16:45 〜 17:00
[SSS11-06] 加速度エンベロープのWasserstein内挿を用いた広帯域地震動合成
Hybrid approaches to broadband (BB) ground motion synthesis combine long-period (LP) and short-period (SP) waveforms calculated by two methods suitable for each period range. They have been applied in research and practice, but it is disadvantageous that simulations are independently carried out under different assumptions, which can lead to incompatible time histories and frequency properties.
This study explores an approach for maintaining consistency between LP and SP components using an empirical relationship of past observation records. We propose a machine learning method that generate SP waveforms from LP waveforms obtained using physics-based simulations. Acceleration envelopes and Fourier amplitude spectra are transformed, and they are combined to produce a broadband waveform. To effectively obtain the relationship of envelopes from limited amount of data, we formulate the problem as the conversion of probability distributions to allow the introduction Wasserstein distance, and embed pairs of LP and SP envelopes into a common latent space to improve the consistency of the entire waveform. An experimental application to the 2008 M7 off Ibaraki earthquake demonstrates that the proposed method exhibits superior performance compared to existing methods and neural network approaches. In particular, the proposed method reproduces global properties in the time domain, which confirms the effectiveness of the embedding approach and the advantage of the Wasserstein distance as a dissimilarity measure of envelopes.
This study explores an approach for maintaining consistency between LP and SP components using an empirical relationship of past observation records. We propose a machine learning method that generate SP waveforms from LP waveforms obtained using physics-based simulations. Acceleration envelopes and Fourier amplitude spectra are transformed, and they are combined to produce a broadband waveform. To effectively obtain the relationship of envelopes from limited amount of data, we formulate the problem as the conversion of probability distributions to allow the introduction Wasserstein distance, and embed pairs of LP and SP envelopes into a common latent space to improve the consistency of the entire waveform. An experimental application to the 2008 M7 off Ibaraki earthquake demonstrates that the proposed method exhibits superior performance compared to existing methods and neural network approaches. In particular, the proposed method reproduces global properties in the time domain, which confirms the effectiveness of the embedding approach and the advantage of the Wasserstein distance as a dissimilarity measure of envelopes.