2024 年 10 巻 31 号 p. 1159-1164
Recovering accelerograms from response spectra can be helpful in several applications in earthquake engineering, such as hazard assessment and earthquake resistant design. However, this inverse problem is challenging because there is insufficient information in the response spectra to determine ground motion time histories. Supervised learning techniques of neural networks have been employed with wavelet packet transform to address this problem. Nevertheless, such approaches still require human-labeled training data, which can lead to memorizing the training set. An alternative approach to overcome the limitation is the use of the generative adversarial network (GAN), which is an unsupervised learning approach. GAN does not require any paired information like human-labeled data that has been used in computer vision fields to generate new data samples by capturing the underlying characteristics of the training data. The method uses random input in the form of latent vectors, which are processed by the generator model of GAN to produce a new data sample. This paper introduces the GAN-based method to produce accelerograms not only triggered by latent vectors but also conditioned on pseudo-spectral acceleration (PSA) that is additionally provided as input to the generator model. Several error metrics are utilized to compute the differences between the PSAs. The PSAs of generated samples generally match well with the input PSAs on which the generator model was conditioned, suggesting the proposed method is effective in solving the inverse problem.
