Ground-air cooperative unmanned aerial vehicle dynamic game decision management planning system

Abstract

The invention discloses a ground-air cooperative unmanned aerial vehicle dynamic game decision management planning system. The system comprises an unmanned aerial vehicle array state sensor group located in the air, a decision management center located on the ground and a planning actuator. The decision management center located on the ground comprises a sensor database, a data preprocessing module, a space-time diagram module and a dynamic game decision module. Historical flight data are preprocessed and input into a space-time diagram module, spatial position changes are analyzed by using a diagram convolutional neural network, and time dynamic characteristics are captured in combination with a gating cycle unit, so that a pre-trained game decision model is generated. The current flight data is preprocessed and then input into the pre-trained game decision model, and the generated state forecast is manually confirmed and adjusted to generate a decision instruction. And an execution module in the planning executor receives the decision instruction and sends a corresponding operation instruction. According to the invention, a space-time diagram module is innovatively introduced in the dynamic game decision-making process of the unmanned aerial vehicle, and the synergistic effect between unmanned aerial vehicle individuals and the time dynamic relationship of the individuals are accurately simulated; ground-air cooperation is realized through a pre-training model of a ground control decision center and real-time flight data of an aerial unmanned aerial vehicle array, and adaptive adjustment and planning of a dynamic game decision are completed.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a ground-to-air coordinated unmanned aerial vehicle dynamic game decision-making management planning system. Background Art

[0002] Due to its flexibility, real-time data transmission and autonomous execution of tasks, drones have become one of the fastest, most cost-effective and most widely used unmanned means of executing tasks in the world. Drones often carry out tasks in a centralized manner in an array. With the popularity of drone arrays, people have begun to pay attention to the changing laws of spatial position relationships and individual dynamic characteristics. Usually, the upper plate of the drone array is equipped with a comprehensive state sensor to monitor the system status. By collecting all drone sensor data, a database reflecting the state of the drone dynamic game system can be established in the ground remote control. Through ground-air collaboration, the ground control center can use the historical flight database to manage and plan the dynamic game of the drone array to achieve more efficient task execution and system safety. In the array, drones often carry out tasks in a collaborative manner. When a drone is the main task executor, the speed and position relationship between it and other auxiliary drones will become complicated. These factors make the spatial and temporal relationship of the drone array fluctuate, uncertain and unstable, thereby affecting the system's dynamic game planning and task execution efficiency, and even leading to safety accidents.

[0003] In order to address this challenge, it is crucial to scientifically and accurately capture the spatial position changes and individual temporal dynamic characteristics of drone arrays. This not only helps to improve the energy efficiency of drones, but also ensures the safety of mission execution and increases the accuracy and adaptability of dynamic game decisions. Artificial intelligence detection planning methods can mine more implicit features in drone sensor data and have stronger learning capabilities. RNN recurrent neural networks and GRU gated recurrent units based on time series properties are widely used to help process long-term time dependencies and time series data.

[0004] As a variant of recurrent neural network, GRU gated recurrent unit adopts special gate structure and memory unit to solve the problem of gradient vanishing or exploding when traditional network processes long time series data. Its hidden layer contains input gate, output gate and forget gate, and the selective memory of information and the update of unit state are controlled by the gated unit system. For high-dimensional and complex drone array data, GRU can mine inherent hidden time dynamic features and deal with long-term time dependency problems. However, in the dynamic game decision of drones, GRU can hardly discover the spatial coupling relationship and mutual influence relationship between drones in the array by relying solely on long-term time dependency. Therefore, it is necessary to combine appropriate methods to mine spatial features and obtain more comprehensive and accurate relationship of drone spatial array changes.

[0005] Since the sensors of drones are distributed in a fixed manner, the position information and other data sent back by the sensors can represent the mutual correlation between individual drones. Based on the mutual influence and collaborative characteristics between array individuals, the GCN architecture based on graph learning has more universal advantages, can explore the coupling relationship between different drone spaces in the array, and make more accurate judgments on the mutual collaborative state. At the same time, the GRU architecture based on recurrent neural networks can capture the temporal dynamic characteristics of individual drones, so that the system detection takes into account the influence of historical flight data. However, there are currently few methods and devices that introduce graph learning architectures to assist in the dynamic game decision management planning of drones. The GRU and other time series models used for dynamic games of drones in the past have failed to combine the particularity of collaborative scenarios in drone array applications for graph learning, resulting in low state detection accuracy and high error rate, which brings economic losses and affects reliability. Therefore, there is an urgent need to invent a better ground-to-air collaborative drone game decision management planning system to complete adaptive and high-precision dynamic game decisions. Summary of the invention

[0006] The purpose of the present invention is to address the deficiencies in the prior art and provide a ground-to-air coordinated unmanned aerial vehicle dynamic game decision-making management planning system.

