A heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method and system

By adopting a heterogeneous UAV swarm asynchronous split zero-order collaborative learning scheduling method, the problems of heavy training burden, slow synchronous training node impact, and insufficient utilization of asynchronous update resources in UAV swarm collaborative learning are solved. This method achieves lightweight training, competitive scheduling, and effective utilization of compensation, thereby improving the system's efficiency and scalability.

CN123113751APending Publication Date: 2026-11-17NORTHEASTERN UNIV CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611349701.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In collaborative learning among drone swarms, the training burden on the drone end is heavy, synchronous training of heterogeneous drone swarms is easily affected by slow nodes, asynchronous updates are difficult to balance update timeliness and resource utilization, and existing asynchronous scheduling methods are difficult to fully utilize high-value heterogeneous data sources.

Method used

A heterogeneous UAV swarm asynchronous split zero-order collaborative learning scheduling method is adopted. The UAV set and model are initialized through the edge server, the comprehensive scheduling score is calculated, the target UAV is selected to perform split zero-order detection training, the training payload data is generated and uploaded, the edge server updates the model parameters, and the scheduling is optimized by effectively utilizing the deficit queue and the comprehensive scheduling score feedback mechanism.

Benefits of technology

This reduces the computational and communication burden on drones, alleviates slow node congestion during synchronous training, reduces the impact of outdated updates, improves the utilization of high-value data sources, enhances fairness for drones that have been out of service for a long time, and improves the actual control of scheduling strategies and system scalability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113751A_ABST
    Figure CN123113751A_ABST
Patent Text Reader

Abstract

This invention discloses an asynchronous splitting zero-order collaborative learning scheduling method and system for heterogeneous UAV swarms, belonging to the fields of artificial intelligence and edge computing technology. The edge server determines the currently schedulable UAV set from the UAV ensemble; calculates the comprehensive scheduling score of each candidate UAV in the candidate pool to determine the target UAV; the target UAV performs splitting zero-order detection training, generating and uploading training payload data; the edge server updates the edge server-side model parameters and the UAV-side edge model parameters; the edge server updates the target UAV's effective utilization deficit queue, and the updated effective utilization deficit queue is fed back into the subsequent comprehensive scheduling score calculation. This achieves a unified closed loop of lightweight training, competitive scheduling, obsolescence suppression, and effective utilization compensation under conditions of limited computing power on the UAV side, heterogeneous data distribution, unstable link states, and outdated asynchronous updates.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and edge computing technology, and in particular relates to a method and system for asynchronous splitting zero-order cooperative learning scheduling of heterogeneous unmanned aerial vehicle swarms. Background Technology

[0002] In scenarios such as disaster emergency response, drone inspection, low-altitude perception, environmental monitoring, and traffic patrol, multiple drones typically need to collect images, videos, or sensor data and collaborate with edge servers to complete intelligent model training.

[0003] Due to differences in flight areas, observation angles, mission objectives, communication links, and equipment performance among various drones, local data from each drone typically exhibits non-independent and identically distributed characteristics. Furthermore, there are differences in computing power, bandwidth, upload latency, and available time among different drones. Therefore, efficiently completing collaborative learning and scheduling under conditions of limited computing power at the drone end, unstable link status, heterogeneous data distribution, and asynchronous arrival is a crucial issue in drone edge intelligent systems.

[0004] Existing synchronous federated learning typically involves the following process: a central server initializes a global model; the server distributes global model parameters to multiple drones or edge devices; each drone performs complete local training based on its local data, including forward and backward propagation; each drone uploads its local model parameters, gradients, or model updates; the server aggregates the results after waiting for uploads from multiple drones; and the server forms a new global model and enters the next round of training. This approach can complete collaborative training without centrally uploading raw data, but it usually requires each drone to perform the complete training process locally, and the server to wait for multiple clients to complete their uploads before entering the aggregation phase.

[0005] To alleviate the slow node waiting problem in synchronous federated learning, existing asynchronous federated learning allows the server to immediately execute model updates after a drone uploads an update, without waiting for all drones to complete simultaneously. Its general process includes: the server maintains the current global model version; the drone trains locally based on the received model version; after training, the drone uploads its local update; the server asynchronously aggregates updates according to their arrival order; and the server continues to distribute the new model version to subsequent drones. This approach can alleviate the synchronous blocking problem, but it still requires drones to perform relatively heavy local training, and asynchronously arriving updates may have staleness issues.

[0006] Split learning divides a complete deep neural network into a terminal-side model and a server-side model. The terminal performs the first half of the model computation and uploads intermediate features to the server; the server performs the second half of the model computation and loss calculation. The general process includes: the terminal performs forward propagation of the first half of the model; the terminal uploads intermediate features; the server performs forward propagation and loss calculation of the second half of the model; the server performs backpropagation and returns intermediate gradients to the terminal; the terminal updates its terminal-side model based on the backpropagated gradients. This approach reduces the burden on the terminal to perform complete model training, but the terminal usually still needs to participate in gradient backpropagation, and its scheduling mechanism generally does not fully consider candidate competition, asynchronous obsolescence, and underutilization issues within a drone swarm.

[0007] Existing drone or wireless edge scheduling solutions consider data value, fairness, link status, or service latency. For example, the scheduler might select drones based on data importance, link quality, latency, or task priority. Such solutions typically involve: collecting status information for each drone; constructing a scheduling score for each drone; selecting the drone with the highest score to execute the task; and updating the status based on the task execution result. While this method introduces scheduling scores, in asynchronous systems, if drones are dispatched immediately upon becoming idle, the scheduler often faces only a single candidate, making it difficult for complex scores to truly change the scheduling outcome.

[0008] In summary, the existing technology still has the following technical shortcomings: (1) Heavy training burden on UAVs: Existing federated learning methods usually require UAVs to complete a relatively complete model training process locally, including forward propagation, backpropagation and parameter updates. For UAVs with limited computing power, battery capacity and payload capacity, performing complete local training for a long time will cause a large computational and energy consumption pressure.

[0009] (2) Synchronous training in heterogeneous UAV swarms is easily affected by slow nodes: There are significant differences in computing power, communication bandwidth, upload latency and availability among different UAVs. Synchronous aggregation usually requires waiting for multiple UAVs to complete local training and upload, which can easily slow down the overall training process due to slow nodes or unstable links.

[0010] (3) Simple asynchronous updates struggle to balance update timeliness and resource utilization: Although asynchronous federated learning can reduce waiting time, the global model version may have changed by the time the updates uploaded by the UAV reach the server. If outdated updates are used directly, it may affect the model's convergence stability; if outdated updates are simply discarded, it will waste the computing and communication resources already consumed by the UAV.

[0011] (4) Existing asynchronous scheduling methods are difficult to fully utilize high-value heterogeneous data sources: UAV local data usually exhibits a non-IID distribution, and different UAV data contribute differently to the global model training. If the scheduling strategy is selected based solely on idle state, waiting time, or polling rules, high-value UAV data may not be fully utilized, thereby affecting the global model training effect. Summary of the Invention

[0012] To address the shortcomings of existing technologies, this invention aims to provide a heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method and system, so as to achieve a unified closed loop of lightweight training, competitive scheduling, obsolescence suppression and effective utilization of compensation under the conditions of limited computing power at the UAV end, heterogeneous data distribution, unstable link state and asynchronous update obsolescence.

[0013] The technical solution of this invention is as follows: On the one hand, this invention provides a heterogeneous unmanned aerial vehicle (UAV) swarm asynchronous splitting zero-order cooperative learning scheduling method, comprising the following steps: S1: Edge server initializes drone collection Drone edge model, server-side model, current global model version number, candidate pool In-transit task collection And the scheduling status of each drone.

[0014] S2: The edge server, at the current moment, determines the scheduling status of each drone and the set of tasks in progress. Determine the currently schedulable drone set from the drone set. .

[0015] S3: Update the candidate pool based on the current set of schedulable drones, calculate the comprehensive scheduling score of each candidate drone in the candidate pool, and determine whether the multi-candidate competitive scheduling trigger condition is met; if it is met, determine the target drone based on the comprehensive scheduling score; if it is not met, keep the candidate pool open and continue to wait for new schedulable drones to join the candidate pool or wait for the multi-candidate competitive scheduling trigger condition to be met.

[0016] S4: The target UAV performs a split zero-order detection training task, generating and uploading training payload data.

[0017] S5: After receiving the training payload data, the edge server updates the model parameters on the edge server side. And drone edge model parameters Then the edge server removes the in-transit training task, namely the split zero-order probe training task, from the in-transit task set. Move out of the middle.

[0018] S6: The edge server updates the effective utilization deficit queue of the target drone, and the updated effective utilization deficit queue is fed back to the subsequent comprehensive scheduling score calculation.

[0019] S7: Repeat S2-S6 until the training time budget or training stop condition is reached, and then send the final updated edge server model parameters to the subsequently scheduled drones.

[0020] Furthermore, the drone ensemble Represented as: (1); in, Indicates the first A drone, Indicates the drone's serial number. This indicates the number of drones, each drone Holding local datasets .

