Self-supervised Learning for Occupant Activity Recognition in Building Environments Using BMS data

Abstract: Balancing energy efficiency and occupant comfort is key to maintaining the sustainability of buildings. Understanding occupant activities is essential for optimising energy use without compromising comfort. This paper proposes a self-supervised learning approach for recognising occupant activity patterns using indoor environmental data from the Building Management System (BMS). A modified Transformer Masked Autoencoder (Ti-MAE) is adopted to extract latent representations of data, followed by the K-means Clustering Algorithm for clustering typical occupant activity patterns. Experiments using real-life building data demonstrate its robust performance in occupant activity recognition, even without specific sensors. The approach optimises energy efficiency while preserving privacy.