基于深度学习的数据中心光通信色散估计与管理  

Estimation and management of fiber dispersion in the data center based on the deep learning

作　　者：瞿国庆[1,2] 于树科 QU Guoqing;YU Shuke(Internet of Things Technology Research Institute,Jiangsu Vocational College of Business,Nantong 226011,China;Nantong Greatwisdom Information Technology Co.,Ltd.,Nantong 226010,China)

机构地区：[1]江苏商贸职业学院物联网技术研究所,江苏南通226011 [2]南通智大信息技术有限公司,江苏南通226010

出　　处：《光学技术》2018年第5期617-621,共5页Optical Technique

基　　金：江苏省第十四批“六大人才高峰”高层次人才培养资助项目(XYDXX-121);南通市崇川区科技创新计划项目(CCY201607)

摘　　要：最大似然估计能够准确地估计光通信的色散,但其对高速符号传输的计算复杂度较高,云计算数据中心的光网络是一种高速光通信的情况,为了降低数据中心中光通信色散估计的计算成本,设计了一种基于深度学习的数据中心光通信均衡器。基于人工神经网络的均衡器分为两个阶段,第一阶段采用光信道的脉冲响应数据对人工神经网络进行训练,对人工神经网络的模型参数进行优化,建立人工神经网络的非线性响应模型;第二阶段采用训练的人工神经网络均衡器对光信道的传输数据进行处理,实现对光信道色散的估计与补偿。按照数据中心的光网络方案进行了仿真实验,结果显示,基于人工神经网络的均衡器提高了光通信的光信噪比,并且延长了光通信的传输距离。Fiber dispersion can be estimated by maximum likelihood sequence estimation accurately,but it suffers from the high computational complexity in the condition of high-speed transmission symbol rate,the optical networks of data centers of cloud computing usually work with high-speed transmission symbol rate,in order to reduce the computational cost of estimation of fiber dispersion in data center,an efficient equalizer of the optical communication of data center based on deep learning is designed.The equalizer based on artificial neural network（ANN）consists of two phases,in the first phase,the impulse response datasets of the optical channel are adopted to train ANN and optimize the parameters of ANN model,and the non-linear response model of ANN is constructed;in the second phase,the trained ANN equalizer is used to process the transmitted data of optical channels,and the fiber dispersion is estimated and compensated.Simulation experimental results according to the optical network scenario of data center show that,the equalizer based on the ANN model improves the optical signal to noise ratio,and extend the transmission distance of optical communication.

关 键 词：人工神经网络 深度学习 数据中心 光纤色散 光纤通信 

分 类 号：TN253[电子电信—物理电子学]