[0021] The global model is a complete deep neural network model used to complete the target learning task. The edge server divides the global model into drone-end edge models deployed on the drone according to the preset model segmentation positions. and server-side models deployed on edge servers The UAV edge model is used to extract front-end features from the input data to obtain front-end features; the server-side model is used to receive the front-end features and complete subsequent feature processing, prediction output, and loss calculation. This represents the parameters of the drone's edge model. This represents the server-side model parameters.

[0022] The candidate pool This is used to temporarily cache drones that enter the schedulable state and meet the scheduling conditions within a short time window, and to use them as candidate drones; the state of the candidate pool includes an open state and a closed state, with the initial state being the closed state. The set of tasks in transit Used for recording Training tasks that have been dispatched but have not yet arrived at the edge server are represented as follows: (4); in, Indicates time, Indicates with drones The corresponding training tasks en route, This indicates the dispatch time for training missions en route. This indicates the estimated time for the training task to arrive at the edge server.

[0023] Each training mission on the way It should include at least the drone identifier, the time of dispatch of the training task en route, the estimated time of arrival of the training task at the edge server, and the version number of the task request corresponding to receiving the training task, represented as: (5); in, Indicates drone The task request version number received when a training task is in transit is determined by the current global model version number maintained by the edge server at the time the training task is dispatched. (6); Global model version number at initial moment It is 0.

[0024] The drone At any moment The scheduling state is represented as : (7); in, For drones The time of the most recent valid update completion. For drones Effective use of deficit queues; For drones The estimated time to restore the schedulable state.

[0025] Furthermore, the currently schedulable set of drones This indicates that the estimated time for restoring a schedulable state has been reached at the current moment, and the set of tasks in transit... The set of drones that do not correspond to any in-transit training missions is represented as: (8); in, Indicates drone The expected time for restoring a schedulable state has been reached; Represents the set of tasks in progress There is no drone in China For training tasks that have not yet arrived at the edge server, the edge server determines the drone's status when both of the above conditions are met. It is in a schedulable state.

[0026] Further, S3 includes the following steps: S3.1: Currently available set of schedulable drones When the candidate pool is not empty, if the candidate pool is currently closed and there are drones in the currently schedulable drone set that meet the scheduling conditions, the edge server will open the candidate pool and record the opening time of the candidate pool, and add the drones in the currently schedulable drone set that meet the scheduling conditions as candidate drones to the candidate pool; if the candidate pool is currently open, the edge server will continue to add the drones that have newly entered the schedulable state at the current time and meet the scheduling conditions to the candidate pool.

[0027] The candidate pool opening time is: (9); in, Open time for the candidate pool.

[0028] The state update process of the candidate pool is represented as follows: (10); in, For the updated candidate pool, This represents the currently schedulable set of drones. A subset of drones that meet the scheduling conditions.

[0029] S3.2: Calculate the static data value, AoU-like state, and link feasibility factor of each candidate UAV, and then determine the comprehensive scheduling score of each candidate UAV in the candidate pool. .

[0030] S3.3: The edge server determines whether the current candidate pool meets any preset multi-candidate competition scheduling trigger condition. If yes, it executes S3.4; otherwise, it keeps the candidate pool open and continues to wait for new schedulable drones to join or waits for the multi-candidate competition scheduling trigger condition to be met.

[0031] The multi-candidate competition scheduling triggering conditions include at least the candidate pool quantity threshold triggering condition, the waiting time timeout triggering condition, and the server idle triggering condition.

[0032] Selection pool quantity threshold trigger condition: (25); in, Candidate pool Number of candidate drones This is the threshold for the size of the candidate pool.

[0033] Waiting timeout trigger condition: (26); in, This is a flexible micro-wait window.

[0034] Server idle trigger condition: (27); in, Indicates time Edge server idle trigger variable; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This indicates that there are currently no training tasks en route that have not yet arrived at the edge server; This indicates that there is at least one candidate drone in the candidate pool; Let represent the minimum contention waiting window, and satisfy . .

[0035] S3.4: Select the candidate UAV with the highest comprehensive scheduling score as the target UAV. .

[0036] (29).

[0037] Furthermore, S3.2 specifically includes: S3.2.1: Predetermine the static data value of each UAV.

[0038] Specifically, setting up drones The local dataset is The total number of categories is drones In the The empirical distribution ratio on the class samples is Then drone The label distribution entropy is expressed as: (11); in, For drones The label distribution entropy; This is the sample category number; To prevent abnormally small constants in logarithmic calculations.

[0039] Edge servers further build lightweight probe models and using a pre-configured public reference training set. The lightweight probe model undergoes a small number of warm-up training rounds; the common reference training set Local datasets independent of each drone This is used to enable lightweight probe models to acquire basic recognition capabilities for target-oriented learning tasks.

[0040] After the warm-up training is completed, the edge server will reduce the probe model parameters. Distributed to all drones; drones Use its local dataset The lightweight probe model is evaluated, and its performance on the local dataset is calculated. The average loss is calculated and returned to the edge server, denoted as . : (12); in, For drones Local dataset The number of samples in For local datasets One of the training samples, For sample labels; Represents the loss function. g ( · ; θp ) represents a lightweight probe model. This represents the parameters of the lightweight probe model. This represents the average loss of the lightweight probe model.

[0041] Then, the drone The average loss and label distribution entropy are fused to obtain the original data value score. : (13); in, Indicates drone The original data value score, This represents the label entropy weight coefficient. This represents the entropy adjustment index. The edge server calculates the coefficient of variation of the value scores of all drone raw data. CV : (14); in, The coefficient of variation represents the value score of all raw drone data; This represents the standard deviation of the value scores of all raw drone data. This represents the average of the raw data value scores for all drones.

[0042] like CV Less than the preset threshold If the data value of each drone is not significantly different, then the static data value of each drone will be set to a uniform distribution. (15); in, The value of static data from drones.

[0043] likeCV Greater than or equal to the preset threshold Then, the value scores of the original data are normalized: (16); in, This indicates the normalized drone. Data value score, Number the drone. The original data value score, This represents the minimum value score among all the raw data scores of drones. This represents the maximum value score among all the raw data scores of all drones.

[0044] Introducing basic protection weights The smoothed, normalized data value scores are obtained as follows: (17); in, Indicates the smoothed drone Normalized data value score, To represent the basic protection weight; This represents the sum of the normalized data value scores of all drones.

[0045] The smoothed, normalized data value scores are then soft-mixed with a uniform distribution to obtain the final static data value: (18); in, ξ This is the mixing coefficient.

[0046] S3.2.2: Calculate the current time of each candidate UAV. AoU-like state.

[0047] Candidate drones The most recent valid update completion time was Then at the current moment The AoU-like state is as follows: (19); in, It is in an AoU-like state. This is a growth index.

[0048] S3.2.3: Combine the static data value and AoU-like state of each candidate UAV to obtain the basic score of each candidate UAV.

[0049] (20); in, The base score.

[0050] S3.2.4: Calculate the comprehensive scheduling score of each candidate UAV based on its basic score.

[0051] Edge servers for each drone Maintain an efficient utilization of the deficit queue Read the current effective utilization deficit queue And this is fed back into the overall scheduling score in a multiplicative manner: (twenty one); in, These are the weighting coefficients.

[0052] Furthermore, S4 specifically includes the following steps: S4.1: The edge server dispatches a split zero-order detection training task to the target drone, clears the candidate pool and sets the candidate pool status to the closed state, and enters the next round of candidate drone collection process.

[0053] Specifically: Order Edge server to target drone Distribute the current edge model parameters of the drone. Task request version number And training control parameters, and based on the target UAV The computing power, uplink bandwidth, and training payload data size are used to calculate the estimated time for the training task en route, i.e., the split zero-order probe training task, to reach the edge server. ; Target drone At the present moment Assigned to perform a zero-order split detection training mission, the forward computation of a single UAV edge model is Target drone The edge computing capability is The estimated edge computing latency required for the UAV to perform the zero-order split detection training task for: (30); Target drone The size of the uploaded training payload data is At present The uplink bandwidth is The estimated uplink communication latency required to upload training payload data is then... for: (31); Therefore, the estimated time to reach the edge server for: (32); Then, the edge server identifies the target drone. Expected arrival time at the edge server and task request version number Record to in-transit task collection .

[0054] S4.2: The target UAV selects local mini-batch training samples from its local dataset and calculates the original intermediate features corresponding to the local mini-batch training samples.

[0055] Target drone In this training, it was derived from its local dataset. The selected local mini-batch training samples are ,in, This represents the input data in this local mini-batch training sample. Indicates the corresponding sample label; target drone Using the current drone edge model parameters For input data Perform forward computation to obtain the original intermediate features. : (33); S4.3: Target UAV A random seed is generated according to the random seed generation rules, and then a perturbation direction is generated. The parameters of the perturbed UAV edge model are then constructed based on the perturbation direction.

[0056] Target drone Based on random seed Generate perturbation directions for each parameter block of the UAV edge model, for the th A parameter block, the perturbation direction is denoted as . The parameters of the disturbed UAV edge model are constructed according to the following formula: (34); in, This indicates the parameter block number of the drone edge model. Indicates the first The original parameters of each parameter block, Indicates the first The parameters after perturbation of each parameter block. Indicates the first The perturbation direction corresponding to each parameter block Indicates the first The perturbation scale corresponding to each parameter block.

