Traditional techniques for generating indoor occupancy maps involve the use of a network of distributed sensors to determine occupancy and/or movement within a facility to derive actionable insights and inform automated control mechanisms. These approaches are constrained by their ability to provide only coarse and often binary data—essentially, determining whether someone is present at a location or not. A more sophisticated approach that can capture data on occupancy with both spatial and temporal details can enable finer control and more nuanced automation triggers. This disclosure describes a large occupancy model that extends large language model (LLM) techniques with an occupancy sensor encoding branch that encodes raw sensor data from sensors within a building. The sensor data are collated, time synchronized, and feature interpolation is performed to enable a model to be trained and used irrespective of the number and type of sensors within an installation. The large occupancy model can provide high-fidelity answers to contextual queries for building control. The response can be provided to upstream building automation systems for improved building control.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Shin, D, "Sensor-encoded Generative Language Modeling for High Resolution Occupancy Sensing and Building Control", Technical Disclosure Commons, (December 19, 2023)