A deep learning-based online planning method for high-speed aircraft descent timing

Abstract

This invention specifically relates to an online planning method for the timing of impact maneuvers in high-speed aircraft based on deep learning. It achieves high-precision long-term Koopman linear modeling of the aircraft's motion state through a deep neural network and constructs a mapping model between the impact state and the hit state. An optimal impact timing function is specifically proposed, enabling online planning of the impact timing after the aircraft's agile avoidance maneuvers. This improves the terminal accuracy, speed, and angle indicators of the aircraft after completing the agile avoidance maneuver. This method overcomes the limitations of existing solutions in terms of design adaptability and efficiency in online computation. It can plan the impact timing online based on the current aircraft state after completing various obstacle avoidance scenarios, significantly improving the terminal accuracy of high-speed aircraft.

Description

Technical Field

[0001] This invention relates to the field of aircraft motion control technology, specifically to an online planning method for the timing of downpressure in high-speed aircraft based on deep learning. Background Technology

[0002] In related technologies, trajectory reconstruction and avoidance technology based on active situational awareness has become an important research direction to improve the survivability of high-speed aircraft. That is, under the premise of determining the basic flight mission of the aircraft, active flight avoidance and close-range combat avoidance are carried out in a targeted manner based on the dynamic information obtained by the aircraft during flight, including refueling and acceleration, longitudinal / lateral skipping, and spiral trajectories. However, existing terminal flight control schemes for high-speed aircraft are difficult to meet the applicability and computational timeliness requirements of planning the timing of downward pressure after flexible avoidance in complex environments. The uncertainty of the timing and scheme of dynamic trajectory changes leads to uncertainty in the aircraft's state after the trajectory change, rendering the original predetermined flight scheme inapplicable. Therefore, it is necessary to re-optimize the aircraft trajectory online based on the aircraft's state after flexible avoidance to improve the accuracy and other performance characteristics of the aircraft's terminal flight control.

[0003] When a high-speed aircraft moves along a planned flight path, it needs to actively avoid or evade moving obstacles or areas that are dynamically perceived during flight. However, in existing technologies, the original flight path becomes inapplicable after the high-speed aircraft dynamically changes its trajectory and avoids obstacles. Due to the poor timeliness of dynamic trajectory calculation, it is difficult to guarantee that the trajectory planning after high-speed maneuvers meets the terminal accuracy, speed, angle, and other performance requirements of the predetermined mission.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] This invention provides an online planning method for the timing of high-speed aircraft's downward pressure based on deep learning, which solves the problems of accuracy and speed in the online optimization of flight trajectory after dynamic avoidance of high-speed aircraft in the prior art, and can effectively overcome the defects existing in the prior art.

[0006] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to a first aspect of the present invention, a deep learning-based online planning method for the timing of downpressure in high-speed aircraft is provided, the method comprising: When the aircraft enters the downpressure selection zone during the unpowered gliding phase, the aircraft's current motion state is acquired. ; to change the current motion state The motion state prediction model is used to predict the motion state of the aircraft and obtain the future motion state sequence. Configure system input state based on future motion state sequence The aircraft terminal state is estimated by using the aircraft down-hit state mapping model to estimate the system input state, and the hit point state prediction sequence is output. The hit point state prediction sequence includes: miss confidence, relative residual velocity and terminal tilt sine. Utilizing the optimal down-pressing timing function Global optimization of the timing of the downslope is performed on the hit point state prediction sequence to obtain the optimal downslope distance. It is used for the down-short phase guidance of aircraft.

[0008] In some exemplary embodiments, the method further includes: During the unpowered gliding phase, the aircraft conducts active situational awareness to acquire situational awareness information; Based on situational awareness information, an evasion strategy is determined for the object to be avoided, so that the aircraft can execute the evasion strategy and perform maneuvering to avoid the object; wherein, the evasion strategy includes: trigger slant range Duration of change Command angle of attack Command tilt angle Fuel ratio Direct control force in each coordinate axis direction under the velocity coordinate system.

[0009] In some exemplary embodiments, the commanded angle of attack is configured with a height damping term for eliminating height oscillations after the maneuvering obstacle avoidance ends, including:

[0010] in, Indicates the angle of attack of the command; Indicates the angle of attack with maximum lift-to-drag ratio; Indicates the height damping coefficient; Indicates the trajectory inclination angle; This represents the speed scalar of the aircraft.

[0011] In some exemplary embodiments, the method further includes: Configuring lateral overload Used to correct heading deviations caused by the lateral component of gravity and lateral maneuvers, including:

[0012] in, Indicates the guidance coefficient; This represents the relative derivative of the distance vector between the aircraft and the target with respect to time. This represents the derivative of the elevation angle of the line of sight with respect to time in the line-of-sight coordinate system. Based on the lateral overload requirement and the overload corresponding to the maximum lift-to-drag ratio angle of attack, configure the command tilt angle. ,include:

[0013] in, This indicates that lateral overload is required; This indicates the overload corresponding to the angle of attack with the maximum lift-to-drag ratio.

[0014] In some exemplary embodiments, the method further includes: pre-training a motion state prediction model, including: Based on the particle motion model and combined with the motion state characteristics of the aircraft, a dynamic model of the aircraft is constructed; whereby the motion state characteristics are used to represent the dynamic characteristics of the aircraft during its motion process. The system configures the aircraft's random initial state and random avoidance strategy during the unpowered gliding phase; and combines this with a dynamic model to configure multiple maneuvering obstacle avoidance trajectories to construct an original training set. The aircraft's state space corresponding to each maneuvering obstacle avoidance trajectory is generated independently within its respective distribution range and satisfies uniformly distributed random initial values, constituting a subset of all motion states during the unpowered gliding phase's programmed flight and maneuvering obstacle avoidance flight. Avoidance strategies include: trigger slant distance... Duration of change Command angle of attack Command tilt angle Fuel ratio Direct control forces in the directions of each coordinate axis in the velocity coordinate system; The original training set is normalized to obtain the corresponding normalized training set; each example in the normalized training set includes a six-dimensional sequence of system state variables and a two-dimensional sequence of control inputs. Constructing a Koopman backpropagation neural network The normalized training set is then input into the neural network to obtain the output sequence of predicted state variables. By concatenating the predicted state variable sequence with the system state variable sequence, the current flight state is obtained in an upgraded dimension. ; Upgrade the current flight status Concatenate with the control input sequence to obtain the Koopman up-dimensional state. ; Construct a neural network to predict output ; Upgrade Koopman to a higher dimension Input the prediction-output neural network and obtain the future flight state predicted by the prediction-output neural network in an upgraded dimension; The neural network is updated until the model converges based on the loss function; the loss function is constructed based on the reconstruction prediction loss, the linear prediction loss, and the infinite norm prediction error loss. Record the neural network parameters, configure the Koopman backpropagation neural network as the dimension-increasing function, and configure the prediction output neural network as the Koopman linear model; A motion state prediction model is constructed by combining the dimension-increasing function and the Koopman linear model.

