QoT- Estimation Assisted by Transfer learning in Extended C-band Network Operating on 400ZR

We propose a transfer learning-based technique that assists in estimating the Quality-of-transmission (QoT) of the lightpaths in an extended C-band network on 400ZR. The proposed scheme develops the cognition using the traditional C-band operating network knowledge.

I. Introduction

The modern optical network exploits various technologies such as Elastic optical networks (EONs) and Software-defined networking (SDN) to allow the dynamic and adaptive pro-visioning of network resources. The introduction of these technologies paved a path towards partially or fully disag-gregated optical networks. The main step towards flexible and disaggregated optical networks is to provide the abstraction of the WDM transport as a topology graph weighted by the Gen-eralized signal-to-noise ratio (GSNR) degradation on transpar-ent Lightpaths (LPs) introduced by each traversed Network element (NE), mainly by Optical line systems (OLSs) [1]. Typically, the OLS controller [2] sets the amplifier operating point and subsequently defines the GSNR degradation. The more correct the nominal operating point is set, the more is the potential to depend on the overall LP GSNR. Thus, a smaller system margin is demanded, and, subsequently, larger traffic can be deployed. The NEs are mainly influenced by variations on the working point vs. the nominal values due to the aging effect, change in spectral load, and different effects of infield operations. These induced fluctuations create a difference between the actual vs. the nominal GSNR computed by the QoT-Estimation engine [3]. The main sources of GSNR uncertainties are ripples on amplifiers' gain and Noise figure (NF), to cater to these uncertainties, a system margin must be deployed to avoid network Out-of-service (OOS) [4]. In this work, we propose a Transfer learning (TL) scheme utilizing the dataset of the traditional C-band fully operational network to train a Machine learning (ML) agent operating together with a reliable QoT- Estimation engine in the network controller of the extended C-band sister network. The ML agent's scope is to correct the GSNR uncertainties due to Erbium-doped fiber amplifiers (EDFAs) ripples and spectral load dependence, for LP on the extended C-band sister network whose nominal NE parameters have been perturbed to include a realistic degree of uncertainty reduced by the TL scheme. The two considered networks have different topologies based on the same hard-ware: fiber type and EDFAs. The perturbed uncertainty in this work is only EDFA ripples and varying spectral load.