| With the continuous development of the information society,technology industries such as big data,cloud computing,and mobile internet have shown strong development potential.The rapid,stable,and secure transmission of massive amounts of information has also become one of the most important factors affecting current social progress.Modern communication network architecture mainly includes core networks,metropolitan area networks,access networks,cellular networks,local area networks,data center networks,and satellite networks.Information interaction between base stations in different networks is achieved through optical fiber links.In addition,in data center networks,information transfer between data centers is also achieved through optical fiber links.With the continuous development of laser technology,optical fiber communication technology has now become the "artery" of modern communication backbone networks.The continuous development of laser technology and the continuous in-depth study of digital signal processing algorithms have led to the continuous development of optical fiber communication systems towards greater capacity,longer distances,and higher speed.Compared to traditional direct modulation/detection optical fiber communication systems,coherent modulation and demodulation techniques have been widely used in high-speed and long-distance optical fiber transmission systems due to their more modulation dimensions,higher sensitivity,and stronger frequency selectivity.Nonlinear damage is the most important factor restricting high-speed and long-distance coherent optical communication systems.Based on the study of the theoretical basis of optical fiber communication systems,this paper describes various linear impairments in optical fiber communication systems,and briefly introduces their mainstream equalization algorithms.In addition,the research focuses on nonlinear equalization algorithms based on artificial intelligence technology.Focusing on the adaptive clustering algorithm for 16QAM modulation format,unsupervised clustering algorithms can be used to achieve the clustering decision of receiver symbols in optical fiber communication systems without system prior conditions;For higher order modulation formats,the emphasis is placed on the study of nonlinear compensation algorithms that combine perturbation theory with convolutional neural networks and complex convolutional neural networks,and the introduction of a migration learning approach verifies that the neural network nonlinear equalizer assisted by migration learning can be quickly trained and converged to the optimal parameters in different scenarios;In addition,for ultra-low complexity nonlinear equalization algorithms,the concept of integrated learning is introduced,and multiple decision trees are formed into a random forest.The random forest algorithm is used to achieve nonlinear compensation for optical fiber communication systems.Finally,build a 975km PDM-16QAM coherent optical transmission experimental platform,build a 375km PDM-64QAM coherent optical transmission experimental platform,and build a wavelength division multiplexing simulation platform to verify the algorithm performance,The main work and innovations of this article are as follows:1.A clustering algorithm based on the cascade of DBSCAN and DW-KNN is proposed.Using DBSCAN to achieve unsupervised clustering of receiver symbols,and then using DW-KNN to classify noise points that DBSCAN cannot correctly cluster,not only solves the noise point problem of DBSCAN algorithm,but also solves the supervised problem of training data required by DW-KNN algorithm,ultimately achieving significant improvement in performance and flexibility of nonlinear equalization algorithm.Based on the optimal parameters,the optimized DBSCAN algorithm improves performance by 0.33 dB when the system is at the optimal transmit power on the experimental platform,and improves system performance by 0.30 dB when the system is at the optimal transmit power on a wavelength division multiplexing system.2.Based on the joint analysis of perturbation theory and nonlinear Schrodinger equation,a novel feature map is created,and a nonlinear equalization algorithm based on convolutional neural networks is proposed.Because the nonlinear feature map preserves the positional relationship between the triple product feature elements,as well as the correspondence between the real and imaginary parts of the triple product feature elements as complex elements.Compared to fully connected neural networks,while ensuring system performance,the time complexity is reduced by about 25%and the spatial complexity is reduced by about 50%.3.Based on a new feature map,a complex convolutional neural network is introduced,and a large convolutional kernel is developed to increase the visual receptive field during the convolution process,improving classification accuracy.Based on complex convolutional neural networks and large convolutional kernels,the required map size at the input of the neural network can be reduced under the same equilibrium performance.Compared to real small kernel convolutional neural networks,the time complexity is reduced by about 50%and the spatial complexity is reduced by about 60%.4.Introducing transfer learning and using the parameters of the optimized and converged neural network nonlinear equalizer as the initial parameters of neural network nonlinear equalizers of the same structure in other scenarios greatly accelerates the process of neural network training optimization and reduces the amount of training data required during the training process.5.A random forest nonlinear equalizer based on ensemble learning is proposed,which combines multiple decision trees through ensemble learning and obtains the final classification results through a voting mechanism.With very low computational complexity,the equalization performance can be comparable to that of neural networks.In addition,a nonlinear equalization scheme based on subarea is proposed.For the receiver symbols,a base classifier is used to divide them into different power regions,and then nonlinear equalizers in different power regions are used to achieve their respective equalization tasks,avoiding the problem of channel condition mismatch encountered by artificial intelligence algorithms in the process of channel equalization. |
PDF Full Text Request