Research On Base Station Service Behaviors Based On Cellular Network Analysis

With the rapid development of wireless communication networks,4G and 5G networks are gradually popularized,and the demand for services in cellular mobile communication networks is exploding.Nowadays,users'demand for communication networks has long expanded from traditional voice and message requirements to social,video,travel and e-commerce in daily life,which has raised the requirements to wireless network architecture and resource utilization efficiency.In order to meet the growing needs of users,how to deeply understand the service usage patterns of base stations in cellular networks becomes an urgent problem to solve.It's crucial to optimize the cellular network architecture,develop reliable and effective resource scheduling strategies and caching strategies,and improve existing network deployments.At present,a lot of researches and optimization strategies for base station deployment and service scheduling cache for cellular networks have been carried out at home and abroad,but the following problems still exist:First,the research in cellular networks revolves around the perspective of base stations and users.There is no ground truth value of labels,and how to effectively take advantage of the algorithms of unsupervised learning for statistical analysis has become the key point and difficulty of researches.Secondly,cellular networks cover hundreds of millions of users,and individual behaviors vary widely.Research on individual behaviors cannot provide constructive advice for network deployment intuitively and effectively.This thesis analyzes the massive measured traffic data of cellular network,combines the external geographical location information,utilizes the unsupervised learning algorithms to conduct in-depth analysis and data mining of the base station traffic patterns,and based on this,improves the forecasting method of traffic throughput.The main research work content and innovations of the thesis include the following three parts:First of all,combines the open source interface of a mapping software to obtain the location information-points of interest(Pols)in the city,the location points in the geographic location information have a two-level granularity classification label.The coverage of the base station is divided by the Thiessen polygon,which is called Voronoi,and the Pols feature vector of the base station is statistically derived.The term frequency-inverse document frequency(TF-IDF)normalization method,which frequently utilized in natural language processing,is applied to process the feature vectors of the base station thus to avoid the impact of the coverage area size and the unevenness of the distribution of points of interest.The unsupervised learning algorithm is used to clustering the base stations,thereby mapping the labels of points of interest to the categories of the base station.The method can predict the base station category when there is no traffic data and estimate the possible traffic patterns.Based on the tagged base station data,the horizontal and vertical comparison analysis is used to analyze and mine the traffic patterns of different types of base stations,and typical traffic patterns are obtained.These modes can be used to optimize the scheduling and caching strategies of traffic resources in cellular networks.Secondly,introduces the social network analysis method to study the cellular network data,and proposes a method to construct the base station relationship network model,which is to construct the base station relational network by defining the base station as the node and the traffic flow similarity between base stations as the weight of the side.Based on this network,the Planar Maximally Filtered Graph algorithm is used to filter the relationship network,and the main relationships in the network are extracted.Then,the Louvain algorithm is used to perform community division and scene recognition on the base station.The feature vector obtained by the time series traffic data can be utilized to mine the traffic flow mode.At the same time,after superimposing to the service category dimension,the feature vector can be extended to the two-dimensional matrix to obtain the usage of different types of application services of the base station.Finally,we combine the location information of urban points of interest mentioned in the previous chapter to verify our methodology.The results show that using the social network analysis method to classify the base stations,typical traffic scenarios are accurately identified,simultaneously the traffic period patterns on the time series and the difference in preferences on the application dimension are distinguished.Finally,based on the results of scene recognition using the social network analysis method,this thesis proposes a method for base station traffic prediction.For a single base station,scene identification is performed first,and the base station can be divided into categories of multiple traffic patterns.Considering that the base station in the same category has similarity in traffic consuming,the base station of the same category is combined with its own time series traffic sequence as its characteristics,and the stacked auto-encoder(SAE)is used for feature compression,and the compressed feature information is input into the Long Short-Term Memory(LSTM)model for training.Then a traffic flow prediction model of the base station is obtained.The method combines the community association of the base stations,effectively amplifies the feature dimension and information amount of the base station,and compared with the traditional time series prediction methods support vector regression(SVR)and Autoregressive Integrated Moving Average(ARIMA)models,it shows better performance.In addition,the average error can be significantly reduced compared to the feature extraction based solely on geographic location.