Unmanned aerial vehicle group network self-healing method and device based on graph learning

Abstract

The embodiment of the invention provides an unmanned aerial vehicle group network self-healing method and device based on graph learning, and relates to the field of unmanned system communication and control. The method is used for solving the problems that an existing unmanned aerial vehicle group self-healing method is large in power adjustment limitation, a moving strategy has defects, and a graph learning method faces topological imbalance and is insufficient in response and generalization ability. According to the method, through a hierarchical residual graph topology reconstruction strategy, the problems of message passing unevaluation and excessive node aggregation occurring in the unmanned aerial vehicle relocation process are effectively relieved, it is ensured that the recovered network topology structure is more balanced and stable, and the overall elasticity and survivability of the network are improved; by introducing task semantic embedding, dynamic diffusion intensity and implicit multi-branch adaptive aggregation, the model can dynamically perceive and adaptively respond to damage scenes distributed by different nodes, and the response efficiency and cross-scene applicability of a self-healing algorithm are improved. By designing a differentiable joint loss function, reasonable constraint and stable gradient guidance are provided for optimization of a graph learning algorithm.

Description

Technical Field

[0001] This invention relates to the field of communication and control of unmanned systems, and more specifically to a method and apparatus for self-healing unmanned aerial vehicle (UAV) swarm networks based on graph learning. Background Technology

[0002] In recent years, unmanned aerial vehicle (UAV) swarm networks have shown broad application prospects in fields such as data acquisition, disaster relief, and environmental monitoring. However, in complex or special environments, UAVs face threats from their own malfunctions or external interference. Large-scale node failures can severely damage network communication links, leading to a collapse in global connectivity. To address this challenge, existing research has explored various self-healing methods for UAV swarms, aiming to restore network connectivity and improve system robustness.

[0003] Current self-healing methods for UAV swarms can be mainly divided into power adjustment-based methods and node movement-based methods. Power adjustment-based methods dynamically adjust the transmission power of UAVs to change the communication range, re-establishing or enhancing communication links after connectivity breaks. Their advantages include no physical movement required and fast response, but disadvantages include limitations imposed by the power of the UAVs themselves and battery life, and the difficulty in effectively restoring global connectivity through power adjustment alone when there is a large-scale node failure. Node movement-based methods plan the movement paths of remaining UAVs to reposition them and compensate for communication blind spots or breaks in the network. Specifically, these methods can be further subdivided into three categories: critical node replacement: identifying nodes in the network that are crucial to connectivity, and assigning nearby or planned UAVs to replace them after their failure. This method is simple in strategy but often faces cascading movement problems, resulting in low recovery efficiency and difficulty in dealing with large-scale node failures; and central aggregation: guiding remaining UAVs to gather towards the network center or the center of the damaged area, forming a denser connectivity cluster to enhance connectivity. This method is effective for localized damage, but it is prone to regional over-aggregation, affecting the robustness and balance of the network topology after recovery. Optimal planning algorithms use complex algorithms to move the remaining drones in the optimal way, achieving goals such as minimizing recovery time and maximizing connectivity. Among these, graph learning-based methods (such as Graph Neural Network, GNN) excel in connectivity recovery due to their consideration of complex node interactions and powerful information aggregation capabilities.

[0004] In recent years, network recovery methods based on graph neural networks have attracted attention. Some studies have optimized the recovery effect by improving graph convolution operators, enhancing input features and topology, and designing bipartite graph convolutional networks, but limitations still exist: excessive node aggregation and network topology imbalance: when guiding UAVs to relocalize, nodes are prone to clustering in the communication interruption center area, resulting in excessively high local node density, disrupting topological balance, wasting resources, and potentially forming new bottlenecks or isolating remote UAVs, reducing the long-term robustness of the network; lack of response efficiency and cross-scenario generalization ability: when facing different scales of damage scenarios, the model needs to store a large number of parameters or spend a long time optimizing, resulting in low response efficiency; and most of them adopt fixed information aggregation mechanisms, which are difficult to adaptively perceive scene changes and cannot meet the needs of rapid self-healing.

[0005] In summary, existing self-healing methods for UAV swarms suffer from limitations in power adjustment and flawed movement strategies, while graph learning methods also face problems such as topological imbalance and insufficient response and generalization capabilities. Summary of the Invention

[0006] This invention provides a graph learning-based self-healing method and apparatus for unmanned aerial vehicle (UAV) swarm networks, which addresses the limitations of existing UAV swarm self-healing methods, such as large limitations in power adjustment and defective mobility strategies, as well as the problems faced by graph learning methods, such as topological imbalance and insufficient response and generalization capabilities.

[0007] This invention provides a graph learning-based self-healing method for unmanned aerial vehicle (UAV) swarm networks, comprising:

[0008] Based on each independent cluster in the residual graph at the moment of communication breakdown of the UAV swarm, construct an intra-cluster relative neighborhood graph, and construct a cluster-level relative neighborhood graph based on the distance between the nearest surviving node pairs of any independent cluster pair; obtain a global reconstructed adjacency matrix based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph.

[0009] Using a parametric encoder, a global description vector and task descriptions at each node level are concatenated to generate node task embedding vectors and task embedding matrices. A normalized node position feature matrix and a task embedding matrix are then concatenated to obtain the input feature matrix for initial graph convolutional encoding and the initial graph convolutional output. The initial graph convolutional output is used to obtain graph convolutional encoded node representation vectors and layer output sequences based on multi-layer core graph convolution. The graph convolutional encoded node representation vectors included in the layer output sequences are weighted and fused based on attention weights to obtain a multi-branch aggregated representation matrix including all surviving nodes. The multi-branch aggregated representation matrix is ​​then fused with the normalized node position feature matrix using gate control vectors to obtain the optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes.

[0010] The optimal target location matrix is ​​used to determine the number of clusters under the target location distribution and whether the UAV swarm has restored connectivity. The connectivity loss is determined based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold. The number of clusters under the target location distribution is traversed, and the inter-cluster distance loss is determined based on the communication distance threshold and the maximum inter-cluster distance. A joint loss function is obtained based on the travel distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss. The joint loss function is iteratively optimized to obtain the final target location matrix that satisfies full connectivity and minimizes loss.

[0011] This invention provides a graph learning-based self-healing device for unmanned aerial vehicle (UAV) swarm networks, comprising:

[0012] The construction unit is used to construct an intra-cluster relative neighborhood graph for each independent cluster in the residual graph at the moment of communication failure of the UAV swarm, construct a cluster-level relative neighborhood graph based on the distance between the nearest surviving nodes of any independent cluster pair, and obtain a global reconstructed adjacency matrix based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph.

[0013] The unit is used to generate node task embedding vectors and task embedding matrices by concatenating the global description vector and the task descriptions at each node level using a parametric encoder; it also concatenates the normalized node position feature matrix and the task embedding matrix to obtain the input feature matrix and the initial graph convolutional output of the graph convolutional encoding; the initial graph convolutional output is used to obtain the node representation vector and the layer output sequence of the graph convolutional encoding based on multi-layer core graph convolution; the node representation vectors of the graph convolutional encoding included in the layer output sequence are weighted and fused based on attention weights to obtain a multi-branch aggregated representation matrix including all surviving nodes; the multi-branch aggregated representation matrix and the normalized node position feature matrix are fused through a gate control vector to obtain the optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes;

[0014] An optimization unit is used to determine the number of clusters under the target location distribution and whether the UAV swarm has restored connectivity based on the optimal target location matrix; determine the connectivity loss based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold; traverse the number of clusters under the target location distribution and determine the inter-cluster distance loss based on the communication distance threshold and the maximum inter-cluster distance; obtain a joint loss function based on the movement distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss; iteratively optimize the joint loss function to obtain the final target location matrix that satisfies full connectivity and minimizes the loss.

