【作者】 王磊；

【导师】 曾祥烨；

【作者基本信息】 河北工业大学 ， 通信与信息系统， 2022， 硕士

【摘要】 近年来,随着视频会议、大数据、物联网和移动数据传输为代表的新业务和新技术的迅猛发展,数据中心光互联的流量呈现指数级增长。与其他传输系统相比,基于脉冲幅度调制(PAM)的强度调制/直接检测(IM/DD)系统具有成本低、结构简单等优点,近年来作为一种有前景的解决方案得到了广泛的研究。对于高速IM/DD系统,不仅会受到收发器件带宽和光纤色散的限制,而且还会受到非线性效应的限制。由于IM/DD系统结构的原因,一定程度上线性失真与非线性失真相互作用,加重非线性失真。因此,有效的非线性均衡技术对于保证系统传输性能至关重要。然而传统非线性均衡技术存在复杂度高,补偿能力有限等缺点。近年来,基于机器学习的均衡技术广泛应用在IM/DD系统中。特别是不同类型的神经网络均衡器取得了优异的性能。本文主要研究神经网络算法在IM/DD系统接收端均衡的应用。主要研究工作如下:(1)针对IM/DD系统中的信号损伤问题,分析了系统主要失真来源。为了验证神经网络均衡器的有效性,搭建IM/DD PAM4光纤传输系统。对比了前馈神经网络和循环神经网络均衡算法和传统均衡算法的均衡效果。结果表明,神经网络均衡器的性能优于传统均衡器。(2)针对神经网络由于不同接收光功率得到的数据集不同,网络参数还需要重新训练的问题,引入迁移学习技术将在不同接收光功率下训练的网络参数进行迁移。结果表明,高接收光功率训练的网络参数向低功率迁移时,泛化效果好。得到的效果与原始单独训练的误码率(BER)性能基本一致。将这种技术应用到神经网络均衡中,可以减少实验的训练开销。(3)针对前馈神经网络非线性均衡器存在复杂度高的问题,提出改进权重剪枝的动态剪枝方法。结果表明,该方法能够动态全局地调整权重,使误剪的权重重新恢复,在降低复杂度的同时显著提高了剪枝后的均衡性能。采用动态剪枝方法相比权重剪枝方法,在7%硬判决-前向纠错(HD-FEC)门限下可将接收机灵敏度提高约1 d B。更多还原

【Abstract】 In recent years,with the rapid development of new businesses and technologies represented by video conferencing,big data,Internet of Things and mobile data transmission,Optical interconnection traffic in data centers is increasing exponentially.Compared with other transmission systems,intensity modulation/direct detection(IM/DD)system based on pulse amplitude modulation(PAM)has the advantages of low cost and simple structure,widely studied as a promising solution in recent years.For high-speed IM/DD systems,not only is limited by the bandwidth of transceiver devices and fiber dispersion,but also is limited by nonlinear effects.Because of IM/DD system structure,linear distortion and nonlinear distortion interact with each other to some extent,aggravating nonlinear distortion.Therefore,effective nonlinear equalization technique is very important to ensure the transmission performance of the system.However,the traditional nonlinear equalization technology has some disadvantages such as high complexity and limited compensation ability.In recent years,the equalization technology based on machine learning is widely used in IM/DD system.especially different types of neural network equalizer has achieved excellent performance.This paper mainly studies the application of neural network algorithm in IM/DD system receiver equalization.The main research work is as follows:(1)In view of the problem of signal distortion in IM/DD system,the main distortion sources of the system are analyzed.In order to verify the effectiveness of neural network equalizer,an IM/DD PAM4 optical fiber transmission system is setup.The equalization effects of feedforward neural network and recurrent neural network and traditional equalization algorithms are compared.The results show that the performance of neural network equalizer is better than that of traditional equalizer.(2)In order to solve the problem that neural network parameters need to be retrained due to different data sets obtained by different received optical powers,transfer learning technology is introduced to transfer network parameters trained at different optical powers.The results show that the generalization effect is good when the network parameters of high receiving optical power training transfer to low power.The obtained results are basically consistent with the BER performance of the original single training.Applying this technique to neural network equalization can reduce the training cost of experiment.(3)Aiming at the high complexity of feedforward neural network nonlinear equalizer,a dynamic pruning method with improved weight pruning is proposed.The results show that this method can dynamically adjust the weight globally and restore the weight of the wrong pruning,which can significantly improve the balancing performance after pruning while reducing the complexity.Compared with the weight pruning method,the proposed method can improve the receiver sensitivity by about 1 d B under the 7% hard decision forward error correction(HD-FEC)threshold.更多还原