A method for drone-assisted federated learning

Abstract

This invention provides a method for drone-assisted federated learning, comprehensively considering vehicle mobility, incentive decision-making, and transfer problems. Addressing the issue of participants' reluctance to unconditionally contribute their resources for local model training, a fair incentive mechanism and trust model are designed to encourage participants to upload reliable model updates and join the federated learning task. Regarding vehicle mobility, this invention proposes a two-way mechanism utilizing drone service migration and drone movement to ensure the integrity of the federated learning task. Simultaneously considering the participation willingness of user vehicles, this invention designs a game theory-based user vehicle incentive method, enabling user vehicles and drones to make corresponding decisions to maximize their respective utility.

Description

Technical Field

[0001] This invention belongs to the field of machine learning, specifically relating to a method for drone-assisted federated learning. Background Technology

[0002] In traditional centralized machine learning, participants send their local data directly to a central server, which may be untrusted. Once uploaded, participants have no way to delete the data and are completely unaware of how the central server uses it. Furthermore, participants are vulnerable to attacks from internal or external attackers during transmission, leading to the risk of privacy breaches on their local data.

[0003] Federated learning solves the "data silo" problem by replacing data sharing with model sharing, ensuring both data privacy and the performance of the global model. However, federated learning still faces significant challenges, such as how to guarantee the privacy of uploaded model updates, how to avoid unreliable model updates, and how to incentivize participants to contribute their resources.

[0004] The patent "Resource Control Method and Device in Federated Learning Model Training Process" proposes a method for UAVs to obtain the total time and total energy consumption models required for a task through model parameters, and then determine the target resources based on these models. This patent considers the impact of noise interference on UAVs and utilizes multiple transmissions to process the federated learning model parameters to ensure model accuracy and security, but this reduces the final model update time. The patent "Data Processing Method for Highway UAV Inspection Based on Federated Adaptive Learning" proposes a data processing method that adaptively adjusts the number of iterations in the next training round using an adaptive algorithm, allocating more computational tasks to edge servers. While this patent standardizes the data received by the UAV to adjust the local iteration count for different user vehicles in a timely manner, it does not consider the participation willingness of the user vehicles themselves, which may reduce model accuracy to some extent. The patent "A Resource Management Method for Distributed Learning of UAV Swarms Based on Two-Way Auction Game Theory" proposes a resource management method for distributed learning of UAV swarms based on two-way auction game theory. It uses two-way auction theory for iterative updates, enabling the market to quickly converge to the point of maximizing social welfare. This patent uses a two-way auction game to enable user vehicles to quickly reach a balance in federated learning and makes a reasonable analysis of whether user vehicles participate. However, the patent does not take into account that drones cannot simultaneously serve user vehicles with different needs, and its application areas are limited. Summary of the Invention

[0005] This invention addresses the problems in existing technologies by providing a method and system for UAV-assisted federated learning. To perform federated learning tasks in an air-to-ground network, a UAV acts as a server, while ground-based user vehicles collaboratively complete model training. This invention first considers the performance parameters of the user vehicles, then generates a trust model based on processed data, selects suitable user vehicles, and takes into account that user vehicles may change their interest in different federated learning tasks based on their own circumstances. Trust values ​​are used to determine whether to participate in model iteration, improving the accuracy of model generation. The UAV's coverage area is incorporated into the overall model, allowing the UAV to consider data migration or following, thus improving the model's rationality.

[0006] To address the above technical problems, this invention provides the following technical solution: a method for drone-assisted federated learning, comprising the following steps:

[0007] S1. After the UAV sends out the global model, it makes a participation judgment on user vehicles within the target range and determines the initial user vehicles that meet the preset conditions.

[0008] S2. Select user vehicles of interest from the initial user vehicles, calculate their channel quality to obtain the trust value, determine the user vehicles to participate in the global model training, and the user vehicles determine the number of iterations to participate in the model training based on their own channel quality and trust value.

