Method and system for multi-stage task allocation of unmanned aerial vehicle cluster oriented to quality of service perception

Abstract

This application belongs to the field of UAV task allocation technology, and particularly relates to a method and system for multi-stage task allocation of UAV swarms based on quality of service awareness. The method includes: receiving a target task and decomposing it into multiple stage sub-tasks; determining the QoS requirements of any stage sub-task; based on the QoS requirements and the global state of the UAV swarm, allocating the stage sub-tasks to candidate UAV groups through a high-level planner, forming a stage-group mapping; for any candidate UAV group, selecting a primary execution UAV and a backup UAV from the candidate UAV group based on the real-time state of the UAVs, to execute the corresponding stage sub-task; obtaining the actual stage QoS utility value and task execution result after task execution; and updating the policy parameters of the high-level planner and the low-level controller based on the actual stage QoS utility value and task execution result to optimize subsequent task allocation decisions. This application improves the efficiency of UAV task allocation.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicle (UAV) task allocation technology, and in particular relates to a multi-stage task allocation method and system for UAV swarms oriented towards service quality perception. Background Technology

[0002] Accelerating the collaborative, service-oriented, and intelligent development of information network infrastructure has become an important strategic direction for the construction of new infrastructure and the development of the digital economy. Against this backdrop, intelligent collaborative systems based on drone swarms are widely used in disaster relief, environmental monitoring, precision agriculture, and emergency communications. By constructing multi-hop wireless links in a Flying Ad-hoc Network (FANET), drone swarms can achieve self-organized task allocation and collaborative execution without ground infrastructure support, demonstrating extremely high flexibility and scalability.

[0003] Currently, research on multi-UAV task allocation mainly focuses on two directions. The first is allocation methods based on static optimization or heuristic algorithms, which allocate tasks and UAVs globally through linear programming or greedy matching to minimize time or energy consumption. However, these methods typically assume stable network topology and link quality, optimizing only at a single Quality of Service (QoS) dimension, and cannot adapt to the highly dynamic channel environment and heterogeneous QoS characteristics of multi-stage tasks in FANETs. The second is dynamic task allocation methods based on Reinforcement Learning (RL). These methods learn the mapping between UAV states and task rewards, enabling a degree of environmental adaptation. However, most studies only optimize end-to-end overall performance metrics, such as average latency or system throughput, while ignoring the sequential dependencies and differentiated QoS requirements between multi-stage tasks. This single-layer strategy makes allocation decisions prone to getting trapped in local optima, failing to achieve cross-stage global performance optimization. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a method and system for multi-stage task allocation in UAV swarms based on quality of service awareness. Through a collaborative mechanism between a high-level planner and a low-level controller, it achieves dynamic perception of QoS requirements for multi-stage tasks and intelligent allocation of UAV resources, thereby improving the efficiency and accuracy of task allocation in UAV swarms.

[0005] Firstly, this application provides a multi-stage task allocation method for UAV swarms based on service quality awareness, the method comprising, Receive the target task and decompose the target task into multi-stage sub-tasks; The QoS requirements of any stage subtask are determined. Based on the QoS requirements and the global state of the UAV cluster, the stage subtask is assigned to candidate UAV groups through a high-level planner, forming a stage-group mapping. For any of the candidate drone groups, the low-level controller selects a primary execution drone and a backup drone from the candidate drone group based on the real-time status of the drones to execute the corresponding phase sub-tasks; The actual QoS utility value and task execution result after the task is executed are obtained. Based on the actual QoS utility value and task execution result, the policy parameters of the high-level planner and the low-level controller are updated to optimize subsequent task allocation decisions.

