High maneuvering target interception point prediction method based on long and short term memory network

Abstract

The invention provides a high-maneuverability target interception point prediction method based on a long-short-term memory network, aiming at a high-speed target which flies across an airspace and a speed domain, is uncontrollable in speed and high in maneuverability, and can improve the interception point prediction precision. An interception missile motion state, a target motion state, the height of the target relative to the ground, the radar target tracking duration and the relative motion state between the interception missile and the target are acquired and cached in the feature sequence; when the length of the feature sequence reaches the length of an input sequence of the prediction network, inputting the feature sequence into the prediction network, and outputting the predicted hit point position and the time when the target arrives at the hit point by the prediction network; the prediction network comprises a long-short term memory (LSTM) network, a multi-head self-attention mechanism network and a full-connection neural network.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft guidance and control, and particularly relates to a method for predicting an interception point of a highly maneuverable target based on a long short-term memory network. Background Art

[0002] For an incoming target, if the hit point can be predicted based on the target information and interceptor information detected by a radar, the interceptor can be first made to fly towards the predicted hit point. When the target enters the field of view of the interceptor's seeker, the interceptor then uses a proportional navigation law to attack the target. At this time, the required overload of the interceptor can be greatly reduced, which helps to improve the interception success rate. Since the relative motion relationship between the interceptor and the target during the attack has time-varying and strong non-linearity, it is very difficult to predict the hit point by an analytical method.

[0003] Currently, most of the technologies adopt a numerical integration method to simulate the trajectory of the interceptor attacking the target. For example, in the prior art 1 (see Hu Zhiheng, Zhou Di, Zou Xinguang. Identification of Guidance Law and Trajectory Prediction of Interceptor Missiles [J]. Systems Engineering and Electronics, 2018, 40(03): 609-614.), first, an extended Kalman filter (EKF) is used to identify the guidance parameters of the missile. Under the condition that the initial parameters of the missile motion are known, the position of the interception point is determined through trajectory simulation based on the identified guidance law. However, the hit point prediction method based on numerical integration has the characteristics of large computational amount and long time consumption, and cannot be applied to the on-line application situation.

[0004] With the development of artificial intelligence technology, a method for obtaining the predicted interception point based on a neural network has been developed later. For example, in the prior art 2 (see Yang Zicheng, Xian Yong, Li Shaopeng, etc. Learning-based Midcourse Missile Defense Interception Time and Interception Point Prediction Method [J]. Journal of Beijing University of Aeronautics and Astronautics, 2021, 47(11): 2360-2368.), taking the shutdown parameters and shutdown time of the active section of the interceptor as input quantities, an interception time and interception point prediction model is established, and a supervised learning algorithm is constructed based on a feedforward neural network to realize the prediction of the interception point. However, currently, most of the methods based on neural networks have good hit point prediction accuracy for targets with low maneuverability. When the target has strong maneuverability, the accuracy of the current methods will drop significantly.

[0005] Currently, the technologies are mainly applicable to predicting the interception point of a target with simple dynamic characteristics and low maneuverability. However, when intercepting a highly maneuverable and high-speed target with complex dynamic characteristics and flying across airspaces and speed ranges, the non-linearity of the relative motion relationship between the interceptor and the target is enhanced, and the difficulty of predicting the hit point increases. Currently, no related technologies with good prediction accuracy have been seen. Summary of the Invention

[0006] In view of this, the present invention provides a method for predicting the interception point of a highly maneuverable target based on a long short-term memory network, which can improve the accuracy of interception point prediction for high-speed targets that fly across airspaces and speed domains, have uncontrollable speed magnitudes, and strong maneuverability.

[0007] To solve the above technical problems, the present invention is implemented as follows.

[0008] A method for predicting the interception point of a highly maneuverable target based on a long short-term memory network includes:

[0009] Step 1: The radar regularly measures the motion states of the target and the interceptor missile to obtain the element x of the feature sequence i , and caches it into the feature sequence X;

[0010] The element x i includes: the motion state S of the interceptor missile D , the motion state S of the target T , the height H of the target relative to the ground T , the radar tracking duration t of the target l , and the relative motion state S between the interceptor missile and the target R ;

[0011] Step 2: When the length of the feature sequence X reaches the input sequence length n of the prediction network, input it into the prediction network, and the prediction network outputs the predicted hit point position and the time t when the target reaches the hit point go ; and continuously update the prediction output of the feature sequence X;

[0012] The prediction network includes a target motion state feature extraction network and a relationship fitting network between features and hit points; the target motion state feature extraction network consists of a long short-term memory network LSTM and a multi-head self-attention mechanism network; the relationship fitting network between features and hit points uses a fully connected neural network.

[0013] Preferably, in the step 1,

[0014] The motion state S of the interceptor missile D = [x D , y D , z D , V D , θ D , ψ vD , where (x D , y D , z D ) is the position of the interceptor missile, V D is the speed of the interceptor missile, and θ D and ψ vD are the ballistic inclination angle and the ballistic deflection angle of the interceptor missile respectively;

[0015] Target motion state S T = [x T , y T , z T , V T , θ T , ψ vT , where (x T , y T , z T ) is the position of the target, V T is the velocity of the target, θ T and ψ vT are the ballistic inclination angle and the ballistic deflection angle of the target respectively;

[0016] The radar tracking time duration t of the target l = t - t l0 , where t is the current time, and t l0 is the time when the radar first detects the target; t l is the timestamp of each element in the input feature sequence of the prediction network;

[0017] The relative motion state between the interceptor and the target where r D is the distance between the interceptor and the target; q yD and q zD are the components of the interceptor's field of view angle in the elevation direction and the azimuth direction respectively; is the derivative of r D , q yD , q zD , and is calculated according to the following system of equations of the relative motion between the interceptor and the target;

[0018]

[0019] where, v M is the velocity of the target, θ M is the ballistic inclination angle of the target in the launch coordinate system, and ψ vM is the ballistic deflection angle of the target in the launch coordinate system.

[0020] Preferably, in the prediction network, the tensors output by the multi-head self-attention mechanism network are concatenated end to end for each row, flattened into a vector, and then input into the fully connected neural network; the fully connected neural network includes an input layer, multiple hidden layers, and an output layer connected in sequence; an activation function layer is provided between the input layer and the hidden layers, and between the hidden layers.

[0021] Preferably, a dropout layer is provided between the input layer and the hidden layers, and between the hidden layers.

[0022] Preferably, the prediction network further includes a hit point position output layer and a target arrival time output layer connected after the output layer, which are used to scale the predicted hit point position and the time for the target to reach the hit point output by the prediction network to appropriate orders of magnitude respectively.

[0023] Preferably, the total loss function L used for training the prediction network is L = ω p L p + ω t L t ; where: L p and L t are the loss function terms for the hit point position prediction and the target arrival hit point time prediction respectively, and ω p and ω t are the corresponding weighted weights.

