Accurate quality of transmission estimation with machine learning

In optical transport networks the quality of transmission (QoT) is estimated before provisioning new connections or upgrading existing ones. Traditionally, a physical layer model (PLM) is used for QoT estimation coupled with high margins to account for the model inaccuracy and the uncertainty in the evolving physical layer conditions. Reducing the margins increases network efficiency but requires accurate QoT estimation. We present two machine learning (ML) approaches to formulate such an accurate QoT estimator. We gather physical layer feedback, by monitoring the QoT of existing connections, to understand the actual physical conditions of the network. These data are used to train either the input parameters of a PLM or a machine learning model (ML-M). The proposed ML methods account for variations and uncertainties in equipment parameters, such as fiber attenuation, dispersion, and nonlinear coefficients, or amplifier noise figure per span, which are typical in deployed networks. We evaluated the accuracy of the proposed methods under various uncertainty scenarios and compared them to QoT estimators proposed in the literature. The results indicate that our estimators yield excellent accuracy with a relatively small amount of data, outperforming other prior estimators.