Research On Optical Fiber Nonlinear Mitigation Technology With Machine Learning

With the development of Internet applications,the demand for network capacity is growing rapidly,so the development of high speed and large capacity optical fiber communication system is extremely urgent.However,the research on high transmission rate,large channel capacity and long-distance transmission is limited by two kinds of factors: linear impairments,including fiber loss,chromatic dispersion,polarization mode dispersion;Nonlinear impairments,including laser phase noise,self-phase modulation,cross-phase modulation,four-wave mixing and so on.Most of all,fiber nonlinearity is one of the major limitations to the achievable capacity and the transmission distance in long distance fiber optic transmission systems.Therefore,the fiber nonlinearity equalization techniques in optical fiber transmission have the great significance of reality.In order to compensate the nonlinear impairments in fiber optic link and realize high speed and large capacity fiber optic communication,it is necessary to introduce nonlinear compensation(NLC)digital signal processing(DSP)algorithm into digital coherent receivers.In recent years,some NLC algorithms have demanded excessive signal processing resources,and accurate knowledge of the transmission system,and it is hard to make breakthroughs in reducing complexity.Therefore,the DSP algorithm with high efficiency and low complexity is the key to the high speed,large capacity,long distance optical fiber transmission system.This has drawn a great attention to applications of machine learning for fiber optic communication.Machine learning algorithm compensates the nonlinear impairments by learning from the received data.Machine learning algorithm does not need to know the specific optical fiber transmission link parameters in advance,and greatly reduces the complexity.Since the machine learning algorithm does not involve the details of fiber optic link,it can be widely applied to all optical fiber communication links,including short distance,long distance,submarine and so on.In our research,we focus on the studies of nonlinear equalization technology based on machine learning.We explore equalization income improved by neural networks(NN).Besides,we apply transfer learning(TL)to single-channel and multi-channel systems,which greatly reduces the training overhead of the NN equalizer.In view of the sensitivity of the probabilistic shaping system to nonlinearity,we proposed TL-NN algorithm to equalize the nonlinear effect in the probabilistic shaping system,and we demonstrate that the scheme compensates for the combined effects of transceiver and fiber propagation nonlinearities.At last,we propose and experimentally demonstrate a simplified transfer learning NN-NLC(TL-NN-NLC)model by employing the most significant intra-channel interference terms as inputs.The inputs features to the NN are the symbols of interest and the triplets,which is selected by the nonlinear perturbation coefficients and the set window size.Meanwhile,we use transfer learning algorithm to apply the hyperparameters learned in previous NN to the test dataset of different launch powers to accelerate the remodeling.