Research On Deep Learning Equalizer For Short-reach Optical Communication Systems

In recent years,with the continuous development of various broadband applications and multimedia services,people have put forward higher demands on the transmission distance and transmission rate of optical fiber communication networks.As a key component of the optical fiber communication network,the demand of data center for increased transmission capacity has also become particularly urgent.At present,the short-reach transmission system based on data center interconnection mainly adopts intensity modulation/direct detection(IM/DD)technology.Compared with other technologies,this technology is widely used in short-reach transmission systems subject to cost constraints with the characteristics of simple system structure,low power consumption,and low cost.Although there are many advantages based on IM/DD technology,device nonlinear effects and link damage problems will severely limit the performance of the transmission system as the transmission rate and distance increase.In the system based on intensity modulation direct detection(IM/DD),the signal is affected by the nonlinearity of the device and the damage of the transmission link,which causes a serious degradation of the signal transmission quality.Since traditional linear equalizers(e.g.,FFE,LMS)cannot achieve device nonlinear equalization.In addition,problems such as high complexity and limited compensation ability exist in traditional nonlinear equalizers(such as Volterra equalizer).In recent years,researchers have discovered that neural network technology has a strong ability to deal with nonlinear problems,and neural network has gradually been introduced into the quality evaluation and performance monitoring of optical fiber communication systems,as well as the structural optimization design of the entire communication system.This work will focus on the damage equalization technology based on neural network.The main research work are as follows:Firstly,theoretically analyze the causes of linear effects and nonlinear effects.And verify their impact on system performance through simulation.Study the equalization performance of traditional signal equalization algorithms(e.g.,FFE equalizer and Volterra equalizer)for link damage.Secondly,to further improve the link damage equalization performance,this work focuses on the damage equalization technology based on neural networks.The equalization principle based on the fully connected neural network equalization algorithm,the one-dimensional convolutional neural network and the two-dimensional convolutional neural network equalization algorithm is analyzed in detail,and the equalization performance of the above algorithm is verified through simulation and experiment.The experimental results show that the transmission distance of 38 km,72 km and 80 km can be respectively achieved under the threshold of 7% FEC.In addition,in order to prevent the neural network from over-fitting caused by pseudo-random numbers,a more random pseudo-random number is prepared by mixing random selection rules,and cross-validation is performed.Finally,an equalization algorithm based on temporal convolutional neural network is proposed,and its equalization principle is analyzed in detail.The damage equalization performance of the fully connected neural network,the one-dimensional and two-dimensional convolutional neural network,and the temporal convolutional neural network equalization algorithm proposed in this paper is analyzed under back-to-back and different optical fiber transmission links based on the 56-Gbit/s PAM4 direct detection experimental transmission platform.Compare and analyze the complexity of various equalization algorithms.Experimental results show that the equalization algorithm proposed in this paper can achieve the best equalization performance with relatively low complexity,and the signal transmission distance can be extended to 100 km under the 7% FEC threshold.