[0015] In some exemplary embodiments, the Koopman backpropagation neural network includes: The first feature processing layer takes a 6-dimensional sequence of state variables as input and outputs a 16-dimensional sequence of the first intermediate state variables. The second feature processing layer takes the first intermediate state variable sequence as input and outputs a 16-dimensional second intermediate state variable sequence. The third feature processing layer takes a 16-dimensional sequence of second intermediate state variables as input and outputs a 16-dimensional sequence of third intermediate state variables. The fourth feature processing layer takes a 16-dimensional sequence of the third intermediate state variables as input and outputs a 16-dimensional sequence of predicted state variables. Each feature processing layer is represented as follows:

[0016] in, This represents the output of the feature processing layer; Indicates the activation function; These are the weight parameters for the feature processing layer; This represents the output of the previous feature processing layer; Indicates the bias parameter; , , which represents the number of hidden layers in a neural network.

[0017] In some exemplary embodiments, the method further includes: pre-training an aircraft downforce hit state mapping model, including: Based on the aircraft kinematic model, construct the aircraft's maneuvering obstacle avoidance trajectory in the unpowered gliding phase for the object to be avoided. Record the initial downward state and the flight state when the aircraft enters the downward state at each feature point to construct the original training set. The original training set is normalized to construct a system input state-based model. The normalized training set; where the system input states include: relative initial velocity Relative slope distance Relative hit time Relative positive distance relative height Relative yaw time Relative downward pressure distance ; Building a backpropagation neural network The normalized training set is input into the backpropagation neural network to obtain the output state variable sequence predicted by the model. ; Calculate the output state variable sequence With system input status The loss between the two is calculated, and the loss function is used to update the neural network until convergence; where the loss function is constructed based on the root mean square error.

[0018] In some exemplary implementations, backpropagation neural networks include: The first feature processing layer takes a 7-dimensional sequence of state variables as input and outputs a 10-dimensional sequence of the first intermediate state variables. The second feature processing layer takes the first intermediate state variable sequence as input and outputs a 10-dimensional second intermediate state variable sequence. The third feature processing layer takes the second intermediate state variable sequence as input and outputs a 10-dimensional third intermediate state variable sequence. The fourth feature processing layer takes the third intermediate state variable sequence as input and outputs a 3-dimensional output state variable sequence. .

[0019] According to a second aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the above-described deep learning-based online planning method for the timing of high-speed aircraft pressure drop is implemented.

[0020] According to a third aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described deep learning-based online planning method for the timing of high-speed aircraft pressure drops.

[0021] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the above-described deep learning-based online planning method for high-speed aircraft downpressure timing by executing the executable instructions.

[0022] The embodiments of this invention provide a deep learning-based online planning method for the timing of a high-speed aircraft's downward thrust. Based on the Deep-Koopman method, a motion state prediction model is obtained and used to predict the aircraft's motion state during the gliding phase. Simultaneously, by constructing a mapping model between the thrust state and the hit state, the aircraft's state at the end of flight is accurately estimated. A targeted optimal thrust timing function is proposed for global optimization of the thrust timing, enabling online planning of the thrust timing after agile maneuvering avoidance, thus improving the terminal accuracy, speed, and angle indicators of the aircraft after completing the agile avoidance maneuver. This method overcomes the limitations of existing solutions in terms of design adaptability and efficiency in online computation. It can plan the thrust timing online based on the current aircraft state after completing various obstacle avoidance scenarios, significantly improving the terminal accuracy of the high-speed aircraft's flight trajectory.

[0023] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0025] Figure 1 The illustration shows a schematic diagram of an online planning method for the timing of a high-speed aircraft's down-pressure maneuver based on deep learning, an exemplary embodiment of the present invention. Figure 2 This diagram illustrates an exemplary embodiment of the present invention of a dynamic evasive flight trajectory of an aircraft. Figure 3 schematically illustrates a comparison of aircraft motion state prediction according to an exemplary embodiment of the present invention; Figure 4 schematically illustrates an exemplary embodiment of the present invention of a prediction of the hit state of an aircraft under different downward slant ranges; Figure 5 schematically illustrates a schematic diagram of the results of a DKDIS Monte Carlo simulation experiment according to an exemplary embodiment of the present invention; Figure 6 The diagram illustrates a simulated three-dimensional maneuvering flight trajectory, an exemplary embodiment of the present invention. Figure 7 The diagram illustrates the composition of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0026] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0027] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0028] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a deep learning-based online planning method for the timing of high-speed aircraft's downward thrust, solving the problem of online trajectory optimization after dynamic avoidance maneuvers in existing high-speed aircraft. (Reference) Figure 1 As shown, the method includes: Step S11: When the aircraft enters the down-pressure selection zone during the unpowered gliding phase, the current motion state of the aircraft is acquired. ; to change the current motion state The motion state prediction model is used to predict the motion state of the aircraft and obtain the future motion state sequence. Step S12: Configure the system input state based on the future motion state sequence. The aircraft terminal state is estimated by using the aircraft down-hit state mapping model to estimate the system input state, and the hit point state prediction sequence is output. The hit point state prediction sequence includes: miss confidence, relative residual velocity and terminal tilt sine. Step S13, utilize the optimal pressing timing function Global optimization of the timing of the downslope is performed on the hit point state prediction sequence to obtain the optimal downslope distance. It is used for the down-short phase guidance of aircraft.

[0029] In this embodiment, a motion state prediction model based on the Deep-Koopman method is used to predict the motion state of a high-speed aircraft during the gliding phase. A DIM model mapping the aircraft's descent timing and landing point is constructed based on deep learning. An optimal descent timing function is proposed based on the aircraft's trajectory features and hit indicators. By combining the two prediction models and the optimal down-press timing function, the down-press timing within the aircraft's down-press selection zone can be globally optimized to improve the aircraft's final hit rate, hit speed, and hit angle.

[0030] The following will describe in more detail each step of the online planning method for the timing of high-speed aircraft pressure drop based on deep learning in this exemplary embodiment, with reference to the accompanying drawings and embodiments.

[0031] In this example implementation, the method described above further includes: Step S101: During the unpowered gliding phase, the aircraft performs active situational awareness to obtain situational awareness information. Step S102: Determine the evasion strategy for the object to be avoided based on situational awareness information, so that the aircraft can execute the evasion strategy and perform maneuvering obstacle avoidance on the object to be avoided.

