An unmanned aerial vehicle network intelligent clustering algorithm based on K-means clustering and differential evolution

Abstract

The application discloses an unmanned aerial vehicle network intelligent clustering algorithm based on K-means clustering and differential evolution, and belongs to the technical field of unmanned aerial vehicle network algorithms. A model of an unmanned aerial vehicle node clustering topological configuration and an optimization clustering problem is established; a K-means clustering algorithm is integrated into a differential evolution algorithm, the intelligent optimization iteration capability of the differential evolution algorithm is combined with the clustering task orientation performance of the K-means clustering algorithm, and the optimization clustering problem is solved; according to the unmanned aerial vehicle node clustering topological configuration, a clustering leader node election rule is designed, connection is established between different clustering leader nodes based on a nearest principle, and a clustering network of an unmanned aerial vehicle cluster is formed. The application aims at intelligently fusing the differential evolution algorithm and the K-means clustering thought, intelligently iteratively optimizing, quickly determining optimal clustering center distribution, electing a leader node and constructing a hybrid topological structure of an "intra-cluster star + inter-cluster mesh", effectively improving network robustness and connectivity, and providing a reliable communication topological basis for network routing design.

Description

Technical Field

[0001] This invention relates to the field of UAV network algorithm technology, specifically to an intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution. Background Technology

[0002] Drone networks are large-scale and their nodes are highly dynamic and mobile. To facilitate network management, reduce network complexity, adapt to high dynamic mobility, enhance topology stability, and improve communication efficiency and throughput, drone networks need to be clustered. By dividing them into multiple logical sub-clusters, hierarchical control and data aggregation can be achieved, enabling large-scale dynamic drone networks to operate efficiently and reliably. The drone swarm clustering problem is essentially a complex combinatorial optimization problem. It requires rationally dividing a large number of drone nodes distributed in three-dimensional space into several clusters, each with a cluster center, such that the sum of the distances from all nodes to their respective cluster centers is minimized. This optimization objective directly relates to the energy consumption, link quality, and network management efficiency of intra-cluster communication, and is one of the core evaluation criteria for the overall drone swarm communication scheme. Raza et al. proposed a FANET topology management method based on adaptive K-means clustering. By dynamically adjusting the transmission range of drone nodes and optimizing the cluster head selection process, it effectively reduces packet loss rate and node energy consumption, improves link quality, and reduces routing overhead. Compared with ant colony optimization and particle swarm optimization algorithms, it has better performance in terms of energy consumption and cluster construction time. Li Peng et al. proposed a location deployment method based on an improved and efficient global optimization algorithm to address the problem of three-dimensional spatial deployment of UAV base stations. By introducing an adaptive selection mechanism based on the offspring population generation strategy to optimize the differential evolution algorithm, the optimization capability is improved. Under multiple constraints, the signal propagation path loss is minimized, verifying the advantages of this method in terms of deployment accuracy.

[0003] The core idea of ​​the clustering scheme based on the K-means clustering algorithm is to divide network nodes into K clusters through iterative optimization. K initial centroids are randomly selected, and the distance from each node to each centroid (usually using Euclidean distance) is calculated. Nodes are then assigned to the cluster containing the nearest centroid. The centroid of each cluster (i.e., the mean of all points within the cluster) is recalculated, and this process is repeated until the change in centroid position is less than a preset threshold or the maximum number of iterations is reached. This algorithm has advantages in clustering tasks, such as fast execution speed and strong task specificity. However, the K-means algorithm itself does not possess intelligent optimization capabilities and is prone to getting trapped in local optima in high-dimensional problems. This weakness is further amplified in drone swarms with hundreds of nodes, and even traditional optimization schemes such as K-means++ struggle to address this issue.

[0004] The core idea of ​​the intelligent optimization clustering scheme based on differential evolution algorithm is to use the difference vectors among individuals in the population for mutation. By continuously performing the mutation-crossover-selection optimization cycle on the population, after multiple iterations, the population's optimal solution will gradually approach the theoretical optimal solution. Even if a candidate solution vector gets stuck in a local optimum, it may find an opportunity to escape the local optimum through individual mutation, exhibiting strong global convergence ability and robustness. However, the lack of targeted correction of the clustering objective leads to low population search efficiency.

[0005] Existing clustering network technologies suffer from low efficiency, mostly relying on traditional intelligent optimization algorithms or traditional clustering algorithms, failing to simultaneously leverage the advantages of both. This results in low search efficiency or insufficient search breadth, making it prone to getting trapped in local optima and difficult to train excellent clustering network schemes under limited time or computing power constraints. Specifically, the iterative process of differential evolution algorithms mainly relies on three operations: mutation, crossover, and selection. Individuals represent a set of candidate cluster center coordinates, which are perturbed and recombined in a continuous search space through difference vectors, gradually approaching the global optimum. However, the update direction of the test vectors generated by mutation and crossover operations depends entirely on the difference information of other individuals in the population, causing individuals to face great directional uncertainty during the search process. Ultimately, a large number of training rounds are wasted on ineffective oscillations and detours, and the fitness of the best individual in the population decreases extremely slowly, resulting in low convergence efficiency. To overcome this fundamental deficiency, this invention combines the K-means clustering algorithm—a local optimization strategy that directly reflects cluster structure characteristics—with the global search capability of the differential evolution algorithm. After each generation of evolution in the differential evolution process, the K-means clustering approach is used to locally correct the center position, adjusting it towards a direction closer to the current cluster structure. This effectively reduces the fitness value of each individual without disrupting population diversity. This method fully leverages the search breadth advantage of the differential evolution algorithm's random mutation and the search efficiency advantage of the K-means clustering algorithm's center migration, achieving more efficient and effective UAV swarm clustering and networking performance.

