CN117311390B - Intelligent combined guidance method for closed-loop tracking of aerospace shuttle aircraft - Google Patents

Abstract

The invention belongs to the technical field of closed-loop tracking intelligent combination guidance of air-to-space aircraft, and specifically relates to a closed-loop tracking intelligent combination guidance method of air-to-space aircraft, which includes the following steps: S1: Design an LSTM aircraft trajectory prediction algorithm; S2: Establish RLV and then Enter the segment error model and transform the constrained control problem; S3: Design an aircraft tracking controller based on adaptive dynamic programming. The present invention adopts a prediction correction framework, and combines prediction and adaptive dynamic programming based on the LSTM method to design a controller, which can ensure that the selected performance index function reaches the optimum within a limited time domain, obtain the optimal feedback guidance law, and improve the performance of the controller. autonomy.

Description

A closed-loop tracking intelligent combined guidance method for air-to-space round-trip aircraft

Technical Field

The present invention belongs to the technical field of closed-loop tracking and intelligent combined guidance of an aerospace round-trip aircraft, and in particular relates to a closed-loop tracking and intelligent combined guidance method for an aerospace round-trip aircraft.

Background Art

Reusable launch vehicle (RLV) refers to a multi-purpose and reusable aircraft that can freely travel between the earth's surface and space orbit. It is an inevitable trend to achieve fast, reliable and cheap access to space in the future, and is also a research hotspot in the current aerospace field. Due to the fast speed change, strong coupling, model uncertainty and external environment of the reentry phase of RLV, the design of the reentry phase control system faces greater challenges. In order to ensure the safe and stable reentry flight of the controlled aircraft, the optimization of the RLV reentry phase trajectory and the design of the guidance law are particularly critical. The goal of RLV reentry trajectory optimization is to achieve flight trajectory control to reach a certain optimal target while satisfying conditions such as state constraints and control quantity constraints. Unlike standard trajectory guidance, the prediction and correction guidance method does not rely on the reference trajectory, but first predicts the flight endpoint during the flight, and designs the controller based on the deviation between the predicted landing point and the expected endpoint. It has higher flexibility and landing point accuracy, and does not rely on the initial state of reentry. It has stronger anti-interference ability to the initial reentry disturbance, and is increasingly becoming the development direction of research in various countries;

Aircraft trajectory prediction is one of the indispensable functional components in intelligent flight systems. In a complex game environment, predicting the trajectory of an aircraft in advance will provide a reference direction for subsequent maneuvering decisions. Trajectory prediction refers to estimating the trajectory at a future moment based on existing information and in accordance with certain rules or methods.

From the current research, the existing technologies can be divided into two categories: kinematic model methods and data-based methods. The former is more widely used. For example, in 2017, Wei Xiqing and others from Harbin Institute of Technology proposed an extended Kalman filter combined with the Singer model for state estimation to further recursively infer the target motion trajectory, in response to the periodic jumping motion problem of hypersonic aircraft. In 2018, Zhang Kai and other scholars from the Air Force Early Warning Academy constructed target motion and flight intention characteristics and used Bayesian theory to iteratively derive the flight model. Subsequently, they used the Monte Carlo sampling method to achieve trajectory prediction. Although the above methods have good interpretability, their limited prediction accuracy and long prediction time make them only applicable to certain specific scenarios.

In summary, the guidance method in the prior art cannot ensure that the selected performance indicator function reaches the optimum within a limited time domain, and cannot effectively obtain the optimal feedback guidance law, resulting in low autonomy of the controller.

Summary of the invention

The purpose of the present invention is to provide an intelligent combined guidance method for closed-loop tracking of an air-to-space round-trip vehicle. It adopts a predictive correction framework, combines prediction with adaptive dynamic programming based on the LSTM method to design a controller, and can ensure that the selected performance indicator function reaches the optimal level within a limited time domain, obtain the optimal feedback guidance law, and improve the autonomy of the controller.

The technical solution adopted by the present invention is as follows:

A closed-loop tracking intelligent combined guidance method for an air-to-space round-trip aircraft comprises the following steps:

S1: Design an LSTM-based aircraft trajectory prediction algorithm;

S2: Establish the RLV reentry error model and transform the constraint control problem;

S3: Design an adaptive dynamic programming based vehicle tracking controller.

