Multi-layer cooperative graph based multi-uav task allocation method, system, device and medium

Abstract

This invention provides a multi-UAV task allocation method, system, device, and medium using a multi-layer cooperative graph, relating to the field of UAV control technology. The method includes: obtaining a multi-layer cooperative graph model based on a set of real-time state information of the multi-UAV system, task description information, and skill description information; encoding the features of nodes in each layer of the multi-layer cooperative graph model, and fusing the node features to obtain global state features; inputting the global state features into a decision network to obtain real-time action information of each graph operation operator; adjusting the connection relationships between adjacent layers of nodes in the multi-layer cooperative graph model based on the real-time action information; and performing multi-UAV task allocation based on the updated connection relationships between adjacent layers of nodes in the multi-layer cooperative graph model. This invention improves the robustness of task allocation and decision convergence efficiency of the multi-UAV system by using dynamic evolution of the graph topology for multi-UAV task allocation.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a multi-UAV task allocation method, system, device, and medium using a multi-layer cooperative graph. Background Technology

[0002] Multi-UAV systems, with their advantages of flexibility, scalability, and low cost, have shown great application potential in fields such as military reconnaissance, disaster relief, area surveillance, communications relay, and coordinated strikes. Efficient task allocation and scheduling are key to maximizing the operational effectiveness of multi-UAV systems.

[0003] To meet the above requirements, existing multi-UAV task allocation technologies typically utilize specific optimization algorithms or fixed rules to perform action optimization and decision output within a predefined, fixed task label space.

[0004] However, this technology is limited by its structured decision-making mechanism, resulting in poor robustness and low decision convergence efficiency in highly dynamic and complex task scenarios. Summary of the Invention

[0005] This invention provides a multi-UAV task allocation method, system, device, and medium with a multi-layer cooperative graph, which addresses the shortcomings of existing technologies in highly dynamic and complex task scenarios, such as poor robustness and low decision convergence efficiency, thereby improving the robustness and decision convergence efficiency of multi-UAV systems in highly dynamic and complex task scenarios.

[0006] This invention provides a multi-UAV task allocation method for multi-layer cooperative graphs, including: Based on the real-time status information set of multiple unmanned aerial vehicle (UAV) systems, the task description information, and the skill description information, a multi-layer cooperation graph model is obtained; the multi-layer cooperation graph model includes a task node layer, a cooperation cluster node layer, a skill node layer, and an UAV node layer. The nodes of each layer in the multi-layer cooperative graph model are feature-encoded to obtain the node features of each layer, and the node features of each layer are fused to obtain the global state features of the multi-layer cooperative graph model. The global state features are input into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete the connection edge between the target node and any adjacent layer node of the target node for the selected target node. Based on the real-time action information, the connection relationships between adjacent layer nodes in the multi-layer cooperative graph model are adjusted to obtain an updated multi-layer cooperative graph model. Based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperative graph model, the multi-UAV system performs task allocation operations.

[0007] According to the multi-layer cooperative graph multi-UAV task allocation method provided by the present invention, the step of performing feature encoding on the nodes of each layer in the multi-layer cooperative graph model to obtain the node features of each layer includes: For each task node in the task node layer, the first preset prompt word guides the large language model to extract semantic features from the task description information corresponding to the task node, thereby obtaining the node features of the task node. For each cooperative cluster node in the cooperative cluster node layer, a topology graph is constructed based on the communication topology relationship between each UAV node in the cooperative cluster node. A graph neural network is used to extract the topology relationship features of the topology graph to obtain the node features of the cooperative cluster node. For each skill node in the skill node layer, the second preset prompt word guides the large language model to extract semantic features from the skill description information corresponding to the skill node, thereby obtaining the node features of the skill node; For each UAV node in the UAV node layer, the real-time status information corresponding to the UAV node in the real-time status information set is feature-encoded using a feature coding network to obtain the node features of the UAV node.

[0008] According to the multi-UAV task allocation method of the multi-layer cooperative graph provided by the present invention, the step of feature fusion of node features of each layer to obtain the global state features of the multi-layer cooperative graph model includes: By using a graph transformation network to aggregate node features between any two adjacent layers, multiple inter-layer aggregated features are obtained. The inter-layer aggregated features include the relationship features between task nodes in the task node layer and cooperative cluster nodes in the cooperative cluster node layer, the relationship features between cooperative cluster nodes and skill nodes in the skill node layer, and the relationship features between skill nodes and drone nodes in the drone node layer. The global state features are obtained by concatenating multiple inter-layer aggregated features.

[0009] According to the multi-UAV task allocation method of the multi-layer cooperative graph provided by the present invention, the step of inputting the global state features into the decision network to obtain the real-time action information of each graph operation operator includes: For each graph operation operator, the global state features are input into the policy subnetwork corresponding to the graph operation operator in the decision network to obtain the set of probability distributions output by the policy subnetwork. Based on the set of probability distributions, determine the index of the target node selected by the graph operation operator; Based on the index of the target node and the operation type corresponding to the target node, obtain the real-time action information of the graph operation operator; The probability distribution set includes the probability values ​​of each node in the target node layer corresponding to the graph operation operator being selected as the target node; each strategy sub-network corresponds to an operation type, which includes adding or deleting connection edges.

[0010] According to a multi-unmanned aerial vehicle (UAV) task allocation method for a multi-layer cooperative graph provided by the present invention, the training steps of the decision network include: Multiple evolutionary training phases are constructed; the multiple evolutionary training phases include a single-type task training phase, a multi-task coordination training phase, and a dynamic task expansion training phase; For the current evolution training stage, if the duration for which the execution performance value of the current evolution training stage reaches the preset convergence threshold is longer than the preset duration, then switch to the next evolution training stage, and use the network parameters trained in the current evolution training stage as the initialization parameters for the next evolution training stage for incremental training. In the single-type task training phase, the scale of drones in the multi-UAV system participating in the training is a first preset scale, and the training scenario includes single-type task description information. In the multi-task coordination training phase, the number of drones in the sample multi-UAV system is a second preset number, and the training scenario includes various types of task description information; the second preset number is larger than the first preset number. In the dynamic task expansion training phase, the number of drones in the sample multi-UAV system is a third preset size, and the training scenario includes at least one task description information randomly added to the training scenario in the multi-task coordination training phase; the third preset size is greater than or equal to the second preset size.

[0011] According to the multi-UAV task allocation method of the multi-layer cooperative graph provided by the present invention, the total reward value of the decision network is obtained by fusing the task completion rate reward value, load balancing reward value and task cost reward value during the training process; The task completion rate reward value is determined based on the total number of tasks completed in the simulation environment and the preset total number of tasks; The load balancing reward value is determined based on the number of drone nodes connected to each cooperative cluster node in the simulation environment, as well as the number of drone nodes deployed in the drone node layer and the number of cooperative cluster nodes deployed in the cooperative cluster node layer in the simulation environment. The task cost reward value is determined based on the simulation resource consumption cost incurred by each drone node in the drone node layer of the simulation environment when the drone performs its assigned task.

