An Autonomous Control Method for Unmanned Aerial Vehicle Swarms Based on VGG16 and Virtual Game Theory

Abstract

This invention provides an autonomous control method for a drone swarm based on VGG16 and virtual game theory. The drones utilize a pre-trained attitude recognition model to identify the historical movement information of their opponents and calculate the opponent's strategy tendency at the previous moment based on this information. Then, the drones select a candidate action from a set of potential actions, calculating the opponent's strategy tendency and probability of choosing that action according to the game theory formula. A reward function is designed, and the drones select their own action based on the obtained opponent's strategy probability using the optimal response decision rule, observing the changes in the opponent's actions until the entire drone swarm reaches a stable state. This enables the analysis and game theory of multiple surrounding drones. When encountering signal interference, this invention allows some drone units to malfunction while the drone itself can adjust its movement attitude, maintaining relative stability for the entire swarm.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically relating to an autonomous control method for UAV swarms based on VGG16 and virtual game theory. Background Technology

[0002] In recent years, due to the continuous development of drone technology, drones have been used to collaboratively perform various complex tasks, including search and rescue, and security patrols. Drone swarms are widely used in military, flood control and disaster relief, earthquake relief, and daily life scenarios. However, the widespread application of drone swarms has also brought about challenges in drone swarm control. On the one hand, drones are increasingly being used in harsh and complex environments where signals are often unstable. On the other hand, even localized signal interference can cause some drones to lose control, leading to crashes. To avoid this, drones need to be able to autonomously control their flight even when briefly losing control due to signal interference.

[0003] Currently, drone security defense technologies are limited to single-domain drone security control. While these technologies can effectively detect and prevent certain abnormal behaviors in drone swarms, a single security defense is insufficient to guarantee the normal operation of a drone swarm in the face of the diverse means that can cause such abnormal behavior. This necessitates that the drone swarm maintain normal operation even after abnormal behavior occurs. Therefore, combining security defense with drone swarm autonomous control technology enables the drone swarm to maintain autonomous control even in the face of signal interference or abnormal behavior. However, traditional drone swarm control technologies suffer from slow feedback and cannot respond promptly to the needs of drone swarm missions. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides an autonomous control method for unmanned aerial vehicle (UAV) swarms based on VGG16 and virtual game theory. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] This invention provides an autonomous control method for drone swarms based on VGG16 and virtual game theory, applicable to drone swarms, where any drone in the swarm is a player and the other drones are players. The method includes:

[0006] Step 1: Each player obtains the motion image of their opponent from the previous moment.

[0007] Step 2: After preprocessing the motion image, it is fed into the trained pose recognition model to identify the opponent's movement direction in the previous moment.

[0008] Step 3: Obtain a pre-designed set of candidate actions for the drone;

[0009] Step 4: Based on the opponent's strategic inclinations and movement directions for all available actions in the previous moment, calculate the opponent's strategic inclinations for all available actions in the current moment.

[0010] Step 5: Based on the strategy tendencies of each player in all candidate actions at the current moment, calculate the joint probability of all players choosing a certain candidate action at the current moment, and convert the joint probability into the most likely action to be performed by all players at the current moment.

[0011] Step 6: Design a reward function based on the action chosen by all opponents at the current moment that maximizes the reward function.

[0012] Step 7: The player flies according to their own actions and continues to observe the changes in their opponent's actions;

[0013] Step 8: Repeat steps 1 through 7 until the drone swarm stabilizes.

[0014] The beneficial effects of this invention are:

[0015] (1) This invention addresses the attack methods of unstable drone signals and signal interference in harsh scenarios. After the drone swarm loses control, it can fly autonomously. At the same time, each drone can perform independent calculations and decisions, which has good autonomy and protects individual privacy.

[0016] (2) This invention uses the VGG16 network as the attitude recognition model, which has the advantages of small recognition limitations, good transferability, and low computational resource consumption. It can realize real-time recognition of UAV attitudes on a large scale, which provides a basis for faster action selection in the autonomous flight control module of UAV based on extended Kalman filter virtual game. In addition, the dataset uses UAV flight attitudes in multiple scenarios, so the flight attitude recognition model based on the VGG16 network has good recognition accuracy in many scenarios.