[0015] This invention provides a computer device including a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs any of the above-described graph learning-based self-healing methods for unmanned aerial vehicle (UAV) swarm networks.

[0016] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform any of the above-described graph learning-based self-healing methods for unmanned aerial vehicle (UAV) swarm networks.

[0017] This invention provides a graph learning-based self-healing method and apparatus for unmanned aerial vehicle (UAV) swarm networks. The method includes: constructing an intra-cluster relative neighborhood graph for each independent cluster in the surviving graph at the moment of communication interruption in the UAV swarm; constructing a cluster-level relative neighborhood graph based on the distance between the nearest surviving node pairs of any independent cluster pair; obtaining a globally reconstructed adjacency matrix based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph; using a parametric encoder φ(·), concatenating a global description vector with each node-level task description to generate a node task embedding vector and a task embedding matrix; concatenating a normalized node position feature matrix with the task embedding matrix to obtain an initial graph convolutional encoding input feature matrix and an initial graph convolutional output; the initial graph convolutional output, based on multi-layer core graph convolution, obtains a graph convolutional encoded node representation vector and a layer output sequence; and weighting the graph convolutional encoded node representation vectors included in the layer output sequence based on attention weights. The following steps are performed: First, a multi-branch aggregated representation matrix including all surviving nodes is obtained. Then, the multi-branch aggregated representation matrix and the normalized node position feature matrix are fused using a gate control vector to obtain an optimal node fusion matrix including all surviving nodes and an optimal target position matrix including all surviving nodes. Based on the optimal target position matrix, the number of clusters under the target position distribution and whether the UAV swarm has restored connectivity are determined. Based on the optimal target position matrix, the number of clusters under the target position distribution, and the communication distance threshold, connectivity loss is determined. The number of clusters under the target position distribution is traversed, and the inter-cluster distance loss is determined based on the communication distance threshold and the maximum inter-cluster distance. Based on the movement distance loss, connectivity loss, inter-cluster distance loss, repulsion loss, and uniformity loss, a joint loss function is obtained. The joint loss function is iteratively optimized to obtain the final target position matrix that satisfies full connectivity and minimizes loss. This method effectively alleviates the message passing imbalance and excessive node aggregation problems that occur during UAV relocation by employing a hierarchical residual graph topology reconstruction strategy. This ensures a more balanced and stable restored network topology, enhancing the overall resilience and survivability of the network. By introducing task semantic embedding, dynamic diffusion intensity, and implicit multi-branch adaptive aggregation, the model can dynamically perceive and adaptively respond to disruptive scenarios with different node distributions, improving the response efficiency and cross-scenario applicability of the self-healing algorithm. Furthermore, by designing a differentiable joint loss function, reasonable constraints and stable gradient guidance are provided for the optimization of the graph learning algorithm. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A schematic diagram of the graph learning-based self-healing method for unmanned aerial vehicle (UAV) swarm networks provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of a graph learning-based self-healing device for unmanned aerial vehicle (UAV) swarm networks provided in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Specifically, Figure 1 This is a schematic diagram of the graph learning-based self-healing method for unmanned aerial vehicle (UAV) swarm networks provided in an embodiment of the present invention; the following uses... Figure 1 Taking an example, the method provided by the embodiments of the present invention will be described in detail, such as Figure 1 As shown, the method includes the following steps:

[0023] Step 101: Based on each independent cluster in the residual graph at the moment of communication breakdown in the UAV swarm, construct a relative neighborhood graph within each cluster. Construct a cluster-level relative neighborhood graph based on the distance between the nearest surviving node pairs of any independent cluster pairs. Based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph Obtain the globally reconstructed adjacency matrix

[0024] Step 102: Using the parametric encoder φ(·), concatenate the global description vector τ. gb With each node-level task description To generate node task embedding vectors and task embedding matrix X emb ; Concatenate the normalized node position feature matrix X norm With task embedding matrix X emb The input feature matrix H of the initial graph convolutional encoding is obtained. (0) and the initial graph convolution output H (1) The initial graph convolution output is based on multi-layer core graph convolution to obtain graph convolutionally encoded node representation vectors and layer output sequences. The graph convolutionally encoded node representation vectors included in the layer output sequences are weighted and fused based on attention weights to obtain a multi-branch aggregated representation matrix J including all remaining nodes. The multi-branch aggregated representation matrix is ​​then fused with the normalized node position feature matrix X using gate control vectors. normThe optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes are obtained.

[0025] Step 103: Determine the number of clusters under the target location distribution and whether the UAV swarm has restored connectivity based on the optimal target location matrix; determine the connectivity loss based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold; traverse the number of clusters under the target location distribution and determine the inter-cluster distance loss based on the communication distance threshold and the maximum inter-cluster distance; obtain the joint loss based on the movement distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss, and iteratively optimize the joint loss to obtain the node target location that satisfies full connectivity and minimizes the loss.

[0026] Before step 101, it is necessary to determine the basic parameters required when using this method. The following uses the UAV swarm network connectivity recovery algorithm framework as an example to introduce in detail the graph learning-based UAV swarm network self-healing method provided by the embodiments of the present invention.

[0027] Specifically, first, collect the 3D coordinates of all UAVs included in the finite 3D space, and organize them into an original position matrix with each UAV corresponding to a row. Determine the total number of UAVs in the original UAV swarm as N, and determine the index set of the UAVs in the original UAV swarm as I, I = {1,2,…,N}. Then, the original position matrix of all UAVs in the original UAV swarm can be determined by the following formula:

[0028]

[0029] Among them, P ori p represents the original position matrix, that is, the original position matrix of all drones included in the drone swarm. i Indicates drone v i The three-dimensional coordinates, where N represents the total number of drones in the drone swarm, and R... N×3 Let p represent an N x 3 matrix space of real numbers, corresponding to the three-dimensional coordinates of N UAVs. i (t0)∈R 3 Indicates drone v i The three-dimensional coordinates at time t0.

[0030] Furthermore, based on the original position matrix of the UAV, the distance between any two nodes in the original position matrix is ​​calculated using the Euclidean distance formula, as follows:

[0031] d ij =‖p i -p j ‖twenty two)

[0032] Where, d ijIndicates drone v i and drones j The Euclidean distance between them, p i Indicates drone v i The three-dimensional coordinates, p j Indicates drone v j The three-dimensional coordinates of the vector are given by ||·|2, which represents the 2-norm of the vector.

[0033] In practical applications, if and only if the drone v i and drones j When the Euclidean distance between them is not greater than a predefined distance threshold, i.e., d ij ≤d tr At that time, the drone v i and drones j Communication connections are established between them. Therefore, the original UAV swarm at time t0 can be represented as an undirected graph, as shown below:

[0034] G={V,E} (3)

[0035] Where G represents the undirected graph of the original UAV swarm at time t0, V represents the node set of the original UAV swarm, E represents the communication edge set of the original UAV swarm, and V = {v i |i∈I},E={(v i ,v j )|v i ∈V,v j ∈V,d ij ≤d tr}

[0036] Furthermore, when some drones in the original drone swarm fail due to signal interference or equipment malfunction in a certain environment, the communication of the original connected network is disrupted. Let there be N drones in this swarm. D If ∈[1,N-2] nodes are destroyed, then the number of surviving nodes is N. R N R =NN D The set of indices of the destroyed nodes is The set of indices for the surviving nodes is then:

[0037]

[0038] Where I represents the set of indices of all drones in the original drone swarm before the destruction occurred, and d tr d represents a predefined communication distance threshold. ij ≤d tr Time node v i and node v j Establish a communication connection between them, I DI represents the set of indices of the destroyed nodes. R r represents the set of indices of the surviving nodes. i The index r represents the surviving node. i ∈I R .