[0006] To address the issues of low search efficiency and long training rounds in existing traditional intelligent optimization algorithms such as differential evolution in clustering networking scenarios, and the problems of insufficient search breadth and easy getting trapped in local optima in traditional clustering networking scenarios such as K-means, this invention proposes an intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution. It integrates the intelligent optimization capability of differential evolution algorithm in autonomously searching for optimal solutions through random mutation with the excellent performance of K-means algorithm in efficiently iterating cluster centers with the goal of optimization in clustering problems, thereby improving the performance of UAV clustering networking. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution, belonging to the field of UAV network algorithm technology. It establishes a model of the UAV node clustering topology and an optimal clustering problem; integrates K-means clustering with the differential evolution algorithm, combining the intelligent optimization iteration capability of differential evolution with the task-oriented performance of K-means clustering to solve the optimal clustering problem; based on the UAV node clustering topology, it designs cluster leader node election rules, establishing connections between different cluster leader nodes based on the proximity principle to form a UAV swarm clustering network. This invention aims to quickly determine the optimal cluster center distribution, elect leader nodes, and construct a hybrid topology structure of "intra-cluster star + inter-cluster mesh" through intelligent iterative optimization by intelligently integrating differential evolution and K-means clustering ideas, effectively improving network robustness and connectivity, and providing a reliable communication topology foundation for network routing design.

[0008] A smart clustering algorithm for unmanned aerial vehicle (UAV) networks based on K-means clustering and differential evolution includes the following steps:

[0009] Step 1: Establish a model for the clustering topology of UAV nodes and the optimal clustering problem;

[0010] Step 2, construct the K-means differential evolution algorithm:

[0011] By integrating the K-means clustering algorithm into the differential evolution algorithm, the intelligent optimization and iteration capability of the differential evolution algorithm is combined with the clustering task-oriented performance of the K-means clustering algorithm to solve the optimal clustering problem.

[0012] Step 3: Construct a clustered network for the drone swarm:

[0013] Based on the cluster topology of UAV nodes, a cluster leader node election rule is designed, and connections are established between different cluster leader nodes based on the proximity principle to form a UAV swarm cluster network.

[0014] Furthermore, in step one, the establishment of the UAV node cluster topology is as follows:

[0015] Unmanned aerial vehicle network recording<GS,NU> ; where GS represents the ground control station, and NU represents the set of all nodes in the UAV swarm; under the action of the clustering algorithm, the nodes in the entire UAV swarm network will be divided into several clusters; for ease of management, the UAV nodes in each cluster are further divided into two categories, namely leader nodes (CH) and ordinary nodes (CM).

[0016] The responsibilities of the ground control station (GS): As the command center of the entire UAV swarm network, the ground control station controls the flight process, flight route, payload mission functions, and normal operation of communication links of each UAV node.

[0017] The responsibilities of a common node (CM): Each UAV common node flies in its own flight mode and performs a predetermined task; in addition, UAV nodes cooperate with each other to form a self-organizing network;

[0018] The responsibilities of the leader node (CH): In addition to performing the functions of ordinary nodes, the drone leader node is also responsible for managing the nodes within its corresponding cluster; for example... Figure 1 As shown.

[0019] Furthermore, in step one, the process of establishing the optimal clustering problem model is as follows:

[0020] In three-dimensional space, given N randomly distributed nodes, the horizontal plane coordinate range is... and Vertical coordinate range ,in , , These represent the maximum values ​​of the coordinate ranges in the X, Y, and Z directions, respectively; the nodes are divided into S clusters according to task requirements; the goal is to determine the coordinates of the S cluster centers, i.e., each center has 3 coordinate parameters, for a total of... A set of decision variables is used to minimize the total distance from all nodes to their nearest cluster center.

[0021] The optimization problem model for clustering drone swarms is as follows:

[0022] Let the set of cluster centers be... Among them, cluster centers ;

[0023] The total decision variable vector P is:

[0024] ;

[0025] Where T represents the transpose; R represents the set of real numbers; x i y i z i These represent the coordinates in the X, Y, and Z directions, respectively.

[0026] Minimizing the objective function f(p) yields the optimization problem model minf(p):

[0027] To minimize the total distance from all nodes to their nearest cluster centers, the constraint is that the X-coordinate of any node does not exceed [0, ... The range, where the Y-axis coordinate does not exceed [0, ... The range, the Z-direction coordinate does not exceed [0, ]scope;

[0028] ;

[0029]

[0030] in, Let be the coordinates of the k-th node; Let i be the coordinates of the i-th cluster center; This represents Euclidean distance.

[0031] Furthermore, in step two, the K-means differential evolution algorithm is constructed as follows:

[0032] Step 2.1, Population Initialization:

[0033] Initialize by assigning initial values ​​to each dimension of each individual in the population. Each individual x in the population... i,G It is expressed as follows:

[0034] ;

[0035] Where, x i,G In the G-th iteration of the population, the th iteration is... Individual, This represents the population size, i.e., NP=3S; Indicates the generation number;

[0036] Considering that the optimal cluster center of a sub-cluster must be inside the polyhedron formed by all member nodes of the subset, when initializing nodes, S nodes are directly selected as the initial cluster center positions according to the number of clusters, and random perturbation is added to give the initial exploration direction.

[0037] Step 2.2, Mutation:

[0038] For each individual goal x i,G After the population initialization is completed, for each iteration of the population, a mutation operation is first performed, randomly selecting two other individuals in the population to generate mutation vectors. Calculate using one of the following two formulas, each with a 50% probability:

[0039] ;

[0040] ;

[0041] in, , and Both are target vectors, representing the currently feasible solutions chosen by the individuals. The individual index "best" indicates the individual with the best current fitness in the entire population, i.e., the individual closest to the optimal solution. It is required that the randomly selected individuals r1, r2, and r3 are all distinct and also distinct from i. Therefore, the population size must contain at least four nodes. ; variable factor It is a real variable that controls the scaling of the deviation variable; This represents the value of the mutation factor for individual i in the Gth iteration;

[0042] An adaptive adjustment strategy is used to change the mutation factor during the iteration process. Value:

[0043] ;

[0044] in, and and represent the minimum and maximum mutation factor values ​​among all individuals in the population at the current iteration round, respectively; the fitness f is determined by the optimization objective, the smaller f is, the better the individual's fitness and the closer it is to the optimal solution; and These represent the minimum and maximum fitness values ​​among all individuals in the population at the current iteration round, respectively. This represents the fitness value of the current individual in the current iteration round. As a regulating factor;

