International Journal of Rock Mechanics & Mining Sciences: David Recio-Gordo, Rafael Jimenez

Download as pdf or txt
Download as pdf or txt
You are on page 1of 7

International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187

Contents lists available at SciVerse ScienceDirect

International Journal of
Rock Mechanics & Mining Sciences
journal homepage: www.elsevier.com/locate/ijrmms

Technical Note

A probabilistic extension to the empirical ALPS and ARMPS systems


for coal pillar design
David Recio-Gordo a,b, Rafael Jimenez a,n
a
Technical University of Madrid, ETSI Caminos, Canales y Puertos, Madrid, Spain
b
ACCIONA Infraestructuras, R&D Geotechnics and Underground Works Department, Spain

a r t i c l e i n f o

Article history:
Received 22 August 2011
Received in revised form
13 February 2012
Accepted 11 March 2012
Available online 17 April 2012

1. Introduction (ii) pillar strength estimation (see below for a discussion); and
(iii) safety factor computation to define pillar stability. Classic
Coal pillar design has attracted much attention to develop coal pillar strength formulas for room and pillar design typically
cost-effective recovery schemes with appropriate safety condi- involve two parameters: (i) coal strength; and (ii) pillar width
tions, as we aim to minimize pillars dimension (i.e. to maximize to height ratio (w/h). Many authors have studied the Uniaxial
the extraction of coal) while sufficient stability is maintained [1]. Compressive Strength (UCS) of coal for pillar design and its
To maintain a stable roof in a room and pillar design, ‘small’ variability with specimen size [10,11], and it is common that
pillars are left regularly, although such pillars can be later mined classic formulas use a UCS related to the specific seam strength,
in retreat when all panels have been fully mined-out. Similarly, the residual strength for the critical specimen size, or even to
the longwall mining system consists of ‘large’ panels which are empirical values [12]. In addition, there is ample evidence that
mined with continuous miners protected under a hydraulic roof pillar strength increases with the w/h ratio and, in particular,
movable support. The main goal of pillars in longwall mining is to experience suggests that pillars with w/h410 are almost ‘indes-
maintain appropriate conditions of serviceability within the road tructible’ [13]; however the rate of increase of pillar strength as a
entry systems until the panel is fully mined. Because the roof over function of the w/h ratio is not clear, and linear and power
the mined-out panel will collapse behind the movable roof expressions have been proposed (see e.g., [6]). However, such
support, the side pillars must be able to resist high loads in methodology supposes an over-simplification of the problem, and
comparison with pillars used in a room and pillar design. experience demonstrated that the pillar failure mechanism
Many parameters affect coal pillar stability, including pillar needed to be considered to study pillar stability.
strength and loads, mine and pillar geometry, surrounding geology, For instance, Wilson [14] (see also [15,16]) proposed a model
etc. In that sense, for instance, significant research efforts have been for pillar strength that involves the role of the w/h ratio, and also
conducted to characterize coal pillar strength using laboratory considered more complex modes of behavior with non-constant
testing, in situ measurements, back analysis or even numerical or stress levels within the pillar. Such stress gradient is due to the
analytical models [2–5]; however, despite such efforts, significant confinement provided by the passive pressures mobilized through
uncertainties for coal strength estimation still remain and there are the external pillar zone (where stress is minimum) when it begins
often significant divergences between predictions provided by each to yield (due to the frictional strength mobilization, via Poisson
approach [6]. effect, that occurs in the pillar–roof and pillar–floor interfaces),
For that reason, room and pillar designs have been tradition- and it results in an increased strength of the central pillar core,
ally based on empirical approaches (see e.g., [7–10]). Such where stress is maximum (see e.g., [4,5]). Barron [17] observed
classical methods involve three steps for pillar design: (i) pillar that confinement around the pillar core and failure type were
load estimation (usually by means of the tributary area method); linked, with low confinements producing brittle failures and high
confinements producing ductile failures. (Such brittle–ductile
failure transitions have also been discussed in other fields of rock
n
Corresponding author. Tel.: þ34 913366710(x)107; fax: þ34 913366774.
mechanics, such as stiff clay behavior [18], rock failure criteria
E-mail address: [email protected] (R. Jimenez). [19,20] or constitutive modeling of porous rocks [21]).

