1 s2.0 S136516092300148X Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

International Journal of Rock Mechanics & Mining Sciences 170 (2023) 105474

Contents lists available at ScienceDirect

International Journal of Rock Mechanics and Mining Sciences


journal homepage: www.elsevier.com/locate/ijrmms

Numerical back analysis method of three-dimensional in situ stress fields


considering complex surface topography and variable collinearity
Huaisheng Xu a, b, Shaojun Li a, *, Dingping Xu a, **, Xiang Huang a, b, Minzong Zheng a,
Jianhua He c, Kai Zhao d
a
State Key Laboratory of Geomechanics and Geotechnical Engineering, Institute of Rock and Soil Mechanics, Chinese Academy of Sciences, Wuhan, Hubei, 430071,
China
b
University of Chinese Academy of Sciences, Beijing, 100049, China
c
PowerChina Chengdu Engineering Corporation Limited, Chengdu, 610072, China
d
Huadian Jinshajiang Upstream Hydropower Development Corporation Limited Yebatan Branch, Ganzi Tibetan Autonomous Prefecture, Sichuan, 627150, China

A R T I C L E I N F O A B S T R A C T

Keywords: Numerical back analysis based on incomplete in situ stress measurements is widely used to estimate the three-
In situ stress field dimensional (3D) in situ stress fields of large deep underground caverns. However, the variable collinearity
Numerical back analysis caused by complex geological environments has rarely been considered. This paper proposes a numerical back
Complex geological environment
analysis method for 3D in situ stress fields based on stepwise regression (BSSR). In BSSR, the collinearity between
Surface topography
Variable collinearity
independent variables is associated with the surface topography, and insignificant variables caused by variable
Stepwise regression collinearity are eliminated through stepwise regression. BSSR was applied to the underground powerhouse of a
hydropower station for validation, and the results showed that it can reliably estimate the 3D in situ stress field
despite the complex geological environments. Multidimensional mathematical models of in situ stress fields were
established with the improved predictive ability and clear physical meaning, which can quantify the contribu­
tions of six geological actions and three stress sources. The findings of this study can help with understanding the
formation mechanism of in situ stress fields for large, deep underground caverns in complex geological envi­
ronments and provide a useful reference for optimizing excavation schemes and support designs.

1. Introduction approach.20–23 In practice, however, it is difficult to obtain sufficient


and reliably representative measurements for large, deep underground
In situ stress is an important parameter for a wide range of endeavors caverns because of the complex geological conditions, high test costs,
in rock mechanics, including rock mass engineering design, hydraulic and poor measurement conditions.9,10,24,25 To compensate for the
fracturing analysis, rock mass permeability, and earthquake risk asses­ limited in situ stress measurements, numerical simulations are routinely
sment.1–4 The magnitude and direction of the in situ stress directly employed to extrapolate the complete three-dimensional (3D) in situ
determine the apparent scale, spatial distribution, and degree of risk of stress field of a study area from incomplete measurements.26–29 Various
deformation and damage to the surrounding rock mass, which are methods have been proposed for estimating 3D in situ stress fields,
especially important for underground engineering.5–10 Therefore, before including multiple linear regression (MLR),9,10,30–34 neural networks
the excavation of large, deep underground caverns such as underground and genetic algorithms,35,36 displacement back analysis,37 and gray
powerhouses at hydropower stations, obtaining an accurate distribution system theory.38 MLR is widely used because it provides a unique co­
of the in situ stress field is critical for support design and safety efficient solution for independent variables. Research on using
assessment. MLR-based back analysis for estimating in situ stress fields has mostly
Because of the complexity and variability of in situ stress fields,11–19 focused on determining the weights of variables; however, such research
in situ stress measurement is still the most direct and effective has generally neglected the collinearity between independent

Abbreviations: 3D, three-dimensional; MLR, multiple linear regression; SR, stepwise regression; EDZ, excavation damage zone; RFD, rock failure degree.
* Corresponding author.
** Corresponding author.
E-mail addresses: sjli@whrsm.ac.cn (S. Li), dpxu@whrsm.ac.cn (D. Xu).

https://doi.org/10.1016/j.ijrmms.2023.105474
Received 13 June 2022; Received in revised form 21 February 2023; Accepted 20 June 2023
Available online 26 June 2023
1365-1609/© 2023 Elsevier Ltd. All rights reserved.
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

variables.9,10,33–38 Although a mathematical model can estimate an in 2.2. Collinearity between independent variables
situ stress field by considering all influencing factors as independent
variables, the variable collinearity can make some variables redundant. MLR is an estimation method based on the principle of linear elastic
This distorts the weights of the variables used for prediction and in­ superposition that can be applied to the back analysis of in situ stress
creases the calculation workload, which ultimately degrades the accu­ fields. The regression model is given by
racy of the analysis on the formation mechanism of in situ stress fields.
σ = UB + e
Some scholars have noticed this problem, but no studies have yet ⎡ ⎤ ⎡ ⎤⎡ ⎤
explored the reasons for the collinearity between independent variables σx U11 U21 U31 U41 U51 U61 B1
⎢ ⎥ ⎢ ⎥⎢ ⎥
in detail.30–32 Therefore, understanding and eliminating variable ⎢ σ y ⎥ ⎢ U12
⎢ ⎥ ⎢ U22 U32 U42 U52 U62 ⎥ ⎢ ⎥
⎥ ⎢ B2 ⎥
collinearity is a critical problem for the MLR-based back analysis of in ⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢ σz ⎥ ⎢ U13 U23 U33 U43 U53 U63 ⎥ ⎢ ⎥
⎥ ⎢ B3 ⎥ (1)
situ stress fields. ⎢ ⎥ ⎢
⎢ ⎥=⎢ ⎥⎢ ⎥ + e
⎢ τxy ⎥ ⎢ U14 U64 ⎥⎢
⎥ ⎥
This study focused on the collinearity between independent variables ⎢ ⎥ ⎢ U24 U34 U44 U54 ⎢ B4 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
and their association with surface topography in the numerical back ⎢ τyz ⎥ ⎢ U15 U25 U35 U45 U55 U65 ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦ ⎣ B5 ⎦
analysis of 3D in situ stress fields; therefore, a new numerical back
τxz U16 U26 U36 U46 U56 U66 B6
analysis method based on stepwise regression (BSSR) was proposed. The
underground powerhouse of Yebatan hydropower station was used for
where σ is the dependent variable matrix and represents the actual in
application and verification. Moreover, the advantages and physical
situ stress field. U is the independent variable matrix; The columns of U
significance of BSSR are discussed, and suggestions for the numerical
(i.e., U1j − U6j (j = 1 − 6)) are the stress field formed by the independent
back analysis of 3D in situ stress fields are provided.
variables Ux ,Uy ,Ug ,Uxy ,Uyz ,and Uxz , respectively, and the rows of U (i.e.,
Ui1 − Ui6 (i = 1 − 6)) are the stress components σ x , σy , σ z , τxy , τyz , and τxz ,
2. Collinearity between independent variables of a stress field
respectively. B is the coefficient matrix of the independent variables,
and e is the error matrix.
2.1. Sources of independent variables
According to MLR theory, an important assumption is that all inde­
pendent variables are unrelated to each other.42–44 If each independent
Generally, the stress state in the crust can be attributed to the su­
variable represents only one stress component in the stress field, the
perposition of gravitational, tectonic, and local sources.3,39 A tectonic
independent variable matrix U is given by
source refers to far-field stresses related to plate-driving forces, which
⎡ ⎤
are typically uniform at the lithospheric scale40 like gravity. The tectonic U11 0 0 0 0 0
and gravitational sources are the main factor for the formation of in situ ⎢
⎢ U22 0 0 0 0 ⎥⎥
stress fields. Various conditions such as the topography, anisotropic U=⎢

