Page 283 - Machine Learning for Subsurface Characterization

Classification of sonic wave Chapter 9 245

geophysics, the investigation scale is from hundred meters to kilometers; in
civil engineering, the scale is from centimeters to tens of meters; in the oil
and gas industry, the investigation scale is from millimeters to meters.
One widely used NDT method is ultrasonic testing, which is a centimeter-
scale measurement to characterize the mechanical discontinuity in material.
Ultrasonic testing relies on the propagation of compressional (P-wave) and
transverse (S-wave) sonic wave through the material. Ultrasonic testing can
be based on reflection, transmission, or refraction of waves as they propagate
from the source to the sensor. Ultrasonic pulse velocity (UPV) measurement
utilizes wave transmission, whereas ultrasonic pulse echo (UPE) measurement
uses wave reflection. Ultrasonic testing method was used to investigate the
development of discontinuities in rock samples under uniaxial compressive test
by Martı ´nez-Martı ´nez et al. [2] and Ramos et al. [3]. They found that the pro-
portion of the high-frequency to low-frequency component decreases with the
increase in the intensity of discontinuities. Another NDT method is acoustic
emission (AE) testing that can be used to characterize the general mode and
3D location of cracks in material, but it may not able to reconstruct the 3D
geometry. CT scanning can be used to reconstruct the 3D geometry of the
discontinuity, but it is more expensive than AE testing. Unlike AE testing,
it is harder to deploy CT scanning to acquire real-time fracture signatures.
Watanabe et al. [4] applied CT scanning to build a 3D numerical model of rock
fractures. Cai et al. [5] used CT scanning to reconstruct the 3D fracture network
in coal samples during cyclic loading. CT scanning and AE testing are often
used in tandem to quantify the fracture geometry in materials during geomecha-
nical experiments, such as laboratory fracturing or triaxial experiments. Unlike
CT, AE, and ultrasonic methods, resistivity and electromagnetic methods have
not shown much potential in the characterization of discontinuities. In our
study, we intend to develop data-driven classification methods to facilitate
static characterization of discontinuities at centimeter scale by processing com-
pressional wavefront travel times recorded at multiple sensors/receivers when
emitted by a single source/transmitter.

1.3 Machine learning for characterization of discontinuities

Most popular supervised machine learning algorithms used for the characteri-
zation of discontinuities are artificial neural network (ANN), random forest,
support vector machine (SVM), and convolutional neural network (CNN). Zhou
et al. [6] applied ANN and SVM to classify rock fracture and blast events based
on AE signals from the rock fracturing experiment. Liu et al. [7] used ANN to
predict rock types from AE measurements because different rocks have differ-
ent failure modes and different failure modes generate different types of AE
signals. Farhidzadeh et al. [8] used SVM to classify the AE signals as being
emitted from tensile fracture or shear fracture. Wang et al. used machine learn-
ing algorithms on laboratory data describing the stress filed, like stress ratio and