Detección de "soft-failures" mediante la implementación de algoritmos basados en Machine Learning.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Las redes de fibra óptica, esenciales en las telecomunicaciones modernas, enfrentan un desafío notable con las soft-failures, degradaciones graduales que comprometen la calidad de la transmisión sin interrupciones abruptas. Su detección tradicional es compleja, y costosa cuando se convierten en hard-failures, requiriendo intervención manual y prolongando los tiempos de diagnóstico. Las SF pueden originarse por el envejecimiento de componentes, desalineaciones espectrales o efectos no lineales, impactando levemente la calidad de servicio. A pesar de esto, el Machine Learning surge como una solución prometedora para automatizar su tratamiento.
Este trabajo propone y valida experimentalmente un sistema de detección de soft failures en un enlace óptico de 200 km, diseñado para ser robusto y adaptable a la dinámica de la red. Se simularon estados operativos normales y tres tipos de soft failures representativas: ganancia reducida en amplificadores, atenuación en el span e interferencia no lineal. Los datos de BER, OSNR y potencia fueron preprocesados extrayendo sus características estadísticas (media y desviación estándar) de ventanas de 3 minutos de operación, y normalizándolos para asegurar su transferibilidad. Se evaluaron tres algoritmos de detección de anomalías no supervisados: Isolation Forest, DBSCAN y Autoencoder.
Los resultados demostraron que el Autoencoder alcanzó la mayor precisión en la detección, logrando un 100% con datos de modulación DP-QPSK. Lo más notable es que este mismo modelo, mantuvo un rendimiento excepcional con una precisión del 100% al ser evaluado con datos de modulación DP-16-QAM, sin necesidad de re entrenamiento. Demostrando una detección de fallas robusta y agnóstica al formato de modulación, siendo un prometedor avance para la implementación de sistemas de gestión de red autónomos y fiables en entornos operativos dinámicos.
Optical fiber networks, essential in modern telecommunications, face a significant challenge with soft-failures, gradual degradations that compromise transmission quality (increasing BER and reducing OSNR) without abrupt interruptions. Their traditional detection is complex and costly, requiring manual intervention and prolonging diagnosis times. Soft-failures can originate from component aging, spectral misalignments, filter failures, or nonlinear effects, significantly impacting quality of service. Despite these challenges, Machine Learning emerges as a promising solution to automate this process. This work proposes and experimentally validates a soft-failure detection system on a 200 km optical link, designed to be robust and adaptable to network dynamics. Normal operational states and three representative SF types were simulated: reduced gain in amplifiers, span attenuation, and nonlinear interference. BER, OSNR, and received power data were preprocessed by extracting their statistical features (mean and standard deviation) from 3-minute operation windows , and normalized to ensure pattern transferability. Three unsupervised anomaly detection algorithms were evaluated: Isolation Forest, DBSCAN, and Autoencoder. Experimental results demonstrated that the Autoencoder model achieved the highest accuracy in soft-failure detection, reaching 100% accuracy with DP-QPSK modulation data. Most notably, this same model, pre-trained exclusively with DP-QPSK, maintained exceptional performance with a 100% accuracy when evaluated with DP-16 QAMmodulation data, without the need for re-training. This capability demonstrates robust and modulation-agnostic fault detection, representing a promising advance for implementing autonomous and reliable network management systems in dynamic operational environments.
Optical fiber networks, essential in modern telecommunications, face a significant challenge with soft-failures, gradual degradations that compromise transmission quality (increasing BER and reducing OSNR) without abrupt interruptions. Their traditional detection is complex and costly, requiring manual intervention and prolonging diagnosis times. Soft-failures can originate from component aging, spectral misalignments, filter failures, or nonlinear effects, significantly impacting quality of service. Despite these challenges, Machine Learning emerges as a promising solution to automate this process. This work proposes and experimentally validates a soft-failure detection system on a 200 km optical link, designed to be robust and adaptable to network dynamics. Normal operational states and three representative SF types were simulated: reduced gain in amplifiers, span attenuation, and nonlinear interference. BER, OSNR, and received power data were preprocessed by extracting their statistical features (mean and standard deviation) from 3-minute operation windows , and normalized to ensure pattern transferability. Three unsupervised anomaly detection algorithms were evaluated: Isolation Forest, DBSCAN, and Autoencoder. Experimental results demonstrated that the Autoencoder model achieved the highest accuracy in soft-failure detection, reaching 100% accuracy with DP-QPSK modulation data. Most notably, this same model, pre-trained exclusively with DP-QPSK, maintained exceptional performance with a 100% accuracy when evaluated with DP-16 QAMmodulation data, without the need for re-training. This capability demonstrates robust and modulation-agnostic fault detection, representing a promising advance for implementing autonomous and reliable network management systems in dynamic operational environments.
Description
Tesis presentada para optar al título de Ingeniero/a Civil en Telecomunicaciones.
Keywords
Fibras ópticas, Algoritmos computacionales, Aprendizaje de máquina