A method and system for emergency drone swarm networking

Abstract

This invention discloses a method and system for emergency drone swarm networking, belonging to the field of drone swarm networking technology. It solves the link prediction lag problem at its source by using deep fusion perception of multi-dimensional graph feature matrices. The node feature matrix, edge tensor matrix, and structural feature matrix are simultaneously incorporated into dynamic attention calculation, with the relative speed difference and battery difference between adjacent drones as core weighting factors, achieving accurate prediction of future link stability. Compared with existing technologies that rely solely on the current state or a single feature for perception, this invention improves topology prediction accuracy and response speed, fundamentally reducing frequent topology reconstruction and communication link jitter caused by perception delays, and improving stability in complex and dynamic emergency rescue environments. Simultaneously, this invention, through a sparse-gated multi-expert fusion decision-making mechanism, reduces computational resource consumption while ensuring decision diversity, achieving resource-efficient intelligent decision-making and reducing decision latency.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) swarm networking technology, and more specifically, relates to an emergency UAV swarm networking method and system. Background Technology

[0002] Emergency rescue sites often face harsh conditions such as lack of public network access, satellite denial, and "three disruptions" (network outage, power outage, and road outage). Traditional ground-based networking methods rely on vehicles, roads, and communication facilities, resulting in inefficient deployment and difficult adjustments. Leveraging the lightweight and easily deployable characteristics of small unmanned aerial vehicle (UAV) nodes, communication networks can be rapidly established at rescue sites, providing stable network support and operational assurance for ground rescue personnel. However, in complex and time-varying rescue environments, the uneven spatial and temporal distribution of operational needs, coupled with highly sudden and unpredictable demands and discretely distributed node resources, makes it difficult for traditional single-UAV-based systems to meet diverse mission requirements. UAV swarm networking, as a distributed heterogeneous network formed by multiple UAVs through self-organization, not only provides flexible communication relay, data fusion, and collaborative sensing services in complex environments but also effectively enhances network robustness and environmental adaptability through intelligent algorithms for joint scheduling of node links and physical locations.

[0003] However, some existing drone swarm networking technologies have significant limitations. For example, while existing multi-hop self-organizing networks and protocol stack optimization techniques have improved swarm connectivity to some extent, they mostly rely on static topology assumptions and fixed routing strategies, making them ill-equipped to handle topology shifts and link oscillations caused by high-speed maneuvers, resulting in poor stability in complex, dynamic emergency rescue environments. Furthermore, the current lack of a unified dynamic perception and decision-making mechanism among drone swarm protocols leads to high decision-making delays. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides an emergency drone swarm networking method and system to solve the technical problems of poor stability and high decision delay in complex and dynamic emergency rescue environments.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for networking emergency unmanned aerial vehicle (UAV) swarms, comprising: Every first preset time period or when a change in the cluster topology is detected, dynamic graph data of the drone cluster at the current moment is acquired in real time. This dynamic graph data includes: a node feature matrix, an edge tensor matrix, and a structural feature matrix. The node feature matrix includes the node features of all drone nodes in the cluster; each node feature includes the drone's position, battery level, speed, and received signal strength. The edge tensor matrix includes the edge features between adjacent drones in the cluster; each edge feature includes the relative distance and relative speed difference between two drones connected by an edge. Battery Difference Signal strength on the communication link; the structural feature matrix includes: the degree, centrality data, and clustering coefficient of each UAV node in the cluster topology; The dynamic graph data at the current moment is fused: the node feature matrix, edge tensor matrix and structural feature matrix in the dynamic graph data are input into the multi-head attention model for weighted aggregation to obtain the embedded features of each UAV in the cluster at the current moment; In each cluster leader drone within the cluster, its current embedded features are input into its action decision model to obtain the current action command of that cluster leader drone, and subsequently, the current action commands of each drone within the cluster to which that cluster leader drone belongs. The action decision model comprises a sparse gating network and an expert module. The expert module includes multiple expert sub-networks connected to the output of the sparse gating network. The sparse gating network is used to calculate the matching degree between the input embedded features and each expert sub-network, activating the K expert sub-networks with the highest matching degrees, and inputting the embedded features into them respectively. The output commands of the K activated expert sub-networks constitute the action commands of the corresponding drone; K is a positive integer; different expert sub-networks correspond to different types of action commands. Control each drone to execute corresponding action commands to achieve network formation.

[0006] More preferably, K is 1 or 2.

[0007] More preferably, the above-mentioned emergency drone swarm networking method further includes: performing the following training operations every second preset time period: Collect data samples from the cluster at each time step and store them in the experience replay pool, including: obtaining the cluster's data at each time step. State observations at time t ; Including: the Dynamic graph data at any given moment ;right Perform a fusion operation to obtain the number of drones in the cluster at the [number]th [period]. Embedded features at time t; each cluster head UAV at the t moment The embedded features at time t are input into the corresponding action decision model to obtain the corresponding cluster-head UAV at time t. The action command at the given time is used to obtain the action command of each drone in the cluster to which the cluster leader drone belongs. The action command at the given time, and record the first... The sum of the number of activated expert subnetworks in all action decision models in the cluster at time t. ; Obtain the status observation values ​​of each UAV in the cluster after executing the corresponding action command. and the cluster in the Reward value at any time ; For the first Networking bonus value of the cluster at any given time With activation reward value The weighted summation result; where, when the first... When the topological connectivity of the cluster is greater than the first preset connectivity at a given time, The first preset positive reward value; otherwise, The first preset negative reward value; when When the quantity is less than the preset quantity, The second preset positive reward value; otherwise, The second preset negative reward value; , , and The quadruple formed as a cluster in the th The data sample obtained at time is denoted as . And store it in the experience replay pool; Includes: each drone in the cluster in the 1st Action commands at any given time; Data samples are drawn from the experience replay pool to form a training sample set; based on the training sample set, a reinforcement learning algorithm is used to train the multi-head attention model and the action decision model.

[0008] More preferably, the second preset time period has the same duration as the first preset time period.

