Unmanned aerial vehicle cluster network agglomeration clustering method and device based on graph attention network

Abstract

The invention provides an unmanned aerial vehicle cluster network agglomeration clustering method and device based on a graph attention network, and relates to the technical field of unmanned aerial vehicle cluster network clustering, and the method comprises the steps: obtaining an original graph of an unmanned aerial vehicle cluster network; based on the node matrix and an encoder of the current graph attention network, calculating a hidden representation of each unmanned aerial vehicle node under the influence of all corresponding neighbor nodes, then predicting a link probability between each unmanned aerial vehicle node and the corresponding neighbor node, and generating a reconstruction graph of the unmanned aerial vehicle cluster network; adjusting the graph attention network according to the reconstruction loss determined by the original graph and the reconstructed graph; combining the unmanned aerial vehicle nodes meeting a preset condition into a plurality of node clusters based on a condensation clustering algorithm and hidden representation, and generating a new unmanned aerial vehicle cluster network; performing agglomeration clustering operation on the new unmanned aerial vehicle cluster network; the technical problem that the clustering efficiency, the clustering accuracy and the graph structure information utilization rate of the unmanned aerial vehicle cluster network are not high is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of drone cluster network clustering, and in particular to a drone cluster network agglomerative clustering method and device based on a graph attention network. Background Art

[0002] UAV swarm networks have rapidly become a hot topic in academia and industry around the world. With the advantages of simple, fast deployment and strong robustness, they are widely used in military and civil fields, such as collaborative reconnaissance, precision agriculture, disaster management and environmental monitoring. Compared with traditional fixed networks, UAV networks have the characteristics of mobility, flexibility and self-organization, which enable them to quickly and spontaneously establish communication links. These advantages enable UAV swarm networks to perform tasks that are difficult to accomplish with traditional fixed networks. However, the inherent characteristics of UAV swarm networks, such as high mobility and limited energy, pose challenges to stable and reliable communication, so it is necessary to design and optimize the network architecture accordingly. In recent years, clustering has become a popular method for building stable UAV swarm networks. It has been widely recognized as an effective method to improve the coordination, scalability and reliability between UAVs in swarm networks. By dividing the UAV group into smaller, more manageable clusters, more efficient resource allocation, stronger communication links and overall control of the UAVs in the cluster can be provided. Clustering is a hierarchical network structure, each cluster contains a cluster head and multiple cluster members. The cluster head plays the role of a leader, responsible for controlling and coordinating the situational information of the cluster to which it belongs, while communicating with other cluster heads. Moreover, cluster members only need to communicate with their own cluster heads. This approach reduces routing maintenance overhead and improves the scalability of the network. In addition, clustering enhances the fault tolerance of the network because the failure of a cluster head or member only affects a small part of the network, not the entire network. This makes clustering very suitable for highly dynamic drone cluster networks.

[0003] For the realization of drone clustering in the drone cluster network scenario, the traditional method mainly adopts centralized clustering algorithm. Specifically, a central node such as a mobile aggregation node or a ground station is required to control the global network information. The central node is responsible for most of the data processing and decision-making. The entire drone cluster network is regarded as a whole cluster, and then the centralized clustering algorithm divides it into subclusters. The existing centralized clustering algorithms mainly include spectral clustering algorithm and fuzzy logic clustering algorithm. These algorithms rely on the central base station to assign low weights to the edges between points with a long distance and high weights to the edges between points with a short distance. The purpose of clustering is achieved by cutting the graph composed of all nodes and edges, so that the sum of the edge weights between different subgraphs after cutting is as low as possible, and the sum of the edge weights within the subgraph is as high as possible. In addition, there are some distributed clustering algorithms that predict link duration. At the same time, due to the rise of neural networks, neural networks are used to learn topological structure relationships in dynamic networks, which has achieved great success. For example, the autoencoder clustering algorithm realizes accurate clustering of dynamic cluster networks.

[0004] However, the inventors have found that the above clustering algorithms all have certain drawbacks in practical applications. For example, the clustering efficiency, clustering accuracy and graph structure information utilization of highly dynamic drone cluster networks are all negatively affected. Summary of the invention

[0005] The purpose of the present invention is to provide a method and device for agglomerative clustering of a drone cluster network based on a graph attention network, so as to alleviate the technical problems of low clustering efficiency, clustering accuracy and low utilization rate of graph structure information of the drone cluster network.

[0006] In a first aspect, an embodiment of the present invention provides a method for agglomerative clustering of a drone cluster network based on a graph attention network, comprising: Obtain an original graph of the drone cluster network; wherein the original graph is a node matrix for characterizing each drone node and at least one neighbor node corresponding to each drone node; Based on the node matrix and the encoder of the current graph attention network, a hidden representation of each drone node under the influence of all corresponding neighbor nodes is calculated; wherein the hidden representation is used to characterize the similarity between each drone node and the corresponding neighbor node; Predicting the link probability between each of the drone nodes and the corresponding neighbor nodes according to the hidden representation, and generating a reconstructed graph of the drone cluster network; A reconstruction loss is determined according to the difference between the original image and the reconstructed image, and a weight coefficient of the image attention network is adjusted to obtain a new current image attention network; Based on the agglomerative clustering algorithm and the hidden representation, the drone nodes that meet the preset conditions are merged into multiple node clusters to generate a new drone cluster network; An agglomerative clustering operation is performed on the new drone cluster network.

