Optimized Machine Learning Techniques for Quality of Transmission Assessment and Fiber Optic Sensing in Optical Networks

Abstract

In recent decades, optical transmission systems have undergone a revolutionary trans formation, driven by the surging demands of global internet traffic and bandwidth intensive applications. Technologies like software-defined networking (SDN) and elastic optical networks (EON) play a pivotal role in optimizing network infrastructures through dynamic resource allocation, enhancing reliability, data rates and capacity. However, deploying SDN and EON introduces complexity in optical network design. Traditional analytical approaches, requiring precise system knowledge, often yield underutilized net work resources and increased costs. To address these challenges, researchers are shifting towards intelligent, autonomously managed optical networks using machine learning (ML) techniques to uncover patterns in massive, dynamic networks. This thesis focuses on improving optical network performance by using ML for quality of transmission (QoT) prediction of lightpaths in software-defined optical networks before deployment. The metric of generalized signal-to-noise ratio (GSNR) is utilized, target ing uncertainties from amplifier gain and noise figure fluctuations. Enhanced GSNR predictions enable reliable lightpath deployment, reducing margins and maximizing re source utilization. We assess the proposed QoT estimation techniques using synthetic datasets from the European (EU) and United States (USA) networks. These datasets are generated using an open-source GNPy simulation tool. In our preliminary research, supervised learning solutions are developed to estimate lightpath QoT using techniques like decision tree, K-nearest neighbours, random for est, linear support vector regression, linear regression, artificial neural networks (ANNs) and convolutional neural networks (CNNs). Simulation data is utilized to train models that reduce system margins. To improve model generalization, transfer learning (TL) is incorporated in ANN, leveraging insights from a fully operational network to aid in QoT estimation within a new extended C-band network adhering to 400ZR standards. Using a transfer learning approach, we’ve achieved an average reduction of 0.63 dB in the deployed margin. To further extend this work, the correction of uncertainties in GSNR in a recently implemented extended C-band network is investigated using two TL techniques, transfer learning partial tuning (TLPT) and transfer learning feature extrac tion (TLFE). Our novel contributions include developing a knowledge distillation (KD) based model for accurate and efficient QoT estimation, crucial for real-time execution in resource-constrained networks. The proposed framework achieved a 2.25% accuracy boost over the baseline model for the EU network and a 3.30% improvement for the USA network compared to the baseline model trained conventionally from scratch. This tailored solution is designed for implementation in real networks with limited computa tional resources. Building upon this foundation, the subsequent proposal introduces a novel framework that synergistically combines knowledge distillation and transfer learn ing, further advancing our pursuit of optimized QoT estimation methodologies. Our findings suggest that the proposed lightweight model utilizing KD+TL outperforms the traditional DNN-based models for GSNR prediction achieving a mean squared error (MSE) of 0.023 dB. In comparison, the standalone KD method yielded an MSE of 0.25 dB, while the standalone TL method resulted in an MSE of 0.12 dB. Furthermore, this integrated framework is smaller, with improved accuracy and faster evaluation speed, making it a potential choice for performing real-time network operations. Furthermore, we leverage machine learning techniques in fiber optic sensing (FOS) to address road traffic and earthquake detection. Using state-of-polarization measurements from real-world and synthetic datasets, we employ an unsupervised autoencoder-based long short-term memory (LSTM) approach to automate road traffic pattern recognition. The model training is performed using data acquired from the metropolitan fiber cable in Turin, achieving a remarkable 97% accuracy in detecting and computing traffic in stances per hour. Additionally, we propose a smart grid fibre sensing technique to enable early earthquake alerts that makes use of a bidirectional gated recurrent unit (Bi-GRU) enhanced with an attention mechanism, utilizing realistic synthetic earthquake waves. The developed model exhibits an impressive ability to detect P-waves in less than a second with an accuracy of approximately 97%. Overall, this thesis introduces optimized machine-learning solutions to unleash the potential of autonomous, intelligent optical network operation and management.