Features for an Activity Recognition Model in Office Spaces

Abstract

Designing new buildings should not only consider building physics and HVAC systems, but also human behavior that depends on energy monitoring and management systems (EMMS). These systems may provide information and advice to occupants about the relevance of their behavior regarding the current state of a dwelling and its connected grids, but also they should modify the dwelling setting accordingly. Therefore, advanced EMMS need to estimate the relevance of occupants' activities and building simulation has to take it into account in order to be able to consider EMMS at design step, reducing the so-called performance gap with reality. This thesis focuses on developing features for the estimation of human activities which cannot be measured directly, with the help of non-intrusive sensors. Algorithms from machine learning have been adapted to an office setting which is a sensor test-bed with a large number of ENOCEAN sensors, with regular/event based data reporting. We start by proposing an occupancy estimation feature that tries to classify the approximate number of people (within a range) in the room. The range is dynamic and is dependent on the occupancy seen from training as well as the prediction accuracy desired. Occupancy is a meta-feature that utilizes low-level features from CO2, temperature, motion and contact sensors. Another meta-feature that has been investigated is the classification of acoustic events from audio recordings. This feature builds up by using various features in the domain of signal processing that are directly derived from the audio signal. These features have been developed with the aim of training a robust knowledge model along with other low-level features for occupant activity classification and performance usage analysis.