A method for intelligent trajectory planning of hypersonic vehicles facing space-time random obstacles

Abstract

The application discloses a kind of hyper-sonic aircraft intelligent trajectory planning methods for space-time random obstacle avoidance, comprising: in the process that hyper-sonic aircraft moves along nominal trajectory, real-time detection burst obstacle, in the case where burst obstacle is detected to appear, repeatedly execute flight control operation until complete all flight process, wherein flight control operation includes: obtaining the current flight state parameter of hyper-sonic aircraft and the space-time geometric parameter of burst obstacle and input to trajectory planning model;Trajectory planning model determines the attitude control quantity of hyper-sonic aircraft according to current flight state parameter and space-time geometric parameter, then controls the flight state of hyper-sonic aircraft according to attitude control quantity.The application realizes the accurate regulation and control of flight state in burst scene by trajectory planning model dynamic output adaptive attitude control quantity, improves the adaptive control ability and safety of aircraft in complex burst scene.

Description

Technical Field

[0001] This invention relates to the field of aircraft control technology, and in particular to an intelligent trajectory planning method for hypersonic aircraft oriented towards spatiotemporal random obstacle avoidance. Background Technology

[0002] Hypersonic vehicles are now widely used in many fields, such as military and high-speed transportation. However, obstacle avoidance is a highly challenging technical problem for hypersonic vehicles. Compared to traditional aircraft, hypersonic vehicles are extremely fast and have a short reaction time, making the obstacle avoidance logic of traditional aircraft unsuitable for hypersonic vehicles.

[0003] In existing technologies, obstacle avoidance for hypersonic vehicles is mainly aimed at avoiding fixed obstacles or not considering obstacles at all. It cannot effectively solve the problem of online path planning for obstacles with spatiotemporal random characteristics, resulting in poor real-time performance, flexibility, and safety.

[0004] Therefore, there is an urgent need for a trajectory planning method for hypersonic vehicles that can take into account obstacles with dynamic characteristics and plan trajectories in real time. Summary of the Invention

[0005] This invention provides an intelligent trajectory planning method for hypersonic vehicles to avoid spatiotemporal random obstacles, in order to solve the problems of poor real-time performance, flexibility and safety of route planning for hypersonic vehicles in the prior art.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide an intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance, comprising: During the movement of a hypersonic vehicle along its nominal trajectory, sudden obstacles are detected in real time. If a sudden obstacle is detected, the flight control operation is repeated until the hypersonic vehicle completes the entire flight process. The flight control operations include: Acquire the current flight status parameters of the hypersonic vehicle and the spatiotemporal geometric parameters of the sudden obstacle; The current flight state parameters and the spatiotemporal geometric parameters are input into the trajectory planning model; wherein, the trajectory planning model includes: a temporal feature extraction module and a global feature extraction module; Based on the temporal feature extraction module, the temporal correlation characteristics of the current flight state parameters and spatiotemporal geometric parameters are extracted to generate hidden states that characterize the predicted trajectory information; Based on the global feature extraction module, global features are extracted from the hidden state to obtain global features; Based on global characteristics, determine the attitude control variables of the hypersonic vehicle, and control the flight state of the hypersonic vehicle based on the attitude control variables.

[0007] It is understood that, compared with the prior art, this invention controls the aircraft to move according to the nominal trajectory when there are no sudden obstacles. When there are sudden obstacles, the flight control operation process is invoked in real time to predict the aircraft control information when there are sudden obstacles. This avoids the risk caused by the aircraft still moving according to the original planned trajectory when there are sudden obstacles. When a sudden obstacle appears, the flight control operation can be invoked to predict the latest path, reducing the safety risk of the aircraft colliding with the sudden obstacle and improving the real-time performance and flexibility of the aircraft trajectory prediction. In addition, the trajectory planning model of this invention adopts a neural network architecture including a temporal feature extraction module and a global feature extraction module, replacing the process of solving the optimal control information by a nonlinear programming algorithm. This solves the potential non-convergence problem of nonlinear algorithms and can guarantee the stability of the solution.

[0008] As a preferred option, the nominal trajectory can be generated in the following ways: Establish a dynamic model for hypersonic vehicles that takes into account the Earth's rotation and oblateness; The constraints in the dynamic model of the hypersonic vehicle are defined, including: mechanical constraints, thermal constraints, boundary constraints, and obstacle constraints. Based on the hypersonic vehicle dynamics model, the aforementioned constraints, and the performance function to be planned, an optimal control problem is constructed. The optimal control problem is solved using the Chebyshev pseudospectral method, and the nominal trajectory is output.

[0009] Understandably, by establishing a dynamic model that considers the Earth's rotation and oblateness, and combining the dynamic model, constraints, and performance functions to construct the optimal control problem, and then solving it using the Chebyshev pseudospectral method, the mechanical environment of the actual flight of the aircraft is closely matched. This avoids the nominal trajectory deviation caused by the simplified model, provides accurate trajectory basis for subsequent flight control, and ensures the feasibility and reliability of the flight control method.

[0010] As a preferred embodiment, the training method for the trajectory planning model includes: Obtain the optimal trajectory sequence of the hypersonic vehicle under each preset flight mission; Based on the optimal trajectory sequence, several training samples are generated; each training sample includes: flight state parameter samples of the hypersonic vehicle at the current moment, spatiotemporal geometric parameter samples of sudden obstacles, and corresponding actual attitude control quantities. The trajectory planning model is trained by taking each training sample as input and the corresponding predicted attitude control quantity as output.

[0011] It is understandable that training the trajectory planning model with the optimal trajectory sequence can enable the model to accurately learn the correspondence between obstacle parameters and attitude control quantities under different flight scenarios and different obstacle interferences, thereby improving the model's fitting accuracy and generalization ability.

[0012] As a preferred embodiment, the step of training the trajectory planning model using each training sample as input and the corresponding predicted attitude control quantity as output includes: Initialize the network parameters of the trajectory planning model; Based on the training samples and the initialized network parameters, the network parameter update operation is repeatedly performed until the preset total number of training rounds is completed, and the final network parameters are saved. The network parameter update operation includes: Obtain training samples; Based on the current network parameters, the corresponding predicted attitude control quantity is output according to the flight state parameter samples of the current training samples and the spatiotemporal geometric parameter samples of sudden obstacles. The predicted attitude control quantity is compared with the corresponding actual attitude control quantity, and the current loss function value is calculated based on the comparison result. The Adam adaptive optimization algorithm is used to backpropagate and update the current network parameters based on the current loss function value to obtain the network parameters for the next round; and when the current training round reaches the preset learning rate decay period, the current learning rate is halved according to the preset decay coefficient.

[0013] Understandably, by adopting a training method that initializes network parameters and iteratively updates them, combined with loss function comparison, backpropagation updates using the Adam adaptive optimization algorithm, and a strategy of halving the learning rate at preset intervals, the network parameters of the trajectory planning model can gradually converge to the optimal state. This improves the accuracy and stability of the model's prediction of attitude control quantities, ensures the model's training precision and efficiency, and enhances the model's learning effect on the mapping relationship between flight state parameters, spatiotemporal geometric parameters, and attitude control quantities.

