Aircraft aerodynamic parameter online identification method of RBF training neural network
By combining RBF-trained neural networks and extended Kalman filters, real-time identification of aircraft aerodynamic parameters is achieved, solving the problem of insufficient robustness caused by the uncertainty of aerodynamic parameters in the aircraft control system, improving the accuracy and efficiency of aerodynamic parameter identification, and enhancing the accuracy and efficiency of aircraft attitude control.
Patent Information
- Application Number
- CN202510646552.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-20
- Publication Date
- 2025-10-17
AI Technical Summary
Existing aircraft control systems are unable to handle the uncertainties in aerodynamic parameters during flight in real time, resulting in insufficient robustness and difficulty in meeting practical requirements.
By employing RBF training of a neural network combined with the extended Kalman filter method, the aerodynamic parameters of the aircraft are identified in real time through online estimation and updating of the neural network weights. The extended Kalman filter is used to perform state prediction and measurement updates on the BP neural network to obtain accurate aerodynamic parameters.
It improves the accuracy and efficiency of online identification of aerodynamic parameters of aircraft, reduces reliance on historical data, achieves higher convergence speed and more accurate aerodynamic model approximation, and enhances the accuracy and efficiency of aircraft attitude control.
Smart Images
Figure CN120805636A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to an aerodynamic parameter online identification method based on an extended Kalman filter training BP neural network, and belongs to the field of aircraft parameter identification. BACKGROUND
[0002] A control system of an aircraft is usually designed based on a pre-established mathematical model of a controlled object, and the accuracy of the established mathematical model will directly affect the performance of the control system, and especially the understanding of the uncertain parameters of the model of the controlled object determines the robustness of the control system. Since there are parameters that cannot be directly measured in the actual flight process, and some parameters change dramatically and exceed the expectation along with the change of the flight environment, the structure of the aircraft and the flight state, the guidance control system designed based on the standard aerodynamic model may not meet the actual requirements, and the uncertain influence generated in the flight process needs to be solved, therefore, the parameter online identification technology is required to obtain a more accurate aerodynamic model in real time, so as to ensure that the control system can be adjusted in time. SUMMARY
[0003] The application aims to provide an aircraft aerodynamic parameter online identification method of RBF training neural network, the neural network weights in the BP neural network are estimated and updated in real time through the extended Kalman filter, the real-time aerodynamic parameter identification result is obtained, and the accuracy and efficiency of the parameter online identification are improved.
[0004] The application is achieved by the following technical scheme:
[0005] The application discloses an aircraft aerodynamic parameter online identification method of RBF training neural network, and comprises the following steps:
[0006] S1, a three-layer BP neural network is established based on the nonlinear mapping relationship between the input and output parameters of the aircraft aerodynamic model.
[0007] S2, the weight matrix and the threshold matrix of the BP neural network are selected and rearranged into vectors as system state variables to construct a state equation.
[0008] S3, a six-degree-of-freedom aircraft motion model is established.
[0009] S4, the attack angle, the sideslip angle, the velocity inclination angle and the angular velocities of the three channels of the pitch, the yaw and the roll of the aircraft in the S3 model are taken as observation values of the extended Kalman filter.
[0010] S5, an observation equation containing observation noise is established according to the observation values of S4.
[0011] S6, the state equation of S2 and the observation equation of S5 are used to perform t k~t k+1 The state prediction at the moment is updated according to the prediction result. The updated state variable is obtained, and the aerodynamic parameters of the aircraft are identified online.
[0012] S7, the state variable obtained by S6 is combined with the aircraft state result to obtain new aerodynamic parameters C L , C N , C mx , C my and C mz The aerodynamic parameter identification is completed, wherein the parameters are lift coefficient, drag coefficient, roll moment coefficient, yaw moment coefficient and pitch moment coefficient, that is, the online identification of the aerodynamic parameters of the aircraft is realized.
[0013] Further, S8, the obtained identification parameters are used to control the attitude of the aircraft, and the accuracy and efficiency of the attitude control of the aircraft are improved.
[0014] Further, the input-output relationship of the three-layer BP neural network established in S1 is:
[0015]
[0016] Wherein, H is the hidden layer neuron, g(·) is the activation function, I is the input layer neuron, is the weight matrix from the input layer to the hidden layer, B [1] is the threshold matrix of the hidden layer. Y is the output layer neuron, is the weight matrix from the hidden layer to the output layer, B [2] is the threshold matrix of the output layer.
