Detección e identificación de “Soft Failures” en enlaces ópticos en base a Machine Learning.
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Esta memoria de título se centra en la investigación, desarrollo e implementación de un algoritmo para la detección e identificación de Soft Failures en enlaces ópticos, utilizando téc- nicas de aprendizaje automático (Machine Learning) y algoritmos matemáticos. El objetivo principal es encontrar alternativas más eficientes y rápidas para llevar a cabo esta tarea. Para alcanzar este objetivo, se realizó una investigación exhaustiva sobre el origen de estos fallos, su comportamiento en el espectro óptico y las técnicas académicas actuales para abordar este problema.
El estudio se divide en cuatro partes: la primera se ocupa del montaje de un conjunto de laboratorio capaz de reproducir estos fallos; la segunda, de la creación de un conjunto de datos lo suficientemente amplio para realizar estudios, abarcando tanto datos experimentales como simulados. La tercera parte se enfoca en la detección de Soft Failures mediante el análisis de la variación del bit error rate (BER) y su comportamiento en escenarios normales y en presencia de anomalías. Finalmente, la cuarta parte aborda la identificación de un Soft Failure a partir del análisis del espectro óptico.
Un aspecto clave de este trabajo fue la elección de los modelos más adecuados, tanto para la detección como para la identificación, que permitan obtener resultados precisos y adaptarse a las diversas situaciones que se presentan en la práctica.
Los resultados de esta investigación indican que, en la etapa de detección, el modelo basado en autoencoder logra una precisión del 100% al identificar datos anómalos en comparación con datos en condiciones normales de operación. Posteriormente, el modelo fue sometido a una revalidación al agregar ruido gaussiano a las muestras de BER normales y, al compararlo con otros modelos matemáticos, mostró una mayor sensibilidad y, por ende, un mejor rendimiento en presencia de estos nuevos datos. Finalmente, en la etapa de identificación, una red neuronal convolucional en una dimensión alcanzó una precisión del 100% en la clasificación de Soft Failures y demostró un rendimiento superior frente a otros modelos de Machine Learning en distintos escenarios donde se modificaron algunos parámetros del enlace.
This thesis focuses on the research, development, and implementation of an algorithm for the detection and identification of Soft Failures in optical links, using machine learning techniques and mathematical algorithms. The main goal is to find more efficient and faster alternatives for carrying out this task. To achieve this goal, a thorough investigation was conducted into the origins of these failures, their behavior in the optical spectrum, and current academic techniques for addressing this problem. The study is divided into four parts: the first deals with setting up a laboratory environment capable of recreating these failures; the second involves creating a sufficiently large dataset for conducting studies, incorporating both experimental and simulated data. The third part focuses on detecting a Soft Failure by analyzing the variation in the bit error rate (BER) and its behavior in normal conditions and in the presence of anomalies. Finally, the fourth part addresses the identification of a Soft Failure through optical spectrum data analysis. A key aspect of this work was the selection of suitable models for both detection and identi- fication, which would yield accurate results and adapt to the diverse situations encountered in practice. The results of this research indicate that, in the detection stage, the autoencoder-based model achieves 100 % accuracy in identifying anomalous data compared to data under normal operating conditions. Subsequently, the model underwent revalidation by adding Gaussian noise to the normal BER samples. When compared to other mathematical models, it showed greater sensitivity and, therefore, better performance in the presence of these new data. Finally, in the identification stage, a one-dimensional convolutional neural network achieved 100 % accuracy in classifying Soft Failures and demonstrated superior performance compared to other Machine Learning models in various scenarios where some link parameters were changed.
This thesis focuses on the research, development, and implementation of an algorithm for the detection and identification of Soft Failures in optical links, using machine learning techniques and mathematical algorithms. The main goal is to find more efficient and faster alternatives for carrying out this task. To achieve this goal, a thorough investigation was conducted into the origins of these failures, their behavior in the optical spectrum, and current academic techniques for addressing this problem. The study is divided into four parts: the first deals with setting up a laboratory environment capable of recreating these failures; the second involves creating a sufficiently large dataset for conducting studies, incorporating both experimental and simulated data. The third part focuses on detecting a Soft Failure by analyzing the variation in the bit error rate (BER) and its behavior in normal conditions and in the presence of anomalies. Finally, the fourth part addresses the identification of a Soft Failure through optical spectrum data analysis. A key aspect of this work was the selection of suitable models for both detection and identi- fication, which would yield accurate results and adapt to the diverse situations encountered in practice. The results of this research indicate that, in the detection stage, the autoencoder-based model achieves 100 % accuracy in identifying anomalous data compared to data under normal operating conditions. Subsequently, the model underwent revalidation by adding Gaussian noise to the normal BER samples. When compared to other mathematical models, it showed greater sensitivity and, therefore, better performance in the presence of these new data. Finally, in the identification stage, a one-dimensional convolutional neural network achieved 100 % accuracy in classifying Soft Failures and demonstrated superior performance compared to other Machine Learning models in various scenarios where some link parameters were changed.
Description
Tesis presentada para optar al título de Ingeniero Civil en Telecomunicaciones
Keywords
Aprendizaje de máquina, Comunicación por fibra óptica, Fibras ópticas Mediciones