Heterogeneous Sensing Federated Learning Method via UAV Sampling and D2D Communication

Abstract

This invention addresses all the issues arising from drone heterogeneity, limited communication resources, and significant overlap in local drone data distribution. It provides a heterogeneous-aware federated learning method using drone sampling and D2D communication. This method utilizes a submodule maximization algorithm and a greedy algorithm to jointly optimize drone sampling and end-to-end communication, effectively utilizing computing resources while protecting drone data privacy. Furthermore, the invention proposes a fast-converging drone sampling and end-to-end resource allocation method that improves the efficiency of federated learning training while ensuring drone participation, thereby enhancing global model performance. The proposed method converges quickly and efficiently when updating all models. It specifically addresses the low utilization of drone sampling and computing resources in dynamic heterogeneous environments. It can converge quickly and meet latency requirements in heterogeneous data environments, while simultaneously improving the algorithm's convergence speed and model accuracy.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and specifically relates to a heterogeneous perception federated learning method through drone sampling and D2D (Device-to-Device) communication. Background Art

[0002] Traditional centralized training methods for AI models require aggregating all raw data onto a central server, which is impractical for wireless networks due to bandwidth consumption and privacy concerns. Federated learning offers a solution that balances data privacy and model performance by supporting iterative model training across different participants, coordinated by a central server, and has proven to be more effective in wireless network environments.

[0003] Federated learning has been applied in many applications, such as optimizing resource allocation in vehicle-to-vehicle communications and content recommendations on smartphones. However, existing federated learning methods cannot effectively address the challenges caused by client heterogeneity, limited communication resources, and significant overlap in local data distribution among drones. This leads to low computational resource utilization and slow convergence. Therefore, an effective method is urgently needed to address these issues. Summary of the Invention

[0004] To address the above-mentioned problems in the prior art, the present invention provides a heterogeneous perception federated learning method using drone sampling and D2D communication. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0005] A heterogeneous perception federated learning method using drone sampling and D2D communication is applied to an edge federated learning system consisting of an edge server and multiple drones. The heterogeneous perception federated learning method using drone sampling and D2D communication includes:

[0006] S100, in the current training round, the edge server transmits the global model updated in the previous training round to all drones;

[0007] S200, in the current training round, the edge server selects sampling drones from the drones that meet the delay constraint, and divides them into first sampling drones for direct communication and second sampling drones for indirect communication based on the total delay;

[0008] S300: In the current training round, both the first sampling drone and the second sampling drone locally train the global model transmitted by the edge server using a local dataset; the second sampling drone transmits the trained model parameters to the matching first sampling drone, and the first sampling drone aggregates its own model parameters with the model parameters sent by the second sampling drone and then uploads them to the edge server;

[0009] S400, in the current training round, the edge server performs weighted average aggregation on the model parameters uploaded by each drone to obtain aggregated model parameters, and uses the aggregated model parameters to update the global model of the previous training round to obtain the global model of the current training round;

[0010] S500 , repeating the process from S100 to S400 until the global model converges.

[0011] Beneficial effects:

[0012] This invention addresses all the issues arising from drone heterogeneity, limited communication resources, and significant overlap in local drone data distribution. It provides a heterogeneous-aware federated learning method using drone sampling and D2D communication. This method utilizes a submodule maximization algorithm and a greedy algorithm to jointly optimize drone sampling and end-to-end communication, effectively utilizing computing resources while protecting drone data privacy. Furthermore, the invention proposes a fast-converging drone sampling and end-to-end resource allocation method that improves the efficiency of federated learning training while ensuring drone participation, thereby enhancing global model performance. The proposed method converges quickly and efficiently when updating all models. It specifically addresses the low utilization of drone sampling and computing resources in dynamic heterogeneous environments. It can converge quickly and meet latency requirements in heterogeneous data environments, while simultaneously improving the algorithm's convergence speed and model accuracy.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of the heterogeneous perception federated learning method through drone sampling and D2D communication provided by the present invention. DETAILED DESCRIPTION

[0015] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0016] The present invention provides a heterogeneous perception federated learning method through drone sampling and D2D communication, which is applied to an edge federated learning system consisting of an edge server and multiple drones. The multiple drones in the edge federated learning system constitute a drone set, which is represented by , each drone has a size of Local dataset ;in, is the total number of drones, The serial number of the drone.

[0017] The goal of this invention is to train a global model that performs well on data from all drones. , each drone uses a size of Local dataset and use gradient descent to train this global model, but in federated learning, each drone computes its local gradients, which are then uploaded to the edge server (or securely aggregated), which then updates the global model.

