High-speed glider control quantity identification method and system based on long short-term memory network

Abstract

The application provides a high-speed glider control variable identification method and system based on a long short-term memory network, comprising the following steps: S1, dividing target attack angle and roll angle intervals, and defining attack angle and roll angle labels; S2, constructing an attack angle identification network based on a long short-term memory network and a roll angle identification network with Self-Attention, and training the attack angle identification network and the roll angle identification network using existing target trajectory data; and S3, identifying control variable labels using the trained attack angle identification network and the roll angle identification network, and outputting control variables according to the identified label results.

Description

Technical Field

[0001] This invention relates to the field of missile guidance and control technology, and more specifically, to a method and system for identifying control quantities of high-speed gliders based on long short-term memory networks. Background Technology

[0002] High-speed gliders fly at high speeds, typically exceeding Mach 5. They often employ aerodynamic shapes with high lift-to-drag ratios, such as waverider bodies, lifting bodies, and blended wing-body designs. They possess extremely high speeds and agile maneuverability, enabling them to perform designated maneuvers such as rapid evasion, no-fly zone maneuvers, maneuvering turns, and high-speed descents within specific airspaces. Their trajectory characteristics exhibit non-inertial trajectory behavior. Classical kinematic models for maneuvering targets are suitable for inertial trajectories. Therefore, when using these models to estimate or predict the trajectory of boost-gliders, model mismatches inevitably occur, leading to decreased estimation accuracy or even divergence. Thus, a prerequisite for high-precision trajectory estimation and prediction of targets is the construction of a high-accuracy target motion model. Accurate estimation of the aerodynamic forces acting on the target is the core issue in constructing the target motion model, and the magnitude of these aerodynamic forces is determined by the control variables angle of attack and roll angle. Therefore, the identification of control variables for high-speed glider targets is a crucial research problem.

[0003] Patent document CN105550402A (application number: 201510888393.9) discloses a design method for a hypersonic stable maneuvering glide trajectory based on frequency conversion of angle of attack or tilt angle. It has eight steps: 1. Constraint modeling of the hypersonic vehicle reentry process; 2. Establishment of the angle of attack corridor for stable glide; 3. Establishment of the tilt angle corridor for stable glide; 4. Design of the reference angle of attack curve for stable maneuvering glide; 5. Design of the reference tilt angle curve for stable maneuvering glide; 6. Solving for the initial glide altitude and trajectory tilt angle; 7. Constant damping differential feedback control scheme; 8. Generation of the stable maneuvering glide trajectory.

[0004] This invention utilizes long short-term memory networks and self-attention mechanisms to perform online identification of target angle of attack and tilt angle. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for identifying control variables of high-speed gliders based on long short-term memory networks.

[0006] A method for identifying control variables of a high-speed glider based on a long short-term memory network, provided by the present invention, includes:

[0007] Step S1: Divide the target angle of attack and heave angle ranges, and define labels for the angle of attack and heave angle;

[0008] Step S2: Construct an angle of attack recognition network based on a long short-term memory network and a tilt angle recognition network that introduces self-attention, and train the angle of attack recognition network and the tilt angle recognition network using existing target ballistic data;

[0009] Step S3: Use the trained angle of attack recognition network and tilt angle recognition network to identify control variable labels, and output control variables based on the identified label results.

[0010] Preferably, the angle-of-attack recognition network adopts:

[0011] P(L α |μ k-N+1:k ) = LSTM α (μ k-N+1:k (1)

[0012] Wherein, the μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k The sequence contains feature vectors μ from time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's altitude and velocity information μ = [h, v] x ,v y ,v z ] T Where h represents the target height; v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label of the angle of attack under the input sequence conditions. α |μ k-N+1:k ).

[0013] Preferably, the angle of attack identification network includes: n fully connected layers, 2 LSTM layers, and one output layer.

[0014] Preferably, the tilt angle recognition network employs:

[0015] P(L γ |μ k-N+1:k ) = Self-att(LSTM γ (μ k-N+1:k (2)

[0016] Where, μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k It contains feature vectors for time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's velocity information μ = [v x ,v y ,v z ]T , where v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label for the tilt angle under the input sequence conditions. γ |μ k-N+1:k ).

