Explosion Monitoring with Machine Learning: A LSTM Approach to Seismic Event Discrimination
Abstract
The streams of seismic data that analysts look at to discriminate natural from man- made events will soon grow from gigabytes of data per day to exponentially larger rates. This is an interesting problem as the requirement for real-time answers to questions of non-proliferation will remain the same, and the analyst pool cannot grow as fast as the data volume and velocity will. Machine learning is a tool that can solve the problem of seismic explosion monitoring at scale. Using machine learning, and Long Short-term Memory (LSTM) models in particular, analysts can become more efficient by focusing their attention on signals of interest. From a global dataset of earthquake and explosion events, a model was trained to recognize the different classes of events, given their spectrograms. Optimal recurrent node count and training iterations were found, and cross validation was performed to evaluate model performance. A 10-fold mean accuracy of 96.92% was achieved on a balanced dataset of 30,002 instances. Given that the model is 446.52 MB it can be used to simultaneously characterize all incoming signals by researchers looking at events in isolation on desktop machines, as well as at scale on all of the nodes of a real-time streaming platform. LLNL-ABS-735911
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2017
- Bibcode:
- 2017AGUFM.S43A0834M
- Keywords:
-
- 1914 Data mining;
- INFORMATICS;
- 1932 High-performance computing;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS;
- 7290 Computational seismology;
- SEISMOLOGY