[0014] As a preferred embodiment, the step of extracting the temporal correlation characteristics of the current flight state parameters and spatiotemporal geometric parameters to generate a hidden state for characterizing the predicted trajectory information includes: Input the candidate hidden state from the previous time step, the current flight state parameters, and the spatiotemporal geometry parameters of the sudden obstacle into the reset gate; The reset weight is calculated by resetting the gate based on the candidate hidden state of the previous time step, the current flight state parameters, and the spatiotemporal geometric parameters of the sudden obstacle. The current candidate hidden state is obtained based on the reset weight and the hidden state of the previous time step; Input the candidate hidden state from the previous time step and the current flight state parameters into the update gate; The update weight is calculated based on the candidate hidden state of the previous time step and the current flight state parameters, using the update gate. The temporal feature extraction module obtains the hidden state of the current time step by linearly interpolating between the hidden state of the previous time step and the current candidate hidden state based on the updated weight.

[0015] Understandably, by inputting the candidate hidden state from the previous time step, the current flight state parameters, and the spatiotemporal geometric parameters of sudden obstacles into the reset gate, calculating the reset weight, and combining it with the hidden state from the previous time step to obtain the current candidate hidden state, effective information in the hidden state from the previous time step and redundant information can be effectively filtered and retained, while suppressing redundant information. Inputting the candidate hidden state from the previous time step and the current flight state parameters into the update gate to calculate the update weight, and then performing linear interpolation between the current candidate hidden state and the hidden state from the previous time step using the update weight, the current hidden state at the current time step can be obtained. This can reasonably integrate current temporal information and historical temporal information, accurately extract temporal correlation characteristics, and thus generate a hidden state that can accurately represent the predicted trajectory information, ensuring the accuracy and rationality of temporal feature extraction and adapting to the needs of subsequent trajectory planning and attitude control quantity determination.

[0016] As a preferred embodiment, the attitude control parameters include: desired angle of attack and desired roll angle; Controlling the flight state of a hypersonic vehicle based on attitude control variables includes: The hypersonic vehicle is controlled to fly according to the desired angle of attack and desired roll angle.

[0017] Understandably, clearly defining attitude control parameters, including the desired angle of attack and the desired roll angle, provides a clear and specific control basis for the flight status control of hypersonic vehicles. By directly controlling the vehicle to fly according to the desired angle of attack and the desired roll angle, precise control of the vehicle's flight attitude can be achieved, ensuring the clarity and executability of attitude control commands, and guaranteeing the pertinence and effectiveness of flight status control.

[0018] As a preferred embodiment, the flight status parameters include: the current position, current angle of attack, current roll angle, and current speed of the hypersonic vehicle; The spatiotemporal geometric parameters of the sudden obstacle include: the discovery time, center longitude, center latitude, and radius of the sudden obstacle.

[0019] Understandably, specifying flight status parameters, including the hypersonic vehicle's current position, current angle of attack, current roll angle, and current velocity, and specifying the spatiotemporal geometric parameters of sudden obstacles, including the obstacle's detection time, center longitude, center latitude, and radius, can provide complete and accurate input for the trajectory planning model, ensuring the accuracy and reliability of attitude control quantity calculations.

[0020] As a preferred embodiment, the establishment of a hypersonic vehicle dynamics model that considers the Earth's rotation and oblateness includes: The dynamic model of a hypersonic vehicle considering Earth's rotation and oblateness is characterized as follows: in, It is the distance from the Earth's center. This is the aircraft's current speed. It is the aircraft's current speed tilt angle. This is the current longitude of the aircraft. It is the aircraft's current speed deflection angle. It is the current latitude of the aircraft. It is aerodynamic drag. It is aerodynamic lift. It is the mass of the aircraft. It is the current tilt angle of the aircraft. , , It is the component of gravitational acceleration in the spacecraft's velocity coordinate system. , , It is the component of centrifugal inertial force in the aircraft's velocity coordinate system. , , It is the component of the Coriolis inertial force in the aircraft's velocity coordinate system.

[0021] Understandably, this model, by incorporating the Earth's oblateness and rotation effects, constructs a hypersonic vehicle dynamics model that better reflects the real flight environment, significantly improving the accuracy of trajectory and attitude calculations. Furthermore, the model fully incorporates gravity, centrifugal inertial force, and Coriolis inertial force, effectively eliminating the long-term accumulated errors caused by traditional simplified models. These errors refer to the small calculation deviations caused by the model simplification and neglect of real physical factors during long-duration, long-distance, high-speed flight. These deviations accumulate and amplify with increasing flight time and distance, ultimately leading to significant deviations in trajectory, attitude, and position calculations. Therefore, this model is particularly suitable for long-distance, long-duration flight missions.

[0022] Accordingly, the present invention also provides a hypersonic vehicle control device, comprising: a sudden obstacle detection module and a flight control module; The sudden obstacle detection module is used to detect sudden obstacles in real time during the movement of a hypersonic vehicle along its nominal trajectory. The flight control module repeatedly executes flight control operations when it detects a sudden obstacle, until the hypersonic vehicle completes the entire flight process; The flight control operations include: Acquire the current flight status parameters of the hypersonic vehicle and the spatiotemporal geometric parameters of sudden obstacles; The current flight state parameters and the spatiotemporal geometric parameters are input into the trajectory planning model; wherein, the trajectory planning model includes: a temporal feature extraction module and a global feature extraction module; Based on the temporal feature extraction module, the temporal correlation characteristics of the current flight state parameters and spatiotemporal geometric parameters are extracted to generate hidden states that characterize the predicted trajectory information; Based on the global feature extraction module, global features are extracted from the hidden state to obtain global features; Based on global characteristics, the attitude control variables of the hypersonic vehicle are determined, and the flight state of the hypersonic vehicle is controlled according to the attitude control variables.

[0023] Accordingly, the present invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described above.

[0024] Accordingly, the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described above. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the steps of an intelligent trajectory planning method for hypersonic vehicles to avoid spatiotemporal random obstacles, as provided in an embodiment of the present invention. Figure 2 A schematic diagram of a scenario for the intelligent trajectory planning method for hypersonic vehicles to avoid spatiotemporal random obstacles provided in an embodiment of the present invention; Figure 3 A flowchart illustrating a trajectory planning model training method provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a trajectory planning model provided in an embodiment of the present invention; Figure 5A three-dimensional angular effect diagram of a hypersonic vehicle intelligent trajectory planning method for avoiding spatiotemporal random obstacles provided in an embodiment of the present invention; Figure 6 A schematic diagram of the effect of a ground angle for a hypersonic vehicle intelligent trajectory planning method for avoiding spatiotemporal random obstacles provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of a hypersonic vehicle control device provided by the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See Figure 1 , Figure 1 This is a flowchart illustrating a method for intelligent trajectory planning of hypersonic vehicles for avoiding spatiotemporal random obstacles, as provided in an embodiment of the present invention.

