A CNN-Based Ground Motion Selection Method Considering the Impact of Frequency and Time Characteristics of Ground Motions

Abstract

The time characteristics (TCs) of ground motions (GMs) significantly affect the seismic response of tall buildings; however, few existing GM selection methods effectively consider the impact of the TCs of GMs. This leads to noticeable uncertainty in the nonlinear response time-history analysis (NLRHA) results of tall buildings and substantial computational demands to secure a reasonable estimation of the structural seismic responses. This paper proposed a GM selection method considering the impact of frequency and time characteristics (SIFT) of GMs based on convolutional neural networks (CNNs). In the proposed SIFT method, the existing two-step GM selection procedure was adopted to select candidate GMs to effectively consider the site condition, GM duration, and impact of GM frequency characteristics. The proposed method developed the response diagram in the time domain (RDTD) to represent the impact of the TCs of GMs, which shows the relative magnitudes of seismic responses of single-degree-of-freedom systems with varying frequencies at any given moment throughout the duration of the earthquake. A CNN model was constructed and trained with transfer learning technique to learn the fuzzy features of the RDTD, establish the mapping relations between features of the RDTD and seismic responses of tall buildings, and finally select GMs from the candidate GMs. The proposed SIFT method and existing spectrum matching-based GM selection method were adopted to select GMs from different GM databases for the NLRHA of structures with different periods to verify the effectiveness of the proposed SIFT method. This method can ensure seismic responses calculated using fewer GMs are close to those calculated using a large number of GMs, thus considerably improving the computational efficiency of seismic assessment of tall buildings.