[0039] Furthermore, the residual network at time t0 is modeled as a residual graph using the following formula:

[0040] G R ={V R E R} (5)

[0041]

[0042] Among them, G R V represents the residual graph model of the residual network at time t0. R Represents the set of remaining nodes. E R This represents the remaining set of communication edges. Indicates index r i The remaining nodes, Indicates index r j The remaining nodes.

[0043] Similarly, based on the original position matrix of all drones included in the original drone swarm, the position matrix of the surviving drones at time t0 (the surviving node position matrix) can be determined:

[0044]

[0045] Where P represents the position matrix of the remaining UAV at time t0. Represents the remaining nodes The three-dimensional coordinates at time t0, N R Indicates the number of remaining drones. N represents R A space of real numbers with 3 rows and 3 columns.

[0046] Furthermore, construct the residual graph G. R adjacency matrix A represents the surviving graph G. R The adjacency matrix; if and only if At that time, (A) ij =1, and the rest of the elements are 0.

[0047] In this embodiment of the invention, after the disruption occurs, the original UAV swarm is divided into several independent connected components. For ease of description, these connected components are referred to as a cluster. Accordingly, the cluster set at time t can be represented as:

[0048]

[0049] Among them, S all (t) represents the set of independent connected components of the surviving network at time t, or the set of independent clusters of the surviving network at time t; S j (t) represents the j-th cluster at time t. Indicates cluster S j (t) contains the set of nodes, V R This represents the set of remaining nodes.

[0050] Furthermore, after the disruption, the number of clusters in the surviving network can be determined using the following formula:

[0051] N S (t)=|S all (t)|∈N + And N S (t)≥1 (9)

[0052] Where, N S (t) represents the number of clusters at time t, N + Represents the set of positive integers.

[0053] In the method provided by this invention, the connectivity restoration problem aims to restore full graph connectivity in the shortest time by moving nodes, which needs to be achieved through iterative optimization.

[0054] Specifically, if the set of remaining nodes includes the remaining nodes Its movement time during the recovery process is The total recovery time is the longest travel time among all drones, i.e. And when the recovery is complete, it satisfies N. S (t0+T rc ) = 1.

[0055] To simplify modeling, it is assumed that each UAV moves at a constant linear speed of maximum velocity V0 during the recovery process, and the recovery time is proportional to the maximum travel distance. Therefore, in this embodiment of the invention, the optimization objective of the connectivity restoration problem can be equivalently transformed into finding the optimal target position to minimize the maximum travel distance of the remaining nodes, that is, using the initial position in the remaining node position matrix P as a reference, and through the total recovery time T... rc By relating the movement distance to time, the core objective of "minimizing the maximum movement distance to quickly restore connectivity" is clarified, guiding subsequent target location calculations. The formulas involved in this process include:

[0056]

[0057] in, This represents the optimal target position matrix for the UAVs corresponding to the remaining nodes. Represents the remaining nodes The optimal target position, This represents the actual location of the drones corresponding to each surviving node when network connectivity is restored, and 'st' represents the constraint subject to. Represents the remaining nodes The three-dimensional coordinates at time t0, I R r represents the set of indices of the surviving nodes. i The index r represents the surviving node. i ∈I R N S (t0+T rc ) indicates the time t0+T when the recovery is complete. rc The number of clusters, T rc V0 represents the total recovery time (the maximum time of movement of all drones) and the maximum movement speed of the drone.

[0058] It should be noted that, in practical applications, the direction of movement of each surviving node can be determined by the optimal target position of the surviving node. It is certain, but not all remaining nodes need to be completely moved to that position, i.e., equation This is not always true because the network often restores connectivity as surviving nodes move, i.e.:

[0059]

[0060] However, due to the fact that the actual positions of the drones corresponding to each remaining node were unknown before the actual mission was executed. It cannot be obtained precisely. Assuming a surviving node can be found... optimal target position (Indicates the drone corresponding to the surviving node when connectivity is restored) The actual location of the drones corresponding to each surviving node when network connectivity is restored (the actual location of the drones). Existence and Between. Based on this assumption, the remaining nodes are used. optimal target position The actual location of the UAV corresponding to each surviving node during the replacement connection. Therefore, the optimization objective is transformed into:

[0061]

[0062] In step 101, for the residual graph G R For each independent cluster in the dataset, determine the set of node indices for the cluster as shown below:

[0063]

[0064] in, Indicates cluster S q The set of surviving node indices, c i Indicates cluster S q The index of the surviving nodes, Indicates cluster S q The number of surviving nodes, S q ∈S all (t0).

[0065] Furthermore, the intra-cluster subgraph and intra-cluster node set are represented by the following formulas:

[0066]

[0067] in, This represents the intra-cluster subgraph of the q-th cluster. Indicates cluster S q The set of surviving nodes within the cluster. Indicates cluster S q Intra-cluster communication edge set, Indicates cluster S q The internal index is c i The remaining nodes.

[0068] Furthermore, based on the cluster node set and the edge set of the cluster relative neighborhood graph, cluster S is determined. q The intra-cluster relative neighborhood graph is shown below:

[0069]

[0070] in, Indicates cluster S q The relative neighborhood graph within the cluster, This represents the edge set of the relative neighborhood graph within a cluster.

[0071] In this embodiment of the invention, when constructing a cluster-level relative neighborhood graph based on the nearest distance between any two independent clusters, the following formula must be satisfied:

[0072]

[0073] in, Indicates cluster S q Internal residual nodes and European distance, Indicates cluster S q Internal residual nodes and European distance, Indicates cluster S q Internal residual nodes and European distance, That is, for any pair of surviving nodes within the cluster and remaining nodes Intra-cluster relative neighborhood graph China retains There is no edge between them, and there is no other surviving node. Its relationship with the remaining nodes Distance between edges With remaining nodes Distance between the edges Simultaneously compared with the remaining nodes and remaining nodes If they are closer to each other, the remaining nodes will be preserved. and remaining nodes The edges between them.

[0074] Furthermore, the topology of each independent cluster is transformed into a reconstructed adjacency matrix. In this embodiment of the invention, cluster S q The reconstructed adjacency matrix of the intra-cluster relative neighborhood graph is obtained through the following formula.

[0075]

[0076] in, Indicates cluster S q Reconstruct the adjacency matrix, This represents the element in the i-th row and j-th column of the reconstructed adjacency matrix. Indicates cluster S q The remaining nodes, This indicates that it does not exist.

[0077] Furthermore, the reconstructed intra-cluster adjacency matrix sequence is obtained as shown below:

[0078]

[0079] After determining the relative neighborhood graph within a cluster and reconstructing the adjacency matrix, for any cluster pair S i and S j The Euclidean distance between the nearest surviving node pairs in a cluster can be determined using the following formula:

[0080]

[0081] Where, d S (S i ,S j ) represents cluster S i and cluster S j The Euclidean distance between the nearest surviving node pairs, where u represents the cluster S. i Any surviving node within S, where v represents cluster Sj Any surviving node within, d uv This represents the Euclidean distance between the surviving node u and the surviving node v.

[0082] Furthermore, each independent cluster can be considered as a supernode. By traversing all surviving node pairs included in any two independent clusters, the Euclidean distance of the surviving node pairs with the smallest distance is determined, and a cluster-level relative neighborhood graph is constructed based on the Euclidean distance of the surviving node pairs with the smallest distance. That is, for any cluster pair S i ,S j If there is no third cluster S k If the following formula is satisfied, a cluster-level relative neighborhood graph can be constructed.