[0045] Step 2.3, Cross:

[0046] The target vector reflects the feasible solution currently chosen by the individual, while the mutation vector reflects the potential direction of change of the individual; the crossover operation combines the target vector and the mutation vector, replacing the values ​​of several dimensions in the target vector with the mutated values ​​to obtain the test vector, thereby realizing the iteration of the individual;

[0047] The target vector and the mutated vector are combined, and the values ​​of several dimensions in the target vector are replaced with the mutated values ​​to obtain the test vector; the test vector :

[0048] ;

[0049] The above formula, ;

[0050] Where D represents the individual dimension, and Let x represent two random sequences with values ​​in the ranges [0,1] and [1,D] respectively in the G-th iteration; ji,G+1 v represents the target vector value of dimension j in the (G+1)th iteration; ji,G+1Represents the mutation vector value of dimension j in the (G+1)th iteration, u ji,G+1 This represents the test vector value of the j-th dimension in the (G+1)-th iteration;

[0051] Crossover probability Let represent the crossover probability of the i-th individual in the G-th iteration, with a value range of [0,1], controlling the probability that the mutated vector is selected as the test vector; an adaptive adjustment strategy is used to change the crossover probability during the iteration process. Value:

[0052] ;

[0053] in, and represents the minimum and maximum mutation factor values ​​among all individuals in the population, respectively; the fitness f is determined by the optimization objective, the smaller f is, the better the individual's fitness and the closer it is to the optimal solution; and These represent the minimum and maximum fitness values ​​among all individuals in the population, respectively. This represents the fitness value of the current individual in the Gth iteration. As a regulating factor;

[0054] Step 2.4, Boundary Condition Handling:

[0055] If a solution outside the feasible region appears during the mutation process, i.e. or Therefore, the test vectors must be restricted to the feasible region, and the corrected test vectors can be calculated using the following formula. :

[0056] ;

[0057] Where, x j (L) represents the lower boundary of the j-th dimension, U j (L) represents the upper boundary of the j-th dimension;

[0058] Step 2.5, select:

[0059] The K-means differential evolution algorithm uses a greedy criterion to process the test vectors... Compared with the target vector in the current population The next generation of target vectors is obtained through comparison and iteration. If the target vector has a lower fitness value and is closer to the optimal solution, the target vector is selected; otherwise, the test vector is selected. The test vector is compared only with the target vector, not with all individuals in the existing population.

[0060] The selection process is represented as follows:

[0061] ;

[0062] in, This represents the fitness value of the test vector of individual i in the population during the G-th iteration. This represents the fitness value of the target vector of individual i in the population during the G-th iteration. Let represent the temporary target vector of individual i in the population during the G-th iteration. After K-means clustering correction, the temporary target vector can be transformed into the target vector, updating the distribution of the cluster center positions that the individual considers to be the best at the moment, and continuing to participate in the next iteration.

[0063] Step 2.6, K-means clustering correction:

[0064] Temporary target vector As the current best cluster center position, calculate the Euclidean distance between all N drone nodes and each cluster center position, assign the drone nodes to the cluster center with the smallest distance, thereby determining the cluster affiliation of each node. Add the local optimization strategy of K-means clustering algorithm, perform a local correction on the center position given by each individual, calculate the average value of the coordinates of all drone nodes in each cluster, and assign it to the target vector of the next iteration. The target vector of the next iteration is used as the current best cluster center position decided by the DE-Kmeans algorithm in this round, thus completing one round of training.

[0065] , ;

[0066] in, This represents the coordinates of the c-th cluster center corresponding to the target vector of individual i in the (G+1)-th iteration. The values ​​of the 3c-2, 3c-1, and 3c dimensions; This represents the number of nodes in cluster c; Let i represent the UAV node coordinates of individual i belonging to the c-th cluster in the G-th iteration;

[0067] Step 2.7, iterate repeatedly:

[0068] Repeat steps 2.2 to 2.6 until the preset number of iterations is completed, then output the target vector of the individual with the best fitness in the current population as the optimal solution. :

[0069] .

[0070] Furthermore, in step three, the process of constructing the UAV swarm cluster network is as follows:

[0071] Based on the cluster centers obtained in step two, each UAV node selects a leader node according to its distance from the cluster center. The closer to the cluster center, the higher the priority of becoming the leader node. If, in a complex battlefield environment, the leader node of a UAV cluster loses its communication capability due to attack or interference, the next leader node for that cluster is determined according to the priority. Each leader node represents a subnet and is responsible for managing the information interaction within the entire subnet. All ordinary nodes automatically join the corresponding subnet. The subnet adopts a star network topology, and each ordinary node only establishes a connection with its leader node.

[0072] Step 3.1, Initialization:

[0073] The ground control station GS informs each UAV node of the cluster center of the cluster to which it belongs, as well as the number and number of other UAV nodes in the same cluster. An empty intra_rankI list is created for each UAV node I to store the distance between other nodes in the cluster and the cluster center.

[0074] Step 3.2, Sending and receiving cluster header declarations:

[0075] Each drone node I calculates its distance d(I) from the cluster center and sends the distance information d(I) to other nodes in the cluster; after receiving the distance information d(J) from other nodes J (J≠I), node I adds the distance information d(J) to intra_rankI;

[0076] Step 3.3, establish cluster head order:

[0077] The elements in intra_rankI are sorted in ascending order. If they are equal, the node with the smaller number is sorted first. After sorting, the node knows its own distance information d(I) from the cluster center and its sort order, which is the order in which it is a candidate to become the cluster head node.

[0078] Step 3.4, Intra-cluster networking request:

[0079] Node I, whose candidate cluster head node has a priority of 0, broadcasts the cluster head declaration CH(I,0) to all other UAV nodes in the same cluster, announcing that it has become the cluster head node with a priority of 0, and that the other nodes in the same cluster are member nodes. After receiving the cluster head declaration, the other UAV nodes confirm their own identity and apply to the cluster head node I to establish a connection. The cluster head node I sends the cluster networking completion information to the ground control station.

