In recent years, equalizer based on Neural Network (NN) has been widely used in optical fiber nonlinear impairment compensation. However, in practical application, it needs to consume a lot of resources to retrain NN to adapt to optical communication system in new environment. Transfer learning applies some parameters of the NN model trained by the initial system (source domain) to the NN model in the new environment (target domain). Only a small amount of training data is needed to achieve rapid reconstruction of the target domain model. However, this method needs to find the best source domain in all source domains for migration to obtain good performance. When the target domain changes, it is necessary to find the best source domain again, which will consume a lot of training resources. This work suggests a solution based on multi-source domain transfer learning to solve this issue.

This method employs Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) as equalizers. It alternately updates network parameters through two processes: specific source domain training and multi-source domain training. Subsequently, the optical communication system in the new environment is fine-tuned, allowing it to adapt to changes in the transmission system using only a small amount of initial training data. Moreover, good performance can be achieved without the need to search for the optimal source domain.

A 5-channel 50-GBaud Wavelength Division Multiplexing (WDM) Dual-Polarization 16-order Quadrature Amplitude Modulation (DP-16QAM) optical transmission system is simulated to verify the effectiveness of the proposed method. The numerical simulation results show that the multi-source domain transfer learning outperforms the retraining method when using just 35% of the target domain data. Meanwhile, the Q-factor of multi-source domain transfer learning are improved by 0.82 and 0.18 dB, respectively, in compared with retraining and single source domain transfer learning.

Therefore, the multi-source transfer learning scheme is suitable for practical optical communication systems.