Comparison of Real- and Complex-Valued NN Equalizers for Photonics-Aided 90-Gbps D-band PAM-4 Coherent Detection
Comparison of Real- and Complex-Valued NN Equalizers for Photonics-Aided 90-Gbps D-band PAM-4 Coherent Detection
复制标题
用于光子学辅助 90 Gbps D 频段 PAM-4 相干检测的实值和复值 NN 均衡器的比较
DOI：
10.1109/jlt.2021.3109126
复制
发表时间：
2021-11-01
影响因子：
4.7
通讯作者：
Yu, Jianjun
中科院分区：
文献类型：
--
作者：
Wen Zhou;Shi, Junting;Yu, Jianjun
5G将100 GHz以下定义为毫米波段，而100 GHz - 3 THz在6 G中被归类为THz波段。深度学习（DL）有望实现6 G无线网络的重大范式转变。在本文中，D波段90-Gbps的单通道PAM-4信号的产生和传输超过10公里的SMF和3米的无线链路在140 GHz的可以实现。提出了一种新的复值神经网络均衡器，该均衡器使用“BARRELU”激活函数从接收到的噪声信号中直接恢复PAM-4信号。实验中有三种DSP选择。在方案1中，首先在发射端进行级联多模算法（CMMA）预均衡（pre-EQ），然后在接收端进行频偏估计（FOE）、载波相位恢复（CPR）和实值神经网络（RVNN）均衡。这里，RVNN均衡器包括具有softmax输出层的DNN、两步联合DNN均衡器和LSTM。实验结果表明，基于LSTM的均衡器优于其他真实的基于NN的均衡器，平均0.5至1.5 dB的BER为10−3的数量级。从选项1开始，在其他两个选项中，发射机处的实值CMMA预均衡对于CVNN是意外的。在选项2中，我们仅将联合收割机下变频和CVNN结合起来，在接收机处被视为“纯数据驱动”训练。这种纯数据驱动的CVNN均衡器在提高误码率的同时，也有较大的计算负担，特别是在n 0 = 571和n1 = 200个训练单元的情况下，误码率低至1 × 10−4，一次迭代的时间复杂度达到350000。由于采用了FOE和CPR等传统的面向物理的模型，CVNN在Opt.3中的计算负担得到了显著的减轻。此外，我们比较了性能的CVNN和RVNN的BER决策精度，时间复杂度和接收机灵敏度。在相同复杂度的相同DSP上，在[371-260-1]结构的11000个样本和300个历元下，对DNN和CVNN进行了比较，结果表明CVNN由于保留了相位信息，性能更好。因此，我们认为，联合使用基于模型的，例如，FOE、CPR步骤和复杂的基于DL的技术具有未来6 G无线物理层算法的潜力。
5G defines below 100 GHz as the millimeter-wave bands, whereas 100 GHz - 3 THz is categorized as THz band in 6G. Deep leraning (DL) is expected to enable a significant paradigm shift in 6G wireless networks. In this paper, D-band 90-Gbps single channel PAM-4 signal generation and transmission over 10-km SMF and 3-m wireless link at 140-GHz can be achieved. A novel complex-valued neural network (CVNN) equalizer using ‘ℂReLU’ activation function to directly recover PAM-4 signals from received noised signals is demonstrated. There are three DSP options in the experiment. In Opt.1, firstly conducted cascaded multi-modulus algorithm (CMMA) pre-equalization (pre-EQ) at transmitter, then processed via frequency offset estimation (FOE), carrier phase recovery (CPR) and finally real-valued neural network (RVNN) equalization at receiver. Here, the RVNN equalizers include DNN with a softmax output layer, two-step joint-DNN equalizer and LSTM. The experimental results show that LSTM-based equalizer outperforms the other real NN-based equalizers by average 0.5 to 1.5 dB at BER of 10−3 magnitude. Differently from Opt. 1, real-valued CMMA pre-EQ at transmitter is unexpected for CVNN in the other two options. In Opt. 2, we only combine down-conversion and CVNN regarded as ‘pure data-driven’ training at receiver. This pure data-driven CVNN equalizer improves BER a lot and also has a larger computation burden, especially BER is as low as 1 × 10−4 with n0 = 571 and n1 = 200 training cells and the time complexity reaches 350000 in one iteration. Thanks to the aid of traditional mathematical-oriented models including FOE and CPR, the computation burden of CVNN in Opt. 3 is released significantly. Furthermore, we compare the performance of CVNN and RVNN in terms of BER decision accuracy, time complexity and receiver sensitivity. Followed by the same DSP with the same complexity, the comparison result between DNN and CVNN in the same structure of [371-260-1] with 11000 samples and 300 epochs shows that CVNN performs better due to its reservation of phase information. Therefore, we believe that the joint use of model-based, e.g., FOE, CPR steps and complex DL-based techniques has a potential for the future 6G wireless physical layer algorithms.