Application of a Paired Neural Network to Aftershock Identification

Abstract:

We train a Paired Neural Network (PNN) with a contrastive loss function to perform aftershock identification based on waveform similarity. Provided with seismic waveforms preprocessed into matching and non-matching example pairs, our network converts each waveform to a lower dimensional latent space representation. The network is then trained such that matching and non-matching example pairs have shorter and longer distances in their respective representations. Such techniques have proven useful in the field of one-shot learning, where the model may need to identify examples from different classes than those on which it was originally trained. In order to have strict control of our labeling, following Zhu et al. (2018) we generate 57,373 constructed waveforms consisting of high signal-to-noise ratio P-arrivals added to randomly selected seismic noises to act as our matching examples. The signals were recorded by the sparse IMS network from 2010-2019, while the noises were recorded by IMS and the University of Utah (UU) network from 2010-2019. In addition to the constructed waveforms, 12,495 real seismic waveforms from 2009 are included in our dataset to act as negative class examples from which non-matching pairs will be constructed. We apply the trained model and waveform cross-correlation on synthetic test datasets and compare the performance of the two approaches.