A method and system for generating real-time penetration trajectory of an aircraft

Abstract

The present invention relates to a method and system for generating a real-time penetration trajectory of an aircraft. The present invention solves the penetration trajectory optimization problem by using a second-order cone programming method to obtain the penetration trajectory under different initial conditions, and after generating a data set, uses the generated data set to train a deep neural network to obtain a trained deep neural network model, and then uses the trained deep neural network model as a real-time controller of the angle of attack and the roll angle to quickly generate a penetration trajectory based on the current state parameters of the hypersonic gliding aircraft. In the actual application process, the penetration trajectory can be obtained by simply inputting the obtained current state parameters, so as to improve the solution efficiency of the convex optimization problem, thereby meeting the real-time requirements of the highly dynamic environment of the HGV.

Description

Technical Field

[0001] The present invention relates to the technical field of trajectory planning, and in particular to a method and system for generating a real-time penetration trajectory of an aircraft. Background Art

[0002] Hypersonic Gliding Vehicle (HGV) is a near-space vehicle with high speed, high maneuverability and long-range gliding capability, and has broad application prospects. In order to counter missile defense systems, it is necessary to improve the penetration capability of HGV. The effective penetration method of HGV is to make full use of its speed advantage. Due to the development of cooperative guidance technology, HGV needs to penetrate multiple interceptors and reach the target area. In addition, the engagement process is highly dynamic, and real-time decision-making performance is crucial for HGV.

[0003] Trajectory optimization is an important area in aerospace. For HGV, trajectory optimization is an effective way to achieve penetration. However, there are few studies on HGV penetration of multiple interceptors, which is an important topic. In addition, the solution process of existing penetration trajectory optimization methods takes several seconds, which does not meet the real-time requirements of engagements. Convex optimization, as a direct numerical method for trajectory optimization, shows good potential in real-time applications. Convex optimization has achieved real-time applications in many fields, but the solution time is still in seconds or sub-seconds, which cannot meet the real-time requirements of HGV highly dynamic environments. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a method and system for generating a real-time penetration trajectory of an aircraft.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for generating a real-time penetration trajectory of an aircraft, comprising:

[0007] Construct the penetration trajectory optimization problem of hypersonic glide vehicle;

[0008] A second-order cone programming method is used to solve the penetration trajectory optimization problem to generate a data set; the data set includes multiple sets of training sample pairs; each of the training sample pairs includes state parameters of the hypersonic gliding vehicle and control commands corresponding to the state parameters;

[0009] Using the data set to train a deep neural network to obtain a trained deep neural network model;

[0010] Acquiring current state parameters of the hypersonic glide vehicle, and inputting the current state parameters of the hypersonic glide vehicle into the trained deep neural network model to obtain control commands; the obtained control commands include angle of attack control commands and roll angle control commands;

[0011] A penetration trajectory of the hypersonic glide vehicle is generated according to the control command.

[0012] Preferably, the construction of the penetration trajectory optimization problem of the hypersonic glide vehicle specifically includes:

[0013] Constructing control constraints based on the dynamic model and aerodynamic model of the hypersonic glide vehicle; the control constraints include angle of attack constraints and roll angle rate constraints;

[0014] Obtaining a first objective function and a second objective function; the first objective function is used to characterize that the hypersonic glide vehicle reaches the target area; the second objective function is used to characterize that the speed direction of the hypersonic glide vehicle will reach the desired penetration direction within a preset time period;

[0015] Generate a third objective function according to the combination of the first objective function and the second objective function;

[0016] A penetration trajectory optimization problem of a hypersonic glide vehicle is constructed based on the third objective function and the control constraint; the penetration trajectory optimization problem is to obtain a control quantity that minimizes the third objective function under the condition of satisfying the dynamic model and the control constraint.

[0017] Preferably, the penetration trajectory optimization problem is:

[0018]

[0019] Among them, J is the third objective function, x pf is the terminal value of the horizontal coordinate position of the hypersonic glide vehicle on the ground, y pf is the terminal value of the ordinate position of the hypersonic glide vehicle on the ground, are constants, t0 and t I are all moments, θ is the flight path angle, ψ is the flight heading angle, θ ex is the desired flight path angle, ψ ex is the desired flight heading angle, dt is the time derivative, is the variable of the state quantity, f() is the state function, x0 is the initial state quantity, x(t0) is the state quantity at time t0, α(t) is the angle of attack at time t, α min is the minimum angle of attack, α max is the maximum angle of attack, is the roll angle rate at time t, is the maximum roll angle rate.

