Intelligent real-time cooperative guidance method for hypersonic gliding aircraft

Abstract

The invention provides an intelligent cooperative guidance method for a hypersonic gliding aircraft. The method comprises the following steps: decoupling an overall guidance problem into a longitudinal channel and a transverse channel; for the longitudinal channels, a neural network and a proportional-differential feedback control law are used for generating control quantities respectively, weighted summation is carried out to control the arrival time, and a weighting coefficient is related to the speed of the aircraft; for a transverse channel, a rapid threat avoidance planner based on a geometric visible graph is adopted to generate a reference path for avoiding a no-fly zone, and path tracking and terminal guidance are performed in combination with a neural network and a proportional-differential feedback control law. Compared with an existing single aircraft guidance method, through channel decoupling and a mixed guidance strategy, the problem of time collaborative strike of multiple aircrafts under the constraint of a no-fly zone is solved, and under the condition that uncertain factors exist, avoidance safety, time collaboration and terminal strike precision can still be considered; and the cooperative combat effectiveness and the system robustness are obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aircraft gliding segment guidance law design, and in particular to an intelligent real-time cooperative guidance method for hypersonic gliding aircraft. BACKGROUND

[0002] As the core and key link of multi-hypersonic gliding aircraft cooperative combat, the accurate time cooperative guidance law design has a decisive influence on breaking through the enemy's defense system and improving the overall task success rate and system fault tolerance. For example, the goal of multi-hypersonic gliding aircraft cooperative combat is to form a saturated attack through the number advantage and cooperative effect, so as to overload the enemy's early warning and interception resources. Not only does each aircraft need to have high-precision attack capability, but also the entire aircraft cluster needs to be controlled to reach the designated target area according to the predetermined time or at the same time under the premise of meeting their own constraints, so as to achieve the multiplication of combat effectiveness.

[0003] Due to the high sensitivity of the flight trajectory of a single aircraft to initial state errors, atmospheric environment uncertainties and other uncertain factors, these small uncertainties will be significantly amplified during the long-time unpowered gliding process, resulting in a large time cumulative error between the aircrafts. Therefore, the traditional guidance algorithm aimed at optimizing the performance of a single aircraft cannot directly cope with the strong coupling and distributed characteristics inherent in the multi-hypersonic gliding aircraft cooperative task, and it is difficult to guarantee the final cooperative attack effect.

[0004] In addition, the existing research on the time cooperative problem of multi-hypersonic gliding aircraft under the constraint of no-fly zones is still insufficient, forming an obvious research gap.

[0005] Therefore, it is also necessary to develop a time cooperative guidance law for multi-hypersonic gliding aircraft under the constraint of no-fly zones. SUMMARY

[0006] In order to solve the time cooperative problem of multi-hypersonic gliding aircraft, the present application proposes an intelligent real-time cooperative guidance method for hypersonic gliding aircraft to meet the multiple constraints in the flight trajectory of the aircraft.

[0007] The method can further specifically involve a numerical-predictive-correction algorithm, a neural network, a path planning algorithm based on a geometric visibility graph, and a combination of a feedback control law for gliding segment time cooperative guidance. The method can be particularly further applied to the case under the constraint of no-fly zones.

[0008] One aspect of the present disclosure provides an intelligent real-time cooperative guidance method for a hypersonic glider. The method comprises: decoupling a guidance problem into a longitudinal channel and a lateral channel; for the longitudinal channel, weighting and summing control quantities generated by a neural network and a feedback control law respectively to obtain a longitudinal control quantity based on a no-fly zone, a target and state information of the glider; wherein the weighting coefficients are related to at least a speed of the glider; for the lateral channel, generating a reference path for avoiding the no-fly zone using a path planning algorithm, and generating a lateral control quantity in combination with at least a neural network, a feedback control law and a target direction discrimination mechanism.

[0009] The method can comprise:

[0010] decoupling a guidance problem into a longitudinal channel and a lateral channel;

[0011] for the longitudinal channel, inputting at least part of the state information of the glider and the target into the neural network and the feedback control law in the longitudinal channel respectively; wherein the neural network in the longitudinal channel is configured to predict a terminal longitudinal control quantity when the glider reaches the target position, and then generate a longitudinal control quantity;

[0012] weighting and summing the longitudinal control quantities generated by the neural network and the feedback control law in the longitudinal channel respectively to obtain a longitudinal control quantity output by the longitudinal channel; wherein the weighting coefficients used in the weighting and summing are related to at least a speed of the glider;

[0013] for the lateral channel, generating a reference path for avoiding the no-fly zone using a path planning algorithm, the reference path comprising: a plurality of intermediate nodes, and a plurality of intermediate paths between the plurality of intermediate nodes;

[0014] based on the plurality of intermediate nodes and the plurality of intermediate paths on the reference path, performing segmented guidance; when guiding the intermediate paths, generating a lateral control quantity via the feedback control law of the lateral channel; when guiding the target, generating a lateral control quantity via the neural network of the lateral channel.

[0015] In an example embodiment of the present disclosure, the neural network in the longitudinal channel is a terminal attack angle prediction network; wherein the proportional-differential control law in the longitudinal channel generates a control quantity comprising: calculating a longitudinal control instruction based on the target and the state information of the glider: wherein k p is a proportional coefficient; k d is a differential coefficient; e H is a difference between a height of the glider and a height constraint. Let α be the rate of change of the difference between the aircraft altitude and the altitude constraint; wherein, the weighted summation formula for the longitudinal channel is as follows: α cmd =ρ(V)α1+(1-ρ(V))α2 where α1 is the control quantity generated by the feedback control law, and α2 is the control quantity generated by the neural network; the weighting coefficient ρ(V) is in the form of: Where V is the current speed of the aircraft, V tran Let k be a preset switching speed threshold, and k be a steepness coefficient; such that: when the aircraft speed V is much higher than V... tran When ρ(V) approaches 1, the weight of the control quantity generated by the feedback control law increases; when the aircraft speed V is much lower than V0... tran When ρ(V) approaches 0, the weights of the control quantities generated by the neural network increase; wherein, the neural network of the lateral channel includes multiple prediction networks for predicting terminal miss distance, remaining range, and target direction; the path planning algorithm is a fast threat avoidance planner based on a geometrically visible graph, which generates a collision-free reference path by recursively finding the external common tangent point pair of the no-fly zone with the deepest penetration depth connecting the start and end points; wherein, the feedback control law in the lateral channel is a proportional-derivative control law, and the control quantities generated by the proportional-derivative control law include: constructing a second-order dynamic control equation based on the reference path and the aircraft's state information: Where ξ is the damping ratio, ω n e is the dimensionless natural frequency. ψ This is the difference between the line-of-sight heading angle and the desired heading angle of the aircraft to the next node on the reference path. For e ψ The first derivative with respect to time; the lateral guidance command is calculated based at least on the second-order dynamic control equations: Where sign(·) is the sign function, σ max σ is the maximum allowable tilt angle. prev The tilt angle of the previous guidance cycle, δ ψ The dead zone threshold is defined as follows: The target direction discrimination mechanism in the lateral channel is based on a dual-predicted trajectory reference mechanism. The generated control variables include: constructing a dual reference system for the target point relative to the current predicted trajectory and the reverse predicted trajectory, where the current predicted trajectory is the predicted flight path generated by the aircraft maintaining the current tilt angle, and the reverse predicted trajectory is another predicted flight path generated by the aircraft performing a tilt angle flip operation and maintaining the same control strategy; determining two critical points v1 and v2 on the horizontal plane that are closest to the target point on the two predicted trajectories, and calculating the relative position vector of the target point relative to these two critical points. and Combined with the horizontal position vector formed by the last two waypoints on the reference path Through calculation Construct a target orientation criterion in a two-dimensional plane using specific components of the cross product of two relative position vectors in the horizontal plane: Where k is the normal vector of the local horizontal plane. Based on the dual-trajectory orientation criterion. It can be determined that the target point is located within the space formed by the current predicted trajectory and the reverse predicted trajectory; at least based on the dual-trajectory orientation criterion, when the difference between the predicted remaining range and the great circle distance from the aircraft to the target is sufficiently small, a roll angle sign command is generated by comparing the predicted terminal miss distance:

