A Multi-Task Distributed Recruitment Method for UAV Swarms

Abstract

The present invention discloses a multi-task distributed recruitment method for UAV swarms, which is applied to the field of UAV autonomous networking and control. Aiming at the traditional consensus-based distributed method, when facing large-scale UAV swarms, there are problems of unstable and low task recruitment efficiency. The present invention conducts a multi-task distributed recruitment process based on the task demand information concentration, uses the task demand information concentration to characterize the task resource requirements, and controls the strategy of each UAV to select tasks through the update of the task demand information concentration, so as to realize a dynamic recruitment process based on the resource supply and demand relationship. And in order to reduce the locality of each UAV's decision-making, different from the distributed competition method based on auction, the present invention combines deep reinforcement learning to design a competitive recruitment mechanism, generates the priority of each UAV by deep reinforcement learning, and competes to join the task group, ensuring the stability in the task recruitment process and improving the task recruitment efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous networking and control of unmanned aerial vehicles, and particularly relates to a multi-task recruitment technology for unmanned aerial vehicle clusters. Background Art

[0002] In recent years, as an emerging autonomous cluster system technology, unmanned aerial vehicle clusters have been widely used in the civilian field, including forest fire monitoring, topographic surveying, earthwork measurement, post-disaster rescue, etc. In these scenarios, due to the limited capabilities of a single unmanned aerial vehicle, it is impossible to collect multi-angle and high-precision perception data for the perceived target, and it is also impossible to process and transmit the perception data within an effective time. Therefore, it is necessary for the unmanned aerial vehicle that discovers the task to recruit neighboring unmanned aerial vehicles to form a multi-unmanned aerial vehicle system to collaboratively complete the tasks of collecting and processing perception data.

[0003] For dynamic and complex task targets, the collaboration among multiple unmanned aerial vehicles becomes a key condition for whether the cluster can complete complex tasks. Due to the high dynamic characteristics of the environment and tasks in the unmanned aerial vehicle cluster, new tasks will continuously be generated and tasks will be executed and completed. Therefore, in the scenario of complex multi-tasks, for the high-dynamic heterogeneous resource requirements of different tasks, it is necessary to dynamically recruit unmanned aerial vehicles for each task to ensure the efficient collaboration of the unmanned aerial vehicle cluster in multi-task execution.

[0004] Existing research results can be divided into two categories: centralized task recruitment and distributed task recruitment. In terms of the design of centralized multi-task recruitment algorithms, current research mainly uses methods such as mixed integer linear programming and the Hungarian algorithm to model and solve combinatorial optimization problems for the matching of unmanned aerial vehicle clusters and fixed task targets; in the research on distributed multi-task recruitment of unmanned aerial vehicle clusters, current work mainly includes decision-making models based on distributed games, auction-based algorithms, and distributed consensus-based algorithms to achieve the distributed assignment of multi-task-oriented unmanned aerial vehicle clusters when the task targets are fixed and the task resource requirements are static.

[0005] However, applying traditional task recruitment methods to large-scale UAV clusters faces three challenges: First, in large-scale UAV clusters, the heterogeneous resource requirements of different tasks change dynamically. For example, the sensing resource requirements and computing resource requirements change dynamically over time. Maintaining global task requirement information on each UAV will introduce a large number of interaction processes. Second, due to the mutual coupling between heterogeneous resources, when allocating heterogeneous resources to different tasks, not only the matching degree between tasks and resources needs to be considered, but also the matching degree between different resources needs to be considered. This makes the traditional allocation scheme for fixed task resource requirements and independent resources unable to effectively carry out the task allocation process. Third, when facing large-scale UAV clusters with multiple heterogeneous resources coupled together, the global decision-making space is large and the decision-making complexity is high. Based on the limited computing power of UAVs, real-time decision-making processes cannot be achieved. At the same time, traditional consensus-based distributed methods, when facing large-scale UAV clusters, their convergence time will become unacceptable as the scale of the UAV cluster increases. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a multi-task distributed recruitment method for UAV clusters, which can achieve dynamic recruitment based on the relationship between resource supply and demand.

