Abstract:
The paper proposes a method for improvement of the quality of geophysical data preparation on the example of geoacoustic observations to train neural networks when solving the problem of identification of pre- and post-seismic anomalies. The method is based on the transformation of geoacoustic emission signal associated with deformation processes in near-surface rocks into three-dimensional images. A series of such images contains the information on signal characteristics dynamics. Thee-dimensional images are the matrices consisting of the the distribution vectors of selected characteristics (spectral, structural, statistical and so on). The structure, data tensor, is formed from a series of such images. It is supplied to the neural network input. Due to external factors impact (weather, industrial), a recorded geoacoustic signal is distorted. Thus, it is necessary to clean the initial data. In order to do this, we suggest using a neural network which clusters the prepared images and removes outliers in the obtained clusters. A new tensor is formed from the remaining images. It undergoes the cleaning procedure again. This process continues until no outliers are observed in the output data as the result of clustering. When the cleaning is over, the second neural network will be trained to identify common features and differences, as well as hidden patterns in the geoacoustic pulse flux. Application of the developed method for tensor cleaning, based on artificial intelligence technologies, allows us to improve significantly the quality of data preparation. The obtained results will be useful for the investigations in the fields of identification and classification of pre- and postseismic anomalies in geoacoutstic emission signals associated with deformation processes in near-surface rocks in a seismically active region.
Fulltext:
PDF file (1657 kB)