Low-altitude unmanned aerial vehicle cluster modeling method

Abstract

The invention discloses a low-altitude unmanned aerial vehicle cluster modeling method, and belongs to the technical field of unmanned aerial vehicle cluster control. The method comprises the following steps: each unmanned aerial vehicle measures link quality indexes of the unmanned aerial vehicle and other unmanned aerial vehicles in a cluster in real time to form a communication quality vector; collecting all communication quality vectors to form a dynamic communication quality matrix; the dynamic communication quality matrix is used as input, a geodesic distance matrix is calculated, centralization processing and characteristic decomposition are carried out, and low-dimensional embedded coordinates of each unmanned aerial vehicle in the communication manifold space are obtained; and fusing the physical space coordinates and the communication manifold coordinates to form a cluster dynamic model. According to the method, a communication quality vector is regarded as a sampling point in a high-dimensional communication manifold space, and low-dimensional embedded representation is extracted in real time, so that unified dynamic modeling of a physical space pose and a communication topological structure is realized, the asymmetry and space-time non-stationary characteristics of a communication link in a low-altitude environment can be accurately depicted, and the communication efficiency is improved. And reliable underlying model support is provided for a cluster upper-layer algorithm.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) swarm control technology, and in particular to a method for modeling low-altitude UAV swarms. Background Technology

[0002] In missions such as collaborative reconnaissance, logistics delivery, and emergency rescue, low-altitude UAV swarms require accurate swarm models to support higher-level algorithms such as formation control, task allocation, and information fusion. Existing swarm modeling methods mainly include continuous models based on fluid dynamics, regional models based on Voronoi diagrams, and topological models based on potential field methods. These methods play an important role in describing the macroscopic behavior and spatial distribution of swarms.

[0003] However, the low-altitude flight environment is significantly complex. Factors such as wind disturbance, building obstruction, multipath effects, and electromagnetic interference cause the communication link quality between UAVs to exhibit strong asymmetric, non-stationary, and anisotropic characteristics. Existing modeling methods typically assume a single functional relationship between communication link quality and the physical distance between UAVs. This simplistic assumption fails to accurately reflect the actual characteristics of communication links in the complex low-altitude environment, leading to insufficient accuracy in predicting the state of swarm communication links. This frequently results in misclassifying actually broken links as available, causing problems such as lost swarm control commands or retransmission storms.

[0004] The core flaw of existing technologies lies in the lack of a dynamic modeling method that can uniformly map physical spatial pose and communication connection quality to the same feature space, making it difficult for upper-level algorithms such as formation maintenance, cooperative perception, and distributed task allocation to obtain reliable underlying model support. Summary of the Invention

[0005] The main objective of this invention is to provide a method for modeling low-altitude unmanned aerial vehicle (UAV) swarms to address the problems raised in related technologies.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for modeling low-altitude unmanned aerial vehicle (UAV) swarms is provided, comprising the following steps: Communication quality vector acquisition: Each drone in the cluster measures its link quality indicators with other drones in the cluster in real time, which constitutes the communication quality vector of that drone. Dynamic communication quality matrix construction: The communication quality vectors of all UAVs are collected to form a dynamic communication quality matrix; Online communication manifold learning: Using the dynamic communication quality matrix as input, calculate the geodesic distance matrix between UAVs, perform eigenvalue decomposition after centering the geodesic distance matrix, and obtain the low-dimensional embedding coordinates of each UAV in the communication manifold space; Joint embedding modeling: The physical space coordinates of each UAV are fused with the communication manifold coordinates to form a cluster dynamic model that represents the physical space relationship and communication topology of the cluster.

[0007] Furthermore, the link quality indicator is a comprehensive link quality indicator, and its calculation formula is as follows: in, This represents the overall link quality index from UAV i to UAV j at time t. To normalize the received signal strength indication value, For packet error rate, For normalized signal-to-noise ratio, This is a normalized link capacity estimate. , , , The weighted coefficients are and satisfy the following conditions: .

[0008] Furthermore, the geodesic distance matrix between the calculated UAVs includes: The short-circuit distance between drones is determined based on the link quality index. When the link quality index is greater than or equal to a preset threshold, the short-circuit distance is equal to 1 minus the link quality index; when the link quality index is less than the preset threshold, the short-circuit distance is set to infinity. A weighted graph is constructed based on the short-circuit distance, and the shortest path algorithm is used to calculate the shortest path distance between all UAV pairs to form a geodesic distance matrix.