[0017] Further, the system state variable X in S2 is represented as:
[0018] X=[W1,W2,B1,B2] T
[0019] Wherein, W1, W2, B1, B2 are represented as follows:
[0020]
[0021] The superscript indicates the value in the weight matrix of the corresponding layer, and the subscript indicates the corresponding neuron position of the front and rear layers. The superscript indicates the value in the threshold matrix of the corresponding layer, and the subscript indicates the corresponding neuron position of the layer
[0022] Further, the system state equation in S2 is:
[0023]
[0024] where W t is the system noise vector, f(X,t) = 0 if the neural network parameters tend to be stable.
[0025] Further, the 6-DOF aircraft motion model described in S3 is:
[0026]
[0027] where,
[0028]
[0029] where: a, b, g are the angle of attack, side slip angle and angle of velocity of the aircraft, m, g, V, q are the mass, gravity coefficient, velocity and angle of trajectory of the aircraft, w x , w y , w z are the roll, yaw and pitch angular velocities of the aircraft, J x , J y , J z are the principal moments of inertia of the aircraft about x, y, z axes, respectively. J xy , J xz are the products of inertia of the aircraft about x, y axes. The aircraft is a plane-symmetric configuration, so the products of inertia of the aircraft about x, z axes and about y, z axes are zero, i.e. J yz = 0. L, N are the aerodynamic lift and drag forces, where is the dynamic pressure, p is the atmospheric density, S ref is the characteristic area of the aircraft. M x , M y , M z are the roll, yaw and pitch moments, respectively, which are expressed as where L ref is the characteristic length of the aircraft. C L , C N , C mx , C my and C mz are functions related to the Mach number Ma, the angle of attack a, the side slip angle b and the equivalent rudder deflection angle.
[0030] Further, the angle of attack, the side slip angle, the angle of velocity, and the angular velocities of the pitch, yaw and roll channels are taken as the observation values of the extended Kalman filter in S4, i.e.
[0031] Z = [a, b, g, w x , w y , w z ] T
[0032] Furthermore, the observation equation in S5 is
[0033] Z=h(X,t)+V t
[0034] Where V t is the observation noise vector. h(X,t) is the calculation function that obtains the observation value through the state variable at time t.
[0035] Furthermore, the implementation method in S6 is:
[0036] The t k ~t k+1 The state prediction at the moment is expressed as:
[0037]
[0038] The subscript k / k in the above formula represents t k The final calculated value at time k+1 / k represents the value obtained by using t k The value of t predicted by k+1 The value at a moment, with only one subscript, indicates that the variable takes the value at the corresponding moment.
[0039] The variance matrix prediction is expressed as:
[0040]
[0041] Among them, t k ~t k+1 The state transition matrix is:
[0042]
[0043] Where Δt is the filtering step size. k is the system process noise variance matrix. I is the identity matrix, and the Jacobian matrix of the state equation is:
[0044]
[0045] Measurement Update:
[0046] State Estimation:
[0047]
[0048] Variance matrix estimation:
[0049]
[0050] Filter gain matrix:
[0051]
[0052] where Rk+1 For observing the noise variance matrix, the Jacobian matrix of the observation equation is:
[0053]
[0054] Advantages:
[0055] 1. The RBF trained neural network aerodynamic parameter online identification method for a vehicle disclosed in the application uses the extended Kalman filter to update the neural network parameters, only needs the flight state in the real-time flight process as an observation quantity to obtain the identified aerodynamic parameters in combination with the estimated aerodynamic parameters, and improves the efficiency of the online identification of the aerodynamic parameters of the vehicle.
[0056] 2. The RBF trained neural network aerodynamic parameter online identification method for a vehicle disclosed in the application uses the extended Kalman filter algorithm to replace the process of correcting the parameters through the reverse calculation of the neural network, adjusts the real-time neural network model, and predicts and adjusts the model, and can obtain aerodynamic parameter identification values with higher precision compared with the single Kalman filter method.
[0057] 3. The RBF trained neural network aerodynamic parameter online identification method for a vehicle disclosed in the application does not need historical data as samples, effectively handles the dependence of the neural network training on a large amount of data samples, and improves the convergence speed and precision of the online identification of the aerodynamic parameters of the vehicle.
