Data missingness in structural health monitoring of civil engineering systems is a common problem caused by a combination of human and environmental factors. Traditional statistical imputation methods struggle to effectively handle large-scale, high-dimensional feature data, while existing imputation approaches based on generative adversarial networks (GANs) often encounter challenges such as gradient vanishing and mode collapse. To address the imputation needs of high-dimensional nonlinear missing data in civil engineering applications, this study proposes a novel Bayesian-optimized dynamic attention generative adversarial network framework (BO-GAN-DA). The framework integrates Bayesian optimization with a dynamic attention mechanism, enabling accurate imputation of missing values, robust handling of outliers, and quantification of imputation uncertainty using only partial observed data, without relying on a complete dataset for supervised training. Furthermore, the method reduces dependence on computational resources. Evaluation tests conducted on a reinforced retaining wall finite element model and a scaled steel frame model demonstrate that the BO-GAN-DA model maintains strong robustness under noise with varying standard deviations and delivers superior imputation performance across different missing rates. Compared with benchmark models, the proposed method achieves an approximately 6.8% improvement in imputation accuracy and a reduction of about 47% in mean absolute error (MAE), thereby validating its effectiveness and superiority.
Gao et al. (Sat,) studied this question.