A method for identifying individual attack intent in a swarm of bullets based on neural networks.
By acquiring the relative characteristics of individual missiles and missile swarms and inputting them into a trained neural network, the system identifies individual missiles with attack intentions within the missile swarm, solving the problem of not being able to identify individual intentions in existing technologies and improving the interception efficiency of air defense systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-23
- Publication Date
- 2026-04-03
AI Technical Summary
Existing neural network-based methods cannot effectively identify the attack intent of each individual missile in a missile swarm, which increases the difficulty of interception in air defense operations.
By acquiring the position and velocity information of individual missiles, calculating their relative characteristics with the missile swarm, and then performing dimensionality reduction and normalization processing, the data is input into a trained neural network for identification, thereby identifying individual missiles with attack intentions.
It enables targeted identification of individual missiles within a missile swarm, improving the interception efficiency of the air defense system, with an identification accuracy rate of 96.60%.
Smart Images
Figure CN115270994B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intent recognition technology, and particularly relates to a method for identifying individual attack intent in a swarm of bullets based on neural networks. Background Technology
[0002] With the rapid advancement of air defense technology, the attack capability of a single missile has become limited. Therefore, researchers worldwide are conducting research on "swarm" missiles capable of coordinated operations, such as the US "Netfire" missile. Swarm attacks make battlefield defense more complex and difficult, and how to intercept swarms in air defense operations has become a research hotspot. If the attack intentions of individual missiles within an incoming swarm can be effectively analyzed and a rapid and effective threat assessment can be conducted, it will be helpful to target and strike the most threatening individuals.
[0003] Currently, commonly used intent recognition methods include evidence reasoning and neural networks. Neural network-based analysis methods are data-driven intelligent algorithms that learn autonomously from large amounts of data, avoiding excessive human intervention in the entire intent recognition process. They have good self-organizing capabilities and the ability to represent the acquired information in complex situations.
[0004] However, existing neural network-based methods only consider identifying the overall intent of all input objects. In missile swarm defense, identifying the intent of individual missiles within the swarm is more important than identifying the intent of the entire swarm. Furthermore, battlefield situations are complex and constantly changing, and the size of incoming missile swarms is never fixed. Therefore, how to determine whether individual missiles within a swarm possess an attack intent, in order to improve defensive capabilities, is a pressing issue that needs to be addressed. Summary of the Invention
[0005] To address the aforementioned shortcomings in the prior art, the present invention provides a neural network-based method for identifying individual attack intentions in a missile swarm, which solves the problem that existing methods cannot identify the intentions of each individual missile in a swarm.
[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:
[0007] A method for identifying individual attack intent in a swarm of projectiles based on a neural network is provided, comprising:
[0008] S1. Obtain the position and velocity characteristics of individual missiles in the missile swarm;
[0009] S2. Perform dimensionality reduction on the positional features of all individual missiles in the missile group to obtain the positional features of the missile group.
[0010] S3. Based on the positional characteristics of the missile group and the positional characteristics and velocity information of individual missiles, obtain the relative characteristics of individual missiles and the missile group, and perform normalization processing.
[0011] S4. After splicing the normalized data of all missiles in the missile group at multiple consecutive moments, input the data into the trained neural network for identification, and identify the missiles in the missile group that have the intention to attack.
[0012] Furthermore, methods for obtaining the relative characteristics of individual missiles and missile groups include:
[0013] S31. Based on the positional characteristics of the missile group and the positional characteristics of the individual missiles, calculate the rate of change of the relative distance between the individual missiles and the missile group at each moment;
[0014]
[0015] in, For individual missile i in a missile swarm, in t j The rate of change of relative distance at any given time; and t j Time and t j-1 The distance of missile i relative to the missile group at any given moment; and These represent the missile group and the individual missile i within the missile group, respectively, at t. j Locational characteristics at any given time;
[0016] S32. Based on the velocity information of the individual missiles and the position information of the missile group, calculate the angle between the velocity vector of the individual missile and the line of sight of the center of the missile group:
[0017]
[0018] in, For t j The angle between the velocity vector of missile individual i and the line of sight of the missile group center at any given moment; For t j The velocity vector of missile individual i at any given moment;
[0019] S33, Calculate t j The velocity of each individual missile i at any given time
[0020]
[0021] S34. Using the rate of change of relative distance Angle of eye line and speed As t j The relative characteristics of individual missile i and missile group at any given moment.