[0007] Furthermore, based on the node matrix and the encoder of the current graph attention network, the step of calculating the hidden representation of each drone node under the influence of all corresponding neighboring nodes includes: Under the action of the encoder in the current graph attention network, a neighbor node vector corresponding to the current drone node is obtained based on the original graph; Calculate the attention coefficient used to characterize the similarity between the current drone node and each neighbor node; Input the attention coefficient, the weight coefficient of the current graph attention network and the neighbor node vector corresponding to the current drone node into a nonlinear function to obtain a hidden representation of the current drone node under the influence of each corresponding neighbor node; The above steps are repeated until the hidden representation of each drone node in the original graph is obtained.

[0008] Furthermore, the step of calculating the attention coefficient used to characterize the similarity between the current drone node and each neighbor node includes: Calculate the distance similarity between the current drone node and each neighbor node according to the Euclidean distance between the current drone node and each neighbor node; Calculate the speed similarity between the current drone node and each neighbor node based on the speed comparison between the current drone node and each neighbor node at the current moment; Preset weights are configured for the distance similarity and the speed similarity, and an attention coefficient is calculated.

[0009] Furthermore, the step of predicting the link probability between each drone node and the corresponding neighbor node according to the hidden representation to generate a reconstructed graph of the drone cluster network includes: The vector set of all neighbor nodes corresponding to each drone node, the transposed matrix of the vector set of all neighbor nodes corresponding to each drone node and the attention coefficient are input into the activation function to generate a reconstructed graph of the drone cluster network; wherein the reconstructed graph is a node matrix used to characterize each drone node and the link probability between each drone node and the corresponding neighbor node.

[0010] Furthermore, based on the agglomerative clustering algorithm and the hidden representation, the drone nodes that meet the preset conditions are merged into multiple node clusters to generate a new drone cluster network, including: Under the action of the agglomerative clustering algorithm, according to the hidden representation of each of the drone nodes, determining the node and / or cluster with the maximum current affinity from the drone nodes; If the current affinity meets the preset affinity threshold requirement, the nodes and / or clusters with the maximum current affinity are merged into a node cluster; The above steps are repeated until all the nodes and / or clusters with the maximum current affinity among the drone nodes and meeting the preset affinity threshold requirements are merged to generate a new drone cluster network.

[0011] Furthermore, the step of performing agglomerative clustering operation on the new drone cluster network includes: Determining whether the reconstruction loss meets expectations; If the reconstruction loss does not meet expectations, return to the step of obtaining the original graph of the drone cluster network; If the reconstruction loss meets expectations, the agglomerative clustering result of the new drone cluster network is calculated.

[0012] Furthermore, the method further comprises: Based on the reconstruction loss and the clustering loss generated in the process of generating a new drone cluster network, the weight coefficient of the graph attention network is jointly adjusted to obtain a new current graph attention network.

[0013] In a second aspect, an embodiment of the present invention further provides a drone cluster network agglomerative clustering device based on a graph attention network, comprising: An acquisition module is used to acquire an original graph of the drone cluster network; wherein the original graph is a node matrix used to characterize each drone node and at least one neighbor node corresponding to each drone node; A calculation module, based on the node matrix and the encoder of the current graph attention network, calculates the hidden representation of each drone node under the influence of all corresponding neighbor nodes; wherein the hidden representation is used to characterize the similarity between each drone node and the corresponding neighbor node; A generation module predicts the link probability between each drone node and the corresponding neighbor node according to the hidden representation, and generates a reconstructed graph of the drone cluster network; An adjustment module, which calculates a reconstruction loss determined according to a difference between the original image and the reconstructed image, and adjusts a weight coefficient of the image attention network to obtain a new current image attention network; A merging module, which merges the drone nodes that meet the preset conditions into multiple node clusters based on the agglomerative clustering algorithm and the hidden representation to generate a new drone cluster network; The clustering module performs agglomerative clustering operations on the new drone cluster network.

[0014] In a third aspect, an embodiment provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the steps of the method described in any of the aforementioned implementation methods are implemented.

[0015] In a fourth aspect, an embodiment provides a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the steps of the method described in any one of the aforementioned implementations.