[0007] The technical solution adopted by the present invention is as follows: A multi-task distributed recruitment method for UAV clusters, based on a UAV cluster including multiple task groups. Each task group includes a task discovery UAV, several sensing UAVs, and several computing UAVs. Each sensing UAV and each computing UAV include an information cache, and the information cache stores a set of task requirement information sensed by the UAV. The set of task requirement information includes multiple task requirement information and the concentration of each task requirement information; The method includes the following steps:

[0008] S1. The task discovery UAV in a certain task group calculates the remaining resource requirements of the task according to the scale of the discovered task target.

[0009] S2. The task group conducts initial recruitment of the task group according to the remaining resource requirements of the task. The specific recruitment process is as follows:

[0010] The task group generates task requirement information according to the remaining resource requirements of the task, and calculates the concentration of the task requirement information according to the remaining resource requirements of the task.

[0011] The task requirement information is locally diffused based on the concentration of the task requirement information.

[0012] The UAVs that sense the task requirement information select candidate tasks according to the concentration of all the task requirement information they sense.

[0013] The UAV makes a decision using Deep Q-Network in the deep reinforcement learning algorithm based on the remaining resource requirements of the current candidate task and the resources it carries, determines the maximum Q value, and competes for the candidate tasks.

[0014] Until the task requirements are met or the task start time expires, the initial recruitment process of the task group is completed.

[0015] S3. The task discovery UAV divides the tasks among all the UAVs in the task group according to the resource status of each UAV in the recruited task group and the heterogeneous resource requirements of the tasks. Each UAV starts to perform the sensing and computing cooperation tasks according to the division of labor.

[0016] During the process of each UAV in a certain task group starting to perform the sensing and computing cooperation tasks according to the division of labor, it also includes the update process of the task group, specifically:

[0017] A1. The task discovery UAV collects the task execution information, monitors the task execution status in real time, and periodically notifies the members in the task group of the task execution status.

[0018] A2. Each UAV dynamically generates task requirement information according to the task execution status, and calculates the concentration of the task requirement information according to the remaining resource requirements of the task.

[0019] A3. Each UAV performs local diffusion, spatio-temporal attenuation, and linear superposition of the task requirement information according to the concentration of the task requirement information to update the strategy of each UAV to select tasks.

[0020] A4. If a UAV outside the current task group senses the task requirement information, it selects a candidate task according to the concentration of the task requirement information in the information cache, and competes to join the current task group based on the deep reinforcement learning algorithm, so as to obtain an updated task group.

[0021] A5. The task discovery UAV divides the tasks among all the UAVs in the task group according to the resource status of each UAV in the updated task group and the heterogeneous resource requirements of the tasks. Each UAV starts to perform the sensing and computing cooperation tasks according to the division of labor.

[0022] A6. If the task is completed, the current task group is disbanded; otherwise, go to step A1.

[0023] Advantages of the present invention: A multi-task distributed recruitment method for an unmanned aerial vehicle (UAV) cluster proposed by the present invention has the following advantages: The present invention can implement a multi-task dynamic recruitment process in the UAV cluster. This method comprehensively considers the resource balance and the dynamic changes in resource requirements, establishes a distributed dynamic task recruitment method according to demand, can adapt to complex and changeable task scenarios, and shows obvious advantages in terms of resource balance and task completion efficiency, enhances the timeliness of the cluster's collaborative task execution, and can support efficient and persistent multi-UAV collaborative task execution. Description of the Drawings

[0024] Figure 1 It is a schematic diagram of the dynamic recruitment process of the task group of the present invention;

[0025] Figure 2 It is a schematic diagram of the initialization recruitment method process of the task group of the present invention;

[0026] Figure 3 It is a block diagram of the multi-task group dynamic recruitment system of the UAV cluster in the present invention. Detailed Embodiments

[0027] To facilitate the understanding of the present invention by those of ordinary skill in the art, the following definitions are first made for the technical terms involved in the present invention:

[0028] 1. Task Requirement Information

[0029] Task requirement information includes the remaining computing resource requirements, remaining sensing resource requirements, and status information during the task execution of the task.

[0030] 2. Task Requirement Information Concentration

[0031] The task requirement information concentration is used to characterize the relative resource requirement size of the task, and the task requirement information in the demand information cache of each UAV characterizes the dynamics of each task requirement in a time-varying manner, so as to update the task selection strategy of the UAV.

