You are on page 1of 8

RockMechanicsContfbutions

andChallenges,
Hustrulid
& Johnson(eds) 1990Balkema,Rotterdam.
ISBN 90 6191 123 0

Threedimensional
stochastic
jointgeometry
modeling
including
a verification:A casestudy
H. S.W.Pinnaduwa,
N.W.Deepa& O.Stephansson
LuleclUniversityof Technology,
Sweden

1 INTRODUCTION

Presence of discontinuities strongly affects the mechanical and hydraulic behavior


of rock masses. Therefore, joint geometry network plays a vital role in civil
mining and petroleum engineering projects associated with jointed rock. The
need for a realistic representation of joint geometry in rock masses has been
recognized.
Since joint geometry pattern can vary from one statistically homogeneousregion
to another, each statistically homogeneousregion should be represented by a
separate joint geometry model. Therefore, the first step in the procedure of
joint geometry modeling in a rock massshould be the identification of statistically
homogeneousregions. To model joint geometry in three dimensional (3D) space,
for a statistically homogeneousregion, it is necessary to know the number of
joint sets, and for each joint set, the intensity, spacing, location, orientation,
shape and dimension distributions. These joint parameters are inherently sta-
tistical. Sample values of joint parameters provided by the field data usually
subject to errors due to sampling biases and represent only one or two dimensional
properties. Therefore, before inferring these parameters from sampling values,
sampling biases should be corrected on field data. In addition, principles of
stereology need to be used in order to infer 3D parameters of the joint sets
from 1D or 2D parameter values. At present, a 3D joint geometry modeling scheme
which incorporates the aforementioned features and also includes a formal ver-
ification procedure is not available in the literature (Kulatilake 1988 & 89).
This paper shows development of 3D joint geometry models which investigates
statistical homogeneity, includes corrections for sampling biases and applications
of stereological principles and also checks validity of such built-up models to
an area in Stripa mine, Sweden. Due to space limitation, only a summarized
presentation is given in this paper.

2 DEVELOPMENT OF 3D JOINT GEOMETRY MODELS

2.1 Investigation of statistical homogeneity

Figure 1 shows the flow chart ued to develop 3D stochastic joint geometry models.
Orientation data from a thirty six meter long stretch of the ventilation drift,
Stripa mine (Rouleau et al. 1981) were used to investigate the statistical
homogeneity of the rock mass. A ten meter long stretch from this thirty six
meter stretch was identified as the largest statistically homogeneous region
(Kulatilake et al. 1990a).

1051
C)u[tt:3'

. .

Step1: Obtaina statistically


homogeneous
region.
...

.. ..

. .

Step 2: Delineate joint sets


Step3: Applycorrection
fororientation
biasforeachjointset.
Step4: Modelthetrueorientation
distribution
foreachjointset, c :.-- ... - ,-
-

c]us: 4 ..... , .
. .

\
.....

Slep5: Determine
jointspacing
distributions
alongscanlines ... , ..... :.
takingintoaccountthe sampling biason spacing. ...................
.

Step6: Inferjointspacingdistribution
alongthemeanpole / ..........
.......

direction(true spacings)for each joint set.


..

,. . ;

Step7: Estimatethetracelengthdistribution
foreachjointset
taking into account the sampling biases. . ---.___ .

Step8: Inferjointsizedistribution
foreachjointsettakinginto .

account Ihe sampling biases (joints are . .

C2uE: 5
considered as circular discs).
Step9: Eslimate
themeanjointcenlerdensilyin 3D Figure 2 Polar equal area projection of
(number/volume)for each joint set. orientation data in upper hemisphere
Step 10: Obtaina distribution
for the randomvariable,"number
of joint centers per chosen volume".
Step11: Suggest
a 3D stochastic
jointgeometry
modelby
describingthe joint geometry parameters:
(a) number of joint sets
(b) orientationdistributionfor each joint set
(c) spacing distributionfor each joint set
(d) distribution
for densityin 3D for each joint set
(e) size (diameler when joint shape is considered
as circular ) dislributionfor each joinl set.
Figure 1 Flow chart of the procedure used
for development of stochastic joint geome- Fibre 3 Distribution of orientation for
try models clter (a) based on raw relative
frequencies; (b) based on corrected rele-
lve frequencies
2.2 Joint orientation modeling