[0028] In one embodiment, the present invention provides an intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance, comprising the following steps S101: Step S101: During the hypersonic vehicle's movement along the nominal trajectory, detect sudden obstacles in real time; Step S102: If a sudden obstacle is detected, repeat the flight control operation until the hypersonic vehicle completes the entire flight process.

[0029] In this embodiment, the hypersonic vehicle moves along a pre-calculated nominal trajectory until it detects a sudden obstacle. During its movement along the nominal trajectory, the vehicle continuously detects whether a sudden obstacle has appeared. Specifically, this is achieved by using multiple sensors to collaboratively collect environmental information. Through cross-verification of multi-source data such as lidar, vision, and millimeter-wave radar, the sudden obstacle target is initially identified. Then, real-time data interaction between the sensors and the control unit is achieved using data link communication, uploading target features and location information. Simultaneously, environmental data and collaborative sensing information from remote nodes and the backend system are received. The multi-source sensing data and remote interaction information are combined for comprehensive analysis, verification, and decision-making to accurately determine whether a sudden obstacle exists and output the determination result. It can be understood that if the hypersonic vehicle does not detect a sudden obstacle during the entire flight process, the flight control operation procedure will not be invoked. Instead, the vehicle will move along the pre-calculated nominal trajectory until the flight process is completed. If a sudden obstacle is detected during the flight process, the flight control operation will be repeated until the hypersonic vehicle completes the entire flight process.

[0030] For easier understanding, please refer to Figure 2 , Figure 2 This is a schematic diagram of a scenario for the intelligent trajectory planning method for hypersonic vehicles to avoid spatiotemporal random obstacles provided in an embodiment of the present invention.

[0031] according to Figure 2 As can be seen from the scenario shown, if the aircraft moves along the nominal trajectory throughout the entire flight, it can avoid fixed obstacles. Therefore, if no sudden obstacles (i.e., random obstacles in the figure) are detected throughout the entire flight, it is safe to move along the nominal trajectory. However, if a sudden obstacle is detected, moving along the nominal trajectory still poses a clear safety risk. Therefore, the embodiments of the present invention propose that if no sudden obstacle is detected, the aircraft moves along the nominal trajectory. If a sudden obstacle is detected, the trajectory is corrected online by repeatedly executing flight control operations until the hypersonic aircraft completes the entire flight process.

[0032] In one embodiment, after detecting a sudden obstacle, the flight control operations that need to be repeatedly performed include: acquiring the current flight state parameters of the hypersonic vehicle through onboard sensing / measurement systems and other means, and acquiring the spatiotemporal geometric parameters of the sudden obstacle through onboard detection systems and other means. Then, the current flight state parameters and the spatiotemporal geometric parameters of the sudden obstacle are input into the trajectory planning model. The trajectory planning model includes a temporal feature extraction module and a global feature extraction module. Based on the temporal feature extraction module of the trajectory planning model, the model extracts the temporal correlation characteristics of the flight state parameters and spatiotemporal geometric parameters to generate a hidden state that represents the predicted trajectory information. The hidden state is then input into the global feature extraction module of the model. The global feature extraction module performs global feature extraction on the input hidden state to obtain global features that can be used to determine the attitude control quantities. Based on the global features, the attitude control quantities of the hypersonic vehicle can be determined, and the flight state of the hypersonic vehicle can be controlled based on the attitude control quantities. It is understood that, compared with the prior art, this invention controls the aircraft to move according to the nominal trajectory when there are no sudden obstacles. When there are sudden obstacles, the flight control operation process is invoked in real time to predict the aircraft control information when there are sudden obstacles. This avoids the risk caused by the aircraft still moving according to the original planned trajectory when there are sudden obstacles. When a sudden obstacle appears, the flight control operation can be invoked to predict the latest path, reducing the safety risk of the aircraft colliding with the sudden obstacle and improving the real-time performance and flexibility of the aircraft trajectory prediction. In addition, the trajectory planning model of this invention adopts a neural network architecture including a temporal feature extraction module and a global feature extraction module, replacing the process of solving the optimal control information by a nonlinear programming algorithm. This solves the potential non-convergence problem of nonlinear algorithms and can guarantee the stability of the solution.

[0033] In one embodiment, the method for generating the nominal trajectory includes: establishing a hypersonic vehicle dynamic model considering the Earth's rotation and oblateness; setting constraints in the dynamic model, including mechanical constraints, thermal constraints, boundary constraints, and obstacle constraints; constructing an optimal control problem based on the dynamic model, constraints, and the performance function to be planned; solving the optimal control problem based on the Chebyshev pseudospectral method; and outputting the nominal trajectory. Specifically, firstly, a hypersonic vehicle dynamic model considering the Earth's rotation and oblateness is established, i.e., a differential equation containing relevant state variables is established, namely: (1) in, It is the distance from the Earth's center. This is the aircraft's current speed. It is the aircraft's current speed tilt angle. This is the current longitude of the aircraft. It is the aircraft's current speed deflection angle. It is the current latitude of the aircraft. It is aerodynamic drag. It is aerodynamic lift. It is the mass of the aircraft. It is the current tilt angle of the aircraft. , , It is the component of gravitational acceleration in the spacecraft's velocity coordinate system. , , It is the component of centrifugal inertial force in the aircraft's velocity coordinate system. , , This refers to the components of the Coriolis inertial force in the spacecraft's velocity coordinate system. The expression for the gravitational acceleration component is: (2) in, It is the projection of gravitational acceleration onto the geocentric vector direction. The projection of gravitational acceleration onto the direction of the Earth's rotation vector is expressed as follows: (3) in, It is the Earth's gravitational constant. It is the Earth's second-order band harmonic coefficient. The expression for the components of the centrifugal inertial force, where the Earth's equatorial radius is given by: (4) in, Given the Earth's rotational angular velocity, the component expression for the Coriolis inertial force is: (5) Based on differential equation (1), the component expressions of gravity, centrifugal inertial force and Gothic inertial force are considered simultaneously. Appropriate force and heat constraints, boundary constraints and obstacle constraints are set. At the same time, the shortest flight time is used as the optimization objective to construct the performance index function and construct the optimal control problem.

[0034] Specifically, the following constraints must be met during the gliding process of the aircraft: First, the following initial constraints need to be met: (6) In the formula, For flight time, The initial geocentric distance of the aircraft. The initial longitude of the aircraft. The initial latitude of the aircraft. The initial velocity of the aircraft. The initial trajectory tilt angle (that is, the aircraft velocity tilt angle at the starting point of the trajectory). The initial trajectory deflection angle (i.e., the vehicle's velocity deflection angle at the trajectory's starting point) is generally given by the handover conditions of the glide phase. Specifically, the flight process of a hypersonic vehicle is usually divided into multiple phases, such as the boost phase, reentry phase, and glide phase. The glide phase is not an independent flight phase, but rather a continuation of the tasks from the previous flight phase (such as the boost phase or reentry phase). At the handover moment between adjacent flight phases (i.e., the handover point), the vehicle's position, velocity, attitude, and other state variables must satisfy the continuity condition, meaning that the initial state at the start of the glide phase is completely consistent with the terminal state at the end of the previous phase. Therefore, the initial conditions of the glide phase are not set independently, but are jointly determined by the terminal constraints, guidance laws, and flight mission requirements of the previous flight phase, and are transmitted through the state continuity conditions at the handover point.

