Missing-aware measurement correction and conformal uncertainty quantification for bridge modal-frequency anomaly detection
Abstract
Modal frequencies identified from bridge vibration measurements are indirect structural measurands whose long-term interpretation is complicated by environmental and operational variability and asynchronous sensor outages. This study develops a missing-aware measurement-correction framework that maintains normal modal-frequency references from incomplete multi-source monitoring data and quantifies their uncertainty for anomaly screening. Physically interpretable temperature, traffic, wind and cyclic-time features are combined with decay-filled values, observation masks and elapsed-time descriptors in a residual temporal convolutional network. Adaptive conformal prediction (CP) constructs mode-specific intervals from normal residuals, with guarded updating to prevent suspected state changes from contaminating calibration. Persistent coordinated upper-bound exceedance of at least two modes for three consecutive samples then triggers an alarm. The framework is evaluated using 30 min records collected from Mingzhou Bridge between January 2018 and May 2020, including a documented deck-pavement replacement. Missing-aware inclusion retained 34 501 labelled endpoints, compared with 17 369 complete-window endpoints, while the normal-period modal references achieved R2 values of 0.864, 0.794 and 0.746 and empirical coverage near the nominal 95%. During the intervention period, adaptive calibration reduced the false-alarm rate from 2.79%–1.31% relative to static CP, and the persistent rule achieved 99.04% recall and an F1 score of 99.24%. Controlled asynchronous-missingness tests confirmed uninterrupted reference availability with graceful accuracy degradation. Site-specific validation on an independent cable-stayed bridge further supported the applicability of the prediction and uncertainty-calibration workflow. These results demonstrate uncertainty-calibrated modal-frequency interpretation under incomplete long-term monitoring records.