Improving the adaptability of the optical performance monitor by transfer learning

Owing to their mighty fitting ability, the supervised learning-based deep-learning (DL) models have been widely used in the area of optical performance monitoring (OPM) to improve the optical monitors' performance. However, the supervised learning-based DL models used in OPM are based on two important premises. The first premise is enough training data with labels; the second premise is the same distribution of the training and test data. Nevertheless, it is hard to meet the two premises in the real-world environment where the optical performance monitors are deployed, since the data are unlabeled and the optical network environment is dynamic (such as component aging caused by slow parameter variation), causing the degradation of the monitoring performance. This is because the supervised learning-based DL models lack the adaptability of the dynamic environment. For the purpose of improving the optical performance monitors' adaptability, we propose a transfer-learning-based convolutional neural network model to maintain the monitoring performance in the dynamic optical network environment. The transfer-learning method can transfer the learned knowledge from the labeled data under an invariant optical network environment to the unlabeled data under a dynamic optical network environment. During the training phase, the maximum mean discrepancy (MMD) is applied to match the features extracted from the source and the target domains. When the trained model is deployed in the OPM monitor, the robustness of the system to the dynamic environment would be enhanced. Four signals (60/100 Gbps 16/64 QAM) under different working environments are used to verify the adaptability of the method. The influence of the MMD's weight rates, batch size, and weight parameters confirmed the effectiveness of our method.