[0009] More preferably, the above-mentioned emergency drone swarm networking method further includes: calculating the average of the broken link representation value of all edges in the swarm topology of the previous period and the normalized communication rate every first preset time period, as the topology jitter index of the swarm in the current period. Specifically, when the communication link corresponding to an edge is disconnected, its disconnection value is 0; when the communication link corresponding to an edge is connected, its disconnection value is 1.

[0010] More preferably, the action command types include: power adjustment, position adjustment, relay switching, and cluster head election.

[0011] More preferably, the cluster head UAV carries a corresponding action decision model; the UAVs within the cluster carry only a sparse gating network. For any i At that moment, when the When the topology jitter index of the cluster is greater than the preset index, any cluster leader drone Any UAV within a cluster of the same cluster In the Methods for obtaining action commands at a given moment include: Controlling drones within the cluster Acquiring cluster head drones The expert module of the action decision model in the system, and the cluster-head UAV. In the Action commands at any time ; In-cluster drones In the middle, the drones within the cluster No. The embedded features at each time step are input into the sparse gating network within it and the acquired cluster-head drone. Action instructions are obtained from the action decision model composed of expert modules. ; when and When the same action instruction type exists, As an intra-cluster drone In the Action commands at any given time; when and When the action command types in the code are all different: For cluster drones In the The current battery level is less than the preset battery level, and and If there is an action command of the power adjustment action type, select... and This includes action commands of the power adjustment type, as for drones within the cluster. In the Action commands at any given time; For cluster drones In the The current battery level is less than the preset battery level, and and There are no action commands of the power adjustment type in the cluster, or, for drones within the cluster... In the If the current battery level is greater than the preset battery level, continue to determine the next step. Is the cluster topology connectivity at any given time greater than the second preset connectivity? For the The cluster topology connectivity at time t is greater than the second preset connectivity, and and If there is an action command for switching action types during relay, select... and This includes action commands for relay switching action types, serving as intra-cluster drones. In the Action commands at any given time; For the The cluster topology connectivity at time t is greater than the second preset connectivity, and and There is no relay action instruction for switching action types, or, in the case where there is no such instruction, the first... If the cluster topology connectivity at a given time is less than the second preset connectivity, continue the judgment. and Does the system contain an action instruction of the position adjustment type? If so, select... and This includes action commands for position adjustment actions, serving as intra-cluster drones. In the The action command at the current moment; otherwise, select and This includes action commands for cluster head election actions, serving as instructions for drones within the cluster. In the Action commands at any given time; The above-mentioned emergency drone swarm networking method also includes: when the topology jitter index of the swarm is greater than the preset index, after obtaining the action command of the drone in the swarm, deleting the expert module obtained from the corresponding swarm head drone in the drone in the swarm.

[0012] Secondly, the present invention provides an emergency drone swarm networking system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the emergency drone swarm networking method provided in the first aspect of the present invention.

[0013] Thirdly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls the device containing the storage medium to execute the emergency unmanned aerial vehicle swarm networking method provided in the first aspect of the present invention.

[0014] Fourthly, the invention also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the emergency drone swarm networking method provided in the first aspect of the invention.

[0015] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: 1. This invention provides a method for networking emergency drone swarms. Through deep fusion perception using multi-dimensional graph feature matrices, it addresses the link prediction lag problem at its source. Specifically, it simultaneously incorporates node feature matrices, edge tensor matrices, and structural feature matrices into dynamic attention calculations, using the relative speed difference and battery level difference between adjacent drones as core weighting factors to achieve accurate prediction of future link stability. Compared to existing technologies that rely solely on the current state or a single feature for perception, this invention significantly improves topology prediction accuracy and response speed, fundamentally reducing frequent topology reconstruction and communication link jitter caused by perception delays, and enhancing stability in complex and dynamic emergency rescue environments. Simultaneously, this invention employs a sparse-gated multi-expert fusion decision-making mechanism, significantly reducing computational resource consumption while ensuring decision diversity, achieving resource-efficient intelligent decision-making and reducing decision latency.

[0016] 2. Furthermore, the emergency drone swarm networking method provided by this invention introduces a penalty term for the activation count of expert subnetworks into the reinforcement learning reward function, forcing the action decision model to prioritize the optimal combination of expert subnetworks while maintaining Top-K sparse routing. Compared to the shortcomings of existing technologies, which either have a single decision expression or excessive computational overhead, this invention ensures decision quality while significantly reducing computational and energy consumption burdens in multi-dimensional heterogeneous scenarios, achieving a balance between decision diversity and resource efficiency.

[0017] 3. Furthermore, the emergency drone swarm networking method provided by this invention designs a distributed self-healing execution mechanism. This mechanism ensures that drone nodes within the cluster only reside in a lightweight gating network and basic execution logic under normal conditions. When the cluster's topology jitter index is detected to be greater than a preset index, the node requests temporary expert subnetwork parameters from the cluster leader drone to load the required expert subnetwork. After completing the self-healing action, the loaded parameters are immediately released, achieving rapid recovery and extremely low overhead self-healing capabilities. Unlike existing technologies that require each drone to fully deploy a complex decision-making model, resulting in severely limited computation and energy consumption, this invention significantly reduces the persistent resource occupation of ordinary nodes, making self-healing recovery faster, overall energy consumption more balanced, and system continuous operation more robust, achieving extremely lightweight operation of nodes within the cluster.

