Research On Artificial Intelligence For IT Operations Of SDN Network Anomaly Management

With the rapid development of computer technology,the difficulty of system operation and maintenance has gradually increased due to the massive network scale and data transmission.Software Defined Networking(SDN),a widely studied network architecture in recent years,effectively achieves centralized network management by separating the data and control layers.Maintaining the long-term stable operation of the system requires real-time monitoring and analysis of Key Performance Indicators(KPI).To achieve a more automated and intelligent operation and maintenance mode,Artificial Intelligence for IT Operations(AIOps)technology can be leveraged to further ensure system quality.This thesis focuses on the anomaly detection and root cause analysis in SDN network based on AIOps.The main contributions of this thesis are as follows:(1)A SDN intelligent operation and maintenance anomaly control system framework is proposed.AIOps technology is introduced into the traditional SDN network structure,and performance monitoring,anomaly detection,and root cause analysis are designed based on the original SDN system.The global control data plane detects anomalies in KPI data within a specified time window,locates the root cause after determining the anomaly,and determines the anomaly source.(2)SDN controller acquisition and monitoring functions are implemented.The SDN controller can obtain real-time KPI information from the data plane,grasp network-wide information in the control plane,and perform real-time log alarms based on the bandwidth utilization rate of each link.By connecting to an external My SQL database,data can be persistently stored as a source for anomaly detection and root cause analysis.(3)A generalized Bagging ensemble learning method is proposed.Traditional anomaly detection and deep learning methods may perform well in specific anomaly scenarios,but in actual environments,KPI anomalies have the characteristics of diversity and unknowns.Therefore,common KPI anomaly types in SDN networks are defined,and different individual learners are selected based on the idea of ensemble learning to obtain a binary judgment result of whether an anomaly exists.By integrating multiple types of anomaly detection algorithms,it can have good noise resistance and universality in different scenarios.(4)An improved root cause localization algorithm for SDN is proposed,which optimizes the potential score calculation and proposes a search strategy based on impact scores.Based on the traditional heuristic search root cause localization algorithm,the root cause analysis conceptual model in the SDN network is defined,and the fault propagation types in the SDN environment are analyzed to determine the derived KPI and multidimensional root cause selection.The customized distance potential score can alleviate the problem of prediction bias to a certain extent,and the search plan based on impact scores can solve the problem of exponential growth in space and improve the accuracy and timeliness of root cause localization.Finally,an experimental simulation environment is constructed using ONOS controller and Mininet custom topology,and functional and performance tests are conducted on performance monitoring,anomaly detection,and root cause analysis.