[0007] The object of the present invention is achieved through the following technical solutions: a ground-air coordinated unmanned aerial vehicle dynamic game decision management planning system, the system comprising: an unmanned aerial vehicle array state sensor group located in the air, a decision management center located on the ground and a planning executor; the decision management center located on the ground comprises a sensor database, a data preprocessing module, a spatiotemporal graph module and a dynamic game decision module;

[0008] The drone array state sensor group located in the air is used to detect the real-time position and individual state of the drone array, and transmit the current flight data obtained by the new monitoring to the sensor database in the decision-making management center located on the ground;

[0009] The drone array is composed of N drones, each of which is equipped with a corresponding state sensor;

[0010] The sensor database is used to receive the current flight data newly monitored and store the historical flight data, and then transfer the current flight data and the historical flight data to the data preprocessing module;

[0011] The data preprocessing module is used to perform preprocessing operations on the current flight data and the historical flight data respectively, wherein the preprocessing operations include missing value filling and data cleaning, to obtain the preprocessed current flight data and the historical flight data, and then to transfer the preprocessed historical flight data to the space-time graph module, and to transfer the preprocessed current flight data to the dynamic game decision module;

[0012] The spatiotemporal graph module is used to perform mutual information evaluation on the pre-processed historical flight data to screen variables, obtain the adjacency matrix using the maximum nearest neighbor method, train the spatiotemporal graph model using the GCN graph learning module and the GRU time series learning module, and save the optimal model as the pre-trained game decision model and save it in the dynamic game decision module;

[0013] The dynamic game decision module is used to use the pre-trained game decision model to predict the pre-processed current flight data, obtain the predicted flight data at the next moment, and generate a state forecast based on it; and manually confirm and adjust the state forecast to generate corresponding decision instructions and send them to the planning executor;

[0014] The planning executor is used to use the internal execution module to receive the decision instructions and send corresponding operation instructions to achieve the synergy effect and adaptive adjustment planning of the drone cluster.

[0015] Furthermore, the sensor database receives the current flight data newly monitored and stores the historical flight data, specifically:

[0016] The sensor database receives the current flight data newly monitored, and the current flight data is in, is the flight data sent back by the state sensor on the jth UAV in the UAV array at the mth moment, j = 1, 2, …, j, …, N, m represents the current moment;

[0017] The sensor database stores historical flight data. The historical flight data is

[0018]

[0019] in, is the flight data set sent back by the state sensor on the jth UAV in the UAV array at time 0 to m-1, is the flight data sent back by the j-th UAV at the i-th moment.

[0020] Furthermore, the data preprocessing module performs preprocessing operations on the current flight data and the historical flight data respectively, and the preprocessing operations include missing value filling and data cleaning to obtain the preprocessed current flight data and historical flight data, specifically:

[0021] The data preprocessing module converts historical flight data Perform preprocessing operations, including missing value filling and data cleaning, to obtain preprocessed historical flight data

[0022]

[0023] in, The historical flight data set after preprocessing operation for the jth UAV; The historical flight data of the j-th UAV after the preprocessing operation at the i-th time;

[0024] The missing value filling is as follows: when the flight data sent back by the jth UAV at the i-th time Any variable in If there is a missing value, the missing value is filled by using the variables of the k moments before and after through mathematical calculation to obtain the variable after the missing value is filled. The formula is as follows:

[0025]

[0026] The data cleaning is specifically: deleting the flight data of each drone at every moment Complete data cleaning based on the erroneous data in the data;

[0027] The data preprocessing module converts the current flight data into Perform preprocessing operations to obtain the preprocessed current flight data in, is the current flight data of the j-th UAV after the preprocessing operation at the m-th moment.

