Topology reconstruction method and device for unmanned aerial vehicle cluster

Abstract

The invention provides a topology reconstruction method and device for an unmanned aerial vehicle cluster, and the method comprises the steps: constructing a network topology structure corresponding to a communication node of a target unmanned aerial vehicle cluster in real time based on a virtual node; performing feature transformation by using the graph attention network model to obtain structure perception embedded features; outputting an importance evaluation result of each communication node based on the structure sensing embedded features by using the policy network model; in response to detection of existence of the attacked damaged communication nodes, selecting key communication nodes from the damaged communication nodes, and selecting redundant communication nodes from normal communication nodes which are not attacked; and under the condition of ensuring the connectivity of the network topology structure, performing position movement replacement and iteration on the redundant communication nodes and the key communication nodes until the damaged communication nodes are all moved, replaced or recovered, thereby obtaining a reconstructed network topology structure. Through the method, the stability and the accuracy of topology reconstruction of the unmanned aerial vehicle cluster are improved.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) swarm technology, and in particular to a method and apparatus for topology reconstruction of UAV swarms. Background Technology

[0002] Drone swarms can achieve collaborative flight and task division based on wireless ad hoc networks. However, in complex and harsh environments, the communication nodes or links of drone swarms are susceptible to interference, damage, or failure, which can lead to network fragmentation and affect the overall mission execution. Therefore, drone swarm systems need to have the ability to quickly restore network topology and connectivity after damage, so as to ensure that drone swarm systems have the ability to withstand attacks or certain events and improve stability.

[0003] Currently, mainstream UAV swarm topology reconstruction technologies include topology reconstruction methods based on power control, node movement, and clustering structures. These methods rely on heuristic or rule-driven approaches for network recovery. Network recovery metrics only reflect the local or static characteristics of nodes and cannot fully characterize the true impact of nodes on resilience and connectivity within the entire topology. Furthermore, relying solely on distance or resource redundancy for allocation during topology repair can lead to unstable network recovery results and even introduce new link breakage issues. In addition, most traditional methods can only handle direct connections between nodes or one-hop neighbor information, easily overlooking the regulatory effect of the global structure on node ordering and reconstruction behavior.

[0004] In summary, existing UAV swarm topology reconstruction methods still suffer from problems such as blind strategy, weak structure awareness, and low information utilization when facing complex, highly dynamic, high-risk, and high-robust network topology damage scenarios in real-world environments. These issues reduce the stability and accuracy of UAV swarm topology reconstruction and make it difficult to fully meet the intelligent reconstruction needs of highly dynamic swarm systems. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method and apparatus for topology reconstruction of UAV swarms. The method constructs a network topology structure by combining virtual nodes with the corresponding communication network of the UAV swarm. A graph attention network model is used to perform feature transformation on the network topology structure to obtain structure-aware embedded features. A policy network model is then used to evaluate the importance of these embedded features, determining the importance of each UAV's corresponding communication node in the network topology. When a damaged communication node appears in the UAV swarm, based on the importance evaluation results, key nodes are prioritized for recovery, and structural connections are dynamically adjusted. By introducing a structure-aware node importance ranking mechanism and an alternative topology reconstruction strategy, the method achieves the identification and adaptive recovery of damaged communication nodes. In complex, highly dynamic, and strongly interfering network environments, this method improves the stability and accuracy of topology reconstruction for UAV swarms, effectively enhancing the network connectivity and system robustness of the UAV swarm.

[0006] This application provides a method for topology reconstruction of a drone swarm, the method comprising: Based on the real-time communication nodes corresponding to each drone in the target drone cluster and the preset virtual nodes, the network topology corresponding to the target drone cluster is constructed in real time. The network topology is transformed using a pre-defined graph attention network model to obtain the structure-aware embedding features corresponding to the network topology. Using a pre-trained policy network model based on the structure-aware embedding features, the importance assessment result of each communication node in the network topology is output. In response to the detection that at least one compromised communication node has been attacked, based on the importance assessment result, a critical communication node is selected from the compromised communication nodes, and a redundant communication node is selected from the normal communication nodes that have not been attacked. When the connectivity of the network topology meets the preset conditions, the redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions, and the next redundant communication node is iteratively selected to move and replace the next key communication node until all the damaged communication nodes are moved, replaced or restored, so as to obtain the reconstructed network topology corresponding to the target UAV cluster.

[0007] Furthermore, the real-time construction of the network topology corresponding to the target drone cluster based on the real-time communication node corresponding to each drone in the target drone cluster and the preset virtual node includes: The system acquires real-time location data and transmission signal power data for each drone in the target drone cluster in real time, and determines the received signal power data for each drone based on the real-time location data and the transmission signal power data, corresponding to the received signal power data of the other drones in the target drone cluster that each drone receives. Determine whether the received signal power data between each pair of drones in the target drone cluster is greater than a preset power threshold to obtain a power determination result; A communication link is established between two drones in the target drone cluster whose power judgment result is greater than a certain value. The real-time location data corresponding to each drone is used as a communication node, and the communication link is used as a communication path to construct the adjacency network structure corresponding to the target drone cluster in real time. A one-way edge connection is established between the preset virtual node and each of the communication nodes to obtain the virtual communication path between the virtual node and each of the communication nodes. Based on the adjacency network structure, the virtual node and the virtual communication path, the network topology structure corresponding to the target UAV cluster is constructed in real time.

[0008] Furthermore, the structure-aware embedding features include node embedding features corresponding to each communication node and graph-level embedding features corresponding to the virtual node; the step of using a preset graph attention network model to perform feature transformation on the network topology to obtain the structure-aware embedding features corresponding to the network topology includes: Based on the initial node features corresponding to each communication node and the initial graph features corresponding to each virtual node, the initial node attribute features corresponding to each communication node and the initial graph attribute features corresponding to each virtual node are obtained using a preset normalized feature encoder. Each of the multiple graph convolutional layers in the preset graph attention network model performs convolution operations on the initial node attribute features and the initial graph attribute features, respectively, to obtain the first convolution output result of each communication node in each graph convolutional layer and the second convolution output result of each virtual node in each graph convolutional layer. For each of the communication nodes, the initial node attribute features corresponding to the communication node are concatenated with the first convolution output results corresponding to the communication node in each graph convolutional layer to obtain the node embedding features corresponding to each of the communication nodes. The initial graph attribute features corresponding to the virtual node are concatenated with the second convolution output results corresponding to the virtual node in each graph convolutional layer to obtain the graph-level embedding features corresponding to the virtual node.

[0009] Furthermore, the method of using a pre-trained policy network model based on the structure-aware embedding features to output an importance assessment result for each communication node in the network topology includes: The structure-aware embedding features are input into a pre-trained policy network model, and the policy network model performs an outer product operation on the node embedding features and graph-level embedding features in the structure-aware embedding features to obtain the dimensional interaction features between the node embedding features and the graph-level embedding features. The dimensional interaction features are vector-mapped using a preset first learnable parameter to obtain a dimensional interaction feature vector. The deep Q-network model set by the policy network model is activated and linearly transformed on the dimensional interaction feature vector based on a preset second learnable parameter, and outputs the Q value corresponding to each communication node; wherein, the Q value represents the expected reward value brought about by selecting the communication node as the action corresponding to the node embedding feature for removal in the state corresponding to the graph-level embedding feature. The Q value corresponding to each of the communication nodes is determined as the importance assessment result of each of the communication nodes in the network topology.

