TY - JOUR
T1 - Learning occupants’ workplace interactions from wearable and stationary ambient sensing systems
AU - for the Wellbuilt for Wellbeing Project Team
AU - Ghahramani, Ali
AU - Pantelic, Jovan
AU - Lindberg, Casey
AU - Mehl, Matthias
AU - Srinivasan, Karthik
AU - Gilligan, Brian
AU - Arens, Edward
N1 - Funding Information:
This study was funded by the U.S. General Services Administration (GSA) under interagency agreement # GX0012829 with the U.S. Department of Energy and Lawrence Berkeley National Laboratory. GSA’s Wellbuilt for Wellbeing Group is a multidisciplinary research project team (GSA Contract # GS-00-H-14-AA-C-0094) consisting of the following members: Kevin Kampschroer, Judith Heerwagen and Brian Gilligan of GSA. Esther Sternberg, Perry Skeath, Casey Lindberg, and Matthias Mehlof the University of Arizona Institute on Place and Wellbeing. Bijan Najafi, Javad Razjouyan, Hyoki Lee, and Hung Nguyen of the Baylor College of Medicine Interdisciplinary Consortium on Advanced Motion Performance (iCAMP). Sudha Ram, Faiz Curim and Karthik Srinivasian of the University of Arizona INSITE Center for Business Intelligence and Analytics. Kelly Canada of LMI Inc. Priya Saha, Rebecca Goldfinger-Fein, Alicia Darbishire, and Mills Wallace of the Federal Occupantional Health Service. Davida Herzl, Reuben Herzl, Melissa Lunden, Nicole Goebel, and Scott Andrews of Aclima Inc.
Publisher Copyright:
© 2018 Elsevier Ltd
PY - 2018/11/15
Y1 - 2018/11/15
N2 - Having access to real-time information on building occupants’ state of interactions enables optimization of building systems for improved energy efficiency, well-being and productivity of the occupants. In this paper, we propose a framework to learn occupant interactions from ambient sensing technologies (e.g., sensing of variables such as sound (dB), CO2 (ppm), light intensity (lux), dry-bulb temperature (°C), relative humidity (RH%), pressure (mbar)) from both stationary and wearable devices and select the technologies and averaging windows which contain the required information for learning. In this framework, several supervised machine learning algorithms are tested on the labeled datasets and the algorithm which outperforms others is selected. Two types of sensing devices were utilized for analyses: wearable devices worn around the neck by the test subjects, and a network of stationary devices located in the test subjects’ working indoor spaces. 221 employees of federal agencies housed in facilities managed by the US. General Services Administration in the mid-Atlantic and Southern states participated in this study, answering questions about their current task every hour. Overall accuracies were observed of 86.72% for wearable and stationary devices, 81.25% for only wearable-only, and 85.16% for stationary-only for prediction of the mixed multi-label classification via Random Forests algorithm. The high prediction allows for identifying subjects’ tasks when training labels are not available. Predicting occupants’ interactions as a main indicator of occupants’ behavior have significant implications for the energy efficiency of building systems (up to 20% savings).
AB - Having access to real-time information on building occupants’ state of interactions enables optimization of building systems for improved energy efficiency, well-being and productivity of the occupants. In this paper, we propose a framework to learn occupant interactions from ambient sensing technologies (e.g., sensing of variables such as sound (dB), CO2 (ppm), light intensity (lux), dry-bulb temperature (°C), relative humidity (RH%), pressure (mbar)) from both stationary and wearable devices and select the technologies and averaging windows which contain the required information for learning. In this framework, several supervised machine learning algorithms are tested on the labeled datasets and the algorithm which outperforms others is selected. Two types of sensing devices were utilized for analyses: wearable devices worn around the neck by the test subjects, and a network of stationary devices located in the test subjects’ working indoor spaces. 221 employees of federal agencies housed in facilities managed by the US. General Services Administration in the mid-Atlantic and Southern states participated in this study, answering questions about their current task every hour. Overall accuracies were observed of 86.72% for wearable and stationary devices, 81.25% for only wearable-only, and 85.16% for stationary-only for prediction of the mixed multi-label classification via Random Forests algorithm. The high prediction allows for identifying subjects’ tasks when training labels are not available. Predicting occupants’ interactions as a main indicator of occupants’ behavior have significant implications for the energy efficiency of building systems (up to 20% savings).
KW - Buildings energy efficiency
KW - Interaction detection
KW - Machine learning
KW - Occupant behavior modeling
KW - Ubiquitous computing
KW - Workplace interaction
UR - http://www.scopus.com/inward/record.url?scp=85051803175&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85051803175&partnerID=8YFLogxK
U2 - 10.1016/j.apenergy.2018.08.096
DO - 10.1016/j.apenergy.2018.08.096
M3 - Article
AN - SCOPUS:85051803175
VL - 230
SP - 42
EP - 51
JO - Applied Energy
JF - Applied Energy
SN - 0306-2619
ER -