Rapid damage state identification of structures using generalized zero-shot learning method

Abstract

Identification of damaged structures after natural disasters, such as earthquakes, is crucial for ensuring public safety and facilitating timely repairs. Recently, machine learning-based models have shown promise in this direction. Traditional machine-learning approaches require a significant amount of labeled data for training. However, obtaining labeled data for damage identification can be challenging because it is time-consuming and expensive. To resolve this issue, this study proposes a generalized zero-shot learning (GZSL) methodology to identify the degree of structural damage in images. The proposed methodology was used for assessing the failure mode of reinforced concrete shear walls involving pixel images on a scale of 0–1. The GZSL model with ResNet18 as its backbone demonstrated good performance, achieving 100% and 86.7% accuracies on training and test sets, respectively. This methodology was also utilized for assessing building damage using wavelet images with a broader color spectrum; the ResNet50-based GZSL model demonstrated excellent performance, achieving an accuracy of 68%, even with a smaller number of samples that included both seen and unseen classes.