[0035] The gliding process also needs to meet the following terminal constraints: (7) In the formula, For the terminal time, For terminal longitude, The terminal dimension is generally given by the task requirements. The above conditions are generally given by the task requirements. It can be understood that the boundary constraints referred to in this embodiment include initial constraints such as: Formula (6) and terminal constraints such as: Formula (7).

[0036] The gliding process also needs to satisfy the following path constraints: Specifically, to ensure the structural safety of the aircraft, the force and heat conditions at various times need to be constrained, mainly considering dynamic pressure, overload, and stagnation point heat flux density constraints: (8) In the formula, For dynamic pressure, For overload, For stagnation heat flux density, For the maximum available dynamic pressure, For maximum available overload, For maximum heat flux density, Atmospheric density is calculated using an exponential density model. It is the heat flux density coefficient, which is determined by the geometry of the stagnation point of the aircraft.

[0037] The gliding process also needs to meet the following obstacle constraints: Specifically, if we consider the obstacle as an infinitely tall cylinder extending vertically upwards from the ground surface, the obstacle constraint can be represented as: (9) In the formula, Let the longitude of the center of the i-th threat zone be . Let i be the latitude of the center of the i-th threat zone. Let be the radius of the i-th threat zone.

[0038] The gliding process also needs to satisfy the following process and control constraints: Specifically, to ensure the physical feasibility of the gliding trajectory and to ensure that the control variables do not exceed the control system's capacity, constraints are imposed on the state and control at each moment: (10) In the formula, For the aircraft state vector, This represents the lower limit of the state. The upper limit of the state. For aircraft control vectors, As the lower limit of control, To control the upper limit, Angle of attack for the aircraft.

[0039] In one possible implementation, the performance function to be planned needs to take the minimum flight time as the optimization objective. The objective function is to minimize the flight time, i.e.: (11) In one possible implementation, after constructing the optimal control problem, the Chebyshev pseudospectral method is used to solve the optimal control problem in the nominal trajectory solution stage. Specifically, the solution process is as follows: First, the Chebyshev-Gauss-Lobatto (CGL) discrete points of the Chebyshev pseudospectral method are defined on the interval [-1, 1]. Therefore, it is necessary to linearly transform the original time domain t from [t0, tf] to [-1, 1]. Thus, the time variable is transformed, and the transformation relationship is as follows: (12) in, It is the initial time. It is the terminal time; Next, discrete node calculations are performed. The discrete nodes of the Chebyshev pseudospectral method are selected as the extreme points of the Nth-order Chebyshev polynomial, i.e., CGL points. They are non-uniformly distributed in the interval [-1, 1]. For the Type II standard CGL points, have: (13) Then, taking the state variables and control variables at the above N+1 discrete nodes, we construct Lagrange interpolation polynomials as approximate expressions for the continuous state and control, respectively. Then the true state variables... With control variables The approximate expression is: (14) (15) The Lagrange interpolation basis functions are: (16) In the formula, For CGL points, according to the properties of Lagrange interpolation, the approximate state value at discrete nodes is equal to the actual state, and the approximate control value is equal to the actual control value.

[0040] Next, regarding and Differentiating the approximate expression, we obtain the state vector in... The approximate expression for the derivative at point is: (17) In the formula, for Differential matrix The element in the k-th row and j-th column, The calculation formula is: (18) The continuous state-time derivatives in the dynamic equations are replaced with algebraic approximations of the state derivatives derived by interpolation. These approximations are then calculated at all interpolation nodes using the Chebyshev pseudospectral method, thus discretizing the continuous differential constraints. In this way, the constraints of the original optimal control problem's dynamic differential equations can be transformed into algebraic constraints, i.e., for... ,have: (19) For optimizations involving integral terms in the performance metrics, the Clenshaw-Curtis numerical integral is used for approximation. For any continuous function in the interval [-1, 1]... Its integral can be approximated by the sum of function values ​​at N+1 CGL discrete points, that is: (20) Therefore, the approximate formula for calculating the performance index by integral is transformed into: (twenty one) In the formula, The weighting for Clenshaw-Curtis can be determined by the following formula: If N is even: (twenty two) If N is odd: (twenty three) Among them, the " " above the summation symbol "" indicates that the first and last two expressions are divided by 2.

[0041] In summary, the optimization problem can be transformed into a nonlinear programming problem as follows, which aims to minimize the performance index function while satisfying constraints such as dynamics, boundary conditions, and path constraints: (twenty four) in, Let these be the initial and terminal constraint functions. Other process constraint functions. After the original optimal control problem is solved, the optimal solution to the trajectory planning problem can be obtained by solving this problem using algorithms such as sequence quadratic programming, and the obtained optimal control sequence can be used. and the corresponding optimal state Construct the optimal state-action pair sequence This serves as the storage format for the nominal trajectory. Understandably, by establishing a dynamic model that considers the Earth's rotation and oblateness, and combining the dynamic model, constraints, and performance functions to construct the optimal control problem, and then solving it using the Chebyshev pseudospectral method, the system closely matches the actual mechanical environment of the aircraft's flight. This avoids nominal trajectory deviations caused by simplified models, providing accurate trajectory data for subsequent flight control and ensuring the feasibility and reliability of the flight control method.

[0042] In one possible implementation, the training method for the trajectory planning model includes: obtaining the optimal trajectory sequence of the hypersonic vehicle under each preset flight mission; generating several training samples based on the optimal trajectory sequence; wherein each training sample includes: flight state parameter samples of the hypersonic vehicle at the current moment, spatiotemporal geometric parameter samples of sudden obstacles, and corresponding actual attitude control quantities; using each training sample as input and the corresponding predicted attitude control quantities as output, the trajectory planning model is trained.

[0043] Specifically, after obtaining the optimal trajectory sequence of the hypersonic vehicle under various preset flight missions, the optimal trajectory sequence can be discretized according to equal time steps to obtain a series of discrete trajectory nodes. Each node corresponds to a unique time point, covering the entire flight cycle of the vehicle from the start to the end of the mission. From the discretized optimal trajectory sequence, the following are extracted for each trajectory node: the flight state parameter sample of the hypersonic vehicle at the current time, the spatiotemporal geometric parameter sample of the sudden obstacle, and the corresponding actual attitude control quantity. After obtaining the flight state parameter sample of the hypersonic vehicle at each time point, the spatiotemporal geometric parameter of the sudden obstacle, and the corresponding actual attitude control quantity, the trajectory planning model can be trained using each training sample as input and the corresponding predicted attitude control quantity as output. It can be understood that training the trajectory planning model with the optimal trajectory sequence allows the model to accurately learn the correspondence between obstacle parameters and attitude control quantities under different flight scenarios and different obstacle interferences, improving the model's fitting accuracy and generalization ability.