[0009] S3. Drones and user vehicles perform federated global model training. During the learning process, a trust mechanism is introduced to exclude malicious user vehicles in each model training iteration. Drones calculate the estimated departure time of user vehicles and determine whether to implement a service migration strategy based on a preset cost threshold. For high-performance user vehicles, it is determined whether their following cost is greater than the migration cost. If so, a service migration strategy is implemented; otherwise, a drone following strategy is implemented. After each round of training, a trust value is fed back to the drone to update the global model until the model training is completed.

[0010] Furthermore, the aforementioned step S1 specifically includes the following sub-steps:

[0011] S101. The similarity and temporal participation of the user vehicle's stored model with the global model calculated by the drone:

[0012]

[0013] Ψ i =Ω i +θ i (3)

[0014] ω i For user vehicle V i Stored model parameters, The latest model parameters for region O; This refers to the gap between the user's vehicle model and the global model. The average interval of the set of wheels is provided by the drones in region O. Ωi∈[0,1],θ i ∈[-1,1); The threshold for determining region O; the drone calculates and estimates the user vehicle V. i The time to leave the current region O is χ1 and χ2 are coefficients determined based on the actual urban traffic environment, g o G represents the hourly traffic volume in area O. o This is the maximum hourly capacity of region O. o / G o The value L represents the level of congestion in the area. i It is the distance to travel before leaving, v i For the speed of the vehicle, For the quality of the calculated update, It's the packet loss rate. It's the CPU frequency. This represents the user's computing cost;

[0015] S102, Calculation Necessity Ψ i Then, define ξ i The decision result indicates whether a user's vehicle is suitable for participating in the model update, as shown in the following formula:

[0016]

[0017] Furthermore, as described above, step S2 includes the following sub-steps:

[0018] S201. After the UAV sends out the global model, it receives the parameters fed back by the initial user vehicle and then identifies the initial user vehicle as the user vehicle of interest.

[0019] S202. Calculate the error level of the user's vehicle of interest:

[0020]

[0021] in, It's the packet loss rate. It's the CPU frequency, θ i These are the vehicle's time parameters. and It is the deviation between the information transmitted between the drone and the user, and f(·) is the trust value conversion function corresponding to the deviation and the model update, which is used to control the trust value of each user within the same range.

[0022] S203. Calculate the trust value as follows:

[0023] 1-|μ i |,

[0024] S204. The vehicle of interest determines the number of iterations to participate in model training based on its trust value and its current channel quality.

[0025] Furthermore, in the aforementioned step S3, excluding vehicles belonging to malicious users specifically involves: calculating the trust value P = (ρ1, ρ2, ..., ρ...) of the vehicles of interest. N ),

[0026]

[0027] Where τ is the number of rounds of global update, Let i be the communication capability of node i during the τth global update. For the calculated update quality, μ i Indicates the degree of error. It's the packet loss rate. It's the CPU frequency, θ i is the vehicle's time parameter, and f(·) is the trust value conversion function corresponding to the deviation and model update, which controls the trust value of each user within the same range.

[0028] Furthermore, in step S2 above, the user vehicle determines the number of iterations for participating in model training based on its own channel quality and trust value; specifically, the UAV informs the user vehicle participating in model training of the reward R,τ before each iteration. i τ represents the user's decision, i.e., the number of rounds in which they participated in the global update. -i =(τ1..τ i-1 ..τ i+1 ..τ N ) indicates that, except for τ i The policy for all users except those mentioned above uses a trust value ρ. i To measure contribution level, a utility function is used to represent the difference between the user vehicle's reward and cost. This is calculated by subtracting the reward paid to the user vehicle from the total revenue gained from training the user vehicle model. The benefits gained by the drone are proportional to its reputation, level of participation, and user participation costs. The drone adjusts its reward R, and the user vehicle's reward V... i Adjust its participation wheel τ i .

[0029] Furthermore, in the aforementioned step S3, the service migration strategy includes the following sub-steps:

[0030] S301, Calculate user vehicle V i The model training task in UAV A o Energy consumption during execution;

[0031] S302. The drone calculates the iteration time for each user vehicle based on its computing power and the allocated model training workload; the total overhead C is calculated based on the iteration time and energy consumption. Vi ,

[0032] S303. Determine if there are any unfinished model training tasks when the user's vehicle leaves area O. If so, move the next model training task to area L to complete it, in the total cost C. Vi The time difference between the two migration services is added; when implementing the server migration strategy, when drone A in region L... l Obtain the model update results for the user's vehicle and transmit them to UAV A. o Afterwards, drone A o Complete the global model update and release the task for the next global model training.