[0018] 4. The emergency drone swarm networking method provided by this invention does not require each drone to be equipped with high-performance hardware. Cluster nodes normally operate under a complex dormant model, only waking up when needed. This results in low deployment costs, strong adaptability, and can be incrementally deployed directly within existing drone swarm systems. Compared to existing technologies that commonly face problems such as insufficient computing power, rapid battery depletion, and high costs in practical applications, this invention provides a more practical and easily promoted complete solution, particularly suitable for large-scale networking applications of micro and small drones. Attached Figure Description

[0019] Figure 1 A block diagram illustrating the principle of adaptive networking of unmanned aerial vehicle (UAV) swarms provided in an embodiment of the present invention; Figure 2 This is a flowchart of the UAV swarm adaptive networking method provided in an embodiment of the present invention; Figure 3 This is a timing diagram of the workflow for adaptive networking of unmanned aerial vehicle (UAV) swarms provided in an embodiment of the present 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 and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] To achieve the above objectives, in a first aspect, the present invention provides a method for networking emergency unmanned aerial vehicle (UAV) swarms, comprising: Every first preset time period or when a change in the cluster topology is detected, dynamic graph data of the drone cluster at the current moment is acquired in real time. This dynamic graph data includes: a node feature matrix, an edge tensor matrix, and a structural feature matrix. The node feature matrix includes the node features of all drone nodes in the cluster; each node feature includes the drone's position, battery level, speed, and received signal strength. The edge tensor matrix includes the edge features between adjacent drones in the cluster; each edge feature includes the relative distance and relative speed difference between two drones connected by an edge. Battery Difference Signal strength on the communication link; the structural feature matrix includes: the degree, centrality data, and clustering coefficient of each UAV node in the cluster topology; The dynamic graph data at the current moment is fused: the node feature matrix, edge tensor matrix and structural feature matrix in the dynamic graph data are input into the multi-head attention model for weighted aggregation to obtain the embedded features of each UAV in the cluster at the current moment; In each cluster leader UAV within the cluster, its current embedded features are input into its action decision model to obtain the current action command of the cluster leader UAV, and then to obtain the current action commands of each UAV within the cluster to which the cluster leader UAV belongs. The action decision model includes a sparse gating network and multiple expert subnetworks connected to the output of the sparse gating network. The sparse gating network is used to calculate the matching degree between the input embedded features and each expert subnetwork, activates the K expert subnetworks with the highest matching degree, and inputs the embedded features into them respectively. The output commands of the activated K expert subnetworks constitute the action commands of the corresponding UAV. K is a positive integer; different expert subnetworks correspond to different types of action commands. Preferably, in one optional implementation, K is 1 or 2. Control each drone to execute corresponding action commands to achieve network formation.

[0022] This invention addresses the link prediction lag problem at its source by employing deep fusion perception using multi-dimensional graph feature matrices. It simultaneously incorporates node feature matrices, edge tensor matrices, and structural feature matrices into dynamic attention calculations, using the relative speed difference and battery level difference between adjacent drones as core weighting factors to achieve proactive perception of potential link breakage risks. Compared to existing technologies that rely solely on the current state or a single feature, this invention significantly improves topology prediction accuracy and fundamentally reduces frequent reconstructions and link oscillations caused by perception delays.

[0023] It should be noted that the expert subnetworks mentioned above can adopt any machine learning model, such as SVM, neural network, MLP, etc., without any restrictions.

[0024] In one optional implementation, the above-mentioned emergency drone swarm networking method further includes: performing the following training operation every second preset time period: Collect data samples from the cluster at each time step and store them in the experience replay pool, including: obtaining the cluster's data at each time step. State observations at time t ; Including: the Dynamic graph data at any given moment ;right Perform a fusion operation to obtain the number of drones in the cluster at the [number]th [period]. Embedded features at time t; each cluster head UAV at the t moment The embedded features at time t are input into the corresponding action decision model to obtain the corresponding cluster-head UAV at time t. The action command at the given time is used to obtain the action command of each drone in the cluster to which the cluster leader drone belongs. The action command at the given time, and record the first... The sum of the number of activated expert subnetworks in all action decision models in the cluster at time t. ; Obtain the status observation values ​​of each UAV in the cluster after executing the corresponding action command. and the cluster in the Reward value at any time ; For the first Networking bonus value of the cluster at any given time With activation reward value The weighted summation result; where, when the first... When the topological connectivity of the cluster is greater than the first preset connectivity at a given time, The first preset positive reward value; otherwise, The first preset negative reward value; when When the quantity is less than the preset quantity, The second preset positive reward value; otherwise, The second preset negative reward value; , , and The quadruple formed as a cluster in the th The data sample obtained at time is denoted as . And store it in the experience replay pool; Includes: each drone in the cluster in the 1st Action commands at any given time; Data samples are drawn from the experience replay pool to form a training sample set; based on the training sample set, a reinforcement learning algorithm is used to train the multi-head attention model and the action decision model.

[0025] It should be noted that the aforementioned first preset connectivity, first preset positive reward value, first preset negative reward value, preset quantity, second preset positive reward value, and second preset negative reward value can be determined based on experience. In one optional implementation, the first preset connectivity is set to 2; the first preset positive reward value is set to 10; the first preset negative reward value is set to -50; the preset quantity is set to 3; the second preset positive reward value is set to 5; and the second preset negative reward value is set to -30.

[0026] The above training method achieves a balance between decision diversity and resource efficiency. This invention introduces a penalty term for the activation count of expert subnetworks into the reinforcement learning reward function, forcing the action decision model to prioritize the optimal expert combination while maintaining Top-K sparse routing. Compared to existing technologies that either suffer from simplistic decision representation or excessive computational overhead, this invention ensures decision quality while significantly reducing computational and energy consumption in multidimensional heterogeneous scenarios.

[0027] It should be noted that the above training method is only a preferred method, not the only one. The above reward value can also be the network reward value, which is not limited here.

[0028] It should be noted that the reinforcement learning algorithm described above can be the Proximal Policy Optimization (PPO) algorithm, SAC algorithm, Q-Learning algorithm, DQN algorithm, etc., and is not limited here. Preferably, in one optional implementation, the reinforcement learning algorithm described above is the Proximal Policy Optimization (PPO) algorithm.

[0029] The durations of the second preset time period and the first preset time period can be determined empirically. Preferably, in one optional implementation, the second preset time period and the first preset time period are the same length to reduce computational and communication overhead. In one optional implementation, the durations of the first and second preset time periods are both 5 seconds.