[0006] Furthermore, The phase subtasks are assigned to candidate drone groups via a high-level planner, specifically including: Obtain global status information of the drone cluster, which includes at least one of the following: location information, energy status information, communication capability information, and computing resource information of each drone, as well as network topology information of the drone cluster; The QoS requirement information and the global status information are input into the task allocation model, and the actual matching degree is calculated through the task allocation model. The actual matching degree is compared with the preset matching degree threshold to obtain the comparison result. Based on the comparison result, candidate UAV groups are determined, and a mapping relationship between the stage sub-task and the selected candidate UAV groups is established to form a stage-group mapping.

[0007] Furthermore, The actual matching degree is calculated through the task allocation model, specifically including: A QoS requirement vector is determined based on the QoS requirement information, and a state feature vector is determined based on the global state information. The QoS requirement vector and the state feature vector are input into the feature coding network to obtain the task feature embedding and the UAV feature embedding, respectively. Using the task feature embedding as the query vector and the UAV feature embedding as the key vector, the association weight between the stage subtask and each UAV is calculated through an attention mechanism. The UAV feature embeddings are weighted and aggregated based on the association weights to obtain a comprehensive feature representation; The actual matching degree is obtained by embedding the task features and performing similarity calculations on the comprehensive feature representation.

[0008] Furthermore, The association weights between the stage sub-tasks and each UAV are calculated using an attention mechanism, specifically including: The task features are embedded and linearly transformed using a first learnable parameter matrix to obtain a query vector. The embedded UAV features are linearly transformed using a second learnable parameter matrix to obtain a key vector. Calculate the dot product between the query vector and each key vector to obtain the initial association score; The initial association score is scaled by dividing it by the square root of the feature dimension to obtain the scaled association score. The scaled association scores are normalized to obtain the association weights between the stage sub-tasks and each UAV.

[0009] Furthermore, The selection of a primary and backup drone from the candidate drone group by a low-level controller based on the real-time status of the drones specifically includes: In the candidate drone group, a group coordinator drone is determined according to predetermined election rules; The group coordinator drone collects real-time status data of all candidate drones in the candidate drone group. Based on the QoS requirements and the real-time status data, calculate the real-time stage QoS utility value of each candidate drone in the candidate drone group corresponding to the stage sub-task; The candidate drones are sorted according to the real-time stage QoS utility value, and the main executing drone is determined based on the sorting result; The real-time stage QoS utility value of the candidate drones other than the main executing drone is compared with a preset utility threshold, and a backup drone is determined based on the comparison result.

[0010] Furthermore, Obtain the actual QoS utility value and task execution result after task execution, specifically including: Collect actual execution data after the completion of the sub-tasks in the aforementioned stage; Based on the actual execution data, determine whether the stage sub-task has been successfully completed and obtain the task completion status; The actual QoS utility value at the actual stage is calculated based on the actual execution data. The stage QoS utility value, the task completion status, and the resource consumption data are used together as the task execution result.

[0011] Furthermore, The actual QoS utility value at the actual stage is calculated based on the actual execution data, specifically including: For each service quality indicator in the QoS requirements, the ratio between the actual capability value of the indicator and the requirement threshold of the indicator for the stage subtask is calculated based on the actual execution data. The ratio is input into a nonlinear mapping function to obtain the normalized utility value of each service quality indicator; A preset weighting coefficient is assigned to each service quality indicator. The normalized utility value of each service quality indicator is multiplied by its corresponding weighting coefficient and then summed in a weighted manner to obtain the actual stage QoS utility value.

[0012] Furthermore, The policy parameters of the high-level planner and the low-level controller are updated based on the actual QoS utility value and task execution results, specifically including: The stage QoS utility value is used as the core reward basis. Combined with the task completion status and the resource consumption data, the high-level planner reward and the low-level controller reward are calculated respectively. The status, actions, rewards, and next status data generated during the task allocation and execution process will be stored as experience samples in the joint experience replay buffer. An attention-integrated hierarchical reinforcement learning algorithm is employed to sample experiential data from the joint experiential replay buffer. With the goal of maximizing long-term cumulative rewards, the policy network and value network parameters of the high-level planner and the low-level controller are asynchronously updated using gradient descent.