[0016] In an exemplary embodiment of the present invention, the training process of the neural network includes:

[0017] Step 1.1: For multiple initial states, generate multiple sets of longitudinal and lateral trajectory data containing the aircraft state and the corresponding terminal state respectively through homotopy guidance law and numerical prediction algorithm;

[0018] Step 2.1: Extract the aircraft state variable set from the trajectory data as input, and extract terminal angle of attack, terminal miss distance, remaining range, target direction, etc. as output to construct a training sample set;

[0019] Step 2.2: training the multiple neural networks of the longitudinal and lateral channels using the training sample set; wherein the structure of the terminal attack angle prediction network is set to 1 input layer, 6 hidden layers, and 1 output layer. The input layer contains 4 neurons for receiving the state variable group of the aircraft, and each hidden layer contains 256 neurons; the output layer contains 1 neuron for outputting the terminal attack angle; wherein the structure of the terminal miss distance prediction network is set to 1 input layer, 5 hidden layers, and 1 output layer. The input layer contains 7 neurons for receiving the state variable group of the aircraft, and each hidden layer contains 256 neurons; the output layer contains 1 neuron for outputting the terminal miss distance; wherein the structure of the remaining range prediction network is set to 1 input layer, 6 hidden layers, and 1 output layer. The input layer contains 6 neurons for receiving the state variable group of the aircraft, and each hidden layer contains 256 neurons; the output layer contains 1 neuron for outputting the remaining range; wherein the structure of the target direction prediction network is set to 1 input layer, 6 hidden layers, and 1 output layer; the input layer contains 7 neurons for receiving the state variable group of the aircraft, and each hidden layer contains 256 neurons; the output layer contains 1 neuron for outputting the target direction; the output layer adds a sigmoid function as an activation function; wherein linear weight connections are used between the input layer and the hidden layers, between the hidden layers and the output layer, and between each hidden layer, and a tanh function is added as an activation function; wherein the network for regression tasks uses MSE as a loss function, and the network for binary classification tasks uses BCE as a loss function; the parameters of the neural network are optimized using the Adam method.

[0020] In the example embodiments of the present application, generating the control quantity using the neural network includes: inputting the state variable group of the aircraft at the current time into the neural network to obtain the predicted values of the terminal attack angle, the terminal miss distance, the remaining range, and the target direction in real time; for the longitudinal channel, using the predicted terminal attack angle constructing an attack angle-altitude profile to generate an attack angle command a2; for the lateral channel, generating a bank angle sign based on at least the predicted remaining range, terminal miss distance, and target direction through the target direction discrimination mechanism based on double-predicted trajectory reference.

[0021] The example embodiments of the present application also provide a flight vehicle control device, which includes a processor and a memory for storing executable instructions. When the processor invokes and runs the executable instructions stored in the memory, the flight vehicle control device performs the intelligent real-time collaborative guidance method for a hypersonic gliding flight vehicle according to any of the above embodiments.

[0022] The present application has the advantages of:

[0023] 1. In the embodiments of the present application, the guidance problem is decoupled into a longitudinal channel and a lateral channel, control quantities are generated using neural networks and feedback control laws respectively for the longitudinal channel, the control quantities generated by the neural networks and the feedback control laws are weighted and summed to obtain the required control quantity at the current time, wherein the weighting coefficients are related to at least the speed of the aircraft, and the lateral channel uses a fast threat avoidance planner to generate a reference path for flying around a no-fly zone, and the lateral channel generates control quantities using neural networks and feedback control laws respectively. In this way, time-coordinated attacks on target positions can be achieved for a multi-hypersonic gliding aircraft, and relevant constraints in the flight process (such as no-fly zones) can be met.

[0024] 2. More specifically, the present application is based on a numerical-predictive correction algorithm, neural networks, a path planning algorithm based on a geometric visibility graph, and a proportional-differential control law, a piecewise optimization strategy based on homotopy theory is used in the longitudinal guidance channel, which can adaptively adjust the control center of gravity according to the flight stage, ensuring the safety and robustness of the entire flight and the coordination of the final attack; a piecewise hybrid guidance strategy driven by situation awareness is designed in the lateral guidance channel, so that the aircraft can efficiently and safely avoid multiple no-fly zones, and at the end of the flight, through intelligent decision-making, it can accurately consume excess range and correct the heading, achieving precise capture of the target and significantly improving the task planning ability and hit accuracy in a complex counter-environment.

[0025] 3. The present application is based on a numerical-predictive correction algorithm, neural networks, a path planning algorithm based on a geometric visibility graph, and a proportional-differential control law, the terminal attack angle obtained by neural network prediction is used to generate an attack angle-altitude profile, and the terminal miss distance, remaining range, and target direction obtained by neural network prediction are used to generate a bank angle direction, taking into account the requirements of time coordination and landing point constraints.

[0026] 4. The present application is based on a numerical-predictive correction algorithm, neural networks, a path planning algorithm based on a geometric visibility graph, and a proportional-differential control law, and proposes an intelligent real-time cooperative guidance framework based on neural network prediction and proportional-differential control law, which simultaneously satisfies the no-fly zone constraints and the time coordination requirements, and also has good guidance accuracy and robustness in the presence of uncertain factors. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The figure is a schematic diagram of the intelligent real-time cooperative guidance framework of the present application based on a numerical-predictive correction algorithm, neural networks, a path planning algorithm based on a geometric visibility graph, and a proportional-differential control law.

[0028] Figure 2Definition of aircraft angle and force for the figure.

