A hypersonic vehicle dynamics correction method based on neural network

Abstract

The present invention discloses a hypersonic aircraft dynamics correction method based on a neural network, which belongs to the field of aircraft dynamics technology. The method solves the problem of low correction accuracy of traditional hypersonic aircraft dynamics correction methods in the prior art; the present invention includes the following steps: S1. Constructing an ideal model of hypersonic aircraft dynamics according to the characteristics of the hypersonic aircraft and actual control requirements; S2. Setting training samples based on actual data obtained from experiments and predicted data obtained from the ideal model of hypersonic aircraft dynamics, and constructing a correction relationship between the input and output of the ideal model of hypersonic aircraft dynamics; S3. Constructing a neural network structure, using the neural network structure to train the ideal model of hypersonic aircraft dynamics, inputting training samples, and obtaining an optimized correction relationship. The present invention improves the correction accuracy of the ideal model of hypersonic aircraft dynamics and can be applied to the correction of aircraft state transformations.

Description

Technical Field

[0001] The present invention relates to a hypersonic aircraft dynamics correction method, in particular to a hypersonic aircraft dynamics correction method based on a neural network, and belongs to the technical field of aircraft dynamics. Background Art

[0002] The dynamic model of a hypersonic aircraft is an important basis for achieving high-precision hypersonic aircraft control, trajectory planning and guidance. However, there are often deviations between the artificially constructed dynamic model of a hypersonic aircraft and the actual model. For aircraft that are often in a highly dynamic physical environment, their dynamic response is more sensitive to deviations, that is, even a small deviation may bring about a large disturbance, which will have an adverse impact on the planning and control of the hypersonic aircraft.

[0003] In the existing technology, the accuracy of dynamic model correction is related to the way of processing experimental data. Relevant practitioners often conduct corresponding experiments to correct the dynamic model. The technical challenges of the dynamic correction method of hypersonic aircraft are as follows: (1) Complex flow conditions: Hypersonic aircraft face extremely complex flow conditions, such as shock waves, boundary layer separation, etc. The above characteristics require the dynamic correction method to consider more nonlinear and unsteady effects; (2) Difficulty in accurate modeling: Accurate simulation of aerodynamic and thermodynamic effects at hypersonic speeds is the prerequisite for accurate modeling. The current correction method has the problem of insufficient accuracy in modeling, especially the limited prediction accuracy under extreme conditions; (3) Complexity of coupling effects: The dynamic correction of hypersonic aircraft not only involves aerodynamics, but also needs to consider thermal effects, structural dynamics and coupling effects of control systems, which increases the complexity of system design and analysis; (4) Difficulty in experimental verification: Since the flow and thermodynamic effects under hypersonic conditions are difficult to fully simulate in ground laboratories, verifying the reliability and accuracy of the dynamic correction method usually requires large-scale numerical simulation and wind tunnel experiments.

[0004] In summary, a hypersonic vehicle dynamics correction method with high correction accuracy is needed. Summary of the Invention

[0005] A brief overview of the present invention is provided below to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify key or important aspects of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is simply to present certain concepts in a simplified form as a prelude to the more detailed description discussed later.

[0006] In view of this, in order to solve the problem of low correction accuracy of traditional hypersonic aircraft dynamics correction methods in the prior art, the present invention provides a hypersonic aircraft dynamics correction method based on a neural network.

[0007] The technical solution is as follows: A method for correcting hypersonic vehicle dynamics based on a neural network, specifically comprising the following steps:

[0008] S1. Construct an ideal dynamics model of a hypersonic vehicle based on its characteristics and actual control requirements.

[0009] S2. Based on the actual data obtained from the experiment and the predicted data obtained from the ideal hypersonic vehicle dynamics model, set the input of the ideal hypersonic vehicle dynamics model as training samples and construct a correction relationship between its input and output;

[0010] S3. Construct a neural network structure, use the neural network structure to train the ideal model of hypersonic aircraft dynamics, input training samples, and obtain the optimized correction relationship.