Furthermore, the S1 comprises the following steps:

S101: Perform data preprocessing operations by combining system state data at each time step and error data from the end point to build an information data set and database for the RLV trajectory prediction problem;

S102: Obtain the coupling information between the environment and the system through input data, build an LSTM network for state prediction, use the back propagation algorithm to update the prediction network weights, obtain the system's predicted state model, and realize real-time prediction of the system state trajectory at each time step;

S103: During the flight, the obtained prediction model is used to continuously predict the flight endpoint, and the deviation between the predicted landing point and the expected endpoint is used as the control error and input to the controller to adjust the control amount.

Furthermore, the data preprocessing in S101 includes information fusion and feature extraction.

Furthermore, the method for establishing the information data set in S101 includes the following steps:

S10101: Obtain information, relying on sensors, radars, etc., to obtain various status information of the aircraft that changes over time;

S10102: Data preprocessing: using zero mean standardization preprocessing method to preprocess the data;

S10103: Construct training samples and decompose trajectory data into training samples and labels.

Furthermore, constructing the training sample comprises the following steps:

Starting from the first trajectory point in the data set, in chronological order, select the aircraft state information corresponding to the time of the first 20 trajectory points to predict the state information of the next trajectory point, where the state information of each time step is used as the input of the corresponding cell of the neural network, and the separation interval is selected as 1. Starting from the second trajectory point, the training samples are selected in the same way.

Furthermore, the S102 includes the following steps:

S10201: taking the state error information of the aircraft as input, and mapping the input data to a new space through an embedding function;

S10202: Taking the historical state information of the aircraft as input, fusing the error information with the historical state information of the aircraft;

S10203: Use LSTM network to predict the future trajectory of the aircraft based on the observed historical state information and fusion information;

S10204: Construct a prediction model through sample data, normalize the training samples and prediction samples, input the training samples into the network model to train the network, adjust the network structure according to the loss size, use the test samples to test the network performance, and obtain the prediction results.

Furthermore, the information data set is 14-dimensional, including state difference information between the current state and the terminal state of the aircraft, including geocentric distance information difference, longitude information difference, latitude information difference, speed information difference, track angle information difference, heading angle information difference and roll angle information.

Further, the S2 comprises the following steps:

An RLV reentry error model is established, and a performance index function that simultaneously reflects the formation error, control quantity and collision avoidance effect is designed. The collision avoidance problem in the scenario is converted into a constraint problem through the safety obstacle function, and the collision avoidance control problem is converted into a stable control problem of the error system.

Further, the S3 comprises the following steps:

Design a control algorithm based on adaptive dynamic programming, build a judgment network to approximate the optimal performance index function and solve the optimal control strategy, use the policy gradient method to update the norm of all neural network values, use the network output iteration, and finally obtain the optimal control strategy

Furthermore, the state of the RLV needs to strictly meet the following constraints:

1). Define the aircraft state x in the control algorithm, satisfying the starting state condition x ₀ and the end state condition x _f ;

2). Affected by the performance of the aircraft, during the reentry process, the control quantity u is defined to satisfy the constraint u _min ≤u≤u _max , and the state quantity x satisfies the constraint x _min ≤x≤x _max , where u _min represents the upper limit of the control quantity and u _max represents the lower limit of the control quantity; where x _min represents the upper limit of the state quantity and x _max represents the lower limit of the state quantity.

The technical effects achieved by the present invention are:

The invention discloses a closed-loop tracking intelligent combined guidance method for a space shuttle vehicle, which adopts a prediction-correction framework and combines prediction and adaptive dynamic programming based on the LSTM method to design a controller, thereby solving the "curse of dimensionality" problem of traditional dynamic programming control algorithms. The guidance law is continuously iterated and updated through learning, and ultimately the selected performance index function is ensured to be optimal within a limited time domain, thereby obtaining the optimal feedback guidance law and improving the autonomy of the controller.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a block diagram of an air-space round-trip vehicle closed-loop tracking intelligent combined guidance system based on LSTM and adaptive dynamic programming according to the present invention;

FIG2 is a diagram of an aircraft prediction model of the LSTM of the present invention;

