Integrated Sensor Data Processing for Occupancy Detection in Residential Buildings

Occupancy information can save energy consumption buildings by 10-50 %.  Many commercial buildings utilize a building management system (BMS) and occupancy sensors to better control HVAC system.  The installation process of occupancy sensors is often complicated and bothersome.  Moreover, their high cost will not ensure short payback period, and thereby has limited the adoption of such technology in the residential sector.  Recently, several research works explored data fusion approach to detect occupancy information from environment sensors.  Many of them derived occupancy information by feeding various sensor data into pre-existing machine learning models.  Such approach requires high sensor density and accuracy of occupancy detection highly depends on the installation condition.  Hence, high accuracy of the model can be achieved in the test site, but scalability and transferability of the approaches are limited.  This paper presents an economical way of occupancy detection utilizing a two-layer detection approach with the data obtained from multiple non-intrusive sensors (temperature and motion).  The data from the sensors were first fed into different white-box detection models for recognizing human activities (door handle touch, water usage, and motion near the door area).  Since our sensors are not intrusive, we need to utilize data fusion to enhance the validity and reliability of occupancy detection.  However, those activities may not happen at the same time, so that we have to consider a series of activities occurred in a short period, called event.  As the event have different durations and cannot be readily applicable to existing machine learning models due to varying input matrix sizes.  Hence, we devised a fixed format to summarize the event regardless of the total duration of the event.  After that, the event data was used to train and test four machine learning models (Random Forest, Decision Tree, K-Nearest Neighbor, and Support Vector Machine).  The proposed occupancy detection system was installed in a 65 m² living lab.  Four temperature sensors and one motion sensor had been used to collect the environmental information for 54 days.  The validity of the proposed two-layer detection system was verified by the accuracy and the f1-score of each models.  In all machine learning models, we found that the proposed system showed significantly improved the accuracy and the f1-score over conventional approach with the same data.  As such, the proposed work demonstrated similar or improved level of the accuracy (95%) and f1-score (95%) over other works, while using much less sensor density.  Moreover, the scalability of the approach was also enhanced due to the generality of human activities.