The Impacts of CSI Temporal Variations on CSI-based Occupancy Monitoring Systems: An Exploratory Study

Abstract

Channel State Information (CSI) has been used for an alternative sensing source for occupancy monitoring systems to classify activities of daily living (ADLs). Previous studies have proposed learning-based activity classification models, which require similar distributions of CSI for training and testing datasets. However, as CSI varies even in a static environment, the activity classification model trained with data collected in a particular day would be invalid for other time frames. In this context, this study examines the impacts of the CSI temporal variations on the learning-based occupant activity monitoring systems. An experiment was performed to collect the CSI data while an occupant performed daily activities for six days. Three learning-based activity classification models reconstructed from the previous studies were trained and tested with time-dependent cross validation. The performances of the benchmark models were greatly degraded (below 60%) with testing data collected at different days than the training data, while their performances with testing data collected at the same day with training data were over 90%. This study also explores the opportunity to address this issue with transfer learning techniques.