Research On Nonlinear Equalization Algorithm Of Coherent Optical Transmission System

The current development of Internet technology is becoming more and more vigorous.Emerging technologies such as 5G mobile communication,artificial intelligence cloud computing,and Internet of Things technology are also in full swing.These new technologies has spawned countless new network services,such as live webcast,4K ultra-high-definition television,navigation,and autonomous driving.This is a huge challenge for the fiber-optic backbone network that bears the main Internet traffic.In view of this,the future evolution of optical communication systems is undoubtedly oriented by large capacity,high speed and long distance.High-order modulation format is one of the main approaches to increase the optical communication rate,but it is sensitive to phase fluctuation.In addition,dense wavelength division multiplexing(DWDM)is an another method to increase bandwidth,but it also limited by technical challenges.The most difficult problem in long-distance optical communication is the nonlinear phase noise accumulated by the fiber Kerr effect,which will distort the optical signal.The nonlinear effects in DWDM systems is more serious,which greatly restricts the transmission distance of the system.Therefore,compensation for the non-linear damage is a difficult problem that must be faced in the development of the backbone optical network.Presently,the main method to combat non-linear effects is the digital backpropagation algorithm,which essentially implements a virtual fiber back propagation of the received impaired signal.However,the extremely high computational complexity makes it difficult to be applied it in real-time high-rate optical transmission system.The fast-evolving artificial intelligence(AI)algorithms in the past few years have been introduced into optical communications to solve various problems.This thesis mainly studies the application of AI algorithms in the nonlinear noise equalization.We explore equalization income improved by deep neural networks(DNN).At the same time,a single-layer functional-link neural network(FLNN)with excellent performance is proposed.The proposed FLNN can be regarded as a flattened DNN and can be fast trained by finding the Moore-Penrose Inverse.The research shows that FLNN is superior to the improved DNN equalizer in terms of equalization performance and processes a lower complexity.Aiming at the high complexity of non-linear effects inWDM systems,the thesis proposes a joint equalization method,which is generally applicable to various neural networks based equalization in WDM systems.We also employ transfer learning(TL)to train deep neural network equalizer,which greatly reduces the training overhead of the equalizer in the WDM system.The equalization by neural networks based on joint equalization and TL is expected to become an optional solution to counter nonlinear effects in future applications.