[0016] The embodiment of the present invention provides a method and device for agglomerative clustering of drone cluster networks based on a graph attention network. According to the attribute relationship between each node in the drone cluster network, the hidden representation of each drone node under the influence of the corresponding neighbor node is calculated, that is, the similarity between each drone node and its corresponding neighbor node can be known through the hidden representation, thereby clarifying the attribute relationship between each node in the drone cluster network; the probability of the existence of a link between each drone node and its corresponding neighbor node is predicted based on the hidden representation, thereby generating a reconstructed graph of the drone cluster network; the original graph and the reconstructed graph of the drone cluster network are difference calculated to obtain the reconstruction loss, thereby The weight coefficient of the attention network is adjusted so that the next time the drone cluster network is clustered using the graph attention network, more accurate results can be obtained; at the same time, based on the hidden representation of each drone node, the drone nodes that meet the preset conditions are merged under the action of the agglomerative clustering algorithm to form a new drone cluster network, and the new drone cluster network is subjected to agglomerative clustering to obtain more accurate clustering results; the present application uses a graph attention network to reconstruct the drone cluster network while taking into account the attributes between the nodes in the drone cluster network, and based on the reconstructed drone cluster network, agglomerative clustering can be performed to obtain more accurate clustering results.

[0017] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 A flow chart of a method for agglomerative clustering of drone cluster networks based on a graph attention network provided by an embodiment of the present invention; Figure 2 A schematic diagram of an application of a graph attention network agglomerative clustering algorithm provided in an embodiment of the present invention; Figure 3 A schematic diagram of the modularity evaluation results of the UAV cluster network agglomerative clustering results based on the graph attention network; Figure 4 Schematic diagram of the silhouette coefficient evaluation results of the agglomerative clustering results of the drone cluster network based on the graph attention network; Figure 5 A schematic diagram of functional modules of a drone cluster network agglomerative clustering device based on a graph attention network provided by an embodiment of the present invention; Figure 6 A schematic diagram of the hardware architecture of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] At present, the traditional clustering algorithm of highly dynamic UAV cluster networks still has some problems, such as: 1. The reliability and scalability of centralized clustering algorithms based on spectral clustering may be poor. In particular, when the central node fails, it will endanger the entire unmanned cluster network and reduce the clustering efficiency of the highly dynamic UAV cluster network.

[0023] 2. For the clustering algorithm for predicting link duration, it is a distributed agglomerative clustering algorithm that takes link duration into consideration. Compared with the centralized clustering method, it is more suitable for highly dynamic drone cluster networks. However, due to the continuous changes in the node positions of highly dynamic drone cluster networks, the clustering results of this type of algorithm are not very accurate.

[0024] 3. Although the traditional agglomerative clustering algorithm based on neural networks has improved the accuracy compared with the clustering algorithm for predicting link duration, it also has problems such as low effective utilization of graph structure information.

[0025] Based on this, an embodiment of the present invention provides a method and device for agglomerative clustering of a drone cluster network based on a graph attention network, which determines the similarity between nodes through the attributes between the graph attention network and the nodes in the drone cluster network, and merges the node clusters to achieve a more accurate agglomerative clustering result for the reconstructed drone cluster network.

[0026] To facilitate understanding of this embodiment, a drone cluster network agglomerative clustering method based on a graph attention network disclosed in an embodiment of the present invention is first introduced in detail. This method can be applied to intelligent control devices such as host computers, servers, and controllers.

[0027] Figure 1 A flow chart of a method for agglomerative clustering of drone cluster networks based on a graph attention network is provided in an embodiment of the present invention.

[0028] Reference Figure 1 As shown, the method comprises the following steps: Step S102, obtaining the original graph of the drone cluster network.

[0029] The original graph can be obtained from the routing table, and the original graph is a node matrix used to characterize each drone node i and at least one neighbor node j corresponding to each drone node i. Here, if two nodes perform information routing (communication), and the information routing jump is no more than 1 hop, then the relationship between the two nodes is neighbor nodes, that is, the communication between the two nodes does not perform multi-hop forwarding, then they are neighbors.

[0030] For example, the original graph G=Z_{ij}, I=J; the node matrix is ​​composed of 0 and 1, for example, Z_{ij}=1 means that there is a topological link between the i node and the j node; it should be noted that the number of neighbor nodes J is set based on the number I of drone nodes i so that the node matrix is ​​square.

[0031] Step S104, based on the node matrix and the encoder of the current graph attention network, calculate the hidden representation of each drone node under the influence of all corresponding neighbor nodes.

[0032] Among them, the hidden representation is used to characterize the similarity between each drone node and the corresponding neighbor node. The encoder in the aforementioned node matrix and graph attention network can obtain the similarity between each drone node and the corresponding neighbor node according to the attribute relationship between each node.

[0033] Step S106, predicting the link probability between each drone node and the corresponding neighbor node based on the hidden representation, and generating a reconstructed graph of the drone cluster network.

[0034] Here, based on the similarity between each drone node and the corresponding neighbor node referred to by the hidden representation, the link probability between each drone node and the corresponding neighbor node can be determined, and then the reconstructed graph of the drone cluster network can be obtained; the reconstructed graph is also a node matrix that can characterize the link relationship between each node, which can be understood as the reconstructed original graph.