[0029] Figure 3 Definition of CAV-H aerodynamic data for the figure.

[0030] Figure 4 Definition of nominal trajectory design related parameters for the figure.

[0031] Figure 5 Definition of constraint condition parameter setting for the figure.

[0032] Figure 6 Definition of feasible flight corridor lower boundary on the altitude-velocity profile for the figure.

[0033] Figure 7 Path planning schematic for the fast threat avoidance planner for the figure.

[0034] Figure 8 Route tracking task scenario schematic for the no-fly zone obstacle avoidance for the figure.

[0035] Figure 9 Geometric principle for the target direction discrimination mechanism based on double prediction trajectory reference for the figure.

[0036] Figure 10 Monte Carlo distribution of deviations for the figure.

[0037] Figure 11A Confusion matrix heat map for the evaluation of the target direction prediction network for the figure. Figure 11B Indicator record table for the evaluation of the target direction prediction network for the figure.

[0038] Figure 12A True value-predicted value scatter plot for the evaluation of the model for the regression task for the figure. Figure 12B Indicator record table for the evaluation of the model for the regression task for the figure.

[0039] Figure 13 Reference path generated by the fast threat avoidance planner under nominal initial conditions for the figure.

[0040] Figure 14A Path deviation curve of the path tracking controller when tracking the straight line segment reference path for the figure. Figure 14B Minimum miss distance curve of the path tracking controller when tracking the circular arc segment reference path for the figure.

[0041] Figure 15 Altitude-velocity profile of all aircraft under nominal initial conditions for the figure.

[0042] Figure 16 Ground trajectory curve of all aircraft under nominal initial conditions for the figure.

[0043] Figure 17 Bank angle command curve of all aircraft under nominal initial conditions for the figure.

[0044] Figure 18 Angle of attack command curve for all vehicles under nominal initial conditions.

[0045] Figure 19 Height-velocity profile for Monte Carlo targeting all vehicles.

[0046] Figure 20 Ground track curve for Monte Carlo targeting all vehicles.

[0047] Figure 21 Synergy time error distribution scatter plot for Monte Carlo targeting.

[0048] Figure 22 Impact point distribution scatter plot for Monte Carlo targeting.

[0049] Figure 23 Key indicator box plot for Monte Carlo targeting. DETAILED DESCRIPTION

[0050] The application will be further described in detail below with examples.

[0051] The application is based on numerical-predictive correction algorithm, neural network, path planning algorithm based on geometric visibility graph and proportional-differential control law, as shown in Figure 1 .

[0052] Figure 1 In the above, the dynamic model is used to represent the characteristics of the vehicle, which can reflect that, under the input control quantity: angle of attack α, roll angle σ, the vehicle will be in what new attitude, state.

[0053] The angle of attack α is the angle between the velocity direction and the longitudinal axis of the vehicle, and the vehicle is positive when it looks up and negative when it looks down; the roll angle σ is defined as positive when it rotates clockwise when viewed from the tail of the fuselage.

[0054] In the vehicle attitude and state information, the geocentric distance r is defined as the straight-line distance from the vehicle center of mass to the center of the earth, the longitude is θ, the latitude is φ, the flight path angle γ is defined as the angle between the velocity vector V and the local horizontal plane (the horizontal plane where the north direction N and the east direction E are located), and the upward direction is positive (U is the vertical upward direction); the velocity deflection angle ψ is defined as the angle between the north direction N and the horizontal component of the velocity (the horizontal projection of V), and the clockwise rotation is positive.

[0055] Through the longitudinal channel and the lateral channel in the guidance law (which can also be referred to as control algorithm, flight control method, etc.), based on at least part of the attitude and state information of the vehicle, the control quantities angle of attack α and roll angle σ can be output again, respectively. The process is described in detail below.

[0056] Firstly, the overall control system architecture of the aircraft is decoupled into two subsystems, the longitudinal channel and the lateral channel.

[0057] In the longitudinal channel, a homotopy guidance strategy is adopted to generate the longitudinal control command. This channel contains two parallel computation paths. One is a proportional-derivative control module based on the reference altitude; the other is an intelligent prediction using a terminal attack angle prediction network, whose output is input into an attack angle-altitude profile module.

[0058] The outputs of the two modules, i.e. the control variables, are finally input into a homotopy guidance law module, which combines them using a dynamically changing weighting factor, and outputs the final longitudinal control command.

[0059] In the lateral channel, a fast threat evasion planner, i.e. a path planning algorithm based on a geometric visibility graph, plans a reference path according to the known no-fly zones and target information. The input information obtained by the aircraft sensors, i.e. part or all of the aircraft state information, is input into three independent prediction neural networks, which output key prediction values of the end state of the flight in real time, respectively. Then, together with the reference path planned by the fast threat evasion planner, they are input into a situation awareness driven piecewise hybrid guidance module, which contains a proportional-derivative control law for tracking the reference path and a numerical-predictive correction algorithm for accurate arrival at the target, and finally generates the lateral guidance command.

[0060] When designing, verifying or adjusting the guidance method, the generated longitudinal and lateral control commands act on a dynamics model. This model is used to simulate or represent the actual motion response of the aircraft, and its state update will be used as the input of the control process at the next time, thus forming a complete closed-loop control system. It should be understood that, in actual flight, the longitudinal and lateral control commands can be used as the input of the aircraft controller.

[0061] Before using the system, each prediction neural network contained therein needs to be trained offline. The data set required for training, i.e. the data pair consisting of a large number of flight state inputs (speed, attitude, etc.) and the corresponding standard results (such as the actual terminal miss distance, the remaining range, etc.), is generated in the following further description. By training on these data sets, it is ensured that the network can make fast and accurate predictions of key parameters during flight, and the role of the prediction value of the neural network will be described below.

[0062] The specific steps for obtaining the final required guidance law are described in further detail as follows:

[0063] Step 1: Find the terminal angle of attack of the angle-height profile of the longitudinal channel using the golden section method, and constitute a homotopy guidance law with the proportional-differential control law, so that the arrival time of all aircrafts is consistent; use the numerical-prediction algorithm in the lateral channel to predict the terminal miss distance, remaining range and target direction, and obtain data of multiple trajectories that satisfy the constraint conditions of guiding the aircrafts to arrive at the target position simultaneously;

[0064] The following are the definitions of each element of the trajectory obtained by the present application.