[0032] Specifically, the aforementioned high-speed aircraft can be manned aircraft, unmanned aircraft, or other aviation flight equipment. For example, the aircraft can be a flexible air-breathing high-speed aircraft. Alternatively, the aircraft can be a bank-to-turn (BTT) high-speed aircraft with mid-course refueling and acceleration capabilities, exhibiting optimal aerodynamic performance in high-speed (approximately Mach 12) gliding mode with a maximum lift-to-drag ratio of approximately 4.2, and a maximum lift-to-drag ratio of approximately 2.0 in low-speed mode (minimum speed before target impact, approximately Mach 3). The aircraft engine generates a maximum axial acceleration of approximately 1.8g, and operates at maximum thrust for approximately 100 seconds.

[0033] Specifically, after completing the rocket boost, the high-speed vehicle enters the unpowered gliding phase. In addition to passive avoidance according to the mission plan, it uses situational awareness to maneuver and avoid temporarily appearing obstacles. These obstacles can be temporary dynamic no-fly zones, temporary dynamic obstacles, or static obstacles. The high-speed vehicle can maneuver and avoid them at close range. During the unpowered gliding phase, the vehicle can use angle-of-attack commands with altitude damping based on the maximum lift-to-drag ratio and a proportional guidance method to provide roll angle commands for programmed flight. During maneuvering and avoidance, it flexibly changes its trajectory by tracking the angle-of-attack and roll angle commands. Upon reaching the vicinity of the target, it determines the conditions for rapid descent and uses proportional guidance to reach the target, stopping when the altitude reaches zero.

[0034] Specifically, during the gliding phase, the aircraft can conduct active situational awareness to acquire situational awareness information in real time and determine evasion strategies based on this information. This situational awareness information can include the aircraft's system state and control inputs. The system state describes the aircraft's rigid body motion state and / or dynamic change characteristics.

[0035] The avoidance strategies include: trigger slant distance Duration of change Command angle of attack Command tilt angle Fuel ratio And, the direct control force in each coordinate axis direction under the velocity coordinate system. The unit is aircraft weight. Trigger slant range refers to the straight-line distance between the aircraft and the target, used to determine the timing of maneuvers. Variation duration can be used to indicate the duration of a maneuver. The formula is expressed as:

[0036] By configuring evasion strategies using the parameters described above, it is possible to adapt to different flight states and scenarios, and to accommodate different evasion ballistic patterns. This manifests as variations in trajectory change strategies, timing, and duration. (Reference) Figure 2 The diagram shown illustrates the dynamic evasive flight trajectory of an aircraft.

[0037] For example, the method further includes: configuring the command angle of attack with a height damping term for eliminating height oscillations after the maneuver obstacle avoidance ends, including:

[0038] in, Indicates the angle of attack of the command; Indicates the angle of attack with maximum lift-to-drag ratio; Indicates the height damping coefficient; Indicates the trajectory inclination angle; This represents the speed scalar of the aircraft.

[0039] Specifically, during the unpowered gliding phase, aircraft typically employ the maximum lift-to-drag ratio. Considering that evasive maneuvers such as vertical jumps can cause significant changes in altitude and vertical velocity, an altitude damping term is configured for the commanded angle of attack to eliminate altitude oscillations after the evasive maneuver, as shown in the formula above, based on... This constitutes a high-damping term. For example, the maximum lift-to-drag ratio and angle of attack. It can be configured to 4.5°.

[0040] For example, the method further includes: Configuring lateral overload Used to correct heading deviations caused by the lateral component of gravity and lateral maneuvers, including:

[0041] in, Indicates the guidance coefficient; This represents the relative derivative of the distance vector between the aircraft and the target with respect to time. This represents the derivative of the elevation angle of the line of sight with respect to time in the line-of-sight coordinate system. Based on the lateral overload requirement and the overload corresponding to the maximum lift-to-drag ratio angle of attack, configure the command tilt angle. ,include:

[0042] in, This indicates that lateral overload is required; This indicates the overload corresponding to the angle of attack with the maximum lift-to-drag ratio.

[0043] Specifically, in order to correct the lateral component of gravity and the heading deviation caused by lateral avoidance during aircraft maneuvers, guidance coefficients are used. The relative derivative of the distance vector between the aircraft and the target with respect to time. The derivative of the elevation angle of the line of sight with respect to time in the line-of-sight coordinate system To optimize lateral overload requirements. For example, guidance coefficient. It can be configured to 3.

[0044] Maximum lift-to-drag ratio angle of attack Overload is represented as Then the command tilt angle can be calculated using the formula mentioned above. ;in, .

[0045] In the above formula for calculating the command tilt angle, The function is represented as:

[0046] For example, in step S11, when the aircraft enters the down-press selection zone during the unpowered gliding phase, the current motion state of the aircraft is acquired. ; to change the current motion state The motion state prediction model is used to predict the motion state of the aircraft and obtain the future motion state sequence.

[0047] Specifically, once the aircraft enters the downforce selection zone during the unpowered gliding phase, a trained motion state prediction model can be used to predict its future motion state sequence. This could mean the aircraft first performs obstacle avoidance maneuvers during the unpowered gliding phase before entering the downforce selection zone, or it could directly enter the downforce selection zone during the unpowered gliding phase. The motion state prediction model is built upon a Koopman backpropagation neural network.

[0048] For example, the method further includes: pre-training a motion state prediction model, including: Step S21: Based on the particle motion model and combined with the motion state characteristics of the aircraft, construct the dynamic model of the aircraft; wherein, the motion state characteristics are used to represent the dynamic characteristics of the aircraft during the motion process. Step S22: Configure the random initial state and random avoidance strategy of the aircraft during the unpowered gliding phase; and configure multiple maneuvering obstacle avoidance trajectories in conjunction with the dynamic model to construct the original training set; wherein, the state space of the aircraft corresponding to the maneuvering obstacle avoidance trajectory is generated independently within its respective distribution range and satisfies uniformly distributed random initial values, constituting a subset of all motion states of the unpowered gliding phase programmed flight and maneuvering obstacle avoidance flight; the avoidance strategy includes: trigger slant distance Duration of change Command angle of attack Command tilt angle Fuel ratio Direct control forces in the directions of each coordinate axis in the velocity coordinate system; Step S23: Normalize the original training set to obtain the corresponding normalized training set; wherein each example in the normalized training set includes a six-dimensional system state variable sequence and a two-dimensional control input sequence; Step S24: Construct a Koopman backpropagation neural network The normalized training set is then input into the neural network to obtain the output sequence of predicted state variables. Step S25: Concatenate the predicted state variable sequence with the system state variable sequence to obtain the current flight state in upgraded dimensions. ; Step S26: Upgrade the current flight status. Concatenate with the control input sequence to obtain the Koopman up-dimensional state. ; Step S27: Construct a neural network for predicting output. ; Upgrade Koopman to a higher dimension Input the prediction-output neural network and obtain the future flight state predicted by the prediction-output neural network in an upgraded dimension; Step S28: Update the neural network based on the loss function until the model converges; wherein, the loss function is constructed based on the reconstruction prediction loss, the linear prediction loss, and the infinite norm prediction error loss; Step S29: Record the neural network parameters, configure the Koopman backpropagation neural network as the dimension-increasing function, and configure the prediction output neural network as the Koopman linear model; combine the dimension-increasing function and the Koopman linear model to construct a motion state prediction model.

