Attention-enhanced convolutional neural network for predicting the full‐profile seismic response of buildings

Data-driven surrogate models can predict the seismic response of buildings in near real-time. However, the current surrogate models for predicting the buildings’ full-profile seismic response (the peak seismic response of all floors) are only applicable to buildings with a specific number of floors due to their output dimensions are immutable. To solve this problem, a data pre-processing method based on spline interpolation is proposed to unify the data dimensions of the full-profile seismic response of different buildings. Then, the attention-enhanced convolutional neural network was established to predict the full-profile seismic response of buildings with arbitrary number of floors. The channel and spatial channel attention mechanism is adopted in the proposed surrogate model to fully extract the time–frequency features of ground motions as supplementary input of seismic intensity measures and building parameters to ensure the prediction accuracy and generalization capability. The excellent performance on numerical simulation and field sensing cases validates the effectiveness of the proposed method, and it is not necessary to retrain the surrogate model when predicting the full-profile seismic response of the new buildings under new earthquakes.