Research On Nonlinear Mitigation Of WDM Channel Based On Machine Learning

In recent years,with the proliferation of various bandwidth-intensive applications,the core communication network faces a dramatic increase in network traffic.The current coherent optical transmission system is facing the challenge of network bandwidth,and wavelength division multiplexing(WDM)technology is a promising technical solution to realize large-capacity long-distance optical fiber transmission system.While the Kerr nonlinear effect of high-speed and long-distance transmission fiber limits the capacity of modern WDM optical communication systems.The traditional DSP nonlinear equalization technology needs to obtain the parameter information of the transmission link and a large amount of computing resources,which makes it difficult to implement in the actual system.This makes the low-complexity,high-performance nonlinear equalization algorithm the focus of research to improve the transmission distance and channel capacity of the system.Machine learning(ML)has excellent solution capabilities for regression and classification problems.The introduction of ML in coherent receivers is expected to improve the performance and reduce the complexity of existing nonlinear equalization algorithms.In this thesis,the nonlinear equalization of ML combined with traditional DSP is studied on the basis of existing.The improvement and research of the deep neural network nonlinear equalizer based on the perturbation model are carried out in the coherent WDM system,the original structure and the improved network are discussed and verified by simulation and experiment.The improved scheme uses the perturbation model to consider the influence of the cross-phase modulation of the WDM system to extract features from the received data preprocessing,and the principal component analysis is used to increase the network feature extraction capability.At the same time,the sliding window is used to optimize the triplet calculation method to reduce the computational complexity.After reducing the central channel power and eliminating the influence of self-phase modulation,the network can obtain 0.61 d B Q gain.Then,the compensation effect of dynamic source domain and fixed source domain of transfer learning on different system parameters is discussed.It is proposed to use meta learning algorithm combined with deep neural network to find the optimal source domain,and compare it with the source domain obtained by traversal transfer learning.Verified by single-channel experiment and multi-channel simulation,the results show that migrating fixed high nonlinear data as the source domain can get better performance and lower complexity for each target domain,but excessive nonlinear effects will lead to the distortion of the received signal and seriously degrade the performance;the meta learning algorithm considers the data within and between the source domains,and is not sensitive to the source domain set data.The performance and complexity of the optimal source domain found by meta learning training is slightly better than that of traversal transfer learning for each target domain.In the transmission power comparison of the singlechannel experimental system,the average number of multiplications of transfer learning is reduced by 94.8% compared with retraining,while the meta-learning is reduced by96.24%.In the multi-channel simulation system,the average number of multiplications of transfer learning is reduced by 90% compared with retraining,the meta-learning algorithms reduce the number of multiplications by 91.3% on average.Meta learning also obtains 0.5d B and 0.14 d B gains in Q factor.