Using machine learning in an open optical line system controller

Abstract

The reduction of system margin in open optical line systems (OLSs) requires the capability to predict the quality of transmission (QoT) within them. This quantity is given by the generalized signal-to-noise ratio (GSNR), including both the effects of amplified spontaneous emission (ASE) noise and nonlinear interference accumulation. Among these, estimating the ASE noise is the most challenging task due to the spectrally resolved working point of the erbium-doped fiber amplifiers (EDFAs), which depend on the spectral load, given the overall gain profile. An accurate GSNR estimation enables control of the power optimization and the possibility to automatically deploy lightpaths with a minimum margin in a reliable manner. We suppose an agnostic operation of the OLS, meaning that the EDFAs are operated as black boxes and rely only on telemetry data from the optical channel monitor at the end of the OLS. We acquire an experimental data set from an OLS made of 11 EDFAs and show that, without any knowledge of the system characteristics, an average extra margin of 2.28 dB is necessary to maintain a conservative threshold of QoT. Following this, we applied deep neural network machine-learning techniques, demonstrating a reduction in the needed margin average down to 0.15 dB.