Structural damage identification based on Wasserstein Generative Adversarial Network with gradient penalty and dynamic adversarial adaptation network.

Transfer learning-based damage identification is widely explored as it can utilize prior damage knowledge from finite element model (FEM) to identify damage without real structure damage labels. However, lacking a substantial amount of real structure data and overlooking the variances in conditional (local) distributions between source domain (data generated by FEM) and target domain (real structure data) severely restrict its practical application. This paper proposes a transfer learning-based damage identification method by combining Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and Dynamic Adversarial Adaptation Network (DAAN). Firstly, WGAN-GP generative model is employed to augment real structure samples to address the issue of insufficient data. Then, an optimization transfer model with the capability of balancing the importance of marginal and conditional distributions between FEM data and real structure data is proposed based on DAAN transfer model. It dynamically learns damage-invariant features for damage identification. Thirdly, damage identification procedures are provided by combining the generative model WGAN-GP and the transfer model DAAN with moving load induced responses as inputs. Numerical and experimental examples indicate that the proposed method achieves superior transfer effect with limited real structure data and maintains high damage identification accuracy compared to the traditional transfer learning-based methods. [ABSTRACT FROM AUTHOR]