A distributed dynamic unmanned aerial vehicle task scheduling method based on digital twinning

Abstract

The application discloses a kind of distributed dynamic unmanned aerial vehicle task scheduling methods based on digital twinning, applied to unmanned aerial vehicle fleet network field for Internet of Things, for the problem that fixed unmanned aerial vehicle fleet network configuration cannot satisfy the dynamic change of task demand in task execution process, the present application is according to the task demand currently issued by coordination unmanned aerial vehicle, unmanned aerial vehicle is converted by means of multi-period distributed dynamic crowd sourcing algorithm, the non-display between task demand and computing resource and communication resource demand is transformed, and distributed intelligence is added to task, to form unmanned aerial vehicle fleet;Unmanned aerial vehicle group builds the digital twin of unmanned aerial vehicle group network, rehearses unmanned aerial vehicle group task execution effect, plans unmanned aerial vehicle cluster configuration and logical topology, estimates network delay upper bound, feedback adjustment fleet crowd sourcing strategy, realizes unmanned aerial vehicle fleet dynamic supplementary configuration, reduces task cost while guaranteeing task demand.

Description

Technical Field

[0001] This invention belongs to the field of drone swarm networks for the Internet of Things, and specifically relates to a supplementary drone task scheduling technology that adapts to dynamic task requirements during task execution. Background Technology

[0002] With the continuous development of wireless communication technology, drone swarms have demonstrated great potential in emergency search and rescue (SAR) applications, aiming to provide satisfactory flexibility and scalable network performance. Members of a drone swarm are equipped with artificial intelligence (AI)-enhanced computing units and various sensors, enabling them to collaboratively complete complex detection and surveillance tasks that a single drone cannot accomplish. However, due to the uncertainty of mission type and distribution within the detection area, drones cannot be pre-deployed to form a fixed swarm to meet mission latency and resource requirements. Coupled with the high mobility of drones and the unpredictable nature of missions, the dynamic matching of complex mission requirements with limited drone resource supply becomes a challenge for drone networks. In this context, drone task allocation is crucial for achieving real-time and mission-driven drone swarm deployment.

[0003] Existing research focuses on designing centralized task allocation schemes. The performance of centralized schemes largely depends on prior information. The high mobility of UAVs and the unpredictable nature of tasks necessitate frequent global information collection by the central control node, creating a communication bottleneck. Distributed task allocation schemes benefit from the limitation of high-frequency interaction range, while also having low communication overhead and allowing for local adjustments to swarm configurations, thus attracting considerable attention, such as auction algorithms and consensus algorithms. However, current distributed schemes neglect the coupling relationship between task allocation and resource scheduling. A single swarm configuration or resource collaboration lacks coordinated feedback adjustments, failing to ensure task execution efficiency. Furthermore, due to the dynamic changes in task requirements during execution, the fixed UAV swarm needs frequent rebuilding and resource reallocation, resulting in time-consuming network performance statistics that significantly lag behind the real-time nature of task execution. Therefore, designing a real-time supplementary swarm configuration method that combines task scheduling and resource allocation without affecting the existing swarm configuration, while ensuring task requirements are met and reducing task costs, remains a challenge. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention employs a distributed dynamic UAV task scheduling method based on digital twins. It constructs a collaborative network architecture for task scheduling and resource allocation based on digital twins to achieve a closed-loop swarm collaborative crowdsourcing process encompassing task allocation, resource allocation, pre-assessment, and feedback adjustment. Distributed intelligent UAVs join the swarm, and a learning model enables the non-explicit conversion between task requirements and UAV resources. By constructing a digital twin of the UAV swarm network, a model for collecting and processing perception data is derived, the network logical topology is allocated and pre-simulated, network task execution time is estimated, and feedback adjustments are made to the crowdsourcing strategy, enabling dynamic multi-cycle replenishment of the UAV swarm.

[0005] The technical solution adopted in this invention is: a distributed dynamic drone task scheduling method based on digital twins. The application scenario is: coordinating drones to crowdsource ordinary drones to form a drone swarm to perform perception tasks. Each perception task needs to achieve a perception accuracy requirement within the end-to-end latency. The drone swarm includes a coordinating drone and several ordinary drones that have joined the perception task. During the task execution, the ordinary drones collect perception data and collaboratively complete the processing of the perception data, and summarize the processed perception results to the coordinating drone. During the formation of the drone swarm, the coordinating drone is located directly above the perception task and is responsible for publishing task request messages. Idle ordinary drones receive task request messages and intelligently select to join the task.