1365-1609/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijrmms.2012.03.009
182 D. Recio-Gordo, R. Jimenez / International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187

Similarly, initial longwall mining experiences showed that pillar of unstable pillars had a safety factor of less than one, and
shape could influence the pillar behavior at failure, and it was that 76% of stable pillars had safety factors greater than one.
observed that many pillars did not fail in the classic sense; instead, They also observed that there was a wide range of cases with
their failure was associated to the serviceability of roads and to the safety factors between 0.5 and 1.2 for which the stability predic-
control of ground. It was evident that the pillar failure mode was tions were uncertain, hence concluding that other parameters
linked with its shape [5], and analytical approaches and new pillar may also control the behavior of that wide range of cases. This is
mechanics concepts emerged. For instance, several authors (see e.g., due to existing uncertainties on the estimation of pillar load and
[22–26]) have also pointed out the importance of other basic of pillar strength, as well as to additional complex phenomena
parameters such as geology, quality of surrounding strata, in situ involved on pillar behavior [11,47]. Mark proposed (see e.g., [48])
stress, gallery geometry, mining operation processes, etc. that roof quality was a key factor for longwall pillar stability,
Numerical methods provide one alternative for coal pillar design and also that other factors such as entry width and primary
considering such factors. Their ability to construct 3D models with support were relevant. To improve the original ALPS system, the
complex geometries and construction sequences, to incorporate Coal Mine Roof Rating (CMRR) [26,49,50] was incorporated. In a
the influence of surrounding strata, to perform parametric analyses similar fashion to RMR [51], the CMRR considers some important
and back-analyses, as well as to consider advanced constitutive factors affecting roof competence such as intact rock strength;
models and stress conditions are, among others, some of their shear strength, spacing and persistence of discontinuities; the
most interesting capabilities [27]. Some codes (mainly based on influence of a strong bed within the bolting interval, or ground-
the boundary-element (BE) method) have been specifically devel- water inflow. CMRR quantifies the competence of the mine roof
oped for coal pillar design, such as BESOL [28], MULSIM/NL [29], and using a numerical scale between 0 and 100, and three general
LAMODEL [30,31] and, of course, other (commercial or in-house) quality levels are defined: weak roof (CMRR o45), moderate roof
geotechnical codes can also be employed. Currently used codes are (45oCMRRo65), and strong roof (CMRR 465).
quite advanced and they can consider complex non-linear constitu- After their analysis of pillar behavior in several USA mines,
tive laws for the pillar and for the overburden (or goaf) materials; with observations of ‘satisfactory’ (same pillar design used for
they can also consider the support system, hence providing realistic successive panels, no tailgate blockages exists, minimal delays
predictions of mine stability and deformations [32,33]. Some recent and good ground controly) and ‘unsatisfactory’ (pillar design
experiences with the application of numerical models to different must be changed frequently, panel mine aborted, existence of
aspects of coal pillar behavior and of other aspects of underground unacceptable conditions in deepest areas, rock falls and delaysy)
coal mining can be found in the technical literature [34–38]. performance, Mark et al. [52] developed a (linear) predictive
However, despite such recent advances in numerical models, model to assess the expected behavior of longwall pillars using
their routine use for practical pillar design, particularly at early the ALPS safety factor (ALPS SF) and the roof mine quality (as
stages of a project, is probably yet not feasible for reasons similar given by CMRR). (Note that, although it is possible nowadays to
to those given elsewhere [39–41]. For that reason, empirical and estimate the CMRR using computer programs or drill core data,
observational-based methods have gained a great relevance as a direct underground observations were employed for CMRR esti-
simple and practical tool for design of such complex problem as mation in this case). Such criterion was initially fitted through a
coal pillar design [16]. (Note that numerical models also need to statistical regression and then modified into a more conservative
be supported by empirical observations to improve the quality of design boundary. Later [5], after a review of some aspects (such as
their predictions; see e.g., [42,43]). Although empirical methods coal UCS, depth of mine cover, pillar shape, entry width) and,
are not an universal solution, they are a powerful tool to capture more importantly, after the introduction of the Mark–Bieniawski
the real (and complex) pillar behavior in a very simple way, and strength formula [53] for large rectangular pillars, the original
they can be calibrated to incorporate insight from numerical stability criteria was updated to consider the revised version of
modeling and to reproduce real field experiences [40]. the factor of safety, ALPS(R) SF, as follows (see Fig. 1):
In this paper, we present a probabilistic model based on the
ALPSðRÞ SF ¼ 20:016CMRR ð1Þ
theory of linear classifiers that can be used to make probabilistic
predictions of pillar behavior in longwall and retreat room and As shown in Fig. 1, higher design values of ALPS(R) SF are needed
pillar mining. The objective is to provide designers with a quali- for satisfactory performance when the CMRR parameter is low
tative tool for (probabilistic) decision making in the context of (i.e., when the roof is ‘weak’), while a lower value of ALPS(R) SF
coal pillar design. Our approach is based on two well known and can be used for pillar dimensioning when we have a roof of higher
widely accepted empirical methods for coal pillar design: the quality (i.e. of higher CMRR). In engineering practice, ALPS can be
Analysis for Longwall Pillar Stability (ALPS) [44,45] and the Analysis used to analyze the stability conditions (as given by ALPS(R) SF) of
of Retreat Mining Pillar Stability (ARMPS) [46]. These approaches pillars with a given geometry or, alternatively, the designer can
are currently a basic reference for preliminary design of coal compute the pillar dimensions needed to achieve a target ALPS(R) SF.
pillars in longwall and (evolutioned) room and pillar mining. Both Building on ALPS, Mark and Chase [47] developed the ARMPS
methods are (briefly) presented below. Additional details can be system to analyze the stability conditions of pillar systems during
found in the original publications. retreat operations and ore recovery in mines operated by room
and pillars. The method is similar to ALPS (it has the same load-
ing and strength formulation), but it is modified to consider
2. The empirical ALPS and ARMPS systems the specifics of retreat operations in room and pillar mining.
(Note that standard practice for retreat operations may also
The ALPS system [44,45] was developed to analyze the change for different mining regions or different countries; for a
stability of gate entry systems with large chain pillars in longwall recent review, see [54]). For instance, ARMPS allows to estimate
mines. Following the framework of classic room and pillar theory, pillar loads for situations such as before retreat, when the active
the ALPS system proposes pillar designs based on (i) loads applied front panel is fully retreated, and during other common config-
to the pillars, (ii) the strength of the pillars and (iii) the safety urations of retreat mining operations. ARMPS was empirically
factor given by the ratio of strength over load. Based on field calibrated with 140 real cases in the USA [47], and experi-
measurements of an extensive pillar monitoring program of ences showed that 77% of the cases labeled as ‘satisfactory’ or
more than a hundred real cases, the authors observed that 88% ‘unsatisfactory’ could be adequately classified using a ARMPS SF
D. Recio-Gordo, R. Jimenez / International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187 183