[0010] Furthermore, in response to detecting that at least one compromised communication node among the communication nodes has been attacked, based on the importance assessment result, selecting a critical communication node from the compromised communication nodes and selecting redundant communication nodes from the unattacked normal communication nodes, includes: In response to the detection that at least one compromised communication node among the communication nodes has been attacked, the importance ranking of the communication nodes is determined based on the descending order corresponding to the importance assessment results. Among the damaged communication nodes, the damaged communication node that ranks first in importance is selected to identify the damaged communication node as a critical communication node. Among the normal communication nodes that have not been attacked, select the normal communication node whose importance assessment result is less than that of the critical communication node, and determine the normal communication node as a candidate communication node. The candidate communication node whose importance ranking result is the last among the candidate communication nodes is selected to determine the candidate communication node as a redundant communication node.

[0011] Furthermore, when the connectivity of the network topology meets preset conditions, moving and replacing the redundant communication nodes and the key communication nodes at corresponding positions includes: The redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions, and it is determined whether the connectivity result of the network topology under the operation of the movement and replacement is greater than or equal to a preset connectivity threshold. If the connectivity result is greater than or equal to a preset connectivity threshold, then the connectivity of the network topology is determined to meet the preset conditions under the move-replace operation.

[0012] Furthermore, the step of moving and replacing the redundant communication nodes and the key communication nodes at corresponding positions when the connectivity of the network topology meets preset conditions further includes: If the connectivity result is less than the preset connectivity threshold, then the move and replacement operation is restored, the redundant communication node is removed from the candidate communication nodes, and the candidate communication node with the last position in the importance ranking result is selected from the updated candidate communication nodes to determine the candidate communication node as the updated redundant communication node. The updated redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions until the connectivity result is greater than or equal to a preset connectivity threshold.

[0013] This application embodiment also provides a topology reconfiguration device for a drone swarm, the topology reconfiguration device comprising: The network topology module is used to construct the network topology of the target drone cluster in real time based on the real-time communication nodes corresponding to each drone in the target drone cluster and the preset virtual nodes. The feature transformation module is used to perform feature transformation on the network topology using a preset graph attention network model to obtain the structure-aware embedding features corresponding to the network topology. The node evaluation module is used to output the importance evaluation result of each communication node in the network topology based on the structure-aware embedding features using a pre-trained policy network model. The damage detection module is used to, in response to detecting that at least one damaged communication node has been attacked, select a critical communication node from the damaged communication nodes based on the importance assessment result, and select a redundant communication node from the normal communication nodes that have not been attacked. The topology module is used to move and replace the redundant communication nodes and the key communication nodes at corresponding positions when the connectivity of the network topology meets the preset conditions, and iteratively select the next redundant communication node to move and replace the next key communication node until all the damaged communication nodes are moved, replaced or restored, so as to obtain the reconstructed network topology corresponding to the target UAV cluster.

[0014] This application also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the topology reconfiguration method for UAV swarm described above are performed.

[0015] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the above-described method for topology reconfiguration of a drone swarm.

[0016] The topology reconstruction method and apparatus for a drone swarm provided in this application include: constructing a network topology corresponding to the target drone swarm in real time based on the real-time communication nodes corresponding to each drone in the target drone swarm and preset virtual nodes; performing feature transformation on the network topology using a preset graph attention network model to obtain structure-aware embedding features corresponding to the network topology; outputting an importance assessment result for each communication node in the network topology based on the structure-aware embedding features using a pre-trained policy network model; in response to detecting at least one damaged communication node that has been attacked, selecting a key communication node from the damaged communication nodes based on the importance assessment result, and selecting redundant communication nodes from the normal communication nodes that have not been attacked; when the connectivity of the network topology meets preset conditions, moving and replacing the redundant communication nodes and the key communication nodes at corresponding positions, and iteratively selecting the next redundant communication node to move and replace the next key communication node, until all damaged communication nodes have been moved and replaced or restored, to obtain the reconstructed network topology corresponding to the target drone swarm.

[0017] Compared with existing topology reconstruction methods based on power control, node movement, and clustering structures, this method constructs a network topology by combining virtual nodes with the communication network corresponding to the UAV swarm. It then uses a graph attention network model to perform feature transformation on the network topology, obtaining structure-aware embedded features. A policy network model is used to evaluate the importance of these embedded features, determining the importance of each UAV's corresponding communication node in the network topology. When a damaged communication node appears in the UAV swarm, based on the importance evaluation results, key nodes are prioritized for recovery, and structural connections are dynamically adjusted. By introducing a structure-aware node importance ranking mechanism and an alternative topology reconstruction strategy, it achieves the identification and adaptive recovery of damaged communication nodes. In complex, highly dynamic, and highly interfering network environments, this method improves the stability and accuracy of topology reconstruction for UAV swarms, effectively enhancing network connectivity and system robustness.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a topology reconstruction method for a drone swarm provided in an embodiment of this application; Figure 2 This is a schematic diagram illustrating a network topology structure provided in an embodiment of this application; Figure 3 This is a schematic diagram of a reconstructed network topology provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a topology reconfiguration device for a drone swarm provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0022] In the fields of power distribution networks and computer networks, topology reconfiguration technology is mainly used to deal with sudden problems such as node failures and link failures. By adjusting the network structure, it enhances the robustness of the communication system, thereby ensuring the continuous and stable operation of the network. With the development of wireless communication and multi-agent systems, topology reconfiguration technology has gradually expanded to the field of unmanned aerial vehicle (UAV) swarm systems.

[0023] Unmanned aerial vehicle (UAV) swarms rely on wireless ad hoc networks for coordinated flight and task allocation. However, in complex and harsh environments, their communication nodes or links are susceptible to interference, damage, or failure, leading to network fragmentation and impacting overall mission execution. Therefore, UAV swarm systems must possess the ability to rapidly restore network topology and connectivity after damage to ensure the system's resilience and continuous operational capability in the face of attacks or other events.

[0024] Research has revealed that current mainstream UAV swarm topology reconfiguration technologies include topology reconfiguration methods based on power control, node movement, and clustering structures. However, when faced with complex, dynamic, and high-risk network topology damage scenarios in real-world environments, these methods still have the following key shortcomings, making it difficult to fully meet the intelligent reconfiguration needs of highly dynamic swarm systems.

[0025] Existing methods lack precise identification and prioritization mechanisms for critical nodes, resulting in highly unreliable structural recovery. These methods generally rely on heuristic or rule-driven approaches for network recovery, such as using static metrics like node degree, connectivity, shortest path, and coverage. However, these metrics only reflect the local or static characteristics of nodes and cannot fully characterize the true impact of nodes on the overall topology in terms of resilience and connectivity. For example, some bridging nodes may have low degrees locally, but play a crucial role in maintaining overall network connectivity. Failure to prioritize the repair of such nodes may lead to limited or even failed network recovery.

[0026] Existing topology reconstruction strategies lack structural constraints, resulting in highly random recovery behavior. Traditional methods typically schedule neighboring nodes to move or adjust power based on predefined rules during topology repair, making it difficult to form a generalizable and structurally controllable reconstruction mechanism. For example, some methods do not consider whether the movement of replacement nodes helps restore the overall network connectivity when selecting replacement nodes, relying solely on distance or resource redundancy for allocation. This approach leads to unstable recovery results and may even introduce new link breakage issues.

