Synth_Seism / README.md
mexalon's picture
Update README.md
1d5bf8f
metadata
task_categories:
  - feature-extraction
size_categories:
  - 100K<n<1M
  • The objective of this work is to construct a method for analyzing waveforms of signals obtained during microseismic monitoring using a neural network in order to localize the coordinates of the sources of seismic events, and their differentiation.
  • Microseismic monitoring is one of the existing methods of analyzing the condition of the studied geophysical object: mineral deposits, large-scale industrial facilities, etc. It includes a system of sensors that detect weak seismic or acoustic signals, a data collection system and algorithms for their processing. The main task of monitoring is to determine the characteristics of a microseismic event: the time of the first entry, magnitude, and its location in space.
  • This dataset contains synthetic waveforms to form a training and validation samples. The advantage ofusing synthetic data is that for it, all the necessary parameters of each seismic event are known in advance (time of entry, coordinates of the source, magnitude, parameters of the source mechanism, velocity model of the medium). This makes it possible to create and train models based on data generated taking into account the features characteristic of a given monitoring area, while the resulting models may have a greater generalizing ability than those trained on real waveforms. In addition, this approach, unlike using banks of real waveforms to train the model, eliminates the possibility of distortion of the results associated with the use of manual data markup. The main disadvantage of using synthetic data for training models is the need to adapt the resulting models to real data. The synthetic waveforms used in this work were created using Pyrocko, an open–source set of libraries for seismological modeling [Heimann et al., 2018]. The propagation of seismic waves was modeled for an elastically viscous layered medium. The velocity model of the medium was taken from [Málek, Horálek, Janský, 2005]. The choice was determined by the freely available pre-calculated bank of Green's functions necessary to obtain waveforms. The sources of seismic signals were modeled by a double pair of forces with a random distribution of displacement directions (strike, deep, rake) and magnitudes uniformly distributed within the specified boundaries (0-2). The epicenters and depths of the sources were randomly set inside an area with a radius of 1.5 km and a depth of 1000 meters. Waveforms (displacement) were obtained for five stations (four symmetrically located at a distance of 500 meters from the origin, and one in the center) for three channels (two horizontal N, E and vertical Z) with a sampling frequency of 100 Hz, the length of each recording is 4 seconds. A priori moments of arrival of p and s waves were obtained for each waveform. As a result of the simulation, training and test samples were formed from 106 and 103 events, respectively (15 waveforms in each).