Many-to-many unmanned aerial vehicle interception method based on neural network

Abstract

The invention discloses a many-to-many unmanned aerial vehicle interception method based on a neural network, and the method comprises the steps: constructing a threat function according to the position, height, course and speed of the unmanned aerial vehicles of a friend and a foe for the unmanned aerial vehicles in a many-to-many combat scene, constructing a threat matrix through the training weight of the neural network, and distributing an optimal interception target for the unmanned aerial vehicles of our part; and many-to-many unmanned aerial vehicle interception is realized. According to the method, the threat matrix based on neural network weight training is used to realize maximization of interception advantages.

Description

Technical Field

[0001] This invention belongs to the field of autonomous collaborative control and intelligent decision-making for unmanned aerial vehicles (UAVs), and particularly relates to a many-to-many UAV interception method based on neural networks. Background Technology

[0002] Traditional multi-UAV cooperative interception methods often employ rule bases or heuristic strategies, which suffer from poor environmental adaptability, limited decision-making dimensions, and difficulty in handling high-dimensional nonlinear situations. With the increasing scale of UAV swarm operations and the growing complexity of battlefield environments, traditional methods face severe challenges in terms of real-time performance, adversarial intelligence, and decision-making accuracy. Neural networks, possessing powerful nonlinear fitting and high-dimensional data processing capabilities, can learn from historical air combat data or simulated environment data to achieve accurate assessment and intelligent decision-making in complex air combat situations, providing a new technical approach for multi-to-multi UAV cooperative interception. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to address the deficiencies mentioned in the background art by providing a multi-to-multi UAV interception method based on neural networks. This method uses neural networks to achieve real-time threat assessment and maneuver decision-making in complex battlefield situations, thereby improving the overall combat effectiveness of our UAVs.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0005] A neural network-based many-to-many drone interception method includes the following steps:

[0006] Step 1): When the enemy sends out n drones, our side sends out n drones to intercept them.

[0007] Step 2), obtain the position, speed, altitude, and heading of each enemy drone, where the position of the j-th enemy drone is X. Bj = (x Bj y Bj , z Bj ), speed is V j The height is h j The heading angle is C Bj j is a natural number greater than or equal to 1 and less than or equal to n;

[0008] Obtain the position, speed, altitude, and heading of each of our drones, where the position of our i-th drone is X. Ri = (x Ri y Ri , z Ri ), speed is V i The height is h i The heading angle is C Ri, where i is a natural number greater than or equal to 1 and less than or equal to n;

[0009] Obtain the relative distance between any friendly drone and any enemy drone, where the relative distance between the i-th friendly drone and the j-th enemy drone is d. ij ;

[0010] Step 3): For each of our drones, calculate its threat feature vector to each of the enemy drones to form a first set of threat feature vectors; for each of the enemy drones, calculate its threat feature vector to each of our drones to form a second set of threat feature vectors; the threat feature vectors include energy threat value, azimuth threat value, entry angle threat value, and range threat value.

[0011] Step 4) Establish a deep threat assessment model. The deep threat assessment model adopts a neural network model based on a multilayer perceptron architecture and uses mean squared error as the loss function to calculate the corresponding comprehensive threat assessment value based on the threat feature vectors of our UAV to enemy UAV or enemy UAV to our UAV.

[0012] Step 5) Generate several threat feature vectors of our UAVs against enemy UAVs and several threat feature vectors of enemy UAVs against our UAVs through an air combat simulation environment as training data. Then, use an expert system and a method based on combat result backtracking to label the training data with comprehensive threat assessment values ​​to form training samples. Train the deep threat assessment model based on the training samples.

[0013] Step 6): Substitute each vector from the first threat feature vector set and the second threat feature vector set into the trained deep threat assessment model to obtain the threat matrix S of our UAV against the enemy and the threat matrix T of the enemy UAV against us.