[0028] Furthermore, the space-time graph module performs mutual information evaluation on the pre-processed historical flight data to screen variables, obtains the adjacency matrix using the maximum nearest neighbor method, trains the space-time graph model using the GCN graph learning module and the GRU time series learning module, and saves the optimal model as the pre-trained game decision model and saves it in the dynamic game decision module, which specifically includes the following sub-steps:

[0029] (a.1) The key component of the spatiotemporal graph model is the spatiotemporal graph neural network, which is divided into three key elements: time domain, space domain, and graph structure; Graph structure: The construction of the graph includes two steps: screening UAVs that are highly relevant to the flight mission, and determining the edge relationship between these UAVs and generating the adjacency matrix required by GCN; Spatial domain: Graph convolutional neural network is a neural network specially designed for processing graph structure data. It learns the mutual correlation between different UAV flight data from a spatial perspective. It captures the spatial dependency between UAVs by propagating information between each UAV and its neighbors in the graph; Time domain: Gated recurrent unit is a recursive neural network for time series analysis. It introduces a gating mechanism to control the information flow from the perspective of the time domain to capture the dynamic changes in the flight data of each UAV in time;

[0030] First, the preprocessed historical flight data The mutual information value is obtained by evaluating the historical flight data set after preprocessing operations corresponding to any two drones A and B. The calculation formula is as follows:

[0031]

[0032] Where A=1,2,…,A,…,N,B=1,2,…,B,…,N,A≠B;αis the historical flight data set after preprocessing operation The number of grids divided in the direction, β is the historical flight data set after preprocessing operation The number of grids divided in the direction; for and The joint probability of for The marginal probability of for The marginal probability of; λ is the variable control parameter;

[0033] Arrange the mutual information values ​​corresponding to all two arbitrary drones from large to small, select the historical flight data set after preprocessing corresponding to the first S drones with the largest mutual information values ​​as the key historical flight data set, and obtain the key historical flight data set.

[0034]

[0035] in, represents the g-th key historical flight data set, g = 1, 2, ..., g, ..., S;

[0036] (a.2) Use the maximum nearest neighbor method to analyze key historical flight data Construct edge relationships and obtain edge weight matrix

[0037]

[0038] Among them, edge f,g Represents a collection of key historical flight data and key historical flight data sets The edge weight between them is f=1,2,…,f,…,S; when f≠g, Where ξ represents the bandwidth of the Gaussian kernel function; when f = g, edge f,g =0;

[0039] Key historical flight data The similarity is used to obtain each key historical flight data set The K other key historical flight data sets with the greatest similarity are in the edge weight matrix Only keep key historical flight data sets The edge weights of the corresponding K other key historical flight data sets are changed to 0, and the graph structure is obtained.

[0040] Then the graph structure Converted to adjacency matrix structure It is best to get key historical flight data The picture in, Shows key historical flight data The time step relationship in

[0041] (a.3) Then the figure Input into the GCN graph learning module, use the graph convolution network to perform convolution calculations, and get the output of the GCN graph learning module

[0042] Among them, σ(·) is the activation function; U is given by The Laplace calculation results are:

[0043]

[0044] Where D is the graph The node degree matrix of

[0045] g θ Is the convolution kernel function:

[0046] Among them, θ k is the polynomial parameter, T k are the Chebyshev polynomial parameters, For key historical flight data The calculated diagonal matrix;

[0047] (a.4) Output It consists of sub-outputs from time 0 to time t-1: in, Represents the output at the i-th moment;

[0048] And key historical flight data Divided from time 0 to time t-1, divided into in, represents the key historical flight data set at the i-th moment;

[0049] Then the output and key historical flight data Input to the GRU temporal learning module to capture temporal dependencies based on the gated recurrent unit;

[0050]

[0051] Where, Θ is the Hadamard symbol,; Represents the output of the update gate at the i-th moment; represents the output of the reset gate at the i-th moment; is the hidden state at the i-1th moment; is the hidden state at the i-th moment; is the candidate hidden state at the i-th moment; W xr , W xz , W xh , W hr , W hz and W hh are learnable parameters; b r 、b z and b h are bias parameters respectively;

[0052] After the above calculation, the time domain convolution output at the m-1th moment is obtained Among them, is the learnable parameter, b γ is the bias parameter;

[0053] (a.5) Then according to the time domain convolution output at the m-1th moment The key historical flight data set at the m-1th moment Calculate the mean square error to get the total loss; based on the total loss, the learnable parameter W xr , W xz , W xh , Whr , W hz , W hh and W γ And the bias parameter b r 、b z 、b h and b γ The network is updated and the state of the entire network when the total loss is minimum is recorded and saved as the optimal model. The optimal model is then used as the pre-trained game decision model and saved in the dynamic game decision module.

