Research And Application Of The Self-learning Technology For KPI Alarming Threshold In Communication Networks

Optical cable communication,cable communication,radio wave and other communication methods escort people's daily communication with the rapid development of communication network technology in our country.At the same time,the network management software provides a large number of KPI(key performance indicator)data,which plays an important role in network anomaly detection,fault diagnosis and so on to remind and urge the staff to adjust the network strategy.The abnormal fluctuation of KPI data is monitored by alarm threshold,which is often set manually,simply and roughly,not targeted.And the alarm effect is not ideal.In order to solve the problem of self-regulation of KPI alarm threshold in communication network,make it alarm accurately and reach the engineering operation standard.The research content and related work of this paper are as follows:(1)By analyzing the characteristics of KPI time series data,a two-stage computing architecture model is proposed:LTSO(Learning Threshold Parameters by Semi-supervised Learning and Optimization),which is based on semi-supervised learning and multi-objective optimization.This method includes:KPI data collection and processing,network element(e.g.base station)similarity module analysis,model threshold/parameter initialization,KPI sample aggregation,semi-supervised learning label expansion,multi-objective optimization threshold by self-learning and thresholds sharing.(2)The threshold sharing of similarity network element/base station can save the labor cost.Therefore,a hierarchical clustering feature selection method for different types of KPI threshold data is proposed.According to the threshold value of RSRP,we choose the strong correlation features such as the height of the base station,the transmission power of the base station,and the proportion of all kinds of buildings,trees,water objects in the geographical scene.For the Traffic threshold,the base station frequency point configuration,base station capacity,base station traffic and user related behavior data are taken as clustering features.On this basis,the network element/base station types are accurately classified from the two aspects of essential features and historical timing features.(3)In order to solve the problem of false negatives and false positives sample tags in the process of threshold optimization,a semi-supervised direct push tag expansion method based on tree model-based learner is proposed.Through the construction of statistical characteristics,entropy characteristics and segmentation characteristics in KPI time series data,we can fully describe the KPI indicators,and use tree model to carry out high-order combination in the feature domain,and construct an integrated base learner with high benchmark accuracy.A simple split semi supervised voting method is proposed to predict the final tag with high accuracy and expand the semi supervised learning samples.(4)A multi-objective threshold optimization algorithm based on NSGA-? is designed and implemented.The number and area of false positives and false negatives are taken as the optimization objectives,and the fitness functions in various application scenarios are constructed.Through the continuous inputting of KPI time series data segment,the optimization iteration is carried out to calculate the final model alarm threshold.According to the above research content,the KPI threshold monitoring software is designed and implemented.The effectiveness of the above work is verified by the KPI data of Nanjing area,and the goal of threshold self-learning adjustment monitoring is achieved.