[0006] The scheduling method includes the following steps:

[0007] S1. Coordinate the drones to determine the latest mission requirements, the current configuration of the drone swarm, the mission execution status and logical topology, and publish the mission requirement message to start the crowdsourcing process. The mission requirements include the perception accuracy requirement and the mission latency requirement. The current configuration of the drone swarm includes the size of the drone swarm to which the mission belongs and the number of perception and computing resources of the drones. The mission execution status refers to the mission execution time based on the current logical topology and the perception accuracy that can be achieved.

[0008] S2. When an idle drone receives a task request message, it makes an intelligent decision based on a multi-agent reinforcement learning model as to whether to join the task and which task to join.

[0009] S3. Based on the existing drone swarm, coordinate the drones to construct a new virtual digital twin of the drone swarm, combining it with the newly added ordinary drones for the mission. In the constructed digital twin, the logical topology allocation algorithm is used to assign the perception business data flow and business flow direction to each of the newly added ordinary drones for the mission, and to rehearse the logical topology of the drone swarm during the mission execution process, and calculate the end-to-end latency of the business flow.

[0010] Based on the improved perception accuracy of the newly added general drone's perception resources and the end-to-end latency of the new business flow logic topology, the drone evaluation is coordinated to determine whether to accept the drone's addition to the task.

[0011] S4. If all the ordinary drones that want to join the task are evaluated, the reward is given to the multi-agent reinforcement learning model for further iterative training, and then the process returns to S1 to enter the next crowdsourcing cycle.

[0012] The beneficial effects of this invention are as follows: Addressing the dynamic matching problem between complex task requirements and limited UAV resource supply, this invention couples task allocation and resource scheduling, designing a closed-loop swarm collaborative crowdsourcing process encompassing "task allocation, resource allocation, pre-assessment, and feedback adjustment." By coordinating UAVs with digital twin technology, a digital twin of the UAV swarm network is constructed, pre-simulating network latency during task execution. Using task execution performance as a comprehensive guide, the swarm size and logical topology are iteratively supplemented, eliminating statistical processes in actual network execution and avoiding swarm reconstruction. By updating reinforcement learning neural network parameters, the non-explicit relationship between task requirements and resources is explored, thereby achieving truly task-demand-driven joint allocation of tasks and resources. Latency assessment results are used as collaborative feedback to adjust swarm configuration, guiding UAVs to achieve convergent distributed decision-making, reducing task costs while ensuring task requirements are met. Attached Figure Description

[0013] Figure 1 This is a flowchart of the closed-loop cluster collaborative crowdsourcing process of the present invention;

[0014] Figure 2 This is a schematic diagram of the learning model structure of the multi-agent reinforcement learning multi-cycle distributed dynamic crowdsourcing algorithm of the present invention;

[0015] Figure 3 This is a flowchart of the multi-agent reinforcement learning multi-cycle distributed dynamic crowdsourcing algorithm in this invention. Detailed Implementation

[0016] To facilitate understanding of this invention by those skilled in the art, the technical terms involved in this invention are first defined as follows:

[0017] 1. Drone swarm

[0018] A drone swarm consists of a coordinating drone and other regular drones already participating in the corresponding mission. k There are a total of k responsible for performing perception tasks. Each of the sensing tasks requires an end-to-end latency τ. kThe system aims to achieve the required perception accuracy. A coordinating drone positioned directly above the mission is responsible for broadcasting mission request messages, attracting nearby drones to join the mission, and assessing the compatibility of the drones with the mission.

[0019] 2. Digital Twin

[0020] Based on the physical formation provided by the formation algorithm, the drones are coordinated to construct a digital twin of a specified drone swarm in virtual space to simulate the mission execution process.

[0021] 3. Task execution time

[0022] The execution time of a drone swarm mission refers to the end-to-end latency of the maximum sensing data stream according to the logical topology. This end-to-end latency includes the time for ordinary drones to collect, transmit, and process sensing data. If the drone swarm U executing mission k... k If each sensing data stream can be processed within a specified time, then task k can be completed within the specified time. The logical topology of the UAV swarm includes the internal flow of sensing data streams. After collecting sensing data, ordinary UAVs process the data locally or select other ordinary UAVs for collaborative processing. Finally, the sensing results are aggregated at the coordinator for unified collection. Assume δ i，j =1 indicates that ordinary drone i selects ordinary drone j for collaborative processing; otherwise, δ i，j =0. Considering the very small amount of data, the process of transmitting the final perception results to the coordinating drone is omitted in this invention.

