Data transmission method and device in federated learning model training

By identifying a set of candidate user terminals in a satellite communication system, comprehensively considering the amount of training data and channel gain, and optimizing the data transmission strategy, the communication quality and efficiency issues in federated learning model training are resolved, and efficient model training is achieved.

CN123123472APending Publication Date: 2026-11-17BEIJING SMARTCHIP SEMICON TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611025740.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In wireless communication scenarios, federated learning model training is limited by strict bandwidth, and existing technologies struggle to improve model training efficiency while simultaneously enhancing communication quality and reducing latency.

Method used

By determining the candidate user terminal set, comprehensively considering the amount of training data, model gradient differences, and channel gain, the total training revenue and communication cost are calculated, and the target user terminal set is iteratively updated to optimize the data transmission strategy.

Benefits of technology

Under the same communication resource consumption, improve the convergence speed and training accuracy of the model, reduce the impact of non-uniform distribution of training data, and improve the overall communication efficiency of the satellite communication system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123472A_ABST
    Figure CN123123472A_ABST
Patent Text Reader

Abstract

This application discloses a data transmission method and apparatus for federated learning model training, belonging to the field of machine learning technology. The method includes: determining a candidate user terminal set from various user terminals; calculating the total training revenue value corresponding to the candidate user terminal set based on the amount of training data of each candidate user terminal and the gradient magnitude value of the local model loss function; calculating the communication cost value of each candidate user terminal based on the channel gain value between each candidate user terminal and the satellite base station and the number of model parameters, and calculating the total communication cost value corresponding to the candidate user terminal set; iteratively updating the candidate user terminal set with the objective of maximizing the ratio of the total training revenue value to the total communication cost value; determining the candidate user terminal set at the convergence point as the target user terminal set; and having each target user terminal in the target user terminal set transmit the parameters of its local model to the satellite base station. This application can improve the training efficiency of the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of machine learning technology, and in particular relates to a data transmission method and apparatus for training a federated learning model. Background Technology

[0002] Federated Learning (FL) is a privacy-preserving computation and machine learning framework for distributed scenarios. It allows each user terminal to independently train a local model using its local dataset, and only uploads the trained model parameters or gradient information to the cloud server. The cloud server then performs aggregation operations on the received model parameters or gradients to generate a new generation of global model parameters and distributes them to each user terminal to update the local model. This enables multiple data holders to collaboratively complete model training without sharing their local raw data, thereby significantly reducing the bandwidth overhead caused by the transmission of raw data while ensuring data privacy and security.

[0003] However, in wireless communication scenarios, uplink communication is typically subject to strict bandwidth limitations. Given limited communication resources, it's necessary to select a subset of user terminals from all available terminals, instructing only these selected terminals to upload model parameters or gradient information to the cloud server. To improve communication quality, reduce latency, and simultaneously increase the efficiency of global model training, a reasonable data transmission strategy needs to be designed. Summary of the Invention

[0004] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a data transmission method and apparatus for federated learning model training to improve the training efficiency of the model.

[0005] In a first aspect, this application provides a data transmission method for training a federated learning model, applied to a satellite communication system, wherein the satellite communication system includes a satellite base station and multiple user terminals, each of which locally deploys a local model; the method includes: A candidate user terminal set is determined from each of the user terminals, the candidate user terminal set including at least one candidate user terminal; Based on the amount of training data used to train the local model for each of the candidate user terminals and the gradient magnitude of the loss function of the local model, the total training gain value corresponding to the candidate user terminal set is calculated. Based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters of the local model of each candidate user terminal, the estimated communication cost value for each candidate user terminal to upload the parameters of the local model to the satellite base station is calculated, and based on the communication cost value corresponding to each candidate user terminal, the total communication cost value corresponding to the set of candidate user terminals is calculated. With the objective of maximizing the ratio of the total training revenue to the total communication cost, the candidate user terminal set is iteratively updated. The candidate user terminal set that converges is determined as the target user terminal set, and each target user terminal in the target user terminal set transmits the parameters of the local model to the satellite base station.

[0006] The data transmission method for federated learning model training provided in this application comprehensively considers the differences in local training data volume and model gradient of each user terminal during the federated learning process, and calculates the total training benefit value corresponding to the candidate user terminal set. This can reduce the impact of non-uniform distribution of training data volume on the global model update efficiency. By aiming to maximize the ratio of total training benefit value to total communication cost value, the target user terminal set participating in data transmission and global model training is iteratively updated. This method can deeply integrate the service volume weighting characteristics of global aggregation in federated learning with satellite communication cost constraints, and simultaneously improve the convergence speed and training accuracy of the model under the same communication resource consumption, thereby improving the training efficiency of the model.

[0007] According to one embodiment of this application, calculating the total training gain value corresponding to the candidate user terminal set based on the amount of training data used to train the local model for each of the candidate user terminals and the gradient magnitude value of the local model's loss function includes: Calculate the ratio of the amount of training data used by each candidate user terminal to train the local model to the total amount of training data of each user terminal, and obtain the benefit evaluation weight corresponding to each candidate user terminal. Based on the revenue evaluation weights corresponding to each candidate user terminal, a weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal is calculated to obtain the total training revenue value corresponding to the candidate user terminal set.

[0008] In this embodiment, by using the ratio of the amount of training data used by each candidate user terminal to train the local model to the sum of the training data of each user terminal as the benefit evaluation weight corresponding to each candidate user terminal, the amount of computation required to calculate the total training benefit value can be reduced and the calculation speed can be improved.

[0009] According to one embodiment of this application, based on the formula:

[0010] The total training revenue value corresponding to the candidate user terminal set is calculated; Where U represents the set of candidate user terminals, and K represents the total amount of training data for each of the user terminals. This represents the amount of training data corresponding to candidate user terminal i. This represents the revenue evaluation weight corresponding to candidate user terminal i. This represents the gradient magnitude value corresponding to candidate user terminal i.

[0011] According to one embodiment of this application, based on the formula:

[0012]

[0013]

[0014]

[0015] The channel gain value between the candidate user terminal and the satellite base station is calculated. Where h represents the channel gain value between the candidate user terminal and the satellite base station. This indicates large-scale decay. d is a scaling factor, and d is the distance between the candidate user terminal and the satellite base station. This is the path loss index. This indicates small-scale fading. Rice factor, This represents the random phase of the line-of-sight component. This indicates Rayleigh's decline. and They are independent Gaussian variables with zero mean.

[0016] According to one embodiment of this application, the step of calculating the estimated communication cost for each candidate user terminal to upload the parameters of its local model to the satellite base station based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters of the local model of each candidate user terminal, includes: Obtain the initial transmission power and initial channel allocation vector for each of the candidate user terminals; Based on the initial transmission power of each candidate user terminal, the initial channel allocation vector, the channel gain value, and the number of model parameters, the estimated communication delay for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated. Based on the initial transmission power and communication delay of each candidate user terminal, calculate the estimated communication energy consumption of each candidate user terminal when uploading the parameters of the local model to the satellite base station; Based on the estimated communication latency and energy consumption of each candidate user terminal uploading the parameters of the local model to the satellite base station, the communication cost value corresponding to each candidate user terminal is determined.

[0017] In this embodiment, by calculating the communication delay and communication energy consumption based on the real-time channel gain value of each candidate user terminal, the user selection and transmission parameters can be adjusted in response to real-time fluctuations in the channel, reducing packet loss and retransmission, thereby improving the efficiency of data transmission.

[0018] According to one embodiment of this application, calculating the total communication cost value corresponding to the set of candidate user terminals based on the communication cost value corresponding to each of the candidate user terminals includes: The maximum value of the communication delay corresponding to each of the candidate user terminals is determined as the communication delay cost value corresponding to the set of candidate user terminals; The sum of the communication energy consumption corresponding to each of the candidate user terminals is determined as the communication energy consumption cost value corresponding to the candidate user terminal set. The total communication cost value corresponding to the candidate user terminal set is obtained by calculating the weighted sum of the communication latency cost value and the communication energy consumption cost value.

[0019] In this embodiment, by using the weighted sum of communication latency cost and communication energy cost as the total communication cost value corresponding to the candidate user terminal set, and then taking the maximization of the ratio of total training revenue value to total communication cost value as the optimization objective, energy consumption and latency during data transmission can be reduced, thereby improving the overall communication efficiency of the satellite communication system.

[0020] According to one embodiment of this application, based on the formula:

[0021]

[0022]

[0023] The estimated communication delay for the candidate user terminal to upload the parameters of the local model to the satellite base station is calculated. in, This represents the estimated communication latency for candidate user terminal i to upload the parameters of the local model to the satellite base station, and S represents the size of the data packet corresponding to the parameters of the local model. Let R be the initial channel allocation vector for candidate user terminal i, where R is the total number of channels. This indicates that candidate user terminal i is assigned to channel n. Indicates according to After channel allocation, the information transmission rate from candidate user terminal i to the satellite base station, where B is the channel bandwidth. Let i be the initial transmission power of the candidate user terminal. Let n be the channel gain value of candidate user terminal i on channel n. This represents the noise power spectral density.

[0024] According to one embodiment of this application, the step of iteratively updating the candidate user terminal set with the objective of maximizing the ratio of the total training revenue to the total communication cost, determining the candidate user terminal set at the convergence point as the target user terminal set, and instructing each target user terminal in the target user terminal set to transmit the parameters of the local model to the satellite base station includes: Initialize the candidate user terminal set for the current iteration round; The candidate user terminal set is updated with the objective function being maximized; the objective function represents the ratio of the total training revenue to the total communication cost. After the update termination condition is met, the target transmit power of each candidate user terminal is calculated with the goal of minimizing the expected communication cost value of each candidate user terminal in the candidate user terminal set. Based on the target transmit power of each candidate user terminal, a target user-channel matching pair set that minimizes the total communication cost value corresponding to the candidate user terminal set is determined; the target user-channel matching pair set includes matching pairs of each candidate user terminal and channel. Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set, calculate the value of the objective function corresponding to the candidate user terminal set in the current iteration round; If the difference between the value of the objective function corresponding to the current iteration round and the value of the objective function corresponding to the previous iteration round is less than the convergence threshold, then the candidate user terminal set of the current iteration round is determined as the target user terminal set. The target channel corresponding to each target user terminal is determined according to the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power.