[0024] Preferably, the prediction network is pre-trained, and the training samples are constructed as follows:

[0025] Step 1.1: For the selected hit point M, according to the target ballistic prediction, determine the height H of the target relative to the ground, the radar tracking target duration t T and the target motion state S l at each sampling moment during the period from the radar's first detection of the target to the interceptor hitting the target; T ;

[0026] Step 1.2: Determine the motion state S D of the interceptor at each sampling moment during the period from the interceptor launch to the interceptor hitting the target through simulation;

[0027] Step 1.3: Correlate the target motion state S T and the interceptor motion state S D in terms of time, and calculate the relative motion state S R of the interceptor-target;

[0028] Step 1.4: Determine the time t go for the interceptor to hit the target at each sampling moment;

[0029] Step 1.5: The S D , S T , H T , t l , S R corresponding to the same sampling moment form an element x i in the feature sequence X; all the elements x i at all sampling moments during the period from the radar's first detection of the target to the interceptor hitting the target constitute the data sequence and perform normalization processing;

[0030] Step 1.6: Use a sliding window of length n and move along the direction of increasing t l to partition the data sequence and obtain multiple feature sequences X;

[0031] Step 1.7: Establish labels: The position corresponding to the hit point M is the hit point position label, and the hit point position labels of the multiple feature sequences X obtained in Step 1.6 are the same; for each feature sequence X, the t corresponding to the last element go is t go label;

[0032] For multiple hit points, execute Step 1.1 to Step 1.7 to obtain a series of training samples.

[0033] Preferably, in Step 1.3, the time correspondence between the target motion state S T and the interceptor motion state S D is as follows:

[0034] The sampling periods of the interceptor and the target are the same, both being T s ;

[0035] When the target samples the k-th cycle, the interceptor is launched and the interceptor starts sampling the first cycle;

[0036] The interceptor samples a total of m cycles to obtain the interceptor motion state

[0037] The target samples a total of k + m - 1 cycles to obtain the target motion state

[0038] Adopt the method of time rollback and sequentially match S go and S D and S T along the direction of increasing t to achieve time correspondence; where sequentially match; both match

[0039] Preferably, in Step 1.4, the time t when the interceptor hits the target at each sampling moment is determined as: go as:

[0040] corresponding to t go = t l ; corresponding to t go = (j - 1)·T s , j = 1, 2,...., k + m - 2.

[0041] Preferably, the selection of the hit points for constructing the training samples is:

[0042] For a launch position, N t trajectory points within the maximum interception altitude and maximum range of the interceptor missile are selected as candidate impact points on the predetermined trajectory of the target, and the time interval between the candidate impact points during flight is T p seconds; for the m-th candidate impact point, m = 1, 2,..., N t , based on the current launch position and firing table of the interceptor missile, obtain the time t am required for the interceptor missile to reach the current candidate impact point and the terminal velocity v am , compare the time t am for the interceptor missile to reach the current candidate impact point with the time t gom for the target to reach the current candidate impact point from the initial position, and the terminal velocity v am of the interceptor missile with the minimum terminal velocity If the formula (I) is satisfied, the m-th candidate impact point is considered as a valid impact point;

[0043]

[0044] Take out 1 valid impact point that earliest meets the launch requirements of formula (I) from all valid impact points as the impact point M to participate in the construction of the training sample;

[0045] For different launch positions, determine the impact point M, obtain multiple impact points, and participate in the construction of the training sample.

[0046] Beneficial effects:

[0047] (1) For the combat scenario of an interceptor intercepting a highly maneuverable target, a hit point prediction network configuration of LSTM network + self-attention mechanism + fully connected neural network is proposed. Among them, the position and time of the interceptor hitting the target are related to the flight sequences of both sides. It is not determined by the states of both sides at a certain moment but by the flight processes of both sides. Because only a certain amount of input data can contain the characteristics of the data in the time series and can extract which actions determine the highly maneuverable state when the highly maneuverable target realizes high maneuverability. Therefore, the hit point prediction network needs to extract the characteristics of the aircraft's motion state changing with time as the prediction basis. So, a long short-term memory network (LSTM) is set up to extract the time characteristics of the aircraft's motion state sequence. The maneuver amplitudes of the target and the interceptor change at each moment. A large maneuver amplitude can significantly change the flight trajectory and has a great impact on the prediction, while a small maneuver amplitude has a small impact. Therefore, a self-attention mechanism is introduced to weigh the importance of each part of the motion state sequence, so that the prediction network can focus on the parts of the motion state sequences of the attacking and defending aircraft that have a greater impact on the hit point prediction. Finally, a fully connected layer is used to fit the mapping relationship between the input features and the hit point, thus realizing the prediction. The hit point prediction network configuration of the present invention can be applied to high-speed targets flying across airspaces and speed domains with uncontrollable speed magnitudes and strong maneuverability, and can improve the accuracy of intercept point prediction.

[0048] (2) The present invention designs characterization features for high-speed targets, including not only the motion state of the interceptor and the motion state of the target, but also the height H of the target relative to the ground T , the duration t of the radar tracking the target l , and the relative motion state quantity S between the interceptor and the target R .

[0049] Among them, the setting of H T is considered because the target maneuver differences are large at different heights. Therefore, H T can, to a certain extent, characterize the ballistic characteristics and high maneuverability of the target at different flight stages. The higher the target maneuverability, the greater the change rate of H T . The introduction of H T can help the network learn and extract high maneuver characteristics. Moreover, compared with the y T coordinate of the target, H T is more deterministic after the earth model is determined, and the atmospheric density closely related to the target maneuverability is a function of H T . Therefore, it is simpler and more direct to use H T to characterize the target ballistic characteristics.

[0050] The introduction of t l is to let the network learn the relationship between t go and t l , that is, tl The larger it is, the t go The smaller it is, which is also to make the network learn to correctly represent t go That is, t go The order of magnitude of should be the same as that of t l is quite comparable. Therefore, introducing t l helps the prediction network make a more reasonable prediction for t go and make a more reasonable prediction;

[0051] S R represents the relative motion amount between S D , S T Although S R can be calculated from S D and S T , since the prediction network lacks prior knowledge about the conversion relationship between S D , S T and S R at the initial stage of training, and the error of the prediction network independently deriving the relationship between them based on a large amount of input data is relatively large, which will increase the learning difficulty of the network. Therefore, S R is directly used as a feature input to give a clear hint to the prediction network, thereby accelerating the exploration speed of the prediction network in the initial stage.

[0052] (3) In a preferred embodiment, in the fully connected neural network part of the prediction network, to further enhance the generalization ability of the network, a random inactivation layer (Dropout) is also placed between each layer, so that during the training process of the network, some neurons are randomly made temporarily inactive, forcing each neuron in the network to learn meaningful features to complete the task, preventing the network from relying too much on certain neurons during training, thereby improving the robustness and generalization ability of the network.

[0053] (4) In a preferred embodiment, the output layer of the fully connected neural network is set to 2 layers, namely the hit point position output layer and the target arrival time output layer. Considering that the order of magnitude of the unit of the hit point position and the target arrival hit point time is quite different, sharing one output layer will cause a large difference in the order of magnitude of the weights of the last few layers of the network, and the network is prone to instability. Therefore, on the basis of the original output layer, two independent linear scaling layers are further connected to scale the output of the network to appropriate orders of magnitude respectively, improving the network stability.