Fig. 3 is a schematic diagram of LSTM of the present invention;

FIG4 is a diagram of aircraft state prediction results of the present invention;

FIG5 is a graph showing a change in absolute percentage error of aircraft state prediction according to the present invention;

6 is a state change curve diagram of tracking control based on adaptive dynamic programming of the present invention;

7 is a graph showing the error variation of the tracking control based on adaptive dynamic programming of the present invention;

FIG8 is a graph showing weight transformation of an aircraft evaluation neural network according to the present invention;

FIG9 is a LSTM model parameter diagram of the present invention;

FIG. 10 is a diagram showing the initial conditions and terminal constraints of the aircraft of the present invention.

DETAILED DESCRIPTION

In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

Embodiment 1:

As shown in FIG1 , a closed-loop tracking intelligent combined guidance method for an air-to-space round-trip aircraft includes the following steps:

S1: Design an LSTM-based aircraft trajectory prediction algorithm;

As shown in Figures 2 and 3, S1 includes the following steps:

S101: Build an information data set and database for RLV trajectory prediction problem by combining system state data and error data from the end point at each time step to perform data preprocessing operations such as information fusion and feature extraction;

In order to provide control direction for the aircraft controller, it is first necessary to predict the state of the aircraft in the future time step. When predicting the state of the aircraft, the future state is related to the historical state information of the aircraft and the state difference between the current state and the terminal state of the aircraft. Therefore, the prediction should be based on a large amount of empirical data to analyze the correlation between these characteristics and future behaviors to establish a prediction model.

Define the observation time step as T _obs , the prediction time step as T _pred , and define the historical state trajectory of the aircraft as O = { p ^t |t = 1, 2, ..., T _obs }, where p ^t represents the state information of the aircraft at time t, including: (1) the distance from the center of the earth r, (2) the longitude information θ, (3) the latitude information (4) speed information θ, (5) track angle information γ, (6) heading angle information χ, (7) roll angle information σ; define the state difference set between the current state and the final state of the aircraft as , the real future state information of the aircraft is represented by Y = {p ^t |t = T _obs +1, T _obs +2,…, T _obs +T _pred }. Therefore, the trajectory prediction problem can be described as:

Y＝f _o (O,O _e ) (1)

The goal of trajectory prediction is to find a mapping function f _o from the set O, O _e to the set Y, so that the function can predict the future trajectory of the aircraft based on its historical trajectory.

Since the prediction method based on deep learning needs to train the network based on a large amount of data, first of all, it is necessary to establish an information data set of aircraft status characteristics, which can be divided into the following three steps:

S10101: Information acquisition part: Relying on sensors, radars, etc., various status information of the aircraft that changes with time is acquired. The data set obtained by simulation by the above method is 14 dimensions, which includes the status characteristic information of the aircraft (geocentric distance information, longitude information, latitude information, speed information, track angle information, heading angle information, roll angle information), and the status difference information between the current state and the terminal state of the aircraft (geocentric distance information difference, longitude information difference, latitude information difference, speed information difference, track angle information difference, heading angle information difference, roll angle information difference).

S10102: Data preprocessing: Since there are order of magnitude differences between data samples of different state feature information of aircraft, when training the network, data samples with larger order of magnitude will dominate, resulting in slower convergence speed and lower accuracy, so data preprocessing is required. The present invention adopts a zero-mean normalization preprocessing method, and its calculation formula is as follows:

In the formula, d _i is the original sample data of each feature information, d ′ _i is the sample data after each feature information is processed, is the mean of all sample data of each feature information, τ _i is the standard deviation of all sample data of each feature information, where i = {1, 2, ..., 14}, representing each dimension of the 14 dimensions of the data set. That is, according to the mean and standard deviation of the state information of each dimension in the 14 dimensions of the data set, the original data of this dimension of all data samples is transformed into normalized data. The processing unifies the data into a specific interval, so that the learning process will not be slow due to the large difference in information of different data samples, resulting in low learning speed and low network accuracy.