[0044] In one possible implementation, the specific method for training the trajectory planning model using each training sample as input and the corresponding predicted attitude control quantity as output is as follows: First, initialize the network parameters of the trajectory planning model. Based on the training samples and the initialized network parameters, repeatedly perform the network parameter update operation until the preset total training rounds are completed, and save the final network parameters. The network parameter update operation includes: acquiring training samples; based on the current network parameters, and according to the flight state parameter samples and spatiotemporal geometric parameter samples of sudden obstacles in the current training samples, outputting the corresponding predicted attitude control quantity; comparing the predicted attitude control quantity with the corresponding actual attitude control quantity, and calculating the current loss function value based on the comparison result; updating the current network parameters through backpropagation using the Adam adaptive optimization algorithm based on the current loss function value to obtain the network parameters for the next round; and halving the current learning rate by a preset decay coefficient when the current training round reaches the preset learning rate decay period.

[0045] For ease of understanding, see Figure 3 , Figure 3 This is a flowchart illustrating a trajectory planning model training method provided in an embodiment of the present invention.

[0046] like Figure 3 As shown, for example, training a trajectory planning model can be performed using the following algorithm: S201: Initialize neural network weights, including... and b: S202: Initialize network parameters. Specifically, initialize the learning rate (lr) to 0.01, the total number of training epochs (n_epoches) to 200, and the current training epoch (epoch) to 1. S203: Determine if epoch ≤ n_epoches. If yes, proceed to S204; otherwise, proceed to step S212. S204: Set i=1 and go to S205; S205: Determine if i ≤ n_batches? If yes, go to S206; otherwise, go to S209, where n_batches is the total batch size. S206: Extract the i-th group of normalized optimal trajectory samples from the training set, use them as network input x and network output y, and then proceed to S207; S207: Calculate the loss function value based on the current network parameters, update the network parameters using the Adam algorithm, and then proceed to S208; S208: Let i = i + 1, and go to S205; S209: Set epoch = epoch + 1, and go to S210; S210: Determine if mod(epoch,20)==0. If yes, go to S211; otherwise, go to S203. S211: Let lr = lr / 2, then go to S203; S212: Save neural network weights and b; Understandably, by adopting a training method that initializes network parameters and iteratively updates them, combined with loss function comparison, backpropagation updates using the Adam adaptive optimization algorithm, and a strategy of halving the learning rate at preset intervals, the network parameters of the trajectory planning model can gradually converge to the optimal state. This improves the accuracy and stability of the model's prediction of attitude control quantities, ensures the model's training precision and efficiency, and enhances the model's learning effect on the mapping relationship between flight state parameters, spatiotemporal geometric parameters, and attitude control quantities.

[0047] In one embodiment, extracting the temporal correlation features of the current flight state parameters of the aircraft and the spatiotemporal geometric parameters of sudden obstacles to generate a hidden state for characterizing the predicted trajectory information includes: inputting the candidate hidden state of the previous time step and the current flight state parameters into a reset gate; calculating a reset weight based on the candidate hidden state of the previous time step and the current flight state parameters through the reset gate; obtaining the current candidate hidden state based on the reset weight and the hidden state of the previous time step; inputting the candidate hidden state of the previous time step and the current flight state parameters into an update gate; calculating an update weight based on the candidate hidden state of the previous time step and the current flight state parameters through the update gate; and the temporal feature extraction module linearly interpolating between the current candidate hidden state and the hidden state of the previous time step according to the update weight to obtain the hidden state of the current time step. Specifically, 10 samples can be extracted as a batch and input into the temporal feature extraction module. After the temporal feature extraction module initially extracts features, it obtains an n-dimensional hidden state containing trajectory history information. The temporal feature extraction module uses a reset gate and an update gate to determine how to fuse the old and new memories. The process is as follows: At time step t, the temporal feature extraction module first combines the candidate hidden state of the previous time step with the state of the current time step. The input is fed to the reset gate, which calculates weights to determine how much previous historical information is forgotten. When the reset gate is closed, it ignores previous memories and focuses on the current state input, which helps to capture short-term dependencies in the time series. The specific update process of the reset gate is as follows: First, the reset gate calculates the reset weight. The specific formula is as follows: (25) In the formula, and It resets the gate weight. This is the sigmoid function.

[0048] Then, the reset gate resets the weights. The hidden state of the previous time step Obtain candidate hidden states The specific formula is as follows: (26) In the formula, and Update the weights for the candidate hidden states.

[0049] Next, the candidate hidden state from the previous time step is... and the flight status parameters at the current time step The input is fed into the update gate, which calculates update weights to determine the extent to which candidate hidden states are utilized. Update hidden status This helps to capture long-term dependencies in time series. The update gate update process is as follows: (27) In the formula, and To reset the gate weights.

[0050] Finally, the temporal feature extraction module utilizes updated weights In the hidden state of the previous time step and candidate hidden state Interlinear interpolation is used to update the hidden state at the current time step. , specific formula: (28) In the formula, This is for element-wise multiplication.

[0051] In one embodiment, the temporal feature extraction module linearly interpolates between the current candidate hidden state and the hidden state of the previous time step based on the updated weights to obtain the hidden state of the current time step. Then, it also includes: based on the global feature extraction module, performing global feature extraction on the hidden state to obtain global features, and then determining the attitude control quantity of the hypersonic vehicle based on the global features, and controlling the flight state of the hypersonic vehicle based on the attitude control quantity.

[0052] For easier understanding, please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of a trajectory planning model provided in an embodiment of the present invention.

[0053] For example, the temporal feature extraction module can be a GRU network, and the global feature extraction module can be a DNN network. For ease of understanding, the following will combine... Figure 4 This section specifically explains the state sequence processing and action decision-making process of the trajectory planning model when the temporal feature extraction module uses a GRU network and the global feature extraction module uses a DNN network. Figure 4 As shown, the model uses the state sequences of the aircraft and sudden obstacles as the network input to the GRU layer. The input can be represented as: (29) In the formula, These are the aircraft's current altitude, current longitude, current latitude, current speed, current speed tilt angle, and current speed deflection angle. These are the time of discovery of the sudden obstacle, the longitude of the center of the sudden obstacle, the latitude of the center of the sudden obstacle, and the radius of the sudden obstacle.

[0054] In practice, 10 samples can be extracted as a batch input. After preliminary feature extraction by the GRU layer, an n-dimensional hidden state containing trajectory history information is obtained. Hide state As input, features of these hidden states are globally extracted through a DNN network, and the output is... Specifically, a DNN network consists of four fully connected (FC) layers. The outputs of the first three FC layers can be calculated using the following formula: (30) In the formula, and For the first Layer weights and biases.

[0055] For the last FC layer: (31) In the formula, and This refers to the weights and biases of the 4th layer. The output... It can be represented as: (32) in, This represents the expected angle of attack of the output. This represents the expected tilt angle of the output. It can be understood that the DNN network is responsible for globally extracting features from the hidden variables, which can enhance the model's non-linear expressive power, reduce network training loss, and suppress high-frequency oscillations in the output action.

