A nonlinear structural pulse-like seismic response prediction method based on pulse-like identification and decomposition learning

Accurate and fast prediction of structural response under seismic action is important for structural performance assessment, however, existing deep learning-based prediction methods do not consider the effect of pulse characteristics of near-fault pulse-like ground motions on structural response. To address the above issues, a new method based on wavelet decomposition and attention mechanism-enhanced decomposition learning, i.e. WD–AttDL, is proposed in this study to predict structural response under pulse-like ground motions. This method innovatively combines a WD-based velocity pulse-identification method with decomposition learning, where decomposed pulses and high-frequency features are used as inputs to the neural-network model, thus simplifying the identification of pulse features for the model. The decomposition learning model integrates several types of neural network components such as convolutional neural network feature extraction submodule, long short-term memory neural network temporal learning submodule and self-attention mechanism submodule. In order to verify the accuracy and validity of the proposed methodology, three sets of case studies were carried out, including elasto-plastic time-history analyses of planar reinforced concrete (RC) frame structures, a three-dimensional RC frame structure, and two types of masonry seismic isolation structures. Compared with existing structural seismic response models, WD–AttDL synergistically integrates the advantages of different modules and thus offers a higher prediction accuracy. In particular, it reduces the peak error of the predicted response, which is important for the evaluation of structural performance. In addition, WD–AttDL has a great potential for application in fast vulnerability and reliability analysis of pulse-like earthquakes in nonlinear structures.