[0054] The beneficial effects of the present invention are: 1. The GCN graph learning module is innovatively introduced to perform graph convolution analysis on the spatial distribution of drone sensors, so that the state detection model can deeply explore the spatial characteristics of sensor signals and improve the decision-making reliability in drone applications; 2. Combining long time series and spatial dimensions, the dynamic game state data of the drone is comprehensively learned, and the dynamically changing spatiotemporal relationship is accurately extracted to improve the real-time and accuracy of decision-making; 3. According to the operating characteristics of the drone, the decision model is adaptively adjusted to improve the response capability to abnormal situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a structural diagram of a ground-air collaborative UAV dynamic game decision-making management planning system;

[0056] Figure 2 A schematic diagram of the locations of the corresponding status sensors installed in the drone;

[0057] Figure 3 It is the spatial position distribution map of the drone array;

[0058] In the figure, 1-state sensor; 2-UAV array; 3-UAV array state sensor group in the air; 4-sensor database; 5-data preprocessing module; 6-space-time graph module; 7-dynamic game decision module; 8-planning executor; 9-decision management center on the ground; 10-execution module. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical scheme and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0060] In order to overcome the challenges of difficult mining of spatial coupling relationships between individual drones and difficult capturing of individual temporal characteristics in the dynamic game decision management planning of drone arrays, the purpose of the present invention is to construct a ground-to-air collaborative decision-making system, collect status data from the air, build a database in the ground decision management center, and combine GCN to mine the spatial coupling relationships between individuals in the drone array, GRU to capture the temporal dynamic characteristics of individuals, and intelligently manage the dynamic game decisions of drones. A ground-to-air collaborative drone dynamic game decision management planning system is proposed to achieve efficient task execution and system security.

[0061] It constructs a graph model to characterize the spatial topological relationship of drones in the array, uses GCN to learn the spatial dependency of the graph, and GRU to learn the temporal dependency, and uses this convolution kernel cyclic splicing to predict future states through historical flight data and current flight data and perform dynamic game decision-making tasks.

[0062] Example 1

[0063] like Figure 1 As shown, the present invention provides a ground-to-air coordinated UAV dynamic game decision management planning system, comprising: a UAV array state sensor group 3 located in the air, a decision management center 9 located on the ground, and a planning executor 8. The decision management center 9 located on the ground includes a sensor database 4, a data preprocessing module 5, a spatiotemporal graph module 6, and a dynamic game decision module 7.

[0064] The drone array status sensor group 3 located in the air is used to detect the real-time position and individual status of the drone array 2, and transmit the current flight data obtained by new monitoring to the sensor database 4 in the decision-making management center 9 located on the ground.

[0065] The drone array 2 is composed of N drones, each of which is equipped with a corresponding state sensor 1, such as Figure 2 As shown. Figure 3 As shown, N drones are distributed in a certain spatial position and head to a designated mission location. The current flight data obtained by the new monitoring includes information such as the position, speed, and energy level of the drones.

[0066] The sensor database 4 is used to receive the current flight data newly monitored and store the historical flight data, specifically:

[0067] The sensor database 4 receives the current flight data obtained by new monitoring, and the current flight data is in, is the flight data transmitted back by the state sensor on the j-th UAV in the UAV array 2 at the m-th moment, j=1,2,...,j,...,N, and m represents the current moment.

[0068] The sensor database 4 stores historical flight data. The historical flight data is

[0069]

[0070] in, is the flight data set sent back by the state sensor on the jth UAV in the UAV array 2 at the time 0 to m-1, is the flight data sent back by the j-th UAV at the i-th moment.

[0071] Then the current flight data and historical flight data are transmitted to the data preprocessing module 5 .

[0072] The data preprocessing module 5 is used to perform preprocessing operations on the current flight data and the historical flight data respectively. The preprocessing operations include missing value filling and data cleaning to obtain the preprocessed current flight data and historical flight data, specifically:

[0073] The data preprocessing module 5 converts the historical flight data Perform preprocessing operations, including missing value filling and data cleaning, to obtain preprocessed historical flight data

[0074]

[0075] in, The historical flight data set after preprocessing operation for the jth UAV; Historical flight data of the j-th UAV after preprocessing operation at the i-th moment.