[0032] 3. UAV Cluster

[0033] It consists of N sensing UAVs and M computing UAVs. The sensing UAVs carry various small sensing devices to collect information such as target images and signals, and the computing UAVs have good microprocessors and are responsible for processing the data collected by the sensing UAVs.

[0034] 4. Heterogeneous Resources

[0035] The heterogeneous resources of the UAVs in the present invention include sensing resources, computing resources, and communication resources. The sensing resources specifically refer to the on-board sensors of each UAV, such as small lidar, cameras, etc.; the computing resources are the remaining CPU frequencies of each UAV.

[0036] 5. Information caching

[0037] The information cache is a set of task requirement information perceived by each drone, including the perceived task requirement information and its concentration. When the drone receives the task requirement information, the task requirement information and its concentration in the information cache are updated.

[0038] 6. Task group

[0039] The task group is a set composed of sensing drones and computing drones participating in the task, which collaboratively execute the corresponding tasks.

[0040] The present invention mainly includes two parts: First, since the sensing ability and computing ability of a single drone are not sufficient to independently complete the task, each task group dynamically generates task requirement information for task group recruitment according to resource requirements and task execution status. And the resource requirements are characterized by the task requirement concentration, so as to realize multi-task on-demand recruitment; Second, in order to achieve a fast recruitment process, reduce the locality of drone decision-making, and ensure the stability of the recruitment process, we implement a task competition mechanism based on the deep reinforcement learning algorithm to improve the recruitment efficiency. Thus, based on the task requirement information concentration and the task competition mechanism, a multi-task iterative recruitment process is designed.

[0041] As Figure 1 shown, it is a schematic flow diagram of the dynamic recruitment method for the drone cluster task of the present invention, including the following steps:

[0042] S1. The task discovery drone calculates the remaining task resource requirements according to the task target scale, including sensing resource requirements and computing resource requirements.

[0043] S2. Generate task requirement information according to the remaining task resource requirements. And calculate the task requirement information concentration to perform the iterative recruitment process of the initial task group. In each round of initial recruitment, it includes task requirement information concentration update, candidate task selection based on task requirement concentration, and task group competition joining based on the deep reinforcement learning algorithm.

[0044] S3. The task discovery drone assigns tasks to all drones within the initial task group according to the resource status of each drone and the heterogeneous task resource requirements. Each drone starts to execute sensing and computing cooperation tasks according to the assignment.

[0045] S4. The task discovery drone collects task execution information, monitors the task execution status in real time, and periodically notifies the task execution status to the members within the group.

[0046] S5. Each drone dynamically generates task requirement information according to the task execution status, and calculates the task requirement information concentration according to the remaining task resource requirements.

[0047] S6. Each unmanned aerial vehicle (UAV) performs local diffusion, spatio-temporal attenuation, and linear superposition of the task requirement information according to the concentration of the task requirement information, so as to update the strategy for each UAV to select tasks.

[0048] S7. The UAV that senses the task requirement information selects candidate tasks according to the concentration of the task requirement information in the information cache, and competes to join the task group based on the deep reinforcement learning algorithm.

[0049] S8. The task discovery UAV adjusts the task division according to the result of the task group resource adjustment.

[0050] S9. If the task is completed, the task group is disbanded; otherwise, step S4 is executed.

[0051] Figure 2 This is a flowchart of the initial task group recruitment method of the present invention, which includes three steps: S1, S2, and S3.

[0052] In step S1, the task discovery UAV calculates the initial resource requirements of the task according to the task target scale. The UAV u that discovers task j i , initializes the task group T j ={u i}, and estimates the computing resource requirements and sensing resource requirements. According to the task target scale, determine the number of task sensing positions and the number of frames of sensing data at each sensing position and the number of data points included in each frame of sensing data Thus, estimate the initial computing resource amount, that is, the sum of the CPU frequencies of the computing UAVs required in the task group T j is where is a calculation function for the total number of CPU calculation cycles theoretically required for the task, that is, it can calculate the total number of CPU cycles theoretically required to process all the sensing data generated by this task by the sensing data processing algorithm. The specific form of this function needs to be determined according to the data processing algorithm specifically adopted by the task. is the limit value of the execution time of this task, that is, if the task execution time is greater than then this task fails.