⎢ U33 0 0 0 ⎥⎥ (2)
strength, deformation, elastic properties, erosion, and manmade exca­ ⎢ U44 0 0 ⎥⎥
vation can result in local sources of stress,41 which are secondary factors
⎣ U55 0 ⎦
for the formation of in situ stress fields. For numerical back analysis, the symmetric U66
gravitational and tectonic sources can be divided into six types of However, because of Poisson’s effect, which is given in Eq. (3), the
geological actions, as shown in Fig. 1. Meanwhile, local sources are variables Ux ,Uy , and Ug can form multiple stress components at the same
considered by constructing a 3D geological model that includes the real time45:
geological information. A key issue is weighting the contributions of ⎧ ( )
these stress sources and geological actions to in situ stress fields. ⎨ σx = ν⋅σ y for Ux and Uy
( ) (3)
⎩ σx = σy = ν σz for Ug
1− ν

Fig. 1. Variables for the back analysis of in situ stress fields: (a) gravity (Ug), (b) compression in the x-direction (Ux), (c) compression in the y-direction (Uy), (d) shear
on the horizontal plane (Uxy), (e) shear on the yz-plane (Uyz), and (f) shear on the xz-plane (Uxz).

2
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

where ν is Poisson’s ratio. The collinearity problem exists between the correlation coefficient between any two columns of the matrix U (i.e.,
independent variables Ux , Uy , and Ug . And the independent variable independent variables) can be calculated as follows:
⃒ ⃒
matrix U is given by ⃒ ∑n ⃒
⃒ (X − X)(Y − Y) ⃒
⃒ ⃒
⎡ ⎤ (6)
i=1 i i
ν r = ⃒√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ √ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ ̅ ⃒
⃒ 2⃒
⎢ U11 νU22 1 − νU33 0 0 0 ⎥ n 2 n
⃒ i=1 (Xi − X) i=1 (Yi − Y) ⃒
⎢ ⎥
⎢ ν ⎥
⎢ νU11 U22 U 0 0 0 ⎥ Fig. 3 shows the correlation coefficient matrix between independent
⎢ 1− ν
33 ⎥
⎢ ⎥
⎢ ⎥ variables for each geological model. To demonstrate the influence of the
U=⎢ ⎢ 0 0 U 0 0 0 ⎥ (4)
surface topography on variable collinearity, a blank control group was
33

⎢ ⎥
⎢ 0

0 0 U44 0 0 ⎥
⎥ added, for which its independent variable matrix U is the diagonal

⎢ 0 0 0 0 U 55 0

⎥ matrix in Eq. (2). The figure shows that the collinearity between inde­
⎣ ⎦
pendent variables becomes more significant as the complexity of the
0 0 0 0 0 U66
surface topography increases. The maximum values of the correlation
The collinearity problem caused by Poisson’s effect only considers an coefficients among variables of the blank control group, Model 1, Model
ideal geological environment (Fig. 2a). In practical engineering, the 2, and Model 3 were 0.2 (no correlation), 0.29 (weak correlation), 0.54
influence of local sources on the in situ stress field cannot be ignored (moderate correlation), and 0.79 (strong correlation), respectively. For
⃒ ⃒
especially for complex geological environments such as deep valleys, example, Model 3 has ⃒rUy ,Uyz ⃒ = 0.79, which indicates a strong corre­
where the influence of the surface topography is most significant.46–48 lation between variables Uy and Uyz ; therefore, Uyz may be redundant.
To identify the influence of complex surface topography on variable
collinearity, the surface topographies of existing cases9,10,30–38 were 3. Proposed method
classified into three generalized geological models: (a) simple, (b)
relatively complex, and (c) complex (Fig. 2). Then, each single geolog­ 3.1. Stepwise regression
ical action (i.e., independent variable) shown in Fig. 1 was applied, and
the six stress components at the model centroid were taken to obtain the Stepwise regression (SR) is an independent variable selection
independent variable matrix U of each geological model: method for linear regression models. The basic concept is to consider
⎡ ⎤ whether the existing variables of the regression model can be eliminated
− 21.34 − 5.34 − 0.83 0 0 0
⎢ − 5.34 − 21.34 − 0.83 when each variable is introduced, and the optimal regression model is
0 0 0 ⎥