[0020] Preferably, a second-order cone programming method is used to solve the penetration trajectory optimization problem, which specifically includes:

[0021] Continuously linearizing and discretizing the kinetic model to obtain a first kinetic model;

[0022] When encountering an interceptor, a slack variable is introduced to discretize the third objective function to obtain a linear function;

[0023] generating a convex optimization problem based on the linear function and the second-order cone constraint with the first dynamic model and the control constraint as constraints;

[0024] Acquire an initial state; the initial state includes: an initial state quantity, an initial trust region, and an isometric quantity;

[0025] Determine the initial trajectory profile of the convex optimization problem according to the initial state by adopting a penetration strategy;

[0026] Determining the current trajectory profile of the convex optimization problem using a penetration strategy according to the initial trajectory profile;

[0027] Determine the next trajectory profile of the convex optimization problem by adopting a penetration strategy according to the current trajectory profile;

[0028] Determining whether a stop condition is satisfied according to the state quantity in the current trajectory profile and the state quantity in the next trajectory profile;

[0029] When the stop condition is met, the next trajectory profile is the optimal solution;

[0030] When the stopping condition is not met, the trust domain is updated according to the variable trust region strategy, and after replacing the initial trust domain with the updated trust domain, the process returns to "determining the initial trajectory profile of the convex optimization problem according to the initial state and the preset trust domain using the penetration strategy".

[0031] Preferably, the updated trust region is δ:

[0032]

[0033] δ0 is the initial trust region, l1 and l2 are preset parameters, and e is the natural logarithm.

[0034] Preferably, before using the data set to train a deep neural network to obtain a trained deep neural network model, the method further includes:

[0035] The data in the data set is preprocessed.

[0036] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0037] The real-time penetration trajectory generation method of the aircraft provided by the present invention solves the penetration trajectory optimization problem by using the second-order cone programming (SOCP) method to obtain the penetration trajectory under different initial conditions, and after generating a data set, the deep neural network is trained using the generated data set to obtain a trained deep neural network model, and then the trained deep neural network model is used as a real-time controller of the angle of attack and the roll angle to quickly generate the penetration trajectory based on the current state parameters of the hypersonic gliding aircraft. In the actual application process, the penetration trajectory can be obtained by only inputting the obtained current state parameters, so as to improve the solution efficiency of the convex optimization problem, thereby meeting the real-time requirements of the highly dynamic environment of the HGV.

[0038] Corresponding to the above-mentioned method for generating a real-time penetration trajectory of an aircraft, the present invention further provides a system for generating a real-time penetration trajectory of an aircraft, the system comprising:

[0039] The penetration trajectory optimization problem construction module is used to construct the penetration trajectory optimization problem of hypersonic glide vehicles;

[0040] An optimization problem solving module, used for solving the penetration trajectory optimization problem by using a second-order cone programming method to generate a data set; the data set includes a plurality of training sample pairs; each of the training sample pairs includes a state parameter of a hypersonic glide vehicle and a control command corresponding to the state parameter;

[0041] A neural network model training module is used to train a deep neural network using the data set to obtain a trained deep neural network model;

[0042] A control command generation module is used to obtain current state parameters of the hypersonic glide vehicle, and input the current state parameters of the hypersonic glide vehicle into the trained deep neural network model to obtain control commands; the obtained control commands include angle of attack control commands and roll angle control commands;

[0043] The penetration trajectory generation module is used to generate a penetration trajectory of the hypersonic glide vehicle according to the control command.

[0044] Preferably, the penetration trajectory optimization problem building module includes:

[0045] A control constraint construction unit, used to construct control constraints based on a dynamic model and an aerodynamic model of the hypersonic glide vehicle; the control constraints include an angle of attack constraint and a roll angle rate constraint;

[0046] An objective function acquisition unit is used to acquire a first objective function and a second objective function; the first objective function is used to characterize that the hypersonic glide vehicle reaches the target area; the second objective function is used to characterize that the speed direction of the hypersonic glide vehicle will reach the desired penetration direction within a preset time period;

[0047] An objective function generating unit, configured to generate a third objective function according to the combination of the first objective function and the second objective function;

[0048] A penetration trajectory optimization problem construction unit is used to construct a penetration trajectory optimization problem of a hypersonic glide vehicle based on the third objective function and the control constraint; the penetration trajectory optimization problem is to obtain a control quantity that minimizes the third objective function under the condition of satisfying the dynamic model and the control constraint.

[0049] Preferably, the optimization problem solving module includes:

[0050] A kinetic model processing unit, used for performing continuous linearization and discretization processing on the kinetic model to obtain a first kinetic model;

[0051] A linear function determination unit, used for introducing a slack variable to discretize the third objective function to obtain a linear function when encountering an interceptor;

[0052] A convex optimization problem generating unit, configured to generate a convex optimization problem based on the linear function and the second-order cone constraint with the first dynamic model and the control constraint as constraints;

[0053] An initial state acquisition unit, used to acquire an initial state; the initial state includes: an initial state quantity, an initial trust region and an isometric quantity;

[0054] An initial trajectory profile determination unit, configured to determine an initial trajectory profile of the convex optimization problem according to the initial state by adopting a penetration strategy;

[0055] A current trajectory profile determination unit, configured to determine a current trajectory profile of the convex optimization problem by adopting a penetration strategy according to the initial trajectory profile;

[0056] A next trajectory profile determination unit, configured to determine the next trajectory profile of the convex optimization problem by adopting a penetration strategy according to the current trajectory profile;

[0057] A judging unit, used for judging whether a stop condition is satisfied according to the state quantity in the current trajectory profile and the state quantity in the next trajectory profile;

[0058] An optimal solution determination unit, used for determining that the next trajectory profile is the optimal solution when a stop condition is met;

[0059] The loop solving unit is used for updating the trust domain according to the variable trust region strategy when the stopping condition is not met, and replacing the initial trust domain with the updated trust domain, and then returning to "determining the initial trajectory profile of the convex optimization problem according to the initial state and the preset trust domain using the penetration strategy".