[0053] In step S2, the task group performs an initial iterative recruitment process according to the remaining resource requirements of the task. Specifically, it includes the following sub-steps:

[0054] S21. The task group T j , generates task requirement information according to the remaining resource requirements of the task, and calculates the concentration of the task requirement information. The task requirement information is expressed as where id is the task identification number, p j∈{1, …, M} is the task priority. The larger the priority, the higher the urgency of executing the task. respectively represent the current initial sensing resource status and computing resource status of the task. The sensing resource requirement of task j has been met; otherwise, it has not been met. Indicates that the initial computing resource requirement of task j has been met; otherwise, it has not been met. And for the unmanned aerial vehicle i ∈ T j The task requirement information D generated init The concentration is

[0055]

[0056] Among them, Represents the current sensing resource status of the task group. Represents the task group T j The number of sensing unmanned aerial vehicles in it. Represents the current computing resource status of the task group. Represents the task group T j The total CPU frequency of the existing computing unmanned aerial vehicles in it. Is the number of neighbors of the unmanned aerial vehicle i that do not belong to the task group T j

[0057] S22. Update the task selection strategy of the unmanned aerial vehicle based on the generated concentration of the task requirement information. The task requirement information diffuses locally based on the concentration, and updates the concentration of the task requirement information in the information cache of each unmanned aerial vehicle according to the spatio-temporal attenuation and linear superposition process, so as to update the task selection strategy of each unmanned aerial vehicle in this round of recruitment process. It mainly includes the following processes:

[0058] S221. Local diffusion process: The unmanned aerial vehicle that generates the task requirement information and the unmanned aerial vehicle that receives the new task requirement information use the unmanned aerial vehicle network topology as the diffusion medium, and execute the local task requirement information diffusion process according to the concentration of the task requirement information and the task status of the neighbor unmanned aerial vehicles. The unmanned aerial vehicle that generates the new task requirement information and the unmanned aerial vehicle that receives the new task requirement information need to perform the local diffusion process of the task requirement information. The k-hop unmanned aerial vehicle node to which the task requirement information diffuses calculates the concentration of the next-hop task requirement information as where α is the spatial attenuation coefficient, and the value range of α is (0.5, 0.8). If then diffuse the task requirement information to the next hop, and the diffused neighbor needs to be outside the current task group and has not received the same task requirement information repeatedly.

[0059] S222. Spatio-temporal attenuation process: During the diffusion process of the task requirement information, the concentration of the task requirement information decays hop by hop with the diffusion hop count, that is

[0060] ​

[0061] Among them, represents the concentration of task demand information at the 0th hop, and α k represents the k-th power of α;

[0062] Moreover, for the task demand information in the UAV information cache, as time goes by, its concentration will decay in the time dimension to dynamically characterize the timeliness of this task information, that is

[0063] η j (kΔt) = βη j ((k - 1)Δt)

[0064] Among them, β is the time decay coefficient, β ∈ (0, 0.9), Δt is the decay period, that is, the concentration of task demand information in all UAV information caches decays in the time dimension every Δt; Δt ∈ {1s, 2s, …, 10s}, which can be flexibly adjusted according to the actual effect. The larger Δt is, the longer the action time of the task demand information is. At the same time, during the spatio-temporal decay process, when the concentration of task demand information η i < η min , it will immediately become invalid and the corresponding task demand information will be removed from the information cache, and η min = 1.

[0065] S223. Linear superposition process: Inspired by the ant colony foraging process, we regard the accumulation of the concentration of task demand information of the same task group at the same node as a linear superposition process to strengthen the influence of this task group on neighboring nodes. That is, during the diffusion process of task demand information, when the UAV receives the task demand information, the update method of the concentration of task demand information in its information cache is

[0066] η j = η j + η' j

[0067] Among them, η' j is the concentration of the demand information of the newly received task j at the node.

