AT-StarGAN-GP: Attention-enhanced multi-domain signal translation for structural health monitoring

This paper presents AT-StarGAN-GP, an enhanced StarGAN architecture with gradient penalty and dual-attention mechanisms for structural health monitoring. The proposed model overcomes limitations of conventional GANs in structural vibration analysis by integrating temporal and channel attention to capture global temporal dependencies and local feature interactions. It employs a refined Wasserstein loss with gradient penalty and adaptive λ -scaling to stabilize training and balance multiple objectives. The model achieves high accuracy in modal frequency detection, producing nearly identical real and synthetic estimations across all modes. Damping ratio estimations show deviations below 5%, confirming precise structural response modeling. Frequency-domain and singular value decomposition (SVD) analyses yield mean magnitude-squared coherence (MMSC) values between 0.973 and 0.989, validating stability and fidelity. In terms of computational efficiency, AT-StarGAN-GP processes all 30 joints in 313.45 ms (∼10.45 ms per joint), maintaining real-time feasibility and outperforming CycleGAN (≈9 s) while matching Vanilla StarGAN performance. Although Eigen perturbation methods (Modal Assurance Criterion (MAC), Damage Localization Vector (DLV), Flexibility, Eigenvalue) are faster (0.25–12.67 ms), they lack generative capability essential for data recovery and nonlinear damage modeling. Sensitivity analysis shows optimal coherence (MMSC > 0.97) when λ -parameters are moderately weighted (10–15), and attention is fully enabled. Attention mechanisms significantly enhance temporal and inter-sensor spectral alignment, confirming AT-StarGAN-GP as a real-time, generative, and accurate solution for vibration-based SHM.