CN-122022191-A - Ext> Industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> methodext> andext> deviceext> basedext> onext> federalext> learningext>

CN122022191ACN 122022191 ACN122022191 ACN 122022191ACN-122022191-A

Abstract

The application provides an industrial Internet 5G-A communication evaluation model training method and device based on federal learning, wherein the method comprises the following steps: the server determines a client identification set participating in training; the server distributes the first initial artificial intelligent model to the client; the client is used for training the first initial artificial intelligent model according to the local simulation communication data, and uploading the obtained model parameters to the server; the server side aggregates the received model parameters and updates the first initial artificial intelligent model; the server iterates until the parameters are converged to obtain a second initial artificial intelligent model and distributes the second initial artificial intelligent model to all clients; the client is used for training the second initial artificial intelligent model according to the local real communication data to obtain a corresponding exclusive artificial intelligent model. And the server side aggregates the received model parameters, replaces original data uploading by parameter uploading, realizes multi-client knowledge sharing on the premise of protecting data privacy, and improves model training efficiency and generalization capability.

Inventors

WU NAIXING
CHEN YU
LENG WEI
LIN WENLIANG
DENG ZHONGLIANG
ZHANG LI
ZHOU ZHONGXIANG
HE TONG
Zou Jingbo
MIAO YAN

Assignees

中国联合网络通信有限公司深圳市分公司
北京邮电大学
郑州航空工业管理学院

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. Ext> theext> industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> methodext> basedext> onext> federalext> learningext> isext> characterizedext> byext> comprisingext> aext> serverext> andext> atext> leastext> oneext> clientext>,ext> whereinext> theext> serverext> comprisesext> aext> firstext> initialext> artificialext> intelligentext> modelext>;ext> The method comprises the following steps: S110, the server determines a client identification set participating in training; s120, the server distributes the first initial artificial intelligent model to the corresponding client according to the client identification set, wherein the client is used for receiving the first initial artificial intelligent model, acquiring local simulation communication data, training the first initial artificial intelligent model according to the local simulation communication data, obtaining model parameters, and uploading the model parameters to the server; S130, when the model parameters received by the server side reach a preset value, the server side aggregates the received model parameters and updates the first initial artificial intelligent model according to the aggregated model parameters; And S140, repeating the steps S110-S130 by the server to enable the parameters of the first initial artificial intelligent model to converge, obtaining a second initial artificial intelligent model, and distributing the second initial artificial intelligent model to all clients, wherein the clients are further used for acquiring local real communication data when receiving the second initial artificial intelligent model, and training the second initial artificial intelligent model according to the local real communication data to obtain a corresponding exclusive artificial intelligent model.
2. The method of claim 1, wherein the model parameters include communication index attention weights, training update information, training durations, training data sizes, and training accuracy.
3. The method of claim 2, wherein the step of the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters comprises: the server side carries out weighted aggregation on the received model parameters according to a federal average algorithm to obtain aggregated global model parameters; and the server updates the first initial artificial intelligent model according to the global model parameters.
4. A method according to claim 3, wherein the step of the server side performing weighted aggregation on the received model parameters according to a federal averaging algorithm to obtain aggregated global model parameters comprises: The calculation formula for weighting and aggregating the received model parameters according to the federal average algorithm is as follows: Wherein, the Representing the parameters of the global model after aggregation, Representing the total number of clients participating in the present round of training, Represent the first The amount of local training data for the individual clients, Representing the total data volume of all clients participating in the training, Represent the first Local model parameters uploaded by the individual clients.
5. The method of claim 1, wherein the step of the server aggregating the received model parameters and updating the first initial artificial intelligence model based on the aggregated model parameters comprises: training at a first learning rate at a first goal stage of federal training; and training at a second learning rate in a second target stage of federal training, wherein the second learning rate is lower than the first learning rate, and the first learning rate or the second learning rate is determined by scaling the basic learning rate by a preset adjustment factor.
