Reliability Assessment of Machine Learning Methods in Seismic Damage Detection of Reinforced Concrete Buildings

Machine learning (ML) techniques, have surfaced as a prospective option for identifyingdamage recently. They excel in swiftly, precisely, and automatically handling extensive datasetsfrom various origins. Evaluating the effectiveness of diverse ML techniques has become essentialdue to the growing adoption of ML in damage identification in structures. These evaluations setstandards for assessing alternative methods and reveal perspectives on the fundamental data andstructures. The current study investigated two ML classifiers: Random Forests (RF) and SupportVector Machine (SVM). The primary objective was to detect damage grades in reinforced concrete(RC) buildings in Nepal, Ecuador, Haiti, and South Korea. Moreover, a new metric was introducedto evaluate the "reliability" of outcomes derived from ML, focusing on the probability ofmisidentifying grades of damage. This approach contributes to a deeper comprehension of thereliability of ML outcomes. Findings demonstrated the superior efficacy of the RF classifier,outperforming the SVM classifier in accuracy across three datasets. The reliability metric indicatedaverage reliabilities of ۸۲% for RF and ۷۸% for SVM. This research underscores the efficacy ofML techniques, specifically highlighting the RF classifier's reliability in damage detection of RCbuildings.