Using the method given in Shanley and Mahtab (1976), orientation data in this
region were delineated into four joint sets as shown in Fig. 2. For each joint
set, high dispersion of the orientation data can be seen very clearly from Fig.
2. For such joint sets orientation bias correction may be significant. Joint
data for each joint set come from the east and west walls and the floor of the
drift. The procedure available for orientation bias correction for finite size
joints intersecting finite size sampling domains (Kulatilake et al. 1990b) is
directly applicable only for vertical sampling planes. Thus a general procedure
applicable for sampling domains of any orientation was developed to correct for
orientation bias (Wathugala et al. 1990). This procedure was applied to each
joint set in order to correct for orientation bias. For example, Fig. 3 shows
the effect of sampling bias correction on cluster 4.
Raw orientation data as well as orientation data corrected for sampling bias
were subjected to chi-square goodness-of-fit tests to check the suitability of
Bingham distribution (Bingham 1964) and the hemispherical normal distribution
(Arnold 1941) in representing the statistical distribution of data. The maximum
significance level at which the tried probability distribution is suitable to
represent the statistical distribution of data was computed for both raw and
corrected data of each cluster. This significance level should be at least 0.05
to accept the tried probability distribution to represent the data. According
to the results, only cluster 4 follows a Bingham distribution for raw data;
neither of the two probability distributions satisfied the corrected data.
Therefore, empirical distributions obtained for corrected data were chosen to
represent the statistical distributions of joint set orientations. For three
clusters, considerable differences were observed between the cht-square values

1052
for raw and corrected data. That reflects the significance of sampling bia-
correction.

2.3 Joint spacing, linear intensity (frequency) and location modeling along
scanlines and mean vectors of joint sets

These tasks were performed according to two methods.

2.3.1 Method 1

For a joint set, meanspacing and meanlinear intensity (average numberper unit
length) estimates dependon the chosen direction as shownin Fig. 4. For a fixed
direction, mean linear intensity is the reciprocal of the mean spacing. The
estimations along the mean vector (mean pole) directions can be considered as
the true values. Figure 5 showsthe flow chart used to obtain spacing and linear
intensity distributions along the mean vector directions starting from the
observations made on several scanlines. Seven scanline directions (Fig. 6) were
chosen to analyze spacing and intensity in different directions to have a good
coverage in 3D. Along each direction, several parallel scanlines were drawn
either on the walls or on the floor of the drift, having joint traces coming
from all four joint sets, to estimate spacing distribution as well as observed
mean spacing. For each direction, observed spacing values were subjected to
chi-square and Kolmogorov-Smirnovgoodness-of-fit tests to check the suitability
of exponential gamma, lognormal, normal, uniform and triangular distributions
in representing the observed spacings. For all seven directions the exponential
distributions were found to be the best distributions to represent the dis-
tributions of observed spacing by satisfying the goodness-of-fit tests at very
high significance levels. A typical fit obtained is shown in Fig. 7.

.
IChocoe
scanline
directions
onthewalls
andfloor
Haanpole

direon IEstimate
average
spacing
each
along
ontheI Determi
the
best
scanline
direction
observed
date
based probability
bution
distri-
torepresent
distribution
of
observed
spacing

Figure &
linear intensity
ki=ki Ces

estimation
e
Influence of direction on mean
i
{Correction
bias

'or
iempling
onaverage
spacin

I
Calculate
along
each
the
stat
along
scan
observedl

corrected
scanline average
direction
' spacing

orlenlation
of along
mean
vectors
of
clusters J I the jointsets

lDetermine
the
best
probability
distribution
.
to representspacing and linear intensity
along each of the meen vector direction

Figure 5 Flow chart used to estimate


Figure 6 Chosenscanline directions for spacing and linear intensity distribution
spacing and intensity analysis along the mean vector directions

Chosen scanlines were of finite size. To obtain unbiased estimates of spacing


and intensity, scanlines shouldbe of infinite size. SenandKazi (1984) pointed
out this samplingbias on spacing and suggestedequations to correct for this
bias when the observed spacing follows an exponential distribution. However,
there is an error in their equation and the corrected equation is given in
Kulatilake(1988& '89). Thelatter equation
wasusedto obtaincorrected
mean
spacing
fromtheobserved
mean
spacing.Thiserrorwasfound
to berespectively
1053
about 0.1, 2 and 5% for length of scanline/mean spacing ratios of 9, 6, and 5.
Hudson and Priest (1983) expressed mean linear intensity resulting from several
joint sets in any arbitrary direction in 3D in terms of the mean linear intensities
along the mean pole directions of the joint sets. Using this expression,
Karzulovic and Goodman(1985) suggested a procedure based on least square method
to estimate the mean linear intensities along the mean vector directions of joint
sets using the mean linear intensities estimated along several scanline directions.
To apply this procedure, the number of scanlines should be greater than or equal
to the number of joint sets. This procedure was used with aforementioned seven
scanlines in order to estimate the mean linear intensities along the mean vectors
of the four joint sets. The relative standard error obtained in the least square
procedure was less than four percent with respect to any mean linear intensity
estimated along scanline directions indicating quality estimations. These
findings lead to the conclusion that joint spacing along mean vector directions
follow exponential distributions with mean spacing values obtained in respective
directions. Thus, according to statistical theory, linear intensity and linear
location distributions along the mean vector direction for each joint set follow
respectively Poisson and uniform distributions.