[0012] According to a multi-unmanned aerial vehicle (UAV) task allocation method based on a multi-layer cooperative graph provided by the present invention, the step of performing task allocation operations on the multi-UAV system based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperative graph model includes: Based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperation graph model, the task execution parameters corresponding to each UAV node in the UAV node layer are determined; the task execution parameters include the skill description information of the skill node corresponding to each UAV node, the cluster information of the cooperation cluster node to which each UAV node belongs, and the task description information of the task node corresponding to each UAV node. Based on the task execution parameters and the correspondence between each UAV node and each UAV in the multi-UAV system, the real-time task allocation information of each UAV in the multi-UAV system is determined. Based on the real-time task allocation information of each UAV, a task allocation operation is performed on each UAV.

[0013] The present invention also provides a multi-unmanned aerial vehicle (UAV) task allocation system with a multi-layered cooperative graph, comprising: The model building unit is used to obtain a multi-layer cooperation graph model based on the real-time status information set of the multi-UAV system, the task description information and the skill description information; the multi-layer cooperation graph model includes a task node layer, a cooperation cluster node layer, a skill node layer and a UAV node layer; The feature extraction unit is used to encode the features of the nodes in each layer of the multi-layer cooperative graph model to obtain the node features of each layer, and to fuse the node features of each layer to obtain the global state features of the multi-layer cooperative graph model. The decision unit is used to input the global state features into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete the connection edge between the target node and any adjacent layer node of the target node for the selected target node. The allocation unit is used to adjust the connection relationship between adjacent layer nodes in the multi-layer cooperative graph model according to the real-time action information to obtain an updated multi-layer cooperative graph model, and to perform task allocation operation on the multi-UAV system according to the connection relationship between adjacent layer nodes in the updated multi-layer cooperative graph model.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-UAV task allocation method of any of the above-described multi-layer cooperative graphs.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-UAV task allocation method of any of the above-described multi-layer cooperative graphs.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a multi-UAV task allocation method as described above for a multi-layer cooperative graph.

[0017] The present invention provides a multi-layer cooperative graph multi-UAV task allocation method, system, device and medium. By constructing a multi-layer cooperative graph model that includes tasks, cooperative clusters, skills and UAV nodes, and using decision network-driven graph operation operators to incrementally add and delete inter-layer connection edges, the traditional global action optimization based on fixed label space is transformed into a dynamic reconstruction process based on graph topology evolution. This significantly improves the robustness of multi-UAV systems in dealing with complex dynamic environments, while effectively improving the convergence efficiency of decision-making through fine-grained adjustment of connection relationships. Attached Figure Description

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

[0019] Figure 1 This is one of the flowcharts illustrating the multi-UAV task allocation method for multi-layer cooperative graphs provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of the multi-layered cooperation graph model provided by the present invention.

[0021] Figure 3 This is the second flowchart illustrating the multi-UAV task allocation method for multi-layer cooperative graphs provided by this invention.

[0022] Figure 4 This is a schematic diagram of the structure of the multi-unmanned aerial vehicle (UAV) task allocation system with a multi-layered cooperative graph provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] All actions involving the acquisition of signal information or data in this application are carried out in accordance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0026] With the rapid development of UAV technology, multi-UAV systems have demonstrated enormous application potential in complex scenarios such as military reconnaissance, disaster relief, area surveillance, communication relay, and coordinated strikes. In these complex and dynamic application scenarios, tasks often exhibit characteristics of multi-type operations, strong heterogeneity, and high concurrency. For example, they may require the simultaneous execution of multiple tasks such as coordinated search, area surveillance, communication relay, and coordinated strikes. Efficient task allocation and scheduling are crucial for maximizing the operational effectiveness of multi-UAV systems. The goal is to rationally allocate UAV resources and capabilities in dynamic and uncertain environments to optimize task completion rates, timeliness, and overall efficiency.

[0027] Existing multi-UAV task allocation technologies typically utilize specific optimization algorithms or fixed rules to perform action optimization and decision output within a predefined, fixed task label space. While this technology can achieve certain results in specific scenarios, its computational burden increases dramatically with the complexity and scale of the tasks, and it struggles to adapt to sudden tasks and dynamic environmental changes.

[0028] Specifically, existing technologies face several key challenges: First, they lack the ability to generalize modeling across multiple task types. Because existing methods are often designed for single or predefined task types, there is a strong coupling between model parameters and task structure. When faced with new or complex tasks described in text, they cannot understand the task semantics, making rapid adaptation difficult and often requiring model redesign or retraining, resulting in poor generalization ability. Second, they lack dynamic adjustment and coordination capabilities. Most existing methods treat UAVs as homogeneous individuals or possess only limited static skills, failing to explicitly model their diverse optional skills and dynamic grouping and collaboration relationships based on skills. When some UAVs are damaged, new tasks are introduced, or the environment changes abruptly, they lack the ability to quickly reconstruct collaboration relationships and reallocate skills and tasks, resulting in poor robustness. Third, they lack the ability to understand human intent. Commanders typically issue task instructions using natural language. Existing technologies struggle to directly parse such high-level, ambiguous intents and translate them into executable decision logic, leading to significant semantic inconsistencies in the decision-making process and making UAV behavior difficult to predict and intervene in.

[0029] To address the shortcomings of existing technologies, this application provides a multi-layer cooperation graph method for multi-UAV task allocation. This method constructs a multi-layer cooperation graph model (MLCGM) to formally decompose the task allocation problem into a dynamic connection optimization problem between different node levels. It utilizes learnable graph operation operators to explicitly manage the associations and reconstructions between layers, achieving fine-grained modeling and efficient adjustment of complex collaborative relationships. Simultaneously, a large language model (LLM) is introduced as a core component for semantic understanding and representation fusion. This deeply integrates the commander's natural language task description with the real-time environment and UAV status, generating a joint representation that drives decision-making, enabling the understanding of new tasks and zero-shot generalization. Furthermore, by employing a hierarchical course learning strategy, progressive training is achieved from simple single-task, small-scale scenarios to complex multi-task, large-scale scenarios, stably and efficiently improving the performance and generalization ability of the policy network. This method significantly enhances the task autonomy, collaborative efficiency, and overall robustness of multi-UAV systems in dynamic open environments, providing core technical support for fully autonomous task collaboration in multi-UAV systems.

[0030] Figure 1 This is one of the flowcharts illustrating a multi-UAV task allocation method for a multi-layered cooperative graph provided by the present invention.

[0031] It should be noted that the multi-UAV task allocation method with multi-layer cooperative graph provided in this embodiment can be widely applied to various highly dynamic and complex task requirements scenarios, including but not limited to urban emergency response, battlefield coordination, and other highly dynamic and complex task requirements scenarios.

[0032] like Figure 1 As shown, the method includes steps 110, 120, 130 and 140.

[0033] Step 110: Based on the real-time status information set of the multi-UAV system, the task description information, and the skill description information, obtain a multi-layer cooperation graph model; the multi-layer cooperation graph model includes a task node layer, a cooperation cluster node layer, a skill node layer, and a UAV node layer.

[0034] Here, a Multi-Unmanned Aerial Vehicle System (MUAVS) refers to a collaborative swarm of multiple unmanned aerial vehicles (UAVs) with communication, sensing, and operational capabilities. The real-time status information set encompasses the real-time status information of each UAV in the MUAV system during the current period. This real-time status information may include real-time physical motion parameters and real-time energy consumption parameters, specifically including but not limited to the UAV's identifier, position coordinates, flight speed, current mission load (or mission status), and remaining battery power. For MUAV systems of different sizes, the data dimensions of the real-time status information set can be dynamically adjusted to adapt to different application scenarios, from small formations to large-scale swarms.