[0054] (5) In a preferred embodiment, considering that the interceptor has a certain damage radius, it is not necessarily required to reach the hit point at the same time as the target. There can be a time difference between the interceptor and the target reaching the hit point, and this time difference is exactly the coverage range of the damage radius of the interceptor. Therefore, when screening for effective hit points, the launch conditions of the interceptor are set according to this situation. All scenarios where the interceptor reaches the hit point within the time difference can be regarded as hitting the target, making the data for constructing the training samples more in line with the real situation, and enabling the network to be applied when facing complex real scenarios, with better generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is the prediction network architecture of the high-maneuver target interception point prediction method based on the long short-term memory network of the present invention;

[0056] Figure 2 It is a schematic diagram for matching S T and S D and determining t go ;

[0057] Figure 3 It is the aircraft motion state feature extraction network based on LSTM;

[0058] Figure 4 Aircraft input feature extraction scaled dot product self-attention network;

[0059] Figure 5 It is the structure diagram of the multi-head self-attention module for aircraft input feature extraction;

[0060] Figure 6 It is the fully connected network for fitting the relationship between aircraft input feature information and the hit point;

[0061] Figure 7 It is the training method for the hit point prediction network;

[0062] Figure 8 It is the working process of the hit point prediction network;

[0063] Figure 9 It is the distribution of the interceptor launch positions;

[0064] Figure 10 It is the flight time firing table of interceptor 1;

[0065] Figure 11 It is the terminal velocity firing table of interceptor 1;

[0066] Figure 12 It is the flight time firing table of interceptor 2;

[0067] Figure 13 It is the terminal velocity firing table of interceptor 2;

[0068] Figure 14 For the launch position of the interceptor missile and the distribution of effective hit points;

[0069] Figure 15 For the loss function curve of the hit point prediction network;

[0070] Figure 16 For the t on the test set go Prediction error;

[0071] Figure 17 For the hit point position prediction error on the test set. Specific implementation manner

[0072] The present invention provides a high-maneuver target interception point prediction method based on a long short-term memory network for high-speed targets with complex dynamic characteristics, complex flight characteristics and strong maneuverability. The prediction network adopted by this method consists of a long short-term memory network, a multi-head self-attention mechanism network and a fully connected neural network. The input features include not only the motion state S of the interceptor missile D , the motion state S of the target T , but also the height H of the target relative to the ground T , the radar tracking target duration t l , and the relative motion state S between the interceptor missile and the target R . This method can improve the accuracy of interception point prediction for high-speed targets flying across airspace and speed domain, with uncontrollable speed magnitude and strong maneuverability.

[0073] The following will describe the embodiments of the present invention in detail.

[0074] Step S1: Interceptor missile motion modeling and guidance law setting.

[0075] 1.1. Interceptor missile motion modeling.

[0076] (1) Particle dynamics model

[0077] Taking the launch coordinate system as the reference coordinate system, the dynamic equations of the interceptor missile are established as

[0078]

[0079] In the formula, V D is the speed magnitude of the interceptor missile; θ D and ψ vD are the ballistic inclination angle and ballistic deflection angle of the interceptor missile respectively; m D is the mass of the interceptor missile; P D and X D are the thrust and drag forces received by the interceptor missile respectively; α D and β Dare the angle of attack and sideslip angle of the interceptor missile respectively; g is the acceleration due to gravity at the position of the interceptor missile; a yD and a zD are the longitudinal and lateral acceleration commands of the interceptor missile respectively.

[0080] (2) Particle kinematic model

[0081] Under the static spherical earth model, the kinematic equations of the interceptor missile are:

[0082]

[0083] In the formula, x D , y D and z D are the positions of the interceptor missile in the launch coordinate system.

[0084] (3) Mass equation

[0085] The mass equation of the interceptor missile is

[0086]

[0087] In the formula, m 0 is the mass of the missile at launch; is the mass flow rate per second of the interceptor missile engine, and t is time.

[0088] 1.2. Guidance law setting and control quantity generation.

[0089] When the interceptor missile intercepts the target, the relative motion relationship equations between the interceptor missile and the target are:

[0090]

[0091] In the formula, r D is the interceptor missile-target relative distance (abbreviated as "missile-target distance"), V M is the speed of the target, V D is the speed of the interceptor missile, θ M is the ballistic inclination angle of the target in the launch coordinate system, ψ vM is the ballistic deflection angle of the target in the launch coordinate system, q yD and q zD are the components of the interceptor missile's field of view angle in the elevation and azimuth directions respectively; is the derivative of r D , q yD , q zD respectively.

[0092] Assume that the interceptor missile uses a three-dimensional proportional navigation law to attack the target, and the guidance command is

[0093]

[0094] Where: a yD and a zD are the normal accelerations of the interceptor missile, N yD and N zD are the longitudinal and lateral proportional navigation coefficients respectively.

[0095] After the interceptor missile is vertically launched and flies uncontrolled, when the altitude reaches the specified altitude h t the program turns until the velocity vector of the missile coincides with the line connecting the missile and the target, and then it conducts proportional navigation flight with the predicted impact point as the target according to (5). When the missile-target distance shrinks to the detection distance of the seeker, the seeker detects the target, and then it conducts proportional navigation flight with the detected real target according to Equation (5).

[0096] For an axisymmetric interceptor missile with an inclination stabilization system, when studying its guidance problem, the control variables are the angle of attack and the sideslip angle. The expression for the normal acceleration of the interceptor missile is

[0097]

[0098] Where, θ uD is the angle between the velocity vector V D of the interceptor missile and the local north-east-down coordinate system O 1 x b y b plane, i.e., the local ballistic inclination angle; P D is the thrust of the interceptor missile; Y D and Z D are the lift and side force received by the interceptor missile respectively. Ignoring minor factors, the calculation formulas for the lift and side force are

[0099]

[0100] Where are the lift coefficient and side force coefficient of the interceptor missile respectively; α and β are the angle of attack and sideslip angle of the interceptor missile respectively; ρ is the air density; S is the relative reference area of the interceptor missile.

[0101] Considering that the angle of attack and sideslip angle of the interceptor missile are small, there are α D ≈sinα D , β D ≈sinβ D , and substituting Equation (7) into Equation (6), the relationship between the angle of attack and sideslip angle and the normal accelerations a yD and a zD is

[0102]

[0103] Where is the dynamic pressure.

[0104] Substituting the normal acceleration command of the proportional navigation law shown in Equation (5) into Equation (8), the angle of attack and sideslip angle for implementing this command can be obtained.

[0105] Step S2: Formulating the interception ejection table and determining the launch conditions.

[0106] When the launch position of the interceptor missile is fixed, the launch parameters are determined, and the guidance mode is determined, its flight trajectory in the air is unique. In this step, the firing table for the interceptor missile to attack targets at different positions in the air is formulated.

[0107] Let the feasible interception altitude range of the interceptor missile be [h smin , h smax , and the feasible interception range be [r smin , r smax . In the area that simultaneously satisfies the interception altitude and interception range (i.e., the area where the interception altitude h s ∈ [h smin , h smax and the interception range r s ∈ [r smin , r smax ), multiple points are selected at certain intervals as target points. The interceptor missile uses the guidance law shown in Equation (5) to attack the target, and the terminal velocity v a and flight time t a of the interceptor missile attacking these targets are obtained, thereby establishing the firing table of the terminal velocity and flight time of the interceptor missile.

