Cross-structure domain translation with structural state awareness: AT-StarGAN-GP for multi-level synthetic signal generation in structural health monitoring

In structural health monitoring (SHM), the scarcity of labeled data representing different structural states and structural configurations remains a key challenge for developing robust and generalizable deep learning models. Building upon a reported attention-enhanced star generative adversarial network with gradient penalty (AT-StarGAN-GP), this study examines its application to generating synthetic structural vibration signals that reflect both cross-structure domain differences and multiple localized structural state scenarios. The framework supports conditional translation across domains, learning from acceleration responses measured on a source structure to generate synthetic responses for a target structure under structural state conditions. The architecture integrates multi-type attention mechanisms, including self-attention, channel-attention, and spatial-attention, and evaluates four conditioning strategies: concatenation (Concat), feature-wise linear modulation (FiLM), adaptive instance normalization (AdaIN), and conditional instance normalization (CIN). A dual-conditional formulation based on joint location and structural state level is employed. The framework is evaluated using two small-scale physical bridge models, each instrumented at six joint locations, with structural state variations simulated at selected joints using three mass levels (no mass, low mass, and high mass). Evaluation across modal, temporal, distributional, and spectral metrics shows that CIN provides improved performance, achieving a modal assurance criterion of 0.894, frequency response assurance criterion of 0.872, the lowest distributional discrepancy (maximum mean discrepancy 0.0006), and an 86% pass rate in Kolmogorov–Smirnov tests. Attention mechanisms improve localization accuracy by 34.35% and reduce entropy by 6.98%, while self-attention reduces cross-domain Kullback-Leibler (KL) divergence by 24.5%. The AdaIN variant shows more consistent behavior under unseen-domain scenarios (MAC 0.8207–0.8405). All evaluated configurations retain over 94% of baseline performance under 25% noise. Five-fold cross-validation confirms training stability, and Friedman–Nemenyi testing indicates statistically significant separation from weaker baselines, although differences relative to mid-tier methods remain limited in some comparisons. Overall, the results indicate that the extended AT-StarGAN-GP framework can generate structurally consistent synthetic vibration responses across different structures and state conditions under controlled laboratory settings.