[0057] The parameters after perturbation of each parameter block Parameters of the UAV edge model after perturbation .

[0058] S4.4: Target UAV Using the perturbed drone edge model parameters For input data Forward calculations are performed to obtain intermediate perturbation features.

[0059] (35); in, Indicates the target drone The intermediate features of the disturbance are calculated based on the edge model of the disturbed UAV. This represents the parameters of the drone edge model after the disturbance.

[0060] S4.5: The target UAV is based on the original intermediate features and intermediate features of perturbation Calculate the intermediate feature differences and generate training payload data.

[0061] The intermediate feature difference is divided into: (36); in, This represents the intermediate feature difference.

[0062] Target drone Generate training payload data: (37); in, Indicates the target drone Training payload data uploaded to the edge server, This indicates the number of parameter blocks in the drone edge model. This represents the set of perturbation scales corresponding to each parameter block.

[0063] Furthermore, S5 specifically includes the following steps: S5.1: Edge server receives target drone Uploaded training payload data Then, based on the original intermediate features and intermediate feature difference Restore the intermediate features of the perturbation.

[0064] S5.2: The edge server calculates the original loss and the perturbation loss under the same edge server model snapshot, and then calculates the loss difference.

[0065] Assume the edge server receives the training payload data. The snapshot of the edge server-side model parameters used at that time is The original loss is: (39); The disturbance loss is: (40); Loss difference is divided into: (41); in, Indicates based on original intermediate features The calculated original loss, Indicates based on perturbation intermediate features The calculated disturbance loss, This represents the difference in loss between the disturbance loss and the original loss.

[0066] S5.3: The edge server reconstructs the perturbation direction based on the random seed and constructs a zero-order gradient estimate for the parameter block of the UAV edge model, thereby obtaining the overall zero-order gradient estimate of the UAV edge model parameters.

[0067] Edge servers based on training payload data random seed Reconstruct the target drone The direction of the perturbation, for the first edge model of the UAV. The zeroth-order gradient estimate of a block of parameters is: (42); in, Indicates the original loss Compared to the edge model of the drone, the first Zero-order gradient estimation of a block of parameters.

[0068] By combining the zero-order gradient estimates of each parameter block, we obtain the overall zero-order gradient estimate of the parameters of the UAV edge model: (43); in, This represents the global zero-order gradient estimate of the parameters of the UAV edge model.

[0069] S5.4: Edge servers based on original losses Calculate its parameters relative to the edge server-side model. The gradient is used for backpropagation to update the model parameters on the edge server.

[0070] (44); in, This represents the updated edge server-side model parameters. This indicates the optimizer update operation of the edge server-side model. Indicates the original loss Relative to edge server-side model parameters The true gradient.

[0071] S5.5: The UAV edge model performs a zero-order update based on the global zero-order gradient estimation of the UAV edge model parameters, and obtains the updated UAV edge model parameters.

[0072] Before updating the parameters of the UAV edge model corresponding to the current training task, the edge server reads the asynchronous valid update coefficients. The drone edge model parameters are updated based on the overall zero-order gradient estimation of the drone edge model parameters and the asynchronous effective update coefficients: (45); in, This represents the updated edge model parameters for the drone. Indicates the zero-order update step size. Indicates the target drone At any moment The corresponding asynchronous effective update coefficient.

[0073] The method for calculating the asynchronous effective update coefficient is as follows: The edge server reads the task request version number corresponding to the task. and the current global version number Calculate the version obsolescence of this update. : (46); Edge servers calculate the asynchronous effective update coefficient based on version staleness: (47); in, This indicates the obsolescence penalty coefficient; This indicates the lower limit of the effective update coefficient; This indicates the maximum allowed version obsolescence.

[0074] S5.6: The edge server removes the in-transit training task from the in-transit task set. Move out of the middle.

[0075] Furthermore, the edge server updates the effective utilization deficit queue of the target UAV, and the updated effective utilization deficit queue is fed back into the subsequent comprehensive scheduling score calculation, specifically as follows: The edge server updates the drone's availability status if... Then, a valid model update is performed, setting the global version number. and update the drones Last successful update time ;like In this case, the training task will not participate in model parameter updates, will not increment the global version number, and will not update. However, the effective utilization deficit queue of the corresponding drone is updated according to the effective update difference.

[0076] Define valid update threshold And calculate the effective update difference: (48); When drones After the training task arrives and completes asynchronous updates, update its effective utilization deficit queue: (49); in, To optimize the use of the deficit queue after the update; The leakage coefficient is the factor indicating a deficit. To effectively utilize the upper limit of the deficit queue; This represents the truncation function.

[0077] On the other hand, the present invention also provides a heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling system for implementing a heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling method, comprising: The UAV-side module is used for the target UAV to perform split zero-order detection training, generate and upload training payload data, and receive the final updated edge server-side model parameters.

[0078] The edge server module is used to initialize the drone ensemble and the drone edge model. Server-side model The current global model version number, candidate pool, in-transit task set, and scheduling status of each drone; at the current moment, based on the scheduling status of each drone and the in-transit task set, determine the currently schedulable drone set from the drone set. Based on the current set of schedulable drones Update the candidate pool, calculate the comprehensive scheduling score of each candidate drone in the pool, and determine whether the multi-candidate competitive scheduling trigger condition is met. If it is met, determine the target drone based on the comprehensive scheduling score. If not, keep the candidate pool open and continue waiting for new schedulable drones to join or for the multi-candidate competitive scheduling trigger condition to be met. After receiving the training payload data, update the model parameters on the edge server. And drone edge model parameters Then, the training task in transit is removed from the set of tasks in transit. Remove from the queue; update the effective utilization deficit queue for the target UAV, and feed the updated effective utilization deficit queue back into the subsequent comprehensive scheduling score calculation. Send the final updated edge server-side model parameters to the UAV-side module.

[0079] Thirdly, this application proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to execute the heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling method.

[0080] Fourthly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the heterogeneous unmanned aerial vehicle swarm asynchronous splitting zero-order cooperative learning scheduling method.

[0081] Fifthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned heterogeneous unmanned aerial vehicle (UAV) swarm asynchronous splitting zero-order cooperative learning scheduling method.

[0082] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Reduce the computational burden on the drone: This invention enables the drone to perform only two forward calculations without having to perform a complete backward propagation, thereby reducing the computational load, energy consumption, and training burden on the drone.

[0083] 2. Reduce communication burden on UAVs: UAVs upload intermediate feature differences, random seeds, and perturbation scales, rather than the complete model gradient or the complete perturbation direction, which helps reduce the communication load.

[0084] 3. Alleviating slow node blocking in synchronous training: This invention adopts an asynchronous update mechanism, which does not require the server to wait for all drones to complete training and upload, thus reducing the impact of slow nodes on the overall training progress.

[0085] 4. Reduce the negative impact of stale updates: This invention suppresses stale updates by using stale delta and effective update coefficients, so that stale updates do not affect edge model updates with the same intensity.

[0086] 5. Improve the actual control of the scheduling strategy: This invention constructs a multi-candidate competitive environment through a competition-first candidate pool, so that the scheduling score can truly change the winner selection and avoid the scheduling strategy from degenerating into first-come, first-served.

[0087] 6. Improve the utilization of high-value data sources: This invention enables the scheduler to prioritize UAV data sources that have a greater potential contribution to the global model by using static data value items, thereby improving the shortcomings of simple averaging or first-come-first-served strategies under strong Non-IID conditions.

[0088] 7. Improve fairness for long-term unserved drones: This invention uses AoU-like status to enable drones that have not been successfully updated for a long time to gradually gain a higher chance of competing in subsequent scheduling.

[0089] 8. Compensation for scheduled but underperforming drones: This invention identifies and compensates drones that underperform due to high obsolescence by using eta gap and deficit queue, rather than simply providing fairness compensation based on the number of scheduling attempts.

[0090] 9. Balancing server utilization and scheduling contention: This invention achieves a balance between candidate contention and server idling through three types of triggering conditions: Watermark, Timeout, and Server-starvation.

[0091] 10. Improve system implementability and scalability: This invention can be implemented as a method, system, electronic device and computer-readable storage medium, and can be extended to different numbers of UAVs, different model structures, different link conditions and different data value assessment methods. Attached Figure Description

[0092] Figure 1 This is a flowchart of the asynchronous splitting zero-order cooperative learning scheduling method for heterogeneous UAV swarms in an embodiment of the present invention.

[0093] Figure 2 This is a schematic diagram of the multi-candidate competition scheduling mechanism in an embodiment of the present invention.

[0094] Figure 3 This is a flowchart of the integrated scheduling score and asynchronous effective utilization compensation in an embodiment of the present invention.

[0095] Figure 4 This is a timing diagram of the lightweight Split-ZO training protocol in an embodiment of the present invention.

[0096] Figure 5 This is a flowchart of the zero-order reconstruction and update process on the edge server side in an embodiment of the present invention.

[0097] Figure 6 This is a structural diagram of the heterogeneous unmanned aerial vehicle (UAV) swarm asynchronous splitting zero-order cooperative learning scheduling system in an embodiment of the present invention. Detailed Implementation

[0098] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0099] Example 1: A heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method, such as Figure 1 As shown, it includes the following steps: S1: Edge server initializes the drone collection and drone-side edge model. Server-side model The current global model version number, candidate pool, in-transit task set, and scheduling status of each UAV.

