AUTOMATED DETECTION OF BUILDING ENVELOPE DEFECTS - A SYSTEMATIC REVIEW

Building envelope serves to protect users and structures, and supply insulation, among other purposes, such as giving an identity to a building. However, due to adverse environmental conditions and physical factors, various deteriorations may occur in the components and materials constituting it. In order for the buildings to have a long service life, it is important to detect the deterioration in the materials and to plan the necessary maintenance and repair. There are various human-based inspection methods, both simple and advanced, for defect detection in a building. However, defect detection through human-based inspection needs expert knowledge in most cases and otherwise might be inaccurate. Additionally, it is usually time-consuming and poses a safety threat in the case of the facades of high-rise buildings. For these reasons, automated defect detection methods are developing to facilitate the process and to improve both the time needed and the objectiveness of the inspection results. Thus, there has been an increasing interest in machine learning-assisted automated detection methods in recent years, mostly concerning image/visualbased methods in addition to other advanced methods. In this study, to put forth the current state of automated defect detection regarding the inspection of the building envelope and to examine the role of machine learning in this process, a systematic literature review was conducted using data from the Scopus database. In the paper, following a brief background on machine learning and the explanation of the review method, research studies were analyzed and discussed considering the material and defect types studied, accuracies achieved for the defects, and some factors affecting accuracy. Regarding the crack, as the most commonly considered defect type in these studies, the accuracy rate of the model was seen to be higher (i.e. 94-98%) when it was inspected as the sole defect rather than as a defect among other defects. However, when more than two defect types were considered together, the accuracy was seen to be ranging between 40-70%.