[0018] like Figure 1 As shown, the present invention provides a heterogeneous perception federated learning method through drone sampling and D2D communication, including:

[0019] S100, in the current training round, the edge server transmits the global model updated in the previous training round to all drones;

[0020] S200, in the current training round, the edge server selects sampling drones from the drones that meet the delay constraint, and divides them into first sampling drones for direct communication and second sampling drones for indirect communication based on the total delay;

[0021] It is worth noting that the allocation to drones The more bandwidth the drone has, the The smaller the delay, the lower the latency. Since the latency of each round of federated learning is determined by the slowest drone, the present invention can reallocate bandwidth from faster drones to slower ones to reduce the latency of federated learning. Therefore, the optimal solution can only be achieved when all bandwidth is allocated and all selected drones have the same completion time. A binary search method can be used to find the optimal latency and bandwidth allocation. This involves determining the sampling drones and dividing them into indirect and direct drones before performing the federated learning process to update the global model.

[0022] The present invention can sample and obtain a set of sampling drones, which then need to be divided into two categories: one is a set of drones that communicate directly with the edge server; , one type is a collection of drones that communicate indirectly with edge servers The purpose of dividing the sampling drones into two categories is to reduce the overall system latency and mitigate the impact of stragglers. The specific implementation details are described below.

[0023] In one embodiment of the present invention, S200 includes:

[0024] S210, in the current training round, the edge server determines a drone that satisfies the delay constraint condition among all drones;

[0025] S220, in the current training round, the edge server selects a sampling drone from the drones that meet the delay constraint condition;

[0026] In this step, in the current training round, the edge server uses the submodule maximization algorithm to select a predetermined number of drones from the drones that meet the delay constraint conditions to form a sampling drone set. ; Among them, the drones that meet the delay constraint conditions are drones within the predetermined delay range.

[0027] After S220, the heterogeneous perception federated learning method through drone sampling and D2D communication further includes:

[0028] In the current training round, the edge server targets the sampled drone set For each sampling drone in The distance between other sampling drones in the sampling process, if the distance is not greater than the maximum coverage range of the sampling drone , then add the number of the other sampling drone to the end-to-end communication drone set corresponding to the sampling drone middle.

[0029] Specifically, for sampling drones Perform the following steps:

[0030] Computational sampling drones The distance between other drones, if the distance between them is less than or equal to the maximum coverage range of the drone , then add each other to their respective end-to-end communication drone sets, i.e. , Indicates sampling drone A collection of end-to-end communication drones, Indicates the numbers of other sampling drones; represents the set of end-to-end communicating drones of other sampling drones, It's a sampling drone. Number.

[0031] S230 , in the current training round, the edge server divides the sampling drones into first sampling drones for direct communication and second sampling drones for indirect communication using the total delay.

[0032] In an optional embodiment of the present invention, S230 includes:

[0033] S231, in the current training round, the edge server initializes the set of drones for direct communication and indirectly communicate with drone collections is empty;

[0034] S232, in the current training round, the edge server uses a greedy algorithm to select a sampling drone with the minimum total delay and add it to the set of drones in direct communication In and from the sampling drone collection Remove the sampling drone number;

[0035] S233, in the current training round, the edge server for any sampling drone , if the sampling drone If the membership conditions and delay conditions are met, the sampling drone will be Added the number to the collection of drones for indirect communication In and from the sampling drone collection Remove the sampling drone The number of; wherein, the membership condition is the sampling drone The number of end-to-end communication drones in other sampling drones The delay condition is the sampling drone Local training latency and sampling drones To the sampling drone The sum of the transmission delays of the sampling drone is no greater than The total delay.

[0036] It is worth noting that for sampling drones If the sampling drone and If so, add a sampling drone arrive and from Remove sampling drone , Indicates sampling drone A collection of end-to-end communication drones, Indicates that in the current training round , sampling drone To the sampling drone The transmission delay, It's a sampling drone. Local training latency, In the current training round , sampling drone The total delay.

[0037] The total delay of federated learning mainly includes the transmission delay of the model parameters transmitted by the drone and the local training delay. All drones share the bandwidth resources of the edge server, so the sampling drone The total delay is expressed as:

[0038] ;

[0039] ;

[0040] ;

[0041] Where, For the current training round , sampling drone The total delay of In the current training round , sampling drone transmission delay; Sampling drones Local training latency; is the size of the model parameters; For the current training round , sampling drone The transmission rate; For the current training round , sampling drone the bandwidth allocated by the edge server; , is the total uplink bandwidth, For the current training round , sampling drone The transmission power; For the current training round , sampling drone Channel gain to the edge server; is the power spectral density of Gaussian noise.

