Study Of Machine Learning Assisted High-Speed Fiber Communication Theory And Technologies

Optical fiber transmission network has huge potential of applying machine learning(ML)technology since it is the infrastructure of information communication.In optical fiber transmission system,traditional analytical models cannot be used in complex links or large-scale dynamic optical networks since they are limited by the accuracy of complex system parameters acquisition and computational complexity.However,benefit from data-driven characterisitics,ML algorithms can realize the dynamic mapping of complex problems without acquiring specific system parameters or models.In the dissertation,we use the theoretical analysis and model construction 'knowledge'of optical fiber system to achieve the combined design of traditional digital signal processing(DSP)algorithms and ML algorithms.Besides,the 'data'obtained from experimental and simulation systems can optimize the algorithm performance.In a word,the propsed 'customized' algorithm,which based on'knowledge-driven' and 'data-driven' characteristics,can not only make full use of the advantages of ML algorithms in solving nonlinear problems and classification problems,but also take adavantage of the reliability and robustness of traditional DSP.In the dissertation,we focuse on the theme of 'Combining machine learning technology and traditional DSP algorithm to achieve the goals of high realibility,high transmission rate and intelligent optical fiber transmission systems.' Based on the theme,we explore the ML technology in the field of fiber transmission impairments compensatation and optical performance monitoring.The details and novelty are as follows:1.Machine learning assisted impairments compensation technology for optical fiber transmission systemA novel receiver side DSP structure and an equalization scheme based on pruned neural network are proposed in the dissertation to compensate impairments in short-reach,high-speed intensity modulation and direct detection(IM/DD)optical fiber transmission system.a)To solve the problem of bandwidth limitation induced by commercial low-cost devices in data centers,a novel receiver side DSP structure which composed of full response equalizer,noise-whitening post-filter and maximum likelihood sequence detection(MLSD)algorithm is proposed.Given the above structure,it is possible to compensate nonlinear impairemnts by providing low-complexity variable step polynomial nonlinear equalizer as the full response equalizer.The experimental results of single-sideband modulation(SSB)PAM4 signal transmission show that when the system's 10dB bandwidth is roughly 13.5GHz,64Gbps PAM4 signals can be transmitted over 80km dispersion-uncompensated standard single-mode fiber(SSMF).Besides that,by combining the above-mentioned DSP structure and bandwidth pre-compensation scheme,100Gbps PAM4 signal transmission over 160m multimode fiber(MMF)is achieved in vertical cavity surface emitting laser(VCSEL)-MMF system,which suffers severer nonlinearity and bandwidth limitation.The study provides an important reference for the engineering implementation of low-cost DSP technology in data center interconnections.b)In view of the limitations of polynomial nonlinear equalizer in compensating nonlinear impairments,a neural network equalizer which based on circulating pruning scheme is proposed to achieve low-complexity,high-roubstness nonlinear equalization.The 112Gbps SSB-PAM4 transmission experiments show that the neural network equalizer can bring about an order of magnitude BER performance improvement compared with Volterra equalizer after 80km dispersion uncompensated SSMF transmission.Besides that,the circulating pruning scheme can gurantee BER below the hard-decision Forward error correction(HD-FEC)threshold when number of neural network connections reduced 90%.Aiming at the abnormal performance improvement brought by neural network equalizer,we explore the reason and real performance improvement through simulation.The study provides an important reference for the further research on the joint design of neural network and equalization technology.2.Machine learning assisted optical performance monitoring technologyA multiple paramaters simultaneous monitoring scheme and a nonlinear region OSNR monitoring scheme are proposed in the dissertation to monitor multiple parameters in low-margin elastic optical networks(EONs).a)Based on multi-task learning artificial neural network(MTL-ANN),a low-complexity,high-accuracy,high-stability modulation format and OSNR simultaneous monitoring scheme is realized.The simulation and experimental results acquired from IM/DD system and coherent transmission system show that modulation format identificaion and OSNR monitoring accuracy can reache 100%and 98.5%,respectively.To solve the problem of computational resource consumption induced by manually adjusting the task loss weight of MTL-ANN,a loss weight adaptive MTL-ANN is proposed.Aiming at the problem induced by inaccurate monitoring results,a two-stage algorithm is proposed to improve monitoring confidence.The experimental results show that the confidence level of doubtful estimation results within 3 dB deviation is 1.The study not only achieves high accuracy multiple parameters simultaneous monitoring,but also greatly reduces the consumption of computing resources,which is helpful to realize low-cost multiple parameters joint monitoring in EON.b)In view of the influence of fiber nonlinearity on the existing OSNR monitoring technology,an OSNR monitoring scheme in nonlinear region assisted by tap coefficients of adaptive filter is proposed.Feature extraction technology is proposed to simplify the neural network structure.The scheme is simulated and verified in a dual-polarization,wavelength division multiplexing coherent optical fiber system.During the simulation,different link impairments and configurations are considered to simulate the actual EON and validate generalization of the algorithm.The simulation results show that when assisted by the adaptive tap coefficients,the mean square error(MSE)of OSNR monitoring is 0.3dB.Compared with the case where adaptive tap coefficients are not used,about 1dB OSNR monitoring MSE is decreased.The study can be combined with the existing coherent receiver side algorithm to achieve flexible and low-complexity OSNR monitoring in nonlinear region.