[0100] The drone collection Represented as: (1); in, Indicates the first A drone, Indicates the drone's serial number. This indicates the number of drones, each drone Holding local datasets Local datasets from different drones typically exhibit non-independent and identically distributed characteristics.

[0101] The global model is a complete deep neural network model used to complete the target learning task. The edge server divides the global model into drone-end edge models deployed on the drone according to the preset model segmentation positions. and server-side models deployed on edge servers The UAV edge model is used to extract front-end features from the input data to obtain front-end features; the server-side model is used to receive the front-end features and complete subsequent feature processing, prediction output, and loss calculation. This represents the parameters of the drone's edge model. This represents the server-side model parameters.

[0102] For input data The end-to-end mapping relationship of the global model is represented as follows: (2); in, For global models, Indicates input data, Let represent the global model parameter set, and satisfy: (3); The candidate pool This is used to temporarily cache drones that enter the schedulable state and meet the scheduling conditions within a short time window, and to treat them as candidate drones; the state of the candidate pool includes an open state and a closed state, with the initial state being the closed state.

[0103] The set of tasks in transit Used for recording Training tasks that have been dispatched but have not yet arrived at the edge server are represented as follows: (4); in, Indicates time, Indicates with drones The corresponding training tasks en route, This indicates the dispatch time for training missions en route. This indicates the estimated time for the training task to arrive at the edge server.

[0104] Each training mission on the way It should include at least the drone identifier, the time of dispatch of the training task en route, the estimated time of arrival of the training task at the edge server, and the version number of the task request corresponding to receiving the training task, represented as: (5); in, Indicates drone The task request version number received when a training task is in transit is determined by the current global model version number maintained by the edge server at the time the training task is dispatched. (6); Global model version number at initial moment It is 0.

[0105] The drone At any moment The scheduling state is represented as : (7); in, For drones The time of the most recent valid update completion. For drones The effective utilization deficit queue is used to record the effective utilization deficit of UAVs that have been scheduled but not fully utilized for updates. It is updated according to the effective update difference corresponding to each training mission of the UAV. For drones The estimated time to restore the schedulable state, initially, if the drone If the training task can be received immediately, then let After the training mission is dispatched en route, the edge server determines the data based on the drone's... The computational latency and uplink communication latency are used to estimate the time required for the system to recover to a schedulable state, and the system is updated accordingly. .

[0106] S2: At the current moment, the edge server determines the set of drones that can be scheduled from the drone set based on the scheduling status of each drone and the set of tasks in progress. .

[0107] The currently schedulable set of drones This indicates that the estimated time for restoring a schedulable state has been reached at the current moment, and the set of tasks in transit... The set of drones that do not correspond to any in-transit training missions is represented as: (8); in, Indicates drone The expected time for restoring a schedulable state has been reached; Represents the set of tasks in progress There is no drone in China The corresponding training task that has not yet arrived at the edge server. When both of the above conditions are met, the edge server determines that the drone is in transit. It is in a schedulable state.

[0108] S3: Update the candidate pool based on the current set of schedulable drones, calculate the comprehensive scheduling score of each candidate drone in the candidate pool, and determine whether the multi-candidate competitive scheduling trigger condition is met; if it is met, determine the target drone based on the comprehensive scheduling score; if it is not met, keep the candidate pool open and continue to wait for new schedulable drones to join the candidate pool or wait for the multi-candidate competitive scheduling trigger condition to be met.

[0109] like Figure 2 As shown, it includes the following steps: S3.1: Currently available set of schedulable drones When the candidate pool is not empty, if the candidate pool is currently closed and there are drones in the currently schedulable drone set that meet the scheduling conditions, the edge server will open the candidate pool and record the opening time. The drones in the currently schedulable drone set that meet the scheduling conditions will be added to the candidate pool as candidate drones for subsequent multi-candidate competitive scheduling. If the candidate pool is currently open, the edge server will continue to add drones that have newly entered the schedulable state and meet the scheduling conditions to the candidate pool.

[0110] The candidate pool opening time is: (9); in, Open time for the candidate pool.

[0111] The state update process of the candidate pool is represented as follows: (10); in, For the updated candidate pool, This represents the currently schedulable set of drones. A subset of drones that meet the scheduling conditions.

[0112] In this embodiment, the candidate pool adopts a multi-candidate competition mechanism. That is, the edge server does not immediately dispatch training tasks after detecting a single candidate drone that is in a schedulable state and meets the scheduling conditions. Instead, it caches multiple candidate drones that meet the scheduling conditions within a short time window, and performs candidate comparison and task dispatch uniformly after the multi-candidate competition scheduling trigger condition is met.

[0113] S3.2: Calculate the static data value, AoU-like state, and link feasibility factor of each candidate UAV, and then determine the comprehensive scheduling score of each candidate UAV in the candidate pool. .

[0114] This step addresses the scheduling scoring problem under conditions of Non-IID data distribution, asynchronous obsolete updates, and insufficient effective utilization. Instead of simply calculating a static score before dispatch, this step connects the "pre-schedule scoring" and "post-update compensation" into a feedback loop: Before scheduling, the edge server calculates a comprehensive scheduling score based on the data value of each UAV, its long-term unsuccessful update status, and the degree of historical underutilization; after the task is completed and arrives at the edge server, the edge server calculates the effective update coefficient based on the obsolescence of the update and feeds back the underutilization to subsequent scheduling.

[0115] like Figure 3 As shown, it includes the following steps: S3.2.1: Predetermine the static data value of each UAV to represent the potential contribution of local data in each UAV's local dataset to the global model training.

[0116] The static data value can be obtained through pre-evaluation, data distribution measurement, marginal contribution estimation, or other data value evaluation methods. In a preferred embodiment, the edge server can pre-calculate the static data value of each UAV based on the label distribution characteristics of each UAV's local data and the evaluation results of the lightweight probe model before the formal asynchronous scheduling training begins.

[0117] Specifically, setting up drones The local dataset is The total number of categories is drones In the The empirical distribution ratio on the class samples is Then drone The label distribution entropy can be expressed as: (11); in, For drones The label distribution entropy is used to characterize the drone. The extent of category coverage or distribution diversity of local data; This is the sample category number; To prevent abnormally small constants in logarithmic calculations.

[0118] Edge servers further build lightweight probe models and using a pre-configured public reference training set. The lightweight probe model undergoes a small number of warm-up training rounds; the common reference training set Local datasets independent of each drone This is used to enable lightweight probe models to acquire basic recognition capabilities for target-oriented learning tasks.

[0119] After the warm-up training is completed, the edge server will reduce the probe model parameters. Distributed to all drones; drones Use its local dataset The lightweight probe model is evaluated, and its performance on the local dataset is calculated. The average loss is calculated and returned to the edge server, denoted as . : (12); in, For drones Local dataset The number of samples in For local datasets One of the training samples, For sample labels; Represents the loss function. g ( · ; θp ) represents a lightweight probe model. This represents the parameters of the lightweight probe model. The average loss of the lightweight probe model is used to characterize the current lightweight probe model's performance on UAVs. The difficulty of fitting local data or its value in supplementing training.

[0120] The label distribution entropy and the lightweight probe model evaluation results are two complementary indicators used to evaluate the value of UAV local data. They are independent of each other and represent the value of UAV local data from two dimensions: data distribution diversity and model fitting difficulty, respectively.

[0121] Then, the drone The average loss and label distribution entropy are fused to obtain the original data value score. : (13); in, Indicates drone The original data value score, This represents the label entropy weight coefficient. The entropy adjustment index is represented by formula (13). It means that when the local data of a certain UAV has a high loss under the current lightweight probe model and its label distribution has a high diversity, the local data of the UAV has a higher potential utilization value for subsequent global model training.

[0122] To avoid excessive skewness of static values ​​due to minor evaluation perturbations in near-IID scenarios, the edge server calculates the coefficient of variation of the value scores of all drone raw data. CV : (14); in, The coefficient of variation represents the raw data value scores of all drones, used to measure the significance of differences in data value between different drones; This represents the standard deviation of the value scores of all raw drone data. This represents the average of the raw data value scores for all drones. The larger the value, the more significant the difference in data value between different drones. The smaller the value, the more similar the data value of each drone.

[0123] like CV Less than the preset threshold If the data value of each drone is not significantly different, then the static data value of each drone will be set to a uniform distribution. (15); in, The value of static data from drones.

[0124] like CV Greater than or equal to the preset threshold If the data is in a relatively obvious Non-IID data distribution state, then the original data value scores are normalized: (16); in, This indicates the normalized drone. Data value score, Number the drone. The original data value score, This represents the minimum value score among all the raw data scores of drones. The value of the raw data of all UAVs is represented by the maximum value. Formula (16) is used to map the raw data value scores of different UAVs to a uniform scale, which is convenient for subsequent calculation of static data value weights.

[0125] To prevent low-resolution drones from losing scheduling opportunities for an extended period, a basic protection weight is introduced. The smoothed, normalized data value scores are obtained as follows: (17); in, Indicates the smoothed drone Normalized data value score, To represent the basic protection weight, used to prevent low-scoring drones from losing scheduling opportunities for an extended period; This represents the sum of the normalized data value scores of all drones. This formula introduces a basic protection weight. While emphasizing high-value drones, a certain basic scheduling probability is reserved for low-value drones.

