International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

Technical Note

Likelihood statistic for interpretation of the stability graph for

open stope design
F.T. Suorinenia,*, P.K. Kaisera, D.D. Tannantb
b

a
Mirarco/Geomechanics Research Centre, Laurentian University, Sudbury, Ont., Canada P3E 2C6
Department of Civil and Environmental Engineering, School of Mining and Petroleum Engineering, University of Alberta, Edmonton, Alta.,
Canada T6G 2G7

Accepted 20 May 2001

1. Introduction
The stability graph method for open stope design is an
empirical method, and its interpretation is highly
subjective. Subjective interpretation of the stability
graph has resulted in unknown risks from human bias
and inherent errors. Users of the stability graph method
for open stope design are for example, given the wrong
impression that if a stope plots in the stable zone, that
stope is denitively stable and its performance in service
presents no risk of instability. Statistical tools exist that
can be applied to interpret the stability graph and
signicantly minimize the subjectivity in the stability
graph method without making it seem more rigorous
than it is currently perceived. This paper identies the
Baysian likelihood method as a powerful tool for a
statistical interpretation of the stability graph, and uses
the extended database based on the Potvin [1] calibrated
stability graph factors to illustrate the method and its
benets. Mathews and his co-workers [2] in Golder
Associates introduced the stability graph method of
open stope design in 1980. The stability graph is a plot
of a stability number N against a shape factor HR. The
stability number and shape factor are dened in Eqs. (1)
and (3) respectively:
N 0 Q0 ABC;
Q0

RQD Jr
Jw ;
Jn Ja

1
2

where Q0 is the modied tunnelling quality index [2]. Q0

is obtained from Barton et al. [3] rockmass quality index
*Corresponding author. Tel.: +1-705-675-1151; fax: +1-705-6754866.
E-mail address: [email protected] (F.T. Suorineni).

Q by setting the stress reduction factor SRF to 1 as

shown in Eq. (2). The water reduction factor Jw is often
set to one in Canadian conditions (dry hard rock
underground mines). In other mining environments,
where this is not the case, the relevant Jw values should
be applied. The parameters A, B and C are the stress
factor, joint orientation factor and gravity factor
respectively. The shape factor HR is dened as the
hydraulic radius:
HR

Area
:
Perimeter

Fig. 1 is the stability graph. The stability graph is

divided into three zones labeled stable, unstable and
caving zones. These zones were originally dened by
visually tting boundaries between clusters of data
points representing stable and unstable stope surfaces,
and between the unstable and caved stope surfaces. The
zone enclosed by the boundaries is the transition zone.
The transition zone contains a mixture of stable,
unstable and caved stopes. A stope surface that plots
in this zone may be stable, unstable or cave in service.
The transition boundaries in Fig. 1 are eye-balled.
The problem with visual demarcation of zones is their
subjective nature and reproducibility in future analysis
and the unknown risk associated with inherent errors.
Since 1980, the boundaries originally dened by
Mathews et al. have changed considerably (see [4]).
The changes are ascribed to the following factors:
*
*

accumulation of more data with time,

re-denition of the stability number by various
authors,
calibration of the boundaries to specic local mine
site conditions,

1365-1609/01/$ - see front matter r 2001 Elsevier Science Ltd. All rights reserved.
PII: S 1 3 6 5 - 1 6 0 9 ( 0 1 ) 0 0 0 3 3 - 8

736

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

Fig. 1. Stability graph (after Nickson [18]).

mixing data from dierent denitions of the stability

number and
human bias.

The stability graph method of open stope design is an

empirical method. Empirical methods are more reliable
when they are based on large databases. The original
database that was used to develop the stability graph
method consisted of only 50 case histories, and potential
users had little condence in its application. Potvin [1]
increased the database to 175 case histories resulting in
re-calibration of factors A, B and C, and a redenition
of the boundaries between stability states, all of which
lead to improved condence in the use of the stability
graph method. Hadjigeorgiou et al. [5] reports a total
database of 228 case histories. A distinction should be
made between the databases based on Potvins denition
of A, B and C, and those based on the original Mathews
factors such as the Golder Associates [6] and the
Australian [7] databases. This is because the two
databases are based on dierent denitions of the
stability number N, and consequently have dierent
transition boundaries between stability states.
There is no justication for continued use of the
original stability graph factors. Empirical design methods can improve and evolve with the availability of more
data and increased experience with the method. There
have been several signicant developments since 1980,
which have improved the reliability of the stability
graph (see [1,819]). The focus of these developments is
on unambiguously dening the stability graph class
boundaries, including omitted factors in N, and recalibrating existing factors. These developments, based
on an increased database, and continuing experience
with the design technique, further enhance the reliability
of the stability graph method and cannot be ignored.
Potvin [1] visually divided the stability graph into two
zones called stable and caved zones, but with an
additional inferred supportable limit, instead of the
three zones dened by Mathews et al. [2]. Potvins two-