[0011] Furthermore, in the above S1, it is assumed that the hypersonic aircraft does not have sideslip and the aerodynamic drag coefficient C D and the aerodynamic lift coefficient C L Expressed as:

[0012] C D =d 00 +d 10 α+d 01 ma+d 11 α·ma+d 20 α 2 +d 02 ma 2

[0013] C L =l 00 +l 10 α+l 01 ma+l 11 α·ma+l 20 α 2 +l 02 ma 2

[0014] Where α is the angle of attack, ma is the Mach number, and d ij is the characterization parameter of the aerodynamic drag coefficient under the polynomial characterization method, l ij is the characterization parameter of the aerodynamic lift coefficient under the polynomial characterization method;

[0015] Considering the Earth's oblateness, rotation and gravitational perturbation, an ideal dynamic model M of a hypersonic vehicle is constructed.

[0016] The ideal dynamic model M of hypersonic aircraft is expressed as:

[0017]

[0018] in, is the dimensionless geocentric distance, R0 is the average radius of the Earth, which represents the dimensionless distance from the center of the Earth The dimensionless parameter of , r is the distance from the center of the earth, θ is the longitude, φ is the latitude, is the dimensionless velocity of the aircraft relative to the earth, g0 is the gravitational acceleration at sea level, γ is the path angle, ψ is the heading angle, is the dimensionless Earth rotation angular velocity, σ is the aircraft roll angle, is the dimensionless lift, is the dimensionless drag acceleration.

[0019] Furthermore, in S3, according to the actual experimental conditions, time t k-1 The actual state quantity is x k-1 , time t k-1 The actual state change in the time interval Δt is Δx k-1,real , time t k The actual state quantity is x k,real , time t k The actual state change in the time interval Δt is Δx k,real , for the ideal model M' of hypersonic vehicle dynamics, at time t k-1 The predicted state change within the time interval Δt is Δx k-1,pre , time t k The predicted state change within the time interval Δt is Δx k,pre , time t k The predicted state change Δx k,pre and the actual state change Δx k,real The deviation is ΔX k ;

[0020] The state change amount includes speed V, path angle γ and heading angle ψ;

[0021] Set time t k-1 The actual state x k-1 , actual state change Δx k-1,real and predicted state change Δx k-1,pre As the input of the ideal model M of hypersonic vehicle dynamics, that is, the training sample, the deviation ΔX k As the output of the hypersonic vehicle dynamics model M;

[0022] The input and output of the hypersonic vehicle dynamics model M are expressed as:

[0023] input:Δx k-1,real ,Δx k-1,pre ,x k-1

[0024] output:ΔX k =Δx k,real -Δx k,pre

[0025] Construct the correction relationship f between the input and output of the hypersonic aircraft dynamics model M * ;

[0026]

[0027] Furthermore, in said S4, the neural network structure includes an input layer F input , fully connected layer F fcn , Tanh activation layer F tan , LayerNorm normalization layer F layernorm , Relu activation layer F relu and the output layer F output ;

[0028] In the input layer F of the neural network structure input Input training samples, pass through the fully connected layer F fcn , Tanh activation layer F tan and LayerNorm normalization layer F layernorm After that, it passes through the fully connected layer F fcn and Relu activation layer F relu The cycle is optimized three times and finally passes through the fully connected layer F tan Then the output layer F output Output optimized correction relationship f * ;

[0029] The process of building a neural network structure is expressed as:

[0030] F input (x) = W input x+b input

[0031] F fcn (x) = W fcn x+b fcn

[0032] F output (x) = W output x+b output

[0033]

[0034] F relu (x)=max(0,x)

[0035]

[0036] μ=mean(x),σ=var(x)

[0037] Among them, W input is the input layer F input The weight, W fcn is the fully connected layer F fcn The weight, W output is the output layer F output The weight of b input is the input layer F input The bias coefficient, b fcn is the fully connected layer F fcn The bias coefficient, b output is the output layer F output Bias coefficient, γ' is the LayerNorm normalization layer F layernorm The scaling factor, γ'=1, β is the LayerNorm normalization layer F layernorm The offset coefficient, β = 0, μ is the mean of the input x, mean is the mean function, σ is the variance of the input x, var is the variance function, max is the maximum value function, ∈ is a small amount to maintain numerical stability;