[0025] In this embodiment, by iteratively optimizing the user terminal set, power configuration, and channel allocation scheme, the computational complexity can be reduced and the solution speed can be improved. After updating the candidate user terminal set, the target transmit power that minimizes the communication cost of each candidate user terminal is found, and the channel allocation scheme that minimizes the total communication cost of the system is solved. This enables precise matching between transmit power and channel status and mission requirements, thereby improving the overall communication efficiency of the satellite communication system while increasing the model training speed.

[0026] According to one embodiment of this application, the iteration termination condition is that the number of candidate user terminals in the candidate user terminal set is greater than or equal to a target number; updating the candidate user terminal set with the objective of maximizing the ratio of the total training revenue to the total communication cost includes: Based on each first user terminal and the candidate user terminal set, a temporary user terminal set corresponding to each first user terminal is constructed; the first user terminal is a user terminal among the user terminals that has not been added to the candidate user terminal set. Calculate the value of the auxiliary function corresponding to each of the aforementioned temporary user terminal sets; wherein the auxiliary function is constructed based on the objective function; The first user terminal corresponding to the temporary user terminal set with the largest value of the auxiliary function is added to the candidate user terminal set.

[0027] In this embodiment, by constructing an auxiliary function based on the objective function and adopting a greedy strategy, the user terminal that maximizes the increment of the auxiliary function is selected to be added to the candidate user set in each round of updates. This avoids solving complex fractional optimization problems, reduces computational complexity, and thus improves the computational speed of updating the candidate user terminal set.

[0028] According to one embodiment of this application, the step of determining the target transmit power of each candidate user terminal with the objective of minimizing the expected communication cost of the candidate user terminals includes: Multiple candidate power values ​​are determined from the target power range; Calculate the estimated communication cost for the candidate user terminal to transmit signals using each candidate power value; The candidate power value with the lowest communication cost is determined as the target transmit power of the candidate user terminal.

[0029] In this embodiment, by using a linear search method to solve for the target transmit power of each candidate user terminal, the computational complexity can be reduced and the speed of solving for the target transmit power can be increased.

[0030] According to one embodiment of this application, the step of determining the target user-channel matching pair set that minimizes the total communication cost value corresponding to the candidate user terminal set based on the target transmit power of each of the candidate user terminals includes: Based on the target transmit power of each candidate user terminal, the expected communication cost value for each candidate user terminal to select each candidate channel is calculated, and a communication cost matrix is ​​obtained. Each candidate user terminal is used as a node in the first subgraph, each candidate channel is used as a node in the second subgraph, the matching pairs of each candidate user terminal and each candidate channel are used as edges between the nodes of the first subgraph and the nodes of the second subgraph, and the communication cost matrix is ​​used as the weight matrix corresponding to each edge to establish a user-channel bipartite graph. The target user-channel matching pair set is obtained by finding the user-channel matching pairs that minimize the total weight value of each edge in the user-channel bipartite graph.

[0031] In this embodiment, by modeling the channel allocation problem as a bipartite graph minimum weight matching problem and then solving it, the computational complexity can be reduced while ensuring that the solution meets the channel allocation constraints, thereby improving the efficiency of solving the target user-channel matching pair set.

[0032] According to one embodiment of this application, the objective function is:

[0033] The auxiliary function is:

[0034] Where U represents the set of candidate user terminals, and K represents the total amount of training data for each of the user terminals. This represents the amount of training data corresponding to candidate user terminal i. Let represent the gradient magnitude value corresponding to candidate user terminal i, E represent the communication energy consumption cost value corresponding to the candidate user terminal set, and T represent the communication latency cost value corresponding to the candidate user terminal set. and It is a proportionality coefficient, and , This represents the value of the objective function corresponding to the previous iteration round.

[0035] Secondly, this application provides a data transmission device for training a federated learning model, comprising: A determining module is used to determine a candidate user terminal set from each user terminal, the candidate user terminal set including at least one candidate user terminal; The first calculation module is used to calculate the total training revenue value corresponding to the set of candidate user terminals based on the amount of training data used to train the local model for each candidate user terminal and the gradient magnitude value of the loss function of the local model. The second calculation module is used to calculate the estimated communication cost of each candidate user terminal uploading the parameters of its local model to the satellite base station based on the channel gain value between each candidate user terminal and the satellite base station and the number of model parameters of the local model of each candidate user terminal, and to calculate the total communication cost value corresponding to the set of candidate user terminals based on the communication cost value corresponding to each candidate user terminal. The update module is used to iteratively update the candidate user terminal set with the objective of maximizing the ratio of the total training revenue value to the total communication cost value, determine the candidate user terminal set when the iteration reaches convergence as the target user terminal set, and instruct each target user terminal in the target user terminal set to transmit the parameters of the local model to the satellite base station.

[0036] According to the data transmission device in the federated learning model training of this application, by comprehensively considering the differences in the amount of local training data and model gradients of each user terminal during the federated learning process, the total training revenue value corresponding to the candidate user terminal set is calculated. This can reduce the impact of the non-uniform distribution of training data on the global model update efficiency. By taking the maximization of the ratio of the total training revenue value to the total communication cost value as the objective, the target user terminal set participating in data transmission and global model training is iteratively updated. This can deeply integrate the service volume weighting characteristics of global aggregation in federated learning with the satellite communication cost constraints, and simultaneously improve the convergence speed and training accuracy of the model under the same communication resource consumption, thereby improving the training efficiency of the model.

[0037] According to one embodiment of this application, the first calculation module is further configured to: Calculate the ratio of the amount of training data used by each candidate user terminal to train the local model to the total amount of training data of each user terminal, and obtain the benefit evaluation weight corresponding to each candidate user terminal. Based on the revenue evaluation weights corresponding to each candidate user terminal, a weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal is calculated to obtain the total training revenue value corresponding to the candidate user terminal set.

[0038] According to one embodiment of this application, the second calculation module is further configured to: Obtain the initial transmission power and initial channel allocation vector for each of the candidate user terminals; Based on the initial transmission power of each candidate user terminal, the initial channel allocation vector, the channel gain value, and the number of model parameters, the estimated communication delay for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated. Based on the initial transmission power and communication delay of each candidate user terminal, calculate the estimated communication energy consumption of each candidate user terminal when uploading the parameters of the local model to the satellite base station; Based on the estimated communication latency and energy consumption of each candidate user terminal uploading the parameters of the local model to the satellite base station, the communication cost value corresponding to each candidate user terminal is determined.

[0039] According to one embodiment of this application, the second calculation module is further configured to: The maximum value of the communication delay corresponding to each of the candidate user terminals is determined as the communication delay cost value corresponding to the set of candidate user terminals; The sum of the communication energy consumption corresponding to each of the candidate user terminals is determined as the communication energy consumption cost value corresponding to the candidate user terminal set. The total communication cost value corresponding to the candidate user terminal set is obtained by calculating the weighted sum of the communication latency cost value and the communication energy consumption cost value.

[0040] According to one embodiment of this application, the updating module is further configured to: Initialize the candidate user terminal set for the current iteration round; The candidate user terminal set is updated with the objective of maximizing the ratio of the total training revenue to the total communication cost. After the update termination condition is met, the target transmit power of each candidate user terminal is calculated with the goal of minimizing the expected communication cost value of each candidate user terminal in the candidate user terminal set. Based on the target transmit power of each candidate user terminal, a target user-channel matching pair set that minimizes the total communication cost value corresponding to the candidate user terminal set is determined; the target user-channel matching pair set includes matching pairs of each candidate user terminal and channel. Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set, the value of the objective function corresponding to the candidate user terminal set in the current iteration round is calculated; the objective function represents the ratio of the total training revenue to the total communication cost. If the difference between the value of the objective function corresponding to the current iteration round and the value of the objective function corresponding to the previous iteration round is less than the convergence threshold, then the candidate user terminal set of the current iteration round is determined as the target user terminal set. The target channel corresponding to each target user terminal is determined according to the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power.

[0041] According to one embodiment of this application, the updating module is further configured to: Based on each first user terminal and the candidate user terminal set, a temporary user terminal set corresponding to each first user terminal is constructed; the first user terminal is a user terminal among the user terminals that has not been added to the candidate user terminal set. Calculate the value of the auxiliary function corresponding to each of the aforementioned temporary user terminal sets; wherein the auxiliary function is constructed based on the objective function; The first user terminal corresponding to the temporary user terminal set with the largest value of the auxiliary function is added to the candidate user terminal set.

[0042] According to one embodiment of this application, the updating module is further configured to: Multiple candidate power values ​​are determined from the target power range; Calculate the estimated communication cost for the candidate user terminal to transmit signals using each candidate power value; The candidate power value with the lowest communication cost is determined as the target transmit power of the candidate user terminal.

[0043] According to one embodiment of this application, the updating module is further configured to: Based on the target transmit power of each candidate user terminal, the expected communication cost value for each candidate user terminal to select each candidate channel is calculated, and a communication cost matrix is ​​obtained. Each candidate user terminal is used as a node in the first subgraph, each candidate channel is used as a node in the second subgraph, the matching pairs of each candidate user terminal and each candidate channel are used as edges between the nodes of the first subgraph and the nodes of the second subgraph, and the communication cost matrix is ​​used as the weight matrix corresponding to each edge to establish a user-channel bipartite graph. The Hungarian algorithm is used to find user-channel matching pairs that minimize the total weight of each edge in the user-channel bipartite graph, thus obtaining the target user-channel matching pair set.

[0044] Thirdly, this application 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 to implement the data transmission method in federated learning model training as described in the first aspect above.

[0045] Fourthly, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data transmission method in federated learning model training as described in the first aspect above.

[0046] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the data transmission method in federated learning model training as described in the first aspect above.

[0047] Sixthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the data transmission method in federated learning model training as described in the first aspect above.