[0035] Mission description information comprises the functional definitions and requirement constraints of a complex mission to be performed by a multi-UAV system. These complex missions include, but are not limited to, cooperative search missions, area surveillance missions, communication support missions, and cooperative strike missions. Mission description information can be organized in natural language text format, facilitating real-time expansion and adjustment based on mission instructions.

[0036] Skill description information refers to the definition of the basic operational capabilities that a UAV can perform, which can be expanded and adjusted according to actual needs. Here, basic operational capabilities include, but are not limited to, autonomous navigation skills, target tracking skills, network maintenance skills, formation movement skills, and target engagement skills. Among them, autonomous navigation skills are the skills to move from a designated starting point to a designated ending point while maintaining collision avoidance; target tracking skills are the skills to continuously track a given target at a certain distance; network maintenance skills are the skills to establish communication relationships with neighboring UAVs within a given communication radius; formation movement skills are the skills to maintain a fixed relative position with neighboring UAVs at all times; and target engagement skills are the skills to launch weapons at targets at a designated distance.

[0037] Figure 2 This is a schematic diagram of the structure of the multi-layer cooperation graph model provided by the present invention; as shown below. Figure 2 As shown, the multi-layered cooperative graph model is a mathematical model that formalizes the dynamic task allocation problem as a graph structure topology optimization. This model is divided vertically into four functionally related node levels: the task node level, the cooperative cluster node level (also known as the cluster node level), the skill node level, and the drone node level.

[0038] Among them, the task node layer Located at the top layer of the model, each task node in the task node layer represents a complex task that needs to be performed, including but not limited to cooperative search, area surveillance, communication support, and cooperative strike, which can be expanded and adjusted according to actual needs.

[0039] Cooperative Cluster Node Layer Located below the task node layer, each cooperative cluster node (also called a cluster node) in the cooperative cluster node layer represents the cooperative relationship of a group of drones (also called a cooperative cluster) formed to perform a specific objective. A cooperative cluster node in the cooperative cluster node layer refers to a group of drones that have been selected and bound with specific skills to perform a specific objective. The generation of a cooperative cluster for each cooperative cluster node in the cooperative cluster node layer depends on the state of the drones and the matching degree between skills and tasks. By establishing cooperative cluster nodes, the cooperative structure between drones can be explicitly managed.

[0040] Skill Node Layer Located below the cooperative cluster node layer, each skill node in the skill node layer corresponds to an atomic operation capability, that is, the specific skill item corresponding to the skill description information mentioned above, including but not limited to autonomous navigation, target tracking, network maintenance, target strike, etc., which can be expanded and adjusted according to implementation requirements.

[0041] Drone node layer Located at the bottom layer, each drone node in the drone node layer corresponds to a drone entity in the physical scene.

[0042] Optionally, after obtaining the real-time status information set of the multi-UAV system, the task description information, and the skill description information, a multi-UAV task allocation model with a four-layer node structure, i.e., a multi-layer cooperative graph model, can be established based on this information. This multi-layer cooperative graph model is a dynamically variable graph structure used to formally describe the mapping relationship between tasks, cooperative organizations, skills, and physical UAVs.

[0043] Specifically, multi-layered collaboration graph models include: Construct the UAV node layer: Instantiate the corresponding UAV node for each UAV in the UAV system, and use the real-time status information of each UAV in the real-time status information set as the original attribute information of its corresponding UAV node.

[0044] Construct a skill node layer: Instantiate the corresponding skill node for each skill description information, and use each skill description information as the original attribute information of the skill node corresponding to each skill description information.

[0045] Constructing a cooperative cluster node layer: Several cooperative cluster nodes are generated based on the task complexity and drone scale of the multi-UAV system. As an intermediate logical layer connecting skills and tasks, the cooperative cluster node initially lacks specific physical attributes; it serves only as a convergence center for topological connections, representing the dynamic combination of drones and skills. The initial connection topology information of the drone nodes within each cooperative cluster node is used as the original attribute information of each cooperative cluster node.

[0046] Construct the task node layer: Instantiate the corresponding task node for each task description information, and use each task description information as the original attribute information of the task node corresponding to each task description information.

[0047] Subsequently, after instantiating the four layers of nodes, an initial set of connection edges can be established between the UAV node layer and the skill node layer, between the skill node layer and the cooperative cluster node layer, and between the cooperative cluster node layer and the task node layer, based on preset initialization rules or historical state memory. Specifically, for the current cycle, if a historical task allocation scheme from the previous cycle already exists in the multi-UAV system, the corresponding nodes are connected according to that scheme; if the multi-UAV system is in its initial startup state, i.e., there is no historical task allocation scheme from the previous cycle, the connection relationship can be set to an empty set or a fully connected or random connection can be established according to preset rules, awaiting dynamic adjustment by the subsequent decision network.

[0048] Finally, the four types of instantiated node sets and their initialized connecting edges are encapsulated into a unified graph object, resulting in a multi-layered cooperative graph model. This model is stored in memory as a dynamic graph structure, supporting efficient node addition, deletion, and edge modification operations.

[0049] In summary, the multi-layered collaborative graph model constructed through the above steps enables highly unified modeling of multi-UAV task allocation problems across physical space, capability space, collaborative space, and task space. Because this model employs a hierarchical graph structure, it possesses inherent scalability. When introducing new UAVs, new skill types, or new task objectives, only nodes need to be added to the corresponding node layers, and potential connecting edges between adjacent layers need to be established. There is no need to redesign the entire model architecture. Therefore, this model can adapt to the expansion and adjustment of the number of UAVs, UAV skills, and UAV tasks. It provides a highly flexible and scalable data structure foundation for solving the dynamic resource scheduling of large-scale UAV systems in heterogeneous task environments, and provides a solid foundation for achieving adaptive task allocation of multiple UAVs in highly dynamic environments. Furthermore, by transforming task allocation operations into the topological evolution of edge sets in the graph model, it is possible to decouple and finely manage complex collaborative processes.

[0050] Step 120: Encode the features of the nodes in each layer of the multi-layer cooperative graph model to obtain the node features of each layer, and fuse the node features of each layer to obtain the global state features of the multi-layer cooperative graph model.

[0051] Optionally, after obtaining the multi-layer cooperation graph model, the original attribute information of the nodes in each layer of the multi-layer cooperation graph model can be extracted to obtain the node features of each layer.

[0052] It should be noted that, since the original attribute information of nodes in each layer of the multi-layer collaborative graph model contains multiple modalities such as semantic information, topological structure information, and physical state information, in order to effectively process the original attribute information of different modalities, a differentiated encoding mechanism can be used to extract features from nodes in each layer.