2.3.2 Method 2

First, joint traces on the two walls and the floor were sorted out into the four
joint sets. Then for each cluster, the following analyses were conducted.
The flow chart shown in Fig. 5 was followed for each cluster until mean corrected
spacings and mean linear intensities were found along the seven scanline
directions. Then the relationship shown in Fig. 4 was used to estimate the
corresponding mean linear intensities along the mean vector direction of the
cluster. Finally, these values were averaged to obtain the final mean linear
intensity estimate for each cluster. However, when the angle between the mean
vector direction and the scanline direction was greater than seventy degrees,
the estimation corresponding to that was not taken into account in the averaging
procedure, since under this case the accuracy of the estimation is questionable.
Again, the conclusions regarding the probability distribution types for spacing,
intensity and location are sae as for method 1.

2.4 Joint trace length and 3D joint size modeling

First, joint traces appearing on the two walls and the floor were sorted out
into the four joint sets. Then for each cluster, the following modeling was
performed; one on the wall data and the other based on the floor data. For
example, Fig. 8 provides the subsequent modeling performed on the wall data for
each joint cluster.
Due to finite size of the walls (10m x 4m) and the floor (10m x 4m), the
observed traces are subject to censored and size biases. Since observed traces
of length less than about 0.4m were neither mapped on the walls nor on the floor,
it was assumed that the truncation limit for joint traces is 0.4m. Area sampling
technique was used in sampling the joint traces on both the walls and on the
floor. As for spacing, goodness-of-fit tests were carried out on the sampled
joint traces. For all four joint clusters, gammadistribution was found to be
the best distribution to represent observed trace length distribution on both
walls and on the floor.
The method given in Kulatilake and Wu (1984a) was used to correct for censoring
error and size bias, and to estimate mean trace length on an infinite 2D exposure
using the observed trace data from a finite 2D exposure. Differences upto sixty
percent were found between the observed mean trace length and the estimated mean
trace length on infinite 2D exposures. Such results imply the importance of the
correction. An attempt was made to express the probability density function for
the trace length distribution on 2D infinite exposure, f(1), from the probability
density function obtained for observed traces on the 2D finite exposure, g(1).
It was necessary to incorporate the estimated corrected mean trace length and

1054
correction factors in f(1) to account for the effect of size, censoring and the
truncation biases on the probability distribution. For example, the following
relationship was obtained between f(1) and g(1) for analysis based on the wall
data

Kg( l)[wh + [x(wsinO + hcosO ) ]


t(O -
[wh + l(wsinO + hcosO)]

where w and h are the width and the height of the wall, is the corrected mean
trace length, 0A is the average apparent dip of the wall data set, 1 is the trace
length and K is a constant.
Assumingjoints as finite circular discs, joint diameter distribution for a
joint cluster was estimated from f(1) using the procedure given in Kulatilake
and Wu (1986). Diameter distributions for all joint clusters were found to be
of Gammatype. Figure 9 shows the diameter distribution obtained for cluster
4 based on wall data.

F11:I
gammaprobabilffy
nio., -. .0, . o.3
mean.t25 m S.D.- 0.62
"t
"=II 1
is,bion.
,-.
Totalsnlinelenglh-60
I [[ Standarddeviation= 0.58 m

20 30
' 0 40
D;ometer
Figure 9 The gammafit obtained for diam-
Figure 7 e exponential fic obtained for eter distribution for cluster 4 (wall
spacing along scanline difction 1 data)

Joint
cluster
xI
censored
error
and
&size
Estimate
mean
tracebias
(nofor[.
length
correcting /
correction
,ortncat Oata ontraces
on wa,s
I IOata
ontrason I

ICo,
recaon
,o,
si,.
bias hPsrbceanbi.YmddlStt;
tOwirtp;z7;asojfiendnitetr%
m]
Alumped
factor
tocorrect
for
Probability
distribution
torepresent
uncensored,
I
censoring
&truncation
error
I--]untruncated
trace
ength
on2Dnilrite
exposures
J

Geemelrlcel
i:m3allty
relationship
between
trace
fenglh
(20)
I8est theoetica!
anddiemelm{30) ecz3unl fo semp biases
Iio relesent probability
Ihe nl distribution
diameter
in 30

