Missing monitoring data detection and imputation of cable-supported bridge using multisource fusion and generative parallel deep learning model

Bridge structural health monitoring (BSHM) has become increasingly widespread due to its critical role in ensuring the safety and reliability of modern bridge structures. However, maintaining the integrity of sensor data remains a significant challenge, as environmental factors and sensor malfunctions often result in missing data and anomalies. To address these issues, this study presents a novel generative deep learning-based data cleaning model, termed the convolutional autoencoder (CAE) enhanced by Transformer (CAT), designed for the detection and imputation of missing acceleration data in the presence of environmental influences. The CAT model combines CAEs for local feature extraction with Transformer encoders to capture global dependencies and long-term temporal correlations. A parallel network architecture is employed to process acceleration data alongside environmental factors, such as wind speed and temperature, enabling the integration of structural and environmental information. Residual skip connections are introduced to ensure that critical features are consistently preserved throughout complex, nonlinear processing layers, ultimately enhancing reconstruction accuracy and the overall robustness. Field data monitored from a long-span suspension bridge is used to evaluate the performance of the proposed model. Experimental results demonstrate the robustness and accuracy of the CAT model, achieving a mean squared error (MSE) of 0.002998 and the Pearson correlation coefficient (PCC) of 0.9831 under moderate data loss conditions. Even under severe scenarios with up to 50% missing data, the model maintains reliable performance, with an MSE of 0.012931 and a PCC of 0.9257. The CAT model outperforms traditional approaches in both reconstruction precision and robustness, addressing challenges associated with sensor anomalies in large-scale monitoring systems.