[0022] Furthermore, the formula for obtaining the positional features of the missile swarm through dimensionality reduction is as follows:
[0023]
[0024] in, For the missile group in t j Positional characteristics at any given moment; m is the total number of individual missiles in the missile swarm; and These are missiles 1 and m in the missile swarm, respectively, at t j Locational characteristics at any given moment.
[0025] Furthermore, the training method for the neural network includes:
[0026] A1. Generate trajectory data of individual missiles during the normal flight and attack phases of the missile swarm, and execute steps S2 to S3 to form the original dataset by normalizing the data from multiple consecutive moments of all missiles.
[0027] A2. Add relative features of normal missile individuals to the original dataset to make the neural network output 0, and make the data of normal missile individuals equal to the data of attacking missile individuals to form a training dataset.
[0028] A3. Label the data in the training dataset. Set the expected value of attacking missile individuals to 1, and the expected value of normal missile individuals to 0;
[0029] A4. Input the training dataset into the neural network, and calculate the loss function based on the actual output and expected output of the neural network;
[0030] A5. Based on the loss function, the network parameters are updated using gradient descent with momentum and adaptive learning rate. When the update meets the termination condition, the network parameter update is terminated, and the trained neural network is obtained.
[0031] Furthermore, the formula for calculating the adaptive learning rate is:
[0032]
[0033] Where α(k+1) and α(k) are the learning rates at the (k+1)th and (k)th iterations, respectively; E(k+1) and E(k) are the mean squared errors at the (k+1)th and (k)th iterations, respectively.
[0034] Furthermore, the termination condition is that the mean square error reaches a set value, the maximum number of training iterations is reached, or the parameters are not updated for multiple consecutive times.
[0035] Furthermore, methods for generating trajectory data for individual missiles during the normal flight and attack phases of a missile swarm include:
[0036] A11. Use follower-leader, artificial potential field, or consensus formation algorithms to simulate and generate trajectory data for each missile in the flight phase of the missile swarm;
[0037] A12. When the missile swarm attacks, a missile swarm guidance algorithm based on proportional guidance law is used to generate the attack trajectory of each individual missile.
[0038] Furthermore, the calculation formula for the formation algorithm of the artificial potential field is as follows:
[0039]
[0040]
[0041] Among them, U rep (q) and U att (q) represents the repulsive potential field and the gravitational potential field, respectively; η and ξ are the constants of the repulsive potential field and the gravitational potential field, respectively; q represents the position information of the individual missile; q obs ρ represents the positional information of adjacent missiles or obstacles; ρ0 represents the range of the repulsive field; ρ(.) is the Euclidean distance calculation function.
[0042] The formula for calculating the consistency of the grouping algorithm is:
[0043]
[0044] Where m is the total number of individual missiles in the missile swarm; u i For the input of missile individual i, u i This ensures that missile individual i and all missile individuals j that have a communication relationship with missile individual i reach a consensus; e ij (k) represents the communication status between missile individual i and missile individual j at time k, e ij =1 indicates that there is communication, e ij =0 indicates no communication;
[0045] The calculation formula for the missile swarm guidance algorithm based on the proportional guidance law is as follows:
[0046]
[0047] Among them, a i Let v be the normal acceleration of missile individual i, N be a constant, and v be the normal acceleration of missile individual i. i Let i be the velocity of missile individual. Let be the rate of change of the angle between the line of sight of missile individual i and the baseline.
[0048] Furthermore, the neural network is a fully connected neural network with an input layer having an input data dimension of 9, 5 hidden layers with 20, 16, 12, 8, and 8 neurons in each layer, and an output layer having an output data dimension of 1.
