RNN-Based Demodulation Framework for Spectrum Narrowing in Ultra-Dense WDM Networks

Abstract

To cope with the continuous traffic growth, spectral efficiency needs to be enhanced. Ultra-dense WDM systems, where narrower guard-bands are inserted between wavelength-division-multiplexed (WDM) signals than those used in conventional dense WDM (DWDM) systems, are expected to offer higher spectral efficiency. However, such systems are hindered by the spectrum narrowing induced by wavelength selective switches (WSSs) at nodes. As a result, intersymbol interference (ISI) can be excessive. Linear digital filters widely used in typical digital coherent receivers can equalize the ISI. However, ISI equalization yields noise enhancement due to the interaction between ISI and amplifier noise. Sequence estimation can alleviate the impact of this interaction; however, to introduce sequence estimation into real systems, an ISI-imposing filter is needed because the adaptive filter used for polarization recovery automatically equalizes the ISI. Although the optimum ISI-imposing filter can be created with the proper spectrum model, its characteristics change frequently in real optical path networks. Therefore, we need to develop a demodulation framework that does not necessitate any spectrum models. In this letter, we propose a novel demodulation framework that uses a recurrent neural network (RNN) to simultaneously realize the ISI-imposing filter and sequence estimator. Extensive computer simulations show that our proposal increases the transmission distance in ultra-dense WDM networks.