[0126] The smoothed, normalized data value scores are then soft-mixed with a uniform distribution to obtain the final static data value: (18); in, ξ This is the mixing coefficient.

[0127] Through the above methods, the value of static data is... Simultaneously considering the diversity of local data distribution, the difficulty of fitting the current lightweight probe model to the local data of this UAV, and scheduling starvation constraints, it is possible to improve the scheduling priority of high-value data sources under strong Non-IID conditions, while avoiding some UAVs being unable to participate in training for a long time due to value estimation bias.

[0128] S3.2.2: Calculate the current time of each candidate UAV. AoU-like state.

[0129] Candidate drones The most recent valid update completion time was Then at the current moment The AoU-like state is as follows: (19); in, It is in an AoU-like state. This is a growth index, used to characterize the increase in scheduling priority caused by drones failing to complete updates for an extended period.

[0130] S3.2.3: Combine the static data value and AoU-like state of each candidate UAV to obtain the basic score of each candidate UAV; (20); in, The base score.

[0131] S3.2.4: Calculate the comprehensive scheduling score of each candidate UAV based on its basic score.

[0132] Edge servers for each drone Maintain an efficient utilization of the deficit queue Read the current effective utilization deficit queue And this is fed back into the overall scheduling score in a multiplicative manner: (twenty one); Expanding, we get: (twenty two); in, This is a weighting coefficient used to control the intensity of the impact of effectively utilizing the deficit queue on the overall scheduling score. When When it is larger, it indicates that the drone If there are instances of insufficient utilization in historical updates, these instances will receive some compensation when they subsequently enter the candidate pool, thereby increasing their chances of being selected.

[0133] In one alternative implementation, the edge server also constructs a link feasibility factor based on the expected end-side computational latency and expected uplink communication latency of the UAV: (twenty three); in, For candidate drones The estimated edge computation latency required to perform the training task after being selected. For candidate drones The estimated uplink communication latency required to upload relevant data for this training task after completing the edge-side calculation. The overall scheduling score can then be further expressed as: (twenty four); S3.3: The edge server determines whether the current candidate pool meets any preset multi-candidate competition scheduling trigger condition. If yes, it executes S3.4; otherwise, it keeps the candidate pool open and continues to wait for new schedulable drones to join or waits for the multi-candidate competition scheduling trigger condition to be met.

[0134] The multi-candidate competition scheduling triggering conditions include at least the candidate pool number threshold triggering condition, the waiting time timeout triggering condition, and the server idle triggering condition. Specifically, the candidate pool number threshold triggering condition indicates that the number of candidate drones in the candidate pool reaches a preset candidate pool size threshold; the waiting time timeout triggering condition indicates that the waiting time after the candidate pool is opened reaches a preset upper limit; and the server idle triggering condition indicates that the edge server currently has no upcoming training tasks en route, and scheduling is triggered in advance to avoid the edge server becoming idle.

[0135] First, the watermark condition triggers the selection pool quantity threshold: (25); in, Candidate pool Number of candidate drones The threshold for the size of the candidate pool is defined as follows: when the number of candidate drones in the candidate pool reaches the threshold, it is considered that sufficient competition has been formed, and dispatch is triggered.

[0136] Second, the timeout trigger condition: (26); in, This is a flexible micro-waiting window used to limit the maximum waiting time of the candidate pool. When the waiting time reaches the upper limit of this flexible micro-waiting window, dispatch is triggered even if the size of the candidate pool has not reached the threshold, in order to avoid excessively long waiting times.

[0137] Third, server-starvation trigger conditions: To avoid prolonged idle time for edge servers and to prevent the candidate pool from degenerating into a single-candidate dispatch immediately after opening, this invention sets a minimum contention waiting window. Triggered when the candidate pool is open, the current in-transit task set is empty, and the candidate pool opening time has reached the minimum contention waiting window: (27); in, Indicates time Edge server idle trigger variable; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This indicates that there are currently no training tasks en route that have not yet arrived at the edge server; This indicates that there is at least one candidate drone in the candidate pool; Let represent the minimum contention waiting window, and satisfy . ;when When the edge server has no training tasks in transit, there are candidate drones in the candidate pool, and the candidate pool has been open for the minimum contention waiting window, candidate comparison and task dispatch are triggered. At this time, server idle dispatch will not be triggered.

[0138] Based on the three types of triggering conditions, define the dispatch trigger variable: (28); in, To dispatch the trigger variable, when When = 1, proceed to winner selection.

[0139] Through this mechanism, the present invention constructs a realistic multi-candidate comparison scenario with a short waiting window, enabling the comprehensive scheduling score to move beyond theoretical calculations and truly influence the dispatching results. Simultaneously, three triggering conditions—Watermark, Timeout, and Server-starvation—work together to ensure that the system does not significantly increase server idle time while creating competition.

[0140] S3.4: Select the candidate UAV with the highest comprehensive scheduling score as the target UAV. ; (29); S4: The target UAV performs a split zero-order detection training task, generating and uploading training payload data.

[0141] like Figure 4 As shown, the specific steps include: S4.1: The edge server dispatches a Split-ZO probing task to the target drone, clears the candidate pool and sets the candidate pool status to the off state, and enters the next round of candidate drone collection process.

[0142] Specifically: For ease of expression, let Edge server to target drone Distribute the current edge model parameters of the drone. Task request version number And the necessary training control parameters, and based on the target UAV The computing power, uplink bandwidth, and training payload data size are used to calculate the estimated time for the training task (split zero-order probe training task) to reach the edge server. The training control parameters include, but are not limited to, mini-batch sample size, random seed generation rule, perturbation scale set, maximum allowable staleness, uploaded data format, and task identifier. Specifically, the mini-batch sample size represents the number of local training samples selected by the target UAV in this training task; the random seed generation rule represents the preset rule used to generate the random seed for this zero-order perturbation, enabling the target UAV and the edge server to generate and reconstruct consistent perturbation directions based on the same random seed; the perturbation scale set represents the set of perturbation amplitudes corresponding to each parameter block of the UAV-side edge model; the maximum allowable staleness represents the maximum version difference that the training payload data is allowed to participate in model updates; the uploaded data format represents the composition, arrangement order, and data precision of each field in the training payload data; and the task identifier represents a unique identifier used to distinguish different training tasks and associate them with the target UAV, task request version number, and in-transit training task records.

[0143] Target drone At the present moment Assigned to perform a zero-order split detection training mission, the forward computation of a single UAV edge model is Target drone The edge computing capability is Since the splitting zero-order detection training task includes one forward computation of the original UAV edge model and one forward computation of the perturbation edge model, the estimated edge-side computation delay required for the UAV to perform the splitting zero-order detection training task is as follows: for: (30); Target drone The size of the uploaded training payload data is At present The uplink bandwidth is The estimated uplink communication latency required to upload training payload data is then... for: (31); Therefore, the estimated time for the in-transit training task (split zero-order probe training task) to reach the edge server is... for: (32); Then, the edge server identifies the target drone. Expected arrival time at the edge server and task request version number Record to in-transit task collection This is used to determine whether the edge server is in a state of having no training tasks in transit.

[0144] S4.2: The target UAV selects local mini-batch training samples from its local dataset and calculates the original intermediate features corresponding to the local mini-batch training samples.

[0145] Target drone In this training, it was derived from its local dataset. The selected local mini-batch training samples are ,in, This represents the input data in this local mini-batch training sample. This indicates the corresponding sample label. Target drone. Using the current drone edge model parameters For input data Perform forward computation to obtain the original intermediate features. : (33); S4.3: Target UAV A random seed is generated according to the random seed generation rules, and then a perturbation direction is generated. The parameters of the perturbed UAV edge model are then constructed based on the perturbation direction.

[0146] Target drone Based on random seed Generate perturbation directions for each parameter block of the UAV edge model, for the th A parameter block, the perturbation direction is denoted as . The parameters of the disturbed UAV edge model are constructed according to the following formula: (34); in, This indicates the parameter block number of the drone edge model. Indicates the first The original parameters of each parameter block, Indicates the first The parameters after perturbation of each parameter block. Indicates the first The perturbation direction corresponding to each parameter block Indicates the first The perturbation scale corresponding to each parameter block.

[0147] The parameters after perturbation of each parameter block Parameters of the UAV edge model after perturbation .

[0148] S4.4: Target UAV Using the perturbed drone edge model parameters For input data Forward calculations are performed to obtain intermediate perturbation features.

[0149] (35); in, Indicates the target drone The intermediate features of the disturbance are calculated based on the edge model of the disturbed UAV. This represents the parameters of the drone edge model after the disturbance.

[0150] S4.5: The target UAV is based on the original intermediate features and intermediate features of perturbation Calculate the intermediate feature differences and generate training payload data.

[0151] The intermediate feature difference is divided into: (36); in, This represents the intermediate feature difference.

[0152] Target drone Generate training payload data: (37); in, Indicates the target drone Training payload data uploaded to the edge server, This indicates the number of parameter blocks in the drone edge model. This represents the set of perturbation scales corresponding to each parameter block.