[0080] Step 3.5, Inter-cluster network initialization:

[0081] Once the ground control station receives the intra-cluster networking completion information for all 21 sub-clusters, it informs each cluster head node I of the numbers of other cluster head nodes, creates an empty CHI list and an inter_rankI list for each cluster head node I. The CHI list is used to store the peer nodes of the inter-cluster links to be connected, and the inter_rankI list is used to store the cluster head node numbers. The elements in the inter_rankI list are sorted in ascending order, and the sorting order of its own number is known, which is the order in which inter-cluster networking begins.

[0082] Step 3.6, Inter-cluster networking request:

[0083] Node I, with a cluster networking priority of 0, searches for connectable neighbor leader node J outwards from itself within the maximum communication radius, broadcasting an inter-cluster link establishment message to request a connection. If cluster leader node J receives the link establishment message from cluster leader node I, and its own connection count has not yet reached the maximum, it sends an agreement link establishment message to cluster leader node I to confirm the connection; otherwise, it sends a rejection link establishment message to deny the connection request. Similarly, if its own connection count has not yet reached the maximum, cluster leader node I can send an confirmation message to establish the connection after receiving the agreement link establishment message. If the link establishment is successful, both parties store each other's node numbers in the CH list. Otherwise, cluster leader node I sends an overflow message indicating that its own connection count has reached the limit, and the link establishment fails.

[0084] Step 3.7, sequential networking between clusters:

[0085] Once the current cluster head node I reaches the maximum number of connections, it stops broadcasting, waits for a period of time, and sends the inter-cluster networking completion information to the next node in the inter_rankI list. After receiving the inter-cluster networking completion information, the cluster head node J jumps to step 3.6 to continue execution until the last node in the inter_rankI list completes the link establishment and sends the inter-cluster networking completion information to the ground control station, thus completing the inter-cluster networking.

[0086] Step 3.8, return the result:

[0087] The intra-cluster and inter-cluster networking is completed, and the K-means differential evolution algorithm returns the topology connection results of the UAV cluster.

[0088] Compared with the prior art, the technical effects of the present invention are as follows:

[0089] This invention integrates the intelligent optimization capability of differential evolution algorithm in autonomously searching for optimal solutions with the efficient iterative clustering characteristics of K-means algorithm in clustering problems. By continuously executing mutation, crossover, selection, and correction iterations, the population alternates between random direction search and cluster center correction strategies during the exploration process. This approach not only enables a wide-ranging search for optimal solution locations through the automatic iterative mechanism of mutation, crossover, and selection, reducing the negative impact of numerous local optima in high-dimensional optimization objectives, but also integrates the advantages of clustering algorithms into the iterative process through mean correction, guiding the population to further reduce its fitness value, effectively improving the clustering quality and convergence speed of large-scale UAV swarms. Simultaneously, a hybrid topology structure of "intra-cluster star + inter-cluster mesh" is constructed based on the elected cluster leader node, providing a reliable communication topology foundation for UAV network construction. Attached Figure Description

[0090] Figure 1 This is a schematic diagram of the drone network of the present invention;

[0091] Figure 2 This is a schematic diagram of the K-means differential evolution algorithm of the present invention;

[0092] Figure 3 This is a simulation diagram of the clustering results of the K-means differential evolution algorithm of this invention;

[0093] Figure 4 This is a simulation diagram of network topology chain building based on the K-means differential evolution algorithm of the present invention. Detailed Implementation

[0094] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0095] A smart clustering algorithm for unmanned aerial vehicle (UAV) networks based on K-means clustering and differential evolution includes the following steps:

[0096] Step 1: Establish a model for the clustering topology of UAV nodes and the optimal clustering problem;

[0097] Step 2, construct the K-means differential evolution algorithm:

[0098] By integrating the K-means clustering algorithm into the differential evolution algorithm, the intelligent optimization and iteration capability of the differential evolution algorithm is combined with the clustering task-oriented performance of the K-means clustering algorithm to solve the optimal clustering problem.

[0099] Step 3: Construct a clustered network for the drone swarm:

[0100] Based on the cluster topology of UAV nodes, a cluster leader node election rule is designed, and connections are established between different cluster leader nodes based on the proximity principle to form a UAV swarm cluster network.

[0101] Furthermore, in step one, the process of establishing the optimal clustering problem model is as follows:

[0102] The nodes in a drone swarm network need to cooperate with each other to complete pre-planned tasks;

[0103] The communication is considered in two types: communication between UAVs and the ground, and communication between UAVs. The communication between UAVs and the ground includes air-to-ground (A2G) and ground-to-air (G2A); the communication between UAVs is air-to-air (A2A).

[0104] Furthermore, each UAV node has different capabilities and resources, and independent flight routes and missions; each UAV node is set to have no resource restrictions in the ground station; if a UAV node is within the communication range of the ground control station GS, the UAV nodes can communicate with each other; each UAV node has a GPS configured with a network time protocol system; the GPS provides the node's location and mobility to calculate the link expiration time between two nodes.

[0105] In three-dimensional space, assuming there are 105 randomly distributed nodes with horizontal coordinates ranging from [0, 100] to [2] and vertical coordinates ranging from [0, 50], the nodes need to be divided into 21 clusters according to the task. The goal is to determine the coordinates of the 21 cluster centers, i.e., each center has 3 coordinate parameters, for a total of 63 decision variables, so as to minimize the total distance from all nodes to their nearest cluster center.

[0106] The optimization problem model for clustering drone swarms is as follows:

[0107] Let the set of cluster centers be... Among them, cluster centers ;

[0108] The total decision variable vector P is:

[0109] ;

[0110] Where T represents the transpose; R represents the set of real numbers; x i y i z i These represent the coordinates in the X, Y, and Z directions, respectively.

[0111] The objective function f(p) minimizes the total distance from all nodes to their nearest cluster center, with the constraint that the XY coordinates of all cluster centers do not exceed the range [0, 100], and the Z coordinates do not exceed the range [0, 50].

[0112] ;

[0113] ;

[0114] ;

[0115] in, Let be the coordinates of the j-th node; Let i be the coordinates of the i-th cluster center; This represents Euclidean distance.