Figure8 Flowchartfor modellin


Jointsizein 3Dfor eachcluster

2.5 Threedimensional
Joint intensity (number
per unit volume)modeling
3Djoint intensitywasestimated
according
to thefollowing
fourmethods.
2.5.1 Method 1

For eachcluster, meanvalue for 3D joint intensity wasestimatedusingthe


following equation:

4(k,), (2)
(ku) nE(D)
where
(k,) and(k,)arerespectively
mean
linearintensityalongits mean
vector
1055
and mean3D intensity of j th joint set and (D2) is the expected or the mean
value of squaredjoint diameter. This equationwasderived basedon an equation
presented by Oda (1982). For each cluster, (D2) can be estimated based on the
diameter distribution obtained either from the wall data or from the floor data.
ValuesobtainedfromSection2.3.1 wereusedfor (k). Table 1 providesthe
results obtained. Since linear intensity follows a Poissondistribution, it is
reasonable to conclude that 3D intensity follows a Poisson distribution with
corresponding
(k). Mean3D intensity for the rock masswas obtainedthrough
equation (3).

Table . 3D meanjoint

Method 1 Method 2 Method 3 Method 4

Wall Floor Wall Floor Wall Floor Wall Floor

Cluster 1 .061 .042 .278 .190 .320 .166 .220 0.133

Cluster 2 .403 .568 .332 0.468 .344 .324 .360 .453

Cluster 3 .978 1.056 .633 0.683 .624 .461 .745 .733

Cluster 4 1.083 2.024 .796 1.266 .766 1.401 0.882 1.564

2.525 3.690 2.039 2.607 2.054 2.352 2.207 2.883

mean
3Djointintensity (3)

2.5.2 Method 2

The same procedure as for Method 1 was used for this case. However, for this
case, values obtained from Section 2.3.2 were used for (k)j in computing
Table 1 provides the results obtained.

2.5.3 Method 3

For each cluster, mean value for 3D Joint intensity was estimated using the
following derived equation:

(k) = (D)(Isinv[) (4)


where(ko); and (k,); are respectively, meanareal intensity (meannumberof joint
centers per unit area) and mean 3D intensity of J th joint set, and, E(D) and
(Isinw})are the expected or meanvalue of the diameter and I$invlrespectively.
The angle w, is the angle between the joint plane and the sampling plane on
which (ko); was estimated. Froma samplingdomain, what we can directly estimate
is the averagenumberof Joints per unit area. The procedure given in Kulatilake
and Wu (1984b) was used to estimate (ko); from the average numberof joints per
unit area. Table 1 provides the results obtained.

2.5.4 Method 4

Values of (k,)/ obtained from the previous three methodswere averaged to obtain
these estimations. Results are given in Table 1.

1056
3 JOINT SYSTEM MODELING

Section two dealt with modeling of joint geometry parameters for the chosen
statistically homogeneous region. Results obtained in Section two were used in
this section to build joint geometry networks in 3D. For each joint cluster the
following statistical models were used to generate joints in 3D: (1) Number of
joints per certain volume is Poisson distributed with mean value (k,)j (2)
Location of joint clusters in 3D is uniformly distributed (3) Orientation is
distributed according to the empirical distribution obtained for corrected data
(4) Diameter is gammadistributed with the parameter values obtained.

4 JOINT GENERATION IN 3D, PREDICTION IN 2D, AND VERIFICATION

Only one of the several verification studies performed for the largest joint set
(cluster 4) is discussed in this paper. The joints were generated in the volume
shown in Fig. 10, according to the statistical model given in Section 3 using
MonteCarlosimulation. In Fig. 10, m,xis the largestmeandiameterof the
four joint sets. In the chosen volume, the vertical plane EFGH of size (10m x
4m) (Fig. 10) was chosen to simulate the tunnel wall. For this case,
obtained from Section 2.5.4 was used. Diameter estimation obtained based on
wall data was used. Joint traces appearing on EFGHwere censored and truncated.
A truncation limit of 0.4m was used. These joint traces were used to estimate
the parameters given in Table 2. Several such simulations were repeated. Mean
predictions were computed using the results obtained through thirty simulations.
Table 2 shows a comparison between predictions and actual field data. In
statistical sense, the agreement is very good.

Table2. A comparison between the field data and


predictions on 2D from Monte Carlo simulation of
the parameters in the model for Joint cluster 4.

