Repository files navigation

Link 1:

https://ieeexplore.ieee.org/document/8704258

Link 2:

https://www.researchgate.net/publication/332814555_Unsupervised_Clustering_of_Seismic_Signals_Using_Deep_Convolutional_Autoencoders

https://drive.google.com/file/d/16itT_IZpM8w8KyFN8eL8iEfYX66Hk6Xb/view?usp=sharing

Reference:

Mousavi, S. M., W. Zhu, W. Ellsworth, G. Beroza (2019).                        Unsupervised Clustering of Seismic Signals Using Deep Convolutional Autoencoders, IEEE Geoscience and Remote Sensing Letters, 1 - 5, doi:10.1109/LGRS.2019.2909218.

BibTeX:

@article{mousavi2019unsupervised, title={Unsupervised Clustering of Seismic Signals Using Deep Convolutional Autoencoders}, author={Mousavi, S Mostafa and Zhu, Weiqiang and Ellsworth, William and Beroza, Gregory}, journal={IEEE Geoscience and Remote Sensing Letters}, year={2019}, publisher={IEEE}}

In this paper, we use deep neural networks for unsupervisedclustering of seismic data.We perform the clustering in afeature space that is simultaneously optimized with the clusteringassignment, resulting in learned feature representations thatare effective for a specific clustering task. To demonstrate theapplication of this method in seismic signal processing, we designtwo different neural networks consisting primarily of full convolutionaland pooling layers and apply them to: (1) discriminatewaveforms recorded at different hypocentral distances and (2)discriminate waveforms with different first-motion polarities. Ourmethod results in precisions that are comparable to those recentlyachieved by supervised methods, but without the need for labeleddata, manual feature engineering, and large training sets. Theapplications we present here can be used in standard singlesiteearthquake early warning systems to reduce the false alertson an individual station level. However, the presented techniqueis general and suitable for a variety of applications includingquality control of the labeling and classification results of other
supervised methods.

Sampel data. a) and b) are two examples of the seismograms with different polarity of first motion.c) and d) are examples of local and teleseismic waveforms respectively while e) and f) are the associated Short-Time Fourier transforms.

The architecture of fully convolutional autoencoder used in our study.

Clustering results.

Visualization of embeded features.