[0116] Furthermore, in step two, the K-means differential evolution algorithm is constructed as follows:

[0117] The K-means differential evolution algorithm incorporates the K-means clustering algorithm into the differential evolution intelligent algorithm. First, it initializes several exploration individuals, each representing a possible cluster center distribution of a drone swarm, i.e., a point in the solution space of the optimization problem. Then, it iteratively optimizes the individuals in the population by repeatedly performing mutation, crossover, and selection processes using the differential evolution algorithm. Based on the K-means clustering idea, it endows the population with the ability for structured fine-tuning, allowing individuals to automatically correct their exploration results towards the cluster centers, thus enabling the population to quickly approach the optimal solution. Figure 2 As shown;

[0118] Step 2.1, Population Initialization:

[0119] Initialize by assigning initial values ​​to each dimension of each individual in the population. Each individual x in the population... i,G It is expressed as follows:

[0120] ;

[0121] Where, x i,G In the G-th iteration of the population, the th iteration is... Individual, This represents the population size, i.e., NP=3S; Indicates the generation number;

[0122] Considering that the optimal cluster center of a sub-cluster must be inside the polyhedron formed by all member nodes of the subset, when initializing nodes, S nodes are directly selected as the initial cluster center positions according to the number of clusters, and random perturbation is added to give the initial exploration direction, rather than randomly selecting the initial population position, so as to help the population approach the ideal solution faster.

[0123] Step 2.2, Mutation:

[0124] For each individual goal x i,G After the population initialization is completed, for each iteration of the population, a mutation operation is first performed, randomly selecting two other individuals in the population to generate mutation vectors. Calculate using one of the following two formulas, each with a 50% probability:

[0125] ;

[0126] ;

[0127] in, , and Both are target vectors, representing the currently feasible solutions chosen by the individuals. The individual index "best" indicates the individual with the best current fitness in the entire population, i.e., the individual closest to the optimal solution. It is required that the randomly selected individuals r1, r2, and r3 are all distinct and also distinct from i. Therefore, the population size must contain at least four nodes. ; variable factor It is a real variable that controls the scaling of the deviation variable; This represents the value of the mutation factor for individual i in the Gth iteration;

[0128] An adaptive adjustment strategy is used to change the mutation factor during the iteration process. Value:

[0129] ;

[0130] in, and and represent the minimum and maximum mutation factor values ​​among all individuals in the population at the current iteration round, respectively; the fitness f is determined by the optimization objective, the smaller f is, the better the individual's fitness and the closer it is to the optimal solution; and These represent the minimum and maximum fitness values ​​among all individuals in the population at the current iteration round, respectively. This represents the fitness value of the current individual in the current iteration round. As a regulating factor;

[0131] For individuals with low fitness, their f is larger, resulting in a larger adaptive mutation factor F, which encourages them to try mutations; for individuals with high fitness, their f is smaller, resulting in a smaller mutation factor F, which inhibits their mutations. This helps the population find a balance between exploration and utilization, fully leveraging the overall exploration capabilities while retaining currently superior individuals to continue participating in evolution.

[0132] Step 2.3, Cross:

[0133] The target vector reflects the feasible solution currently chosen by the individual, while the mutation vector reflects the potential direction of change of the individual; the crossover operation combines the target vector and the mutation vector, replacing the values ​​of several dimensions in the target vector with the mutated values ​​to obtain the test vector, thereby realizing the iteration of the individual;

[0134] The target vector and the mutated vector are combined, and the values ​​of several dimensions in the target vector are replaced with the mutated values ​​to obtain the test vector; the test vector :

[0135] ;

[0136] The above formula, ;

[0137] Where D represents the individual dimension, and Let x represent two random sequences with values ​​in the ranges [0,1] and [1,D] respectively in the G-th iteration; ji,G+1 v represents the target vector value of dimension j in the (G+1)th iteration; ji,G+1 Represents the mutation vector value of dimension j in the (G+1)th iteration, u ji,G+1 This represents the test vector value of the j-th dimension in the (G+1)-th iteration;

[0138] The above formula means: test vector The value is either equal to the mutation vector. It is either equal to the target vector. Specifically, based on the crossover probability Random probability and Determined; Test Vector It equals the mutation vector in the following two cases. Scenario 1: The first The test vector of dimension 1 is assigned the corresponding value of the mutated vector; Case 2, if Less than Then The value is assigned to the corresponding value of the mutation vector for that dimension. In other cases, the test vector... equal to the target vector This ensures that while the target vector mutates according to the crossover probability, at least one dimension value will mutate; otherwise, the individual will lose the meaning of this iteration.

[0139] Crossover probability Let represent the crossover probability of the i-th individual in the G-th iteration, with a value range of [0,1], controlling the probability that the mutated vector is selected as the test vector; an adaptive adjustment strategy is used to change the crossover probability during the iteration process. Value:

[0140] ;

[0141] in, and represents the minimum and maximum mutation factor values ​​among all individuals in the population, respectively; the fitness f is determined by the optimization objective, the smaller f is, the better the individual's fitness and the closer it is to the optimal solution; and These represent the minimum and maximum fitness values ​​among all individuals in the population, respectively. This represents the fitness value of the current individual in the Gth iteration. As a regulating factor;

[0142] For individuals with low fitness, a larger f value results in a smaller crossover probability CR value, reducing the probability of the mutated vector being selected. For individuals with high fitness, a smaller f value results in a larger CR value, increasing the probability of the mutated vector being selected. Like the adaptive mutation factor F, the adaptive crossover probability CR also helps the population find a balance between exploration and utilization, fully leveraging the overall exploration capability while retaining currently superior individuals to continue participating in evolution, which conforms to the "survival of the fittest" logic of collective intelligent search.

[0143] Step 2.4, Boundary Condition Handling:

[0144] Step 2.5, select:

[0145] Step 2.6, K-means clustering correction:

[0146] Step 2.7, iterate repeatedly:

[0147] Repeat steps 2.2 to 2.6 until the preset number of iterations is completed, then output the target vector of the individual with the best fitness in the current population as the optimal solution. :

[0148] ;

[0149] In three-dimensional space, given 105 randomly distributed nodes, the horizontal coordinates range from [0, 100]. 2 The vertical coordinates are [0, 50]. The goal is to determine the coordinates of 21 cluster centers, i.e., each center has 3 coordinate parameters, for a total of 63 decision variables, so as to minimize the total distance from all nodes to their nearest cluster center.