[0065] ① Dynamic equation

[0066] The partial angle and force definition of the aircraft is shown in Figure 2 . Considering that the atmosphere is relatively stationary relative to the earth, a plane earth geographic coordinate system is constructed, the geocentric distance r is defined as the straight line distance from the aircraft mass center to the earth center, the longitude is θ, the latitude is φ, the flight path angle γ is defined as the angle between the velocity vector V and the local horizontal plane (the horizontal plane where the north direction N and the east direction E are located), and the upward direction is positive (U is the vertical upward direction); the velocity deflection angle ψ is defined as the angle between the north direction N and the horizontal component of the velocity (the horizontal projection of V), and the clockwise rotation is positive; the lift is L, the upward direction is positive, the drag is D, the backward direction is positive, m is the mass of the aircraft, g is the gravitational acceleration; the angle of attack α (not shown) is the angle between the velocity direction and the longitudinal axis of the aircraft, the upward direction of the aircraft is positive, and the downward direction is negative; σ (not shown) is the roll angle, and the direction is defined as positive when looking from the tail of the fuselage in the clockwise direction, then the gliding segment dynamic equation of the aircraft can be written as:

[0067]

[0068] wherein, is the rate of change of the geocentric distance of the aircraft, is the rate of change of the longitude of the aircraft, is the rate of change of the latitude of the aircraft, is the linear acceleration of the aircraft, is the flight path angle acceleration of the aircraft, is the velocity deflection angle acceleration of the aircraft. The reference area of the aircraft is S, the air density is ρ, the dynamic pressure is q, the Mach number is Ma, the lift coefficient (Coefficient of Lift) is C L , the drag coefficient (Coefficient of Drag) is C D, the flight altitude is set at 50 km, the standard atmosphere model (US Standard Atmosphere 1976) published by the National Bureau of Standards of the United States in 1976 is selected for the variation law of air density, the physical quantities involved are all in international standard units and derived units, the vehicle model is the widely used hypersonic vehicle model CAV-H, for example, for the vehicle mass m = 907.185 kg, S = 0.484 m 2 (General Vehicle CAV-H Aerodynamic Reference Area), the dynamic pressure q and the Mach number Ma are:

[0069]

[0070] Since Figure 3 The lift coefficient and the drag coefficient are discretized with respect to the angle of attack and the Mach number, that is, for a specific combination of the angle of attack and the Mach number recorded in the table, the available coefficients have been given. However, for other combinations of angles of attack and Mach numbers that are not recorded, corresponding calculations need to be performed. For example, for the data in Figure 3 A second-order fitting is performed using the curve_fit function in the scipy.optimize package of python to obtain the coefficients required for the current vehicle state (especially the specific angle of attack and Mach number):

[0071] The aerodynamic data of CAV-H are shown in Figure 3 The aerodynamic coefficients of the vehicle in the low layer dense atmosphere can be fitted using the following general formula:

[0072]

[0073] The coefficients in the given formula (the coefficients in the formula depend on the actual application, and their values can be determined by calculation simulation or experimental verification) are:

[0074]

[0075] For example, C L0 = -0.047, C L1 = -0.0068…, the CL coefficient has a positive and negative sign. It should be understood that there can be multiple sets of different coefficients.

[0076] Then, the lift and the drag are:

[0077]

[0078] For the sake of unity, the related parameter settings of the nominal trajectory design are shown in Figure 4 .

[0079] ② Constraint conditions

[0080] To ensure the safety and the performance of the flight control system of a hypersonic glide vehicle (HGV) during the reentry phase, the thermal rate overload n and dynamic pressure q three typical constraints:

[0081]

[0082] where the maximum thermal rate the maximum overload and the maximum dynamic pressure q max is selected according to the specific vehicle. Q is a constant for calculation, which is a dimensionless empirical constant determined according to the specific vehicle and the specific position of the vehicle. Considering the maneuverability of the hypersonic vehicle, the attack angle is required to be maintained within a certain range, and the selected roll angle control quantity in this paper complies with the "bang-bang" control form, i.e. the amplitude is always constant, and the constraint form of the control quantity is as follows:

[0083]

[0084] The subscript i = 1, 2, …, n represents the i-th control quantity. t represents time.

[0085] α min is the minimum value of the control quantity α, and α max is the maximum value of the control quantity α. The value of the control quantity σ is always equal to σ max .

[0086] For the avoidance of conflicts and the guarantee of safety, the vehicle needs to fly around the no-fly zone on the path from the starting point to the target. Since the hypersonic glide vehicle usually flies in the extremely high airspace, the no-fly zone is set as an infinitely high cylinder, and the no-fly zone constraint is established:

[0087] d[P(θ,φ),P NFZ,j (θ NFZ,j ,φ NFZ,j )] > R NFZ,j

[0088] where the subscript j = 1, 2, …, n represents the j-th no-fly zone, n is the number of no-fly zones, P(θ,φ) represents the horizontal position of the vehicle determined by the longitude and latitude, P NFZ (θ NFZ ,φ NFZ ) and R NFZ represent the center position (horizontal position determined by longitude and latitude) and the radius of the no-fly zone respectively. d(·) represents the spherical distance between two points P1(θ1,φ1), P2(θ2,φ2), which is calculated by the following formula:

[0089] d[P1(θ1,φ1),P2(θ2,φ2)]=R e cos -1 (sinφ1sinφ2+cosφ1cosφ2cos(θ2-θ1))

[0090] where R e = 6378000 m is the earth radius.

[0091] Considering the position of the terminal and the cooperative time constraint:

[0092]

[0093] where H is the altitude, subscript f denotes the preset terminal variable, t co is the cooperative time, defined as the average of the arrival times of all aircraft released from multiple different positions within a certain range in the air to the same target.

[0094] The range of the angle of attack, the angle of roll and the control variable is:

[0095]

[0096] The related parameter settings for the constraint conditions are shown in Table 1. Figure 5 It should be understood that the above specific steps, values, etc. are only examples for illustrating the present application and are not limiting, and other different steps, values, etc. can also be applied.

[0097] ③Homotopy guidance law of longitudinal channel

[0098] During the reentry gliding process of a hypersonic gliding vehicle, its flight state is limited by the above-mentioned multiple dynamic constraints, which are strongly related to the altitude and speed of the vehicle. These constraints can be inversely calculated to obtain the minimum safe altitude H min (V) that the vehicle must maintain at any given speed V. Only above this altitude can all dynamic constraints be satisfied. This relationship constitutes the lower boundary of the feasible flight corridor in the altitude-speed plane, as shown in FIG. 1. Figure 6

[0099] To take into account the satisfaction of constraints throughout the flight and the accurate arrival of the terminal state, the longitudinal channel adopts a segmented optimization design. In the high-speed phase of flight, the primary task is to ensure that the vehicle safely flies within the feasible flight envelope. For this purpose, a proportional-differential control strategy based on the reference altitude is adopted:

[0100]

[0101] where k p and k d ​respectively, are the proportional and derivative gains, e H = H - H ref ref (V) is the deviation of current altitude from reference altitude H ref (V) is defined as the dynamic constraint boundary H min (V) is the margin amplification of H ref (V) = H min (V) * μ, where μ > 1 and its value is a function negatively related to the speed value. This strategy not only makes the aircraft far away from the dangerous constraint boundary, but also reserves enough maneuvering space and energy for the subsequent terminal profile correction phase. When the aircraft speed is reduced to a certain extent, or close to the terminal area, the focus of the guidance law is shifted to meet the terminal altitude constraint and flight time constraint, and the attack angle-altitude profile guidance law is used in this phase:

[0102]

[0103] where H f is the target altitude, H0 is the initial altitude of the current guidance period, is the initial attack angle of the current guidance period, is the terminal attack angle to be optimized determined by the initial altitude and initial attack angle of the current guidance period. Considering that the task scenario of the present application aims to simultaneously reach the target area as much as possible for multiple hypersonic gliding vehicles rather than strictly following the preset specified arrival time, by setting the reference coordinated time t co,ref , the objective function is constructed as a performance index for optimizing of each hypersonic gliding vehicle, i.e., to make the arrival times of all vehicles as close as possible. For example, the golden section method can be used to iteratively converge to

[0104] In order to realize the smooth transition between the above two control strategies and ensure the adaptive adjustment capability of the guidance system in different flight phases, a speed-dependent dynamic weight parameter ρ(V) is introduced. The final homotopy guidance instruction α is as follows:

[0105] α = ρ(V) α1+ (1 - ρ(V)) α2

[0106] where the weight parameter ρ(V) is designed as a sigmoid function, and the specific form is:

[0107]

[0108] where κ is the steepness coefficient set, e is the base of natural logarithm, and V tran ​This is a preset transition altitude threshold. Through this homotopy approach, the system can automatically adjust its control strategy according to the flight phase, while retaining the algorithm's ability to control the process constraints of altitude and the terminal's time constraints.

[0109] ④ Hybrid segmented lateral guidance strategy for lateral channels

[0110] The core task of lateral guidance is to guide aircraft to safely avoid no-fly zones and accurately reach the target point while satisfying all constraints. This invention proposes a hybrid segmented lateral guidance strategy, including a rapid threat avoidance planner, a path tracking controller, and situational awareness-based terminal guidance, which are used offline or low-frequency online to generate a global collision-free path, accurately track a reference path, and achieve high-precision target acquisition and range attrition in the terminal phase of flight, respectively.

[0111] To address the problem of rapid avoidance by aircraft in environments with multiple no-fly zones, this invention proposes a path planning algorithm based on geometric visibility maps. The algorithm accepts a starting point S, a target point E, and a set C = {c1, c2, ..., cn} consisting of n circular no-fly zones. n Using this as input, we progressively construct a collision-free path P that avoids all given no-fly zones. The path planning diagram is shown below. Figure 7 As shown.

[0112] Specifically, the first step is to detect the straight line segment connecting the starting point S and the target point E. If the line intersects with any no-fly zone, return the line segment as the desired path if there is no intersection; otherwise, find the specific no-fly zone c that causes the greatest penetration depth to the line. d Then calculate the distance from the starting point S to the no-fly zone c. d The two external common tangent points e1 and e2, and the distance from target point E to the no-fly zone c d The two external common tangent points are s1 and s2. In all possible pairs of tangent points (e... i ,s j In ), through heuristic functions Estimate the total length of each potential path from the starting point S to the ending point E, where Represents circle c d From e i to s j The algorithm uses a greedy strategy to select the pair of tangent points that minimize L. Then, it is assumed that this pair of tangent points is determined to be e. i s j The original path planning problem S→E is then decomposed into two independent subproblems: planning the path from S to e. i The path and from s jFind the path to E and recursively apply the same strategy until the termination condition is met. To prevent infinite recursion, the algorithm can set a maximum recursion depth limit d. max This limits the possible number of recursions and terminates the process early if the length of the new path is not significantly better than the original path. Finally, by merging the sub-paths and arc segments, a complete obstacle avoidance path P = {p1, p2, ..., p...} is formed. l}, where l represents the number of nodes. Since the aircraft does not need to strictly follow the path planned by the algorithm at all times, and the path planning algorithm is not required to provide an absolutely globally optimal solution, the design goal of the fast threat avoidance planner focuses more on quickly generating a feasible bypass path that can effectively avoid all no-fly zones, rather than guaranteeing the global optimality of the path length. In practical applications, to ensure that the aircraft has sufficient safety margin when performing lateral avoidance maneuvers, the radius of the original no-fly zone can be appropriately expanded to take into account the aircraft's maneuver envelope and uncertainties.

[0113] A reference path P = {P0, P1, ..., P2} consisting of l waypoints is planned using the fast threat avoidance planner. l Following this, the present invention designs a path tracking controller suitable for hypersonic gliders. The path tracking task scenario for obstacle avoidance in no-fly zones is as follows: Figure 8 As shown.

[0114] In the diagram, v and v' represent the positions of the aircraft when tracking the straight line segment and the circular arc segment, respectively; ψ LOS ψ is the line-of-sight azimuth angle of the aircraft at position v pointing to the next node on the reference path; desired For P i-1 To P i The azimuth angle (i.e., between two nodes on the reference path). Define v and the next node P. i The spherical distance between them is r hori The difference between the line-of-sight heading angle and the desired heading angle of the aircraft to the next node is:

[0115] e ψ =ψ LOS -ψ desired

[0116] The current error heading angular rate is calculated as follows:

[0117]

[0118] Construct the second-order dynamic control equations:

[0119]

[0120] Where ξ is the damping ratio, ω nare the dimensionless natural frequencies, which together regulate the error dynamic response characteristics. They can be designed and selected according to the desired tracking accuracy and the maneuverability of the vehicle. Once selected, they are given values during guidance. For the lateral maneuvering requirement of the glide phase, a bank angle control law based on the second-order dynamic control equation is designed:

[0121]

[0122] where sign(·) is the sign function, δ ψ is the dead zone threshold, which is used to avoid high-frequency oscillation of the bank angle and help maintain stable flight of the vehicle; σ max is the maximum bank angle (it should be understood that this control quantity can be set to the maximum bank angle actually allowed by the vehicle, or to a set value less than the maximum bank angle actually allowed); σ prev is the bank angle of the last guidance period. That is, outside the dead zone range, when the sign of the required error heading angle acceleration changes (i.e., when the direction of the aerodynamic force acting on the vehicle needs to change), the direction of the bank angle can be changed. In short, this can achieve the effect that the vehicle flies to the left or right side according to whether the current target is on the left or right side of the vehicle.

[0123] For the circular arc phase, considering the limited turning ability of the hypersonic gliding vehicle, a constant bank angle direction is set in advance according to the relative position of the restricted area:

[0124]

[0125] where is a vector pointing from the vehicle to the center of the restricted area, is the velocity vector of the vehicle. As an example, this formula can correspond to Figure 8 p i to p i+1 stage guidance.