[0013] Furthermore, The method also includes introducing a dynamic adjustment mechanism during task allocation, specifically including: Real-time monitoring of global status changes in the drone swarm and environmental changes during mission execution; When a global state or environmental change is detected to exceed a preset threshold, the high-level planner is triggered to reassess the rationality of the current task allocation scheme. If the reassessment results indicate that the current solution is no longer optimal, the task reassignment process will be initiated.

[0014] Secondly, based on the same inventive concept, this application provides a multi-stage task allocation system for UAV swarms oriented towards service quality awareness, the system comprising: The receiving module is used to receive the target task and decompose the target task into multi-stage sub-tasks; The task allocation module is used to determine the QoS requirements of any stage sub-task. Based on the QoS requirements and the global state of the UAV cluster, the stage sub-task is allocated to candidate UAV groups through a high-level planner, forming a stage-group mapping. The drone selection module is used to select a primary drone and a backup drone from any candidate drone group based on the real-time status of the drones by a low-level controller, so as to execute the corresponding stage sub-task; An optimization module is used to obtain the actual stage QoS utility value and task execution result after the task is executed, and update the policy parameters of the high-level planner and the low-level controller based on the actual stage QoS utility value and task execution result to optimize subsequent task allocation decisions.

[0015] Compared with the prior art, this application has the following advantages: 1. This application employs a hierarchical task allocation framework, establishing a global mapping from task phases to UAV groups at a high level, and enabling real-time autonomous selection of UAVs within the group at a low level, thus complementing global planning with local execution. This hierarchical structure not only significantly reduces the state space dimension but also achieves policy coordination across multiple time scales, enabling the system to simultaneously consider long-term QoS optimization and short-term real-time response.

[0016] 2. The QoS utility function modeling method proposed in this application unifies multi-dimensional QoS indicators into a single utility space through nonlinear mapping and adaptively adjusts indicator weights based on task type. This method realizes the transformation from qualitative constraints to quantitative rewards, enabling reinforcement learning agents to adaptively weigh different QoS objectives during training. By dynamically updating weights based on task execution results, this application achieves joint optimization of multiple QoS objectives, effectively improving the energy efficiency and resource utilization of UAV swarms while ensuring real-time performance and reliability.

[0017] 3. This application addresses the aforementioned challenges using the Attention Integration Hierarchical Reinforcement Learning (AIHRL) algorithm. The algorithm introduces a multi-head attention mechanism into the high- and low-level reinforcement learning structure, utilizing task stage features as query vectors and UAV state features as key-value vectors. It automatically focuses on the UAV nodes with the greatest impact on QoS, significantly enhancing the feature representation capability of the policy. Furthermore, a two-layer Soft Actor-Critic policy update mechanism is employed to achieve cross-level joint training and information sharing, thereby maintaining the stability and generalization ability of policy convergence in dynamic environments.

[0018] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a service quality-aware multi-stage task allocation method for drone swarms according to an embodiment of this application is shown. Figure 2 A schematic diagram of an AIHRL algorithm design according to an embodiment of this application is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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 some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Figure 1 A flowchart illustrating a service quality-aware multi-stage task allocation method for UAV swarms according to an embodiment of this application is shown, such as... Figure 1 As shown in the figure, a multi-stage task allocation method for UAV swarms based on service quality awareness according to an embodiment of this application includes, S1, Receive the target task and decompose the target task into multi-stage sub-tasks; In this embodiment, after the target task Ti is generated, the central controller decomposes it into an ordered set of task stages, denoted as T = {s1, ..., sp}. Each task stage is considered an atomic unit of resource allocation, indexed by its parent task identifier and stage order.