[0042] S234, the edge server repeats S232 to S233 until the sampling drone set Is an empty set, get the set of drones that communicate directly and indirectly communicate with drone collections ; wherein the drone set Contains at least one first sampling drone, the drone set Includes at least one second sampling drone.

[0043] S300: In the current training round, both the first sampling drone and the second sampling drone locally train the global model transmitted by the edge server using a local dataset; the second sampling drone transmits the trained model parameters to the matching first sampling drone, and the first sampling drone aggregates its own model parameters with the model parameters sent by the second sampling drone and then uploads them to the edge server;

[0044] The model parameters include the gradient obtained by training the global model using the gradient descent method.

[0045] S400, in the current training round, the edge server performs weighted average aggregation on the model parameters uploaded by each drone to obtain aggregated model parameters, and uses the aggregated model parameters to update the global model of the previous training round to obtain the global model of the current training round;

[0046] It is worth noting that the edge server does not participate in the training of the global model. Its main function is to perform weighted average aggregation of the model parameters trained by each drone to obtain the aggregated parameters of the global model for the current training round, and provide an updated global model for the next training round of distributed (drone) training. The following details the aggregation and update process:

[0047] After receiving updates from multiple drones, the edge server calculates a weighted average based on the amount of data from each drone. The weighted formula is , It is The amount of data from drones, is the total number of drones participating in the update, yes One of the drones with updated parameters;

[0048] The server uses the aggregated weights to update the global model parameters. , the model parameters of the updated global model It can be calculated by the following formula:

[0049] ;

[0050] in It is The updated model parameters of the UAV itself.

[0051] For example, assume that there are three drones, A, B, and C, with 100, 200, and 300 data records, respectively. After a round of global model training, each drone uploads its own updated model parameters. The model parameters of drone A are , the data volume is 100. The model parameters of UAV B are , the data volume is 200. The model parameters of UAV C are , the data volume is 300. First, the model parameters of each drone are calculated, expressed as:

[0052] ;

[0053] ;

[0054] ;

[0055] The edge server then performs weighted aggregation on these model parameters to generate aggregated parameters, which are used to generate the model parameters of the new global model. :

[0056] ;

[0057] In this way, the edge server obtains aggregated model parameters that combine the contributions of all participating drones and more accurately reflect the overall data distribution. Knowing these aggregated model parameters, they can be used to update the global model from the previous training round, obtaining the global model for the current training round, i.e., the updated global model.

[0058] S500 , repeating the process from S100 to S400 until the global model converges.

[0059] Because each round of federated learning is synchronous, the latency depends on the slowest drone, so

[0060] In the current training round The total delay used to update all models is expressed as ,in, In the current training round , sampling drone transmission delay; The value of is 1; Sampling drones Local training latency; In the current training round , sampling drone The total delay.

[0061] This invention utilizes a submodular maximization algorithm and a greedy algorithm to jointly optimize drone sampling and end-to-end communication. This method addresses the inability of existing federated learning methods to effectively address all of the issues arising from drone heterogeneity, limited communication resources, and significant overlap in local drone data distribution, leading to low computational resource utilization and slow convergence. The technical solution of this invention addresses these issues. Through adaptive drone sampling and end-to-end transmission, it effectively utilizes computational resources while protecting drone data privacy. It also proposes a fast-converging drone sampling and end-to-end allocation method, which improves the efficiency of federated learning training while ensuring drone participation, thereby enhancing model performance. This novel method offers fast convergence and high efficiency when updating the global model. It particularly addresses the low utilization of drone sampling and computational resources in dynamic, heterogeneous environments. It can quickly converge and meet latency requirements in heterogeneous data environments, while simultaneously improving algorithm convergence and model accuracy.

[0062] It is worth noting that the terms "first" and "second" in this disclosure are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of this disclosure, "plurality" means two or more, unless otherwise specifically defined.