[0153] Due to the intermediate characteristics of the perturbation It can be determined by the edge server based on the original intermediate features. and intermediate feature difference The result is: (38); Therefore, the target drone No need for training payload data Repeatedly uploading complete perturbation intermediate features reduces uplink communication load.

[0154] S5: After receiving the training payload data, the edge server updates the model parameters on the edge server side. And drone edge model parameters Then the edge server removes the in-transit training task, namely the split zero-order probe training task, from the in-transit task set. Move out of the middle.

[0155] like Figure 5 As shown, it includes the following steps: S5.1: Edge server receives target drone Uploaded training payload data Then, based on the original intermediate features and intermediate feature difference Restore the intermediate features of the perturbation.

[0156] S5.2: The edge server calculates the original loss and the perturbation loss under the same edge server model snapshot, and then calculates the loss difference.

[0157] Assume the edge server receives the training payload data. The snapshot of the edge server-side model parameters used at that time is The original loss is: (39); The disturbance loss is: (40); Loss difference is divided into: (41); in, Indicates based on original intermediate features The calculated original loss, Indicates based on perturbation intermediate features The calculated disturbance loss, This represents the difference in loss between the disturbance loss and the original loss.

[0158] S5.3: The edge server reconstructs the perturbation direction based on the random seed and constructs a zero-order gradient estimate for the parameter block of the UAV edge model, thereby obtaining the overall zero-order gradient estimate of the UAV edge model parameters.

[0159] Edge servers based on training payload data random seed Reconstruct the target drone The direction of the perturbation, for the first edge model of the UAV. The zeroth-order gradient estimate of a block of parameters is: (42); in, Indicates the original loss Compared to the edge model of the drone, the first Zero-order gradient estimation of a block of parameters.

[0160] By combining the zero-order gradient estimates of each parameter block, we obtain the overall zero-order gradient estimate of the parameters of the UAV edge model: (43); in, This represents the global zero-order gradient estimate of the parameters of the UAV edge model.

[0161] S5.4: Edge servers based on original losses Calculate its parameters relative to the edge server-side model. The gradient is used for backpropagation to update the model parameters on the edge server. (44); in, This represents the updated edge server-side model parameters. This indicates the optimizer update operation of the edge server-side model. Indicates the original loss Relative to edge server-side model parameters The true gradient.

[0162] S5.5: The UAV edge model performs a zero-order update based on the global zero-order gradient estimation of the UAV edge model parameters, and obtains the updated UAV edge model parameters.

[0163] Before updating the parameters of the UAV edge model corresponding to the current training task, the edge server reads the asynchronous valid update coefficients. The drone edge model parameters are updated based on the overall zero-order gradient estimation of the drone edge model parameters and the asynchronous effective update coefficients: (45); in, This represents the updated edge model parameters for the drone. Indicates the zero-order update step size. Indicates the target drone At any moment The corresponding asynchronous effective update coefficient.

[0164] Using the above methods, the target drone By performing only two forward computations and uploading training payload data on the edge, the edge server can complete the recovery of perturbation intermediate features, loss difference calculation, zero-order gradient reconstruction, and model parameter update, thereby significantly reducing the computational and communication burden of training on the UAV.

[0165] The method for calculating the asynchronous effective update coefficient is as follows: The edge server reads the task request version number corresponding to the task. and the current global version number Calculate the version obsolescence of this update. : (46); when At that time, it indicates the target drone. The version of the drone-side edge model used is consistent with the global model version currently maintained by the edge server, and the training payload data does not contain outdated versions; when This indicates that from the time the task was dispatched until the training payload data arrived at the edge server, the edge server had already completed training based on other training tasks. This is the second effective model update. With... As the version of the training payload increases, the staleness of the data relative to the current global model version also increases.

[0166] Edge servers calculate the asynchronous effective update coefficient based on version staleness: (47); in, This indicates the obsolescence penalty coefficient; This indicates the lower limit of the effective update coefficient; This represents the maximum allowed version staleness. When the version staleness exceeds the maximum allowed version staleness, the asynchronous effective update coefficient for this update is set to 0, indicating that this update will not participate in the update of the UAV edge model parameters. The asynchronous effective update coefficient... Corresponding to the currently arriving training task, it is used to adjust the update magnitude when the training task participates in the current update of the edge model parameters of the UAV, and is not used for the next training task of the target UAV.

[0167] S5.6: The edge server removes the in-transit training task from the in-transit task set. Move out of the middle.

[0168] S6: The edge server updates the effective utilization deficit queue of the target drone, and the updated effective utilization deficit queue is fed back to the subsequent comprehensive scheduling score calculation.

[0169] After the task completes server-side reception, zero-order reconstruction, and asynchronous processing, the edge server updates the drone's availability status; if Then, a valid model update is performed, setting the global version number. and update the drones Last successful update time ;like In this case, the training task will not participate in model parameter updates, will not increment the global version number, and will not update. However, the effective utilization deficit queue of the corresponding drone can be updated based on the effective update difference.

[0170] Define valid update threshold And calculate the effective update difference: (48); when hour, This indicates that the validity of the update meets the requirements and no additional compensation is generated; when This indicates that although the update was received and processed by the server, its actual contribution was weakened due to its high obsolescence, thus resulting in a positive deficit increment.

[0171] When drones After the training task arrives and completes asynchronous updates, update its effective utilization deficit queue: (49); in, To optimize the use of the deficit queue after the update; This is the deficit leakage coefficient, used to prevent the long-term accumulation of historical compensation from excessively affecting subsequent scheduling; To effectively utilize the upper limit of the deficit queue, it is used to limit the intensity of compensation; This represents the truncation function.

[0172] Through the above mechanism, this invention enables the scheduling score to simultaneously consider the potential value of the UAV's local data, the state of the UAV having not been updated for a long time, insufficient historical effective utilization, and the impact of asynchronous obsolescence on the update contribution. Specifically, the effective utilization deficit queue does not simply compensate for "few selections," but rather compensates for "scheduled but ineffective contributions." This allows the scheduler to more precisely distinguish the actual training contribution of the UAV and feed the training update results back to the next round of scheduling scoring, thus forming a closed loop of "comprehensive scoring—task assignment—asynchronous update—effectiveness evaluation—deficit feedback—next round scoring."

[0173] S7: Repeat S2-S6 until the training time budget or training stop condition is reached, and then send the final updated edge server model parameters to the subsequently scheduled drones.

[0174] Example 2: A heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling system is used to implement the asynchronous splitting zero-order cooperative learning scheduling method for heterogeneous UAV swarms, such as... Figure 6 As shown, it includes: The UAV-side module is used for the target UAV to perform split zero-order detection training, generate and upload training payload data, and receive the final updated edge server-side model parameters.

[0175] The edge server module is used to initialize the drone ensemble and the drone edge model. Server-side model The current global model version number, candidate pool, in-transit task set, and scheduling status of each drone; at the current moment, based on the scheduling status of each drone and the in-transit task set, determine the currently schedulable drone set from the drone set. Based on the current set of schedulable drones Update the candidate pool, calculate the comprehensive scheduling score of each candidate drone in the pool, and determine whether the multi-candidate competitive scheduling trigger condition is met. If it is met, determine the target drone based on the comprehensive scheduling score. If not, keep the candidate pool open and continue waiting for new schedulable drones to join or for the multi-candidate competitive scheduling trigger condition to be met. After receiving the training payload data, update the model parameters on the edge server. And drone edge model parameters Then, the training task in transit is removed from the set of tasks in transit. Remove from the queue; update the effective utilization deficit queue for the target UAV, and feed the updated effective utilization deficit queue back into the subsequent comprehensive scheduling score calculation. Send the final updated edge server-side model parameters to the UAV-side module.

[0176] The core of this invention lies in unifying candidate pool construction, comprehensive scoring, UAV selection, splitting zero-order detection, asynchronous edge server updates, and effective utilization of compensation feedback into a closed loop. This closed loop has the following characteristics: (1) Scheduling is not done in isolation, but directly determines which UAV will perform the next Split-ZO probing.

[0177] (2) Training is not conducted in isolation; the effectiveness of training will be fed back to subsequent scheduling.

[0178] (3) The edge server not only performs model updates, but is also responsible for candidate pool management, version management and deficit feedback.

[0179] (4) The UAV participates in training in a low-burden manner, while the server undertakes the main computation.

[0180] (5) Simultaneously address the issues of Non-IID data value, asynchronous obsolescence, candidate competition, and insufficient effective utilization.

[0181] This invention also provides a lightweight edge-splitting zero-order training protocol, the key features of which include: (1) The UAV only performs two forward calculations: the original edge model and the perturbation edge model.

[0182] (2) The drone does not perform a complete backpropagation.

[0183] (3) The UAV uploads the difference between the original intermediate features and the perturbation features, rather than the complete gradient.

[0184] (4) The UAV uploads a random seed and perturbation scale, enabling the edge server to reconstruct the perturbation direction.

[0185] (5) The edge server completes the zero-order gradient estimation based on the loss difference and the reconstruction perturbation direction.

[0186] (6) The server model is updated through backpropagation, and the edge model is updated through zero-order estimation.