[0027] Existing methods have limited sensing range and struggle to capture global structural relationships within the topology. Most traditional methods can only handle direct connections between nodes or one-hop neighbor information, making it difficult to learn complex topological nesting relationships from multi-hop structures. In large-scale UAV swarms or irregular topologies, structural coupling is high, and local policy adjustments may cause chain reactions in remote structures. Therefore, if policy formulation is based solely on local information, it is easy to overlook the regulatory effect of the global structure on node ordering and reconstruction behavior.

[0028] In summary, existing UAV swarm topology reconstruction methods still suffer from problems such as blind strategy, weak structure awareness, and low information utilization when facing complex, highly dynamic, high-risk, and high-robust network topology damage scenarios in real-world environments. These issues reduce the stability and accuracy of UAV swarm topology reconstruction.

[0029] Based on this, this application provides a topology reconstruction method for UAV swarms. It constructs a network topology structure by combining virtual nodes with the corresponding communication network of the UAV swarm. A graph attention network model is used to perform feature transformation on the network topology structure to obtain structure-aware embedded features. A policy network model is then used to evaluate the importance of these embedded features, determining the importance of each UAV's corresponding communication node in the network topology. When a damaged communication node appears in the UAV swarm, based on the importance evaluation results, key nodes are prioritized for recovery, and structural connections are dynamically adjusted. By introducing a structure-aware node importance ranking mechanism and an alternative topology reconstruction strategy, the method achieves the identification and adaptive recovery of damaged communication nodes. In complex, highly dynamic, and strongly interfering network environments, this method improves the stability and accuracy of topology reconstruction for UAV swarms, effectively enhancing the network connectivity and system robustness of the UAV swarm.

[0030] Please see Figure 1 , Figure 1 This is a flowchart illustrating a topology reconstruction method for a drone swarm provided in an embodiment of this application. Figure 1 As shown in the figure, the topology reconstruction method for a drone swarm provided in this application includes: S101. Based on the real-time communication nodes corresponding to each drone in the target drone cluster and the preset virtual nodes, construct the network topology structure corresponding to the target drone cluster in real time.

[0031] In this step, the target drone cluster in three-dimensional space is modeled, with each drone serving as a communication node. The modeling method is used to obtain communication links to establish an adjacency network structure, representing the actual connection relationships in the target drone cluster and forming the original communication network.

[0032] Furthermore, in previous studies, global features were mostly obtained by averaging the features of all communication nodes. Here, the embodiments of this application adopt the concept of virtual nodes. All communication nodes in the adjacency network structure form a one-way connection with the virtual node. That is, the communication node only transmits information to the virtual node, but does not receive information from the virtual node. The virtual node does not affect the graph topology of the adjacency network structure and plays the role of extracting global features.

[0033] In one possible implementation of this application, step S101 may include: S1011. Real-time location data and transmission signal power data of each drone in the target drone cluster are acquired in real time, and based on the real-time location data and the transmission signal power data, the received signal power data corresponding to the transmission signals received by each drone from other drones in the target drone cluster are determined.

[0034] It should be noted that, in order to model and optimize the communication network of the target UAV cluster, it is necessary to construct a topology model of the target UAV cluster in three-dimensional space that conforms to the actual deployment characteristics. In this embodiment, the physical location information and communication capabilities of the UAVs distributed in three-dimensional space are considered to establish a communication connectivity graph (i.e., network topology) in three-dimensional geometric space so as to facilitate the input processing of the graph attention network model and policy network model.

[0035] In this embodiment of the application, a three-dimensional Cartesian coordinate system (XYZ) is established in the target drone cluster, and the expression for the real-time position data of each drone in the target drone cluster is as follows.

[0036] .

[0037] in, Indicates the first A drone in time step Corresponding real-time location data; They represent the first A drone in time step exist , , Components on the axis; This represents the number corresponding to the drone cluster, with a total number of N drones.

[0038] Furthermore, each drone can transmit signals to other drones at a fixed power level. In the initial state, the target drone cluster forms a connected communication network, in which any two drones can transmit data through a multi-hop communication link, and the communication channel between drones is modeled as an air-to-air communication link.

[0039] In this embodiment of the application, the received signal power data corresponding to the transmitted signals received by each drone from other drones in the target drone cluster is determined by the following formula.

[0040] .

[0041] in, Indicates at each time step , No. The drone received the first The received signal power data corresponding to the signal transmitted by the drone; Indicates the antenna gain at the receiving end; This indicates the antenna gain at the transmitting end; This indicates the preset transmission signal power data for each drone; Indicates the first A drone in time step Corresponding real-time location data; Indicates the first A drone in time step Corresponding real-time location data; This indicates a large-scale fading effect; This indicates a small-scale fading effect.

[0042] Here, since there is no ground obstruction in the network of the drone swarm, the large-scale fading effect is represented as follows.

[0043] .

[0044] in, This indicates a large-scale fading effect; This represents the preset path loss index (usually greater than 0). Indicates the preset electromagnetic wave frequency; This represents the speed of light. Indicates the first A drone in time step With the The relative distance between the drones.

[0045] Furthermore, the small-scale fading effect is modeled using the Rice distribution, and the representation of the small-scale fading effect is shown below.

[0046] .

[0047] in, This indicates a small-scale fading effect; Indicates the first A drone in time step With the The relative distance between the drones; Rice factor; Indicates the strength of the main path; Indicates the scattering path intensity; This represents the first kind of 0th order modified Bessel function.

[0048] Here, the Rice factor The representation of is shown below.

[0049] .

[0050] S1012. Determine whether the received signal power data between each pair of drones in the target drone cluster is greater than a preset power threshold, and obtain the power determination result.

[0051] It should be noted that, since the received signal power is only related to the relative distance between the two drones, the definition is... Therefore, the received signal power data has symmetry, that is: At each time step , No. The drone received the first The received signal power data corresponding to the transmitted signal of the drone and the first The drone received the first The received signal power data corresponding to the signals transmitted by the drones are the same.

[0052] In this step, the received signal power data of each pair of drones in the target drone cluster are compared to determine whether the received signal power data is greater than a preset power threshold, and the power judgment result is obtained.

[0053] S1013. Establish a communication link between two drones in the target drone cluster whose power judgment result is greater than a certain value, and use the real-time location data corresponding to each drone as a communication node and the communication link as a communication path to construct the adjacency network structure corresponding to the target drone cluster in real time.

[0054] Here, when establishing a communication link between two drones in a target drone swarm, based on the concept that power judgment and received signal power are only related to the relative distance between the two drones, for any two drones to establish a communication link, the relative distance between the two drones should satisfy the following expression for the conditions for establishing a communication link.

[0055] .

[0056] in, Indicates the first The drone and the first The relative distance between the drones; Indicates the strength of the main path; Indicates the scattering path intensity; This represents a first-order modified Bessel function of the 0th order; Indicates the preset electromagnetic wave frequency; This represents the speed of light. Rice factor; Indicates the antenna gain at the receiving end; This indicates the antenna gain at the transmitting end; This indicates the preset transmission signal power data for each drone; This indicates the preset path loss index; This indicates the preset power threshold.

[0057] In this embodiment of the application, any two drones that establish a communication link are called "neighbors". The real-time location data of each drone is used as the communication node and the communication link is used as the communication path to build the adjacency network structure corresponding to the target drone cluster in real time.

[0058] Here, the adjacency network structure is used to define the information transmission path between nodes in the graph structure. When the UAV is in dynamic flight, the real-time location data is updated over time, and the adjacency network structure corresponding to the target UAV cluster also changes with the change of the UAV's real-time location data.

