Modèles de classification et de prédiction de performance des signaux optiques WDM basés sur les algorithmes d’apprentissage machine

Abstract

The constant growth of the traffic is forcing telecom operators to deploy Wavelength-Division Multiplexing (WDM) optical transmission systems with constantly evolving data rates, capacity, and flexibility. However, as these systems carry a multitude of applications (video, cloud, Internet of Things …); the impact of performance degradations is becoming increasingly important. A potential solution would be to implement proactive network monitoring and management tools that can exploit performance metrics collected on the network. In this context, Machine Learning (ML) is gradually positioning itself as a promising solution to manage the complexity and scalability of heterogeneous optical networks. Indeed, tools based on ML techniques would be very useful in the physical layer of optical networks in a Software Defined Network (SDN) context, in particular for performance prediction. This thesis aims at proposing new methods based on ML algorithms that can be integrated into network control system in order to manage the performance of the lightpaths deployed in optical networks which carry traffic up to several Terabits per second. In addition, each of the proposed methods is validated by simulations using field performance data collected from lightpaths deployed in production networks. Aiming at addressing different problems, these new proposed methods are presented in three parts of the thesis. In the first part, the proposed ML method is a classifier built with Bit Error Rate (BER) data from a lightpath carried in a production network during a 31-day observation period. The classifier, based on the K-nearest neighbor (K-NN) algorithm, aims at characterizing the BER quality level and addresses the problem of estimating the quality of transmission (QoT) of an established lightpath. Moreover, in this part, a module for assessing the impact of different features used for lightpath QoT estimation has been implemented using the Support Vector Machine (SVM) algorithm. In the second part, the proposed ML models aim at answering the problem of predicting the performance, namely the BER and signal-to-noise ratio (SNR) of established lightpaths, based on historical field data. The models are built using two variants of neural networks, namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). This part is divided into three sections. The first section presents univariate SNR forecast models using historical lightpath BER data only to implement the SNR forecast module. The second section presents multivariate SNR forecast models using both historical BER data and additional features such as WDM channel power, temperature and others in the forecast module. The third section explores the potential of transfer learning by evaluating the multivariate and univariate SNR forecast models with field data from a different set of deployed lightpaths. Finally, in the third part, the proposed ML method is an anomaly detection tool based on the SVM and the Inter Quartile Range (IQR) methods. First, the tool defines and extracts the features that characterize the anomalies observed in the SNR time series of several lightpaths carried in a production network. Then, the tool performs an early detection of the anomalies in the SNR time series. The anomaly detection tool is tested on field performance data collected in a production network.

Date	17 Nov 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Christine Tremblay (Supervisor)