[0033] Furthermore, in the aforementioned step S3, for high-performance user vehicles, determining whether their following overhead is greater than their migration overhead, if so, executes the service migration strategy; otherwise, the drone following strategy includes the following steps: S3.1, based on user vehicle V... i Time parameter θ i When θ i When it is less than 0, it means that the user vehicle V i The departure time is less than the time of one iteration, and the drone A o For V i Implement a tracking and movement strategy, and calculate the cost of the drone's movement.

[0034] S3.2 When executing the tracking movement strategy, the UAV needs to determine the following speed and the distance the UAV has moved. Then, compute the user set S = [V1, ... V1] for service migration. i ,...V N Total cost:

[0035]

[0036] The drone following overhead involved in service migration is:

[0037]

[0038] Among them, C i，O For user V i Overhead within range O,

[0039] S3.3 In order to reduce communication power consumption, when In drone A o Within a certain range, the mobile drone follows the user's vehicle; when C vi <Cf In drone A o A L Within the specified range, service migration can be selected.

[0040] Another aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program as steps of the method of the present invention.

[0041] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method described in the present invention.

[0042] Compared to existing technologies, the beneficial technical effects of the above technical solution in this invention are as follows: A method for drone-assisted federated learning, in order to execute federated learning tasks in an air-to-ground network, uses a drone as a server, with ground-based user vehicles collaboratively completing model training. First, the model similarity and time parameters of the user vehicle are used to determine whether the user vehicle should participate in the federated learning task. Then, a trust value model is used to unify the contribution of user vehicles to the model, thereby attracting high-quality user vehicles to participate. Furthermore, considering vehicle mobility, a two-way mechanism utilizing drone service migration and drone movement is proposed to ensure the complete execution of the federated learning task. Simultaneously, considering the user vehicle's willingness to participate, this invention designs a game theory-based user vehicle incentive method, enabling user vehicles and drones to make corresponding decisions to maximize their respective utility. Attached Figure Description

[0043] Figure 1 This is a ground scene image assisted by drones in existing technologies.

[0044] Figure 2 A flowchart for drones to determine vehicles within their range.

[0045] Figure 3 Flowchart of the user vehicle trust mechanism.

[0046] Figure 4 This is a flowchart of the drone migration process.

[0047] Figure 5 A strategy diagram for drones to follow high-performance vehicles.

[0048] Figure 6 This is a flowchart of the method of the present invention.

[0049] Figure 7 This is a communication map of unmanned aerial vehicles (UAVs) within region O.

[0050] Figure 8 This is a diagram of drone tracking and data migration. Detailed Implementation

[0051] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0052] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0053] In existing technologies, drone-assisted ground network scenarios such as Figure 1 As shown in the background, and considering the shortcomings of existing technologies, this invention constructs a drone-assisted ground network scenario consisting of drones and ground user terminals, whereby the drones provide auxiliary capabilities for ground communication. Specific application scenarios are as follows... Figure 7 As shown, the mobile drone acts as a server, responsible for task unloading, model updates, etc. The vehicle can act as a user terminal to perform tasks and connect to the drone via wireless communication.

[0054] refer to Figure 6 The present invention provides a method for unmanned aerial vehicle-assisted federated learning, comprising the following steps: S1, after the unmanned aerial vehicle distributes the global model, it performs participation judgment on user vehicles within the target range and determines the initial user vehicles that meet the preset conditions;

[0055] refer to Figure 2 Step S1 specifically includes the following sub-steps:

[0056] S101. The similarity and temporal participation of the user vehicle's stored model with the global model calculated by the drone:

[0057]

[0058] Ψ i =Ω i +θ i (3)

[0059] ω i For user vehicle V i Stored model parameters, The latest model parameters for region O; This refers to the gap between the user's vehicle model and the global model. The average interval of the set of wheels is provided by the drones in region O. Ωi∈[0,1],θi ∈[-1,1); It is the threshold that determines region O;

[0060] Based on the current traffic conditions in the area, the drone calculates and estimates the user vehicle V. i The time to leave the current region O is χ1 and χ2 are coefficients determined based on the actual urban traffic environment. o This indicates the hourly traffic volume in area O. G o This is the maximum hourly capacity of region O. o / G o The value represents the level of congestion in the area; the higher the value, the longer it takes for vehicles to leave the area. i It is the distance to travel before leaving, v i For the speed of the vehicle, For the quality of the calculated update, It's the packet loss rate. It's the CPU frequency. This represents the user's computing cost.

