A predictive correction guidance method for a reentry vehicle based on neural network feedback correction

Abstract

The present invention discloses a predictive correction guidance method for a reentry vehicle based on neural network feedback correction, belonging to the technical field of guidance correction. It solves the problem that the actual effect of the traditional predictive correction guidance method for reentry vehicles in the prior art is poor and difficult to achieve the expectation; the present invention generates the reentry trajectory of the vehicle under aerodynamic disturbance and the predictive trajectory of the vehicle without aerodynamic disturbance through Monte Carlo shooting, and calculates and obtains the aerodynamic disturbance identification network data set; constructs the aerodynamic disturbance identification neural network structure, inputs the aerodynamic disturbance identification network data set into the aerodynamic disturbance identification neural network structure for training to obtain the trained neural network structure; uses the trained neural network structure to correct the aerodynamic model, that is, sets the aerodynamic model correction conditions and makes correction judgments to realize the predictive correction guidance of the vehicle. The present invention improves the accuracy of the predictive correction guidance of the reentry vehicle, avoids the loss of calculation efficiency, and can be applied to the correction of the aerodynamic model.

Description

Technical Field

[0001] The present invention relates to a predictive correction guidance method for a reentry vehicle, and particularly to a predictive correction guidance method for a reentry vehicle based on neural network feedback correction, belonging to the technical field of guidance correction. Background Art

[0002] The principle of the predictive correction guidance algorithm for a reentry vehicle is that the vehicle predicts the terminal state according to the current flight state with a certain guidance command, compares it with the expected terminal state bound by the vehicle to obtain the terminal state error, and corrects the guidance command through the terminal state error to generate the guidance command at each moment to guide the vehicle towards the expected terminal state; the predictive correction guidance algorithm does not depend on the nominal trajectory, only needs to generate the corresponding guidance command according to the terminal state error and combined with various path constraints and control quantity constraints, and at the same time it does not have the assumption of the small deviation theory, and can adjust the trajectory within a large range. The adaptability and robustness of the predictive correction guidance algorithm are relatively strong.

[0003] However, since the essence of predictive correction is to correct the guidance command through the error between the predicted terminal state and the expected terminal state, the accuracy of the predicted state is directly related to the final guidance accuracy. The flight airspace of the reentry vehicle has a large span and the flight speed range changes widely, and its aerodynamic model inevitably has a large deviation. The deviation of the aerodynamic model will cause a large difference between the predicted state and the actual flight state during the guidance process, especially in the dense atmosphere area with a lower airspace. Therefore, how to design and improve the predictive correction algorithm to improve the robustness of the predictive correction guidance algorithm to the perturbation of the aerodynamic model is the key technology that needs to be solved urgently.

[0004] In the prior art, drawing on the idea of the control system to eliminate the deviation according to the error feedback, the aerodynamic disturbance amount of the vehicle is identified through the deviation between the predicted state of the vehicle and the actual flight state, so as to correct the model used in the vehicle prediction, which can alleviate the problem of poor guidance accuracy caused by the non-uniform model to a certain extent, but the actual effect is poor and it is difficult to achieve the expected result.

[0005] In summary, a predictive correction guidance method for a reentry vehicle with better correction effect based on aerodynamic disturbance identification and improved robustness of the predictive correction guidance algorithm is needed. Summary of the Invention

[0006] A brief overview of the present invention is given 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 the key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Its purpose is only to present certain concepts in a simplified form as a prelude to a more detailed description to follow.

[0007] In view of this, in order to solve the problem that the actual effect of the traditional reentry vehicle prediction and correction guidance method in the prior art is poor and difficult to meet the expectations, the present invention provides a reentry vehicle prediction and correction guidance method based on neural network feedback correction.

[0008] The technical solution is as follows: A reentry vehicle prediction and correction guidance method based on neural network feedback correction specifically includes the following steps:

[0009] S1. Generate the reentry trajectory of the vehicle under aerodynamic disturbance and the predicted trajectory of the vehicle without aerodynamic disturbance through Monte Carlo shooting, and calculate to obtain the aerodynamic disturbance identification network data set;

[0010] S2. Construct the aerodynamic disturbance identification neural network structure, input the aerodynamic disturbance identification network data set into the aerodynamic disturbance identification neural network structure for training, and obtain the trained neural network structure;

[0011] S3. Use the trained neural network structure to correct the aerodynamic model, that is, set the aerodynamic model correction conditions, and perform correction judgment on the aerodynamic model in each correction instruction guidance cycle to achieve the prediction and correction guidance of the vehicle.