[0049] Specifically, the motion state prediction model predicts the motion state of a high-speed aircraft under a given flight plan, mainly focusing on the position and velocity information required in the guidance problem. In order to facilitate modeling and data normalization, this invention adopts a target launch system, that is, the origin of the coordinate system is located at the target, the x-axis points to the gliding start point of the aircraft, and the y-axis points to the sky, forming a right-handed system.

[0050] Based on the particle motion model and combined with the motion state characteristics of the aircraft, a dynamic model of the aircraft is constructed. The formula is expressed as:

[0051] in, For speed, For the trajectory inclination angle, For the ballistic deflection angle, Latitude Longitude It is the geocentric oblique distance. For the mass of the aircraft, The three components of the direct control force in the ballistic system are expressed in units of the aircraft's weight. The three components of thrust in a ballistic system. Throttle ratio; These are the three components of the aerodynamic force in the velocity system.

[0052] (1) Construct the original training set. Based on the mission planning of the aircraft, generate multiple ballistic trajectories based on the motion model of the aircraft mass by using the initial state of the gliding segment of the high-speed aircraft and the randomly generated evasion strategy parameters. The state space of the velocity, position and motion states of these ballistic trajectories independently generates random initial values ​​that satisfy uniform distribution within their respective distribution ranges, forming a subset of all motion states of the gliding segment program flight and dynamic evasion flight.

[0053] The state variables and flight commands of the aircraft are configured using a random initial state generation algorithm, as expressed by the formula:

[0054]

[0055] in, This represents the state variables of the aircraft, including the aircraft's position coordinates in the target launch system. and the three components of the aircraft's velocity ; The flight commands for the aircraft are the angle of attack of the aircraft. and tilt angle . This refers to the `random.uniform` function in Python. Flight commands serve as the control input for the aircraft. This is represented as a longitudinal plane model of a high-speed aircraft.

[0056] Simulating the agile obstacle avoidance maneuvers of a high-speed aircraft under active situational awareness, as described in the above embodiments, the avoidance strategy is described based on the following parameters, including: trigger slant range. Duration of change Command angle of attack Command tilt angle Fuel ratio ; and the direct control force in each coordinate axis direction under the velocity coordinate system. .

[0057] Referring to Table 1, the specific parameters corresponding to different trajectory patterns are shown. Scheme A represents a longitudinal skip flight scheme with constant angle of attack and fuel ratio under powered conditions; Scheme B represents a longitudinal / lateral trajectory variation flight scheme with constant angle of attack and roll angle under unpowered conditions. The upper / lower bounds of the parameters indicate the selection range of the trajectory variation parameters, and the random numbers are uniformly distributed.

[0058] Table 1

[0059] Using the Runge-Kutta fourth-order algorithm with a step size of 0.01s, 150 complete ballistic trajectories were obtained by further calculations on the randomly generated initial state and evasion strategy. Based on the above trajectories, the trajectory of the downforce selection zone before downforce was selected, that is, the trajectory within the slant distance range of 200-60km. Feature points were taken every 2km, and the motion state of the i-th feature point was inherited. Based on this, 5 sets of flight commands are randomly generated. This resulted in an original training set of 52,500 cases.

[0060] (2) Constructing a normalized training set For the original training set, a normalization method is selected to obtain a normalized training set, where each example includes a five-dimensional sequence of system state variables and a two-dimensional sequence of control inputs. The normalization method formula is expressed as:

[0061]

[0062] in, ; Represents the state variables of the aircraft; Indicates the control inputs of the aircraft; and This refers to the initial value or average value.

[0063] (3) Calculate the current flight state dimensionality upgrade Constructing a Koopman backpropagation neural network Input the normalized training set .

[0064] In each example, a 6-dimensional sequence of state variables is input. The first layer yields a 16-dimensional sequence of state variables. This 16-dimensional sequence of state variables is then input... The second layer yields a 16-dimensional sequence of state variables, which is then input into... The third layer yields a 16-dimensional sequence of state variables, which is then input into... The fourth layer yields a 16-dimensional sequence of state variables.

[0065] Among them, the Koopman backpropagation neural network Each feature processing layer in the middle is represented as follows:

[0066] in, This represents the output of the feature processing layer; Indicates the activation function; These are the weight parameters for the feature processing layer; This represents the output of the previous feature processing layer; Indicates the bias parameter; , , which represents the number of hidden layers in a neural network.

[0067] Among them, activation function The formula is expressed as:

[0068] Where e is the natural constant, and x is the input to the activation function. This is the output of the activation function.

[0069] The 6-dimensional state variable sequence in the normalized training set is summed with the Koopman backpropagation neural network. The output 16-dimensional state variable sequence is concatenated to obtain a 22-dimensional current flight state. The formula is expressed as:

[0070] Then, upgrade the current flight state to a higher dimension. With two-dimensional control input By splicing them together, we obtain the complete 24-dimensional Koopman up-dimensional state. The formula is expressed as:

[0071] (4) Calculate the prediction of the upgraded state Construct a neural network to predict output Koopman backpropagation neural network The output is a 24-dimensional upscaling neural network that predicts the current flight state from the input. This yields a 24-dimensional up-dimensional state prediction. The prediction output neural network... The formula is expressed as:

[0072] in, This represents the output parameters of the output layer of the predictive output neural network; Indicates the weights of the output layer; This indicates the output of the previous layer.

[0073] (5) Calculate the future flight state dimensionality upgrade The six-dimensional state variable sequence of each example in the normalized training set is input into the longitudinal plane model of the high-speed aircraft to obtain a six-dimensional predicted state variable sequence.

[0074] Simultaneously, the six-dimensional state variable sequence of each example in the normalized training set is input into the Koopman backpropagation neural network. This yields a 16-dimensional sequence of state variables.

