University of Virginia researchers are using machine learning to help automate the mapping of sensor and control "points" in commercial buildings.
January 2, 2018
minute read time
Lead Performer: University of Virginia – Charlottesville, Virginia
Project Term: January 1, 2018 – December 31, 2020
DOE Total Funding: $500,000
Cost Share: $67,713
Funding Type: Buildings Energy Efficiency Frontiers & Innovation Technologies (BENEFIT) – 2017 (DE-FOA-0001632)
Related Projects: Brick
Project Objective
Building automation systems (BAS) interact with large numbers of sensors and device controller "points". Within the BAS, each of these is identified by a unique name. Most installed BAS use bespoke point naming schemes that are specific to vendors, manufacturers, and installers and are often inconsistent within a single installation. Although standard naming conventions (e.g., Project Haystack) exist, these lack some important metadata. Point mapping—acquiring a semantic understanding of points and their relationships—is a manual and labor intensive process often requiring multiple days on site. The cost of point mapping represents a significant barrier to installation of advanced monitoring and analytics systems, upgrading of control algorithms, and fault detection and diagnostics.
Automating point mapping and semantic inference is an active area of interest in the building research community. The University of Virginia will develop techniques to automatically extract contextual information for sensing and control points based on point names and the raw time series values to allow integration of building analytics engines to commercial BAS with minimal or no manual point mapping.
Preliminary exploration of automated sensor metadata interference and mapping has shown promise. This effort will explore both active learning to cluster points that are similar in their point names and time series data. It will also explore transfer learning to leverage similarities in the structure across different buildings to propagate metadata learning from one building to other similar but unannotated buildings. This technique will be applied to building datasets that have already been manually mapped and evaluated for accuracy and cost savings by comparing against the manual effort used to create the mapping. The team is targeting a payback period of less than one year.
Contacts
DOE Technology Manager: Amir Roth and Harry Bergmann
Lead Performer: Kamin Whitehouse, University of Virginia