[0012] Further, in the S1, the aerodynamic disturbance identification network data set includes the aerodynamic disturbance coefficient deviation and the state quantity deviation;

[0013] Select the vehicle altitude h, vehicle speed v, flight path angle γ, and vehicle mechanical energy e as the predicted vehicle states. The actual flight state is (h1, v1, γ1, e1), where h1 is the actual vehicle altitude, v1 is the actual vehicle speed, γ1 is the actual flight path angle, and e1 is the actual vehicle mechanical energy. Calculate the state quantity deviation (δh, δv, δγ, δe) between the predicted vehicle state and the actual flight state;

[0014] The state quantity deviation (δh, δv, δγ, δe) is expressed as:

[0015]

[0016] Among them, δh is the altitude state quantity deviation, δv is the speed state quantity deviation, δγ is the path angle state quantity deviation, and δe is the mechanical energy state quantity deviation;

[0017] The vehicle mechanical energy e is expressed as:

[0018]

[0019] Among them, r is the current geocentric distance of the vehicle;

[0020] Select the aerodynamic drag coefficient perturbation deviation δC dAs the deviation of the aerodynamic disturbance coefficient.

[0021] Further, in the step S2, the aerodynamic disturbance identification neural network structure includes an input layer, two hidden layers and an output layer. L2 regularization is used to train the aerodynamic disturbance identification neural network to avoid overfitting. The state quantity deviation is used as the input, and the aerodynamic disturbance coefficient deviation is used as the output.

[0022] L2 regularization L L2 It is expressed as:

[0023]

[0024] L data = δC d - Φ(δh, δv, δγ, δe)

[0025]

[0026] Wherein, L data is the error between the neural network predicted value Φ(δh, δv, δγ, δe) and the aerodynamic disturbance coefficient deviation δC d , λ is the regularization parameter, is the square of the L2 norm of the neural network parameter set w, and w i is the neural network parameter;

[0027] The input layer includes 4 nodes, each hidden layer includes 10 nodes, and both hidden layers use the Relu activation function.

[0028] The Relu activation function f(x) is expressed as:

[0029] f(x) = max(0, x)

[0030] Wherein, max is the limiting function, and x represents the output data of the input layer or the first hidden layer.

[0031] Further, in the step S3, the correction judgment process is expressed as: if the aerodynamic model correction condition is satisfied, the aerodynamic disturbance evaluation result is obtained through the neural network structure according to the current state deviation amount to correct the aerodynamic model, and the corrected model is used for the prediction and correction guidance algorithm; if the aerodynamic model correction condition is not satisfied, the current aerodynamic model is continued to be used for the prediction and correction guidance algorithm.

[0032] The prediction and correction guidance algorithm is a guidance mode of longitudinal and lateral profile separation. The longitudinal profile guidance adopts the double correction logic of the angle of attack and the bank angle. The longitudinal profile adopts the bank angle correction strategy based on the range and the angle of attack profile correction strategy based on the altitude. The lateral profile adopts the bank angle corridor flipping strategy.

[0033] The prediction and correction guidance algorithm is expressed as:

[0034] minJ(Δs togo )→σ cmd

[0035] minJ(h)→(α max ,α min )

[0036]

[0037] where J(·) is the objective function, Δs togo is the remaining flight range to be flown, σ cmd is the bank angle command, Δψ is the heading angle deviation, Δψ threshold is the heading angle deviation threshold, sign(·) is the sign function used to obtain the positive or negative value of a variable, is the bank angle command for the k-th guidance cycle, is the bank angle command for the (k - 1)-th guidance cycle, α max is the maximum angle of attack, α min is the minimum angle of attack.

[0038] The beneficial effects of the present invention are as follows: The present invention constructs a neural network, whose input and output layers can meet the dimensional requirements of the aircraft state deviation and aerodynamic deviation. The re-entry trajectory of the aircraft under aerodynamic bias is generated by the Monte Carlo shooting method, and the corresponding predicted trajectory of the aircraft under the predictive correction guidance command is recorded. The aircraft state deviation is obtained by disturbing the re-entry trajectory and the predicted trajectory, so as to generate a data set for training the neural network to train and verify the network. The aerodynamic identification under small online computational load is achieved through offline neural network training; The trained neural network of the present invention is used in the predictive correction algorithm, the aerodynamic model correction condition is set, and the aerodynamic model correction is realized only under the condition of severe aerodynamic disturbance through the correction condition. When the non-corrected predictive correction guidance algorithm can meet the guidance accuracy, no additional aerodynamic model correction is performed, avoiding the loss of computational efficiency and further improving the correction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The schematic 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:

[0040] Figure 1 is a schematic flow chart of the predictive correction guidance method for a re-entry vehicle based on neural network aerodynamic evaluation;

[0041] Figure 2 is a schematic diagram of the neural network structure for aerodynamic disturbance identification;

[0042] Figure 3Schematic diagram of the process of an embodiment of the predictive correction guidance method;

[0043] Figure 4 Schematic diagram of the process of an embodiment of the longitudinal profile bank angle correction;

[0044] Figure 5 Schematic diagram of the influence of different aerodynamic disturbances on the altitude - speed profile of the aircraft;

[0045] Figure 6 Schematic diagram of the process of an embodiment of the predictive correction guidance method for a re - entry vehicle based on neural network aerodynamic evaluation;

[0046] Figure 7 Schematic diagram of the change of the loss function in the aerodynamic model training;

[0047] Figure 8 Schematic diagram of the altitude - speed profile result of the aircraft guidance;

[0048] Figure 9 Schematic diagram of the altitude result of the aircraft guidance;

[0049] Figure 10 Schematic diagram of the speed result of the aircraft guidance;

[0050] Figure 11 Schematic diagram of the altitude - range result of the aircraft guidance;

[0051] Figure 12 Schematic diagram of the result of the heading angle corridor flip of the aircraft guidance;

[0052] Figure 13 Schematic diagram of the result of the angle of attack correction of the aircraft guidance;

[0053] Figure 14 Schematic diagram of the result of the bank angle correction of the aircraft guidance. Detailed implementation mode

[0054] In order to make the technical solutions and advantages in the embodiments of the present invention clearer and more understandable, the following further details the exemplary embodiments of the present invention with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than an exhaustive list of all embodiments. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0055] Refer to Figures 1 - 14 This embodiment is described in detail. A predictive correction guidance method for a re - entry vehicle based on neural network feedback correction specifically includes the following steps:

[0056] S1. Generate the reentry trajectory of the aircraft under aerodynamic disturbances and the predicted trajectory of the aircraft without aerodynamic disturbances through the Monte Carlo shooting method, and calculate the aerodynamic disturbance identification network data set;

[0057] S2. Construct the aerodynamic disturbance identification neural network structure, input the aerodynamic disturbance identification network data set into the aerodynamic disturbance identification neural network structure for training, and obtain the trained neural network structure;

[0058] S3. Use the trained neural network structure to correct the aerodynamic model, that is, set the aerodynamic model correction conditions, and perform correction judgment on the aerodynamic model in each correction command guidance cycle to achieve aircraft prediction and correction guidance;

[0059] Further, in the S1, the aerodynamic disturbance identification network data set includes the aerodynamic disturbance coefficient deviation and the state quantity deviation;

[0060] Select the aircraft altitude h, aircraft speed v, flight path angle γ, and aircraft mechanical energy e as the predicted aircraft states. The actual flight state is (h1, v1, γ1, e1), where h1 is the actual aircraft altitude, v1 is the actual aircraft speed, γ1 is the actual flight path angle, and e1 is the actual aircraft mechanical energy. Calculate the state quantity deviation (δh, δv, δγ, δe) between the predicted aircraft state and the actual flight state;

[0061] The state quantity deviation (δh, δv, δγ, δe) is expressed as:

[0062]

[0063] Among them, δh is the altitude state quantity deviation, δv is the speed state quantity deviation, δγ is the path angle state quantity deviation, and δe is the mechanical energy state quantity deviation;

[0064] The aircraft mechanical energy e is expressed as:

[0065]

[0066] Among them, r is the current geocentric distance of the aircraft;

[0067] Select the aerodynamic drag coefficient perturbation deviation δC d as the aerodynamic disturbance coefficient deviation;

[0068] Specifically, in this embodiment, the pseudocode description of step S1 is as follows:

[0069] Begin Monte Carlo shooting

[0070] Define the number of shooting times num;

[0071] Select the initial value of the aircraft state and the typical control profile;

[0072] for i in num:

[0073] Integrate to generate a flight trajectory under pneumatic disturbances;

[0074] Integrate to generate a flight trajectory without pneumatic disturbances;

[0075] Calculate the terminal state deviation (δh, δv, δγ, δe) and the pneumatic pull deviation coefficient δC d

[0076] end

[0077] The terminal state deviation (δh, δv, δγ, δe) is the state quantity deviation, and the pneumatic pull deviation coefficient δC d is the pneumatic drag coefficient perturbation deviation.

