Wearable devices and Machine Learning algorithms to assess indoor thermal sensation: metrological analysis

Personal comfort modeling is considered the most promising solution for indoor thermal comfort management in buildings. The use of wearable sensors is investigated for the real-time measurement of physiological signals to train comfort models for buildings monitoring and control. To achieve the required reliability, different uncertainty sources should be considered and weighted in the measurement results evaluation. This study presents an example of personal comfort model (PCM) development based on wearable sensors (i.e., Empatica E4 smartband and MUSE headband) acquiring multimodal signals (i.e., photoplethysmographic – PPG, electrodermal activity – EDA, skin temperature – SKT, and electroencephalographic – EEG ones), together with a metrological characterization of the modeling procedure. Starting from the data collected within an experimental campaign on 76 subjects, different Machine Learning (ML) algorithms were exploited to create comfort models capable of predicting the human thermal sensation (TS). The most accurate model was considered to investigate the impact of sensors uncertainty through a Monte Carlo simulation. Results showed that the Random Forest model is the best performing one (accuracy: 0.86). Monte Carlo simulation method proved that the model is very robust towards measurement uncertainties of input features (expanded uncertainty of the model accuracy: ± 0.04, k = 2). This confirms the possibility to derive the subject’s TS exploiting only physiological signals; measurement uncertainty is influenced mostly by PPG and EDA signals. This kind of investigation could lead to the development of PCMs, exploitable within control systems to optimize subjects’ well-being and building energy efficiency.