[0030] In one optional implementation, the above-mentioned emergency drone swarm networking method further includes: calculating the average of the broken link representation values ​​of all edges in the swarm topology of the previous period and the normalized communication rate every first preset time period, as the topology jitter index of the swarm in the current period. Specifically, when the communication link corresponding to an edge is disconnected, its disconnection value is 0; when the communication link corresponding to an edge is connected, its disconnection value is 1.

[0031] In one alternative implementation, the action types include: power adjustment, position adjustment, relay switching, and cluster head election.

[0032] It should be noted that the drones in the aforementioned drone swarm are divided into multiple clusters, each with a cluster leader drone and the rest being intra-cluster drones; communication between clusters occurs through the cluster leader drone; the cluster leader drone carries a corresponding action decision model; and the intra-cluster drones only carry a sparse gating network. When the topology jitter index is less than or equal to a preset index (0.85 in one optional implementation), the topology is normal, and the intra-cluster drones and the cluster leader drone have a fixed cooperative relationship, with their behavior pre-set according to their cooperative relationship with the cluster leader drone. When the topology jitter index is greater than the preset index, the topology jitters abnormally, and the swarm enters a self-healing phase, performing cluster leader adjustment and wake-up scheduling operations to ensure transparent connection of action execution between the multi-node system. Specifically, in one optional implementation, for any... At that moment, when the When the topology jitter index of the cluster at a given time exceeds a preset index, any cluster leader drone... Any UAV within a cluster of the same cluster In the Methods for obtaining action commands at a given moment include: Controlling drones within the cluster Acquiring cluster head drones The expert module of the action decision model in the system, and the cluster-head UAV. In the Action commands at any time ; In-cluster drones In the middle, the drones within the cluster No. The embedded features at each time step are input into the sparse gating network within it and the acquired cluster-head drone. Action instructions are obtained from the action decision model composed of expert modules. ; when and When the same action instruction type exists, As an intra-cluster drone In the Action commands at any given time; when and When the types of action instructions in the code are all different: For cluster drones In the The current battery level is less than the preset battery level (in one optional implementation, this is taken as 30% of the rated battery level of the drones in the cluster), and and If there is an action command of the power adjustment action type, select... and This includes action commands of the power adjustment type, as for drones within the cluster. In the Action commands at any given time; For cluster drones In the The current battery level is less than the preset battery level, and and There are no action commands of the power adjustment type in the cluster, or, for drones within the cluster... In the If the current battery level is greater than the preset battery level, continue to determine the next step. Does the cluster topology connectivity at any given time exceed the second preset connectivity (which is 1.5 in one optional implementation)? For the The cluster topology connectivity at time t is greater than the second preset connectivity, and and If there is an action command for switching action types during relay, select... and This includes action commands for relay switching action types, serving as intra-cluster drones. In the Action commands at any given time; For the The cluster topology connectivity at time t is greater than the second preset connectivity, and and There is no relay action instruction for switching action types, or, in the case where there is no such instruction, the first... If the cluster topology connectivity at a given time is less than the second preset connectivity, continue the judgment. and Does the system contain an action instruction of the position adjustment type? If so, select... and This includes action commands for position adjustment actions, serving as intra-cluster drones. In the The action command at the current moment; otherwise, select and This includes action commands for cluster head election actions, serving as instructions for drones within the cluster. In the Action commands at any given moment.

[0033] The above-mentioned emergency drone swarm networking method also includes: when the topology jitter index of the swarm is greater than the preset index, after obtaining the action command of the drone in the swarm, deleting the expert module obtained from the corresponding swarm head drone in the drone in the swarm.

[0034] This invention enables UAV nodes within a cluster to reside only in a lightweight gated network and basic execution logic under normal conditions. When the cluster's topology jitter index is detected to exceed a preset index, the node requests temporary expert subnetwork parameters from the cluster leader UAV to load the required expert subnetwork. After completing the self-healing process, the loaded parameters are immediately released. Unlike existing technologies that require each UAV to deploy a complete and complex decision-making model, resulting in severely limited computation and energy consumption, this invention significantly reduces the persistent resource occupation of ordinary nodes, enabling faster self-healing recovery, more balanced overall energy consumption, and stronger system continuous operation capabilities, achieving extremely lightweight operation of nodes within the cluster.

[0035] To further illustrate the emergency drone swarm networking method provided by this invention, a specific embodiment is described in detail below: This embodiment designs a drone swarm networking method based on dynamic attention and multi-expert decision-making, the architecture of which is as follows: Figure 1 As shown, a mobile networking system for emergency rescue is constructed by integrating dynamic topology perception, intelligent decision fusion, and distributed self-healing execution throughout the entire process.

[0036] Specifically, the dynamic perception layer, which collects real-time status data from the drone swarm, serves as the interface connecting the physical environment and intelligent decision-making. In actual operation, the dynamic perception layer supports periodic acquisition of multi-dimensional states and dynamic graph construction. It can proactively sense information such as node position, speed, energy (i.e., battery level), and link quality, generating a topology jitter index to trigger upper-layer decisions. The intelligent decision-making layer is the "brain" of the drone swarm, primarily implementing multi-expert routing and policy fusion. In practice, based on the real-time dynamic graph uploaded by the perception layer, the intelligent decision-making layer extracts look-ahead embedding vectors using a dynamic attention mechanism and performs routing decisions for heterogeneous scenarios through sparse multi-expert gating. It dynamically selects the optimal expert sub-policy to construct a decision family to support automated execution of action generation and resource scheduling. The distributed execution layer focuses on cluster head selection and self-healing regulation, forming a crucial foundation for achieving highly robust topology maintenance. In practice, the distributed execution layer provides a unified execution view and feedback capability to the upper layer. It can generate node labels for various node entities, such as highly maneuverable rotary-wing UAVs, resource-rich fixed-wing platforms, energy-constrained micro UAVs, and edge relay nodes, and collect their execution status in real time, thereby dynamically constructing a self-healing feedback graph to provide data support for upper-layer decision-making.

