A Layer-Reduced Neural Network Based Digital Backpropagation Algorithm for Fiber Nonlinearity Mitigation

Abstract

A layer-reduced neural network based digital backpropagation algorithm called smoothing learned digital backpropagation (smoothing-LDBP), is proposed in this paper. The smoothing-LDBP smooths the power terms in nonlinear activation functions to limit the bandwidth. The limited bandwidth of the power terms generates fewer in-band distortions, thus reduces the required layer for a given equalization performance. Simulation results show that the required layers of smoothing-LDBP are reduced by approximately 62% at 6.7% HD-FEC compared with learned digital backpropagation. Owing to the layer reduction, the latency and the complexity are reduced by 69% and 51%, respectively. The layer-reduced property of smoothing-LDBP is also validated by a proof-of-concept experiment.

Modified Mean-Power-Distribution-Based Nonlinear Coefficient Optimization for Digital Back-Propagation in High-Baud-Rate Optical Systems

“…D. Rafique et al optimized the nonlinear coefficient based on the correlation of signal power in neighboring symbols [19]. M. Secondini et al proposed an enhanced DBP algorithm that combined the benefits of split-step and perturbation-based approaches to reduce computational complexity [20] and more recently, a smoothing learned DBP using a neural network was proposed to reduce the computational complexity of DBP algorithm [21]. In our previous work, a nonlinear coefficient optimization scheme based on the peak power distribution of Gaussian pulse was proposed to improve the NLC performance of the CS-DBP algorithm for a 2400 km SSMF 32 GBaud polarization-multiplexing (PM) 16 quadrature amplitude modulation (16QAM) optical transmission system [22].…”

A Fiber Nonlinearity Compensation Scheme With Complex-Valued Dimension-Reduced Neural Network

“…To improve the compensation performance, a helpful strategy is designing ML-NLC by utilizing the theory of fiber nonlinearity. One kind of ML-NLC algorithm following this strategy treats linear and nonlinear steps of DBP as layers and activation functions of NN [18]- [22]. Same as DBP, these approaches require at least two samples per symbol.…”