986 km Field Trial of Cascaded ANN-based Link-Penalty Models for QoT Prediction

Quality of transmission (QoT) prediction, one of the essential technologies to enable low-margin automatic optical networks, has recently attracted increasing interest among researchers [1]. Analytical model-based QoT estimation, such as the GNPy software, provides system-level optimization for static operations [2]. However, in low-margin dynamic optical networks, QoTs should be predicted based on traffic payload, link configurations, and other parameters to obtain precise QoTs for dynamic configurations with further reduced link margins. In [3], a point-to-point multi-channel QoT prediction based on artificial neural networks(ANN) was proposed, however, lacking end-to-end solutions in terms of training and deployment. P. Safari, etc., developed a network-wide model
for QoT classification based on a deep convolutional neural network (DCNN) model [4]. However, the complex DCNN model strongly depends on network typologies and requires a huge amount of data from the whole network. which makes it infeasible to train and deploy the model in a practical optical network. In this paper, two cascaded multi-channel ANN models are developed, co-trained, and demonstrated to predict link penalties over a 986-km field-trial testbed consisting of two links and ROADMs. With the two-cascaded ANN setting, the QoT of a channel over the two links is predicted considering traffic payloads in two links with precision of ± 0.16 dB. The co-training of the two cascaded penalty prediction models also provides a feasible solution to provide network-level QoT predictions.