Deep Weighted Fusion Learning (DWFL)-based multi-sensor fusion model for accurate building occupancy detection

Abstract

With the advancement of artificial intelligence, the dominance of deep learning (DL) models over ordinary machine learning (ML) algorithms has become a reality in recent years due to its capability of handling complex pattern recognition without manual feature pre-definition. With the growing demands for power savings, building energy loss reduction could benefit from DL techniques. For buildings/rooms with the varying number of occupants, heating, ventilation, and air conditioning (HVAC) systems are often found in operations without much necessity. To reduce the building’s energy loss, accurate occupancy detection/prediction (ODP) results could be used to control the proper operations of HVACs. However, ODP is a challenging issue due to multiple reasons, such as improper selection/deployment of sensors, inefficient learning algorithms for pattern recognition, varying room conditions, etc. To overcome the above challenges, we propose a DL-based framework, i.e., Deep Weighted Fusion Learning (DWFL), to detect and predict occupancy counts with optimal multi-sensor fusion structure. DWFL fuses the extracted features from multiple types of sensors with the priority/weight assignment to each sensor. Such weight assignment considers different room conditions and the pros/cons of each type of sensor. To evaluate DWFL model in terms of occupancy prediction accuracy, we have set up an experimental testbed with low-cost cameras, carbon dioxide (${\mathrm{CO}}_2$), and passive infrared (PIR) sensors. Among the recently proposed occupancy detection models, DeepFusion utilized deep learning model on heterogeneous sensor data and achieved 88% accuracy in occupancy count estimation (Xue et al., 2019). Another deep learning-based model MI-PIR achieved 91% accuracy on raw analog data from PIR sensors (Andrews et al., 2020). Our research outcome is 94%. Therefore, the experiment results show that our DWFL scheme outperforms the state-of-the-art ODP methods by 3%.