Parameter Field value Predicted Average pre-


range diction

Ro 0.01 0-0.02 0
R, 0.18 0.07-0.33 0.17
Rz 0.81 0.67-0.93 0.83
mean 4p 14.9 7.09-18.6 13.2 Figure 10 Volume and vertical section used
mean 8 245.1 223.3-257.0 238.8 in verification
lobs (m) 1.38 0.87-1.39 1.14
1o' (m) 0.79 0.23-1.27 0.61
Number of 40 30-49 41
joints

Note: Ro- No/N. 2' N2/N and R t - Ni/N


where
where
8 - dip direction (deg.)
- dip angle (deg.) No - expected number of joints with both ends
censored intersecting the window
lobs - mean trace length on the
finite window N - expected number of joints
with only one end
observed intersecting the window
lcoxx - corrected mean trace
N2 - expected number of joints with both ends
length
observed intersecting the window
N - No + N + Nz

5 SUMMARY AND CONCLUSIONS

Developmentof 3D stochastic joint geometry models including corrections for


samplingbiases andapplications of stereological techniquesandcheckingvalidity
of such built up models were addressed in the paper. Verification studies
performed so far has indicated the need to tryout different schemesin modeling

1057
joint geometry parameters and to consider a number of combinations of these
schemes in order to come up with a realistic 3D Joint geometry model which
provides a good comparison with field data during verification. The verification
reported in the paper shows that the modeling scheme used for the verification
has very good capability in producing 2D predictions which provide very good
agreement with 2D field data for the cluster studied. Further verification
studies for other clusters of the same site as well as for different sites are
recommended to check the validity of the suggested modeling schemes.

ACKNOWLEDGEMENTS

The Swedish Natural Science Research Council and the Arizona Mining and Mineral
Resources Research Institute under USBM Grant No. G 1194104 provided partial
financial support for this study.

REFERENCES

Arnold, K.J. 1941. On spherical probability distributions. Ph.D. dissertation,


Massachusetts Institute of Technology, Cambridge, Massachusetts: 42.
Bingham, C. 1964. Distributions on the sphere and on the projective plane.
Ph.D. dissertation, Yale University, New Haven, Connecticut: 93.
Hudson, J.A. and Priest, S.D. 1983. Discontinuity frequency in rock masses.
Int. J. Rock Mech. Min. Sci. 20: 73-89.
Karzulovic, A. and Goodman, R.E. 1985. Determination of principal joint fre-
quencies. Int. J. Rock Mech. Min. Sci. 22: 471-473.
Kulatilake, P.H.S.W. 1988 & 1989. Stochastic joint geometry modelling:
state-of-the-art. Proc. 29th U.S. Symp. on Rock Mech., Minneapolis, Minnesota:
215-229. Also invited papers at the conferences held in Switzerland, Sweden
and Mexico.
Kulatilake, P.H.S.W., Wathugala, D.N., Poulton, M. and Stephansson, O. 1990a.
Analysis of structural homogeneity of rock masses. Int. J. Engineering Geology
(in press).
Kulatilake, P.H.S.W. and Wu, T.H. 1984a. Estimation of mean trace length of
discontinuities. Rock Mech. and Rock Engineering. 17: 215-232.
Kulatilak, P.H.S.W., and Wu, T.H. 1984b. The density of discontinuity traces
in sampling windows. Int. J. Rock Mech. Min. Sci. 21:345-347.
Kulatilake, P.H.S.W. and Wu, T.H. 1986. Relation between discontinuity size and
trace length. Proc. 27th U.S. Symp. on Rock Mech., Tuscaloosa, Alabama:
130-133.
Kulatilake, P.H.S.W. andWu, T.H. and Watugala, D.N. 1990b. Probabilistic modelling
of joint orientation. Int. J. Numerical and Analytical Methods in Geomechanics
(in press).
0da, M. 1982. Fabric tensor for discontinous geological materials. Soils and
foundations. 22: 96-108.
Rouleau, A., Gale, J.E. and Baleshta, J. 1981. Fracture mapping in the ventilation
drift at Stripa: procedures and results. Research report LBL-13071, Lawrence
Berkeley Laboratory, Berkeley, California.
Sen, Z. and Kazi, A. 1984. Discontinuity spacing and RQD estimates from finite
length scanlines. Int. J. Rock Mech. Min. Sci. 21: 203-212.
Shanley, R.J. and Mahtab, M.A. 1976. Delineation and analysis of centers in
orientation data. Math. Geology. 8: 9-23.
Terzaghi, R. 1965. Sources of error in joint surveys. Geotechnique. 15: 287-304.
Wathugala,D.N., Kulatilake, P.H.S.W., Wathugala,G.W.andStephansson,
O. 1990.
A general procedureto correct samplingbias on Joint orientation using a
vector approach. Computersand Geotechnics(submitted for publication).

1058

You might also like