Deep learning based digital backpropagation demonstrating SNR gain at low complexity in a 1200 km transmission link

Abstract

A deep learning (DL) based digital backpropagation (DBP) method with a 1 dB SNR gain over a conventional 1 step per span DBP is demonstrated in a 32 GBd 16QAM transmission across 1200 km. The new DL-DPB is shown to require 6 times less computational power over the conventional DBP scheme. The achievement is possible due to a novel training method in which the DL-DBP is blind to timing error, state of polarization rotation, frequency offset and phase offset. An analysis of the underlying mechanism is given. The applied method first undoes the dispersion, compensates for nonlinear effects in a distributed fashion and reduces the out of band nonlinear modulation due to compensation of the nonlinearities by having a low pass characteristic. We also show that it is sufficient to update the elements of the DL network using a signal with high nonlinearity when dispersion or nonlinearity conditions changes. Lastly, simulation results indicate that the proposed scheme is suitable to deal with impairments from transmission over longer distances.

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

“…After the data pre-processing procedure above, the data forward propagates through the smoothing-LDBP and the following operations, i.e., the downsampling and the phase recovery, to the cost function. These operations are treated as static layers, which can pass the gradients back to the smoothing-LDBP [19]. The gradients of the cost function with respect to the parameters are utilized for training, i.e., iteratively updating linear taps and nonlinear phase factors κ i .…”

“…The convolution process of E and E 2 in the frequency domain is shown in the left side of the figure, which indicates that the inband spectrum of σ i is also distorted. Using the gradient-descent-based optimizer, LDBP learns a set of "M"-shape linear layers to limit the broadened bandwidth [18], [19]. However, the in-band distortion of σ i can not be completely removed by learned linear layers.…”

“…A variant of the NN-based NLC algorithm, called learned DBP (LDBP), treated the linear steps and the nonlinear steps of SSFM as the linear layers and the nonlinear activation functions of an NN [15], [16]. LDBP showed a significant performance gain compared with the DBP in the experimental settings [17]- [19]. The taps in linear layers can be extremely short by gradually pruning them during training [17].…”

“…To reduce the complexity and latency, a straightforward approach is reducing layers in LDBP for fewer FFT and IFFT. However, performance degradation of LDBP was observed when the layers are fewer than the number of fiber spans [19].…”

Performance Versus Complexity Study of Neural Network Equalizers in Coherent Optical Systems

“…However, the test of similar NN architectures in coherent optical systems has been carried out, mainly, numerically [17]- [20], or in short-haul experiments [21]- [24]. It is worth noticing that some very recent works evaluated the functioning of NN-based equalizers in metro/long-haul trials [4], [5], [8]- [10].…”