Navegando por Assunto "Aprendizado não-supervisionado"

Agora exibindo 1 - 1 de 1

Acesso aberto (Open Access)
Unsupervised learning algorithms for data-driven fault management in optical networks
(Universidade Federal do Pará, 2024-12-09) RIBEIRO, Andrei Nogueira; LOBATO, Fabrício Rossy de Lima; http://lattes.cnpq.br/6344884902408613; COSTA, João Crisóstomo Weyl Albuquerque; http://lattes.cnpq.br/9622051867672434
Over the past years, the emergence of more complex and bandwidth-hungry applications has charged efforts to ensure the reliability of optical networks. The occurrence of faults, for instance, can directly affect the quality of transmission of these optical systems, leading to several implications, including packet losses and service disruption. Hence, it is vital to mitigate faults in optical networks to guarantee the availability of the system and meet the service level agreement requirements. Moreover, as the complexity of optical networks evolves constantly, machine learning-based approaches have been proposed to deal with the system dynamics while providing automated fault management. In that regard, most proposed approaches are based on supervised learning (SL) models, which require large amounts of fault data to be properly trained. However, data from fault conditions are typically scarce in practical scenarios, which poses limitations for deploying SL-based models. Therefore, this work explores several unsupervised learning algorithms to perform fault management in optical networks. As fault data are absent in several real-world scenarios, unsupervised strategies trained with only data from normal operating conditions can be an effective alternative. These strategies disregard the need for data from abnormal network conditions and thus require much less data for model training. In this work, the fault detection and localization performances of cluster-based algorithms (K-means, Fuzzy C-means, Mahalanobis Squared-Distance-based model, and Gaussian Mixture Model) and dimensionality reduction-based approaches (Principal Component Analysis and Autoencoder) are compared leveraging a dataset derived from an optical testbed. The techniques are evaluated in terms of Type I (false-positive) and Type II (false-negative) error trade-offs. Ultimately, all techniques demonstrated satisfactory fault detection results when trained with only data from normal conditions, achieving an average accuracy of more than 90%. Such results suggest their applicability to real-world optical network fault management scenarios.