S10103: Construct training samples Construction method: Trajectory prediction is a supervised learning problem, which requires decomposing trajectory data into training samples and labels. Starting from the first trajectory point in the data set, in chronological order, select the aircraft state information corresponding to the time of the first 20 trajectory points to predict the state information of the next trajectory point, where the state information of each time step is used as the input of the corresponding cell of the neural network. Then, in order to ensure the continuity of the samples in time, select the separation interval as 1, that is, starting from the second trajectory point, select the training samples in the same way.

S102: Obtain the coupling information between the environment and the system through input data, build an LSTM network for state prediction, use the back propagation algorithm to update the prediction network weights, obtain the system's predicted state model, and realize real-time prediction of the system state trajectory at each time step.

After obtaining the data set, the present invention uses the mechanism of deep learning to learn the mapping relationship based on the LSTM algorithm, considers the temporal nature of the aircraft state information, designs a time-based back propagation algorithm, determines the number of sample iterations, learning rate and other hyperparameters in the network training process based on experience and offline experiments, and updates the network weights; finally, in the actual online prediction process, the trained aircraft state prediction network is used to realize real-time online prediction of the aircraft state. The prediction process is shown in Figure 2-3.

First, the data containing the aircraft status and error information is input into the embedding layer to extract the error information, and the Then the historical status information of the aircraft is extracted through the data processing layer and combined with the error information The merged data is used as the input of the LSTM layer, and its output is the predicted trajectory of the aircraft.

Here, S102 includes the following steps:

S10201: Embedding layer: embedding the state error information of the aircraft As input, the input data is mapped to the new space through the embedding function φ(·). Its mathematical expression is:

In the formula, is the predicted state error information of the aircraft at time step t. is the spatial representation of the state error information of the aircraft at time step t, and W ₁ is the weight of the embedding function. The embedding function φ(·) is defined as follows:

S10202: Data processing layer: The historical state information of the aircraft p ^t is taken as input, and the error information Fusion with the aircraft's historical state information is as follows:

Where ⊙ represents the Hadamard product of the matrix and W ₂ is the weight of the embedding function. is the predicted state information of the aircraft at time step t. Therefore, the fusion information ^qt depends not only on the historical state information of the aircraft, but also on the error information of the aircraft state from the end point.

S10203: LSTM layer: The LSTM network is used to predict the future trajectory of the aircraft based on the observed historical state information and fusion information q ^t , as shown in the following formula:

In the formula, dp ^t represents the derivative of the aircraft state generated at time t; W ₃ is the network weight of LSTM; W ₄ is the weight of the embedding function. ^{o t} and h ^t-1 are the output and hidden state of LSTM respectively, and the input of LSTM is [q ^t , dp ^t-1 ]. Here, dp ^t-1 reflects the differential amount of data information, indicating that the network considers the dynamics of data changes in the prediction process. The LSTM network structure is shown in Figure 3:

It can be seen that the input of the LSTM network at time t not only includes the input data x ^t , but also includes the hidden state h ^t-1 from the previous moment. x ^t and h ^t-1 obtain the probability of discarding the hidden cell state C ^t-1 of the previous layer through the activation function in the forget gate, and the product with C ^t-1 becomes part of the cell state C ^t at time t; x ^t and h ^t-1 determine the information that needs to be updated for the cell state C ^t at time t through the activation function in the input gate, and this information becomes another part of C ^t . Therefore, the cell state C ^t at time t contains useful information from the past moment and new information at the current moment. Finally, the output result is determined at the output gate based on x ^t , h ^t-1 and the cell state. Therefore, the output of the LSTM at time t takes into account the useful information of the past moment and the state of the current moment. Such a network structure is conducive to learning the relationship between time series data and improving the ability to handle time series prediction problems.

S10204: Training network: After the sample data is constructed, the prediction model first normalizes the training samples and prediction samples. Then the training samples are input into the network model to train the network, and the network structure is adjusted according to the loss size. Finally, the test sample is used to test the network performance and obtain the prediction result.

The network training process uses the mean square error function as the loss function, that is:

In the formula, n represents the number of batch samples in each training process. is the future state of the aircraft predicted by the LSTM model, and Y _k is its actual future state. The optimization goal of the deep neural network is to make the MSE approach 0. The process of deep neural network training is the process of updating the network weights. According to the above loss function, the back propagation algorithm is used to update the network weights, and finally the prediction result is obtained through the weight update.