can be used as a ‘soft’ probabilistic design criterion (i.e., as a ‘soft’


Satisfactory boundary) within the ALPS(R) SF vs. CMRR framework for design
80 Unsatisfactory of longwall operations; or within the depth of cover-ARMPS SF
Mark’s Criterion
model in retreat room and pillar.
In other words, we extend the (deterministic) capabilities of
70 ALPS and ARMPS to consider prediction uncertainties. The reason is
that ALPS establishes a ‘hard’ linear boundary within an ALPS(R) SF
vs. CMRR space (see Eq. (1)) to determine the occurrence (or not) of
60 pillar ‘failures’ (in the wide sense of ‘unsatisfactory’ performance).
CMRR

However, this ‘hard-boundary’ linear approach fails to separate


completely the two performance classes considered (‘satisfactory’
50 vs. ‘unsatisfactory’) and, at the same time, it does not provide
estimations of the uncertainty of the predictions. Similarly,
our statistical approach can also be used to establish a (previously
unavailable) linear classifier of ‘satisfactory’ vs. ‘unsatisfactory’
40
conditions in retreat room and pillar mining using depth of cover
vs. ARMPS SF space variables.

30

0.5 1.0 1.5 2.0


ALPS (R) SF
3. The linear classifier
Fig. 1. Data set used in the ALPS(R) design method (from [5]).
Classifiers are statistical techniques to assign labels (Y) to a
given a set of observations (X). Once trained, they can be used to
estimate the behavior (i.e., the outcome) of future observations
for which the outcome (i.e. label) is unknown. The key idea is that
Satisfactory
Unsatisfactory such analysis can be conducted probabilistically, so that given a
2.5 set of observations (X), we can compute the conditional prob-
abilities of the different outcomes given the observation (p(Y/X));
consequently, we can also assign the label to the outcome that
maximizes such conditional probability. Our description of linear
2.0 classifiers below is mainly based on the work of Jordan [56];
additional useful references are [57] and [58].
ARMPS SF

We use the logistic function (see Fig. 3) to define soft boundaries


1.5 between the two categories (or classes) considered (‘satisfactory’ vs.
‘unsatisfactory’). To that end, we use a binary classification (0,1) for
‘satisfactory’ and ‘unsatisfactory’ cases, respectively. Given one
observation of the feature vector, X¼x, the conditional probability
1.0 of each outcome p(Y¼ y/X¼x), with y¼0 or y¼1, is modeled by
a mapping function F(yTx), where F is the logistic function in
this case and y is the vector of coefficients of the linear classifier
0.5 (note that yTx is a linear transformation).