6. Ext> theext> industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> methodext> basedext> onext> federalext> learningext> isext> characterizedext> byext> comprisingext> aext> serverext> andext> atext> leastext> oneext> clientext>,ext> whereinext> theext> serverext> comprisesext> aext> firstext> initialext> artificialext> intelligentext> modelext>,ext> aext> clientext> identificationext> setext> usedext> forext> determiningext> participationext> inext> trainingext> andext> distributingext> theext> firstext> initialext> artificialext> intelligentext> modelext> toext> theext> correspondingext> clientext> accordingext> toext> theext> clientext> identificationext> setext>;ext> The method comprises the following steps: When the client receives the first initial artificial intelligent model, local simulation communication data are obtained, and training is carried out on the first initial artificial intelligent model according to the local simulation communication data to obtain model parameters; The client side uploads the model parameters to the server side, wherein the server side is used for aggregating the received model parameters when the received model parameters reach a preset value, updating the first initial artificial intelligent model according to the aggregated model parameters, and circularly iterating until the parameters are converged to obtain a second initial artificial intelligent model; And when the client receives the second initial artificial intelligent model, acquiring local real communication data, and training the second initial artificial intelligent model according to the local real communication data to obtain a corresponding exclusive artificial intelligent model.
7. The method of claim 6, wherein the model parameters include communication index attention weights, training update information, training durations, training data sizes, and training accuracy, and wherein the step of training the first initial artificial intelligence model based on the local simulation communication data to obtain model parameters comprises: The client extracts local association features among communication indexes of the local simulation communication data through the first initial artificial intelligent model to obtain a corresponding feature map; The client determines a time sequence feature vector of the communication index changing along with time according to the feature map; and the client determines the communication index attention weight according to the time sequence feature vector.
8. The method of claim 6, wherein the step of training the second initial artificial intelligence model based on the local real communication data to obtain a corresponding proprietary artificial intelligence model comprises: the client inputs the local real communication data into the exclusive artificial intelligent model to obtain exclusive attention weight of a communication index corresponding to the client, wherein the local real communication data comprises the index real weight; The client determines KL divergence between the exclusive attention weight of the communication index and the real weight of the index; And if the KL divergence is smaller than a preset threshold, the client outputs the exclusive artificial intelligent model and the corresponding exclusive attention weight of the communication index.
9. Ext> theext> industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> deviceext> basedext> onext> federalext> learningext> isext> characterizedext> byext> comprisingext> aext> serverext> andext> atext> leastext> oneext> clientext>,ext> whereinext> theext> serverext> comprisesext> aext> firstext> initialext> artificialext> intelligentext> modelext>;ext> The device comprises: A client module is selected and used for determining a client identification set participating in training by the server; The system comprises a client, a model distribution module, a model parameter acquisition module and a server, wherein the client is used for receiving a first initial artificial intelligent model and acquiring local simulation communication data, training the first initial artificial intelligent model according to the local simulation communication data to obtain model parameters, and uploading the model parameters to the server; the parameter aggregation module is used for aggregating the received model parameters when the model parameters received by the server side reach a preset value, and updating the first initial artificial intelligent model according to the aggregated model parameters; The iterative training module is used for repeatedly triggering the client selecting module, the model distributing module and the parameter aggregation module by the server to enable the parameters of the first initial artificial intelligent model to converge, obtain a second initial artificial intelligent model and distribute the second initial artificial intelligent model to all clients, wherein the clients are further used for acquiring local real communication data when receiving the second initial artificial intelligent model, and training the second initial artificial intelligent model according to the local real communication data to obtain a corresponding exclusive artificial intelligent model.
10. Ext> aext> computerext> electronicext> deviceext> comprisingext> aext> processorext>,ext> aext> memoryext> andext> aext> computerext> programext> storedext> onext> theext> memoryext> andext> executableext> onext> theext> processorext>,ext> theext> computerext> programext> whenext> executedext> byext> theext> processorext> implementingext> theext> stepsext> ofext> aext> federallyext> learnedext> industrialext> internetext> 5ext> Gext> -ext> aext> communicationext> assessmentext> modelext> trainingext> methodext> accordingext> toext> anyext> oneext> ofext> claimsext> 1ext> toext> 8ext>.ext>

