Faculty
Science
Supervisor Name
Cristián Bravo Roman
Keywords
deep neural networks, time series analysis, volcanic seismicity
Description
Monitoring volcanic events as they occur is a task that, to this day, requires significant human capital. The current process requires geologists to monitor seismographs around the clock, making it extremely labour-intensive and inefficient. The ability to automatically classify volcanic events as they happen in real-time would allow for quicker responses to these events by the surrounding communities. Timely knowledge of the type of event that is occurring can allow these surrounding communities to prepare or evacuate sooner depending on the magnitude of the event. Up until recently, not much research has been conducted regarding the potential for machine learning (ML) models to supplement or substitute human monitoring of volcanoes. Recent initiatives in this field have demonstrated that it is possible to classify volcanic events using ML techniques. Additionally, recent research in general signal processing has shown that the novel technique of multi-head self-attention (MHSA), used in natural language processing (NLP), is also useful in signal analysis. In this report, we seek to apply MHSA to create a deep neural network (DNN) that can automatically classify volcanic events. Our proposed model architecture provides minor improvements over existing approaches on pre-processed data. When considering raw signals coming directly from monitoring stations, our model outperforms existing approaches by a great margin.
Acknowledgements
Thank you Dr. Cristián Bravo Roman, Dr. Cindy Mora Stock, and the Western USRI program for their support.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Document Type
Poster
Included in
Data Science Commons, Geophysics and Seismology Commons, Longitudinal Data Analysis and Time Series Commons, Statistical Models Commons, Volcanology Commons
A Transformer-Based Classification System for Volcanic Seismic Signals
Monitoring volcanic events as they occur is a task that, to this day, requires significant human capital. The current process requires geologists to monitor seismographs around the clock, making it extremely labour-intensive and inefficient. The ability to automatically classify volcanic events as they happen in real-time would allow for quicker responses to these events by the surrounding communities. Timely knowledge of the type of event that is occurring can allow these surrounding communities to prepare or evacuate sooner depending on the magnitude of the event. Up until recently, not much research has been conducted regarding the potential for machine learning (ML) models to supplement or substitute human monitoring of volcanoes. Recent initiatives in this field have demonstrated that it is possible to classify volcanic events using ML techniques. Additionally, recent research in general signal processing has shown that the novel technique of multi-head self-attention (MHSA), used in natural language processing (NLP), is also useful in signal analysis. In this report, we seek to apply MHSA to create a deep neural network (DNN) that can automatically classify volcanic events. Our proposed model architecture provides minor improvements over existing approaches on pre-processed data. When considering raw signals coming directly from monitoring stations, our model outperforms existing approaches by a great margin.