The prediction result information of each state is expressed as (1) geocentric distance information r _ref , (2) longitude information θ _ref , (3) latitude information (4) velocity information θ _ref , (5) track angle information γ _ref , (6) heading angle information χ _ref , (7) roll angle information σ _ref .

S103: During the flight, the obtained prediction model is used to continuously predict the flight endpoint, and the deviation between the predicted landing point and the expected endpoint is used as the control error and input to the controller to adjust the control amount.

S2: Establish the RLV reentry error model and transform the constraint control problem;

Establishing the RLV reentry error model and transforming the constraint control problem in S2 includes the following steps:

An RLV reentry error model is established, and a performance indicator function that simultaneously reflects the formation error, control quantity and collision avoidance effect is designed. The collision avoidance problem in the scenario is converted into a constraint problem through the safety obstacle function, thereby converting the collision avoidance control problem into the optimal stable control problem of the error system to ensure the safety of the system.

In the RLV reentry phase, it is assumed that the aircraft is a point mass flying without power, and when the earth is considered to be a rotating ellipsoid, the side force and the influence of the earth's rotation during the reentry process are ignored, and the sideslip angle is taken as zero. Then the RLV reentry phase dynamic system is:

In the formula, r,θ, v, γ, χ represent the distance from the center of the earth, longitude, latitude, flight speed, track angle and heading angle respectively; σ is the roll angle; m is the mass of the aircraft; g is the acceleration of gravity, g = μ _g /r ² , where μ _g is the gravity parameter; L is the lift, L = q _d SC _L ; D is the drag, D = q _d SC _D , where S is the RLV aerodynamic reference area, q _d is the dynamic pressure, q _d = 0.5ρv ² ; ρ is the atmospheric density, Where _ρ0 is the atmospheric density at sea level, _Re is the radius of the earth, and β is a constant coefficient; the lift coefficient _CL and the drag coefficient _CD are expressed as functions of the angle of attack α, which will be given in subsequent simulations.

If only the existing control quantity is processed, high-frequency chattering will be introduced and the convergence of the problem will be affected, so a new control variable needs to be introduced. The new auxiliary control variable is introduced to decouple the control quantity from the state quantity, and the new control quantity is:

Then the RLV reentry phase dynamics model can be rewritten as:

In the formula, is the new state quantity. The state matrix f(x') and control matrix B are:

B＝[0,0,0,0,0,0,1] ^T

Define the next step trajectory prediction state quantity given by the last LSTM prediction model of S1 as Then the error system of the aircraft state and reference input is as follows:

Where e represents the error between the current state of the aircraft and the target.

Combining equations , , and , we can get the RLV reentry error system:

In order to ensure the safe and stable flight of the aircraft, it is necessary to ensure that the state and control quantity of the aircraft meet the constraints and ensure the safety of the system. Based on this, this step will design a constraint item based on the safety barrier function to provide a basis for the constraint requirements in the subsequent performance indicator function.

In order to ensure the safe and stable flight of the aircraft, the state of the RLV needs to strictly meet some constraints:

1). Define the aircraft state x in the control algorithm, satisfying the starting state condition x ₀ and the end state condition x _f .

2). Affected by the performance of the aircraft, during the reentry process, the control quantity u is defined to satisfy the constraint u _min ≤u≤u _max , and the state quantity x satisfies the constraint x _min ≤x≤x _max , where u _min represents the upper bound of the control quantity and u _max represents the lower bound of the control quantity; where x _min represents the upper bound of the state quantity and x _max represents the lower bound of the state quantity.

According to the state constraints of the aircraft, the aircraft state safety domain and state obstacle function can be designed as follows:

Security domain _Dx :

D _x ＝{x∈R ⁿ ∣x _min ≤x≤x _max } (14)

Barrier function μ _x :

In the formula, when the state of the aircraft is outside the safety domain _Dx at a certain moment, it is judged that the constraint condition is not met, and η is a positive real constant that satisfies η＞1.

It can be seen that _μx is a safety barrier function. The relevant performance index function is designed through the function _μx . When the state of the aircraft is close to the boundary of the safety domain _Dx at a certain moment, the performance index function will approach infinity. At the next moment, the controller will change in the direction of minimizing the performance index function. Therefore, the existence of _μx ensures the safety of the system state.