[0014] ;

[0015] In the formula, S ij T represents the overall threat assessment value of our i-th UAV against the j-th enemy UAV. ji The overall threat assessment value of the enemy's j-th drone to our i-th drone;

[0016] Step 7) Assign interception targets to our drones based on the threat matrix S and T;

[0017] Step 7.1) Find the element T with the largest overall threat assessment value in the threat matrix T. yp The most threatening enemy drone at present is determined to be the y-th enemy drone, where y and p are both natural numbers greater than or equal to 1 and less than or equal to n;

[0018] Step 7.2) Find the element S with the largest overall threat assessment value in the y-th column of the threat matrix S. xy The drone that poses the greatest threat to the enemy's y-th drone is our x-th drone;

[0019] Step 7.3) Assign the enemy's y-th drone to our x-th drone as the interception target, delete the y-th row of threat matrix T, and delete the x-th row of threat matrix S;

[0020] Step 7.4), repeat steps 7.1) to 7.3) until all our drones are assigned to intercept targets;

[0021] Step 8) Control our drones to intercept enemy drones according to the assigned interception targets.

[0022] As a further optimization of the neural network-based many-to-many UAV interception method of the present invention, in step 3), the energy threat value of our i-th UAV to the j-th enemy UAV... Azimuth Threat Value Threat value at the entrance angle Distance Threat Value The calculation steps are as follows:

[0023] Step 3.1) Calculate the energy threat value according to the following formula. :

[0024] ;

[0025] In the formula, The preset energy sensitivity coefficient has the following physical dimensions: Used to adjust energy difference Energy Threat Value The extent of the impact; The energy difference , , g is the gravitational constant, h i The altitude of the drone;

[0026] Step 3.2) Calculate the angle between the line connecting our i-th UAV and the enemy j-th UAV and due north, and use this angle as the azimuth angle C between our i-th UAV and the enemy j-th UAV. fij ;

[0027] Step 3.3) Calculate the azimuth angle between our i-th UAV and the enemy j-th UAV according to the following formula. and entry angle :

[0028] ;

[0029] ;

[0030] Step 3.4) Calculate the azimuth threat value according to the following formula. and entry angle threat value :

[0031] ;

[0032] Step 3.5), calculate the distance threat value according to the following formula. :

[0033] ;

[0034] In the formula, This is the preset maximum effective threat distance; This is a preset scaling factor used to adjust the relative weight of the distance term in the overall threat vector.

[0035] As a further optimization of the neural network-based many-to-many UAV interception method of the present invention, the deep threat assessment model described in step 4) includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer.

[0036] The input layer is used to receive the threat feature vector X of our drone against the enemy drone or the enemy drone against our drone;

[0037] The first hidden layer and the input layer are fully connected, containing 128 neurons, each configured with a ReLU activation function; the output of the first hidden layer is represented as H1=ReLU(W1·X+b1), where the weight matrix... Weight matrix ;

[0038] The second hidden layer is fully connected to the first hidden layer and contains 64 neurons, each configured with a ReLU activation function; the output of the second hidden layer is represented as H2=ReLU(W2·H1+b2), where the weight matrix... Weight matrix W;

[0039] The third hidden layer and the second hidden layer are fully connected, containing 32 neurons, each configured with a ReLU activation function; the output of the third hidden layer is represented as H3 = ReLU(W3·H2 + b3), where the weight matrix... , ;

[0040] The output layer and the third hidden layer are fully connected and each contains one neuron equipped with a Sigmoid activation function. The final output of the output layer is represented as U = Sigmoid(W₄·H₃ + b₄), where the weight matrix... Weight matrix U is the overall threat assessment value between 0 and 1.

[0041] As a further optimization of the neural network-based many-to-many UAV interception method of the present invention, in step 3) .

[0042] As a further optimization of the neural network-based many-to-many UAV interception method of the present invention, the training samples in step 5) are stored in a structured record format and divided into a training set of 80%, a validation set of 10%, and a test set of 10%.

[0043] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0044] This invention utilizes a pre-trained deep threat assessment model to accurately extract and fuse high-dimensional nonlinear battlefield features, outputting continuous and accurate comprehensive threat values. This overcomes the limitations of traditional methods, which rely on manually designed functions for threat calculation and have limited accuracy. Furthermore, it introduces an enemy-friendly threat matrix and designs an iterative optimal allocation algorithm to achieve rational scheduling of friendly UAV resources and precise matching with high-threat enemy targets. Attached Figure Description

[0045] Figure 1 This is a schematic diagram illustrating the principle and flow of the present invention;

[0046] Figure 2 This is a schematic diagram illustrating the change in the comprehensive threat assessment value during the countermeasures of this invention;

[0047] Figure 3 This is a diagram comparing the trajectories of drones when using different methods of countermeasures. Detailed Implementation

[0048] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0049] This invention can be implemented in many different forms and should not be considered limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully express the scope of the invention to those skilled in the art. In the drawings, components are enlarged for clarity.

