Virtual sensing of room activity using passive environmental sensors

A data-driven approach has been developed to classify indoor activities using only commonly available passive environmental sensors, such as CO2, temperature, humidity, and passive infrared (PIR). An integrated IoT system comprising of sensor nodes, edge node and an intelligent server is designed and developed to provide real-time activity classification. Spectral Clustering and bidirectional long short-term memory (BiLSTM) are employed to achieve automatic labeling and room state prediction. The results show that the overall classification accuracy ranges from 88% to 96% for five target states across three distinct environments using CO2 and PIR values as input variables. Additionally, incorporating more input variables has been evaluated to access the ability of real-time classification of proposed model. An innovative monitoring mode can provide a different approach for detecting activities and occupancy in the future.