[0108] It should be noted that since the interceptor missile is vertically launched and program turns to align with the target point after reaching a certain altitude (i.e., the velocity vector coincides with the line of sight), therefore, regardless of the absolute position of the target point, as long as its altitude h t = h s and the distance r D between it and the interceptor missile = r s , the motion law and parameters of the interceptor missile are the same, that is, the flight time and the final velocity are equal. Therefore, on the premise that the position and initial velocity of the interceptor missile are certain, a firing table for determining the final velocity v s , flight time t s by h a , r a can be established.

[0109] Assume that the motion law of the target has been obtained through ground radar detection and tracking. When the altitude h t of the target ∈ [h smin , h smax and the distance r D between the target and the interceptor missile ∈ [r smin , r smaxWhen it is in this situation, it may all be intercepted by the interceptor missile. Therefore, a certain point within the interception range is used as the hit point of the interceptor missile's interception target. At this time, it is assumed that the time required for the target to fly to this point is t go , and then according to h s =h t , r s =r D Interpolate the firing table to obtain the time t a required for the interceptor missile to fly to the selected hit point and the terminal velocity v a . Theoretically, when t a =t go , the interceptor missile and the target reach the target point simultaneously, that is, the interceptor missile successfully intercepts the target. However, considering factors such as the damage radius of the warhead of the interceptor missile, it is considered that when |t a -t go |≤Δt a (Δt a is the threshold), the interceptor missile successfully intercepts the incoming target. In order to ensure the damage effect on the target, usually there is a limit on the minimum terminal velocity of the interceptor missile, that is, it is necessary to ( is the required minimum terminal velocity). Therefore, the launch condition of the interceptor missile is:

[0110]

[0111] Since the interceptor missile will turn on the seeker near the target and guide with the real target, therefore, for Δt a within a certain range, the interceptor missile can still hit the target.

[0112] Step S3: Input feature setting and dataset generation of the hit point prediction network.

[0113] 3.1. Construction of the input features of the hit point prediction network.

[0114] The position of the hit point is related to the motion states of the interceptor missile and the target. Therefore, the following state variables are selected as the input features of the hit point prediction network.

[0115] (1) The motion state S D =[x D , y D , z D , V D , θ D , ψ vD of the interceptor missile and the motion state S T =[x T , y T , z T , V T , θ T , ψ vT of the target. Among them, (xD , y D , z D ) is the position of the interceptor, V D is the velocity of the interceptor, θ D and ψ vD are the ballistic inclination angle and the ballistic deflection angle of the interceptor respectively; (x T , y T , z T ) is the position of the target, V T is the velocity of the target, θ T and ψ vT are the ballistic inclination angle and the ballistic deflection angle of the target respectively.

[0116] (2) The height H of the target relative to the ground T , which is used to prompt the current ballistic characteristics of the target to the prediction network. Taking a hypersonic glide reentry target as an example, when H T is very large, due to the thin air and large heat flux, the target basically does not make large maneuvers and the ballistic is relatively straight; when the target descends to a certain height where the air density is large, the target has a large maneuvering ability and can make large maneuvers both longitudinally and laterally. Laterally, it is mainly a swinging maneuver similar to an "S" shape, and longitudinally, it is a jumping maneuver with gradually decreasing amplitude; in the downward pressure section, in order to achieve a large impact angle attack, the target descends rapidly in height while the lateral maneuver amplitude is relatively small. Therefore, H T can characterize the ballistic characteristics of the target at different flight stages to a certain extent. The additional introduction of H T is because when considering the curvature of the Earth's surface, when the target is far from the radar (more than 200 km), H T and y T have a large difference, and for the same target, the value of y T is also related to the origin position of the north celestial east coordinate system, while H T is determined after the Earth model is determined and the atmospheric density, which is closely related to the target maneuverability, is a function of H T . Therefore, using H T to characterize the target ballistic characteristics is simpler and more direct.

[0117] (3) The duration t of the radar tracking the target l = t - t l0 , where t is the current time and t l0 is the time when the radar first detects the target. t l is the timestamp of each element in the input sequence of the prediction network, indicating the time relationship before and after each element in the input sequence. The introduction of t l is to let the network learn the relationship between t go and t l , that is, the larger t l is, the larger t goThe smaller it is, the more the network is required to learn to correctly represent t go That is, t go should be of the same order of magnitude as t l Therefore, introducing t l helps the prediction network to make more reasonable predictions for t go

[0118] (4) Relative motion state variables between the interceptor and the target where r D is the distance between the interceptor and the target; q yD and q zD are the components of the interceptor's field of view angle in the vertical and azimuth directions respectively; is the derivative of r D , q yD , q zD and can be calculated from the relative motion relationship equations of the interceptor and the target in formula (4). S R is used to characterize the relative position relationship and changes between the offensive and defensive aircraft. Although S R can be calculated from S D and S T , due to the lack of prior knowledge about the conversion relationship between S D , S T and S R at the initial stage of training the prediction network, and the error of the prediction network independently deriving the relationship between them from a large amount of input data is relatively large, which will increase the learning difficulty of the network. Therefore, S R is directly used as a feature input to give a clear hint to the prediction network.

[0119] 3.2. Input sample setting.

[0120] Let the time for the target to reach the hit point be t go . The prediction network needs to predict the hit point position and t go based on the motion state sequences of the offensive and defensive aircraft input for a period of time. Therefore, an input sample of the prediction network is defined as S = (X, Y), where X = {x 1 , x 2 ,..., x n} is the feature sequence, and Y = [x p , y p , z p , t go is the prediction label, where (x p , y p , z p ) is the hit point position and tgo is the time for the target to reach the hit point.

[0121] The elements in X are arranged according to t l ​Arranged in increasing order, defined as

[0122]

[0123] Where: i = 1, 2, ..., n (n is the number of elements in X) represents the position index of the element x i in X. t go Take the last element x in X n corresponding t go value, because x n is the input of the network at the current moment, so x n corresponding t go is the time from the current moment for the target to reach the hit point.

[0124] 3.3. Generation of the sample data set for the hit point prediction network.

[0125] The generation of the sample data set for the hit point prediction network includes the following steps:

[0126] Step ①: Selection of the predetermined hit point.

[0127] For a launch position, select N t trajectory points within the maximum interception altitude and maximum range of the interceptor missile on the predetermined trajectory of the target as candidate hit points, and the flight time interval between candidate hit points is T p seconds. For the m-th (m = 1, 2, ..., N t ) target point, based on the current launch position and firing table of the interceptor missile, obtain the time t am required for the interceptor missile to reach this target point and the terminal velocity v am , compare the time t am for the interceptor missile to reach the predicted hit point and the time t gom for the target to reach the predicted hit point from the initial position, and the terminal velocity v am of the interceptor missile with the minimum terminal velocity If the formula (9) is satisfied, then the m-th hit point is considered a valid hit point. Make this judgment for all N t hit points to obtain N y (N y ≤ N t ) valid hit points, that is, the interceptor missile can hit the target at these N y hit points when launched at different times. However, according to the principle of "detect early and intercept early", select the earliest 1 valid hit point that meets the requirements (i.e., the point where N y = 1) as the predetermined hit point of this type of interceptor missile at this launch position.

[0128] Execute the operation of Step ① for different launch positions to obtain multiple predetermined hit points.

[0129] Step ②: Generation of input feature sequence data.