zone stability graph has been criticized by Stewart and

Forsyth [6] who argue that the stability graph method is
non-rigorous, and that Potvins division of the graph
into just two zones gave the misconception that a stope
is either stable or caved, possibly implying that the
method is rigorous. Nickson [18] corrected this misconception of the Potvins stability graph by including a
transition boundary between unstable and caved stopes,
which he referred to as the supportable limit (Potvins
inferred support limit). The PotvinNickson form of the
stability graph is the accepted form of the stability
graph, even though Neumann [20] argues that it is the
Potvins design curves that are more familiar to most
operators. The PotvinNickson stability graph consists
of three zones, namely stable, transition zone (a mixture
of stable, unstable and caved stopes) and caved. The
word caved in the three-zone PotvinNickson stability graph actually reects the true meaning of the word
in the mining terminology, i.e. continuous cave propagation that is dicult to arrest.

2. Limitations and objectives

Limitations in the use of the stability graph are
discussed in Suorineni et al. [9]. Prominent among the
limitations were the lack of a fault factor and an
appropriate gravity factor for stope walls dipping less
than 701. Suorineni et al. [9] introduced a fault factor to
correct for faults and Hadjigeorgiou et al. [5] presented a
new chart for sliding failure in stope walls dipping less
than 701. Experience with the gravity factor shows that
the stability of low dipping stope surfaces is often
underestimated by the Potvin chart for sliding failure
[20].
Bawden et al. [21] used data from Greer [22] to
illustrate the necessity for calibration of the stability
graph zones to local mine conditions implying that for
each mining environment, the stability graph should
have somewhat dierent boundaries. While this is ideal,
it does not allow for the accumulation and transfer of
experience from one mining environment to another,
which is the fundamental advantage of empirical design.
Trueman et al. [7] observed that it is advantageous to
reduce the site-specic component of the stability graph
method to negate the need for developing zones of
the graph for individual sites. An alternative solution is
to have multiple transition boundaries in the form of
probability of failure such that any mine would choose
its design line (probability contour) based on its
acceptable risk.
Human bias and unknown inherent errors are
introduced into the stability graph when the stability
graph zones are dened visually. The limitation of
visually dening the stability graph zones was recognized by Potvin [1], who recommended that statistical

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

tools be used to dene these zones. Some eorts have

been made by Nickson [18] and Hadjigeorgiou et al. [5]
to use statistics to verify the Potvin [1] stability graph
boundary between stable and unstable stopes (unstable
and caved stopes). Both Hadjigeorgiou et al. [5] and
Nickson [18] used the Mahalanobis distance method of
discriminant analysis, and concluded that there was no
need to change the Potvin boundary. However, both
recommended an additional boundary between unstable
and caving stopes. The statistical boundaries of
Hadjigeorgiou et al. [5] and Nickson [18] are single
lines, which may give the wrong impression that there
are distinctively dened boundaries between stable and
unstable stopes, and between unstable and caving
stopes.
Statistical tools can be used to better interpret
boundaries on the stability graph without making it
seem more rigorous than it is currently perceived. The
objectives of this paper are to use the Baysian likelihood
statistic in discriminant analysis to
*

show that the stability graph zones dened as stable,

unstable and caving are not absolute or denitive but
overlap,
dene general stability graph transition boundaries
between stable and unstable stopes, and between
unstable and caved stopes,
estimate inherent predictive errors in stability graphs
as there are no denitive boundaries between the
stability graph zones,
estimate the risk/cost involved in making predictions
with the stability graph,
introduce a stability graph with multiple design
curves based on a stopes likely performance and
present a robust statistical procedure for the interpretation of the stability graph.