[0075] The six-dimensional predicted state variable sequence output from the physical model of the high-speed aircraft and the Koopman backpropagation neural network are used. The output 16-dimensional state variable sequence is concatenated to obtain a 22-dimensional future flight state.

[0076] (6) Updating the neural network based on the loss function During training, a multi-step prediction loss function is constructed based on reconstruction prediction loss, linear prediction loss, and infinite norm prediction error loss.

[0077] Among them, the reconstruction prediction loss function Represented as:

[0078] in, Indicates the length of the prediction time domain during training; This represents the weight decay coefficient; This represents the system state of the aircraft at time k+i; This indicates that the flight state of the aircraft at time k is upgraded to a higher dimension.

[0079] in, .in, It is the identity matrix. It is a zero matrix.

[0080] Linear prediction loss function The formula is expressed as:

[0081] in, This represents the upgraded state of the aircraft's actual state.

[0082] in, .in, It is the identity matrix. and It is a zero matrix.

[0083] Infinite norm prediction error loss function The formula is expressed as:

[0084] in, This represents the coefficient matrix.

[0085] The total loss function is defined as follows:

[0086] in, This is the total loss function; , , These represent the weighting coefficients of each loss term.

[0087] Among them, when and When the quantity is small, it can also be simplified to:

[0088] in, The formula for mean square error is... To upgrade the dimensionality of the true state at step n, For the n-step linear prediction of the upgraded state at time k.

[0089] Using the Adam optimizer to update the parameters of the backpropagation neural network at a learning rate of 1e-3, under the current method, the model iterates 2 million times, and the residuals of the loss function tend to converge, which is considered as training completion.

[0090] (7) Obtain the Koopman linearized model Record neural network parameters, This is the Koopman dimension-raising function. This is the Koopman linear model.

[0091] By outputting the neural network weight The system matrix of the linear model of the high-speed aircraft is obtained. and control matrix ;Right now,

[0092] in, This indicates an upgrade in the flight state; This indicates a control input.

[0093] The trained deep learning-based Koopman linearization modeling method was applied to predict the state of a high-speed aircraft within its down-press selection zone. The simulation step size was 0.01 s, and the prediction duration was 20 s. Referring to Figures 3(a)-3(f), the comparisons of the predicted trajectory and reference trajectory in the x-coordinate direction, and the comparison of the x-direction velocity components are shown in the figures. Similarly, the comparisons of the predicted trajectory and reference trajectory in the y-coordinate direction, and the comparison of the y-direction velocity components are shown in the figures. The figures also show that, compared with the simulation results of the aircraft's mathematical model, the motion state prediction model achieves high-precision long-term prediction of the aircraft's motion state within the down-press selection zone. The prediction accuracy and prediction duration meet the requirements of the down-press timing optimization design.

[0094] For example, in step S12, the system input state is configured based on the future motion state sequence. The aircraft terminal state is estimated by using the aircraft down-hit state mapping model to estimate the system input state and output the hit point state prediction sequence. The hit point state prediction sequence includes: miss confidence, relative residual velocity and terminal tilt sine.

[0095] Specifically, a Dive-Impact Mapping Model (DIM) can be pre-trained to obtain the Dive-Impact Mapping Model (DIM) which shows the mapping relationship between the timing of the Dive-Impact and the landing point.

[0096] For example, the method further includes: pre-training an aircraft downforce hit state mapping model, including: Step S31: Based on the aircraft kinematic model, construct the aircraft's maneuvering obstacle avoidance trajectory for the object to be avoided in the unpowered gliding phase. Record the initial downward state and the flight state at the time of impact when the aircraft enters the downward state at each feature point to construct the original training set. Step S32: Normalize the original training set and construct a system input state-based training set. The normalized training set; where the system input states include: relative initial velocity Relative slope distance Relative hit time Relative positive distance relative height Relative yaw time Relative downward pressure distance ; Step S33: Construct a backpropagation neural network The normalized training set is input into the backpropagation neural network to obtain the output state variable sequence predicted by the model. ; Step S34, calculate the output state variable sequence With system input status The loss between the two is calculated, and the loss function is used to update the neural network until convergence; where the loss function is constructed based on the root mean square error.

[0097] Specifically, based on the motion prediction model, a high-precision prediction of the aircraft's motion state within the down-press selection zone is obtained. This section establishes a down-press hit state mapping model based on a deep neural network to establish the relationship between the aircraft's down-press point motion state and its motion state at the time of impact. This enables the prediction of the miss distance, hit velocity, and trajectory inclination angle at the time of impact corresponding to different down-press timings.

[0098] (1) Constructing the original training set Based on 150 avoidance trajectories generated from the aircraft motion state prediction model, the aircraft enters a down-pressure state at each feature point. The down-pressure phase simulation is completed using the Runge-Kutta fourth-order algorithm with a step size of 0.01s. The initial down-pressure state and the flight state at the time of impact are recorded, generating 10,500 original training cases.

[0099] (2) Constructing a normalized training set To improve the learning performance of neural networks, a special normalization method is adopted for the original training set. Based on the guidance method used in the compression phase, a method is proposed that uses relative initial velocity. Relative slope distance Relative hit time Relative positive distance relative height Relative yaw time Relative downward pressure distance The input states of the DIM system are expressed by the following formula: .in,

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] in, and This represents the three components of position and velocity when the aircraft under the target launch system begins to descend; Indicates the slope distance of the pressing point. This indicates the magnitude of the downward pressure point velocity.

[0107] Meanwhile, the terminal state of the aircraft's flight is represented as follows: The off-target confidence level is proposed. Relative residual velocity and the sine of the end inclination angle The output states that make up the DIM model ;in,

[0108]

[0109]

[0110] in, Indicates the magnitude of the final velocity. Indicates the magnitude of the downward pressure point velocity. Indicates the final miss distance. Indicates the off-target confidence distance.

[0111] (3) Constructing a backpropagation neural network

[0112] Building a backpropagation neural network Input the normalized training set In each example, a 7-dimensional sequence of state variables is input. The first layer yields a 10-dimensional sequence of state variables; the 10-dimensional sequence of state variables is then input... The second layer yields a 10-dimensional sequence of state variables; the 10-dimensional sequence of state variables is then input... The third layer yields a 10-dimensional sequence of state variables; the 10-dimensional sequence of state variables is then input... The fourth layer ultimately yields a 3D prediction of the system output state variable sequence. ,Right now .

