A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks

Abstract

This invention relates to the field of missile maneuvering, and provides a method for predicting the maneuvering intentions of hypersonic vehicles (HGVs) based on a remote sensing constellation. A model is established for the maneuvering intention sequence of HGVs; based on the model, the types of maneuvers of the target in the longitudinal and lateral directions are described, and the maneuver amplitude and frequency are quantified; based on the model, a parallel stacked long short-term memory (PSLSTM) neural network is designed; based on the quantified maneuver amplitude and frequency and the parallel stacked PSLSTM, a process for identifying and predicting the maneuvering intentions of HGVs is designed; based on the identification and prediction process of different maneuvering parameters, a sensor switching strategy is designed to improve the adaptability of remote sensing satellites to the uncertainty of target intentions and increase the overall tracking benefit. This invention addresses the uncertainty of target maneuvering intentions during continuous tracking and observation of HGVs by a constellation tracking observation system, improving the adaptability of remote sensing satellites to this problem.

Description

Technical Field

[0001] This invention specifically relates to the field of missile maneuvering, and more specifically to a method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks. Background Technology

[0002] Remote sensing satellite tracking and observation systems rely on the calculation results of hypersonic missile (HGV) trajectory prediction algorithms for mission planning. However, the accuracy of HGV trajectory prediction is still affected by the uncertainty of target maneuvering intentions. The gliding process of HGVs exhibits non-ballistic characteristics, and the maneuvering intention at each moment is uncertain. A sequence of maneuvering intentions from multiple moments is called a maneuvering intention sequence, which also possesses significant uncertainty. Although research on HGV trajectory prediction has matured, the uncertainty of maneuvering sequences still significantly impacts trajectory prediction and constellation tracking and observation mission planning. Therefore, research on the uncertainty of HGV maneuvering intention sequences is gradually unfolding. Traditional methods use Bayesian estimation theory to infer the target intention sequence. Among deep learning methods, using stacked LSTM networks to predict target maneuver types is currently a maneuvering intention prediction method with strong adaptability and high prediction accuracy. Furthermore, LSTM networks can also approximate the dynamic relationship between aerodynamic parameters and displacements under varying Mach numbers and average angles of attack through aerodynamic analysis of HGVs. In summary, the current HGVs maneuver intent model is mainly based on maneuver intent type modeling. The consideration of longitudinal and lateral maneuver amplitude mainly relies on manually dividing the trajectory envelope into several categories, which has a significant impact on subsequent accuracy. Summary of the Invention

[0003] This invention provides a method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks. It addresses the uncertainty of target maneuvering intent during continuous tracking and observation of HGVs by satellite constellation tracking systems, thereby improving the adaptability of remote sensing satellites to the uncertainty of target intent.

[0004] This invention is achieved through the following technical solution:

[0005] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, the prediction method comprising the following steps:

[0006] Step 1: Model the maneuver intent sequence of HGVs;

[0007] Step 2: Based on the model from Step 1, describe the target's maneuver types in the longitudinal and lateral directions, and quantify the maneuver amplitude and frequency;

[0008] Step 3: Based on the model in Step 1, design a parallel stacked long short-term memory neural network PSLSTM;

[0009] Step 4: Based on the quantization of maneuver amplitude and frequency in Step 2 and the parallel stacked long short-term memory neural network PSLSTM in Step 3, design the maneuver intent recognition and prediction process for HGVs.

[0010] Step 5: Based on the identification and prediction process of different maneuver parameters in Step 4, design a sensor switching strategy to improve the adaptability of remote sensing satellites to the uncertainty of target intent and increase the overall tracking benefits.

[0011] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, wherein step 1 specifically involves establishing that the target's maneuvering intent at the i-th moment can be described by an intent combination consisting of four elements, in the following specific form:

[0012]

[0013] In the formula, the superscript represents the sequence number of the maneuver intention, and the subscript represents the sequence number of the intention parameter;

[0014] Based on formula (1), a vertical intent model and a horizontal intent model are established.

