?
Data Centers HVAC abnormal operating behaviour Detection with semi-supervised data-driven approaches
The paper investigates semi-supervised anomaly detection in data center heating, ventilation, and air conditioning (HVAC) systems using multivariate time-series data under limited data availability. A one-year dataset with 10-minute resolution, including nine HVAC operational parameters and two exogenous variables (outdoor temperature and IT load), was analyzed across normal and several abnormal operating regimes. A rulebased threshold aggregation approach was compared with a oneclass Support Vector Machine (OSVM) model with RBF kernel. Temperature influence mitigation and window-based statistical processing were applied to enhance change sensitivity. The baseline method demonstrated stable performance on normal regimes but exhibited delayed detection and scalability limitations due to manual threshold tuning. The OSVM approach provided a unified anomaly scoring framework and eliminated per-variable threshold configuration, improving adaptability under cold-start conditions while showing increased sensitivity to gradual regime shifts. The results highlight the trade-off between interpretability and scalability in HVAC anomaly detection and confirm the feasibility of transitioning toward automated data-driven monitoring under constrained datasets.