Namatad: Inferring occupancy from building sensors using machine learning

Abstract

Driven by the need to improve efficiency, modern buildings are instrumented with numerous sensors to monitor utilization and regulate environmental conditions. While these sensor systems serve as valuable tools for managing the comfort and health of occupants, there is an increasing need to expand the deployment of sensors to provide additional insights. Because many of these desired insights have high temporal value, such as occupancy during emergency situations, such insights are needed in real time. However, augmenting buildings with new sensors is often expensive and requires a significant capital investment. In this paper, we propose and describe the real-time, streaming system called Namatad that we developed to infer insights from many sensors typical of Internet of Things (IoT) deployments. We evaluate the effectiveness of this platform by leveraging machine learning to infer new insights from environmental sensors within buildings. We describe how we built the components of our system leveraging several open source, streaming frameworks. We also describe how we ingest and aggregate from building sensors and sensing platforms, route data streams to appropriate models, and make predictions using machine learning techniques. Using our system, we have been able to predict the occupancy of rooms within a building on the University of Washington campus over the last three months, in real time, at accuracies of up to 95%.