A neural network based approach to classify VLF signals as rock rupture precursors

Electromagnetic signals have been increasingly investigated in the last decade in association to natural earthquakes and laboratory rock fractures. Studies on this type of signals are hampered by  the lack of continuous recordings and, when data are available, the sampling rates (> kHz) is such to require an efficient and systematic processing of large data sets. Despite this limitation, previous studies performed under controlled conditions in the laboratory seem to suggest that electromagnetic signals exhibit characteristic patterns, called OIS - Ordered Impulsive Sequences, on a specific frequency band (the very low frequency, VLF) that correlate uniquely with the paroxistic rupture of rocks specimens under uniaxial tests. Importantly, these characteristic patterns were also detected in the atmosphere in association to moderate magnitude earthquakes occurring within a few days (up to 5) from their detection. The similarity of laboratory and atmospheric VLF offers a unique opportunity to study the relation between VLF and rock deformation on at least two different scales and to enlarge the dataset by combining laboratory and atmospheric data. Here, we deployed tools for a systematic monitoring of electromagnetic signals in the atmosphere and we show that the enlarged VLF dataset, which comprises both laboratory and natural electromagnetic signals, can be successfully processed using a neural network approach. Our neural network architecture was designed to deal with time series and is structured using a recurrent neural networks (RNN) and a Long Short Term memory (LSTM) as a state variable. After a careful data collection, signal sequences were classified as rock rupture precursors (“RUPTURE”) and some of them, including those composing the background noise, as no rupture precursors (“QUIET”). A deep BI-LSTM neural network with 1000 hidden units has been trained in order to fit the known classification and to implicitly acquire the most important features and cut offs to split the potential events to not events. Our main results are 1. laboratory and atmospheric OIS signals are similar and scalable; 2. the similarity is such that it can be successfully used to train a neural network for signal detection in the atmosphere; 3. the neural network is capable of detecting OIS from the huge data set which is made of all the atmospheric background; 4. the extracted signals are those which were typically recorded in association to earthquakes in a temporal window of a few days. The above results show that LSTM neural networks are effective “automatic detectors” for characteristic spectral patterns revealed in the VLF both in the laboratory and in atmospheric signals recorded in association with transient natural events involving fracturing of rock volumes (e.g. earthquakes). The above results suggest that the electromagnetic radiation in the very low frequency band is a promising and valuable signal to probe the deformation of the seismically active Earth crust.