⎢ 0
⎥ obtained when no variables can be introduced.42–44
0 − 2.50 0 0 0 ⎥
U1 = ⎢⎢ 0
⎥ (5a) In SR, the significance of variables is tested using two methods: F-test
⎢ 0 0 − 7.38 0 0 ⎥ ⎥
⎣ 0 0 0 0 − 7.38 0 ⎦ and t-test, respectively. The F-test verifies whether the regression model
0 0 0 0 0 − 7.38 is significant; its null (H0) and alternative hypotheses (H1) are given as
follows:
⎡ ⎤
− 21.08 − 4.30 − 0.45 0 − 0.19 − 0.13
H0 : B1 = B2 = ... = Bi = ... = Bm = 0
⎢ − 4.38
⎢ − 17.52 − 0.46 0 − 3.15 0 ⎥ ⎥ (7)
⎢ 0 H1 : at least one Bi ∕
=0
0.33 − 1.33 0 − 2.49 0 ⎥
U2 = ⎢ ⎥ (5b)
⎢ 0
⎢ 0 0 − 6.93 0 − 0.15 ⎥
⎥ For a certain confidence level α, H0 is rejected and the regression
⎣ − 0.79 0 ⎦
− 3.18 − 0.24 0 − 4.82 model is significant when the F-value satisfies the following condition:
0 0 0 − 0.29 0 − 5.46
ESS/m
⎡ ⎤ F= > Fα (m, n − m − 1) (8)
− 20.29 − 3.74 − 0.34 1.20 − 0.34 − 1.82 RSS/(n − m − 1)
⎢ − 3.64 − 14.93 − 0.38 0.60 − 5.18 − 0.15 ⎥
⎢ ⎥
⎢ − 0.11 − 0.41 − 0.96 0 − 2.34 − 0.87 ⎥ where ESS is the explained sum of squares, RSS is the residual sum of
U3 = ⎢ ⎥ (5c)
⎢ 0.97
⎢ 0.84 0.1 − 6.46 − 0.12 − 0.51 ⎥
⎥ squares, m is the number of independent variables, and n is the number
⎣ − 0.89 − 3.92 − 0.22 − 0.49 − 4.39 0.10 ⎦ of samples.
− 1.35 0 0 − 0.75 − 0.10 − 4.37 Meanwhile, the t-test verifies whether a variable is significant; its
null (H0) and alternative hypotheses (H1) are given as follows:
where U1 corresponds to Eq. (4). Eq. (5) shows that an increasingly
complex surface topography also increases the complexity of the stress H0 : B1 = 0, H1 : B1 ∕
=0 (9)
field formed by the independent variables; in other words, the number of For a certain confidence level α, H0 is rejected and the variable is
non-zero values in the matrix U increases significantly. The Pearson significant when the t-value satisfies the following condition:

Fig. 2. Generalized geological models of the surface topography: (a) Model 1, (b) Model 2, and (c) Model 3. Model 1 has a horizontal surface and is simple. In Model
2, the surface changes only in inclination, and it is relatively complex. In Model 3, the surface changes in both trend and inclination, and it is complex. Note: An
elastic model was adopted for numerical simulations with an elastic modulus of 20 GPa, Poisson’s ratio of 0.25, and density of 2500 kg/m3.

3
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

Fig. 3. Pearson correlation coefficients between independent variables of different generalized geological models of the surface topography: (a) blank control group,
(b) Model 1, (c) Model 2, and (d) Model 3.

x
t= √̅̅̅ > tα2 (n − 1) (10)
s/ n

where x is the sample mean value and s is the sample standard deviation.
After introducing an independent variable in SR, the first step is to
conduct the F-test on the regression equation to ensure that the estab­
lished mathematical model is significant. However, passing the F-test
does not guarantee that all variables in the regression model are sig­
nificant; therefore, the t-test needs to be implemented to ensure that
each variable retained in the model is significant. As a result, in the
stepwise regression method, when each variable is introduced and
eliminated, the F-test and t-test should be repeated. The stepwise
regression method is illustrated in Fig. 4.

3.2. BSSR

The analysis in Section 2 indicates that variable collinearity cannot


be neglected when establishing a mathematical model of in situ stress
fields. Fig. 5 shows the proposed method BSSR for addressing this issue,
which has four main steps:

(1) 3D geological modeling: Establish a 3D geological model that


includes real information such as the surface topography, strati­
graphic classification, and faults, which is fundamental to the
back analysis of in situ stress fields.
(2) Correlation analysis and SR: Analyze the collinearity between
independent variables and contribution of independent variables
using the Pearson correlation coefficient. Then, obtain the
optimal mathematical model using SR.
(3) Modification of variable coefficients: Eliminate the constant in Fig. 4. The stepwise regression process.
the regression model, and modify the coefficients of significant
variables. Specifically, sensitivity analysis is performed on the
95% confidence interval of the regression coefficients, and the

4
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

Fig. 5. Main steps of the proposed BSSR.

optimal variable coefficients are obtained at the minimum RSS. deep valley is more than 2000 m. The underground powerhouse of the
The constant is eliminated from the regression model because its station is on the right bank of Jinsha River with a horizontal depth of
presence would cause problems such as distorted calculation re­ 270–480 m and vertical burial depth of 240–460 m.
sults and non-convergence. The Jinsha River fault zone is dominated by dextral strike-slip faults,
(4) Verification of the back analysis results: The results of the nu­ which have a displacement rate of approximately 3 mm/a. According to
merical back analysis are verified in terms of three aspects: the preliminary analysis, the maximum principal stress orientations are
magnitude of the stress components, orientation of the maximum mainly E–W and N–E. The local faults are well developed around the
principal stress, and locations of excavation damage zones (EDZs) underground powerhouse. There are three dominant fault groups each
in the underground cavern. comprising one of the second-order faults F2, F3, and F4 with many
third-order faults, as shown in Fig. 6. The F2 fault group has an orien­
BSSR is a general numerical back analysis method of 3D in situ stress tation of EW/S50–60◦ . The F3 fault group has an orientation of
fields; it analyzes the collinearity and significance of independent vari­ N65–75◦ W/SW38–45◦ . The F4 fault group has an orientation of
ables and then eliminates the nonsignificant ones. BSSR adopts more N50–60◦ E/NW70–85◦ .
mathematical than empirical methods. Theoretically, there are no re­
strictions on the application of BSSR in deep underground engineering.
If the target in situ stress field can be back-analyzed, the BSSR method 4.2. In situ stress measurements
can be used therein.
Twelve groups of in situ stress measurements were performed in the
4. Application for verification PD08 tunnel on the right bank of Yebatan hydropower station. The lo­
cations of the measurement points are shown in Fig. 6, and the mea­
4.1. Project overview surement data are given in Table 1. The maximum principal stress
around the underground powerhouse was 8.77–37.57 MPa. Meanwhile,
Yebatan hydropower station is the seventh of 13 cascade hydro­ the maximum principal stress orientation was N82◦ E–N54◦ W, with a
power stations in the upper reaches of the Jinsha River, and it has a total dominant orientation of N80.4◦ W. Additionally, the maximum principal
installed capacity of 2240 MW. The elevation is generally high in the stress plunge ranged from − 3◦ to 47◦ , with a dominant plunge of 27.3◦ ,
northwest and low in the southeast, and the elevation difference in the and the maximum principal stress generally tended to follow the valley.
The azimuth of the powerhouse axis was NW50◦ , and the included angle

5
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

Fig. 6. Geological plan for the underground powerhouse of Yebatan hydropower station at an elevation of 2742 m.