[0083] max(d S (S i ,S k ),d S (S j ,S k )) <d S (S i ,S j (17)

[0084] Where, d S (S i ,S j ) represents cluster S i and cluster S j The Euclidean distance between the nearest surviving node pairs, where u represents the cluster S. i Any surviving node within S, where v represents cluster S j Any surviving node within S k Indicates the absence of cluster S i ,S j The third cluster besides S i ,S j ∈S all (t0).

[0085] Then in the cluster-level relative neighborhood graph Establish cluster-level connections (S) i ,S j This yields the cross-cluster candidate connection set E. S E S ∈{(S i ,S j )|S i ∈S all (t0),S j ∈S all (t0),S i ≠S j}

[0086] Furthermore, for cluster-level relative neighborhood graphs Any cluster-level connection (S) on i ,S j The number of cross-cluster redundant edges is allocated according to their local relative size:

[0087]

[0088] Among them, E S Let represent the set of cross-cluster candidate connections in the cluster-level relative neighborhood graph, where α represents the redundancy ratio coefficient, α∈(0,1). This indicates rounding up.

[0089] Establish inter-cluster bridging matrix Take the m closest distance between the surviving nodes across clusters. ij Establish virtual connections for the remaining node pairs and connect the inter-cluster bridging matrix. Set the corresponding element to 1, and set the rest of the elements to 0.

[0090] In this embodiment of the invention, the intra-cluster reconstructed relative neighborhood graph is merged with the inter-cluster virtual bridge edges to obtain the final reconstructed residual topology. The corresponding global reconstructed adjacency matrix is:

[0091]

[0092] in, This represents the global reconstruction connection matrix corresponding to the remaining topology after final reconstruction, where ∨ represents the element-wise logical OR operation. The inter-cluster bridging matrix is ​​represented by diag(·), and the block diagonal stacking operator is represented by diag(·). This represents the intra-cluster reconstructed adjacency matrix of cluster S1. Represents the Nth digit in the residual network. S Clusters.

[0093] In step 102, scene-level statistical information is collected, such as the proportion of surviving nodes, edge density, average node degree, and cluster ratio. The global description vector is... N gb This represents the dimension of the global description vector. N represents gb A dimensional real vector space.

[0094] Collect statistical information for each node in the local topology, such as intra-cluster degree, inter-cluster degree, and percentage of the size of its own cluster, and summarize it into a residual node-level task description matrix:

[0095]

[0096] in, Represents the remaining nodes The corresponding node-level description vector, Nnd T represents the dimension of the remaining node-level task description. nd This represents the remaining node-level task description matrix.

[0097] Furthermore, task embeddings for each surviving node are extracted based on global and surviving node-level task descriptions to jointly encode the global and local roles of the surviving node in the current surviving topology. Any surviving node... The corresponding task embedding vector is:

[0098]

[0099] in, Represents the remaining nodes The task embedding vector, [·||·] represents feature concatenation, N τ φ(·) represents the dimension of the task embedding vector, φ(·) represents the parameterized encoder, which can be implemented using a small feedforward neural network, and τ gb Represents the global description vector. Represents a node Node-level description vector, N represents τ A dimensional real vector space.

[0100] Furthermore, the task embedding matrix is ​​obtained:

[0101]

[0102] Among them, X emb This represents the task embedding matrix.

[0103] Furthermore, after obtaining the task embedding matrix, the remaining node position matrix is ​​first normalized to obtain a normalized node position feature matrix. Then, the normalized node position feature matrix is ​​concatenated with the task embedding matrix to obtain the input feature matrix for the initial graph convolutional encoding. The normalized node position feature matrix and the input feature matrix for the initial graph convolutional encoding are shown below:

[0104]

[0105] Among them, X norm Let P represent the normalized node position feature matrix, and let P represent the damaged node position matrix. cen Let |t0| represent the location center of the surviving network at time t0, and ||·|2| represent the 2-norm of the vector. Represents the remaining nodes The three-dimensional coordinates at time t0, I R H represents the set of indices of the surviving nodes. (0) Let X represent the input feature matrix for the initial graph convolutional encoding. emb This represents the task embedding matrix.

[0106] In this embodiment of the invention, after obtaining the input feature matrix of the initial graph convolutional encoding, the input feature matrix of the initial graph convolutional encoding and the reconstructed adjacency matrix can be... and the task embedding matrix X emb The input is an Adaptive Graph Convolution Operator (AGCO), which is then passed through a non-linear activation function (such as the LeakyReLU function) to obtain the node representation of the initial graph convolutional encoding.

[0107] Before performing the above operations, a unique diffusion intensity needs to be assigned to each surviving node included in the surviving node location matrix.

[0108] Then, based on the task embedding matrix, the offset vector and scale vector corresponding to the surviving nodes are determined through a small feedforward network. The diffusion intensity, the offset vector corresponding to the surviving node, and the scale vector corresponding to the surviving node are shown below:

[0109]

[0110] in, Represents the remaining nodes The diffusion intensity, sigmoid(·) represents the sigmoid activation function, Represents the remaining nodes The scale factor, Represents the remaining nodes The offset factor, ∈0 represents a predefined baseline, ∈0∈(0,1), Indicates global reconstruction of the adjacency matrix The infinite norm of , where δ represents the offset vector of all surviving nodes. F represents the scale vector of all surviving nodes. eps (·) indicates a small feedforward network. Represents the scale vector The natural logarithm of X emb Represents the task embedding matrix, N R ε represents the total number of surviving nodes, and ε represents the diffusion factor matrix.

[0111] It should be noted that this structure can achieve node-level modulation of graph convolution diffusion intensity, thereby enabling adaptive feature propagation under different task semantics.

[0112] After determining the diffusion factor matrix, the input feature matrix of the initial graph convolutional encoding and the reconstructed adjacency matrix can be used. and the task embedding matrix X embThe input is an adaptive graph convolution operator, which is then passed through a nonlinear activation function to obtain the node representation of the initial graph convolutional encoding, as follows:

[0113]

[0114] Among them, H (1) AGCO represents the node representation of the initial graph convolutional encoding. (1) This represents the adaptive graph convolution operator in the first layer of graph convolution.

[0115] In this embodiment of the invention, the adaptive graph convolution operator, as the core operation of the graph convolution network, can be defined based on the graph convolution operation in the first-order Chebyshev approximate spectral domain. The adaptive graph convolution operator AGCO in the l-th layer of graph convolution... (l) The definition is as follows:

[0116]

[0117] Among them, AGCO (l) This represents the adaptive graph convolution operator in the l-th layer graph convolution. Let ε represent the global reconstructed adjacency matrix, and let ε represent the diffusion factor matrix. Represents the reconstruction of the adjacency matrix The corresponding Laplace matrix,

[0118] W (l) X represents the learnable linear transformation parameters of the graph convolution at layer l. in This represents the input feature matrix of the adaptive graph convolution operator. For a size of N R ×N R The identity matrix.

[0119] Furthermore, in determining the node representation H of the initial graph convolutional encoding... (1) Based on this, the core graph convolutional encoding is further performed through Q-1 layers to obtain the node representation vector and layer output sequence of the graph convolutional encoding. The processing of the graph convolutional unit in the l-th layer [2, Q] can be represented as follows:

[0120]

[0121] Among them, H (l) Let H represent the node representation in the graph convolution of the l-th layer, σ(·) represent the element-wise nonlinear activation function, and H (l -1) This represents the node representation in the (l-1)th layer graph convolution. AGCO represents the intermediate result in the graph convolution of the l-th layer. (l) This represents the adaptive graph convolution operator in the l-th layer graph convolution. Let ε represent the global reconstructed adjacency matrix, and let ε represent the diffusion factor matrix. Represents the reconstruction of the adjacency matrix The corresponding Laplace matrix, W ( l ) Let represent the learnable linear transformation parameters of the l-th layer graph convolution, and Q represent the total number of layers in the core graph convolution. The output sequence of the representation layer.

