Deep learning model for prediction of non-linear cyclic hysteresis of seismic isolation devices: Full-scale experimental validation

Abstract

Recurrent neural networks predicting the non-linear hysteresis behavior of specific triple pendulum bearing (TPB) and lead rubber bearing (LRB) seismic isolation devices, are developed, and tested in this paper. Experimental datasets of full-scale isolators were derived from a shake-table test program of a five-story building specimen, performed at the Hyogo Engineering Research Center of Miki, Japan. Data measured during several table motions of different amplitudes, frequency content, and durations, are appropriately processed to construct a substantial TPB/LRB dataset of 158/55 samples. A comprehensive framework is proposed to process the data, to optimize the network architecture, and to train, validate, and to test the machine learning models. The comparisons with reference experimental data showed that developed models could predict the two-dimensional hysteresis behavior of studied isolators with a very good accuracy. The R2 value of the TPB/LRB model was of 0.83/0.96 on new unseen dataset. The illustration of representative predictions showed the power of a single model to capture different and irregular hysteresis patterns, performed by a typical isolator under variable and realistic loading conditions and that can't be described by conventional analytical models. The generalization capability of these surrogate models on such a substantial data, revealed the benefit of applying machine learning to solve complex structural engineering problems.