[0060] Preferably, it also includes:

[0061] The preprocessing module is used to preprocess the data in the data set.

[0062] The technical effect achieved by the aircraft real-time penetration trajectory generation system provided by the present invention is the same as the technical effect achieved by the aircraft real-time penetration trajectory generation method provided above, so it will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0064] Figure 1 A flow chart of a method for generating a real-time penetration trajectory of an aircraft provided by the present invention;

[0065] Figure 2 A schematic diagram of a penetration scenario provided by an embodiment of the present invention;

[0066] Figure 3 A deep neural network training framework diagram provided by an embodiment of the present invention;

[0067] Figure 4 A trajectory diagram generated by the second-order cone programming provided in an embodiment of the present invention;

[0068] Figure 5 A simulation diagram of the training process of the first network provided in an embodiment of the present invention;

[0069] Figure 6 A simulation diagram of the training process of the second network provided by an embodiment of the present invention;

[0070] Figure 7 A penetration process diagram of tasks 1-3 provided in an embodiment of the present invention;

[0071] Figure 8 An angle of attack curve diagram for tasks 1-3 provided by an embodiment of the present invention;

[0072] Fig. 9 A roll angle curve diagram for tasks 1-3 provided in an embodiment of the present invention;

[0073] Fig.10 A schematic structural diagram of the aircraft real-time penetration trajectory generation system provided by the present invention. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The purpose of the present invention is to provide a method and system for generating real-time penetration trajectories for an aircraft, which can improve the efficiency of solving convex optimization problems and thus meet the real-time requirements of HGV highly dynamic environments.

[0076] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0077] like Figure 1 As shown, the method for generating a real-time penetration trajectory of an aircraft provided by the present invention comprises:

[0078] Step 100: Construct a penetration trajectory optimization problem for a hypersonic glide vehicle. The constructed penetration trajectory optimization problem is:

[0079]

[0080] Among them, J is the third objective function, x pf is the terminal value of the horizontal coordinate position of the hypersonic glide vehicle on the ground, y pf is the terminal value of the ordinate position of the hypersonic glide vehicle on the ground, are constants, t0 and t I are all moments, θ is the flight path angle, ψ is the flight heading angle, θ ex is the desired flight path angle, ψ ex is the desired flight heading angle, dt is the time derivative, is the variable of the state quantity, f() is the state function, x0 is the initial state quantity, x(t0) is the state quantity at time t0, α(t) is the angle of attack at time t, α min is the minimum angle of attack, α max is the maximum angle of attack, is the roll angle rate at time t, is the maximum roll angle rate.

[0081] Step 101: A second-order cone programming method is used to solve the penetration trajectory optimization problem to generate a data set. The data set includes multiple sets of training sample pairs. Each training sample pair includes state parameters of the hypersonic glide vehicle and control commands corresponding to the state parameters.

[0082] Step 102: Use the data set to train the deep neural network to obtain a trained deep neural network model.

[0083] Step 103: Obtain the current state parameters of the hypersonic glide vehicle, and input the current state parameters of the hypersonic glide vehicle into the trained deep neural network model to obtain control commands. The obtained control commands include angle of attack control commands and roll angle control commands.

[0084] Step 104: Generate a penetration trajectory of the hypersonic glide vehicle according to the control command.

[0085] The following takes the penetration trajectory optimization problem when a hypersonic glide vehicle encounters two interceptors as an example to illustrate the specific implementation process of the above-mentioned aircraft real-time penetration trajectory generation method.

[0086] The penetration scenario in this embodiment mainly focuses on the gliding phase of the HGV. After gliding to the airspace near the stationary target, the HGV can follow a downward pressure trajectory to strike the stationary target. Figure 2 The penetration scenario of the HGV in the taxiing phase is shown. And the HGV needs to reach the target area after the penetration so that the vehicle can effectively attack the stationary target in the downward phase.

[0087] Since the penetration process of HGV encountering the interceptor is the main research object, the penetration trajectory optimization problem can be regarded as a short-range trajectory optimization problem, so the curvature and rotation of the earth are ignored. Based on this, the dynamic model of HGV is as follows:

[0088]

[0089] Where h is the flight altitude of the HGV, x p and p represents the coordinate position of the aircraft on the ground, v is the speed of the aircraft, θ and ψ are the flight path angle and heading angle, α and σ are the angle of attack and roll angle, L and D are the lift and drag, g is the acceleration of gravity, and m is the mass of the aircraft. ye and q ze are the elevation and azimuth of the HGV to the interceptor's line of sight, respectively. Figure 2 Demonstrates clear geometric meaning. (·) ,1 and(·) ,2 Respectively represent the corresponding variables of the two interceptors. i is the relative distance between the ith interceptor and the HGV, qy and q z are the elevation and azimuth of the interceptor’s line of sight to the HGV. i is the interceptor's velocity, θ i and ψ i They are the flight trajectory angle and heading angle of the interceptor respectively.

