Detection of damages caused by earthquake and reinforcement corrosion in RC buildings with Deep Transfer Learning

Abstract

The Reinforced Concrete (RC) buildings in countries within earthquake zones like Turkey are generally damaged more than anticipated during earthquakes. Corrosion of reinforcement and other damages caused by the corrosion are also widely encountered in relatively old structures that have not received adequate engineering services. Although it is not easy to differentiate the earthquake and corrosion damages in some cases, the formation mechanisms of these damages and their consequences in terms of structural safety may be quite different. For this reason, in post-earthquake damage detections, it is important to determine the cause of the damage in a realistic way to plan the future interventions on the building (repair methodology, demolition / reconstruction, etc.) properly and to accurately decide on the financial schemes for required interventions (insurance, government support, owner, etc.). In this sense, smart systems are needed to be activated to speed up the decisionmaking process of the engineers/technical staff to be involved in field damage assessment surveys after earthquakes. Based on this motivation, a Deep Transfer Learning algorithm was developed in this study, which allowed distinguishing of damages caused by corrosion from earthquake-induced damages in structural elements of RC buildings. This study, which was conducted for the first time in the literature, tested the performance of Deep Transfer Learning method in damage detection and classification efforts of RC structural members. For development of the algorithm, a data pool was used consisting of images belonging to 1040 damaged RC elements, mostly columns, obtained from field surveys conducted after Istanbul-Silivri (Mw = 5.8) earthquake dated 26.09.2019 and Elazig-Sivrice (Mw = 6.8) earthquake dated 24.01.2020. The developed algorithm can determine whether the damage was caused by earthquake or reinforcement corrosion with an estimate success of 90.62 % based on damage image. Additionally, at the end of the study, classification performance of developed algorithm was also tested by using different data pool from Samos Earthquake dated 30.10.2020 (Mw = 6.6). It has been seen that the algorithm makes damage type predictions with high success in the new data set.