[0053] For example, in one possible implementation, feature encoding is performed on the nodes of each layer in the multi-layer collaboration graph model to obtain the node features of each layer, including: For each task node in the task node layer, the first preset prompt word guides the large language model to extract semantic features from the task description information corresponding to the task node, thereby obtaining the node features of the task node. For each cooperative cluster node in the cooperative cluster node layer, a topology graph is constructed based on the communication topology relationship between each UAV node in the cooperative cluster node. A graph neural network is used to extract the topology relationship features of the topology graph to obtain the node features of the cooperative cluster node. For each skill node in the skill node layer, the second preset prompt word guides the large language model to extract semantic features from the skill description information corresponding to the skill node, thereby obtaining the node features of the skill node; For each UAV node in the UAV node layer, the real-time status information corresponding to the UAV node in the real-time status information set is feature-encoded using a feature coding network to obtain the node features of the UAV node.

[0054] Figure 3 This is the second flowchart illustrating the multi-UAV task allocation method for multi-layered cooperative graphs provided by the present invention; like Figure 3 As shown, before outputting the real-time action information of each graph operation operator through the decision network, the nodes of each layer in the multi-layer cooperative graph model can be pre-extracted and encoded.

[0055] Specifically, for the task node layer To achieve deep understanding and generalization of different natural language description tasks, a large language model can be used to extract semantic features from the original attribute information of each task node, i.e., the task description information. Specifically, for each task node in the task node layer... , and its corresponding task description information The input is fed into the large language model, and the first preset prompt word guides the large language model to output the task description information. Encoding as fixed-length task features That is, task node The node features are calculated using the following formula: ; The first preset prompt here is a preset text instruction used to guide the large language model in parsing task features such as core actions, collaboration scale requirements, and spatial constraints.

[0056] For cooperative cluster node layer Its coding focuses on extracting the interaction relationships between members within the collaborative group. Specifically, for each collaborative cluster node in the collaborative cluster node layer... First, a topology graph is constructed based on the communication topology relationships between the UAV nodes within the cooperative cluster, resulting in an undirected graph structure. Subsequently, a graph neural network (GNN) was used to analyze this undirected graph structure. To extract and encode the collaborative association features between nodes, this undirected graph structure is used. Encoded as fixed-length cluster features That is, cooperative cluster nodes The node features are calculated using the following formula: .

[0057] For skill node layer To achieve a deep understanding and generalization of skills described in different natural language, a large language model can also be used to extract semantic features from the original attribute information of each skill node, i.e., the skill description information. Specifically, for each skill node in the skill node layer... The corresponding skill description information The input is fed into a large language model, and the second preset prompt word guides the large language model to output skill description information. Skill features encoded as fixed length That is, skill nodes The node features are calculated using the following formula: ; The second preset prompt is a preset text instruction used to guide the large language model to understand the technical connotations and execution logic of skills such as autonomous navigation and target strike.

[0058] For drone node layer The encoding of each drone node aims to extract the state features of each drone. Specifically, for each drone node... First, the location coordinates, flight speed, current task load (or task status), and remaining battery power of the corresponding drone node in the real-time status information set can be concatenated into a vector. Then, feature encoding networks (such as fully connected neural networks) can be used. For vectors Perform a non-linear encoding mapping to encode it into a fixed-length feature vector. This yields the state characteristics reflecting the current operational status of the drone, i.e., the drone node. The node features. The feature encoding network is a fully connected neural network. At that time, drone nodes The specific calculation formula for the node features is as follows: .

[0059] Therefore, by encoding the features of each node in each layer of the multi-layer cooperative graph model through the aforementioned feature extraction method, the feature encoding of each node in the multi-layer cooperative graph model can be completed, which is also the abstract expression of the node features of each node in the multi-layer graph structure. In this process, by selectively using a large language model to parse the semantic features of tasks and skills, and combining it with a graph neural network to capture the topological relationships of cooperative clusters, this solution effectively eliminates the semantic gap between the commander's high-level intentions and the underlying physical execution logic, significantly improving the understanding depth of multi-UAV systems for complex heterogeneous tasks and their zero-shot generalization ability across scenarios.

[0060] It should be noted that the large language model mentioned in this embodiment is actually an existing pre-trained language model that has been pre-trained on massive general corpora, such as the Bidirectional Encoder Representations from Transformers (BERT) model.

[0061] Subsequently, after obtaining the node features of each layer, feature fusion can be performed on the node features of each layer to obtain global state features. Feature fusion here refers to the interaction and integration of feature information scattered across each layer of nodes to form a comprehensive representation vector that can characterize the current environmental situation, mission requirements, resource distribution, and existing cooperative relationships of the entire multi-UAV system.

[0062] For example, in one possible implementation, the feature fusion of node features from each layer to obtain the global state features of the multi-layer cooperative graph model includes: By using a graph transformation network to aggregate node features between any two adjacent layers, multiple inter-layer aggregated features are obtained. The inter-layer aggregated features include the relationship features between task nodes in the task node layer and cooperative cluster nodes in the cooperative cluster node layer, the relationship features between cooperative cluster nodes and skill nodes in the skill node layer, and the relationship features between skill nodes and drone nodes in the drone node layer. The global state features are obtained by concatenating multiple inter-layer aggregated features.

[0063] Optionally, to extract the interaction relationships between nodes in different layers, a graph transform network (also known as a graph transformer network) can be used to aggregate the node features of any two adjacent layers to obtain three inter-layer aggregated features: ; ; ; in, This represents the aggregation of node features for all task nodes. This represents the aggregation of node characteristics for all cooperative cluster nodes. This represents the aggregation of node features for all skill nodes. This represents the aggregation of node features for all drone nodes. This represents the relationship characteristics between task nodes and cooperative cluster nodes. This represents the relationship characteristics between cooperative cluster nodes and skill nodes. This indicates the relationship characteristics between skill nodes and drone nodes.

[0064] Furthermore, by concatenating the above three inter-layer aggregation features, the global state features of the entire multi-layer cooperation graph can be obtained. Extraction: .

[0065] In summary, the cross-layer aggregation mechanism of graph transformation networks can accurately characterize the complex coupling relationships between task requirements, collaborative organization, operational capabilities, and entity resources. The generated global state features not only include the independent attributes of nodes at each layer but also contain the constraints and interaction logic between layers. This fusion scheme provides multi-dimensional, high-fidelity information input for the subsequent decision network to accurately output graph operation instructions, ensuring the global consistency and scientific nature of task allocation decisions in highly dynamic environments.

[0066] Here, feature fusion can capture the inter-layer correlation information between task nodes and cooperative cluster nodes, between cooperative cluster nodes and skill nodes, and between skill nodes and drone nodes. These heterogeneous features are mapped to a unified feature space and converged, enabling the simultaneous understanding of the commander's high-level intentions, the mid-level collaborative topology, and the low-level physical constraints, providing multi-dimensional information support for decision-making.

[0067] Step 130: Input the global state features into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete the connection edge between the target node and any adjacent layer node of the target node for the selected target node.

[0068] Optionally, the multi-UAV task allocation problem can be transformed into a problem of adjusting the connection edges of multi-layer graph nodes through the modeling in step 110. To achieve real-time adjustment of the connection relationships between nodes at different levels, corresponding graph operation operators can be added for each adjacent layer to dynamically adjust the connection relationships between adjacent layers. These graph operation operators are used to adjust the connection relationships between nodes at different levels in the multi-layer cooperative graph model in real time. Their core logic lies in achieving flexible resource allocation through the dynamic addition and deletion of connection edges; that is, by setting operations to add and delete connection edges, flexible allocation and release of resources are achieved. Adding a connection edge means establishing a new allocation relationship, while deleting a connection edge breaks rigid or inefficient cooperative relationships, allowing resources to return to an idle state. This operator-based dynamic adjustment mechanism can quickly reorganize the global task allocation scheme by adding and deleting local edges when facing sudden tasks or UAV malfunctions.

