Nuevos enfoques de aprendizaje profundo en la deteccion de anomalias. Aplicaciones

This thesis investigates new deep learning approaches for anomaly detection, with a particular focus on multivariate time series data and reconstruction-based unsupervised methods. Anomaly detection in time series is a critical task in many real-world applications, including industrial monitoring, cybersecurity, healthcare, and large-scale information systems, where abnormal behaviors may indicate faults, attacks, or system degradation. The problem is especially challenging due to the rarity and diversity of anomalies, the high dimensionality of modern data, the presence of noise and non-stationarity, and the limited availability of labeled data. The thesis addresses these challenges through a combination of methodological rigor and model design innovation. First, it provides a comprehensive and unified experimental study of reconstruction-based deep learning methods for time series anomaly detection. A wide range of representative models—covering Autoencoders (AE), Variational Autoencoders (VAE), GAN-based reconstructors, and Transformer-based architectures—are evaluated under a common experimental protocol. All models are tested on heterogeneous multivariate datasets using consistent preprocessing, windowing, and evaluation criteria. To ensure scientific robustness, the comparison is supported by non-parametric statistical tests (Friedman and Nemenyi), allowing performance differences to be assessed beyond anecdotal results. This study clarifies the strengths, weaknesses, and typical failure modes of different reconstruction paradigms across diverse scenarios. Building on the insights obtained from this unified evaluation, the thesis proposes a new Two-Phase anomaly detection framework based on a time–frequency Transformer architecture. The proposed approach combines temporal and spectral representations of time series while preserving timestamp-level anomaly localization. A key contribution is the introduction of a diagnostics-driven design philosophy, in which additional model components are activated only when justified by measurable data characteristics. Specifically, the framework includes three optional modules: (i) an inter-metric embedding mechanism to capture cross-variable dependencies, (ii) a contamination-aware pseudo-filtering strategy to mitigate the effect of anomalous samples within nominal training data, and (iii) an attention-based fusion mechanism to adaptively combine time-domain and frequency-domain anomaly scores. The Two-Phase framework separates diagnostic analysis from model training and inference, reducing the risk of unnecessary model complexity and improving robustness across datasets with different statistical properties. Extensive experiments demonstrate that the proposed method achieves strong average performance across multiple benchmarks and obtains favorable statistical rankings compared to established baselines. At the same time, the thesis explicitly discusses the limitations of reconstruction-based approaches, including sensitivity to threshold selection, training data contamination, and challenging datasets with complex long-range dependencies. Overall, this thesis contributes both a rigorous empirical foundation for understanding reconstruction-based deep anomaly detection methods and a practical, data-aware modeling framework that improves robustness and interpretability. The results provide valuable guidance for researchers and practitioners working on anomaly detection in complex multivariate time series, and they open several directions for future research, including streaming settings, concept drift adaptation, uncertainty estimation, and interpretability of anomaly scores.