[0059] S1014. Establish unidirectional edge connections between the preset virtual nodes and each of the communication nodes to obtain the virtual communication paths between the virtual nodes and each of the communication nodes, and construct the network topology corresponding to the target UAV cluster in real time based on the adjacency network structure, the virtual nodes and the virtual communication paths.

[0060] In this embodiment of the application, in order to improve the expressive power and global perception capability of the communication network corresponding to the UAV in the graph attention network model, a virtual node aggregation mechanism is introduced to enhance the non-local information propagation capability between nodes in the graph structure and provide more stable full-graph context features in the policy network model.

[0061] In the actual communication network of UAV swarms, the connection between communication nodes is usually limited by the communication radius, resulting in a graph structure with strong locality and significant asymmetry. When graph neural networks rely solely on the transmission of information from local neighbors, the model depth is limited and the receptive field is restricted, making it difficult to capture the overall structural features and global information of the network. To address this, a virtual aggregation node is introduced to serve as a hub for information fusion and global perception by establishing edge connections with all real communication nodes.

[0062] In the network topology structure corresponding to the target drone cluster constructed in real time, the information of all communication nodes is aggregated into virtual nodes, but the virtual nodes themselves do not propagate information to other communication nodes. That is, the information flow of the network topology structure is only from the adjacent network structure to the virtual nodes, rather than a bidirectional information flow.

[0063] For example, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating a network topology structure provided in an embodiment of this application. Figure 2 As shown, a three-dimensional spatial model of the target drone cluster is performed. A communication link is established between two drones in the target drone cluster whose power judgment result is greater than a certain value. Then, the adjacency network structure corresponding to the target drone cluster is established. Finally, a unidirectional edge connection is established between the preset virtual node and each communication node to construct the network topology structure corresponding to the target drone cluster.

[0064] S102. The network topology is transformed using a preset graph attention network model to obtain the structure-aware embedding features corresponding to the network topology.

[0065] In this embodiment, the structure-aware embedding feature includes the node embedding feature corresponding to each communication node and the graph-level embedding feature corresponding to the virtual node.

[0066] In this step, the initial node features corresponding to the communication nodes and the initial graph features corresponding to the virtual nodes in the network topology are input into a preset graph attention network model. The graph attention network model first increases the dimension of the node features through a preset normalized feature encoder. Then, the node embedding features corresponding to each communication node and the graph-level embedding features corresponding to the virtual nodes in the network topology are extracted through multiple graph convolutional layers set by the graph attention network model.

[0067] In one possible implementation of this application, step S102 may include: S1021. Based on the initial node features corresponding to each communication node and the initial graph features corresponding to the virtual node, the initial node attribute features corresponding to each communication node and the initial graph attribute features corresponding to the virtual node are obtained using a preset normalized feature encoder.

[0068] In this embodiment of the application, the expression for the preset normalized feature encoder to process the initial node features corresponding to each communication node and the initial graph features corresponding to each virtual node is as follows.

[0069] .

[0070] in, This represents the initial node attribute characteristics corresponding to each communication node or the initial graph attribute characteristics corresponding to each virtual node. Indicates the first Initial node features corresponding to each communication node or initial graph features corresponding to each virtual node; and This represents the weights corresponding to the normalized feature encoder; and This represents the basic coefficients corresponding to the normalized feature encoder.

[0071] Here, when Indicates the first When the initial node characteristics corresponding to each communication node are... That is, initial node features Including the The three-dimensional location coordinates, node degree, and remaining power of each communication node; when When representing the initial graph features corresponding to virtual nodes, the initial graph features include the adjacency network structure, virtual nodes, and virtual communication paths.

[0072] S1022. Using a preset graph attention network model, each of the multiple graph convolutional layers performs convolution operations on the initial node attribute features and the initial graph attribute features to obtain the first convolution output result corresponding to each communication node in each graph convolutional layer and the second convolution output result corresponding to each virtual node in each graph convolutional layer.

[0073] It should be noted that the core of the graph attention network model lies in the aggregation and updating of information. Two key issues need to be considered: how to select nodes or edges for information aggregation, and how to define the aggregation function to achieve the updating of node states.

[0074] In this embodiment, a random sampling method is used to select which neighboring node information should be aggregated. This method helps to improve the efficiency of the model when dealing with large-scale drone swarms, as well as the model's generalization ability to swarms of different sizes. In addition, since the importance of neighboring nodes in information aggregation varies with task and location, an attention mechanism is introduced to measure the importance of neighbors to the current node, and a dot product attention mechanism is used as part of the convolution kernel to process the graph structure.

[0075] Here, for communication nodes and its "neighbor" communication nodes The attention weights for these two communication nodes are defined as follows.

[0076] .

[0077] in, Represents communication node and its "neighbor" communication nodes The corresponding attention weights, i.e., communication nodes For communication nodes Importance weights; and Representing communication nodes and communication nodes Initial node attribute characteristics; For communication nodes The set of "neighbor" communication nodes; and These represent the corresponding weights.

[0078] Furthermore, communication nodes The aggregation feature representation of the "neighbor" communication nodes is shown below.

[0079] .

[0080] in, Represents communication node The aggregation characteristics of "neighbor" communication nodes; For communication nodes The set of "neighbor" communication nodes; Represents communication node Initial node attribute characteristics; Represents communication node and its "neighbor" communication nodes The corresponding attention weights; This indicates the weight of the "neighbor" communication node.

[0081] Furthermore, in terms of node information updates, each of the multiple graph convolutional layers in the preset graph attention network model adopts an independent multi-head attention mechanism. The multi-head attention mechanism is an enhanced form of attention mechanism. Its core idea is to use multiple independent attention heads simultaneously to learn the attention relationships between nodes in different subspaces, thereby improving the model's ability to express diverse information in the graph structure. Compared with single attention, the multi-head mechanism can understand the relationship between neighboring nodes from multiple perspectives.

[0082] In this step, each graph convolutional layer is equipped with multiple attention heads. Each graph convolutional layer uses multiple attention heads and outputs the first convolutional output result corresponding to each communication node and the second convolutional output result corresponding to each virtual node based on the feature update function.

[0083] In this embodiment of the application, each graph convolutional layer communicates with nodes. The expression for the feature update function is shown below.

[0084] .

[0085] in, Indicates the communication node after information update The embedded features, namely, communication nodes The first convolution output result corresponding to each graph convolutional layer; It is the first in each graph convolutional layer The neighbor representation obtained by aggregating attention heads; This represents the concatenation operation of multiple attention head results in each graph convolutional layer; It is the number of attention heads in each graph convolutional layer; For communication nodes The set of "neighbor" communication nodes; This represents the initial node attribute characteristics corresponding to each communication node; Represents the weights of the feature update function; Represents the norm function; This represents the activation function.

[0086] Furthermore, based on the above expression, the second convolution output corresponding to the virtual node can be obtained similarly.

[0087] S1023. For each of the communication nodes, the initial node attribute features corresponding to the communication node are concatenated with the first convolution output results corresponding to the communication node in each graph convolutional layer to obtain the node embedding features corresponding to each of the communication nodes.

[0088] In this embodiment of the application, the representation of the node embedding feature corresponding to each communication node is as follows.

[0089] .

[0090] in, Represents communication node Corresponding node embedding features; This represents the initial node attribute characteristics corresponding to each communication node; Represents communication node In the The first convolution output result corresponding to the graph convolutional layer; This indicates a concatenation operation along the feature dimension.

