OSNR Estimation Providing Self-Confidence Level as Auxiliary Output From Neural Networks

Abstract

Accurate optical monitors are critical for automating operations of fiber-optic networks. Deep neural network (DNN) based optical monitors have been investigated as accurate optical monitors to leverage a large amount of data obtained from fiberoptic networks. Although DNN-based optical monitors have been trained and tested to ensure the given accuracy criteria, this does not ensure sufficient accuracy under unexpected conditions, that is, out of test conditions, e.g., a newly developed modulation format that is not included in the test dataset. Thus, it is necessary to prepare a monitor to assess the current accuracy of a DNN-based optical monitor's output for robust automation of networks. We present a DNN-based optical monitor that simultaneously outputs an optical signal-to-noise ratio and its uncertainty information using a dropout method at the inference phase. This monitor was evaluated in cases in which the DNNs were trained with either a limited number of records or partially missing records in a training dataset. The proposed monitor successfully informed that own output has large uncertainties due to a limited amount of training data or a missing part in training dataset. Additionally, to improve an accuracy of estimated uncertainty, the number of partial neural networks by dropout at the inference phase was optimized. This is a valuable step toward designing robust "self-driving" optical networks.

Link State Aware Dynamic Routing and Spectrum Allocation Strategy in Elastic Optical Networks

“…Besides, asynchronous sampling is used to preprocess the data, then the data is inputted to the convolutional neural network (CNN) for OSNR estimation [36], [37]. To improve the robustness of parameter estimation, Tanimura et al propose an OSNR estimation model based on DNN, while assessing the current accuracy and providing the uncertainty information [38]. Besides these various neural network models, LightGBM [39] and XGBoost [40] models can also achieve the parameters estimation.…”