[0122] In the above formula, by outputting H from the upper layer (l-1) Convolution result with the current graph The addition achieves residual connections at each level. Since message passing in each layer of graph convolution occurs within the first-order neighborhood of a node, the processing at different levels can be viewed as multi-scale message passing with different propagation radii: shallow representations maintain local geometric consistency, while deep representations capture global connectivity across clusters. Thus, the layer output sequence constitutes an implicit multi-branch representation set, representing node representations with different receptive fields.

[0123] Furthermore, a dynamic aggregation mechanism based on multi-head attention is introduced on the multi-branch representation set. First, for each surviving node... Its multi-layer characterization This reflects the information hierarchy from local to global. It is based on node-level task description. Through a shared feedforward network F query The · operator can generate query vectors, as shown below:

[0124]

[0125] in, Represents the remaining nodes The query vector, F query (·) indicates a shared feedforward network. Represents the remaining nodes Node-level description vectors.

[0126] Then, based on the query vector and the node representation vector of the graph convolutional encoding, the attention scores of all surviving nodes for different branches and the attention weights of all surviving nodes for different branches are determined, as shown below:

[0127]

[0128] in, Represents the remaining nodes For the attention weights of the l-th layer branch, exp(·) denotes the exponential function. Represents the remaining nodes The attention score for the l-th layer, where Q represents the total number of convolutional layers in the core graph. Represents the query vector transpose, Represents the residual nodes in the l-th layer graph convolution. The representation vector, D hid Indicates the hidden layer dimension.

[0129] Finally, the node representation vectors of the graph convolutional encoding are weighted and fused according to the attention weights of all surviving nodes for different branches, resulting in a multi-branch aggregated representation including all surviving nodes and a multi-branch aggregated representation matrix including all surviving nodes, as shown below:

[0130]

[0131] in, Represents the remaining nodes Multi-branch aggregation representation. Represents the residual nodes in the l-th layer graph convolution. The representation vector, J represents the multi-branch aggregation representation matrix, N R This represents the total number of remaining nodes.

[0132] Furthermore, based on the multi-branch aggregation representation matrix, normalized node position features, and task embedding, a gating vector is dynamically generated. This gating vector is then used to fuse the multi-branch aggregation representation matrix and the normalized node position feature matrix X. norm The optimal node fusion matrix, which includes all surviving nodes, is obtained, where the gate control vector and the optimal node fusion matrix for all surviving nodes are shown below:

[0133] μ = sigmoid(W gate [J;X] norm W proj ;X emb (33)

[0134] Y=μ⊙J+(1-μ)⊙(X norm W proj (34)

[0135] Where μ represents the gate vector, W gate W represents the learnable linear transformation matrix generated by the gated vector. proj Let represent the learnable linear transformation matrix of the original position feature projection, ⊙ denotes element-wise multiplication, Y represents the node representation after gated fusion, and 1-μ denotes the element-wise complement of the gate vector μ.

[0136] In practical applications, W proj and W gate Both are learnable linear transformation matrices, used to transform the original normalized positional features X. norm and splicing features [J; X norm ;Xemb ] matrix reshaping.

[0137] Furthermore, the node representation Y, which is injected with the original features, is input into the final output decoder F. out And use the tanh(·) function to limit the range of values ​​to (0,2):

[0138] X out =tanh(F out (Y))+1 (35)

[0139] Among them, X out This represents the decoder output, tanh(·) represents the hyperbolic tangent activation function, and F... out (·) indicates the final output decoder.

[0140] Furthermore, regarding X out Inverse normalization yields the optimal target positions and the optimal target position matrix, which includes all surviving nodes, as shown below:

[0141]

[0142] in, Represents the remaining nodes The optimal target position, Represents the remaining nodes At the original coordinates of time p cen Indicates the location center of the remaining network. This represents the maximum original distance from the node to the center.

[0143] In this embodiment of the invention, after obtaining the optimal target location matrix including all surviving nodes according to the above steps, it is necessary to identify the independent clusters under the optimal target location distribution based on the optimal target location matrix, and then determine whether the drone swarm has restored connectivity.

[0144] First, identify and record the independent clusters under the optimal target location distribution, as follows:

[0145]

[0146] in, This indicates the number of clusters distributed at the target location. Represents independent clusters under the target location distribution. This represents the i-th cluster in the target location distribution.

[0147] Then, for any two remaining nodes in the optimal target position matrix Based on the coordinates in the target position matrix Calculate Euclidean distance

[0148] If the Euclidean distance is less than the predefined communication distance threshold d tr Then, edges are established in the target undirected graph. Breadth-first search (BFS) or depth-first search (DFS) is used to traverse all connected components of the target undirected graph. Each connected component is a cluster. The number of clusters is counted. If the number of clusters is... This indicates that the UAV swarm can restore connectivity under the optimal target location distribution; if the number of clusters... This indicates that the optimal target location cannot restore connectivity for the drone swarm; when the judgment... This indicates that the drone swarm can restore connectivity under the optimal target location distribution.

[0149] In this embodiment of the invention, a cluster number of 1 means that all nodes are connected into a whole through communication edges, satisfying the global connectivity requirement; a cluster number greater than 1 indicates the existence of isolated clusters, which require further optimization.

[0150] In this embodiment of the invention, the movement distance loss can be defined using the Euclidean distance between the optimal target position and the original position, as shown below:

[0151]

[0152] Among them, L P The distance traveled is represented by ρ, which represents the density normalization factor used to eliminate spatial scale differences under different residual scales. ρ = (N R / N) 1 / 3 R(t0) represents the set of surviving nodes at time t0.

[0153] In this loss, the first term directly penalizes the maximum movement distance, corresponding to minimizing the network recovery time; the second term uses L1-norm to minimize the movement distance of each surviving node, which on the one hand provides a smoother gradient for optimizing node movement constraints, and on the other hand saves the additional movement cost of the entire drone swarm.

[0154] Furthermore, a soft adjacency matrix is ​​defined based on the relative distance between optimal target locations.

[0155] in, Let represent the soft adjacency matrix, β represent the sharpness parameter in the soft adjacency matrix, and ReLU(·) represent the activation function. Represents the remaining nodes The target location d represents the Euclidean distance between two nodes at the target location. tr This indicates the communication distance threshold.

[0156] In practical applications, β>0 is the sharpness parameter, used to control the sensitivity of adjacency weights to distance. Under this definition, the weight between two different surviving nodes within the communication range is always 1, while the weight of surviving nodes outside the range decays smoothly with distance, thus ensuring the continuous differentiability of adjacency relationships.

[0157] The normalized Laplace matrix is ​​constructed based on the soft adjacency matrix, as shown below:

[0158]

[0159] in, Represents a soft adjacency matrix The corresponding normalized Laplace matrix, Represents a soft adjacency matrix The corresponding weighted degree matrix, Represents the weighted degree matrix The inverse square root matrix, N represents R ×N R 3D identity matrix.

[0160] In practical applications, Perform eigenvalue decomposition to obtain an ascending sequence of eigenvalues. Among them, the second smallest eigenvalue Algebraic connectivity, in other words, characterizes the strength of global connectivity. Represents the Laplace matrix The k-th ascending characteristic value.

[0161] To further obtain a differentiable approximation of the number of connected components, the differentiable approximation of the number of connected components is defined as follows:

[0162]

[0163] in, The differentiable approximation of the number of connected components, |λ k | represents the eigenvalue λ k The absolute value of N R This represents the total number of remaining nodes. K and b are both hyperparameters that control the smoothness and eigenvalue threshold, respectively. This can be understood as a continuous estimate of the number of eigenvalues ​​close to zero, thereby achieving a smooth approximation of the number of discrete connected components.