[0049] The beneficial effects of this invention are as follows: This solution obtains the position and velocity information of each individual missile in the missile group, and based on this, obtains the position characteristics of each individual missile relative to the missile group. Then, based on the connected position characteristics, a neural network is used to identify the intention of the individual missile. This breaks through the limitation of traditional methods that can only identify the overall intention of air targets. Targeted identification of individual missiles is conducive to quickly distinguishing individuals with attack capabilities in the missile group, which facilitates the efficient execution of interception missions by the subsequent air defense system. Attached Figure Description
[0050] Figure 1 This is a flowchart of a method for identifying individual attack intentions in a swarm of projectiles based on neural networks.
[0051] Figure 2 This refers to the relative motion between the missile and the target.
[0052] Figure 3 This is a schematic diagram of the structure of the fully connected neural network model in this scheme.
[0053] Figure 4 The graph shows the change in the loss function during network training for this scheme.
[0054] Figure 5 The confusion matrix obtained when verifying the recognition effect of this scheme.
[0055] Figure 6 This diagram illustrates the individual attack intent identification method used in this scheme for incoming missile swarms. Detailed Implementation
[0056] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0057] refer to Figure 1 , Figure 1 A flowchart illustrating a method for identifying individual attack intent in a swarm of projectiles based on neural networks is shown, such as... Figure 1 As shown, the method S includes steps S1 to S4.
[0058] In step S1, the position and velocity characteristics of individual missiles in the missile swarm are obtained; Figure 2 The relative motion between the missile and the target is shown.
[0059] In step S2, the positional features of all individual missiles in the missile swarm are reduced in dimensionality to obtain the positional features of the missile swarm:
[0060]
[0061] in, For the missile group in t j Positional characteristics at any given moment; m is the total number of individual missiles in the missile swarm; and These are missiles 1 and m in the missile swarm, respectively, at t j Locational characteristics at any given moment.
[0062] In step S3, based on the positional characteristics of the missile group and the positional characteristics and velocity information of individual missiles, the relative characteristics of individual missiles and the missile group are obtained and normalized.
[0063] In one embodiment of the present invention, the method for obtaining the relative characteristics of an individual missile and a missile swarm includes:
[0064] S31. Based on the positional characteristics of the missile group and the positional characteristics of the individual missiles, calculate the rate of change of the relative distance between the individual missiles and the missile group at each moment;
[0065]
[0066] in, For individual missile i in a missile swarm, in t j The rate of change of relative distance at any given time; and t j Time and t j-1 The distance of missile i relative to the missile group at any given moment; and These represent the missile group and the individual missile i within the missile group, respectively, at t. j Locational characteristics at any given time;
[0067] S32. Based on the velocity information of the individual missiles and the position information of the missile group, calculate the angle between the velocity vector of the individual missile and the line of sight of the center of the missile group:
[0068]
[0069] in, For t j The angle between the velocity vector of missile individual i and the line of sight of the missile group center at any given moment; For t j The velocity vector of missile individual i at any given moment;
[0070] S33, Calculate t j The velocity of each individual missile i at any given time
[0071]
[0072] S34. Using the rate of change of relative distance Angle of eye line and speed As t j The relative characteristics of individual missile i and missile group at any given moment.
[0073] In implementation, the preferred normalization formula for this scheme is:
[0074]
[0075] in, It is t j The relative characteristics of individual missile i and the missile group at any given moment; For set The maximum value in; For set The minimum value in.
[0076] In step S4, the normalized data of all missiles in the missile swarm at multiple consecutive moments are spliced together and input into the trained neural network for identification, thereby identifying missiles in the missile swarm that have the intention to attack.
[0077] Accurate identification of attack intent relies on the temporal information of the target. Inputting information from only a single moment would greatly complicate the identification process. Therefore, this scheme uses multi-moment data as the input to the neural network, i.e., features from three moments. The data are then concatenated to form the final input of the neural network.
[0078] The neural network in this scheme is preferably a fully connected neural network, and its network structure is referenced from... Figure 3 The input layer has a data dimension of 9, the hidden layers have 5 layers, and the number of neurons in each layer is 20, 16, 12, 8, and 8 respectively. The output layer has a data dimension of 1.