[0130] In this step, each selected predetermined hit point is traversed, and Steps 3.2 to 3.5 are executed to obtain a series of training samples.

[0131] Select one of the hit points M. According to the target trajectory prediction, determine the height H of the target relative to the ground at each sampling moment during the period from the radar's first detection of the target to the interceptor hitting the target. T , the radar tracking duration t of the target l and the target motion state S T = [x T , y T , z T , V T , θ T , ψ vT , which form H T , t l and S T in the feature sequence X.

[0132] Through simulation, determine the motion state S of the interceptor at each sampling moment during the period from the launch of the interceptor to the interceptor hitting the target. D = [x D , y D , z D , V D , θ D , ψ vD ;

[0133] Time-align the target motion state S T and the interceptor motion state S D and calculate the relative motion state of the interceptor-target. It should be noted that: the initial moment of S T is the moment when the radar first detects the target, and the initial moment of S D is the moment when the interceptor is launched. Since the observation distance of the radar is usually greater than the range of the interceptor, the interceptor is often launched after the radar has tracked the target for a period of time. That is, the initial moments of S T and S D are not the same. Therefore, it is necessary to process S T and S D into the same time length according to the time sequence, and then calculate S R . The specific method is as follows:

[0134] Adopt the method of time regression, and sequentially match S go and S D along the direction of increasing t T , and at the same time determine the corresponding tgo Labels, such as Figure 2 as shown Figure 2 in is the motion state of the target when the radar first detects the target. is the motion state at the time of interceptor missile launch, corresponding to the target state is the state when the interceptor missile hits the target, corresponding to the target state Let S D and S T have a sampling period of T s , along t go looking, then corresponds to t go = 0, corresponds to t go = T s , and so on, corresponds to t go = t l . Since the radar tracking time of the target is greater than the flight time of the interceptor missile, i.e., t l > t a , in order to align S T and S D in chronological order and considering the actual situation, let t go > t a During the time period, the interceptor missile states corresponding to the target state are all (During this period, the interceptor missile has not been launched, so its states are all ). Matches with in sequence. Then, according to S D and S T after the matching is completed, calculate S R .

[0135] The S D , S T , H T , t l , S R corresponding to the same sampling moment form an element x i in the feature sequence X; during the period from when the radar first detects the target to when the interceptor missile hits the target, the elements x i at all sampling moments constitute the data sequence At this time, a data sequence with a length of q = k + m - 1 is obtained

[0136] Step ③: Standardize the input feature sequence data.

[0137] Since the features S D , S T , S R , HT and t l Some of the quantities in it have a large difference in order of magnitude. Directly inputting them into the network is likely to cause numerical instability in the network or the characteristics with a large order of magnitude to overwhelm those with a small order of magnitude. Therefore, the feature sequence is standardized to convert all data to the same scale and eliminate the influence of the dimension on the network analysis of features. The z-score standardization method is used for the given input sequence The standardization formula is

[0138]

[0139] In the formula: is the value after standardization, and μ and σ are respectively the mean and standard deviation of

[0140] Step ④: The feature sequence is segmented by a sliding window to generate the feature sequence X in the sample.

[0141] Select a valid hit point. The standardized feature sequence generated through Steps 3.2 and 3.3 has a length of q = k + m - 1. The designed length of the sample feature sequence X is n (n < q). Therefore, a sliding window with a step size of 1 and a length of n is used to segment it, and the sliding direction is along t l increasing direction, then q - n + 1 short feature sequences X 1 , X 2 ,..., X q are obtained, that is

[0142] X j ={x j , x j+1 ,..., x j+n-1} (12)

[0143] In the formula: j = 1, 2,..., q, x j represents the element in X j , and the definition is the same as that in formula (10).

[0144] Step ⑤: Establish the label Y.

[0145] Because is generated by simulating target shooting for a hit point M, the hit point position labels corresponding to X j are the same. The t go label takes the t j corresponding to the last element x j+n-1 in X go .

[0146] Select different hit points, and different input data feature sequences can be generated. More input feature samples can be generated through a sliding window.

[0147] Step S4: Construct a hit point prediction network.

[0148] 4.1. Overall structure diagram of the prediction network.

[0149] As Figure 1 shown, the prediction network includes a target motion state feature extraction network and a relationship fitting network between features and hit points. The target motion state feature extraction network consists of a long short-term memory network (LSTM) and a multi-head self-attention mechanism network. The relationship fitting network between features and hit points uses a fully connected neural network.

[0150] The design principle of the prediction network is as follows: The position and time of the interceptor hitting the target are related to the flight sequences of both sides (not determined by the states of both sides at a certain moment but determined by the flight processes of both sides). Therefore, the hit point prediction network needs to extract the features of the aircraft motion state changing with time as the prediction basis. Thus, a long short-term memory network (LSTM) is set up to extract the time features of the aircraft motion state sequence. The maneuver amplitudes of the target and the interceptor change at each moment. A large maneuver amplitude can significantly change the flight trajectory and has a great impact on the prediction, while a small maneuver amplitude has a small impact. Therefore, a self-attention mechanism is introduced to weigh the importance of each part of the motion state sequence, so that the prediction network can focus on the parts of the aircraft motion state sequences of both the offensive and defensive sides that have a greater impact on the hit point prediction. Finally, a fully connected layer is used to fit the mapping relationship between the input features and the hit points, thereby achieving the prediction.

[0151] Figure 1 In D , the input of the prediction network is the aircraft feature sequences X of both the offensive and defensive sides, including the motion state vector S of the interceptor T , the motion state S of the target T , the flight altitude H of the target relative to the ground l , the duration t of the radar tracking the target R , and the relative motion state vector S of the aircraft of both the offensive and defensive sides 1 , C 2 ,..., C n is the cell state of the LSTM, and h 1 , h 2 ,..., h nis the LSTM hidden layer state. The LSTM outputs in a continuous multi-time-step manner, that is, it outputs a vector at each time step, and the output is also a sequence of length n. The Self-Attention network extracts deeper features by weighted summation of elements in the input data sequence, where the weighted weights are dynamically calculated based on the correlation between elements in the sequence. If some elements in the sequence have a relatively large correlation with other elements, it indicates that these elements have a greater impact on the generation of the sequence and need to be focused on. Conversely, their impact on the output should be reduced. This mechanism enables the prediction network to weigh the importance of each element in the input sequence and dynamically adjust their impact on the prediction result of the hit point by adjusting the weights. For input sequences with a higher element dimension, the self-attention operation can be repeated multiple times in parallel, that is, in the form of Multi-Head Attention, to enhance the effect of the self-attention mechanism. Finally, a fully connected neural network is used to fit the mapping relationship between the features output by the multi-head self-attention mechanism network and the hit point position (x p , y p , z p ), and the time tgo to reach the target hit point, and the prediction result is output.

[0152] Figures 3 to 6 The structures of the LSTM network, the multi-head self-attention mechanism network, and the fully connected network are given respectively.

[0153] 4.2. The LSTM network for extracting the temporal features of the aircraft motion state.

