The immense increase in the number of devices generating data and the development of new technologies, such as cloud computing, video streaming, Internet of Things (IoT), and 5G/6G, are drawing network operators to develop Machine Learning (ML) - based solutions to automate the management and control of communication systems. Moreover, these technologies are leading to increasingly stringent requirements for optical networks in terms of capacity, latency, and reliability. One of the promising solutions to meet these challenges is Space-Division Multiplexing (SDM), and in particular the use of Multi-core fibers (MCFs), which improves transmission capacity, scalability, and spectral efficiency. Nevertheless, these new techniques introduce complexity in network monitoring and main- tenance, particularly for timely and accurate fault detection and localization, which are crucial since when an optical network suffers a failure, an immense loss of data will occur. This thesis focuses in localizing physical-layer failures in uncoupled MCF-based optical networks, exploiting ML solutions. Two strategies are explored: a one-step multi-class classification approach, and a two-step approach where failure detection (binary classifi- cation) is followed by failure localization (multi-class classification). This study exploits Quality of Transmission (QoT) metrics, such as Bit Error Rate (BER), Optical Signal-to- Noise Ratio (OSNR), Error Vector Magnitude (EVM), and Throughput, gathered from a real MCF network located in L’Aquila, Italy. We extract statistical features from measured QoT metrics, using temporal sliding windows, and to ensure robustness and generalize the results, we adopt a temporal validation strategy. The experimental results obtained from a realistic dataset simulating different failure scenarios and modulation for- mats, such as 4-QAM, 16-QAM, 64-QAM (directly related to the severity of the failure), show that the proposed solutions can reach up to 96% classification Accuracy. Moreover, we highlight the behavior of the QoT metrics and how they affect the classification, and utilize approaches based on eXplainable Artificial Intelligence (XAI) to quantify which features does the ML classifier consider to take the prediction. The numerical results, as the first of their kind, show that the proposed framework achieves a high Accuracy in both failure detection and localization, demonstrating its potential as a key enabler for intelligent monitoring in future MCF-based optical transport networks.
L’enorme aumento del numero di dispositivi che generano dati e lo sviluppo di nuove tec- nologie, come il cloud computing, lo streaming video, l’Internet of Things (IoT) e le reti 5G/6G, stanno spingendo gli operatori di rete a sviluppare soluzioni basate su Machine Learning (ML) per automatizzare la gestione e il controllo dei sistemi di comunicazione. Inoltre, queste tecnologie stanno portando a requisiti sempre più stringenti per le reti ot- tiche in termini di capacità, latenza e affidabilità. Una delle soluzioni più promettenti per affrontare queste sfide è lo Space-Division Multiplexing (SDM), e in particolare l’uso delle fibre multi-core (MCF), che migliorano la capacità trasmissiva, la scalabilità e l’efficienza spettrale. Tuttavia, queste nuove tecniche introducono complessità nel monitoraggio e nella manuten- zione della rete, in particolare per quanto riguarda il rilevamento e la localizzazione tem- pestiva e accurata dei guasti, aspetti fondamentali considerando che un guasto in una rete ottica può causare enormi perdite di dati. Questa tesi si concentra sulla localizzazione di guasti a livello fisico in reti ottiche basate su MCF non accoppiate, sfruttando soluzioni di ML. Sono state esplorate due strategie: un approccio a classificazione multi-classe diretta (one-step) e un approccio a due stadi (two-step), in cui il rilevamento del guasto (classificazione binaria) è seguito dalla local- izzazione del guasto (classificazione multi-classe). Lo studio sfrutta metriche di qualità di trasmissione (Quality of Transmission, QoT), come il Bit Error Rate (BER), l’Optical Signal-to-Noise Ratio (OSNR), l’Error Vector Magnitude (EVM) e il Throughput, raccolte da una rete MCF reale situata a L’Aquila, Italia. Le metriche QoT vengono elaborate estraendo caratteristiche statistiche mediante finestre temporali scorrevoli e, per garantire la robustezza e la generalizzabilità dei risultati, viene adottata una strategia di validazione temporale. I risultati sperimentali, ottenuti da un dataset realistico che simula diversi scenari di guasto e formati di modulazione come 4-QAM, 16-QAM e 64-QAM (direttamente legati alla severità del guasto), mostrano che le soluzioni proposte possono raggiungere fino al 96% di accuratezza nella classificazione. Inoltre, si analizza il comportamento delle metriche QoT e il loro impatto sulla classificazione, e si utilizzano approcci basati su eXplainable Artificial Intelligence (XAI) per quantificare quali caratteristiche vengono considerate dal classificatore ML nel processo decisionale. I risultati numerici, primi nel loro genere, mostrano che il framework proposto raggiunge un’elevata accuratezza sia nel rilevamento che nella localizzazione dei guasti, dimostrando il suo potenziale come elemento chiave per un monitoraggio intelligente nelle future reti di trasporto ottico basate su MCF.
Data-driven failure localization in multi-core optical networks
Lobaccaro, Pasquale
2024/2025
Abstract
The immense increase in the number of devices generating data and the development of new technologies, such as cloud computing, video streaming, Internet of Things (IoT), and 5G/6G, are drawing network operators to develop Machine Learning (ML) - based solutions to automate the management and control of communication systems. Moreover, these technologies are leading to increasingly stringent requirements for optical networks in terms of capacity, latency, and reliability. One of the promising solutions to meet these challenges is Space-Division Multiplexing (SDM), and in particular the use of Multi-core fibers (MCFs), which improves transmission capacity, scalability, and spectral efficiency. Nevertheless, these new techniques introduce complexity in network monitoring and main- tenance, particularly for timely and accurate fault detection and localization, which are crucial since when an optical network suffers a failure, an immense loss of data will occur. This thesis focuses in localizing physical-layer failures in uncoupled MCF-based optical networks, exploiting ML solutions. Two strategies are explored: a one-step multi-class classification approach, and a two-step approach where failure detection (binary classifi- cation) is followed by failure localization (multi-class classification). This study exploits Quality of Transmission (QoT) metrics, such as Bit Error Rate (BER), Optical Signal-to- Noise Ratio (OSNR), Error Vector Magnitude (EVM), and Throughput, gathered from a real MCF network located in L’Aquila, Italy. We extract statistical features from measured QoT metrics, using temporal sliding windows, and to ensure robustness and generalize the results, we adopt a temporal validation strategy. The experimental results obtained from a realistic dataset simulating different failure scenarios and modulation for- mats, such as 4-QAM, 16-QAM, 64-QAM (directly related to the severity of the failure), show that the proposed solutions can reach up to 96% classification Accuracy. Moreover, we highlight the behavior of the QoT metrics and how they affect the classification, and utilize approaches based on eXplainable Artificial Intelligence (XAI) to quantify which features does the ML classifier consider to take the prediction. The numerical results, as the first of their kind, show that the proposed framework achieves a high Accuracy in both failure detection and localization, demonstrating its potential as a key enabler for intelligent monitoring in future MCF-based optical transport networks.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Lobaccaro.pdf
accessibile in internet per tutti
Dimensione
9.14 MB
Formato
Adobe PDF
|
9.14 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary___Lobaccaro.pdf
accessibile in internet per tutti
Dimensione
2.1 MB
Formato
Adobe PDF
|
2.1 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/240042