Identification of Warm Discomfort-Related Action Segments using Deep Clustering for Establishing Personalized Energy Saving Measures

The creation of energy-efficient and thermally comfortable environments in buildings is essential to achieve the goal of national greenhouse gas reduction. The integration of deep learning algorithms and vision sensors into thermal environments enables the categorization of occupants’ discomfort-related poses into several representative action segments, thus making it possible to understand their behavioral steps. Considering that demographically diverse individuals employ different ways to restore thermal comfort, the human pose clustering can help identify their demographic characteristics. However, it is still unknown whether the identified action segments are different from each other. Therefore, this study categorized occupants’ warm discomfort-related poses into several representative segments and visualized the identified segments. A deep learning-based clustering analysis was conducted using video data on 20 college students’ warm discomfort-related actions. The clustering results found that individuals’ representative segments observed during the execution of the actions vary depending on which direction they moved and where key body joints are located. These results demonstrate that decomposing occupants’ warm discomfort-related actions enables the identification of their demographics based on the presence of certain segments and, in turn, contribute to the design of personalized energy-saving measures.