[0154] The structure of the LSTM network is as Figure 3 shown. It is an LSTM network with multiple hidden layers. Its input contains three dimensions. The first dimension is the number of samples, the second dimension is the sequence length, and the third dimension is the number of features. At t = 1, an LSTM cell is equivalent to a fully connected neural network. At t = 2, it is also equivalent to a fully connected neural network. After unfolding along the time axis, the hidden layer state h 1 and the cell state C 1 trained at t = 1 will be passed to the LSTM cell corresponding to t = 2, and t = 2 will be passed to the LSTM cell corresponding to t = 3, and so on, until finally passed to the LSTM cell corresponding to the terminal time t = n of the input sequence. Therefore, in the LSTM, the hidden layer state and the cell state act as "memory", which contains information related to the data seen by the network before. This transmission mechanism enables the LSTM to capture the temporal dependence relationship in the sequence.

[0155] The transmission of the LSTM network hidden layer state and the cell state is controlled by the forget gate, the input gate, and the output gate.

[0156] The unit output h at time t of the LSTM network t depends on the sequence element X of the current input t , the output h at time t-1 t-1 and the cell state C passed between the units t . The unit output at time t-1 in turn depends on the sequence element X of the input t-1 , the output h at the previous time t-2 and the cell state C t-1 , and so on. Through such a mechanism, the sequence element input at the previous time affects the transmission of the cell state and the hidden layer state, thereby affecting the unit output at the next time, that is, the information transmission of the sequence element input at the next time. This influence can characterize the correlation between the sequence elements at two times. Therefore, the LSTM can model the temporal dependence relationship between the elements of the input sequence.

[0157] 4.3. The multi-head self-attention mechanism network for extracting the importance features between input elements

[0158] On the premise of retaining the temporal features of the sequence, the self-attention mechanism network further aggregates the features at important moments in the aircraft motion, enabling the prediction network to also focus on the important information existing locally in the sequence. This patent adopts the scaled dot-product self-attention mechanism network, which uses W q , W k and W v three weight matrices to calculate the attention weights and extract information. These matrices project the aircraft motion state input x (i) into the query q (i) , key k (i) and value v (i) components of the sequence respectively, that is, q (i) =x (i) W q , k (i) =x (i) W k and v (i) =x (i) W v , where i is the index position of the element x (i) in the sequence. If the dimension of x (i) is d, then the shapes of the projection matrices W q and W k are d×d q and d×d k respectively, while the shape of W v is d×d v . In order to calculate the dot product of the query and the key, there is d q =d k, and the dimension of the value can be arbitrarily selected, which determines the dimension of the output vector (referred to as the context vector). An input sequence containing n elements, after self-attention operation, obtains n context vectors. Taking the second element x of the sequence (2) as an example, the corresponding context vector z (2) is calculated as shown in Figure 4 .

[0159] Figure 4 Among them, the query q (2) is the self-generated prompt information generated according to x (2) , and the key k (j) (j = 1, 2,..., n, n is the length of the input sequence) is the reference vector used to match the query. Because the attention weight ω 2,j is obtained by the dot product of q (2) and k (j) , according to the property of the dot product, the more similar q (2) and k (j) are, the larger ω 2,j is, that is, the element x at the position index j in the sequence (j) has a stronger correlation with x (2) . The value v (j) represents the information that needs to be obtained from x (j) . The context vector z (2) corresponding to x (2) is obtained by weighted summation of the value v (j) through the normalized attention weight α 2,j . Among them, α 2,j is obtained by scaling ω 2,j by d k and normalizing it through the softmax function. Normalization is to prevent the attention weight from becoming too small or too large, resulting in numerical instability of the network or affecting the ability of the model to converge during training. The method of normalizing ω 2,j using softmax is

[0160]

[0161] In the formula: is the scaled attention weight. By reprocessing the motion state sequences of both the attacking and defending sides processed by LSTM through the self-attention mechanism, important local information can be focused on when extracting the global temporal features of the sequence, enabling the prediction network to better handle the maneuvers of the aircraft of both the attacking and defending sides.

[0162] Input a sequence, Figure 4The self-attention module shown outputs only one sequence of context vectors, called a single-head self-attention module. To enhance the focusing effect of self-attention, for an input sequence with a higher element dimension, multiple self-attention operations can be performed on it in parallel to obtain multiple sequences of context vectors, called the multi-head self-attention mechanism. The number of output sequences of context vectors is the number of heads (Head Number). Taking the second element x (2) as an example, its multi-head self-attention operation process is as Figure 5 shown.

[0163] Figure 5 In , Q represents , K represents , and V represents (2) For the m context vectors corresponding to x , they are usually concatenated end to end to form a vector as the output vector of the multi-head self-attention module. When the dimension of x (2) is high, different heads of the multi-head self-attention module can focus on different parts of x (2) , thus capturing more important information, such as respectively paying attention to the state changes of the interceptor and the target through different self-attention heads.

[0164] 4.4. A fully connected network that fits the relationship between the input feature information and the hit point.

[0165] After extracting the sequence features of the motion states of the aircraft on both the offensive and defensive sides through the LSTM and multi-head self-attention modules, a fully connected neural network is used to fit the mapping relationship between the features and the hit point position (x p , y p , z p ), and the time t go when the target reaches the hit point, so as to achieve prediction. The structure of the fully connected neural network is as Figure 6 shown.

[0166] Figure 6 In The tensors, while the fully connected neural network only accepts vector inputs. Therefore, the rows of the tensor output by the multi-head self-attention module are concatenated end to end and flattened into a vector, which is then input into the fully connected neural network. The fully connected neural network includes an input layer, multiple hidden layers, and an output layer connected in sequence; an activation function layer is set between the input layer and the hidden layers, and between the hidden layers. To further enhance the generalization ability of the network, a dropout layer is also placed between each layer, causing the network to randomly deactivate some neurons temporarily (i.e., their output values are zero) during training, forcing each neuron in the network to learn meaningful features to complete the task and preventing the network from relying too much on certain neurons during training, thereby improving the robustness and generalization ability of the network.

[0167] The present invention also designs the output layer. Considering that the unit of the hit point position (x p , y p , z p ) is m, while the unit of the time t go for the target to reach the hit point is s, and the order of magnitude of the two is quite different. Sharing one output layer will cause a large difference in the order of magnitude of the weights in the last few layers of the network, and the network is prone to instability. Therefore, on the basis of the original output layer, two independent linear scaling layers are connected, and the output of the network is scaled to an appropriate order of magnitude respectively, so as to obtain the prediction results of the hit point position (x p , y p , z p ) and the time t go for the target to reach the hit point, and the original output layer is used as the last hidden layer.

[0168] Step S5: Training of the hit point prediction network.

[0169] The training method of the hit point prediction network is as Figure 7 shown.

[0170] Figure 7 In, the feature sequence X of the training sample is input into the hit point prediction network to obtain the hit point prediction output Combined with the true label Y to calculate the loss function L, and the update gradient ΔW ij of the weight parameter is obtained by differentiating the loss function, and then the gradient descent method is used to update the network weight parameter, where α is the learning rate of gradient descent. To simultaneously achieve the prediction of the hit point position (x p , y p , z p ) and the time t goFor accurate prediction, it is necessary to fuse the loss function terms corresponding to the two prediction targets into a total loss function L to generate gradients, and then backpropagate to optimize the weights of each layer of the prediction network, so that the loss functions of the two prediction tasks are minimized. A commonly used method for fusing multi-task loss functions is to assign a weight to each task's loss function term and then perform weighted summation, that is

[0171] L = ω p L p + ω t L t (14)

[0172] In the formula: L is the total loss function, L p and L t are the loss functions for predicting the position of the hit point and the time for the target to reach the hit point respectively, ω p and ω t are the corresponding weighted weights, which are used to adjust the relative proportion of the loss function terms, so that the network can be fully optimized for each prediction target.