[0017] Preferably, the tilt angle recognition network includes: n fully connected layers, 2 LSTM layers, a Self-Attention layer, and an output layer.

[0018] Preferably, step S3 involves: using the trained angle-of-attack recognition network to identify the angle of attack, and obtaining the target angle-of-attack label probability P(L) at the current moment. α |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current angle of attack;

[0019]

[0020] The trained tilt angle recognition network is used to identify the tilt angle, and the target tilt angle label probability P(L) at the current time is obtained. γ |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current bank angle.

[0021]

[0022] A control input identification system for a high-speed glider based on a long short-term memory network, provided by the present invention, includes:

[0023] Module M1: Divide the target angle of attack and heave angle ranges, and define labels for angle of attack and heave angle;

[0024] Module M2: Construct an angle of attack recognition network based on a long short-term memory network and a roll angle recognition network that introduces self-attention, and train the angle of attack recognition network and roll angle recognition network using existing target ballistic data;

[0025] Module M3: Uses the trained angle of attack recognition network and tilt angle recognition network to identify control variable labels, and outputs control variables based on the identified label results.

[0026] Preferably, the angle-of-attack recognition network adopts:

[0027] P(L α |μ k-N+1:k ) = LSTM α (μ k-N+1:k (1)

[0028] Wherein, the μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k The sequence contains feature vectors μ from time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's altitude and velocity information μ = [h, v] x ,v y ,v z ] T Where h represents the target height; v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label of the angle of attack under the input sequence conditions. α |μ k-N+1:k );

[0029] The angle of attack identification network includes: n fully connected layers, 2 LSTM layers, and one output layer.

[0030] Preferably, the tilt angle recognition network employs:

[0031] P(L γ |μ k-N+1:k ) = Self-att(LSTM γ (μ k-N+1:k (2)

[0032] Where, μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k It contains feature vectors for time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's velocity information μ = [v x ,v y ,v z ] T , where v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label for the tilt angle under the input sequence conditions. γ |μ k-N+1:k );

[0033] The tilt angle recognition network includes: n fully connected layers, 2 LSTM layers, a Self-Attention layer, and an output layer.

[0034] Preferably, module M3 employs the following method: using a trained angle-of-attack recognition network to identify the angle of attack, and obtaining the target angle-of-attack label probability P(L) at the current moment. α |μ k-N+1:kThe angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current angle of attack;

[0035]

[0036] The trained tilt angle recognition network is used to identify the tilt angle, and the target tilt angle label probability P(L) at the current time is obtained. γ |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current bank angle.

[0037]

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. By dividing the target control parameters into intervals for angle of attack and bank angle, labels for the target control parameters were defined;

[0040] 2. A target control variable recognition network was designed based on a long short-term memory network, and Self-Attention was introduced into the control variable recognition network;

[0041] 3. The target control variable identification algorithm based on long short-term memory network of the present invention can accurately and quickly identify target control variable labels and has good generalization ability, effectively improving the accuracy of target dynamics modeling. Attached Figure Description

[0042] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0043] Figure 1 This is a flowchart illustrating the target control variable identification based on a long short-term memory network involved in this invention.

[0044] Figure 2 This is a schematic diagram of an implementation scheme for the high-speed gliding target angle of attack identification technology based on long short-term memory networks involved in this invention.

[0045] Figure 3 This is a schematic diagram of an implementation scheme for the high-speed gliding target tilt angle recognition technology incorporating Self-Attention, as described in this invention. Detailed Implementation

[0046] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0047] This invention proposes a method for identifying control variables of high-speed gliding targets based on a long short-term memory (LSTM) network. The target control variables include angle of attack and bank angle. Based on deep learning theory, this invention defines labels for the angle of attack and bank angle of the target control variables and designs and trains a target control variable identification network based on LSTM. The network identifies the target control variables based on radar measurement information, providing accurate information support for target trajectory estimation.

[0048] This invention identifies the angle of attack using a Long Short-Term Memory (LSTM) network, and further incorporates a self-attention mechanism into the pan angle recognition network. The LSTM network possesses the ability to describe the long-term correlation of target maneuvers, while self-attention allows the neural network to focus on certain features of the input data, making the model's interpretation of the data more accurate and resolving the problem of information redundancy. This invention achieves accurate identification of target control variables and outputs target control variable information that meets the requirements of ballistic estimation.