[0015] A method for predicting the maneuvering intent of a hypersonic missile based on deep neural networks, wherein step 2, establishing a longitudinal intent model, specifically classifies longitudinal maneuvering intent types into two categories: ski-jump trajectory (QEG) and quasi-balanced trajectory (SG). The first maneuvering intent parameter describes the target's longitudinal maneuvering intent type, in the following form:

[0016]

[0017] Because the ski-jump track QEG and the quasi-balanced track SG are very different, therefore The calculation methods for intent parameters are different:

[0018]

[0019] Where γ is the trajectory angle, L is the lift, m is the mass, v is the velocity, and σ is the velocity tilt angle.

[0020] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, wherein step 2, establishing a lateral intent model, specifically involves the target velocity tilt angle being an intuitive description of the target's maneuvering direction. Furthermore, as shown in equation (1), the sign of the velocity tilt angle σ is opposite to the maneuvering direction of the trajectory. Therefore, the lateral maneuvering intent parameters... Recorded as,

[0021]

[0022] Lateral maneuvering is achieved by changing the target's attitude and thus its velocity tilt angle. The lateral component of the lift force acting on the target is a direct reflection of the target's lateral maneuvering intention. Therefore, the lateral maneuvering intention parameter... Recorded as,

[0023]

[0024] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, wherein step 3, designing a parallel stacked long short-term memory neural network (PSLSTM), specifically includes the following steps:

[0025] Step 3.1: Sequence data is input into the network through the Sequence layer;

[0026] Step 3.2: Connect stacked LSTM networks, where the number of parallel networks Pn is a variable value;

[0027] Step 3.3: Each stacked layer consists of one fully connected layer and three LSTM layers connected in series, and the three stacked layers eventually merge into one fully connected layer;

[0028] Step 3.4: The output layer, activated by Softmax, finally reaches the classification layer.

[0029] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, wherein step 4, designing the maneuvering intent recognition and prediction process for HGVs, specifically includes the following steps:

[0030] Step 4.1: Preprocess the observation information according to the requirements of the maneuver intent recognition network.

[0031] Step 4.2: Identify the intent labels and intent parameters of HGVs based on the designed PSLSTM neural network.

[0032] Step 4.3: Integrate the preprocessed observation information and maneuver intent recognition information into the joint maneuver intent prediction network to obtain the final maneuver intent prediction result.

[0033] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, wherein step 5 specifically involves calculating the shortest time required for antenna backscan.

[0034]

[0035] In the formula, α tur ω is the required rotation angle for antenna retracement, and ω is the angular velocity of antenna rotation.

[0036] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, during the target maneuver interval T. inThe time interval between the two maneuvers of the target.

[0037]

[0038] In the formula, Let k be the end time of the k-th maneuver. Let k be the start time of the (k+1)th maneuver, where k is the number of maneuvers.

[0039] During the period T when the tracking load accuracy decreases ac The current tracking load observation angle is less than the street landing angle σ under the current accuracy requirement. cur ≤σ min And the observation distance is greater than the critical value ρ cur >ρ max The earliest arrival time T within the time period be Described as

[0040] T be =T f_start +T fb

[0041] In the formula, T f_start This is the moment when the antenna begins to retrace.

[0042] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, considering the target maneuvering interval T. in During the period T when the tracking load accuracy decreases ac and earliest arrival time T be Calculate the optimal switching time period T sw

[0043]

[0044] In the formula, The optimal start and end times for the switching time period. For the start and end times of the target maneuver gap period, To track the start time of the period when the load accuracy decreases.

[0045] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, according to the optimal switching time period T sw The relay remote sensing satellite completes the tracking switch process before the tracking accuracy of the tracking satellite decreases, reducing the probability of tracking failure due to the maneuvering intentions of HGVs and improving the stability of the remote sensing satellite's continuous tracking of HGVs.

[0046] The beneficial effects of this invention are:

[0047] This invention addresses the uncertainty problem of HGVs maneuver sequences by describing the maneuver types of HGVs in the longitudinal and lateral directions, and quantifying the maneuver amplitude and frequency.

[0048] This invention fully utilizes the temporal classification and regression capabilities of Long Short-Term Memory (LSTM) networks to address the issue of limited types of HGV observation information, thereby enabling the recognition and prediction of maneuver intentions.

