A Contactless Non-Intrusive Approach for Machine Learning-Based Personalized Thermal Comfort Prediction
Analytics
29 views ◎7 downloads ⇓
Abstract
Indoor thermal environmental conditions play a significant role in protecting occupants’ well-being. In this regard, schedule-based and predefined environmental control is one of the main reasons for the current discomfort and dissatisfaction with the thermal environment. These general standards make it impossible to consider people’s differences in thermal sensation and personal preference. Recent research is attempting to leverage occupants’ demand in the control loop of the buildings to consider the well-being of each individual based on their own physiological properties. These thermal comfort models are called "personalized comfort models". In this regard, studies are trying to utilize skin temperature recorded by infrared thermal cameras for developing personal comfort models through machine learning prediction algorithms. However, there are some critical gaps in the current methods that have limited the application of this platform in real buildings. Some of the main shortcomings of the current approaches are the limited distance from the camera, the absence of automated and accurate detection of facial areas, and the limitation on the detectable facial positions. To capitalize on the potential and address the existing constraints, new solutions are required that take a more holistic approach to non-intrusive thermal scanning by integrating the benefits of sensor fusion, image processing, and machine learning.The contribution of this dissertation is in the three main aspects of literature review, data collection, and model development. This study presents a comprehensive and systematic review of the current machine learning-based personalized thermal comfort studies. In addition, we introduce "Charlotte-ThermalFace", our recently developed dataset, and how it addresses some of the existing gaps in the subject. Charlotte-ThermalFace contains more than 10,000 infrared thermal images in varying thermal conditions, several distances from the camera, and at different head positions. The data is fully annotated with the facial landmarks, ambient temperature, relative humidity, air velocity, distance to the camera, and subject thermal sensation at the time of capturing each image. By using this dataset, we have developed a personalized comfort model for subjects at a farther distance in a completely non-intrusive method. We have accomplished this by incorporating both visual and thermal images to create a multi-modal sensing platform. Through this interconnected system, we use visual images and the deep learning based HR-net algorithm for localizing facial landmarks, and thermal images to measure the temperature values of the detected areas. This research implements an automated approach to register simultaneous thermal and visual frames and read the facial temperature accurately for subjects at a distance from the camera. Through this method, we could extract facial skin temperature at a distance and in several head positions. For creating machine learning-based personalized thermal comfort models, we have implemented two powerful classification algorithms: Random Forest and K-Nearest Neighbor. The prediction results indicate an average accuracy of 86\% for the Random Forest and 74\% for the K-Nearest Neighbor algorithm. This study presents promising findings for the creation of automated thermal comfort prediction platforms from a distance through the utilization of thermal cameras.