The Research Of Mobile Network Status Based On Multidimensional Association

In Recent years,the architecture and technology of mobile network are developing with the rapid increase of mobile users,a variety of terminal access and the explosion of data service.The increase in the size and complexity of current cellular networks is complicating their operation and maintenance tasks.Therefore,in order to reduce the labor costs in network operation and maintenance,intelligent network operation and maintenance becomes the necessary trend in the future to improve the performance of network.In this paper,we discuss two key problems with respect to the analysis of mobile network status in intelligent network operation and maintenance based on association analysis.The problems include the anomaly recognition of mobile network and traffic prediction,which are introduced in detail as following.Firstly,for the problem of detecting anomaly of KPI in mobile network,we introduce the density-based clustering algorithm,and propose a novel kernel density-based local outlier factor(KLOF)to assign a degree of being an outlier to each object.Comparing with statistical models,our algorithm can solve the outliers in training data,which makes our model more stable.In addition,this algorithm calculates the weight for each neighbor point by kernel density distance,which overcomes the defects caused by characteristics of mobile networks,such as the direction of anomaly,the sparsity of high-performance points and different variation ranges of different parameters.The experiment also proves that KLOF can find outliers efficiently and the accuracy of detection is improved.Secondly,for the problem of automatic root cause analysis in mobile network,a two-layer clustering framework of automatic root cause analysis based on unsupervised techniques is designed.At first,this framework introduces an iterative hybrid algorithm with self-organizing map(SOM)and K-mediods to cluster the anomaly automatically without labels.The categories are distinguished by the distribution of parameters,and they are used to the next step of feature selection.And then,we propose a feature selection method based on KS test and mutual information entropy,combing with the physical properties of KPI,to select the KPI parameters related to each type of anomaly,which also utilize the difference of parameters distribution between normal class and abnormal class.Further,we use clustering algorithm to classify the selected KPIs into different classes,which represent different root causes.At the same time,the classes of anomaly and corresponding root cause form the expert knowledge to assist the root cause analysis online.The experiments show that our framework can solve the problem of sparsity of anomaly and correspoding root causes.Lastly,for the problem of fine-grained traffic prediction in mobile network,we introduce the method of multi-dimension discrete traffic prediction.Based on data fusion,a two-layer local correlation based data-driven prediction method is proposed to improve the accuracy of fine-grained traffic burst prediciton.Comparing with traditional traffic prediction based on time series,our method considers the impact of external features except traffic itself.In the first phrase of determining the features related to traffic,we design a framework with multiple independent binary classifiers to disscuss features with local correlation in different traffic states.Each binary classifier is built to find the related features corresponding to specific traffic states.And then,in the phrase of determining the prediction function,we design an ensemble framework based on transfer learning.This framework is combined by multiple independent predictors,and each predictor is the prediction model for a traffic state.When we need predict one of these states,the left models of other states are used to assist our prediction,which make the final result obtianed by voting.The prediction models in dfferent time segments are various,then during this procedure,we consider the impact of user behavior,which repsepect the periodicity in adjacent hours instead of the same hour.Therefore,we design a hierarchical clustering to merge adjacent and similar hours based on their distribution of traffic.And the prediction model is built for each time segments.The experiments prove that our data-driven method can improve the accuracy of burst traffic prediction,which lead to the increasement of precision in traffic prediction.In this dissertation,we disscuss the application of unsupervised learning,supervised learning,transfer learning and some other multidimensional association analysis methods in the analysis of mobile network status.Through the data collected in current cellular network,we analyze the characteristics of data in mobile network and explore the application performance of above methods.Therefore,our work provides a feasible research methods for itelligent mobile network operation and maintenance.