Thursday March 10, 2022
Local distributed learning architecture applied to smart buildings

Buildings both residential and commercial together consume close to half of the world’s total energy produced and is growing at a non decreasing pace. So efficient resource utilization forms the primary motivation behind integrating smartness into a brick and mortar structure. Although active from early 2000's, literature survey reveals that there are significant business gaps that bottlenecks smart building development. Data privacy, high capital investments and obscure monetary benefits are the major factors that impede the motivation to integrate smartness in a building.

This work introduces the idea of a zero sensor intelligence by embedding human-space interaction models on a graph based abstraction of a building. Spectral decomposition of the semantically enriched connected graph helps in ranking multiple spaces with regard to temporal importance or likely energy dissipation. Next, the work investigates the problem of optimal sensor placement and proposes a Virtual Sensor Field which is a fabric of real and machine-learnt sensors. Distributed learning technique is applied to find robust sensor placement configurations with incremental data over time. This culminates in a novel pre-integration platform to bring clarity on at-least how many sensors are to be installed and where in a building. Once sensors are installed, the system explores levels of data privacy, that supports the philosophy of edge computing: “Process data as close as possible to the generation site.” In a nutshell, the work lays the blueprint of a generic smart building solution with less sensors, lower carbon footprint and auto-updating models with strictly localised raw data at edge/in-house.
Mis à jour le 1 March 2022