The overall goal of the paper is to give more value and

meaning to the various aspects of the stability graph
through quantitative interpretation, and to inform users
of the potential misclassications in stope performance,
and their implications to a mine. As part of the overall
objective, the paper uses the stability graph to demonstrate the power of the likelihood statistic to quantify
subjective classications in empirical engineering design.

3. Transition boundaries
The transition zones statistically dened by Nickson
[18] and Hadjigeorgiou et al. [5] are single curves while
the eye-balled transition boundaries are bands that
can be dened by lower and upper limits. This dierence
between the statistical and eye-balled transition
boundaries is signicant. Single-curve transition zones
may imply denitive boundaries without errors, while
the eye-balled bands imply that there is no sharp

737

boundary between the stability states of stopes (there

are grey areas), and therefore carry some risk. The
mixture of points at boundaries between the stability
graph classes shows that there are no denitive
boundaries between classes, and that boundaries accounting for grey areas are more realistic.

4. Baysian discrimination
The boundaries of the stability graph can be examined
statistically with the Baysian likelihood statistic in
discriminant analysis, using equi-probability contours
and likelihood ratio approaches. The errors associated
with these boundaries can also be assessed, and risk cost
associated with misclassications determined. The Baysian likelihood method optimizes the stability graph
class boundaries, which can be further corrected for
data inequality bias, and seriousness of misclassications. Baysian discrimination has several advantages
over other methods such as the logistic regression and
Mahalanobis distance methods.
In logistic regression [23], it is not easy to account for
cost [24] and it is dicult to correct for misclassication
errors without introducing additional bias. Furthermore, when the number of data points in the various
groups in the database are not equal (i.e. unequal a
priori probabilities), the discriminant is bias towards the
data group with most points.
The general Baysian likelihood ratio also has advantages over the Mahalanobis distance method, which
assumes that a priori probabilities are equal. The
Mahalanobis distance method is a special case of the
Baysian likehood ratio in which the likelihood ratio is
unity, and it is incapable of dening the class of an
object for which the likelihood ratio is not unity.
The principle of likelihood in discriminant analysis is
related to Bayes probability theorem. Consider the
stable S and unstable U (unstable plus caved) stopes in
the stability graph, then Bayes theorem is stated in the
following forms:
PUjX

PXjUPU
;
PXPXjX

PSjX

PXjSPS
;
PXPXjX

where PUjX and PSjX are a posteriori probabilities,

and U and S represent the matrices of unstable and
stable stopes. PU and PS are the a priori probabilities, and PXjU and PXjS are conditional probabilities. X and X are observed data and sample population
respectively.
Bayes discrimination rule states that assign the
object to the group with the highest conditional
probability (i.e. if as shown in Eq. (5) the conditional

738

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

probability of the stope wall X to be unstable PXjU is

greater than that of its being stable PXjS; we assign X
to the unstable group)
5

PXjU > PXjS:

From Eqs. (4) and (5), Bayes theorem can be related

to Bayes discrimination rule as shown below:
PUjX PU PXjU

:
6
PSjX
PS PXjS
The left-hand side of Eq. (6) is the ratio of a posteriori
probabilities. The rst factor on the right is the ratio of a
priori probabilities and the second factor is dened as
the likelihood ratio L; and is the ratio of the conditional
probabilities of the data.
The equiprobability and likelihood ratio methods
assume that the two data groups are normally distributed. Two likelihood functions are dened for the
two data groups (unstable and stable stopes) as
1
p
fu X
2psHR sN 0 1  rHR N 0


1
% u T S1 Xu  X
% u ;
 exp  Xu  X
7
u
2
fs X

boundaries, and to show that the stability graph class

boundaries overlap and are therefore not denitive or
absolute.