[0035] Step S108, calculate the reconstruction loss determined according to the difference between the original image and the reconstructed image, and adjust the weight coefficient of the image attention network to obtain a new current image attention network.

[0036] In order to enable each drone node in the next drone cluster network using the graph attention network to have more accurate hidden representation extraction, the weight coefficient of the current graph attention network is adjusted here.

[0037] Step S110, based on the agglomerative clustering algorithm and hidden representation, the drone nodes that meet the preset conditions are merged into multiple node clusters to generate a new drone cluster network.

[0038] Here, under the action of the distributed agglomerative clustering algorithm, each drone node can be merged into several node clusters according to the similarity of the hidden representation of each drone node, thus realizing the new construction of the drone cluster network.

[0039] Step S112, performing agglomerative clustering operation on the new drone cluster network.

[0040] The new drone cluster network constructed based on the steps of the aforementioned embodiment can obtain more accurate agglomerative clustering results.

[0041] In a preferred embodiment of practical application, according to the attribute relationship between each node in the drone cluster network, the hidden representation of each drone node under the influence of the corresponding neighbor node is calculated, that is, the similarity between each drone node and its corresponding neighbor node can be known through the hidden representation, and then the attribute relationship between each node in the drone cluster network is clarified; the probability of the existence of a link between each drone node and its corresponding neighbor node is predicted based on this hidden representation, and then a reconstructed graph of the drone cluster network is generated; the original graph and the reconstructed graph of the drone cluster network are differenced to obtain the reconstruction loss, so as to adjust the weight coefficient of the graph attention network, so that the drone cluster network clustered by the graph attention network next time can obtain more accurate results; at the same time, according to the hidden representation of each drone node, under the action of the agglomerative clustering algorithm, the drone nodes that meet the preset conditions are merged to form a new drone cluster network, and the new drone cluster network is subjected to agglomerative clustering processing to obtain more accurate clustering results; the present application reconstructs the drone cluster network through the graph attention network, taking into account the attributes between each node in the drone cluster network, and agglomerative clustering can be performed based on the reconstructed drone cluster network to obtain more accurate clustering results.

[0042] UAV node in the embodiment of the present invention It only needs to have a partial view of neighboring nodes, use the graph attention network to learn the hidden representation of drone nodes, and use agglomerative clustering to merge drone nodes based on the learned hidden representation, thus achieving the optimal clustering of drone cluster networks. Figure 2 As shown: The clustering algorithm architecture provided by the embodiment of the present invention mainly includes two parts. One part is the graph attention network, which takes the mobile features (speed similarity) of the drone nodes and the graph structure (distance similarity) of the drone cluster network as input, and together constitutes the similarity of the drone nodes. In fact, the mobile features of the drone nodes represent the current speed of the drone nodes, and the graph structure of the drone cluster network represents the topological distance between the drone nodes. The graph attention network learns the hidden representation of the drone nodes by minimizing the reconstruction loss. The reconstruction loss is calculated based on the difference between the input original graph matrix and the reconstructed graph matrix obtained from the hidden representation. The graph attention network aims to learn a low-dimensional representation of the drone movement similarity to retain the key information of clustering.

[0043] Specifically, step S104 in the above embodiment can be implemented by the following steps, including: In step 1.1, under the action of the encoder in the current graph attention network, the neighbor node vector corresponding to the current drone node is obtained based on the original graph.

[0044] like Figure 2As shown in the figure, the encoder in the current graph attention network processes the original graph G to generate the neighbor node vector corresponding to the current drone node .

[0045] In step 1.2), the attention coefficient used to characterize the similarity between the current drone node and each neighbor node is calculated.

[0046] Exemplarily, the attention coefficient can be calculated by including the following steps: Step 1.2.1), based on the Euclidean distance between the current drone node and each neighbor node, calculate the distance similarity between the current drone node and each neighbor node; distance similarity It is expressed as:

[0047] in, Represents the drone node and The Euclidean distance between .

[0048] Step 1.2.2), based on the speed comparison between the current drone node and each neighbor node at the current moment, calculate the speed similarity between the current drone node and each neighbor node; speed similarity It can be expressed as:

[0049] in, is the speed of the current drone node i at time t, is the speed of neighbor node j corresponding to the current drone node i at time t.

[0050] Step 1.2.3), configure the preset weights for distance similarity and speed similarity, and calculate the attention coefficient. Calculated by the following formula:

[0051] in, Indicates the preset weights for speed similarity and distance similarity.

[0052] Step 1.3), the attention coefficient , the weight coefficient W of the current graph attention network and the neighbor node vector corresponding to the current drone node Input nonlinear function to get the hidden representation of the current drone node under the influence of each corresponding neighbor node , which is achieved by:

[0053] Each drone node learns a hidden representation by focusing on the mobility features and graph structure of its neighboring nodes. The hidden representation can be understood as the With neighbor nodes The similarity between them is measured by the speed similarity and topological distance similarity of the drone nodes.