[0187] This invention also provides a multi-candidate contention scheduling mechanism, the key points of which include: (1) The edge server does not immediately dispatch tasks to idle drones, but first establishes a candidate pool.

[0188] (2) The candidate pool aggregates UAVs that enter the idle state at similar times through a micro-waiting window.

[0189] (3) The number of candidates in the candidate pool is controlled by the Watermark condition.

[0190] (4) The candidate pool controls the waiting limit through the Timeout condition.

[0191] (5) The candidate pool avoids the server from being completely idle by using the Server-starvation condition.

[0192] (6) Winner selection is carried out in a multi-candidate set, rather than being passively assigned in a single-candidate state.

[0193] This invention also provides a comprehensive scheduling scoring and effective utilization compensation mechanism, the key points of which include: (1) The scheduling score simultaneously integrates static data value, AoU-like state and effective utilization deficiency.

[0194] (2) Static data value is used to reflect the potential contribution of different UAV local data to global training.

[0195] (3) AoU-like status is used to increase the competitive chances of drones that have not been successfully updated for a long time.

[0196] (4) stale delta is used to measure how old the version is when the update arrives.

[0197] (5) The effective update coefficient ηi(t) is used to suppress the influence of old updates.

[0198] (6) eta gap is used to characterize the extent to which the actual effective update coefficient is below the threshold.

[0199] (7) The deficit queue is used to compensate for drones that are "scheduled but have insufficient effective contribution".

[0200] (8) The deficit queue is merged into the subsequent scheduling score through multiplication.

[0201] Example 3: This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling method.

[0202] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0203] The processor is used to execute all or part of the steps in the heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling method as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0204] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling method described in the above embodiments.

[0205] Example 4: This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0206] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the heterogeneous UAV swarm asynchronous split zero-order cooperative learning scheduling method described in the various embodiments of this application.

[0207] The aforementioned storage media include: flash memory, hard disks, multimedia cards, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disks, optical discs, servers, APP (Application) application stores, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method.

[0208] Example 5: This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method.

[0209] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0210] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0211] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A heterogeneous unmanned aerial vehicle (UAV) swarm asynchronous splitting zero-order cooperative learning scheduling method, characterized in that, Includes the following steps: S1: Edge server initializes drone collection Drone edge model, server-side model, current global model version number, candidate pool In-transit task collection And the scheduling status of each drone; S2: The edge server, at the current moment, determines the scheduling status of each drone and the set of tasks in progress. Determine the currently schedulable drone set from the drone set. ; S3: Update the candidate pool based on the current set of schedulable drones, calculate the comprehensive scheduling score of each candidate drone in the candidate pool, and determine whether the multi-candidate competitive scheduling trigger condition is met. If the conditions are met, the target UAV is determined based on the comprehensive scheduling score; If the conditions are not met, the candidate pool remains open, and the system continues to wait for new schedulable drones to join the candidate pool or for the multi-candidate competition scheduling trigger condition to be met. S4: The target UAV performs a split zero-order detection training task, generates and uploads training payload data; S5: After receiving the training payload data, the edge server updates the model parameters on the edge server side. And drone edge model parameters Then the edge server removes the in-transit training task, namely the split zero-order probe training task, from the in-transit task set. Remove from the middle; S6: The edge server updates the effective utilization deficit queue of the target drone, and the updated effective utilization deficit queue is fed back to the subsequent comprehensive scheduling score calculation. S7: Repeat S2-S6 until the training time budget or training stop condition is reached, and then send the final updated edge server model parameters to the subsequently scheduled drones.

2. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 1, characterized in that, The drone collection Represented as: (1); in, Indicates the first A drone, Indicates the drone's serial number. This indicates the number of drones, each drone Holding local datasets ; The global model is a complete deep neural network model used to complete the target learning task. The edge server divides the global model into drone-end edge models deployed on the drone according to the preset model segmentation positions. and server-side models deployed on edge servers The UAV edge model is used to extract front-end features from the input data to obtain front-end features; the server-side model is used to receive the front-end features and complete subsequent feature processing, prediction output, and loss calculation. This represents the parameters of the drone's edge model. Indicates server-side model parameters; The candidate pool This is used to temporarily cache drones that enter the schedulable state and meet the scheduling conditions within a short time window, and to use them as candidate drones; the state of the candidate pool includes an open state and a closed state, with the initial state being the closed state. The set of tasks in transit Used for recording Training tasks that have been dispatched but have not yet arrived at the edge server are represented as follows: (4); in, Indicates time, Indicates with drones The corresponding training tasks en route, This indicates the dispatch time for training missions en route. This indicates the estimated time for the training task to arrive at the edge server. Each training mission on the way It should include at least the drone identifier, the time of dispatch of the training task en route, the estimated time of arrival of the training task at the edge server, and the version number of the task request corresponding to receiving the training task, represented as: (5); in, Indicates drone The task request version number received when a training task is in transit is determined by the current global model version number maintained by the edge server at the time the training task is dispatched. (6); Global model version number at initial moment =0; The drone At any moment The scheduling state is represented as : (7); in, For drones The time of the most recent valid update completion. For drones Effective use of deficit queues; For drones The estimated time to restore the schedulable state.

3. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 1, characterized in that, The currently schedulable set of drones This indicates that the estimated time for restoring a schedulable state has been reached at the current moment, and the set of tasks in transit... The set of drones that do not correspond to any in-transit training missions is represented as: (8); in, Indicates drone The expected time for restoring a schedulable state has been reached; Represents the set of tasks in progress There is no drone in China For training tasks that have not yet arrived at the edge server, the edge server determines the drone's status when both of the above conditions are met. It is in a schedulable state.

4. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 1, characterized in that, S3 includes the following steps: S3.1: Currently available set of schedulable drones When the candidate pool is not empty, if the candidate pool is currently closed and there are drones in the currently schedulable drone set that meet the scheduling conditions, the edge server will open the candidate pool and record the opening time of the candidate pool, and add the drones in the currently schedulable drone set that meet the scheduling conditions as candidate drones to the candidate pool; if the candidate pool is currently open, the edge server will continue to add the drones that have newly entered the schedulable state at the current time and meet the scheduling conditions to the candidate pool. The candidate pool opening time is: (9); in, Open time for the candidate pool; The state update process of the candidate pool is represented as follows: (10); in, For the updated candidate pool, Indicates the currently schedulable set of drones A subset of UAVs that meet the scheduling conditions; S3.2: Calculate the static data value, AoU-like state, and link feasibility factor of each candidate UAV, and then determine the comprehensive scheduling score of each candidate UAV in the candidate pool. ; S3.3: The edge server determines whether the current candidate pool meets any preset multi-candidate competition scheduling trigger condition. If yes, it executes S3.4; otherwise, it keeps the candidate pool open and continues to wait for new schedulable drones to join or waits for the multi-candidate competition scheduling trigger condition to be met. The multi-candidate competition scheduling triggering conditions include at least the candidate pool number threshold triggering condition, the waiting time timeout triggering condition, and the server idle triggering condition; Selection pool quantity threshold trigger condition: (25); in, Candidate pool Number of candidate drones This is the threshold for the candidate pool size; Waiting timeout trigger condition: (26); in, For flexible micro-wait windows; Server idle trigger condition: (27); in, Indicates time Edge server idle trigger variable; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This indicates that there are currently no training tasks en route that have not yet arrived at the edge server; This indicates that there is at least one candidate drone in the candidate pool; Let represent the minimum contention waiting window, and satisfy . ; S3.4: Select the candidate UAV with the highest comprehensive scheduling score as the target UAV. ; (29)。 5. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 4, characterized in that, S3.2 specifically includes: S3.2.1: Predetermine the static data value of each UAV; Specifically, setting up drones The local dataset is The total number of categories is drones In the The empirical distribution ratio on the class samples is Then drone The label distribution entropy is expressed as: (11); in, For drones The label distribution entropy; This is the sample category number; To prevent abnormally small constants in logarithmic calculations; Edge servers further build lightweight probe models and using a pre-configured public reference training set. The lightweight probe model undergoes a small number of warm-up training rounds; the common reference training set Local datasets independent of each drone This is used to enable lightweight probe models to acquire basic recognition capabilities for target-oriented learning tasks; After the warm-up training is completed, the edge server will reduce the probe model parameters. Distributed to all drones; drones Use its local dataset The lightweight probe model is evaluated, and its performance on the local dataset is calculated. The average loss is calculated and returned to the edge server, denoted as . : (12); in, For drones Local dataset The number of samples in For local datasets One of the training samples, For sample labels; Represents the loss function. g ( · ; θp ) represents a lightweight probe model. This represents the parameters of the lightweight probe model. This represents the average loss of the lightweight probe model; Then, the drone The average loss and label distribution entropy are fused to obtain the original data value score. : (13); in, Indicates drone The original data value score, This represents the label entropy weight coefficient. Indicates the entropy adjustment index; The edge server calculates the coefficient of variation of the value scores of all drone raw data. CV : (14); in, The coefficient of variation represents the value score of all raw drone data; This represents the standard deviation of the value scores of all raw drone data. This represents the average of the raw data value scores for all drones. like CV Less than the preset threshold If the data value of each drone is not significantly different, then the static data value of each drone will be set to a uniform distribution. (15); in, The value of static data from drones; like CV Greater than or equal to the preset threshold Then, the value scores of the original data are normalized: (16); in, This indicates the normalized drone. Data value score, Number the drone. The original data value score, This represents the minimum value score among all the raw data scores of drones. This represents the maximum value score among all the raw data of all drones. Introducing basic protection weights The smoothed, normalized data value scores are obtained as follows: (17); in, Indicates the smoothed drone Normalized data value score, To represent the basic protection weight; This represents the sum of the normalized data value scores for all drones; The smoothed, normalized data value scores are then soft-mixed with a uniform distribution to obtain the final static data value: (18); in, ξ The mixing coefficient; S3.2.2: Calculate the current time of each candidate UAV. AoU-like states under; Candidate drones The most recent valid update completion time was Then at the current moment The AoU-like state is as follows: (19); in, It is in an AoU-like state. For growth index; S3.2.3: Combine the static data value and AoU-like state of each candidate UAV to obtain the basic score of each candidate UAV; (20); in, Based on the score; S3.2.4: Calculate the comprehensive scheduling score for each candidate UAV based on its base score; Edge servers for each drone Maintain an efficient utilization of the deficit queue Read the current effective utilization deficit queue And this is fed back into the overall scheduling score in a multiplicative manner: (21); in, These are the weighting coefficients.

6. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 1, characterized in that, S4 specifically includes the following steps: S4.1: The edge server dispatches a split zero-order detection training task to the target drone, clears the candidate pool and sets the candidate pool status to the closed state, and enters the next round of candidate drone collection process; Specifically: Order Edge server to target drone Distribute the current edge model parameters of the drone. Task request version number And training control parameters, and based on the target UAV The computing power, uplink bandwidth, and training payload data size are used to calculate the estimated time for the training task en route, i.e., the split zero-order probe training task, to reach the edge server. ; Target drone At the present moment Assigned to perform a zero-order split detection training mission, the forward computation of a single UAV edge model is Target drone The edge computing capability is The estimated edge computing latency required for the UAV to perform the zero-order split detection training task for: (30); Target drone The size of the uploaded training payload data is At present The uplink bandwidth is The estimated uplink communication latency required to upload training payload data is then... for: (31); Therefore, the estimated time to reach the edge server for: (32); Then, the edge server identifies the target drone. Expected arrival time at the edge server and task request version number Record to in-transit task collection ; S4.2: The target UAV selects local mini-batch training samples from its local dataset and calculates the original intermediate features corresponding to the local mini-batch training samples; Target drone In this training, it was derived from its local dataset. The selected local mini-batch training samples are ,in, This represents the input data in this local mini-batch training sample. Indicates the corresponding sample label; target drone Using the current drone edge model parameters For input data Perform forward computation to obtain the original intermediate features. : (33); S4.3: Target UAV A random seed is generated according to the random seed generation rule, and then a perturbation direction is generated. The parameters of the perturbation UAV edge model are then constructed based on the perturbation direction. Target drone Based on random seed Generate perturbation directions for each parameter block of the UAV edge model, for the th A parameter block, the perturbation direction is denoted as . The parameters of the disturbed UAV edge model are constructed according to the following formula: (34); in, This indicates the parameter block number of the drone edge model. Indicates the first The original parameters of each parameter block, Indicates the first The parameters after perturbation of each parameter block. Indicates the first The perturbation direction corresponding to each parameter block Indicates the first The perturbation scale corresponding to each parameter block; The parameters after perturbation of each parameter block Parameters of the UAV edge model after perturbation ; S4.4: Target UAV Using the perturbed drone edge model parameters For input data Perform forward computation to obtain intermediate perturbation features; (35); in, Indicates the target drone The intermediate features of the disturbance are calculated based on the edge model of the disturbed UAV. This represents the parameters of the drone edge model after the disturbance; S4.5: The target UAV is based on the original intermediate features and intermediate features of perturbation Calculate the intermediate feature differences and generate training payload data; The intermediate feature difference is divided into: (36); in, Indicates the intermediate feature difference; Target drone Generate training payload data: (37); in, Indicates the target drone Training payload data uploaded to the edge server, This indicates the number of parameter blocks in the drone edge model. This represents the set of perturbation scales corresponding to each parameter block.

7. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 1, characterized in that, S5 specifically includes the following steps: S5.1: Edge server receives target drone Uploaded training payload data Then, based on the original intermediate features and intermediate feature difference Recover the intermediate features of the perturbation; S5.2: The edge server calculates the original loss and the perturbation loss under the same edge server model snapshot, and then calculates the loss difference; Assume the edge server receives the training payload data. The snapshot of the edge server-side model parameters used at that time is The original loss is: (39); The disturbance loss is: (40); Loss difference is divided into: (41); in, Indicates based on original intermediate features The calculated original loss, Indicates based on perturbation intermediate features The calculated disturbance loss, This represents the difference in loss between the disturbance loss and the original loss; S5.3: The edge server reconstructs the perturbation direction based on the random seed and constructs a zero-order gradient estimate for the parameter block of the UAV edge model, thereby obtaining the overall zero-order gradient estimate of the UAV edge model parameters. Edge servers based on training payload data random seed Reconstruct the target drone The direction of the perturbation, for the first edge model of the UAV. The zeroth-order gradient estimate of a block of parameters is: (42); in, Indicates the original loss Compared to the edge model of the drone, the first Zero-order gradient estimation of a block of parameters; By combining the zero-order gradient estimates of each parameter block, we obtain the overall zero-order gradient estimate of the parameters of the UAV edge model: (43); in, This represents the global zeroth-order gradient estimate of the parameters of the UAV edge model; S5.4: Edge servers based on original losses Calculate its parameters relative to the edge server-side model. The gradient is used for backpropagation to update the model parameters on the edge server. (44); in, This represents the updated edge server-side model parameters. This indicates the optimizer update operation of the edge server-side model. Indicates the original loss Relative to edge server-side model parameters The true gradient; S5.5: The UAV edge model performs a zero-order update based on the global zero-order gradient estimation of the UAV edge model parameters, and obtains the updated UAV edge model parameters. Before updating the parameters of the UAV edge model corresponding to the current training task, the edge server reads the asynchronous valid update coefficients. The drone edge model parameters are updated based on the overall zero-order gradient estimation of the drone edge model parameters and the asynchronous effective update coefficients: (45); in, This represents the updated edge model parameters for the drone. Indicates the zero-order update step size. Indicates the target drone At any moment The corresponding asynchronous effective update coefficient; The method for calculating the asynchronous effective update coefficient is as follows: The edge server reads the task request version number corresponding to the task. and the current global version number Calculate the version obsolescence of this update. : (46); Edge servers calculate the asynchronous effective update coefficient based on version staleness: (47); in, This indicates the obsolescence penalty coefficient; This indicates the lower limit of the effective update coefficient; Indicates the maximum allowed version obsolescence; S5.6: The edge server removes the in-transit training task from the in-transit task set. Move out of the middle.

8. The heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to claim 1, characterized in that, The edge server updates the effective utilization deficit queue of the target drone, and the updated effective utilization deficit queue is fed back into the subsequent comprehensive scheduling score calculation, specifically as follows: The edge server updates the drone's availability status if... Then, a valid model update is performed, setting the global version number. and update the drones Last successful update time ;like In this case, the training task will not participate in model parameter updates, will not increment the global version number, and will not update. However, the effective utilization deficit queue of the corresponding drone is updated according to the effective update difference. Define valid update threshold And calculate the effective update difference: (48); When drones After the training task arrives and completes asynchronous updates, update its effective utilization deficit queue: (49); in, To optimize the use of the deficit queue after the update; The leakage coefficient is the factor indicating a deficit. To effectively utilize the upper limit of the deficit queue; This represents the truncation function.

9. A heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling system, used to implement the heterogeneous UAV swarm asynchronous splitting zero-order cooperative learning scheduling method according to any one of claims 1-8, characterized in that, include: The UAV-side module is used to perform split zero-order detection training on the target UAV, and to generate and upload training payload data. Receive the final updated edge server-side model parameters; The edge server module is used to initialize the drone ensemble and the drone edge model. Server-side model The current global model version number, candidate pool, in-transit task set, and scheduling status of each UAV; Based on the scheduling status and on-the-go tasks of each drone at the current moment, determine the set of drones that can be scheduled at the moment. Based on the current set of schedulable drones Update the candidate pool, calculate the comprehensive scheduling score of each candidate UAV in the candidate pool, and determine whether the multi-candidate competition scheduling trigger condition is met; If the conditions are met, the target UAV is determined based on the comprehensive scheduling score; If the conditions are not met, the candidate pool remains open, and the system continues to wait for new schedulable drones to join or for the multi-candidate competitive scheduling trigger condition to be met; after receiving the training payload data, the model parameters on the edge server are updated. And drone edge model parameters Then, the training task in transit is removed from the set of tasks in transit. Remove from the middle; Update the effective utilization deficit queue of the target UAV, and feed the updated effective utilization deficit queue back to the subsequent comprehensive scheduling score calculation; send the final updated edge server-side model parameters to the UAV-side module.

10. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the heterogeneous unmanned aerial vehicle swarm asynchronous splitting zero-order cooperative learning scheduling method according to any one of claims 1-8.