1
p
2psHR sN 0 1  rHR N 0


1
% s T S1 Xs  X
% s ;
 exp  Xs  X
s
2

where X is the matrix of the corresponding stability state

hydraulic radii and stability numbers, S1
u is the inverse
of the pooled variancecovariance of the unstable cases,
r is the correlation coecient and s is the variance.
Subscripts u and s represent unstable and stable states of
stopes. The ratio fu X=fs X corresponds to the likelihood ratio in Eq. (6). Eqs. (7) and (8) can be rewritten
as follows:
% u S1 Xu  X
% u c;
Xu  X
9a
u

% s S1 Xs  X
% s k:
Xs  X
s

9b

Eq. (9) is a second degree polynomial that denes

ellipses representing two sets of equiprobability contours for unstable and stable open stope walls. For the
special case when the likelihood ratio is 1, we have two
intersecting equiprobability contours.
The constants, c and k in Eq. (9) are dened as chisquare w2p a distributions [24] because they represent
sums of independently derived squares of estimates of
unknown parameters [25]. w2p a is the (100a)th percentile
of a chi-squared distribution with p-degrees of freedom.
Eq. (9) can be rewritten as
% T S1 X  X
% pw2 a:
X  X
10
p

The following section uses the equiprobabilitycontours concept to estimate the stability graph class

4.1. Equiprobability-contour method (approximate

method)
The equiprobability-contours concept is an approximate method but has the advantage of determining the
degree of overlap between data groups, which is
important for estimating prediction errors. The success
of a discriminating rule in discriminant analysis depends
on the amount of overlap of the data groups. Increasing
overlap of data diminishes the success and usefulness of
the discrimination. If the overlap or mix of data is large,
many wrong predictions will be made. On the other
hand, if the overlap is small, only a few decisions will be
in error. A data separability index (amount of overlap)
is dened as the ratio of the distance between the
% s of two data groups and the pooled
% u and X
centroids X
variance S; expressed as
%u X
% sj
jX
:
11
W
S
The data separability index increases with decreasing
overlap.
Eq. (10) leads to groups of contours, each containing
(1a)100% of the probability of an observation
occurring. The point L in Fig. 2 is a stope surface
that has 40% probability of stability, 20% probability
of being unstable and 40% probability of caving.
Similarly, the stability states of other points in the
stability graph can be dened. A curve dened by
joining the intersections of equal equiprobability contours for unstable and stable stopes optimally partitions
the two data groups. This partition represents a
transition boundary in the stability graph. The technique is used in SYSTAT statistical package to estimate
the transition boundaries in the stability graph as shown
in Fig. 2.
In Fig. 2, the lines AA, BB and CC are discriminant
curves obtained by joining the intersections of corresponding equiprobability contours by considering stable
and unstable stopes, stable and caved stopes and
unstable and caved stopes, respectively. aa, bb and cc
are linear extrapolations of the straight portion of the
actual discriminant curves AA, BB and CC, respectively.
Linear portions of AA, BB and CC occur within the
region of high data density and therefore high condence. In the stability graph terminology, aa is the
upper limit of the upper boundary of the transition zone
that separates stable from unstable stopes. The line bb is
the lower limit of the upper boundary of the transition
zone between stable and unstable stopes, and cc is the
lower limit of the lower boundary of the transition zone
between the unstable and caved classes.

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

739

4.2. Likelihood ratio method

The likelihood ratio L is dened as the ratio of the
likelihood functions of two data groups or normal
density functions of a bivariate system, or of any pair of
data groups in a multivariate data set (Eq. (12)). For the
special case when c k; the likelihood ratio is equal to 1
and we have intersecting equiprobability contours as
discussed earlier:
fu X
:
12
L
fs X
Eq. (12) can be expanded and rewritten as
% S1 X
% S1 X
%u X
% s ;
lnL XT S1 d  12 X
u
s

13

where d is the mean dierence vector given by

% s:
%u X
dX

14

For the special case when the likelihood ratio is unity,

Eq. (13) becomes zero and can be rewritten as Eq. (15),
and any object with this characteristic can neither be
assigned to the unstable nor to the stable group
(transition object). A curve dening all transition objects
is the discriminating curve between stable and unstable
stope clusters (the transition boundary):
%s X
% u K 0;
2XT S1 X
15
i

where
% T S1 Xs :
% T S1 Xu  X
KX
u
s

16

Eq. (16) is a scalar, and Eq. (15) is equivalent to a

straight-line equation, similar to Eq. (17), whose coecients b0 and b1 can be determined:
b0 log10 HR b1 log N 0 k 0

17

and
0

HR 10b0 b1 log10 N :

Fig. 2. (a) Interpretation of points in the stability graph: Point L

interprets as a stope surface with 40% probability of being stable, 20%
probability of being unstable and 40% probability of caving.
(b) Estimated stability graph class boundaries using equiprobability
contours.

18

Following this procedure, and using the calibration

database, two limits of the transition boundary between
stable and unstable stopes have been determined. The
upper limit of the transition boundary between stable
and unstable stopes is
0

HR 100:56660:3471 log N

19

and the lower limit of transition boundary between

unstable and stable stopes is
The advantage of the equiprobability contour method
is that it gives relative likelihoods or probability of a
stope wall being stable versus unstable (e.g. the stope
wall is 40% stable and 60% unstable) and therefore
identies potential trouble stopes for support. The
method also proves that the three zones identied in
the stability graph as stable, unstable and cave have
considerable overlap. A stope plotting in a given zone
only has a higher probability of performing as such, in
service.

HR 100:41050:3915 log N :

20

The procedure was also used to determine the lower

boundary of the transition zone (i.e. the supportable
limit) from the support database.
The upper limit of the lower boundary is
0

HR 100:82780:1578 log N

21

while the lower limit of lower boundary is

0

HR 100:90160:1578 log N :

22

740

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

5. Estimating prediction errors

The apparent error rate APER [24] is a statistical
index that can be used to estimate prediction errors in
the stability graph. APER is dened as the ratio of the
total number of misclassied objects to the total number
of objects, and is determined from a given classication
criterion by the use of a confusion matrix. The following
denes a confusion matrix for the stability graph
database.
APER is determined from the following equation:


nsm num
APER
100%;
ns nu
Fig. 3. Stability graph class boundaries: stable/unstable boundary
(bold black curves mark upper and lower boundaries) from
unsupported stopes database, supportable limit (unstable/cave boundary) from unsupported stopes database (broken grey curves mark
upper and lower boundaries) and supportable limit from supported
stopes database (continuous grey curves mark upper and lower limits).
Boundaries developed from Baysian likelihood ratio discrimination.

23

where nsm is the number of stable points misclassied,

num is the number of unstable objects misclassied, ns
and nu are the total number of stable and unstable
objects respectively, and nsc and nuc are the numbers of
stable and unstable stopes correctly classied.
Based on Table 1 and Fig. 3, an APER of 15% is
determined for the calibration database stability graph.
This implies that 15% of predictions using this stability
graph will be in error. A classication criterion that
minimizes Eq. (23), and therefore the APER of 15%, is
the best to use in dening the boundaries between
classes of objects.
Misclassication costs are based on probabilities of
correctly and incorrectly classifying objects. These
probabilities are described by a priori and a posteriori
probabilities. These a priori and a posteriori probabilities are dened from the confusion matrix in Table 1 in
Eqs. (24) and (25) respectively:

The boundary between unstable and caved stopes in the

unsupported database represents the supportable limit
based on our current support technology. This boundary was established following the same procedure, and
nearly coincides with that derived from the supported
database. Fig. 3 shows the transition zone boundaries
from the statistical analysis using the calibration
database, and includes the supportable limits from both
the support database and the unsupported database. It
is concluded that the boundaries of the stability graph
zones can be established statistically, and that the
supportable limit is actually the boundary between
unsupported unstable and caved stopes.
In deriving these optimized boundaries (Eqs. (19)
(22)) it was assumed that a priori probabilities of stable
and unstable stopes, and unstable and caved stopes in
the database are equal. It was also assumed that the
seriousness of misclassifying an unstable stope as stable
is equal to that of misclassifying a stable stope as
unstable. These assumptions may not be valid, and
procedures for further adjustment of the stability graph
class boundaries to account for data bias and inequality
in misclassication cost are presented next.

PU

nu
;
nt

PUjS

PS

num
;
nu

ns
;
nt

PSjU

24

num
:
ns

25

In the following sections, Eqs. (24) and (25) are used to

determine the risk cost, and to account for inequality in
consequences of misclassifying a stope walls state of
stability.

Table 1
Example confusion matrix for stability graph
Predicted membership
Stable
Actual membership
a

Stable zone
Unstable zone

nsc 88
num nu  nuc 19

Numbers in brackets refer to cases in the stability graph database presented in Fig. 3.

Total
Unstable
nsm ns  nsc 12
nuc 82

ns 100
nu 101

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

741

6. Risk cost associated with misclassication

The overlap of the three classes in the stability graph
means that there is room for error in predicting the
performance of a stope. Misclassication of stope
performance comes with cost. The severity of the cost
depends on the relative degrees of seriousness [26] due to
the misclassications. For example, is it more serious to
classify a stable stope as unstable or to classify an
unstable stope as stable? The answer to this question
forms the basis for further adjustment of the optimized
stability graph class boundaries to further minimize risk
in misclassifying a case that will result in high risk cost.
In the rst optimization process it is assumed that
misclassifying a stable stope as unstable is just as serious
as misclassifying an unstable stope as stable.
A misclassication cost matrix [24], is dened in terms
of stability graph classes as
True population

Classify in stable zone

Classify in unstable zone

Stable
Unstable

0
c * stablejunstable

c * unstablejstable
0

26
where c is the cost.
The expected cost of misclassication ECM due to
data bias, and inequality in relative seriousness is given by
ECM cunstablejstablePunstablejstablePS
cstablejunstablePstablejunstablePU:
27
Risk is dened as the product of probability of
occurrence of an unwanted event and the consequences
or cost due to the event occurring:
Risk probability of failureconsequence:

28

In the context of this study, misclassication is the

undesirable event, and misclassication cost is the
consequence. Eq. (28) can be re-written as
Risk misclassification probability
 misclassification cost:

to adjust the stability graph zones dened in Eqs. (19)

(22) to suite their local conditions.
The complete equation for the optimization and
adjustment of the stability graph boundaries is
lnL lnx XT S1 d  12 XTu S1 Xu  XTs S1 Xs :
32
An alternative approach is to develop a multiple
likelihood contours stability graph that denes
the stability of each stope in terms of its performance.
Such a graph is presented and discussed in the next
section.
The likelihood ratio concept was applied to 112
actual case histories in a Canadian hard rock
underground mine where depths of stope wall
failures were determined from cavity survey data.
The likelihood ratios were back computed from
Eq. (13) and plotted against depths of failure
(Fig. 4). This gure shows that the higher the likelihood ratios, the less the depth of failure, and therefore the less likely is the stope to fail. Consequently,
the likelihood ratio can be used to predict stope
performance.

29

A risk cost ratio x (Eq. (30)) is introduced for the

adjustment of the optimized stability graph boundaries
to account for dierences in costs of misclassication
and data bias:
Risk of misclassifying a stable stope as unstable
x
:
Risk of misclassifying an unstable stope as stable
30
Eq. (30) is rewritten as
cunstablejstablePunstablejstablePstable
x
:
cstablejunstablePstablejunstablePstable

Fig. 4. Likelihood ratio versus depth of failure (112 case histories).

31

The cost ratio term c(unstable|stable)/c(stable|unstable)

is mine specic, and forms the basis for individual mines

7. Design curves based on the likelihood ratio

An additional advantage of using the likelihood ratio
method is that a stability graph can be developed with
curves of varying likelihoods of failure. Such a graph
gives meaning to each point plotted on the stability
graph with regard to its expected performance. The
graph is useful for instant stope assessment in existing
mines. It is also a starting point for new mines that want
to plan new stope sizes.
The following equation is statistically developed by
dividing the unsupported database into stable and failed
stopes, with stopes in the unstable and caved classes

742

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744
Table 2
Confusion matrix for a hard rock Canadian mine
Predicted
membership

Actual membership

Stable zone
Unstable zone

Total

Stable

Unstable

46
6

31
29

77
35

Fig. 5. Multiple likelihood ratio (L) contour curves stability graph.

considered as unstable:


fs X
log
0:3878 log N 0  log HR 0:4905;
fu X

33

where the ratio of likelihood functions in brackets is the

likelihood ratio L; and compares the likelihood of a
stope being stable to it being unstable.
Eq. (33) is the basis for generating a stability graph
with multiple likelihood contours and can be rewritten
as
0

HR 100:4905log L0:3738 log N :

34

By substituting various levels of likelihood ratios

the stability graph can be contoured into various
levels of risk in stope surface performance. Fig. 5 is a
stability graph with varying likelihood ratios.
Each contour species the relative likelihood for a
stope surface to be stable versus unstable. The line
representing a likelihood ratio of 1 implies that a stope
surface falling on this line has equal chance of being
stable and unstable. Individual mines can select their
own design curve according to the level of risk they
are willing to accept. Such a level of risk could be in
terms of dilution, compensations for accidents or
support costs, or some combinations of these. Diederichs and Kaiser [27] note that with the risk to
human safety reduced by remotely controlled removal
of blasted ore from non-entry stopes, stability concerns within stopes are primarily driven by the
economics of acceptable risks of instability, and
the optimization of designs by balancing risk costs
with productivity potential. The approach presented
in this section provides a basis for including acceptable
risk and optimizing designs in the open stope design
process.
Eqs. (32) and (34) are applied to a Canadian hard
rock mine that experienced dilution. The cost matrix

Fig. 6. Optimized (bold black curve) and adjusted (grey curve) design
curves based on dilution.

and confusion matrix for the mine are given below

(see Table 2):
True population

Classify in stable zone

Classify in unstable zone

Stable
Unstable

0
300; 000

174; 000
0

35
Costs of dilution and cablebolt support were obtained
from a study by Tannant and Diederichs [28]. A total of
112 stope surfaces were analyzed.
The equation to account for inequality in data and
relative consequences in terms of dilution is given as
0

HR 100:4905log Llog x0:3738 log N :

36

Fig. 6 shows the optimized and adjusted design curves.

The adjustment gives a saving of CDN $17,203 per
stope.

8. Conclusions and recommendations

The stability graph method of open stope design is a
non-rigorous empirical method. However, the database
is amenable to statistical treatment without aecting the
robust nature of the method. Various options exist for
dening the boundaries between the classes in the

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

stability graph. A statistical denition of the stability

graph boundaries is the best option. Logistic regression,
Mahalanobis distance and Baysian likelihood discrimination methods are all possible statistical tools for use in
dening the stability graph boundaries. It is dicult to
include risk cost in logistic regression, and Mahalanobis
distance is only applicable when a prior probabilities of
the objects are equal.
Baysian likelihood discrimination is the ideal tool for
use in the statistical interpretation of the stability graph
because of its power to
*

*
*

show that the stability graph zones dened as

stable, unstable and caving have considerable
overlap,
estimate the error rate in a stability graph,
dene general stability graph transition boundaries
between stable and unstable stopes, and between
unstable and caved stopes,
estimate inherent predictive errors in stability graphs
as there are no denitive boundaries between the
stability graph zones,
estimate the risk/cost involved in making predictions
with the stability graph and
introduce a multiple design curves stability graph in
terms of a stopes likelihood of performance.

This ability of the Baysian likelihood discrimination

has been used to give more meaning to the stability
graph class boundaries and individual stope walls
plotting in each class.
It is recommended that the Baysian likelihood
discriminant be used when updating the stability graph
with the addition of more data and new stability
graph factors. The statistical boundaries dened by the
likelihood method are to be preferred to existing
boundaries in stope design. Use of multiple likelihood
design curves stability graph, which is based on the
general stability graph database for open stope design,
at specic mine sites is encouraged, as this allows the
specic mines to choose design curves based on their
acceptable risk level.

References
[1] Potvin Y. Empirical open stope design in Canada. PhD thesis,
Department of Mining and Mineral Processing, University of
British Columbia, 1988. 343pp.
[2] Mathews KE, Hoek E, Wyllie DC, Stewart SBV. Prediction of
stable excavations for mining at depth below 1000 metres in hard
rock. CANMET Report DSS Serial No. OSQ80-00081, DSS File
No. 17SQ.23440-0-9020, Department of Energy, Mines and
Resources, Ottawa, 1980. 39pp.
[3] Barton N, Lien R, Lunde J. Engineering classication of rock
masses for the design of tunnel support. Rock Mech 1974;6:
188236.