[0091] S1024. The initial graph attribute features corresponding to the virtual node are concatenated with the second convolution output results corresponding to the virtual node in each graph convolutional layer to obtain the graph-level embedding features corresponding to the virtual node.

[0092] In this embodiment of the application, graph-level embedding features can be represented as a global semantic representation of the network topology, and the representation of the graph-level embedding features corresponding to virtual nodes is shown below.

[0093] .

[0094] in, This represents the graph-level embedding feature corresponding to the virtual node; Indicates the virtual node at the 1st The second convolution output result corresponding to the convolutional layer of each graph; This indicates a concatenation operation along the feature dimension. This represents the initial graph attribute characteristics corresponding to the virtual node.

[0095] S103. Using a pre-trained policy network model based on the structure-aware embedding features, output the importance assessment result of each communication node in the network topology.

[0096] It should be noted that, in order to identify key nodes in the topology reconstruction of UAV swarms, this embodiment of the application models the node ranking problem as a Markov decision process, constructs a policy network model to learn and reason about the importance of nodes, thereby providing an executable ranking basis for the subsequent topology reconstruction stage.

[0097] Specifically, ranking the importance of nodes in a drone swarm first requires determining an importance metric—that is, what evaluation standard should be used to assess the importance of a node in the network. Among numerous swarm tasks, network connectivity is a factor affecting task completion. Therefore, the maximally connected subgraph in the swarm network is chosen as the evaluation metric for network importance, which can be expressed in the following form.

[0098] .

[0099] in, Represents the largest connected subgraph in a cluster network; This represents the number of nodes in the largest connected subgraph in the original network; This indicates removal from the original network. The maximum number of nodes in a connected subgraph after a given node.

[0100] Thus, the connectivity of the cluster network can be evaluated using the metric of the maximum connected subgraph in the cluster network, thereby determining the importance of nodes. Therefore, the node importance ranking process can be modeled as a Markov decision process, and a quadruple can be used to determine the importance of nodes. The key components of the Markov decision-making process are represented in the form of .

[0101] in, Let be the state space, representing the topological state of the communication network during the failure evolution of the communication node corresponding to the UAV. Each state can correspond to a communication graph, i.e., ,in, Includes node connection information and its embedding features .

[0102] Let be the action space, representing the set of all drone nodes that are not yet invalid in the current state. The goal of the node importance ranking strategy is to select the most critical communication node from the set of all drone communication nodes that are not yet invalid and remove it.

[0103] The reward function is defined as the change in network connectivity after performing a certain action, reflecting the degree to which the communication node disrupts the overall network connectivity.

[0104] This is a terminated state, used to terminate the program when all drone communication nodes in the network fail or the specified removal limit is reached.

[0105] Thus, under the above definition, the node importance ranking process can be modeled as a "state-action-reward trajectory," that is, The goal is to learn a policy function that ensures that, in any state... Select action The goal is to maximize the cumulative reward by finding a sorted sequence of nodes to remove that results in the fastest decline in network connectivity, thereby identifying the most critical nodes with the greatest impact.

[0106] In this embodiment, the node embedding feature representation corresponding to each communication node is used as the semantic representation of the communication node encoded in the network topology, and is regarded as a candidate action. The graph-level embedding features corresponding to the virtual nodes are used as the state. The expression.

[0107] Furthermore, considering the inherently complex graph structure and long-range dependency characteristics of the node importance ranking problem in the pre-trained policy network model, an N-step temporal difference deep Q network (N-step DQN) is introduced as the core training algorithm in the policy network model framework. By introducing N-step temporal difference (N-step TD) learning, the learning process is accelerated and the guidance is reduced, thereby improving learning efficiency and final performance while maintaining training stability, thus enhancing the stability and convergence efficiency of policy evaluation.

[0108] Specifically, during the training of the policy network model, the trajectory of each N-step segment is... Periodically stored in the experience replay pool, and then randomly sampled from it for training, the loss function used to train the deep Q network model in the N-step temporal difference deep Q network is shown below.

[0109] .

[0110] in, This indicates the current deep Q-network model's state-action pairs. The predicted value; Indicates the target value; This represents the mean squared error loss between the predicted Q value and the N-step temporal difference, serving as a supervision signal during the training process.

[0111] Here, the target value The representation of is shown below.

[0112] .

[0113] in, , indicating the future Cumulative instant rewards; This is a discount factor used to balance short-term gains and long-term returns; This represents a target depth Q-network model used for stable target value estimation; Indicates the state The action selected below.

[0114] Furthermore, to enhance the model's ability to model the topology of UAV swarm communication networks, a structure-preserving reconstruction loss term is introduced during the training of the policy network. This ensures that the node embedding features learned by the graph attention network model retain the local connectivity relationships in the original graph structure as much as possible in space. The loss function corresponding to the reconstruction loss term is shown below.

[0115] .

[0116] in, Indicates the reconstruction loss item; This represents the total number of nodes in the graph; It is an element in the adjacency network structure, if the communication node With communication nodes If there is an edge connecting them, then =1, otherwise =0; and Communication nodes With communication nodes Node embedding features; To embed network parameters into the graph; It is the squared L2 norm, used to measure the Euclidean distance between the embedded feature vectors of two nodes.

[0117] Thus, the training objective of the policy network model should not only optimize the node ranking strategy, but also consider the structure preservation objective. Therefore, the expression of the joint loss function of the policy network model is as follows.

[0118] .

[0119] in, This represents the joint loss function of the policy network model; This represents the mean squared error loss between the predicted Q value and the multi-step N-step time series differencing. Indicates the reconstruction loss item; This indicates the weight of the reconstruction loss term.

[0120] In one possible implementation of this application, step S103 may include: S1031. The structure-aware embedding features are input into a pre-trained policy network model. The policy network model performs an outer product operation on the node embedding features and graph-level embedding features in the structure-aware embedding features to obtain the dimensional interaction features between the node embedding features and the graph-level embedding features.

[0121] In this embodiment, to capture the higher-order interaction relationship between node embedding features and graph-level embedding features, the policy network model does not use the conventional concatenation operation, but instead uses the outer product operation. The expression for the dimensional interaction features between node embedding features and graph-level embedding features obtained by the outer product operation is as follows: ;in, and Representing communication nodes The node embedding features and the graph-level embedding features of virtual nodes.

[0122] S1032. The dimensional interaction features are vector-mapped using a preset first learnable parameter to obtain a dimensional interaction feature vector.

[0123] In this embodiment of the application, the dimensional interaction features are vector-mapped using the following formula.

[0124] .

[0125] in, Represents the dimensional interaction feature vector; Represents dimensional interaction features; This represents the first preset learnable parameter.

[0126] S1033. Using the deep Q-network model set by the policy network model, based on the preset second learnable parameter, the dimensional interaction feature vector is activated and linearly transformed, and the Q value corresponding to each communication node is output.

[0127] The Q value represents the expected reward value resulting from selecting the communication node as the action corresponding to the node embedding feature and removing it under the state corresponding to the graph-level embedding feature.

[0128] In this embodiment of the application, the Q value corresponding to each communication node is calculated using the following formula.

[0129] .

[0130] in, Represents communication node The corresponding Q-value, that is, the state corresponding to the graph-level embedding feature, determines the selection of the communication node. The expected reward value resulting from removing the action corresponding to the node's embedded features; Represents the dimensional interaction feature vector; This represents the preset second learnable parameter.

