Data‐driven modeling and prediction on hysteresis behavior of flexure RC columns using deep learning networks

Abstract

Hysteresis behavior of structural components has been one of the research focus for the structural engineering community for decades, comprehensively determines the structural performance and safety, and plays an important role in structural disaster mitigation. It is of great significance to continuously monitor structural responses and accurately characterize structural hysteresis. Currently, the nonlinear properties of real‐world structural components cannot be obtained before its failure. Thus, a historical database is collected firstly. Then, a data‐driven analysis method is proposed for predicting hysteresis behaviors of reinforced concrete (RC) columns. A bidirectional LSTM (BLSTM) network is developed to model and predict hysteresis curves. The data with unfixed lengths are specially processed to simultaneously guarantee a uniform size and avoid data loss, and the clipping layers are inserted in the model to clip off inferior predictions and improve the accuracy. This methodology is systematically studied and validated by employing a sythetic database generated by numerical simulation and the full‐scale experiment database named PEER database. Result shows that proposed BLSTM can predict the hysteresis curves of the RC components with acceptable accuracy and robustness. Moreover, the interpretability analysis is performed on identifying the learning and prediction principle of the BLSTM model on hysteresis prediction and its accuracy variation under different model architectures.