[0079] In one embodiment of the present invention, the training method for the neural network includes:
[0080] A1. Generate trajectory data of individual missiles during the normal flight and attack phases of the missile swarm, and execute steps S2 to S3 to form the original dataset by normalizing the data from multiple consecutive moments of all missiles.
[0081] In implementation, the preferred method for generating trajectory data of individual missiles during the normal flight and attack phases of the missile swarm includes steps A11 and A12:
[0082] In step A11, follower-leader, artificial potential field, or consensus formation algorithms are used to simulate and generate trajectory data for each missile in the flight phase of the missile swarm.
[0083] The implementation process of the leader-follower missile swarm formation algorithm is as follows: Assuming that an individual in the missile swarm formation is the leader and the other individuals are followers, the algorithm needs to control the movement of the leader. This scheme uses relative distance and relative angle to represent the relative position, that is, the relative position between the leader and the followers.
[0084] The formula for calculating the formation algorithm of the artificial potential field is:
[0085]
[0086]
[0087] Among them, U rep (q) and U att (q) represents the repulsive potential field and the gravitational potential field, respectively; η and ξ are the constants of the repulsive potential field and the gravitational potential field, respectively; q represents the position information of the individual missile; q obs ρ represents the positional information of adjacent missiles or obstacles; ρ0 represents the range of the repulsive field; ρ(.) is the Euclidean distance calculation function.
[0088] The formula for calculating the consistency of the grouping algorithm is:
[0089]
[0090] Where m is the total number of individual missiles in the missile swarm; u i For the input of missile individual i, u i This ensures that missile individual i and all missile individuals j that have a communication relationship with missile individual i reach a consensus; e ij (k) represents the communication status between missile individual i and missile individual j at time k, e ij =1 indicates that there is communication, e ij =0 indicates no communication;
[0091] In step A12, when the missile swarm attacks, a missile swarm guidance algorithm based on proportional guidance law is used to generate the attack trajectory of each individual missile.
[0092] The calculation formula for the missile swarm guidance algorithm based on the proportional guidance law is as follows:
[0093]
[0094] Among them, a i Let v be the normal acceleration of missile individual i, N be a constant, and v be the normal acceleration of missile individual i. i Let i be the velocity of missile individual. Let be the rate of change of the angle between the line of sight of missile individual i and the baseline.
[0095] A2. Add relative features of normal missile individuals to the original dataset to make the neural network output 0, and make the data of normal missile individuals equal to the data of attacking missile individuals to form a training dataset.
[0096] This solution uses the above method to achieve a balanced distribution of samples in the training dataset. This avoids the situation where the network is unable to learn how to identify the attack intentions of individuals in a swarm due to the limited number of attack data.
[0097] A3. Label the data in the training dataset. Set the expected value of attacking missile individuals to 1, and the expected value of normal missile individuals to 0;
[0098] The output of the hidden layer of the neural network is obtained through formula (1), where the function f is the logsigmoid activation function, as shown in formula (2). The output layer uses the same activation function and outputs the recognition result y of the attack intent of missile individual i. i The output range is 0 to 1.
[0099] h i (x)=f(x T w i +b i )……(1)
[0100] A4. Input the training dataset into the neural network, and calculate the loss function based on the actual output and expected output of the neural network;
[0101]
[0102] in, y represents the expected output of the neural network. i This is the time output of the neural network;
[0103] A5. Based on the loss function, the network parameters are updated using gradient descent with momentum and adaptive learning rate. When the update meets the termination condition (the termination condition is that the mean square error reaches a set value, the maximum number of training iterations is reached, or the parameters have not been updated for several consecutive times), the network parameter update is terminated, and the trained neural network is obtained.
[0104] The momentum-based weight and bias update process is shown below:
[0105] v dW =β·v dw +(1-β)·W
[0106] v db =β·v db +(1-β)·db,W new =W old -α·v dW b new =v old -α·v db
[0107] In the formula, dW and db are the partial derivatives of the weights and biases, respectively, and α and β are hyperparameters that need to be set manually.