[0009] Furthermore, the preset threshold is set to 0.3, and the shortest path algorithm adopts either the Floyd-Warshall algorithm or the Dijkstra algorithm.

[0010] Furthermore, the formula for centering the geodesic distance matrix is ​​as follows: in, The centered matrix, For the geodesic distance matrix, This indicates that the square operation is performed on each element of the geodesic distance matrix. For a centered matrix, It is an N-order identity matrix. It is an N-dimensional column vector with all elements being 1, where N is the number of drones.

[0011] Furthermore, the execution feature decomposition includes: Perform eigenvalue decomposition on the centered matrix and arrange the eigenvalues ​​in descending order; Select the top d largest positive eigenvalues ​​and their corresponding eigenvectors, and calculate the low-dimensional embedding coordinates of each UAV based on the eigenvalues ​​and eigenvectors, where the embedding dimension d ranges from 4 to 6.

[0012] Furthermore, the online communication manifold learning step also includes incremental online updates, specifically: Maintain the eigenvector matrix and eigenvalue diagonal matrix of the previous time step as prior information; When the dynamic communication quality matrix of the new moment arrives, calculate the change of the centralized matrix relative to the previous moment; The subspace tracking algorithm is used to incrementally update the feature vector and feature value to obtain the low-dimensional embedding coordinates at the current time.

[0013] Furthermore, in the joint embedding modeling step, the formula for calculating the joint embedding feature vector is: in, Let i be the joint embedding feature vector of UAV i at time t. Let i be the physical space coordinates of the drone. Let i be the communication manifold coordinates of UAV i. and These are the weighting coefficients.

[0014] Furthermore, the physical space coordinates are acquired in real time via an airborne RTK-GNSS module or a UWB positioning system.

[0015] Furthermore, it also includes a model application step: based on the cluster dynamic model, performing formation maintenance, leader election, or distributed information fusion tasks in the communication manifold space or joint embedding space.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a high-dimensional vector representing the communication quality between each UAV and other UAVs in the cluster, and treats this vector as a sampling point of the UAV in the high-dimensional communication manifold space. It then uses the online dynamic Isomap manifold learning algorithm to extract its inherent low-dimensional structure in real time, thereby obtaining the low-dimensional embedded coordinates of each UAV in the communication manifold space. This achieves unified dynamic modeling of physical space pose and communication topology.

[0017] This method can accurately characterize the asymmetry and spatiotemporal nonstationarity of communication links in low-altitude environments. The accuracy of trunk communication topology modeling can reach over 95%. The false alarm rate of link interruption prediction is reduced by 87% compared with existing technologies, and it can predict the drastic deterioration of link quality 1.8 to 3.2 seconds in advance.

[0018] By employing an incremental feature tracking and update algorithm, the complete feature decomposition operation is avoided for each frame, and the embedded coordinate update frequency can reach 50Hz, meeting the real-time requirements of low-altitude clusters. The distributed information fusion accuracy based on this model is improved by 41.7% compared to the traditional distance-weighted method, providing reliable underlying model support for upper-layer cluster algorithms. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the low-altitude unmanned aerial vehicle (UAV) swarm modeling method provided in an embodiment of the present invention; Figure 2 A detailed flowchart illustrating the online communication manifold learning steps provided in this embodiment of the invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] Low-altitude unmanned aerial vehicle (UAV) swarms face complex and ever-changing communication environments when performing collaborative tasks. Factors such as wind disturbance, building obstruction, multipath effects, and electromagnetic interference cause the communication link quality between UAVs to exhibit significant asymmetry, non-stationarity, and anisotropy. Traditional modeling methods typically assume that link quality has a single functional relationship with physical distance. This simplistic assumption often leads to inaccurate link state predictions in the complex low-altitude environment, thereby affecting the reliable transmission of swarm control commands.

[0022] To address the aforementioned problems, this invention proposes a low-altitude UAV swarm modeling method based on dynamic communication manifold learning. The core idea of ​​this method is to construct a high-dimensional vector representing the communication quality between each UAV and other UAVs in the swarm, and to treat this vector as a sampling point of the UAV in the high-dimensional communication manifold space. Through an online manifold learning algorithm, the intrinsic low-dimensional structure of this high-dimensional space is extracted in real time, thereby obtaining the low-dimensional embedded coordinates of each UAV in the communication manifold space. This low-dimensional embedded representation not only preserves the essential characteristics of the swarm communication topology but also can be fused with physical space coordinates to form a unified dynamic swarm model.