[0150] The drone cluster size was set to 105 drones. The horizontal coordinates of the drone nodes in the XY direction were all within the range of [0, 100], and the vertical coordinates in the Z direction were within the range of [0, 50]. Randomly generated drone node coordinate information was read and written in file format. These nodes were divided into 21 clusters according to the task requirements. The initial population size was set to 500, the maximum number of algorithm iterations was set to 5000, and the mutation factor F and crossover probability CR were both adaptively adjusted, with upper and lower limits of [0.4, 0.9] and [0.1, 0.9], respectively. The optimization results obtained by using the K-means clustering and differential evolution fusion algorithm proposed in this invention, after multiple iterations and final position migration, can effectively reduce the fitness value of the optimal individual in the population, and quickly and effectively obtain relatively ideal clustering results, such as... Figure 3 As shown in the figure, the 21 pentagrams represent the cluster centers of 21 drone clusters.

[0151] Furthermore, in step three, the process of constructing the UAV swarm cluster network is as follows:

[0152] Based on the cluster centers obtained in step two, each UAV node selects a leader node according to its distance from the cluster center. The closer to the cluster center, the higher the priority of becoming the leader node. If, in a complex battlefield environment, the leader node of a UAV cluster loses its communication capability due to attack or interference, the next leader node for that cluster is determined according to the priority. Each leader node represents a subnet and is responsible for managing the information interaction within the entire subnet. All ordinary nodes automatically join the corresponding subnet. The subnet adopts a star network topology, and each ordinary node only establishes a connection with its leader node.

[0153] Inter-cluster communication between UAV clusters relies on the leader node of each cluster to establish link connections. First, based on the randomly preset cluster number, starting from the leader node of the cluster with the smallest number, the search for connectable neighbor nodes expands outward within the communication radius, centered on itself. Once a still connectable neighbor node is detected, a HELLO message is sent to it to request the establishment of a connection. After receiving the message, if the neighbor node has not yet reached its maximum connection limit, it immediately returns a HELLO message to confirm that the connection is effective. The search stops when the leader node in the search reaches its maximum connection limit. The next leader node with the corresponding number then begins the search, until all leader nodes have been searched. At this point, a snapshot topology of the UAV cluster network for the current time slot is completed.

[0154] Step 3.1, Initialization:

[0155] The ground control station GS informs each UAV node of the cluster center of the cluster to which it belongs, as well as the number and number of other UAV nodes in the same cluster. An empty intra_rankI list is created for each UAV node I to store the distance between other nodes in the cluster and the cluster center.

[0156] Step 3.2, Sending and receiving cluster header declarations:

[0157] Each drone node I calculates its distance d(I) from the cluster center and sends the distance information d(I) to other nodes in the cluster; after receiving the distance information d(J) from other nodes J (J≠I), node I adds the distance information d(J) to intra_rankI;

[0158] Step 3.3, establish cluster head order:

[0159] The elements in intra_rankI are sorted in ascending order. If they are equal, the node with the smaller number is sorted first. After sorting, the node knows its own distance information d(I) from the cluster center and its sort order, which is the order in which it is a candidate to become the cluster head node.

[0160] Step 3.4, Intra-cluster networking request:

[0161] Node I, whose candidate cluster head node has a priority of 0, broadcasts the cluster head declaration CH(I,0) to all other UAV nodes in the same cluster, announcing that it has become the cluster head node with a priority of 0, and that the other nodes in the same cluster are member nodes. After receiving the cluster head declaration, the other UAV nodes confirm their own identity and apply to the cluster head node I to establish a connection. The cluster head node I sends the cluster networking completion information to the ground control station.

[0162] Step 3.5, Inter-cluster network initialization:

[0163] Once the ground control station receives the intra-cluster networking completion information for all 21 sub-clusters, it informs each cluster head node I of the numbers of other cluster head nodes, creates an empty CHI list and an inter_rankI list for each cluster head node I. The CHI list is used to store the peer nodes of the inter-cluster links to be connected, and the inter_rankI list is used to store the cluster head node numbers. The elements in the inter_rankI list are sorted in ascending order, and the sorting order of its own number is known, which is the order in which inter-cluster networking begins.

[0164] Step 3.6, Inter-cluster networking request:

[0165] Node I, with a cluster networking priority of 0, searches for connectable neighbor leader node J outwards from itself within the maximum communication radius, broadcasting an inter-cluster link establishment message to request a connection. If cluster leader node J receives the link establishment message from cluster leader node I, and its own connection count has not yet reached the maximum, it sends an agreement link establishment message to cluster leader node I to confirm the connection; otherwise, it sends a rejection link establishment message to deny the connection request. Similarly, if its own connection count has not yet reached the maximum, cluster leader node I can send an confirmation message to establish the connection after receiving the agreement link establishment message. If the link establishment is successful, both parties store each other's node numbers in the CH list. Otherwise, cluster leader node I sends an overflow message indicating that its own connection count has reached the limit, and the link establishment fails.

[0166] Step 3.7, sequential networking between clusters:

[0167] Once the current cluster head node I reaches the maximum number of connections, it stops broadcasting, waits for a period of time, and sends the inter-cluster networking completion information to the next node in the inter_rankI list. After receiving the inter-cluster networking completion information, the cluster head node J jumps to step 3.6 to continue execution until the last node in the inter_rankI list completes the link establishment and sends the inter-cluster networking completion information to the ground control station, thus completing the inter-cluster networking.

[0168] Step 3.8, return the result:

[0169] The intra-cluster and inter-cluster networking is completed, and the K-means differential evolution algorithm returns the topology connection results of the UAV cluster.

[0170] Simulation results of network topology building based on K-means differential evolution algorithm are as follows: Figure 4 As shown, the simulation assumes that, under the premise of satisfying the maximum number of connections not exceeding 4 and the maximum communication radius, the order of the cluster head node is determined according to the distance between each cluster member and the cluster center from near to far, and the network topology connection relationship between the cluster head nodes is constructed. Within each cluster, a star topology is used to connect the member nodes and the cluster head node.