Table 1
In situ stress measurements around the underground powerhouse (Upward dip angle has a positive value).
Point NO. Depth(m) σ1 σ2 σ3
Magnitude(MPa) Azimuth (◦ ) Plunge (◦ ) Magnitude(MPa) Azimuth(◦ ) Plunge(◦ ) Magnitude(MPa) Azimuth(◦ ) Plunge(◦ )

08–1 183 16.51 288.2 30.0 12.36 25.3 12.1 6.32 134.7 57.2
08–2 260 37.57 274.9 13.8 17.98 174.8 35.7 15.03 22.5 50.9
08–3 163 18.51 305.7 12.0 9.46 40.8 22.8 4.34 9.9 − 63.9
08–4 292 26.30 283.0 − 6.7 19.51 199.8 45.1 3.85 6.5 44.1
08-5a 300 12.62 293.4 53.4 6.91 123.1 36.2 3.96 29.7 4.6
08-6a 312 23.30 262.1 28.4 8.06 10.8 30.8 5.21 138.5 45.7
08–7 228 27.19 279.4 47.2 12.59 23.2 12.4 5.57 123.9 40.1
08–8 320 30.29 287.6 39.5 13.65 150.1 41.8 6.76 37.9 22.8
08–9 350 24.54 278.0 12.7 12.18 153.3 68.5 8.51 12.0 17.1
08–10 310 22.37 286.5 − 2.8 10.99 206.5 74.4 5.1 15.8 15.4
08-11a 425 8.77 261.5 14.9 5.97 1.4 32.7 3.87 330.6 − 53.2
08–12 460 16.54 250.6 36.3 9.54 91.4 51.9 4.18 348.2 10.2
a
: Measurement point was not included in the back analysis.

with the maximum principal stress was about 30◦ . displacement boundary conditions, and the variable Ug was loaded ac­
Notably, the measuring points 08-5, 08-6, and 08–11 are excluded cording to the acceleration of gravity in the finite difference software
from the numerical back analysis of in situ stress fields because these and FLAC3D.49 Table 2 presents the physical and mechanical properties
the underground powerhouse are on both sides of the secondary fault adopted for the numerical simulations.
F2, which have limited significance for the back analysis. Meanwhile, all
the other measuring points have been retained because the extent to 4.3.2. Step 2
which these measuring points are affected by faults is unclear. And ac­ Fig. 8 shows the Pearson correlation coefficients between the inde­
curate modeling can simulate the stress state variation caused by faults pendent and dependent variables. The correlation between each inde­
at the measuring points. pendent variable and dependent variable (σ) is ranked as follows:
Ux (0.72) > Ug (0.61) > Uy (0.48) > Uxz (0.47) > Uxy (0.38) > Uyz (0.02)
4.3. Application of BSSR Moreover, Uyz is a nonsignificant variable with little contribution to
the actual in situ stress field and can be ignored in the back analysis.
4.3.1. Step 1 Furthermore, the collinearity problem mainly exists in two groups of
The study area was divided into six strata according to the classifi­ independent variables; one group comprises significant variables Ux and
cation of the surrounding rock mass, and 17 faults near the powerhouse nonsignificant variables Uy and Uxz , while the other group comprises
and measurement points were selected to establish a 3D finite element significant variables Ug and nonsignificant variables Uxy and Uxz ; how­
model (Fig. 7). In total, there were 1.8 million zones and 750,000 nodes. ever, it is still difficult to determine which variables should be elimi
The variables Ux , Uy , Uxy , Uyz , and Uxz were loaded according to the

6
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

nated. Therefore, the stepwise regression is further adopted to eliminate


this collinearity problem.
Table 3 presents the regression parameters obtained for each SR step.
R2Adj became 0.514 after the introduction of Ux and increased to 0.882
after the introduction of Ug . It remained mostly unchanged with the
introduction of the other variables, which indicates that Uy ,Uxy ,Uyz , and
Uxz had extremely small contributions. Thus, only Ux and Ug were
retained in the regression model:

Modelregression = 2.020⋅Ug + 2.860⋅Ux + 2.362, R2Adj = 0.882 (11)

Furthermore, the proof by contradiction can be used to verify the


rationality of the variable selection for SR. First, the null hypothesis
(H0): Ux is eliminated. Then, σx is mainly provided by Uy and Ug , and it
can be calculated using Eq. (3): σ y ⋅ν + σz ⋅ν/(1 − ν) = − 5.82MPa (ν =
0.215, σy = − 10.64 MPa, and σ z = − 12.88 MPa), which is far less than
σ x = − 20.67 MPa. Therefore, Ux must be retained to form σ x , and H0 is
rejected. Here, σx , σy , and σ z are the mean measured stress components
calculated according to Table 1. The same process can be used to prove
the rationality of other variable selections.

