Conditional generation of artificial earthquake waveforms based on adversarial networks

Abstract

Earthquake waveforms are necessary for simulation, verification, and validation during the development of structural and earthquake engineering. However, due to the fact that the processes of earthquake preparation and generation are extremely complex and the human observations cover a relatively short period compared to large earthquake cycles, recorded earthquake waveforms are often limited and quite imbalanced. Consequently, the generation of artificial earthquake waveforms is meaningful and can produce useful earthquake data. In this study, a new method based on a generative approach is proposed by leveraging machine learning (ML) techniques. The generative models based on generative adversarial network (GAN) and conditional GAN (cGAN) are introduced, and the adversarial training between the generator and the discriminator is utilized to learn how to generate artificial earthquake waveforms. In addition, several criteria are exploited to terminate the adversarial training early in order to prevent modal collapse. Besides, different earthquake datasets and several earthquake features are used to verify the proposed method. The preliminary generation is carried out using the dataset recorded during the Chi-chi earthquake, and the generated samples show similar earthquake features to those extracted from the earthquake dataset. Subsequently, conditional generation is conducted using several sub-datasets recorded in different areas around Taiwan. The results demonstrate that the well-trained model can efficiently generate waveforms according to a pre-specified seismic intensity, with an overall accuracy of up to 90%. Although the model is reliable for Taiwan and in general to restricted subdomain of the probability distribution support, the proposed method shows its ability to produce new samples for a given class, even when the number of generated samples exceeds the number of training samples.