[0108] An adaptive learning rate can accelerate the overall pace of network parameter updates. At time t, when the degree of agreement between the actual output and the expected output, E(t), increases compared to the previous time step, the learning rate α is decreased; when the degree of agreement decreases compared to the previous time step, the learning rate α is increased. The formula for calculating the adaptive learning rate is:
[0109]
[0110] Where α(k+1) and α(k) are the learning rates at the (k+1)th and (k)th iterations, respectively; E(k+1) and E(k) are the mean squared errors at the (k+1)th and (k)th iterations, respectively.
[0111] The following is combined Figures 2-5 To illustrate the effectiveness of the missile individual attack intent identification method in this scheme:
[0112] The network is trained using different missile swarm formation algorithms and missile swarm formation motion patterns. The datasets used for network training exhibit the following diversity:
[0113] (1) The size of the missile swarm formation varies, ranging from 6 to 9; (2) The number of individuals with attack capabilities in the missile swarm formation varies, ranging from 1 to 3; (3) The initial positions of the missile swarm formation vary.
[0114] Using steps A1-A2 of this scheme, 121,666 training data points are generated, with 50% representing data from individuals with attack intent and 50% representing data from individuals without attack intent. These data are then used for network training. The change in the loss function during training is shown below. Figure 4 As shown. To verify the generalization ability of the current network, the trained network was used to identify individual attackers within newly generated missile swarm data. The overall identification success rate on the test set reached 96.60%.
[0115] To further observe the network's intent recognition performance, 11,976 test data points were generated, with each group having 50% of the data points containing individuals with malicious intent and 50% containing individuals without malicious intent. Experiments were conducted to obtain the recognition confusion matrix for this dataset, with the diagonal representing the recognition accuracy. (See details below.) Figure 5 As shown. By Figure 5 It can be seen that the accuracy rate of this method in identifying individuals with attack intent can reach 97.29%, which is higher than the accuracy rate of 95.92% in identifying normal individuals. The main reason is that attacking individuals have movement characteristics that are significantly different from missile formations, and can be more accurately identified by the network.
[0116] The experimental results also show that identifying normal individuals is more likely to result in misclassification than identifying attacking individuals. This is because the calculation of missile formation characteristics includes attacking individuals. The unusual maneuvers of attacking individuals will be reflected in the movement characteristics of the missile formation, which will make normal individuals have some movement characteristics that are different from those of the missile formation, thus making them more likely to be misclassified.
[0117] When using this method to identify missile groups, such as Figure 6 In the area indicated by the dashed line, when no individual missile is found to have an attack motive, the attack missile and the decoy missile exhibit the same dynamic characteristics, and the algorithm uniformly identifies them as decoys; when the attack missile maneuvers, such as Figure 6 In the solid line section, the algorithm of this scheme accurately distinguishes between attack missiles and decoy missiles.
[0118] This solution abandons the strategy of using the kinematic characteristics of individual missiles as input data for the identification network. Instead, it integrates missile swarm information and inputs the relative characteristics between individual missiles and the swarm. Ultimately, it achieves the task of identifying the attack intentions of individuals in a missile swarm with varying scales, with an accuracy rate of up to 96.60%.
Claims
1. A method for identifying individual attack intent in a swarm of projectiles based on neural networks, characterized in that, include: S1. Obtain the position and velocity characteristics of individual missiles in the missile swarm; S2. Perform dimensionality reduction on the positional features of all individual missiles in the missile group to obtain the positional features of the missile group. S3. Based on the positional characteristics of the missile group and the positional characteristics and velocity information of individual missiles, obtain the relative characteristics of individual missiles and the missile group, and perform normalization processing; the methods for obtaining the relative characteristics of individual missiles and the missile group include: S31. Based on the positional characteristics of the missile group and the positional characteristics of the individual missiles, calculate the rate of change of the relative distance between the individual missiles and the missile group at each moment; ， in, For individual missiles in a missile swarm i exist The rate of change of relative distance at any given time; and They are respectively Time and Individual missiles in a swarm i The distance relative to the missile salvo; , , )and( , , ( ) represent the missile group and individual missiles within the missile group, respectively. i exist Locational characteristics at any given time; S32. Based on the velocity information of the individual missiles and the position information of the missile group, calculate the angle between the velocity vector of the individual missile and the line of sight of the center of the missile group: in, for individual missiles at all times i The angle between the velocity vector and the line of sight at the center of the missile group; for individual missiles at all times i The velocity vector; S33, Calculation Every missile individual i speed : ； S34. Using the rate of change of relative distance Angle between eyes and line of sight and speed As individual missiles at all times i The relative characteristics of the missile swarm; S4. After splicing the normalized data of all missiles in the missile group at multiple consecutive moments, input the data into the trained neural network for identification, and identify the missiles in the missile group that have the intention to attack.
