Deep Time Series Anomaly Detection

TSAD is critical in domains like cybersecurity, industrial monitoring, finance, and healthcare, but deployment is hindered by scarce labels, diverse anomalies, and distribution shift. This thesis addresses these challenges through a consolidated Time Series Anomaly Detection (TSAD) foundation and four methods. CARLA learns generalisable representations via self-supervised contrastive learning. DACAD transfers normal-pattern structure across domains with minimal labels. GenIAS generates realistic anomaly variations to strengthen training under scarcity. CEDL provides calibrated, imbalance-robust classification using a radial decision function. Together, these contributions deliver a practical, label-efficient TSAD framework supported by open-source code and benchmarks.