Similarly, according to the control quantity constraints of the aircraft, the safety domain of the aircraft control quantity and the control quantity obstacle function can be designed as follows:

Security domain _Du :

D _u ＝{u∈R ⁿ ∣u _min ≤u≤u _max } (16)

Barrier function μ _u :

Next, define the system's performance indicator function J as:

J＝∫ _t ^∞ U(e,u)dt (18)

In the formula, the instantaneous performance index U(e,u) is defined as:

U(e,u)＝e ^T Qe+u ^T Ru+μ _x e ^T e+μ _u u ^T u (19)

Where Q and R are positive definite skew-symmetric matrices.

It can be seen that the performance index function consists of four parts: the first term e ^T Qe represents the aircraft state error cost, the second term u ^T Ru represents the aircraft control quantity cost, the third term μ _x e ^T e represents the aircraft state constraint, and the third term μ _u u ^T u represents the aircraft control quantity constraint.

Then the optimal performance index function J ^* can be expressed as:

In order to achieve the smooth flight and constraint conditions of the RLV reentry error system, the control goal is to find a set of control strategies u that can minimize the performance index function and limit the system state to the safety domain D _x and the control quantity to the safety domain _Du . The above control problem can be written as:

S3: Design an adaptive dynamic programming based vehicle tracking controller;

Designing an adaptive dynamic programming-based vehicle tracking controller in S3 includes the following steps:

A control algorithm based on adaptive dynamic programming is designed, and an evaluation network is constructed to approximate the optimal performance index function and solve the optimal control strategy. The policy gradient method is used to update the norm of all valued values of the neural network, and the network output is iterated to finally obtain the optimal control strategy.

Because the designed performance index function is continuously differentiable, the following Lyapunov equation can be obtained:

In the formula It represents the partial derivative of the performance index function J with respect to the error system e, and J(0) = 0. The Hamiltonian equation of the problem can be defined as:

When the performance index function is optimal, the Hamiltonian equation will become the Hamilton-Jacobi-Bellman equation, that is,

The optimal performance index function J ^* is obtained by solving the Hamilton-Jacobi-Bellman equation above, and it has a unique solution. When the solution J ^* exists and is continuously differentiable, the optimal control strategy can be solved It is obtained that the aircraft control strategy that minimizes the performance index function is

In order to solve the problem that the Hamilton-Jacobi-Bellman equation is difficult to solve in practical applications, the present invention uses the principle of single-layer judgment neural network approximation to approximate the value of the optimal performance indicator function J ^* :

J ^* = W _c ^T σ _c (e) + ε _c (26)

In the formula, W _c ∈R ⁿ is the ideal weight vector of the judgment neural network, σ _c (e) ∈R ⁿ represents the activation function of the judgment neural network, and ε _c ∈R represents the approximate error of the judgment neural network. Taking the partial derivative, we can get:

In the formula, and Denote the partial derivatives of the activation function and the approximate error respectively. Substituting the above formula into, we can get the Hamiltonian equation:

Where e _cH is the residual error generated by the approximation of the neural network. In the framework of adaptive dynamic programming, considering the fact that the ideal weights are unknown, it is usually based on the estimated weight vector Establish a judgment neural network to approximate the optimal performance index function, that is:

Therefore, the approximate Hamiltonian can be written as:

definition Weight estimation error At this time there are:

In order to adjust the critical judgment neural network weight vector Define the error function The policy gradient method is used to minimize the error function, so the weight adjustment rule of the judgment neural network is set to:

Where α _c > 0 is the learning rate of the neural network. Therefore, the ideal control strategy can be described as:

Its approximate value is expressed as:

Based on the above three steps, the entire process of closed-loop tracking and intelligent combined guidance of aerospace round-trip vehicles based on LSTM and adaptive dynamic programming is completed.

This technical solution adopts a predictive correction framework, combines prediction based on the LSTM method with adaptive dynamic programming to design a controller, solves the "curse of dimensionality" problem of traditional dynamic programming control algorithms, and continuously iterates and updates the guidance law through learning. Ultimately, it ensures that the selected performance indicator function reaches the optimal level within a limited time domain, obtains the optimal feedback guidance law, and improves the autonomy of the controller.

