Enhanced dual-stage iterative recovery framework for incomplete data using compressed sensing and convolutional neural networks

In the structural health monitoring, missing data from faulty sensors significantly undermines the reliability of real-time structural condition evaluation. Existing researches primarily address this issue by employing either compressed sensing algorithms or deep learning-based spatiotemporal models to recover missing signals. However, the former approach is often constrained by strong sparsity assumptions, while the latter fails to fully utilize prior history response, thereby limiting the performance of the network. To address these limitations, this study proposes an enhanced dual-stage iterative recovery framework (EDIRF) for incomplete data using both compressed sensing (CS) and convolutional neural networks (CNN). In the first stage, the CS-based imputation was calculated according to the prior history response of the faulty sensor. In the second stage, CNN is employed to establish a nonlinear spatiotemporal mapping model for the possible response prediction. Finally, the recovery data could be obtained through the proposed EDIRF. Validation on two real bridges demonstrates that the proposed EDIRF achieves high recovery accuracy for both single-channel and multi-channel faulty types, maintaining superior performance with R² consistently above 0.85 under patterns of missing rates of 90% or complex block missing. Additionally, taking the seasonal effect into consideration could enhance the reconstruction performance of EDIRF. Moreover, the surrogate channel strategy is proposed to solve the unavailability of faulty channel in providing prior history responses. Among the current mainstream methods, EDIRF outperforms in both reconstruction accuracy and training efficiency. In summary, the proposed EDIRF provides reliable technical support for bridge health monitoring, demonstrating practical value in engineering.