A Group-Aware Data Quality Stress Testing Framework for Bridge Structural Health Monitoring: Evidence from the Vänersborg and Z24 Datasets

Abstract

Bridge structural health monitoring (SHM) depends on sensor streams that may contain missing blocks, noise, spikes, drift, and recorded faults. We evaluated a frozen, group-aware stress testing protocol on 64 Vänersborg bridge opening events and 153 processed Z24 scenario × setup traces. Models used healthy training data only; degradations were restricted to held-out groups. Area under the receiver operating characteristic curve (ROC-AUC) and average precision (AP) were co-primary metrics. Vänersborg principal component analysis (PCA) baseline AUC and AP values were 0.500/0.454, respectively. In addition, 20% missingness yielded 0.188 and 0.327 (paired changes −0.313 and −0.126), and 0.5% spikes yielded 0.058 and 0.298 (−0.442 and −0.156), respectively. Z24 PCA baseline AUC and AP values were 0.731 and 0.929, respectively, and the studied perturbations changed both metrics only slightly. Under event power-scaled 5 dB noise, Vänersborg PCA AUC decreased by 0.221, while the other detectors increased. A new label-blind sensitivity analysis fixed one robust channel variance from the healthy training events. This equalized injected power across evaluation classes and changed the PCA ΔAUC to +0.075, showing that the original scaling asymmetry contributed materially to the detector-specific directions. Results are exact conditional summaries of two observed finite archives and do not establish universal quality, severity, or maintenance thresholds.