[0056] In one possible implementation, the optimal trajectory sequence of a hypersonic vehicle under various preset flight missions can be obtained by employing the Chebyshev pseudospectral method. It is understood that to ensure the training effect of the trajectory planning model, the training samples need to satisfy flight mission constraints and possess good generalization and physical rationality. To maximize the guarantee that the training samples themselves are feasible trajectory sequences that satisfy flight mission constraints and possess good generalization and physical rationality (for ease of understanding, this invention refers to them as the optimal trajectory sequence), this embodiment obtains the optimal trajectory sequence by calculating the theoretically optimal trajectory sequence for the vehicle to avoid obstacles under different flight state parameters and different spatiotemporal geometric parameters of sudden obstacles. Through a reasonable calculation process, the final generated samples can satisfy both physical laws and achieve mathematical optimality. Specifically, in this embodiment, the following method is used to obtain the optimal trajectory sequence: Step 1: Constructing the Optimal Control Problem. First, construct a dynamic model of the hypersonic vehicle considering the Earth's rotation and oblateness. Define the constraints in the dynamic model, including mechanical constraints, thermal constraints, and boundary or obstacle constraints. Based on the dynamic model, constraints, and the performance function to be planned, construct the optimal control problem. Solve the optimal control problem using the Chebyshev pseudospectral method and output the nominal trajectory. Specifically, first establish a dynamic model of the hypersonic vehicle considering the Earth's rotation and oblateness, i.e., establish differential equations containing relevant state variables, namely: (33) in, It is the distance from the Earth's center. This is the aircraft's current speed. It is the aircraft's current speed tilt angle. This is the current longitude of the aircraft. It is the aircraft's current speed deflection angle. It is the current latitude of the aircraft. It is aerodynamic drag. It is aerodynamic lift. It is the mass of the aircraft. This is the current tilt angle of the aircraft, measured in radians (rad), representing the angle of inclination of the lift vector within the aircraft's body coordinate system. , , It is the component of gravitational acceleration in the spacecraft's velocity coordinate system. , , It is the component of centrifugal inertial force in the aircraft's velocity coordinate system. , , It is the component of the Coriolis inertial force in the spacecraft's velocity coordinate system. The expression for the gravitational acceleration component is: (34) in, It is the projection of gravitational acceleration onto the geocentric vector direction. The projection of gravitational acceleration onto the direction of the Earth's rotation vector is expressed as follows: (35) in, It is the Earth's gravitational constant. It is the Earth's second-order band harmonic coefficient. The expression for the components of the centrifugal inertial force, where the Earth's equatorial radius is given by: (36) in, It is the Earth's rotational angular velocity, and the Coriolis inertial force component is: (37) Based on the differential equation, considering the component expressions of gravity, centrifugal inertial force, and Gothic inertial force, and setting appropriate mechanical and thermal constraints, boundary constraints, and obstacle constraints, while taking the shortest flight time as the optimization objective, a performance index function is constructed to form the optimal control problem.

[0057] Specifically, the following constraints must be met during the gliding process of the aircraft: First, the following initial constraints need to be met: (38) In the formula, For flight time, The initial geocentric distance of the aircraft. The initial longitude of the aircraft. The initial latitude of the aircraft. The initial velocity of the aircraft. The initial trajectory tilt angle (that is, the aircraft velocity tilt angle at the starting point of the trajectory). The initial trajectory deflection angle (i.e., the vehicle's velocity deflection angle at the trajectory's starting point) is generally given by the handover conditions of the glide phase. Specifically, the flight process of a hypersonic vehicle is usually divided into multiple phases, such as the boost phase, reentry phase, and glide phase. The glide phase is not an independent flight phase, but rather a continuation of the tasks from the previous flight phase (such as the boost phase or reentry phase). At the handover moment between adjacent flight phases (i.e., the handover point), the vehicle's position, velocity, attitude, and other state variables must satisfy the continuity condition, meaning that the initial state at the start of the glide phase is completely consistent with the terminal state at the end of the previous phase. Therefore, the initial conditions of the glide phase are not set independently, but are jointly determined by the terminal constraints, guidance laws, and flight mission requirements of the previous flight phase, and are transmitted through the state continuity conditions at the handover point.

[0058] The gliding process also needs to meet the following terminal constraints: (39) In the formula, For the terminal time, For terminal longitude, The terminal dimension is generally given by the task requirements. The above conditions are generally given by the task requirements. It can be understood that the boundary constraints referred to in this embodiment include initial constraints such as: formula (38) and terminal constraints such as: formula (39).

[0059] The gliding process also needs to satisfy the following path constraints: Specifically, to ensure the structural safety of the aircraft, the force and heat conditions at various times need to be constrained, mainly considering dynamic pressure, overload, and stagnation point heat flux density constraints: (40) In the formula, For dynamic pressure, For overload, For stagnation heat flux density, For the maximum available dynamic pressure, For maximum available overload, For maximum heat flux density, Atmospheric density is calculated using an exponential density model. It is the heat flux density coefficient, which is determined by the geometry of the stagnation point of the aircraft.

[0060] The gliding process also needs to meet the following obstacle constraints: Specifically, if we consider the obstacle as an infinitely tall cylinder extending vertically upwards from the ground surface, the obstacle constraint can be represented as: (41) In the formula, Let the longitude of the center of the i-th threat zone be . Let i be the latitude of the center of the i-th threat zone. Let be the radius of the i-th threat zone.

[0061] The gliding process also needs to satisfy the following process and control constraints: Specifically, to ensure the physical feasibility of the gliding trajectory and to ensure that the control variables do not exceed the control system's capacity, constraints are imposed on the state and control at each moment: (42) In the formula, For the aircraft state vector, This represents the lower limit of the state. The upper limit of the state. For aircraft control vectors, As the lower limit of control, To control the upper limit, Angle of attack for the aircraft.

[0062] In one possible implementation, the performance function to be planned needs to take the minimum flight time as the optimization objective. The objective function is to minimize the flight time, i.e.: (43) Step 2: Solve the optimal control problem using the Chebyshev pseudospectral method. Specifically, this includes the following steps: First, the original time domain t is linearly transformed from [t0, tf] to [-1, 1]. Therefore, the time variable is transformed as follows: (44) in, It is the initial time. It is the terminal time; Then, discrete node calculations are performed. The discrete nodes of the Chebyshev pseudospectral method are selected as the extreme points of the Nth-order Chebyshev polynomial, i.e., CGL points. They are non-uniformly distributed in the interval [-1,1]. For the Type II standard CGL points, have: (45) Next, we approximate the state and control variables through interpolation. Taking the state and control variables at the above N+1 discrete nodes, we construct Lagrange interpolation polynomials as approximate expressions for the continuous state and control, respectively. Then, the true state variables... With control variables The approximate expression is: (46) (47) The Lagrange interpolation basis functions are: (48) In the formula, For CGL points, according to the properties of Lagrange interpolation, the approximate state value at discrete nodes is equal to the actual state, and the approximate control value is equal to the actual control value.