[0054] from Figure 2 It can be seen that the node vector and attention coefficient The product is then combined with the weight coefficient of the graph attention network and processed by the nonlinear function RULE to obtain the hidden representation of the current drone node. .

[0055] Step 1.4), repeat the above steps 1.1) - step 1.3) until the hidden representation of each drone node in the original graph is obtained.

[0056] Here, the original graph corresponds to a drone cluster network, that is, when each drone node in the drone cluster network obtains a hidden representation through the steps of the above embodiment, the loop is terminated.

[0057] like Figure 2 As shown, the inner product decoder of the graph attention network combines the attention coefficient and the set Z of node vectors to generate a reconstructed graph. Further, step S106 of the above embodiment can be implemented by the following steps, specifically including: Step 2.1), the vector set of all neighbor nodes corresponding to each drone node , the transposed matrix of the vector set of all neighbor nodes corresponding to each drone node and attention coefficient Input the activation function to generate the reconstructed graph of the drone cluster network.

[0058] The reconstructed graph is a node matrix used to characterize each drone node and the link probability between each drone node and the corresponding neighbor node.

[0059] The inner product decoder predicts the links between drone nodes based on the hidden representation, expressed as:

[0060] in, Reconstructed graph representing the drone swarm network, is an activation function that maps the dot product to a value between 0 and 1, representing the probability of a link between a drone node and its corresponding neighbor node.

[0061] Based on the above embodiment, the loss function can be reconstructed in step S108 to measure the original image and reconstruction graph The difference between can be expressed as:

[0062] Among them, this application uses the binary cross entropy loss function as the reconstruction loss in the graph attention network. Minimizing the reconstruction loss enables the clustering algorithm to accurately reconstruct the drone graph structure, thereby enhancing the generalization ability of the graph attention network.

[0063] Further, such as Figure 2 As shown, the hidden representation of each drone node is clustered and merged to obtain a new drone cluster network; specifically, step S110 includes: In step 3.1), under the action of the agglomerative clustering algorithm, the nodes and / or clusters with the largest current affinity are determined from the drone nodes according to the hidden representation of each drone node.

[0064] The UAV nodes and / or the clusters formed by merging the UAV nodes are merged according to the learned hidden representation. In the distributed agglomerative clustering algorithm proposed in this application, the UAV nodes are clustered using a bottom-up approach. Specifically, UAV nodes as Initial clusters, the cluster after merging operation in the clustering process may be composed of multiple drone nodes; Agglomerative clustering can select two clusters with the largest affinity at a time based on the affinity between any two drone nodes / clusters and The merger can be expressed as:

[0065] Among them, the similarity between each drone node and the corresponding neighbor node represented by the hidden representation output by the graph attention network is transformed into a formula for calculating the cluster combination with the maximum similarity; is a function that measures the affinity between two clusters. and for any two nodes and / or clusters, respectively.

[0066] In step 3.2), if the current affinity meets the preset affinity threshold requirement, the nodes and / or clusters with the largest current affinity are merged into a node cluster.

[0067] Since the agglomerative clustering algorithm in this application is a distributed algorithm and is unsupervised, in order to avoid all nodes becoming a cluster and causing cluster imbalance, a preset affinity threshold for merging drone nodes is pre-set. , this threshold ensures that only similarity greater than The drone nodes / clusters will be merged to form a new node cluster. This node cluster can be understood as being formed by the merger of two drone nodes, two clusters, or one cluster and one drone node.

[0068] Step 3.3), repeat the above steps 3.1) - step 3.2) until all the nodes and / or clusters with the maximum affinity among the drone nodes and meeting the preset affinity threshold requirements are merged to generate a new drone cluster network.

[0069] It can be understood that when the node with the greatest affinity in the drone node cluster does not meet the preset affinity threshold requirement, the merger is terminated at this time, that is, the nodes and / or clusters with the greatest affinity that meet the preset affinity threshold requirement have all been merged.

[0070] In order to further ensure the accuracy of the agglomerative clustering results, the drone cluster architecture can be optimized through the aforementioned embodiment steps, and agglomerative clustering can be performed on this basis; exemplary S112 can be implemented through the following steps: Step 4.1) determines whether the reconstruction loss meets expectations.

[0071] In step 4.2), if the reconstruction loss does not meet expectations, the current drone cluster architecture can continue to be optimized, and the process returns to step S102, where the current drone cluster architecture obtains a new original graph and performs the operations in the aforementioned embodiment.

[0072] In step 4.3), if the reconstruction loss meets expectations, the agglomerative clustering result of the new drone cluster network is calculated.

[0073] As an optional embodiment, in order to further make the hidden representation of each drone node output by the graph attention network more accurate, the method of the present application also includes: In step 5.1), based on the reconstruction loss and the clustering loss generated in the process of generating a new drone cluster network, the weight coefficient of the graph attention network is jointly adjusted to obtain the new current graph attention network.