[0173] Step S6: The overall working process of the hit point prediction network.

[0174] After the hit point prediction network is trained, the prediction results can be used to determine the launch timing of the interceptor missile and guide the interceptor missile to intercept the target. During actual prediction, the radar regularly measures the motion states of the target and the interceptor missile to obtain the elements x i of the feature sequence and cache them into the feature sequence X; when the length of the feature sequence X reaches the input sequence length n of the prediction network, it is input into the prediction network, and the prediction network outputs the predicted hit point position and the time t go for the target to reach the hit point.

[0175] Figure 8 shows the complete process of real-time prediction of the hit point prediction network. As shown in the figure, the working process of the hit point prediction network is divided into the following steps:

[0176] Step 6.1: The radar regularly measures the motion states of the target and the interceptor missile to obtain the elements x i of the feature sequence and cache them into X.

[0177] Step 6.2: Check whether the length of X is less than the network input sequence length n. If so, go to Step 6.1; if it is greater than n, discard the oldest x i in X to ensure that the length of X is equal to n, and then input it into the prediction network to predict the position of the hit point and the time t go for the target to reach the hit point.

[0178] Step 6.3: Interpolate the numerical firing table of the interceptor according to the predicted result of the impact point to obtain the flight time t for the interceptor to reach the impact point along the standard trajectory a and the terminal velocity v a , when the launch condition is met, i.e., |t go - t a | ≤ Δt a and , the interceptor can be launched, otherwise go to Step 6.1

[0179] Step 6.4: After the interceptor is launched, first use the predicted impact point as the target for mid-course guidance. If the distance between the interceptor and the target meets the interceptor seeker activation condition, i.e., the missile-target distance r is less than the interceptor seeker activation distance r t , then switch to terminal guidance and guide with the real target, otherwise go to Step 6.1

[0180] Step 6.5: If the interceptor hits the target, i.e., r < r b , or loses the target, i.e., r > r b and , then end, where r b is the damage radius of the interceptor

[0181] The interceptor switches to terminal guidance when it is close to the target because there are errors in the impact point prediction. Guiding with the real target can reduce the errors and enable the interceptor to better cope with target maneuvers and accurately hit the target in the terminal stage

[0182] The following verifies the method for predicting the impact point of a highly maneuverable and high-speed target

[0183] I. Parameter Settings

[0184] 1) Interception scenario parameter settings

[0185] According to the public literature, two types of interceptors are set, and their parameters are shown in Table 1

[0186] Table 1 Interceptor parameter settings

[0187]

[0188]

[0189] Suppose the launch positions of the interceptor missiles are distributed on both sides of the ground trajectory of the aircraft. The starting point of the reference ground trajectory of the aircraft is 75° east longitude and 60.66° north latitude, and the ending point is 90° east longitude and 62° north latitude. Moreover, the latitude difference between the launch position of the interceptor missile and the ground trajectory point at the same longitude does not exceed ±0.2°. The purpose is to avoid the situation where the interceptor missile fails to intercept because the aircraft has left the range of the interceptor missile before the interceptor missile is launched. In order to enable the hit point prediction network to adapt to the changes in the launch positions of the interceptor missiles, the launch area of the interceptor missiles is discretized so that there are significant differences between the selected launch positions. Among them, the longitude discretization granularity of the launch position distribution area of the interceptor missile is 5°, and the latitude discretization granularity is 0.1°. Then there are a total of 12 interceptor missile launch positions, numbered (1)-(12), as shown in the figure.

[0190] 2) Hit point prediction parameter settings

[0191] The structure and training hyperparameter settings of the hit point prediction network are shown in Table 2.

[0192] Table 2 Hit point prediction network training parameter settings

[0193]

[0194]

[0195] In Table 2, the learning rate decay node refers to when the number of training Epochs reaches the node value, multiplying the current learning rate by the learning rate decay factor as the learning rate for the next stage. The training set ratio refers to the ratio of the number of samples in the training set to the number of samples in the total data set.

[0196] II. Simulation results

[0197] 1) Numerical firing table of the interceptor missile

[0198] Suppose N s = 10 groups of fixed points h s ∈ [h smin , h smax and r s ∈ [r smin , r smax are selected during the Monte Carlo simulation. The flight time and terminal velocity firing tables of interceptor missile 1 are shown in Figure 10 and Figure 11 respectively, and the flight time and terminal velocity firing tables of interceptor missile 2 are shown in Figure 12 and Figure 13 respectively.

[0199] The position of the hit point and the time t go when the target arrives at the hit point are predicted through the prediction network. According to Figures 10 to 13For a determined numerical firing table, substitute the local altitude and the missile-target distance corresponding to the impact point into the firing table and interpolate to obtain the flight time \(t\) for the interceptor missile to reach the impact point. a and the terminal velocity \(v\). a , and according to Equation (9), it can be determined whether the interceptor missile is launched.

[0200] 2) Training results of the impact point prediction network

[0201] When generating the dataset, there are two types of interceptor missiles available for each launch position, namely interceptor missile 1 and interceptor missile 2. During the glide section of the high-speed aircraft, the selection time interval for candidate impact points is \(T\) p = 5 s. Since the duration of the downward pressure section is relatively short, \(T\) p is set to 1 s, so there are a total of 238 candidate impact points. After screening according to Equation (9), 101 effective impact points are obtained. The distribution of the selected impact points and the interceptor missile launch positions is as Figure 14 shown. For the interceptor missile at a certain launch position, if the earliest effective impact point that meets the launch requirements is selected as the predetermined impact point for this type of interceptor missile at this launch position, then 13 predetermined impact points are obtained. The total number of dataset samples \(S=(X, t\) go ) generated based on these predetermined impact points is 15096. According to the given training set and test set ratio, the total number of training set samples is 12076, and the total number of test set samples is 3020. After the training is completed, the loss function curve of the prediction network is as Figure 15 shown, and the \(t\) go on the test set and the prediction error of the impact point position (i.e., the distance between the predicted impact point position and the actual position) are respectively as and 16 and Figure 17 shown.

[0202] As Figure 14 shown, the launch positions of the interceptor missiles are distributed on both sides of the ground trajectory of the aircraft. Among them, the launch positions that can meet the launch conditions of interceptor missile 1 are the launch positions numbered (1)-(9) in Figure 14 , and the corresponding predetermined impact point altitude range is 30 - 35 km, that is, interceptor missile 1 mainly conducts interception during the glide section of the aircraft; the launch positions that can meet the launch conditions of interceptor missile 2 are the launch positions numbered (9)-(12) in Figure 14 , and the corresponding predetermined impact point altitude range is 10 - 30 km, that is, interceptor missile 2 mainly conducts interception at the end of the glide section or during the downward pressure section. From Figure 15 , it can be seen that both the training loss and the test loss of the prediction network converge well. As Figure 16 shown, the maximum prediction error of \(t\) go on the test set is 5.26 s, and the average prediction error is 1.01 s. As Figure 17 shown, the maximum prediction error of the impact point position on the test set is 1364.76 m, and the average prediction error is 359.39 m. The Monte Carlo method is used to determine \(t\)go When the prediction error is less than 3 s and the prediction error of the impact point position is less than 600 m, the interceptor enters the terminal guidance and has a probability of more than 90% of hitting the target. Then, combining Figure 16 and Figure 17 it can be seen that the prediction efficiency of the impact point prediction network on the test set is 87.9%, indicating that the trained impact point prediction network has good prediction accuracy under the given interception scenario.