[0061] S102, Calculation Necessity Ψ i Then, define ξ i The decision result indicates whether a user's vehicle is suitable for participating in the model update, as shown in the following formula:

[0062]

[0063] The necessity of participation is influenced by both the model similarity Ωi and the estimated departure time parameter θ. i The shadow, if ξ i If it is 0, then the user's vehicle V i They will not participate in this mission.

[0064] S2. Select user vehicles of interest from the initial user vehicles, calculate their channel quality to obtain a trust value, and determine the user vehicles to participate in the global model training. Each user vehicle determines the number of iterations it will participate in the model training based on its own channel quality and trust value. (Reference) Figure 3 Step S2 includes the following sub-steps:

[0065] S201. After the UAV sends out the global model, it receives the parameters fed back by the initial user vehicle and then identifies the initial user vehicle as the user vehicle of interest.

[0066] S202. Calculate the error level of the user's vehicle of interest:

[0067]

[0068] in, It's the packet loss rate. It's the CPU frequency, θ i These are the vehicle's time parameters. and It is the deviation between the information transmitted between the drone and the user, and f(·) is the trust value conversion function corresponding to the deviation and the model update, which is used to control the trust value of each user within the same range.

[0069] S203. Calculate the trust value as follows:

[0070] 1-|μi|,

[0071] S204. The vehicle of interest determines the number of iterations to participate in model training based on its trust value and its current channel quality.

[0072] Specifically, before each iteration, the drone informs the users participating in model training of the vehicle's reward R,τ. i τ represents the user's decision, i.e., the number of rounds in which they participated in the global update. -i =(τ1..τ i-1 ..τ i+1 ..τ N ) indicates that, except for τ i The policy for all users except those mentioned above uses a trust value ρ. i To measure contribution level, a utility function is used to represent the difference between the user vehicle's reward and cost. This is calculated by subtracting the reward paid to the user vehicle from the total revenue gained from training the user vehicle model. The benefits gained by the drone are proportional to its reputation, level of participation, and user participation costs. The drone adjusts its reward R, and the user vehicle's reward V... i Adjust its participation wheel τ i .

[0073] S3, drones, and user vehicles perform federated global model training, introducing a trust mechanism during the learning process to exclude malicious user vehicles in each model training iteration. (Reference) Figure 8 The drone tracking and data migration graph is used to calculate the estimated departure time of user vehicles and determine whether to implement a service migration strategy based on a preset cost threshold. For high-performance user vehicles, it is determined whether the tracking cost is greater than the migration cost. If so, the service migration strategy is executed; otherwise, the drone tracking strategy is implemented. After each round of training, the trust value is fed back to the drone to update the global model until the model training is completed.

[0074] In step S3, excluding vehicles belonging to malicious users specifically involves: calculating the trust value P = (ρ1, ρ2, ..., ρ...) for vehicles belonging to users of interest. N ),

[0075]

[0076] Where τ is the number of rounds of global update, Let i be the communication capability of node i during the τth global update. For the calculated update quality, μ i Indicates the degree of error. It's the packet loss rate. It's the CPU frequency, θ i is the time parameter of the user's vehicle, and f(·) is the trust value conversion function corresponding to the deviation and model update. Its function is to control the trust value of each user within the same range.

