PSI - Issue 6

Chebyshev Igor et al. / Procedia Structural Integrity 6 (2017) 252–258 Chebyshev I.S./ Structural Integrity Procedia 00 (2017) 000–000

255

4

Fig. 3. a. P, S – waves are not registering on piezosensor in weak consolidated sample b. P-waves are registering on piezosensor in consolidated sample of rock. One of the problems is the sensor destruction when instant loading of strength samples with mostly fragile properties. This problem, as a rule, can be observed during rocks with hard-to-recover (HTR) reserves research. Usually the number of rock samples for geomechanical studies is limited and amounts to 10-20 samples. There is a need to restore the dynamic elastic moduli through the properties determined on all samples – density and porosity. The first method is linear regression (Fig.4.a,b). The second method uses machine learning, artificial neural networks (ANN) (Haykin, 1999), and one hot encoding (Brownlee, 2017) ( Fig.5.a,b). Lithological features were chosen as the input data. A total of 36 features are used in the description of rock samples. Some of them are the following: siltstone, limestone, dolomite, argillite, fine-grained, coarse-grained, medium-grained, heterogeneous, light gray, dense, fine-crystalline, medium-fine-grained, crystal, bio-clast, oil-saturated, smell, sandy, micaceous, silty, interlayer, weakly cemented, siderite, clayman, massive, coalhead, oolite, wackestone, packstone, detritus, carbonate. There were 1083 samples with density, porosity and lithology used for predicting acoustic properties.

2 − 2 V S 2 − V

2 )

V P

υ dyn =

2( V P

2 )

(1)

S

2 (3 V

P 2 − 4 V S

2 )

ρ V S

E dyn =

V P

2 − V

2

(2)

S

a.

b.

R 2 =0.91

R 2 =0.89

Fig. 4. a. P-, b. S – wave velocities predicted with neural networks using one hot encoding.