Research On Network Fault Prediction Based On Alarm Data Analysis In Telecommunication Network Management

Network fault management is the key to the network management system to ensure the stability and reliability of the entire communication network.Alarm correlation analysis and network fault prediction are two important research contents in current network fault management.Combining data mining and machine learning methods to analyze the potential relationships in the alarm data and track the occurrence of predicted failures has an important impact and significance on improving the quality of network services and promoting the intelligent transformation of the telecommunications industry.This thesis relies on enterprise cooperation project,researches on alarm data analysis and fault data prediction in telecommunication network management systems,and proposes a comprehensive system from alarm data correlation analysis to network fault prediction.The main work of the thesis is as follows:1)This thesis summarizes the concept,characteristics and basic processing steps of alarms and faults,describs the theory of related algorithms involved in this thesis,and finally analyzes the research results and application of alarm data correlation analysis and network fault prediction at home and abroad.2)Aiming at the problems of redundant alarms and overlapping alarms between different sub-networks in the entire network,a correlation analysis method is proposed to effectively analyze the alarm timing relationship and locate the root alarm.First,an improved K-Modes clustering algorithm based on the network topology is proposed to obtain alarm clusters with similar topology.Then,a dynamic sliding time window is designed and implemented to select the alarm sequence set and dynamically adapt to the different alarm data distribution in each cluster,to ensure that the associated alarms are contained in the same time window as much as possible.Taking into account the sequence of the alarm time,sequential pattern mining is used to mine the correlation between the alarms.Experiments have effectively verified the correlation between alarms,which can be used to guide the filtering of redundant and accompanying alarms,locate the root cause alarms,and finally realize the association between root cause alarms and faults.3)Furthermore,based on the correlation data of alarms and faults,a network fault prediction method based on multi-model integration is proposed.First,through statistics of associated data to observe the distribution of the data and the characteristics of each attribute,a ensemble model called FBCX_Ensemble model is built,integrating Factorization Machine(FM),Extreme Gradient Boosting(XGBoost),Bidirectional Long-Short Time Memory(BI-LSTM)and Convolutional Neural Networks(CNN),which can fully learn richer expressed information.In addition,this thesis proposes a head fault evaluation index based on the true and false positive rates.This index focuses on the fault prediction of the head position in the ROC curve,and is used to evaluate the accuracy of the positive samples when the rate of false positives is small in the model evaluation.The FBCX_Ensemble model is applied to the actual data set for alarm fault prediction and the performance is compared with the single model.The results show that the ensemble model has better fault recognition.