Caracterização e previsão de falhas em serviços de conectividade à Internet

Abstract

The Ipê Network is fundamental to the Brazilian scientific community, being responsible for interconnecting universities and research centers throughout the country. The network also presents international connections, allowing Brazilian cooperation with foreign research entities. It is an extensive network, producing a high volume of data and presenting challenges related to its operation. This work is divided into two parts, the first being responsible for presenting an analysis of the network through the characterization of failure behavior. The second attempt consists of constructing learning models to predict the occurrence of failures, allowing for planning on how to mitigate the problems caused by the occurrence of failures. Data is collected through the Via Ipê web app and corresponds to the period of November 2020 through November 2021. The problem is modeled as supervised learning for binary classification and recurrent neural networks (LSTMs) are used. The Ipê Network presents heterogeneous behavior, manifesting great variety on the dependability of its connectivity services in its different PoPs. Different models considering the network’s characteristics are proposed to deal with this scenario, from more general to more restricted models. The models’ performance metrics reveal different types of failures, complementing the initial analysis of the data. The problem is shown to be difficult, but the proposed methodology shows promise, with acceptable results in some cases.