Fig. 7. 3D geological model (IV, III2, III1, and II: classification of the sur­ 4.3.3. Step 3
rounding rock mass; B1: clastic rock; B3: rock debris containing mud; B4: mud Once the constant in the regression model was eliminated, the
containing rock debris; B5: mud). sensitivity analysis was performed on the regression coefficients of the
retained variables Ux and Ug . The results are presented in Table 4. The
optimal values of the variable coefficients were close to the estimated
Table 2
values (i.e., B) by SR, which indicates that the regression model had
Physical and mechanical properties adopted for numerical simulations.
relatively good prediction accuracy.
Geological Type Elastic modulus Poisson’s Density (kg/
unit (GPa) ratio m3)
4.3.4. Step 4
Strata IV 5.5 0.275 2550 Fig. 9(a)–(f) show histograms comparing the measured and predicted
III2 9.5 0.255 2650
stress components at the measurement points. The normal stresses were
III1 13.5 0.235 2650
II 19.0 0.215 2750 relatively consistent, but the shear stresses had relatively large differ­
Fault B1 3.0 0.28 2150 ences, which can be attributed to the complex geological environment.
B3 2.0 0.29 2250 Overall, the stress field predicted by BSSR reflected the distribution
B4 2.0 0.31 2350
characteristics described of the measurements.
B5 1.0 0.33 2450
Fig. 10 shows the measured and predicted orientations of the

Fig. 8. Pearson correlation coefficients between var­


iables in back analysis of the in situ stress field of
Yebatan underground powerhouse. The lower-left
panels show the scatter diagrams between two vari­
ables, which were used for linear fitting. The upper-
right panels show the Pearson correlation co­
efficients (**: confidence level of 0.01, *: confidence
level of 0.05). The histograms along the diagonal do
not show a good normal distribution because of the
complex geological environment and limited mea­
surement data.

7
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

Table 3
Regression parameters obtained at each step of SR.
Step Constant Ux Ug Uy Adjusted R2 Variable introduced Variable eliminated

Step 1 B − 3.491 2.830 – – 0.514 Ux –


P 0.001 0.000 – –
Step 2 B 0.340 2.820 1.967 – 0.882 Ug –
P 0.564 0.000 0.000 –
Step 3 B 0.317 2.876 1.989 − 0.099 0.880 Uy –
P 0.596 0.000 0.000 0.610
Step 4 B 0.338 2.819 1.966 0.003 0.879 Uxz Uy
P 0.582 0.000 0.000 0.990
Step 5 B − 0.169 2.778 1.874 1.213 0.885 Uxy Uxz
P 0.801 0.000 0.000 0.128
Step 6 B 0.257 2.807 1.982 0.114 0.880 Uyz Uxy
P 0.687 0.000 0.000 0.715
Step 7 B 0.340 2.820 1.967 – 0.882 – Uyz
P 0.564 0.000 0.000 –

B: unstandardized coefficient (variable coefficient); P: P-value of t-test. P < 0.05 indicates that the independent variable has a significant impact on the dependent
variable.

brittle failure risk of the underground caverns. The results showed that
Table 4 the EDZ locations as indicated by the RFD distribution match the actual
Modification of regression coefficients for the variables Ug and Ux.
spalling locations (Fig. 11(c)). In summary, the BSSR results for the
Variable Coefficients Ug Ux underground powerhouse of Yebatan hydropower station accurately
95% confidence interval for B [1.46 2.27] [2.33 3.29] characterized the 3D in situ stress field.
B 1.97 2.82
Optimal value 2.09 2.65 5. Discussion

maximum principal stress at the measurement points. The predicted 5.1. Comparison of BSSR and MLR
orientations were NW80◦ –EW and tended to follow the valley with an
inclination of about 30◦ . This is in good agreement with the measure­ The in situ stress field of Yebatan underground powerhouse was used
ments. The predicted stress field reflected the azimuthal characteristics to compare the advantages of the selected model (i.e., BSSR) and full
of the maximum principal stress described by the measurements. model (i.e., traditional MLR). Table 5 presents the regression parameters
Fig. 11 compares the real and predicted EDZ locations. The sur­ for both models, which led to the following observations.
rounding rock mass of the PD08 tunnel had large-scale spalling (Fig. 11
(a)) concentrated in the upstream spandrel, which basically corresponds • The explanatory abilities of the models were equivalent. Both models
to the orientation of the maximum principal stress (Fig. 11(b)). In this had R2Adj = 0.882, which indicates that the elimination of
study, the rock failure degree (RFD)50 was used as an indicator of the

Fig. 9. Comparison between the measured and predicted stress components: (a) σ x , (b) σ y , (c) σ z , (d) τxy , (e) τyz , and (f) τxz .

8
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

Fig. 11. Comparison between the real and predicted locations of EDZs: (a)
Fig. 10. Comparison between the measured and predicted maximum principal Spalling in PD08 tunnel, (b) relationship between EDZ locations and the
stress orientations: (a) azimuth and (b) plunge. maximum principal stress, (c) EDZs and RFD distribution in the under­
ground powerhouse.

independent variables did not affect the explanatory ability of the


5.2. Physical significance of mathematical model of the in situ stress field
regression model for dependent variables.
• The selected model had greater significance. The F value was about
In general, the stress field in a deep valley has three kinds of sources:
three times higher for the selected model than for the full model, and
gravitational, tectonic, and local. For the back analysis of in situ stress
all of its variables passed the significance test. Thus, the selected
fields, these three types of stress sources can be characterized by six
model was more suitable for explaining the dependent variables.
geological actions, as shown in Fig. 1. The contributions of different
• The selected model estimated the variable coefficients more accu­
stress sources were weighted by establishing special and general math­
rately. In contrast, the full model had variable coefficients with
ematical models of in situ stress fields (Table 6).
negative values, which contradicted the theoretical estimation.
• The selected model showed better predictive ability for the confi­
5.2.1. Special mathematical model
dence intervals of variables. The 95% confidence interval was nar­
rower for the variables in the selected model. For example, Ux and Ug ModelSpecial = g2 ⋅Ug + x2 ⋅Ux + y2 ⋅Uy + xy2 ⋅Uxy + yz2 ⋅Uyz + xz2 ⋅Uxz (12)
were common between the two models; the 95% confidence intervals
for these two variables were about 4.1 and 1.8 times wider, respec­ This model can be used to evaluate the contribution weight of the six
tively, in the full model than in the selected model. geological actions to the in situ stress field. Note that after a variable
• The number of eliminated variables is positively correlated with the coefficient (i.e., B) is modified, the weight of its contribution will
workload. Thus, the calculation workload of the selected model was change, but the degree of change is small.
about a third that of the full model.
5.2.2. General mathematical model
In summary, the selected model showed significant advantages over The general model can be used to evaluate the contribution weight of
the full model, which demonstrates the applicability of BSSR to the the three stress sources to the in situ stress field macroscopically:
numerical back analysis of 3D in situ stress fields.