500 1000 1500 2000 1.0


Depth Cover (ft)

Fig. 2. Data set used in the ARMPS design method (from [5]).
0.8

value of 1.46 (see Fig. 2). In this case, ‘unsatisfactory’ refers to


cases with significant entry closure due to pillar squeezing, to
0.6
sudden collapses of groups of pillars, or to violent failures, among
others. The CMRR parameter does not seem to be so critical
µ (X)

within the ARMPS framework as it was with ALPS and, similarly,


the UCS of coal has little influence on the safety of the pillar 0.4
within the ARMPS system [12]. (Note, however, that in this case
the author did not propose a criterion for separation between
acceptable and non-acceptable performances).
0.2
In this work, and following the recent trend for development
of simple tools for complex rock engineering design (see e.g.,
[40,55]), we propose a new probabilistic approach to the ALPS and
ARMPS systems, in which we use advanced statistical techniques 0.0
to develop classifiers that help us not only to estimate the
−4 −2 0 2 4
expected performance of a given design (i.e. ‘satisfactory’ vs.
X
‘unsatisfactory’), but also to estimate probabilities for each out-
come. In particular, we present a logistic-type linear classifier that Fig. 3. The logistic function.
184 D. Recio-Gordo, R. Jimenez / International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187

In general, we will have N observations as the training set of the changes at each iteration, hence the name of ‘‘iteratively
classifier, and so that we can produce a matrix with observed reweighted’’.) Substituting in Eq. (8) the expressions for the gradient
(input) features X¼(X1, X2,y, XN), and a vector of observed and the Hessian matrix given by Eqs. (9) and (10), we obtain the
outcomes Y¼(Y1, Y2,y, YN). Such set of values will be denoted as simplified expression of the IRLS algorithm for our logistic linear
D¼ {(Xn, Yn);n¼ 1,y, N)}. Similarly, we use m(x) to denote to the classifier as
expected value of Y for a given X¼x; i.e., m(x)9p(Y¼1/x). Then, the
probability of the Bernoulli distribution is given by yðt þ 1Þ ¼ yðtÞ þðX T W ðtÞ XÞ1 X T ðymðtÞ Þ ð11Þ

pðy=xÞ ¼ mðxÞy ½1mðxÞ1y ð2Þ The algorithm converges fast, and only a few of iterations are
usually needed to obtain an adequate result, hence providing
As mentioned above, the mapping between observations and quick solutions with low computational cost, even for cases in
probabilities is performed using the logistic function. We start by which large databases are involved.
defining an auxiliary parameter to denote the following linear
transformation:

ZðxÞ9yT x, ð3Þ 5. Results and discussion

which is translated into a probability scale via the logistic func- For our pillar design case, each observation of the training set
tion (see Fig. 3). Therefore, we get will be composed by (i) the feature vector Xn ¼(1, x1, x2) of input
1 variables, where the constant 1 is introduced to provide an
mðxÞ ¼ ð4Þ
1 þeZðxÞ independent term to the classifier. For the ALPS(R)-based classi-
and, inverting this equation, we get fier, x1 corresponds to the pillar safety factor given by ALPS(R) SF,
  whereas x2 corresponds to the CMRR indicating roof quality. For
mðxÞ the ARMPS-based classifier, x1 corresponds to the depth of cover
ZðxÞ ¼ ln : ð5Þ
1-mðxÞ and x2 to the pillar safety factor given by ARMPS SF; and (ii) the
Given the set of N observations from the database, and using observed outcome, Yn, which takes the value of 1 for ‘unsatis-
Eq. (2), we can compute the likelihood of our observations for a factory’ (or pillar ‘failure’) cases, while it takes the value of
given set of parameters, y, as the product of N Bernoulli prob- 0 otherwise.
abilities: Once that the dataset of training cases given by Figs. 1 and 2 is
Y y considered, we can use the IRLS algorithm given by Eq. (11) to
pðy1 . . .yn =x1 . . .xn , yÞ ¼ mnn ð1mn Þ1yn ð6Þ compute the parameters of the classifier. The results of such
n
analysis are presented in Tables 1 and 2. (Results presented in this
and, taking logarithms to compute the log-likelihood, we get paper have been computed with free statistical software R [60]).
X As noted previously, the IRLS algorithm converges quickly, reach-
log-likelihood ¼ lðy=DÞ ¼ fyn log mn þ ð1yn Þlogð1mn Þg: ð7Þ
n ing the solution in the sixth iteration for the ALPS(R) SF vs. CMRR
classifier and in the fifth iteration for the depth of cover vs. ARMPS
To calibrate the model, we maximize Eq. (6) with respect to y;
SF classifier (for a tolerance of 1  E-12).
that is, we maximize the likelihood of our model or, in other
Once the classifier has been trained, it is possible to develop
words, the probability that our model produces the available
new separation class boundaries (of ‘satisfactory’ vs. ‘unsatisfac-
(training) data. The optimization algorithm is described below.
tory’ performance), so that they match the 50% probability line of
our previously defined model p(Y¼1/X, y) (in other words, so that
4. The optimization algorithm
Table 1
To be able to use the linear classifier, it needs to be ‘learned’ for Iterative theta calibration for the ALPS vs. CMRR classifier.
each specific case. That is, as indicated above, we need to identify
ALPS
the set of parameters y that maximize the likelihood for the
available dataset of ‘adequate’ vs. ‘inadequate’ performance of Iteration y1 y2 y3
the coal pillar designs (although, for mathematical convenience, it
is preferable to work with Eq. (7) to maximize the likelihood). 0 0.000000 0.000000 0.000000
1 7.161885  3.673658  0.060643
Such optimization can be performed using the IRLS (Iteratively
2 11.708455  6.093545  0.099460
Reweigthed Least Squares) method [59]; the IRLS is a Newton– 3 14.104186  7.396625  0.119346
Raphson algorithm, which is expressed by 4 14.582125  7.663022  0.123185
5 14.597859  7.671972  0.123308
yðt þ 1Þ ¼ yðtÞ H1 ry l ð8Þ 6 14.597876  7.671982  0.123308
(t) (t þ 1)
where y and y are the parameter estimates at the t-th and
(tþ 1)-st iterations, H is the Hessian matrix (square matrix of
second partial derivatives) and ryl is the gradient of the log- Table 2
likelihood function that we aim to maximize. Taking derivatives Iterative theta calibration for the ARMPS vs. depth of cover classifier.