[0048] The above-described one or more technical solutions in the embodiments of this application have at least one of the following technical effects: According to the data transmission method in federated learning model training of this application, by comprehensively considering the differences in local training data volume and model gradient of each user terminal during the federated learning process, the total training benefit value corresponding to the candidate user terminal set is calculated. This can reduce the impact of non-uniform distribution of training data volume on the global model update efficiency. By taking the maximum ratio of total training benefit value to total communication cost value as the objective, the target user terminal set participating in data transmission and global model training is iteratively updated. This can deeply integrate the service volume weighting characteristics of global aggregation in federated learning with the satellite communication cost constraint, and simultaneously improve the convergence speed and training accuracy of the model under the same communication resource consumption, thereby improving the training efficiency of the model.

[0049] In some embodiments, by using the ratio of the amount of training data used by each candidate user terminal to train the local model to the sum of the training data of each user terminal as the benefit evaluation weight corresponding to each candidate user terminal, the amount of computation required to calculate the total training benefit value can be reduced and the computation speed can be improved.

[0050] In some embodiments, by calculating communication latency and communication energy consumption based on the real-time channel gain value of each candidate user terminal, it is possible to adjust user selection and transmission parameters in response to real-time channel fluctuations, thereby reducing packet loss and retransmission and improving data transmission efficiency.

[0051] In some embodiments, by using the weighted sum of communication latency cost and communication energy cost as the total communication cost value corresponding to the candidate user terminal set, and then taking the maximization of the ratio of total training revenue to total communication cost value as the optimization objective, energy consumption and latency during data transmission can be reduced, thereby improving the overall communication efficiency of the satellite communication system.

[0052] In some embodiments, by iteratively optimizing the user terminal set, power configuration, and channel allocation scheme, computational complexity can be reduced and solution speed can be improved. After updating the candidate user terminal set, the target transmit power that minimizes the communication cost of each candidate user terminal is found, and the channel allocation scheme that minimizes the total communication cost of the system is solved. This enables precise matching between transmit power and channel status and mission requirements, thereby improving the overall communication efficiency of the satellite communication system while increasing the model training speed.

[0053] In some embodiments, by constructing an auxiliary function based on the objective function and employing a greedy strategy to select user terminals that maximize the increment of the auxiliary function in each round of updates to add to the candidate user set, it is possible to avoid solving complex fractional optimization problems, reduce computational complexity, and thus improve the computational speed of updating the candidate user terminal set.

[0054] In some embodiments, solving for the target transmit power of each candidate user terminal using a linear search method can reduce computational complexity and increase the speed of solving for the target transmit power.

[0055] In some embodiments, by modeling the channel allocation problem as a bipartite graph minimum weight matching problem and then solving it, the solution can satisfy the channel allocation constraints while reducing computational complexity, thereby improving the efficiency of solving the target user-channel matching pair set.

[0056] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of a federated learning process based on a satellite communication system provided in an embodiment of this application; Figure 2 This is a flowchart illustrating the data transmission method in the training of the federated learning model provided in this application embodiment; Figure 3 This is a schematic diagram of the data transmission process during the training of the federated learning model provided in this application embodiment; Figure 4 This is a schematic diagram of the data transmission device in the training of the federated learning model provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0060] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0061] The data transmission method and apparatus for training federated learning models provided in this application will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0062] The data transmission method for training a federated learning model provided in this application embodiment can be executed by an electronic device or a functional module or entity within an electronic device that can implement the data transmission method for training a federated learning model. The electronic devices mentioned in this application embodiment include, but are not limited to, satellite base stations, edge computing gateways, distributed computing nodes, and user terminals. The data transmission method for training a federated learning model provided in this application embodiment is described below using an electronic device as the execution subject.

[0063] The data transmission method for federated learning model training provided in this application is applied to a satellite communication system. For example... Figure 1As shown, a satellite communication system includes satellite base stations and multiple user terminals. Satellite base stations refer to base stations deployed on Earth-orbiting satellite platforms, and can be on-board GNB (Next-Generation NodeB) base stations based on technologies such as 5G-NTN (Non-Terrestrial Network) and 6G-NTN. User terminals refer to terminals deployed on the ground, and can be servers, industrial control computers, edge computing gateways, personal computers, workstations, intelligent inspection equipment, field monitoring hosts, embedded processors of intelligent power equipment, programmable logic controllers, etc. Each user terminal establishes a communication connection with the satellite base station through a wireless network.

[0064] In this embodiment of the application, the satellite base station is used as the master station to undertake the task of training the global model for federated learning. For example... Figure 1 As shown, the satellite base station broadcasts the initial global model parameters to each user terminal. The user terminal initializes its local model based on the received initial global model parameters and independently trains the local model using its local training dataset to update the local model parameters. After completing a certain number of rounds of local training, each user terminal uploads the trained model gradient information and other local model parameters to the satellite base station. The satellite base station aggregates the received local model parameters, calculates the global loss function, and updates the global model parameters using algorithms such as gradient descent. The updated global model parameters are then distributed to all user terminals to guide the subsequent local model update process.

[0065] In wireless communication scenarios, uplink communication is typically subject to strict bandwidth limitations. To improve communication quality, reduce communication latency, and simultaneously increase the efficiency of global model training, a reasonable data transmission strategy needs to be designed.

[0066] like Figure 2 As shown, the data transmission method in the training of this federated learning model includes steps 210, 220, 230, and 240.

[0067] Step 210: Determine a candidate user terminal set from each user terminal. The candidate user terminal set includes at least one candidate user terminal.

[0068] When communication resources are limited, it is necessary to select a subset of user terminals from all user terminals and have these selected terminals upload model parameters or gradient information to the cloud server. In this embodiment, the process begins by selecting from each user terminal... Determine the set of candidate user terminals The candidate user terminal set includes at least one candidate user terminal. The candidate user terminal set can be randomly selected from all user terminals and added to the candidate user terminal set; alternatively, the user terminals that participated in the previous round of data transmission and global model update can be determined as the candidate user terminals for this round, or the candidate user terminal set can be determined through other means. This application embodiment does not limit this approach.

[0069] Step 220: Calculate the total training revenue value corresponding to the candidate user terminal set based on the amount of training data used to train the local model for each candidate user terminal and the gradient magnitude value of the local model loss function.

[0070] User terminals participating in federated learning may differ in storage capacity, computational performance, data acquisition performance, and workload, leading to variations in the amount of training data used to train the local model. The amount of training data affects the generalization performance of the trained local model. For example, models trained on small datasets are more prone to overfitting, and their gradients may contain significant noise. If user terminals participating in global model parameter updates are selected solely based on the gradient magnitude of the local model's loss function, terminals with small training data, those heavily influenced by noise, and those with abnormally high gradient magnitudes may be prioritized. This results in a significant deviation between the aggregated global gradient and the ideal global gradient, impacting the convergence speed and training efficiency of the global model.

[0071] Therefore, in this embodiment of the application, considering the impact of non-uniform distribution of training data on the global model update efficiency, the total training revenue value corresponding to the candidate user terminal set is calculated based on the amount of training data used by each candidate user terminal to train the local model and the gradient magnitude value of the local model loss function.

[0072] During the local training phase of the federated learning process, candidate user terminal i utilizes its local training dataset. The local model is trained independently. The training dataset for candidate user terminal i includes... The number of data samples, the amount of training data corresponding to candidate user terminal i is . Candidate user terminal i for each data sample Calculate the current local model parameters loss function And calculate the gradient vector of the local model loss function with respect to the local model parameters. gradient vector The magnitude is the gradient magnitude of the loss function. .

[0073] In the subsequent global model parameter update phase, the satellite base station can aggregate the gradient vectors uploaded by all selected target user terminals and perform gradient descent to update the global model parameters. For example, the update formula for the global model parameters in round t is:

[0074] in, Let be the global model parameters for round t. These are the global model parameters for round t+1. Let U' be the learning rate, and U' be the target user terminal set. Let represent the gradient vector of the local model loss function of the target user terminal j in round t.

[0075] For federated learning tasks, if we assume that the training sample labels are uniformly distributed across devices, the differences in the gradient direction of the local loss function for each device with respect to the model parameters can be approximated, thus simplifying the analysis of gradient aggregation. According to the global model parameter update formula, in each round of global model parameter updates, the magnitude of the change in global model parameters towards the minimum point of the global model loss function is proportional to the magnitude of the gradient vector of the local model loss function of each target user terminal after weighting by the amount of training data. This means that user terminals with large amounts of training data and large gradient magnitudes of the loss function have a more significant impact on global model updates; therefore, selecting these user terminals can improve the speed of global model training.

[0076] Therefore, the amount of training data used to train the local model based on each candidate user terminal i can be determined. and the gradient magnitude of the local model loss function Calculate the total training reward value corresponding to the candidate user terminal set. For example, it can be... and The ratio as The total training reward of the candidate user terminal set is obtained by linearly weighting the values ​​of the candidate terminals. Of course, other methods can also be used to calculate the total training reward, such as approximating the sum of the marginal rewards of each candidate user terminal using the Shapley value, etc. This application does not limit this method.

[0077] Step 230: Based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters of the local model of each candidate user terminal, calculate the estimated communication cost value for each candidate user terminal to upload the parameters of the local model to the satellite base station, and calculate the total communication cost value corresponding to the set of candidate user terminals based on the communication cost value corresponding to each candidate user terminal.

[0078] Given limited communication resources, it is necessary to consider the estimated communication costs incurred by user terminals in uploading parameters of their local models to satellite base stations. Communication costs represent the communication resources consumed during the communication process, such as time, energy, and channel bandwidth.

[0079] In satellite communication, the space-to-ground link exhibits significant time-varying fading characteristics, with parameters such as channel signal-to-noise ratio and available bandwidth potentially changing dynamically over time and space. Simply assuming a fixed channel state or using only average channel conditions to estimate communication costs and plan data transmission strategies may lead to selected user terminals struggling to efficiently upload gradient data during periods of poor channel quality, potentially resulting in packet loss or retransmissions. This not only increases communication latency but may also affect the accuracy of global model aggregation due to missing gradient information.