[0023] This invention mainly comprises two parts: First, addressing the problem that fixed UAV swarm network configurations cannot meet dynamically changing task requirements during task execution, this invention constructs a joint network collaboration architecture based on digital twins for task scheduling and resource allocation to achieve a closed-loop UAV swarm collaborative crowdsourcing process encompassing "task allocation, resource allocation, pre-evaluation, and feedback adjustment." The digital twin of the UAV swarm network is coordinated to allocate computing and communication resources, adjust the logical topology, estimate the upper bound of network latency, and provide feedback to adjust the crowdsourcing strategy, enabling dynamic replenishment of the UAV swarm, ensuring task latency requirements are met, and avoiding frequent swarm reorganization and reconstruction. Second, to achieve joint optimization of task configuration and resource allocation, a multi-period distributed dynamic crowdsourcing algorithm based on multi-agent reinforcement learning is proposed to dynamically build the UAV swarm. This algorithm achieves a non-explicit conversion between task latency requirements and computing and communication resource requirements, reducing task costs while ensuring task resource satisfaction. The algorithm predicts the task execution time under the current network configuration and rewards and corrects the intelligent construction strategy of the coordinated UAVs to explore dynamically task-demand-driven UAV swarms.

[0024] likeFigure 1 The diagram shown is a flowchart of the closed-loop swarm collaborative crowdsourcing process of the present invention. A distributed dynamic UAV task scheduling method based on digital twins includes the following steps:

[0025] S1. Coordinate the drones to determine the latest mission requirements, the current configuration of the drone swarm, the mission execution status and logical topology, and publish the mission requirement message to start the crowdsourcing process. The mission requirements include the perception accuracy requirement and the mission latency requirement. The current configuration of the drone swarm includes the size of the drone swarm to which the mission belongs and the number of perception and computing resources of the drones. The mission execution status refers to the mission execution time based on the current logical topology and the perception accuracy that can be achieved.

[0026] S2. When an idle drone receives a task request message, it uses a multi-period distributed dynamic crowdsourcing algorithm based on a multi-agent reinforcement learning model to intelligently decide whether to join and which task to join.

[0027] S3. Based on the existing drone swarm, coordinate the drones to construct a new virtual digital twin of the drone swarm, combining it with the newly added ordinary drones for the mission. In the constructed virtual digital twin, the logical topology allocation algorithm is used to assign the perception business data flow and business flow direction to each of the newly added ordinary drones for the mission, and to rehearse the logical topology of the drone swarm during the execution of the mission, and calculate the end-to-end latency of the business flow.

[0028] Based on the improved perception accuracy of the newly added general drone's perception resources and the end-to-end latency of the new business flow logic topology, the drone evaluation is coordinated to determine whether to accept the drone's addition to the task.

[0029] S4. If all the ordinary drones that want to join the task are evaluated, the reward is given to the multi-agent reinforcement learning model for further iterative training, and then the process returns to S1 to enter the next crowdsourcing cycle.

[0030] In step S1, the drones are coordinated to determine the latest mission requirements, the current configuration of the drone swarm, the mission execution status and logical topology, and to publish the mission requirement message. This includes the following sub-steps:

[0031] S11. When task requirements change, including changes in task perception accuracy requirements or task latency requirements, the coordinating drone will remove service flows and related drones that exceed the task latency requirements from the existing drone swarm configuration, retaining the remaining drone swarm configuration to form a preserved logical topology. If the preserved swarm configuration cannot meet the latest task requirements, the coordinating drone will issue a task requirement message to attract drones to join the task.

[0032] The mission request message includes the following: the current perception accuracy of the mission, the mission execution time, the upper limit of the mission execution latency, the size of the current drone swarm to which the mission belongs, and the mission coordinates.

[0033] In step S2, the idle drone, based on the received task request message and a multi-period distributed dynamic crowdsourcing algorithm using a multi-agent reinforcement learning model, intelligently decides whether to join and which task to join. This includes the following sub-steps:

[0034] Figure 2 This is a schematic diagram of the learning model structure of the multi-agent reinforcement learning multi-period distributed dynamic crowdsourcing algorithm of the present invention. Each ordinary UAV acts as an agent to maintain and iteratively train the neural network parameters.

