Deep HystereticNet to predict hysteretic performance of RC columns against cyclic loading

Abstract

This article attempts to reproduce hysteretic performance of HRB600 bar reinforced concrete columns under cyclic loading by adopting multivariate deep learning methods. Bidirectional long short-term memory (BiLSTM) and bidirectional gated recurrent units network (BiGRU) have been modified for temporal sequence related strength prediction with the consideration of multivariate parameters. To this end, a database with a 79,244 (row) × 12 (column) matrix is prepared for training, validation, and testing purposes. The inputs to the model include time intervals, loading displacement, concrete strength, axial compression ratio, longitudinal reinforcement tensile grade, longitudinal reinforcement diameter, longitudinal reinforcement ratio, stirrup tensile strength, stirrup diameter, spacing, and stirrup reinforcement ratio while the target is the cyclic load. By comparing the energy dissipation, accumulated energy dissipation, stiffness degradation and backbone curve of the hysteretic performance, the capacity of the adopted methods is compared with that of the finite element analysis. The comparison shows that the adopted method can recovery the hysteretic curves very well and the GRU method exhibits the most accurate prediction performance because of the highest R2 and the lowest MAE, MSE, and RMSE. The originality and innovation are mainly reflected in the use of time series data-driven deep learning model to replace the existing empirical formula and finite element method to reconstruct the nonlinear hysteretic behavior, so as to predict the seismic performance of the structure.