[0171] As UAV missions progress, the relative positions and communication conditions between nodes change, and link interruptions and recovery occur frequently. Therefore, it is necessary to construct a dynamically changing network topology to enable UAV nodes to adaptively network. Every so often, all UAV nodes will re-exchange information with their neighbors, run clustering algorithms to re-plan the topology, construct a new network topology map, and complete routing and forwarding tasks based on the new network topology. When the UAV swarm is attacked, resulting in damage to some nodes, or when network topology failures occur due to changes in the geographical environment or signal interference, each UAV node only needs to re-execute the clustering algorithm to rebuild the network topology. Considering the high dependence of strategic and tactical execution on UAV swarm formation in the battlefield environment, the spatial position information of UAV nodes corresponding to different formations for different tactical missions and external environments is constructed in advance, and the topology planning is executed in advance and the results are saved so as to quickly realize intelligent clustering and topology reconstruction of UAVs in complex and ever-changing battlefield environments.

Claims

1. A smart clustering algorithm for unmanned aerial vehicle (UAV) networks based on K-means clustering and differential evolution, characterized in that, The steps include the following: Step 1: Establish a model for the clustering topology of UAV nodes and the optimal clustering problem; Step 2, construct the K-means differential evolution algorithm: By integrating the K-means clustering algorithm into the differential evolution algorithm, the intelligent optimization and iteration capability of the differential evolution algorithm is combined with the clustering task-oriented performance of the K-means clustering algorithm to solve the optimal clustering problem. Step 3: Construct a clustered network for the drone swarm: Based on the cluster topology of UAV nodes, a cluster leader node election rule is designed, and connections are established between different cluster leader nodes based on the proximity principle to form a UAV swarm cluster network.

2. The intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution as described in claim 1, characterized in that, In step one, the establishment of the UAV node cluster topology is as follows: Unmanned aerial vehicle network recording<GS,NU> ; Where GS represents the ground control station and NU represents the set of all nodes in the UAV swarm. Under the action of the clustering algorithm, the nodes in the entire UAV swarm network will be divided into several clusters. For ease of management, the UAV nodes in each cluster are further divided into two categories: leader nodes CH and ordinary nodes CM. The responsibilities of the ground control station (GS): As the command center of the entire UAV swarm network, the ground control station controls the flight process, flight route, payload mission functions, and normal operation of communication links of each UAV node. The responsibilities of a regular CM node: Each UAV regular node flies in its own flight mode and performs its predetermined tasks; in addition, UAV nodes cooperate with each other to form a self-organizing network. The responsibilities of the leader node (CH): In addition to the functions of a regular node, the drone leader node also manages the nodes within its corresponding cluster.

3. The intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution as described in claim 1, characterized in that, In step one, the process of establishing the optimal clustering problem model is as follows: In three-dimensional space, given N randomly distributed nodes, the horizontal plane coordinate range is... and Vertical coordinate range ,in , , These represent the maximum values ​​of the coordinate ranges in the X, Y, and Z directions, respectively; the nodes are divided into S clusters according to task requirements; the goal is to determine the coordinates of the S cluster centers, i.e., each center has 3 coordinate parameters, for a total of... A set of decision variables is used to minimize the total distance from all nodes to their nearest cluster center. The optimization problem model for clustering drone swarms is as follows: Let the set of cluster centers be... Among them, cluster centers ; The total decision variable vector P is: ； Where T represents the transpose; R represents the set of real numbers; x i y i z i These represent the coordinates in the X, Y, and Z directions, respectively. Minimizing the objective function f(p) yields the optimization problem model minf(p): ； ； in, Let be the coordinates of the k-th node; Let i be the coordinates of the i-th cluster center; This represents Euclidean distance.

