An unmanned cluster distributed cooperative task decision method, system and device

Abstract

This invention discloses a method, system, and device for distributed collaborative task decision-making in unmanned swarms, relating to the field of data management technology. The method includes: low-dimensional encoding of local observations during swarm startup or topology changes; selective state synchronization by combining spatial and temporal attention to filter key nodes; dynamic calculation of type-individual joint decision weights based on node type, individual capabilities, and task stage dependencies; task collaboration based on these decision weights; formation of self-organized task groups and their weights by incorporating target escape risk; dynamic adjustment of group weights by representing intra-group node relationships through virtual axes, virtual nests, and virtual keys; and adaptive adjustment of group configurations in response to environmental or task changes. This invention addresses the technical problem of insufficient adaptability of unmanned swarms in collaborative decision-making under dynamic task environments, achieving the technical effect of improving the dynamic adaptability and task execution efficiency of unmanned swarm collaborative decision-making.

Description

Technical Field

[0001] This invention relates to the field of data management technology, specifically to a method, system, and device for unmanned cluster distributed collaborative task decision-making. Background Technology

[0002] Unmanned swarms typically consist of unmanned nodes of different types and capabilities. During dynamic missions such as reconnaissance, search, penetration, and withdrawal, the swarm topology, node states, target movement states, and mission phases may continuously change. As the mission environment and node states dynamically evolve, the roles of each node in perception, computation, communication, and mission execution also change accordingly. If relatively fixed cooperative relationships and decision weights are still used, it is difficult to reflect the actual capabilities and importance of different nodes in the current mission phase in a timely manner. When the risk of target escape increases, local mission requirements change abruptly, or the swarm configuration is adjusted, it is also difficult to quickly complete the reorganization of cooperative relationships and the redistribution of decision weights, which can easily lead to untimely mission response and decreased node cooperative efficiency, thus limiting the adaptability of unmanned swarms in collaborative decision-making under dynamic mission environments. Summary of the Invention

[0003] This application provides a method, system, and device for distributed collaborative task decision-making in unmanned clusters, which addresses the technical problem of insufficient adaptability of unmanned clusters in collaborative decision-making under dynamic task environments in the prior art.

[0004] In view of the above problems, this application provides a method, system and device for decision-making of unmanned cluster distributed collaborative tasks.

[0005] A first aspect of this application provides a method for decision-making in a distributed collaborative task of an unmanned cluster, the method comprising: When the unmanned cluster starts up or undergoes topology changes, each node encodes the collected local observation data into a low-dimensional state embedding vector. Using the current subtask objective as the query and the neighbor embedding as the key, a compatibility score is calculated through scaled dot product attention. Key nodes are then selected and selective state synchronization is performed based on the attention weights output by the time attention. The unmanned cluster is divided into subgroups according to node type and type weight coefficients are configured. Based on the state of key nodes after selective state synchronization, the overall capability value of each node is measured and multiplied by the type weight coefficient of its subgroup to obtain the decision weight for type-individual joint ranking. This decision weight is based on the stage dependency matrix. The system updates the dependency intensity type weight coefficients of each task role in the current task phase; it performs task coordination processing based on the decision weights, synchronously calculates the target escape risk in real time, broadcasts a distress signal when the risk exceeds the threshold, and forms a self-organizing task group based on the response. The nodes in the group multiply the decision weight by the risk weight coefficient to obtain the group's decision weight; for the self-organizing task group, each node defines a virtual axis and a virtual nest and calculates the virtual bond strength, and updates the group's decision weight based on the number of bond connections and spatial centrality. When environmental changes or task switching trigger the self-organizing task group configuration adjustment, the bond breakage condition or bond formation condition is activated.

[0006] A second aspect of this application provides an unmanned cluster distributed collaborative task decision-making system, the system comprising: The synchronization module, when the unmanned cluster starts up or undergoes topology changes, encodes the collected local observation data into low-dimensional state embedding vectors by each node. Using the current subtask objective as the query and neighbor embeddings as the key, it calculates a compatibility score through scaled dot product attention, filters key nodes, and performs selective state synchronization based on the attention weights output by temporal attention. The weight configuration module divides the unmanned cluster into subgroups according to node type and configures type weight coefficients. Based on the state of key nodes after selective state synchronization, it measures the comprehensive capability value of each node and multiplies it by the type weight coefficient of its subgroup to obtain the decision weight for type-individual joint ranking. The weight is determined based on the stage dependency matrix. The current task phase updates the dependency intensity of each task role with type weight coefficients; the collaborative processing module is used to perform task collaborative processing according to the decision weights, synchronously calculate the target escape risk in real time, broadcast a distress signal when the risk exceeds the threshold, and form a self-organizing task group based on the response. The nodes in the group multiply the decision weight by the risk weight coefficient as the group's decision weight; the update module is used to define virtual axes and virtual nests for each node in the self-organizing task group and calculate the virtual bond strength, and update the group's decision weights according to the number of bond connections and spatial centrality. When environmental changes or task switching trigger the self-organizing task group configuration adjustment, the bond breaking condition or bond formation condition is activated.

[0007] Thirdly, this application also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the unmanned cluster distributed collaborative task decision-making method described in any one of the first aspects above.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: In this application, when an unmanned cluster starts up or undergoes topology changes, each node encodes the collected local observation data into a low-dimensional state embedding vector. Using the current subtask objective as the query and the neighbor embedding as the key, a compatibility score is calculated through scaled dot product attention. Key nodes are then selected and selective state synchronization is performed based on the attention weights output by time attention. The unmanned cluster is divided into subgroups according to node type and type weight coefficients are configured. Based on the state of key nodes after selective state synchronization, the comprehensive capability value of each node is measured and multiplied by the type weight coefficient of its subgroup to obtain the decision weight for type-individual joint ranking. The decision weight is based on the stage dependency moments. The system updates the dependency strength type weight coefficients of each task role in the current task phase within the cluster; based on the decision weights, it performs task coordination processing, synchronously calculates the target escape risk in real time, broadcasts a distress signal when the risk exceeds a threshold, and forms self-organizing task groups based on the response. Within each group, nodes multiply their decision weights by the risk weighting coefficients to obtain the group's internal decision weights. For each self-organizing task group, each node defines virtual axes and virtual nests and calculates virtual bond strength. The internal decision weights are updated based on the number of bond connections and spatial centrality. When environmental changes or task switching trigger configuration adjustments in the self-organizing task group, bond breakage or bond formation conditions are activated. This invention addresses the technical problem of insufficient adaptability of collaborative decision-making in dynamic task environments in existing unmanned swarms. By constructing dynamic decision weights based on key node states, task phases, and node capabilities, and combining them with a risk-driven self-organizing task group and configuration dynamic adjustment mechanism, it achieves the technical effect of improving the dynamic adaptability and task execution efficiency of collaborative decision-making in unmanned swarms. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of a distributed collaborative task decision-making method for unmanned clusters provided in an embodiment of this application; Figure 2 This is a schematic diagram of the stage dependency matrix in an unmanned cluster distributed collaborative task decision-making method provided in an embodiment of this application; Figure 3 This application provides a schematic diagram of the structure of an unmanned cluster distributed collaborative task decision-making system. Figure 4 This is a schematic diagram of the structure of an exemplary electronic device of this application.

[0011] Explanation of reference numerals in the attached diagram: Synchronization module 11, Weight configuration module 12, Collaborative processing module 13, Update module 14, Bus 300, Receiver 301, Processor 302, Transmitter 303, Memory 304, Bus interface 305. Detailed Implementation

[0012] This application provides a method, system, and device for distributed collaborative task decision-making in unmanned clusters. It addresses the technical problem of insufficient adaptability of unmanned clusters in collaborative decision-making under dynamic task environments in existing technologies. By constructing dynamic decision weights based on the status of key nodes, task stages, and node capabilities, and combining them with risk-driven self-organizing task groups and dynamic configuration adjustment mechanisms, it achieves the technical effect of improving the dynamic adaptability and task execution efficiency of unmanned cluster collaborative decision-making.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a method for decision-making in a distributed collaborative task of an unmanned cluster, the method comprising: Step S100: When the unmanned cluster starts or the topology changes, each node encodes the collected local observation data into a low-dimensional state embedding vector. Using the current subtask target as the query and the neighbor embedding as the key, the compatibility score is calculated by scaling dot product attention. Key nodes are selected and selective state synchronization is performed based on the attention weights output by the time attention.