[0035] S21. If a regular drone is not currently performing any task, does not belong to any drone swarm, and is not waiting for a response from a coordinating drone, then the regular drone is currently in an idle state. The regular drone collects task request messages, forms an observation state, and intelligently decides whether to join and, if so, which task's drone swarm to join. The state, action, and reward rules for the multi-agent reinforcement learning multi-period distributed dynamic crowdsourcing algorithm are as follows:

[0036] state space Including the drone's own state {λ i C i I i x i y i , z i} and task requirement status Where μ i C represents the sensor parameters of the UAV, used to characterize the UAV's perception performance. i I represents the CPU frequency of the drone itself, and the amount of computing resources available to the drone. i Indicates whether the drone is currently in an idle state, x i y i , z i P represents the drone's own coordinates. k For the task Current perception accuracy, d k Let τ be the execution time of task k. k Let |U be the upper bound of the execution delay of task k. k | represents the current drone swarm size for task k, x k y k , z k Let k be the coordinate of the task.

[0037] Action space a i∈{0, ..., K}, where a i =0 indicates that the drone will not participate in the mission. Indicates that the drone has joined the mission.

[0038] Reward r i (t) consists of the task benefits generated in this crowdsourcing cycle, defined as the quotient of the improved perception accuracy in this cycle and the number of drones added to the drone swarm in this cycle. P k (t), P k (t-1) represents the perception accuracy of task k in the current period and the previous period, respectively, |U k (t)|,|U k (t-1) represents the swarm size of task k in the current period and the previous period, respectively, and α is the parameter coefficient.

[0039] Figure 3 This is a flowchart of the multi-agent reinforcement learning-based multi-cycle distributed dynamic crowdsourcing algorithm in this invention. A typical UAV takes the current observation state and the action from the previous crowdsourcing cycle as input, and selects the function with the maximum state-action value Q. π The action execution of (s, a) is performed. After completing this crowdsourcing cycle and receiving feedback from the coordinating drone, the reward is calculated and the neural network parameters are iteratively updated.

[0040] In step S3, the coordinated drones, based on the existing drone swarm, combine newly added general drones to construct a new virtual digital twin of the drone swarm. Within this virtual digital twin, a logical topology allocation algorithm is used to assign perception service data flow directions to each newly added general drone, and the logical topology during the drone swarm's task execution is rehearsed, calculating the end-to-end latency of the service flow. Based on the improved perception accuracy of the newly added general drones and the end-to-end latency of the new service flow logical topology, the coordinated drones evaluate whether to accept the addition of that drone to the task, specifically including the following sub-steps:

[0041] S31. To describe the generation, forwarding, and processing of sensing data traffic in a drone swarm, we established a data collection model, a link model, and a processing model. To realistically simulate actual task execution, we constructed a drone sensing data collection and processing model and a sensing data transmission model between drones, and derived the task execution latency.

[0042] Data collection model: We assume that the sensor perception accuracy of each drone is μ i The higher the value, the more sensory data is collected. Therefore, the amount of sensory data for drone i is λ. i =βμ i λ s, where β is a constant, λ s Let be the amount of data collected per unit of precision. Then, the time it takes for the UAV to collect sensing data can be expressed as:

[0043]

[0044] Among them, T s The time required to collect unit-sensing data.

[0045] Link Model: To avoid mutual interference between drone swarms, each drone swarm occupies a non-overlapping channel, but drones within each swarm operate on the same channel. Within the drone swarm, line-of-sight (LoS) communication is the primary consideration. The average signal-to-noise ratio (SNR) between drone i and drone j can be expressed as:

[0046]

[0047] Where, p i，j For transmission power, For channel gain, g i，j For channel power gain, L i，j =20log d i，j +20log f0+20log(4π / c) is the free space propagation loss, d i，j Where f0 is the distance between drones, c is the carrier frequency of the channel, and N is the speed of light. gw It is the power of Gaussian white noise. Therefore, the transmission rate from drone i to drone j can be obtained through r. i，j =B log2(1+γ) i，j The transmission time can be calculated as follows.

[0048] Processing Model: We use the same model as the link model to simulate the service process of the UAV handling sensor data traffic. For the Central Processing Unit (CPU) with a frequency of C... i The processing speed of drone i can be... Where C represents the CPU computation cycles required per bit of sensor data. When drone i selects drone j for collaborative processing, other drones in the swarm may also select drone j for collaborative computation. Assuming that drone j's computing power is evenly distributed among each drone requesting its collaborative computation, the computation time can be expressed as... δ i，j This indicates that drone i selects drone j for collaborative processing; δ i，j=1 indicates that drone i selects drone j for collaborative processing; otherwise, δ i，j =0.