4. The intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution according to claim 1, characterized in that, In step two, the K-means differential evolution algorithm is constructed as follows: Step 2.1, Population Initialization: Initialize by assigning initial values ​​to each dimension of each individual in the population. Each individual x in the population... i,G It is expressed as follows: ； Where, x i,G In the G-th iteration of the population, the th iteration is... Individual, This represents the population size, i.e., NP=3S; Indicates the generation number; Based on the number of clusters, S nodes are directly selected as the initial cluster centers, and random perturbations are added to provide the initial exploration direction; Step 2.2, Mutation: For each individual objective, after population initialization, for each iteration of the population, a mutation operation is first performed, randomly selecting two other individuals in the population to generate mutation vectors. Calculate using one of the following two formulas, each with a 50% probability: ； ； in, , and Both are target vectors, representing the currently feasible solutions chosen by the individuals. The individual index "best" indicates the individual with the best current fitness in the entire population, i.e., the individual closest to the optimal solution. It is required that the randomly selected individuals r1, r2, and r3 are all distinct and also distinct from i. Therefore, the population size must contain at least four nodes. ; variable factor It is a real variable that controls the scaling of the deviation variable; This represents the value of the mutation factor for individual i in the Gth iteration; An adaptive adjustment strategy is used to change the mutation factor during the iteration process. Value: ； in, and and represent the minimum and maximum mutation factor values ​​among all individuals in the population at the current iteration round, respectively; the fitness f is determined by the optimization objective, the smaller f is, the better the individual's fitness and the closer it is to the optimal solution; and These represent the minimum and maximum fitness values ​​among all individuals in the population at the current iteration round, respectively. This represents the fitness value of the current individual in the current iteration round. As a regulating factor; Step 2.3, Cross: The target vector and the mutated vector are combined, and the values ​​of several dimensions in the target vector are replaced with the mutated values ​​to obtain the test vector; the test vector : ； The above formula, ; Where D represents the individual dimension, and Let x represent two random sequences with values ​​in the ranges [0,1] and [1,D] respectively in the G-th iteration; ji,G+1 v represents the target vector value of dimension j in the (G+1)th iteration; ji,G+1 Represents the mutation vector value of dimension j in the (G+1)th iteration, u ji,G+1 This represents the test vector value of the j-th dimension in the (G+1)-th iteration; Crossover probability Let represent the crossover probability of the i-th individual in the G-th iteration, with a value range of [0,1], controlling the probability that the mutated vector is selected as the test vector; an adaptive adjustment strategy is used to change the crossover probability during the iteration process. Value: ； in, and represents the minimum and maximum mutation factor values ​​among all individuals in the population, respectively; the fitness f is determined by the optimization objective, the smaller f is, the better the individual's fitness and the closer it is to the optimal solution; and These represent the minimum and maximum fitness values ​​among all individuals in the population, respectively. This represents the fitness value of the current individual in the Gth iteration. As a regulating factor; Step 2.4, Boundary Condition Handling: If a solution outside the feasible region appears during the mutation process, i.e. or Therefore, the test vectors must be restricted to the feasible region, and the corrected test vectors can be calculated using the following formula. : ； Where, x j (L) represents the lower boundary of the j-th dimension, U j (L) represents the upper boundary of the j-th dimension; Step 2.5, select: The K-means differential evolution algorithm uses a greedy criterion to process the test vectors... Compared with the target vector in the current population The next generation of target vectors is obtained through comparison and iteration. If the target vector has a lower fitness value and is closer to the optimal solution, the target vector is selected; otherwise, the test vector is selected. The test vector is compared only with the target vector, not with all individuals in the existing population. The selection process is represented as follows: ； in, This represents the fitness value of the test vector of individual i in the population during the G-th iteration. This represents the fitness value of the target vector of individual i in the G-th iteration of the population; Let represent the temporary target vector of individual i in the population during the G-th iteration. After K-means clustering correction, the temporary target vector can be transformed into the target vector, updating the distribution of the cluster center positions that the individual considers to be the best at the moment, and continuing to participate in the next iteration. Step 2.6, K-means clustering correction: Temporary target vector As the current best cluster center position, calculate the Euclidean distance between all N drone nodes and each cluster center position, assign the drone nodes to the cluster center with the smallest distance, thereby determining the cluster affiliation of each node. Add the local optimization strategy of K-means clustering algorithm, perform a local correction on the center position given by each individual, calculate the average value of the coordinates of all drone nodes in each cluster, and assign it to the target vector of the next iteration. The target vector of the next iteration is used as the current best cluster center position decided by the DE-Kmeans algorithm in this round, thus completing one round of training. , ； in, This represents the coordinates of the c-th cluster center corresponding to the target vector of individual i in the (G+1)-th iteration; This represents the number of nodes in cluster c; Let i represent the UAV node coordinates of individual i belonging to the c-th cluster in the G-th iteration; Step 2.7, iterate repeatedly: Repeat steps 2.2 to 2.6 until the preset number of iterations is completed, then output the target vector of the individual with the best fitness in the current population as the optimal solution. : 。 5. The intelligent clustering algorithm for UAV networks based on K-means clustering and differential evolution according to claim 1, characterized in that, In step three, the process of constructing the UAV swarm cluster network is as follows: Step 3.1, Initialization: The ground control station GS informs each UAV node of the cluster center of the cluster to which it belongs, as well as the number and number of other UAV nodes in the same cluster. An empty intra_rankI list is created for each UAV node I to store the distance between other nodes in the cluster and the cluster center. Step 3.2, Sending and receiving cluster header declarations: Each drone node I calculates its distance d(I) from the cluster center and sends the distance information d(I) to other nodes in the cluster; after receiving the distance information d(J) from other nodes J (J≠I), node I adds the distance information d(J) to intra_rankI; Step 3.3, establish cluster head order: The elements in intra_rankI are sorted in ascending order. If they are equal, the node with the smaller number is sorted first. After sorting, the node knows its own distance information d(I) from the cluster center and its sort order, which is the order in which it is a candidate to become the cluster head node. Step 3.4, Intra-cluster networking request: Node I, whose candidate cluster head node has a priority of 0, broadcasts the cluster head declaration CH(I,0) to all other UAV nodes in the same cluster, announcing that it has become the cluster head node with a priority of 0, and that the other nodes in the same cluster are member nodes. After receiving the cluster head declaration, the other UAV nodes confirm their own identity and apply to the cluster head node I to establish a connection. The cluster head node I sends the cluster networking completion information to the ground control station. Step 3.5, Inter-cluster network initialization: Once the ground control station receives the intra-cluster networking completion information for all 21 sub-clusters, it informs each cluster head node I of the numbers of other cluster head nodes, creates an empty CHI list and an inter_rankI list for each cluster head node I. The CHI list is used to store the peer nodes of the inter-cluster links to be connected, and the inter_rankI list is used to store the cluster head node numbers. The elements in the inter_rankI list are sorted in ascending order, and the sorting order of its own number is known, which is the order in which inter-cluster networking begins. Step 3.6, Inter-cluster networking request: Node I, with a cluster networking priority of 0, searches for connectable neighbor leader node J outwards from itself within the maximum communication radius, broadcasting an inter-cluster link establishment message to request a connection. If cluster leader node J receives the link establishment message from cluster leader node I, and its own connection count has not yet reached the maximum, it sends an agreement link establishment message to cluster leader node I to confirm the connection; otherwise, it sends a rejection link establishment message to deny the connection request. Similarly, if its own connection count has not yet reached the maximum, cluster leader node I can send an confirmation message to establish the connection after receiving the agreement link establishment message. If the link establishment is successful, both parties store each other's node numbers in the CH list. Otherwise, cluster leader node I sends an overflow message indicating that its own connection count has reached the limit, and the link establishment fails. Step 3.7, sequential networking between clusters: Once the current cluster head node I reaches the maximum number of connections, it stops broadcasting, waits for a period of time, and sends the inter-cluster networking completion information to the next node in the inter_rankI list. After receiving the inter-cluster networking completion information, the cluster head node J jumps to step 3.6 to continue execution until the last node in the inter_rankI list completes the link establishment and sends the inter-cluster networking completion information to the ground control station, thus completing the inter-cluster networking. Step 3.8, Results: Intra-cluster and inter-cluster networking are completed, and the K-means differential evolution algorithm returns the UAV cluster topology connection results.

6. An electronic device, characterized in that, include: One or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 5.

Images