[0016] In this embodiment, when the unmanned cluster starts or the topology changes, each node collects local observation data including the effective range of sensors, communication link quality, current task execution progress, and spatial coordinates, and maps it into a low-dimensional state embedding vector through a variational autoencoder network. Then, using the target vector of its current subtask as the query and the received low-dimensional state embedding vector of the neighboring node as the key, the compatibility score between the query and the key is calculated by scaling dot product attention and key nodes are marked. Then, combined with the attention weight ranking output by time attention, selective state synchronization is performed on key nodes.

[0017] Furthermore, in the method provided in the application embodiments, the process of filtering key nodes and performing selective state synchronization based on the attention weights of the time attention output also includes: Each node collects its own local observation data, which is mapped into a low-dimensional state embedding vector through a variational autoencoder network. The local observation data includes the effective range of the sensor, the quality of the communication link, the current task execution progress, and spatial coordinates. Each node uses the target vector of its current subtask as the query and the received low-dimensional state embedding vector of the neighboring node as the key value to perform selective state synchronization by marking key nodes.

[0018] In this embodiment, each node collects its own local observation data, which is then mapped into a low-dimensional state embedding vector via a variational autoencoder network. The local observation data includes the effective range of the sensor, the quality of the communication link, the current task execution progress, and the spatial coordinates. Specifically, each node collects local observation data according to a preset sampling period. At each sampling moment, it reads the maximum effective detection distance currently set by the sensor module and uses the maximum effective detection distance as the effective range of the sensor. It reads the number of data packets received by the communication module from neighboring nodes and the total number of data packets sent by neighboring nodes in the current sampling period. The ratio of the number of received data packets to the total number of sent data packets is used as the communication link quality. It reads the amount of work completed in the current subtask and the total amount of work in the current subtask recorded by the task management module. The ratio of the amount of work completed to the total amount of work is used as the current task execution progress. It reads the current position of the node output by the positioning module. Using the reference position preset in the task area as the coordinate origin, it subtracts the coordinate values ​​of the corresponding directions of the coordinate origin from the coordinate values ​​of the current position in the horizontal, vertical, and vertical directions to obtain the horizontal distance, vertical distance, and vertical distance of the node relative to the coordinate origin. The horizontal distance, vertical distance, and vertical distance are then combined to form spatial coordinates. The effective range of the sensor, the quality of the communication link, the current task execution progress and the spatial coordinates obtained at the same sampling time are arranged in a fixed order to form local observation data, and each data item is normalized. For any data item, the corresponding preset minimum value is subtracted from the data item, and the difference is divided by the difference between the corresponding preset maximum value and preset minimum value. The result is used as the normalized value of the data item. The normalized values ​​are combined in the original order to obtain normalized local observation data.

[0019] A variational autoencoder network (VAC) is constructed and trained based on normalized local observation data obtained from multiple historical sampling times. The VAC consists of an encoder and a decoder. The encoder is configured with an input layer, a hidden layer, a mean output layer, and a variance output layer, while the decoder is configured with a hidden layer and an output layer. A set of normalized local observation data is input into the encoder. The input layer passes each data point to the hidden layer. The hidden layer multiplies each input data point by its corresponding connection weight, sums the products, adds the corresponding bias value, and then performs activation processing to obtain the hidden feature data. The hidden feature data is then input into the mean output layer and the variance output layer. The mean output layer multiplies each hidden feature data point by its corresponding connection weight, sums the products, adds the corresponding bias value, and obtains the mean data corresponding to the latent variable. The variance output layer obtains the variance data corresponding to the latent variable using the same calculation method. The standard deviation data is calculated based on the variance data. Random data with the same dimension as the mean data is obtained from the standard normal distribution. Each random data is multiplied by the corresponding standard deviation data, and then added to the corresponding mean data to obtain the latent variable data. The latent variable data is input into the decoder. The decoder performs weighted multiplication, product summation, bias stacking and activation processing on the latent variable data in sequence to obtain reconstructed local observation data with the same dimension as the input local observation data.

[0020] The variational autoencoder network is trained using reconstructed local observation data. The difference between each data point in the reconstructed local observation data and its corresponding normalized local observation data is calculated. These differences are squared, summed, and divided by the number of data points to obtain the reconstruction error. The distribution error between the latent variable distribution and the standard normal distribution is calculated using the mean and variance data obtained from the encoder. The reconstruction error and distribution error are then added to obtain the network loss value. The adjustment amounts corresponding to the connection weights and biases in the encoder and decoder are calculated in reverse based on the network loss value. The connection weights and biases are updated according to a preset learning step size. Historical normalized local observation data is repeatedly input and parameters are updated until the difference between the network loss values ​​obtained from two consecutive training iterations is less than a preset convergence threshold. The connection weights and biases are then fixed at this point, resulting in the trained variational autoencoder network. During the operation of the unmanned cluster, the normalized local observation data of the node is input into the trained variational autoencoder network. The encoder calculates the corresponding mean and variance data according to the fixed connection weights and bias values. The latent variable data is obtained by following the calculation process of the random data, standard deviation data and mean data mentioned above. The latent variable data with a dimension lower than that of the local observation data is arranged in a fixed order to obtain the low-dimensional state embedding vector of the corresponding node.

[0021] Next, each node uses its current subtask's target vector as the query and the received low-dimensional state embedding vector of its neighboring nodes as the key. It calculates the compatibility score between the query and the key value through scaled dot product attention and marks neighboring nodes with compatibility scores higher than a preset spatial attention threshold as key nodes. Subsequently, for key nodes, a time attention mechanism is used to perform weighted aggregation of the state embeddings from multiple consecutive historical moments to obtain the attention weight ranking of each key node. Selective state synchronization is then performed based on the attention weight ranking. Specifically, complete local observation data is sent to neighboring nodes with attention weights exceeding a first preset synchronization threshold, and low-dimensional state embedding vectors are sent to neighboring nodes with attention weights lower than the first preset synchronization threshold but higher than the second preset threshold. This filters the communication objects of a single node from the total number of cluster N to the number of key nodes k.

[0022] Furthermore, the method provided in the application embodiments, which performs selective state synchronization by marking key nodes, further includes: By using a scaled dot product attention approach, each node calculates a compatibility score between the query and the key value, and marks neighboring nodes with compatibility scores higher than a preset spatial attention threshold as key nodes. Through a temporal attention mechanism, the state embeddings of each key node at multiple consecutive historical moments are weighted and aggregated to capture state change trends and output the attention weight ranking of each key node. Each node performs selective state synchronization according to the attention weight ranking, wherein complete local observation data is sent to neighboring nodes with attention weights exceeding a first preset synchronization threshold, and low-dimensional state embedding vectors are sent to neighboring nodes with attention weights lower than the first preset synchronization threshold but higher than the second preset threshold, so that the number of communication objects of a single node is reduced from the total number of clusters N to the number of key nodes k.

