A novel type of neural networks for feature engineering of geological data:
Case studies of coal and gas hydrate-bearing sediments
-
Abstract
The nature of the measured data varies among different disciplines of geosciences. In rock engineering, features of
data play a leading role in determining the feasible methods of its proper manipulation. The present study focuses
on resolving one of the major deficiencies of conventional neural networks (NNs) in dealing with rock engineering
data. Herein, since the samples are obtained from hundreds of meters below the surface with the utmost difficulty,
the number of samples is always limited. Meanwhile, the experimental analysis of these samples may result in
many repetitive values and 0s. However, conventional neural networks are incapable of making robust models in
the presence of such data. On the other hand, these networks strongly depend on the initial weights and bias
values for making reliable predictions. With this in mind, the current research introduces a novel kind of neural
network processing framework for the geological that does not suffer from the limitations of the conventional
NNs. The introduced single-data-based feature engineering network extracts all the information wrapped in every
single data point without being affected by the other points. This method, being completely different from the
conventional NNs, re-arranges all the basic elements of the neuron model into a new structure. Therefore, its
mathematical calculations were performed from the very beginning. Moreover, the corresponding programming
codes were developed in MATLAB and Python since they could not be found in any common programming
software at the time being. This new kind of network was first evaluated through computer-based simulations of
rock cracks in the 3DEC environment. After the model’s reliability was confirmed, it was adopted in two case
studies for estimating respectively tensile strength and shear strength of real rock samples. These samples were
coal core samples from the Southern Qinshui Basin of China, and gas hydrate-bearing sediment (GHBS) samples
from the Nankai Trough of Japan. The coal samples used in the experiments underwent nuclear magnetic resonance
(NMR) measurements, and Scanning Electron Microscopy (SEM) imaging to investigate their original micro
and macro fractures. Once done with these experiments, measurement of the rock mechanical properties,
including tensile strength, was performed using a rock mechanical test system. However, the shear strength of
GHBS samples was acquired through triaxial and direct shear tests. According to the obtained result, the new
network structure outperformed the conventional neural networks in both cases of simulation-based and case
study estimations of the tensile and shear strength. Even though the proposed approach of the current study
originally aimed at resolving the issue of having a limited dataset, its unique properties would also be applied to
larger datasets from other subsurface measurements.
-
-