5:15 PM - 6:45 PM
[SSS10-P09] Sample generation of ground-motion time-series waveforms
Keywords:input ground motion, sample generation, ground-motion simulation
For reliable risk assessment reflecting regional characteristics at specific locations or structures, considering the variability in input ground motion due to uncertainties in earthquake scenarios such as fault location, geometry, and rupture process is essential. However, conducting earthquake motion simulations for each of the numerous fault models with various scenarios, taking into account the complex ground structures and fault rupture, is not necessarily feasible.
In this study, aiming to ensure diversity in the input ground motion waveforms, we attempt to generate a new set of earthquake waveforms by using existing seismic simulation waveforms for a certain number of scenarios. The generated set of earthquake waveform data should exhibit "seismic waveform-like" characteristics in both the time and frequency domains and possess an "appropriate distribution" as part of the input ground motion waveform ensemble. To achieve this, we consider methods for sample generation based on feature extraction from the existing waveform data set and probability distribution estimation, ensuring that the newly generated waveforms inherit characteristics from the existing set.
First, utilizing an inverse fault-type fault model following the standard ground-motion prediction approach (recipe) with a heterogeneous medium model, we computed approximately 1500 seismic waveforms considering the variability in the spatial relationship between faults and observation points, as well as fault rupture styles (6000-dimension data per waveform). This set is referred to as the "existing waveform data set." As a quantitative representation of the features in the time and frequency domains, we calculated wavelet packet (WP) coefficients using WP transform, which reconciles time and frequency resolutions.
Next, we extracted dimension-reduced features from the WP coefficients and performed sample generation of these features following an assumed probability density function. The generated features were then decoded, and the WP coefficients were inverse-transformed to produce seismic waveform data. This collection is termed the "generated seismic waveform data set."
In this presentation, we explore two approaches for assuming the probability density function of the features: one based on a normal distribution using Variational Auto Encoder (VAE) and the other based on an empirical copula (Imai et al. 2021). We report the differences in the characteristics of the generated seismic waveform data sets resulting from these two distributions. We will further investigate the variability that should be present in the input ground motion ensemble, aligning with the goal of creating detailed building damage functions.
In this study, aiming to ensure diversity in the input ground motion waveforms, we attempt to generate a new set of earthquake waveforms by using existing seismic simulation waveforms for a certain number of scenarios. The generated set of earthquake waveform data should exhibit "seismic waveform-like" characteristics in both the time and frequency domains and possess an "appropriate distribution" as part of the input ground motion waveform ensemble. To achieve this, we consider methods for sample generation based on feature extraction from the existing waveform data set and probability distribution estimation, ensuring that the newly generated waveforms inherit characteristics from the existing set.
First, utilizing an inverse fault-type fault model following the standard ground-motion prediction approach (recipe) with a heterogeneous medium model, we computed approximately 1500 seismic waveforms considering the variability in the spatial relationship between faults and observation points, as well as fault rupture styles (6000-dimension data per waveform). This set is referred to as the "existing waveform data set." As a quantitative representation of the features in the time and frequency domains, we calculated wavelet packet (WP) coefficients using WP transform, which reconciles time and frequency resolutions.
Next, we extracted dimension-reduced features from the WP coefficients and performed sample generation of these features following an assumed probability density function. The generated features were then decoded, and the WP coefficients were inverse-transformed to produce seismic waveform data. This collection is termed the "generated seismic waveform data set."
In this presentation, we explore two approaches for assuming the probability density function of the features: one based on a normal distribution using Variational Auto Encoder (VAE) and the other based on an empirical copula (Imai et al. 2021). We report the differences in the characteristics of the generated seismic waveform data sets resulting from these two distributions. We will further investigate the variability that should be present in the input ground motion ensemble, aligning with the goal of creating detailed building damage functions.