[0023] In this embodiment, each node first calculates a compatibility score between the query and the key value using a scaled dot product attention approach. Neighboring nodes with compatibility scores higher than a preset spatial attention threshold are marked as key nodes. Specifically, each node uses its current subtask's target vector as the query and reads the low-dimensional state embedding vectors of its neighboring nodes that it can currently receive. Each neighboring node's low-dimensional state embedding vector is used as a key value, where the target vector and the low-dimensional state embedding vectors share the same vector dimension. For any neighboring node, the elements at the same position in the query and the corresponding key value of that neighboring node are multiplied one by one, and all product results are summed to obtain the dot product of the query and the key value. The vector dimension of the key value is read, the square root of the vector dimension is calculated, and the square root is divided by the dot product result to obtain the compatibility score corresponding to that neighboring node. The compatibility scores of each neighbor node are calculated sequentially using the same method and compared with a preset spatial attention threshold. Neighbor nodes with compatibility scores higher than the preset spatial attention threshold are marked as key nodes, and their node identifiers, current low-dimensional state embedding vectors, and compatibility scores are recorded. Neighbor nodes with compatibility scores lower than the preset spatial attention threshold are not included in the subsequent temporal attention calculation process. After comparing the compatibility scores of all neighbor nodes, the number of marked key nodes is counted as k, thus reducing the scope of nodes targeted for subsequent state synchronization from the total number of cluster nodes N to the number of key nodes k.

[0024] Next, a time attention mechanism is used to weight and aggregate the state embeddings of each key node across multiple consecutive historical moments, capturing state change trends and outputting the attention weight ranking of each key node. For each key node, the low-dimensional state embedding vectors corresponding to multiple consecutive historical moments before the current moment are read according to a preset sampling period and arranged in chronological order of sampling time. The low-dimensional state embedding vector corresponding to the historical moment closest to the current moment is used as the time comparison benchmark. The time comparison benchmark is multiplied element-wise with the low-dimensional state embedding vectors of each of the other historical moments, and all product results corresponding to the same historical moment are summed to obtain the time relevance score of each historical moment. An exponential operation is performed on each time relevance score, and the exponential operation result corresponding to any historical moment is divided by the sum of the exponential operation results corresponding to all historical moments to obtain the time attention weight of that historical moment, so that the sum of the time attention weights corresponding to multiple consecutive historical moments is 1. Each element in the low-dimensional state embedding vector at each historical moment is multiplied by the corresponding temporal attention weight. The products at the same dimensional position across different historical moments are then summed to obtain the weighted aggregated state embedding for that key node. This allows the state embeddings from different historical moments to participate in the current state representation according to their corresponding temporal attention weights. Subsequently, the weighted aggregated state embedding of each key node is multiplied one by one by the elements at the same position in the target vector of the current subtask, and the sum is obtained to get the initial attention score for that key node. The initial attention scores of k key nodes are then exponentially calculated. The result of the exponential calculation for any key node is divided by the sum of the exponential calculation results for the k key nodes to obtain the attention weight for that key node. The k key nodes are then ranked from largest to smallest attention weight to obtain the attention weight sorting for each key node.

[0025] Each node performs selective state synchronization among k key nodes according to the attention weight. Specifically, complete local observation data is sent to key nodes whose attention weight exceeds the first preset synchronization threshold, and low-dimensional state embedding vectors are sent to key nodes whose attention weight is lower than the first preset synchronization threshold but higher than the second preset threshold. Specifically, a first preset synchronization threshold and a second preset synchronization threshold are preset, with the first preset synchronization threshold being greater than the second preset synchronization threshold. Each node sequentially reads the attention weights corresponding to each key node in descending order of attention weight, and first compares the attention weights with the first preset synchronization threshold. When the attention weight exceeds the first preset synchronization threshold, the effective range of the sensor, the communication link quality, the current task execution progress, and the spatial coordinates obtained by the current node at the current sampling time are read, and complete local observation data is formed according to the preset field order. The complete local observation data is then sent to the corresponding key node. When the attention weight does not exceed the first preset synchronization threshold, the attention weight is further compared with the second preset synchronization threshold. When the attention weight is lower than the first preset synchronization threshold but higher than the second preset synchronization threshold, the low-dimensional state embedding vector obtained by the current node through the variational autoencoder network is read, and the low-dimensional state embedding vector is sent to the corresponding key node. When the attention weight is not higher than the second preset synchronization threshold, the state data of the current sampling period is not sent to the corresponding key node. Therefore, by scaling dot product attention, the range of nodes involved in the subsequent state synchronization of a single node is first reduced from the total number of nodes N in the cluster to the number of key nodes k. Then, based on the attention weights obtained from time attention and the first and second preset synchronization thresholds, the method of sending complete local observation data or low-dimensional state embedding vectors is determined among the k key nodes.

[0026] Step S200: Divide the unmanned cluster into subgroups according to node type and configure type weight coefficients. Based on the state of key nodes after selective state synchronization, measure the comprehensive capability value of individual nodes and multiply it with the type weight coefficient of the subgroup to obtain the decision weight of type-individual joint sorting. The type weight coefficients are updated according to the dependence strength of the current task stage on each task role in the stage dependency matrix.

[0027] In this embodiment, the unmanned swarm is first divided into multiple subgroups according to node type, and type weight coefficients are configured for each subgroup to characterize the degree of differentiated contribution of the corresponding node type to the global decision mean field. Next, based on the current task stage, the type weight coefficients of each subgroup are updated using the dependence strength of each task stage on different task roles in the stage dependency matrix. Subsequently, based on the state of key nodes after selective state synchronization, the node comprehensive capability of each key node is calculated, and the node comprehensive capability of each key node is matched and multiplied with the type weight coefficients corresponding to its subgroup to obtain the decision weights that take into account both the contribution of node type and the individual capability of node.

[0028] Furthermore, the method provided in the application embodiments, in obtaining the decision weights for the joint ranking of type and individual, further includes: The unmanned swarm is divided into multiple subgroups according to node type. Each subgroup has a defined and initialized type weight coefficient, which quantifies the degree of differential contribution of node type within each subgroup to the global decision mean field. Based on the task stage, the type weight coefficients of multiple subgroups are updated according to the stage dependency matrix. Based on the state of key nodes after selective state synchronization, the node comprehensive capability of each key node is calculated. By matching the type weight coefficients and performing multiplication, the decision weight based on type-individual joint ranking is obtained.

[0029] Furthermore, the method provided in the application embodiments also includes: The stage dependency matrix is ​​constructed using the dependency strength of the task stage on the task role as matrix items. The task stages include the reconnaissance stage, the cooperative search stage, the penetration stage, and the withdrawal stage. The task roles include perception, computing, communication, and attack. The dynamically evolving node comprehensive capability is calculated as Σ(role capability value × stage role dependency weight).

[0030] In this embodiment, the unmanned swarm is first divided into multiple subgroups according to node type. Each subgroup has a defined and initialized type weight coefficient, which quantifies the differentiated contribution of node types within each subgroup to the global decision-making average field. Specifically, during unmanned swarm initialization, the pre-configured node type identifiers of each node are read, and nodes with the same node type identifier are grouped into the same subgroup. The number of subgroups included in the unmanned swarm is counted. The result of dividing 1 by the number of subgroups is used as the initial type weight coefficient for each subgroup, ensuring that each subgroup has the same initial type weight before the task begins. For each subgroup, the capability parameters of the four task roles (sensing, computation, communication, and strike) corresponding to each node within that subgroup are read. Each capability parameter is divided by the pre-set maximum capability parameter for the corresponding task role to obtain a role capability value in the range of 0 to 1. For the same task role within the same subgroup, the role capability values ​​corresponding to all nodes in that subgroup are summed and then divided by the number of nodes in that subgroup to obtain the average perception, computation, communication, and strike capability values ​​for that subgroup.

