D-Team

Our Team

How can we effectively manage massive distributed acoustic sensing (DAS) datasets using artificial intelligence (AI)?

DAS can easily generate terabytes of data per day of recording; therefore, automated tools for anomaly (i.e., seismic event) detection are essential. The main motivation of this research is to develop an autoencoder-based deep learning model that helps detect seismic events on multichannel DAS recordings. By classifying data into anomaly-free and anomalous categories, the algorithm can significantly reduce data storage needs by recording only when anomalies are detected, facilitating one’s ability to manage large DAS datasets efficiently.

Anomaly detection is one of the leading applications in unsupervised learning that aims to identify statistical outliers and is useful in seismic analysis because data containing events are much rarer than data containing only background noise. Autoencoders are a type of convolutional neural network that are often used for unsupervised learning and are particularly useful for anomaly detection. Autoencoders compress input data into a lower-dimensional latent space and reconstruct it using this reduced representation. As shown in the figure, for anomaly detection in images, the autoencoder model is trained solely on anomaly-free data to capture its essential features. During testing, the autoencoder reconstructs new images based on the learned model weights. Normal images are reconstructed with minor errors, while anomalous images, featuring patterns absent from the training data, show higher reconstruction errors, enabling anomaly detection. The developed algorithm has potential applications in various monitoring projects, including induced seismicity (to detect seismic events), environmental seismology (to identify environment-induced variations), and ambient noise interferometry (to exclude anomalous data in advance), among others. Check out this paper for more information.

Creating fast algorithms for seismic data analysis