[0038] The input layer F input The sample nodes are determined by the input sample dimension, which is 24-dimensional. The 24-dimensional dimension is determined by the six states of the ideal model of hypersonic vehicle dynamics and the time t k-1 The actual state x k-1 , actual state change Δx k-1,real and predicted state change Δx k-1,pre Composition, fully connected layer F fcn It is a network structure in which each node is connected to the next layer of nodes. The output layer F output The sample nodes are determined by the output sample dimension, which is 6 dimensions.

[0039] The beneficial effects of the present invention are as follows: the present invention first establishes a three-degree-of-freedom dynamic model of a hypersonic aircraft, namely, an ideal model of hypersonic aircraft dynamics, and then obtains real data of the aircraft through experiments, and compares the real data with the output data of the ideal model of hypersonic aircraft dynamics, namely, the state change quantity. Finally, the deviation relationship between the experimental data and the model data is mined using an independently constructed neural network structure, and the ideal model of hypersonic aircraft dynamics is corrected. The output of the hypersonic aircraft dynamics model, namely, the corrected state change quantity, can be obtained through the correction relationship, thereby realizing precise hypersonic aircraft dynamics correction; the present invention further verifies the effectiveness of the correction by constructing an artificial model M2 of hypersonic aircraft dynamics for numerical simulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 The figure is a flowchart of a hypersonic vehicle dynamics correction method based on a neural network;

[0042] Figure 2 Schematic diagram of neural network structure;

[0043] Figure 3 Schematic diagram of training loss;

[0044] Figure 4 Schematic diagram of verification loss;

[0045] Figure 5 This is a schematic diagram of speed correction comparison;

[0046] Figure 6 This is a schematic diagram of path angle correction comparison;

[0047] Figure 7 This is a schematic diagram of the comparison of heading angle correction;

[0048] Figure 8 It is a schematic diagram of speed deviation;

[0049] Figure 9 is a schematic diagram of path angle deviation;

[0050] Figure 10 Schematic diagram of heading angle deviation. DETAILED DESCRIPTION

[0051] To make the technical solutions and advantages of the embodiments of the present invention more clearly understood, exemplary embodiments of the present invention are further described in detail below with reference to the accompanying drawings. It should be noted that the embodiments described are only a portion of the embodiments of the present invention, and are not an exhaustive list of all embodiments. It should be noted that the embodiments of the present invention and the features thereof may be combined with each other unless they conflict.

[0052] refer to Figures 1-10 The present embodiment is described in detail. A method for correcting hypersonic vehicle dynamics based on a neural network specifically includes the following steps:

[0053] S1. Construct an ideal dynamics model of a hypersonic vehicle based on its characteristics and actual control requirements.

[0054] S2. Based on the actual data obtained from the experiment and the predicted data obtained from the ideal hypersonic vehicle dynamics model, set the input of the ideal hypersonic vehicle dynamics model as training samples and construct a correction relationship between its input and output;

[0055] S3. Construct a neural network structure, use the neural network structure to train the ideal model of hypersonic aircraft dynamics, input training samples, and obtain the optimized correction relationship.

[0056] Furthermore, in the above S1, it is assumed that the hypersonic aircraft does not have sideslip and the aerodynamic drag coefficient C D and the aerodynamic lift coefficient C L Expressed as:

[0057] C D =d 00 +d 10 α+d 01 ma+d 11 α·ma+d 20 α 2 +d 02 ma 2

[0058] C L =l 00 +l 10 α+l 01 ma+l 11 α·ma+l 20 α 2 +l 02 ma 2

[0059] Where α is the angle of attack, ma is the Mach number, and d ij is the characterization parameter of the aerodynamic drag coefficient under the polynomial characterization method, l ijis the characterization parameter of the aerodynamic lift coefficient under the polynomial characterization method;

[0060] Considering the Earth's oblateness, rotation and gravitational perturbation, an ideal dynamic model M of a hypersonic vehicle is constructed.