[0049] This invention calculates and analyzes the optimal relay window for remote sensing satellites, enabling them to have a higher continuous tracking capability when tracking HGVs. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the remote sensing satellite's intention to predict HGVs according to the present invention.

[0051] Figure 2 This is a schematic diagram of the motor intent recognition and prediction process of the present invention.

[0052] Figure 3 This is a schematic diagram of the sensor switching strategy of the present invention.

[0053] Figure 4 This is a diagram of the PSLSTM neural network structure of the present invention.

[0054] Figure 5 This is a schematic diagram illustrating the area of ​​uncertainty regarding the intent of this invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Tracking and observing hypersonic missiles (HGVs) via remote sensing constellations is a critical issue in space-based early warning systems. Unlike inertial targets, HGVs fly at lower altitudes and at higher speeds, resulting in narrower observation angles and observation windows for remote sensing satellites. Furthermore, multiple satellite switches are required to complete continuous tracking of HGVs. The high maneuverability of HGVs means that the uncertainty of their maneuvering intentions can significantly impact the consistency of satellite tracking.

[0057] Traditional remote sensing satellite tracking algorithms typically switch tracking satellites between the completion of one subtask and the start of the next during mission planning. However, this switching method is less adaptable to targets like HGVs, which have highly uncertain maneuvering intentions and a wide maneuvering range. Towards the end of a subtask, sensors often approach their limit in terms of observation angle or distance, leading to a decrease in tracking capability. If the target makes a large maneuver at this time, the tracking satellite may lose track of the target prematurely, causing the relay satellite to fail to track.

[0058] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, the prediction method comprising the following steps:

[0059] Step 1: Model the maneuver intent sequence of HGVs;

[0060] Step 2: Based on the model from Step 1, describe the target's maneuver types in the longitudinal and lateral directions, and quantify the maneuver amplitude and frequency;

[0061] Step 3: Based on the model in Step 1, design a parallel stacked long short-term memory neural network PSLSTM;

[0062] Step 4: Based on the quantization of maneuver amplitude and frequency in Step 2 and the parallel stacked long short-term memory neural network PSLSTM in Step 3, design the maneuver intent recognition and prediction process of HGVs, and use the hierarchical network to complete the recognition and prediction process of different maneuver parameters of HGVs in sequence.

[0063] Step 5: Based on the identification and prediction process of different maneuver parameters in Step 4, design a sensor switching strategy to improve the adaptability of remote sensing satellites to the uncertainty of target intent and increase the overall tracking benefits.

[0064] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, wherein step 1 specifically involves establishing that the target's maneuvering intent at the i-th moment can be described by an intent combination consisting of four elements, in the following specific form:

[0065]

[0066] In the formula, the superscript represents the sequence number of the maneuver intention, and the subscript represents the sequence number of the intention parameter;

[0067] Based on formula (1), a vertical intent model and a horizontal intent model are established.

[0068] A method for predicting the maneuvering intent of a hypersonic missile based on deep neural networks, wherein step 2, establishing a longitudinal intent model, specifically classifies longitudinal maneuvering intent types into two categories: ski-jump trajectory (QEG) and quasi-balanced trajectory (SG). The first maneuvering intent parameter describes the target's longitudinal maneuvering intent type, in the following form:

[0069]

[0070] Because the ski-jump track QEG and the quasi-balanced track SG are very different, therefore The calculation methods for intent parameters are different:

[0071]

[0072] Where γ is the trajectory angle, L is the lift, m is the mass, v is the velocity, and σ is the velocity tilt angle.

[0073] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, wherein step 2, establishing a lateral intent model, specifically involves the target velocity tilt angle being an intuitive description of the target's maneuvering direction. Furthermore, as shown in equation (1), the sign of the velocity tilt angle σ is opposite to the maneuvering direction of the trajectory. Therefore, the lateral maneuvering intent parameters... Recorded as,

[0074]

[0075] Lateral maneuvering is achieved by changing the target's attitude and thus its velocity tilt angle. The lateral component of the lift force acting on the target is a direct reflection of the target's lateral maneuvering intention. Therefore, the lateral maneuvering intention parameter... Recorded as,

[0076]

[0077] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, wherein step 3, designing a parallel stacked long short-term memory neural network (PSLSTM), specifically includes the following steps:

[0078] Step 3.1: Sequence data is input into the network through the Sequence layer;

[0079] Step 3.2: Connect stacked LSTM networks. The number of parallel networks Pn is a variable value. In the figure, Pn = 3.

