Unsupervised deep learning on seismic data to detect volcanic unrest

The significant efforts of the last years in new monitoring techniques and networks have led to large datasets and improved our capabilities to measure volcano conditions.  Thus nowadays the challenge is to retrieve information from this huge amount of data to significantly improve our capability to automatically recognize signs of potentially hazardous unrest.
Unrest detection from unlabeled data is a particularly challenging task, since the lack of annotations on the temporal localization of these phenomena makes it impossible to train a machine learning model in a supervised way. The proposed approach, therefore, aims at learning unsupervised low-dimensional representations of the input signal during normal volcanic activity by training a variational autoencoder (VAE) to compress, reconstruct and synthesize input signals. Thanks to the internal structure of the proposed VAE architecture, with 1-dimensional convolutional layers with residual blocks and attention mechanism, the representation learned by the model can be employed to detect deviations from normal volcanic activity. In our experiments, we test and evaluate two techniques for unrest detection: a generative approach, with a bank of synthetic signals used to assess the degree of correspondence between normal activity and an input signal; and a discriminative approach, employing unsupervised clustering in the VAE representation space to identify prototypes of normal activity for comparison with an input signal.