Enhancing rubber rupture detection in rubber bearing through generative adversarial network and feature-bagging zero-shot methodology

Abstract

Base isolation technology is a design strategy developed to protect buildings from the direct impact of seismic forces, utilizing base isolation devices, with rubber bearings being the most commonly used type. After an earthquake, manually inspecting rubber bearings for damage is inefficient, unable to reveal internal damages, and carries significant risks. Consequently, there is a pressing need for an innovative damage detection method. The difficulty of obtaining and labeling data related to rubber rupture damage makes it hard to apply supervised learning methods to construct damage detection models. In response to this, this study combined the active sensing method with unsupervised learning based on feature bagging, establishing a robust rubber damage detection model that successfully addressed the zero-shot problem faced in rubber damage detection processes. To increase the proportion of data on rubber damage, a generative adversarial network based data augmentation methods were applied. The research findings demonstrated that the developed model achieved an average precision of 0.9216 and an area under the ROC curve (Receiver Operating Characteristic curve) of 0.9788 for rupture damage detection, outperforming other machine learning models.