[0077] In step S3, the drone migration strategy is as follows: Drones in region O issue tasks. In addition to uploading their own parameters, participating user vehicles must also upload their estimated travel paths. The drone calculates the iteration time for each user vehicle based on its computing power and the assigned task load. Then, based on the vehicles' travel paths, it determines which vehicles are suitable to complete the task within the departure time. For user vehicles that can complete the task, the normal process continues. For user vehicles that cannot complete the task, it is determined whether service migration (calculated in regions O and L) is necessary based on their paths. A cost threshold is set; if the cost is less than the threshold, service migration occurs; if it is greater than the threshold, calculation is not performed in region O.

[0078] refer to Figure 4 In step S3, the service migration strategy includes the following sub-steps:

[0079] S301, Calculate user vehicle V i The model training task in UAV A o Energy consumption during execution;

[0080] S302. The drone calculates the iteration time for each user vehicle based on its computing power and the allocated model training workload; the total overhead C is calculated based on the iteration time and energy consumption. Vi ,

[0081] S303. Determine if there are any unfinished model training tasks when the user's vehicle leaves area O. If so, move the next model training task to area L to complete it, in the total cost C. Vi The time difference between the two migration services is added; when implementing the server migration strategy, when drone A in region L... l Obtain the model update results for the user's vehicle and transmit them to UAV A. o Afterwards, drone A o Complete the global model update and release the task for the next global model training.

[0082] refer to Figure 5Service migration requires the coordination of multiple drones, which is not always a good option. Considering the mobility of drones, users with high computing power can be identified. In a given iteration, to ensure that a user can continue to complete the task without leaving the drone's range, the drone can decide to follow that user.

[0083] In step S3, for high-performance user vehicles, it is determined whether their following overhead is greater than their migration overhead. If so, a service migration strategy is executed; otherwise, a drone following strategy is implemented, including the following steps:

[0084] S3.1, Based on user vehicle V i Time parameter θ i When θ i When it is less than 0, it means that the user vehicle V i The departure time is less than the time of one iteration, and the drone A o For V i Implement a tracking and movement strategy, and calculate the cost of the drone's movement.

[0085] S3.2 When executing the tracking movement strategy, the UAV needs to determine the following speed and the distance the UAV has traveled. Then, compute the user set S = [V1, ... V1] for service migration. i ,...V N Total cost:

[0086]

[0087] The drone following overhead involved in service migration is:

[0088]

[0089] Among them, C i，O For user V i Overhead S3.3 within range O: To reduce communication power consumption, when... In drone A O Within a certain range, the mobile drone follows the user's vehicle; when C Vi <C f In drone A O A L Within the specified range, service migration can be selected.

[0090] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for drone-assisted federated learning, characterized in that, Includes the following steps: S1. After the drone distributes the global model, it performs participation judgment on user vehicles within the target range to determine the initial user vehicles that meet the preset conditions; specifically, it includes the following sub-steps: S101. The similarity and temporal participation of the user vehicle's stored model with the global model calculated by the drone: P i =Oh i +θ i (3) ω i For user vehicle V i Stored model parameters, The latest model parameters for region O; This refers to the gap between the user's vehicle model and the global model. The average interval of the set of wheels is provided by the drones in region O. Ωi∈[0,1],θ i ∈[-1,1); It is the threshold that determines region O. Calculate and estimate the user vehicle V for the drone i Time to leave the current region O χ1 and χ2 are coefficients determined based on the actual urban traffic environment, g o G represents the hourly traffic volume in area O. o It is the maximum capacity of region O per hour, g o / G o The value L represents the level of congestion in the area; the higher the value, the longer it takes for vehicles to leave the area. i It is the distance to travel before leaving, v i For the speed of the vehicle, S102, Calculation Necessity Ψ i Then, define ξ i The decision result indicates whether a user's vehicle is suitable for participating in the model update, as shown in the following formula: S2. Select user vehicles of interest from the initial user vehicles, calculate their channel quality to obtain the trust value, determine the user vehicles to participate in the global model training, and the user vehicles determine the number of iterations to participate in the model training based on their own channel quality and trust value. S3. Drones and user vehicles perform federated global model training. During the learning process, a trust mechanism is introduced to exclude malicious user vehicles in each model training iteration. Drones calculate the estimated departure time of user vehicles and determine whether to implement a service migration strategy based on a preset cost threshold. For high-performance user vehicles, it is determined whether their following cost is greater than the migration cost. If so, a service migration strategy is implemented; otherwise, a drone following strategy is implemented. After each round of training, a trust value is fed back to the drone to update the global model until the model training is completed.