[0023] S2, determine the QoS requirements of any stage subtask, and based on the QoS requirements and the global state of the UAV cluster, assign the stage subtask to candidate UAV groups through a high-level planner to form a stage-group mapping; In this embodiment of the application, the QoS requirement is a requirement vector containing multiple predefined service quality dimensions, including but not limited to minimum bandwidth requirement to ensure data transmission capacity; maximum latency requirement to constrain the real-time performance of task execution; maximum packet loss rate requirement to ensure data integrity; minimum link reliability requirement to reduce the risk of communication interruption; and minimum number of cooperating drones requirement to meet the requirements of parallel task execution.

[0024] In this embodiment of the application, step S2 specifically includes: S21, Obtain the global status information of the UAV cluster. The global status information includes at least one of the following: location information, energy status information, communication capability information, and computing resource information of each UAV, as well as the network topology information of the UAV cluster. S22, input the QoS requirement information and the global status information into the task allocation model, and calculate the actual matching degree through the task allocation model; S23, compare the actual matching degree with the preset matching degree threshold to obtain the comparison result, determine the candidate UAV group based on the comparison result, establish the mapping relationship between the stage sub-task and the selected candidate UAV group, and form a stage-group mapping.

[0025] In this embodiment of the application, the location information is obtained by using GPS or an inertial navigation system to obtain the real-time coordinates of each UAV; Energy status information is obtained by monitoring the remaining battery power and power consumption of each drone; Communication capability information includes available bandwidth, signal strength, and communication module status; Computing resource information includes CPU utilization, memory usage, and storage space; Network topology information is used to construct a connection graph between drones through link quality detection, and parameters such as link latency and packet loss rate are recorded.

[0026] In this embodiment of the application, step S22 specifically includes: S221, determine the QoS requirement vector based on the QoS requirement information, and determine the state feature vector based on the global state information; S222, the QoS requirement vector and the state feature vector are input into the feature coding network to obtain the task feature embedding and the UAV feature embedding, respectively; S223, using the task feature embedding as the query vector and the UAV feature embedding as the key vector, calculate the association weight between the stage subtask and each UAV through an attention mechanism; S224, Based on the association weights, the UAV feature embeddings are weighted and aggregated to obtain a comprehensive feature representation; S225, The task features are embedded and the comprehensive feature representation are used to calculate the similarity to obtain the actual matching degree.

[0027] In this embodiment of the application, the QoS requirement vector is a standardized vector formed by extracting parameters such as minimum bandwidth, maximum latency, maximum packet loss rate, minimum link reliability, and minimum number of cooperating drones for each stage from the task description. The state feature vector is constructed by normalizing the data of each dimension in the global state information to create a feature vector with a unified format.

[0028] In this embodiment, a multilayer perceptron (MLP) is used to map the QoS requirement vector to a high-dimensional latent space to obtain a task feature embedding representation; By processing the UAV state feature vectors using convolutional neural networks (CNNs) or microprocessors (MLPs), a UAV feature embedding representation can be obtained.

[0029] In this embodiment, the association weights corresponding to each UAV are multiplied element-wise with the corresponding UAV feature embeddings to obtain weighted UAV feature vectors; all weighted UAV feature vectors are summed to obtain a comprehensive feature representation, which integrates UAV feature information that is highly correlated with the stage sub-tasks.

[0030] In this embodiment of the application, the cosine similarity calculation method is used to calculate the cosine value of the angle between the task feature embedding vector and the comprehensive feature representation vector. This cosine value is used as the actual matching degree. The value range of the actual matching degree is [-1, 1]. The closer the value is to 1, the higher the matching degree between the stage sub-task and the candidate UAV group.

[0031] In this embodiment of the application, step S223 specifically includes: S2231, The task features are embedded and linearly transformed through the first learnable parameter matrix to obtain the query vector; S2232, The UAV features are embedded and linearly transformed using a second learnable parameter matrix to obtain a key vector; S2233, Calculate the dot product between the query vector and each key vector to obtain the initial association score; S2234, The initial association score is divided by the square root of the feature dimension and scaled to obtain the scaled association score; S2235, normalize the scaled association scores to obtain the association weights between the stage sub-tasks and each UAV.

