An IoT Assisted Machine Learning Framework for Monitoring the Thermal Variation on Building Envelope

Abstract

Periodic or non-periodic thermal fluctuations can lead to significant damages to both the interior and exterior of buildings. These damages may result in poor insulation, excessive energy consumption, and the need for expensive repairs. Regular thermal inspections of building surfaces can prevent these issues. Traditional methods of building inspection, such as blower door tests and air filtration measurements, often involve complex parameter tuning and are carried out in controlled environments with intrusive equipment deployment. Besides, these formal inspections by professionals can be expensive, inconclusive, and inconvenient for frequent monitoring. Alternative cost-effective, non-intrusive and flexible approach is using IoT technology based sensors and thermal cameras to monitor the structural health of buildings. These methods provide data-driven insights on the building's thermal performance to the residents and assist professionals in expediting the inspection process. While various research approaches have been proposed to detect and analyze thermal variations on building surfaces, most of them rely on visible damage and controlled settings. Few studies have focused on the longitudinal thermal characteristics of indoor environments, mainly due to challenges of automated data collection, absence of building metadata, scalability, and interpreting contextual factors like weather conditions and indoor activities. With this motivation, we proposed thermal condition monitoring framework based on temperature-humidity sensors and thermal images which can provide spatial and temporal thermal characteristics of uncontrolled inside built environment without using building metadata. First, we used low-cost temperature and humidity sensor data and proposed an unsupervised temporal clustering method to identify the thermal patterns of indoor surfaces with respect to outside weather conditions. We quantified the thermal characteristics of each building based on their responses to outside weather conditions using thermal variables. We also depicted the thermal changes in the inside built environment for human activities, i.e., cooking. Moreover, we attempted to deal with the existing challenges of collecting thermal images in inside built environment for longer period of time. We analyzed the spatial and temporal thermal dynamics over various building elements (i.e., walls, windows, doors, etc.) using thermal image series of indoor scenes collected over a long period of time. We postulated a graph based network that combines the spatial and temporal features of image regions from the graphical representation of spatial and temporal relations among image regions in the sequential thermal images. We extended the thermal variation analysis by incorporating the structural orientation of indoor scenes and focusing only on the building components. We eliminated the surrounding non-building components in indoor scenes and identified the thermal status of a specific building component by considering thermal changes of its' surrounding building components. Our proposed unsupervised graph based framework, we determined the spatial and temporal thermal anomalies for different building components and discovered potential damage prone areas in approximately 70% of the cases.