[0049] The logical topology of a drone swarm includes the internal flow of the sensing data stream. Ordinary drones process the sensing data locally or collaborate with other drones, and the final sensing results are aggregated by the coordinating drone for unified collection. Considering the very small data volume, we omit the process of transmitting the final sensing results to the coordinating drone. Therefore, assuming drone i selects drone j for collaborative processing of the collected sensing data, the end-to-end latency of the sensing data stream from drone i to drone j can be expressed as: The task execution latency is the end-to-end latency of the maximum sensing data stream of the UAV swarm, which can be expressed as:

[0050] S32. Construct an optimization model with the goal of minimizing task cost while ensuring task requirements are met. For task k, the drone swarm U k and collaborative computing scheduling δ i，j The optimization problem is:

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] C5: I i =1, i∈U k

[0057]

[0058] C7: δ i，j ∈{0,1}, i,j∈U k

[0059] In this framework, C1 represents the required perception accuracy for the drone swarm, P is the perception accuracy threshold, and C2 is the upper bound of the end-to-end latency. C1 and C2 ensure that the drone swarm's task requirements are met, C3 prevents the swarm from becoming too large, and C4 and C5 ensure that only idle drones can join a single swarm. C6 and C7 are variable constraints. Based on the optimization problem objective and constraints, a reward function for a multi-period distributed dynamic crowdsourcing algorithm based on multi-agent reinforcement learning was designed. The specific reward function has been described in detail above.

[0060] The perception accuracy threshold P and the drone swarm size U are set according to the mission requirements. Considering the high perception accuracy requirements of the mission, P is set to 0.9. The drone swarm size U is set to the number of ordinary drones divided by the number of missions to avoid assigning too many drones to a single mission.

[0061] S33. The above problem is non-convex and difficult to solve using traditional optimization methods. Therefore, a multi-period distributed dynamic crowdsourcing algorithm and a logical topology allocation algorithm based on multi-agent reinforcement learning are used to iteratively solve the UAV swarm U. k and collaborative computing scheduling δ i，j The coordination drones, based on the received information about drones wishing to join, evaluate the benefits of adding drones to the drone swarm according to the logical topology allocation algorithm, determine whether to accept the addition of the corresponding drone, and establish the business flow for the joining drones, forming the drone swarm's logical topology. In each iteration of the logical topology allocation algorithm, at most one drone is allowed to join the task. Each round updates the remaining drones wishing to join and supplements the drone swarm configuration based on the previously formed physical and logical topologies. The specific steps are as follows:

[0062] S331, If ​​the current drone swarm size |U k | Greater than or equal to the threshold u or the perception accuracy of the current drone swarm If the value is greater than or equal to the threshold P, the coordinated drone will not accept any ordinary drones to join the fleet during this crowdsourcing cycle, and will send a rejection message to all ordinary drones that want to join during this crowdsourcing cycle; otherwise, proceed to step S342.

[0063] S332. Based on the formation algorithm, coordinate the drones to pre-plan unified formation positions for the remaining ordinary drones that want to join. Here, we will not describe the specific formation algorithm in detail. Any formation algorithm can be nested into our algorithm. Since most formation algorithms aim to maintain the formation of the drone swarm, they mainly change the physical topology of the swarm. Our algorithm plans the formation topology based on the physical topology.

[0064] S333, Calculate the perception accuracy of the remaining ordinary drones to be added. The ordinary drones are sorted from largest to smallest according to their perception accuracy. The ordinary drone i with the largest perception accuracy is added to the drone swarm in this round to form a temporary drone swarm.

[0065] S334. Sort the ordinary drones in the temporary drone swarm from largest to smallest according to the size of computing resources, and take the ordinary drones one by one in order.

[0066] S335. Assuming the currently acquired ordinary drone is ordinary drone j, then ordinary drone i selects ordinary drone j for collaborative computation, let δ i，j =1, forming a new perception service data flow from ordinary drone i to ordinary drone j, adding this service flow to the original logical topology to form a temporary logical topology;

[0067] S336. Evaluate the end-to-end latency of the newly added service flow in the temporary logical topology and the intersecting service flow. If the end-to-end latency is less than or equal to the upper limit of task execution latency, proceed to step S337; otherwise, return to S334. The intersecting service flow refers to the data flow processed by the ordinary UAV j.

