Research On Performance Monitoring And Resource Allocation For Large-scale Optical Networks Based On Machine Learning

The emergence of new technologies such as cloud computing,edge computing,Internet of Things,virtual reality,artificial intelligence and 5G,has led to an explosive increase in network traffic.Optical network as one of the most important infrastructures for high-speed data transmission always confronts with the development requirements of high bandwidth and low latency.With great progress of machine-learning(ML)technologies in recent years,it is imperative to improve intelligence in optical networks with ML.Therefore,this dissertation focuses on the theme of network performance monitoring and resource allocation with advanced ML technology to improve the utilization of optical network resources.The main works and innovations of this dissertation are as follows:1.A new feature vector called "link-adjacent channel" is introduced to characterize the nonlinear interference from the neighboring channels and is combined into the artificial neural network(ANN)algorithm to estimate the quality of transmission(QoT)of optical transport network(OTN).Our simulation shows that the estimation accuracy can be improved by 0.7 dB compared with the traditional representation method based on the end to end network characteristic parameters.To further reduce the time expense of the training process,we put forward and verify the extreme learning machine(ELM)algorithm as a ML method of low computational complexity,whose time expenxe is only one tenth of ANN's,with the same performance.It is also shown by simulation that,the ML-based QoT estimating tools are useful for detection and localization of soft failures on a per link level.2.As for the issue of resource allocation in OTN,a physical-layer-impairment(PLI)-based routing and wavelength assignment(RWA)scheme combining ML-based QoT prediction technology with the genetic algorithm(GA)is proposed and its advantage in reducing the blocking probability of optical networks over the traditional RWA algorithms is shown by simulation.We also present a soft failure aware resources allocation algorithm based on GA(SFA-GA).Simulation shows that,when soft failure exists in OTN,the proposed SFA-GA can support the highest traffic load among the relevant algorithms at any given blocking ratio.Typically,at the blocking ratio of 0.01,the SFA-GA allows a traffic load as high as 280 Erlangs,which is about 1.5 times of the commonly used Dijkstra routing plus first fit for spectrum assignment.3.For low-latency dynamic bandwidth allocation(DBA)algorithm,a novel predictive DBA method based on Long Short-Term Memory(LSTM)neural network is proposed to reduce the latency of uplink data transmission in optical access network.By predicting the number of packets arriving at the optical network unit buffer based on LSTM and allocating bandwidth in advance,the round-trip time delay in traditional DBAs can be eliminated for reducing network latency.Our study shows that LSTM has better performance than feed-forward neural network in terms of delay,jitter and packet loss ratio.For lower time cost in the training process,the online sequential extreme learning machine(OS-ELM)-based DBA algorithm,an ML algorithm based on non-iterative online learning,is proposed.Simulation proves that the OS-ELM algorithm has a very small training time,less than one thousandth of that for the LSTM algorithm.