[0068] The UAV that senses the task demand information selects candidate tasks according to the concentration of all the sensed task demand information, and the probability that each task in its information cache is selected as a candidate task, that is, the strategy for the UAV i to select a task is expressed as:

[0069]

[0070] Among them, Q i is the task set in the information cache of the UAV i, and P i，j represents the probability that the UAV selects task j as a candidate task.

[0071] S24. The drone makes a decision using Deep Q-Network in the deep reinforcement learning algorithm based on the remaining resource requirements of the current candidate task and the resources it carries, determines the maximum Q value, and competes for the candidate task. Its state set, action set, and reward value are as follows:

[0072] State set: Each drone uses the remaining resource requirements of candidate task i, the remaining resource requirements of the current task j where the drone is located, its own CPU frequency, and sensing ability as the state set of the deep reinforcement learning algorithm. Among them, the remaining task resources include task priority, the remaining number of sensing targets of the task, and the remaining CPU frequency requirements of the task.

[0073] Action set: The action set is A = {0, 1}, that is, when the action is 1, it means that the current drone joins the task group to be added, and when the action is 0, it means that the current drone maintains the task allocation status of the previous time slot.

[0074] Reward value: The reward is defined as the task switching efficiency, that is where β k ∈ {0, 1} represents the type of drone k. When β k = 1, it means that drone k is a sensing drone, and vice versa, it is a computing drone.

[0075] The reward in the DQN of the present invention is an index to judge the quality of the current task adjustment strategy in the current multi-task state. The higher the resource allocation efficiency, the better the strategy is considered to be selected in this case. Through continuous exploration and learning, the deep reinforcement learning algorithm will find a better strategy to minimize the task completion time.

[0076] The competition process mainly includes the following two steps:

[0077] S241. Each drone makes a decision based on DQN according to the candidate task state and the current task state, and chooses whether to participate in the competition process of the candidate task.

[0078] S242. The drones that choose to join the same candidate task carry out the task competition and joining process. First, a competition application is sent to the task discovery drone of the task group. The application content uses the Q value corresponding to participating in the task as the drone competition priority and the sensing resources and computing resources of the drone itself. The task discovery drone sorts according to the Q value size and selects the drone with the largest Q value that meets the remaining resource requirements of the task to join the task group.

[0079] S25. If the resource requirements of the task group have been met, that is, δ c = 0 and δ s= 0, or if the task start time expires, the initial recruitment process of the task group is completed. Otherwise, return to step S21 to continue the iterative recruitment process of resources.

[0080] In step S3, the task discovery drone, based on the resource status of each drone and the heterogeneous resource requirements of the task, assigns tasks to all drones within the initial task group, and each drone starts to execute sensing and computing collaboration tasks according to the assignment. It includes the following steps:

[0081] S31. In task group j, the task discovery drone performs sensing division according to the set of required sensing positions and the number M of available sensing drones in the current group, numbers the sensing drones to obtain the set Then the set of sensing positions for drone i to perform the sensing task can be expressed as where is the number of sensing positions assigned to sensing drone and

[0082]

[0083] S32. The task discovery drone, according to the total amount of task data and the computing capabilities of each drone, performs computing task allocation, calculates the amount of sensing data that each drone needs to process, and calculates the amount of sensing data assigned to drone as:

[0084]

[0085] where, is the number of target sensing positions of task j, is the number of frames to be collected for each sensing target of task j, is the data size of each frame in task j.

[0086] As Figure 3 shown, there are multiple task groups in the drone cluster. Each group contains one task discovery drone, multiple sensing drones and computing drones. The task discovery drone collects the task execution status and notifies the group members. Each computing drone collaboratively executes the computing task, and the sensing drones collect data of each sensing target. In step S5, the task discovery drone collects task execution information and notifies the task status; in steps S5 and S6, the drones in each group generate and spread task requirement information according to the task execution status; in step S7, the drones perform the dynamic adjustment process of the task group according to the sensed task requirement information; finally, in step S8, each task discovery drone adjusts the task division of the corresponding task group after dynamic adjustment.

[0087] S4. The task discovery UAV collects task execution information, monitors the task execution status in real time, and periodically notifies the members in the group of the task execution status. It mainly includes the following steps:

[0088] S41. The discovery UAV u of task j i With a period , periodically collect the task execution status, including the remaining amount of sensed data to be processed β j (k) at the current time t = kΔt, and the set loc′ of the positions of the sensed targets for which information has not been collected yet j .