[0050] like Figure 1 As shown, this invention discloses a many-to-many UAV interception method based on neural networks, comprising the following steps:

[0051] Step 1): When the enemy sends out n drones, our side sends out n drones to intercept them.

[0052] Step 2), obtain the position, speed, altitude, and heading of each enemy drone, where the position of the j-th enemy drone is X. Bj = (x Bj y Bj , z Bj ), speed is V j The height is h j The heading angle is C Bj j is a natural number greater than or equal to 1 and less than or equal to n;

[0053] Obtain the position, speed, altitude, and heading of each of our drones, where the position of our i-th drone is X. Ri = (x Ri y Ri , z Ri ), speed is V i The height is h i The heading angle is C Ri , where i is a natural number greater than or equal to 1 and less than or equal to n;

[0054] Obtain the relative distance between any friendly drone and any enemy drone, where the relative distance between the i-th friendly drone and the j-th enemy drone is d. ij ;

[0055] Step 3): For each of our drones, calculate its threat feature vector to each of the enemy drones to form a first set of threat feature vectors; for each of the enemy drones, calculate its threat feature vector to each of our drones to form a second set of threat feature vectors; the threat feature vectors include energy threat value, azimuth threat value, entry angle threat value, and range threat value.

[0056] Energy threat value of our i-th drone to the enemy j-th drone Azimuth Threat Value Threat value at the entrance angle Distance Threat Value The calculation steps are as follows:

[0057] Step 3.1) Calculate the energy threat value according to the following formula. :

[0058] ;

[0059] In the formula, The preset energy sensitivity coefficient has the following physical dimensions: Used to adjust energy difference Energy Threat Value The extent of the impact; The energy difference , , g is the gravitational constant, h i The altitude of the drone;

[0060] Step 3.2) Calculate the angle between the line connecting our i-th UAV and the enemy j-th UAV and due north, and use this angle as the azimuth angle C between our i-th UAV and the enemy j-th UAV. fij ;

[0061] Step 3.3) Calculate the azimuth angle between our i-th UAV and the enemy j-th UAV according to the following formula. and entry angle :

[0062] ;

[0063] ;

[0064] Step 3.4) Calculate the azimuth threat value according to the following formula. and entry angle threat value :

[0065] ;

[0066] Step 3.5), calculate the distance threat value according to the following formula. :

[0067] ;

[0068] In the formula, This is the preset maximum effective threat distance; The preset scaling factor is used to adjust the relative weight of the distance term in the comprehensive threat vector; priority is given to... ;

[0069] Step 4) Establish a deep threat assessment model. The deep threat assessment model adopts a neural network model based on a multilayer perceptron architecture and uses mean squared error as the loss function to calculate the corresponding comprehensive threat assessment value based on the threat feature vectors of our UAV to enemy UAV or enemy UAV to our UAV.

[0070] The deep threat assessment model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer.

[0071] The input layer is used to receive the threat feature vector X of our drone against the enemy drone or the enemy drone against our drone;

[0072] The first hidden layer and the input layer are fully connected, containing 128 neurons, each configured with a ReLU activation function; the output of the first hidden layer is represented as H1=ReLU(W1·X+b1), where the weight matrix... Weight matrix ;

[0073] The second hidden layer is fully connected to the first hidden layer and contains 64 neurons, each configured with a ReLU activation function; the output of the second hidden layer is represented as H2=ReLU(W2·H1+b2), where the weight matrix... Weight matrix W;

[0074] The third hidden layer and the second hidden layer are fully connected, containing 32 neurons, each configured with a ReLU activation function; the output of the third hidden layer is represented as H3 = ReLU(W3·H2 + b3), where the weight matrix... , ;

[0075] The output layer and the third hidden layer are fully connected and each contains one neuron equipped with a Sigmoid activation function. The final output of the output layer is represented as U = Sigmoid(W₄·H₃ + b₄), where the weight matrix... Weight matrix U is the overall threat assessment value between 0 and 1;

[0076] Step 5) Generate several threat feature vectors of our UAVs against enemy UAVs and several threat feature vectors of enemy UAVs against our UAVs through an air combat simulation environment as training data. Then, use an expert system and a method based on combat result backtracking to label the training data with comprehensive threat assessment values ​​to form training samples.

