Deep Neural Network for Joint Nonlinearity Compensation and Polarization Tracking in the Presence of PDL

This article presents a novel deep linear and nonlinear compensation network (DLNCN) that effectively addresses linear and nonlinear distortion in conjunction with polarization-dependent loss (PDL), while also accounting for changes caused by the joint impact of PDL and time-varying rotation of the state of polarization (RSOP). To accomplish this, we introduce neural network layers dedicated for PDL compensation, and we devise a transfer learning approach that selectively updates weights in layers affected by the variations while keeping the remaining weights unchanged. To monitor RSOP with PDL, we employ a pilot-based acquisition and a pilot-aided decision-directed tracking technique. Our numerical tests demonstrate successful RSOP tracking in the presence of PDL impairments, outperforming state-of-the-art schemes by an average of over 0.75 dB in Q-factor for a dual-polarized 960 km 32 Gbaud 64-QAM transmission with a polarization linewidth of 3 kHz. These results highlight the effectiveness of our proposed deep neural network structure, which includes a dedicated layer for PDL compensation, and its ability to work seamlessly with RSOP tracking.