[0113] Among them, neural networks Each layer in the matrix can be represented as:

[0114] in, This represents the output of the feature processing layer; Indicates the activation function; These are the weight parameters for the feature processing layer; This represents the output of the previous feature processing layer; Indicates the bias parameter; , , which represents the number of hidden layers in a neural network.

[0115] Among them, activation function The formula is expressed as:

[0116] Where e is the natural constant, and x is the input to the activation function. This is the output of the activation function.

[0117] (4) Update the neural network based on the loss function During the training process, the training residual formula adopts the root mean square error formula. The formula is expressed as:

[0118] Using the Adam optimizer to update the parameters of the backpropagation neural network at a learning rate of 1e-3, under the current method, the model iterates 2 million times, and the residuals of the loss function tend to converge, which is considered as training completion.

[0119] Using the simulation results of the aircraft's particle dynamics model as a reference trajectory, the aircraft state predictions from the DIM model were compared with the reference state by referring to Figures 4(a)-4(c), which respectively show the comparison between the model predictions and the reference state for the miss confidence, relative remaining velocity, and terminal trajectory inclination sine at different down-pressure slant ranges. As shown in the figures, the DIM model can make high-precision predictions of the miss confidence, relative remaining velocity, and trajectory inclination sine at different down-pressure timings, providing a foundation for subsequent down-pressure timing optimization design.

[0120] For example, in step S13, the optimal pressing timing function is used. Global optimization of the timing of the downslope is performed on the hit point state prediction sequence to obtain the optimal downslope distance. It is used for the down-short phase guidance of aircraft.

[0121] Specifically, the predicted hit point state sequence obtained from the DIM model input can be input into the optimal pressure timing function. The optimal downslope range is obtained. The aircraft will then perform downslope guided flight at this range.

[0122] Specifically, based on the predictions of the motion state prediction model and the down-pressure hit state mapping model, an optimal down-pressure timing function is proposed to evaluate the performance index of the down-pressure results at each moment within the down-pressure selection zone of the aircraft.

[0123] First, whether the target is ultimately hit is a decisive indicator for aircraft trajectory optimization, therefore it needs to be considered separately, and the current hit confidence distance needs to be set. The terminal velocity is set to 1m. If the distance at the end of the flight path exceeds this range, the downward slant range is considered unsuitable. Furthermore, considering the requirements of rapid downward movement and high-speed strike, it is desirable to increase the terminal velocity and trajectory angle while ensuring the predicted target hit. Therefore, an optimal downward timing function is proposed. The formula is expressed as:

[0124] in, , These represent the weighting coefficients; for example, their values ​​can be configured to be 0.9 and 0.1. This indicates the maximum tolerable off-target confidence level; for example, it can be configured to have a value of 0.5. It is the downslope margin; for example, it can be configured to have a value of 5.

[0125] All of the above coefficients can be adjusted appropriately according to the needs of the scenario.

[0126] For example, when an aircraft enters a roll-down phase within the roll-down selection zone, in order to ensure the accuracy of the aircraft's flight trajectory at the end, maintain the error between the aircraft and the target / landing point, and with the rapid descent in altitude, the normal direction needs to be overloaded. And lateral overload required Calculated by the following formula:

[0127] in, Let be the derivative of the elevation angle of the line of sight with respect to time in the line-of-sight angular coordinate system. This is the derivative of the line-of-sight azimuth angle with respect to time in the line-of-sight angular coordinate system.

[0128] When the aircraft enters a roll-down phase within the roll-down selection zone, this normal-direction overload is required. And lateral overload required Convert to angle of attack and yaw angle commands. The formula is expressed as:

[0129]

[0130] in, Angle of attack command; This indicates that the bisection method is used to solve the problem through an aerodynamic model. This is the tilt angle command. The guidance coefficient is selected in the same way as the general proportional guidance coefficient selection method to adjust the shape of the aircraft's flight trajectory, thereby raising or lowering the trajectory to achieve the terminal landing speed, landing angle, and accuracy requirements. In this embodiment, a fixed value of 3 is chosen to simplify the analysis of the relationship between the timing of the downward pressure and the landing point state.

[0131] For example, when the aircraft enters the down-pressure selection zone, i.e., the slant distance to the target is less than 200km, the current motion state of the aircraft is acquired. The input vector of the motion state prediction system is obtained by normalization. Through motion state prediction network Obtain the future state sequence Thus, the corresponding DIM model input state is obtained. ,pass The model obtains the hit point prediction state sequence. ,Will Substituting the optimal down-press timing function, we obtain the optimal down-press slant distance. The aircraft will perform down-slope guided flight at this slant range.

[0132] Referring to Figures 5(a)-5(h), the following are presented sequentially: comparison of the downward slant range planning results, DKDIS miss confidence prediction, optimized miss confidence, comparison of miss frequency graphs, optimized terminal velocity, optimized terminal tilt angle, optimized terminal velocity, and optimized terminal tilt angle. As shown in the figures, the optimized performance of the DKDIS algorithm under different aircraft scenarios and evasion patterns is demonstrated. In 1000 Monte Carlo simulation tests, compared to a fixed downward slant range of 140km, the hit rate of the optimized DKDIS algorithm increased from 84.4% to 99.3%. Comparing the terminal velocity and trajectory tilt angle of the hit examples reveals that the optimized trajectory has a larger terminal velocity and terminal tilt angle. Compared to the conservative use of the maximum downward slant range of 220km, the average terminal velocity increased by 410.3m / s, and the average terminal trajectory tilt angle improved by 18.6°.

[0133] For example, this method can also train a Deep-Koopman-based maneuvering flight control model for attitude prediction and maneuver avoidance during the gliding phase. Addressing the uncertainties and strong coupling in high-speed aircraft modeling and control system design, the data-driven Deep-Koopman method captures the influence of the aircraft's elastic modes on the rigid body's motion state through the dynamic changes in the rigid body's motion state.

[0134] Specifically, define the aircraft system state vector. and control input vector .in,

[0135]

[0136] To improve modeling accuracy, it is necessary to consider the effects of states such as velocity, altitude, mass, and fuel ratio on the dynamics of the aircraft around its center of mass. Therefore, the original system state vector is augmented to obtain the system input vector. , is represented as:

[0137] in, Indicates speed, Indicates the trajectory inclination angle, Indicates the trajectory inclination angle, Indicates latitude, Indicates longitude. Indicates altitude, Indicates angle of attack. Indicates the sideslip angle. Indicates the tilt angle. Indicates the roll angular velocity. Indicates yaw rate, Indicates pitch angular velocity, Indicates quality, Indicates the roll deflection angle. Indicates the yaw angle. Indicates the pitch deflection angle. This indicates the fuel ratio.