[0077] The deep threat assessment model is trained based on training samples;

[0078] The training samples are stored in a structured record format and divided into three sets: 80% training set, 10% validation set, and 10% test set.

[0079] Step 6): Substitute each vector from the first threat feature vector set and the second threat feature vector set into the trained deep threat assessment model to obtain the threat matrix S of our UAV against the enemy and the threat matrix T of the enemy UAV against us.

[0080] ;

[0081] In the formula, S ij T represents the overall threat assessment value of our i-th UAV against the j-th enemy UAV. ji The overall threat assessment value of the enemy's j-th drone to our i-th drone;

[0082] Step 7) Assign interception targets to our drones based on the threat matrix S and T;

[0083] Step 7.1) Find the element T with the largest overall threat assessment value in the threat matrix T. yp The most threatening enemy drone at present is determined to be the y-th enemy drone, where y and p are both natural numbers greater than or equal to 1 and less than or equal to n;

[0084] Step 7.2) Find the element S with the largest overall threat assessment value in the y-th column of the threat matrix S. xy The drone that poses the greatest threat to the enemy's y-th drone is our x-th drone;

[0085] Step 7.3) Assign the enemy's y-th drone to our x-th drone as the interception target, delete the y-th row of threat matrix T, and delete the x-th row of threat matrix S;

[0086] Step 7.4), repeat steps 7.1) to 7.3) until all our drones are assigned to intercept targets;

[0087] Step 8) Control our drones to intercept enemy drones according to the assigned interception targets.

[0088] To verify the effectiveness of this invention, digital simulation analysis was conducted in an air combat simulation. The simulation involved three drones on each side (n=3), for a total of six drones. The friendly drones were numbered R1, R2, and R3, and the enemy drones were numbered B1, B2, and B3. First, based on steps 3 and 4, the basic threat characteristic values ​​and the neural network-comprehensive threat values ​​S and T for all friendly and enemy pairs were calculated to form a threat matrix. Then, using the allocation algorithm in step 5, the optimal interception targets were assigned to the friendly drones. The initial allocation result in this embodiment was: R1 intercepts B2, R2 intercepts B3, and R3 intercepts B1. Figure 3 As shown, compared with the traditional proportional navigation (PN) method, the method of this invention, by introducing situation prediction and decision-making based on neural networks, can generate more agile and advanced flanking and interception paths, effectively compressing the enemy's maneuver space and demonstrating higher intelligence and adaptability.