[0017] (3) The method of the present invention enables the UAV to observe the movement status of the opponent in real time and continuously iterate the model. Therefore, in the case of signal interference, the UAV can make rapid dynamic adjustments through the autonomous control module.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the autonomous flight control model framework for unmanned aerial vehicles provided by the present invention;

[0020] Figure 2 This is a flowchart illustrating an autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory provided by the present invention.

[0021] Figure 3 This is a schematic diagram of a simulated autonomous flight scenario of 30 drones encountering interference, provided by the present invention.

[0022] Figure 4 This is a schematic diagram of the loss function of the VGG16 network provided by the present invention;

[0023] Figure 5 This is a schematic diagram illustrating the accuracy of the VGG16 network provided by this invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0025] This invention pre-collects drone attitude datasets from numerous scenarios and trains a model within a VGG16 network. During flight, the drone's camera collects attitude datasets of other drones in the cluster. The pre-processed datasets are then fed into the VGG16 network for attitude recognition. The recognition results are transmitted to the drone's autonomous flight control module. Based on analysis and interaction with multiple surrounding drones, even if signal interference causes some drone units to malfunction, the malfunctioning unit can adjust its movement attitude, maintaining relative stability for the entire cluster.

[0026] This invention mainly consists of two parts: a UAV swarm attitude recognition module based on a VGG16 network and a UAV swarm autonomous flight control module based on extended Kalman filter virtual game theory. The UAV swarm attitude recognition module is responsible for information collection and recognition, while the UAV autonomous flight control module is responsible for the autonomous control of the UAVs after encountering interference or attacks. The UAV autonomous flight model framework is as follows: Figure 1 As shown.

[0027] Example 1

[0028] This invention provides an autonomous control method for drone swarms based on VGG16 and virtual game theory, applicable to drone swarms, where any drone in the swarm is a player and other neighboring drones are players, including:

[0029] Step 1: Each player obtains the motion image of their opponent from the previous moment.

[0030] Step 2: After preprocessing the motion image, it is fed into the trained pose recognition model to identify the opponent's movement direction in the previous moment.

[0031] This invention uses the movement direction of the previous moment as historical movement information, preprocesses it into contrast transformation, and then sends the contrast-transformed motion image into the pose recognition model.

[0032] Steps 1 and 2 of this invention are completed by a drone swarm attitude recognition module based on a VGG16 network. (Drone) Record its opponent during the move. That is, the direction of movement of nearby drones is recorded and saved.

[0033] Step 3: Obtain a pre-designed set of candidate actions for the drone;

[0034] Step 4: Based on the opponent's strategic inclinations and movement directions for all available actions in the previous moment, calculate the opponent's strategic inclinations for all available actions in the current moment.

[0035] Step 5: Based on the strategy tendencies of each player in all candidate actions at the current moment, calculate the joint probability of all players choosing a certain candidate action at the current moment, and convert the joint probability into the most likely action to be performed by all players at the current moment.

[0036] Step 6: Design a reward function based on the action chosen by all opponents at the current moment that maximizes the reward function.

[0037] Step 7: The player flies according to their own actions and continues to observe the changes in their opponent's actions;

[0038] Step 8: Repeat steps 1 through 7 until the drone swarm stabilizes.

[0039] In this invention, the players Repeat steps 1-7 until the drone swarm reaches a stable state, and obtain the distribution of all possible action strategies of all opponents at the current moment.

[0040] Steps 3 to 8 of this invention are completed by the UAV swarm autonomous flight control module based on extended Kalman filter virtual game theory. For example... Figure 3 As shown, this module uses Repast-Simphony simulation software to construct 30 UAVs on a two-dimensional plane, and tests and implements an autonomous flight control module for UAV swarm based on virtual game theory using extended Kalman filters.

[0041] This invention provides an autonomous control method for a drone swarm based on VGG16 and virtual game theory. The drones utilize a pre-trained attitude recognition model to identify the historical movement information of their opponents and calculate the opponent's strategy tendency at the previous moment based on this information. Then, the drones select a candidate action from a set of potential actions, calculating the opponent's strategy tendency and probability of choosing that action according to the game theory formula. A reward function is designed, and the drones select their own action based on the obtained opponent's strategy probability using the optimal response decision rule, observing the changes in the opponent's actions until the entire drone swarm reaches a stable state. This enables the analysis and game theory of multiple surrounding drones. When encountering signal interference, this invention allows some drone units to malfunction while the drone itself can adjust its movement attitude, maintaining relative stability for the entire swarm.