[0076] The missing value filling is as follows: when the flight data sent back by the jth UAV at the i-th time Any variable in If there is a missing value, the missing value is filled by using the variables of the k moments before and after through mathematical calculation to obtain the variable after the missing value is filled. The formula is as follows:

[0077]

[0078] The data cleaning is specifically: deleting the flight data of each drone at every moment Complete data cleaning by removing the erroneous data in the data.

[0079] The data preprocessing module 5 converts the current flight data into Perform preprocessing operations to obtain the preprocessed current flight data in, is the current flight data of the j-th UAV after the preprocessing operation at the m-th moment.

[0080] The preprocessed historical flight data is then transferred to the space-time diagram module 6, and the preprocessed current flight data is transferred to the dynamic game decision module 7.

[0081] The space-time graph module 6 is used to perform mutual information evaluation on the pre-processed historical flight data to screen variables, obtain the adjacency matrix using the maximum nearest neighbor method, train the space-time graph model using the GCN graph learning module and the GRU time series learning module, and save the optimal model as the pre-trained game decision model and save it in the dynamic game decision module 7, which specifically includes the following sub-steps:

[0082] (a.1) The key component of the spatiotemporal graph model is the spatiotemporal graph neural network, which is divided into three key elements: time domain, space domain, and graph structure. Graph structure: The construction of the graph includes two steps: screening drones that are highly relevant to the flight mission (based on mutual information evaluation), and determining the edge relationships between these drones and generating the adjacency matrix required by GCN (based on the maximum nearest neighbor method). Spatial domain: Graph convolutional neural network (GCN) is a neural network specifically designed to process graph structured data. It learns the mutual correlation between different drone flight data from a spatial perspective. It captures the spatial dependency between drones by propagating information between each drone and its neighbors in the graph. Temporal domain: Gated recurrent unit (GRU) is a recursive neural network for time series analysis. It introduces a gating mechanism to control the information flow from the perspective of the time domain to capture the dynamic changes in the flight data of each drone in time.

[0083] Evaluate pre-processed historical flight data using mutual information (MIC) The correlation between the historical flight data sets after the preprocessing operations corresponding to any two UAVs in the game can be used to select the historical flight data sets after the preprocessing operations with the largest information contribution, ensuring that the decision model obtains the richest input information. This method can identify the S historical flight data sets after the preprocessing operations with the strongest collaborative relationship in the current game state, use these most relevant data for modeling, eliminate data redundancy, and reduce unnecessary information processing, thereby improving the accuracy of game decisions and the intelligence level of the system. At the same time, this targeted sensor selection also enhances the flexibility and adaptability of the system in complex and changing environments. These selected sensor data are used as input variables of the spatiotemporal graph model, providing more effective support for subsequent decision management.

[0084] First, the preprocessed historical flight data The mutual information value is obtained by evaluating the historical flight data set after preprocessing operations corresponding to any two drones A and B. The calculation formula is as follows:

[0085]

[0086] Where A=1,2,...,A,…,N,B=1,2,…,B,…,N,A≠B;αis the historical flight data set after preprocessing operation The number of grids divided in the direction, β is the historical flight data set after preprocessing operation The number of grids divided in the direction; for and The joint probability of for The marginal probability of for The marginal probability of ; λ is the variable control parameter.

[0087] Arrange the mutual information values ​​corresponding to all two arbitrary drones from large to small, select the historical flight data set after preprocessing corresponding to the first S drones with the largest mutual information values ​​as the key historical flight data set, and obtain the key historical flight data set.

[0088]

[0089] in, represents the g-th key historical flight data set, g=1,2,…,g,…,S.

[0090] (a.2) Use the KNN method to select key historical flight data The core of constructing edge relationships is to calculate the maximum similarity between the selected key historical flight data sets, and record the maximum similarity as the edge weight relationship between the key historical flight data sets, so as to characterize the key historical flight data. The interdependencies and synergies between each of the key historical flight data sets in the

[0091] Use the maximum nearest neighbor method to analyze key historical flight data Construct edge relationships and obtain edge weight matrix

[0092]

[0093] Among them, edge f,g Represents a collection of key historical flight data and key historical flight data sets The edge weight between them is f=1,2,…,f,…,S; when f≠g, Where ξ represents the bandwidth of the Gaussian kernel function; when f = g, edge f,g =0.