[0078] Furthermore, in S2, the pneumatic disturbance identification neural network structure includes an input layer, two hidden layers, and an output layer. L2 regularization is used to train the pneumatic disturbance identification neural network to avoid overfitting. The state quantity deviation is used as the input, and the pneumatic disturbance coefficient deviation is used as the output;

[0079] L2 regularization L L2 is expressed as:

[0080]

[0081] L data = δC d - Φ(δh, δv, δγ, δe)

[0082]

[0083] where L data is the error between the neural network predicted value Φ(δh, δv, δγ, δe) and the pneumatic disturbance coefficient deviation δC d and λ is the regularization parameter, is the square of the L2 norm of the neural network parameter set w, and w i is the neural network parameter;

[0084] The input layer includes 4 nodes, each hidden layer includes 10 nodes, both hidden layers use the Relu activation function, and the output layer includes 1 node;

[0085] The Relu activation function f(x) is expressed as:

[0086] f(x) = max(0, x)

[0087] where max is the limiting function and x represents the output data of the input layer or the first hidden layer;

[0088] Reference Figure 2 , the Input Layer is the input layer, which represents 4 nodes, the Hidden Layer is the hidden layer, which represents 10 nodes, and the Output Layer is the output layer, which represents 1 node;

[0089] Furthermore, in the S3, the correction judgment process is expressed as: if the aerodynamic model correction condition is satisfied, the aerodynamic disturbance evaluation result is obtained through the neural network structure according to the current state deviation to correct the aerodynamic model, and the corrected model is used for the predictive correction guidance algorithm; if the aerodynamic model correction condition is not satisfied, the current aerodynamic model continues to be used for the predictive correction guidance algorithm;

[0090] The predictive correction guidance algorithm is a guidance mode of longitudinal and lateral profile separation. The longitudinal profile guidance adopts the double correction logic of angle of attack and bank angle. The longitudinal profile adopts the bank angle correction strategy based on range and the angle of attack profile correction strategy based on altitude. The lateral profile adopts the bank angle corridor flipping strategy;

[0091] The predictive correction guidance algorithm is expressed as:

[0092] minJ(Δs togo )→σ cmd

[0093] minJ(h)→(α max ,α min )

[0094]

[0095] where J(·) is the objective function, Δs togo is the remaining range to fly, σ cmd is the bank angle command, Δψ is the course angle deviation, Δψ threshold is the course angle deviation threshold, sign(·) is the sign function used to take the positive or negative of the variable, is the bank angle command in the k-th guidance cycle, is the bank angle command in the (k - 1)-th guidance cycle, α max is the maximum angle of attack, α min is the minimum angle of attack;

[0096] Specifically, the aerodynamic model correction condition can be a height condition, a speed condition, or a range condition, aiming to ensure the stability of the aircraft guidance method by making the aerodynamic correction occur only in certain flight segments, while reducing the computational load;

[0097] Reference Figure 4 andFigure 6 , the present invention can effectively identify the aerodynamic disturbances of the aircraft in a complex disturbance environment, enabling the aircraft to correct the aerodynamic model error of the aircraft in real time during the predictive correction guidance process, thereby ensuring the guidance accuracy. The current state of the aircraft (h, θ, φ, v, γ, ψ), where θ is the current geocentric longitude of the aircraft, φ is the current geocentric latitude of the aircraft, ψ is the flight heading angle, and the predicted terminal state (h fe , θ fe , φ fe , v fe , γ fe , ψ fe ), where h fe is the predicted altitude, θ fe is the predicted geocentric longitude, φ fe is the predicted geocentric latitude, v fe is the predicted speed, γ fe is the predicted path angle, ψ fe is the predicted heading angle, and the desired terminal state (h c , θ c , φ c , v c , γ c , ψ c ), where h c is the desired altitude, θ c is the desired geocentric longitude, φ c is the desired geocentric latitude, v c is the desired speed, γ c is the desired path angle, ψ fe is the desired heading angle;

[0098] Reference Figures 8 - 14 , in the initial state, the initial altitude H0 = 50, in km, the initial speed V0 = 2500, in m / s, the initial path angle γ0 = -1°, and in the terminal state, the output altitude H f = 20, in km, the output speed V f = 1500, in m / s, and the aircraft range Range = 680, in km;

[0099] Among them, Velocity is the speed, in m / s, Height is the altitude, in km, Time is the time, in s, Range to Go is the remaining range, Aziumuth Deviation is the heading angle deviation, Alpha Angle is the angle of attack, and Correct Bank Angle is the corrected bank angle.

