Data Streaming Proportions: a Framework for Monitoring and Anomaly Detection

Abstract: A framework for monitoring HVAC data streaming and detecting anomalies is presented. Latent Dirichlet Allocation identifies three states corresponding to heating, cooling, and baseline system behaviors, while the Dirichlet distribution detects proportion anomalies independent of total data volume. Using 34 months of sensor data from an academic-residential building, the framework reveals seasonal and operational trends, with notable anomalies linked to maintenance events. This method enables facility managers to monitor system states and diagnose deviations efficiently but requires sufficient historical data and expert state interpretation. This low-complexity approach provides a practical tool for real-time HVAC monitoring in dynamic environments.