[0094] Key historical flight data The similarity is used to obtain each key historical flight data set The K other key historical flight data sets with the greatest similarity are in the edge weight matrix Only keep key historical flight data sets The edge weights of the corresponding K other key historical flight data sets are changed to 0, and the graph structure is obtained.

[0095] Then the graph structure Converted to adjacency matrix structure Finally get the key historical flight data The picture in, Shows key historical flight data The time step relationship in .

[0096] (a.3) Using graph convolutional networks (GCN) to perform convolution calculations on graph structures can effectively obtain key historical flight data The deep coupling relationship between the S key historical flight data sets during collaboration provides rich spatiotemporal information for decision management. In the dynamic environment of ground-air collaboration, drones need to quickly adapt to changing conditions, resulting in complex and changeable relative position relationships. The graph convolutional network in the GCN graph learning module can specifically extract feature information of different drone nodes, such as speed, energy, etc., and enhance the individual drone's perception of the overall environment through the aggregation of neighbor node information. This is crucial for collaborative decision-making and helps drone groups form a global perspective. In addition, the GCN model has good scalability and can adapt to drone groups of different sizes.

[0097] Then the figure Input into the GCN graph learning module, use the graph convolution network to perform convolution calculations, and get the output of the GCN graph learning module

[0098] Among them, σ(·) is the activation function; U is given by The Laplace calculation results are:

[0099]

[0100] Where D is the graph The node degree matrix of

[0101] g θ Is the convolution kernel function:

[0102] Among them, θ k is the polynomial parameter, T k are the Chebyshev polynomial parameters, For key historical flight data The computed diagonal matrix.

[0103] (a.4) Using gated recurrent units (GRUs) to capture the temporal dependencies of individual drones can effectively enhance the performance of dynamic game decision management systems. GRUs are good at modeling the dynamic changes in time series data and can accurately capture the evolution of drone states over time, which is crucial for decision management in dynamic environments. In addition, the gating mechanism of GRU can selectively retain and forget historical information, improving the efficiency of modeling long-term dependencies, helping individual drones to maintain their memory of previous states and make more coherent decisions. Combining GRU with GCN, which is good at spatial modeling, can fully capture the dynamic characteristics of drone groups in the temporal and spatial dimensions, provide richer information for decision management, and further enhance the continuity and robustness of the system.

[0104] The output It consists of sub-outputs from time 0 to time t-1: in, Represents the output at the i-th moment;

[0105] And key historical flight data Divided from time 0 to time t-1, divided into in, represents the key historical flight data set at the i-th moment;

[0106] Then the output and key historical flight data Input to the GRU temporal learning module to capture temporal dependencies based on the gated recurrent unit;

[0107]

[0108] Where, Θ is the Hadamard symbol,; Represents the output of the update gate at the i-th moment; represents the output of the reset gate at the i-th moment; is the hidden state at the i-1th moment; is the hidden state at the i-th moment; is the candidate hidden state at the i-th moment; Wxr , W xz , W xh , W hr , W hz and W hh are learnable parameters; b r 、b z and b h are bias parameters respectively;

[0109] After the above calculation, the time domain convolution output at the m-1th moment is obtained Among them, is the learnable parameter, b γ is the bias parameter;

[0110] (a.5) Then according to the time domain convolution output at the m-1th moment The key historical flight data set at the m-1th moment Calculate the mean square error to get the total loss; based on the total loss, the learnable parameter W xr , W xz , W xh , W hr , W hz , W hh and W γ And the bias parameter b r 、b z 、b h and b γ The network is updated and the state of the entire network when the total loss is minimum is recorded and saved as the optimal model. The optimal model is then used as the pre-trained game decision model and saved in the dynamic game decision module 7.

[0111] The dynamic game decision module 7 is used to use the pre-trained game decision model to predict the pre-processed current flight data, obtain the predicted flight data at the next moment, and generate a state forecast based on it; and manually confirm and adjust the state forecast to generate corresponding decision instructions and send them to the planning executor 8.

[0112] The predicted flight data at the next moment can be used to update the state information of each individual drone at the next moment, which is the state forecast; including: position: adjust the expected position; speed: update the speed according to the current mission requirements and environmental information; energy level: reflect the power and endurance of the drone to ensure the effective management of resources during the mission execution; other key parameters: such as sensor status, mission priority, etc., accurately reflect the dynamic environment.