[0164] Based on the above process of determining the soft adjacency matrix, the normalized Laplacian matrix, and the differentiable approximation of the number of connected components, the connectivity loss can be obtained, as shown below:

[0165]

[0166] Among them, LC C represents connectivity loss. true γ represents the actual number of connected components in the remaining network under the target location distribution, and γ>0 represents the attenuation coefficient. The former term guides the enhancement of the overall network connectivity by penalizing excessively small algebraic connectivity, while the latter term imposes additional penalties when multiple isolated clusters exist to encourage inter-cluster fusion.

[0167] Furthermore, if the number of clusters under the optimal target location distribution Based on exponential decay, the penalty term L for the shortest distance between clusters can be calculated. CD Explicitly modeling the attraction between isolated clusters:

[0168]

[0169] Among them, L CD Indicates the inter-cluster distance loss. This represents the number of clusters under the optimal target location distribution.

[0170] If the number of clusters under the optimal target location distribution Then L CD =0.

[0171] Furthermore, to avoid excessive node clustering and structural imbalance, and to maintain a safe distance, a local repulsion loss L based on an exponential distance penalty is calculated. R The details are as follows:

[0172]

[0173] Among them, L R It indicates rejection of loss.

[0174] Furthermore, to maintain the balance of roles of remaining nodes and the stability of the network topology after connectivity is restored, the uniformity loss L based on degree centrality and proximity centrality is calculated. U The uniformity of the distribution is measured by the sum of the coefficients of variation of the two centrality indices, and the uniformity loss is shown below:

[0175] L U =CV(C D )+CV(C C (45)

[0176] Among them, L U C represents the uniformity loss, CV(·) represents the coefficient of variation of the centrality index, and C D Let C represent the set of degree centralities of all surviving nodes. C This represents the set of proximity centrality of all surviving nodes.

[0177] Specifically:

[0178]

[0179] in, Indicates the remaining node v i Degree centrality, Indicates the remaining node v i Proximity centrality, C D (·) indicates degree centrality, C C (·) denotes the proximity centrality operator, as shown below:

[0180]

[0181] in, Represents the remaining nodes in the soft adjacency matrix and The weights.

[0182] In this embodiment of the invention, according to a predefined weighting coefficient η c ,η CD ,η R ,η U The weighted sum of each loss term >0 yields the joint loss function L. joint :

[0183] L joint =L P +η c L C +η CD L CD +η R L R +η U L U (48)

[0184] Among them, L joint Let η represent the joint loss function (total loss). c η CD η R η U L represents the weighting coefficient of each loss. P L represents the distance traveled. CD L represents the inter-cluster distance loss. R L represents the rejection of loss. U L represents the uniformity loss. C This indicates connectivity loss.

[0185] In practical applications, this joint loss function, within a differentiable framework, simultaneously considers recovery efficiency, connectivity guarantee, topological rationality, and structural balance, providing systematic guidance and constraints for model training and online optimization.

[0186] In this embodiment of the invention, iterative optimization continuously adjusts the model parameters through gradient descent, gradually reducing the joint loss function to ensure that the final target location is not only connected, but also optimal in terms of movement cost and topology.

[0187] The iterative optimization process provided in this embodiment of the invention includes:

[0188] First, the Adam optimizer is used to perform gradient optimization on all learnable parameters in the process of obtaining the joint loss function in steps 102 to 103, based on a predefined learning rate.

[0189] Then, repeat step 101 to generate a large number of random samples and perform data preprocessing; iterate the joint loss function in steps 102 to 103 several times on each sample, and finally obtain the pre-trained model weights M.

[0190] Finally, when a UAV swarm network failure occurs, node location and topology information are collected; step 101 is executed to preprocess the data; steps 102 and 103 are iteratively executed to obtain the joint loss function until the stopping condition is met: the number of iterations reaches the maximum limit, or the number of clusters under the target location distribution is reached. The cumulative number of times exceeds the set number and has been satisfied in at least one iteration

[0191] It should be noted that during the process of obtaining the joint loss function in steps 102 to 103 of the iterative execution, if the following condition is met in at least one iteration of the iterative process: From satisfaction The final target position is obtained by taking the output of the optimal target position corresponding to the iteration with the smallest joint loss value from the results of the iterations; otherwise, it is determined that the algorithm cannot converge and the self-healing task fails.

[0192] This invention provides a graph learning-based self-healing method and apparatus for unmanned aerial vehicle (UAV) swarm networks. The method reconstructs a global adjacency matrix based on intra-cluster and inter-cluster relative neighborhood graphs. It uses intra-cluster rules to remove redundant edges and inter-cluster rules to select high-priority cluster pairs and adaptively allocate redundant edges, avoiding excessive node aggregation and reducing local competition for network resources. The reconstructed topology combines sparsity and global reachability, reducing the risk of local bottlenecks, preventing the isolation of remote surviving nodes, ensuring a balanced and robust topology after recovery, and improving network resilience. By concatenating global description vectors and node-level task descriptions, and jointly encoding scene statistics and local node roles, the model intuitively perceives the damaged scene without requiring extensive pre-setting. Parameters are fine-tuned over a long period; a multi-branch aggregation representation matrix combined with attention weights dynamically filters different receptive field node vectors, solving the performance degradation problem of traditional fixed aggregation mechanisms across scenarios, and adapting to a small number of, regional, and large-scale random failure scenarios; the optimal target position matrix obtained by fusion of gate control vectors can be directly converted into 3D coordinate navigation commands without additional conversion, making it highly practical; a joint loss function constructed from multiple losses provides stable gradients using spectral metrics such as algebraic connectivity, avoiding gradient vanishing or oscillations and ensuring efficient convergence; the synergy of various losses achieves multi-objective balance by reducing energy consumption, ensuring strong global connectivity, and maintaining node spacing and topological equilibrium. Pre-training and online optimization are combined to quickly access preprocessing results, reducing online latency after network failures, meeting the response requirements of hazardous environments, and satisfying the need for self-healing response speed in hazardous environments.

[0193] Based on the same inventive concept, this invention provides a graph learning-based self-healing method device for unmanned aerial vehicle (UAV) swarm networks. Since the principle by which this device solves the technical problem is similar to that of the graph learning-based UAV swarm network self-healing method, the implementation of this device can refer to the implementation of the method, and the repeated parts will not be described again.

[0194] like Figure 2 As shown, the device includes a construction unit 201, a obtaining unit 201, and an optimization unit 203.

[0195] Construction unit 201 is used to construct a relative neighborhood graph within each independent cluster in the residual graph at the moment of communication breakdown of the UAV swarm. Construct a cluster-level relative neighborhood graph based on the distance between the nearest surviving node pairs of any independent cluster pairs. Based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph Obtain the globally reconstructed adjacency matrix

[0196] Unit 202 is obtained, which is used to concatenate the global description vector τ using the parametric encoder φ(·). gb With each node-level task description To generate node task embedding vectors and task embedding matrix X emb ; Concatenate the normalized node position feature matrix X norm With task embedding matrix X emb The input feature matrix H of the initial graph convolutional encoding is obtained. (0) and the initial graph convolution output H (1) The initial graph convolution output is based on multi-layer core graph convolution to obtain graph convolutionally encoded node representation vectors and layer output sequences. The graph convolutionally encoded node representation vectors included in the layer output sequences are weighted and fused based on attention weights to obtain a multi-branch aggregated representation matrix J including all remaining nodes. The multi-branch aggregated representation matrix is ​​then fused with the normalized node position feature matrix X using gate control vectors. norm The optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes are obtained.

