Research On Fault Root Cause Analysis On Topology Alarm Text

Root cause of failure localization refers to locating the root cause of failure of current services quickly through abnormal alarm return information,and how to locate accurately and quickly is crucial for maintaining and managing the network system environment.This paper designs a series of supervised and unsupervised algorithms based on labeled alarm event data and node topology data,and proposes a three-stage implementation of fault root cause location,which is finally implemented on the ground.In this paper,we firstly implement fault root cause time period detection based on classification algorithms.In the first stage,each time period of the original data set is intercepted by sliding time window construction,and the data set is constructed while data enhancement is achieved.The first step is to extract features based on L1 regularized logistic regression,reduce the features from 101 dimensions to 42 dimensions,and then construct weighted metrics to measure the merits of the classification algorithms.Finally,the SVM algorithm has the best classification result,which reflects the unique advantage of SVM algorithm in solving the small sample,non-linear and high-dimensional pattern recognition like fault root cause analysis.For predicting the time period in which the root cause exists,this paper next determines the alarm type of the fault root cause,and then determines the alarm node of the fault root cause.In order to determine the structure of Bayesian network graph,this paper first extracts the relationship graph of alarm types based on association rule mining,and at this stage,constructs data sets based on alarm type features from both temporal and spatial dimensions,and then performs frequent pattern mining based on Apriori algorithm in temporal dimension and Prefix Span algorithm in spatial dimension.Then,Apriori algorithm is used for frequent pattern mining in temporal dimension,and Prefix Span algorithm is used for sequential pattern mining in spatial dimension.The Bayesian network results show that the Bayesian network using the BDeu distribution as a parameter to learn the prior distribution has the highest accuracy of78.2%,and the accuracy of the model with a reasonable graph structure set manually is on average better than that of the structure learning,and the graph structure is more streamlined and the model is more efficient.Based on this,this paper improves the prediction results by voting to increase the accuracy to 85%.The root cause type of alarms is known,and the suspected root cause nodes are obtained by reverse search.Combined with the conclusion of the feature analysis:88% of the root causes fall at nodes with 2 or 3 types of alarms and on average only half of the nodes generate alarms in each time period,the index is constructed to rank the possibility of the suspected nodes being the root cause,and the node with the highest possibility is selected as the root cause of the failure with 89% accuracy.