[0080] Step 3.3: Each stacked layer consists of one fully connected layer and three LSTM layers connected in series, and the three stacked layers eventually merge into one fully connected layer;

[0081] Step 3.4: The output layer, activated by Softmax, finally reaches the classification layer.

[0082] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, wherein step 4, designing the maneuvering intent recognition and prediction process for HGVs, specifically includes the following steps:

[0083] Step 4.1: Preprocess the observation information according to the requirements of the maneuver intent recognition network.

[0084] Step 4.2: Identify the intent labels and intent parameters of HGVs based on the designed PSLSTM neural network.

[0085] Step 4.3: Integrate the preprocessed observation information and maneuver intent recognition information into the joint maneuver intent prediction network to obtain the final maneuver intent prediction result.

[0086] A method for predicting the maneuvering intent of a hypersonic missile based on a deep neural network, wherein step 5 specifically involves, as follows: Figure 5 The figure shows the uncertainty region (UR) of maneuver intent. The solid line in the figure is the predicted trajectory of HGVs maneuver, and the area between the dotted line and the dashed line is the uncertainty region after considering the HGVs maneuver intent.

[0087] Based on the target's maneuvering intent sequence, the target trajectory is divided into maneuvering time periods and maneuvering intervals. Targets maneuvering continuously are a special case; in this case, the UR width can be compared, and the shorter interval is considered the maneuvering interval. The tracking accuracy of remote sensing satellites performing tracking operations is constrained by the payload's own performance and affected by factors such as observation angle and observation distance. Tracking accuracy degrades towards the end of the visible time window. When considering the timing of new mission bundles for bidding satellites, step 5 specifically involves calculating the shortest time required for antenna retracement.

[0088]

[0089] In the formula, α tur ω is the required rotation angle for antenna retracement, and ω is the angular velocity of antenna rotation.

[0090] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, during the target maneuver interval T. in The time interval between the two maneuvers of the target.

[0091]

[0092] In the formula, Let k be the end time of the k-th maneuver. Let k be the start time of the (k+1)th maneuver, where k is the number of maneuvers.

[0093] During the period T when the tracking load accuracy decreases ac The current tracking load observation angle is less than the street landing angle σ under the current accuracy requirement. cur ≤σ min And the observation distance is greater than the critical value ρ cur >ρ max The earliest arrival time T within the time periodbe Described as

[0094] T be =T f_start +T fb

[0095] In the formula, T f_start This is the moment when the antenna begins to retrace.

[0096] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, considering the target maneuvering interval T. in During the period T when the tracking load accuracy decreases ac and earliest arrival time T be Calculate the optimal switching time period T sw

[0097]

[0098] In the formula, The optimal start and end times for the switching time period. For the start and end times of the target maneuver gap period, To track the start time of the period when the load accuracy decreases.

[0099] A method for predicting the maneuvering intent of hypersonic missiles based on deep neural networks, according to the optimal switching time period T sw The relay remote sensing satellite completes the tracking switch process before the tracking accuracy of the tracking satellite decreases, reducing the probability of tracking failure due to the maneuvering intentions of HGVs and improving the stability of the remote sensing satellite's continuous tracking of HGVs.

Claims

1. A method for predicting the maneuvering intention of a hypersonic missile based on a deep neural network, characterized in that, The prediction method comprises the following steps: Step 1: a model is established for a maneuvering intention sequence of a hypersonic guided vehicle (HGV); Step 2: based on the model of step 1, a longitudinal overtake ground track (QEG) and a quasi-equilibrium ground track (SG) of the target and a target speed inclination angle in the lateral direction are described, and a maneuvering amplitude and frequency are quantified; Step 3: based on the model of step 1, a parallel stack long short-term memory neural network (PSLSTM) is designed; Step 4: based on the quantified maneuvering amplitude and frequency of step 2 and the parallel stack long short-term memory neural network (PSLSTM) of step 3, a maneuvering intention recognition and prediction process of the HGV for the target is designed; Step 5: based on the different maneuvering parameter target recognition and prediction process of step 4, a sensor switching strategy is designed to improve the adaptability of a remote sensing satellite to a target intention uncertainty problem and improve the total tracking benefit; The step 5 is specifically to calculate the shortest time required for antenna back-sweeping: wherein α tur is the required rotation angle for antenna flyback, and ω is the angular velocity of antenna rotation. Target maneuver gap period T in is the interval period for the target two-time maneuver process, In the formula, is the end time of the kth maneuver, is the start time of the k+1th maneuver, k is the number of maneuvers. Tracking load accuracy degradation period T ac For the current tracking load observation angle is less than or equal to the accuracy degradation threshold angle, that is σ cur ≤σ min And the observation distance is greater than the threshold value ρ cur >ρ max The earliest arrival time T be Description as T be = T f_start + T fb In the formula, T f_start is the time at which the antenna begins to retrace; considering the target maneuver gap period T in , tracking the payload precision drop period T ac and the earliest arrival time T be , calculating the optimal switching time period T sw In the formula, is a start time and an end time of the best switching time period, is a start time and an end time of the target maneuver gap time period, is a start time of the tracking load precision reduction time period.

2. The method of claim 1, wherein, The step 1 is specifically to establish a target intention combination consisting of four elements at an i-th maneuvering intention, and the specific form is as follows: In the formula, the upper index represents a maneuvering intention sequence number, and the lower index represents an intention parameter sequence number; Based on the formula (1), a longitudinal intention model and a lateral intention model are established.

3. The method of claim 2, wherein, The step 2 of establishing the longitudinal intention model is specifically that the longitudinal maneuvering intention type is divided into two types of overtake ground track (QEG) and quasi-equilibrium ground track (SG), and a first maneuvering intention parameter is used to describe the target longitudinal maneuvering intention type, and the form is as follows: Since the difference between the overflights QEG and the quasi-equilibrium path SG is very clear, therefore The calculation of the intention parameters is different: Wherein, γ is a trajectory inclination angle, L is a lift, m is a mass, v is a speed, and σ is a speed inclination angle.

4. The method of claim 3, wherein, The step 2 establishes the lateral intention model, specifically, the target speed inclination angle is a direct description of the target maneuvering direction, and according to formula (1), the sign of the speed inclination angle σ is opposite to the maneuvering direction of the track, therefore, the lateral maneuvering intention parameter is denoted as, The action of lateral maneuver is realized by changing the target attitude and then the velocity tilt angle. The lateral component of the lift force on the target is the direct reflection of the target's lateral maneuver intention, so the lateral maneuver intention parameter is denoted as 5. The method of claim 2, wherein, The step 3 of designing the parallel stack long short-term memory neural network (PSLSTM) specifically comprises the following steps: Step 3.1: sequence data is input into the network through a Sequence layer; Step 3.2: a stack LSTM network is connected, and the network parallel number Pn is a variable value; Step 3.3: each stack layer is composed of a full connection layer and three LSTM layers in series, and the three stack layers finally converge into a full connection layer; Step 3.4: an output layer with an activation function of Softmax is finally reached to a classification layer.

6. The method of claim 2, wherein, The step 4 of designing the maneuvering intention recognition and prediction process of the HGV specifically comprises the following steps: Step 4.1: according to the requirements of the maneuvering intention recognition network, the observation information is preprocessed Step 4.2: based on the designed PSLSTM neural network, an intention label and an intention parameter of the HGV are recognized Step 4.3: the preprocessed observation information and the maneuvering intention recognition information are integrated and input into a maneuvering intention joint prediction network to obtain a final maneuvering intention prediction result.

7. The method of claim 1, wherein, According to the optimal switching time period T sw , the relay remote sensing satellite completes the tracking switching process before the tracking accuracy of the tracking satellite decreases, reduces the tracking failure probability caused by the maneuvering intention of the hypersonic missile HGVs, and improves the stability of the remote sensing satellite in continuous tracking of the hypersonic missile HGVs.

Images