[0061] The ideal dynamic model M of hypersonic aircraft is expressed as:

[0062]

[0063] in, is the dimensionless geocentric distance, R0 is the average radius of the Earth, which represents the dimensionless distance from the center of the Earth The dimensionless parameter of , r is the distance from the center of the earth, θ is the longitude, φ is the latitude, is the dimensionless velocity of the aircraft relative to the earth, g0 is the gravitational acceleration at sea level, γ is the path angle, ψ is the heading angle, is the dimensionless Earth rotation angular velocity, σ is the aircraft roll angle, is the dimensionless lift, is the dimensionless drag acceleration.

[0064] Furthermore, in S3, according to the actual experimental conditions, time t k-1 The actual state quantity is x k-1 , time t k-1 The actual state change in the time interval Δt is Δx k-1,real , time t k The actual state quantity is x k,real , time t k The actual state change in the time interval Δt is Δx k,real , for the ideal model M' of hypersonic vehicle dynamics, at time t k-1 The predicted state change within the time interval Δt is Δx k-1,pre , time t k The predicted state change within the time interval Δt is Δx k,pre , time t k The predicted state change Δx k,pre and the actual state change Δx k,real The deviation is ΔX k ;

[0065] The state change amount includes speed V, path angle γ and heading angle ψ;

[0066] Set time t k-1 The actual state x k-1 , actual state change Δx k-1,real and predicted state change Δxk-1,pre As the input of the ideal model M of hypersonic vehicle dynamics, that is, the training sample, the deviation ΔX k As the output of the hypersonic vehicle dynamics model M;

[0067] The input and output of the hypersonic vehicle dynamics model M are expressed as:

[0068] input:Δx k-1,real ,Δx k-1,pre ,x k-1

[0069] output:ΔX k =Δx k,real -Δx k,pre

[0070] Construct the correction relationship f between the input and output of the hypersonic aircraft dynamics model M * ;

[0071]

[0072] Furthermore, in said S4, the neural network structure includes an input layer F input , fully connected layer F fcn , Tanh activation layer F tan , LayerNorm normalization layer F layernorm , Relu activation layer F relu and the output layer F output ;

[0073] In the input layer F of the neural network structure input Input training samples, pass through the fully connected layer F fcn , Tanh activation layer F tan and LayerNorm normalization layer F layernorm After that, it passes through the fully connected layer F fcn and Relu activation layer F relu The cycle is optimized three times and finally passes through the fully connected layer F fcn Then the output layer F output Output optimized correction relationship f * ;

[0074] The process of building a neural network structure is expressed as:

[0075] F input (x) = W input x+b input

[0076] F fcn (x) = W fcn x+bfcn

[0077] F output (x) = W output x+b output

[0078]

[0079] F relu (x)=max(0,x)

[0080]

[0081] μ=mean(x),σ=var(x)

[0082] Among them, W input is the input layer F input The weight, W fcn is the fully connected layer F fcn The weight, W output The output layer F output The weight of b input is the input layer F input The bias coefficient, b fcn is the fully connected layer F fcn The bias coefficient, b output is the output layer F output Bias coefficient, γ' is the LayerNorm normalization layer F layernorm The scaling factor, γ'=1, β is the LayerNorm normalization layer F layernorm The offset coefficient, β = 0, μ is the mean of the input x, mean is the mean function, σ is the variance of the input x, var is the variance function, max is the maximum value function, ∈ is a small amount to maintain numerical stability;

[0083] The input layer F input The sample nodes are determined by the input sample dimension, which is 24-dimensional. 24-dimensional is determined by the dimensionless distance from the center of the earth of the ideal model of hypersonic vehicle dynamics. Longitude θ, latitude φ, dimensionless speed Path angle γ, heading angle ψ and time t k-1 The actual state x k-1 , actual state change Δx k-1,real and predicted state change Δx k-1,pre Composition, fully connected layer F fcn It is a network structure in which each node is connected to the next layer of nodes. The output layer F output The sample nodes are determined by the output sample dimension, which is 6 dimensions.