[0126] After the vehicle completes all the necessary restricted area avoidance maneuvers, the guidance strategy will switch to the terminal phase. Considering that the bank angle adopts a “bang-bang” control form, its contribution to the lift has a constant characteristic. Specifically, given the initial state and the longitudinal guidance algorithm, regardless of the subsequent bank angle reversal strategy, the remaining range of the vehicle remains unchanged. When the remaining range of the vehicle exceeds the great circle distance to the target, it indicates that there is redundant range to be consumed. First, the last segment of the reference path is tracked using the path tracking guidance law described above, and the redundant range is consumed in a gradual manner through a periodic oscillation flight mode. During this process, the numerical prediction algorithm is continuously used to accurately estimate the real-time remaining range.

[0127] To achieve accurate heading correction and guide the aircraft to finally approach the target, especially in the complex scenario that still needs to retain the ability of multiple trajectory adjustment after the roll angle reversal operation, the present application proposes a target direction discrimination mechanism based on double-predicted trajectory reference, whose geometric principle is shown in Figure 9 For the aircraft currently at position v between the last two path segments (denoted as ), first, a double-reference system of the target point relative to the current predicted trajectory and the reverse predicted trajectory is established, where the current predicted trajectory is the predicted flight path generated based on the current flight state and maintaining the current roll angle, and the reverse predicted trajectory is another predicted flight path generated by immediately performing a roll angle reversal operation on the aircraft and always maintaining this control strategy. Then, on the current predicted trajectory and the reverse predicted trajectory respectively, two critical points v1 and v2 closest to the target point in the horizontal plane are determined, and the relative position vectors of the target point relative to the two critical points are calculated and The horizontal position vector composed of the last two waypoints is calculated The target direction criterion in the two-dimensional plane is constructed by calculating the specific component of the cross product of and the two relative position vectors in the horizontal plane:

[0128]

[0129] where k is the normal vector of the local horizontal plane. Based on the double-trajectory orientation criterion , it can be determined that the target point is located in the space formed by the current predicted trajectory and the reverse predicted trajectory (as shown in Figure 9 ). Even if there is a directional deviation in the initial roll angle instruction, the control system can still reach the target through a limited number of roll angle reversal operations.

[0130] When the difference between the remaining distance and the straight-line distance from the current position to the target point is less than a certain value, and the double-trajectory orientation criterion is met at the same time, the aircraft should fly towards the target with the shortest path. To improve the terminal position accuracy, a double-channel prediction model is constructed:

[0131]

[0132] where M(·) is the mapping relationship between the roll angle and the terminal miss distance, σ + and σ - represent the forward and reverse roll angle control profiles, and represent the terminal miss distances obtained by the forward predicted trajectory and the reverse predicted trajectory respectively. The adaptive decision of the roll angle sign is realized by comparing the terminal miss distances:

[0133]

[0134] That is, if the terminal miss distance obtained using the forward predicted trajectory is smaller than the terminal miss distance obtained using the backward predicted trajectory, then continue to use the forward roll angle control. Otherwise, in other cases, use the backward predicted trajectory. For example, if at the v position in Figure 9 the terminal miss distance obtained using the backward predicted trajectory is smaller, then use the backward predicted trajectory, and vice versa.

[0135] To make the applicable scenarios of the trained neural network general, for the initial conditions listed in Figure 4 , the pull-off items listed in Figure 10 generate a plurality of initial conditions conforming to a Gaussian distribution (conforming to the 3σ distribution principle, and limited to the maximum deviation values listed in the table), and after calculating the attack angle and roll angle commands using the guidance laws of the longitudinal channel and the lateral channel respectively, the trajectory propagation is performed (i.e., for a plurality of different initial conditions, different trajectories corresponding thereto are obtained), and a real sample data set (initial condition-flight trajectory) containing 5*10 5 samples is obtained, which is divided into a training set, a test set and a validation set according to a ratio of 8:1:1, for training the neural network. It should be understood that other quantities or ratios can also be used.

[0136] Step 2: Train the neural network

[0137] The numerical-predictive correction algorithm has a relatively heavy online calculation burden, which cannot meet the real-time demand of hypersonic vehicle guidance. The purpose of using the neural network is to serve as a proxy model for the computationally intensive module, to improve the computational efficiency while not losing too much guidance accuracy.

[0138] To improve the training effect of the neural network, define the height difference ΔH between the vehicle and the target:

[0139] ΔH = H - H f

[0140] Therefore, the state variable group (r, θ, φ, V, γ, ψ, sign(σ)), i.e., the geocentric distance r of the vehicle, the longitude θ, the latitude φ, the speed V, the flight path angle γ, the speed deviation angle ψ, and the sign of the roll angle sign(σ) at the current time, etc., can be equivalently changed according to the height at the time of reaching the target. For example, it is equivalently transformed to and part or all of them are used as inputs of the neural network, and the outputs of the four neural networks are the terminal attack angle, the terminal miss distance, the remaining range and the target direction respectively. That is, the real sample data set here is the pairing of the plurality of transformed state variables and the terminal attack angle, the terminal miss distance, the remaining range and the target direction.

[0141] The structure of the terminal attack angle prediction network is set to one input layer, six hidden layers, and one output layer. The input layer includes four neurons for receiving the state variable group of the aircraft, and each hidden layer includes 256 neurons. The output layer includes one neuron for outputting the terminal attack angle. The structure of the terminal miss distance prediction network is set to one input layer, five hidden layers, and one output layer. The input layer includes seven neurons for receiving the state variable group of the aircraft, and each hidden layer includes 256 neurons. The output layer includes one neuron for outputting the terminal miss distance. The structure of the remaining range prediction network is set to one input layer, six hidden layers, and one output layer. The input layer includes six neurons for receiving the state variable group of the aircraft, and each hidden layer includes 256 neurons. The output layer includes one neuron for outputting the remaining range. The structure of the target direction prediction network is set to one input layer, six hidden layers, and one output layer. The input layer includes seven neurons for receiving the state variable group of the aircraft, and each hidden layer includes 256 neurons. The output layer includes one neuron for outputting the target direction. The sigmoid function is added as the activation function of the output layer. Linear weight connections are used between the input layer and the hidden layer, between the hidden layers, and between the hidden layers and the output layer of all neural networks. The tanh function is added as the activation function between the hidden layers. It should be understood that the specific neural network structure, the activation function used, and the like are examples, and other structures or activation functions can also be used.

[0142] For the three neural networks for predicting the terminal attack angle, the terminal miss distance, and the remaining range, since a regression task is performed, the loss function is MSE, i.e., the Mean Squared Error. For the neural network for predicting the target direction, since a binary classification task is performed, the loss function is BCE, i.e., the Binary Cross-Entropy. The Adam method is used to optimize the parameters of the neural network. The data batch size for each learning is 512, the learning rate is 0.0001, and the training is stopped when the loss function does not decrease within 10 rounds.