[0031] Next, a stage dependency matrix is ​​constructed based on the task phases. This matrix uses the dependency strength of each task phase on a specific task role as its matrix elements. The task phases include reconnaissance, cooperative search, penetration, and withdrawal. The task roles include perception, computation, communication, and strike. Specifically, the reconnaissance, cooperative search, penetration, and withdrawal phases are represented as the four rows of the stage dependency matrix, and perception, computation, communication, and strike are represented as the four columns. For each task phase, the task requirements for perception, computation, communication, and strike are retrieved from the task planning data, along with the standard processing capacity of each task role within a unit task cycle. For the perception role, the perception task requirement for that task phase is divided by the standard processing capacity of the perception role to obtain the dependency strength. For the computation role, the computation task requirement is divided by the standard processing capacity of the computation role to obtain the dependency strength. For the communication role, the communication task requirement is divided by the standard processing capacity of the communication role to obtain the dependency strength. For the strike role, the strike task requirement is divided by the standard processing capacity of the strike role to obtain the dependency strength. The four dependency strengths obtained from the same task phase are summed to obtain the total dependency strength value for that task phase. Then, the dependency strengths of the sensing, computing, communication, and strike roles are divided by the total dependency strength value to obtain the corresponding phase-specific role dependency weights for the sensing, computing, communication, and strike phases, respectively, ensuring that the sum of the four phase-specific role dependency weights for the same task phase is 1. The same process is then used to calculate the four phase-specific role dependency weights for the reconnaissance, cooperative search, penetration, and withdrawal phases, and these weights are written into the corresponding rows and columns of the phase dependency matrix to obtain the phase dependency matrix. For example... Figure 2 As shown, the stage dependency matrix uses reconnaissance, cooperative search, penetration, and withdrawal stages as mission stage dimensions, and perception, computation, communication, and strike as mission role dimensions. Each matrix item represents the stage-specific role dependency weight of each mission role within the corresponding mission stage. In this example of the stage dependency matrix, the stage-specific role dependency weights for perception, computation, communication, and strike in the reconnaissance stage are 0.42, 0.18, 0.30, and 0.10, respectively; for the cooperative search stage, they are 0.34, 0.26, 0.30, and 0.10; for the penetration stage, they are 0.16, 0.20, 0.24, and 0.40; and for the withdrawal stage, they are 0.24, 0.18, 0.43, and 0.15. The sum of the four stage-specific role dependency weights corresponding to the same mission stage is 1. When the mission stage changes, the stage-specific role dependency weight in the row corresponding to the mission stage in the stage dependency matrix is ​​read, and the subsequent type weight coefficients and node comprehensive capabilities are calculated accordingly.

[0032] During task execution, the current task stage identifier is read, and the corresponding row in the stage dependency matrix is ​​determined based on the current task stage identifier. From this row, the perception stage role dependency weight, calculation stage role dependency weight, communication stage role dependency weight, and strike stage role dependency weight corresponding to the current task stage are read sequentially. For any subgroup, the average perception role capability value of the subgroup is multiplied by the perception stage role dependency weight, the average calculation role capability value is multiplied by the calculation stage role dependency weight, the average communication role capability value is multiplied by the communication stage role dependency weight, and the average strike role capability value is multiplied by the strike stage role dependency weight. The four products are then added together to obtain the type contribution value corresponding to the subgroup in the current task stage. The type contribution values ​​of all subgroups are obtained in the same way, and the type contribution values ​​of all subgroups are added together to obtain the total type contribution value. The type contribution value of any subgroup is then divided by the total type contribution value, and the result is used as the updated type weight coefficient of the subgroup. Therefore, the type weight coefficients of multiple subgroups are updated according to the stage role dependency weights corresponding to the current task stage; when the current task stage changes, the four stage role dependency weights corresponding to the changed task stage in the stage dependency matrix are read, and the type contribution value and type weight coefficients of each subgroup are recalculated according to the same process.

[0033] Based on the state of the key nodes after selective state synchronization, the overall node capability of each key node is calculated. For any key node, the state parameters of the four task roles corresponding to that key node are read. Specifically, the perception role capability value is obtained by dividing the current effective range of the key node's sensors by the maximum effective range of the node's sensors; the computing role capability value is obtained by dividing the current available computing resources of the key node by the total computing resources of the node; the communication role capability value is obtained by dividing the number of data packets successfully received in the current sampling period by the total number of data packets that should be received; and the strike role capability value is obtained by dividing the current remaining executable strike task resources of the key node by the initial total strike task resources. The system reads the four stage-specific role dependency weights corresponding to the current task stage in the stage dependency matrix. It then multiplies the perception role's capability value with the perception stage-specific role dependency weight, the calculation role's capability value with the calculation stage-specific role dependency weight, the communication role's capability value with the communication stage-specific role dependency weight, and the attack role's capability value with the attack stage-specific role dependency weight. Finally, it adds up the four product results to obtain the node's comprehensive capability in the current task stage. That is, the dynamically evolving node's comprehensive capability is equal to the sum of the products of each role's capability value and its corresponding stage-specific role dependency weight. When the key node's state or the current task stage changes, the corresponding role capability value and stage-specific role dependency weight are reread, and the node's comprehensive capability is updated according to the same calculation process.

[0034] The decision weights based on type-individual joint ranking are obtained by matching type weight coefficients and performing multiplication. For any key node, the node type identifier of the key node is read, the subgroup to which the key node belongs is determined based on the node type identifier, and the type weight coefficient of the subgroup corresponding to the current task stage is read. The overall node capability of the key node is multiplied by the type weight coefficient of its subgroup to obtain the decision weight of the key node. The decision weights corresponding to all key nodes are calculated in the same way, and then the key nodes are sorted from largest to smallest according to the decision weights, ensuring that each decision weight corresponds to the node identifier, node type, overall node capability, and type weight coefficient of the corresponding key node, thus obtaining the decision weights based on type-individual joint ranking.

[0035] Step S300: Perform task coordination processing based on the decision weights, synchronously calculate the target escape risk in real time, broadcast a distress signal when the risk exceeds the threshold, form a self-organizing task group based on the response, and multiply the decision weights of the nodes in the group by the risk weighting coefficients to obtain the decision weights of the group.

[0036] In this embodiment, task coordination is performed based on decision weights. Specifically, the decision weights corresponding to the current subtask to be executed and each key node participating in the current subtask are read. The key nodes are sorted from largest to smallest according to their decision weights. Based on the pre-set number of nodes required for the current subtask to be executed, the corresponding task execution nodes are determined sequentially starting from the top-ranked key nodes. The target location, task type, and task execution parameters corresponding to the current subtask to be executed are sent to the task execution nodes. The task execution nodes execute the current subtask to be executed according to the received task execution parameters and update the current task execution progress, node spatial coordinates, and target motion state according to a preset sampling period, thereby completing the task coordination based on decision weights.

[0037] During the execution of tasks at each node, the escape risk of the target is calculated synchronously in real time, and the escape risk level is determined based on the target's velocity vector and the rate of change of direction. When the escape risk level is greater than the preset distress threshold, the current task execution node broadcasts a distress signal containing the target's location, escape risk level, and response time window in the preset area where the target is located. The node receiving the distress signal calculates the response priority based on its remaining energy and the spatial distance to the target, marks the nodes whose response priority exceeds the preset response threshold as candidate response nodes, and sends a response confirmation to the current task execution node. Based on the response, a self-organizing task group is formed.

[0038] After the self-organizing task group is formed, each node in the group uses the decision weight based on the type-individual joint ranking as a benchmark, multiplies the decision weight by the risk weighting coefficient to obtain the corresponding intra-group decision weight; at the same time, non-intra-group nodes that are not part of the self-organizing task group among the key nodes of the current task execution node transfer their decision weight to the intra-group response node according to a preset ratio, and the transferred decision weight is incorporated into the intra-group decision weight of the corresponding intra-group response node to form the intra-group decision weight corresponding to the self-organizing task group.

[0039] Furthermore, in the method provided in the application embodiments, forming a self-organizing task group further includes: During the task execution process of each node, the escape risk level is calculated based on the target's velocity vector and direction change rate. When the escape risk level exceeds a preset distress threshold, the current task execution node broadcasts a distress signal in a preset area where the target is located. The distress signal includes the target's location, escape risk level, and response time window. The node receiving the distress signal calculates its response priority based on its remaining energy and the spatial distance to the target. When the priority exceeds a preset response threshold, it is marked as a candidate response node and sends a response confirmation to the task execution node, forming a self-organizing task group. After the self-organizing task group is formed, each node in the group uses a decision weight based on type-individual joint ranking as a benchmark, multiplied by a risk weighting coefficient, as the group's decision weight. Nodes outside the group transfer their decision weight to response nodes within the group according to a preset ratio. Nodes outside the group are key nodes of the task execution node that do not belong to the self-organizing task group.