[0042] Example 2

[0043] This invention relates to a drone swarm attitude recognition module based on a VGG16 network for collecting, preprocessing, and classifying data.

[0044] The training process of the pose recognition model is as follows:

[0045] (1) Collect historical motion videos of the UAV in multiple scenarios in advance, and preprocess the historical motion videos;

[0046] Data collection: Videos of drones moving in various scenarios including cities, grasslands, deserts, Gobi, and oceans were captured. The captured drone videos were categorized into nine motion postures based on the drone's direction of movement: "forward," "backward," "left," "right," "hovering," "left-forward," "left-backward," "right-forward," and "right-backward." Partial datasets for each of the nine postures across different scenarios are provided.

[0047] Dataset preprocessing: The captured drone motion video is converted into frame-by-frame still images. Histogram equalization is then used to transform the contrast of these still images; the formula can be expressed as:

[0048]

[0049] in, This represents the total number of pixels in the image. This indicates that the gray level in the image is The number of pixels, This indicates the number of gray levels in an image.

[0050] (2) Divide the preprocessed motion videos into training datasets and test datasets;

[0051] This invention shuffles the preprocessed UAV motion attitude dataset and divides it into two categories: training dataset and validation dataset.

[0052] (3) Obtain the pre-built VGG16 network and use the VGG16 network as the pose recognition model;

[0053] The VGG16 network consists of 13 convolutional layers, 3 fully connected layers, and 5 pooling layers; the loss rate of the VGG16 network for model training on the dataset is as follows: Figure 4 As shown, the accuracy is as follows Figure 5 As shown.

[0054] The VGG16 network consists of 13 convolutional layers, 3 fully connected layers, and 5 pooling layers. First, the data passes through two convolutional layers consisting of 64 3x3 kernels each with a stride of 1, followed by a 2x2 max pooling layer with a stride of 2. Then, it passes through two more convolutional layers consisting of 128 3x3 kernels each with a stride of 1, followed by another 2x2 max pooling layer with a stride of 2. Next, it passes through three more convolutional layers consisting of 256 3x3 kernels each with a stride of 1, followed by another 2x2 max pooling layer with a stride of 2. Then, it passes through three more convolutional layers consisting of 512 3x3 kernels each with a stride of 1, followed by another 2x2 max pooling layer with a stride of 2. Finally, it passes through three fully connected layers. The classification result is then output through a SoftMax regression classifier.

[0055] 1) Convolutional layer: Implements local connectivity and weight sharing, greatly reducing the number of parameters in the convolution process. The convolution process can be expressed by the formula:

[0056] ,

[0057] in, Indicates the input parameters. Indicates the weight parameter. Indicates the bias and weight parameters. and bias The backpropagation algorithm continuously iterates and updates the parameters until the parameter changes reach a threshold and then stop.

[0058] 2) Pooling Layer: The feature matrix obtained from the convolutional layer is subjected to max pooling, then passed through the ReLU activation function. The output is used as the input to the next layer. The ReLU function formula can be expressed as:

[0059] ,

[0060] in, Represented as the output parameters obtained after passing through the previous layer, the purpose of the ReLU function is to avoid negative values ​​in the parameters;

[0061] 3) Fully Connected Layer: The feature matrix obtained after 13 convolutional layers and 5 pooling layers is input into the fully connected layer, which performs a fully connected operation, as shown below:

[0062] ,

[0063] in, This represents the result obtained after convolution and pooling operations. 1 eigenvector Represented as the first The neuron in the first Weights on each feature parameter Indicates the first The bias of each neuron;

[0064] 4) SoftMax Regression Classification: The feature vector obtained after passing through the fully connected layer is fed into the classifier, which performs the classification and determination of the UAV's attitude type. The probability distribution obtained after passing through the SoftMax regression classifier is then calculated. The output can be expressed by the formula:

[0065] ,

[0066] in, Represented as the feature vector obtained from the fully connected layer, These are represented as the model's weight parameters. Represented as a probability distribution The normalized coefficient, This represents the output type corresponding to the fully connected layer;

[0067] 5) Iterative optimization: The loss function is calculated using cross-entropy during backpropagation, and the weight parameters are optimized using the Adam gradient descent algorithm. and bias Iterative optimization:

[0068] loss function Represented as:

[0069] ,

[0070] in, This is represented as the probability distribution of the predicted values. This represents the probability distribution of the target value. for sub-items, for Sub-items;

[0071] The Adam gradient descent optimization algorithm is optimized according to the following formula:

[0072] ,

[0073] in, Represented as the number of iterations, These are the parameters to be optimized. Represented as a loss function, Represented as the learning rate, This represents the gradient value of the loss function. Describing first-order momentum, This represents the first-order momentum parameter. Represented as second momentum, Expressed as second-order momentum parameters, This is represented as the gradient at the current time step.

[0074] (4) The training dataset is fed into the VGG16 network to train the pose recognition model, and the training process is iteratively optimized according to the loss function;

[0075] (5) The test dataset is sent into the VGG16 network for model verification, and the recognition results are compared with the actual labels to determine the accuracy of the VGG16 network.

[0076] The test dataset was fed into the VGG16 network for model validation. The recognition results were compared with the actual labels. The accuracy of the UAV flight attitude recognition model trained by VGG16 reached 99.83%.

[0077] (6) Repeat steps 4 to 5 until the VGG16 network reaches the accuracy standard and obtains the trained VGG16 network.

[0078] Example 3

[0079] As an optional embodiment of the present invention, step 4 includes:

[0080] Step 41: The player iteratively calculates the opponent's strategic tendencies in taking all actions in the previous moment.

[0081] At the current time t, the players... Iterative calculation of the game opponent up to time t-1 using an autoregressive model based on the continuous Boltzmann equation. The strategic tendency of all actions:

[0082] ,

[0083] in, , Indicates a collection of drones. Indicates the number of drones, express Time-based gamers The opponent in the game The strategic tendency of all actions Represents the correlation coefficient at time h. This represents the random error term.

[0084] Step 42: Select any action as a candidate action;

[0085] Step 43: Calculate the opponent's strategy tendency to choose the candidate action at the current moment based on the movement direction and strategy tendency of the previous moment.

[0086] Step 43 includes:

[0087] Gamers Calculation based on extended Kalman filter Constantly playing against opponents Select a candidate action strategic tendencies :

[0088] ,

[0089] in, A game player exist The opponent in the game of time The strategic tendency of all actions; This represents policy bias noise (i.e., interference factors) that contains attitude information.

[0090] The game process between the player and the opponent can be represented by a state-space model as follows:

[0091] ,

[0092] in, For the game process, For continuous Boltzmann components, It is an indicator function with a mean of zero. Tendency error, obey N(0,P) distributed, P Covariance matrix 。

[0093] Example 4

[0094] In one specific embodiment, step 5 includes:

[0095] Step 51, based on strategy preference Speculators Through continuous Boltzmann components Estimate the opponent exist Select a candidate action at any time strategy probability The formula can be expressed as:

[0096] ,

[0097] in, This represents the entropy parameter in the continuous Boltzmann equations. Representing the players The set of optional actions, mathematically described as , Representing the players The number of selectable actions, Indicates a candidate action; Representing the players The set of policy probabilities over the set of optional actions is mathematically described as follows: , Representing the players Select Action The probability of;

[0098] Step 52, calculate the drone All game opponents Choose actions at all times joint probability The formula can be expressed as:

[0099]

[0100] Among them, actions It includes nine selectable actions: "forward", "backward", "left", "right", "hover", "left forward", "left backward", "right forward", and "right backward".

[0101] Step 53, targeting the game opponent j Each candidate action undergoes a process of calculating the joint probability to obtain the game opponent. j The joint probability of all candidate actions;

[0102] Step 54, select joint probability Biggest candidate action The most likely action that the opponent in the game will choose at the current moment. Thus, the joint probability Converted to a pure policy, it can be expressed as the following formula:

[0103]

[0104] in, For the optional action set of the drone unit, for Any optional action within the system. This conversion method increases the probability that unmanned system clusters will choose the same action strategy.