[0074] Here, when clustering stops in the aforementioned embodiment, a clustering loss function that considers not only the attention coefficient between the two clusters but also the local structure around the clusters can be calculated. It is expressed as:

[0075] in, Indicates that each iteration is associated with the drone node Move the neighbor node with the highest similarity. Therefore, Represents a drone node About Neighbor Nodes The hidden representation of Represents a drone node Its nearest neighbor node The difference between the mobile similarity of and the mobile similarity of other neighboring nodes; parameter Proper weightage is given to both parts; Represents a drone node The number of neighbor nodes; the nearest neighbor node closest to the drone node can be found by comparing the distance .

[0076] At this time, the reconstruction loss of the graph attention network is jointly optimized and the clustering loss of agglomerative clustering , and the total loss function Defined as:

[0077] in, is to weigh the reconstruction loss and clustering loss preset ratio.

[0078] The embodiment of the present invention can effectively utilize the mutual relationship between the inherent attributes of nodes in the unmanned platform cluster network, introduce a graph attention network, which pays attention to the neighbors of each node to learn the hidden representation of each node, and combines the node attribute value with the topological structure of the drone cluster network. The optimal clustering result is obtained by jointly optimizing agglomerative clustering and graph attention network. Since drone nodes only need to pay attention to the local view of neighboring nodes, there is no need to pay attention to the global situation information of the network, and the graph attention neural network introduces an attention mechanism, which can more effectively utilize the topological structure relationship. Therefore, this distributed clustering algorithm is more suitable for highly dynamic drone cluster networks, and can effectively utilize the mobile feature relationship between unmanned cluster networks to achieve the optimal clustering effect.

[0079] In practical applications, the embodiments of the present invention simulated drone cluster networks with 200, 400, 600, 800, and 1000 drone nodes, and used silhouette coefficient and modularity to evaluate clustering effects.

[0080] Among them, the SC algorithm is the spectral clustering algorithm, which mainly adopts the idea of ​​graph cutting. The edge weight value between two points that are far away is lower, while the edge weight value between two points that are close is higher. By cutting the graph composed of all data points, the sum of the edge weights between different subgraphs after cutting is made as low as possible, while the sum of the edge weights within the subgraph is made as high as possible, thereby achieving the purpose of clustering.

[0081] The GT algorithm is an algorithm that uses coalition game theory to predict the link duration of unmanned cluster networks. In the coalition game, a group of players hope to cooperate by forming a coalition to improve the efficiency of themselves and the entire cluster network. Coalition game theory aims to model cluster formation by identifying drones to players and clusters to coalitions. Each drone makes a decision based on the value of the alliance, allowing drones to join or leave the alliance according to their preferences. In this way, the proposed coalition formation algorithm converges to the Nash equilibrium solution, effectively improving the stability of the network.

[0082] like Figure 3 As shown in the figure, GCCA is the application of the "UAV cluster network agglomerative clustering algorithm based on graph attention network" proposed in this application. Modularity is a commonly used network clustering effect evaluation coefficient, which is defined as the ratio of the total number of edges within the cluster to the total number of edges in the cluster network. The larger the value, the better the clustering effect. Figure 3 As shown, the modularity values ​​of the spectral clustering algorithm and the game theory method are both below 0.5, while the modularity value of GCCA is above 0.7 on average, indicating that the clustering effect of the embodiment of the present invention is the best. The clustering algorithm of the embodiment of the present invention is distributed and does not need to be classified according to global information, and is more suitable for highly dynamic drone cluster networks.

[0083] like Figure 4 As shown in the figure, GCCA is the application of the "UAV cluster network agglomerative clustering algorithm based on graph attention network" proposed in this application. The silhouette coefficient is also a commonly used network clustering effect evaluation indicator. It can be understood as a measure of the clarity of the outline of each cluster after clustering. The larger the value, the better the clustering effect. Figure 4 As shown, the modularity values ​​of the spectral clustering algorithm and the game theory method are both below 0.45, while the modularity value of GCCA is above 0.55 on average, indicating that the clustering effect of the embodiment of the present invention is the best.

[0084] In the embodiment of the present invention, the drone node only needs to autonomously select the drone node for agglomerative clustering based on the movement similarity and distance similarity of the neighboring nodes. The core idea of ​​agglomerative clustering is to merge two clusters at each step based on the movement similarity between nodes. Each unmanned node is initially regarded as a cluster, and then the two clusters with the highest similarity are merged, iteratively, to solve the clustering problem in the highly dynamic drone cluster network scenario.

