Nuclear Blast Discrimination using a Convolutional Neural Network

Abstract:

Recent developments in machine learning techniques opens doors for scientists to take advantage of historical data to create algorithms that augment or replace repeating tasks with information that share similarities. Due to ongoing nuclear threats, it is important to develop robust methods to automatically detect, classify and locate sources, preferably from diverse and widely distributed global sensors. Using a rich list of globally-reported declassified nuclear events that were generated from the SIPRI catalog, the latest DOE and USGS reports for the North Korea nuclear tests, we created a catalog of 1,463 underground nuclear explosions and retrieved waveforms from all available International Federation of Digital Seismograph Network (FDSN) stations. Our dataset is supplemented by the Lawrence Livermore National Laboratory nuclear explosion dataset, the former Soviet Union’s digitized analog seismic records, as well as the Japan National Research Institute for Earth Science and Disaster Resilience (NIED) Hi-Net data for the 6 North Korea nuclear tests. Since nuclear explosions frequently do not have a discernable S-wave arrival, we use only vertical channel P-wave arrival picks for the labeled dataset. A seismic and nuclear event classifier was built focusing initially on seismograms from earthquakes and underground nuclear explosions recorded at regional and teleseismic stations. We built a 10-layer convolutional neural network (CNN), and trained for three classes: earthquake P-wave, nuclear P-wave, and noise. In order to address the effect of low signal-to-noise ratio on CNN performance, an energy filter was developed to calculate the ratio of mean energy of a P-wave to the noise and an energy ratio threshold was applied during pre-processing, which both increased the quality of the seismic traces but also significantly limits the size of the training dataset. Our initial results are promising, with the accuracy of the validation set rising rapidly to accuracy of over 90% after a few epochs, even with limited training data. We anticipate that this classifier system will be useful beyond automatic nuclear detections, with application potentially for other seismic signals associated with geological hazards including pre-eruptive volcanic tremor, tsunami earthquakes, ice-quakes or landslides.