in Eq. (7), the gradient vector is computed as ARMPS


X
ry l ¼ ðyn mn Þxn ¼ X T ðymÞ, ð9Þ
Iteration y1 y2 y3
n

and taking the second derivatives, the Hessian matrix is computed as 0 0.000000 0.000000 0.000000
X 1 3.562409  0.001051  2.216315
H¼ mn ð1mn Þxn xTn ¼ X T WX ð10Þ 2 5.262699  0.001653  3.270681
n 3 5.923046  0.001892  3.681194
4 6.000258  0.001921  3.729237
where W is a diagonal matrix of weights given by W¼diag{m1(1 5 6.001177  0.001921  3.729809
 m1),y, mN(1 mN)}. (Note that W is a function of y, so that, since y
D. Recio-Gordo, R. Jimenez / International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187 185

they provide estimates of 50% of unsatisfactory performance). for a given (constant) roof strength, relatively small changes on
From Eqs. (3) and (5), we have the safety factor can produce relatively important variations on
  the probability of ‘failure’; for instance, for a case of ‘moderate’
mðxÞ
yT x ¼ ZðxÞ ¼ ln ð12Þ roof strength of CMRRE55, the probability of ‘failure’ changes
1mðxÞ
from approximately PfE80% for ALPS(R) SFE0.8; to PfE50%
from where, replacing the trained parameter vector (see Tables 1 for ALPS(R) SFE1.0; to PfE20% for ALPS(R) SFE1.2. It can be
and 2) and using a 50% probability value for the occurrence of emphasized that such type of probabilistic analyses, that are crucial
‘unsatisfactory’ performance, for the ALPS(R) vs. CMRR classifier, for preliminary risk assessments of designs, cannot be performed
we get with the traditional (deterministic) approach.
  Following the same procedure presented above, we can com-
0:5
y1 1þ y2 ALPSðRÞ SF þ y3 CMRR ¼ ZðxÞ ¼ ln ð13Þ pute the class separation line for the depth of cover vs. ARMPS SF
ð10:5Þ
model. The results are as follows (see Fig. 5):
ALPSðRÞ SF ¼ 1:900:016CMRR ð14Þ
ARMPS SF ¼ 1:615:1UE4 depth of cover ½ft ð15Þ
Similarly, we can compute other equi-probability lines to assess
uncertainties within our predictive model of coal pillar behavior Similarly, Fig. 5 shows equi-probability lines for other values of
based on ALPS(R) SF and CMRR. Such equi-probability lines have the probability of unsatisfactory performance. It is interesting to
been plotted, in addition to the class separation line (50% of note that, for this case, the separation capabilities of the classifier
probability of occurrence), in Fig. 4. As shown in such figure, our are not as good as before, since we have a total of 40 (¼17þ23)
new class separation boundary is very similar (although slightly less misclassified datapoints. This is due to the particularities of the
conservative) to Mark’s criterion (shown with a dashed line). The database, which show a high level of ‘mixing’, hence suggesting that
new criterion compares well with respect to the previous criterion; the assignment uncertainties are very large. (Alternatively, it could
for instance, Mark’s criterion is observed to misclassify a total of also happen that there are additional factors that need consideration
nine case histories (8þ1¼9; with 8 false-positives or cases that are to better explain this problem.) In any case, the fact that prediction
incorrectly labeled as ‘failures’, and 1 false-negative) which is one uncertainties are much larger in this case is reflected by the logistic
less than the ten case histories misclassified by our criterion given function behind our classifier, since it shows a much more wide-
by the linear classifier (5þ5¼10). However, our linear classifier is spread form than in the previous case. (Also, remember that Mark
unbiased (note that Mark’s criterion is biased, albeit conservatively, did not provide a separation line for this case due to the difficulties
and that corresponds to approximately a 30–35% probability of for prediction of behavior).
‘failure’) at the same time it allows us to provide uncertainty The linear classifier for the depth of cover vs. ARMPS SF
estimations of our predictions (given by the equi-probability lines model provides additional information about the sensitivity of
shown in Fig. 4). the results to changes in the input variables involved. In that
Furthermore, with respect to model sensitivity, we can observe sense, for instance, it is observed that variations in the depth of
that both Mark’s criterion and our linear classifier are parallel, cover are not as important as changes in the value of the ARMPS
hence suggesting a similar sensitivity of changes in ALPS(R) and in safety factor. (Note that the separation line is quite ‘horizonta-
CMRR values. In addition, the model emphasizes the importance of lized’). Similarly, results suggest that lower values of the ARMPS
roof quality on the probabilistic predictions of performance and, for SF can be accepted, for a given value of probability of ‘failure’, as
instance, we note that the probability of ‘failure’ conditions, Pf, can the depth of cover is increased. In addition, and given the wide-
change significantly for a given classification step within the CMRR spread nature of the data, it is also observed that a large variation
system. (For instance, keeping a constant value of SFE1, we get a in the safety factor is needed in this case to obtain a relevant
PfE80% for a roof at the transition between ‘weak’ and ‘moderate’ variation in the probability of failure for a fixed value of the
(CMRRE45); and a PfE25% for a roof at the transition between deep of cover. For instance, if we have a depth of cover of around
‘moderate’ and strong (CMRRE65).) Similarly, we observe that, 500 ft and we consider a design with a safety factor of around

Satisfactory Satisfactory
80 1% Unsatisfactory Unsatisfactory
Class Separation Line (50%) Class Separation Line (50%)
10% Equi−Probability Lines (see labels) 2.5 Equi−Probability Lines (see labels)
Mark’s Criterion
25%

70

2.0 1%
50%
ARMPS SF

60 75%
CMRR

1.5
10%

50
25%

90% 1.0
50%
40 99%

75%
0.5
90%
30

0.5 1.0 1.5 2.0 500 1000 1500 2000


ALPS (R) SF Depth Cover (ft)

Fig. 4. Linear classifier for the ALPS(R) SF vs. CMRR predictive model. Fig. 5. Linear classifier for the ARMPS SF vs. depth of cover predictive model.
186 D. Recio-Gordo, R. Jimenez / International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187