[0058] 4. The RBF trained neural network aerodynamic parameter online identification method for a vehicle disclosed in the application uses the observable flight data of the vehicle to correct the BP neural network weight and threshold value, makes the BP neural network model approach the actual aerodynamic model in real time, and obtains a more accurate aerodynamic model that can replace the original interpolation aerodynamic model of the vehicle. BRIEF DESCRIPTION OF DRAWINGS
[0059] Figure 1 A flowchart of the RBF trained neural network aerodynamic parameter online identification method for a vehicle is shown.
[0060] Figure 2 The prediction effect of the estimated values of the lift coefficient, the drag coefficient and the side force coefficient on the true values and the comparison curve with the ideal values in Example 1 are shown.
[0061] Figure 3 The prediction effect of the estimated values of the pitch moment coefficient, the roll moment coefficient and the yaw moment coefficient on the true values and the comparison curve with the ideal values in Example 1 are shown.
[0062] Figure 4 The prediction effect of the estimated values of the lift coefficient, the drag coefficient and the side force coefficient on the true values and the comparison curve with the ideal values in Comparative Example 1 are shown.
[0063] Figure 5 The figure shows the prediction effect of the estimated values of the pitch moment coefficient, the roll moment coefficient and the yaw moment coefficient in Comparative Example 1 on the true values and the comparison curve with the ideal values;
[0064] Figure 6 The figure shows the angle of attack, sideslip angle and velocity dip angle tracking curves obtained by using the identified aerodynamic coefficients for attitude control design. DETAILED DESCRIPTION
[0065] Example 1: Simulation experiment is carried out to identify the real-time aerodynamic parameters of the aircraft during flight, and the initial value of the aircraft attitude angle tracking control simulation in the attitude control design simulation process is set as: [a0, b0, g0, w x0 , w y0 , w z0 ] T = [15°, 2°, 0°, 0° / s, 0° / s, 0° / s] T . The aircraft control command is given as [a d , b d , g d ] T = [5°, 0°, 5°] T : The initial height and flight Mach number of the aircraft are h0=30km and V0=3500m / s respectively. The simulation time is 20s and the step size At=0.01s.
[0066] As shown in Figure 1 , the embodiment discloses an aircraft aerodynamic parameter online identification method of RBF training neural network, and the specific implementation steps are as follows,
[0067] S1, based on the nonlinear mapping relationship between the input and output parameters of the aircraft aerodynamic model, a three-layer BP neural network is established.
[0068] S2, the weight matrix and threshold matrix of the BP neural network are selected and rearranged as a vector to serve as a system state variable to construct a state equation.
[0069] S3, a 6-DOF aircraft motion model is established.
[0070] S4, the aircraft angle of attack, sideslip angle, velocity dip angle, and angular velocity of the three channels of pitch, yaw and roll in the S3 model are taken as the observation values of the extended Kalman filter.
[0071] S5, an observation equation containing observation noise is established according to the observation values of S4.
[0072] S6, based on the state equation of S2 and the observation equation of S5, the extended Kalman filter method is used to identify the aerodynamic parameters of the aircraft. k~t k+1 The state prediction at time t is updated according to the prediction result. The updated state variable is obtained.
[0073] S7, the state variable obtained in S6 is combined with the aircraft state result to obtain new aerodynamic parameters C L , C N , C mx , C my and C mz The aerodynamic parameter identification is completed, wherein the parameters are lift coefficient, drag coefficient, roll moment coefficient, yaw moment coefficient and pitch moment coefficient respectively.
[0074] It also includes S8, using the obtained identification parameters to design the aircraft attitude control.
[0075] The input-output relationship of the three-layer BP neural network established in S1 is:
[0076]
[0077] Wherein, H is the hidden layer neuron, g(·) is the activation function, I is the input layer neuron, is the weight matrix from the input layer to the hidden layer, B [1] is the threshold matrix of the hidden layer. Y is the output layer neuron, is the weight matrix from the hidden layer to the output layer, B [2] is the threshold matrix of the output layer.
[0078] The system state variable X in S2 is represented as:
[0079] X = [W1, W2, B1, B2] T
[0080] Wherein, W1, W2, B1, B2 represent as follows:
[0081]
[0082] The superscript indicates the value in the weight matrix of the corresponding layer, and the subscript indicates the corresponding neuron position of the front and rear layers. The superscript indicates the value in the threshold matrix of the corresponding layer, and the subscript indicates the corresponding neuron position of the layer
[0083] The system state equation in S2 is:
[0084]
[0085] Wherein W t is the system noise vector, and if the neural network parameters tend to be stable, f(X, t) = 0.