[0085] In some embodiments, Figure 5 As shown, an embodiment of the present invention provides a drone cluster network agglomerative clustering device based on a graph attention network, comprising: An acquisition module is used to acquire an original graph of the drone cluster network; wherein the original graph is a node matrix used to characterize each drone node and at least one neighbor node corresponding to each drone node; A calculation module, based on the node matrix and the encoder of the current graph attention network, calculates the hidden representation of each drone node under the influence of all corresponding neighbor nodes; wherein the hidden representation is used to characterize the similarity between each drone node and the corresponding neighbor node; A generation module predicts the link probability between each drone node and the corresponding neighbor node according to the hidden representation, and generates a reconstructed graph of the drone cluster network; An adjustment module, which calculates a reconstruction loss determined according to a difference between the original image and the reconstructed image, and adjusts a weight coefficient of the image attention network to obtain a new current image attention network; A merging module, which merges the drone nodes that meet the preset conditions into multiple node clusters based on the agglomerative clustering algorithm and the hidden representation to generate a new drone cluster network; The clustering module performs agglomerative clustering operations on the new drone cluster network.

[0086] Furthermore, the calculation module is used to obtain a neighbor node vector corresponding to the current drone node based on the original graph under the action of the encoder in the current graph attention network; calculate an attention coefficient used to characterize the similarity between the current drone node and each neighbor node; input the attention coefficient, the weight coefficient of the current graph attention network and the neighbor node vector corresponding to the current drone node into a nonlinear function to obtain a hidden representation of the current drone node under the influence of each corresponding neighbor node; repeat the above steps until the hidden representation of each drone node in the original graph is obtained.

[0087] Furthermore, the calculation module is used to calculate the distance similarity between the current drone node and each neighbor node based on the Euclidean distance between the current drone node and each neighbor node; calculate the speed similarity between the current drone node and each neighbor node based on the speed comparison between the current drone node and each neighbor node at the current moment; configure preset weights for the distance similarity and the speed similarity, and calculate the attention coefficient.

[0088] Furthermore, the generation module is used to input the vector set of all neighbor nodes corresponding to each drone node, the transposed matrix of the vector set of all neighbor nodes corresponding to each drone node and the attention coefficient into the activation function to generate a reconstructed graph of the drone cluster network; wherein the reconstructed graph is a node matrix used to characterize each drone node and the link probability between each drone node and the corresponding neighbor node.

[0089] Furthermore, the merging module is used to determine the nodes and / or clusters with the highest current affinity from the drone nodes according to the hidden representation of each of the drone nodes under the action of the agglomerative clustering algorithm; if the current affinity meets the preset affinity threshold requirement, the nodes and / or clusters with the highest current affinity are merged into a node cluster; the above steps are repeated until all the nodes and / or clusters with the highest current affinity in the drone nodes and meeting the preset affinity threshold requirement are merged, thereby generating a new drone cluster network.

[0090] Furthermore, the clustering module is used to determine whether the reconstruction loss meets expectations; if the reconstruction loss does not meet expectations, the step of obtaining the original graph of the drone cluster network is returned to execution; if the reconstruction loss meets expectations, the agglomerative clustering result of the new drone cluster network is calculated.

[0091] Furthermore, the device is also used to jointly adjust the weight coefficient of the graph attention network based on the reconstruction loss and the clustering loss generated in the process of generating a new drone cluster network to obtain a new current graph attention network.

[0092] An embodiment of the present invention provides an electronic device for implementing an electronic device. In this embodiment, the electronic device may be, but is not limited to, a personal computer (PC), a notebook computer, a monitoring device, a server, or other computer device with analysis and processing capabilities.

[0093] As an exemplary embodiment, see Figure 6 The electronic device 110 includes a communication interface 111, a processor 112, a memory 113 and a bus 114. The processor 112, the communication interface 111 and the memory 113 are connected via the bus 114. The memory 113 is used to store a computer program that supports the processor 112 to execute the method. The processor 112 is configured to execute the program stored in the memory 113.

[0094] The machine-readable storage medium mentioned in this article can be any electronic, magnetic, optical or other physical storage device that can contain or store information, such as executable instructions, data, etc. For example, the machine-readable storage medium can be: RAM (Radom Access Memory), volatile memory, non-volatile memory, flash memory, storage drive (such as hard disk drive), any type of storage disk (such as CD, DVD, etc.), or similar storage medium, or a combination thereof.

[0095] The non-volatile medium may be a non-volatile memory, a flash memory, a storage drive (such as a hard drive), any type of storage disk (such as a CD, DVD, etc.), or a similar non-volatile storage medium, or a combination thereof.

[0096] It can be understood that the specific operation methods of each functional module in this embodiment can refer to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0097] The computer-readable storage medium provided in the embodiment of the present invention stores a computer program. When the computer program code is executed, the method described in any of the above embodiments can be implemented. For specific implementation, please refer to the method embodiment, which will not be described in detail here.

[0098] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0099] In addition, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0100] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0101] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the aforementioned embodiments, those of ordinary skill in the art should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the aforementioned embodiments within the technical scope disclosed by the present invention, or can easily conceive of changes, or make equivalent replacements for some of the technical features therein. Such modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention.