SF E1.1, we can see in Fig. 5 that the model would predict proposed criteria for estimation of (probabilities of) pillar perfor-
approximately a 75% probability of ‘failure’ for such design; if mance is an empirical method and, therefore, it is only valid within
we wanted to reduce such probability of failure (say, to less than the range of the variables for which data was available. That means
35%), then the model suggests that we would have to significantly that, at least until more case histories can be included in the
increase the ARMPS safety factor to values of SF 41.5. database, our empirical criterion should be used with care outside
This behavior reflects the inherent inseparability of classes of the USA and for values of the input variables outside the ranges
within the depth of cover-ARMPS SF space considered and, accord- employed to fit the parameters of the classifiers.
ingly, the high level of uncertainty of the model. However, even
in this case, the probabilistic information provided by the linear
classifier is still useful for preliminary designs and for risk assess-
References
ments, as the logistic function reduces the problem of insepar-
ability of the outcomes through a mapping of probabilities of
[1] Wagner H. Pillar design in coal mines. J S Afr Inst Min Metall 1980;80:37–45.
occurrence. In any case, and as noted by [47], additional research
[2] Hustrulid WA. A review of coal strength formulas. Rock Mech 1976;8:43–50.
on this topic is probably needed to obtain better predictive models [3] Mark C, Iannacchione AT. Coal pillar mechanics: theoretical models and field
of pillar stability in retreat room and pillar coal mine design. measurements compared. In: Proceedings of the workshop on coal pillar
mechanics and design. Pittsburgh, PA: U.S. Department of the Interior, Bureau
of Mines, IC 9315; 1992. p 78–93.
[4] Gale WJ. Geological issues relating to coal pillar design. In: McNally GH,
6. Conclusions Ward CR, editors. Symposium of Geology in Longwall Mining; 1996. p. 185–91.
[5] Mark C. The state-of-the-art in coal pillar design. SME Trans 1999;308:8
(originally preprint 99-86).
Pillar design for room and pillar and longwall mining is a [6] Mark C. The evolution of intelligent coal pillar design: 1981–2006.
crucial issue in coal mining. In this paper, we present a novel In: Proceedings of 25th international conference on ground control in mining.
method for probabilistic prediction of coal pillar performance Morgantown, West Virginia; 2006. p. 325–34.
[7] Holland CT, Gaddy FL. The strength of coal in mine pillars. In: Proceedings of
using the ALPS and ARMPS methods for empirical pillar design the sixth U.S. symposium on rock mechanics. Rolla, MO: University of
of (respectively) longwall and retreat room and pillar operations. Missouri; 1964. p. 450–66.
The methodology is based on the statistical theory of linear classi- [8] Obert L, Duvall WI. Rock mechanics and the design of structures in rock.
New York.Wiley; 1967 pp. 542–545.
fiers, in which a logistic function is employed to perform the
[9] Salamon MDG, Munro AH. A study of the strength of coal pillars. J S Afr Inst
mapping between (a linear transformation of) input variables Min Metall 1967;68:55–67.
(i.e., observations) and conditional probabilities. The calibration of [10] Bieniawski ZT. The effects of specimen size on the compressive strength of
coal. Int J Rock Mech Min Sci 1968;5:325–35.
the model is conducted using a maximum likelihood approach
[11] Du X, Lu J, Morsy K, Peng S. Coal pillar design formulae review and analysis.
and the IRLS algorithm. The convergence of the algorithm is fast, In: Proceedings of 27th international conference on ground control in
and the solution is obtained after only a few iterations. mining. Morgantown, WV: West Virginia University; 2008. p. 254–61.
Our model represents an extension with respect to previously [12] Mark C, Barton TM. Pillar design and coal strength. In: Proceedings of new
technology for ground control in retreat mining. NIOSH IC 9446; 1996. p. 49–59.
available (deterministic) criteria, as it allows to define ‘soft’ decision [13] Cook N, Hood M. The stability of underground coal mines workings. In:
boundaries for preliminary design and to compute the uncertainties Proceedings international symposium on stability in coal mining. Vancouver;
of the predictions, so that it can also be used for quantitative risk 1978. p. 135–47.
[14] Wilson AH. An hypothesis concerning pillar stability. Min Eng (London)
analyses and for decision-making under uncertain conditions. 1976;131(141):409–17.