[0090] The aerodynamic model of the HGV is shown below:

[0091]

[0092] in is the atmospheric density, ρ0 is the atmospheric density at sea level, and h s =6700m is a constant, S is the reference area of ​​the aircraft. The aerodynamic lift coefficient and drag coefficient are expressed as C L =C L0 +C Lα α and Among them C L0 , C Lα , C D0 and is an aerodynamic parameter.

[0093] Since the range of roll angle is very wide and the direct numerical method may lead to sudden changes in roll angle within a single time step, the HGV may not be able to track the obtained trajectory well. Therefore, the roll angle rate is selected as the control variable of the SOCP method. Since the range of angle of attack is usually small, the angle of attack is directly used as another control variable. During flight, the control variables are limited. The control constraints are as follows:

[0094]

[0095] Based on the above, in the process of constructing the penetration trajectory optimization problem, the HGV needs to reach the target area. Since the target is the origin of the penetration coordinate system, the first objective function is defined as:

[0096] J1=||(x pf ,y pf )|| (4)

[0097] Among them, x pf and pf They are x p and p The terminal value of .

[0098] Once the interceptor is found, the HGV will maneuver in time. Then, the speed direction should be close to the desired penetration direction. Therefore, the following second objective function is defined:

[0099]

[0100] Among them, θ ex and ψ ex are the desired flight trajectory angle and the desired heading angle, respectively, which are determined by the penetration strategy proposed below.

[0101]

[0102]

[0103] in

[0104]

[0105]

[0106] q ye0 and q ze0 are the initial sight elevation angle and initial sight azimuth angle of HGV to interceptor, respectively, with subscript (·) ,1 and(·) ,2 They represent the corresponding variables for the two interceptors respectively, and χ is a small positive constant.

[0107] The above formula (5) indicates that the speed direction of the HGV will be at the time [t0,t I ] is close to the expected penetration direction. Since the HGV must reach the target area after penetration, (4) and (5) are combined into the following objective function (i.e., the third objective function) by adding a penalty term:

[0108]

[0109] Among them, the constant

[0110] make and x=[h,x p ,y p ,v,θ,ψ,σ] T represent the control quantity and state quantity respectively, and the state equation (1) can be expressed as follows.

[0111]

[0112] The mathematical description of the HGV penetration trajectory optimization problem is as follows:

[0113]

[0114] The goal of this problem is to find the optimal control that minimizes the objective function (10) while satisfying the dynamics (1) and control constraints (3).

[0115] Based on the above construction, the penetration trajectory optimization problem of the hypersonic gliding vehicle is as follows: Figure 3 As shown, the second-order cone programming (SOCP) method is used to solve different penetration trajectory optimization problems to generate the dataset.

[0116] Specifically, in a penetration scenario, the terminal flight time of the HGV is difficult to determine, so a conversion of variables is performed.

[0117] x p The derivation can be obtained

[0118]

[0119] Because the velocity v is always positive, and cosθ>0, cosψ<0 are easily satisfied in the penetration scenario. Therefore, x p It decreases monotonically with time, so it can be regarded as the independent variable of the state equation.

[0120] The kinetic equation (1) can be transformed into

[0121]

[0122] where x=[h,y p ,v,θ,ψ,σ] T is the state vector, and the control vector is still In this embodiment, and ·′ represent the time and x respectively. p The derivative of . We can get p The state equation with is the independent variable is:

[0123]

[0124] One of the main tasks of convexification of penetration problem is how to convexify nonlinear dynamics. (k) ,u (k) ) is linearized successively, and a convex constraint is obtained:

[0125] x′=f(x,u,x p )≈A(x (k) ,u (k) ,x p )x+B(x (k) ,u (k) ,x p )u+c(x (k) ,u (k) ,x p ) (16)

[0126] in, And c = f(x,u,t)-Ax-Bu.

[0127] In order to ensure the convergence of the continuous linear approximation, the search space must be restricted to the trust region, that is:

[0128] |xx (k) |≤δ 0. (17)

[0129] In the formula, is a constant vector.

