Fiber nonlinear impairments compensation using CLDNN with transfer learning for long-haul fiber-optic transmission systems

Abstract

Fiber nonlinear impairments (NLIs) are one of the major limitations for long-haul fiber-optic transmission systems. In this paper, we propose a low-complexity NLIs compensation (NLC) scheme using CLDNN with transfer learning (TL). The proposed NLC scheme combines the advantages of deep neural networks (DNN), convolutional neural networks (CNN), and long short-term memory (LSTM). To reduce training complexity and accelerate the NLIs estimation process, we introduce the TL based on multi-source domain to fast remodel target tasks by learning knowledge from the previous tasks. In the experiment, we compare the proposed NLC scheme with w/o NLC, DNN-NLC, and digital back propagation with modified 10 step-per-span (10-StPs MDBP). The experimental results show the proposed NLC scheme can achieve 11.05 dB $Q$ factor at 5 dBm signal-launched power for single-channel 30 GBaud dual-polarization 16QAM (DP-16QAM) coherent fiber-optic transmission system over 1200 km standard single mode fiber, while the $Q$ factor gains are 3.23 dB, 1.76 dB, and 0.61 dB compared to the w/o NLC, DNN-NLC, and 10-StPs MDBP, respectively. Meanwhile, the implementation complexity of the proposed TL-CLDNN-NLC is approximately 92% of DNN-NLC, and 57% of 10-StPs MDBP. Additionally, the experimental results of three-channel wavelength division multiplexing 30 GBaud DP-64QAM systems show that the proposed TL-CLDNN-NLC can also achieve a better performance with lower complexity compared with DNN-NLC and 10-StPs MDBP.