CN-122021965-A - MSE parameter aggregation method and system for air computing assisted federal learning

CN122021965ACN 122021965 ACN122021965 ACN 122021965ACN-122021965-A

Abstract

The invention relates to the technical field of federal learning of artificial intelligence, in particular to a method and a system for aggregating MSE parameters of federal learning assisted by aerial computation, wherein the method comprises the steps of training a model through multiple iterations, wherein each iteration comprises the steps of receiving global parameters by mobile equipment, calculating gradients based on local partial data, and uploading the gradients according to self amplification coefficients; and the base station minimizes the mean square error MSE of the restored signal and the ideal aggregation gradient through two layers of optimization, and distributes the updated parameters to the mobile equipment until training converges. According to the invention, through collaborative optimization of the amplification coefficients of the mobile equipment and the base station and the training data quantity of the mobile equipment, the MSE is reduced in depth, the equipment calculation load is reduced, the calculation complexity is controlled, the universality is high, the mobile equipment and the base station can be compatible with a mainstream model, and the training performance is superior to that of the traditional method.

Inventors

LIN HUI
ZHU FUJIAN
WEI XI
WU ZIPING
FAN RONGFEI
HU HAN
JIN HAO
LI HUIBO
ZHANG XUEYING
Wen Yalin

Assignees

中国电子科技集团有限公司电子科学研究院
北京理工大学

Dates

Publication Date: 20260512
Application Date: 20251216

Claims (10)

1. An MSE parameter aggregation method for air computing auxiliary federal learning is suitable for a mobile edge computing network consisting of a base station and a plurality of mobile devices, and is characterized in that each iteration comprises the following steps: S1, each mobile device receives the current global parameters of a base station, selects partial data from a local data set and calculates a local gradient; S2, each mobile device uploads the gradient to the base station according to the amplification coefficient of the mobile device, and the base station recovers an aggregate gradient signal according to the amplification coefficient of the mobile device after receiving the superimposed signal; s3, the base station uses the recovered aggregation gradient signal to update the global parameter according to a preset step length; S4, the base station performs the MSE minimization between the recovered signal and the ideal aggregation gradient through two layers of optimization; and S5, the base station distributes the optimized and updated parameters and the new global parameters to the mobile equipment, and enters the next iteration until the training converges.
2. The method of claim 1, wherein the step of determining the position of the substrate comprises, In step S4, the two-layer optimization specifically comprises the steps of fixing the amplification factor of the base station, classifying and optimizing the amplification factor and the data quantity weight of the mobile equipment according to the characteristics of the mobile equipment, and optimizing the amplification factor of the base station to obtain the updated amplification factor and the data training quantity of the mobile equipment.
3. The method of claim 2, wherein the step of determining the position of the substrate comprises, The characteristic classification standard of the mobile equipment is that whether the product of the maximum amplification coefficient of the equipment, the channel coefficient and the base station amplification coefficient is higher than the data volume weight upper limit of the mobile equipment or not.
4. The method of claim 2, wherein the step of determining the position of the substrate comprises, The data volume weight refers to the proportion of the data volume of the single mobile device participating in training to the total data volume of all the mobile devices participating in training, and the weight sets an upper limit threshold value.
5. The method of claim 1, wherein the step of determining the position of the substrate comprises, In step S4, the ideal aggregation gradient is the result of weighted averaging of the local gradients of the respective mobile devices by the respective data amounts.
6. The method of claim 1, wherein the step of determining the position of the substrate comprises, In step S1, a portion of data, specifically a selected amount less than or equal to the size of the own data set and a total amount greater than or equal to the training minimum total data amount, is selected from the local data set.
7. The method of claim 1, wherein the step of determining the position of the substrate comprises, In step S2, the amplification factor of the mobile device sets an upper threshold, and the value of the amplification factor of the base station is greater than 0.
8. The method of claim 1, wherein the step of determining the position of the substrate comprises, In step S2, when the base station recovers the aggregate gradient signal, the interference influence of noise on the superimposed signal during transmission needs to be eliminated.
9. The method of claim 1, wherein the step of determining the position of the substrate comprises, The preset step length is a preset fixed value or a dynamic value dynamically adjusted according to the training loss of each iteration.
10. An air-computing federally-learned MSE parameter aggregation system for implementing the method of any one of claims 1 to 9, the system comprising a base station and a plurality of mobile devices; The base station comprises a parameter distribution module, a signal recovery module, a global updating module, an MSE optimization module and a data storage module, wherein: the parameter distribution module is configured to synchronously send global model parameters of the current round to all the mobile devices, and optimize the updated amplification coefficient and training data quantity configuration parameters; The signal recovery module is configured to receive the superimposed gradient signals uploaded by the plurality of mobile devices, recover the aggregate gradient signals based on the amplification coefficients of the base stations, and eliminate interference of transmission noise on the signals; the global updating module is configured to update global model parameters according to a preset step length according to the restored aggregation gradient signal; The MSE optimization module is configured to execute a two-layer optimization algorithm to respectively optimize the amplification factor of the mobile device, the data volume weight and the amplification factor of the base station, and minimize the mean square error of the recovery signal and the ideal aggregation gradient; The data storage module is configured to store key data such as global model parameters, channel coefficients of each mobile device, amplification coefficient threshold values, training data quantity threshold values and the like; The mobile device comprises a parameter receiving module, a gradient calculating module, a signal uploading module and a local storage module, wherein: the parameter receiving module is configured to receive global model parameters, optimized amplification factors and training data volume configuration parameters issued by the base station; the gradient calculation module is configured to select partial data meeting the requirements from the local data set according to the received global parameters, and calculate the local gradient; The signal uploading module is configured to upload the local gradient signal obtained by calculation to the base station according to the amplification coefficient of the signal uploading module; The local storage module is configured to store data such as a local data set, received global parameters, self amplification factors, training data volume configuration and the like.