[0080] Therefore, in this embodiment, the communication cost value corresponding to each candidate user terminal is calculated based on the real-time channel gain value between each candidate user terminal and the satellite base station. For example, in some embodiments, the distance d between the candidate user terminal and the satellite base station can be calculated based on the real-time positioning data of the candidate user terminal and the satellite, and then calculated according to the formula:

[0081]

[0082]

[0083]

[0084] The channel gain value h between the candidate user terminal and the satellite base station is calculated.

[0085] Where h represents the channel gain between the candidate user terminal and the satellite base station, including large-scale fading. and small-scale fading Two parts.

[0086] Large-scale decay This describes the slow variation in average received signal power due to electromagnetic wavelength propagation distance, environmental obstruction, and atmospheric conditions. The timescale of this variation is on the order of milliseconds to minutes, and it determines the baseline received power level of the communication link. Large-scale fading follows a power-law attenuation. d is the scaling factor; d is the distance between the candidate user terminal and the satellite base station; This is the path loss index, which is typically set to 2 in satellite communication scenarios.

[0087] Small-scale fading This describes the rapid fluctuations in the instantaneous power of the received signal caused by the coherent superposition and cancellation of multipath signals. The time scale of these fluctuations is on the microsecond to millisecond level, determining the instantaneous fluctuations based on the average power. In this embodiment, the Ricean fading model is used to model small-scale fading, wherein... The Rice factor represents the ratio of the line-of-sight (Los) component to the non-line-of-sight (NLos) component in the signal. In satellite communication scenarios, it is related to the elevation angle from the user terminal to the satellite base station. Represents the random phase of the line-of-sight component; This indicates Rayleigh's decline. and They are independent Gaussian variables with zero mean.

[0088] Besides calculations based on distance d and the channel model, in some embodiments, the channel gain value between the candidate user terminal and the satellite base station can also be calculated using other methods. For example, each user terminal can send an SRS (Sounding Reference Signal) to the satellite base station, and the real-time channel gain value can be calculated based on the change in the signal received by the satellite base station relative to the original signal. Of course, other methods can also be used to determine the channel gain value, and this application embodiment does not limit this method.

[0089] After determining the channel gain value, based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters in each candidate user terminal's local model, the estimated communication cost for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated. The number of model parameters in the local model can be the trainable parameters in the local model. The number of data packets to be uploaded can be calculated based on the channel gain, channel bandwidth, and signal transmit power of candidate user terminal i. The uplink data transmission rate from candidate user terminal i to the satellite base station can be calculated, and the volume of the data packets to be uploaded can be estimated based on the number of model parameters and the number of bytes per parameter in the local model of candidate user terminal i. The estimated time for candidate user terminal i to upload the parameters of its local model to the satellite base station can then be calculated, and this calculated time value can be used as the communication cost value corresponding to candidate user terminal i.

[0090] Of course, communication cost can also be evaluated in other ways, such as by combining latency, packet loss rate, and communication energy consumption to calculate communication cost, but this application does not limit this.

[0091] After calculating the communication cost value corresponding to each candidate user terminal, the total communication cost value corresponding to the candidate user terminal set is calculated based on the communication cost value corresponding to each candidate user terminal. The sum of the communication cost values ​​corresponding to all candidate user terminals in the candidate user terminal set can be used as the total communication cost value corresponding to the candidate user terminal set. Of course, other methods can also be used to calculate the total communication cost value, such as using the average of the communication cost values ​​corresponding to each candidate user terminal as the total communication cost value, etc., and this application embodiment does not limit this.

[0092] Step 240: With the goal of maximizing the ratio of total training revenue to total communication cost, iteratively update the candidate user terminal set, determine the candidate user terminal set at the convergence point as the target user terminal set, and instruct each target user terminal in the target user terminal set to transmit the parameters of the local model to the satellite base station.

[0093] In this embodiment, the candidate user terminal set is iteratively updated with the objective of maximizing the ratio of total training revenue to total communication cost, i.e., maximizing the training revenue achievable per unit of communication cost.

[0094] For example, a greedy strategy can be adopted. Starting from the initial set of candidate user terminals, in each iteration, all unselected user terminals are evaluated one by one. The increment of the total ratio of the new set {U, x} relative to the ratio of the current set U is calculated if the current user terminal x is added to the candidate user terminal set U. The user terminal with the largest increment is selected and added to the candidate user terminal set, and the next iteration begins. The iteration can be considered to have converged if the maximum increment is less than a preset threshold, and the current set of candidate user terminals is determined as the target user terminal set.

[0095] Of course, in addition to using a greedy strategy to solve for the target user terminal set, other methods can also be used, such as using metaheuristic algorithms like genetic algorithms and simulated annealing algorithms, or using Lagrange relaxation and dual decomposition methods. This application does not limit these methods.

[0096] After obtaining the target user terminal set U', each target user terminal in the target user terminal set U' is instructed to transmit the parameters of its local model to the satellite base station, so that the satellite base station can update the global model parameters based on the received local model parameters.

[0097] The data transmission method for federated learning model training provided in this application comprehensively considers the differences in local training data volume and model gradient of each user terminal during the federated learning process, and calculates the total training revenue value corresponding to the candidate user terminal set. This can reduce the impact of non-uniform distribution of training data volume on the global model update efficiency. By taking the maximization of the ratio of total training revenue value to total communication cost value as the objective, the target user terminal set participating in data transmission and global model training is iteratively updated. This can deeply integrate the service volume weighting characteristics of global aggregation in federated learning with satellite communication cost constraints, and simultaneously improve the convergence speed and training accuracy of the model under the same communication resource consumption, thereby improving the training efficiency of the model.

[0098] In some embodiments, based on the amount of training data used to train the local model for each candidate user terminal and the gradient magnitude of the local model's loss function, the total training gain value corresponding to the candidate user terminal set is calculated, including: Calculate the ratio of the amount of training data used by each candidate user terminal to train the local model to the total amount of training data of each user terminal, and obtain the corresponding benefit evaluation weight for each candidate user terminal. Based on the revenue evaluation weights corresponding to each candidate user terminal, the weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal is calculated to obtain the total training revenue value corresponding to the candidate user terminal set.

[0099] In this embodiment of the application, the amount of training data used to train the local model for candidate user terminal i is... The ratio of the total amount of training data K of all user terminals in the satellite communication system. This serves as the revenue evaluation weight corresponding to candidate user terminal i. The revenue evaluation weight calculated in this way represents the proportion of the training data of candidate user terminal i in the total amount of global training data, and is not affected by changes in the size of the target user terminal set. That is, there is no need to recalculate the revenue evaluation weight after each update of the candidate user terminal set, thereby reducing the amount of computation and increasing the speed of calculating the total training revenue value.

[0100] After determining the revenue evaluation weights for each candidate user terminal, the following formula can be used:

[0101] Calculate the weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal to obtain the total training reward value corresponding to the candidate user terminal set.

[0102] Where U represents the set of candidate user terminals, and K represents the total amount of training data for each user terminal. This represents the amount of training data corresponding to candidate user terminal i. This represents the revenue evaluation weight corresponding to candidate user terminal i. This represents the gradient magnitude value corresponding to candidate user terminal i.

[0103] In this embodiment, by using the ratio of the amount of training data used by each candidate user terminal to train the local model to the sum of the training data of each user terminal as the benefit evaluation weight corresponding to each candidate user terminal, the amount of computation required to calculate the total training benefit value can be reduced and the calculation speed can be improved.

[0104] In some embodiments, based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters in the local model of each candidate user terminal, the estimated communication cost value for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated, including: Obtain the initial transmission power and initial channel allocation vector for each candidate user terminal; Based on the initial transmission power, initial channel allocation vector, channel gain value and model parameter quantity of each candidate user terminal, calculate the expected communication delay for each candidate user terminal to upload the parameters of the local model to the satellite base station. Based on the initial transmission power and communication latency of each candidate user terminal, calculate the estimated communication energy consumption of each candidate user terminal when uploading the parameters of the local model to the satellite base station; Based on the estimated communication latency and energy consumption of each candidate user terminal uploading the parameters of the local model to the satellite base station, the communication cost value corresponding to each candidate user terminal is determined.

[0105] In this embodiment, the communication cost is determined based on communication latency and communication energy consumption. Communication latency refers to the estimated time required for a candidate user terminal to upload parameters from its local model to the satellite base station, and communication energy consumption refers to the estimated energy consumed by the candidate user terminal to upload the parameters from its local model to the satellite base station.

[0106] In this embodiment, the initial transmission power and initial channel allocation vector of each candidate user terminal are first obtained. Transmission power refers to the average power of the transmitted signal when a user terminal uploads data via a wireless channel; the channel allocation vector represents the channel assigned to the user terminal for uploading data. For candidate user terminals that have been allocated channels and had their transmission power set during the historical update process of federated learning, their historical transmission power and historical channel allocation vector can be directly used to calculate the communication cost value. For subsequent user terminals that have not yet been allocated channels, they can be temporarily assigned the available channel with the highest channel gain, and their initial transmission power can be set to the maximum value of their available power or other preset values. Of course, other methods can also be used to determine the initial transmission power and initial channel allocation vector of each candidate user terminal, and this embodiment does not limit this method.

[0107] Based on the initial transmission power, initial channel allocation vector, real-time calculated channel gain, and model parameter quantity of each candidate user terminal, the estimated communication delay for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated. For example, in some embodiments, assuming the satellite communication system uses Orthogonal Frequency Division Multiple Access (OFDMA), and there are R channels with bandwidth B available for allocation in the satellite communication system, then the following formula can be used:

[0108]

[0109]

[0110] The calculated communication delay is the estimated time required for a candidate user terminal to upload the parameters of its local model to the satellite base station. ; in, This represents the estimated communication latency for candidate user terminal i to upload the parameters of its local model to the satellite base station, and S represents the size of the data packet corresponding to the parameters of the local model. Let R be the initial channel allocation vector for candidate user terminal i, where R is the total number of channels. This indicates that candidate user terminal i is assigned to channel n. Indicates according to After channel allocation, the information transmission rate from candidate user terminal i to the satellite base station, where B is the channel bandwidth. Let i be the initial transmission power of the candidate user terminal. Let n be the channel gain value of candidate user terminal i on channel n. This represents the noise power spectral density.

