In structural health monitoring (SHM), crack‐width measurements are critical indicators of damage severity and serviceability in concrete structures. However, field data are frequently affected by sensor drift, environmental noise, and transmission errors, resulting in anomalous or missing segments that undermine safety assessments. To overcome this challenge, a two‐stage deep learning framework is developed that couples anomaly detection with data repair. First, one‐dimensional crack‐width time series are transformed into Gramian angular field (GAF) images to enable cross‐domain feature representation. A convolutional neural network (CNN) then extracts spatiotemporal features from the GAF images, with the baseline model achieving an anomaly detection accuracy of 98.50%. Second, a conditional residual transformer (C‐R‐Transformer) incorporates auxiliary variables such as ambient temperature and a residual‐correction mechanism to reconstruct corrupted or missing data with high fidelity. Compared with classical interpolation and single‐stage deep models, the proposed framework exhibits greater robustness to complex anomaly patterns, significantly improving data integrity and accuracy. Field experiments further confirm its effectiveness and generalizability for multisensor bridge monitoring. By enabling reliable and interpretable data recovery, this study advances fault diagnosis, supports informed decision‐making, and strengthens the resilience of civil infrastructure monitoring.
Fan et al. (Thu,) studied this question.