[0037] In the adaptive networking architecture of UAV swarms based on dynamic attention and multi-expert fusion, the dynamic attention perception layer, the multi-expert fusion decision layer, and the distributed self-healing execution layer work together in a closed loop. Through a vertical interaction mechanism, each layer constructs a closed loop of perception prediction, decision routing, and self-healing feedback, supporting highly robust and differentiated stable interconnection requirements in highly dynamic and heterogeneous scenarios. Specifically, the dynamic attention perception layer performs multi-dimensional fusion of node features, edge features, and structural features, and introduces weighted weighting based on relative speed and energy differences to achieve forward-looking prediction of future link stability and generate weighted embedding vectors. The multi-expert fusion decision layer employs a sparse gating mechanism and imposes constraints on the number of expert activations in the reward function, achieving efficient routing and policy fusion of experts in heterogeneous scenarios. The distributed self-healing execution layer ensures that ordinary nodes normally retain only a lightweight gating network. When a topology anomaly is detected, it requests temporary expert parameters from the cluster head for loading, and releases them immediately after execution, thus achieving rapid self-healing with low overhead. The specific process is as follows: Every first preset time period or when a change in cluster topology is detected, the dynamic perception layer first collects multi-dimensional states and generates a dynamic graph, which is then sent to the intelligent decision layer. The intelligent decision-making layer performs expert routing mapping based on the state graph and service requirements, generating a decision-making group strategy that includes action policies and activation penalties. It dynamically selects the optimal combination of expert subnetworks to support the decision, achieving on-demand resource routing and differentiated resource protection. The decision strategy is then distributed to the distributed execution layer, which, based on the received strategy information, schedules corresponding node components to complete tasks such as action execution, self-healing wake-up, and connectivity maintenance, thus completing the closed loop. Through the deep coupling and feedback interaction of these three layers, a complete perception-decision-execution closed loop is formed, effectively solving the prominent problems of existing UAV swarm networking technology, such as lagging topology perception, insufficient decision expression capabilities, high computational and energy consumption, and low self-healing recovery efficiency.

[0038] The UAV swarm networking method designed in this embodiment, based on dynamic attention and multi-expert decision-making, is as follows: Figure 2As shown, the system is completed collaboratively by three functional modules integrated within the UAV node: a dynamic attention perception module, a multi-expert fusion decision-making module, and a distributed self-healing execution module, which jointly support stable interconnection in highly dynamic environments. To address the issues of topology shifts and link prediction lag in UAV swarms, a dynamic attention perception module is designed to perform forward embedding of multi-dimensional states and stability weighting, enabling predictive perception of future link loss risks. Considering the characteristics of swarm decision expression being singular and poor adaptability to heterogeneous scenarios, a multi-expert fusion decision-making module is designed. It constructs a routing model based on sparse gating, introduces reward shaping with expert activation penalty, and dynamically routes to dedicated experts, achieving a unified improvement in decision diversity and computational efficiency. In the distributed self-healing execution module, to address the issues of high recovery latency and energy overload during topology oscillations, an expert wake-up package strategy is introduced. Ordinary nodes borrow experts on demand, improving self-healing success rate and robustness while controlling additional load, significantly enhancing the swarm's ability to withstand harsh maneuvering conditions.

[0039] The dynamic attention perception module is integrated into the information processing unit of each UAV, primarily responsible for constructing a dynamic heterogeneous graph in real time and generating look-ahead embedding vectors. The UAV first collects its own and its neighbors' multi-dimensional states, including node features (such as the UAV's position, speed, remaining energy, RSSI / SNR, etc.), edge features (relative distance, relative speed difference, energy difference, signal strength, etc. between adjacent UAVs), and structural features (degree, centrality data, and clustering coefficients of each UAV node in the cluster topology), constructing a dynamic graph containing node feature matrices, edge tensor matrices, and structural feature matrices. This module uses a multi-head attention mechanism to deeply fuse and weightedly aggregate these three types of matrices, generating embedding vectors with link future stability prediction capabilities, and acquiring the cluster's topology jitter index in real time for subsequent module triggering. This module does not require recalculating the entire graph each time, only incrementally updating the changed parts, thus achieving efficient look-ahead topology perception under the limited computing power of the UAV. The dynamic graph constructed in this embodiment is a temporal dynamic heterogeneous graph. Real-time fusion of node feature matrices (Including drone position, battery level, speed, RSSI / SNR, etc.), side tensors (Including the relative distance and relative speed difference between two adjacent drones) Energy difference (Signal strength, etc.) and structural feature matrix (Including: the degree, centrality data and clustering coefficient of each drone node in the cluster topology), and generate a look-ahead stability embedding vector through a multi-head attention mechanism to achieve predictive perception of the risk of future link breakage, breaking through the perception bottleneck of traditional GAT based only on single-dimensional features.

[0040] The multi-expert fusion decision-making module is integrated into the main decision-making unit of the UAV, responsible for generating the optimal action strategy based on the embedding vectors output by the dynamic attention perception module. The cluster-head UAV has multiple pre-built lightweight expert sub-networks (each targeting typical scenarios such as high maneuverability, strong interference, and energy constraints). This module routes the embedding vectors through a sparse gating network, activating only the one or two best-matching expert sub-networks, and applying a penalty constraint on the number of expert activations in the reward function to ensure a balance between decision diversity and computational cost. The actions output by this module include power adjustment, position fine-tuning, relay switching, and cluster head election, which can be executed directly or used as the strategy basis for subsequent self-healing modules, thereby achieving efficient and intelligent decision-making under resource-constrained UAV conditions. This embodiment uses sparse Top-K (K≤2) gating routing for the multi-expert sub-networks and adds a penalty term for the number of expert activations to the reinforcement learning reward function, forcing the system to reduce computational cost while ensuring decision diversity, achieving optimal strategy fusion in multi-dimensional heterogeneous scenarios such as high maneuverability, strong interference, and extremely uneven energy.

