Paper Conference

Proceedings of Building Simulation 2021: 17th Conference of IBPSA


End-to-end Model for Estimating Heat loss of Building envelope using Deep learning with Infrared image

Heesung Park 1, Sangho Son 1, Ji Young Kang 1, Haemin Jung 1, Seungeon Lee 2, Deuk-Woo Kim 2, Wooju Kim 1
1 Department of Industrial Engineering, Yonsei University, Seoul, Republic of Korea
2 Korea Institute of Civil Engineering and Building Technology, Goyang-Si, Republic of Korea

Abstract: Estimating heat loss on the envelope of buildings is important in efficient management of energy within buildings. To determine the degree of heat loss, various methods which use infrared images as input data have been suggested. However, the studies have limitations; they require consideration of environmental conditions and lots of additional metadata inputs to describe them. This paper presents a novel method structured as an endto-end process using a convolutional neural network (CNN) to estimate heat loss, only with infrared images of the building envelope. The process is divided into three steps: CNN-based object detection to identify building components from infrared images, edge detection to calculate actual area, and estimation of relative heat loss among objects with identical material. The results of our process are estimated heat loss types of the objects with their classes and locations. We first generate a temperature distribution using a histogram of pixel colors of object areas, decide a threshold to determine whether the image shows heat loss or not and assign heat loss types based on the threshold. We experimented our method for windows and showed its usability. In the future, we will generalize our method by applying it to other parts of building envelope.
Keywords: Infrared image object detection, Deep learning, Neural Network, Heat loss estimation, End-to-end model
Pages: 3111 - 3118