Data Driven Analysis And Prediction For Wireless Network Performance

With the development of wireless technology and the increase of mobile users,the focus of operators has shifted from wireless network construction to network operation and maintenance.Operators are eager to know real-time performance of networks and experience of users.Widely used big data analysis methods and machine learning algorithms can be used as a solution.However,there are some challenges in intelligent network management,especially in data analysis of mobile network,such as accuracy of prediction,acquisition of user experience and timeliness of trouble shooting.At present,5G commercialization was gradually carried out.5G requires high-frequency spectral efficiency,large capacity and low delay.Among them,user experience rate,connection density and delay are three basic performance indicators.5G network greatly improves the efficiency of network deployment and operation.At the same time,in the face of diversified scenarios and extremely differentiated performance requirements,5G oriented network management is difficult to deal with all problems on the basis of manual or single solution as before.With the rise and maturity of big data,machine learning and other methods,these methods are gradually used in wireless network management.In 4G network,concept of intelligent network management has been widely concerned and recognized.For 5G with more complex scenarios and higher performance requirements,network management needs to consider balance of network performance,user experience and network energy consumption.Intelligent network management is essential.Among them,data-driven wireless network performance analysis is particularly important.This thesis takes data-driven as the main line,and data used in research was collected by the operator at base station.In the second,third and fourth chapters,different data processing and modeling methods are adopted to analyze the data,find information that plays a decisive role in the wireless network traffic from the data,and carry out data analysis and prediction on this basis.The fifth and sixth chapters focus on relationship between wireless indicators through method of data mining.Compared with traditional methods,data-driven methods can analyze data more deeply and mine deep information in the data.The performance of wireless network can be obtained through the statistical quantification of various key performance indicator(KPI)and key quality indicator(KQI)parameters.Firstly,this thesis carried research on prediction by taking downlink traffic as one of the KPI indicators;For KQI data,it is often inferred from KPI data,so in the last two parts of the research content,the relationship between KPI and KQI is deeply studied.For data-driven network performance analysis,this thesis studies from the aspects of traffic data temporal-spatial modeling,network performance prediction,user experience estimation and network fault diagnosis.Combined with statistical theory,causality test,graph theory and machine learning,several solutions for these aspects are proposed.Main contributions of the thesis are as follows:1.Based on Granger causality test,the correlation between grid level regional traffic in urban wireless network is studied.On this basis,the spatio-temporal traffic model is combined with multi variable long short term memory(LSTM)neural network to predict traffic of urban wireless network.Tthe prediction algorithm is applied to real data of different urban scenes to verify prediction accuracy and universal adaptability.2.Study the causal verification method of bi-directional traffic based on crowd flow model,and put forward a further solution to the prediction of wireless traffic based on graph convolutional network(GCN),gated recurrent unit(GRU)and transfer entropy theory.The final verification results show that this method significantly improves the accuracy of cellular traffic prediction.3.Under the framework of time series decomposition and integration,cellular traffic are deeply mined,and a method to extract trend component,periodic component and basic component from complex traffic time series is proposed.On this basis,combined with GCN,a spatial temporal time series prediction based on decomposition and integration system with causal structure learning(DIC-ST)is proposed.This method can not only ensure the accuracy of traffic prediction,but also play an important role in network optimization,base station location and energy consumption.4.This paper proposes causal structure learning for KPI and KQI data,embeds the correlation between multiple KPIs and KQI into the prediction model,and designs a prediction method for user experience.5.Based on kernel theory and naive Bayesian classifier,a method of combining user experience estimation with network fault diagnosis is proposed,and the evaluation standard of cell level wireless sensing ability is given.The results show that this method can accurately estimate the user experience and diagnose outliers in time.This method can improve the efficiency of network management.From the prediction of wireless network performance to the estimation of user experience,and then to the diagnosis of network faults.This thesis completes the research of three main steps in the data analysis of network performance.For the problem of wireless network traffic prediction,we give a cellular traffic prediction method from the perspectives of grid level traffic analysis,base station level traffic analysis,traffic mode analysis based on crowd flow model and time series decomposition and integration.Aiming at the research of user experience in wireless networks,a prediction method suitable for user experience is proposed.Aiming at the research of network fault diagnosis,this thesis proposes a method of combining user experience with fault diagnosis,in order to improve the timeliness of fault diagnosis.These methods can be used for intelligent network management.This thesis combs increasingly complex application environment and gives the key solutions to the increasingly complex application environment.