[0063] Next and Differentiating the approximate expression, we obtain the state vector in... The approximate expression for the derivative at point is: (49) In the formula, for Differential matrix The element in the k-th row and j-th column, The calculation formula is: (50) The continuous state-time derivatives in the dynamic equations are replaced with algebraic approximations of the state derivatives derived by interpolation. These approximations are then calculated at all interpolation nodes using the Chebyshev pseudospectral method, thus discretizing the continuous differential constraints. In this way, the constraints of the original optimal control problem's dynamic differential equations can be transformed into algebraic constraints, i.e., for... ,have: (51) in, These are the differential matrix elements of the Chebyshev pseudospectral method. It is the value of the function on the right side of the dynamic equation at the k-th node.

[0064] Then, an integral approximation of the performance index is performed. For cases where the optimized performance index contains an integral term, the Clenshaw-Curtis numerical integral is used to approximate it. For any continuous function on the interval [-1,1]... Its integral can be approximated by the sum of function values ​​at N+1 CGL discrete points, that is: (52) Therefore, the approximate formula for calculating the performance index by integral is transformed into: (53) In the formula, The weighting for Clenshaw-Curtis can be determined by the following formula: If N is even: (54) If N is odd: (55) Among them, the " " above the summation symbol "" indicates that the first and last two expressions are divided by 2.

[0065] In summary, the optimization problem can be transformed into the following nonlinear programming problem, which minimizes the performance index function under constraints such as dynamics, boundary conditions, and path constraints: (56) in, Let these be the initial and terminal constraint functions. For other process constraint functions.

[0066] After the original optimal control problem is solved, the optimal solution to the trajectory planning problem can be obtained by solving the problem using algorithms such as sequential quadratic programming.

[0067] Step 3: Generate training samples.

[0068] The optimal trajectories (including state variables and control variables at each discrete node) obtained using the Chebyshev pseudospectral method are used as ground truth samples, i.e., the labels of the training samples. Simultaneously, the corresponding initial conditions, boundary conditions, constraints, and performance function parameters are used as input features of the samples. The above steps are repeated to generate multiple sets of optimal trajectories under different initial conditions, boundary conditions, and constraints, thereby constructing a training sample set for the hypersonic vehicle trajectory planning model. It should be noted that obtaining the optimal trajectories used to create training samples is similar to obtaining the nominal trajectories, both obtained through the Chebyshev pseudospectral method. The difference lies in the fact that the nominal trajectory is a flight trajectory conforming to real constraints obtained based on actual operating conditions, while the optimal trajectories used to create training samples are obtained by simulating a large number of different initial conditions, boundary conditions, and constraints, inputting them into the optimal control problem, and solving for the optimal trajectories under different operating conditions to construct a large number of training samples. It is understandable that, through the above steps 1-3, it can be guaranteed that the training samples obtained are rigorously calculated, and the samples themselves are the optimal control solutions that satisfy dynamic constraints, aerodynamic constraints and boundary conditions. They have clear physical meaning, high numerical accuracy, and no invalid or non-compliant motion data. They can provide reliable, accurate and optimal ground truth labels for trajectory planning neural network models, which is conducive to improving the convergence speed of model training and the reliability and optimality of output trajectories.

[0069] In one possible implementation, the attitude control quantities output by the model include the desired angle of attack and the desired roll angle. Based on these, the flight state of the hypersonic vehicle is controlled according to the attitude control quantities, including controlling the hypersonic vehicle to fly according to the desired angle of attack and the desired roll angle. In actual operation, the process of controlling the hypersonic vehicle to fly according to the desired angle of attack and the desired roll angle can be as follows: calculate the corresponding control deflection command based on the desired angle of attack and the desired roll angle, drive the elevator, ailerons, and other control surfaces to adjust the vehicle's attitude, collect the actual angle of attack and the actual roll angle in real time and compare them with the command, eliminate deviations through feedback control, and finally achieve stable and precise attitude tracking flight. It is understandable that clearly defining the attitude control quantities, including the desired angle of attack and the desired roll angle, can provide a clear and specific control basis for the flight state control of the hypersonic vehicle. By directly controlling the vehicle to fly according to the desired angle of attack and the desired roll angle, precise control of the vehicle's flight attitude can be achieved, ensuring the clarity and executability of the attitude control command, and guaranteeing the pertinence and effectiveness of flight state control.

[0070] In practice, the entire process of aircraft motion control can be achieved through the online application of the model. Specifically, the online planning process during the online application of the model involves the following steps: Step 1: Initialize the aircraft status and control quantity Let t=0, where, The initial state of the aircraft, specifically including , ; The initial control variables for the aircraft specifically include: The above conditions are given by the handover point of the gliding segment; Step 2: Determine if t ≤ If yes, proceed to step 3; otherwise, proceed to step 6. This is the time it takes to detect a sudden obstacle; Step 3: Extract the nominal trajectory state-action pair based on the current time. This allows the aircraft to follow the nominal trajectory. Step 4: Perform the action Input into the dynamics model to update the aircraft state. The action pair is obtained from the storage format of the nominal trajectory. The method of obtaining the nominal trajectory has been explained above and will not be repeated here. Step 5: Let t = t + t, proceed to step 2; Step 6: Determine If yes, proceed to step 7; otherwise, end trajectory generation. This is the error of the current geographical location. Given an error tolerance range, where e is the geographical distance between the current aircraft and the destination. When the distance between the aircraft and the destination is less than the error tolerance range, it means that the aircraft has reached the destination and trajectory generation can be terminated. Step 7: Based on the current aircraft status The vehicle states from the previous 9 steps form the network input of the trajectory planning model. Utilize network to generate output ; Step 8: Through Update aircraft status Let t = t + t, proceed to step 6.

[0071] For ease of understanding, the effects of implementing this invention will be described below. See also Figure 5 , Figure 5This is a three-dimensional perspective illustration of an intelligent trajectory planning method for hypersonic vehicles to avoid spatiotemporal random obstacles, provided by an embodiment of the present invention. The figure shows the effects of controlling the vehicle's motion using the nominal trajectory, the DGRU trajectory (i.e., the flight trajectory simulated based on the trajectory planning model including GRU network and DNN network provided by the embodiment of the present invention), and the optimal trajectory (the calculated theoretical optimal trajectory). It can be seen that if the nominal trajectory is used, it can obviously only avoid fixed obstacles and cannot avoid random obstacles (i.e., sudden obstacles). If the optimal trajectory and the DGRU trajectory are used, both fixed obstacles and random obstacles (i.e., sudden obstacles) can be avoided. Thus, it can be seen that the present invention can achieve a better effect in avoiding sudden obstacles.