[0068] S337. If, after traversing the temporary drone swarm, there is always a perception data stream formed by a regular drone i or an end-to-end latency exceeding the latency limit, then the coordinating drone rejects the addition of regular drone i. Otherwise, it accepts the addition of regular drone i and replaces the original drone swarm with a temporary drone swarm, retaining δ. i，j =1, replace the drone swarm logical topology with a temporary logical topology, and return to S331.

[0069] S4. If all ordinary drones seeking to join the task are evaluated, the reward is given to the multi-agent reinforcement learning model for further iterative training, and the process returns to S1 to enter the next crowdsourcing cycle. Specifically, the drone swarm's rewards are based on the number of drones joining in this crowdsourcing cycle and changes in the swarm's task execution performance, i.e., the task benefits generated in this crowdsourcing cycle. Further iterative learning is then conducted. By executing numerous crowdsourcing cycles, the convergence of parameters in the deep reinforcement learning algorithm is ensured. The deep reinforcement learning algorithm is trained centrally, and the parameters gradually converge. Parameter convergence is a known existing technique and will not be elaborated upon in this invention.

[0070] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the scope of the claims of the invention.

Claims

1. A distributed dynamic UAV task scheduling method based on digital twins, characterized in that, The application scenario is as follows: coordinating crowdsourced drones to form a drone swarm to perform perception tasks. Each perception task needs to achieve a perception accuracy requirement within the end-to-end latency. The drone swarm includes a coordinating drone and several ordinary drones that have joined the perception task. During the task execution, the ordinary drones collect perception data and collaboratively process the perception data, and summarize the processed perception results to the coordinating drone. During the formation of the drone swarm, the coordinating drone is located directly above the perception task and is responsible for publishing the task request message. Idle ordinary drones receive the task request message and intelligently select to join the task. The scheduling method includes the following steps: S1. Coordinate the drones to determine the latest mission requirements, the current configuration of the drone swarm, the mission execution status and logical topology, and publish the mission requirement message to start the crowdsourcing process. The mission requirements include the perception accuracy requirement and the mission latency requirement. The current configuration of the drone swarm includes the size of the drone swarm to which the mission belongs and the number of perception and computing resources of the drones. The mission execution status refers to the mission execution time based on the current logical topology and the perception accuracy that can be achieved. S2. When an idle drone receives a task request message, it uses a multi-period distributed dynamic crowdsourcing algorithm based on a multi-agent reinforcement learning model to intelligently decide whether to join and which task to join. S3. Based on the existing drone swarm, coordinate the drones to construct a new virtual digital twin of the drone swarm, combining it with the newly added ordinary drones for the mission. In the constructed virtual digital twin, the logical topology allocation algorithm is used to assign the perception business data flow and business flow direction to each of the newly added ordinary drones for the mission, and to rehearse the logical topology of the drone swarm during the execution of the mission, and calculate the end-to-end latency of the business flow. Based on the improved perception accuracy of the newly added general drone's perception resources and the end-to-end latency of the new business flow logic topology, the drone evaluation is coordinated to determine whether to accept the drone's addition to the task. S4. If all the ordinary drones that want to join the task are evaluated, the multi-agent reinforcement learning model will be rewarded and further iterated and trained. Then, return to S1 and enter the next crowdsourcing cycle.

2. The distributed dynamic UAV task scheduling method based on digital twins according to claim 1, characterized in that, The state space representation of the multi-period distributed dynamic crowdsourcing algorithm based on the multi-agent reinforcement learning model in step S2 is as follows: The action space is represented as: a i ∈{0, ..., K} The reward is represented as follows: Where, μ i For the sensor parameters of a typical drone, C i For the CPU frequency of a typical drone, I i Indicates whether a regular drone is currently in an idle state, x i y i , z i P represents the coordinates of a typical drone. k Let be the current perception accuracy for task k. d k Let τ be the execution time of task k. k Let |U be the upper bound of the execution delay of task k. k | represents the current drone swarm size for task k, x k y k , z k Let k be the coordinate of task k, and a i =0 indicates that the drone will not participate in the mission. This indicates that the drone has joined mission k, P k (t) represents the perceptual accuracy of task k in this cycle, P k (t-1) represents the perception accuracy of task k in the previous cycle, |U k (t) represents the swarm size of the UAVs for task k in this cycle, |U k (t-1) represents the size of the drone swarm for task k in the previous cycle.