[0131] S1034. The Q value corresponding to each of the communication nodes is determined as the importance assessment result of each of the communication nodes in the network topology.

[0132] In this step, a pre-trained policy network model outputs the Q-value of each communication node based on the node embedding features of each communication node and the graph-level embedding features of the virtual node. This reflects the importance of each communication node in the current network topology and determines the importance assessment result of each communication node in the network topology.

[0133] In this way, the network parameters of the graph attention network model and the policy network model are fixed, and no gradient updates are performed. They have efficient and fast reasoning capabilities and can adapt to dynamically changing UAV swarm networks. The policy network model enables the embodiments of this application to have real-time evaluation and decision-making capabilities, providing key support for communication self-healing and structural optimization in complex environments.

[0134] In summary, the embodiments of this application have good scalability and adaptability, and can be adapted to UAV network environments of different sizes and structures. The graph attention network model and policy network model can be continuously trained and updated to adapt to new attack patterns and network disturbances, thereby improving the network's self-healing ability and system robustness.

[0135] S104. In response to detecting that at least one damaged communication node among the communication nodes has been attacked, based on the importance assessment result, a key communication node is selected from the damaged communication nodes, and a redundant communication node is selected from the normal communication nodes among the communication nodes that have not been attacked.

[0136] In this step, by comprehensively considering the damage status of communication nodes and their importance in the network topology, a limited number of normal nodes are used to replace the damaged critical node positions in order to maximize the overall connectivity and stability of the network structure.

[0137] In one possible implementation of this application, step S104 may include: S1041. In response to detecting that at least one compromised communication node among the communication nodes has been attacked, the importance ranking result of the communication nodes is determined based on the descending order of the importance assessment result.

[0138] In this step, when at least one compromised communication node is detected to be under attack, the importance assessment results of each communication node are sorted in descending order, and the importance ranking result of the communication nodes is determined based on the descending order of the communication node ranking sequence.

[0139] S1042. Select the damaged communication node that ranks first in importance among the damaged communication nodes, and identify the damaged communication node as a critical communication node.

[0140] S1043. Select a normal communication node whose importance assessment result is less than that of the critical communication node from among the normal communication nodes that have not been attacked, and determine the normal communication node as a candidate communication node.

[0141] S1044. Select the candidate communication node whose importance ranking result is the last among the candidate communication nodes, so as to determine the candidate communication node as a redundant communication node.

[0142] S105. When the connectivity of the network topology meets the preset conditions, the redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions, and the next redundant communication node is iteratively selected to move and replace the next key communication node until all the damaged communication nodes are moved, replaced or restored, so as to obtain the reconstructed network topology corresponding to the target UAV cluster.

[0143] In this embodiment of the application, when the network of the target drone cluster is attacked and a damaged communication node appears, the redundant communication node with low importance is selected to replace the disguise of the key communication node with high importance according to the obtained node importance ranking result. The topology reconstruction and replacement operation is performed to balance the structural restoration effect and the cost of node replacement, so as to ensure that the network connectivity of the network topology is quickly restored and the final reconstructed network topology is formed.

[0144] In this step, redundant communication nodes are attempted to be moved and replaced with critical communication nodes at their corresponding positions. It is then determined whether this movement and replacement operation can improve or maintain the connectivity level of the network topology. If connectivity is improved or maintained, the movement and replacement operation is executed, and the sets corresponding to damaged and normal communication nodes are updated. If connectivity decreases, the redundant communication node is removed from the candidate communication nodes, and the process continues to attempt to select redundant communication nodes from the updated candidate communication nodes and perform movement and replacement operations. This process continues iteratively until all damaged communication nodes have been moved, replaced, or restored, forming a new reconstructed network topology. This approach balances node importance with structural adaptability, enabling effective adaptive reconstruction of the network topology under limited resource conditions.

[0145] For example, please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating a reconstructed network topology provided in an embodiment of this application. Figure 3As shown, in response to the detection of at least one damaged communication node that has been attacked, a key communication node is selected from the damaged communication nodes, and a redundant communication node is selected from the normal communication nodes that have not been attacked. The redundant communication node and the key communication node are moved and replaced at their corresponding positions to obtain the reconstructed network topology corresponding to the target UAV cluster.

[0146] In this way, an integrated training-inference-reconstruction self-healing process for UAV swarm networks was constructed, realizing a closed-loop operation from environment modeling, model learning, node sorting to topology reconstruction, and supporting online network structure optimization in complex battlefield environments.

[0147] Furthermore, by introducing a topology reconstruction strategy to achieve online recovery, in scenarios where attacks occur or communication nodes fail, key communication nodes are prioritized for recovery based on importance assessment results, and the network topology is dynamically adjusted to improve network connectivity and resilience, significantly enhancing the target UAV cluster's ability to continue executing missions under damaged conditions.

[0148] Furthermore, the embodiments of this application realize task-oriented network adaptive evolution, which ensures network connectivity while taking into account node resource consumption and structural adjustment costs, making the network topology reconstruction process more intelligent and efficient, meeting the real-time and stability requirements of UAV swarm collaborative communication in complex combat environments, and also has a universal and scalable strategy framework that can be applied to UAV swarms of different sizes and communication models, and has strong portability and engineering implementation foundation.

[0149] In one possible implementation of this application, in specific implementation, step S105, when the connectivity of the network topology meets preset conditions, involves moving and replacing the redundant communication nodes with the critical communication nodes at corresponding positions, and may include: S1051. Move and replace the redundant communication node and the key communication node at their corresponding positions, and determine whether the connectivity result of the network topology under the operation of the move and replacement is greater than or equal to a preset connectivity threshold.

[0150] S1052. If the connectivity result is greater than or equal to a preset connectivity threshold, then it is determined that the connectivity of the network topology meets the preset conditions under the move-replace operation.

[0151] Furthermore, step S105, which involves moving and replacing the redundant communication nodes with the critical communication nodes at corresponding positions when the connectivity of the network topology meets preset conditions, further includes: S1053. If the connectivity result is less than the preset connectivity threshold, then the move replacement operation is restored, the redundant communication node is removed from the candidate communication nodes, and the candidate communication node with the last position in the importance ranking result is selected from the updated candidate communication nodes to determine the candidate communication node as the updated redundant communication node.

[0152] S1054. Move and replace the updated redundant communication node and the key communication node at their corresponding positions until the connectivity result is greater than or equal to a preset connectivity threshold.

[0153] The topology reconstruction method for UAV swarms provided in this application constructs a network topology structure by combining virtual nodes with the communication network corresponding to the UAV swarm. It then uses a graph attention network model to perform feature transformation on the network topology structure to obtain structure-aware embedded features. A policy network model is used to evaluate the importance of these embedded features, determining the importance assessment result of the communication node corresponding to each UAV in the network topology structure. When a damaged communication node appears in the UAV swarm, based on the importance assessment result, key nodes are prioritized for recovery and the structural connections are dynamically adjusted. By introducing a structure-aware node importance ranking mechanism and an alternative topology reconstruction strategy, the method achieves the identification and adaptive recovery of damaged communication nodes. In diverse network environments such as complex, highly dynamic, and strongly interfered environments, it improves the stability and accuracy of topology reconstruction for UAV swarms, effectively enhancing the network connectivity and system robustness of the UAV swarm.

