Determining the Availability of Analysis Using CNN with Earthquake Ground Motion Time History

Abstract

The analysis of earthquake ground motion is extremely important for improving safety in structural design. However, some ground motion data contain various noises such as aftershocks that make them unsuitable for analysis, and the work cost of manually determining the availability of huge amounts of data has become an issue. In this paper, we aimed to improve work efficiency by automatically determining the availability of ground motion data using image recognition technology. We applied ResNet34 to images of acceleration time history graphs of ground motions and performed image classification. The four classification labels are as follows: 1. Normal ground motion, 2. Events different from the main motion (aftershock events), 3. Large noise relative to the main motion (SN failure), and 4. Difficult to confirm the rise and convergence of the main motion (insufficient recording time). The accuracy rate was 97.3%. We also investigated the pixels that contribute to the results. This showed that normal ground motions showed high activation near the main motion, while aftershock events showed high activation in the valley shape between peaks, and for SN failure or insufficient recording time data, high activation throughout the data. From these results, we confirmed that the event features of each ground motion were captured by the image recognition technique using deep learning.