Spatiotemporal dependency data imputation for long-term health monitoring of concrete arch bridges

The structural health monitoring of bridge infrastructures is imperative for ensuring their uninterrupted functionality and mitigating potential hazards. Nevertheless, challenges arise Due to missing data and intricate latent dynamics embedded in field sensing measurements, complicating the forecasting efforts. Addressing these challenges, this paper introduces a pioneering CCF-BiGRU model designed for the imputation of missing data in bridge strain monitoring. This model harnesses the spatial correlations among sensor data, analyzed through cross-correlation algorithms, combined with the predictive capabilities of the BiGRU neural network. By conducting experiments with 5%-20% missing completely at random data in authentic bridge monitoring contexts, the study substantiates the efficacy of this approach. The CCF-BiGRU model surpasses its counterparts— BiGRU, BiLSTM, GRU, and LSTM models—in several critical performance metrics (root mean square error, correlation coefficient, and relative accuracy). Notably, its performance metrics remain consistent even as the proportion of missing data escalates. Specifically, in scenarios with 5% to 10% data omission, the CCF-BiGRU model consistently exhibits a uniform error distribution in MASE evaluations, highlighting its robustness. Although CCF-BiLSTM model shows similar interpolation performance, its computational cost is much higher. These compelling results validate the efficiency and reliability of the CCF-BiGRU method, a data-driven solution that not only shows promise but also excels in computational efficiency. It adeptly predicts and fills gaps in bridge strain monitoring data, thereby ensuring precise evaluations of bridge health.