An Adaptive Multi-Method Approach to Enhancing the Quality of Anomalous Data in Bridge Structural Health Monitoring

Bridge structural health monitoring (SHM) data often contain isolated outliers and consecutive gaps that can bias condition assessment. This study develops an adaptive framework combining outlier screening with multi-method gap reconstruction. Outliers are screened using sliding-window value and first-difference interquartile-range criteria and cross-checked against correlated channels to avoid suppressing genuine structural responses. Missing data are reconstructed using time-feature-enhanced recursive XGBoost with endpoint correction, amplitude-restored daily-profile reconstruction, and cross-channel XGBoost. Instead of imposing a fixed gap-length rule, leakage-controlled five-fold temporal validation selects a feasible method by target channel, gap length, and auxiliary-channel availability. Six months of 10 min temperature, deflection, and strain measurements from an in-service cable-stayed bridge were evaluated using artificially masked gaps of 3–432 samples. The best method varied with channel and gap scale: cross-channel reconstruction consistently performed best for deflection, whereas self-series and daily-profile methods remained necessary when cross-channel relationships were weak or auxiliary channels were unavailable. Across four channels, weighted root mean square error decreased by 32.6–96.0% relative to the best-performing conventional baseline. In the supplementary comparison with modern baselines, the proposed method reduced RMSE by 9.8%, 73.7%, and 36.0% for temperature, deflection, and strain ε1, respectively, relative to the best-performing modern baseline for each channel, but did not outperform the modern baselines for strain ε2. Partial-availability tests showed that auxiliary-channel quality mattered more than quantity. The framework provides a transparent, deployment-oriented strategy for bridge SHM data-quality enhancement, although long gaps without reliable synchronous evidence remain difficult to reconstruct.