[0049] Example 1

[0050] According to the present invention, a method for identifying control variables of a high-speed glider based on a long short-term memory network is provided, such as... Figures 1 to 3 As shown, it includes:

[0051] Step S1: Divide the target angle of attack and heave angle ranges, and define labels for the angle of attack and heave angle;

[0052] Step S2: Construct an angle-of-attack (AOA) recognition network based on a Long Short-Term Memory (LSTM) network and a roll angle recognition network incorporating Self-Attention. Train both networks using existing target ballistic data. Introducing Self-Attention into the roll angle recognition network, building upon the LSA network, effectively describes the long-term correlation of target gliding maneuvers. A training dataset is constructed using a high-speed gliding target, considering longitudinal balanced gliding flight, variable angle-of-attack jump gliding flight, and lateral S-shaped and C-shaped maneuvers. The training / validation / test set accuracy of each recognition network is greater than 90%, demonstrating high accuracy and good generalization ability.

[0053] Specifically, the angle-of-attack recognition network employs:

[0054] P(L α |μ k-N+1:k ) = LSTM α (μ k-N+1:k (1)

[0055] Wherein, the μ k-N+1:kRepresents the network input sequence; the network input sequence μ k-N+1:k The sequence contains feature vectors μ from time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's altitude and velocity information μ = [h, v] x ,v y ,v z ] T Where h represents the target height; v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label of the angle of attack under the input sequence conditions. α |μ k-N+1:k ).

[0056] Specifically, the angle of attack identification network includes: n fully connected layers, 2 LSTM layers, and one output layer.

[0057] Specifically, the tilt angle recognition network employs:

[0058] P(L γ |μ k-N+1:k ) = Self-att(LSTM γ (μ k-N+1:k (2)

[0059] Where, μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k It contains feature vectors for time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's velocity information μ = [v x ,v y ,v z ] T , where v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label for the tilt angle under the input sequence conditions. γ |μ k-N+1:k ).

[0060] Specifically, the tilt angle recognition network includes: n fully connected layers, 2 LSTM layers, a Self-Attention layer, and an output layer.

[0061] Step S3: Use the trained angle of attack recognition network and tilt angle recognition network to identify control variable labels, and output control variables based on the identified label results.

[0062] Specifically, step S3 involves: using the trained angle-of-attack recognition network to identify the angle of attack, and obtaining the target angle-of-attack label probability P(L) at the current moment. α |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current angle of attack;

[0063]

[0064] The trained tilt angle recognition network is used to identify the tilt angle, and the target tilt angle label probability P(L) at the current time is obtained. γ |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current bank angle.

[0065]

[0066] A control input identification system for a high-speed glider based on a long short-term memory network, provided by the present invention, includes:

[0067] Module M1: Divide the target angle of attack and heave angle ranges, and define labels for angle of attack and heave angle;

[0068] Module M2: Constructs an angle-of-attack (AOA) recognition network based on a Long Short-Term Memory (LSTM) network and a roll angle recognition network incorporating Self-Attention. These networks are trained using existing target ballistic data. The introduction of Self-Attention into the roll angle recognition network, building upon the LSA network, effectively describes the long-term correlation of target gliding maneuvers. A training dataset is constructed using a high-speed gliding target, considering longitudinal balanced gliding flight, variable angle-of-attack jump gliding flight, and lateral S-shaped and C-shaped maneuvers. The training / validation / test set accuracy of each recognition network is greater than 90%, demonstrating high accuracy and good generalization ability.

[0069] Specifically, the angle-of-attack recognition network employs:

[0070] P(L α |μ k-N+1:k ) = LSTM α (μ k-N+1:k (1)

[0071] Wherein, the μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k The sequence contains feature vectors μ from time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's altitude and velocity information μ = [h, v] x ,v y ,v z ]T Where h represents the target height; v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label of the angle of attack under the input sequence conditions. α |μ k-N+1:k ).

[0072] Specifically, the angle of attack identification network includes: n fully connected layers, 2 LSTM layers, and one output layer.

