Non-intrusive infrared facial thermography for personal thermal sensation prediction models: A field study

• Developed occupant-centric thermal model using ensemble ML and SMOTE. • Used infrared thermography for non-intrusive facial temperature monitoring. • Model outperforms Predicted Mean Vote (PMV) in predicting comfort. • Identified nose temperature as key feature for thermal sensation. Ensuring thermal comfort requires maintaining indoor environmental conditions based on individual thermal sensation, as predefined zone set point temperatures often lead to discomfort due to inter- and intra-individual differences. Thus, understanding individual thermal sensation is vital for implementing occupant-centric control systems. This study explores the potential of using an Ensemble Machine Learning approach with the Synthetic Minority Oversampling Technique (SMOTE), combined with non-intrusive infrared facial thermography to predict thermal sensation votes (TSV). The non-intrusive nature of infrared thermography allowed for seamless and continuous monitoring of facial skin temperatures without disrupting the participants' natural behavior. Conducted in a controlled facility with climate zone D characteristics, the study obtained data of environmental conditions and subjective comfort feedback from twenty participants over six days towards developing personal thermal sensation models. Feature engineering was applied to identify and generate optimal and important features. The SMOTE method expanded the training data to address sample imbalance issues. Finally, ensemble models were employed to construct the personal thermal sensation prediction model. The results show that the personal thermal sensation model outperforms the Predicted Mean Vote (PMV) index. They also reveal the importance of nose temperature compared to other facial skin temperatures (forehead, chin, and cheek) as a significant feature among participants. As the experiment was conducted over a single season, further research with data from varied environmental conditions to generalize the results and enhance the robustness of personalized comfort models.