Study on modeling for damping devices using deep learning

(Verification of analytical accuracy for seismic response analysis of a Single-degree-of-freedom System)

Abstract

This paper presents a data-driven modeling approach for damping devices using deep learning, which can be applied to the seismic response analysis. The proposed model consists of four-layer stacked Long Short-Term Memory and is capable of creating hysteresis loops for the damper whose damping coefficient switches at a specific value. The training data for the proposed deep learning model are analytical values of displacement, velocity, and force obtained from the fundamental waves, such as sinusoidal wave inputs. Dynamic analysis is performed using a Single-Degree-of-Freedom system seismic response analysis model in which the proposed model is adapted to both damping and spring elements, and the results are compared with those obtained using the Maxwell model by using response waveforms and response spectrum. This paper demonstrates that the proposed damping model can capture the force-displacement relationship of oil damper, which has nonlinear characteristics without using equations, and that the seismic response analysis results obtained with the proposed model are as accurate as those obtained with the existing methods.