Point-Adaptive Detector Selection and Alert Confirmation for Dam GNSS Displacement Monitoring

Thresholds calibrated on historical data produce excessive false alarms when conditions change, and field-confirmed abnormal events are too scarce to supervise alarm calibration. This study develops a point-adaptive detector selection and alert confirmation workflow for dam Global Navigation Satellite System (GNSS) displacement monitoring. At each of nine monitoring points, a dual gate screens fifteen candidate anomaly detectors into a small point-specific ensemble, and a four-level confirmation policy—M-of-N temporal persistence, weighted detector support, an independent displacement-band escalation path, and event aggregation—converts the score into auditable alerts for engineering verification rather than a prediction of future instability. On an eighteen-month, nine-point GNSS record, the confirmation policy applied to the single detector retained at all nine points (TCN-AE) detects 68.1% of chronologically injected events, and the point-adaptive fused ensemble detects 66.7% across all nine points (block-bootstrap 95% confidence interval (CI) 51.4–81.9%), with a cleaned-negative alarm rate of 0.0045% and 99.98% precision on the cleaned negative class (full-negative-class false-positive rate (FPR) 0.09% as an upper bound; precision 99.68% as a lower bound, dominated by correctly flagged operational episodes); the prevalence-adjusted positive predictive value is about 41% at a 0.1% operational prevalence. Fusing the point-adaptive ensemble did not improve on a single detector granted the same policy (TCN-AE alone 68.1% and the point-wise best single 69.4% versus the fused ensemble 66.7%), so fusion is reported as a negative finding: the operating point is set by the confirmation policy, not by the choice or combination of score sources. Claims are restricted to the five of nine points with operable orange-level recall; the remaining four are routed to manual review. The workflow is an operational decision layer that turns anomaly scores into a plannable verification workload for dams where abnormal labels are scarce.