Secondary collapse of semi-ruined reinforced concrete frame structures (RCFSSS) during aftershocks poses severe risks to rescuers. However, existing data-driven models are hindered by limited experimental data, domain discrepancies, and unstable performance under complex seismic perturbations, leaving a clear gap in achieving reliable and generalizable early-warning predictions under small-sample conditions. This study proposes a transfer-learning–guided hybrid Dilated CNN–BiLSTM–Transformer (DCBT) model that integrates extensive numerical simulations with limited high-fidelity shaking-table experiments. Predicted vertical-velocity (VV) histories are converted into cumulative absolute vertical-velocity (CAVV) curves to derive interpretable risk thresholds. Comprehensive ablation and comparative experiments show that the proposed model outperforms conventional hybrid models and baseline methods in both accuracy and robustness. Transfer learning effectively bridges simulation–experiment gaps, yielding substantial improvements in stability and generalization under varying seismic conditions. Overall, the proposed framework provides a practical and scalable early-warning tool for rapid secondary-collapse risk assessment and informed decision-making in post-earthquake rescue scenarios.
Xu et al. (Sun,) studied this question.