[0032] S3, for any candidate drone group, the low-level controller selects a primary execution drone and a backup drone from the candidate drone group based on the real-time status of the drones to execute the corresponding stage sub-task; In this embodiment, the lower-level controller first obtains the real-time status information of each drone in the candidate drone group. The real-time status information includes, but is not limited to, the current remaining energy, the current task load, and the current communication quality. Then, according to the preset selection rules, such as prioritizing drones with more remaining energy, lighter task load, and better communication quality, the main execution drone is determined from the candidate drone group. The real-time stage QoS utility value of the candidate drones other than the main execution drone is compared with the preset utility threshold. Based on the comparison result, a backup drone is determined. When the main execution drone fails or cannot complete the task normally, the backup drone can take over the task in a timely manner.

[0033] In this embodiment of the application, step S3 specifically includes: S31, In the candidate drone group, a group coordinator drone is determined according to a predetermined election rule; S32, collect real-time status data of all candidate drones in the candidate drone group through the group coordinator drone; S33, Based on the QoS requirements and the real-time status data, calculate the real-time stage QoS utility value of each candidate drone in the candidate drone group corresponding to the stage sub-task; S34, sort the candidate drones according to the real-time stage QoS utility value, and determine the main executing drone based on the sorting result; S35, compare the real-time stage QoS utility value of the candidate drones other than the main execution drone with a preset utility threshold, and determine the backup drone based on the comparison result.

[0034] S4. Obtain the actual stage QoS utility value and task execution result after the task is executed, and update the policy parameters of the high-level planner and the low-level controller based on the actual stage QoS utility value and task execution result to optimize subsequent task allocation decisions.

[0035] In this embodiment of the application, step S4 specifically includes: S41, after the execution of the stage sub-task is completed, collect the actual execution data; S42, based on the actual execution data, determine whether the stage sub-task has been successfully completed, and obtain the task completion status; S43, calculate the actual stage QoS utility value based on the actual execution data; S44, the stage QoS utility value, the task completion status, and the resource consumption data are used together as the task execution result.

[0036] In this embodiment, actual execution data is collected after the completion of the stage sub-task. The actual execution data includes various performance indicators during the task execution process, such as actual bandwidth usage, actual latency, actual packet loss rate, actual link reliability, and the actual number of participating drones. Based on the actual execution data, it is determined whether the stage sub-task has been successfully completed, and the task completion status is obtained. For example, a qualified threshold is set for each performance indicator. When all indicators in the actual execution data reach or exceed the qualified threshold, the task completion status is determined to be successful; otherwise, it is considered a failure.

[0037] In this embodiment of the application, step S43 specifically includes: S431, for each service quality indicator in the QoS requirement, calculate the ratio between the actual capability value of the indicator and the requirement threshold of the indicator for the stage subtask based on the actual execution data; S432, The ratio is input to a nonlinear mapping function to obtain the normalized utility value of each service quality indicator; S433, assign a preset weight coefficient to each service quality indicator, multiply the normalized utility value of each service quality indicator by its corresponding weight coefficient, and then perform a weighted sum to obtain the actual stage QoS utility value.

[0038] In this application embodiment, in order to uniformly measure the performance of multi-stage tasks under different QoS dimensions, this application designs a weighted QoS utility function to measure the matching effect of a specific task stage and the UAV combination.

[0039] Assuming the task phase The corresponding QoS requirements are:

[0040] in, It is the minimum bandwidth required for the stage; This is the maximum acceptable latency; This is the maximum allowable packet loss rate; It represents the minimum link reliability; It is the minimum number of collaborative drones.

[0041] To comprehensively consider the contribution of different QoS indicators, a QoS capability ratio function is defined:

[0042] in This represents the capability value of drone u on the index q. This represents the threshold for task requirements. The above ratio is mapped to normalized utility, and non-linear weighting is achieved through a sigmoid function:

[0043] in To balance the threshold, Control the steepness of the mapping.

