Design And Implementation Of Network Data Processing System For B5G

With the rapid development and widespread deployment of 5G mobile communication technology,a large amount of mobile communication data has been generated,which requires telecom operators to have strong data collection,processing and analysis capabilities.At the same time,artificial intelligence technology is also developing rapidly.The integration of artificial intelligence technology can realize intelligent management of 5G mobile communication networks and improve network service quality.3GPP introduces the Network Data Analytics Function(NWDAF)into the 5G core network.As the engine of the core network "AI+BIG data",it is responsible for the collection and intelligent analysis of network big data,and studies the B5G-oriented network data processing system.is of great significance.Aiming at the problem that the operator’s data exists in the form of isolated islands and it is difficult to train a reasonable model due to privacy protection considerations,this paper combines deep learning with federated learning and proposes a method for predicting the CPU usage of DeepAR(Deep Autoregressive Recurrent Neural Network)network elements based on federated learning,breaking the island of data,solve the short board of transmission,and achieve the prediction effect close to the centralized method.The paper designs and develops a B5G-oriented network data processing system,which realizes the closed-loop automation process of core network data collection,data storage,data analysis,data prediction and result feedback,and meets the functional and performance requirements of 3 GPP specifications.In addition,the elastic scaling module has been further developed,which can effectively combine the prediction algorithm with the Kubernetes system.The specific work of the paper is as follows:(1)A method for predicting the CPU usage of DeepAR network elements based on federated learning is proposed.In the feature engineering,the holiday feature covariates are designed,and a reasonable neural network structure and related hyperparameters are designed for the scenario of CPU usage prediction.Then I proposed a general federated learning algorithm framework and federated learning process.Experiments have verified that the centralized DeepAR model is better than the ARIMA model and LSTM model,and that the federated learning model has a faster convergence speed than the centralized model when the prediction accuracy is guaranteed.(2)Design and implement a B5G-oriented network data processing system,in which the data collection and coordination module realizes the collection of three types of data:network element performance index data,log data and event data;the data storage module realizes a standardized access interface to the database;The data analysis model supports the event analysis and prediction of "network element load level information" and the event analysis of "slice load level information";the model training module realizes offline training and online service;the elastic scaling module realizes a custom control The controller dynamically adjusts the number of core network element copies in the Kubernetes cluster based on the prediction algorithm.