2. The method for UAV-assisted federated learning according to claim 1, characterized in that, Step S2 includes the following sub-steps: S201. After the UAV sends out the global model, it receives the parameters fed back by the initial user vehicle and then identifies the initial user vehicle as the user vehicle of interest. S202. Calculate the error level of the user's vehicle of interest: in, It's the packet loss rate. It's the CPU frequency, θ i These are the vehicle's time parameters. and It is the deviation between the information transmitted between the drone and the user, and f(·) is the trust value conversion function corresponding to the deviation and the model update, which is used to control the trust value of each user within the same range. S203. Calculate the trust value as follows: 1-|m i |, S204. The vehicle of interest determines the number of iterations to participate in model training based on its trust value and its current channel quality.

3. The method for UAV-assisted federated learning according to claim 2, characterized in that, In step S3, excluding vehicles belonging to malicious users specifically involves: calculating the trust value P = (ρ1, ρ2, ..., ρ...) of the vehicles of interest. N ), Where τ is the number of rounds of global update, Let i be the communication capability of node i during the τth global update. For the calculated update quality, μ i Indicates the degree of error. It's the packet loss rate. It's the CPU frequency, θ i is the vehicle's time parameter, and f(·) is the trust value conversion function corresponding to the deviation and model update, which controls the trust value of each user within the same range.

4. The method for UAV-assisted federated learning according to claim 3, characterized in that, In step S2, the user vehicle determines the number of iterations to participate in model training based on its own channel quality and trust value; Specifically, before each iteration, the drone informs the users participating in model training of the vehicle's reward R,τ. i τ represents the user's decision, i.e., the number of rounds in which they participated in the global update. -i =(τ1..τ i-1 ..τ i+1 ..τ N ) indicates that, except for τ i The policy for all users except those mentioned above uses a trust value ρ. i To measure contribution level, a utility function is used to represent the difference between the user vehicle's reward and cost. This is calculated by subtracting the reward paid to the user vehicle from the total revenue gained from training the user vehicle model. The benefits gained by the drone are proportional to its reputation, level of participation, and user participation costs. The drone adjusts its reward R, and the user vehicle's reward V... i Adjust its participation wheel τ i .

5. A method for drone-assisted federated learning according to claim 4, characterized in that, In step S3, the service migration strategy includes the following sub-steps: S301, Calculate user vehicle V i The model training task in UAV A o Energy consumption during execution; S302. The drone calculates the iteration time for each vehicle based on the user's vehicle's computing power and the allocated model training task. The total overhead C is calculated based on the iteration time and the energy consumption of the execution. Vi , S303. Determine if there are any unfinished model training tasks when the user's vehicle leaves area O. If so, move the next model training task to area L to complete it, in the total cost C. Vi The time difference between the two migration services is added; when implementing the server migration strategy, when drone A in region L... l Obtain the model update results for the user's vehicle and transmit them to UAV A. o Afterwards, drone A o Complete the global model update and release the task for the next global model training.

6. A method for drone-assisted federated learning according to claim 5, characterized in that, In step S3, for high-performance user vehicles, it is determined whether their following overhead is greater than their migration overhead. If so, a service migration strategy is executed; otherwise, a drone following strategy is implemented, including the following steps: S3.1, Based on user vehicle V i Time parameter θ i When θ i When it is less than 0, it means that the user vehicle V i The departure time is less than the time of one iteration, and the drone A o For V i Implement a tracking and movement strategy, and calculate the cost of the drone's movement. S3.2 When executing the tracking movement strategy, the UAV needs to determine the following speed and the distance the UAV has moved. Then, compute the user set S = [V1, ... V1] for service migration. i ,...V N Total cost: The drone following overhead involved in service migration is: Among them, C i，O For user V i Overhead within range O S3.3 In order to reduce communication power consumption, when In drone A o Within a certain range, the mobile drone follows the user's vehicle; when C vi <C f In drone A o A L Within the specified range, service migration can be selected.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images