[0089] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0090] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A many-to-many UAV interception method based on neural networks, characterized in that, Includes the following steps: Step 1): When the enemy sends out n drones, our side sends out n drones to intercept them. Step 2), obtain the position, speed, altitude, and heading of each enemy drone, where the position of the j-th enemy drone is X. Bj = (x Bj y Bj , z Bj ), speed is V j The height is h j The heading angle is C Bj j is a natural number greater than or equal to 1 and less than or equal to n; Obtain the position, speed, altitude, and heading of each of our drones, where the position of our i-th drone is X. Ri = (x Ri y Ri , z Ri ), speed is V i The height is h i The heading angle is C Ri i is a natural number greater than or equal to 1 and less than or equal to n; Obtain the relative distance between any friendly drone and any enemy drone, where the relative distance between the i-th friendly drone and the j-th enemy drone is d. ij ; Step 3): For each of our drones, calculate its threat feature vector to each of the enemy drones to form a first set of threat feature vectors; for each of the enemy drones, calculate its threat feature vector to each of our drones to form a second set of threat feature vectors; the threat feature vectors include energy threat value, azimuth threat value, entry angle threat value, and range threat value. Step 4) Establish a deep threat assessment model. The deep threat assessment model adopts a neural network model based on a multilayer perceptron architecture and uses mean squared error as the loss function to calculate the corresponding comprehensive threat assessment value based on the threat feature vectors of our UAV to enemy UAV or enemy UAV to our UAV. Step 5) Generate several threat feature vectors of our UAVs against enemy UAVs and several threat feature vectors of enemy UAVs against our UAVs through an air combat simulation environment as training data. Then, use an expert system and a method based on combat result backtracking to label the training data with comprehensive threat assessment values ​​to form training samples. Train the deep threat assessment model based on the training samples. Step 6): Substitute each vector from the first threat feature vector set and the second threat feature vector set into the trained deep threat assessment model to obtain the threat matrix S of our UAV against the enemy and the threat matrix T of the enemy UAV against us. ； In the formula, S ij T represents the overall threat assessment value of our i-th UAV against the j-th enemy UAV. ji The overall threat assessment value of the enemy's j-th drone to our i-th drone; Step 7) Assign interception targets to our drones based on the threat matrix S and T; Step 7.1) Find the element T with the largest overall threat assessment value in the threat matrix T. yp The most threatening enemy drone at present is determined to be the y-th enemy drone, where y and p are both natural numbers greater than or equal to 1 and less than or equal to n; Step 7.2) Find the element S with the largest overall threat assessment value in the y-th column of the threat matrix S. xy The drone that poses the greatest threat to the enemy's y-th drone is our x-th drone; Step 7.3) Assign the enemy's y-th drone to our x-th drone as the interception target, delete the y-th row of threat matrix T, and delete the x-th row of threat matrix S; Step 7.4), repeat steps 7.1) to 7.3) until all our drones are assigned to intercept targets; Step 8) Control our drones to intercept enemy drones according to the assigned interception targets.

2. The many-to-many UAV interception method based on neural networks as described in claim 1, characterized in that, Step 3) Energy threat value of our i-th UAV to the j-th enemy UAV Azimuth Threat Value Threat value at the entrance angle Distance Threat Value The calculation steps are as follows: Step 3.1) Calculate the energy threat value according to the following formula. : ； In the formula, The preset energy sensitivity coefficient has the following physical dimensions: Used to adjust energy difference Energy Threat Value The extent of the impact; The energy difference , , g is the gravitational constant, h i The altitude of the drone; Step 3.2) Calculate the angle between the line connecting our i-th UAV and the enemy j-th UAV and due north, and use this angle as the azimuth angle C between our i-th UAV and the enemy j-th UAV. fij ; Step 3.3) Calculate the azimuth angle between our i-th UAV and the enemy j-th UAV according to the following formula. and entry angle : ； ； Step 3.4) Calculate the azimuth threat value according to the following formula. and entry angle threat value : ； Step 3.5), calculate the distance threat value according to the following formula. : ； In the formula, This is the preset maximum effective threat distance; This is a preset scaling factor used to adjust the relative weight of the distance term in the overall threat vector.

3. The many-to-many UAV interception method based on neural networks as described in claim 1, characterized in that, The deep threat assessment model described in step 4) includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer; The input layer is used to receive the threat feature vector X of our drone against the enemy drone or the enemy drone against our drone; The first hidden layer and the input layer are fully connected, containing 128 neurons, each configured with a ReLU activation function; the output of the first hidden layer is represented as H1=ReLU(W1·X+b1), where the weight matrix... Weight matrix ; The second hidden layer is fully connected to the first hidden layer and contains 64 neurons, each configured with a ReLU activation function; the output of the second hidden layer is represented as H2=ReLU(W2·H1+b2), where the weight matrix... Weight matrix W; The third hidden layer and the second hidden layer are fully connected, containing 32 neurons, each configured with a ReLU activation function; the output of the third hidden layer is represented as H3 = ReLU(W3·H2 + b3), where the weight matrix... , ; The output layer and the third hidden layer are fully connected and each contains one neuron equipped with a Sigmoid activation function. The final output of the output layer is represented as U = Sigmoid(W₄·H₃ + b₄), where the weight matrix... Weight matrix U is the overall threat assessment value between 0 and 1.

4. The neural network-based many-to-many UAV interception method according to claim 2, characterized in that, Step 3) .

5. The many-to-many UAV interception method based on neural networks as described in claim 1, characterized in that, Step 5) The training samples are stored in a structured record format and divided into three sets: 80% training set, 10% validation set, and 10% test set.

Images