Research And Implementation Of Autoencoder Technology Based On End-to-end Learning In Optical Fiber Communication System

In recent years,the rapid development of information technology has brought about a surge in the amount of data.With the advantages of high bandwidth,large capacity and low loss,optical fiber communication has become an important way of wired communication to meet the growing demand of data transmission.However,the optical fiber communication system is limited by the traditional communication system design based on block structure,and different modules have different local optimal objectives,so it is difficult to achieve the global optimal performance.The autoencoder technology in deep learning,which is based on the joint optimization goal of end-to-end learning and fits the structure of optical fiber communication system,becomes a possible way to realize the global optimization of optical fiber communication system.This paper focuses on the intensity modulation direct detection(IM-DD)optical fiber communication system,aiming at the research and implementation of end-to-end learning based autoencoder technology applied to the system,proposes and solves some problems,and finally achieves the overall improvement of system performance.The main innovations of this paper are as follows:Firstly,due to the complex physical characteristics of optical fiber communication channel,such as dispersion,nonlinearity,attenuation and noise,the traditional mathematical modeling method not only relies on a lot of professional knowledge,but also hinders the back propagation of gradient in end-to-end training.In this paper,based on the data-driven deep learning method,we use BiLSTM neural network to model the optical transmission devices including modulator,optical fiber and optical detector as channel.In addition,by adding distance labels to the training data,the channel model is suitable for different transmission distances.After analyzing the time-domain signal waveform and eye diagram,it can be concluded that the channel model can accurately reflect the response characteristics of optical fiber communication channel.R-square also shows that the average fitting degree of the transmission response of different modulation formats is as high as 99.6%.Secondly,aiming at the problem that the traditional optical fiber communication system can not be globally optimized based on the modular design structure,the end-to-end learning autoencoder is introduced to optimize the whole IM-DD optical transmission system,including transmitter,channel and receiver as a whole.The simulation results show that the autoencoder has formed a robust and one-to-one mapping mechanism through end-to-end training.It can realize 12Gb/s system transmission at a bit error rate(BER)lower than 6.7%of forward error correction hard decision threshold within 50km transmission distance,which is better than traditional pulse amplitude modulation.In addition,the introduction of multi task learning enables the system to be applied to simultaneous interpreting scenarios with different transmission distances,which increases the mobility of the system and enriches the application scope of the self-encoding system of optical fiber communication.