[0069] like Figure 2 As shown, based on the different hierarchical relationships in which graph operations are applied, graph operations can be divided into the following three categories: The first category is skill selection operators; this category of skill selection operators includes two operators. and Operators used to adjust the connection relationship between drone nodes and skill nodes. and The action space is Among them, operators Used from Select one skill node from the skill nodes. And randomly delete one of the drone nodes connected to it (that is, randomly delete one drone node and skill node). (Existing connected edges) to release the drone from its current skill, returning it to a free resource state; operator Used from Select one skill node from the skill nodes. And randomly add a new drone node to it (that is, randomly add a drone node and a skill node). (New connection edge), enabling the assignment of specific atomized tasks to drones. Based on this, through operators and Different values ​​can be assigned to dynamically adjust the connection edges between drone nodes and skill nodes.

[0070] The second category is skill assembly operation operators; this category of skill assembly operation operators includes two operation operators. and Operators used to adjust the connection relationship between skill nodes and cooperative cluster nodes. and The action space is Among them, operators Used from Select one cluster node from the cluster nodes. And randomly delete one of the skill nodes connected to it (randomly delete one skill node and cluster node). (existing connected edges), operator Used from Select one cluster node from the cluster nodes. And randomly add a new skill node (randomly add a skill node and cluster node) to it. (New connecting edges), thereby completing the dynamic adjustment of connecting edges between skill nodes and cluster nodes.

[0071] The third category is task allocation operators; this category of task allocation operators includes two operators. and Operators are used to adjust the connection relationship between cluster nodes and task nodes. and The action space is Among them, the operator Used from Select a task point from the task nodes. And randomly delete one cluster node connected to it (randomly delete one cluster node and task node). (existing connected edges), operator Used from Select a task point from the task nodes. And randomly add a new cluster node to it (randomly add a cluster node and task node) (New connection edges), thereby completing the dynamic adjustment of the connection edges between cluster nodes and task nodes.

[0072] The Decision Network (DN) receives the global state features extracted from the above steps and maps them to topology adjustment actions for the multi-layer cooperative graph model. It is responsible for outputting the optimal decision instructions, i.e., real-time action information, for each graph operation operator (GOO).

[0073] Optionally, after obtaining the global state features of the multi-layer cooperative graph model, these features can be input into a decision network. The decision network then provides the optimal decision instruction for each graph operation operator, thereby obtaining the real-time action information of each graph operation operator. That is, the output of the policy network at each time step is the optimal action currently selected by the operation operator.

[0074] In some embodiments, the decision network may consist of multiple layers of policy subnetworks with independent parameters (such as fully connected neural networks), each policy subnetwork being configured for each operator. Figure 3 As shown, when the policy subnetwork is a fully connected neural network, considering that there are a total of six graph operation operators, the decision network is configured with six independent fully connected (FC) neural networks. Through this parameter-independent network architecture, each operation operator can share global situation information while maintaining the independence of decision logic, thereby capturing the unique dynamic features of its own level more accurately.

[0075] Accordingly, the step of inputting the global state features into the decision network to obtain the real-time action information of each graph operation operator includes: For each graph operation operator, the global state features are input into the policy subnetwork corresponding to the graph operation operator in the decision network to obtain the set of probability distributions output by the policy subnetwork. Based on the set of probability distributions, determine the index of the target node selected by the graph operation operator; Based on the index of the target node and the operation type corresponding to the target node, obtain the real-time action information of the graph operation operator; The probability distribution set includes the probability values ​​of each node in the target node layer corresponding to the graph operation operator being selected as the target node; each strategy sub-network corresponds to an operation type, which includes adding or deleting connection edges.

[0076] like Figure 3 As shown, the fused global state features can be input into the policy subnetworks corresponding to each graph operation operator in the decision network. Through nonlinear transformation of the policy subnetworks, the probability distribution set output by each policy subnetwork is obtained. Subsequently, based on these probability values, the decision network selects the index of a node as the index of the target node selected by each graph operation operator, and obtains the real-time action information of each graph operation operator by combining the index of the target node and the operation type corresponding to the target node.

[0077] For example, when the policy subnetwork is a fully connected neural network, the specific calculation formula for the real-time action information of each graph operation operator is as follows: ; ; ; ; ; ; in This represents the parameters of a fully connected neural network. This represents the real-time action information output by the policy subnetwork of 6 operators for each operator.

[0078] Here, by inputting global state features in parallel to policy subnetworks with independent parameters, deep decoupling and efficient parallel processing of decision logic of graph operation operators at each level are achieved. While ensuring accurate action assignment for different level nodes and operation types, this greatly improves the efficiency of real-time action generation and the flexibility of resource scheduling and reconfiguration in highly dynamic environments.

[0079] Step 140: Based on real-time action information, adjust the connection relationships between adjacent layer nodes in the multi-layer cooperation graph model to obtain an updated multi-layer cooperation graph model, and perform task allocation operations on the multi-UAV system based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperation graph model.

[0080] Optionally, after obtaining the real-time action information of each graph operation operator, the topology of the multi-layer cooperative graph model can be updated based on the real-time action information.

[0081] Specifically, based on the decision results of each graph operation operator, operations such as adding or deleting connecting edges are performed between the corresponding node levels, thereby evolving the topology of the multi-layer cooperative graph model and obtaining an updated multi-layer cooperative graph model that reflects the optimal allocation scheme at the current moment. This update process is a high-frequency iterative process, ensuring that the task allocation scheme can keep up with changes in the battlefield situation in real time.

[0082] After obtaining the updated multi-layered cooperation graph model, the abstract logical topology connections need to be transformed into physically executable instructions to achieve closed-loop control of the multi-UAV system.

[0083] For example, in one possible implementation, the step of performing task allocation operations on the multi-UAV system based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperation graph model includes: Based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperation graph model, the task execution parameters corresponding to each UAV node in the UAV node layer are determined; the task execution parameters include the skill description information of the skill node corresponding to each UAV node, the cluster information of the cooperation cluster node to which each UAV node belongs, and the task description information of the task node corresponding to each UAV node. Based on the task execution parameters and the correspondence between each UAV node and each UAV in the multi-UAV system, the real-time task allocation information of each UAV in the multi-UAV system is determined. Based on the real-time task allocation information of each UAV, a task allocation operation is performed on each UAV.

[0084] Optionally, during closed-loop control of a multi-UAV system, the updated multi-layered cooperation graph model can be traversed to trace the complete path from each UAV node at the bottom layer, upwards through the skill node layer, the cooperation cluster node layer, and finally to the task node layer. Based on this path, the functional attributes and organizational affiliation of each UAV node can be identified to obtain the corresponding task execution parameters. These task execution parameters include: the skill description information of the skill node corresponding to each UAV node (i.e., the atomic operation capabilities that the UAV currently needs to invoke), the cluster information of the cooperation cluster node to which each UAV node belongs (i.e., the cooperation group number to which the UAV belongs and the communication topology of other members within the group), and the task description information of the task node that each UAV node ultimately points to through the path (i.e., the global task objective that the UAV ultimately serves).

