Cross-domain damage identification of bridge based on generative adversarial and deep adaptation networks

Abstract

Model-driven bridge damage identification based on video measurement frequently encounters challenges due to low data quality, which leads to significant discrepancy between numerical model and actual bridge, exacerbating the cross-domain problem. To address this issue, the generative adversarial network (GAN) and Deep Adaptation Network (DAN) are introduced in this paper. The principles and utilizations of the two networks are introduced, while GAN is used to generate a large number of high-quality artificial samples to enhance the learning, and DAN is used to extract generic damage features to cross the source and target domains. A damage experiment of a steel truss model bridge is conducted to validate the proposed method, and the damage identification results of the proposed method and the traditional method are compared. The identification results illustrate that the proposed method can generate a substantial number of artificial samples only using a limited number of real target samples. These artificial samples facilitate the learning of cross-domain damage features, which ultimately contributes to achieving higher accuracy in damage identification.