[0073] Specifically, the tilt angle recognition network employs:

[0074] P(L γ |μ k-N+1:k ) = Self-att(LSTM γ (μ k-N+1:k (2)

[0075] Where, μ k-N+1:k Represents the network input sequence; the network input sequence μ k-N+1:k It contains feature vectors for time k and the preceding N-1 times; the total length of the sequence is N; the feature vector μ at each time moment contains the target's velocity information μ = [v x ,v y ,v z ] T , where v x ,v y ,v z The network output represents the velocity components of the target in the ballistic coordinate system; the network output is the probability P(L) of each label for the tilt angle under the input sequence conditions. γ |μ k-N+1:k ).

[0076] Specifically, the tilt angle recognition network includes: n fully connected layers, 2 LSTM layers, a Self-Attention layer, and an output layer.

[0077] Module M3: Uses the trained angle of attack recognition network and tilt angle recognition network to identify control variable labels, and outputs control variables based on the identified label results.

[0078] Specifically, module M3 employs the following method: using a trained angle-of-attack recognition network to identify the angle of attack, and obtaining the target angle-of-attack label probability P(L) at the current moment. α |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current angle of attack;

[0079]

[0080] The trained tilt angle recognition network is used to identify the tilt angle, and the target tilt angle label probability P(L) at the current time is obtained. γ |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current bank angle.

[0081]

[0082] Example 2

[0083] Example 2 is a preferred example of Example 1.

[0084] The high-speed gliding target control variable identification technology based on long short-term memory network provided by the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0085] Figure 1 This is a schematic diagram of an implementation scheme for high-speed gliding target control variable recognition based on a long short-term memory network, as involved in this invention. It includes target control variable label definition, control variable recognition network design and training, and target control variable recognition. Target control variable label definition involves dividing the target control variable range and defining the control variable labels and corresponding nominal values ​​used in subsequent steps. Control variable recognition network design and training involves constructing an angle-of-attack recognition network based on a long short-term memory network and a roll angle recognition network incorporating self-attention, and training the model using existing target trajectory data. Target control variable recognition involves using the trained control variable recognition network to identify the control variable labels and outputting the control variable based on the identified label results. The specific implementation scheme is as follows.

[0086] Part 1: Angle of attack identification.

[0087] Step 1: Define Angle of Attack Labels

[0088] Angle of attack labels are defined by dividing the target angle of attack α into intervals, as shown in Table 1.

[0089] Table 1 Control Variable Angle of Attack Labels

[0090]

[0091] Step 2: Angle of Attack Recognition Network Design

[0092] Network input sequence μ k-N+1:k The sequence contains feature vectors μ from the current time (time k) and the previous N-1 times. The total length of the sequence is N. The feature vector μ at each time step contains the target's altitude and velocity information, i.e., μ = [h, v]. x ,v y ,v z ] TWhere h is the target height, v x ,v y ,v z Let L be the velocity component of the target in the ballistic coordinate system. The network output is the probability P(L) of each label for the angle of attack given the input sequence. α |μ k-N+1:k That is, the angle of attack identification network is represented as...

[0093] P(L α |μ k-N+1:k ) = LSTM α (μ k-N+1:k (1)

[0094] The angle-of-attack recognition network is based on a long short-term memory (LSTM) network structure. The network consists of n fully connected (Dense) layers, two LSTM layers, and one output layer (Softmax). Figure 2 Connect as shown.

[0095] After optimization using neural network hyperparameters, the parameter design of the angle of attack recognition network is as follows:

[0096] Table 2 Network Parameter Design

[0097]

[0098] Step 3: Training the Angle of Attack Recognition Network

[0099] The parameters of the target control quantity recognition network shown in equation (1) are trained. The network training settings are as follows:

[0100] Loss function: Cross-entropy loss function.

[0101] Optimizer: Adam, learning rate 0.0001.

[0102] Number of iterations (epochs): 200

[0103] After sufficient training, the network loss function converged, and the training results are shown in Table 3. It can be seen that the recognition accuracy exceeds 90% on the training, validation, and test sets, and the accuracy is approximately the same. This indicates that the designed network did not overfit, can accurately identify the target angle of attack, and has good generalization ability.