2. The method for identifying individual attack intent in a swarm of projectiles according to claim 1, characterized in that, The formula for obtaining the positional features of the missile swarm through dimensionality reduction is: ， ， in,( , , ) for the missile group in Locational characteristics at any given time; m The total number of individual missiles in the missile swarm; , , )and( , , ( ) are individual missiles 1 and 2 in the missile swarm. m exist Locational characteristics at any given moment.
3. The method for identifying individual attack intent in a swarm of projectiles according to claim 1, characterized in that, The training method for the neural network includes: A1. Generate trajectory data of individual missiles during the normal flight and attack phases of the missile swarm, and execute steps S2~S3 to form the original dataset by normalizing the data from multiple consecutive moments of all missiles. A2. Add relative features of normal missile individuals to the original dataset to make the neural network output 0, and make the data of normal missile individuals equal to the data of attacking missile individuals to form a training dataset. A3. Label the data in the training dataset. The expected value of attacking missile individuals is set to 1, and the expected value of normal missile individuals is set to 0. A4. Input the training dataset into the neural network, and calculate the loss function based on the actual output and expected output of the neural network; A5. Based on the loss function, the network parameters are updated using gradient descent with momentum and adaptive learning rate. When the update meets the termination condition, the network parameter update is terminated, and the trained neural network is obtained.
4. The method for identifying individual attack intent in a swarm of projectiles according to claim 3, characterized in that, The formula for calculating the adaptive learning rate is: in, and The first k +1 iterations and the k Learning rate at the next iteration; and The first k +1 iterations and the k Mean squared error at the next iteration.
5. The method for identifying individual attack intent in a swarm of projectiles according to claim 3, characterized in that, The termination conditions are: the mean square error reaches a set value, the maximum number of training iterations is reached, or the parameters are not updated for several consecutive times.
6. The method for identifying individual attack intent in a swarm of projectiles according to claim 3, characterized in that, Methods for generating trajectory data for individual missiles during the normal flight and attack phases of a missile swarm include: A11. Use follower-leader, artificial potential field, or consensus formation algorithms to simulate and generate trajectory data for each missile in the flight phase of the missile swarm; A12. When the missile swarm attacks, a missile swarm guidance algorithm based on proportional guidance law is used to generate the attack trajectory of each individual missile.
7. The method for identifying individual attack intent in a swarm of projectiles according to claim 6, characterized in that, The formula for calculating the formation algorithm of the artificial potential field is: in, and These are the repulsive potential field and the gravitational potential field, respectively; These are the repulsive potential field constant and the gravitational potential field constant, respectively; For the location information of individual missiles; This refers to the positional information of adjacent missiles or obstacles. The range of the repulsive field is set; The function for solving Euclidean distance; The formula for calculating the consistency of the grouping algorithm is: in, The total number of individual missiles in the missile swarm; For individual missiles i Input; for individual missiles at all times i and individual missiles j The communication status, e ij =1 indicates that there is communication. e ij =0 indicates no communication; The calculation formula for the missile swarm guidance algorithm based on the proportional guidance law is as follows: in, For individual missiles i normal acceleration, It is a constant. For individual missiles i speed, For individual missiles i The rate of change of the angle between the line of sight and the baseline.
8. The method for identifying individual attack intent in a swarm of projectiles according to any one of claims 1-7, characterized in that, The neural network is a fully connected neural network with an input layer having an input data dimension of 9, 5 hidden layers with 20, 16, 12, 8, and 8 neurons in each layer, and an output layer having an output data dimension of 1.