Claims

1. A UAV cluster network agglomerative clustering method based on graph attention network, characterized in that: include: Obtain an original graph of the drone cluster network; wherein the original graph is a node matrix for characterizing each drone node and at least one neighbor node corresponding to each drone node; Based on the node matrix and the encoder of the current graph attention network, a hidden representation of each drone node under the influence of all corresponding neighbor nodes is calculated; wherein the hidden representation is used to characterize the similarity between each drone node and the corresponding neighbor node; Predicting the link probability between each of the drone nodes and the corresponding neighbor nodes according to the hidden representation, and generating a reconstructed graph of the drone cluster network; A reconstruction loss is determined according to a difference between the original image and the reconstructed image, and a weight coefficient of the image attention network is adjusted to obtain a new current image attention network; Based on the agglomerative clustering algorithm and the hidden representation, the drone nodes that meet the preset conditions are merged into multiple node clusters to generate a new drone cluster network; An agglomerative clustering operation is performed on the new drone cluster network.

2. The method according to claim 1, characterized in that The step of calculating the hidden representation of each of the drone nodes under the influence of all corresponding neighbor nodes based on the node matrix and the encoder of the current graph attention network comprises: Under the action of the encoder in the current graph attention network, a neighbor node vector corresponding to the current drone node is obtained based on the original graph; Calculate the attention coefficient used to characterize the similarity between the current drone node and each neighbor node; Input the attention coefficient, the weight coefficient of the current graph attention network and the neighbor node vector corresponding to the current drone node into a nonlinear function to obtain a hidden representation of the current drone node under the influence of each corresponding neighbor node; The above steps are repeated until the hidden representation of each drone node in the original graph is obtained.

3. The method according to claim 2, characterized in that The steps for calculating the attention coefficient used to characterize the similarity between the current drone node and each neighbor node include: Calculate the distance similarity between the current drone node and each neighbor node according to the Euclidean distance between the current drone node and each neighbor node; Calculate the speed similarity between the current drone node and each neighbor node based on the speed comparison between the current drone node and each neighbor node at the current moment; Preset weights are configured for the distance similarity and the speed similarity, and an attention coefficient is calculated.

4. The method according to claim 3, characterized in that The step of predicting the link probability between each drone node and the corresponding neighbor node according to the hidden representation to generate a reconstructed graph of the drone cluster network includes: The vector set of all neighbor nodes corresponding to each drone node, the transposed matrix of the vector set of all neighbor nodes corresponding to each drone node and the attention coefficient are input into the activation function to generate a reconstructed graph of the drone cluster network; wherein the reconstructed graph is a node matrix used to characterize each drone node and the link probability between each drone node and the corresponding neighbor node.

5. The method according to claim 1, characterized in that The step of merging the drone nodes that meet the preset conditions into multiple node clusters based on the agglomerative clustering algorithm and the hidden representation to generate a new drone cluster network includes: Under the action of the agglomerative clustering algorithm, according to the hidden representation of each of the drone nodes, determining the node and / or cluster with the maximum current affinity from the drone nodes; If the current affinity meets the preset affinity threshold requirement, the nodes and / or clusters with the maximum current affinity are merged into a node cluster; The above steps are repeated until all the nodes and / or clusters with the maximum current affinity among the drone nodes and meeting the preset affinity threshold requirements are merged to generate a new drone cluster network.

6. The method according to claim 1, characterized in that The step of performing agglomerative clustering operation on the new drone cluster network comprises: Determining whether the reconstruction loss meets expectations; If the reconstruction loss does not meet expectations, return to the step of obtaining the original graph of the drone cluster network; If the reconstruction loss meets expectations, the agglomerative clustering result of the new drone cluster network is calculated.

7. The method according to claim 1, characterized in that The method further comprises: Based on the reconstruction loss and the clustering loss generated in the process of generating a new drone cluster network, the weight coefficient of the graph attention network is jointly adjusted to obtain a new current graph attention network.

8. A UAV swarm network agglomerative clustering device based on graph attention network, characterized in that: include: An acquisition module is used to acquire an original graph of the drone cluster network; wherein the original graph is a node matrix used to characterize each drone node and at least one neighbor node corresponding to each drone node; A calculation module, based on the node matrix and the encoder of the current graph attention network, calculates the hidden representation of each drone node under the influence of all corresponding neighbor nodes; wherein the hidden representation is used to characterize the similarity between each drone node and the corresponding neighbor node; A generation module predicts the link probability between each drone node and the corresponding neighbor node according to the hidden representation, and generates a reconstructed graph of the drone cluster network; An adjustment module, which calculates a reconstruction loss determined according to a difference between the original image and the reconstructed image, and adjusts a weight coefficient of the image attention network to obtain a new current image attention network; A merging module, which merges the drone nodes that meet the preset conditions into multiple node clusters based on the agglomerative clustering algorithm and the hidden representation to generate a new drone cluster network; The clustering module performs agglomerative clustering operations on the new drone cluster network.

9. An electronic device, characterized in that: The method comprises a memory, a processor, and a program stored in the memory and capable of being run on the processor, wherein the processor implements the method according to any one of claims 1 to 7 when executing the program.

10. A computer-readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 is implemented.

Images