Leakage-controlled Bayesian structural health monitoring of earthquake-damaged RC frames using mechanism-aware evidence and experimental demonstration

Reliable post-earthquake structural health monitoring (SHM) requires interpretable, uncertainty-aware condition inference frameworks that are resistant to inflated performance estimates. In this study, a leakage-controlled Bayesian SHM framework for reinforced concrete frames was developed using simulation-derived diagnostic evidence and an experimental demonstration. The damage state, stiffness loss, and residual capacity were defined using mechanically distinct response quantities representing the maximum interstory drift, secant stiffness degradation, and postpeak strength retention, respectively. The simulation database comprised 528 nonlinear finite element pushover analyses covering regular full-scale frames, regular laboratory-scale frames, and vertically irregular stilted frames with different heights. The diagnostic evidence was organized into design, mechanism, and response descriptors, with all empirically derived categorization thresholds estimated within each training fold. A target-specific, same-scalar leakage-control procedure was introduced by withholding descriptors derived from the same response scalar used to define each target condition. For the Bayesian network, the macro F1 decreased from 0.911 to 0.626 for the damage state, from 0.933 to 0.741 for stiffness loss, and from 0.907 to 0.645 for residual capacity after leakage control. The leave-one-geometry-out validation produced mean macro F1 values of 0.501, 0.601, and 0.552, indicating configuration-dependent transferability. Posterior calibration yielded expected calibration errors of 0.138, 0.104, and 0.070 for the damage state, stiffness loss, and residual capacity, respectively. In the experimental demonstration, the four tested reinforced concrete frames produced significant damage probabilities of 0.888–0.943 and substantial stiffness-loss probabilities of 0.955–0.996. These predictions were consistent with the severe damage observed but provided limited discrimination between individual specimens. The proposed framework provides an auditable basis for uncertainty-aware postearthquake screening while avoiding direct target-defining shortcuts.