[0104] Table 3 Network training results

[0105]

[0106] Step 4: Angle of attack identification

[0107] The angle of attack identification network shown in the formula is used to identify the angle of attack, and the target angle of attack label probability P(L) at the current time is obtained. α|μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current angle of attack.

[0108]

[0109] Part Two: Tilt Angle Recognition.

[0110] Step 5: Define the tilt angle label.

[0111] The tilt angle label is defined by dividing the target tilt angle γ into intervals, as shown in Table 4.

[0112] Table 4 Tilt Angle Labels

[0113]

[0114]

[0115] Step 6: Tilt Angle Recognition Network Design

[0116] Network input sequence μ k-N+1:k The sequence contains feature vectors μ from the current time (time k) and the previous N-1 times, with a total length of N. Each feature vector μ contains the target's velocity information, i.e., μ = [v...]. x ,v y ,v z ] T , where v x ,v y ,v z Let L be the velocity components of the target in the ballistic coordinate system. The network output is the probability P(L) of each label for the pan angle given the input sequence. γ |μ k-N+1:k That is, the tilt angle recognition network is represented as...

[0117] P(L γ |μ k-N+1:k ) = Self-att(LSTM γ (μ k-N+1:k (3)

[0118] The tilt angle recognition network introduces a Self-Attention mechanism based on the Long Short-Term Memory (LSTM) network structure to address the problem of information redundancy. The network consists of n fully connected (Dense) layers, two LSTM layers, a Self-Attention layer, and an output layer (Softmax). Figure 3 Connect as shown.

[0119] After optimization using neural network hyperparameters, the parameter design of the tilt angle recognition network is as follows:

[0120] Table 5 Network Parameter Design

[0121]

[0122] Step 7: Training the Tilt Angle Recognition Network

[0123] The parameters of the target control variable recognition network shown in equation (2) are trained. The network training settings are as follows:

[0124] Loss function: Cross-entropy loss function.

[0125] Optimizer: RMSProp, learning rate 0.0001.

[0126] Number of iterations (epochs): 300.

[0127] After sufficient training, the loss functions of each network converged, and the training results are shown in Table 6. It can be seen that the recognition network achieves over 90% recognition accuracy on all datasets, and the accuracy on the training / validation / test sets is approximately the same. This indicates that the designed network exhibits no overfitting, has good generalization ability, and can accurately identify the target angle of attack.

[0128] Table 6 Network Training Results

[0129]

[0130]

[0131] Step 8: Tilt Angle Recognition

[0132] Using the angle-of-attack identification network shown in equation (3), the tilt angle is identified, and the target tilt angle label probability P(L) at the current moment is obtained. γ |μ k-N+1:k The angle of attack corresponding to the target control variable label with the highest probability is taken as the target's current roll angle.

[0133]

[0134] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0135] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for control variable identification of a high-speed glider based on a long short-term memory network, characterized in that, The application relates to a method for identifying control variables of a target trajectory, and belongs to the field of trajectory control. The method comprises the following steps: Step S1: dividing target attack angle and roll angle intervals, and defining attack angle and roll angle labels; Step S2: constructing an attack angle identification network based on a long short-term memory network and a roll angle identification network with Self-Attention, and training the attack angle identification network and the roll angle identification network by using existing target trajectory data; wherein, the Self-Attention is introduced into the roll angle identification network structure on the basis of the long short-term memory network; Step S3: identifying control variable labels by using the trained attack angle identification network and the roll angle identification network, and outputting control variables according to the identified label results. P(L α |μ k-N+1:k )=LSTM α (μ k-N+1:k ) (1) wherein the μ k-N+1:k represents the network input sequence; the network input sequence μ k-N+1:k contains the feature vector μ at time k and the previous N-1 time; the total length of the sequence is N; the feature vector μ at each time contains the height and speed information of the target μ = [h, v x , v y , v z ] T , wherein h represents the target height; v x , v y , v z represents the velocity component of the target in the ballistic coordinate system; the output of the network is the angle of attack label probability P(L α | μ k-N+1:k ) under the condition of the input sequence; The attack angle identification network adopts: P(L γ |μ k-N+1:k )=Self-att(LSTM γ (μ k-N+1:k )) (2) wherein μ k-N+1:k represents the network input sequence; the network input sequence μ k-N+1:k contains the feature vector of time k and the previous N-1 time; the total length of the sequence is N; the feature vector μ of each time contains the speed information of the target μ = [v x , v y , v z ] T , wherein v x , v y , v z represents the velocity component of the target in the ballistic coordinate system; the output of the network is the probability of each label of the roll angle under the condition of the input sequence P (L γ | μ k-N+1:k ); The step S3 adopts: using the trained angle of attack recognition network to recognize the angle of attack, to obtain a target angle of attack label probability P(L α |μ k-N+1:k ) of the current moment, and taking the angle of attack corresponding to the target control variable label with the largest probability as the target current angle of attack; The trained roll angle recognition network is used to recognize the roll angle, and a target roll angle label probability P(L γ |μ k-N+1:k ) of a current moment is obtained; and an angle of attack corresponding to a target control amount label with the maximum probability is taken as a target current roll angle.