[0044] The final definition of stage-level QoS utility is:

[0045] This function is used to comprehensively evaluate the overall performance of the UAV combination across multiple QoS dimensions and is a key reward criterion in the reinforcement learning algorithm of this application.

[0046] In this embodiment of the application, step S4 further includes: S45, using the stage QoS utility value as the core reward basis, and combining the task completion status and the resource consumption data, calculate the high-level planner reward and the low-level controller reward respectively. S46, store the status, actions, rewards and next status data generated during the task allocation and execution process as experience samples into the joint experience replay buffer; S47. An attention-integrated hierarchical reinforcement learning algorithm is used to sample experience data from the joint experience replay buffer. With the goal of maximizing long-term cumulative reward, the policy network and value network parameters of the high-level planner and the low-level controller are updated asynchronously through gradient descent.

[0047] This application proposes an attention ensemble hierarchical reinforcement learning algorithm, AIHRL. Figure 2 A schematic diagram of an AIHRL algorithm design according to an embodiment of this application is shown, such as... Figure 2 The algorithm, as shown, builds upon the traditional Hierarchical Reinforcement Learning (HRL) framework by introducing an attention mechanism and a two-layer Discrete Soft Actor-Critic (DSAC) structure to enhance feature association modeling capabilities and policy training stability. The AIHRL algorithm consists of two parts: a high-level agent (Global Agent) and a low-level agent (Local Agent), corresponding to the high- and low-level sub-tasks of HMDP, respectively. The two agents share a portion of the shared encoder network and achieve information exchange and policy sharing through a joint replay buffer.

[0048] In terms of implementation, the system adopts an asynchronous alternating update mechanism, first training low-level agents to stabilize local QoS, and then jointly optimizing high-level policies.

[0049] To address the challenges of a large number of drones and high feature dimensionality, AIHRL introduces an attention module to achieve dynamic matching between tasks and drone features.

[0050] The specific calculation process is as follows: Feature embedding make As the feature matrix of the task phase, let As a state feature matrix of the UAV:

[0051] Attention weight calculation Will Mapping to queries transforms the features of each drone into a query vector:

[0052] The requirements of each task phase will be transformed into a key vector by mapping to keys.

[0053] Mapping to values ​​transforms resources into value vectors:

[0054] The resulting attention score matrix is:

[0055] Where d is the feature dimension.

[0056] Feature aggregation The resource information is aggregated as follows, encoding the compatibility of each resource with each task stage:

[0057] This aggregation feature is input into the policy network to achieve a dynamic correlation between QoS requirements and UAV capabilities.

[0058] The introduction of the attention module can automatically focus on the drone nodes that have the greatest impact on task QoS, thereby improving the selectivity and generalization of the allocation strategy.

[0059] At each decision-making step, the high-level agent receives the global state. And use the attention output to generate the probability distribution of the drone swarm for each available stage. Strategy Parameterized by the DSAC network, through entropy regularization terms Balance exploration and utilization. The goal is:

[0060] The evaluation network Qhigh is updated by minimizing the time difference loss between the predicted reward aggregated from the low-level agent and the target reward.

[0061] Given a fixed high-level allocation Then, the lower-level agent will assign attributes. Considerations related to mission phases, and the attributes of all drones in the group. Then, an attention module is used to generate a compatibility probability distribution for all drones in the swarm; the highest compatibility probability distribution is selected. As the main drone for the mission The remainder are reserved. DSAC objective optimization is used. :

[0062] Train a low-level critic to predict short-term QoS rewards and balance network performance with drone energy.

[0063] In this embodiment of the application, the method further includes introducing a dynamic adjustment mechanism during the task allocation process, specifically including: Real-time monitoring of global status changes in the drone swarm and environmental changes during mission execution; When a global state or environmental change is detected to exceed a preset threshold, the high-level planner is triggered to reassess the rationality of the current task allocation scheme. If the reassessment results indicate that the current solution is no longer optimal, the task reassignment process will be initiated.

