Damage detection plays a pivotal role in structural health monitoring. As indicated in FEMA and ASCE, structural damage relies on story drifts as a fundamental criterion for categorizing damage states and assessing risk levels. In addition, past studies showed that structural stiffness changes were closely linked to the extent of structural damage due to earthquakes. However, both story drifts and stiffness changes are rarely evaluated concurrently to determine structural damage. In this study, three multi-target neural networks are developed using floor accelerations of buildings under seismic excitation to estimate story drifts and remaining stiffness ratios. Notably, all network architectures are identical. The three neural networks in this study differ in applying distinct loss functions and training strategies to assess and compare the performance of the models. All networks are compared through numerical investigation using a three-story finite element model and experimentally verified using a seismically excited full-scale building.
|