[0086] The 6-DOF aircraft motion model described in S3 is:
[0087]
[0088] where,
[0089]
[0090] where: α, β, γ are the angle of attack, angle of sideslip, and angle of velocity, respectively. m, g, V, θ are the mass of the aircraft, gravitational constant, velocity, and angle of trajectory, respectively. ω x , ω y , ω z are the roll, yaw, and pitch angular velocities of the aircraft, respectively. J x , J y , J z are the principal moments of inertia of the aircraft about the x, y, z axes, respectively. J xy is the product of inertia of the aircraft about the x, y axes. The aircraft is a planar symmetric configuration, so the product of inertia of the aircraft about the x, z axes and the product of inertia about the y, z axes are zero, i.e., J xz = J yz = 0. L, N are the aerodynamic lift and drag forces, respectively, where is the dynamic pressure, and ρ is the air density. S ref is the characteristic area of the aircraft. M x , M y , M z are the roll, yaw, and pitch moments, respectively, and are expressed as where L ref is the characteristic length of the aircraft. C L , C N , C mx , C my , and C mz are functions related to the Mach number Ma, angle of attack α, angle of sideslip β, and equivalent rudder angle δ.
[0091] In S4, the angle of attack, angle of sideslip, angle of velocity, and angular velocities of the three channels of pitch, yaw, and roll are taken as the observations of the extended Kalman filter, i.e.,
[0092] Z = [α, β, γ, ω x , ω y , ω z ] T
[0093] The observation equation in S5 is
[0094] Z = h(X, t) + V t
[0095] where V t is the observation noise vector. h(X, t) is the calculation function of the observation value through the state variable at time t.
[0096] The implementation method in S6 is,
[0097] The t k ~t k+1 state prediction is expressed as:
[0098]
[0099] The subscript k / k of the variable in the above formula indicates the final calculated value at t k , the subscript k+1 / k indicates the value predicted at t k using the value at t k+1 , and only one subscript indicates that the variable takes the value at the corresponding time.
[0100] The variance matrix prediction is expressed as:
[0101]
[0102] where the state transition matrix from t k ~t k+1 is:
[0103]
[0104] where Δt is the filtering step length. Q k is the system process noise variance matrix. I is the unit matrix, and the Jacobian matrix of the state equation is:
[0105]
[0106] Measurement update:
[0107] State estimation:
[0108]
[0109] Variance matrix estimation:
[0110]
[0111] Filtering gain matrix:
[0112]
[0113] where R k+1 is the observation noise variance matrix, and the Jacobian matrix of the observation equation is:
[0114]
[0115] Finally, the real-time aerodynamic parameters obtained are used to design the attitude control of the aircraft.
[0116] The simulation results are shown in Figures 2-3 The figure shows that the estimated aerodynamic coefficients effectively eliminate the influence of aerodynamic bias and effectively estimate the true values, which are very close to the ideal values.
[0117] Comparative Example 1: The same simulation experiment as in Example 1 is performed, except that the state objects of the extended Kalman filter are directly selected as the aerodynamic coefficients, and the neural network is not used for training.
[0118] As shown in Figures 4-5 The figure shows that the estimated aerodynamic coefficients can still estimate the true values to a certain extent, but due to the lack of neural network approximation, there is a certain fluctuation and error under the influence of noise.
[0119] As shown in Figure 6 The figure shows that the aircraft attitude after the attitude control design using the aerodynamic parameters obtained by aerodynamic identification can achieve rapid tracking of the control command.
[0120] The above specific description further details the purpose, technical solutions and beneficial effects of the application. It should be understood that the above description is only a specific embodiment of the application and is not intended to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.