Description

MSE parameter aggregation method and system for air computing assisted federal learning Technical Field The invention relates to the technical field of federal learning of artificial intelligence, in particular to a method and a system for aggregating MSE parameters of federal learning assisted by aerial computation. Background With the rapid development of mobile internet and internet of things technology, mass data generated from mobile devices such as smart phones and sensors are exponentially increased, and a data base is provided for machine learning applications such as disaster early warning based on digital twinning and health monitoring based on wearable devices. However, conventional centralized machine learning requires uploading local raw data of the device to a central server, which not only consumes a large amount of wireless communication resources, but also presents a significant risk of user privacy disclosure. To address privacy and communication resource issues, federal learning (FEDERATED LEARNING, FL) has emerged as a distributed model training architecture. Under the coordination of a Base Station (BS), a plurality of mobile devices realize cooperative training through iterative exchange model parameters without uploading original data. However, the federal learning parameter iteration involves frequent multiple access communication, and in a scenario where the number of devices is large and the spectrum resources are limited, there is a bottleneck that the spectrum efficiency is low and the energy consumption is high. The Over-the-Air computing (Over-the-Air computing) technology supports a plurality of mobile devices to transmit parameter signals to a base station simultaneously by utilizing superposition characteristics of channel signals, and can directly realize summation of data on communication signals, so that frequency spectrum efficiency of parameter aggregation is greatly improved, and communication bottleneck is broken through. However, in practical applications, due to heterogeneous channel fading and transmission noise, a deviation between the base station received signal and the ideal aggregate signal is generated, which is generally represented by Mean Square Error (MSE), thereby reducing the accuracy of the aggregation. In addition, the prior art only optimizes the signal amplification factor, does not consider the influence of the local training data quantity of the equipment on the MSE, easily causes overfitting and overload calculation, and is difficult to meet the requirement of high-precision training. Disclosure of Invention The invention aims to provide an MSE parameter aggregation method and system for air computing assisted federal learning, which are used for solving the problems of insufficient MSE reduction and large equipment computing load caused by optimizing only a single amplification factor in the prior art, and minimizing Mean Square Error (MSE), reducing computing load and controlling computing complexity through collaborative optimization. In order to achieve the above purpose, the present invention provides the following technical solutions: According to an aspect of the present invention, there is provided a method for aggregating MSE parameters for air-computing assisted federal learning, adapted to a mobile edge computing network consisting of a base station and a plurality of mobile devices, by training a model through multiple iterations, each iteration comprising the steps of: S1, each mobile device receives the current global parameters of a base station, selects partial data from a local data set and calculates a local gradient; S2, each mobile device uploads the gradient to the base station according to the amplification coefficient of the mobile device, and the base station recovers an aggregate gradient signal according to the amplification coefficient of the mobile device after receiving the superimposed signal; s3, the base station uses the recovered aggregation gradient signal to update the global parameter according to a preset step length; S4, the base station performs the MSE minimization between the recovered signal and the ideal aggregation gradient through two layers of optimization; and S5, the base station distributes the optimized and updated parameters and the new global parameters to the mobile equipment, and enters the next iteration until the training converges. According to one embodiment of the invention, in the step S4, the two-layer optimization specifically comprises the steps of fixing the amplification factor of the base station, classifying and optimizing the amplification factor and the data quantity weight of the mobile equipment according to the characteristics of the mobile equipment, and optimizing the amplification factor of the base station to obtain the updated amplification factor and the data training quantity of the mobile equipment. According to one embodiment of the invention, the characteristic clas