Nonlinear Phase Noise Compensation Method Based On Machine Learning

Optical fiber nonlinear effect damage has become a bottleneck restricting the improvement of optical communication transmission performance as the communication transmission rate becomes higher and the transmission distance becomes longer and longer.The classic traditional fiber nonlinear damage algorithms such as DBP algorithm,Volterra series algorithm,perturbation theory,have the disadvantages of high compensation complexity,needing a lot of known link information,and inability to adapt to the dynamic changes of the link.The problem that these algorithms is difficult to implement in communication device production,how to maintain a good compensation effect while reducing the complexity of the algorithm has become the goal of current research on optical fiber nonlinear damage compensation algorithms.With the wide application of machine learning in various fields,its characteristics of selfadaptability and excellence learning ability are used in the field of fiber nonlinearity to further improve the performance of fiber damage compensation algorithms,reduce the algorithm complexity,and promote the commercial implementation of fiber nonlinear compensation algorithms.Using the tool of machine learning,this paper has completed the following work about the compensation of optical fiber nonlinear effects:1.Based on the Manakov equation,this paper exploits the excellent learning ability of neural network to design a custom hybrid network classification model combining the characteristics of signal damage and the basic principles of fully connected network,convolutional network and bidirectional long-term memory network in classical networks.The classification model is training and predicting using the data obtained on the single-channel long-distance transmission simulation platform of DP16QAM,and the bit error rate and Q value are selected as the performance indicators of the model.Compared with the compensation effect of the dispersion compensation algorithm and the 2Sps DBP algorithm,the custom hybrid network classification model improves the Q value by up to 1.43dB and 0.99dB respectively in the range of-4dBm to 4dBm transmit power.The complexity of custom model is analyzed in detail and the calculation formula of complexity is given.2.The proposed self-defined hybrid network classification model is optimized,and a simple fully connected classification model and convolution classification model are derived,which can better compensate the nonlinear damage of optical fibers with fewer complexity.The fully connected classification model has better compensation effect than the convolutional classification model,which has the characteristics of higher learning efficiency and faster iterative convergence.In practical applications,one can choose between the two models according to the importance of compensation effect and model training efficiency.The compensation effect of both the fully connected layer model and convolutional layer model is improved comparing to the dispersion compensation algorithm and the 2Sps DBP algorithm.Compared with only the dispersion compensation algorithm,the Q values of the fully connected model and the convolution model are improved by up to 1.66dB and 1.36dB,respectively;compared with the double sampling DBP algorithm,the Q values are improved by up to 1.34dB and 1.33dB.3.For the three improved algorithms based on neural network proposed in this paper,all gives the corresponding calculation formulas of complexity,and the comparison shows that compared with the traditional 2-span DBP compensation algorithm the three improved classification models are all improved.It has better compensation effect and lower compensation complexity.Among them,the fully connected layer has the best performance in compensation effect and complexity.Although the performance of the convolutional layer is slightly weaker than that of the fully connected layer,its characteristics of fast convergence and high training efficiency are conducive to its efficiency improvement of production in practical applications.4.The TPE network hyperparameter optimization algorithm is used to optimize the network parameter values of the model proposed in this paper,which effectively improves the compensation signal performance of the model and reduces the model complexity.By further analyzing the optimized hyperparameter value,we can derive that the target symbol can be distorted by a certain extent of symbols around it,and with the knowledge of the certain extent,we can decrease the complexity of compensation model without damaging the compensation effect.The neural network model is further analyzed in this paper.By visualizing the network weights,it can be found that the influence of the front and rear symbols on the target symbol does not change linearly with the distance,but has certain jumping and interval characteristics.Through the simulation experiments and analysis in this paper,on the one hand,the feasibility and effectiveness of the neural network in assisting the improvement of the performance of the nonlinear compensation algorithm are verified,and on the other hand,it can further deepen the knowledge of nonlinear damage by analyzing the characteristics of the weight of the network model.The understanding of network has guiding significance for the compensation of nonlinear damage in optical fiber.