9
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

Table 5
Comparison between regression parameters of the selected and full models.
Model Variable Coefficients Model Summary

Unstandardized Standardized Coefficients T Sig. 95% confidence interval R2Adj RSS F Sig.
Coefficients for B

B Std. Error Beta Lower Upper

Selected model Constant 0.34 0.59 – 0.58 0.564 − 1.72 1.22 0.882 604.76 198.38 0.000
Ug 1.97 0.15 0.60 12.74 0.000 1.46 2.27
Ux 2.82 0.18 0.72 15.25 0.000 2.33 3.29
Full model Constant − 0.25 0.73 – − 0.34 0.735 − 0.50 0.30 0.882 553.86 67.26 0.000
Ug 1.86 0.20 0.57 9.21 0.000 − 0.39 2.97
Ux 2.81 0.24 0.72 11.76 0.000 − 0.96 0.81
Uy − 0.10 0.20 − 0.03 − 0.50 0.621 − 0.38 0.54
Uxy 1.29 0.84 0.08 1.54 0.130 − 0.84 1.52
Uyz − 0.07 0.44 − 0.01 − 0.16 0.873 2.45 3.19
Uxz 0.08 0.23 0.02 0.33 0.741 1.66 2.28

coupling effects on each other. Furthermore, the tectonic actions with


Table 6
collinearity problems (caused owing to the above two cases) can be
Weights representing the contributions of geological actions and stress sources.
ignored for the back analysis of in situ stress fields because their net
Source Variable Variable Variable Source contributions are very small, and their elimination will not affect the
coefficients weight weight
final back analysis result.
Gravitational Ug g1 g2 g2/g1
source
Tectonic source Ux x1 x2 1 − g2
Uy y1 y2 5.3. Suggestions for numerical back analysis of in situ stress fields
Uxy xy1 xy2
Uyz yz1 yz2 Based on the results of this study, the following is suggested for the
Uxz xz1 xz2 numerical back analysis of in situ stress fields. First, the projections of
Local source – – – g2 − g2/g1
the principal axes for the maximum and minimum stresses on the hor­
Variable weight: Change in R2Adj after the introduction of an independent vari­ izontal plane should correspond to the x- and y-axes of the constructed
able (ΔR2Adj ). 3D geological model. For example, the x- and y-axes of the geological
model correspond to the principal axes of the maximum stress (E–W
direction) and minimum stress (S–N direction) stress in the case of
( ) Yebatan hydropower station, respectively. This minimizes the rotation
g2 g2
ModelGeneral = ⋅ SG + (1 − g2 ) ⋅ ST + g2 − ⋅SL (13) angle of the principal axis of stress on the horizontal plane, and the
g1 g1
compressive tectonic stress in the y-direction. Thus, Uxy and Uy become
where SG , ST , and SL represent the gravitational, tectonic, and local redundant variables, which increases the probability of their
sources respectively. However, the influence of the local sources results elimination.
in the coefficient of Ug usually not being equal to 1. Thus, the weight of Second, BSSR can be used to determine the significant variables and
the gravitational source SG = g2 /g1 and weight of the variable Ug minus their 95% confidence intervals before the application of neural net­
works, genetic algorithms, or other methods for the nonlinear back
the weight of the gravitational source can be regarded as the weight of
analysis of variable coefficients to avoid local optimums and to reduce
the local sources. Local sources also have a significant influence on the
the computational workload significantly. Additionally, although BSSR
variables Ux ,Uy ,Uxy ,Uyz , and Uxz , but the degree of influence is difficult
is based on the principle of linear elasticity, the modification of variable
to quantify. Therefore, for actual evaluations, the contribution weight of
coefficients via sensitivity analysis can also implement the nonlinear
the local sources is given by SL > g2 − g2 /g1 , and the contribution
back analysis.
weight of the tectonic source is given by ST < 1 − G3 − L3 = 1 − g2 .
For the underground powerhouse of Yebatan hydropower station,
the mathematical models of the in situ stress field are as follows: 6. Conclusions

ModelSpecial = 0.51⋅Ux + 0.37⋅Ug (14) This study focused on analyzing the collinearity between indepen­
dent variables caused by complex surface topography and proposed
ModelGeneral = 0.19⋅SG + 0.63⋅ST + 0.18⋅SL (15) BSSR to eliminate this problem. BSSR was then applied to the under­
According to Eq. (14), the EW compressional tectonic action (i.e., ground powerhouse of Yebatan hydropower station to demonstrate its
variable Ux ) has a weight of 51%, and gravity (i.e., variable Ug ) has a feasibility. The main conclusions were as follows:
weight of 37%. According to Eq. (15), the gravitational source has a
weight of 19%, the tectonic source has a weight of 63%, and the local (1) BSSR comprises four steps: 3D geological modeling, variable
sources have a weight of 18%. These models indicate significant tectonic correlation analysis and SR, modification of variable coefficients,
movement at the Yebatan site. and verification of the back analysis results. BSSR effectively
Notably, there are two cases that may cause collinearity problems: 1) solves the problem of variable collinearity for numerical back
the contribution weight of tectonic action to the actual in situ stress field analysis of in situ stress fields, and it has good applicability and
is very small, and 2) the contribution weight is replaced with other more reliability.
important tectonic actions, resulting in its net contribution weight being (2) The collinearity between variables is reflected by the indepen­
very small or even zero. If it is case 1, these tectonic actions can be dent variable matrix U; it originates from Poisson’s effect, and it
ignored, even in geological analysis; however, if it is case 2, these cannot is significantly affected by geological conditions. The collinearity
be simply ignored because the collinearity problem was caused by their between variables increases with the complexity of the surface
topography.