[0041] The distributed self-healing execution module is integrated into the UAV's execution control unit, responsible for topology maintenance and rapid recovery. Normally, UAV nodes only retain a lightweight gating network and basic execution logic, without residing in the complete expert subnetwork. When the dynamic attention perception module detects that the topology jitter index exceeds a threshold, this module sends a lightweight expert request packet to the cluster head. The cluster head then sends the corresponding expert subnetwork model parameter packet, which the node temporarily loads and completes self-healing actions (including cluster head election, relay reconstruction, power adjustment, and position adjustment). Upon completion, the loaded expert subnetwork model parameter packet is immediately released. This module, in collaboration with the cluster head, implements a layered execution mechanism that is "extremely simple under normal conditions and instantly awakens in case of anomalies," significantly reducing the persistent storage and computational overhead of ordinary nodes and ensuring the cluster has millisecond-level recovery capabilities during drastic topology changes. This embodiment enables nodes within a cluster to reside only in a lightweight gated network under normal circumstances. When the topology jitter index is detected to exceed a threshold, a lightweight expert parameter package is requested from the cluster head to temporarily load the required expert subnetwork. After completing the self-healing action, the loaded parameters are released immediately. This allows ordinary nodes to maintain extremely low resource consumption under normal conditions while still achieving rapid distributed recovery when the topology is abnormal, significantly reducing overall computational and energy consumption overhead and improving the self-healing efficiency and robustness of the system in resource-constrained environments.

[0042] This embodiment employs a combined centralized training and distributed lightweight execution approach during actual deployment. This allows cluster nodes to reside only in a lightweight gated network and basic execution logic under normal conditions. When a topology anomaly is detected, the nodes request temporary expert parameters from the cluster head to load the necessary expert subnetwork. After completing a self-healing decision, the loaded parameters are immediately released. This significantly reduces the resident storage requirements and computational energy consumption of ordinary nodes while maintaining global decision consistency. This operational mode overcomes the limitation of existing technologies requiring the complete deployment of complex decision-making models on each UAV, enabling this invention to achieve long-term stable networking on micro-UAV platforms with extremely limited computing power and energy. It exhibits excellent resource adaptability and large-scale deployment capabilities.

[0043] The three functional modules mentioned above work closely together within the drone swarm to form a complete perception-decision-execution closed loop. This enables each drone to independently complete local adaptive adjustments and achieve globally optimal networking through cluster head collaboration, thereby achieving long-term, efficient, robust, and stable interconnection on a drone platform with extremely limited resources.

[0044] This embodiment proposes a lightweight implementation of a dynamic attention and multi-expert fusion-based adaptive networking method for UAV swarms on a resource-constrained platform. It aims to address the topology maintenance challenges in highly dynamic and heterogeneous environments, improving network stability, decision-making efficiency, and robustness. The method employs a hybrid execution paradigm of "centralized training—distributed lightweight execution + expert wake-up package," comprising three core stages: dynamic attention perception, multi-expert fusion decision-making, and distributed self-healing execution. It supports various flexible technical solutions. For example, the dynamic attention stage supports multi-head attention and LeakyReLU activation; the multi-expert fusion decision-making stage supports Top-K sparse gating, PPO / SAC, and other reinforcement learning variants. The overall performance of the method does not depend on a specific algorithm implementation. The techniques described in this embodiment are preferred implementations, designed to address the topology maintenance challenges in highly dynamic and heterogeneous environments, improving network stability, decision-making efficiency, and robustness. The following section combines... Figure 3 (Workflow sequence diagram) This document details the complete execution flow of the method provided in this embodiment on a drone node.

[0045] Upon receiving the drone swarm status, the system first processes the data through a dynamic attention perception stage. This stage divides the multidimensional input into a node feature matrix (including position, remaining energy, velocity, RSSI / SNR, queue length, etc.), an edge tensor matrix (including relative distance, relative velocity difference Δv, energy difference Δe, signal strength, link delay, etc.), and a structural feature matrix (including node degree, centrality data, clustering coefficients, etc.). A multi-head dynamic attention mechanism deeply fuses and weights these three types of matrices to generate an embedding vector with forward-looking capabilities for future link stability. Simultaneously, it adaptively adjusts the attention focus allocation based on the real-time calculated topology jitter index, thereby improving prediction accuracy without requiring a full map recalculation each time.

[0046] Subsequently, in the multi-expert fusion decision-making stage, a sparse gated routing model is constructed based on the real-time embedded vectors and the requirements of the current heterogeneous scenario, and Top-K expert activation and policy fusion are executed. This stage supports dynamic adjustment of expert weights according to the activation cycle, effectively avoiding decision overload or local optima. At the same time, the multi-expert fusion decision-making stage integrates a reward closed-loop mechanism, which predicts changes in decision performance in real time through periodic or event-driven penalty feedback, automatically adjusts the routing strategy or initiates expert switching, significantly enhancing the method's adaptability in complex and maneuvering environments.

[0047] Subsequently, the distributed self-healing execution phase completes cluster head control and wake-up scheduling, ensuring transparent connection of action execution across multi-node systems. Furthermore, during action execution, the distributed self-healing execution phase encapsulates critical feedback (such as connectivity alarms) in lightweight packages and flexibly selects the wake-up method based on the current topology state, achieving robust and low-overhead topology maintenance, greatly improving execution security while ensuring interconnectivity.

[0048] After the state feedback reaches the decision-making stage, the topology changes are first identified through the dynamic attention perception stage, and the embedded vector is iteratively updated. If the self-healing stage is in progress at this time, the distributed self-healing execution stage will simultaneously extract and process the feedback. After that, the multi-expert fusion decision-making stage performs routing analysis and expert redistribution on the feedback data. Finally, after accumulating enough feedback, the dynamic attention perception stage completes the closed-loop iteration of topology prediction, realizing long-term reliable adaptive networking.

[0049] The UAV swarm adaptive networking method proposed in this embodiment, which combines dynamic attention and multi-expert fusion, achieves forward-looking prediction and seamless decision-making in heterogeneous maneuvering environments through efficient collaboration in stages such as dynamic perception, multi-expert sparse routing fusion, and distributed wake-up self-healing. It also significantly improves the robustness, shock resistance, and energy efficiency of topology maintenance. It is suitable for high-dynamic scenarios in emergency rescue and has excellent adaptability, scalability, and engineering feasibility. It is especially suitable for the large-scale deployment of micro-UAV platforms with limited computing power and battery capacity.