[0130] The penetration process is averaged through N+1 discrete points, and the penetration process is discretized into N equidistant points. Then, the infinite-dimensional optimal control problem is transformed into a finite-dimensional optimal control problem, and constraints are imposed on each discrete point. The step size is Δx p =(x pf -x p0 ) / N and the discrete points are represented by {x p0 ,x p1 ,x p2 ,…,x pN-1 ,x pN}. The corresponding state and control are discretized into x i =x(x pi ) and u i =u(x pi ), i = 0, 1, ..., N. Then, the kinetic equation can be numerically integrated as follows:

[0131]

[0132] in, and

[0133] make and Formula (18) can be further written as:

[0134]

[0135] Where I is the identity matrix with appropriate dimensions. The optimization variable vector can be expressed as

[0136] After continuous linearization and discretization of the dynamics, the dynamic equation can be transformed into:

[0137] Mz=F (20)

[0138] in,

[0139] It should be noted that the above formula (20) also includes the initial state constraint x(xp0 )=x0.

[0140] In the penetration coordinate system, we can get x p0 > 0. When encountering an interceptor, the HGV's primary goal is to penetrate. Set x pf =0, the objective function in problem P0 can be converted to:

[0141]

[0142] Since the objective function can also be discretized, by introducing two slack variables and The above nonlinear objective function can be easily converted into a linear function form:

[0143]

[0144] Subject to the following inequality constraints:

[0145]

[0146] The constraint in formula (23) is a second-order cone constraint and is convex.

[0147] The original problem P0 can be reformulated as a convex optimization problem P1, namely:

[0148]

[0149] Subject to the following constraints:

[0150] Mz=F(25)

[0151] α min ≤α i ≤α max (26)

[0152]

[0153]

[0154]

[0155] Where i = 1, ..., N, and is a user-defined slack variable and the superscript k is the current iteration.

[0156] The subsequent SOCP method for solving problem P1 is described in detail as follows.

[0157] Step 1: Input the initial state x0, δ0 and N. Calculate θ according to the penetration strategy ex and ψ ex . Set k = 0 and select the initial trajectory profile (x(0) ,u (0) ) and determine parameters l1 and l2.

[0158] Step 2: At iteration k+1, problem P1 is solved by using the result of the previous iteration (x (k) ,u (k) ). By solving problem P1, we can get (x (k+1) ,u (k+1) ).

[0159] Step 3: Check whether the following stop conditions are met. If the convergence conditions are met, the optimal solution is (x (k+1) ,u (k +1) ), otherwise, go to step 4. The stopping condition is:

[0160]

[0161] in, is a user-defined constant vector.

[0162] Step 4: Set and update the trust zone according to the following flexible trust zone policy settings. Then return to step 2.

[0163] Among them, the variable trust region strategy is:

[0164] δ=sgm(k)δ0(31)

[0165]

[0166] in, is a user-defined constant vector.

[0167] A large dataset containing different penetration trajectories was generated for the above HGV penetration problem. The initial state of the HGV is disturbed by uncertain parameters, as shown in the following equation:

[0168]

[0169] Among them, ξ n is a set of random parameters with known distribution. This data set can be obtained by repeatedly solving the penetration trajectory optimization problem using the above SOCP method until n reaches N d Therefore, each ξ n will correspond to a set of trajectories (x n ,u n ).

[0170] Based on the above constructed data set, in this embodiment, the angle of attack and the roll angle are selected as the control variables, that is, u = [α, σ] T , x=[h,xp ,y p ,v,θ,ψ] T .like Figure 3 As shown, the two networks are used to output the angle of attack and roll angle control commands respectively. The inputs of these two networks are the normalized state after data processing, the normalized expected flight path angle, and the normalized expected heading angle. The input of each network is an 8×1 vector, and the output of each network is a scalar.

[0171] Although two networks are used in the control scheme in this embodiment, the structures of the two networks are the same. Generally speaking, a DNN consists of an input layer, several hidden layers, and an output layer. In this embodiment, the DNN has 8 fully connected hidden layers, each with 128 units. In addition, a rectified linear unit (ReLU) is used as the activation function.

[0172] κ j (x)=max(x,0),j=1,2,...,N L -1 (34)

[0173]

[0174] where κ j represents the activation function of the jth layer, and the bounded function (35) is used to impose control constraints. and u are the upper and lower bounds of the control.

[0175] Based on the above network structure, there are various data types in the original input state vector, such as height, speed, and angle. Since there is no prior knowledge about which features are relevant, it is necessary to avoid assigning weights to some features that are greater than the weights of other features. In order to achieve satisfactory results during training, the data needs to be normalized. Input data normalization is shown below.

[0176]

[0177] Where E(x) and De(x) represent the mean and standard deviation of x. The input features are rescaled to have zero mean and unit variance.

[0178] The processed data can be used as training data for DNN. DNN can be updated in each round of training so that the loss function decreases. The mean absolute error (MAE) loss (37) is selected to evaluate the difference between the predicted value and the target value.

[0179]

[0180] Among them, N b is the number of training samples, is the final output of DNN, and y* is the corresponding target output.

[0181] During the training process, the Adam optimization algorithm is used as the optimizer.

[0182] Based on the above-trained deep neural network model, this embodiment performs simulation verification on it.

