MACHINE LEARNING BASED NONLINEAR CHARACTERIZATION IN HETEROGENEOUS OPTICAL NETWORKS

Abstract

Optical networks have seen exponential increases in data traffic in recent years due to advanced bandwidth hungry communication services such as 5G, the internet of things, and cloud computing. Such increases cause a great demand for increased capacities within optical networks, where the major factor limiting achievable capacity is fiber nonlinearities. As such, an essential aspect of optical transmission networks is the ability to characterize and subsequently compensate for the effects of fiber nonlinearities. This work investigates the use of analytical, numerical, and machine learning based models in the characterization of the effects of nonlinear distortions.

A novel method for building a machine learning based nonlinear signal-to-noise ratio estimator is presented. The proposed model building method aggregates important features from three ensemble models with boosting and shows universal application to all considered training cases. A vast set of heterogeneous system configurations are considered for model training and demonstrate improved model generalization without the need for frequent retraining.

A comparative analysis is presented which investigates the agreement of three approximate XPM models in terms of modeling the characteristics of XPM induced phase noise. From this analysis, it is shown that simpler approximate models, while attractive for machine learning dataset generation, struggle when modeling the statistical characteristics of XPM induced phase noise. Investigation into the symbol pattern dependence, manifestation of artifacts in approximate models, and alterations made to each approximate model targeting their limitations is also explored.

A method is presented to separate the XPM induced phase noise from a cumulative noise observation which can sufficiently separate a representation of the true XPM induced phase noise. The separation method can be used in machine learning model dataset generation to improve model applicability to practical systems. Results from this work show promise in characterizing nonlinear distortions to assist machine learning based nonlinear compensation models in optical networks.