[0040] In this embodiment, each node continuously acquires the spatial coordinates of the target and the corresponding sampling time according to a preset sampling period during task execution. For two adjacent sampling times, the horizontal, vertical, and longitudinal coordinates of the previous sampling time are subtracted from the horizontal, vertical, and longitudinal coordinates of the subsequent sampling time to obtain the position change of the target in the three coordinate directions. Then, the three position changes are divided by the time interval between two adjacent sampling times to obtain the velocity components of the target in the horizontal, vertical, and longitudinal directions. The three velocity components are arranged in a fixed order to form the target motion velocity vector. The three velocity components of the target motion velocity vector are squared, and the sum of the three squared results is taken as the square root to obtain the target motion velocity. For two adjacent sampling periods, the magnitude of the corresponding target motion velocity vector is calculated. Each velocity component is divided by the magnitude of the corresponding target motion velocity vector to obtain two unit velocity vectors. The elements at the same position of the two unit velocity vectors are multiplied and summed. The results are then subjected to inverse cosine calculation to obtain the direction angle corresponding to the two adjacent sampling periods. The direction angle is then divided by the time interval between the two sampling periods to obtain the direction change rate. The velocity risk value is obtained by dividing the target's velocity by the preset maximum target velocity, and the direction change risk value is obtained by dividing the direction change rate by the preset maximum direction change rate. The velocity risk value is multiplied by a preset velocity risk weight, and the direction change risk value is multiplied by a preset direction change risk weight. The two products are then added together to obtain the escape risk value, where the sum of the preset velocity risk weight and the preset direction change risk weight is 1. A series of risk intervals are pre-defined according to the escape risk values ​​from smallest to largest, and a corresponding escape risk level is assigned to each risk interval. The current escape risk value is compared with the upper and lower boundaries of each risk interval to determine the risk interval in which the current escape risk value belongs, and the corresponding escape risk level for that risk interval is read. The escape risk level is compared with a preset distress threshold. When the escape risk level is greater than the preset distress threshold, the current task execution node reads the target's current spatial coordinates, uses the target's current location as the center and a preset communication distance as the area radius to determine the target's location in a preset area, and broadcasts a distress signal within the preset area. When generating a distress signal, the current target location, escape risk level, distress signal broadcast time, and preset response duration are written into the distress signal. The distress signal broadcast time is used as the start time of the response time window, and the sum of the broadcast time and the preset response duration is used as the end time of the response time window.

[0041] The node receiving the distress signal reads its current remaining energy and initial energy. It divides the current remaining energy by the initial energy to obtain the remaining energy ratio. It then reads the target's location and its own current spatial coordinates from the distress signal. It calculates the differences between its own lateral coordinates and the target's lateral coordinates, its own longitudinal coordinates and the target's longitudinal coordinates, and its own vertical coordinates and the target's vertical coordinates. It squares each of these three coordinate differences, sums them, and then takes the square root of the sum to obtain the spatial distance based on the target. It divides the spatial distance by a preset communication distance to obtain the distance ratio, and subtracts the distance ratio from 1 to obtain the distance response value. It multiplies the remaining energy ratio by a preset energy weight and the distance response value by a preset distance weight, and adds the two products to obtain the corresponding node's response priority. The sum of the preset energy weight and the preset distance weight is 1. The response priority is compared with a preset response threshold, and the current time is compared with the end time of the response time window. When the response priority exceeds the preset response threshold and the current time does not exceed the end time of the response time window, the corresponding node is marked as a candidate response node, and the candidate response node sends a response confirmation to the current task execution node. The response confirmation includes the node identifier, current spatial coordinates, current remaining energy, and response priority of the candidate response node. When the response priority does not exceed the preset response threshold or the current time exceeds the end time of the response time window, no response confirmation is sent. The current task execution node receives response confirmations from each candidate response node sequentially within the response time window. After the response time window ends, the corresponding candidate response node is determined based on the received node identifier, and the current task execution node and the candidate response node are jointly determined as nodes within the self-organizing task group, thus forming the self-organizing task group.

[0042] After the self-organizing task group is formed, the escape risk value and the preset maximum escape risk value corresponding to the formation of the self-organizing task group are read. The escape risk value is divided by the preset maximum escape risk value to obtain the normalized escape risk value. Then, the normalized escape risk value is added to 1 to obtain the risk weighting coefficient. For each node in the self-organizing task group, the decision weight of the node in the group based on the type-individual joint sorting is read. The decision weight is multiplied by the risk weighting coefficient to obtain the initial intra-group decision weight of the node in the group. All key nodes corresponding to the current task execution node are read, and the node identifier of each key node is read one by one. The node identifier of the key node is compared with the node identifier of each intra-group node. Key nodes with inconsistent node identifiers and not appearing in the self-organizing task group are identified as non-intra-group nodes. For each non-group node, its current decision weight and preset transfer ratio are read. The current decision weight is multiplied by the preset transfer ratio to obtain the transfer decision weight corresponding to the non-group node. Then, the original decision weight of the non-group node is subtracted from the corresponding transfer decision weight to obtain the decision weight of the non-group node after weight transfer. The transfer decision weights of all non-group nodes are added together to obtain the total transfer decision weight. For each group response node, its response priority is read. The response priority of the group response node is divided by the sum of the response priorities of all group response nodes to obtain the corresponding weight receiving ratio. The total transfer decision weight is multiplied by the weight receiving ratio of the group response node to obtain the decision weight received by the group response node. Then, the received decision weight is added to the initial group decision weight of the group response node to obtain the group decision weight of the group response node. For the current task execution node, the decision weight obtained based on the type-individual joint sorting is multiplied by the risk weighting coefficient to obtain the corresponding group decision weight.

[0043] Step S400: For the self-organizing task group, each node defines a virtual axis and a virtual nest and calculates the virtual bond strength. The decision weights within the group are updated based on the number of bond connections and spatial centrality. When environmental changes or task switching trigger the self-organizing task group configuration adjustment, the bond breaking condition or bond formation condition is activated.

[0044] In this embodiment, for a self-organizing task group, each node defines a virtual axis representing the desired motion direction vector and a virtual nest representing the desired spatial position of the formation. Based on the geometric relationship between the virtual axis and the virtual nest, and combined with the linear weighted result of the spatial distance and directional angle between nodes, the virtual bond strength between adjacent nodes is calculated. Then, based on the number of virtual bond connections of each node and the spatial centrality determined by the reciprocal of the average geometric distance from the node to other nodes in the group, the decision weights within the group are updated. When environmental changes or task switching trigger the configuration adjustment of the self-organizing task group, the bond breaking condition or bond formation condition is activated according to the spatial distance between nodes, the directional angle, and the geometric relationship between the new node and the existing virtual nest, and the weight structure of the self-organizing task group is readjusted as virtual bonds break or form.

[0045] Furthermore, in the method provided in the application embodiments, updating the decision weights within the group further includes: In a self-organizing task group, each node defines its own virtual axis and virtual nest. The virtual axis represents the node's desired motion direction vector, and the virtual nest represents the node's desired spatial position in the formation. Each node calculates the virtual bond strength with its neighboring nodes based on the geometric relationship between the virtual axis and virtual nest. The virtual bond strength is determined by a weighted combination of the spatial distance and the angle between the directions of the nodes, and the strength of the virtual bond is quantified using a linear weighting method. The decision weights within the group are updated based on the number of virtual bond connections and the spatial centrality, where the spatial centrality is measured as the reciprocal of the average geometric distance from a node to other nodes in the group.

