Time series anomaly detection and diagnosis via manifold learning and normalizing flows

• Ano de defesa: 2024
• Autor(a) principal: Dias, Madson Luiz Dantas
• Orientador(a): Mattos, César Lincoln Cavalcante
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.ufc.br/handle/riufc/81121

The emergence of data collection and storage technologies has allowed the accumulation of extensive data over time. Consequently, high-dimensional time series data sets have become prevalent across diverse fields, including sensor networks, security, healthcare, manufacturing, and finances. Although it represents a substantial challenge, detecting rare events in such data sets is a fundamental task in several applications, including cyber-intrusion detection, defect analysis, fault detection, credit card fraud analysis, suspicious trajectory detection, etc. Various methods have been employed to handle this challenge, ranging from traditional approaches such as classification, clustering, distance metrics, density estimation, and statistical techniques to more contemporary solutions involving deep learning models. In this thesis, we provide a comprehensive overview of the field, present state-of-the-art methods of multivariate time series anomaly detection and diagnosis, and introduce two new approaches for solving problems involving anomaly detection in time series. The first one called aggregated anomaly detection with normalizing flows (GRADINGS) is a framework for anomaly detection in time series database, applied to trajectory data that is based on estimating the density for each trajectories segments and aggregating segments’ likelihoods into a single anomaly score. Such a strategy enables the handling of possibly large sequences of different lengths. The second approach called robust anomaly detection on multivariate time series (RANDOMS), which uses normalization flows and manifold learning techniques to solve the anomaly detection and diagnosis problems. Extensive evaluations of the proposed methods have been conducted across various applications, comparing them against several models. The results of our computational experiments demons- trate the efficacy of our approaches, consistently outperforming existing state-of-the-art anomaly detection and diagnosis methods in numerous cases.