Investigation on employment of time and frequency domain data for predicting nonlinear seismic responses of structures

This study examines methods of introducing time and frequency domain data for predicting structures' nonlinear seismic responses. Ground motion records or time history structural responses, widely used in existing seismic response prediction methods, and continuous wavelet transform (CWT) data for time domain response were used to analyze the effects of frequency and time domain data on nonlinear structural behavior predictions that change over time. A convolutional neural network (CNN), one of the machine learning techniques, was introduced to predict seismic responses. This research proposes four CNN models in which time or frequency domain data and combinations of time-frequency domain data were utilized. In the four models, the time history nonlinear displacement response of the target structure is set as the CNN's output. The proposed models were applied to the nonlinear seismic response prediction of 3D frame structures to assess their prediction performance, and the valid types of input data for predicting nonlinear behavior were discussed based on the assessment. To apply and evaluate the presented method, numerical and experimental studies were conducted on three-dimensional reinforced concrete structures.