[0046] In this embodiment, within a self-organizing task group, each node defines its own virtual axis and virtual nest. The virtual axis represents the node's desired motion direction vector, and the virtual nest represents the node's desired spatial position within the formation. Specifically, each node reads its current spatial position and the target position corresponding to the current task. The coordinates of the target position in the horizontal, vertical, and longitudinal directions are subtracted from the coordinates of the node's current spatial position to obtain the position differences in the three directions. The squares of the three position differences are calculated, summed, and the square root is taken to obtain the spatial distance from the node's current position to the target position. The three position differences are then divided by the spatial distance to obtain the unit direction components in the three directions, which are then arranged in a fixed order to form virtual axes. For the virtual nest, the desired position offset relative to the formation center position is pre-configured for each node within the group based on the formation format adopted by the self-organizing task group. The current formation center position of the self-organizing task group is read, and the coordinates of the formation center position in the three coordinate directions are added to the desired position offset of the corresponding node to obtain the node's desired spatial position within the formation. This desired spatial position is then determined as the virtual nest of the corresponding node. When the position of the self-organizing task group changes, the virtual nest of each node is recalculated based on the updated formation center position; when the target position corresponding to the current task changes, the corresponding virtual axis is recalculated based on the updated target position and the current spatial position of the node.

[0047] Each node calculates the virtual bond strength with its adjacent nodes based on the geometric relationship between the virtual axis and the virtual nest. The virtual bond strength is determined by a weighted combination of the spatial distance and the directional angle between nodes, using a linear weighting method to quantify the strength of the virtual bond. For any node and its adjacent nodes, the spatial coordinates of the corresponding virtual nests are read. The coordinates of the two virtual nests in the horizontal, vertical, and longitudinal directions are subtracted, and the squares of the sum of these three differences are taken to obtain the spatial distance between the corresponding virtual nests. The virtual axes corresponding to the two nodes are read, and the directional components at the same position on the two virtual axes are multiplied and the products are added together. The inverse cosine of the result is used to calculate the directional angle between the two virtual axes. The spatial distance is divided by a preset maximum virtual bond distance to obtain a distance ratio. When the distance ratio is greater than 1, it is set to 1. The distance correlation value is then obtained by subtracting the distance ratio from 1. The directional angle is divided by a preset maximum directional angle to obtain an angle ratio. When the angle ratio is greater than 1, it is set to 1. The directional correlation value is then obtained by subtracting the angle ratio from 1. The distance association value and direction association value are multiplied by preset distance weights and preset direction weights respectively, and the two products are added together to obtain the virtual bond strength between two adjacent nodes. The sum of the preset distance weights and preset direction weights is 1. The virtual bond strength between the current node and each adjacent node is calculated in the same way. Adjacent nodes with virtual bond strength higher than a preset effective bond strength threshold are counted as one effective virtual bond connection. The number of effective virtual bond connections is accumulated to obtain the number of virtual bond connections of the current node.

[0048] Based on the virtual key connection count and spatial centrality, update the decision weights within the group. Spatial centrality is measured as the reciprocal of the average geometric distance from a node to all other nodes in the group. For any node within the group, read the current spatial coordinates of that node and all other nodes in the self-organizing task group. Calculate the coordinate differences between that node and each of the other nodes in the three coordinate directions. Square each of the three coordinate differences, sum them, and take the square root to obtain the geometric distance between the corresponding two nodes. Add the geometric distances from that node to all other nodes in the group, and divide by the number of other nodes in the group to obtain the average geometric distance from that node to all other nodes in the group. Divide this average geometric distance by 1 to obtain the spatial centrality of that node. Divide the virtual key connection count of each node by the maximum number of virtual key connections that can be formed in the self-organizing task group to obtain the corresponding normalized connection count value. Divide the spatial centrality of each node by the sum of the spatial centralities of all nodes in the group to obtain the corresponding normalized spatial centrality value. Multiply the normalized connection number by the preset connection weight, multiply the normalized spatial centrality by the preset centrality weight, and add the two products to obtain the weight adjustment value corresponding to the node. The sum of the preset connection weight and the preset centrality weight is 1. Multiply the weight adjustment value by the intra-group decision weight obtained after the formation of the aforementioned self-organizing task group to obtain the adjusted intra-group decision weight. Obtain the adjusted intra-group decision weights for all intra-group nodes using the same process. Then, divide the adjusted intra-group decision weight of any node by the sum of the adjusted intra-group decision weights of all intra-group nodes to obtain the updated intra-group decision weight for that node. Finally, redetermine the weights corresponding to each intra-group node according to the updated intra-group decision weights.

[0049] Furthermore, the method provided in the application embodiments also includes: The bond breakage condition is that the distance between nodes exceeds a preset breakage threshold or the directional angle exceeds a preset angle threshold. The bond formation condition is that a new node enters the group's perception range and its geometric relationship with the existing virtual nest meets a preset matching threshold. When a bond breakage occurs, the node exits the weight structure of the self-organizing task group. When a bond formation occurs, the new node enters the weight structure of the group. The weight distribution is automatically redistributed with bond breakage and bond reconstruction.

[0050] In this embodiment, each node detects the spatial distance and directional angle between nodes corresponding to the current virtual key according to a preset sampling period. The spatial distance between nodes is the spatial distance between adjacent nodes obtained when calculating the virtual key strength, and the directional angle is the directional angle between the virtual axes of adjacent nodes obtained when calculating the virtual key strength. The spatial distance between nodes is compared with a preset breakage threshold, and the directional angle is compared with a preset angle threshold. When the spatial distance between nodes exceeds the preset breakage threshold or the directional angle exceeds the preset angle threshold, the corresponding virtual key is determined to meet the key breakage condition, and the virtual key is updated from a connected state to a broken state. After a key breakage occurs, the virtual key connection relationship between the corresponding node and other nodes in the self-organizing task group is re-determined. When the node no longer maintains a valid virtual key connection with other nodes in the group, the node is removed from the weight structure of the self-organizing task group, and the node is removed from the group it belonged to before leaving. The intra-group decision weights are determined as the decision weights to be reallocated. For nodes that are still retained in the weight structure of the self-organizing task group, the weight adjustment values ​​of each node are obtained again according to the aforementioned calculation process of virtual key connectivity and spatial centrality. The weight adjustment value of any retained node is divided by the sum of the weight adjustment values ​​of all retained nodes to obtain the corresponding weight allocation ratio. The decision weights to be reallocated are allocated to each retained node according to the weight allocation ratio, and then added to the original intra-group decision weights of each retained node. The added intra-group decision weights are then normalized to obtain the intra-group decision weights of each node after the bond breaks.

[0051] For a new node entering the vicinity of a self-organizing task group, the current spatial position of the new node is read, and the spatial distance between the new node and each node in the group is calculated. The obtained spatial distance is compared with the corresponding preset perception distance. When the spatial distance between the new node and at least one node in the group is not greater than the preset perception distance, it is determined that the new node has entered the group formation perception range. Then, the expected spatial position corresponding to each existing virtual nest is read, and the geometric distance between the current spatial position of the new node and the expected spatial position corresponding to each existing virtual nest is calculated. The geometric distance is compared with a preset matching threshold. When the geometric distance between the new node and at least one existing virtual nest is not greater than the preset matching threshold, it is determined that the geometric relationship between the new node and the existing virtual nest meets the preset matching threshold. When a new node enters the group's perception range and its geometric relationship with the existing virtual nest meets the preset matching threshold, the key formation condition is determined to be met. A virtual key is established between the new node and the corresponding node in the group, and the new node is added to the weight structure of the group. After the new node is added, it is included in the node range of the self-organizing task group. The number of virtual key connections and spatial centrality of the new node and the original nodes in the group are re-determined according to the aforementioned calculation process, and the weight adjustment value of each node is obtained accordingly. The current intra-group decision weight of each node is multiplied by its respective weight adjustment value to obtain the corresponding intra-group decision weight to be normalized. Then, the intra-group decision weight to be normalized of any node is divided by the sum of the intra-group decision weights to be normalized of all nodes in the group to obtain the intra-group decision weight of each node after key formation. This causes the weight distribution to be redistributed with key breakage and key reconstruction.