[0085] Subsequently, using a node-to-entity mapping table, the task execution parameters of the logical drone nodes are mapped to each physical drone in the physical scene, obtaining real-time task allocation information for each drone. This mapping ensures that the complex collaborative logic generated in the graph model can be accurately transmitted to the corresponding hardware platform. The real-time task allocation information here is the customized set of operation instructions generated for each specific drone.

[0086] Finally, real-time task allocation information is sent to the onboard computing unit of each UAV in the form of data packets via wireless communication link. After receiving the instructions, the UAV can call its flight control system and mission payload system to carry out actual operations according to the assigned skills, cooperation relationships and mission objectives.

[0087] Here, by deeply parsing the updated graph topology into task execution parameters that include skills, cluster information, and task semantics, not only is the richness and accuracy of the assigned instructions guaranteed, but the cluster can also perfectly reproduce the complex collaborative intent constructed by the graph model at the physical execution level. This path-tracing-based parameter determination method enables each UAV to clearly know its position in the global task and its collaborative relationship with friendly UAVs, significantly improving the collaborative accuracy and operational efficiency of multi-UAV systems when performing large-scale heterogeneous tasks.

[0088] The method provided in this embodiment constructs a multi-layer cooperative graph model that includes tasks, cooperative clusters, skills, and UAV nodes, and uses decision network-driven graph operation operators to incrementally add, delete, and adjust the inter-layer connection edges in real time. This transforms the traditional global action optimization based on a fixed label space into a dynamic reconstruction process based on graph topology evolution, thereby significantly improving the robustness of multi-UAV systems in dealing with complex dynamic environments. At the same time, it effectively improves the convergence efficiency of decision-making through fine-grained adjustment of connection relationships.

[0089] Based on the above embodiments, in order to efficiently and stably train the policy network of graph operation operators... A tiered learning strategy can be adopted, based on the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm for training, to decompose the complex allocation task into multiple stages from easy to difficult, guiding the model to gradually master the collaboration logic.

[0090] The specific training steps include: Multiple evolutionary training phases are constructed; the multiple evolutionary training phases include a single-type task training phase, a multi-task coordination training phase, and a dynamic task expansion training phase; For the current evolution training stage, if the duration for which the execution performance value of the current evolution training stage reaches the preset convergence threshold is longer than the preset duration, then switch to the next evolution training stage, and use the network parameters trained in the current evolution training stage as the initialization parameters for the next evolution training stage for incremental training. In the single-type task training phase, the scale of drones in the multi-UAV system participating in the training is a first preset scale, and the training scenario includes single-type task description information. In the multi-task coordination training phase, the number of drones in the sample multi-UAV system is a second preset number, and the training scenario includes various types of task description information; the second preset number is larger than the first preset number. In the dynamic task expansion training phase, the number of drones in the sample multi-UAV system is a third preset size, and the training scenario includes at least one task description information randomly added to the training scenario in the multi-task coordination training phase; the third preset size is greater than or equal to the second preset size.

[0091] Optionally, multiple evolutionary training phases can be constructed first. At the training scenario level, these evolutionary training phases gradually increase task complexity and resource scale, enabling the decision network to smoothly learn strategies from basic collaboration to complex dynamic reconstruction. These evolutionary training phases include multiple training phases with progressively increasing complexity: a single-type task training phase, a multi-task coordination training phase, and a dynamic task expansion training phase.

[0092] For the single-task training phase, the training scenario contains only task description information for a single task type (e.g., only area surveillance), and the number of drones in the multi-UAV system participating in the training is small, such as less than 5. The goal of this phase is to enable the decision network to learn the most basic graph topology adjustment logic, that is, how to establish connections between drones and specific skills and simple tasks, and to initially establish cooperative awareness among drones.

[0093] For the multi-task coordination training phase, the scale of the multi-UAV systems participating in the training is increased to a second preset scale, for example, the number of UAVs is increased to more than 20, and various types of task description information are introduced into the training scenario, such as simultaneous collaborative search and communication support tasks. At this time, the decision network needs to learn how to optimize resource allocation in a multi-task concurrent environment and solve the problems of resource competition and conflict resolution between different tasks.

[0094] For the dynamic task-expanded training phase, the scale of the multi-UAV systems participating in the training is further expanded to a third preset scale. The training scenarios in this phase, based on the multi-task coordination training phase, will randomly add at least one task description or simulate abnormal situations such as UAV damage or sudden communication link interruptions. By forcing the model to simulate UAV failures or sudden environmental changes during task execution, the policy learning is compelled to quickly reconstruct graph connections to cope with dynamic changes.

[0095] Secondly, the system implements phase switching and incremental training. Execution performance values ​​(such as average task completion rate or total reward value) are monitored in real time during each evolutionary training phase. If the execution performance value of the current phase reaches a preset convergence threshold, such as an average task completion rate exceeding 90% and a duration longer than a preset duration, the model is deemed to have fully mastered the allocation strategy for the current complexity. At this point, the system automatically switches to the next evolutionary training phase with higher complexity, reloading the network parameters trained in the current phase as the initialization parameters for the next phase. This incremental training mode, by using neural network parameters trained in simple task scenarios as initialization conditions for the next complex task, and through parameter reloading, achieves the reuse of learned knowledge and ensures the stability of learned knowledge. It avoids the waste of computational resources caused by training from scratch and ensures that the learned collaborative knowledge can be stably transferred to more complex scenarios.

[0096] The method provided in this embodiment effectively solves the problems of slow training convergence and large policy fluctuations in deep reinforcement learning when facing large-scale heterogeneous task allocation through a hierarchical course learning strategy. Furthermore, through a progressive task complexity evolution and parameter reloading mechanism, it significantly improves the generalization performance and training efficiency of decision networks in handling complex dynamic environments.

[0097] Based on the above embodiments, in order to better guide the model to learn and generate allocation strategies that meet actual engineering needs, this embodiment designs a multi-dimensional comprehensive reward mechanism. In one possible implementation, the total reward value of the decision network is obtained by fusing the task completion rate reward value, load balancing reward value, and task cost reward value during the training process; The task completion rate reward value is determined based on the total number of tasks completed in the simulation environment and the preset total number of tasks; The load balancing reward value is determined based on the number of drone nodes connected to each cooperative cluster node in the simulation environment, and the number of drone nodes deployed in the drone node layer of the simulation environment. The task cost reward value is determined based on the simulation resource consumption cost incurred by each drone node in the drone node layer of the simulation environment when the drone performs its assigned task.

[0098] Optionally, the total reward value of the decision network It can be calculated using the following formula: ; in , , Indicates the weighting coefficient. , , These represent the task completion rate reward value, load balancing reward value, and task cost reward value, respectively, and are defined as follows: in This indicates the number of tasks completed in a single simulation within the simulation environment, i.e., the total number of tasks completed. This indicates the total number of pre-set tasks included in this task batch, which is also the total number of task nodes contained in the task node layer under the simulation environment. The task completion rate reward value, through direct result feedback, incentivizes the decision network to prioritize adjusting the graph topology connections to complete more task objectives, which is the foundation for ensuring the operational efficiency of multi-UAV systems.