[0113] The planning executor 8 is used to use the internal execution module 10 to receive the decision instructions and send corresponding operation instructions to achieve the synergy effect and adaptive adjustment planning of the drone cluster.

[0114] Decision instructions include: Synergy effect: comprehensively consider the synergy effect of drones to optimize task allocation and paths; Task allocation: dynamically adjust tasks according to drone status and cluster needs to improve efficiency; Dynamic adjustment: allow real-time adjustment to flexibly respond to environmental changes and emergencies.

[0115] Assign location tasks to the best coordinated drones; put drones that are not involved in the coordination into low-power mode and on standby to respond to emergencies. Adjust the speed of the drones to an appropriate level based on the assigned tasks and locations.

[0116] The above description is only a preferred 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 in the scope of protection of the present invention.

Claims

1. A ground-air coordinated UAV dynamic game decision-making management planning system, characterized by: The system includes: a drone array state sensor group located in the air, a decision management center and a planning executor located on the ground; the decision management center located on the ground includes a sensor database, a data preprocessing module, a spatiotemporal graph module and a dynamic game decision module; The drone array state sensor group located in the air is used to detect the real-time position and individual state of the drone array, and transmit the current flight data obtained by the new monitoring to the sensor database in the decision-making management center located on the ground; The drone array consists of N drones, each of which is equipped with a corresponding state sensor; The sensor database is used to receive the current flight data newly monitored and store the historical flight data, and then transfer the current flight data and the historical flight data to the data preprocessing module; The data preprocessing module is used to perform preprocessing operations on the current flight data and the historical flight data respectively, wherein the preprocessing operations include missing value filling and data cleaning, to obtain the preprocessed current flight data and the historical flight data, and then to transfer the preprocessed historical flight data to the space-time graph module, and to transfer the preprocessed current flight data to the dynamic game decision module; The spatiotemporal graph module is used to perform mutual information evaluation on the pre-processed historical flight data to screen variables, obtain the adjacency matrix using the maximum nearest neighbor method, train the spatiotemporal graph model using the GCN graph learning module and the GRU time series learning module, and save the optimal model as the pre-trained game decision model and save it in the dynamic game decision module; The dynamic game decision module is used to use the pre-trained game decision model to predict the pre-processed current flight data, obtain the predicted flight data at the next moment, and generate a state forecast based on it; and manually confirm and adjust the state forecast to generate corresponding decision instructions and send them to the planning executor; The planning executor is used to use the internal execution module to receive the decision instructions and send corresponding operation instructions to achieve the synergy effect and adaptive adjustment planning of the drone cluster.

2. According to claim 1, a ground-air coordinated UAV dynamic game decision-making management planning system is characterized in that: The sensor database receives the current flight data newly monitored and the stored historical flight data, specifically: The sensor database receives the current flight data newly monitored, and the current flight data is in, is the flight data sent back by the state sensor on the jth UAV in the UAV array at the mth moment, j = 1, 2, …, j, …, N, m represents the current moment; The sensor database stores historical flight data, which is in, is the flight data set sent back by the state sensor on the jth UAV in the UAV array at time 0 to m-1, is the flight data sent back by the j-th UAV at the i-th moment.

3. The ground-air coordinated UAV dynamic game decision-making management planning system according to claim 2 is characterized in that: The data preprocessing module performs preprocessing operations on the current flight data and the historical flight data respectively, and the preprocessing operations include missing value filling and data cleaning to obtain the preprocessed current flight data and historical flight data, specifically: The data preprocessing module converts historical flight data Perform preprocessing operations, including missing value filling and data cleaning, to obtain preprocessed historical flight data in, The historical flight data set after preprocessing operation for the jth UAV; The historical flight data of the j-th UAV after the preprocessing operation at the i-th time; The missing value filling is as follows: when the flight data sent back by the jth UAV at the i-th time Any variable in If there is a missing value, the missing value is filled by using the variables of the k moments before and after through mathematical calculation to obtain the variable after the missing value is filled. The formula is as follows: The data cleaning is specifically: deleting the flight data of each drone at every moment Complete data cleaning based on the erroneous data in the data; The data preprocessing module converts the current flight data into Perform preprocessing operations to obtain the preprocessed current flight data in, is the current flight data of the j-th UAV after the preprocessing operation at the m-th moment.