Description

Ext> Industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> methodext> andext> deviceext> basedext> onext> federalext> learningext> Technical Field Ext> theext> inventionext> relatesext> toext> theext> fieldext> ofext> federalext> learningext>,ext> inext> particularext> toext> anext> industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> methodext> andext> deviceext> basedext> onext> federalext> learningext>.ext> Background Ext> withext> theext> deepext> integrationext> ofext> industrialext> Internetext> andext> 5ext> Gext> -ext> Aext> andext> otherext> newext> generationext> communicationext> technologiesext>,ext> theext> performanceext> requirementsext> ofext> industrialext> scenesext> onext> aext> communicationext> systemext> areext> increasinglyext> diversifiedext> andext> refinedext>.ext> The traditional industrial Internet has complex and changeable application scenes, namely time-sensitive industries of mechanical manufacturing and energy-consumption-sensitive industries of light detection, and different industrial scenes have obvious differences on specific requirements of the same parameters in a communication system. Because of different object-oriented, the specific requirements for the same parameters in the communication system are obviously different, and the characteristic requirements of various industrial scenes are difficult to match by directly applying the general communication performance evaluation method. Ext> theext> existingext> evaluationext> systemext> hasext> theext> problemsext> ofext> singleext> indexext> dimensionext>,ext> lackext> ofext> quantitativeext> analysisext> standardext>,ext> poorext> sceneext> suitabilityext> andext> theext> likeext>,ext> andext> cannotext> accuratelyext> reflectext> theext> actualext> performanceext> ofext> theext> novelext> communicationext> technologyext> suchext> asext> 5ext> Gext> -ext> Aext> andext> theext> likeext> inext> aext> complexext> industrialext> environmentext>.ext> In addition, federal learning has the problem of conflict between communication cost and performance in an industrial scene, namely, although the improvement of the communication frequency between the edge node and the central server can enhance the synchronization effect of the global model and improve the model precision, the problem also means that the communication turn and the local training turn are higher, so that larger calculation resource consumption and communication expenditure are brought. In practical industrial deployment, the increase of local training rounds can result in higher equipment energy consumption and training time cost, and frequent model uploading and downloading can also significantly promote transmission energy consumption and network delay. Therefore, it is needed to construct a dynamic evaluation method with scene self-adaptation capability, which provides theoretical basis and technical support for intelligent evaluation and optimization of an industrial communication system, so as to promote accurate configuration and standardized management of communication resources and application of the communication resources in industrial Internet. Disclosure of Invention Ext> inext> viewext> ofext> theext> foregoingext>,ext> theext> presentext> applicationext> hasext> beenext> madeext> toext> provideext> aext> federallyext> learnedext> basedext> industrialext> internetext> 5ext> Gext> -ext> aext> communicationext> assessmentext> modelext> trainingext> methodext> andext> apparatusext> thatext> overcomesext> theext> foregoingext> orext> atext> leastext> partiallyext> solvesext> theext> foregoingext> problemsext>,ext> includingext>:ext> Ext> theext> industrialext> Internetext> 5ext> Gext> -ext> Aext> communicationext> evaluationext> modelext> trainingext> methodext> basedext> onext> federalext> learningext> relatesext> toext> aext> serverext> andext> atext> leastext> oneext> clientext>,ext> whereinext> theext> serverext> comprisesext> aext> firstext> initialext> artificialext> intelligentext> modelext>;ext> The method comprises the following steps: S110, the server determines a client identification set participating in training; s120, the server distributes the first initial artificial intelligent model to the corresponding client according to the client identification set, wherein the client is used for receiving the first initial artificial intelligent model, acquiring local simulation communication data, training the first initial artificial intelligent model according to the local simulation communication data, obtaining model parameters, and uploading the model parameters to the server; S130, when the model parameters received by the server side reach a preset value, the server side aggregates the received model parameters and updates the first initial artificial intelligent model according to the aggregated model parameters; And S140, repe