Embodiment 2:

In order to verify the effectiveness of the algorithm proposed in this embodiment, the algorithm is integrated and designed in MATLAB/Simulink, and a simulation experiment is performed. The main simulation process is as follows:

LSTM prediction model and network training parameter settings:

The configuration of the trajectory prediction model is shown in Figure 9. The input of the embedding layer network is the aircraft's geocentric distance information r, longitude information θ, and latitude information The output of the prediction network is the future position information of the aircraft, including the velocity information θ, the track angle information γ, the heading angle information χ, the roll angle information σ, and the 14-dimensional information formed by the error corresponding to the final state.

The experiment first trained the prediction network offline with a large amount of trajectory data, and then used the trained model to predict the trajectory in the actual online prediction process. In this way, the proposed prediction model requires a shorter prediction time, ensuring the real-time prediction. The experiment uses 20 historical time step information to predict future information, that is, T _obs = 20, and each time uses the 20*14 vector of the overall state information of the aircraft as input, and the T _pred *2*3 dimensional vector of the future state information of the aircraft is the output. In this way, 42970*20*14 training samples n are formed using the aircraft trajectory data. The LSTM model randomly selects 80% of the data set for training, retains the remaining 20% of the data set for testing, and then uses 15% of the training set as the validation set to verify and fine-tune the network during training. The algorithm uses single-step trajectory prediction for simulation experiments, that is, T _pred = 1, and the learning rate is uniformly set to 0.01 during back propagation. In order to prevent control fluctuations, smooth interpolation processing is performed.

Adaptive dynamic programming tracking controller parameter settings:

The basic parameters of RLV are as follows: vehicle mass m = 104.305 kg; gravity parameter μ _g = 3.986 × 10 ¹⁴ m ³ / s ² ; vehicle area S = 391.22 m ² ; earth radius _Re = 6.37 × 10 ³ m, sea level atmospheric density ρ ₀ = 0.00238 kg/m ³ ; lift coefficient C _L = -0.207 + 2.04α, drag coefficient C _D = 0.0785 - 0.3529α. System constant coefficient β = 1.3875 × 10 ^-4 , k _Q = 9.44 × 10 ^-5 . Vehicle, quantity constraints The initial conditions and terminal constraints of the aircraft are shown in Figure 10. The state constraints of the aircraft are:

The neural network activation function is The initial value of the neural network weights is W _c = [90, 80, 30, 20, 40, 75, 60, 15] ^T , Q = R = I, η = 1.5, and the learning rate of the neural network is α _c = 0.5.

Among them, Figure 4 shows the prediction results of various states of the aircraft changing with time steps. It can be seen from Figure 4 that the prediction model has achieved accurate state prediction for the aircraft's altitude, speed, latitude, longitude, track angle, heading angle and roll angle. Therefore, the designed LSTM-based state prediction model can realize accurate observation of the aircraft's state.

Among them, Figure 5 shows the absolute percentage error change curve between the aircraft state prediction and the actual state. It can be seen from Figure 5 that the model errors predicted under different states are all below 6%, which verifies the accuracy of the prediction model.

Among them, as shown in Figure 6, it shows the state change curves of the aircraft in terms of altitude, speed, latitude, longitude, track angle, heading angle and roll angle. It can be seen from the figure that the state tracking is basically formed in about 500s, and then maintains a stable tracking flight, which verifies the effectiveness and stability of the designed tracking control algorithm based on adaptive dynamic programming.

Among them, Figure 7 shows the change curves of the state errors of the aircraft in altitude, speed, latitude, longitude, track angle, heading angle and roll angle. It can be seen from the figure that after about 500s, the state error of the aircraft gradually converges to 0, which further verifies the effectiveness of the designed tracking control algorithm based on adaptive dynamic programming.

Among them, Figure 8 shows the change process of the weight parameters of the aircraft evaluation neural network. It can be seen from Figure 8 that with the change of time, within a limited time, the weight parameters of the evaluation neural network are stably converged and approach the corresponding optimal value.

The above are only preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to conventional means in the art unless otherwise specified and limited.