[0197] The optimization unit 203 is used to determine the number of clusters under the target location distribution and whether the UAV swarm has restored connectivity based on the optimal target location matrix; determine the connectivity loss based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold; traverse the number of clusters under the target location distribution and determine the inter-cluster distance loss based on the communication distance threshold and the maximum inter-cluster distance; obtain a joint loss function based on the movement distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss; iteratively optimize the joint loss function to obtain the final target location matrix that satisfies full connectivity and minimizes loss.

[0198] It should be understood that the units included in the above-described graph learning-based UAV swarm network self-healing method device are merely a logical division based on the functions implemented by the device. In practical applications, the units can be superimposed or split. Furthermore, the functions implemented by the graph learning-based UAV swarm network self-healing method device provided in this embodiment correspond one-to-one with the graph learning-based UAV swarm network self-healing method provided in the above embodiments. The more detailed processing flow implemented by this device has been described in detail in the first embodiment of the method above, and will not be described in detail here.

[0199] Another embodiment of the present invention provides a computer device, the computer device including: a processor and a scene database; the scene database is used to store computer program code, the computer program code including computer instructions; when the processor executes the computer instructions, the electronic device executes each step of the graph learning-based UAV swarm network self-healing method shown in the above method embodiment.

[0200] Another embodiment of the present invention provides a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform the various steps of the graph learning-based self-healing method for unmanned aerial vehicle swarm networks shown in the above-described method embodiments.

[0201] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

[0202] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A graph learning-based self-healing method for unmanned aerial vehicle (UAV) swarm networks, characterized in that, include: Based on each independent cluster in the residual graph at the moment of communication breakdown of the UAV swarm, construct an intra-cluster relative neighborhood graph, and construct a cluster-level relative neighborhood graph based on the distance between the nearest surviving node pairs of any independent cluster pair; obtain a global reconstructed adjacency matrix based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph. Using a parametric encoder, a global description vector and task descriptions at each node level are concatenated to generate node task embedding vectors and task embedding matrices. A normalized node position feature matrix and a task embedding matrix are then concatenated to obtain the input feature matrix for initial graph convolutional encoding and the initial graph convolutional output. The initial graph convolutional output is used to obtain graph convolutional encoded node representation vectors and layer output sequences based on multi-layer core graph convolution. The graph convolutional encoded node representation vectors included in the layer output sequences are weighted and fused based on attention weights to obtain a multi-branch aggregated representation matrix including all surviving nodes. The multi-branch aggregated representation matrix is ​​then fused with the normalized node position feature matrix using gate control vectors to obtain the optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes. The optimal target location matrix is ​​used to determine the number of clusters under the target location distribution and whether the UAV swarm has restored connectivity. The connectivity loss is determined based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold. The number of clusters under the target location distribution is traversed, and the inter-cluster distance loss is determined based on the communication distance threshold and the maximum inter-cluster distance. A joint loss function is obtained based on the travel distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss. The joint loss function is iteratively optimized to obtain the final target location matrix that satisfies full connectivity and minimizes loss.

2. The method as described in claim 1, characterized in that, For each independent cluster in the surviving graph at the moment of communication breakdown in the UAV swarm, an intra-cluster relative neighborhood graph is constructed. A cluster-level relative neighborhood graph is constructed based on the distance between the nearest surviving node pairs of any pair of independent clusters. Specifically, this includes: Based on each independent cluster in the residual graph, the set of node indices of the cluster, and the Euclidean distance between any two nodes within each independent cluster, construct an intra-cluster subgraph. Based on the intra-cluster node set and the edge set of the intra-cluster relative neighborhood graph, determine the intra-cluster relative neighborhood graph of the cluster. Traverse all node pairs included in any two independent clusters, determine the Euclidean distance of the node pair with the smallest distance, and construct a cluster-level relative neighborhood graph based on the Euclidean distance of the node pair with the smallest distance. The intra-cluster subgraph and intra-cluster relative neighborhood graph are shown below: For any pair of nodes in the relative neighborhood graph within the cluster, the following formula is satisfied: The nearest distance between any two independent clusters is determined by the following formula: When constructing a cluster-level relative neighborhood graph based on the nearest distance between any two independent clusters, the following formula is satisfied: max(d S (S i ,S k ),d S (S j ,S k ))<d S (S i ,S j ) The globally reconstructed adjacency matrix is ​​shown below: in, This represents the intra-cluster subgraph of the q-th cluster. Indicates cluster S q The set of nodes within a cluster, Indicates cluster S q Intra-cluster communication edge set, Indicates cluster S q The relative neighborhood graph within the cluster, Represents the edge set of the relative neighborhood graph within a cluster. Indicates cluster S q The internal index is c i The node, Indicates cluster S q The set of node indexes Indicates cluster S q The number of nodes, Indicates cluster S q internal nodes With nodes Euclidean distance, d S (S i ,S j ) represents cluster S i and cluster S j The Euclidean distance between the nearest surviving node pairs, where u represents the cluster S. i Any node within S, where v represents the cluster S j For any node within d, uv This represents the Euclidean distance between node u and node v. Describing the variable cluster S i and cluster S j For all pairs of nodes, take the minimum distance S. k Indicates the absence of cluster S i ,S j Any third cluster other than This indicates a global reconstruction of the adjacency matrix, and diag(·) represents the block diagonal stacking operator. This represents the intra-cluster reconstructed adjacency matrix of cluster S1. This represents the inter-cluster bridging matrix, where ∨ represents the element-wise logical OR operation. Represents the Nth digit in the residual network. S Clusters.

3. The method as described in claim 1, characterized in that, Before obtaining the initial graph convolution output, the process also includes: A unique diffusion intensity is assigned to each surviving node included in the surviving node position matrix, and the task embedding matrix determines the offset vector and scale vector corresponding to the surviving node through a small feedforward network. The diffusion intensity, the offset vector corresponding to the surviving node, and the scale vector corresponding to the surviving node are as follows: in, Represents the remaining nodes The diffusion intensity, sigmoid(·) represents the sigmoid activation function, Represents the remaining nodes The scale factor, Represents the remaining nodes The offset factor, ∈0 represents a predefined baseline, ∈0∈(0,1), Indicates global reconstruction of the adjacency matrix The infinite norm of , where δ represents the offset vector of all surviving nodes. F represents the scale vector of all surviving nodes. eps (·) indicates a small feedforward network. Represents the scale vector The natural logarithm of X emb Represents the task embedding matrix, N R Indicates the total number of remaining nodes. Represents the remaining nodes The task embedding vector, [·||·] represents the feature concatenation operator, N τ φ(·) represents the dimension of the task embedding vector, τ represents the parameterized encoder, and φ(·) represents the dimension of the task embedding vector. gb Represents the global description vector. Represents a node Node-level description vector, N represents τ A real vector space.

4. The method as described in claim 1, characterized in that, The initial graph convolution output is based on multi-layer core graph convolution to obtain the node representation vector and layer output sequence of graph convolution encoding, specifically including: The initial graph convolution is then encoded by a core graph convolution layer of Q-1 layers to obtain the node representation vector and layer output sequence of the graph convolution encoding. The processing procedure of the graph convolution unit in the l ∈ [2, Q] layer is as follows: H (0) =[X norm ||X emb ] The layer output sequence is shown below: Among them, H (l) Let H represent the node representation in the graph convolution of the l-th layer, σ(·) represent the element-wise nonlinear activation function, and H (l-1) This represents the node representation in the (l-1)th layer graph convolution. AGCO represents the intermediate result in the graph convolution of the l-th layer. (l) This represents the adaptive graph convolution operator in the l-th layer graph convolution. Let ε represent the global reconstructed adjacency matrix, and let ε represent the diffusion factor matrix. Represents the reconstruction of the adjacency matrix The corresponding Laplace matrix, W (l) The learnable linear transformation parameters of the graph convolution at the l-th layer are represented. N represents R ×N R 3D identity matrix, H (1) AGCO represents the node representation of the initial graph convolutional encoding. (1) X represents the adaptive graph convolution operator in the first layer of graph convolution. in Let X represent the input feature matrix of the adaptive graph convolution operator. emb H represents the task embedding matrix. (0) Let X represent the input feature matrix for the initial graph convolution. norm X represents the normalized node position feature matrix. emb Let [·||·] represent the task embedding matrix, [·||·] represent the feature concatenation operator, and Q represent the total number of convolutional layers in the core graph. The output sequence of the representation layer.

