ·α and C ·β represents the aerodynamic derivative of static stability, C ·p , C ·q and C ·r represents the dynamic aerodynamic derivative term, and represents the static aerodynamic derivative term related to the control surface deflection, and · represents the symbols D, L, m, y, l and n.

Images