Dual Learning Based PAM4 Short-reach Optical Interconnection

In recent years,with the popularity of cloud computing technology,hardware virtualization,and e-commerce,content distribution network and other cloud services accelerate the demand of network bandwidth in data centers.Relevant data shows that the shortreach communication between data centers has become the dominant part in the current communication networks.In short-reach optical interconnection system,PAM4 modulation based intensity modulation direct detection(IM/DD)system is the mainstream choice due to its low system cost and low power consumption.In IM/DD-PAM4 short-reach optical interconnection systems,the degradation of communication signal mainly comes from the dispersion effect of optical fiber channel.Dispersion effect belongs to linear signal damage,which can be recovered by digital signal processing technology.But in IM-DD system,due to the square law detection of the photodiode,the dispersion effect will result in nonlinear signal impairment,which will reduce the system capacity.Conventional linear equalization algorithms can only compensate the linear signal impairments.For the nonlinear impairments,nonlinear filters have limited compensation ability.Deep learning based neural networks have stronger nonlinear expression ability,and have been proved to have better performance in many signal equalization applications.However,deep learning usually needs a lot of training data to perform well in generalization,while a lot of training data will greatly increase the training cost.This paper explores the application of dual learning technology in signal equalization of IM-DD shortreach optical interconnection system.Firstly,this paper analyzes the influence of dispersion on the system peformance under different transmission distances and data rates.We then validate the mechanism of the signal impairment in a simulation system.Based on numerical studies,we propose the dual learning based equalization methods.Different from the conventional netural networks based equalizers,our method can explicity use the dual consistency regularization to improve the model's generalization ability,which can reduce the dependence on datasets.We analyze and compare the performance of the traditional equalization algorithms and the neural network based equalization algorithms under different transmission scenarios.We find that the performance of dual learning based neural networks performs better than the traditional linear and nonlinear equalizers under the same dataset.In addition,we also compare the equalizatiom performance of dual learning model and traditional neural network models under different training dataset size and different input dimensions of the neural network.The contradiction between generalization and training cost of neural network in signal equalization task and the intrinsic reason why dual learning can improve the equalization performance are explained.Finally,experiments are conducted and the results show that the dual learning based equalization performs better than tranditional equalizers.Under a small-size dataset,dual learning can improve the equalization performance of the neural network and the capacity of the short-reach optical interconnection systems.