5. The method as described in claim 1, characterized in that, The weighted fusion of the node representation vectors of the graph convolutional encoding included in the layer output sequence based on attention weights yields a multi-branch aggregated representation matrix including all surviving nodes, specifically including: The node-level task description obtains a query vector based on the feedforward network, and determines the attention scores and attention weights of all surviving nodes for different branches based on the query vector and the node representation vector of the graph convolutional encoding. The node representation vectors of the graph convolutional encoding are weighted and fused according to the attention weights of all surviving nodes to different branches to obtain a multi-branch aggregated representation including all surviving nodes and a multi-branch aggregated representation matrix including all surviving nodes. The query vector, the attention scores of all surviving nodes for different branches, the attention weights of all surviving nodes for different branches, the multi-branch aggregate representation including all surviving nodes, and the multi-branch aggregate representation matrix including all surviving nodes are shown below: in, Represents the remaining nodes The query vector, F query (·) indicates a shared feedforward network. Represents the remaining nodes Node-level description vector, Represents the remaining nodes For the attention weights of the l-th layer branch, exp(·) denotes the exponential function. Represents the remaining nodes The attention score for the l-th layer, where Q represents the total number of convolutional layers in the core graph. Represents the query vector transpose, Represents the residual nodes in the l-th layer graph convolution. The representation vector, D hid Indicates the hidden layer dimension. Represents the remaining nodes Multi-branch aggregation representation. Represents the residual nodes in the l-th layer graph convolution. The representation vector, J represents the multi-branch aggregation representation matrix, N R This represents the total number of remaining nodes.

6. The method as described in claim 1, characterized in that, The process of fusing the multi-branch aggregation representation matrix and the normalized node position feature matrix using gate control vectors to obtain the optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes specifically includes: The gate control vector is obtained based on the multi-branch aggregation representation matrix, the normalized node position feature matrix, and the task embedding matrix; The optimal node fusion matrix, which includes all surviving nodes, is obtained by fusing the multi-branch aggregation representation matrix with the normalized node position feature matrix using gate control vectors. The decoder output determined based on the optimal node fusion matrix and the final output decoder is denormalized to obtain the optimal target position matrix including all surviving nodes. The gating vector, optimal node fusion matrix, final output decoder, and optimal target position matrix are shown below: μ=sigmoid(W gate [J;X norm W proj ;X emb ]) Y=µ⊙J+(1-µ)⊙(X norm W proj ) X out = tanh(F out (Y))+1 Where μ represents the gate vector, W gate Let X represent the learnable linear transformation matrix generated by the gated vector, J represent the multi-branch aggregation representation matrix, and X represent the multi-branch aggregation representation matrix. norm W represents the normalized node position feature matrix. proj X represents the learnable linear transformation matrix of the projection of the original positional features. emb Let X represent the task embedding matrix, ⊙ represent the element-wise multiplication operator, Y represent the node representation after gated fusion, 1-μ represent the element-wise complement of the gate vector μ, and X represent the task embedding matrix. out This represents the decoder output, tanh(·) represents the hyperbolic tangent activation function, and F... out (·) indicates the final output decoder. Represents the remaining nodes The optimal target position, Represents the remaining nodes At the original coordinates of time p cen Indicates the location center of the remaining network. Let P represent the maximum original distance from the node to the center, and let P represent the remaining node position matrix. Represents the remaining nodes The original coordinates.

7. The method as described in claim 1, characterized in that, The connectivity loss is determined based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold. A joint loss is obtained based on the travel distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss. Specifically, this includes: Based on the relative distance between the target locations of the nodes, the communication distance threshold yields a soft adjacency matrix with continuous weights; based on the soft adjacency matrix and the weighted degree matrix determined according to the soft adjacency matrix, a normalized Laplace matrix and an ascending sequence of eigenvalues ​​are obtained. The connectivity loss is determined based on the ascending eigenvalue sequence, the differentiable approximation of the number of connected components, and the number of clusters under the target location distribution. The soft adjacency matrix, normalized Laplacian matrix, number of connected components, differentiable approximation, connectivity loss, and joint loss are as follows: L joint =L P +n c L C +n CD L CD +n R L R +n U L U Where β represents the sharpness parameter, Represents the remaining nodes The target location d represents the Euclidean distance between two nodes at the target location. tr Indicates the communication distance threshold. Represents a soft adjacency matrix The corresponding normalized Laplace matrix, Represents a soft adjacency matrix The corresponding weighted degree matrix, Represents the weighted degree matrix The inverse square root matrix, N represents R ×N R 3D identity matrix L represents the soft adjacency matrix. C Indicates connectivity loss. The number of connected components is a differentiable approximation, K represents the hyperparameter, and |λ| represents the number of connected components. k | represents the eigenvalue λ k The absolute value of N R This represents the total number of surviving nodes, and γ represents the decay coefficient. C represents the second smallest eigenvalue of the Laplace matrix. true L represents the number of actual connected components under the target location distribution. joint L represents the joint loss function. P L represents the distance traveled. CD L represents the inter-cluster distance loss. R L represents the rejection of loss. U This indicates a loss of uniformity.

8. A graph learning-based self-healing device for unmanned aerial vehicle (UAV) swarm networks, characterized in that, include: The construction unit is used to construct an intra-cluster relative neighborhood graph for each independent cluster in the residual graph at the moment of communication failure of the UAV swarm, construct a cluster-level relative neighborhood graph based on the distance between the nearest surviving nodes of any independent cluster pair, and obtain a global reconstructed adjacency matrix based on the intra-cluster relative neighborhood graph and the cluster-level relative neighborhood graph. The unit is used to generate node task embedding vectors and task embedding matrices by concatenating the global description vector and the task descriptions at each node level using a parametric encoder; it also concatenates the normalized node position feature matrix and the task embedding matrix to obtain the input feature matrix and the initial graph convolutional output of the graph convolutional encoding; the initial graph convolutional output is used to obtain the node representation vector and the layer output sequence of the graph convolutional encoding based on multi-layer core graph convolution; the node representation vectors of the graph convolutional encoding included in the layer output sequence are weighted and fused based on attention weights to obtain a multi-branch aggregated representation matrix including all surviving nodes; the multi-branch aggregated representation matrix and the normalized node position feature matrix are fused through a gate control vector to obtain the optimal node fusion matrix including all surviving nodes and the optimal target position matrix including all surviving nodes; An optimization unit is used to determine the number of clusters under the target location distribution and whether the UAV swarm has restored connectivity based on the optimal target location matrix; determine the connectivity loss based on the optimal target location matrix, the number of clusters under the target location distribution, and the communication distance threshold; traverse the number of clusters under the target location distribution and determine the inter-cluster distance loss based on the communication distance threshold and the maximum inter-cluster distance; obtain a joint loss function based on the movement distance loss, the connectivity loss, the inter-cluster distance loss, the repulsion loss, and the uniformity loss; iteratively optimize the joint loss function to obtain the final target location matrix that satisfies full connectivity and minimizes the loss.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the graph learning-based self-healing method for unmanned aerial vehicle swarm networks as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the graph learning-based self-healing method for unmanned aerial vehicle (UAV) swarm networks as described in any one of claims 1-7.

Images