[0099] Represents each cooperative cluster node The number of drone nodes connected in the middle, This represents the number of drone nodes deployed in the drone node layer under the simulation environment, which is the total number of all available drone entities participating in the task allocation of the current training phase in the multi-drone system. This represents the total number of cooperative cluster nodes generated in the cooperative cluster node layer under the simulation environment. The load balancing reward value aims to optimize the distribution efficiency of resources and prevent resource polarization, where some drones are overworked while others are idle, thereby effectively improving the overall redundancy of the drone swarm and the long-term stability of the system.

[0100] Indicates drone node Corresponding drone execution task nodes The corresponding simulation resource consumption cost, Indicates drone node Corresponding drones and mission nodes The corresponding task allocation relationship, if the drone node The corresponding drones were assigned to perform the task nodes. The corresponding task is... ,otherwise The simulated resource consumption costs here include, but are not limited to, the UAV's battery energy consumption, flight time required to execute the mission, and communication bandwidth usage. The mission cost reward value, through a negative feedback mechanism, forces the decision-making network to seek a topology connection scheme with higher energy efficiency and lower time overhead while ensuring the mission completion rate.

[0101] The method provided in this embodiment constructs a multi-objective reward function that integrates task completion rate, load balancing, and operation cost. This guides the decision network to effectively avoid the polarized allocation of cluster resources in a multi-UAV system and minimize the ineffective loss of system resources, thereby achieving a globally optimal balance between operational efficiency, resource utilization, and operational economy in complex dynamic scenarios.

[0102] The following describes the multi-UAV task allocation system of the multi-layer cooperative graph provided by the present invention. The multi-layer cooperative graph multi-UAV task allocation system described below and the multi-layer cooperative graph multi-UAV task allocation method described above can be referred to in correspondence.

[0103] Figure 4 This is a schematic diagram of the multi-UAV task allocation system with a multi-layered cooperative graph provided by the present invention; as shown below. Figure 4 As shown, the system includes: The model building unit 410 is used to obtain a multi-layer cooperation graph model based on the real-time status information set of the multi-UAV system, the task description information and the skill description information; the multi-layer cooperation graph model includes a task node layer, a cooperation cluster node layer, a skill node layer and a UAV node layer. The feature extraction unit 420 is used to encode the features of the nodes in each layer of the multi-layer cooperative graph model to obtain the node features of each layer, and to fuse the node features of each layer to obtain the global state features of the multi-layer cooperative graph model. The decision unit 430 is used to input the global state features into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete the connection edge between the target node and any adjacent layer node of the target node for the selected target node. The allocation unit 440 is used to adjust the connection relationship between adjacent layer nodes in the multi-layer cooperative graph model according to the real-time action information to obtain an updated multi-layer cooperative graph model, and to perform task allocation operation on the multi-UAV system according to the connection relationship between adjacent layer nodes in the updated multi-layer cooperative graph model.

[0104] The system provided in this embodiment constructs a multi-layer cooperative graph model that includes tasks, cooperative clusters, skills, and UAV nodes. It uses decision network-driven graph operation operators to incrementally add, delete, and adjust the inter-layer connection edges in real time. This transforms the traditional global action optimization based on a fixed label space into a dynamic reconstruction process based on graph topology evolution. This significantly improves the robustness of multi-UAV systems in dealing with complex dynamic environments. At the same time, it effectively improves the convergence efficiency of decision-making through fine-grained adjustment of connection relationships.

[0105] The system provided by this invention is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0106] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. The processor 510 can call logical instructions in the memory 530 to execute a multi-UAV task allocation method using a multi-layer cooperative graph. This method includes: obtaining a multi-layer cooperative graph model based on a set of real-time state information of the multi-UAV system, task description information, and skill description information; the multi-layer cooperative graph model includes a task node layer, a cooperative cluster node layer, a skill node layer, and a UAV node layer; performing feature encoding on the nodes of each layer in the multi-layer cooperative graph model to obtain node features of each layer, and performing feature fusion on the node features of each layer to obtain global state features of the multi-layer cooperative graph model; inputting the global state features into a decision network to obtain real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete connection edges between the target node and any adjacent layer node of the target node for a selected target node; adjusting the connection relationships between adjacent layer nodes in the multi-layer cooperative graph model according to the real-time action information to obtain an updated multi-layer cooperative graph model, and performing task allocation operations on the multi-UAV system according to the connection relationships between adjacent layer nodes in the updated multi-layer cooperative graph model.

[0107] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-UAV task allocation method with a multi-layer cooperative graph provided by the above methods. This method includes: obtaining a multi-layer cooperative graph model based on a set of real-time status information of the multi-UAV system, task description information, and skill description information; the multi-layer cooperative graph model includes a task node layer, a cooperative cluster node layer, a skill node layer, and a UAV node layer; and performing feature encoding on the nodes of each layer in the multi-layer cooperative graph model to obtain the node features of each layer. Feature fusion is performed on the node features of each layer to obtain the global state features of the multi-layer cooperative graph model; the global state features are input into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete connection edges between the target node and any adjacent layer node of the target node for a selected target node; according to the real-time action information, the connection relationship between adjacent layer nodes in the multi-layer cooperative graph model is adjusted to obtain an updated multi-layer cooperative graph model, and task allocation operation is performed on the multi-UAV system according to the connection relationship between adjacent layer nodes in the updated multi-layer cooperative graph model.

[0109] Furthermore, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements a multi-UAV task allocation method with a multi-layer cooperative graph provided by the methods described above. This method includes: obtaining a multi-layer cooperative graph model based on a set of real-time status information of the multi-UAV system, task description information, and skill description information; the multi-layer cooperative graph model includes a task node layer, a cooperative cluster node layer, a skill node layer, and a UAV node layer; performing feature encoding on the nodes of each layer in the multi-layer cooperative graph model to obtain node features of each layer, and performing feature fusion on the node features of each layer. The global state features of the multi-layer cooperative graph model are obtained; the global state features are input into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete connection edges between the target node and any adjacent layer node of the target node for a selected target node; according to the real-time action information, the connection relationship between adjacent layer nodes in the multi-layer cooperative graph model is adjusted to obtain an updated multi-layer cooperative graph model, and the task allocation operation is performed on the multi-UAV system according to the connection relationship between adjacent layer nodes in the updated multi-layer cooperative graph model.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-UAV task allocation method for a multi-layered cooperative graph, characterized in that, include: Based on the real-time status information set of multiple unmanned aerial vehicle systems, the task description information and the skill description information, a multi-layer cooperation graph model is obtained. The multi-layered cooperation graph model includes a task node layer, a cooperation cluster node layer, a skill node layer, and a drone node layer; The nodes of each layer in the multi-layer cooperative graph model are feature-encoded to obtain the node features of each layer, and the node features of each layer are fused to obtain the global state features of the multi-layer cooperative graph model. The global state features are input into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete the connection edge between the target node and any adjacent layer node of the target node for the selected target node. Based on the real-time action information, the connection relationships between adjacent layer nodes in the multi-layer cooperation graph model are adjusted to obtain an updated multi-layer cooperation graph model. Then, based on the connection relationships between adjacent layer nodes in the updated multi-layer cooperation graph model, the task execution parameters corresponding to each UAV node in the UAV node layer are determined. The task execution parameters include the skill description information of the skill node corresponding to each UAV node, the cluster information of the cooperation cluster node to which each UAV node belongs, and the task description information of the task node corresponding to each UAV node. Based on the task execution parameters and the correspondence between each UAV node and each UAV in the multi-UAV system, the real-time task allocation information of each UAV in the multi-UAV system is determined. Based on the real-time task allocation information of each UAV, perform task allocation operations on each UAV. For each task node in the task node layer, the node feature of the task node is obtained by using a first preset prompt word to guide the large language model to extract semantic features of the task description information corresponding to the task node; For each skill node in the skill node layer, the node feature of the skill node is obtained by using the second preset prompt word to guide the large language model to extract semantic features of the skill description information corresponding to the skill node; The training steps for the decision network include: Multiple evolutionary training phases are constructed; the multiple evolutionary training phases include a single-type task training phase, a multi-task coordination training phase, and a dynamic task expansion training phase; For the current evolution training phase, if the duration for which the execution performance value of the current evolution training phase reaches the preset convergence threshold is longer than the preset duration, then switch to the next evolution training phase, and use the network parameters trained in the current evolution training phase as the initialization parameters for the next evolution training phase for incremental training; the execution performance value includes the average task completion rate or the total reward value.

