dStructAD: Domain-Level Structured Normality Representation with Variation Calibration for Time Series Anomaly Detection

Time Series Anomaly Detection (TSAD) is critical for ensuring reliability in real-world systems such as industrial monitoring, healthcare, and online services. However, learning-based TSAD methods trained on a single normal-only TargetSet, namely the target dataset, often suffer from incomplete and biased knowledge of normality, causing domain-level false alarms when unseen but valid normal patterns appear at test time. A natural extension is to leverage DomainSet, composed of datasets from one domain, to enrich domain-level normality. Yet, our analysis shows that simply incorporating DomainSet could make domain-admissible normal variations and TargetSet-specific anomalies highly entangled in the learned representations, weakening anomaly separability. Motivated by human experts who establish normal patterns, calibrate admissible variations, and identify true anomalies, we propose dStructAD, a domain-level structured representation framework for time series anomaly detection, to address this challenge. Specifically, dStructAD first builds structured domain-level normality based on the Kolmogorov-Arnold network, and then calibrates admissible variations via two phases. Phase 1 learns shared domain-level semi-structural normality knowledge from DomainSet, while Phase 2 performs structure-preserving TargetSet calibration to absorb admissible variations without collapsing anomaly separability. Across six benchmarks, dStructAD delivers consistent improvements over SOTAs, with 3% average gain in overall performance. Code could be available at https://anonymous.4open.science/r/dStructAD-8F03.