[0089] S42. The task discovery UAV obtains the current sensed data processing rate Thereby estimating the remaining execution time of the current task And calculating the task status quantity The task execution status s j (k) = (μ j (k), |loc′ j |), and notify all UAVs in the group.

[0090] S5. Each UAV generates new task demand information based on the status of neighboring UAVs and the task execution status. The new task demand information is represented as D j (id j , μ j , L, p j ). Where id j is the identification number of task j, μ j is the task status quantity, L = |loc′ j | is the number of positions of the sensed targets for which the current task has not been sensed, and p is the priority of task j; According to the new task demand information, the concentration of the task demand information generated by UAV i in task group j can be calculated as:

[0091]

[0092] S6. Similar to step S22, based on the concentration of the generated task demand information, perform local diffusion, spatio-temporal attenuation, and linear superposition of the task demand information to update the strategy for each UAV to select tasks;

[0093] S7. The UAV that senses the task demand information selects candidate tasks according to the concentration of the task demand information and the task status information in the demand information cache, and performs a competition joining process based on step S24 to adjust the task group it belongs to.

[0094] S8. The task discovery UAV adjusts the task division according to the sensing UAVs and computing UAVs included in the current task group. It mainly includes the following steps:

[0095] S81. For the unperceived task objectives Similar to step S31, reallocate the perceived task objectives;

[0096] S82. Evenly distribute the unprocessed data to all the computing UAVs in the current task group. The amount of data that the UAV i in the task group can obtain is:

[0097]

[0098] S9. If the task is completed, that is, the sensing data of all the sensing targets in the task has been collected and all the sensing data has been processed, then disband the task group; otherwise, execute step S4.

[0099] The innovations of the present invention are as follows: First, based on the task demand information concentration, a multi-task distributed recruitment process is carried out. The task resource demand is characterized by the task demand information concentration, and the strategy for each UAV to select tasks is controlled by the update of the task demand information concentration, so as to realize the dynamic recruitment process according to the resource supply and demand relationship; Second, in order to reduce the locality of the decision-making of each UAV, different from the distributed competition method based on auctions, we design a competitive recruitment mechanism by combining deep reinforcement learning, generate the priorities of each UAV by deep reinforcement learning, and competitively join the task group to ensure the stability in the task recruitment process and improve the task recruitment efficiency.

[0100] Those of ordinary skill in the art will realize that the embodiments described herein are for helping the reader understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A multi-task distributed recruitment method for UAV swarms, characterized in that The UAV cluster based on it includes multiple task groups. Each task group includes a task discovery UAV, several sensing UAVs, and several computing UAVs. Each sensing UAV and each computing UAV include an information cache, and the information cache stores a set of task requirement information sensed by the UAV. The set of task requirement information includes multiple task requirement information and the concentration of each task requirement information; the method includes the following steps: S1. The task discovery UAV in a certain task group calculates the remaining resource requirements of the task according to the scale of the discovered task target. S2. The task group conducts initial recruitment according to the remaining resource requirements of the task. The specific recruitment process is as follows: The task group generates task requirement information according to the remaining resource requirements of the task, and calculates the concentration of task requirement information according to the remaining resource requirements. The task requirement information is locally diffused based on the concentration of task requirement information; and the concentration of task requirement information in the information cache of each UAV is updated according to the spatio-temporal decay and linear superposition process, so as to update the task selection strategy of each UAV in this round of recruitment. The spatio-temporal decay process is as follows: The k-th hop UAV node to which the task requirement information spreads calculates the concentration of the next-hop task requirement information as: where α is the spatial decay coefficient; and the task requirement information in the UAV information cache decays once every Δt in the time dimension, that is η j (kΔt) = βη j ((k - 1)Δt) where β is the time decay coefficient and Δt is the decay period; Meanwhile, during the spatio-temporal attenuation process, the concentration η of task requirement information i <η min When, it fails immediately. Otherwise, the task requirement information is diffused to the next hop, and the diffused neighbor is not within the current task group, and the same task requirement information is not received repeatedly. η min represents the threshold of the concentration of task requirement information; The linear superposition process is as follows: During the diffusion process of the task requirement information, when the UAV receives the task requirement information, the update method of the concentration of task requirement information in its information cache is η j = η j + η' j where η′ j is the concentration of the demand information of the newly received task j by the node; The UAV that senses the task requirement information selects candidate tasks according to the concentration of all the sensed task requirement information. The UAV makes a decision using Deep Q-Network in the deep reinforcement learning algorithm according to the remaining resource requirements of the current candidate task and the resources it carries, determines the maximum Q value, and competes for the candidate task. Until the task requirements are met or the task start time expires, the initial recruitment process of the task group is completed. S3. The task discovery UAV assigns tasks to all UAVs in the task group according to the resource status of each UAV and the task heterogeneous resource requirements in the recruited task group, and each UAV starts to perform sensing and computing cooperation tasks according to the assignment.