[0064] Based on the above method, this application also provides a service quality-aware multi-stage task allocation system for UAV swarms, corresponding to the above method, the system comprising: The receiving module is used to receive the target task and decompose the target task into multi-stage sub-tasks; The task allocation module is used to determine the QoS requirements of any stage sub-task. Based on the QoS requirements and the global state of the UAV cluster, the stage sub-task is allocated to candidate UAV groups through a high-level planner, forming a stage-group mapping. The drone selection module is used to select a primary drone and a backup drone from any candidate drone group based on the real-time status of the drones by a low-level controller, so as to execute the corresponding stage sub-task; An optimization module is used to obtain the actual stage QoS utility value and task execution result after the task is executed, and update the policy parameters of the high-level planner and the low-level controller based on the actual stage QoS utility value and task execution result to optimize subsequent task allocation decisions.

[0065] Based on the same inventive concept disclosed above, this application also provides an electronic device. The electronic device of this application includes at least one processor and at least one memory electrically connected to the processor. The memory is electrically connected to the processor, wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described above.

[0066] It should be noted that the electrical connections between the above-mentioned units do not necessarily represent the connections between lines. Indirect connections are applicable to the embodiments of this application as long as they achieve the purpose of this application.

[0067] Based on the same inventive concept, this application also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the above method.

[0068] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multi-stage task allocation method for UAV swarms oriented towards service quality awareness, characterized in that, The method includes, Receive the target task and decompose the target task into multi-stage sub-tasks; The QoS requirements of any stage subtask are determined. Based on the QoS requirements and the global state of the UAV cluster, the stage subtask is assigned to candidate UAV groups through a high-level planner, forming a stage-group mapping. For any of the candidate drone groups, the low-level controller selects a primary execution drone and a backup drone from the candidate drone group based on the real-time status of the drones to execute the corresponding phase sub-tasks; The actual QoS utility value and task execution result after the task is executed are obtained. Based on the actual QoS utility value and task execution result, the policy parameters of the high-level planner and the low-level controller are updated to optimize subsequent task allocation decisions.

2. The method according to claim 1, characterized in that, The phase subtasks are assigned to candidate drone groups via a high-level planner, specifically including: Obtain global status information of the drone cluster, which includes at least one of the following: location information, energy status information, communication capability information, and computing resource information of each drone, as well as network topology information of the drone cluster; The QoS requirement information and the global status information are input into the task allocation model, and the actual matching degree is calculated through the task allocation model. The actual matching degree is compared with the preset matching degree threshold to obtain the comparison result. Based on the comparison result, candidate UAV groups are determined, and a mapping relationship between the stage sub-task and the selected candidate UAV groups is established to form a stage-group mapping.

3. The method according to claim 2, characterized in that, The actual matching degree is calculated through the task allocation model, specifically including: A QoS requirement vector is determined based on the QoS requirement information, and a state feature vector is determined based on the global state information. The QoS requirement vector and the state feature vector are input into the feature coding network to obtain the task feature embedding and the UAV feature embedding, respectively. Using the task feature embedding as the query vector and the UAV feature embedding as the key vector, the association weight between the stage subtask and each UAV is calculated through an attention mechanism. The UAV feature embeddings are weighted and aggregated based on the association weights to obtain a comprehensive feature representation; The actual matching degree is obtained by embedding the task features and performing similarity calculations on the comprehensive feature representation.

4. The method according to claim 3, characterized in that, The association weights between the stage sub-tasks and each UAV are calculated using an attention mechanism, specifically including: The task features are embedded and linearly transformed using a first learnable parameter matrix to obtain a query vector. The embedded UAV features are linearly transformed using a second learnable parameter matrix to obtain a key vector. Calculate the dot product between the query vector and each key vector to obtain the initial association score; The initial association score is scaled by dividing it by the square root of the feature dimension to obtain the scaled association score. The scaled association scores are normalized to obtain the association weights between the stage sub-tasks and each UAV.