[0111] After calculating the communication latency of each candidate user terminal, the estimated communication energy consumption for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated based on the initial transmission power and communication latency of each candidate user terminal. It can be assumed that the transmission power of candidate user terminal i remains constant at its initial transmission power during the data upload process. If it remains unchanged, then it can be directly... and The product of these two values ​​represents the estimated communication energy consumption of candidate user terminal i. .

[0112] The communication delay of candidate user terminal i is calculated. and communication energy consumption Subsequently, based on communication latency and communication energy consumption Determine the communication cost value corresponding to candidate user terminal i. Communication latency and communication energy consumption can be directly included as the communication cost value. Alternatively, the weighted sum of communication latency and communication energy consumption can be used as the communication cost value. Of course, other methods can also be used to calculate the communication cost value. In addition to communication latency and communication energy consumption, other parameters can be introduced to calculate the communication cost value. This application does not limit this.

[0113] In this embodiment, by calculating the communication delay and communication energy consumption based on the real-time channel gain value of each candidate user terminal, the user selection and transmission parameters can be adjusted in response to real-time fluctuations in the channel, reducing packet loss and retransmission, thereby improving the efficiency of data transmission.

[0114] In some embodiments, the total communication cost value corresponding to the set of candidate user terminals is calculated based on the communication cost value corresponding to each candidate user terminal, including: The maximum value of the communication latency corresponding to each candidate user terminal is determined as the communication latency cost value corresponding to the candidate user terminal set. The sum of the communication energy consumption of each candidate user terminal is determined as the communication energy consumption cost value corresponding to the candidate user terminal set. The total communication cost value corresponding to the candidate user terminal set is obtained by calculating the weighted sum of the communication delay cost value and the communication energy consumption cost value.

[0115] In this embodiment, the total communication cost value corresponding to the set of candidate user terminals is calculated by taking into account the latency and energy consumption required for each candidate user terminal to upload data.

[0116] The satellite base station needs to wait until all target user terminals have completed data uploads before updating the global model parameters. Therefore, in this embodiment, the maximum communication latency corresponding to each candidate user terminal is used. The communication delay cost value T corresponding to the candidate user terminal set is determined.

[0117] The sum of communication energy consumption corresponding to each candidate user terminal The communication energy consumption cost value E corresponding to the candidate user terminal set is determined. After determining the communication latency cost value and communication energy consumption cost value corresponding to the candidate user terminal set, a preset ratio coefficient can be used. and Calculate the weighted sum of communication latency cost and communication energy consumption cost. This yields the total communication cost value corresponding to the candidate user terminal set.

[0118] In this embodiment, by using the weighted sum of communication latency cost and communication energy cost as the total communication cost value corresponding to the candidate user terminal set, and then taking the maximization of the ratio of total training revenue value to total communication cost value as the optimization objective, energy consumption and latency during data transmission can be reduced, thereby improving the overall communication efficiency of the satellite communication system.

[0119] In some embodiments, with the objective of maximizing the ratio of total training revenue to total communication cost, the candidate user terminal set is iteratively updated. The candidate user terminal set at the convergence point is determined as the target user terminal set, and each target user terminal in the target user terminal set transmits the parameters of its local model to the satellite base station, including: Initialize the candidate user terminal set for the current iteration round; The candidate user terminal set is updated with the objective function as the goal; the objective function represents the ratio of the total training revenue to the total communication cost. After the update termination condition is met, the target transmit power of each candidate user terminal is calculated with the goal of minimizing the expected communication cost of each candidate user terminal in the candidate user terminal set. Based on the target transmit power of each candidate user terminal, the set of target user-channel matching pairs that minimizes the total communication cost value corresponding to the set of candidate user terminals is determined; the set of target user-channel matching pairs includes the matching pairs of each candidate user terminal and the channel. Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set, calculate the value of the objective function corresponding to the candidate user terminal set in the current iteration round; If the difference between the value of the objective function corresponding to the current iteration round and the value of the objective function corresponding to the previous iteration round is less than the convergence threshold, then the set of candidate user terminals in the current iteration round is determined as the set of target user terminals. The target channel corresponding to each target user terminal is determined based on the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power.

[0120] In this embodiment, the set of candidate user terminals participating in the global model parameter update, the target power configuration of each candidate user terminal, and the channel allocation scheme are jointly optimized. Solving the joint optimal solution for all three dimensions at once involves a huge combinatorial search space, resulting in excessively high computational complexity. Therefore, in this embodiment, the joint optimization problem is decomposed into three sub-problems: user selection, power configuration, and channel allocation. Each sub-problem is optimized sequentially in each iteration round.

[0121] At the start of each iteration, the candidate user terminal set for the current iteration is first initialized. For example, at the start of the m-th iteration, the entire set of user terminals in the satellite communication system can be selected. A user terminal is randomly selected and added to the candidate user terminal set for the current iteration round. In this context, the empty set serves as the starting point for this iteration. In some embodiments, the empty set can also be used as the starting point directly.

[0122] Complete the candidate user terminal set After initialization, the candidate user terminal set is updated with the objective function as the goal. The objective function represents the ratio of the total training revenue to the total communication cost. For example, in some embodiments, the objective function is:

[0123] Where U represents the set of candidate user terminals, and K represents the total amount of training data for each user terminal. This represents the amount of training data corresponding to candidate user terminal i. Let represent the gradient magnitude value corresponding to candidate user terminal i, E represent the communication energy consumption cost value corresponding to the candidate user terminal set, and T represent the communication latency cost value corresponding to the candidate user terminal set. and It is a proportionality coefficient, and .

[0124] During the update process, when calculating the communication energy consumption cost and communication delay cost corresponding to the candidate user terminal set, for user terminals that have been allocated transmission power and channels in previous iterations, the transmission power and channels calculated in the most recent iteration can be reused to calculate the communication cost; for user terminals that have not been allocated transmission power and channels, the communication cost can be calculated based on the initial transmission power and the initial channel allocation vector.

[0125] The candidate user terminal set can be updated using greedy algorithms, metaheuristic algorithms, etc., and after the update, it should be checked whether a preset update termination condition has been met. The update termination condition can be, for example, the convergence of the objective function or reaching the maximum number of updates. If the update termination condition has not been met, the candidate user terminal set should be updated again.

[0126] After the update termination condition is met, the candidate user terminal set is used. The objective is to minimize the expected communication cost for each candidate user terminal, and then calculate the target transmit power for each candidate user terminal. When calculating the communication cost, for candidate user terminals that have been allocated a channel in previous iterations, the channel cost calculated in the most recent iteration can be reused; for candidate user terminals that have not been allocated a channel, the communication cost can be calculated based on the initial channel allocation vector.

[0127] The target transmit power can be solved using the Laplace algorithm and KKT conditions, or it can be solved using the ADMM (Alternating Direction Method of Multipliers) algorithm, metaheuristic algorithms, etc. The embodiments of this application do not limit this.

[0128] After obtaining the target transmit power of each candidate user terminal, based on the target transmit power of each candidate user terminal, a solution is found that makes the set of candidate user terminals... The set of target user-channel matching pairs that minimizes the corresponding total communication cost. This set includes matching pairs between each candidate user terminal and its corresponding channel.

[0129] The user-channel matching pair set can be used as the variable for channel allocation optimization, the total communication cost can be used as the objective function for channel allocation optimization, and constraints can be set according to the performance parameters of each user terminal and the interference between users. Then, the channel allocation optimization problem can be solved by algorithms such as mathematical programming, game theory, heuristic algorithms, and reinforcement learning to obtain the target user-channel matching pair set.

[0130] Of course, other algorithms can also be used to solve the problem, but this application does not limit the specific algorithms used in this embodiment.

[0131] After updating the candidate user terminal set and obtaining the target transmit power and target user-channel matching pair set for each candidate user terminal, the objective function value corresponding to the candidate user terminal set in the current iteration round m is calculated based on the target transmit power and target user-channel matching pair set for each candidate user terminal. The value of the objective function corresponding to the current iteration is calculated. Then, calculate The value of the objective function corresponding to the previous iteration, i.e., the (m-1)th iteration. The difference .like If the convergence threshold is greater than or equal to the threshold, then the next iteration round begins and the candidate user terminal set is reinitialized; if... If the value is less than the convergence threshold, the iteration stops.

[0132] Of course, other iteration termination conditions can be set, such as stopping iteration when the preset maximum number of iterations is reached, etc., but this application does not limit this.

[0133] The set of candidate user terminals in the current iteration round is determined as the set of target user terminals. Based on the target user-channel matching pair set calculated in the m-th iteration round, the target channel corresponding to each target user terminal is determined, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power calculated in the m-th iteration round, so that the satellite base station updates the global model parameters based on the received local model parameters.

[0134] In this embodiment, by iteratively optimizing the user terminal set, power configuration, and channel allocation scheme, the computational complexity can be reduced and the solution speed can be improved. After updating the candidate user terminal set, the target transmit power that minimizes the communication cost of each candidate user terminal is found, and the channel allocation scheme that minimizes the total communication cost of the system is solved. This enables precise matching between transmit power and channel status and mission requirements, thereby improving the overall communication efficiency of the satellite communication system while increasing the model training speed.

[0135] In some embodiments, the iteration termination condition is that the number of candidate user terminals in the candidate user terminal set is greater than or equal to the target number; the candidate user terminal set is updated with the objective of maximizing the ratio of total training revenue to total communication cost, including: Based on each first user terminal and the candidate user terminal set, a temporary user terminal set corresponding to each first user terminal is constructed; the first user terminal is the user terminal among the user terminals that has not been added to the candidate user terminal set. Calculate the value of the auxiliary function corresponding to each set of temporary user terminals; wherein the auxiliary function is constructed based on the objective function; Add the first user terminal corresponding to the temporary user terminal set with the largest auxiliary function value to the candidate user terminal set.

