Equalization Of XPM Effect In WDM Optical Fiber Communication System Based On Recurrent Neural Network

With the development and application of new technologies such as cloud computing,5G,and artificial intelligence,global IP data has grown exponentially.The optical fiber communication system carries most of the data traffic transmission tasks.In order to meet the rapidly growing network demand,the development of high-speed,long-distance,and large-capacity optical fiber communication transmission networks has become an inevitable trend.However,the performance of optical fiber communication system is mainly limited by two factors,one is linear damage,such as loss,dispersion,polarization mode dispersion;the other is nonlinear damage,such as self-phase modulation,cross-phase modulation,four-wave mixing and so on.For long-distance high-speed optical fiber communication systems,nonlinear damage has become the most important factor restricting system capacity and transmission distance.Therefore,it is of great significance to study nonlinear damage equalization technology.By introducing a Nonlinear Equalizer(NLE)algorithm at the coherent receiving end for digital signal processing,the equalization of the damaged signal can be achieved.The survey found that the existing nonlinear compensation algorithms mainly have the following problems.First,the specific parameters of the transmission link are required,which are often not known in the actual optical fiber transmission system.Second,the algorithm has high computational complexity,and the improvement of algorithm accuracy is still more channel resources are required,and the complexity is difficult to reduce.Therefore,researching low-complexity and efficient algorithms is the key to realizing nonlinear damage equalization.The NLE algorithm based on machine learning can autonomously mine the potential nonlinear mapping relationship from the damaged signal at the receiving end.And it can realize the modeling of nonlinear channel without knowing the specific parameters of the link.In addition,with the powerful computing power of machine learning,the complexity of compensation algorithm is greatly reduced,so the research of equalization algorithm based on machine learning has received great attention.This paper studies the realization of XPM equalization based on recurrent neural networks(RNN)in machine learning.By studying the machine learning algorithm,it is found that RNN is suitable for solving the problem of time-series dependent scenarios.For WDM coherent optical communication systems,the damage of the signal not only depends on the current moment,but also is affected by the transmission signals in the previous and subsequent sequences,which is exactly the problems that RNNs are good at solving.In this paper,different optical fiber communication simulation transmission scenarios are built to evaluate the performance of the proposed RNN-NLE algorithm.In the 2400km-28GBaud-16QAM single-channel scenario,the RNN-NLE algorithm can increase the fiber input power by about 1 dB compared with only the CD algorithm.In the 960 km-28GBaud-16QAM three-channel and nine-channel WDM scenarios,the fiber input power are increased by about 2.8dB and 2.2dB respectively,and only 150 real multiplications are required,which verifies that the compensation scheme can achieve nonlinear damage equalization with low computational complexity.Through the application of transfer learning,the training overhead of the network is further reduced to 30%of the retraining amount.And in the multi-channel scenario,the RNN-NLE+DBP cascaded solution performs better.Just by cascading a single-step DBP,the fiber input power of the three-channel WDM system can be increased by about 3.2dB,and the nine-channel system can be increased by about 2.2dB.In conclusion,RNN-NLE algorithm has good portability,and can be applied to transmission systems with different link parameters only by parameter fine-tuning,and achieve better performance improvement.At the same time,the algorithm has low computational complexity and multiplication.The number of times is only on the order of a hundred.Therefore,the nonlinear compensation scheme based on RNN proposed in this paper has important reference significance.