2. The multi-UAV task allocation method for multi-layer cooperative graphs according to claim 1, characterized in that, The steps for obtaining the node characteristics of the cooperative cluster node layer and the node characteristics of the UAV node layer include: For each cooperative cluster node in the cooperative cluster node layer, a topology graph is constructed based on the communication topology relationship between each UAV node connected to the cooperative cluster node in the UAV node layer. The topology relationship features are extracted from the topology graph using a graph neural network to obtain the node features of the cooperative cluster node. For each UAV node in the UAV node layer, the real-time status information corresponding to the UAV node in the real-time status information set is feature-encoded using a feature coding network to obtain the node features of the UAV node.

3. The multi-UAV task allocation method for multi-layer cooperative graphs according to claim 1, characterized in that, The feature fusion of node features at each layer to obtain the global state features of the multi-layer cooperative graph model includes: By using a graph transformation network to aggregate node features across any two adjacent layers, multiple inter-layer aggregated features are obtained. The inter-layer aggregated features include the relationship features between task nodes in the task node layer and cooperative cluster nodes in the cooperative cluster node layer, the relationship features between cooperative cluster nodes and skill nodes in the skill node layer, and the relationship features between skill nodes and drone nodes in the drone node layer. The global state features are obtained by concatenating multiple inter-layer aggregated features.

4. The multi-UAV task allocation method for multi-layer cooperative graphs according to any one of claims 1-3, characterized in that, The step of inputting the global state features into the decision network to obtain the real-time action information of each graph operation operator includes: For each graph operation operator, the global state features are input into the policy subnetwork corresponding to the graph operation operator in the decision network to obtain the set of probability distributions output by the policy subnetwork. Based on the set of probability distributions, determine the index of the target node selected by the graph operation operator; Based on the index of the target node and the operation type corresponding to the target node, obtain the real-time action information of the graph operation operator; The probability distribution set includes the probability values ​​of each node in the target node layer corresponding to the graph operation operator being selected as the target node; each strategy sub-network corresponds to an operation type, which includes adding or deleting connection edges.

5. The multi-UAV task allocation method for multi-layer cooperative graphs according to any one of claims 1-3, characterized in that, In the single-type task training phase, the number of drones in the multi-UAV system participating in the training is a first preset number, and the training scenario includes single-type task description information. In the multi-task coordination training phase, the number of drones in the sample multi-UAV system is a second preset number, and the training scenario includes various types of task description information. The second preset size is larger than the first preset size; In the dynamic task expansion training phase, the number of drones in the sample multi-UAV system is a third preset number, and the training scenario includes at least one task description information randomly added to the training scenario in the multi-task coordination training phase. The third preset size is greater than or equal to the second preset size.

6. The multi-UAV task allocation method for multi-layer cooperative graphs according to claim 5, characterized in that, The total reward value of the decision network is obtained by fusing the task completion rate reward value, load balancing reward value, and task cost reward value during the training process; The task completion rate reward value is determined based on the total number of tasks completed in the simulation environment and the preset total number of tasks; The load balancing reward value is determined based on the number of drone nodes connected to each cooperative cluster node in the simulation environment, as well as the number of drone nodes deployed in the drone node layer and the number of cooperative cluster nodes deployed in the cooperative cluster node layer in the simulation environment. The task cost reward value is determined based on the simulation resource consumption cost incurred by each drone node in the drone node layer of the simulation environment when the drone performs its assigned task.

7. A multi-layered cooperative graph multi-UAV task allocation system, characterized in that, include: The model building unit is used to obtain a multi-layered cooperation graph model based on the real-time status information set of multiple UAV systems, the task description information and the skill description information. The multi-layered cooperation graph model includes a task node layer, a cooperation cluster node layer, a skill node layer, and a drone node layer; The feature extraction unit is used to encode the features of the nodes in each layer of the multi-layer cooperative graph model to obtain the node features of each layer, and to fuse the node features of each layer to obtain the global state features of the multi-layer cooperative graph model. The decision unit is used to input the global state features into the decision network to obtain the real-time action information of each graph operation operator; the real-time action information is used to instruct each graph operation operator to add or delete the connection edge between the target node and any adjacent layer node of the target node for the selected target node. The allocation unit is used to adjust the connection relationships between adjacent layer nodes in the multi-layer cooperation graph model according to the real-time action information to obtain an updated multi-layer cooperation graph model, and to determine the task execution parameters corresponding to each UAV node in the UAV node layer according to the connection relationships between adjacent layer nodes in the updated multi-layer cooperation graph model; the task execution parameters include the skill description information of the skill node corresponding to each UAV node, the cluster information of the cooperation cluster node to which each UAV node belongs, and the task description information of the task node corresponding to each UAV node. Based on the task execution parameters and the correspondence between each UAV node and each UAV in the multi-UAV system, the real-time task allocation information of each UAV in the multi-UAV system is determined; based on the real-time task allocation information of each UAV, a task allocation operation is performed on each UAV. For each task node in the task node layer, the node feature of the task node is obtained by using a first preset prompt word to guide the large language model to extract semantic features of the task description information corresponding to the task node; For each skill node in the skill node layer, the node feature of the skill node is obtained by using the second preset prompt word to guide the large language model to extract semantic features of the skill description information corresponding to the skill node; The training steps for the decision network include: Multiple evolutionary training phases are constructed; the multiple evolutionary training phases include a single-type task training phase, a multi-task coordination training phase, and a dynamic task expansion training phase; For the current evolution training phase, if the duration for which the execution performance value of the current evolution training phase reaches the preset convergence threshold is longer than the preset duration, then switch to the next evolution training phase, and use the network parameters trained in the current evolution training phase as the initialization parameters for the next evolution training phase for incremental training; the execution performance value includes the average task completion rate or the total reward value.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the multi-UAV task allocation method of any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-UAV task allocation method of the multi-layer cooperative graph as described in any one of claims 1 to 6.

Images