Neural Network-based Equalization In High-speed Passive Optical Network

In these years,increasing bandwidth-consuming applications and market-demanding factors have been driving the need for higher-speed optical access network.Due to the requirement of low-cost optical devices,the bandwidth limitation in the channel is inevitable,resulting severe inter symbol interference.Moreover,the distortion can mix with other effects like nonlinearities of the optical devices to further degrade the signal quality during transmission.Hence,advanced digital signal processing is required to compensate the distortion before or after transmission.In face of severe ISI and nonlinearities,conventional algorithms may not be efficient for equalization.Nowadays machine learning,especially neural network(NN),has been considered as a powerful equalization tool to mitigate both linear and nonlinear distortions in optical communications,which is applied as equalizer to achieve high-speed transmission.In this thesis,we provide a complete research on NN-based equalizer in Passive optical network(PON).In some researches about NN-based equalizer,pseudo-random bit sequence(PRBS)is used to construct the signal.However,it has been revealed that NN can learn the generation rules of PRBS during the training process,resulting in an abnormally high performance.To address the problem of training data,we theoretically analyze the PRBS rule detection of NN and its influence on the performance.Then,we provide a mutual verification strategy to verify the training effectiveness and experimentally validate such strategy.Finally,we propose a combination strategy to construct a strong random sequence that will not be learned by NN.Based on the conclusion above,we investigate the application of NNbased equalizer in PON.We found NN-based equalizer has the same performance with feed-forward equalizer and Volterra nonlinear equalizer in the case of linear distortion only,but outperforms them in strong nonlinearity case.In the experiment,to improve the loss budget,we increase the launch power to 18 d Bm,achieving a 30-d B loss budget for 33Gbaud/s PAM8 signal high-speed intensity modulation and direct detection(IMDD)system with a frequency response of 16.2 GHz,attributed to the strong nonlinear equalization capability of NN.Besides,we study the way to improve the performance of NN-based equalizer based on model structure optimization and data feature processing.The experimental results show that data feature processing may be a more feasible way to improve the performance.A novel decision feedback neural network(DFNN)architecture is proposed to handle the data features and enhance the system performance without inducing additional computational overhead.Compared with the traditional NN,DFNN can achieve 1-d B sensitivity improvement without increasing the complexity.Considering the practical application of NN-based equalizer,for the first time,we proposed an unsupervised learning training method for neural network(NN)-based blind equalization.This scheme consists of labeling and training.The whole scheme can train a NN-based equalizer without given original symbols.Therefore,it can directly train an NN model with only the received signal.Besides,the unsupervised learning can also help a well-trained NN to keep its performance in face of varying system status,such as wavelength shift and bias fluctuation in practical applications.This thesis provides a complete research of NN-based equalizer in PON,which is significant for the application,and also provides a guidance for the future work about NN-based equalizer and production.