| Abstract |
As one of the major disasters on earth, earthquakes and their impacts cover a wide range of social, economic, and environmental aspects. However, forecasting earthquakes is currently impracticable, so many researchers have adopted various measures to cope with possible earthquake effects, such as earthquake early warning (EEW), structural health monitoring (SHM), earthquake-resistant structures, etc. This development allows us to respond to events and reduce impacts quickly. Although those advances are successful, they heavily rely on the availability and variety of earthquake data, which is often limited for large earthquakes or areas that are not earthquake-prone. Therefore, not much earthquake data can be used for structural analysis due to the deficiency of observation. In order to address this issue, this study introduces the variational autoencoder (VAE), a machine learning (ML) based approach. VAE is a generative model capable of automatically extracting the seismic features and reproducing the earthquake data. Moreover, it can generate artificial earthquake waveforms with diversity by using the extracted features, which provides a new way to synthesize waveforms. In this study, the 921 earthquake was first adopted, and a total of 293 waveforms were used for training. The preliminary results show that VAE is great while generating artificial earthquake waveforms. Subsequently, VAE is applied to the earthquake data from seven regions in Taiwan, and the final results verify the feasibility. As a result, VAE can provide merits for the development of structural and earthquake engineering, and the paper ends by suggesting future research.
|