[0183] During the verification process, the trajectory dataset was generated by the SOCP method. Table 1 lists the initial values ​​and parameters of the SOCP method. The parameters of formula (32) are selected as l1 = 2.5 and l2 = 5. The constants in formula (10) are selected as The initial trust region radius and stopping condition are as follows:

[0184]

[0185]

[0186] The random parameters and their distributions are listed in Table 2. The guess u is controlled by giving a constant (0) =[2°,0] T , and obtain the initial trajectory of the SOCP method iteration. The SOCP method generates n = 5000 penetration trajectories for training and 160 trajectories for testing. N = 200 state-control variables are selected along each trajectory. Figure 4 A total of 5160 randomly sampled trajectories with varying initial conditions are shown.

[0187] Table 1 Initial conditions for second-order cone programming

[0188]

[0189] Table 2 Distribution of ξ

[0190]

[0191] During training, the batch size is chosen to be N b = 128. The learning rate is set to 10 during training -4 , and adjusted to 10 in the later stage of training -5 , to obtain better training results. The test results and simulations were obtained on a desktop computer equipped with an Intel Core i9-10900K 3.70GHz processor and 32GB of memory. The DNN was trained on an Nvidia GeForce RTX3080.

[0192] for Figure 3 Network in Net α and NetworkNet σ , when the learning rate is 10-4 500 trainings were performed under the condition of , and then the learning rate was adjusted to 10 -5 , and continue training for 100 times to allow the training process to converge better. After each training stage, the average loss of the training set and the test set is calculated. Figure 5 and Figure 6 As shown in Figure 2, after 600 trainings, the network Net α The training loss is about 2×10 -4 , the test loss is about 3×10 -4 , and the network Net σ The training loss is about 1.5×10 -5 , the test loss is about 3×10 -5 .

[0193] During this verification process, this embodiment simulated three tasks (tasks 1-3) with different initial states to verify the effectiveness of the above-mentioned method for generating real-time penetration trajectories for aircraft. Details of the three tasks are shown in Table 3. The initial conditions of Task 1 are in the generated data set. The initial conditions of Task 2 and Task 3 are located on the boundary of the data set and outside the data set, respectively. All interceptors in the simulation use the proportional guidance law (PNG), the guidance constant is 5, the speed of the interceptor is 1500m / s, and the maximum overload limit is 6g. Figure 7 As shown in Table 3, the results of DNN control show that the HGV can penetrate the interceptor through longitudinal maneuvers. The control instructions generated by the two DNNs are as follows: Figure 8 and Fig. 9 The miss distances of the two interceptors are shown in Table 3, and the average CPU time for generating control commands at each step is less than 0.6 milliseconds. Therefore, the DNN-based method can generate penetration trajectories in real time.

[0194] Table 3 Simulation details of tasks 1-3

[0195]

[0196] Based on the above description, the present embodiment provides a method for generating a real-time penetration trajectory for an aircraft. The HGV penetration problem is a nonlinear trajectory optimization problem. A second-order cone programming (SOCP) method is used to solve the penetration trajectory under different initial conditions and generate a data set. Cai Yaozong trained two DNNs using this data set. The two DNNs are used as real-time controllers of the angle of attack and the roll angle to generate the penetration trajectory. Finally, the simulation results show that the control commands generated by the two DNNs can enable the HGV to penetrate the two interceptors and meet the real-time requirements.

[0197] In addition, corresponding to the above-mentioned aircraft real-time penetration trajectory generation method, the present invention also provides an aircraft real-time penetration trajectory generation system, such as Fig.10 As shown, the system includes:

[0198] The penetration trajectory optimization problem construction module 1 is used to construct the penetration trajectory optimization problem of the hypersonic glide vehicle.

[0199] The optimization problem solving module 2 is used to solve the penetration trajectory optimization problem by using the second-order cone programming method to generate a data set. The data set includes multiple sets of training sample pairs. Each training sample pair includes the state parameters of the hypersonic gliding vehicle and the control commands corresponding to the state parameters.

[0200] The neural network model training module 3 is used to train the deep neural network using the data set to obtain a trained deep neural network model.

[0201] The control command generation module 4 is used to obtain the current state parameters of the hypersonic glide vehicle and input the current state parameters of the hypersonic glide vehicle into the trained deep neural network model to obtain the control command. The obtained control command includes the angle of attack control command and the roll angle control command.

[0202] The penetration trajectory generation module 5 is used to generate the penetration trajectory of the hypersonic glide vehicle according to the control command.

[0203] In order to further improve the accuracy of trajectory generation, as an embodiment of the present invention, the penetration trajectory optimization problem construction module 1 adopted above includes:

[0204] The control constraint building unit is used to build control constraints based on the dynamic model and aerodynamic model of the hypersonic gliding vehicle. The control constraints include angle of attack constraints and roll angle rate constraints.

[0205] The objective function acquisition unit is used to acquire a first objective function and a second objective function. The first objective function is used to characterize that the hypersonic glide vehicle reaches the target area. The second objective function is used to characterize that the speed direction of the hypersonic glide vehicle will reach the desired penetration direction within a preset time period.

