Research On Optical Performance Monitoring Technology Of Coherent Optical Communication System Based On Deep Learning

With the development of optical fiber communication technology in the direction of high speed,large capacity and long distance,the influence of fiber nonlinearity and other damages in optical fiber communication systems is becoming more and more serious.In order to ensure the reliability and stability of the system,optical performance monitoring(OPM)technology is very important.At present,the single optical signal-to-noise ratio(OSNR)monitoring technology is difficult to meet the requirements of system operation and maintenance.Therefore,the joint monitoring of OSNR,nonlinearity and modulation format is an important development direction of OPM,which has important practical significance.This thesis focuses on the multi-parameter joint monitoring technology based on deep learning in coherent optical communication system.The main work is as follows:Firstly,the basic principle of damage monitoring of coherent optical communication system based on deep learning is analyzed,and the multi-parameter joint monitoring technology of OSNR,nonlinear noise power and modulation format based on different deep learning models is studied.Asynchronous complex histograms(ACH)is designed as the input feature data of the neural network,and proposed a multi-task deep neural network(MT-DNN)based on ACH,a multi-task convolution neural network(MT-CNN)based on ACH,MT-DNN based on asynchronous amplitude histogram(AAH),and MT-CNN based on AAH joint monitoring schemes.The joint damage monitoring algorithm of multi-channel optical transmission system is studied,and a 5-channel WDM coherent optical communication simulation system with a single channel rate of 28 GBaud is built.The results show that the four schemes can realize the joint monitoring of the three parameters.Among them,the performance of the joint monitoring of MT-DNN based on ACH is the best,and the total number of learnable parameters is reduced by about 88.6% compared with those of MT-CNN based on ACH and AAH,thus saving computing resources and memory overhead,which verifies the superiority of this scheme in performance and computing efficiency.This thesis optimizes and analyzes the performance of the joint monitoring scheme of MT-DNN based on ACH,studies the setting and optimization of hyperparameter under different transmission conditions,and obtains the best monitoring performance while saving computing resources and reducing training complexity as much as possible.The monitoring results show that for QPSK,16 QAM and 32 QAM signals,this scheme can realize highprecision modulation format identification,and the root mean square error of nonlinear noise power and OSNR monitoring value can reach 0.21 dB and 0.22 dB,respectively,and it can be robust to varying transmission distance,transmission power and laser linewidth.Compared with the traditional single-parameter DNN monitoring model,the performance of MT-DNN joint monitoring model has been effectively improved,and the redundancy of network parameters and training complexity have been reduced.