Dynamic domain adversarial neural network for population-based structural health monitoring

Abstract

Structural health monitoring (SHM) is critical for ensuring the safety and reliability of civil infrastructure, but its practical implementation is limited by challenges such as data scarcity and structural heterogeneity. Population-based SHM (PBSHM) aims to overcome the lack of labeled data for a single structure by leveraging information from a group of similar structures. This paper proposes a dynamic domain adversarial neural network (DDANN) to address the PBSHM problem, which is formulated as a multidomain transfer learning task. The DDANN architecture integrates domain-specific projection modules to handle inconsistent feature dimensions across structures, a shared feature encoder to learn domain-invariant representations, and a classification head to identify damage states. The training process uses a composite loss function that includes domain adversarial loss, maximum mean discrepancy loss for feature distribution alignment, and a pseudo-labeling strategy to leverage unlabeled target data. The efficiency of the proposed approach is validated through three case studies: a numerically simulated population of heterogeneous bridges for damage severity classification, a diverse group of shear structures for damage localization, and an experimental benchmark bridge dataset with environmental effects. Results consistently show that DDANN achieves robust cross-domain generalization, particularly in scenarios with sparse data and structural dissimilarity. The proposed DDANN offers a powerful framework for advancing the practical application of data-driven SHM across populations of structures.