[0136] In this embodiment, the objective function represents the ratio of the total training revenue to the total communication cost. However, the calculation of the fraction is non-linear, meaning that the value of the objective function may not change linearly with the update process of the candidate user terminals, making it difficult to judge the changing trend of the global objective through the increment of the objective function.

[0137] Therefore, in this embodiment, an auxiliary function is constructed based on the objective function, and the candidate user terminal set is updated according to the value of the auxiliary function. For example, in some embodiments, the Dinkelbach method can be used to construct the auxiliary function:

[0138] in, This represents the value of the objective function corresponding to the previous iteration round.

[0139] In the global optimization process across iterations, finding the solution that maximizes the objective function is equivalent to finding the solution that maximizes the auxiliary function. The solution with the largest intercept on the x-axis. Within each iteration, we can... Treating it as a constant, at this time It exhibits linear additivity. Therefore, in the embodiments of this application, a greedy strategy can be used to update the candidate user terminal set, selecting the option that enables the auxiliary function in each update. The user terminal with the largest increase is added to the candidate user set.

[0140] Suppose that after the nth update, the set of candidate user terminals is Then the set Each user terminal in the above is a first user terminal. For each first user terminal... It can be calculated and The union of, as The corresponding set of temporary user terminals .

[0141] After constructing the set of temporary user terminals corresponding to each first user terminal, calculate the auxiliary function corresponding to each set of temporary user terminals. The value of the auxiliary function is determined, and the first user terminal corresponding to the temporary user terminal set with the largest value of the auxiliary function is formally added to the candidate user terminal set. In the middle, the candidate user terminal set is updated to .

[0142] In this embodiment, the iteration termination condition is that the number of candidate user terminals in the candidate user terminal set is greater than or equal to the target number. The target number can be a preset fixed value; of course, it can also be dynamically adjusted, for example, determined based on the number of currently available channels, etc., and this embodiment does not limit this.

[0143] Then, after the (n+1)th update, check the set. If the number of candidate user terminals is less than the target number, the (n+2)th update begins; if it is greater than or equal to the target number, the update ends, and the process proceeds to the next step of solving for the target transmit power.

[0144] In this embodiment, by constructing an auxiliary function based on the objective function and adopting a greedy strategy, the user terminal that maximizes the increment of the auxiliary function is selected to be added to the candidate user set in each round of updates. This avoids solving complex fractional optimization problems, reduces computational complexity, and thus improves the computational speed of updating the candidate user terminal set.

[0145] In some embodiments, the target transmit power of each candidate user terminal is calculated with the objective of minimizing the expected communication cost of the candidate user terminal, including: Multiple candidate power values ​​are determined from the target power range; Calculate the estimated communication cost for candidate user terminals to transmit signals using each candidate power value; The candidate power value with the lowest communication cost is determined as the target transmit power of the candidate user terminal.

[0146] To improve the speed of solving for the optimal transmission power, in this embodiment of the application, the target transmission power of each candidate user terminal is solved by a linear search method.

[0147] For candidate user terminal i, multiple candidate power values ​​are first determined from the target power range. The target power range can be determined based on the available energy of candidate user terminal i, the performance of the transmitting antenna, etc. The search step size is set based on the desired solution accuracy, and then the candidate power values ​​are calculated based on the target power range and the search step size. For example, the target power range is... The search step size is The calculated candidate power value is ,in , .

[0148] After determining the candidate power values, calculate the estimated communication cost for candidate user terminal i to transmit signals using each candidate power value. The candidate power value with the lowest communication cost is determined as the target transmit power of candidate user terminal i. .

[0149] Of course, in addition to determining the candidate power value by a single search step size, two different search step sizes can be set. First, the power range is narrowed down by the larger search step size, and then the final result is determined by the smaller search step size. In addition to determining the candidate power value by the search step size, other methods can also be used to determine the candidate power value, such as selecting the power value located in the target power range from a series of preset power values ​​as the candidate power value. This application does not limit this.

[0150] The solution process of linear search mainly involves multiplication and addition operations, and hardware processors can quickly complete a large number of multiplication and addition calculations through pipelined and parallel computing methods. Using linear search to solve for the target transmit power of each candidate user terminal can avoid complex operations such as matrix inversion and differentiation, thereby improving the speed of solving for the target transmit power.

[0151] In this embodiment, by using a linear search method to solve for the target transmit power of each candidate user terminal, the computational complexity can be reduced and the speed of solving for the target transmit power can be increased.

[0152] In some embodiments, based on the target transmit power of each candidate user terminal, the set of target user-channel matching pairs that minimizes the total communication cost value corresponding to the candidate user terminal set is determined, including: Based on the target transmit power of each candidate user terminal, the expected communication cost value for each candidate user terminal to select each candidate channel is calculated, and the communication cost matrix is ​​obtained. Each candidate user terminal is taken as a node in the first subgraph, each candidate channel is taken as a node in the second subgraph, the matching pairs of each candidate user terminal and each candidate channel are taken as the edges between the nodes of the first subgraph and the nodes of the second subgraph, and the communication cost matrix is ​​taken as the weight matrix corresponding to each edge to establish a user-channel bipartite graph. Find the user-channel matching pairs that minimize the total weight of each edge in the user-channel bipartite graph to obtain the target user-channel matching pair set.

[0153] In this embodiment of the application, to reduce interference between users, the constraint is set that a channel can be allocated to at most one user terminal at a time, and each user terminal can be allocated to at most one channel at a time. Assuming there are R channels available for allocation in the satellite communication system, one channel is allocated to each candidate user terminal in the candidate user terminal set U.

[0154] If channel r is assigned to candidate user terminal i, the resulting user-channel matching pair is (i, r). Then, based on the target transmit power of candidate user terminal i and the available bandwidth of channel r, the expected communication cost c(i, r) for candidate user terminal i to transmit data through channel r can be calculated. Based on the target transmit power of each candidate user terminal, the expected communication cost for each candidate user terminal to select each candidate channel is calculated, resulting in the communication cost matrix C[i][R].

[0155] The channel allocation problem and the bipartite graph matching problem share the same mathematical structure. Therefore, in this embodiment, the channel allocation problem is modeled as a minimum weight matching problem in a bipartite graph, which can then be solved using mature bipartite graph matching algorithms.

[0156] In this embodiment, each candidate user terminal is taken as a node of the first subgraph, each candidate channel is taken as a node of the second subgraph, the matching pair of each candidate user terminal and each candidate channel is taken as the edge between the node of the first subgraph and the node of the second subgraph, and the communication cost matrix is ​​taken as the weight matrix corresponding to each edge. Then c(i, r) is the weight of edge (i, r).

[0157] Therefore, a user-channel bipartite graph is constructed. Finding the target user-channel matching pair set that minimizes the total communication cost corresponding to the candidate user terminal set is equivalent to finding the edge matching combination that minimizes the sum of edge weights. The minimum weight matching problem in the bipartite graph can be quickly solved using methods such as the Hungarian algorithm, auction algorithm, and shortest augmenting path algorithm to obtain the target user-channel matching pair set.

[0158] In this embodiment, by modeling the channel allocation problem as a bipartite graph minimum weight matching problem and then solving it, the computational complexity can be reduced while ensuring that the solution meets the channel allocation constraints, thereby improving the efficiency of solving the target user-channel matching pair set.

[0159] The following scenario illustrates the data transmission method in federated learning model training provided in this application's embodiments. For example... Figure 3 As shown in this example scenario, after the local model training phase ends, the set of target user terminals participating in the global model parameter update begins iteratively. At the beginning of each iteration, the candidate user terminal set for the current iteration is first initialized.

[0160] In each iteration, the first step is to find the candidate user terminal set that maximizes the auxiliary function value. From the user terminals not yet added to the candidate user terminal set, the first user terminal is determined, a corresponding temporary user terminal set is constructed, and the auxiliary function value for this temporary user terminal set is calculated. If the auxiliary function value is greater than the historical maximum value, the historical best user terminal is updated to the first user terminal. It is then checked whether all unselected user terminals have been traversed; otherwise, the next first user terminal is selected. If so, the historical best user terminal is added to the candidate user terminal set, and the number of candidate user terminals in the set is checked to see if it is less than the target number. If so, the historical best user terminal is reset, and the selection of the next first user terminal begins; otherwise, the transmit power of each candidate user terminal is optimized.

[0161] For each candidate user terminal in the candidate user terminal set, the target transmit power of each candidate user terminal is obtained through a linear search method with the goal of minimizing the communication cost. After calculating the target transmit power of each candidate user terminal, the target user-channel matching pair set that minimizes the total communication cost corresponding to the candidate user terminal set is solved based on the target transmit power of each candidate user terminal.

[0162] Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set corresponding to the candidate user terminal set, the objective function value corresponding to the candidate user terminal set in the current iteration round is calculated, and the difference between the objective function value corresponding to the current iteration round and the objective function value corresponding to the previous iteration round is calculated. If the difference is greater than or equal to the convergence threshold, the next iteration round is entered. If the difference is less than the convergence threshold, the iteration ends, the candidate user terminal set in the current iteration round is determined as the target user terminal set, the target channel corresponding to each target user terminal is determined according to the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station with the target transmit power and through the target channel.

[0163] The data transmission method for training a federated learning model provided in this application can be executed by a data transmission device for training a federated learning model. This application uses an example of a data transmission device for training a federated learning model executing the data transmission method to illustrate the data transmission device for training a federated learning model provided in this application.

[0164] This application also provides a data transmission device for training a federated learning model.

[0165] like Figure 4 As shown, the data transmission device in the training of this federated learning model includes: The determining module 410 is used to determine a candidate user terminal set from among the user terminals, the candidate user terminal set including at least one candidate user terminal; The first calculation module 420 is used to calculate the total training revenue value corresponding to the candidate user terminal set based on the amount of training data used to train the local model for each candidate user terminal and the gradient magnitude value of the local model loss function. The second calculation module 430 is used to calculate the expected communication cost value for each candidate user terminal to upload the parameters of the local model to the satellite base station based on the channel gain value between each candidate user terminal and the satellite base station and the number of model parameters of the local model of each candidate user terminal, and to calculate the total communication cost value corresponding to the set of candidate user terminals based on the communication cost value corresponding to each candidate user terminal. The update module 440 is used to iteratively update the candidate user terminal set with the goal of maximizing the ratio of the total training revenue value to the total communication cost value. The candidate user terminal set that converges is determined as the target user terminal set, and each target user terminal in the target user terminal set transmits the parameters of the local model to the satellite base station.

