Automatic classification with an autoencoder of seismic signals on a distributed acoustic sensing cable

This study probes the association between fluid injection in enhanced geothermal systems and certain kinds of seismicity that may result from hydraulic fracturing occurring at depth using unsupervised machine learning. In April and May 2019, a distributed acoustic sensing borehole array at the Frontier Observatory for Research in Geothermal Energy site near Milford, Utah recorded seismic data during hydraulic injection stimulation of a nearby well. Using an autoencoder, a type of deep neural network, we reduce the dimensionality of spectrograms of the detected signals to a lower-dimensional latent feature space with just nine dimensions. Next, Gaussian mixture model clustering is performed on this latent feature space, assigning each detected signal to one of 7 classes. For each signal class, we examine spatiotemporal distributions of the clustered results and find that total detections exhibit a bimodal distribution with respect to channel depth. The shallow mode occurs between 250 and 500 m, and the deep mode is centered around 750 m. In the temporal distribution, clustering results show the two best-clustered signal classes exhibit weak or no correlation with injection -related activities. More generally, we demonstrate the ability to discern not just when and where signals are detected, but also what kind, thus enabling rapid and targeted data exploration and providing constraints on source mechanisms.

Automated summary

This study investigates the relationship between fluid injection in enhanced geothermal systems and induced seismicity from hydraulic fracturing using unsupervised machine learning. Spectrograms of the detected signals are dimensionally reduced to a lower-dimensional latent feature space with a deep neural network called autoencoder. Gaussian mixture model clustering is then performed on this feature space to assign each signal to one of 7 classes. The results show bimodal spatiotemporal distributions, with a shallow mode occurring between 250 and 500 m and a deep mode centered around 750 m. The correlation between the clustered signal classes and injection-related activities is weak or non-existent. This study demonstrates the capability to analyze not only when and where signals are detected, but also their types, facilitating rapid and targeted data exploration and providing insights into source mechanisms.