[0023] See Figure 1 The low-altitude unmanned aerial vehicle (UAV) swarm modeling method provided by this invention includes the following steps: S1: Real-time acquisition of communication quality vector The goal of this step is to acquire real-time communication link quality data between each drone and other drones in the cluster, providing raw input for subsequent manifold learning.

[0024] Suppose a low-altitude unmanned aerial vehicle (UAV) swarm consists of N UAVs, numbered 1, 2, ..., N. At any time t, each UAV i measures the link quality metrics between itself and the other N-1 UAVs in real time via its onboard communication module. These metrics, after processing, constitute the original communication quality vector of UAV i: Among them, vector elements This represents the overall link quality index from UAV i to UAV j. This index is a weighted fusion of multiple underlying measurement parameters, and the specific calculation formula is as follows: The meanings of the parameters in the above formula are as follows: The normalized received signal strength indicator value reflects the power intensity of the signal received by UAV i from UAV j. Packet error rate (PRR) represents the proportion of data packets that are corrupted during transmission. Used to characterize the transmission reliability of the link; Normalized signal-to-noise ratio (SNR) measures the relative intensity of the useful signal to the background noise. The normalized link capacity estimate reflects the theoretical maximum transmission rate of the link. The weighting coefficients of each indicator satisfy... Its specific value can be learned and optimized through offline data-driven methods.

[0025] After normalization A higher value indicates better link quality. It's worth noting that due to asymmetric effects such as building obstruction and multipath propagation in low-altitude environments, the link quality from UAV i to UAV j is affected. Link quality from drone j to drone i They are usually not equal, a characteristic that naturally reflects the asymmetry of low-altitude communication links.

[0026] S2: Constructing a dynamic communication quality matrix This step aggregates the communication quality vectors distributed across each UAV to form a global matrix describing the overall communication status of the cluster.

[0027] Each drone in the cluster will use the communication quality vector acquired in step S1. The communication quality vectors are broadcast to other nodes via inter-drone data links. The ground control station or designated leader node receives and aggregates the communication quality vectors reported by all UAVs in real time, thereby constructing a dynamic communication quality matrix. In certain application scenarios, to facilitate subsequent calculations, it can be Extended to symmetric form Its diagonal elements are set to 1 (indicating that the "link quality" between the drone and itself is optimal), and the off-diagonal elements are... Desirable and The mean or minimum value, the specific strategy is determined according to the application requirements.

[0028] The dynamic communication quality matrix It comprehensively describes the communication state of the cluster at time t, including asymmetric information on link quality, providing a complete data foundation for subsequent manifold learning.

[0029] S3: Online Communication Manifold Learning This step is the core technical aspect of this invention. By performing manifold learning on the high-dimensional communication quality vector, its inherent low-dimensional geometric structure is extracted, and the embedding coordinates of each UAV in the communication manifold space are obtained.

[0030] See Figure 2 This step is implemented using the online dynamic Isomap manifold learning algorithm, and specifically includes the following sub-steps: S3.1: Calculation of Geodesic Distance Matrix With dynamic communication quality matrix As input, a graph structure based on communication quality between UAVs is first constructed, and the geodesic distance matrix is ​​calculated. .

[0031] For any two drones i and j, based on their link quality Determine the short-circuit distance : in, The threshold value for link validity is taken as [value] in this embodiment. .when At that time, it is assumed that there is a valid communication link between UAV i and UAV j, and its short-circuit distance is set as The better the link quality, the smaller the short-circuit distance; when If the link is deemed unusable, the short-circuit distance is set to infinity.

[0032] Based on the weighted graph constructed using the aforementioned short-circuit distances, the Floyd-Warshall algorithm or Dijkstra's algorithm is employed to calculate the shortest path distance between all UAV pairs. The shortest path distance reflects the geodesic distance between two UAVs on the communication manifold; even if there is no direct effective link between two UAVs, a geodesic connection can be established through a relay node. The geodesic distances of all UAV pairs constitute a geodesic distance matrix. .

[0033] S3.2: Centralized processing For the geodesic distance matrix By performing a dual-centering process to eliminate the translation components in the distance matrix, the core matrix is ​​obtained. : in, Indicates to Squaring each element. The centered matrix is ​​defined as follows: In the formula, It is an N-order identity matrix. This is an N-dimensional column vector with all elements equal to 1. The purpose of centering is to transform the data into a coordinate system with the centroid as the origin, which facilitates subsequent feature decomposition.