[0063] Although the present application is described herein with reference to various embodiments, those skilled in the art will be able to understand and implement other variations of the disclosed embodiments in practicing the claimed application by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0064] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A heterogeneous perception federated learning method based on drone sampling and D2D communication, characterized by: Applied to an edge federated learning system consisting of an edge server and multiple drones, the heterogeneous perception federated learning method through drone sampling and D2D communication includes: S100, in the current training round, the edge server transmits the global model updated in the previous training round to all drones; S200, in the current training round, the edge server selects sampling drones from the drones that meet the delay constraint, and divides them into first sampling drones for direct communication and second sampling drones for indirect communication based on the total delay; S300: In the current training round, both the first sampling drone and the second sampling drone locally train the global model transmitted by the edge server using a local dataset; the second sampling drone transmits the trained model parameters to the matching first sampling drone, and the first sampling drone aggregates its own model parameters with the model parameters sent by the second sampling drone and then uploads them to the edge server; S400, in the current training round, the edge server performs weighted average aggregation on the model parameters uploaded by each drone to obtain aggregated model parameters, and uses the aggregated model parameters to update the global model of the previous training round to obtain the global model of the current training round; S500, repeating the process from S100 to S400 until the global model converges; S200 includes: S210, in the current training round, the edge server determines a drone that satisfies the delay constraint condition among all drones; S220, in the current training round, the edge server selects a sampling drone from the drones that meet the delay constraint condition; S230, in the current training round, the edge server divides the sampling drones into a first sampling drone for direct communication and a second sampling drone for indirect communication using the total delay; The S230 includes: S231, in the current training round, the edge server initializes the set of drones for direct communication and indirectly communicate with drone collections is empty; S232, in the current training round, the edge server uses a greedy algorithm to select a sampling drone with the minimum total delay and add it to the set of drones in direct communication In and from the sampling drone collection Remove the sampling drone number; S233, in the current training round, the edge server for any sampling drone , if the sampling drone If the membership conditions and delay conditions are met, the sampling drone will be Added the number to the collection of drones for indirect communication In and from the sampling drone collection Remove the sampling drone Number; S234, the edge server repeats S232 to S233 until the sampling drone set Is an empty set, get the set of drones that communicate directly and indirectly communicate with drone collections ; wherein the drone set Contains at least one first sampling drone, the drone set Includes at least one second sampling drone.

2. The heterogeneous perception federated learning method through drone sampling and D2D communication according to claim 1 is characterized in that Multiple drones in the edge federated learning system form a drone set, which is represented by , each drone has a size of Local dataset ;in, is the total number of drones, is the serial number of the drone; all drones share the bandwidth resources of the edge server.

3. The heterogeneous perception federated learning method through drone sampling and D2D communication according to claim 2 is characterized in that The S220 includes: In the current training round, the edge server uses the submodule maximization algorithm to select a predetermined number of drones from the drones that meet the delay constraint conditions to form a sampling drone set. ; Among them, the drones that meet the delay constraint conditions are drones within the predetermined delay range.

4. The heterogeneous perception federated learning method through drone sampling and D2D communication according to claim 3 is characterized in that After S220, the heterogeneous perception federated learning method through drone sampling and D2D communication further includes: In the current training round, the edge server targets the sampled drone set For each sampling drone in The distance between other sampling drones in the sampling process, if the distance is not greater than the maximum coverage range of the sampling drone , then add the number of the other sampling drone to the end-to-end communication drone set corresponding to the sampling drone middle.

5. The heterogeneous perception federated learning method through drone sampling and D2D communication according to claim 4 is characterized in that The membership condition is the sampling drone The number of end-to-end communication drones in other sampling drones The delay condition is the sampling drone Local training latency and sampling drones To the sampling drone The sum of the transmission delays of the sampling drone is no greater than The total delay.

6. The heterogeneous perception federated learning method through drone sampling and D2D communication according to claim 5 is characterized in that The sampling drone The total delay is expressed as: ； ； ； Where, For the current training round , sampling drone The total delay of In the current training round , sampling drone transmission delay; Sampling drones Local training latency; is the size of the model parameters; For the current training round , sampling drone The transmission rate; For the current training round , sampling drone the bandwidth allocated by the edge server; , is the total uplink bandwidth, For the current training round , sampling drone The transmission power; For the current training round , sampling drone Channel gain to the edge server; is the power spectral density of Gaussian noise.

7. The heterogeneous perception federated learning method through drone sampling and D2D communication according to claim 1 is characterized in that The total delay used to update all models in the current training round is expressed as ,in, In the current training round , sampling drone transmission delay; The value of is 1; Sampling drones Local training latency; In the current training round , sampling drone The total delay.

8. The heterogeneous sensing federated learning method through drone sampling and D2D communication according to claim 1 is characterized in that: The model parameters include gradients obtained by training the global model using a gradient descent method.

Images