用于带限DML-IMDD系统中的深度神经网络均衡器简化方案

Simplified scheme of deep neural network equalizer for limited-bandwidth DML-IMDD system

为提升基于带限光电器件及直接调制激光器(DML)的强度调制直接检测(IMDD)系统性能,解决传统均衡器计算复杂度过高的问题,提出深度神经网络(DNN)均衡器简化方案。首先,利用自适应动量估计(Adam)算法更新DNN的权重系数,优化了传统梯度下降算法的迭代速度和收敛性能;然后,在此基础上引入丢弃层和剪切操作以降低DNN的高计算复杂度,减少网络结构的冗余连接,并避免过拟合现象的产生。最后,在80 Gb/s带限DML-IMDD仿真系统中验证了DNN均衡器简化方案的有效性和可行性。

In order to improve the performance of intensity modulation direct detection(IMDD)systems based on band-limited optoelectronic devices and direct modulation lasers(DML),solve the problem of high computational complexity of traditional equalizers,a simplified scheme for deep neural network(DNN)equalizers is proposed.Firstly,the adaptive momentum estimation(Adam)algorithm is used to update the weight coefficients of DNN,optimizing the iteration speed and convergence performance of traditional gradient descent algorithm.Then,based on this,discard layers and pruning operations are introduced to reduce the high computational complexity of DNN,reduce redundant connections in the network structure,and avoid overfitting phenomenon.Finally,the effectiveness and feasibility of the simplified scheme of DNN equalizer are verified in an 80 Gb/s bandlimited DML-IMDD simulation system.