Story drift and damage level estimation of buildings using relative acceleration responses with multi‐target deep learning models under seismic excitation

Abstract

Damage detection is one of the primary purposes of structural health monitoring to inform catastrophic risks of structures right after extreme loadings such as earthquakes and hurricanes. In structural design codes, story drifts are considered as an indicator to estimate the damage states of structures. For instance, when the story drift ratios achieve 0.2-0.4%, light damage may be present in a building. In addition, the remaining stiffness ratios can also reveal the damage levels of a structure. Previous studies have shown that structural stiffness changes can affect the frequency responses of structures, for example, changing the locations of poles in frequency response functions. In this research, two multi-target neural network models are developed to concurrently estimate story drifts and remaining stiffness ratios using floor accelerations under seismic excitation. One of the multi-target neural network models focuses on developing a physics-guided loss function with a parallel model combination. Meanwhile, the other neural network model sequentially integrates two deep learning approaches by transfer learning. For example, the long short-term memory units estimate story drift responses from floor accelerations. Then, the short-time Fourier transform layers of floor accelerations yield the remaining stiffness ratio estimation. The proposed models are numerically investigated and experimentally verified. As a result, both models can estimate story drift and remaining stiffness ratio using the proposed neural network models.

Hybrid surrogate model combining physics and data for seismic drift estimation of shear‐wall structures

“…Zhang et al 24 introduced a physics-informed neural network that embeds structural dynamics equations into the loss function, thereby significantly improving the prediction accuracy of the seismic responses of building structures. Chou et al 25 formulated a novel neural network that could simultaneously estimate the story drift and the remaining stiffness ratio of building structures through a physics-guided multi-target loss function.…”

“…Compared with data-driven models, hybrid ones offer enhanced performance with limited data by using physical principles to reduce the reliance on data. However, existing hybrid models either fail to consider structural design parameters 24,25,28,29 or are only capable of considering preliminary ones (e.g., number of stories and overall dimensions). 8,27 The urgent necessity for developing a novel hybrid model is highlighted due to a lack of proper consideration for detailed design parameters (e.g., component layout and section sizes), as required by the optimization tasks.…”