743

[4] Suorineni FT, Kaiser PK, Tannant DD. Unifying application of

the stability graph for open stope design. CIM Bull 2000,
submitted for publication.
[5] Hadjigeorgiou J, leClair J, Potvin Y. An update of the stability
graph method for open stope design. 97th CIM-AGM, Rock
Mechanics and Strata Control Session, Halifax, Nova Scotia,
1995.
[6] Stewart SBV, Forsyth WW. The Mathews method for open stope
design. CIM Bull 1995;88(992):4553.
[7] Trueman R, Mikula P, Mawdesley C, Haries N. Experience in
Australia with the Mathews method for open stope design. CIM
Bull 2000;93(1036):1627.
[8] Suorineni FT, Tannant DD, Kaiser PK. Determination of faultrelated sloughage in open stopes. Int J Rock Mech Min Sci
1999;36:891906.
[9] Suorineni FT, Tannant DD, Kaiser PK. Fault factor for the
stability graph method of open-stope design. Trans Inst Min
Metall A 1999;108:A92A104.
[10] Diederichs MS, Kaiser PK. Tensile strength and abutment
relaxation as failure control mechanisms in underground excavations. Int J Rock Mech Min Sci 1999;36:6996.
[11] Suorineni FT. Eects of faults and stress on open stope
design. PhD thesis, Department of Earth Sciences, University of
Waterloo, Waterloo, Canada, 1998. 345pp.
[12] Germain P, Hajigergiou J. Inuence of stope geometry on
mining performance. 100th AGM-CIM, Montreal, 1998,
CD-ROM.
[13] Clark LM, Pakalnis RC. An empirical design approach for
estimating unplanned dilution from open stope hangingwalls and
footwalls. 99th CIM-AGM, Vancouver, 1997.
[14] Kaiser PK, Falmagne V, Suorineni FT, Diederichs MS,
Tannant DD. Incorporation of rockmass relaxation and degradation into empirical stope design. CIM-AGM, Vancouver, 1997.
15pp.
[15] Milne D, Pakalnis RC. Theory behind empirical design techniques. 12e Colloque en Controle de Terrain de lAssociation
Minierre du Quebec, 1997. 20pp.
[16] Germain P, Hadjigeorgiou J, Lesard JF. On the relationship
between stability prediction and observed stope overbreak. In:
Proceedings of the Second NARMS, Rock Mechanics Symposium on Rock Mechanics Tools and Techniques, vol. 1, 1996.
p. 22783.
[17] Milne D, Pakalnis RC, Felderer M. Surface geometry assessment
for open stope design. In: Proceedings of Rock Mechanics.
Rotterdam: A.A. Balkema, 1996. p. 31522.
[18] Nickson SD. Cable support guidelines for underground hard rock
mine operations. MASc thesis, Department of Mining and
Mineral Processing, University of British Columbia, 1992.
223pp.
[19] Potvin Y, Milne D. Empirical cablebolt design. In: Kaiser,
McCreath, editors. Rock support. Rotterdam: A.A. Balkema,
1992. p. 26975.
[20] Neumann MF. Stability graph design methodFa mining
operators guide. CIM Mine Operators Conference, Bathurst,
New Brunswick, paper no. 5, 1999. 8pp.
[21] Bawden WF, Nantel J, Sprott D. Practical rock engineering in
the optimization of stope dimensionsFapplication and cost
eectiveness. CIM Bull 1989;82(296):6370.
[22] Greer GJ. Empirical modelling of open stope stability in a vertical
crater retreat application at Inco Thompsons mine. 91st AGMCIM, Quebec City, Canada, 1989. 12pp.
[23] Hosmer Jr DW, Lemeshow S. Applied logistic regression.
New York, NY: Wiley, 1989. 307pp.
[24] Johnson RA, Wichern DW. Applied multivariate statistical
analysis. Englewood Clis, NJ: Prentice-Hall, 1992. 642pp.

744

F.T. Suorineni et al. / International Journal of Rock Mechanics & Mining Sciences 38 (2001) 735744

[25] Beinstein IH, Garbin CP, Teng GK. Applied multivariate

analysis. Berlin: Springer, 1987.
[26] Van de Geer JP. Introduction to multivariate analysis for social
sciences. San Francisco: W.H. Freeman and Company, 1971.
293pp.

[27] Diederichs MS, Kaiser PK. Rock instability and risk analysis in
open stope mine design. Can Geotech J 1996;33:4319.
[28] Tannant DD, Diederichs MS. Cablebolt optimization in #3 Mine.
Report to Shawn Seldon, Kidd Mines Division, Timmins,
Ontario, 1997. 65pp.