[0138] For the upgraded state To augment the definition, the state of the maneuvering flight control system is defined as follows:

[0139] Therefore, the maneuvering flight control model satisfies the following equation:

[0140]

[0141] Based on the aforementioned dynamic model, sampling was performed in a manner similar to that described in the previous embodiments to generate the original training set. The establishment of the maneuvering flight control model primarily employed sampling between trajectory feature points. Using the initial state of the high-speed aircraft's glide phase and randomly generated maneuvering strategy parameters, multiple ballistic trajectories were generated based on the aircraft's point mass motion model. The state space comprised of the velocity and positional motion states contained in these ballistic trajectories constitutes a subset of all motion states during the aircraft's glide phase programmed flight and maneuvering flight. The simulation initial state settings are shown in Table 2 below.

[0142] Table 2

[0143] The specific parameters of the two-stage maneuver scheme are shown in Table 1. A simulation step size of 0.01s and a fourth-order Runge-Kutta method were used for numerical simulation, generating 10 trajectories. The simulation results are as follows: Figure 6 As shown.

[0144] Based on the above trajectory, the trajectory before compression is taken, and feature points are selected at slant intervals of 10km, inheriting the motion state of the i-th feature point. Each feature point generates 100 sets of random combinations of attitude states and control inputs, as shown in Table 3 below, which displays the random initial state settings at the feature points.

[0145] Table 3

[0146] Using a six-degree-of-freedom mathematical model of a high-speed aircraft, dynamic calculations were performed for each combination in 0.001s steps. The resulting state sequences were used as examples to obtain a total of 100,000 original training sets.

[0147] By analyzing and normalizing the data required in the original training set, a total of 100,000 state training sets were obtained. The distribution range of each state and the normalization parameters are shown in Table 4.

[0148] Table 4

[0149] Similar to the above embodiments, a Deep-Koopman-based maneuvering flight control model was constructed. To balance accuracy and control solution complexity, the linear system dimension was set to 64. During training, the computing platform used Windows 11, an i5-13600kf CPU, and an RTX 4080 GPU. The training time for the current task was approximately 1.86 days. Finally, the system state prediction was obtained. The root mean square error is The loss function can be a total loss function constructed based on reconstruction prediction loss, linear prediction loss, and infinite norm prediction error loss. Using the aircraft dynamics mathematical simulation results as a reference trajectory, the aircraft state predictions from the maneuvering flight control model are compared. The simulation step size is 0.001s, and the prediction time is 0.1s.

[0150] For the trained maneuvering flight control model, the DKMPC method is used for control system design, and the controlled state... .

[0151] To improve control system performance and address the complex parameter optimization design problems inherent in multidimensional systems, a command shaping and matching optimization method for attitude control is proposed. For target attitudes with different maneuver patterns, they are discretized into unsteady reference states, as expressed by the following formula:

[0152]

[0153]

[0154] in, , and This represents the reference state input to the MPC control system at time t. This means taking the maximum of the two values. This means taking the minimum of the two values. and Indicates the current state. and This indicates the maximum allowable state tracking deviation.

[0155] To further improve the control performance of the control system, an angular velocity reference state function is introduced, expressed by the formula:

[0156]

[0157]

[0158] In the formula, , and This represents the tracking coefficient, which is a constant value, where the yaw rate is... Since the weight parameter of the control system is 0, its reference state is... The default value is 0.

[0159] In summary, the reference state vector is formed as follows: .

[0160] The method provided in this invention offers a high-precision and high-speed online optimization method for high-speed aircraft descent timing based on deep learning and Koopman operator theory (Deep-Koopman-Based Approach for Dive-Impact Start-point, DKDIS). It establishes a high-precision linear surrogate model using the Deep-Koopman method to obtain a motion state prediction model. Simultaneously, it employs deep learning to obtain a descent hit state mapping model (DIM) that maps the aircraft's descent timing to its impact point state. Based on this, an online optimization method for descent timing is proposed. This method calculates the optimal descent timing function to achieve online optimization of the aircraft's descent timing, thereby improving the aircraft's terminal accuracy, angle, and velocity.

[0161] During the gliding phase, the aircraft can perform close-range maneuvering to avoid dynamic or static targets. Upon entering the downforce selection zone (i.e., when the slant distance to the target is less than 200 km), the aircraft's current motion state is acquired. This state is then normalized to obtain the input vector for the motion state prediction system, which is then processed by the motion state prediction network. The future state sequence is obtained, and from this, the corresponding DIM model input states are obtained. The predicted state sequence of the hit point is obtained, and it is substituted into the optimal down-press timing function to obtain the optimal down-press slant distance. The aircraft will perform guided flight in the down-slope phase at this slant distance. This ensures that the high-speed aircraft, after dynamically changing its trajectory and achieving avoidance maneuvers, maintains the terminal accuracy, speed, angle, and other performance requirements of the predetermined mission.

[0162] It should be noted that the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may, for example, be executed synchronously or asynchronously in multiple modules.

[0163] It should be noted that although several modules or units of the device for performing actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0164] Figure 7 A schematic diagram of an electronic device suitable for implementing embodiments of the present invention is shown.

[0165] It should be noted that, Figure 7 The electronic device 1000 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0166] like Figure 7 As shown, the electronic device 1000 includes a Central Processing Unit (CPU) 1001, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 1002 or programs loaded from storage section 1008 into Random Access Memory (RAM) 1003. The RAM 1003 also stores various programs and data required for system operation. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An Input / Output (I / O) interface 1005 is also connected to the bus 1004. Furthermore, the electronic device 1000 also includes an FPGA device and a System-on-a-Chip (SoC) device.

[0167] The following components are connected to I / O interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to I / O interface 1005 as needed. Removable media 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1010 as needed so that computer programs read from them can be installed into storage section 1008 as needed.

[0168] In particular, according to embodiments of the present invention, the processes described below with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by central processing unit (CPU) 1001, it performs various functions defined in the system of this application.

[0169] Specifically, the aforementioned electronic devices can be airborne intelligent electronic devices or intelligent terminal devices that communicate with the aircraft. The model can be pre-trained locally on the terminal device or on a server, and then deployed on the intelligent terminal device or airborne electronic device, executing the methods described above.

[0170] It should be noted that the storage medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, wherein computer-readable program code is carried. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0171] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0172] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0173] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The aforementioned storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments. For example, the electronic device may perform... Figure 1 The steps of the method shown.

