Estimating Quality of Transmission in a Live Production Network using Machine Learning

Abstract

We demonstrate QoT estimation in a live network utilizing neural networks trained on synthetic data spanning a large parameter space. The ML-model predicts the measured lightpath performance with <0.5dB SNR error over a wide configuration range.

QoT estimation using EGN-assisted machine learning for multi-period network planning

“…The analytical models are implemented using Python and the Numpy library [26] while ML uses the XGBoost library [21]. In conclusion, the proposed ML-based QoT estimator is orders of magnitude faster than the EGN model with a prediction accuracy that is comparable to previous QoT estimators [18].…”

“…In an attempt to limit the complexity of the parameter space and therefore the number of data points needed for training, note that the WDM grid layouts vary widely while the number of link layouts is limited by the condition of homogeneous span lengths and fixed fiber parameters (standard single-mode fiber). The work can be extended to a more general scenario by considering additional input features as we found [18] that links of heterogeneous span lengths and varying attenuation can be well represented by using combined parameters, such as the average of the cumulative sum of the effective span lengths, as the input into an ML model. This generalization requires a substantially larger data set due to the added degrees of freedom in the parameter space, thereby also increasing the training time.…”

“…To ensure extensive validation, the ML model was tested in a network planning task and full error distributions were considered when evaluating its accuracy on the simulation data. A neural-network-based NLI regression model has been demonstrated to accurately predict QoT in a live production network [18].…”

A QoT Estimation Method using EGN-assisted Machine Learning for Network Planning Applications

“…Note that the data set has widely varying WDM grid layouts but only a limited number of link layouts. We found [10] that links of heterogeneous span lengths and varying attenuation can be well represented by using combined parameters, such as the average of the cumulative sum of the span lengths, as the input into an ML model.…”

“…Regression approaches, on the other hand, have been used for Q-factor prediction of multiple channels on a testbed link [7] , in the context of modeling parameter uncertainty [8] , and recently for the estimation of a generalized signal-to-noise ratio (GSNR) 978-1-6654-3868-1/21/$31.00 ©2021 IEEE distribution, assuming imperfect representation of physical parameters by the ML features [9] . A neural-network-based NLI regression model has been demonstrated to accurately predict QoT in a live production network [10] .…”