面向短距光互连的隐藏层拓展循环神经网络均衡器

Hidden-layers Extended Recurrent Neural Network Equalizer for Short Reach Optical Interconnects

为探究光强度调制/直接检测系统中神经网络非线性均衡器的最优结构,实现更优的光通信性能,通过搭建20 km的112 Gbps四电平脉冲振幅调制无色散补偿的光传输系统,分析了多隐藏层循环神经网络均衡器在该系统中的误码率及算法复杂度性能。结果表明,在同一误码率阈值下,增加隐藏层层数可有效降低单个隐藏层中神经元的数目,从而降低均衡器的算法复杂度。但随着隐藏层数目的进一步增加,均衡器的收敛性能下降,导致误码率及复杂度性能的恶化。通过量化研究隐藏层数目对循环神经网络均衡器性能的影响,发现在112 Gbps容量和20 km传输距离下,2层隐藏层RNNE具有最优的误码率及算法复杂度性能。与单隐藏层均衡器相比,2层隐藏层均衡器在1×10^(-3)误码率阈值下算法复杂度降低23.3%。

There has been a lot of interest in the installation of high-speed short-reach optical interconnect systems recently because of the growth of 5G and the Internet of Things(IoT),which have caused the data traffic between and within data centres to expand quickly.In data centres,optical transmission systems frequently use optical Intensity Modulation and Direct Detection(IM/DD)to save cost and power consumption.However,loss of optical phase from square law detection and fiber dispersion cause a nonlinear distortion in the optical IM/DD system.Moreover,the nonlinear responses of modulator and driver/amplifier also cause serious nonlinear distortions at the same time,which seriously reduce the optical IM/DD system’s transmission range and capacity.Various equalization algorithms have been proposed to eliminate them.A classical equalization scheme is the combination of feedforward and decision feedback equalizer,but the nonlinear distortions can not be effectively equalized.Volterra Nonlinear Equalizer(VNE)can correct for nonlinear distortions,nevertheless,higher-order VNE items in strongly nonlinear settings result in a significant increase in complexity.On the other hand,nonlinear equalizers based on neural networks were also widely investigated in optical communication recently,which includes feedforward neural network,radial basis function neural networks,convolutional neural network and recurrent neural network.In contrast to the feedforward equalizer and VNE,feedforward neural network equalizer exhibits stronger equalization performances,but also brings a higher complexity in order to compensate for strong nonlinear impairments in optical IM/DD system.Moreover,equalizers based on auto-regressive recurrent neural network have higher complexity,however,better performance thanks to the involvement of additional feedback neurons.These equalisers,however,only employ one or two hidden-layers.In optical IM/DD systems,the influence of the number of hidden-layers as well as the number of neurons in every hidden layer on the performance of the equalizer remains unknown.Also,the optimal structure of neural network equalizer is worth exploring.Thus,we constructed a 112-Gbps 20-km four-level pulse-amplitude modulation optical IM/DD transmission simulation platform to investigate the influence of the number of hidden-layers and the number of neurons in every hidden layer on Recurrent Neural Network Equalizer(RNNE)performance.Also,to seek the most efficient equalization scheme with better complexity and Bit Error Rate(BER)performance.The effects of the number of hidden layers and the number of hidden neurons on the performance of RNNE are studied quantitatively to determine the ideal structure for RNNE.Initially,the performance of the RNNE with different numbers of neuron in the second hidden layer has been compared when the number of neurons in the first hidden layer is fixed.The results show that when RNNE has a comparable number of neurons in each hidden layer,the BER and complexity performance is optimized.Then,as for the RNNE with multiple hidden layers,we quantitatively examined the influence of the number of hidden-layer on the BER and complexity of RNNE.According to the results,the two-hidden-layer RNNE outperform RNNE with three-hidden-layer.The complexity of two-hidden-layer RNNE is 23.3%less complex than a single-hidden-layer RNNE.With similar algorithm complexities,the power budget of the two-hidden-layer RNNE is approximately 1 dB higher as compared to the single-hidden-layer RNNE at 7%-OH FEC threshold.This optimization strategy provides a reference for the selection of the number of hidden-layer number as well as the number of hidden neuron while using RNNE to compensate for nonlinear distortions in the optical IM/DD system.