[0203] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in this description can be different and are not restricted. Therefore, those skilled in the art of the present invention can modify or equivalently replace the technical solutions recorded in the foregoing embodiments; and these modifications and replacements do not depart from the spirit and technical solutions of the present invention and should all fall within the protection scope of the present invention.

Claims

1. A method for predicting interception points of high-maneuverable targets based on long short-term memory networks, characterized in that: include: Step 1: The radar periodically measures the motion state of the target and the interceptor missile to obtain the element x of the characteristic sequence i , and cache it into the feature sequence X; The element x i Including: interceptor missile motion state S D , target motion state S T 、Height of target relative to the ground H T 、Radar tracking target time t l , and the relative motion state S between the interceptor missile and the target R ; Step 2: When the length of the feature sequence X reaches the input sequence length n of the prediction network, it is input into the prediction network, and the prediction network outputs the predicted hit point position and the time t when the target arrives at the hit point go ; And continuously update the feature sequence X prediction output; The prediction network includes a target motion state feature extraction network and a feature and hit point relationship fitting network; the target motion state feature extraction network is composed of a long short-term memory network LSTM and a multi-head self-attention mechanism network; the feature and hit point relationship fitting network adopts a fully connected neural network.

2. The method according to claim 1, characterized in that In the step 1, Interceptor missile motion state S D =[x D ,y D ,z D ,V D ,θ D ,ψ vD ], where (x D ,y D ,z D ) is the position of the interceptor missile, V D is the speed of the interceptor missile, θ D and ψ vD are the ballistic inclination and ballistic deviation of the interceptor missile respectively; Target motion state S T =[x T ,y T ,z T ,V T ,θ T ,ψ vT ], where (x T ,y T ,z T ) is the target position, V T is the speed of the target, θ T and ψ vT are the target's ballistic inclination and ballistic deviation respectively; Radar tracking target time t l =tt l0 , where t is the current time, t l0 is the moment when the radar first detects the target; t l It is the timestamp of each element in the prediction network input feature sequence; Relative motion state between interceptor missile and target where r D is the distance between the projectile and the target; q yD and q zD are the components of the interceptor missile's field of view angle in the elevation direction and the azimuth direction respectively; For r D ,q yD ,q zD The derivative of is calculated based on the following relative motion equations between the interceptor missile and the target; Among them, v M is the speed of the target, θ M is the ballistic inclination angle of the target in the launch coordinate system, ψ vM is the ballistic deviation angle of the target in the launch coordinate system.

3. The method according to claim 1, characterized in that In the prediction network, each row of the tensor output by the multi-head self-attention mechanism network is connected end to end, flattened into a vector, and then input into a fully connected neural network; the fully connected neural network includes an input layer, multiple hidden layers, and an output layer connected in sequence; an activation function layer is set between the input layer and the hidden layer, and between the hidden layers.

4. The method according to claim 1, characterized in that A random dropout layer, Dropout, is set between the input layer and the hidden layer, and between the hidden layers.

5. The method according to claim 3 or 4, characterized in that The prediction network further includes a hit point position output layer and a target arrival time output layer connected after the output layer, which are used to scale the predicted hit point position and the target arrival time of the hit point output by the prediction network to appropriate orders of magnitude respectively.

6. The method according to claim 5, characterized in that The total loss function L = ω used in the prediction network training p L p +ω t L t ; Where: L p and L t are the loss function terms for the prediction of the hit point position and the time when the target arrives at the hit point, ω p and ω t is the corresponding weighted weight.

7. The method according to claim 1, characterized in that The prediction network is trained in advance, and the training samples are constructed as follows: Step 1.1: For the selected impact point M, according to the target trajectory forecast, determine the target height H relative to the ground at each sampling time from the time the radar first detects the target to the time the interceptor missile hits the target T 、Radar tracking target time t l and the target motion state S T ; Step 1.2: Determine the motion state S of the interceptor missile at each sampling time from the launch of the interceptor missile to the interceptor missile hitting the target through simulation D ; Step 1.3: Set the target motion state S T and the interceptor missile motion state S D Perform time correspondence and calculate the relative motion state S between the interceptor missile and the target R ; Step 1.4: Determine the time t at which the interceptor missile hits the target at each sampling moment go ; Step 1.5: S corresponding to the same sampling time D , S T , H T ,t l , S R An element x that constitutes the feature sequence X i ; The element x at all sampling times from the time the radar first detects the target to the time the interceptor missile hits the target i Constructing data sequence And standardize it; Step 1.6: Use a sliding window of length n and move along t l Increasing direction, for the data series Perform segmentation to obtain multiple feature sequences X; Step 1.7: Create a label: The position corresponding to the hit point M is the hit point position label, and the hit point position labels of multiple feature sequences X obtained in step 1.6 are the same; the last element in each feature sequence X corresponds to t go t go Label; For multiple hit points, execute steps 1.1 to 1.7 to obtain a series of training samples.

8. The method according to claim 7, characterized in that In step 1.3, the target motion state S T and the interceptor missile motion state S D The time correspondence is: The sampling period of the interceptor missile and the target is the same, both are T s ; When the target is sampled to the kth cycle, the interceptor missile is launched and starts sampling the first cycle; The interceptor missile samples m cycles in total to obtain the motion state of the interceptor missile. The target is sampled for a total of k+m-1 cycles to obtain the target motion state Take the time-reversal method and follow t go The increasing direction matches S in turn D and S T , to achieve time correspondence; and Match in sequence; Match all 9. The method according to claim 8, characterized in that In step 1.4, the time t at which the interceptor missile hits the target at each sampling moment is determined. go for: Corresponding to t go =t l ; Corresponding to t go =(j-1)·T s , j=1,2,....,k+m-2.

10. The method according to claim 7, characterized in that The selection of hit points for training sample construction is: For a launch position, select N on the target's predetermined trajectory. t The trajectory points within the maximum interception altitude and maximum range of the interceptor missile are taken as candidate hit points, and the flight time interval between the candidate hit points is T. p seconds; for the mth candidate hit point, m=1,2,...,N t , based on the current launch position and firing table of the interceptor missile, obtain the time t required for the interceptor missile to reach the current candidate hit point am and the terminal velocity v am , compared with the time t when the interceptor missile reaches the current candidate hit point am and the time t for the target to reach the current candidate hit point from the initial position gom , and the terminal velocity v of the interceptor am Minimum terminal speed If formula (I) is satisfied, the mth candidate hit point is considered to be a valid hit point; From all valid hit points, the first valid hit point that meets the emission requirements of formula (I) is selected as the hit point M to participate in the construction of training samples; For different launch positions, the hit point M is determined, and multiple hit points are obtained to participate in the construction of training samples.

Images