[0143] After the training is completed, the prediction performance of the neural network is evaluated using the test set. For the target direction prediction network performing a binary classification task, the confusion matrix heat map as shown in Figure 11A is drawn according to the true labels of the data set and the predicted values output by the network, and the evaluation indicators are as shown in Figure 11BThe accuracy (Accuracy) is the proportion of the number of samples correctly classified by the model to the total number of samples, the recall (Recall) is the proportion of the number of positive samples correctly identified by the model to the total number of actual positive samples, the F1 score (F1 Score) is the harmonic mean of the precision (Precision) and the recall, the receiver operating characteristic curve (AUC-ROC) is the false positive rate (False Positive Rate, FPR) - true positive rate (True Positive Rate, TPR) curve, and the area under the PR-AUC is the area under the precision-recall curve. The closer these indicators are to 1, the better.

[0144] For the terminal attack angle, terminal miss distance, and remaining range prediction network for the regression task, the focus is on measuring the closeness between the predicted value and the true value. Figure 12A The scatter plots of the true value and the predicted value of the three regression models on the test set are shown. In these scatter plots, the horizontal axis represents the true value in the data set, and the vertical axis represents the predicted value output by the model. Ideally, all data points should be closely clustered around the ideal curve, which indicates that the model's prediction results are highly consistent with the true situation. Figure 12B The coefficient of determination (Coefficient of Determination, R 2 ), root mean squared error (Root Mean Squared Error, RMSE), and mean absolute error (Mean Absolute Error, MAE) of each model's prediction results are shown. Among them, the R 2 value measures the degree of explanation of the model for the variability of the data, and its value range is usually between 0 and 1, the closer to 1, the higher the goodness of fit of the model.

[0145] In the guidance process, after the trained neural network receives the state data of the aircraft, it can output the terminal attack angle, terminal miss distance, remaining range, and target direction for guidance.

[0146] Step 3: Use the predicted values output by the neural network to form an intelligent real-time collaborative guidance framework for hypersonic gliding aircraft.

[0147] Combine the above baseline guidance algorithm composed of longitudinal and lateral channels with the trained neural network to build the final intelligent real-time collaborative guidance framework. The core idea is to replace the time-consuming numerical iteration process in the above baseline algorithm with the fast forward inference of the neural network, so as to realize the real-time generation of guidance instructions while ensuring accuracy.

[0148] Specifically, after extracting features from the real-time acquired aircraft state parameters and performing corresponding normalization processing, a four-branch network inference is executed in parallel. The longitudinal channel uses the terminal angle of attack prediction network outputting the terminal angle of attack. By combining the aforementioned homotopy guidance law, real-time angle-of-attack commands are rapidly generated. The lateral channel uses a neural network to predict the terminal miss distance, remaining range, and target direction, achieving real-time intelligent situational awareness. This, combined with the reference path planned by the rapid threat avoidance planner, generates real-time billing angle commands. This approach not only retains the constraint handling capabilities and logical completeness of the original algorithm framework but also accelerates key computational processes through the introduction of neural networks.

[0149] In the process of predicting or verifying the trajectory, after obtaining the angle of attack and the slope angle, the trajectory curve can be obtained by using the Euler integral.

[0150] To verify the effectiveness of the algorithm, we conducted tests on... Figure 4 The guidance law performance under the nominal initial conditions shown is simulated. Figure 13 This demonstrates a reference path planned by the rapid threat avoidance planner. Figure 13 It can be seen that for aircraft with different starting positions, they can reach the target position by taking different reference paths. Figure 14A , 14B The tracking performance of the path tracking controller is demonstrated. Figures 15-18 The trajectory and control quantity curves obtained by the above-mentioned intelligent real-time collaborative guidance framework are shown. The terminal status is recorded, and the terminal height deviation is guaranteed to be within 100 meters, the terminal horizontal position deviation within 3 kilometers, meeting the requirements, and the coordination time error within 0.5 seconds.

[0151] To further illustrate the effectiveness and robustness of the guidance algorithm, in Figure 4 Based on the nominal initial conditions shown, Figure 10 The pull amount shown was randomly distributed according to the initial conditions, and 1000 Monte Carlo shooting experiments were conducted.

[0152] For the aerodynamic pull-off model of an aircraft, a deviation is added to the standard aerodynamic coefficients:

[0153]

[0154] Where, ΔC L ,ΔC D C is the pull coefficient of the lift coefficient and drag coefficient. L C D This is the standard aerodynamic coefficient.

[0155] The altitude-velocity profile and trajectory curves are as follows: Figure 19 , Figure 20As shown, all aircrafts meet all process constraints during flight. Figures 21-23 The simulation results show that the algorithm can still produce good time coordination guidance effect under uncertainty conditions, and the terminal state accuracy can achieve the expected effect.

[0156] In summary, the embodiments of the present application can solve the time coordination guidance problem of multiple hypersonic gliding vehicles under multiple no-fly zone constraints, and an intelligent real-time coordination guidance method for hypersonic gliding vehicles is proposed to improve the ability of multiple vehicle saturation attack.

[0157] The various steps, methods and functions disclosed herein can be implemented by means of function units or modules in one or more computing devices, and a typical example is an aircraft control device.

[0158] Specifically, the aircraft control device can be built-in processor and memory for storing executable instructions. When the processor calls and runs the relevant instructions in the memory, the control device can be driven to complete the functions of each embodiment of the present application.

[0159] Further, each function unit in the computing device can be realized by a processing circuit. The circuit can integrate one or more microprocessors, microcontrollers, and can also include digital signal processors (DSP), special-purpose digital logic, etc. The core task of the processing circuit is to be configured to execute program codes stored in the memory. The memory can be one or more types of read-only memory (ROM), random access memory (RAM), cache memory, flash memory device or optical memory. The program codes stored therein not only include instructions for executing communication protocols, but more importantly, include specific instructions for implementing various advanced technologies described herein. In a specific implementation, the processing circuit drives the corresponding function unit to perform its preset function in this way.

[0160] In this technical background, the term "unit" should be understood broadly. It can refer to physical entities such as electrical and electronic circuits, devices, modules, logic solid-state devices, etc., or non-physical entities such as computer programs or instruction sets written for performing various tasks, processes, calculations and output functions described herein.

Claims

1. A smart real-time cooperative guidance method for hypersonic glide vehicles, characterized in that, The method includes: Decouple the guidance problem into a longitudinal channel and a lateral channel; For the longitudinal channel, at least a portion of the state information about the target and the aircraft is input to the neural network and feedback control law in the longitudinal channel, respectively; wherein, the neural network in the longitudinal channel is configured to predict the terminal longitudinal control quantity when the aircraft reaches the target position, and then generate the longitudinal control quantity; The longitudinal control quantities generated by the neural network and feedback control law in the longitudinal channel are weighted and summed to obtain the longitudinal control quantity output by the longitudinal channel; wherein the weighting coefficients used in the weighted summation are at least related to the speed of the aircraft. For the lateral passage, a path planning algorithm is used to generate a reference path for avoiding the no-fly zone. The reference path includes: multiple intermediate nodes, and multiple intermediate paths between the multiple intermediate nodes. Based on multiple intermediate nodes and multiple intermediate paths on the reference path, segmented guidance is performed; when guiding an intermediate path, a lateral control quantity is generated via the feedback control law of the lateral channel; when guiding a target, a lateral control quantity is generated via the neural network of the lateral channel.

