ANOMALY DETECTION IN INDOOR ENVIRONMENTAL SENSOR DATA BASED ON A DIGITAL TWIN

Abstract: Anomaly detection in building sensor networks often relies on static thresholds that fail to capture temporal dependencies and system changes. This work combines a baseline model of expected indoor environmental behavior with anomaly detection to form a data-driven Digital Twin. In this framework, a regression model learns expected behavior from historical data, while deviations are quantified as multivariate residuals and analyzed using unsupervised and semi-supervised methods. Preliminary analysis on a representative campus sensor shows detection of anomalous behavior and associated distributional shifts, with sensor relocation serving as an illustrative system change.