5. The method according to claim 1, characterized in that, The selection of a primary and backup drone from the candidate drone group by a low-level controller based on the real-time status of the drones specifically includes: In the candidate drone group, a group coordinator drone is determined according to predetermined election rules; The group coordinator drone collects real-time status data of all candidate drones in the candidate drone group. Based on the QoS requirements and the real-time status data, calculate the real-time stage QoS utility value of each candidate drone in the candidate drone group corresponding to the stage sub-task; The candidate drones are sorted according to the real-time stage QoS utility value, and the main executing drone is determined based on the sorting result; The real-time stage QoS utility value of the candidate drones other than the main executing drone is compared with a preset utility threshold, and a backup drone is determined based on the comparison result.

6. The method according to claim 5, characterized in that, Obtain the actual QoS utility value and task execution result after task execution, specifically including: Collect actual execution data after the completion of the sub-tasks in the aforementioned stage; Based on the actual execution data, determine whether the stage sub-task has been successfully completed and obtain the task completion status; The actual QoS utility value at the actual stage is calculated based on the actual execution data. The stage QoS utility value, the task completion status, and the resource consumption data are used together as the task execution result.

7. The method according to claim 6, characterized in that, The actual QoS utility value at the actual stage is calculated based on the actual execution data, specifically including: For each service quality indicator in the QoS requirements, the ratio between the actual capability value of the indicator and the requirement threshold of the indicator for the stage subtask is calculated based on the actual execution data. The ratio is input into a nonlinear mapping function to obtain the normalized utility value of each service quality indicator; A preset weighting coefficient is assigned to each service quality indicator. The normalized utility value of each service quality indicator is multiplied by its corresponding weighting coefficient and then summed in a weighted manner to obtain the actual stage QoS utility value.

8. The method according to claim 7, characterized in that, The policy parameters of the high-level planner and the low-level controller are updated based on the actual QoS utility value and task execution results, specifically including: The stage QoS utility value is used as the core reward basis. Combined with the task completion status and the resource consumption data, the high-level planner reward and the low-level controller reward are calculated respectively. The status, actions, rewards, and next status data generated during the task allocation and execution process will be stored as experience samples in the joint experience replay buffer. An attention-integrated hierarchical reinforcement learning algorithm is employed to sample experiential data from the joint experiential replay buffer. With the goal of maximizing long-term cumulative rewards, the policy network and value network parameters of the high-level planner and the low-level controller are asynchronously updated using gradient descent.

9. The method according to claim 1, characterized in that, The method also includes introducing a dynamic adjustment mechanism during task allocation, specifically including: Real-time monitoring of global status changes in the drone swarm and environmental changes during mission execution; When a global state or environmental change is detected to exceed a preset threshold, the high-level planner is triggered to reassess the rationality of the current task allocation scheme. If the reassessment results indicate that the current solution is no longer optimal, the task reassignment process will be initiated.

10. A multi-stage task allocation system for UAV swarms oriented towards quality of service awareness, characterized in that, The system includes: The receiving module is used to receive the target task and decompose the target task into multi-stage sub-tasks; The task allocation module is used to determine the QoS requirements of any stage sub-task. Based on the QoS requirements and the global state of the UAV cluster, the stage sub-task is allocated to candidate UAV groups through a high-level planner, forming a stage-group mapping. The drone selection module is used to select a primary drone and a backup drone from any candidate drone group based on the real-time status of the drones by a low-level controller, so as to execute the corresponding stage sub-task; An optimization module is used to obtain the actual stage QoS utility value and task execution result after the task is executed, and update the policy parameters of the high-level planner and the low-level controller based on the actual stage QoS utility value and task execution result to optimize subsequent task allocation decisions.

Images