10
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

(3) The physical meaning of the selected model established by BSSR 10 Zhang C, Feng XT, Zhou H. Estimation of in situ stress along deep tunnels buried in
complex geological conditions. Int J Rock Mech Min Sci. 2012;52:139–162. https://
is well defined. The contributions of six geological actions (Ux ,Uy ,
doi.org/10.1016/j.ijrmms.2012.03.016.
Ug , Uxy , Uyz , and Uxz ) and three stress sources (gravitational, tec­ 11 CD M. Characterizing in situ stress domains at the AECL underground research
tonic, and local) can be evaluated quantitatively according to the laboratory. Can Geotech J. 1990;27:631–646. https://doi.org/10.1139/t90-077.
12 Day-Lewis AD. Characterization and Modeling of in Situ Stress Heterogeneity. California,
special and general mathematical models of the in-situ stress
USA: Stanford University; 2008.
field. The establishment of multidimensional mathematical 13 Gao K, Harrison JP. Multivariate distribution model for stress variability
models will further improve the understanding of in situ stress characterisation. Int J Rock Mech Min Sci. 2018;102:144–154. https://doi.org/
fields for underground engineering structures. 10.1016/j.ijrmms.2018.01.004.
14 Gao K, Harrison JP. Scalar-valued measures of stress dispersion. Int J Rock Mech Min
(4) Compared with the full model established by traditional MLR, the Sci. 2018;106:234–242. https://doi.org/10.1016/j.ijrmms.2018.04.008.
selected model established by BSSR had several obvious advan­ 15 Lei Q, Gao K. Correlation between fracture network properties and stress variability
tages. The mathematical model is more significant, and the pre­ in geological media. Geo. 2018;45(9):3994–4006. https://doi.org/10.1002/
2018gl077548.
diction accuracy of variable coefficients and their confidence 16 Lei Q, Gao K. A numerical study of stress variability in heterogeneous fractured
intervals is greater. Meanwhile, the predictive ability does not rocks. Int J Rock Mech Min Sci. 2019;113:121–133. https://doi.org/10.1016/j.
decrease with the elimination of variables, while the computa­ ijrmms.2018.12.001.
17 Obara Y, Sugawara K. Updating the use of the CCBO cell in Japan: overcoring case
tional workload is significantly reduced. studies. Int J Rock Mech Min Sci. 2003;40(7-8):1189–1203. https://doi.org/10.1016/
j.ijrmms.2003.07.007.
Notably, the classification of surface topography is relatively rough 18 Stephansson O, Ljunggren C, Jing LJT. Stress measurements and tectonic
implications for Fennoscandia. Tectp. 1991;189(1-4):317–322, 0040-1951(91)
in this study. For extremely complex geological environment (e.g., un­ 90504-L.
even surface topography and independent mountains with denudation 19 Yale D. Fault and stress magnitude controls on variations in the orientation of in situ
around), strong local sources increase the variability and complexity of stress. In: Ameen M, ed. London: Geological Society London, Special Publications. 2003:
55–64. https://doi.org/10.1144/GSL.SP.2003.209.01.06, 209(1).
in situ stress fields, which makes it possible for some significant vari­
20 Christiansson R, Hudson JA. ISRM Suggested Methods for rock stress
ables to be eliminated by mistake. Therefore, the rationality of variable estimation—Part 4: quality control of rock stress estimation. Int J Rock Mech Min Sci.
selection should be fully discussed and verified in the practice. In 2003;40(7-8):1021–1025. https://doi.org/10.1016/j.ijrmms.2003.07.004.
addition, the influence mechanism of complex geological environment 21 Haimson BC, Cornet FH. ISRM Suggested Methods for rock stress estimation—Part 3:
hydraulic fracturing (HF) and/or hydraulic testing of pre-existing fractures (HTPF).
(e.g., surface topography and faults) on variable collinearity and in situ Int J Rock Mech Min Sci. 2003;40(7-8):1011–1020. https://doi.org/10.1016/j.
stress field will be explored in subsequent studies. ijrmms.2003.08.002.
22 Hudson JA, Cornet FH, Christiansson R. ISRM Suggested Methods for rock stress
estimation—Part 1: strategy for rock stress estimation. Int J Rock Mech Min Sci. 2003;
40(7-8):991–998. https://doi.org/10.1016/j.ijrmms.2003.07.011.
Declaration of competing interest 23 Sjöberg J, Christiansson R, Hudson JA. ISRM Suggested Methods for rock stress
estimation—Part 2: overcoring methods. Int J Rock Mech Min Sci. 2003;40(7-8):
The authors declare that they have no known competing financial 999–1010. https://doi.org/10.1016/j.ijrmms.2003.07.012.
24 Li S, Feng XT, Li Z, Chen B, Zhang C, Zhou H. In situ monitoring of rockburst
interests or personal relationships that could have appeared to influence
nucleation and evolution in the deeply buried tunnels of Jinping II hydropower
the work reported in this paper. station. Eng Geol. 2012;137–138:85–96. https://doi.org/10.1016/j.
enggeo.2012.03.010.
Data availability 25 Lee H, Ong SH. Estimation of in situ stresses with hydro-fracturing tests and a
statistical method. Rock Mech Rock Eng. 2017;51(3):779–799. https://doi.org/
10.1007/s00603-017-1349-1.
Data will be made available on request. 26 Feng J, Shang L, Li X, Luo P. 3D numerical simulation of heterogeneous in situ stress
field in low-permeability reservoirs. Petrol Sci. 2019;16(5):939–955. https://doi.org/
10.1007/s12182-019-00360-w.
Acknowledgements 27 Liu J, Ding W, Yang H, et al. 3D geomechanical modeling and numerical simulation
of in-situ stress fields in shale reservoirs: a case study of the lower Cambrian
Niutitang formation in the Cen’gong block, South China. Tectp. 2017;712–713:
This work was supported by the National Natural Science Foundation
663–683. https://doi.org/10.1016/j.tecto.2017.06.030.
of China [grant numbers 51879261, 51979268] and the Innovation 28 Matsuki K, Kaga N, Yokoyama T, Tsuda N. Determination of three dimensional in situ
Group Project of the Natural Science Foundation of Hubei Province stress from core discing based on analysis of principal tensile stress. Int J Rock Mech
[grant number ZRQT2020000114]. Min Sci. 2004;41(7):1167–1190. https://doi.org/10.1016/j.ijrmms.2004.05.002.
29 Stephansson O, Zang A. ISRM suggested methods for rock stress estimation—part 5:
establishing a model for the in situ stress at a given site. The ISRM Suggested Methods
References for Rock Characterization, Testing and Monitoring: 2007-2014. Springer. 2012:187–201.
30 Ning Y, Tang H, Smith JV, Zhang B, Shen P, Zhang G. Study of the in situ stress field
in a deep valley and its influence on rock slope stability in Southwest China. Bull Eng
1 Amadei B, Stephansson O. Rock Stress and its Measurement. Springer Science &
Geol Environ. 2021;80(4):3331–3350. https://doi.org/10.1007/s10064-020-02094-
Business Media; 1997.
1.
2 Zoback MD. Reservoir Geomechanics. Cambridge university press; 2010.
31 Zhang SR, Hu AK, Wang C. Three-dimensional inversion analysis of an in situ stress
3 Zang A, Stephansson O. Stress Field of the Earth’s Crust. Springer Science & Business
field based on a two-stage optimization algorithm. J Zhejiang Univ - Sci A. 2016;17
Media; 2009.
(10):782–802. https://doi.org/10.1631/jzus.A1600014.
4 Latham JP, Xiang J, Belayneh M, Nick HM, Tsang CF, Blunt MJ. Modelling stress-
32 Guo HZ, Ma QC, Xue XC, et al. The analytical method of the initial stress field for
dependent permeability in fractured rock including effects of propagating and
rock masses. J Geotech. 1983;(3):64–75. https://doi.org/10.3321/j.issn:1000-
bending fractures. Int J Rock Mech Min Sci. 2013;57:100–112. https://doi.org/
4548.1983.03.006 [in Chinese].
10.1016/j.ijrmms.2012.08.002.
33 Men W, Hei C, Chen ZQ, et al. Application of ridge regression in the inversion
5 Feng XT, Zhou YY, Jiang Q. Rock mechanics contributions to recent hydroelectric
analysis of the initial geo-stress field of rock masses. Rock Soil Mech. 2021;42(4):
developments in China. J Rock Mech Geotech. 2019;11(3):511–526. https://doi.org/
1156–1169. https://doi.org/10.16285/j.rsm.2020.1317 [in Chinese].
10.1016/j.jrmge.2018.09.006.
34 Yuan FB, Liu J, Li PJ, Qiao LP, Li P. Back analysis and multiple-factor influencing
6 Ljunggren C, Chang Y, Janson T, Christiansson R. An overview of rock stress
mechanism of high geostress field for river valley region of Laxiwa Hydropower
measurement methods. Int J Rock Mech Min Sci. 2003;40(7-8):975–989. https://doi.
Engineering. Rock Soil Mech. 2007;(4):836–842. https://doi.org/10.16285/j.
org/10.1016/j.ijrmms.2003.07.003.
rsm.2007.04.040 [in Chinese].
7 Zhao JS, Jiang Q, Lu JF, Chen BR, Pei SF, Wang ZL. Rock fracturing observation
35 Jian Q, Feng XT, Chen JL, Zhang CS, Huang SL. Nonlinear inversion of 3D initial
based on microseismic monitoring and borehole imaging: in situ investigation in a
geostress field in Jinping II Hydropower Station region. Rock Soil Mech. 2008;(11):
large underground cavern under high geostress. Tunn Undergr Space Technol. 2022;
3003–3010. https://doi.org/10.16285/j.rsm.2008.11.015 [in Chinese].
126. https://doi.org/10.1016/j.tust.2022.104549.
36 Zhang LW, Zhang DY, Qiu DH. Application of radial basis function neural network to
8 Zheng MZ, Li SJ, Yao Z, Zhang AD, Xu DP, Zhou JF. Core discing characteristics and
geostress field back analysis. Rock Soil Mech. 2012;33(3):799–804. https://doi.org/
mitigation approach by a novel developed drill bit in deep rocks. J Cent South Univ.
10.16285/j.rsm.2012.03.038 [in Chinese].
2020;27(10):2822–2833. https://doi.org/10.1007/s11771-020-4512-x.
37 Sakurai S. Lessons learned from field measurements in tunnelling. Tunn Undergr
9 Xu DP, Huang X, Jiang Q, et al. Estimation of the three-dimensional in situ stress field
Space Technol. 1997;12(4):453–460. S0886-7798(98)00004-2.
around a large deep underground cavern group near a valley. J Rock Mech Geotech.
2021;13(3):529–544. https://doi.org/10.1016/j.jrmge.2020.11.007.