[0052] In summary, the embodiments of this application have at least the following technical effects: In this application, when an unmanned cluster starts up or undergoes topology changes, each node encodes the collected local observation data into a low-dimensional state embedding vector. Using the current subtask objective as the query and the neighbor embedding as the key, a compatibility score is calculated through scaled dot product attention. Key nodes are then selected and selective state synchronization is performed based on the attention weights output by time attention. The unmanned cluster is divided into subgroups according to node type and type weight coefficients are configured. Based on the state of key nodes after selective state synchronization, the comprehensive capability value of each node is measured and multiplied by the type weight coefficient of its subgroup to obtain the decision weight for type-individual joint ranking. The decision weight is based on the stage dependency moments. The system updates the dependency strength type weight coefficients of each task role in the current task phase within the cluster; based on the decision weights, it performs task coordination processing, synchronously calculates the target escape risk in real time, broadcasts a distress signal when the risk exceeds a threshold, and forms self-organizing task groups based on the response. Within each group, nodes multiply their decision weights by the risk weighting coefficients to obtain the group's internal decision weights. For each self-organizing task group, each node defines virtual axes and virtual nests and calculates virtual bond strength. The internal decision weights are updated based on the number of bond connections and spatial centrality. When environmental changes or task switching trigger configuration adjustments in the self-organizing task group, bond breakage or bond formation conditions are activated. This invention addresses the technical problem of insufficient adaptability of collaborative decision-making in dynamic task environments in existing unmanned swarms. By constructing dynamic decision weights based on key node states, task phases, and node capabilities, and combining them with a risk-driven self-organizing task group and configuration dynamic adjustment mechanism, it achieves the technical effect of improving the dynamic adaptability and task execution efficiency of collaborative decision-making in unmanned swarms.

[0053] Example 2, based on the same inventive concept as the unmanned cluster distributed collaborative task decision-making method in the foregoing examples, such as... Figure 3 As shown, this application provides an unmanned cluster distributed collaborative task decision-making system. The system and method embodiments in this application are based on the same inventive concept. The system includes: Synchronization module 11 is used to encode the collected local observation data into low-dimensional state embedding vectors by each node when the unmanned cluster starts or undergoes topology changes. Using the current subtask objective as the query and the neighbor embedding as the key, a compatibility score is calculated through scaled dot product attention. Key nodes are then selected and selective state synchronization is performed based on the attention weights output by the time attention. Weight configuration module 12 is used to divide the unmanned cluster into subgroups according to node type and configure type weight coefficients. Based on the state of key nodes after selective state synchronization, the comprehensive capability value of each node is measured and multiplied by the type weight coefficient of its subgroup to obtain the decision weight for type-individual joint ranking. The weights are determined based on the stage dependency matrix. The current task phase updates the dependency intensity of each task role with the type weight coefficient; the collaborative processing module 13 is used to perform task collaborative processing according to the decision weight, synchronously calculate the target escape risk in real time, broadcast a distress signal when the risk exceeds the threshold, and form a self-organizing task group according to the response. The nodes in the group multiply the decision weight by the risk weight coefficient as the decision weight of the group; the update module 14 is used to define virtual axes and virtual nests for each node of the self-organizing task group and calculate the virtual bond strength, and update the decision weight of the group according to the number of bond connections and spatial centrality. When environmental changes or task switching trigger the self-organizing task group configuration adjustment, the bond breakage condition or bond formation condition is activated.

[0054] Furthermore, the system is also used to implement the following functions: Each node collects its own local observation data, which is mapped into a low-dimensional state embedding vector through a variational autoencoder network. The local observation data includes the effective range of the sensor, the quality of the communication link, the current task execution progress, and spatial coordinates. Each node uses the target vector of its current subtask as the query and the received low-dimensional state embedding vector of the neighboring node as the key value to perform selective state synchronization by marking key nodes.

[0055] Furthermore, the system is also used to implement the following functions: By using a scaled dot product attention approach, each node calculates a compatibility score between the query and the key value, and marks neighboring nodes with compatibility scores higher than a preset spatial attention threshold as key nodes. Through a temporal attention mechanism, the state embeddings of each key node at multiple consecutive historical moments are weighted and aggregated to capture state change trends and output the attention weight ranking of each key node. Each node performs selective state synchronization according to the attention weight ranking, wherein complete local observation data is sent to neighboring nodes with attention weights exceeding a first preset synchronization threshold, and low-dimensional state embedding vectors are sent to neighboring nodes with attention weights lower than the first preset synchronization threshold but higher than the second preset threshold, so that the number of communication objects of a single node is reduced from the total number of clusters N to the number of key nodes k.

[0056] Furthermore, the system is also used to implement the following functions: The unmanned swarm is divided into multiple subgroups according to node type. Each subgroup has a defined and initialized type weight coefficient, which quantifies the degree of differential contribution of node type within each subgroup to the global decision mean field. Based on the task stage, the type weight coefficients of multiple subgroups are updated according to the stage dependency matrix. Based on the state of key nodes after selective state synchronization, the node comprehensive capability of each key node is calculated. By matching the type weight coefficients and performing multiplication, the decision weight based on type-individual joint ranking is obtained.

[0057] Furthermore, the system is also used to implement the following functions: The stage dependency matrix is ​​constructed using the dependency strength of the task stage on the task role as matrix items. The task stages include the reconnaissance stage, the cooperative search stage, the penetration stage, and the withdrawal stage. The task roles include perception, computing, communication, and attack. The dynamically evolving node comprehensive capability is calculated as Σ(role capability value × stage role dependency weight).

[0058] Furthermore, the system is also used to implement the following functions: During the task execution process of each node, the escape risk level is calculated based on the target's velocity vector and direction change rate. When the escape risk level exceeds a preset distress threshold, the current task execution node broadcasts a distress signal in a preset area where the target is located. The distress signal includes the target's location, escape risk level, and response time window. The node receiving the distress signal calculates its response priority based on its remaining energy and the spatial distance to the target. When the priority exceeds a preset response threshold, it is marked as a candidate response node and sends a response confirmation to the task execution node, forming a self-organizing task group. After the self-organizing task group is formed, each node in the group uses a decision weight based on type-individual joint ranking as a benchmark, multiplied by a risk weighting coefficient, as the group's decision weight. Nodes outside the group transfer their decision weight to response nodes within the group according to a preset ratio. Nodes outside the group are key nodes of the task execution node that do not belong to the self-organizing task group.

[0059] Furthermore, the system is also used to implement the following functions: In a self-organizing task group, each node defines its own virtual axis and virtual nest. The virtual axis represents the node's desired motion direction vector, and the virtual nest represents the node's desired spatial position in the formation. Each node calculates the virtual bond strength with its neighboring nodes based on the geometric relationship between the virtual axis and the virtual nest. The virtual bond strength is determined by a weighted combination of the spatial distance and the directional angle between nodes, and the strength of the virtual bond is quantified using a linear weighting method. The decision weights within a group are updated based on the number of virtual key connections and spatial centrality, wherein the spatial centrality is measured as the reciprocal of the average geometric distance from a node to other nodes in the group.

[0060] Furthermore, the system is also used to implement the following functions: The bond breakage condition is that the distance between nodes exceeds a preset breakage threshold or the directional angle exceeds a preset angle threshold. The bond formation condition is that a new node enters the group's perception range and its geometric relationship with the existing virtual nest meets a preset matching threshold. When a bond breakage occurs, the node exits the weight structure of the self-organizing task group. When a bond formation occurs, the new node enters the weight structure of the group. The weight distribution is automatically redistributed with bond breakage and bond reconstruction.

[0061] Example 3: Based on the inventive concept of the unmanned swarm distributed collaborative task decision-making method in the foregoing embodiments, this application also provides an electronic device, including: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the unmanned swarm distributed collaborative task decision-making method described in any one of the above embodiments.

[0062] Appendix Figure 4 This is a schematic diagram of the structure of an exemplary electronic device of this application. Figure 4 In this document, the bus architecture is represented by bus 300. Bus 300 may include any number of interconnected buses and bridges, and bus 300 connects various circuits including one or more processors represented by processor 302 and memory represented by memory 304. Bus 300 may also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 305 provides an interface between bus 300 and receiver 301 and transmitter 303. Receiver 301 and transmitter 303 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 302 is responsible for managing bus 300 and general processing, while memory 304 can be used to store data used by processor 302 during operation.

