Learning for Perturbation-Based Fiber Nonlinearity Compensation Shenghang Luo1 Sunish Kumar Orappanpara Soman2 Lutz Lampe1 Jeebak Mitra3 and Chuandong Li3

2025-04-29 0 0 1023.4KB 4 页 10玖币

侵权投诉

Learning for Perturbation-Based Fiber Nonlinearity Compensation

Shenghang Luo(1), Sunish Kumar Orappanpara Soman(2), Lutz Lampe(1), Jeebak Mitra(3),

and Chuandong Li(3)

(1) Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver,

BC V6T 1Z4, Canada, shenghang@ece.ubc.ca,lampe@ece.ubc.ca

(2) School of Engineering, Ulster University, Newtownabbey, BT37 0QB, United Kingdom,

s.orappanpara soman@ulster.ac.uk

(3) Huawei Technologies Canada, Ottawa, ON K2K 3J1, Canada, jeebak.mitra@huawei.com,

chuandong.li@huawei.com

Abstract Several machine learning inspired methods for perturbation-based ﬁber nonlinearity (PB-

NLC) compensation have been presented in recent literature. We critically revisit acclaimed beneﬁts of

those over non-learned methods. Numerical results suggest that learned linear processing of perturba-

Introduction

The Kerr-induced ﬁber nonlinearity puts a cap

on the maximum achievable information capac-

ity in long-haul coherent optical transmission sys-

tems[1]–[3]. Over the last few years, the advances

in deep learning algorithms initiated the develop-

ment of several system-agnostic machine learn-

ing (ML) approaches to compensate for the ﬁber

Kerr nonlinearity effects. On this ground, learned

versions of the perturbation theory-based ﬁber

nonlinearity compensation (PB-NLC) technique

have widely been investigated and demonstrated

its effectiveness in estimating the complex nonlin-

ear distortion ﬁeld with the perturbation triplets as

the input features[4]–[10].

Since the overall computational complexity of

the PB-NLC techniques has often been consid-

ered a limitation for practical implementation, one

direction of learned PB-NLC has been to lower

the complexity by, for example, reducing the input

feature vector size and pruning for neural network

(NN) based PB-NLC[5],[6],[9]. On the other hand,

the conventional (CONV) PB-NLC technique has

also notably been optimized in terms of nonlin-

earity compensation capability and overall com-

putational complexity. Nonlinearity compensation

performance improvements have been achieved

by using realistic pulse shapes and the inclu-

sion of the power proﬁle in the coefﬁcient com-

putation[11],[12]. Signiﬁcant complexity reductions

have been obtained by coefﬁcient quantization[13].

Furthermore, the introduction of a cyclic buffer

(CB) in the triplet computation stage permits fur-

ther complexity savings for both learned and non-

learned PB-NLC[10].

In this paper, we revisit the acclaimed

performance beneﬁts of learned PB-NLC ap-

proaches[4],[5] over their non-learned counter-

parts. For this, we carefully evaluate the overall

computational complexity of existing learned and

non-learned PB-NLC techniques, considering the

available advancements for all of them. We note

that such a comparison has not been done in the

references proposing learned PB-NLC. As one

important ﬁnding, our results show that feedfor-

ward NN (FNN)-based PB-NLC has hardly ad-

vantages over non-learned PB-NLC. On the other

hand, the approach of learning PB-NLC coefﬁ-

cients from a least-squares (LS) optimization is

shown to outperform the best non-learned PB-

NLC variant and to provide the best performance-

complexity trade-off of all tested methods.

PB-NLC Techniques

In CONV PB-NLC, the nonlinear distortion ﬁeld

is numerically calculated and subtracted from the

symbol of interest. However, the truncation of the

perturbation approximation at ﬁrst-order in PB-

NLC leads to a power overestimation problem

and degrades the NLC performance. This can

be overcome using an additive-multiplicative PB-

NLC (CONV-AM PB-NLC) technique[15].

ML-based solutions utilizing the existing PB-

NLC distortion model have been investigated

in several literatures. In[4],[5], fully connected

FNNs with perturbation triplets as the input fea-

tures have been proposed to estimate the non-

linear distortion ﬁeld instead of numerically com-

puting it. This technique is referred to as

FNN PB-NLC. NNs can also be used to predict

the distortions in the above-mentioned additive-

multiplicative model, which we refer to as FNN-

arXiv:2210.03440v1 [eess.SP] 7 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

LearningforPerturbation-BasedFiberNonlinearityCompensationShenghangLuo(1),SunishKumarOrappanparaSoman(2),LutzLampe(1),JeebakMitra(3),andChuandongLi(3)(1)DepartmentofElectricalandComputerEngineering,TheUniversityofBritishColumbia,Vancouver,BCV6T1Z4,Canada,shenghang@ece.ubc.ca,lampe@ece.ubc.ca(2)Schoo...

展开>> 收起<<

Learning for Perturbation-Based Fiber Nonlinearity Compensation Shenghang Luo1 Sunish Kumar Orappanpara Soman2 Lutz Lampe1 Jeebak Mitra3 and Chuandong Li3.pdf

共4页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Learning for Perturbation-Based Fiber Nonlinearity Compensation Shenghang Luo1 Sunish Kumar Orappanpara Soman2 Lutz Lampe1 Jeebak Mitra3 and Chuandong Li3

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: