Rock Mass Properties For Engineering Design: R.J. Pine & J.P. Harrison

You might also like

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

Downloaded from http://qjegh.lyellcollection.

org/ at Orta Dogu Teknik Universitesi on May 21, 2016

Rock mass properties for engineering design


R.J. Pine1 & J.P. Harrison2
Camborne School of Mines, University of Exeter, Redruth, Cornwall TR11 4PN, UK
Imperial College, Exhibition Road, London SW7 2AZ, UK

Abstract of this paper, along with example usage of the classifi-


cation approach to some mine pillar examples.
he determination of rock mass properties for The classification approach has an in-built assump-

T engineering design is considered from two


perspectives. These are in-situ measurement,
including classification-based methods, and the
limitations of the classification approach.
Several measurement methods are available which
tion whose implications should be appreciated. Specifi-
cally, the same ‘score’ or rating from a classification
procedure can be obtained by quite different values of
the contributing components. The consolidation of
complex behaviour characteristics into a single rating
will give useful results, if used appropriately. The choice can lead to a loss of understanding of actual behaviour.
of different methods will depend on the nature of the This problem is examined in the second part of the paper
rock mass. The required accuracy should be considered through the use of multi-dimensional space and
realistically and, in many cases, high levels of accuracy multivariate analyses, which are compared with the
are not, in fact, necessary. Examples are included, which customary formula-based rating approach.
show how the mass strengths of mine pillars were
determined with acceptable accuracy using well known
rock mass classification schemes, modified as necessary Current practice and guidelines
to accommodate local conditions.
Rock mass classification is a widely used, economical The overall approach needed for determining rock mass
and extremely useful basis for determining properties, strength and deformability is to upscale and integrate
but there are dangers in uncritical application. The clas- results from measurements within small volumes (e.g.
sification methodology is critically examined, and the use boreholes) and exposures (surface or underground).
of multivariate statistics applied in a multi-dimensional The elements of the required approach include proper
space is considered to optimize the usefulness of the consideration of spacing and continuity of discontin-
measured data. uities within identified sets; quantification of useful
properties (these are contained within the classification
Keywords: case studies, mechanical properties, pillars, rock methods); selection of appropriate measurement
mechanics, site investigation
methods (taking account of time and cost); and
integration of the measurement within the geological
framework of the site (usually recognizing different
Introduction geomechanical domains within a geological overview).

Rock mass properties are intrinsically difficult to Scale effects


determine. The scale of investigation, typified by the
representative elementary volume (REV) (e.g. Hudson & The common quantification problem for all rock masses
Harrison 1997), can be much larger than manageable is scale dependency. Unless the mass is very closely
investigation techniques can accommodate. Typically jointed or almost completely unjointed, the behaviour is
the testing of mass strength and deformability needs to not uniform and is affected by the orientation, spacing
apply load and monitor the response of different sets and persistence of the discontinuities (Hoek & Brown
of discontinuities within a local volume of, for example, 1980). Some of these factors are recognized in the
10 to 100 m3. This may need to be repeated at many common rock mass classification systems (Barton et al.
locations throughout a site with only moderately 1974; Bieniawski 1989), but not all are accounted for
variable geology. Geophysical methods are capable of in a complete and consistent manner. A particular
sensing the deformability of large volumes of rock mass, difficulty arises with persistence of the discontinuities
but cannot be used alone and cannot provide strength due to both limited opportunity for gathering relevant
data. data and for subsequent modelling. The most developed
Current practice relies heavily on rock mass classifi- treatment of persistence in modelling is in discrete
cation methods which alone accommodate the problems fracture networks, typically applied to hydrogeological
of scale effects at reasonable cost. However, other problems. In the program FracMan (Dershowitz et al.
methods are available for specific circumstances. These 1998, see also details at http://fracman.golder.com.)
methods are compared in terms of accuracy, appli- discontinuity data sampled from exposures in out-
cations, indicative costs and limitations, in the first part crops (two-dimensional surface) and borehole
Quarterly Journal of Engineering Geology and Hydrogeology, 36, 5–16 1470-9236/03 $15.00  2003 Geological Society of London
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

6 PINE & HARRISON

Table 1. Suggested levels of accuracy required for rock mass properties in different applications.

Application Strength Deformability Permeability/hydraulic


conductivity

Mining pillars shafts (25%) total inflow (50%) rates


walls (10%)
roofs

Civil excavations tunnels (25%) tunnels (25%) total inflow rates (50%)
caverns (10%) caverns (25%)
pressure tunnels (10%) total leakage (25%)
and shafts rates

Nuclear/radwaste mass transport (factor


of 102 to 102)

Oil and gas borehole stability (10%) reservoir subsidence connectivity/


(25%) transmissivity (50%)

Civil foundations/pile sockets settlement (25%)

(one-dimensional linear sampling) can be used to syn- Required accuracy of rock mass properties
thesize a three-dimensional stochastic discontinuity and available measurement methods
model which shares the statistics of the samples and
allows for specific (deterministic) discontinuities to be With these difficulties in mind it is important to ensure
incorporated. This can be tested by ‘drilling’ further that a realistic approach is taken to the required
boreholes or creating further ‘excavated’ surfaces in the accuracy of rock mass properties and that the optimum
model, from which the discontinuity statistics can be methods are used wherever possible. Table 1 includes
re-calculated and compared with the original data. This some typical applications and an opinion on the
approach has been widely used in the interpretation of required accuracy. Whilst it may appear desirable to
fluid flow in fractured rock for potential radioactive achieve high levels of accuracy, the inherent spatial
waste disposal sites (e.g. Heathcote et al. 1996) and, variability of jointed rock masses means that this can be
more recently, linked to effects of in-situ stress, in illusory. Most practical projects can be managed within
hydrocarbon reservoirs (Jolly et al. 2000). the limits suggested.
In geomechanical models, as opposed to hydro- The main methods available for determining rock
geological models, the difficulty of creating true stress mass mechanical properties, and their inputs and
interaction between discontinuities in a stochastic measurement limitations are shown in Table 2. The
model is severe and requires substantial computational methods are listed in approximate order of cost (lower
capacity, which can seldom be justified in typical cost at the top). The terms GSI, RMR and Q are the
Geological Strength Index, Rock Mass Rating, and
civil and mining engineering projects. However, there
Rock Mass Quality, respectively. The limitations of the
are programs such as UDEC and 3DEC (see details
RMR (or GSI) and Q systems are examined in detail in
at http://www.itascacg.com/udec.html), which allow
the next main section. The properties Em and cm are the
simplified incorporation of joint orientations, spacings
rock mass values of deformation modulus and com-
and persistence, with reasonably sophisticated treatment pressive strength respectively. These designations and
of mechanical discontinuity and block behaviour. parameters are all discussed in detail by Hoek et al.
Very commonly in civil and mining engineering (1995). The terms [field], Vp and Vs are the (in-situ) field
projects, recourse is made to simplified models which use stresses and P- and S-wave acoustic velocities within the
rock mass classification to generate uniform, but poss- rock mass. The term ‘geometry’ applies to the openings
ibly anisotropic, rock mass properties. It is important within which various measurements are made. The
when using such approaches to fully appreciate the scale existence of an opening means that the in-situ stresses
effect (e.g. reducing unit mass strength with increasing are disturbed and this must be accounted for in the
volume up to a REV). This is probably best understood measurement technique.
for mine pillars, due to the wide range and large The methods shown in Table 2 are not equally appli-
numbers of such pillars in existence. Some examples of cable in all circumstances. The technical limitations
this are presented below. can be related to the strength of the intact rock (e.g.
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

ROCK MASS PROPERTIES 7

Table 2. Available methods contributing to the determination of rock mass mechanical properties

Method Inputs and limitations

(a) Strength and deformability


Rock mass classification Em, cm=f(GSI/RMR/Q)
Laboratory tests E,  =f(specimen condition, equipment, procedure)
Pressuremeter and borehole jack Em=f(borehole condition, equipment, procedure)
Back analysis of trial excavations Em=f([field], geometry, deformations)
cm=f([field], geometry, damage)
(b) Deformability only
Seismic methods (e.g. petite sismique) Em =f(Vp, Vs, access, rock mass condition)
Large compression (plate jacking) tests Em=f(geometry, deformations, access, rock mass condition)
Pressure test chambers Em=f(geometry, deformations, access, rock mass condition)

GSI, Geological Strength Index; RMR, Rock Mass Rating; Q, Rock Mass Quality; Em and cm, rock mass values of
deformation modulus and compressive strength, respectively; [field], Vp and Vs, in-situ field stresses and P- and S-wave
acoustic velocities within the rock mass, respectively.

Table 3. Applicability of methods depending on UCS and RQD.

Low UCS Medium UCS High UCS


(<25 MPa) (25–100 MPa) (>100 MPa)

High RQD (>75%) Laboratory tests Laboratory tests Laboratory tests


Seismics Seismics Seismics
Pressuremeter
Borehole jack

Medium RQD (50–75%) Laboratory tests Classification Classification


Pressuremeter Seismics Seismics
Borehole jack Plate jacking tests Plate jacking tests
Back analysis Back analysis
Pressure test chamber Pressure test chamber

Low RQD (<50%) Plate jacking tests Classification Classification


Seismics Seismics
Plate jacking tests Plate jacking tests
Back analysis Back analysis
Pressure test chamber Pressure test chamber

unconfined compressive strength – UCS) and the inten- allowing higher extraction. In civil engineering projects
sity of jointing. A useful and common measure for the it is easier to justify the very safe approach and conse-
latter is Rock Quality Designation (RQD). quently the rock mass is not tested to the same extent.
In Table 3 the UCS and RQD are each divided into We conclude that methods that work well for mine
high, medium and low categories, based on divisions pillars are a sound basis for rock masses in general.
within the RMR system, giving a 33 matrix of suggested The first case example is at Middleton Mine,
applicability. The allocation of methods in this matrix Derbyshire, an underground room and pillar limestone
should be seen as the most likely to yield acceptable results mine in the UK. The key design issues are the stability of
and the category numerical values are approximate. pillars and roofs. Because of the large number of existing
Again, within each category the methods are listed in pillars, considerable insight can be gained from both
approximate order of cost (lower costs at the top). rock mass classification and back analysis.
Using the rock mass classification approach, a typical
Mine pillar case examples RMR value of 77 was determined. The UCS of the intact
rock is reasonably uniform with a value of about 60 MPa.
Mine pillars provide a classic example of where it is Assuming a value for mi, the intact Hoek and Brown
important to determine mass strength values as accu- parameter m, in the range 10 to 15 the uniaxial compressive
rately as possible. The pressures of mining economics strength of the rock mass cm was determined as 18 MPa
are such that a balance must be struck between leav- using the approach presented by Hoek et al. (1995).
ing very safe pillars with high factors of safety and For the same rock mass, back analyses were made of
adequately safe pillars with lower factors of safety, an extensively pillared area, where the tributary area
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

8 PINE & HARRISON

Fig. 1. Limestone pillar with RMR value of 70. Fig. 2. Limestone pillar with RMR value of 60.

analysis (Brady & Brown 1993) was probably reliable in tion modulus of up to 30% lower than for the higher
determining imposed stresses. Making adjustments for RMR values and this could affect the load distribution
pillar width to height ratios and pillar volumes (Hoek & between pillars (but not by as much as 30%). The stiffer,
Brown 1980) a minimum mass strength of 12 MPa was high RMR pillars would attract more than average
required. This indicated that the pillar factors of safety loading as the pillared area became more extensive and
were typically about 1.5, which is reasonable in this this may have contributed to more spalling.
case since the pillars appeared to be generally in good The second case example is taken from Roberts et al.
condition. This example broadly confirms the validity of (1998) at Doe Run Mine in Missouri. Based on a
rock mass classification in determining mass strengths. consistent initial rock mass condition observed in the
However, looking closely at individual pillars there mine pillars, they developed an approach which is
were some interesting variations. In some pillars where simpler in terms of classification but more sophisticated
RMR was c. 70, with a higher than average inferred in treating observed pillar conditions. This can be
mass strength, the pillar profiles were very ragged adapted to other mines for analysis and design of pillar
(Fig. 1). In adjacent pillars where the RMR was c. 60, development and pillar mining. It requires location-
with lower inferred strengths, the pillar profiles were specific classification and back analysis input.
much smoother (Fig. 2) and apparently more stable. The approach to pillar condition observation is shown
The main difference in the RMR values was due to the in Figure 3, which includes a range of six visual ratings
intensity and continuity of the jointing. The more widely from 1 (no stress-induced fracturing) to 6 (failed pillar).
spaced jointing (RMR=70) gave rise to larger individual The pillar stresses were typically determined using a
blocks than for the closer spaced jointing (RMR=60). displacement discontinuity program and the strength
When blocks peeled off, mostly during pillar formation model was based on a common pillar formula, which
(blasting and barring down), the missing larger blocks incorporates a strength scale effect:
gave a worse appearance. Since the pillars were large
(161616 m) the loss of quite large individual blocks Mass strength = K  UCS  W1/2  H3/4 (1)
had little effect on the overall pillar integrity, but could
have been serious in smaller pillars. At an even larger Here, K is a constant, W and H are pillar width and
scale, unaffected by the wider joint spacing, the higher height, respectively.
strength inferred by the higher RMR values would A site-specific relationship was developed between
clearly be applicable. pillar condition, pillar stress/UCS ratio and pillar height.
Another possible factor is that the more jointed pillars The pillar widths vary between about 10 and 20 m.
with the lower RMR values could have a mass deforma- Typically for pillars where the stress is less than 20% of
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

ROCK MASS PROPERTIES 9

were estimated at 53 MPa. This was consistent with


observations. The strength including the bedding in the
classifications was much lower. In this case an uncritical
use of classification would have been misleading.
These pillar examples demonstrate that rock mass
classification can be used to provide acceptably accurate
empirical assessments of mass strength, provided that
local conditions are well understood and appropriate
adjustments made to the standard form of the classifi-
cation schemes. How these adjustments should be made
objectively is a task for which there is little guidance,
and hence is an area that requires further investigation.
One approach to this problem is to make observations
of the type utilized above and, through mathematical
analysis, objectively determine both the critical param-
eters and a bespoke method of combining them to
obtain a classification value. This is described below in
terms of the application of multivariate statistics.

Objective customization of rock


mass classification schemes
As shown above, rock mass classification schemes are a
common and useful part of the design process, especially
at the preliminary stage when data are probably both
limited and of poor quality. These schemes are attractive
in such circumstances because they seem to encapsulate
the entire rock engineering problem in only a few
parameters.
Fig. 3. Pillar condition observation at Doe Run mine (after However, the general nature of these schemes means
Roberts et al. 1998). that they should not be applied uncritically and, to this
end, perhaps we should ask (i) why are there so many
the UCS, the pillar condition rating is 2 (spalling on different rock mass classification schemes? and (ii)
corners, etc.). Where the stress is greater than 30% of the why do these schemes have different formulations, and
UCS, pillar condition rating is 4 or higher (extensive include different parameters? Ultimately, of course, we
fracturing to failure). Pillar height is critical for the more ought to determine which of the various schemes avail-
highly stressed pillars. Note that the only geologically able to us is the best for a particular application, both in
controlled variable is the UCS; in all other respects terms of the parameters used and the results obtained.
(principally joint intensity, orientation and condition) It is possible that, as indicated in the previous section,
the rock mass is presumed to be consistent. Our inter- a customary scheme may not be sufficiently discrimina-
pretation of Figure 3, which is potentially useful for tory in a given application to be of practical benefit. In
pillars elsewhere, is that rating 1 indicates a factor of such an event, it is appropriate that we should use
safety of about 1.5 or more and rating 4 is close to 1.0 records of actual performance in order to help modify
(effective failure). such a scheme and thereby improve its usefulness. Here,
A third case example was presented by Trueman we use a synthetic example to show how the customary
et al.(1992), based on pillar analysis at Tara Mine, use of rock mass classification schemes may be mis-
Ireland. A pillar visually close to failure, with significant leading, and how, through the use of multivariate
spalling, was analysed for applied stress and strength statistics, we can improve the situation. However, this
inferred from both RMR and Q classifications. The approach will require further case studies relating the
pillar was in a strongly bedded limestone and loaded rock mass behaviour to the individual parameters before
sub-normal to bedding. Two values of both RMR and Q it can be used routinely.
were determined based on whether the bedding dis-
continuities were included or not. From these values Functional form of common rock mass
estimates of mass strength were made in a similar classification schemes
manner to that described above in the first example. The
average imposed stress on the pillar was estimated at Two of the most widely used schemes are Rock Mass
54 MPa, and the mass strengths without the bedding Rating (RMR) (Bieniawski 1989) and Rock Mass
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

10 PINE & HARRISON

Table 4. RMR and Q parameters. a technique for casting these schemes into a two-
dimensional form. One simple way of doing this is to
RMR parameters
consider the situation commonly found on a specific site,
Rstrength intact rock strength rating
RRQD RQD rating where some of the parameters will not vary appreciably
Rspacing mean discontinuity set rating across the region of interest and can therefore be con-
Rcondition weathering condition of discontinuity sidered as constant values. For example, a tunnel
surfaces through a limestone rock mass containing three discon-
Rwater groundwater flow and pressure rating
adjustments adjustments to account for specific tinuity sets could be assigned a value of Jn=9, and this
types of construction and rock mass may not change over the length of the tunnel. Thus,
anisotropy an appropriate two-parameter (i.e. graphically three-
Q parameters dimensional) case is that where only two of the RMR
RQD RQD percentage and Q parameters vary at a given location.
Jn number corresponding to the number Now, to illustrate the geometry of RMR in two
of discontinuity sets in the rock mass dimensions, let us assume we are working at a location
Jr number relating to discontinuity where only the rock strength and RQD vary. In
surface roughness
this case, we can rewrite the RMR formula, given by
Ja alteration (weathering) condition of
discontinuity surfaces equation [2], as
Jw groundwater flow and pressure rating
SRF Stress Reduction Factor, which RMR = Rstrength + RRQD + ~Rspacing + Rcondition + Rwater +
accounts for rock strength, in-situ stress
and engineering environment adjustments! (4)
or, more concisely, as
Quality (Q) (Barton et al. 1974). These two schemes each RMR = Rstrength + RRQD + constants (5)
seek to generate a single number that summarizes the
engineering quality and behaviour of the rock mass, and Similarly for Q in two dimensions, let us examine the
each does so on the basis of a defining formula that case where the mechanical properties of the rock mass
combines a number of fundamental rock mass proper- (e.g. degree of weathering, groundwater conditions) do
ties. The two schemes have the well-known defining not change, but that the degree of fracturing (i.e. RQD
formulae of and Jn) does. On this basis the formula for Q, given by
equation [3], can be written as
RMR = Rstrength + RRQD + Rspacing + Rcondition +

and
Rwater + adjustments (2) Q=
RQD
Jn
 S Ja
Jr

Jw
SRF
D =
RQD
Jn
 constants (6)

Jr Jw We can think of equations [5] and [6] as the site-


RQD
Q=   (3) specific relations
Jn Ja SRF
RMR# = RMR  constants = Rstrength + RRQD (7)
These various parameters are briefly described in
Table 4. and
Although these formulae are simple to apply, their use
obscures the fact that, mathematically, there is an under- Q RQD
Q# = = (8)
lying geometry associated with the two schemes, which is constants Jn
defined by the formulae themselves. Indeed, once we
As these relations are equations in two variables, they
investigate these schemes in terms of their geometry, we
can be easily plotted in order to examine their graphical
see that the engineering behaviour they predict is highly
form. Thus, Figure 4 shows that curves of equal RMR#
constrained by these formulae.
value are parallel straight lines with a negative slope,
and Figure 5 shows that curves of equal Q# are straight
RMR and Q in graphical form lines with different positive gradients, but which all pass
through the origin. The different form of these graphs is
RMR and Q rely on five and six parameters respectively, due entirely to the forms of the equations used to
which means in terms of geometry they exist in six- and generate them.
seven-dimensional spaces (one additional dimension is In addition to the general disposition of lines of equal
required for the resulting classification value). Because classification values, Figures 4 and 5 also show, on the
such high dimensional spaces are not easily visualized on lines representing RMR#=18 and Q#=8 respectively,
the two-dimensional medium of paper, we need to find various parameter pairs that lie on these particular lines.
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

ROCK MASS PROPERTIES 11

strength, such that closely spaced discontinuities are


associated with weaker rocks, and vice versa), and that
as such the extreme conditions considered above are
unlikely to occur. Conversely, we suggest that although
such dependence between parameters may occasionally
be seen in the field, it cannot currently be taken as a
universal. Indeed, before accepting the hypothesis that
interdependence between parameters exists (or any other
such hypotheses), there should be both theoretical and
substantial documented field evidence to support it.
The graphical analysis presented in Figure 4 high-
lights a deficiency that exists in all classification schemes
Fig. 4. Plot of the two-dimensional relation RMR#=
Rstrength+RRQD. that rely on a formula to combine a number of classifi-
cation parameters into a single value: namely, so much
information is lost (i.e. knowledge of the individual
parameter values) that similarity between rock masses
can be implied when none exists. This problem of ‘false
similarity’, or lack of discrimination, is one that most
geotechnical engineers are familiar with. A common, but
invalid, means of dealing with it is by resorting to
unwarranted precision in classification values (e.g. by
suggesting that Q values of 3.15 and 3.20 represent
different behaviour).
Having identified that this problem is due to the use of
a singlevalued function in computation of a classifi-
Fig. 5. Plot of the two-dimensional relation Q#=RQD/Jn. cation value, as a means of surmounting it we suggest
that methods that embrace the underlying geometry of
the scheme be considered instead. Indeed, the recently
These parameter pairs represent RQD and strength introduced Geological Strength Index (Hoek 1999)
values in the first case, and discontinuity set geometry adopts just this approach, and is defined only in a
and RQD in the second. These pairs deserve close graphical form. Unfortunately for the RMR and Q
inspection, especially in the context of the rock masses schemes, however, this means that – as noted above – we
they purport to represent. will need to operate in spaces of six or seven dimensions,
In Figure 4, the line representing RMR#=18 encom- which is all but impossible to manage graphically. It is,
passes rock masses that vary from highly fractured, however, practical to work computationally in such
extremely strong rock at the upper left end (where the high-dimension spaces, using the concepts associated
parameter values are RQD=0% and compressive with multivariate statistics, particularly discriminant
strength of 250 MPa), through to extremely weak, intact analysis and principal components analysis (see, for
rock (RQD=100% and zero compressive strength) at the example, Davis 1986). By these means we may be able to
lower right end. Hence, on the basis of a rock mass identify, quantify and hence use, regions of similar
classification value of RMR#=18, we can see that highly engineering behaviour in the appropriate parameter
fractured, extremely strong rocks are regarded as being space, without resorting to a specific defining function.
identical in engineering behaviour to extremely weak
intact rocks; this is, however, clearly not true. Similarly, Multivariate approach to rock mass
the line representing Q#=8 in Figure 5 encompasses rock
classification
masses that range from those containing a single, closely
spaced discontinuity set (Jn=2, RQD=16%), through Multivariate statistics allows us to explore the inter-
to rock masses that contain three discontinuity sets dependence between parameters in a model through the
together with a randomly orientated component, all very use of their variances and covariances. Although these
widely spaced (Jn=12, RQD=96%). As a consequence, statistical descriptors are easily computed for datasets, it
we see that both the RMR and Q schemes consider rock is through a geometrical approach that their relation to
masses of completely different geometry (in particular, the data themselves is most readily appreciated. In
anisotropy) to have identical engineering behaviour. essence, and as will be shown below, variance and
Of course, it could be argued that, in reality, the covariance describe the geometrical attributes of both
various parameters used in a classification scheme are spread and preferred orientation of a group of data
not mutually independent (e.g. discontinuity spacing about their mean. Furthermore, this link with the geom-
tends to be inversely proportional to intact rock etry of a dataset allows these statistical descriptors to be
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

12 PINE & HARRISON

used to compute distance from the centroid of a group


to any particular data point on the basis of the point’s
similarity to the group. Informally, this means that data
that fit well within a group will have smaller distances
from the mean than those data that do not fit well. In
effect, the geometry of the parameter space is distorted
so that, in two dimensions for example, contours of
equal distance from the group’s mean are elliptical,
rather than circular. In the context of rock mass classi-
fication, this means that we can use observed similarity
in engineering behaviour to identify relations between
the classification parameters, rather than use a pre-
defined formula to identify similarity of behaviour.
The basis of this technique is the use of the
Mahalanobis distance (see Leti 1979), which is a
generalization of the more familiar Euclidean distance,
Fig. 6. Example classification data.
to determine distance from the mean of a group to
any data point in the group. In general terms, the
terms in the identity matrix I is to prevent the for-
Mahalanobis distance is a measure of the coherence of
mation of a link between the x- and y- coordinate values.
data groups, such that small values indicate a close knit
In terms of multivariate statistics, this represents
group, and large values a dispersed group, or data that
the case when the x- and y- coordinates are independent,
may more properly be described as forming more than
a case that is also indicated by the covariance being
one group. Here, we consider a group to be data
zero.
represented by the components of rock mass classifi-
This association between the off-diagonal elements of
cation schemes within geological domains. The concept
the matrix I and zero covariance leads to the notion of a
of the Mahalanobis distance is most clearly described
more general distance, whereby the matrix I in equation
through the use of a two-dimensional numerical
[12] is replaced by one that is based on the variances and
example. We begin with Pythagoras’ theorem, which
covariances of the data. One such distance is the gener-
shows that the square of the Euclidean distance between
alized, or Mahalanobis, distance (Leti 1979), and the
two points (x1, y1) and (x2, y2) is computed as
square of this is computed as
d E2 = ~x1  x2!2 + ~y1  y2!2. (9)
d 2M = x̄ , C  1 , x̄T (14)
If we assume that the second of these points represents
the mean of a dataset whose coordinates are~x̄,ȳ!, then where C is the covariance matrix associated with the
the Euclidean distance of the first point from the mean is data, and C1 is its inverse. Thus, the data themselves
are used to generate the matrix that is used to compute
simply given by
their individual distances from their group mean.
In general, the covariance matrix is computed as
d E2 = ~x1  x̄!2 + ~y1  ȳ!2 (10)

Although equation [10] represents a basic approach to 1


C= ~x̄T , x̄!, (15)
computing Euclidean distance, a more mathematically n1
convenient method is to use the matrix form
where n is the number of data, and the matrix x̄ has n
d E2 = @~x1  x̄!~y1  ȳ! #
F GF
10
01
~x1  x̄!
~y1  ȳ!
G (11)
rows and p columns, with p being the number of
dimensions to the data. The covariance matrix is square,
of size pp. As with equation [12], the presence of the
which can written in abbreviated form as overbar indicates that x̄ is a centroid-adjusted matrix,
which means that each of its columns has a mean value
of zero. Writing the covariance matrix in this form
d E2 = x̄ , I , x̄T (12)
shows how readily it may be computed for the case of
If equation [11] is multiplied out in full, then before data with a large number of dimensions (for example,
simplification we find that the six or seven dimensions required with a rock mass
classification scheme).
d E2 = ~x1  x̄!~x1  x̄!  1 + ~y1  ȳ!~x1  x̄!  0 + Rather than the Mahalanobis distances associated
~x1  x̄!~y1  ȳ!  0 + ~y1  ȳ!~y1  ȳ!  1 (13) with the data themselves, it is further computations
based on the concept of the Mahalanobis distance and,
which shows that the presence of the off-diagonal zero in particular, the matrix C1, that are of particular
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

ROCK MASS PROPERTIES 13

interest here. First, it is reasonably straightforward (at two columns of Table 5, with the centroid-adjusted data
least in two dimensions) to compute and plot contours values shown in columns three and four. The centroid
of equal Mahalanobis distance, which will reveal those adjustment is simply made by subtraction of the mean
data that are outliers to the group as a whole. Secondly, values from the individual values in each of columns one
and most importantly, these contours take the form of and two. The precision of these data is much greater
ellipses whose major and minor axes are aligned with the than could (or should) be obtained in practice, but is
eigenvectors of the inverted covariance matrix, with the used here for the purposes of illustration. The centroid-
lengths of the axes being proportional to the eigenvalues adjusted matrix for these data is therefore
of this matrix. As the shape and orientation of the data
group may be defined in terms of the eigenvectors and
eigenvalues, they may be used to derive a relation
between the group parameters themselves. With modern
computer software it is easy to extract the eigenvalues
and eigenvectors of the inverted covariance matrix,
which helps bring this approach within reach of non-
mathematicians.
We now give an example of how this technique can be
used in the interpretation of rock mass classification
data. Let us assume that a number of RMR# data pairs
(Fig. 6) that represent similar engineering behaviour
have been identified. These values are given in the first and the transpose of which is

x̄T = F 0.19  0.61  0.51


 1.3 0.5
0.79 0.09 0.59 0.49  0.21  0.71  0.11
0.3  1.4  1.2 0.1  0.2 0.2 1.7 1.3
G (17)

Multiplying these two matrices together leads to to obtain, as the vector X, the eigenvectors of the system.
Hence, for =0.792, we have
x̄T , x = F 2.449  3.35
 3.35 10.1
G (18)

from which, using equation [15], the covariance F 5.933 2.2313


2.2313 0.837
GF G x1
x2 = 0 (22)
matrix is given by
from which we derive
C=
1
(10  1) 
F
2.449  3.35
3.35 10.1
=
 0.372
G F
0.272  0.372
1.122
G x1 =  0.376x2 (23)
(19)
In this case x1 and x2 represent the centroid-adjusted
with its inverse being parameters RRQD and Rstrength respectively, and so equa-
tion [23] defines the slope of the major axis of an ellipse
C1 = F 6.7272 2.2313
2.2313 1.6312
G (20) that passes through the point (8.61, 9.50), which is the
mean of the raw data as given in Table 5. The equation
This matrix is used in equation [14], with each of the of this axis, which in effect is a best-fit relation between
ten pairs of centroid-adjusted data in turn, to compute the parameters RRQD and Rstrength, is therefore found
dM2, the squares of the Mahalanobis distances, for all of to be
the data pairs in Table 5: these and their square roots are
given in columns 5 and 6 of the table. A property of the Rstrength + 2.66RRQD = 32.4 (24)
Mahalanobis distance, and one that forms a useful
check of the arithmetic, is that the sum of the squared and is shown, together with the raw data, in Figure 7.
Mahalanobis distances should equal p(n1): for 10 Using the eigenvalue =7.566, an equation of similar
two-dimensional data this value is 18, which is the value form may be obtained for the minor axis, but this
computed at the foot of column 5 of the table. equation is of no practical interest in the context of
Continuing with the analysis, using a standard ap- determining a relation between the parameters.
proach of equating the determinant of matrix @C1I# Figure 7 also shows two contours of equal
to zero, the eigenvalues of C1 are obtained as =0.792 Mahalanobis distance, which have been plotted with the
and 7.566. These are then used, in turn, in the equation help of equation [14]. Generally, for multi-dimensional
data it is not possible to derive explicit formulae for
[C  1  I]X = 0 (21) these contours, and so points on a contour have to be
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

14 PINE & HARRISON

Table 5. Example multivariate analysis of RMR# parameters.

raw data centroid adjusted Mahalanobis distance


RRQD Rstrength RRQD Rstrength (dM)2 dM

8.8 8.2 0.19 1.3 1.90 1.38


8.0 10.0 0.61 0.5 1.55 1.24
8.1 9.8 0.51 0.3 1.21 1.10
9.4 8.1 0.79 1.4 2.46 1.57
8.7 8.3 0.09 1.2 1.92 1.39
9.2 9.6 0.59 0.1 2.62 1.62
9.1 9.3 0.49 0.2 1.24 1.11
8.4 9.7 0.21 0.2 0.17 0.42
7.9 11.2 0.71 1.7 2.72 1.65
8.5 10.8 0.11 1.3 2.20 1.48
means: 8.61 9.50 0.00 0.00 sum=18.0

determined either by trial-and-error or by interpolation tation of the mathematics is clearer if we begin by


from a grid of computed distances. For two-dimensional writing equation [14] in the form
data, however, it is possible to determine points on the
ellipse directly. One way to do so is to make use of
the eigenvalues and eigenvectors, on the basis that the d 2M = @xM yM# F GF G
ab
bc
xM
yM (25)
matrices containing these can be multiplied together to
form the inverted covariance matrix. However, interpre- where ~xM,yM! are the coordinates of a point on the
elliptical contour. Considering this point in terms of it
being located on a radial line extending from the centre
of the ellipse, then these coordinates can be written as

~xM,yM! = ~rcos,rsin! (26)

where  is the angle to the horizontal axis subtended


by the radial line. Substituting these coordinates into
equation [25], multiplying out and rearranging, we
arrive at

d 2M
= r2 (27)
2 2
@acos  + 2bsincos + csin #

This shows how, for any particular value of , to-


gether with appropriate values of dM, a, b, and c, a value
of r can be computed, and from this values of ~xM,yM!
Fig. 7. Principal axis and ellipses of Mahalanobis distance for can be determined through the use of equation [26].
the example data. Performing these calculations for a range of angles

Table 6. Example computation of contours of Mahalanobis distance used in Figure 7.

angle , Cartesian components ellipse parameters dM=1.0 dM=1.5


degrees cos() sin() A B C radius x y radius x y

0 1.000 0.000 6.727 0.000 0.000 0.39 9.00 9.50 0.58 9.19 9.50
60 0.500 0.866 1.682 1.932 1.223 0.45 8.84 9.89 0.68 8.95 10.09
120 0.500 0.866 1.682 1.932 1.223 1.01 8.10 10.38 1.52 7.85 10.82
180 1.000 0.000 6.727 0.000 0.000 0.39 8.22 9.50 0.58 8.03 9.50
240 0.500 0.866 1.682 1.932 1.223 0.45 8.38 9.11 0.68 8.27 8.91
300 0.500 0.866 1.682 1.932 1.223 1.01 9.12 8.62 1.52 9.37 8.18
360 1.000 0.000 6.727 0.000 0.000 0.39 9.00 9.50 0.58 9.19 9.50

ellipse parameters: A=acos2, B=2bsincos, and C=csin2, where F G


ab
bc
=C1. Radius=dM/~A + B + C!1/2.
x=~radius  cos! + x̄; y=~radius  sin! + ȳ
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

ROCK MASS PROPERTIES 15

Table 7. Example computation of confidence contours used in Figure 8.

angle , =0.01, T2=2.43 =0.10, T2=1.98 =0.20, T2=1.73


degrees radius x y radius x y radius x y

0 0.94 9.55 9.50 0.76 9.37 9.50 0.67 9.28 9.50


60 1.11 9.16 10.46 0.90 9.06 10.28 0.79 9.00 10.18
120 2.47 7.38 11.64 2.00 7.61 11.24 1.75 7.73 11.02
180 0.94 7.67 9.50 0.76 7.85 9.50 0.67 7.94 9.50
240 1.11 8.06 8.54 0.90 8.16 8.72 0.79 8.22 8.82
300 2.47 9.84 7.36 2.00 9.61 7.76 1.75 9.49 7.98
360 0.94 9.55 9.50 0.76 9.37 9.50 0.67 9.28 9.50

radius=T/~A + B + C!.

between zero and 360( (Table 6), allows a contour of


equal Mahalanobis distance to be produced. It should
be noted that only a small number of example calcu-
lations are shown in this table, and that points should be
evaluated at intervals of about 5( in order to compute
the contour.
Because of the relationship between the Mahalanobis
distance and the underlying covariance matrix, the con-
tour produced when dM=s represents a curve that is s
standard deviations away from the data centroid. As
Figure 7 shows, all but three of the data in the example
lie within 1.5 standard deviations of the mean, which
gives a qualitative idea of the degree of clustering in
the group. This concept can be extended into formal
hypothesis testing (Bolch & Huang 1974), which allows
us to determine whether all of the data in a group such
as that shown in Figure 7 belong, statistically, to a Fig. 8. Significance contours around the example data.
particular engineering behaviour.
The hypothesis test used in multivariate statistics is As Figures 7 and 8 show, an approach based on multi-
Hotelling’s T2 test, where the squared Mahalanobis variate statistics allows groups of classification values that
distances are compared to critical T2 values. In the case correspond to the same engineering behaviour to be used
of testing for outliers, the critical values are computed as to generate relations that are independent of any custom-
(De Maesschalck et al. 2000) ary equation. As a result, we can see that there is scope for
exploring the wealth of information associated with rock

T2 =
~n  1!2
n
S p ~n  p  1!
 ; ,
2 2
, D (28)
mass classification data in a completely new fashion, using
the concepts outlined here. The calculations can be readily
extended to the higher dimensions required for rock mass
where the function ~+! represents the inverse beta classification schemes, and the geometry of groups of
distribution, and the parameter  is the significance level similar engineering behaviour determined in these higher-
at which we wish to test. Equation [27] and the method dimensional spaces.
outlined above are used to compute contours of sig- There are advantages and disadvantages to both
nificance level, with appropriate critical values for T 2 the customary, formula-based approach and the use
from equation [28] being used in place of dM. Table 7 of multivariate statistics, and some of these are
shows example calculations for the 1%, 10% and 20% summarized in Table 8. One issue that is not addressed
significance contours, which are plotted on Figure 8. in this table is the fact that geotechnical engineers are
This figure shows how all of the data lie within the probably most comfortable with the customary ap-
contour representing 20% significance, indicating that proach, whereas multivariate statistics may be some-
they can be taken to lie within one group. Thus, once a thing unknown. This should not stand in the way of
group of classification values that identify a specific adopting a new approach, if the new approach is found
engineering behaviour have been identified, this gives us to be superior. In the context of the approach described
a powerful and quantitative test for determining whether here, this will involve trials of the customary and
other values can be considered to possess the same proposed approaches over a range of specific rock
behaviour. engineering projects, and in each case will require that
Downloaded from http://qjegh.lyellcollection.org/ at Orta Dogu Teknik Universitesi on May 21, 2016

16 PINE & HARRISON

Table 8. Comparison of formula-based and multivariate approaches to rock mass classification.

Customary, formula-based schemes Multivariate approach

fixed algebra, and hence fixed relations algebra determined by the data themselves

similarity of engineering behaviour defined in use engineering behaviour to determine


terms of algebra algebra

single value used to quantify behaviour location and shape in n-dimensional space
classifies behaviour

some parameters could be redundant on allows identification of critical parameters for


particular engineering schemes specific engineering schemes

allows global comparisons site-, and behaviour-specific

actual rock structure performance be recorded in abso- B, Z.T. 1989. Engineering Rock Mass Classifications.
lute terms of serviceability, rather than simply as a Wiley, New York.
deviation from the behaviour predicted on the basis of a B, B.W. & H, C.J. 1974. Multivariate statistical
methods for business and economics. Prentice-Hall, New
classification assessment. It is only through such trials Jersey.
and records that our use of rock mass classification B, B.H.G. & B, E.T. 1993. Rock mechanics for
schemes can be improved. underground mining (2nd edn). Chapman & Hall, London.
D, J.C. 1986. Statistics and data analysis in geology (2nd
edn). Wiley, New York.
Conclusions D M, R., J-R, D. & M, D.L.
2000. The Mahalanobis distance. Chemometrics and Intel-
Rock mass properties can be determined at reasonable ligent Laboratory Systems, 50, 1–18.
cost provided the right method is used and a realistic D, W., L, G., G, J. & L, P.R. 1998.
approach is taken to accuracy (Tables 1–3). It is clearly FracMan: Interactive Discrete Feature Data Analysis,
Geometric Modelling, and Exploration Simulation. User
important also to make the most appropriate measure- Documentation. Golder Associates Inc., Seattle,
ments within an investigation, modelling and design Washington.
framework. H, J.A., J, M.A. & H, A.W. 1996. Mod-
Rock mass classification is the most widely applicable elling groundwater flow in the Sellafield area. Quarterly
method for rock masses with medium to high UCS and Journal of Engineering Geology, 29, S59–S81.
H, E. 1999. Putting numbers into geology – an engineer’s
the strengths inferred from classification have been viewpoint. Second Glossop Lecture, November 1998.
validated by some of the mine pillar examples presented. Quarterly Journal of Engineering Geology, 32, 1–19.
However, care must be taken with most rock masses to H, E. & B, E.T. 1980. Underground excavations in
ensure that scale effects and discontinuity anisotropy rock. IMM, London.
effects are properly considered. H, E., K, P.K. & B, W.F. 1995. Support of
underground excavations in hard rock. A.A. Balkema,
The application of rock mass classification methods Rotterdam.
can be improved, and some thought should always be H, J.A. & H, J.P. 1997. Engineering rock me-
given to the variability and nature of the key parameters chanics, an introduction to the principles. Pergamon Press,
used in the classification. In specific cases particular Elsevier Science, Oxford.
parameters (e.g. RQD and UCS – Figures 4, 6 & 7) will J, R.J.H., W, L. & P, R.J. 2000. ‘Stress-Sensitive
have a dominant effect in characterizing the rock mass Fracture-Flow Modelling in Fractured Reservoirs.’ Paper
presented at the SPE International Petroleum Conference
variability. The use of multivariate statistics applied to and Exhibition, Villahermosa, Mexico.
rock mass characterization data, as described in this L, G. 1979. The distances in descriptive statistics. Metron,
paper, provides a way to maximize the usefulness of this 37, 3–52.
type of data in design models. R, D.P., L, W.L. & Y, T.R. 1998. Pillar
extraction at the Doe Run Company, 1991–1998.
Proceedings of the Conference ‘The Mining Cycle’,
References Australia. IMM, Mt Isa, 227–233.
T, R., P, R.J., T, D.B., V, P. & F, R.
B, N.R., L, R. & L, J. 1974. Engineering 1992. Rock mass characterisation at Tara zinc-lead mine,
classification of rock masses for the design of tunnel Eire. Proceedings of the ISRM International Symposium
support. Rock Mechanics, 6, 189–239. on Rock Characterisation (Eurock ‘92), 251–256.

Received 25 May 2001; accepted 5 April 2002.

You might also like