[0050] Specifically, this embodiment is applicable to post-disaster search and rescue and emergency communication support scenarios. It is mainly used for multiple drones to quickly enter the disaster area to establish a temporary communication network, coordinate the search for survivors, and transmit audio, video, and location information in real time. Addressing the challenges of complex disaster area environments, the paralysis of existing base stations, and the vulnerability of drones to aftershocks or dust, this embodiment can achieve self-organizing networking and highly robust communication under extreme conditions.

[0051] Multi-dimensional feature perception in extreme environments: UAVs in disaster areas need to cope with multiple disturbances such as strong winds, dust, and electromagnetic interference. The dynamic attention perception stage integrates features such as drastic changes in node speed, rapid energy consumption, and drastic fluctuations in RSSI / SNR with changes in structural centrality to generate high-precision look-ahead embedding vectors. This enables the cluster to identify links that are about to break in advance and make preventative adjustments to ensure that the search formation always maintains complete connectivity.

[0052] Sparse experts address multiple threats: Disaster areas simultaneously face complex threats such as high mobility (avoiding collapsed buildings), strong interference (signal scattering caused by aftershocks), and extremely uneven energy distribution (some drones need to hover for extended periods). In this embodiment, the multi-expert fusion decision-making stage automatically selects and activates 1-2 optimal experts based on the embedding vector (e.g., mobility experts lead path fine-tuning, interference experts lead frequency switching, and energy experts lead load balancing). Through an activation count penalty mechanism, the decision-making overhead is kept to a very low level, enabling drones to continue performing search and rescue missions even when battery power is low.

[0053] On-demand wake-up for self-healing and rapid reconstruction: When an earthquake aftershock or a drone is blown off course by strong winds, causing a local link break, the drone node within the cluster immediately requests an expert wake-up packet from the nearest cluster head. This temporarily loads a self-healing expert to complete rapid topology reconstruction (such as re-electing a cluster head and establishing a backup relay link), and releases resources immediately after completion. This mechanism enables the cluster to achieve second-level recovery even under extreme disturbances, ensuring that search and rescue audio / video and vital signs data are transmitted back to the command center in real time, buying precious time for the critical 72-hour rescue operation.

[0054] The three core technologies work in deep synergy to form a complete closed loop of forward-looking perception, efficient decision-making, and on-demand self-healing. These three technologies are highly coupled: dynamic attention perception provides precise embedding vectors for multi-expert fusion decision-making; multi-expert fusion decision-making outputs action strategies while generating parameter loading instructions; and distributed self-healing execution feeds the execution results back to the perception layer in real time, forming a positively reinforcing closed-loop interaction. Compared to the shortcomings of existing technologies where functional modules are relatively independent and lack effective collaboration, this invention significantly improves connectivity stability, real-time decision-making, and resource utilization efficiency under the same conditions.

[0055] In summary, this embodiment proposes an adaptive networking method for emergency UAV swarms based on dynamic attention and multi-expert fusion. It is a vertically layered closed-loop architecture that achieves deep integration of the entire process from perception to decision-making to execution. A vertically integrated networking system is constructed, comprising a dynamic attention perception layer, a multi-expert fusion decision-making layer, and a distributed self-healing execution layer, completely breaking the limitations of existing technologies that only optimize at a single layer of perception or decision-making. Through real-time interaction and feedback across these three layers, a true online closed loop is formed, enabling the system to maintain extremely high connectivity stability and robustness in high-speed three-dimensional maneuvering environments. Specifically, this embodiment systematically solves key problems such as lagging topology perception, insufficient decision expression, and excessive computational energy consumption in highly dynamic heterogeneous environments from three levels: multi-dimensional graph feature look-ahead perception, sparse multi-expert decision-making, and distributed wake-up self-healing execution. This achieves the invention's objective of long-term stable interconnection on extremely resource-constrained platforms. This embodiment forms a compact, functionally complementary, and dynamically adaptive complete closed-loop solution through the close coupling of three technologies: deep fusion of node, edge, and structure feature matrices using dynamic attention, multi-expert sparse routing and activation count penalty mechanism, and distributed self-healing execution driven by "expert wake-up packages." This scheme ensures that ordinary nodes only retain a lightweight gating network under normal circumstances, and only instantaneously load experts in case of anomalies, thus achieving globally optimal decision-making while minimizing computation and energy consumption.

[0056] Secondly, the present invention provides an emergency drone swarm networking system, comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the emergency drone swarm networking method provided in the first aspect of the present invention.

[0057] The related technical solutions are the same as the emergency drone swarm networking method provided in the first aspect of this invention, and are not limited here.

[0058] Thirdly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls the device containing the storage medium to execute the emergency drone swarm networking method provided in the first aspect of the present invention.

[0059] The related technical solutions are the same as the emergency drone swarm networking method provided in the first aspect of this invention, and are not limited here.

[0060] Fourthly, the invention also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the emergency drone swarm networking method provided in the first aspect of the invention.

[0061] The related technical solutions are the same as the emergency drone swarm networking method provided in the first aspect of this invention, and are not limited here.

[0062] Those skilled in the art will readily understand that the above description is merely 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 scope of protection of the present invention.