[0034] S3.3: Eigenvalue Decomposition and Low-Dimensional Embedding For the centered matrix Perform eigenvalue decomposition to obtain low-dimensional embedding coordinates. Let... The eigenvalues ​​are arranged in descending order as follows: The corresponding feature vectors are respectively .

[0035] Select the first d largest positive eigenvalues ​​and their corresponding eigenvectors to construct a low-dimensional embedding coordinate matrix: Wherein, the communication manifold coordinates of the i-th UAV at time t Determined by the following formula: In the formula, Represents the k-th eigenvector The i-th component. The selection of the embedding dimension d needs to strike a balance between information retention and computational efficiency. This invention preferably... This dimensional range effectively captures the main geometric features of the communication manifold while maintaining low computational complexity.

[0036] S3.4: Incremental Online Update Mechanism To meet the stringent real-time requirements of low-altitude UAV swarms and avoid performing a complete feature decomposition operation in every sampling period, this invention designs an incremental feature tracking and update algorithm.

[0037] The basic idea of ​​this algorithm is to maintain the eigenvector matrix of the previous time step t-1. and eigenvalue diagonal matrix As prior information, the communication quality matrix at the new time t. Upon arrival, first calculate the change in the centralized matrix. Because the changes in communication quality between adjacent moments are usually relatively gradual, They often exhibit low-rank characteristics.

[0038] Leveraging this characteristic, this invention employs a subspace tracking algorithm based on projection approximation (such as the FastBlock-BasedSubspaceTracking algorithm) to incrementally update the feature vector and eigenvalues. Specifically, it will... Projecting onto the subspace spanned by the current eigenvectors and its orthogonal complement space, the updated eigenvectors are obtained by solving a low-dimensional optimization problem. and This leads to the low-dimensional embedding coordinates at the new moment. .

[0039] The computational complexity of this incremental update algorithm is O(n). Compared to complete eigenvalue decomposition The complexity is significantly reduced. Real-world testing shows that in a swarm of 50 drones, the embedded coordinate update frequency can stably reach 50Hz, fully meeting the real-time modeling requirements of low-altitude swarms.

[0040] S4: Physical-Communication Joint Embedding Modeling After manifold learning in step S3, each UAV obtains its low-dimensional embedded coordinates in the communication manifold space. This step fuses the communication manifold coordinates with the UAV's physical space coordinates to form a complete swarm dynamic model.

[0041] Physical coordinates of each drone Positioning can be acquired in real time via airborne RTK-GNSS modules or UWB positioning systems, with positioning accuracy typically reaching the centimeter level. Physical space coordinates and communication manifold coordinates are concatenated and fused to form a joint embedded feature vector. In the formula, and This is a weighting coefficient used to adjust the relative importance of physical space information and communication space information. In scenarios where physical distance significantly impacts task execution, it can be increased. The value of can be increased in scenarios where communication quality is the primary constraint. The value of .

[0042] In certain application scenarios, if cluster tasks are mainly constrained by communication topology, communication manifold coordinates can be used directly. As the unique state representation of the drone, this scheme can further simplify the design and implementation of upper-level algorithms.

[0043] In summary, the complete dynamic model of the entire drone swarm at time t is represented as follows: This model uniformly represents the physical spatial relative positions and real-time communication topology of each UAV in the cluster, and evolves and updates in real time as the communication environment changes, providing reliable underlying model support for upper-level cluster algorithms.

[0044] S5: Model Application and Validation Cluster dynamic model built based on S4 steps The upper-level clustering algorithm can be executed directly in the communication manifold space or joint embedding space without relying on the traditional simplification assumptions based on physical distance. The following are examples of how this model is applied in typical clustering tasks: During formation keeping missions, each UAV does not need to strictly maintain physical spatial distance from its neighbors, but rather maintains its communication manifold coordinates. The manifold distance between the communication manifold coordinates of the neighboring machines is within a preset range. This communication manifold-based formation strategy can adaptively respond to changes in the communication environment and achieve flexible formation while ensuring communication connectivity.

[0045] In the leader election task, the leader is selected by calculating the centrality index (e.g., distance to the centroid of the manifold) of each drone in the communication manifold space. Compared with the traditional leader election strategy based on physical location, the leader selected by this scheme has better overall communication quality with other members in the cluster, which is beneficial to improving the efficiency and reliability of cluster command distribution.

