A Deep Learning Framework for Seismic Response Reconstruction and Spatial Damage Assessment of Concrete Gravity Dams

To address the critical problem that sparse coverage of dynamic monitoring instruments in gravity dam projects limits the rapid identification of local damage during earthquakes, this study proposes an analytical framework that integrates virtual reconstruction of seismic responses with regional damage sensing. First, a detailed finite-element model of the validated Huangdeng dam–reservoir–foundation system was established. Subsequently, a time–frequency-constrained multiscale Wasserstein generative adversarial network (MWGAN-TF) was developed to reconstruct seismic responses at unmeasured locations using responses from monitoring locations. Time-varying displacement, acceleration, and normalized mode-shape-difference features were then extracted from the monitored and reconstructed responses. Finally, the developed response–modal hybrid neural network was used to accurately estimate the time histories of damage volume ratios in different subregions of the dam. The results show that the finite element model reasonably reproduces the overall dynamic characteristics and major seismic damage patterns of the dam. MWGAN-TF provides good overall performance in reconstructing time-domain responses and frequency-domain features. The combination of time-varying response and modal features improves regional damage-volume-ratio prediction, while virtual responses facilitate finer damage zoning. The proposed framework provides a systematic approach for seismic response reconstruction and rapid regional damage assessment of concrete gravity dams under sparse monitoring.