Claims
1. A method for online identification of aircraft aerodynamic parameters using an RBF trained neural network, characterized by: The following steps are included: S1. Based on the nonlinear mapping relationship between the input and output parameters of the aircraft aerodynamic model, a three-layer BP neural network is established; S2. Select the weight matrix and threshold matrix of the BP neural network, rearrange them into vectors as system state variables, and construct the state equation; S3, establishing a 6-DOF aircraft motion model; S4, using the aircraft's angle of attack, sideslip angle, velocity inclination, and angular velocity of the three channels of pitch, yaw and roll in the S3 model as observation values of the extended Kalman filter; S5, establish the observation equation containing observation noise based on the observation value of S4; S6, based on the state equation of S2 and the observation equation of S5, the extended Kalman filter method is used to perform t k ~t k+1 The state prediction at the moment is performed, and the measurement is updated according to the prediction result; the updated state variables are obtained; S7, using the state variables obtained in S6 and combining them with the aircraft state results to obtain the new aerodynamic parameters C L 、C N 、C mx 、C my and C mz The aerodynamic parameter identification is completed, where the parameters are lift coefficient, drag coefficient, rolling moment coefficient, yaw moment coefficient and pitching moment coefficient, that is, the online identification of the aerodynamic parameters of the aircraft is realized.
2. The method according to claim 1 or 2, wherein: The method further includes S8, using the obtained identification parameters to perform attitude control of the aircraft to improve the accuracy and efficiency of the attitude control of the aircraft.
3. The method according to claim 1, wherein: The input-output relationship of the three-layer BP neural network established in S1 is: Among them, H is the hidden layer neuron, g(·) is the activation function which is the sigmoid function, I is the input layer neuron, is the weight matrix from the input layer to the hidden layer, B [1] is the threshold matrix of the hidden layer; Y is the output layer neuron, is the weight matrix from the hidden layer to the output layer, B [2] is the threshold matrix of the output layer.
4. The method according to claim 3, wherein: The system state variable X in S2 is expressed as: X=[W1,W2,B1,B2] T Where W1, W2, B1, and B2 are represented as follows: The superscript indicates the value in the weight matrix of the corresponding layer, and the subscript indicates the corresponding neuron position in the previous and next layers; The superscript represents the value in the threshold matrix of the corresponding layer, and the subscript represents the corresponding neuron position of the layer.
5. The method according to claim 4, wherein: The system state equation described in S2 is: Where W t is the system noise vector. If the neural network parameters tend to be stable, then f(X,t)=0.
6. The method according to claim 5, wherein: The 6-DOF aircraft motion model described in S3 is: in, Where: α, β, γ are the aircraft's angle of attack, sideslip angle, and velocity inclination, respectively; m, g, V, θ are the aircraft's mass, gravity coefficient, velocity, and trajectory inclination, respectively; ω x 、ω y 、ω z are the roll, yaw and pitch angular velocities of the aircraft, J x 、J y 、J z are the principal moments of inertia of the aircraft around the x, y, and z axes respectively; J xy is the inertia product of the aircraft about the x and y axes; the aircraft is a plane-symmetrical configuration, so the inertia product of the aircraft about the x and z axes and the inertia product about the y and z axes are zero, that is, J xz =J yz =0; L and N are aerodynamic lift and drag respectively, in is the dynamic pressure, ρ is the atmospheric density, S ref is the characteristic area of the aircraft; M x 、M y 、M z are the rolling moment, yaw moment and pitching moment respectively, expressed as Among them L ref is the characteristic length of the aircraft; C L 、C N 、C mx 、C my and C mz It is a function related to the Mach number Ma, angle of attack α, sideslip angle β and equivalent rudder deflection angle.
7. The method according to claim 6, wherein: In S4, the angle of attack, sideslip angle, velocity inclination, and angular velocity of the three channels of pitch, yaw and roll are used as the observation values of the extended Kalman filter, that is, Z=[α,β,γ,ω x ,oh y ,oh z ] T 8. The method according to claim 7, wherein: The observation equation described in S5 is Z=h(X,t)+V t Where V t is the observation noise vector; h(X,t) is the calculation function for obtaining the observation value through the state variable at time t.
9. The method according to claim 8, wherein: The implementation method in S6 is: The t k ~t k+1 The state prediction at the moment is expressed as: The subscript k / k of the variables in the above formula represents t k The final calculated value at time k+1 / k represents the value obtained by using t k The value of t predicted by k+1 The value at a moment, if there is only one subscript, it means that the variable takes the value at the corresponding moment; The variance matrix prediction is expressed as: Among them, t k ~t k+1 The state transition matrix is: Among them, Δt is the filtering step size; Q k is the system process noise variance matrix; I is the identity matrix, and the Jacobian matrix of the state equation is: Measurement Update: State Estimation: Variance matrix estimation: Filter gain matrix: where R k+1 is the observation noise variance matrix, and the Jacobian matrix of the observation equation is: