Research On Machine Learning Based Traffic Analysis Strategy And Its Applications In Optical Networks

With the rapid development of fifth-generation mobile communications(5G),big data,artificial intelligence,cloud computing,etc.,the explosive growth of information demand brought about by various new services has put forward higher requirements for optical networks.Software-defined optical network(SDON),with its flexible networking mode,high resource utilization rate and flexible service supply capability,is widely regarded by academia as the main solution for future optical networks.Due to the limitation of the fixed wavelength grid distribution mechanism,the traditional optical network transmission plane using WDM technology has become more and more incapable in the face of the flexible and changing business environment.The flexible optical network(EON)with high flexibility and high scalability has emerged as the times require Raw.In the face of profound changes in the control plane and transmission plane in the future optical network,it is particularly important to use technologies including machine learning to analyze optical network traffic and allocate resources efficiently.In this paper,through theoretical analysis and modeling and simulation methods,the problem of abnormal traffic detection in the control plane of optical networks,and the spectrum allocation problem combined with traffic prediction in flexible optical networks are studied in depth.This thesis first introduced the development process of intelligent optical networks,the basic concepts and infrastructure of software-defined optical networks.Secondly,it analyzes the basic principles and key technologies of elastic optical network.The paper discussed in depth the typical machine learning algorithms and their applications,focusing on the traffic analysis based on machine learning and its applications in optical networks.Aiming at the problems of optical network anomaly detection and intrusion recognition,this paper proposes an abnormal traffic detection(AFD)strategy based on machine learning combined with SDON.The AFD strategy is designed to detect point anomalies and sequence anomalies by designing a detection system structure.This strategy uses the isolated forest algorithm to detect point anomalies,while using the EWMA algorithm to detect sequence anomalies.Theoretical research and simulation results show that in a typical network topology,the average detection accuracy of the strategy proposed in the paper is 90% and 85% for point anomaly detection and sequence anomaly detection,respectively,and the comprehensive detection accuracy can reach 87%,which can effectively identify Abnormal flow.Aiming at the problem of spectrum allocation in time-varying elastic optical networks,the paper proposes to introduce and apply the traffic prediction model based on LSTM algorithm to spectrum allocation,thereby obtaining a spectrum reallocation strategy(TP-SA)combined with traffic prediction.This algorithm not only considers the traffic demand of the current time slot,but also considers the spectrum allocation of the previous time slot and the traffic forecast of the future time slot through a more reasonable and accurate traffic forecast,and expands the transmission path of the sub-carrier in advance for resource allocation.Using this TP-SA strategy can prevent unnecessary network reconstruction and save network operating costs.Theoretical analysis and simulation results show that,when the traffic volume reaches 600 Erl,the bandwidth blocking rate is only slightly higher by 6.2%、5.2% and 5.1% in the three typical network topologies of NSFNET,CERNET and USNET when the TP-SA strategy is compared with the full spectrum reallocation strategy,but the number of reconstructions can be reduced by 63.2%,60.6%,and 54.3%,respectively.