3. The distributed dynamic UAV task scheduling method based on digital twins according to claim 2, characterized in that, Step S3 includes establishing a data collection model, a link model, and a processing model respectively to construct a virtual digital twin of the drone swarm; The data collection model assumes that the sensor accuracy of each typical UAV is μ. i The higher the value, the more sensory data is collected; therefore, the amount of sensory data for a typical drone i is λ. i =βμ i λ s , where β is a constant, λ s The amount of data collected per unit precision; then the time it takes for the UAV to collect sensing data is expressed as: Among them, T s The time required to collect unit-sensing data; The link model is as follows: each drone swarm occupies a non-overlapping channel, but all drones in each swarm operate on the same channel; within the drone swarm, line-of-sight communication is the primary consideration; the average signal-to-noise ratio between drone i and drone j is expressed as: Where, p i，j For transmission power, h i，j For channel gain, g i，j For channel power gain, L i，j For free space propagation loss, L i，j =20log d i，j +20log f0+20log(4π / c), d i，j Where f0 is the distance between drones, c is the carrier frequency of the channel, and N is the speed of light. gw It is the power of Gaussian white noise; The formula for calculating the transmission rate from drone i to drone j is: r i，j =B log2(1+γ i，j ) The transmission time from drone i to drone j is expressed as The processing model uses the same model as the link model to simulate the service process of the UAV processing sensor data traffic; for a CPU frequency of C i The processing speed of drone i is... Where C represents the CPU computation cycles required per bit of sensor data; when drone i selects drone j for collaborative processing, if other drones in the cluster also select drone j for collaborative computation, assuming that drone j's computing power will be evenly distributed to each drone requesting its collaborative computation, the computation time is expressed as... δ i，j =1 indicates that drone i selects drone j for collaborative processing; otherwise, δ i，j =0; Ordinary drones process the sensed data locally or collaboratively with other drones, and the final sensed results are aggregated by the coordinating drone for unified collection. Considering the very small data volume, the process of transmitting the final sensed results to the coordinating drone is ignored. Assuming drone i selects drone j for collaborative processing of the collected sensed data, the end-to-end latency of the sensed data stream from drone i to drone j is expressed as: The task execution latency is the end-to-end latency of the maximum sensing data stream of the UAV swarm, expressed as:

4. The distributed dynamic UAV task scheduling method based on digital twins according to claim 3, characterized in that, The implementation process of the logical topology allocation algorithm is as follows: A1. If the current drone swarm size |U k | Greater than or equal to the threshold U or the perception accuracy of the current drone swarm 1- If the value is greater than or equal to the threshold P, the coordinating drone will not accept any ordinary drones to join the fleet during this crowdsourcing cycle, and will send a rejection message to all ordinary drones that want to join during this crowdsourcing cycle; otherwise, proceed to step A2. A2. Based on the formation algorithm, coordinate the drones to pre-plan a unified formation position for the remaining ordinary drones that want to join; A3. Calculate the perception accuracy of the remaining ordinary drones to be added. The ordinary drones are sorted from largest to smallest according to their perception accuracy. The ordinary drone i with the largest perception accuracy is added to the drone swarm in this round to form a temporary drone swarm. A4. Sort the ordinary drones in the temporary drone swarm from largest to smallest according to the amount of computing resources available, and take the ordinary drones one by one in order. A5. Assuming the currently acquired ordinary drone is ordinary drone j, then ordinary drone i selects ordinary drone j for collaborative computation, let δ i，j =1, forming a new perception service data flow from ordinary drone i to ordinary drone j, adding this service flow to the original logical topology to form a temporary logical topology; A6. Evaluate the end-to-end latency of the newly added business data flow in the temporary logical topology and the business data flow that intersects with it. If the end-to-end latency is less than or equal to the upper limit of task execution latency, proceed to step A7; otherwise, return to A4. A7. If, after traversing the temporary drone swarm, there is always a perception data stream formed by a regular drone i or an end-to-end latency exceeding the latency limit, then the coordinating drone rejects the addition of regular drone i. Otherwise, it accepts the addition of regular drones and replaces the original drone swarm with a temporary drone swarm, retaining δ. i，j =1, replace the drone swarm logical topology with a temporary logical topology, and return A1.

Images