4. The ground-to-air coordinated UAV dynamic game decision-making management planning system according to claim 3 is characterized in that: The space-time graph module performs mutual information evaluation on the pre-processed historical flight data to screen variables, obtains the adjacency matrix using the maximum nearest neighbor method, trains the space-time graph model using the GCN graph learning module and the GRU time series learning module, and saves the optimal model as the pre-trained game decision model and saves it in the dynamic game decision module, specifically including the following sub-steps: (a.1) The key component of the spatiotemporal graph model is the spatiotemporal graph neural network, which is divided into three key elements: time domain, space domain, and graph structure; Graph structure: The construction of the graph includes two steps: screening drones that are highly relevant to flight missions, and determining the edge relationships between these drones and generating the adjacency matrix required by GCN; Spatial domain: Graph convolutional neural network is a neural network specifically designed to process graph structured data. It learns the interrelationships between different drone flight data from a spatial perspective. It captures the spatial dependencies between drones by propagating information between each drone and its neighbors in the graph; Time domain: The gated recurrent unit is a recurrent neural network used for time series analysis. It introduces a gating mechanism to control the information flow from the perspective of the time domain to capture the dynamic changes in the flight data time of each drone; First, the preprocessed historical flight data The mutual information value is obtained by evaluating the historical flight data set after preprocessing operations corresponding to any two drones A and B. The calculation formula is as follows: Where A=1,2,…,A,…,N,B=1,2,…,B,…,N,A≠B;αis the historical flight data set after preprocessing operation The number of grids divided in the direction, β is the historical flight data set after preprocessing operation The number of grids divided in the direction; for and The joint probability of for The marginal probability of for The marginal probability of; λ is the variable control parameter; Arrange the mutual information values ​​corresponding to all two arbitrary drones from large to small, select the historical flight data set after preprocessing corresponding to the first S drones with the largest mutual information values ​​as the key historical flight data set, and obtain the key historical flight data set. in, represents the g-th key historical flight data set, g = 1, 2, …, g, …, S; (a.2) Use the maximum nearest neighbor method to analyze key historical flight data Construct edge relationships and obtain edge weight matrix Among them, edge f,g Represents a collection of key historical flight data and key historical flight data sets The edge weight between them is f=1,2,…,f,…,S; when f≠g, Where ξ represents the bandwidth of the Gaussian kernel function; when f = g, edge f,g =0; Key historical flight data The similarity is used to obtain each key historical flight data set The K other key historical flight data sets with the greatest similarity are in the edge weight matrix Only keep key historical flight data sets The edge weights of the corresponding K other key historical flight data sets are changed to 0, and the graph structure is obtained. Then the graph structure Converted to adjacency matrix structure It is best to get key historical flight data The picture in, Shows key historical flight data The time step relationship in (a.3) Then the figure Input into the GCN graph learning module, use the graph convolution network to perform convolution calculations, and get the output of the GCN graph learning module Among them, σ(·) is the activation function; U is given by The Laplace calculation results are: Where D is the graph The node degree matrix of g θ Is the convolution kernel function: Among them, θ k is the polynomial parameter, T k are the Chebyshev polynomial parameters, For key historical flight data The calculated diagonal matrix; (a.4) Output It consists of sub-outputs from time 0 to time t-1: in, Represents the output at the i-th moment; And key historical flight data Divided from time 0 to time t-1, divided into in, represents the key historical flight data set at the i-th moment; Then the output and key historical flight data Input to the GRU temporal learning module to capture temporal dependencies based on the gated recurrent unit; Where, Θ is the Hadamard symbol,; Represents the output of the update gate at the i-th moment; represents the output of the reset gate at the i-th moment; is the hidden state at the i-1th moment; is the hidden state at the i-th moment; is the candidate hidden state at the i-th moment; W xr , W xz , W xh , W hr , W hz and W hh are learnable parameters; b r 、b z and b h are bias parameters respectively; After the above calculation, the time domain convolution output at the m-1th moment is obtained Among them, is the learnable parameter, b γ is the bias parameter; (a.5) Then according to the time domain convolution output at the m-1th moment The key historical flight data set at the m-1th moment Calculate the mean square error to get the total loss; based on the total loss, the learnable parameter W xr , W xz , W xh , W hr , W hz , W hh and W γ And the bias parameter b r , b z , b h and b γ The network is updated and the state of the entire network when the total loss is minimum is recorded and saved as the optimal model. The optimal model is then used as the pre-trained game decision model and saved in the dynamic game decision module.

Images