The inference capabilities of the model are adequate. In the [15] Heasley KA. The forgotten denominator, pillar loading. In Proceedings of the
ALPS(R) SF vs. CMRR model, results show that the new separation fouth North American rock mechanics symposium. Seattle; 2000. p. 457–64.
[16] Oraee K, Hosseini N, Gholinejad M. Optimization of coal pillar design in
criterion is unbiased, and also that Mark’s original criterion is longwall mining method. In: Proceedings of the 29th International con-
biased to correspond with approximately a 30–35% probability of ference on ground control in mining. Morgantown, WV; 27–29 July 2010.
‘unsatisfactory’ performance. In addition, both criteria are parallel, [17] Barron K. An analytical approach to the design of coal pillars. Can Inst Min
Bull 1984;77(868):37–44.
hence suggesting that the sensitivity of the model to changes of the [18] Dehandschutter B, Vandycke S, Sintubin M, Vandenberghe N, Gaviglio P,
input parameters are similar in both cases. Similarly, results show Sizun J-P, et al. Microfabric of fractured Boom Clay at depth: a case study
that the predictive capabilities of the ARMPS SF vs. depth of cover of brittle–ductile transitional clay behavior. Appl Clay Sci 2004;26(1–4):389–401.
[19] Hoek E. Strength of jointed rock masses. Geotechnique 1983;33(3):187–223.
model are weaker due to the inherently inseparable nature of
[20] Raynaud S, Ngan-Tillard D, Desrues J, Mazerolle F. Brittle-to-ductile transi-
observations in such space; although the logistic function reduces tion in Beaucaire marl from triaxial tests under the CT-scanner. Int J Rock
the problem of inseparability of the outcomes through a more Mech Min Sci 2008;45(5):653–71.
widespread mapping of probabilities of occurrence. The linear [21] Ketan R. An elasto-plastic constitutive model for brittle–ductile transition in
porous rocks. Int J Rock Mech Min Sci 1997;34(3–4):283.
classifier for the depth of cover vs. ARMPS SF model also suggests [22] Mark C, Bieniawski ZT. Field measurements of chain pillar response to
that variations in the depth of cover are not as important as longwall abutment loads. In: Proceedings of the fifth conference on ground
changes in the value of the ARMPS safety factor and, similarly, it control in mining; 1986. pp 114–22.
[23] Hasenfus GJ, Su DWH. A comprehensive integrated approach for longwall
suggests that lower values of the ARMPS SF can be accepted, for a development design. In: Proceedings of the workshop on coal pillar
given value of probability of ‘failure’, as the depth of cover is mechanics and design. Pittsburgh: U.S. Department of the Interior, Bureau
increased. In addition, and given the widespread nature of the data, of Mines, IC 9315; 1992. p. 225–37.
[24] Maleki H. In situ pillar strength and failure mechanisms for U.S. coal seams.
it is also observed that a large variation in the safety factor is In: Proceedings of the workshop on coal pillar mechanics and design.
needed in this case to obtain a relevant variation in the probability Pittsburgh, PA: U.S. Department of the Interior, Bureau of Mines, IC 9315;
of failure for a fixed value of the deep of cover. 1992. p. 73–7.
[25] Parker J. Mine pillar design in 1993: computers have become the opiate of
Finally, it has to be emphasized that (although useful for mining engineering, parts I and II Mining Engineering; 1993. July: 714-17
quantitative risk analyses and for decision making under uncer- and August: 1047-50.
tainty), the presented probabilistic solutions are themselves [26] Mark C, Molinda GM. The coal mine roof rating (CMRR)—a decade of
experience. Int J Coal Geol 2005;64(1–2):85–103.
‘uncertain’ given the complexity of the problem, the non comple-
[27] Cassie J, Altounyan P, Cartwright P. Coal pillar design for longwall gate
tely random nature of the sampling, the presence of confounding entries. In: Mark C. et al, editors. Proceedings of the second international
variables, etc.; in other words, they could be (of course) further workshop on coal pillar mechanics and design. Pittsburgh: NIOSH. Vail,
improved in the future as more case histories of ‘satisfactory’ Colorado; 1999. p. 115–37.
[28] Crouch Research, Inc. The BESOL system: boundary element solutions for
(or ‘unsatisfactory’) pillar behavior are included in the training rock mechanics problems: user’s guide, version 2.01. St. Paul, MN: Crouch
database. In addition, the reader should note that the newly Research, Inc.; 1988.
D. Recio-Gordo, R. Jimenez / International Journal of Rock Mechanics & Mining Sciences 52 (2012) 181–187 187

