Speaker
Description
Various nuclear test monitoring techniques require numerous waveforms and expect good spatial coverage of seismic source. The examples, inter alia, are master-event based location and event discrimination, needing data augmentation due to class imbalance and synthetic seismogram simulation can make a job. Generative Networks (GN) are a novel way of producing realistic, high-quality synthetic seismograms.
Here, we mostly concentrate on the waveform’s generation for the master-event based location. For the single component data, we use conventional temporal GN. For multicomponent seismic data, we adopted some multivariate representations of the neural network, which can treat such data simultaneously, considering neural network as multidimensional object, like hypercomplex-valued networks with 4D neurons. Hypercomplex Generative Adversarial Networks (QGAN) and Variational Autoencoders (QVAE) are the variants of generative networks that uses quaternion-valued inputs, weights and intermediate representations. In hypercomplex-valued time series, the components are intrinsically tied together by the algebraic rules of the hypercomplex domain (e.g. quaternion or octonion multiplication rules). These relationships make hypercomplex representations particularly useful for modeling phenomena with inherent coupling between dimensions, while quaternionic convolution in QCGAN are less resource-demanding than their non-quaternionic counterparts, with four times smaller parameter set size. Transformer-based multivariate models are also considered here.
| [email protected] |