[0174] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0175] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0176] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0177] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A deep learning-based online planning method for the timing of downpressure in high-speed aircraft, characterized in that, The method includes: When the aircraft enters the downpressure selection zone during the unpowered gliding phase, the aircraft's current motion state is acquired. ; to change the current motion state The motion state prediction model is used to predict the motion state of the aircraft and obtain the future motion state sequence. Configure system input state based on future motion state sequence The aircraft terminal state is estimated by using the aircraft down-hit state mapping model to estimate the system input state, and the hit point state prediction sequence is output. The hit point state prediction sequence includes: miss confidence, relative residual velocity and terminal tilt sine. Utilizing the optimal down-pressing timing function Global optimization of the timing of the downslope is performed on the hit point state prediction sequence to obtain the optimal downslope distance. It is used for the down-short phase guidance of aircraft.

2. The method according to claim 1, characterized in that, The method further includes: During the unpowered gliding phase, the aircraft conducts active situational awareness to acquire situational awareness information; Based on situational awareness information, an evasion strategy is determined for the object to be avoided, so that the aircraft can execute the evasion strategy and perform maneuvering to avoid the object; wherein, the evasion strategy includes: trigger slant range Duration of change Command angle of attack Command tilt angle Fuel ratio Direct control force in each coordinate axis direction under the velocity coordinate system.

3. The method according to claim 2, characterized in that, The command angle of attack is equipped with a height damping term to eliminate height oscillations after the maneuver obstacle avoidance is completed, including: in, Indicates the angle of attack of the command; Indicates the angle of attack with maximum lift-to-drag ratio; Indicates the height damping coefficient; Indicates the trajectory inclination angle; This represents the speed scalar of the aircraft.

4. The method according to claim 2, characterized in that, The method further includes: Configuring lateral overload Used to correct heading deviations caused by the lateral component of gravity and lateral maneuvers, including: in, Indicates the guidance coefficient; This represents the relative derivative of the distance vector between the aircraft and the target with respect to time. This represents the derivative of the elevation angle of the line of sight with respect to time in the line-of-sight coordinate system. Based on the lateral overload requirement and the overload corresponding to the maximum lift-to-drag ratio angle of attack, configure the command tilt angle. ,include: in, This indicates that lateral overload is required; This indicates the overload corresponding to the angle of attack with the maximum lift-to-drag ratio.

5. The method according to claim 1, characterized in that, The method further includes: pre-training a motion state prediction model, including: Based on the particle motion model and combined with the motion state characteristics of the aircraft, a dynamic model of the aircraft is constructed; whereby the motion state characteristics are used to represent the dynamic characteristics of the aircraft during its motion process. The system configures the aircraft's random initial state and random avoidance strategy during the unpowered gliding phase; and combines this with a dynamic model to configure multiple maneuvering obstacle avoidance trajectories to construct an original training set. The aircraft's state space corresponding to each maneuvering obstacle avoidance trajectory is generated independently within its respective distribution range and satisfies uniformly distributed random initial values, constituting a subset of all motion states during the unpowered gliding phase's programmed flight and maneuvering obstacle avoidance flight. Avoidance strategies include: trigger slant distance... Duration of change Command angle of attack Command tilt angle Fuel ratio Direct control forces in the directions of each coordinate axis in the velocity coordinate system; The original training set is normalized to obtain the corresponding normalized training set; each example in the normalized training set includes a six-dimensional sequence of system state variables and a two-dimensional sequence of control inputs. Constructing a Koopman backpropagation neural network The normalized training set is then input into the neural network to obtain the output sequence of predicted state variables. By concatenating the predicted state variable sequence with the system state variable sequence, the current flight state is obtained in an upgraded dimension. ; Upgrade the current flight status Concatenate with the control input sequence to obtain the Koopman up-dimensional state. ; Construct a neural network to predict output ; Upgrade Koopman to a higher dimension Input the prediction-output neural network and obtain the future flight state predicted by the prediction-output neural network in an upgraded dimension; The neural network is updated until the model converges based on the loss function; the loss function is constructed based on the reconstruction prediction loss, the linear prediction loss, and the infinite norm prediction error loss. Record the neural network parameters, configure the Koopman backpropagation neural network as the dimension-increasing function, and configure the prediction output neural network as the Koopman linear model; A motion state prediction model is constructed by combining the dimension-increasing function and the Koopman linear model.

6. The method according to claim 5, characterized in that, Koopman backpropagation neural networks include: The first feature processing layer takes a 6-dimensional sequence of state variables as input and outputs a 16-dimensional sequence of the first intermediate state variables. The second feature processing layer takes the first intermediate state variable sequence as input and outputs a 16-dimensional second intermediate state variable sequence. The third feature processing layer takes a 16-dimensional sequence of second intermediate state variables as input and outputs a 16-dimensional sequence of third intermediate state variables. The fourth feature processing layer takes a 16-dimensional sequence of the third intermediate state variables as input and outputs a 16-dimensional sequence of predicted state variables. Each feature processing layer is represented as follows: in, This represents the output of the feature processing layer; Indicates the activation function; These are the weight parameters for the feature processing layer; This represents the output of the previous feature processing layer; Indicates the bias parameter; , , which represents the number of hidden layers in a neural network.

7. The method according to claim 1, characterized in that, The method further includes: pre-training an aircraft downforce hit state mapping model, including: Based on the aircraft kinematic model, construct the aircraft's maneuvering obstacle avoidance trajectory in the unpowered gliding phase for the object to be avoided. Record the initial downward state and the flight state when the aircraft enters the downward state at each feature point to construct the original training set. The original training set is normalized to construct a system input state-based model. The normalized training set; where the system input states include: relative initial velocity Relative slope distance Relative hit time Relative positive distance relative height Relative yaw time Relative downward pressure distance ; Building a backpropagation neural network The normalized training set is input into the backpropagation neural network to obtain the output state variable sequence predicted by the model. ; Calculate the output state variable sequence With system input status The loss between the two is calculated, and the loss function is used to update the neural network until convergence; where the loss function is constructed based on the root mean square error.

8. The method according to claim 7, characterized in that, Backpropagation Neural Network include: The first feature processing layer takes a 7-dimensional sequence of state variables as input and outputs a 10-dimensional sequence of the first intermediate state variables. The second feature processing layer takes the first intermediate state variable sequence as input and outputs a 10-dimensional second intermediate state variable sequence. The third feature processing layer takes the second intermediate state variable sequence as input and outputs a 10-dimensional third intermediate state variable sequence. The fourth feature processing layer takes the third intermediate state variable sequence as input and outputs a 3-dimensional output state variable sequence. .

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the online planning method for the timing of high-speed aircraft pressure-down based on deep learning, as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the online planning method for high-speed aircraft downpressure timing based on deep learning, as described in any one of claims 1 to 8, by executing the executable instructions.

Images