[0206] The objective function generating unit is used to generate a third objective function according to the combination of the first objective function and the second objective function.

[0207] The penetration trajectory optimization problem construction unit is used to construct the penetration trajectory optimization problem of the hypersonic glide vehicle based on the third objective function and the control constraint. The penetration trajectory optimization problem is to obtain the control quantity that minimizes the third objective function under the condition of satisfying the dynamic model and the control constraint.

[0208] Similarly, in order to further improve the accuracy, as another embodiment of the present invention, the optimization problem solving module adopted above includes:

[0209] The dynamic model processing unit is used to perform continuous linearization and discretization processing on the dynamic model to obtain a first dynamic model.

[0210] The linear function determination unit is used to introduce a slack variable to discretize the third objective function to obtain a linear function when encountering an interceptor.

[0211] The convex optimization problem generating unit is used to generate a convex optimization problem based on a linear function and a second-order cone constraint with the first dynamic model and the control constraint as constraints.

[0212] The initial state acquisition unit is used to acquire the initial state. The initial state includes: initial state quantity, initial trust region and isometric quantity.

[0213] The initial trajectory profile determination unit is used to determine the initial trajectory profile of the convex optimization problem according to the initial state by adopting the penetration strategy.

[0214] The current trajectory profile determination unit is used to determine the current trajectory profile of the convex optimization problem by adopting a penetration strategy according to the initial trajectory profile.

[0215] The next trajectory profile determination unit is used to determine the next trajectory profile of the convex optimization problem by adopting a penetration strategy according to the current trajectory profile.

[0216] The judging unit is used to judge whether the stop condition is satisfied according to the state quantity in the current trajectory profile and the state quantity in the next trajectory profile.

[0217] The optimal solution determination unit is used to determine that the next trajectory profile is the optimal solution when the stop condition is met.

[0218] The loop solving unit is used for updating the trust domain according to the variable trust region strategy when the stopping condition is not met, and replacing the initial trust domain with the updated trust domain, and then returning to "determining the initial trajectory profile of the convex optimization problem according to the initial state and the preset trust domain by adopting the penetration strategy".

[0219] In addition, in order to reduce errors and improve data training efficiency, as another embodiment of the present invention, the above-mentioned aircraft real-time penetration trajectory generation system preferably also includes: a preprocessing module for preprocessing the data in the data set.

[0220] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0221] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for generating a real-time penetration trajectory of an aircraft, characterized in that: include: Construct the penetration trajectory optimization problem of hypersonic glide vehicle; A second-order cone programming method is used to solve the penetration trajectory optimization problem to generate a data set; the data set includes multiple sets of training sample pairs; each of the training sample pairs includes state parameters of the hypersonic gliding vehicle and control commands corresponding to the state parameters; Using the data set to train a deep neural network to obtain a trained deep neural network model; Acquiring current state parameters of the hypersonic glide vehicle, and inputting the current state parameters of the hypersonic glide vehicle into the trained deep neural network model to obtain control commands; the obtained control commands include angle of attack control commands and roll angle control commands; generating a penetration trajectory of the hypersonic glide vehicle according to the control command; Among them, constructing the penetration trajectory optimization problem of hypersonic gliding vehicle specifically includes: Constructing control constraints based on the dynamic model and aerodynamic model of the hypersonic gliding vehicle; the control constraints include angle of attack constraints and roll angle rate constraints; Obtaining a first objective function and a second objective function; the first objective function is used to characterize that the hypersonic glide vehicle reaches the target area; the second objective function is used to characterize that the speed direction of the hypersonic glide vehicle will reach the desired penetration direction within a preset time period; Generate a third objective function according to the combination of the first objective function and the second objective function; A penetration trajectory optimization problem of a hypersonic glide vehicle is constructed based on the third objective function and the control constraint; the penetration trajectory optimization problem is to obtain a control quantity that minimizes the third objective function under the condition of satisfying the dynamic model and the control constraint.

2. The method for generating a real-time penetration trajectory of an aircraft according to claim 1, characterized in that: The penetration trajectory optimization problem is: Among them, J is the third objective function, x pf is the terminal value of the horizontal coordinate position of the hypersonic glider vehicle on the ground, y pf is the terminal value of the ordinate position of the hypersonic glide vehicle on the ground, is a constant, t0 and tI are both moments, θ is the flight path angle, ψ is the flight heading angle, θ ex is the desired flight path angle, ψ ex is the desired flight heading angle, dt is the time derivative, is the variable of the state quantity, f() is the state function, x0 is the initial state quantity, x(t0) is the state quantity at time t0, α(t) is the angle of attack at time t, α min is the minimum angle of attack, α max is the maximum angle of attack, is the roll angle rate at time t, is the maximum roll angle rate.

