Reconstructing the full-profile seismic time-history response of buildings based on deep learning and sparse sensor monitoring data

Methods for assessing structural damage state based on sensor monitoring data have been widely studied, but existing methods cannot achieve sufficient accuracy when the sensor arrangement is very sparse. In most cases, only a few sensors are installed on the buildings due to cost issues, so this study proposes a data-driven method to reconstruct the full-profile seismic response of buildings from sparse monitoring data. A multi-scale CNN-based feature extractor is developed to fully extract multi-scale features from monitoring data and seismic excitations. The extracted multi-scale features and building features are then fused using the attention-enhanced LSTM neural network to accurately reconstruct the full-profile seismic response of buildings. In addition, a data processing method is proposed to unify the sample data dimensions, which makes the proposed surrogate model applicable to reconstruct the full-profile seismic time-history response of buildings of arbitrary number of stories. The excellent performance on numerical simulation and field sensing cases validates the effectiveness of the proposed method. In all testing cases, the average Pearson correlation coefficient corresponding to the time response reconstruction results is greater than 0.90, and the average peak error is less than 20 %.