[0063] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for decision-making in a distributed collaborative task involving unmanned clusters, characterized in that, The method includes: When the unmanned cluster starts up or the topology changes, each node encodes the collected local observation data into a low-dimensional state embedding vector. Using the current subtask target as the query and the neighbor embedding as the key, the compatibility score is calculated by scaling dot product attention. Key nodes are selected and selective state synchronization is performed based on the attention weight output by time attention. The unmanned cluster is divided into subgroups according to node type and a type weight coefficient is configured. Based on the state of key nodes after selective state synchronization, the comprehensive capability value of individual nodes is measured and multiplied by the type weight coefficient of the subgroup to obtain the decision weight of type-individual joint ranking. The type weight coefficient is updated according to the dependence strength of the current task stage on each task role in the stage dependency matrix. Task collaboration is performed based on the decision weights, the target escape risk is calculated synchronously in real time, a distress signal is broadcast when the risk exceeds the threshold, and a self-organizing task group is formed based on the response. The decision weights of the nodes in the group are multiplied by the risk weighting coefficients to obtain the decision weights of the group. For self-organizing task groups, each node defines virtual axes and virtual nests and calculates virtual bond strength. The decision weights within the group are updated based on the number of bond connections and spatial centrality. When environmental changes or task switching trigger configuration adjustments in the self-organizing task group, bond breaking conditions or bond formation conditions are activated.

2. The unmanned cluster distributed collaborative task decision-making method as described in claim 1, characterized in that, Selective state synchronization is performed by filtering key nodes and ranking them according to the attention weight of time attention output, including: Each node collects its own local observation data, which is mapped into a low-dimensional state embedding vector through a variational autoencoder network. The local observation data includes the effective range of the sensor, the quality of the communication link, the current task execution progress, and the spatial coordinates. Each node uses its current subtask's target vector as the query and the received low-dimensional state embedding vector of its neighboring nodes as the key, and performs selective state synchronization by marking key nodes.

3. The unmanned cluster distributed collaborative task decision-making method as described in claim 2, characterized in that, Selective state synchronization is performed by marking key nodes, including: By using a scaled dot product attention method, each node calculates a compatibility score between the query and the key value, and marks neighboring nodes with compatibility scores higher than a preset spatial attention threshold as key nodes. By using a time attention mechanism, the state embeddings of each key node at multiple consecutive historical moments are weighted and aggregated to capture the state change trend and output the attention weight ranking of each key node. Each node performs selective state synchronization according to the attention weight sorting. Specifically, it sends complete local observation data to neighboring nodes whose attention weight exceeds the first preset synchronization threshold, and sends low-dimensional state embedding vectors to neighboring nodes whose attention weight is lower than the first preset synchronization threshold but higher than the second preset threshold. This reduces the number of communication objects for a single node from the total number of cluster N to the number of key nodes k.

4. The unmanned cluster distributed collaborative task decision-making method as described in claim 1, characterized in that, The decision weights for the type-individual joint ranking are obtained, including: The unmanned cluster is divided into multiple subgroups according to node type. Each subgroup has a defined and initialized type weight coefficient, which quantifies the degree of differential contribution of node type in each subgroup to the global decision mean field. Based on the task phase, update the type weight coefficients of multiple subgroups according to the phase dependency matrix; Based on the state of the key nodes after selective state synchronization, the overall node capability of each key node is calculated. By matching the type weight coefficients and performing multiplication, the decision weight based on type-individual joint ranking is obtained.

5. The unmanned cluster distributed collaborative task decision-making method as described in claim 4, characterized in that, The stage dependency matrix is ​​constructed with the degree of dependency of the task stage on the task role as matrix items. The task stages include the reconnaissance stage, the cooperative search stage, the penetration stage and the withdrawal stage, and the task roles include perception, computing, communication and attack. Among them, the dynamic evolution of node comprehensive ability = Σ(role ability value × stage role dependence weight).

6. The unmanned cluster distributed collaborative task decision-making method as described in claim 1, characterized in that, Form self-organizing task forces, including: During the task execution process of each node, the escape risk level is calculated based on the target's motion velocity vector and direction change rate. When the escape risk level is greater than the preset distress threshold, the current task execution node broadcasts a distress signal in the preset area where the target is located. The distress signal includes the target location, escape risk level, and response time window. The node receiving the distress signal calculates its response priority based on its remaining energy and the spatial distance to the target. When the priority exceeds a preset response threshold, it is marked as a candidate response node and sends a response confirmation to the task execution node, thus forming a self-organizing task group. After the self-organizing task group is formed, each node in the group uses the decision weight based on the type-individual joint ranking as a benchmark, multiplied by the risk weighting coefficient, as the group's decision weight. Among them, non-group nodes transfer their decision weight to the group's response nodes according to a preset ratio. Non-group nodes are key nodes in the task execution nodes that do not belong to the self-organizing task group.

7. The unmanned cluster distributed collaborative task decision-making method as described in claim 1, characterized in that, Update the decision weights within the group, including: In a self-organizing task group, each node defines its own virtual axis and virtual nest. The virtual axis represents the node's desired motion direction vector, and the virtual nest represents the node's desired spatial position in the formation. Each node calculates the virtual bond strength with its neighboring nodes based on the geometric relationship between the virtual axis and the virtual nest. The virtual bond strength is determined by a weighted combination of the spatial distance and the directional angle between nodes, and the strength of the virtual bond is quantified by a linear weighting method. The decision weights within a group are updated based on the number of virtual key connections and spatial centrality, wherein the spatial centrality is measured as the reciprocal of the average geometric distance from a node to other nodes in the group.

8. The unmanned cluster distributed collaborative task decision-making method as described in claim 7, characterized in that, The condition for a bond breaking is that the distance between nodes exceeds a preset breaking threshold or the directional angle exceeds a preset angle threshold. The condition for a bond forming is that a new node enters the group formation perception range and its geometric relationship with the existing virtual nest meets a preset matching threshold. Among them, the weight structure for nodes leaving the self-organizing task group when a bond breaks occurs, the weight structure for new nodes joining the formation when a bond forms, and the weight distribution are automatically redistributed with bond breaking and bond reconstruction.

9. An unmanned cluster distributed collaborative task decision-making system, characterized in that, The system is used to execute the unmanned cluster distributed collaborative task decision-making method as described in any one of claims 1-8, and the system includes: The synchronization module is used to encode the collected local observation data into low-dimensional state embedding vectors when the unmanned cluster starts up or the topology changes. It calculates the compatibility score by scaling dot product attention with the current subtask target as the query and the neighbor embedding as the key value, filters key nodes, and performs selective state synchronization according to the attention weight output by the time attention. The weight configuration module is used to divide the unmanned cluster into subgroups according to node type and configure type weight coefficients. Based on the state of key nodes after selective state synchronization, the module measures the comprehensive capability value of individual nodes and multiplies it with the type weight coefficient of the subgroup to obtain the decision weight for joint sorting of type and individual. The type weight coefficient is updated according to the dependence strength of the current task stage on each task role in the stage dependency matrix. The collaborative processing module is used to perform collaborative task processing based on the decision weights, synchronously calculate the target escape risk in real time, broadcast a distress signal when the risk exceeds the threshold, and form a self-organizing task group based on the response. The nodes in the group multiply the decision weights by the risk weighting coefficients to obtain the decision weights within the group. The update module is used to define virtual axes and virtual nests for each node of a self-organizing task group and calculate the virtual bond strength. It updates the decision weights within the group based on the number of bond connections and spatial centrality. When environmental changes or task switching trigger configuration adjustments in the self-organizing task group, it activates bond breaking conditions or bond formation conditions.

10. An electronic device, characterized in that, include: At least one processor; A memory that is communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the steps of the unmanned cluster distributed collaborative task decision-making method according to any one of claims 1 to 8.

Images