[0084] Specifically, in this embodiment, the hyperparameters of the neural network structure include a learning rate of 0.0001, an optimization algorithm using the Adam optimizer, 500 hidden layer nodes, where the hidden layer is all fully connected layers except the input layer and the output layer, a batch size of 256, a training sample size of 10,000, and a training iteration number of 100.

[0085] The ideal hypersonic aircraft dynamics model M constructed in step S1 is used as the actual hypersonic aircraft dynamics model M1, and the state quantity and state quantity transformation quantity returned by the actual hypersonic aircraft dynamics model M1 are set to the actual data obtained from the experiment;

[0086] Construct an artificial model M2 of hypersonic aircraft dynamics, which does not consider the Earth's rotation and gravitational perturbations. The aerodynamic lift coefficient is 0.8 times the nominal value of the actual model M1 of hypersonic aircraft dynamics, and the aerodynamic drag coefficient is 1.3 times the nominal value of the actual model M1 of hypersonic aircraft dynamics. Due to the influence of mass deviation, the velocity increment δV(mass), path angle increment δγ(mass), and heading angle increment δψ(mass) of the artificial model M2 of hypersonic aircraft dynamics are increased by an offset of 0.1 based on the actual model M1 of hypersonic aircraft dynamics.

[0087] The artificial model M2 of hypersonic vehicle dynamics is expressed as:

[0088]

[0089] The present invention is used to correct the artificial model M2 of hypersonic aircraft dynamics. Since the deviation of the artificial model M2 of hypersonic aircraft dynamics from the actual model M1 of hypersonic aircraft dynamics only affects the state change of the velocity V, the path angle γ, and the heading angle ψ, only the above three quantities are analyzed. The present invention can effectively correct the state change of the artificial model M2 of hypersonic aircraft dynamics, and the data of the corrected artificial model M2 of hypersonic aircraft dynamics is extremely close to that of the original dynamic model.

[0090] refer to Figure 3 , the training loss result is obtained by training the training samples in step S2 through the neural network structure;

[0091] refer to Figure 4 ,The validation loss result is obtained by training the neural network structure using the test set constructed by real-time experimental data;

[0092] refer to Figure 5 , V real is the velocity of the actual model M1 of the hypersonic aircraft dynamics, V fakeis the predicted speed after neural network correction, V fakewithoutoffset is the predicted speed without neural network correction;

[0093] refer to Figure 6 , γ real is the path angle of the actual hypersonic vehicle dynamics model M1, γ fake is the predicted path angle after neural network correction, γ fakewithoutoffset is the predicted path angle without neural network correction;

[0094] refer to Figure 7 , ψ real is the heading angle of the actual model M1 of the hypersonic vehicle dynamics, ψ fake is the predicted heading angle after neural network correction, ψ fakewithoutoffset is the predicted heading angle without neural network correction;

[0095] refer to Figure 8 , V bias withoutoffset is the uncorrected velocity of the artificial model M2 of the hypersonic aircraft dynamics, V bias withoffset is the corrected velocity of the artificial model M2 of the hypersonic vehicle dynamics;

[0096] refer to Figure 9 , γ bias Withoutoffset is the uncorrected path angle of the artificial model M2 of the hypersonic vehicle dynamics, γ bias withoffsetThe corrected path angle of the artificial model M2 of hypersonic vehicle dynamics;

[0097] refer to Figure 10 , ψ bias withoutoffset is the uncorrected heading angle of the artificial model M2 of the hypersonic vehicle dynamics, ψ bias withoffset is the corrected heading angle of the hypersonic aircraft dynamics artificial model M2.

[0098] Although the present invention has been described with respect to a limited number of embodiments, it will be apparent to those skilled in the art, having benefit of the foregoing description, that other embodiments are contemplated within the scope of the invention thus described. Furthermore, it should be noted that the language used in this specification has been selected primarily for readability and didactic purposes, rather than for the purpose of explaining or limiting the subject matter of the present invention. Consequently, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is intended to be illustrative rather than restrictive of the scope of the invention, which is defined by the appended claims.