[0100] Although the present invention has been described in terms of a limited number of embodiments, those skilled in the art will appreciate, in light of the foregoing description, that other embodiments can be contemplated within the scope of the invention as thus described. Further, it should be noted that the language used in this specification has been principally selected for readability and instructional purposes and not to limit or define the inventive subject matter. Accordingly, many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the appended claims. For the scope of the present invention, the disclosure of the present invention is illustrative and not restrictive, and the scope of the present invention is defined by the appended claims.

Claims

1. A predictive correction guidance method for a reentry vehicle based on neural network feedback correction, characterized in that, It includes the following steps: S1. Generate the reentry trajectory of the aircraft under aerodynamic disturbances and the predicted trajectory of the aircraft without aerodynamic disturbances through Monte Carlo shooting, and calculate to obtain the aerodynamic disturbance identification network dataset; S2. Construct the aerodynamic disturbance identification neural network structure, input the aerodynamic disturbance identification network dataset into the aerodynamic disturbance identification neural network structure for training, and obtain the trained neural network structure; S3. Use the trained neural network structure to correct the aerodynamic model, that is, set the aerodynamic model correction conditions, and perform correction judgment on the aerodynamic model in each correction command guidance cycle to achieve aircraft prediction and correction guidance; In S3, the correction judgment process is expressed as: if the aerodynamic model correction conditions are met, obtain the aerodynamic disturbance evaluation result through the neural network structure according to the current state deviation to correct the aerodynamic model, and use the corrected model for the prediction and correction guidance algorithm; if the aerodynamic model correction conditions are not met, continue to use the current aerodynamic model for the prediction and correction guidance algorithm; The prediction and correction guidance algorithm is a guidance mode with longitudinal and lateral section separation. The longitudinal section guidance adopts the double correction logic of angle of attack and bank angle. The longitudinal section adopts the bank angle correction strategy based on range and the angle of attack profile correction strategy based on altitude. The lateral section adopts the bank angle corridor flipping strategy; The prediction and correction guidance algorithm is expressed as: minJ(Δs togo )→σ cmd minJ(h)→(α max ,α min ) Among them, J(·) is the objective function, Δs togo is the remaining flight range, σ cmd is the bank angle command, Δψ is the course angle deviation, Δψ threshold is the course angle deviation threshold, sign(·) is the sign function used to take the positive or negative of the variable, is the bank angle command in the k-th guidance cycle, is the bank angle command in the (k - 1)-th guidance cycle, α max is the maximum angle of attack, α min is the minimum angle of attack.

2. The predictive correction guidance method for a reentry vehicle based on neural network feedback correction according to claim 1, wherein In S1, the aerodynamic disturbance identification network dataset includes the aerodynamic disturbance coefficient deviation and the state quantity deviation; Select the aircraft altitude h, aircraft speed v, flight path angle γ, and aircraft mechanical energy e as the predicted aircraft states. The actual flight state is (h1, v1, γ1, e1), where h1 is the actual aircraft altitude, v1 is the actual aircraft speed, γ1 is the actual flight path angle, and e1 is the actual aircraft mechanical energy. Calculate the state quantity deviation (δh, δv, δγ, δe) between the predicted aircraft states and the actual flight state; The state quantity deviation (δh, δv, δγ, δe) is expressed as: Among them, δh is the altitude state quantity deviation, δv is the speed state quantity deviation, δγ is the path angle state quantity deviation, and δe is the mechanical energy state quantity deviation; The aircraft mechanical energy e is expressed as: Among them, r is the current geocentric distance of the aircraft; Select the aerodynamic drag coefficient perturbation deviation δC d as the aerodynamic perturbation coefficient deviation.

3. A predictive correction guidance method for a reentry vehicle based on neural network feedback correction according to claim 2, characterized in that In S2, the aerodynamic disturbance identification neural network structure includes an input layer, two hidden layers, and an output layer. L2 regularization is used to train the aerodynamic disturbance identification neural network to avoid overfitting. The state quantity deviation is used as the input, and the aerodynamic disturbance coefficient deviation is used as the output; L2 Regularization L L2 Expressed as: L data = δC d - Φ(δh, δv, δγ, δe) Among them, L data is the error between the neural network prediction value Φ(δh, δv, δγ, δe) and the deviation δC of the aerodynamic disturbance coefficient d , λ is the regularization parameter, is the square of the L2 norm of the neural network parameter set w, and w i is the neural network parameter; The input layer includes 4 nodes, each hidden layer includes 10 nodes, both hidden layers adopt the Relu activation function, and the output layer includes 1 node; The Relu activation function f(x) is expressed as: f(x) = max(0, x) Among them, max is the limiting function, and x represents the output data of the input layer or the first hidden layer.

Images