[0046] In distributed information fusion tasks, a weighted consensus fusion algorithm based on communication manifold distance is adopted to replace the traditional physical distance weighted algorithm. Since the communication manifold distance more accurately reflects the actual information transmission quality between nodes, the accuracy of the fusion result is significantly improved. Experimental verification shows that the fusion accuracy using the model of this invention is improved by 41.7% compared to traditional methods.

[0047] The comprehensive performance test results show that the trunking communication topology modeling accuracy of the method of the present invention reaches more than 95%, the false alarm rate of link interruption prediction is reduced by 87% compared with the prior art, and it can predict the severe deterioration of link quality 1.8 to 3.2 seconds in advance, leaving sufficient time window for the early adjustment of trunking control strategy.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for modeling low-altitude unmanned aerial vehicle (UAV) swarms, characterized in that, Includes the following steps: Communication quality vector acquisition: Each drone in the cluster measures its link quality indicators with other drones in the cluster in real time, which constitutes the communication quality vector of that drone. Dynamic communication quality matrix construction: The communication quality vectors of all UAVs are collected to form a dynamic communication quality matrix; Online communication manifold learning: Using the dynamic communication quality matrix as input, calculate the geodesic distance matrix between UAVs, perform eigenvalue decomposition after centering the geodesic distance matrix, and obtain the low-dimensional embedding coordinates of each UAV in the communication manifold space; Joint embedding modeling: The physical space coordinates of each UAV are fused with the communication manifold coordinates to form a cluster dynamic model that represents the physical space relationship and communication topology of the cluster.

2. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 1, characterized in that, The link quality metric mentioned is a comprehensive link quality metric, and its calculation formula is as follows: in, This represents the overall link quality index from UAV i to UAV j at time t. To normalize the received signal strength indication value, For packet error rate, For normalized signal-to-noise ratio, This is a normalized link capacity estimate. , , , The weighted coefficients are and satisfy the following conditions: .

3. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 1, characterized in that, The geodesic distance matrix between the calculated UAVs includes: The short-circuit distance between drones is determined based on the link quality index. When the link quality index is greater than or equal to a preset threshold, the short-circuit distance is equal to 1 minus the link quality index; when the link quality index is less than the preset threshold, the short-circuit distance is set to infinity. A weighted graph is constructed based on the short-circuit distance, and the shortest path algorithm is used to calculate the shortest path distance between all UAV pairs to form a geodesic distance matrix.

4. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 3, characterized in that, The preset threshold value is 0.3, and the shortest path algorithm adopts either the Floyd-Warshall algorithm or the Dijkstra algorithm.

5. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 1, characterized in that, The formula for centering the geodesic distance matrix is ​​as follows: in, The centered matrix, For the geodesic distance matrix, This indicates that the square operation is performed on each element of the geodesic distance matrix. For a centered matrix, It is an N-order identity matrix. It is an N-dimensional column vector with all elements being 1, where N is the number of drones.

6. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 1, characterized in that, The execution feature decomposition includes: Perform eigenvalue decomposition on the centered matrix and arrange the eigenvalues ​​in descending order; Select the top d largest positive eigenvalues ​​and their corresponding eigenvectors, and calculate the low-dimensional embedding coordinates of each UAV based on the eigenvalues ​​and eigenvectors, where the embedding dimension d ranges from 4 to 6.

7. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 1, characterized in that, The online communication manifold learning step also includes incremental online updates, specifically: Maintain the eigenvector matrix and eigenvalue diagonal matrix of the previous time step as prior information; When the dynamic communication quality matrix of the new moment arrives, calculate the change of the centralized matrix relative to the previous moment; The subspace tracking algorithm is used to incrementally update the feature vector and feature value to obtain the low-dimensional embedding coordinates at the current time.

8. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 1, characterized in that, In the joint embedding modeling step, the formula for calculating the joint embedding feature vector is: in, Let i be the joint embedding feature vector of UAV i at time t. Let i be the physical space coordinates of the drone. Let i be the communication manifold coordinates of UAV i. and These are the weighting coefficients.

9. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to claim 8, characterized in that, The physical space coordinates are obtained in real time through an airborne RTK-GNSS module or a UWB positioning system.

10. The low-altitude unmanned aerial vehicle (UAV) swarm modeling method according to any one of claims 1 to 9, characterized in that, It also includes model application steps: based on the cluster dynamic model, perform formation maintenance, leader election or distributed information fusion tasks in the communication manifold space or joint embedding space.

Images