Time Series Sensor Dataset from Real-World Smart Spaces Deployments

Indoor Cyber-Physical Systems (CPS) and smart spaces, such as smart homes and offices, increasingly rely on dense networks of heterogeneous sensors—including motion, light, temperature, humidity, door, sonar, and vibration sensors—to continuously monitor occupant activities and environmental conditions. These sensors generate massive volumes of multivariate time-series data that capture the temporal dynamics of human interactions with the physical environment. Such data serve as the foundation for understanding system behavior, enabling the discovery of sensor dependencies, occupancy patterns, and environmental changes over time. In safety-critical applications, correctly interpreting which sensor produced a given reading and where that sensor is physically located is a prerequisite for downstream tasks such as relational scene-graph construction, automation policy generation, human activity recognition, anomaly detection, and fault diagnosis. However, despite their richness, the sheer volume, heterogeneity, and temporal complexity of these time-series streams make extracting meaningful spatial and semantic relationships highly challenging. To address this gap, we deployed five real-world smart home and office environments comprising varying numbers of heterogeneous sensors, collecting several months of continuous time-series data that capture diverse occupant behaviors (i.e., single to 5-6 inhabitants), environmental conditions, and deployment characteristics. This dataset was collected across five distinct indoor environments, ranging from a single compact office to a multi-room residential unit, as part of an NSF-funded research effort on safety-aware smart space monitoring. To support anonymization and consistent schema design across environments, each physical sensor was assigned a short numeric identifier (S1, S2, S3, ...) at the time of data collection.