3. The method for generating a real-time penetration trajectory of an aircraft according to claim 1, characterized in that: The second-order cone programming method is used to solve the penetration trajectory optimization problem, which specifically includes: Continuously linearizing and discretizing the kinetic model to obtain a first kinetic model; When encountering an interceptor, a slack variable is introduced to discretize the third objective function to obtain a linear function; generating a convex optimization problem based on the linear function and the second-order cone constraint with the first dynamic model and the control constraint as constraints; Acquire an initial state; the initial state includes: an initial state quantity, an initial trust region, and an isometric quantity; Determining an initial trajectory profile of the convex optimization problem according to the initial state using a penetration strategy; Determining the current trajectory profile of the convex optimization problem by adopting a penetration strategy according to the initial trajectory profile; Determine the next trajectory profile of the convex optimization problem by adopting a penetration strategy according to the current trajectory profile; Determining whether a stop condition is satisfied according to the state quantity in the current trajectory profile and the state quantity in the next trajectory profile; When the stop condition is met, the next trajectory profile is the optimal solution; When the stopping condition is not met, the trust domain is updated according to the variable trust region strategy, and after replacing the initial trust domain with the updated trust domain, the process returns to "determining the initial trajectory profile of the convex optimization problem according to the initial state and the preset trust domain using the penetration strategy".

4. The method for generating a real-time penetration trajectory of an aircraft according to claim 3, characterized in that: The updated trust region is δ: δ0 is the initial trust region, l1 and l2 are preset parameters, and e is the natural logarithm.

5. The method for generating a real-time penetration trajectory of an aircraft according to claim 1, characterized in that: Before using the data set to train the deep neural network to obtain a trained deep neural network model, the method further includes: The data in the data set is preprocessed.

6. A real-time aircraft penetration trajectory generation system, characterized in that: include: The penetration trajectory optimization problem construction module is used to construct the penetration trajectory optimization problem of hypersonic glide vehicles; An optimization problem solving module, used for solving the penetration trajectory optimization problem by using a second-order cone programming method to generate a data set; the data set includes a plurality of training sample pairs; each of the training sample pairs includes a state parameter of a hypersonic glide vehicle and a control command corresponding to the state parameter; A neural network model training module is used to train a deep neural network using the data set to obtain a trained deep neural network model; A control command generation module is used to obtain current state parameters of the hypersonic glide vehicle, and input the current state parameters of the hypersonic glide vehicle into the trained deep neural network model to obtain control commands; the obtained control commands include angle of attack control commands and roll angle control commands; A penetration trajectory generation module, used to generate a penetration trajectory of the hypersonic glide vehicle according to the control command; Among them, the penetration trajectory optimization problem construction module includes: A control constraint construction unit, used to construct control constraints based on a dynamic model and an aerodynamic model of the hypersonic glide vehicle; the control constraints include an angle of attack constraint and a roll angle rate constraint; An objective function acquisition unit is used to acquire a first objective function and a second objective function; the first objective function is used to characterize that the hypersonic glide vehicle reaches the target area; the second objective function is used to characterize that the speed direction of the hypersonic glide vehicle will reach the desired penetration direction within a preset time period; An objective function generating unit, configured to generate a third objective function based on the combination of the first objective function and the second objective function; A penetration trajectory optimization problem construction unit is used to construct a penetration trajectory optimization problem of a hypersonic glide vehicle based on the third objective function and the control constraint; the penetration trajectory optimization problem is to obtain a control quantity that minimizes the third objective function under the condition of satisfying the dynamic model and the control constraint.

7. The real-time aircraft penetration trajectory generation system according to claim 6, characterized in that: The optimization problem solving module includes: A kinetic model processing unit, used for performing continuous linearization and discretization processing on the kinetic model to obtain a first kinetic model; A linear function determination unit, used for introducing a slack variable to discretize the third objective function to obtain a linear function when encountering an interceptor; A convex optimization problem generating unit, configured to generate a convex optimization problem based on the linear function and the second-order cone constraint with the first dynamic model and the control constraint as constraints; An initial state acquisition unit, used to acquire an initial state; the initial state includes: an initial state quantity, an initial trust region and an isometric quantity; An initial trajectory profile determination unit, configured to determine an initial trajectory profile of the convex optimization problem according to the initial state by adopting a penetration strategy; A current trajectory profile determination unit, configured to determine a current trajectory profile of the convex optimization problem by adopting a penetration strategy according to the initial trajectory profile; A next trajectory profile determination unit, configured to determine the next trajectory profile of the convex optimization problem by adopting a penetration strategy according to the current trajectory profile; A judging unit, used for judging whether a stop condition is satisfied according to the state quantity in the current trajectory profile and the state quantity in the next trajectory profile; An optimal solution determination unit, used for determining that the next trajectory profile is the optimal solution when a stop condition is met; The loop solving unit is used for updating the trust domain according to the variable trust region strategy when the stopping condition is not met, and replacing the initial trust domain with the updated trust domain, and then returning to "determining the initial trajectory profile of the convex optimization problem according to the initial state and the preset trust domain using the penetration strategy".

8. The real-time aircraft penetration trajectory generation system according to claim 6, characterized in that: Also includes: The preprocessing module is used to preprocess the data in the data set.

Images