2. The method according to claim 1, Its features are: in, The longitudinal control value is the angle of attack of the aircraft, and the lateral control value is the roll angle of the aircraft; The neural network in the longitudinal channel is a terminal angle of attack prediction network; The feedback control law in the longitudinal channel is a proportional-derivative control law, and the process of generating the longitudinal control quantity includes: Based on the target and the state information of the aircraft, calculate the longitudinal control command: Where, k p k is the proportionality coefficient. d e is the differential coefficient; H This is the difference between the aircraft's altitude and the altitude constraint. This represents the rate of change of the difference between the aircraft's altitude and the altitude constraint. The weighted summation formula for the longitudinal channels is as follows: a cmd =ρ(V)α1+(1-ρ(V))α2 Where α1 is the control quantity generated by the feedback control law, and α2 is the control quantity generated by the neural network; the weighting coefficient ρ(V) is in the form of Where V is the current speed of the aircraft, V tran The preset switching speed threshold is defined by k, the steepness coefficient is set, and e is the base of the natural logarithm; such that: as the aircraft speed V gradually exceeds V... tran When ρ(V) approaches 1, the weight of the control quantity generated by the feedback control law increases; when the aircraft speed V gradually decreases below V... tran When ρ(V) approaches 0, the weights of the control quantity generated by the neural network increase; The neural network of the lateral channel includes multiple prediction networks for predicting terminal miss distance, remaining range, and target direction; the path planning algorithm is a geometrically visible graph-based avoidance planner that generates a collision-free reference path by recursively finding the external common tangent point pair of the no-fly zone with the deepest penetration depth from the start point to the end point. The feedback control law in the lateral channel is a proportional-derivative control law, and the control quantity generated by the proportional-derivative control law includes: Based on the reference path and the aircraft's state information, a second-order dynamic control equation is constructed: Where ξ is the set damping ratio, ω n For the dimensionless natural frequency e, ψ This is the difference between the line-of-sight heading angle and the desired heading angle of the aircraft to the next node on the reference path. For e ψ The first derivative with respect to time; Based at least on the aforementioned second-order dynamic control equations, calculate the lateral guidance command: Where sign(·) is the sign function, σ max σ is the maximum allowable tilt angle. prev δ is the tilt angle of the previous guidance cycle. ψ This is the dead zone threshold; The prediction of the target direction in the lateral channel is based on a target direction discrimination mechanism with dual prediction trajectory reference. The generation of lateral control quantities via the neural network through the lateral channel includes: The neural network of the lateral channel generates a current predicted trajectory and a reverse predicted trajectory for the target point, wherein the current predicted trajectory is the predicted flight path generated by the aircraft while maintaining the current tilt angle, and the reverse predicted trajectory is another predicted flight path generated by the aircraft while maintaining the opposite angle of the current tilt angle. On the two predicted trajectories, determine the two critical points v1 and v2 that are closest to the target point on the horizontal plane, and calculate the relative position vector of the target point with respect to these two critical points. and The horizontal position vector formed by combining the last intermediate node on the reference path and the target point Through calculation The target orientation criterion in the two-dimensional plane is constructed from the selected components of the cross product of the two relative position vectors in the horizontal plane: Where k is the normal vector of the horizontal plane; if the criterion for the orientation of the two trajectories... It can be determined that the target point is located within the space formed by the current predicted trajectory and the reverse predicted trajectory; Based at least on the dual-track orientation criterion, when the difference between the predicted remaining range and the great circle distance from the aircraft to the target is sufficiently small, a roll angle sign command is generated by comparing the predicted terminal miss distance:

3. The method according to claim 1, characterized in that, The training process of a neural network includes: Step 1.1: For multiple initial states of the aircraft, generate multiple sets of longitudinal and lateral trajectory data containing the aircraft's state and corresponding terminal state using homotopy guidance law and numerical prediction algorithm; Step 2.1: Extract the aircraft state variable set from the trajectory data as input, and the terminal angle of attack, terminal miss distance, remaining range, target direction, etc. corresponding to the extracted aircraft state variable set as output, to construct a training sample set; Step 2.2: Train the multiple neural networks of the vertical and horizontal channels using the training sample set; The prediction network for the terminal angle of attack is structured as follows: 1 input layer, 6 hidden layers, and 1 output layer. The input layer contains 4 neurons to receive the state variables of the aircraft, each hidden layer contains 256 neurons, and the output layer contains 1 neuron to output the terminal angle of attack. The network used to predict the off-target amount of the terminal is structured as follows: 1 input layer, 5 hidden layers, and 1 output layer. The input layer contains 7 neurons to receive the state variables of the aircraft, each hidden layer contains 256 neurons, and the output layer contains 1 neuron to output the off-target amount of the terminal. The network used to predict the remaining flight range is structured as follows: one input layer, six hidden layers, and one output layer. The input layer contains six neurons to receive the set of state variables of the aircraft, each hidden layer contains 256 neurons, and the output layer contains one neuron to output the remaining flight range. The network used for predicting the target direction is structured as follows: one input layer, six hidden layers, and one output layer. The input layer contains seven neurons to receive the state variables of the aircraft, and each hidden layer contains 256 neurons. The output layer contains one neuron to output the target direction. The output layer is activated by adding a sigmoid function. In this neural network, the input layer and hidden layer, and the hidden layer and output layer are all connected by linear weights, and each hidden layer is connected by linear weights. A hyperbolic tangent tanh function is added as the activation function. The network used for regression tasks uses MSE as the loss function, while the network used for binary classification tasks uses BCE as the loss function; the Adam method is used to optimize the parameters of the neural network.

4. The method as described in claim 3, characterized in that, The step of generating control quantities using the trained neural network during guidance includes: The current state variables of the aircraft are input into the neural network to obtain real-time predictions of terminal angle of attack, terminal miss distance, remaining range and target direction; For the longitudinal channel, the predicted terminal angle of attack is used. Construct an angle-of-attack-height profile to generate the angle-of-attack command α2; For the lateral channel, the tilt angle symbol is generated by a target direction discrimination mechanism based on dual predicted trajectory reference, based at least on the predicted remaining range, terminal miss distance, and target direction.

5. An aircraft control device, comprising: processor, and Memory used to store executable instructions; When the processor retrieves and runs the executable instructions stored in the memory, the aircraft control device executes the intelligent real-time cooperative guidance method for hypersonic gliding vehicles according to any one of claims 1 to 4.

Images