[0166] According to the data transmission device in the federated learning model training of this application, by comprehensively considering the differences in the amount of local training data and model gradients of each user terminal during the federated learning process, the total training revenue value corresponding to the candidate user terminal set is calculated. This can reduce the impact of the non-uniform distribution of training data on the global model update efficiency. By taking the maximization of the ratio of the total training revenue value to the total communication cost value as the objective, the target user terminal set participating in data transmission and global model training is iteratively updated. This can deeply integrate the service volume weighting characteristics of global aggregation in federated learning with the satellite communication cost constraints, and simultaneously improve the convergence speed and training accuracy of the model under the same communication resource consumption, thereby improving the training efficiency of the model.

[0167] In some embodiments, the first computing module 420 is further configured to: Calculate the ratio of the amount of training data used by each candidate user terminal to train the local model to the total amount of training data of each user terminal, and obtain the corresponding benefit evaluation weight for each candidate user terminal. Based on the revenue evaluation weights corresponding to each candidate user terminal, the weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal is calculated to obtain the total training revenue value corresponding to the candidate user terminal set.

[0168] In some embodiments, the second computing module 430 is further configured to: Obtain the initial transmission power and initial channel allocation vector for each candidate user terminal; Based on the initial transmission power, initial channel allocation vector, channel gain value and model parameter quantity of each candidate user terminal, calculate the expected communication delay for each candidate user terminal to upload the parameters of the local model to the satellite base station. Based on the initial transmission power and communication latency of each candidate user terminal, calculate the estimated communication energy consumption of each candidate user terminal when uploading the parameters of the local model to the satellite base station; Based on the estimated communication latency and energy consumption of each candidate user terminal uploading the parameters of the local model to the satellite base station, the communication cost value corresponding to each candidate user terminal is determined.

[0169] In some embodiments, the second computing module 430 is further configured to: The maximum value of the communication latency corresponding to each candidate user terminal is determined as the communication latency cost value corresponding to the candidate user terminal set. The sum of the communication energy consumption of each candidate user terminal is determined as the communication energy consumption cost value corresponding to the candidate user terminal set. The total communication cost value corresponding to the candidate user terminal set is obtained by calculating the weighted sum of the communication delay cost value and the communication energy consumption cost value.

[0170] In some embodiments, the update module 440 is further configured to: Initialize the candidate user terminal set for the current iteration round; The candidate user terminal set is updated with the objective function as the goal; the objective function represents the ratio of the total training revenue to the total communication cost. After the update termination condition is met, the target transmit power of each candidate user terminal is calculated with the goal of minimizing the expected communication cost of each candidate user terminal in the candidate user terminal set. Based on the target transmit power of each candidate user terminal, the set of target user-channel matching pairs that minimizes the total communication cost value corresponding to the set of candidate user terminals is determined; the set of target user-channel matching pairs includes the matching pairs of each candidate user terminal and the channel. Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set, calculate the value of the objective function corresponding to the candidate user terminal set in the current iteration round; If the difference between the value of the objective function corresponding to the current iteration round and the value of the objective function corresponding to the previous iteration round is less than the convergence threshold, then the set of candidate user terminals in the current iteration round is determined as the set of target user terminals. The target channel corresponding to each target user terminal is determined based on the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power.

[0171] In some embodiments, the update module 440 is further configured to: Based on each first user terminal and the candidate user terminal set, a temporary user terminal set corresponding to each first user terminal is constructed; the first user terminal is the user terminal among the user terminals that has not been added to the candidate user terminal set. Calculate the value of the auxiliary function corresponding to each set of temporary user terminals; wherein the auxiliary function is constructed based on the objective function; Add the first user terminal corresponding to the temporary user terminal set with the largest auxiliary function value to the candidate user terminal set.

[0172] In some embodiments, the update module 440 is further configured to: Multiple candidate power values ​​are determined from the target power range; Calculate the estimated communication cost for candidate user terminals to transmit signals using each candidate power value; The candidate power value with the lowest communication cost is determined as the target transmit power of the candidate user terminal.

[0173] In some embodiments, the update module 440 is further configured to: Based on the target transmit power of each candidate user terminal, the expected communication cost value for each candidate user terminal to select each candidate channel is calculated, and the communication cost matrix is ​​obtained. Each candidate user terminal is taken as a node in the first subgraph, each candidate channel is taken as a node in the second subgraph, the matching pairs of each candidate user terminal and each candidate channel are taken as the edges between the nodes of the first subgraph and the nodes of the second subgraph, and the communication cost matrix is ​​taken as the weight matrix corresponding to each edge to establish a user-channel bipartite graph. Find the user-channel matching pairs that minimize the total weight of each edge in the user-channel bipartite graph to obtain the target user-channel matching pair set.

[0174] In this application's embodiments, the data transmission device in the federated learning model training can be an electronic device or a component within an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a satellite base station, server, industrial control computer, edge computing gateway, personal computer, workstation, intelligent inspection equipment, field monitoring host, intelligent power equipment, etc., and this application's embodiments do not impose specific limitations.

[0175] The data transmission device in the federated learning model training of this application embodiment can be a device with an operating system. This operating system can be a Microsoft (Windows) operating system, an Android operating system, an iOS operating system, or other possible operating systems; this application embodiment does not specifically limit it.

[0176] In some embodiments, such as Figure 5 As shown, this application embodiment also provides an electronic device 500, including a processor 501, a memory 502, and a computer program stored in the memory 502 and executable on the processor 501. When the program is executed by the processor 501, it implements the various processes of the data transmission method embodiment in the above-described federated learning model training and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0177] It should be noted that the electronic devices in the embodiments of this application include the aforementioned mobile electronic devices and non-mobile electronic devices.

[0178] This application also provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the data transmission method embodiment in the above-described federated learning model training and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0179] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0180] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the data transmission method in the training of the federated learning model described above.

[0181] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0182] This application also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the data transmission method embodiment in the above federated learning model training, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0183] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0184] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0185] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0186] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0187] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0188] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for data transmission in federated learning model training, applied to a satellite communication system, the method comprising the steps of: The satellite communication system includes a satellite base station and multiple user terminals, each user terminal having a local model deployed locally; the method includes: A candidate user terminal set is determined from each of the user terminals, the candidate user terminal set including at least one candidate user terminal; Based on the amount of training data used to train the local model for each of the candidate user terminals and the gradient magnitude of the loss function of the local model, the total training gain value corresponding to the candidate user terminal set is calculated. Based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters of the local model of each candidate user terminal, the estimated communication cost value for each candidate user terminal to upload the parameters of the local model to the satellite base station is calculated, and based on the communication cost value corresponding to each candidate user terminal, the total communication cost value corresponding to the set of candidate user terminals is calculated. With the objective of maximizing the ratio of the total training revenue to the total communication cost, the candidate user terminal set is iteratively updated. The candidate user terminal set that converges is determined as the target user terminal set, and each target user terminal in the target user terminal set transmits the parameters of the local model to the satellite base station.

2. The method of claim 1, wherein, The step of calculating the total training gain value corresponding to the candidate user terminal set based on the amount of training data used to train the local model for each candidate user terminal and the gradient magnitude value of the local model's loss function includes: Calculate the ratio of the amount of training data used by each candidate user terminal to train the local model to the total amount of training data of each user terminal, and obtain the benefit evaluation weight corresponding to each candidate user terminal. Based on the revenue evaluation weights corresponding to each candidate user terminal, a weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal is calculated to obtain the total training revenue value corresponding to the candidate user terminal set.

3. The method of claim 2, wherein, According to the formula: The total training revenue value corresponding to the candidate user terminal set is calculated; Where U represents the set of candidate user terminals, and K represents the total amount of training data for each of the user terminals. This represents the amount of training data corresponding to candidate user terminal i. This represents the revenue evaluation weight corresponding to candidate user terminal i. This represents the gradient magnitude value corresponding to candidate user terminal i.

4. The method of claim 1, wherein, According to the formula: The channel gain value between the candidate user terminal and the satellite base station is calculated. Where h represents the channel gain value between the candidate user terminal and the satellite base station. This indicates large-scale decay. d is a scaling factor, and d is the distance between the candidate user terminal and the satellite base station. This is the path loss index. denotes small scale fading, is the Rician factor, denotes the random phase of the line of sight component, denotes the Rayleigh fading, and are independent zero mean Gaussian variables.

5. The method of claim 1, wherein, The step of calculating the estimated communication cost for each candidate user terminal to upload its local model parameters to the satellite base station, based on the channel gain value between each candidate user terminal and the satellite base station, and the number of model parameters of each candidate user terminal's local model, includes: Obtain the initial transmission power and initial channel allocation vector for each of the candidate user terminals; Based on the initial transmission power of each candidate user terminal, the initial channel allocation vector, the channel gain value, and the number of model parameters, the estimated communication delay for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated. Based on the initial transmission power and communication delay of each candidate user terminal, calculate the estimated communication energy consumption of each candidate user terminal when uploading the parameters of the local model to the satellite base station; Based on the estimated communication latency and energy consumption of each candidate user terminal uploading the parameters of the local model to the satellite base station, the communication cost value corresponding to each candidate user terminal is determined.

6. The method of claim 5, wherein, The step of calculating the total communication cost value corresponding to the set of candidate user terminals based on the communication cost value corresponding to each of the candidate user terminals includes: The maximum value of the communication delay corresponding to each of the candidate user terminals is determined as the communication delay cost value corresponding to the set of candidate user terminals; The sum of the communication energy consumption corresponding to each of the candidate user terminals is determined as the communication energy consumption cost value corresponding to the candidate user terminal set. The total communication cost value corresponding to the candidate user terminal set is obtained by calculating the weighted sum of the communication latency cost value and the communication energy consumption cost value.