[0105] Example 5

[0106] As an optional embodiment of the present invention, step 6 includes:

[0107] Design Game Players reward function :

[0108] ,

[0109] in, Representing the players Action set The available actions, Indicates the penalty coefficient. Representing the players Select Action The opponent chooses to act. The game payoff at that time Representing the players Select Action The value function of , when players tend to choose the same action. When players tend to choose different actions, .

[0110] Example 6

[0111] As an optional embodiment of the present invention, step 7 includes:

[0112] Step 71: The Player Apply the best response decision rule to select a self-action The optimal response decision formula is expressed as:

[0113]

[0114] Step 72: Based on the reward function and the optimal response decision formula, obtain the optimal decision using the Adam gradient descent algorithm:

[0115]

[0116] Step 73: Take the action corresponding to the optimal decision. Execute the strategy and continue to observe changes in the opponent's actions.

[0117] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0118] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0119] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An autonomous control method for a drone swarm based on VGG16 and virtual game theory, applied to a drone swarm, wherein any drone in the swarm is a player and the other drones are players, characterized in that... include: Step 1: Each player obtains the motion image of their opponent from the previous moment. Step 2: After preprocessing the motion image, it is fed into the trained pose recognition model to identify the opponent's movement direction in the previous moment. Step 3: Obtain a pre-designed set of candidate actions for the drone; Step 4: Based on the opponent's strategic inclinations and movement directions for all available actions in the previous moment, calculate the opponent's strategic inclinations for all available actions in the current moment. Step 5: Based on the strategy tendencies of each player in all candidate actions at the current moment, calculate the joint probability of all players choosing a certain candidate action at the current moment, and convert the joint probability into the most likely action to be performed by all players at the current moment. Step 6: Design a reward function based on the action chosen by all opponents at the current moment that maximizes the reward function. Step 7: The player flies according to their own actions and continues to observe the changes in their opponent's actions; Step 8: Repeat steps 1 through 7 until the drone swarm stabilizes.

2. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 1, characterized in that, The training process of the pose recognition model is as follows: (1) Collect historical motion videos of the UAV in multiple scenarios in advance, and preprocess the historical motion videos; (2) Divide the preprocessed motion videos into training datasets and test datasets; (3) Obtain the pre-built VGG16 network and use the VGG16 network as the pose recognition model; (4) The training dataset is fed into the VGG16 network to train the pose recognition model, and the training process is iteratively optimized according to the loss function; (5) The test dataset is sent into the VGG16 network for model verification, and the recognition results are compared with the actual labels to determine the accuracy of the VGG16 network. (6) Repeat steps 4 to 5 until the VGG16 network reaches the accuracy standard and obtains the trained VGG16 network.

3. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 2, characterized in that, The VGG16 network consists of 13 convolutional layers, 3 fully connected layers, and 5 pooling layers. (1) Convolutional layer: It realizes two functions: local connectivity and weight sharing, which greatly reduces the number of parameters in the convolution process. The convolution process can be expressed by the formula: ， in, Indicates the input parameters. Indicates the weight parameter, Indicates the bias and weight parameters. and bias The backpropagation algorithm continuously iterates and updates the parameters until the parameter changes reach a threshold and then stop. (2) Pooling layer: The feature matrix obtained from the convolutional layer is subjected to max pooling, and then the ReLU activation function is applied to the output. The output is used as the input to the next layer. The ReLU function formula can be expressed as: ， in, Represented as the output parameters obtained after passing through the previous layer, the purpose of the ReLU function is to avoid negative values ​​in the parameters; (3) Fully connected layer: The feature matrix obtained after 13 convolutional layers and 5 pooling layers is input into the fully connected layer, which performs a fully connected operation, as shown below: ， in, This represents the result obtained after convolution and pooling operations. 1 eigenvector Represented as the first The neuron in the first Weights on each feature parameter Indicates the first The bias of each neuron; (4) SoftMax Regression Classification: The feature vector obtained after the fully connected layer is fed into the classifier, which performs the classification and judgment of the UAV attitude type. The probability distribution obtained by the SoftMax regression classifier is then used. The output can be expressed by the formula: ， in, Represented as the feature vector obtained from the fully connected layer, These are represented as the model's weight parameters. Represented as a probability distribution The normalized coefficient, This represents the output type corresponding to the fully connected layer; (4) Iterative optimization: The loss function is calculated using cross-entropy during backpropagation, and the weight parameters are optimized using the Adam gradient descent algorithm. and bias Iterative optimization: loss function Represented as: ， in, This is represented as the probability distribution of the predicted values. This represents the probability distribution of the target value. for sub-items, for Sub-items; The Adam gradient descent optimization algorithm is optimized according to the following formula: ， in, Represented as the number of iterations, These are the parameters to be optimized. Represented as a loss function, Represented as the learning rate, This represents the gradient value of the loss function. Describing first-order momentum, This represents the first-order momentum parameter. Represented as second momentum, Expressed as second-order momentum parameters, This is represented as the gradient at the current time step.

4. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 1, step 4 includes: Step 41: The player iteratively calculates the opponent's strategic tendencies in taking all actions in the previous moment. Step 42: Select any action as a candidate action; Step 43: Calculate the opponent's strategy tendency in choosing the candidate action based on the previous movement direction and strategy tendency.

5. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 4, step 41 includes: At the current time t, the players... Iterative calculation of the game opponent up to time t-1 using an autoregressive model based on the continuous Boltzmann equation. The strategic tendency of all actions: ， in, , Indicates a collection of drones. Indicates the number of drones, express Time-based gamers The opponent in the game The strategic tendency of all actions Represents the correlation coefficient at time h. This represents the random error term.

6. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 5, step 43 includes: Gamers Calculation based on extended Kalman filter Constantly playing against opponents Select a candidate action strategic tendencies : ， in, A game player exist The opponent in the game of time The strategic tendency of all actions; This represents policy bias noise containing attitude information; The game process between the player and the opponent can be represented by a state-space model as follows: ， in, For the game process, For continuous Boltzmann components, It is an indicator function with a mean of zero. Tendency error, obey N(0,P) distributed, P Covariance matrix 。 7. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 6, step 5 includes: Step 51, based on strategy preference Speculators Through continuous Boltzmann components Estimate the opponent exist Select a candidate action at any time strategy probability The formula can be expressed as: ， in, This represents the entropy parameter in the continuous Boltzmann equations. Representing the players The set of optional actions, mathematically described as , Representing the players The number of selectable actions, Indicates a candidate action; Representing the players The set of policy probabilities over the set of optional actions is mathematically described as follows: , Representing the players Select Action The probability of; Step 52, calculate the drone All game opponents Choose actions at all times joint probability The formula can be expressed as: Step 53, targeting the game opponent j Each candidate action undergoes a process of calculating the joint probability to obtain the game opponent. j The joint probability of all candidate actions; Step 54, select joint probability The most likely action to be taken is the action that the opponent is most likely to choose at the current moment. Thus, the joint probability Converted to a pure policy, it can be expressed as the following formula: in, For the optional action set of the drone unit, for Any optional action in the process; this conversion method is conducive to unmanned system clusters having a higher probability of choosing the same action strategy.

8. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 7, step 6 includes: Design Game Players reward function : ， in, Representing the players Action set The available actions, Indicates the penalty coefficient. Representing the players Select Action The opponent chooses to act. The game payoff at that time Representing the players Select Action The value function of , when players tend to choose the same action. When players tend to choose different actions, .

9. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 8, step 7 includes: Step 71: The Player Apply the best response decision rule to select a self-action The optimal response decision formula is expressed as: Step 72: Based on the reward function and the optimal response decision formula, obtain the optimal decision using the Adam gradient descent algorithm: Step 73: Take the action corresponding to the optimal decision. Execute the strategy and continue to observe changes in the opponent's actions.

10. The autonomous control method for unmanned aerial vehicle swarms based on VGG16 and virtual game theory according to claim 9, step 8 includes: Gamers Repeat steps 1-7 until the drone swarm reaches a stable state, and obtain the distribution of all possible action strategies of all opponents at the current moment.

Images