11
H. Xu et al. International Journal of Rock Mechanics and Mining Sciences 170 (2023) 105474

38 Qi L, Cui W, Xiong KZ, Huang XD. Application of grey theory to analysis of in-situ 45 Jaeger JC, Cook NG, Zimmerman R. Fundamentals of Rock Mechanics. John Wiley &
stress field. J Rock Mech Eng. 2002;(10):1547–1550. https://doi.org/10.3321/j.issn: Sons; 2009.
1000-6915.2002.10.023 [in Chinese]. 46 Zhu HC, Tao ZY. Preliminary analysis of topography and ground stress distribution.
39 Yang S, Huang L, Xie F, Cui X, Yao R. Quantitative analysis of the shallow crustal Water Resour Hydropower Eng. 1994;1:29–34 ([in Chinese]).
tectonic stress field in China mainland based on in situ stress data. J Architect Educ. 47 Zhu W, Li G, Wang KJRm, engineering r. Analyses of disking phenomenon and stress
2014;85:154–162. https://doi.org/10.1016/j.jseaes.2014.01.022. field in the region of an underground powerhouse. Rock Mech Rock Eng. 1985;18(1):
40 Zoback ML. First- and second-order patterns of stress in the lithosphere: the world 1–15. https://doi.org/10.1007/BF01020412.
stress map Project. JGR. 1992;97(B8). https://doi.org/10.1029/92jb00132. 48 Goodman RE. Introduction to Rock Mechanics. New York: Wiley; 1989.
41 Zoback ML, Zoback MD, Adams J, et al. Global patterns of tectonic stress. Nature. 49 Itasca HJCP. FLAC3D,(Fast Lagrangian Analysis of Continua in 3 Dimensions). 2002.
1989;341(6240):291–298. 50 Feng XT, Wang Z, Zhou Y, Yang C, Pan PZ, Kong R. Modelling three-dimensional
42 Muirhead RJ. Aspects of Multivariate Statistical Theory. John Wiley & Sons; 2009. stress-dependent failure of hard rocks. Acta Geotech. 2021;16(6):1647–1677. https://
43 Timm NH. Applied Multivariate Analysis. New York, NY: Springer; 2002. doi.org/10.1007/s11440-020-01110-8.
44 Draper NR, Smith H. Applied Regression Analysis. John Wiley & Sons; 1998.

12

You might also like