Claims

1. A hypersonic vehicle dynamics correction method based on a neural network, characterized in that: The following steps are involved: S1. Construct an ideal dynamics model of a hypersonic vehicle based on its characteristics and actual control requirements. S2. Based on the actual data obtained from the experiment and the predicted data obtained from the ideal hypersonic vehicle dynamics model, set the input of the ideal hypersonic vehicle dynamics model as training samples and construct a correction relationship between its input and output; S3. Construct a neural network structure, use the neural network structure to train the ideal model of hypersonic vehicle dynamics, input training samples, and obtain an optimized correction relationship; In the above S1, it is assumed that the hypersonic aircraft does not have sideslip and the aerodynamic drag coefficient and aerodynamic lift coefficient Expressed as: ； ； in, is the angle of attack, is the Mach number, is the characterization parameter of the aerodynamic drag coefficient under the polynomial characterization method, is the characterization parameter of the aerodynamic lift coefficient under the polynomial characterization method; Construct an ideal model of hypersonic vehicle dynamics by taking into account the Earth's oblateness, rotation, and gravitational perturbations ; Ideal model of hypersonic vehicle dynamics Expressed as: ； ； ； ； ； in, is the dimensionless geocentric distance, , is the average radius of the Earth, which represents the dimensionless distance from the center of the Earth The dimensionless parameter of is the distance from the center of the Earth, is the longitude, is the latitude, is the dimensionless velocity of the spacecraft relative to the earth, is the acceleration due to gravity at sea level, is the path angle, is the heading angle, is the dimensionless angular velocity of the Earth's rotation, is the aircraft roll angle, is the dimensionless lift, is the dimensionless drag acceleration.

2. The method for correcting hypersonic vehicle dynamics based on a neural network according to claim 1, characterized in that: In S3, according to the actual experimental conditions, the time The actual state quantity is ,time In time interval The actual state change in ,time The actual state quantity is ,time In time interval The actual state change in , for the ideal model of hypersonic vehicle dynamics ,time In time interval The predicted state change within is ,time In time interval The predicted state change within is ,time The predicted state change and the actual state change The deviation is ; The state change amount includes speed , path angle and heading angle ; The moment The actual state of , actual state change and predicted state change As an ideal model of hypersonic vehicle dynamics The input is the training sample, and the deviation As a dynamic model of hypersonic vehicles Output; Hypersonic vehicle dynamics model The input and output are expressed as: ； Constructing a hypersonic vehicle dynamics model The modified relationship between input and output ; 。 3. The method for hypersonic vehicle dynamics correction based on neural network according to claim 2, characterized in that: The neural network structure includes the input layer , fully connected layer , Tanh activation layer , LayerNorm normalization layer , Relu activation layer and output layer ; In the input layer of the neural network structure Input training samples, through the fully connected layer , Tanh activation layer and LayerNorm normalization layer After that, it passes through the fully connected layer and Relu activation layer The loop is optimized three times and finally passes through the fully connected layer Then the output layer Output optimized correction relationship ; The process of building a neural network structure is expressed as: ； ； ； ； ； ； ； in, For the input layer The weight of is a fully connected layer The weight of For the output layer The weight of For the input layer The bias coefficient of is a fully connected layer The bias coefficient of For the output layer The bias coefficient of LayerNorm normalization layer The scaling factor, , LayerNorm normalization layer The offset coefficient, , For input The mean of To find the mean function, For input The variance of To obtain the variance function, To find the maximum function, is a small amount to maintain numerical stability; The input layer The sample nodes are determined by the input sample dimension, which is 24-dimensional. The 24-dimensional dimension is determined by the 6 states and moments of the ideal model of hypersonic vehicle dynamics. The actual state of , actual state change and predicted state change Composition, fully connected layer For a network structure where each node is connected to the next layer of nodes, the output layer The sample nodes are determined by the output sample dimension, which is 6 dimensions.

Images