[0154] Please see Figure 4 , Figure 4 This is a schematic diagram of a topology reconfiguration device for a drone swarm provided in an embodiment of this application. Figure 4 As shown, the topology reconfiguration device 400 includes: The network topology module 410 is used to construct the network topology structure corresponding to the target drone cluster in real time based on the real-time communication node corresponding to each drone in the target drone cluster and the preset virtual node. The feature transformation module 420 is used to perform feature transformation on the network topology using a preset graph attention network model to obtain the structure-aware embedding features corresponding to the network topology. The node evaluation module 430 is used to output the importance evaluation result of each communication node in the network topology based on the structure-aware embedding features using a pre-trained policy network model. The damage detection module 440 is configured to, in response to detecting that at least one damaged communication node has been attacked, select a critical communication node from the damaged communication nodes based on the importance assessment result, and select a redundant communication node from the normal communication nodes that have not been attacked. The topology module 450 is used to move and replace the redundant communication nodes and the key communication nodes at corresponding positions when the connectivity of the network topology meets the preset conditions, and iteratively select the next redundant communication node to move and replace the next key communication node until all the damaged communication nodes are moved, replaced or restored, so as to obtain the reconstructed network topology corresponding to the target UAV cluster.

[0155] Furthermore, when the network topology module 410 is used to construct the network topology structure corresponding to the target drone cluster in real time based on the real-time communication node corresponding to each drone in the target drone cluster and the preset virtual node, the network topology module 410 is used to: The system acquires real-time location data and transmission signal power data for each drone in the target drone cluster in real time, and determines the received signal power data for each drone based on the real-time location data and the transmission signal power data, corresponding to the received signal power data of the other drones in the target drone cluster that each drone receives. Determine whether the received signal power data between each pair of drones in the target drone cluster is greater than a preset power threshold to obtain a power determination result; A communication link is established between two drones in the target drone cluster whose power judgment result is greater than a certain value. The real-time location data corresponding to each drone is used as a communication node, and the communication link is used as a communication path to construct the adjacency network structure corresponding to the target drone cluster in real time. A one-way edge connection is established between the preset virtual node and each of the communication nodes to obtain the virtual communication path between the virtual node and each of the communication nodes. Based on the adjacency network structure, the virtual node and the virtual communication path, the network topology structure corresponding to the target UAV cluster is constructed in real time.

[0156] Furthermore, the structure-aware embedding features include node embedding features corresponding to each communication node and graph-level embedding features corresponding to the virtual node; when the feature transformation module 420 performs feature transformation on the network topology using a preset graph attention network model to obtain the structure-aware embedding features corresponding to the network topology, the feature transformation module 420 is used to: Based on the initial node features corresponding to each communication node and the initial graph features corresponding to each virtual node, the initial node attribute features corresponding to each communication node and the initial graph attribute features corresponding to each virtual node are obtained using a preset normalized feature encoder. Each of the multiple graph convolutional layers in the preset graph attention network model performs convolution operations on the initial node attribute features and the initial graph attribute features, respectively, to obtain the first convolution output result of each communication node in each graph convolutional layer and the second convolution output result of each virtual node in each graph convolutional layer. For each of the communication nodes, the initial node attribute features corresponding to the communication node are concatenated with the first convolution output results corresponding to the communication node in each graph convolutional layer to obtain the node embedding features corresponding to each of the communication nodes. The initial graph attribute features corresponding to the virtual node are concatenated with the second convolution output results corresponding to the virtual node in each graph convolutional layer to obtain the graph-level embedding features corresponding to the virtual node.

[0157] Furthermore, when the node evaluation module 430 outputs an importance evaluation result for each communication node in the network topology based on the structure-aware embedding features using a pre-trained policy network model, the node evaluation module 430 is used to: The structure-aware embedding features are input into a pre-trained policy network model, and the policy network model performs an outer product operation on the node embedding features and graph-level embedding features in the structure-aware embedding features to obtain the dimensional interaction features between the node embedding features and the graph-level embedding features. The dimensional interaction features are vector-mapped using a preset first learnable parameter to obtain a dimensional interaction feature vector. The deep Q-network model set by the policy network model is activated and linearly transformed on the dimensional interaction feature vector based on a preset second learnable parameter, and outputs the Q value corresponding to each communication node; wherein, the Q value represents the expected reward value brought about by selecting the communication node as the action corresponding to the node embedding feature for removal in the state corresponding to the graph-level embedding feature. The Q value corresponding to each of the communication nodes is determined as the importance assessment result of each of the communication nodes in the network topology.

[0158] Furthermore, when the damage detection module 440, in response to detecting at least one damaged communication node that has been attacked, selects a critical communication node from the damaged communication nodes and a redundant communication node from the normal communication nodes that have not been attacked, based on the importance assessment result, the damage detection module 440 is configured to: In response to the detection that at least one compromised communication node among the communication nodes has been attacked, the importance ranking of the communication nodes is determined based on the descending order corresponding to the importance assessment results. Among the damaged communication nodes, the damaged communication node that ranks first in importance is selected to identify the damaged communication node as a critical communication node. Among the normal communication nodes that have not been attacked, select the normal communication node whose importance assessment result is less than that of the critical communication node, and determine the normal communication node as a candidate communication node. The candidate communication node whose importance ranking result is the last among the candidate communication nodes is selected to determine the candidate communication node as a redundant communication node.

[0159] Furthermore, when the topology module 450 moves and replaces the redundant communication nodes with the key communication nodes at corresponding positions when the connectivity of the network topology meets preset conditions, the topology module 450 is used to: The redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions, and it is determined whether the connectivity result of the network topology under the operation of the movement and replacement is greater than or equal to a preset connectivity threshold. If the connectivity result is greater than or equal to a preset connectivity threshold, then the connectivity of the network topology is determined to meet the preset conditions under the move-replace operation.

[0160] Furthermore, when the topology module 450 moves and replaces the redundant communication nodes with the key communication nodes at corresponding positions when the connectivity of the network topology meets preset conditions, the topology module 450 is also used to: If the connectivity result is less than the preset connectivity threshold, then the move and replacement operation is restored, the redundant communication node is removed from the candidate communication nodes, and the candidate communication node with the last position in the importance ranking result is selected from the updated candidate communication nodes to determine the candidate communication node as the updated redundant communication node. The updated redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions until the connectivity result is greater than or equal to a preset connectivity threshold.

[0161] The topology reconstruction device for UAV swarms provided in this application constructs a network topology structure by combining virtual nodes with the communication network corresponding to the UAV swarm. It uses a graph attention network model to perform feature transformation on the network topology structure to obtain structure-aware embedded features, and uses a policy network model to evaluate the importance of the structure-aware embedded features. It determines the importance evaluation result of the communication node corresponding to each UAV in the network topology structure. When a damaged communication node appears in the UAV swarm, based on the importance evaluation result, it prioritizes the recovery of key nodes and dynamically adjusts the structural connections. By introducing a structure-aware node importance ranking mechanism and an alternative topology reconstruction strategy, it achieves the identification and adaptive recovery of damaged communication nodes. In complex, highly dynamic, and strongly interfered network environments, it improves the stability and accuracy of topology reconstruction for UAV swarms, and effectively enhances the network connectivity and system robustness of UAV swarms.

[0162] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 500 includes a processor 510, a memory 520, and a bus 530.

