Faculty
Joint Actuarial and Statistical Sciences/Computer Science and Physics and Astronomy
Supervisor Name
Cristian Roman Bravo and Cindy Mora Stock
Keywords
audio signals, deep learning, classification, time series, computer hearing
Description
Volcanic seismic signals are a key element in volcano monitoring to assess the state of unrest and a possible eruption style and timing. Different sources such as brittle fracture (volcano-tectonic - VT) or fluid movement (long period - LP) generate signals with distinct characteristics in frequency content and shape, but site effects such as attenuation or background noise make their determination difficult to the untrained eye. In cases of unrest or an eminent eruption, the amount of data would require a fast and reliable source of pre-classification to classify and catalogue to aid in the job usually done by a human.
To model the problem, we will develop a custom-made Transformer model. Transformers are state-of-the-art deep learning methodologies that work with sequence-based data such as audio, text or, in this case, volcanic signals. The power of transformers lies in their ability to identify complex, disconnected patterns and then use them to identify phenomena in a very effective manner. We will be building the model architecture in TensorFlow and will be running them through SHARCNET.
Unfiltered continuous data from seismic stations in Villarrica volcano will be used as train dataset and catalogued from at least these two types of events (VT and LP). The model will be then tested with a different set of stations to assess changes in the signal due to attenuation at the site. This will allow to discriminate the same event in different stations.
Acknowledgements
The work is still in progress and will require further work to get it to a place suitable for publishing in an academic journal
Creative Commons License
This work is licensed under a Creative Commons Attribution-No Derivative Works 4.0 License.
Document Type
Poster
Included in
A Transformer-Based Classification System for Volcanic Seismic Signals
Volcanic seismic signals are a key element in volcano monitoring to assess the state of unrest and a possible eruption style and timing. Different sources such as brittle fracture (volcano-tectonic - VT) or fluid movement (long period - LP) generate signals with distinct characteristics in frequency content and shape, but site effects such as attenuation or background noise make their determination difficult to the untrained eye. In cases of unrest or an eminent eruption, the amount of data would require a fast and reliable source of pre-classification to classify and catalogue to aid in the job usually done by a human.
To model the problem, we will develop a custom-made Transformer model. Transformers are state-of-the-art deep learning methodologies that work with sequence-based data such as audio, text or, in this case, volcanic signals. The power of transformers lies in their ability to identify complex, disconnected patterns and then use them to identify phenomena in a very effective manner. We will be building the model architecture in TensorFlow and will be running them through SHARCNET.
Unfiltered continuous data from seismic stations in Villarrica volcano will be used as train dataset and catalogued from at least these two types of events (VT and LP). The model will be then tested with a different set of stations to assess changes in the signal due to attenuation at the site. This will allow to discriminate the same event in different stations.