Lag-informed thermal response modeling for uncertainty-aware bridge girder-end displacement anomaly detection: A field case study

Bridge girder-end displacement (GED) is strongly influenced by nonlinear, multichannel, and delayed thermal effects, which can mask weak structural anomalies and undermine fixed residual thresholds. This study develops an uncertainty-aware anomaly-detection framework that combines lag-informed response prediction, dynamic conformal calibration, and sequential evidence fusion. A squeeze-and-excitation temporal convolutional network (SE-TCN) learns the mapping from historical multichannel temperatures to GED, while the Geometric Mean Optimizer (GMO) jointly selects the lag order and architecture-specific hyperparameters under a strictly separated training-validation protocol. Dynamic conformal prediction then constructs an empirically calibrated, time-adaptive normal-response envelope, while anomaly-aware residual admission mitigates calibration contamination by excluding samples classified as anomalous. A fused sequential conformal evidence detector (FSCE) integrates normalized interval exceedance, smoothed conformal evidence, residual bias, and temporal persistence. The framework is evaluated using the field-monitoring data from a single-tower cable-stayed bridge and nine injected anomaly scenarios involving offset, drift, and thermal-sensitivity degradation. The final predictor achieves an RMSE of 0.0479 mm, a MAPE of 0.1580%, and an R 2 of 0.9924. At a nominal coverage level of 95%, the empirical coverage is 94.83% with a mean interval width of 0.2199 mm. Across the nine anomaly scenarios, Proposed-FSCE yields zero false alarms and F1-scores ranging from 0.7950 to 0.9993. Anomaly-aware updating limits the end-inflation ratio to 1.00-1.08, compared with 2.03-7.66 under unconditional updating. These results support the proposed framework as an empirically calibrated approach to bridge-displacement anomaly detection.