2. The multi-task distributed recruitment method for an unmanned aerial vehicle cluster according to claim 1, wherein During the process that each UAV in a certain task group starts to perform sensing and computing cooperation tasks according to the assignment, the task group is also updated. Specifically: A1. The task discovery UAV collects task execution information, monitors the task execution status in real time, and periodically notifies the members in the task group of the task execution status. A2. Each UAV updates the task requirement information and the concentration of task requirement information in its own information cache according to the task execution status. A3. The task requirement information updated in step A2 is locally diffused based on the concentration of task requirement information updated in step A2, and the concentration of task requirement information in the information cache of each UAV is updated according to the spatio-temporal decay and linear superposition process, so as to update the task selection strategy of each UAV in this round of recruitment. A4. If a drone outside the current task group perceives the task demand information, it selects candidate tasks according to the concentration of the task demand information updated in step A3 in its information cache, and makes a decision based on Deep Q-Network in the deep reinforcement learning algorithm to determine the maximum Q value and conduct competition for the candidate tasks, thereby obtaining an updated task group. A5. The task discovery drone divides tasks among all the drones in the task group according to the resource status and task heterogeneous resource requirements of each drone in the updated task group, and each drone starts to perform sensing and computing cooperation tasks according to the division of labor. A6. If the task is completed, the current task group is disbanded; otherwise, step A1 is executed.

3. A multi-task distributed recruitment method for an unmanned aerial vehicle cluster according to claim 1 or 2, characterized in that, The local diffusion process is as follows: The drone that generates the task demand information and the drone that receives the task demand information use the drone network topology as the diffusion medium and execute the local task demand information diffusion process according to the concentration of the task demand information and the task status of the neighboring drones.

4. A multi-task distributed recruitment method for an unmanned aerial vehicle cluster according to claim 1 or 2, characterized in that, The specific process of candidate task competition is as follows: Each drone makes a decision based on DQN according to the candidate task status and the current task status it is in, and chooses whether to participate in the competition process of the candidate tasks; The drones that choose to join the same candidate task conduct a task competition joining process: Submit a competition application to the task discovery drone of this task group, and the application content includes: the Q value corresponding to participating in this task as the drone competition priority, and the amount of the drone's own sensing resources and computing resources. The task discovery drone sorts according to the magnitude of the Q value and selects the drone with the largest Q value that meets the remaining resource requirements of the task to join the task group.

5. A multi-task distributed recruitment method for an unmanned aerial vehicle cluster according to claim 1 or 2, characterized in that The process of the task discovery drone dividing tasks among all the drones in the task group is: C1. In task group j, the task discovery UAV performs sensing division according to the required sensing position set and the number M of available sensing UAVs in the current group, numbers the sensing UAVs to obtain a set sensing UAV The number of sensing positions assigned to is determined according to the following formula: where i = 1, 2,..., M; C2. The task discovery UAV allocates the computing tasks based on the total amount of task data and the computing capabilities of each UAV, calculates the amount of sensing data that each UAV needs to process, and calculates the allocated sensing data volume is: Among them, is the number of target perception positions for task j, the number of frames to be collected for each perceived target of task j, is the data size per frame in task j, f i represents the calculation of the drone remaining computing resources.

Images