2. The long short-term memory network-based control variable identification method for a high-speed glider according to claim 1, characterized in that, The roll angle identification network adopts:

3. The long short-term memory network-based control variable identification method for a high-speed glider according to claim 1, characterized in that, The attack angle identification network comprises n full connection layers, 2 LSTM layers and an output layer.

4. A high-speed glider control quantity identification system based on a long short-term memory network, characterized in that, The roll angle identification network comprises n full connection layers, 2 LSTM layers, a Self-Attenttion layer and an output layer. The application further relates to a method for identifying control variables of a target trajectory. The method comprises the following steps: Module M1: dividing target attack angle and roll angle intervals, and defining attack angle and roll angle labels; Module M2: constructing an attack angle identification network based on a long short-term memory network and a roll angle identification network with Self-Attention, and training the attack angle identification network and the roll angle identification network by using existing target trajectory data; wherein, the Self-Attention is introduced into the roll angle identification network structure on the basis of the long short-term memory network; P(L α |μ k-N+1:k )=LSTM α (μ k-N+1:k ) (1) wherein the μ k-N+1:k represents the network input sequence; the network input sequence μ k-N+1:k contains the feature vector μ of the time k and the previous N-1 times; the total length of the sequence is N; the feature vector μ of each time contains the height and speed information of the target μ = [h, v x , v y , v z ] T , wherein h represents the target height; v x , v y , v z represents the speed component of the target in the ballistic coordinate system; the output of the network is the angle of attack label probability P (L α | μ k-N+1:k ) under the condition of the input sequence. Module M3: identifying control variable labels by using the trained attack angle identification network and the roll angle identification network, and outputting control variables according to the identified label results. P(L γ |μ k-N+1:k )=Self-att(LSTM γ (μ k-N+1:k )) (2) wherein μ k-N+1:k represents the network input sequence; the network input sequence μ k-N+1:k contains the feature vector at time k and the previous N-1 time; the total length of the sequence is N; the feature vector μ at each time contains the velocity information of the target μ = [v x , v y , v z ] T wherein v x , v y , v z represents the velocity component of the target in the ballistic coordinate system; the output of the network is the label probability P(L γ | μ k-N+1:k ) of the roll angle under the condition of the input sequence. The module M3 adopts: using the trained angle of attack recognition network to recognize the angle of attack, obtaining the target angle of attack label probability P(L α |μ k-N+1:k ) of the current moment, and taking the angle of attack corresponding to the target control variable label with the largest probability as the target current angle of attack; The trained roll angle recognition network is used to recognize the roll angle, to obtain a target roll angle label probability P(L γ |μ k-N+1:k ) of a current moment; and an angle of attack corresponding to a target control amount label with the maximum probability is taken as a target current roll angle.

5. The long short-term memory network-based control variable identification system for high-speed gliding vehicles according to claim 4, characterized in that, The attack angle identification network adopts:

6. The long short-term memory network-based control variable identification system for high-speed gliding vehicles of claim 4, wherein, The roll angle identification network adopts: The attack angle identification network comprises n full connection layers, 2 LSTM layers and an output layer. The roll angle identification network comprises n full connection layers, 2 LSTM layers, a Self-Attenttion layer and an output layer.

Images