Claims (7)

1. A closed-loop tracking intelligent combined guidance method for an air-to-space round-trip aircraft, characterized in that it comprises the following steps: S1: designing an LSTM aircraft trajectory prediction algorithm; The S1 comprises the following steps: S101: Data information fusion and feature extraction are performed by combining the system state data and the error data from the end point at each time step to build an information data set and database for the RLV trajectory prediction problem; S102: Obtain the coupling information between the environment and the system through input data, build an LSTM network for state prediction, use the back propagation algorithm to update the prediction network weights, obtain the system's predicted state model, and realize real-time prediction of the system state trajectory at each time step; S103: During the flight, the obtained prediction model is used to continuously predict the flight endpoint, and the deviation between the predicted landing point and the expected endpoint is used as a control error and input to the controller to adjust the control amount; S2: Establish the RLV reentry error model and transform the constraint control problem; S3: Design an adaptive dynamic programming based vehicle tracking controller; The S3 comprises the following steps: A control algorithm based on adaptive dynamic programming is designed, and an evaluation network is constructed to approximate the optimal performance index function and solve the optimal control strategy. The policy gradient method is used to update the norm of all valued values of the neural network, and the network output is iterated to finally obtain the optimal control strategy. 2. The closed-loop tracking intelligent combined guidance method for an air-to-space round-trip aircraft according to claim 1, characterized in that: the method for establishing the information data set in S101 comprises the following steps: S10101: Obtain information, relying on sensors and radars to obtain various status information of the aircraft that changes over time; S10102: Data preprocessing: using zero mean standardization preprocessing method to preprocess the data; S10103: Construct training samples and decompose trajectory data into training samples and labels. 3. The closed-loop tracking intelligent combined guidance method for an air-to-space shuttle according to claim 2, characterized in that: the construction of training samples comprises the following steps: Starting from the first trajectory point in the data set, in chronological order, select the aircraft state information corresponding to the time of the first 20 trajectory points to predict the state information of the next trajectory point, where the state information of each time step is used as the input of the corresponding cell of the neural network, and the separation interval is selected as 1. Starting from the second trajectory point, the training samples are selected in the same way. 4. The closed-loop tracking intelligent combined guidance method for an air-to-space shuttle according to claim 1, characterized in that: said S102 comprises the following steps: S10201: taking the state error information of the aircraft as input, and mapping the input data to a new space through an embedding function; S10202: Taking the historical state information of the aircraft as input, fusing the error information with the historical state information of the aircraft; S10203: Use LSTM network to predict the future trajectory of the aircraft based on the observed historical state information and fusion information; S10204: Construct a prediction model through sample data, normalize the training samples and prediction samples, input the training samples into the network model to train the network, adjust the network structure according to the loss size, use the test samples to test the network performance, and obtain the prediction results. 5. According to claim 1, a closed-loop tracking intelligent combined guidance method for an air-to-space round-trip aircraft is characterized in that: the information data set is 14-dimensional, and the information data set includes state characteristic information of the aircraft and state difference information between the current state and the terminal state of the aircraft, the state characteristic information of the aircraft includes geocentric distance information, longitude information, latitude information, speed information, track angle information, heading angle information and roll angle information, and the state difference information between the current state and the terminal state of the aircraft includes geocentric distance information difference, longitude information difference, latitude information difference, speed information difference, track angle information difference, heading angle information difference and roll angle information difference. 6. The closed-loop tracking intelligent combined guidance method for an air-to-space shuttle according to claim 1, characterized in that: S2 comprises the following steps: An RLV reentry error model is established, and a performance index function that simultaneously reflects the formation error, control quantity and collision avoidance effect is designed. The collision avoidance problem in the scenario is converted into a constraint problem through the safety obstacle function, and the collision avoidance control problem is converted into a stable control problem of the error system. 7. According to claim 1, a closed-loop tracking intelligent combined guidance method for a space shuttle vehicle is characterized in that: the state of the RLV needs to strictly meet the following constraints: 1) Define the state of the aircraft in the control algorithm Satisfy the starting state conditions and the end state condition 2). Affected by the performance of the aircraft, during the reentry process, the control amount is defined Satisfy constraints State quantity Satisfy constraints in represents the upper bound of the control quantity, represents the lower bound of the control quantity; represents the upper bound of the state quantity, Indicates the lower bound of the state quantity.