Machine learning–driven structural health monitoring of a high-rise building on thick sediments via seismic ambient noise

Abstract

Seismic ambient noise enables non-invasive monitoring of high-rise buildings, but existing indicators trade floor-level localization for operational simplicity. Resonance-frequency changes track whole-building stiffness but poorly localize structural variations, whereas inter-floor interferometry localizes them through computationally intensive comparisons with reference records. Here we use interpretable machine learning to retain floor-level localization without repeated interferometric analysis. From 22 days of recordings on every floor of a 21-storey building on thick sediments in Shanghai, interferometry identifies wind-associated velocity reductions of about 2% at F6–F7 and F9–F10. An XGBoost model using only basement and roof resonance and polarization parameters predicts intermediate-floor dynamics, and its attribution profile shows transitions at the same levels. This correspondence indicates that the model retains floor-level spatial sensitivity. The candidate zones may reflect modular stiffness and load-transfer contrasts. Once trained, the model predicts the resonance frequency, vibration azimuth, and polarization dip of intermediate floors using only basement and roof recordings, without repeated inter-floor interferometry or pre-event reference records. Linpeng Qin and colleagues combine seismic ambient noise analysis with interpretable machine learning to monitor floor-level dynamics in a high-rise building on thick sediments. The model identifies transitions near floors F6–F7 and F9–F10 using only basement and roof recordings, consistent with interferometric velocity reductions.