[0163] The memory 520 stores machine-readable instructions executable by the processor 510. When the electronic device 500 is running, the processor 510 and the memory 520 communicate via the bus 530. When the machine-readable instructions are executed by the processor 510, they can perform the operations described above. Figure 1 The steps of the topology reconstruction method for the UAV cluster in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0164] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of the topology reconstruction method for the UAV cluster in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0165] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0166] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0167] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0168] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0169] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0170] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for topology reconstruction of a drone swarm, characterized in that, The topology reconstruction method includes: Based on the real-time communication nodes corresponding to each drone in the target drone cluster and the preset virtual nodes, the network topology corresponding to the target drone cluster is constructed in real time. The network topology is transformed using a pre-defined graph attention network model to obtain the structure-aware embedding features corresponding to the network topology. Using a pre-trained policy network model based on the structure-aware embedding features, the importance assessment result of each communication node in the network topology is output. In response to the detection that at least one compromised communication node has been attacked, based on the importance assessment result, a critical communication node is selected from the compromised communication nodes, and a redundant communication node is selected from the normal communication nodes that have not been attacked. When the connectivity of the network topology meets the preset conditions, the redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions, and the next redundant communication node is iteratively selected to move and replace the next key communication node until all the damaged communication nodes are moved, replaced or restored, so as to obtain the reconstructed network topology corresponding to the target UAV cluster.

2. The method according to claim 1, characterized in that, The method of constructing the network topology of the target drone cluster in real time based on the real-time communication nodes corresponding to each drone in the target drone cluster and preset virtual nodes includes: The system acquires real-time location data and transmission signal power data for each drone in the target drone cluster in real time, and determines the received signal power data for each drone based on the real-time location data and the transmission signal power data, corresponding to the received signal power data of the other drones in the target drone cluster that each drone receives. Determine whether the received signal power data between each pair of drones in the target drone cluster is greater than a preset power threshold to obtain a power determination result; A communication link is established between two drones in the target drone cluster whose power judgment result is greater than a certain value. The real-time location data corresponding to each drone is used as a communication node, and the communication link is used as a communication path to construct the adjacency network structure corresponding to the target drone cluster in real time. A one-way edge connection is established between the preset virtual node and each of the communication nodes to obtain the virtual communication path between the virtual node and each of the communication nodes. Based on the adjacency network structure, the virtual node and the virtual communication path, the network topology structure corresponding to the target UAV cluster is constructed in real time.

3. The method according to claim 1, characterized in that, The structure-aware embedding features include node embedding features corresponding to each communication node and graph-level embedding features corresponding to the virtual nodes; The step of using a preset graph attention network model to perform feature transformation on the network topology to obtain the structure-aware embedding features corresponding to the network topology includes: Based on the initial node features corresponding to each communication node and the initial graph features corresponding to each virtual node, the initial node attribute features corresponding to each communication node and the initial graph attribute features corresponding to each virtual node are obtained using a preset normalized feature encoder. Each of the multiple graph convolutional layers in the preset graph attention network model performs convolution operations on the initial node attribute features and the initial graph attribute features, respectively, to obtain the first convolution output result of each communication node in each graph convolutional layer and the second convolution output result of each virtual node in each graph convolutional layer. For each of the communication nodes, the initial node attribute features corresponding to the communication node are concatenated with the first convolution output results corresponding to the communication node in each graph convolutional layer to obtain the node embedding features corresponding to each of the communication nodes. The initial graph attribute features corresponding to the virtual node are concatenated with the second convolution output results corresponding to the virtual node in each graph convolutional layer to obtain the graph-level embedding features corresponding to the virtual node.

4. The method according to claim 1, characterized in that, The pre-trained policy network model, based on the structure-aware embedding features, outputs an importance assessment result for each communication node in the network topology, including: The structure-aware embedding features are input into a pre-trained policy network model, and the policy network model performs an outer product operation on the node embedding features and graph-level embedding features in the structure-aware embedding features to obtain the dimensional interaction features between the node embedding features and the graph-level embedding features. The dimensional interaction features are vector-mapped using a preset first learnable parameter to obtain a dimensional interaction feature vector. The deep Q-network model set by the policy network model is activated and linearly transformed on the dimensional interaction feature vector based on a preset second learnable parameter, and outputs the Q value corresponding to each communication node; wherein, the Q value represents the expected reward value brought about by selecting the communication node as the action corresponding to the node embedding feature for removal in the state corresponding to the graph-level embedding feature. The Q value corresponding to each of the communication nodes is determined as the importance assessment result of each of the communication nodes in the network topology.

5. The method according to claim 1, characterized in that, In response to detecting at least one compromised communication node that has been attacked, the method involves selecting a critical communication node from the compromised communication nodes and selecting redundant communication nodes from the unattacked normal communication nodes, based on the importance assessment result, including: In response to the detection that at least one compromised communication node among the communication nodes has been attacked, the importance ranking of the communication nodes is determined based on the descending order corresponding to the importance assessment results. Among the damaged communication nodes, the damaged communication node that ranks first in importance is selected to identify the damaged communication node as a critical communication node. Among the normal communication nodes that have not been attacked, select the normal communication node whose importance assessment result is less than that of the critical communication node, and determine the normal communication node as a candidate communication node. The candidate communication node whose importance ranking result is the last among the candidate communication nodes is selected to determine the candidate communication node as a redundant communication node.

6. The method according to claim 5, characterized in that, When the connectivity of the network topology meets preset conditions, the redundant communication nodes and the key communication nodes are moved and replaced at corresponding positions, including: The redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions, and it is determined whether the connectivity result of the network topology under the operation of the movement and replacement is greater than or equal to a preset connectivity threshold. If the connectivity result is greater than or equal to a preset connectivity threshold, then the connectivity of the network topology is determined to meet the preset conditions under the move-replace operation.

7. The method according to claim 6, characterized in that, The step of moving and replacing the redundant communication nodes and the key communication nodes at corresponding positions when the connectivity of the network topology meets preset conditions further includes: If the connectivity result is less than the preset connectivity threshold, then the move and replacement operation is restored, the redundant communication node is removed from the candidate communication nodes, and the candidate communication node with the last position in the importance ranking result is selected from the updated candidate communication nodes to determine the candidate communication node as the updated redundant communication node. The updated redundant communication nodes and the key communication nodes are moved and replaced at their corresponding positions until the connectivity result is greater than or equal to a preset connectivity threshold.

8. A topology reconfiguration device for a drone swarm, characterized in that, The topology reconfiguration device includes: The network topology module is used to construct the network topology of the target drone cluster in real time based on the real-time communication nodes corresponding to each drone in the target drone cluster and the preset virtual nodes. The feature transformation module is used to perform feature transformation on the network topology using a preset graph attention network model to obtain the structure-aware embedding features corresponding to the network topology. The node evaluation module is used to output the importance evaluation result of each communication node in the network topology based on the structure-aware embedding features using a pre-trained policy network model. The damage detection module is used to, in response to detecting that at least one damaged communication node has been attacked, select a critical communication node from the damaged communication nodes based on the importance assessment result, and select a redundant communication node from the normal communication nodes that have not been attacked. The topology module is used to move and replace the redundant communication nodes and the key communication nodes at corresponding positions when the connectivity of the network topology meets the preset conditions, and iteratively select the next redundant communication node to move and replace the next key communication node until all the damaged communication nodes are moved, replaced or restored, so as to obtain the reconstructed network topology corresponding to the target UAV cluster.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. The machine-readable instructions are executed by the processor to perform the steps of the topology reconfiguration method for a drone swarm as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the topology reconfiguration method for a drone swarm as described in any one of claims 1 to 7.

Images