Machine Learning Applications and Deployments in Optical Networks

Abstract

The growing demand for higher data capacity and efficient fibre network management has driven the evolution towards dynamic and low-margin optical networks, which aim to maximize resource utilization while operating closer to performance limits. Achieving such networks requires real-time monitoring and telemetry, accurate quality of transmission (QoT) prediction, and rapid model updates to adapt to network changes. Traditional QoT prediction models based on mathematical analysis are inadequate for modern networks due to their reliance on numerous input parameters and computational inefficiency. As a result, machine learning (ML) has emerged as a promising alternative. Long Short-Term Memory (LSTM) models, with their ability to capture nonlinear dynamics in time-series data, have proven effective for short-term performance forecasting. Similarly, Artificial Neural Networks (ANNs) are widely used for multi-channel QoT prediction, excelling in modelling complex nonlinearities and dynamic changes. However, training separate ML models for each scenario is time-intensive and costly, highlighting the need for advanced techniques like transfer learning (TL) and domain adaptation (DA) to improve model generalization. Another challenge in deploying ML is the limited availability of real-time data for model training and updating. This challenge is further amplified by the dynamic nature of optical fibre networks, where acquiring large volumes of real-world data is often impractical. Addressing this requires integrated solutions that can generate high-quality synthetic data and manage the entire ML model lifecycle.
This thesis investigates four mainstream ML-based time-series prediction models for lightpath short-term performance forecasting, including Multi-Layer Perceptron, LSTM, Gated Recurrent Unit, and Transformer. Cross-validation results show that LSTM performs well across different datasets, revealing similar link behaviours. Further experiments explore the impact of transmitters' power supply methods and modulation formats on these behaviours. The thesis also proposes a domain adversarial adaptation (DAA) framework that enhances the generalization of ML models across different optical networks, significantly improving prediction accuracy compared to traditional TL methods. Based on the initial narrow-band model, the proposed DAA framework enables predictions for new bands in multi-channel optical systems with wavelength shifts, thereby achieving full-spectrum predictions. Additionally, a unified monitoring and telemetry platform is introduced, designed to integrate real-time data collection with ML model deployment. This platform enables scalable and low-latency data streaming using technologies such as Apache Kafka and InfluxDB. The practical deployment of ML models in optical networks is demonstrated using an OpenFaaS-based AI engine to manage the entire ML lifecycle encompassing training, deployment, inference and monitoring. A key highlight is the implementation of a GNPy-based digital twin, which leverages real-time network configuration information to generate reliable synthetic data for AI engines. This thesis contributes to the development of intelligent optical networks through ML technologies, real-time monitoring and telemetry, and advanced network intelligence to enhance network resilience, efficiency, and scalability.

Date of Award	9 Dec 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Shuangyi Yan (Supervisor) & Dimitra Simeonidou (Supervisor)