Claims

1. A method for networking emergency unmanned aerial vehicle (UAV) swarms, characterized in that, include: Every first preset time period or when a change in cluster topology is detected, the dynamic graph data of the drone cluster at the current moment is acquired in real time. The dynamic graph data includes: a node feature matrix, an edge tensor matrix, and a structural feature matrix; the node feature matrix includes: node features of all drone nodes in the cluster; the node features include: the position, battery level, speed, and received signal strength of the corresponding drone; the edge tensor matrix includes: edge features between pairs of adjacent drones in the cluster; the edge features include: the relative distance and relative speed difference between two drones connected by the edge. Battery difference The signal strength on the communication link; the structural feature matrix includes: the degree, centrality data and clustering coefficient of each UAV node in the cluster topology; The dynamic graph data at the current moment is fused: the node feature matrix, edge tensor matrix and structural feature matrix in the dynamic graph data are input into the multi-head attention model for weighted aggregation to obtain the embedded features of each UAV in the cluster at the current moment; In each cluster leader drone within the cluster, its current embedded features are input into its action decision model to obtain the current action command of that cluster leader drone, and subsequently, the current action commands of each drone within the cluster to which that cluster leader drone belongs are obtained. The action decision model comprises a sparse gating network and an expert module. The expert module includes multiple expert sub-networks connected to the output of the sparse gating network. The sparse gating network is used to calculate the matching degree between the input embedded features and each expert sub-network, activating the K expert sub-networks with the highest matching degrees, and inputting the embedded features into them respectively. The output commands of the K activated expert sub-networks constitute the action commands of the corresponding drones; K is a positive integer; different expert sub-networks correspond to different types of action commands. Control each drone to execute corresponding action commands to achieve network formation.

2. The emergency drone swarm networking method according to claim 1, characterized in that, K is 1 or 2.

3. The emergency drone swarm networking method according to claim 1, characterized in that, Also includes: Every second preset time period, perform the following training operations: Collect data samples from the cluster at each time step and store them in the experience replay pool, including: obtaining the cluster's data at each time step. State observations at time t ; Including: the Dynamic graph data at any given moment ;right Perform a fusion operation to obtain the number of drones in the cluster at the [number]th [period]. Embedded features at time t; each cluster head UAV at the t moment The embedded features at time t are input into the corresponding action decision model to obtain the corresponding cluster-head UAV at time t. The action command at the given time is used to obtain the action command of each drone in the cluster to which the cluster leader drone belongs. The action command at the given time, and record the first... The sum of the number of activated expert subnetworks in all action decision models in the cluster at time t. ; Obtain the status observation values ​​of each UAV in the cluster after executing the corresponding action command. and the cluster in the Reward value at any time ; For the first Networking bonus value of the cluster at any given time With activation reward value The weighted summation result; where, when the first... When the topological connectivity of the cluster is greater than the first preset connectivity at a given time, The first preset positive reward value; otherwise, The first preset negative reward value; when When the quantity is less than the preset quantity, The second preset positive reward value; otherwise, The second preset negative reward value; , , and The quadruple formed as a cluster in the th The data sample obtained at time is denoted as . And store it in the experience replay pool; Includes: each drone in the cluster in the 1st Action commands at any given time; Data samples are extracted from the experience replay pool to form a training sample set; based on the training sample set, a reinforcement learning algorithm is used to train the multi-head attention model and the action decision model.

4. The emergency drone swarm networking method according to claim 3, characterized in that, The second preset time period has the same duration as the first preset time period.

5. The emergency drone swarm networking method according to claim 1, characterized in that, Also includes: Every first preset time period, the average of the broken link representation value of all edges in the cluster topology of the previous period and the normalized communication rate is calculated as the topology jitter index of the cluster in the current period. Specifically, when the communication link corresponding to an edge is disconnected, its disconnection value is 0; when the communication link corresponding to an edge is connected, its disconnection value is 1.

6. The emergency unmanned aerial vehicle (UAV) swarm networking method according to any one of claims 1-5, characterized in that, The types of action commands include: power adjustment, position adjustment, relay switching, and cluster head election.

7. The emergency drone swarm networking method according to claim 6, characterized in that, The cluster-head UAV carries a corresponding action decision model; the UAVs within the cluster only carry a sparse gating network. For any i At that moment, when the When the topology jitter index of the cluster is greater than the preset index, any cluster leader drone Any UAV within a cluster of the same cluster In the Methods for obtaining action commands at a given moment include: Controlling drones within the cluster Acquiring cluster head drones The expert module of the action decision model in the system, and the cluster-head UAV. In the Action commands at any time ; In-cluster drones In the middle, the drones within the cluster No. The embedded features at each time step are input into the sparse gating network within it and the acquired cluster-head drone. Action instructions are obtained from the action decision model composed of expert modules. ; when and When the same action instruction type exists, As an intra-cluster drone In the Action commands at any given time; when and When the action command types in the code are all different: For cluster drones In the The current battery level is less than the preset battery level, and and If there is an action command of the power adjustment action type, select... and This includes action commands of the power adjustment type, as for drones within the cluster. In the Action commands at any given time; For cluster drones In the The current battery level is less than the preset battery level, and and There are no action commands of the power adjustment type in the cluster, or, for drones within the cluster... In the If the current battery level is greater than the preset battery level, continue to determine the next step. Is the cluster topology connectivity at any given time greater than the second preset connectivity? For the The cluster topology connectivity at time t is greater than the second preset connectivity, and and If there is an action command for switching action types during relay, select... and This includes action commands for relay switching action types, serving as intra-cluster drones. In the Action commands at any given time; For the The cluster topology connectivity at time t is greater than the second preset connectivity, and and There is no relay action instruction for switching action types, or, in the case where there is no such instruction, the first... If the cluster topology connectivity at a given time is less than the second preset connectivity, continue the judgment. and Does the file contain action instructions of the position adjustment type? and If there is an action instruction of the position adjustment action type, then select... and This includes action commands for position adjustment actions, serving as intra-cluster drones. In the The action command at the current moment; otherwise, select and This includes action commands for cluster head election actions, serving as instructions for drones within the cluster. In the Action commands at any given time; The emergency drone swarm networking method further includes: when the topology jitter index of the swarm is greater than a preset index, after obtaining the action command of the drone in the swarm, deleting the expert module obtained from the corresponding swarm head drone in the drone in the swarm.

8. An emergency unmanned aerial vehicle (UAV) swarm networking system, characterized in that, include: The system includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the emergency drone swarm networking method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, it controls the device where the storage medium is located to execute the emergency drone swarm networking method according to any one of claims 1-7.

10. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the emergency drone swarm networking method according to any one of claims 1-7.

Images