Building an occupancy model from sensor networks in office environments

The work presented here aims to answer this question: Using just binary occupancy sensors is it possible to build a behaviour occupancy model over long-term logged data? Sensor measurements are grouped to form artificial words (activities) and documents (set of activities). The goal is to infer the latent topics which are assumed to be the common routines from the observed data. An unsupervised probabilistic model, namely the Latent Dirichlet Allocation (LDA), is applied to automatically discover the latent topics (routines) in the data. Experimental results using real logged data of 24 weeks from an office building, with different number of topics, are shown. The results show the power of the LDA model in extracting relevant patterns from sensor network data.