7. The method of claim 5, wherein, According to the formula: The estimated communication delay for the candidate user terminal to upload the parameters of the local model to the satellite base station is calculated. wherein, denotes the communication latency that the candidate user terminal i is expected to take to upload the parameters of the local model to the satellite base station, S denotes the volume of the data packet corresponding to the parameters of the local model, S represents the size of the data packet corresponding to the parameters of the local model. R is the total number of channels, denotes that candidate user terminal i is assigned to channel n, denotes the transmission rate of the candidate user terminal i to the satellite base station after the channel has been allocated, B is the channel bandwidth, denotes the transmission rate of the candidate user terminal i to the satellite base station after the channel has been allocated, B is the channel bandwidth, denotes the initial transmission power of the candidate user terminal i, denotes the channel gain value of the candidate user terminal i on the channel n, denotes the noise power spectral density.

8. The method of claim 1, wherein, The step of iteratively updating the candidate user terminal set with the objective of maximizing the ratio of the total training revenue to the total communication cost, determining the candidate user terminal set at the convergence point as the target user terminal set, and instructing each target user terminal in the target user terminal set to transmit the parameters of the local model to the satellite base station includes: Initialize the candidate user terminal set for the current iteration round; The candidate user terminal set is updated with the objective function being maximized; the objective function represents the ratio of the total training revenue to the total communication cost. After the update termination condition is met, the target transmit power of each candidate user terminal is calculated with the goal of minimizing the expected communication cost value of each candidate user terminal in the candidate user terminal set. Based on the target transmit power of each candidate user terminal, a target user-channel matching pair set that minimizes the total communication cost value corresponding to the candidate user terminal set is determined; the target user-channel matching pair set includes matching pairs of each candidate user terminal and channel. Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set, calculate the value of the objective function corresponding to the candidate user terminal set in the current iteration round; If the difference between the value of the objective function corresponding to the current iteration round and the value of the objective function corresponding to the previous iteration round is less than the convergence threshold, then the candidate user terminal set of the current iteration round is determined as the target user terminal set. The target channel corresponding to each target user terminal is determined according to the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power.

9. The method of claim 8, wherein, The iteration termination condition is that the number of candidate user terminals in the candidate user terminal set is greater than or equal to the target number; updating the candidate user terminal set with the objective of maximizing the ratio of the total training revenue to the total communication cost includes: Based on each first user terminal and the candidate user terminal set, a temporary user terminal set corresponding to each first user terminal is constructed; the first user terminal is a user terminal among the user terminals that has not been added to the candidate user terminal set. Calculate the value of the auxiliary function corresponding to each of the aforementioned temporary user terminal sets; wherein the auxiliary function is constructed based on the objective function; The first user terminal corresponding to the temporary user terminal set with the largest value of the auxiliary function is added to the candidate user terminal set.

10. The method of claim 8, wherein, The step of determining the target transmit power of each candidate user terminal with the objective of minimizing the expected communication cost of the candidate user terminals includes: Multiple candidate power values ​​are determined from the target power range; Calculate the estimated communication cost for the candidate user terminal to transmit signals using each candidate power value; The candidate power value with the lowest communication cost is determined as the target transmit power of the candidate user terminal.

11. The method of claim 8, wherein, The step of finding the target user-channel matching pair set that minimizes the total communication cost value corresponding to the candidate user terminal set based on the target transmit power of each candidate user terminal includes: Based on the target transmit power of each candidate user terminal, the expected communication cost value for each candidate user terminal to select each candidate channel is calculated, and a communication cost matrix is ​​obtained. Each candidate user terminal is used as a node in the first subgraph, each candidate channel is used as a node in the second subgraph, the matching pairs of each candidate user terminal and each candidate channel are used as edges between the nodes of the first subgraph and the nodes of the second subgraph, and the communication cost matrix is ​​used as the weight matrix corresponding to each edge to establish a user-channel bipartite graph. The target user-channel matching pair set is obtained by finding the user-channel matching pairs that minimize the total weight value of each edge in the user-channel bipartite graph.

12. The method according to claim 9, characterized in that, The objective function is: The auxiliary function is: Where U represents the set of candidate user terminals, and K represents the total amount of training data for each of the user terminals. This represents the amount of training data corresponding to candidate user terminal i. Let represent the gradient magnitude value corresponding to candidate user terminal i, E represent the communication energy consumption cost value corresponding to the candidate user terminal set, and T represent the communication latency cost value corresponding to the candidate user terminal set. and It is a proportionality coefficient, and , This represents the value of the objective function corresponding to the previous iteration round.

13. A data transmission device for training a federated learning model, characterized in that, include: A determining module is used to determine a candidate user terminal set from each user terminal, the candidate user terminal set including at least one candidate user terminal; The first calculation module is used to calculate the total training revenue value corresponding to the set of candidate user terminals based on the amount of training data used to train the local model for each candidate user terminal and the gradient magnitude value of the loss function of the local model. The second calculation module is used to calculate the estimated communication cost of each candidate user terminal uploading the parameters of its local model to the satellite base station based on the channel gain value between each candidate user terminal and the satellite base station and the number of model parameters of the local model of each candidate user terminal, and to calculate the total communication cost value corresponding to the set of candidate user terminals based on the communication cost value corresponding to each candidate user terminal. The update module is used to iteratively update the candidate user terminal set with the objective of maximizing the ratio of the total training revenue value to the total communication cost value, determine the candidate user terminal set when the iteration reaches convergence as the target user terminal set, and instruct each target user terminal in the target user terminal set to transmit the parameters of the local model to the satellite base station.

14. The apparatus according to claim 13, characterized in that, The first calculation module is further configured to: Calculate the ratio of the amount of training data used by each candidate user terminal to train the local model to the total amount of training data of each user terminal, and obtain the benefit evaluation weight corresponding to each candidate user terminal. Based on the revenue evaluation weights corresponding to each candidate user terminal, a weighted sum of the gradient magnitude values ​​corresponding to each candidate user terminal is calculated to obtain the total training revenue value corresponding to the candidate user terminal set.

15. The apparatus according to claim 13, characterized in that, The second calculation module is also used for: Obtain the initial transmission power and initial channel allocation vector for each of the candidate user terminals; Based on the initial transmission power of each candidate user terminal, the initial channel allocation vector, the channel gain value, and the number of model parameters, the estimated communication delay for each candidate user terminal to upload the parameters of its local model to the satellite base station is calculated. Based on the initial transmission power and communication delay of each candidate user terminal, calculate the estimated communication energy consumption of each candidate user terminal when uploading the parameters of the local model to the satellite base station; Based on the estimated communication latency and energy consumption of each candidate user terminal uploading the parameters of the local model to the satellite base station, the communication cost value corresponding to each candidate user terminal is determined.

16. The apparatus according to claim 15, characterized in that, The second calculation module is also used for: The maximum value of the communication delay corresponding to each of the candidate user terminals is determined as the communication delay cost value corresponding to the set of candidate user terminals; The sum of the communication energy consumption corresponding to each of the candidate user terminals is determined as the communication energy consumption cost value corresponding to the candidate user terminal set. The total communication cost value corresponding to the candidate user terminal set is obtained by calculating the weighted sum of the communication latency cost value and the communication energy consumption cost value.

17. The apparatus according to claim 13, characterized in that, The update module is also used for: Initialize the candidate user terminal set for the current iteration round; The candidate user terminal set is updated with the objective function being maximized; the objective function represents the ratio of the total training revenue to the total communication cost. After the update termination condition is met, the target transmit power of each candidate user terminal is calculated with the goal of minimizing the expected communication cost value of each candidate user terminal in the candidate user terminal set. Based on the target transmit power of each candidate user terminal, a target user-channel matching pair set that minimizes the total communication cost value corresponding to the candidate user terminal set is determined; the target user-channel matching pair set includes matching pairs of each candidate user terminal and channel. Based on the target transmit power of each candidate user terminal and the target user-channel matching pair set, calculate the value of the objective function corresponding to the candidate user terminal set in the current iteration round; If the difference between the value of the objective function corresponding to the current iteration round and the value of the objective function corresponding to the previous iteration round is less than the convergence threshold, then the candidate user terminal set of the current iteration round is determined as the target user terminal set. The target channel corresponding to each target user terminal is determined according to the target user-channel matching pair set, and each target user terminal transmits the parameters of the local model to the satellite base station through the target channel with the target transmit power.

18. The apparatus according to claim 17, characterized in that, The update module is also used for: Based on each first user terminal and the candidate user terminal set, a temporary user terminal set corresponding to each first user terminal is constructed; the first user terminal is a user terminal among the user terminals that has not been added to the candidate user terminal set. Calculate the value of the auxiliary function corresponding to each of the aforementioned temporary user terminal sets; wherein the auxiliary function is constructed based on the objective function; The first user terminal corresponding to the temporary user terminal set with the largest value of the auxiliary function is added to the candidate user terminal set.

19. The apparatus according to claim 17, characterized in that, The update module is also used for: Multiple candidate power values ​​are determined from the target power range; Calculate the estimated communication cost for the candidate user terminal to transmit signals using each candidate power value; The candidate power value with the lowest communication cost is determined as the target transmit power of the candidate user terminal.

20. The apparatus according to claim 17, characterized in that, The update module is also used for: Based on the target transmit power of each candidate user terminal, the expected communication cost value for each candidate user terminal to select each candidate channel is calculated, and a communication cost matrix is ​​obtained. Each candidate user terminal is used as a node in the first subgraph, each candidate channel is used as a node in the second subgraph, the matching pairs of each candidate user terminal and each candidate channel are used as edges between the nodes of the first subgraph and the nodes of the second subgraph, and the communication cost matrix is ​​used as the weight matrix corresponding to each edge to establish a user-channel bipartite graph. The target user-channel matching pair set is obtained by finding the user-channel matching pairs that minimize the total weight value of each edge in the user-channel bipartite graph.