[0072] For a clearer explanation of the effects of this invention, please refer to [link to relevant documentation]. Figure 6 , Figure 6 This is a ground-angle schematic diagram of the effect of a hypersonic vehicle intelligent trajectory planning method for avoiding spatiotemporal random obstacles provided in an embodiment of the present invention. It can also be seen from the ground angle that if the vehicle moves according to the nominal trajectory, it can only avoid fixed obstacles and cannot avoid random obstacles (i.e., sudden obstacles). Only by moving according to the optimal trajectory and the DGRU trajectory can it avoid both fixed and random obstacles.

[0073] Accordingly, see Figure 7 , Figure 7 This is a schematic diagram of a hypersonic vehicle control device provided by the present invention, including: a sudden obstacle detection module 701 and a flight control module 702. The sudden obstacle detection module 701 is used to detect whether a sudden obstacle occurs during the movement of a hypersonic vehicle along its nominal trajectory. The flight control module 702 is used to repeatedly execute flight control operations when a sudden obstacle is detected, until the hypersonic vehicle completes the entire flight process. The flight control operations include: acquiring the current flight state parameters of the hypersonic vehicle and the spatiotemporal geometric parameters of the sudden obstacle; inputting the current flight state parameters and spatiotemporal geometric parameters into a trajectory planning model; the trajectory planning model includes: a temporal feature extraction module and a global feature extraction module; based on the temporal feature extraction module, extracting the temporal correlation characteristics of the current flight state parameters and spatiotemporal geometric parameters to generate hidden states representing predicted trajectory information; based on the global feature extraction module, performing global feature extraction on the hidden states to obtain global features; determining the attitude control variables of the hypersonic vehicle based on the global features, and controlling the flight state of the hypersonic vehicle based on the attitude control variables.

[0074] Accordingly, the present invention also provides a terminal device, comprising: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in any of the above embodiments.

[0075] An embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance.

Claims

1. A method for intelligent trajectory planning of hypersonic vehicles for avoiding spatiotemporal random obstacles, characterized in that, include: During the movement of a hypersonic vehicle along its nominal trajectory, sudden obstacles are detected in real time. If a sudden obstacle is detected, the flight control operation is repeated until the hypersonic vehicle completes the entire flight process. The flight control operations include: Acquire the current flight status parameters of the hypersonic vehicle and the spatiotemporal geometric parameters of sudden obstacles; The current flight state parameters and the spatiotemporal geometric parameters are input into the trajectory planning model; wherein, the trajectory planning model includes: a temporal feature extraction module and a global feature extraction module; Based on the temporal feature extraction module, the temporal correlation characteristics of the current flight state parameters and spatiotemporal geometric parameters are extracted to generate hidden states that characterize the predicted trajectory information; Based on the global feature extraction module, global features are extracted from the hidden state to obtain global features; Based on global characteristics, determine the attitude control variables of the hypersonic vehicle, and control the flight state of the hypersonic vehicle based on the attitude control variables.

2. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 1, characterized in that, The nominal trajectory is generated in the following ways: Establish a dynamic model for hypersonic vehicles that takes into account the Earth's rotation and oblateness; The constraints in the dynamic model of the hypersonic vehicle are defined, including: mechanical constraints, thermal constraints, boundary constraints, and obstacle constraints. Based on the hypersonic vehicle dynamics model, the aforementioned constraints, and the performance function to be planned, an optimal control problem is constructed. The optimal control problem is solved using the Chebyshev pseudospectral method, and the nominal trajectory is output.

3. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 1, characterized in that, The training method for the trajectory planning model includes: Obtain the optimal trajectory sequence of the hypersonic vehicle under each preset flight mission; Based on the optimal trajectory sequence, several training samples are generated; each training sample includes: flight state parameter samples of the hypersonic vehicle at the current moment, spatiotemporal geometric parameter samples of sudden obstacles, and corresponding actual attitude control quantities. The trajectory planning model is trained by taking each training sample as input and the corresponding predicted attitude control quantity as output.

4. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 3, characterized in that, The process of training the trajectory planning model using each training sample as input and the corresponding predicted attitude control quantity as output includes: Initialize the network parameters of the trajectory planning model; Based on the training samples and the initialized network parameters, the network parameter update operation is repeatedly performed until the preset total number of training rounds is completed, and the final network parameters are saved. The network parameter update operation includes: Obtain training samples; Based on the current network parameters, the corresponding predicted attitude control quantity is output according to the flight state parameter samples of the current training samples and the spatiotemporal geometric parameter samples of sudden obstacles. The predicted attitude control quantity is compared with the corresponding actual attitude control quantity, and the current loss function value is calculated based on the comparison result. The Adam adaptive optimization algorithm is used to backpropagate and update the current network parameters based on the current loss function value to obtain the network parameters for the next round; and when the current training round reaches the preset learning rate decay period, the current learning rate is halved according to the preset decay coefficient.

5. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 4, characterized in that, The step of extracting the temporal correlation characteristics of the current flight state parameters and spatiotemporal geometric parameters to generate a hidden state for characterizing the predicted trajectory information includes: Input the candidate hidden state from the previous time step, the current flight state parameters, and the spatiotemporal geometry parameters of the sudden obstacle into the reset gate; The reset weight is calculated by resetting the gate based on the candidate hidden state of the previous time step, the current flight state parameters, and the spatiotemporal geometric parameters of the sudden obstacle. The current candidate hidden state is obtained based on the reset weight and the hidden state of the previous time step; Input the candidate hidden state from the previous time step, the current flight state parameters, and the spatiotemporal geometry parameters of the sudden obstacle into the update gate; The update weight is calculated based on the candidate hidden state of the previous time step and the current flight state parameters, using the update gate. The temporal feature extraction module obtains the hidden state of the current time step by linearly interpolating between the hidden state of the previous time step and the current candidate hidden state based on the updated weight.

6. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 5, characterized in that, The attitude control parameters include: desired angle of attack and desired roll angle; Controlling the flight state of a hypersonic vehicle based on attitude control variables includes: The hypersonic vehicle is controlled to fly according to the desired angle of attack and desired roll angle.

7. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 6, characterized in that, The flight status parameters include: the hypersonic vehicle's current position, current angle of attack, current roll angle, and current speed; The spatiotemporal geometric parameters of the sudden obstacle include: the discovery time, center longitude, center latitude, and radius of the sudden obstacle.

8. The intelligent trajectory planning method for hypersonic vehicles oriented towards spatiotemporal random obstacle avoidance as described in claim 7, characterized in that, The establishment of a hypersonic vehicle dynamics model that takes into account the Earth's rotation and oblateness includes: The dynamic model of a hypersonic vehicle considering Earth's rotation and oblateness is characterized as follows: in, It is the distance from the Earth's center. This is the aircraft's current speed. It is the aircraft's current speed tilt angle. This is the current longitude of the aircraft. It is the aircraft's current speed deflection angle. It is the current latitude of the aircraft. It is aerodynamic drag. It is aerodynamic lift. It is the mass of the aircraft. It is the current tilt angle of the aircraft. , , It is the component of gravitational acceleration in the spacecraft's velocity coordinate system. , , It is the component of centrifugal inertial force in the aircraft's velocity coordinate system. , , It is the component of the Coriolis inertial force in the aircraft's velocity coordinate system.

Images