Cheng Yijun, Yang Zheng, Yan Zhijun, Liu Deming, Fu Songnian, Qin Yuwen
Opt Lett. 2022 May 1;47(9):2218-2221. doi: 10.1364/OL.456877.
We experimentally demonstrate meta-learning-enabled accurate optical signal-to-noise ratio (OSNR) monitoring of directly detected 16QAM signals with extremely few training data. When one-shot training, where one amplitude histogram (AH) for each OSNR value includes only 2000 data samples, is implemented for a 16QAM signal within a variable OSNR range of 15-24 dB, the experimental root mean squared error (RMSE) of the retraining technique is 1.53 dB. For transfer learning from the 16QAM simulation to the experimentally generated AH, the RMSE can be reduced to 1.11 dB. In comparison with both the retraining and transfer learning techniques, the RMSE of meta-learning-enabled OSNR monitoring can be further reduced by 42.8% and 22.3%, respectively. In order to reach the optimal accuracy with an RMSE of 0.66 dB, the meta-learning technique requires only 15 AHs for each OSNR value to be monitored, while the retraining and the transfer learning techniques need 20 and 25 AHs, respectively.
我们通过实验证明了，利用元学习能够在极少训练数据的情况下，对直接检测的16QAM信号进行精确的光信噪比（OSNR）监测。当在15 - 24 dB的可变OSNR范围内对16QAM信号进行一次性训练时，每个OSNR值的一个幅度直方图（AH）仅包含2000个数据样本，再训练技术的实验均方根误差（RMSE）为1.53 dB。对于从16QAM模拟到实验生成的AH的迁移学习，RMSE可降至1.11 dB。与再训练和迁移学习技术相比，基于元学习的OSNR监测的RMSE可分别进一步降低42.8%和22.3%。为了达到0.66 dB的RMSE的最佳精度，元学习技术对于每个要监测的OSNR值仅需要15个AH，而再训练和迁移学习技术分别需要20个和25个AH。
