Adversarial Discriminative Domain Adaptation Transfer Learning for Bridge Damage Detection

Authors:
Publisher website
Google Scholar
Abstract
This study introduces a new bridge damage detection method using adversarial discriminative domain adaptation transfer learning. The method aims to transfer damage information between numerical models and real-world bridges across domains. The method uses an adversarial approach to teach the target domain’s feature extractor to mimic the feature mapping of a pretrained source domain extractor, aligning target domain features with the source’s feature space. The innovation is asymmetric mapping for adversarial learning, which represents real-world bridge data features within the finite element model data space and promotes cross-domain damage information transfer. The method’s efficacy was tested through numerical simulations and steel beam-cutting experiments. Moreover, the simulation tested accuracy and stability as domain shifts increased. The lab experiment cut steel beams to induce damage, treating the damaged structure as the target model. Both confirmed the method’s strong performance amidst errors and uncertainties.
Funding Information
  • National Natural Science Foundation of China (52178289)