[29] Zipf RK. MULSIM/NL theoretical and programmer’s manual. Pittsburgh, PA: [45] Mark C. Analysis of longwall pillar stability: An Update. In: Proceedings of the
U.S. Department of the Interior, Bureau of Mines, IC 9321; 1992. workshop on coal pillar mechanics and design. U.S. Department of the
[30] Heasley KAA. New laminated overburden model for coal mine design. In: Interior. U.S. Bureau of Mines, IC 9315; 1992. p. 238–49.
Proceedings of the new technology for ground control in retreat mining. [46] Mark C, Chase F. Analysis of retreat mining pillar stability. In: Proceedings of
NIOSH Inform. Circular 9446; 1997. pp 60–73. the new technology for ground control in retreat mining. NIOSH IC 9446;
[31] Heasley KA. Numerical modeling of coal mines with a laminated displace- 1997. p. 17–34.
ment-discontinuity code. PhD dissertation. Colorado School of Mines; 1998. [47] Ghasemi E, Shahriar K, Sharifzadeh M, Hashemolhosseini H. Quantifying the
187 pp. uncertainty of pillar safety factor by Monte Carlo simulation—a case study.
[32] Mark C, Gale W, Oyler D, Chen J. Case history of the response of a longwall Arch Min Sci 2010;55(3):623–35.
entry subjected to concentrated horizontal stress. Int J Rock Mech Min Sci [48] Iannacchione AT. The effects of roof and floor interface slip on coal pillar
2007;44(2):210–21. behavior. In: Proceedings of the 31st US rock mechanics symposium; 1990.
[33] Larson M, Whyatt J. Deep coal longwall panel design for strong strata: p. 153–60.
The influence of software choice on results. In: Proceedings of the international [49] Molinda GM, Mark C. Coal mine roof rating (CMRR): a practical rock mass
workshop on numerical modeling for underground mine excavation design; classification for coal mines. Pittsburgh, PA: U.S. Department of the Interior,
2009. p. 75–89. Bureau of Mines, IC 9387; 1994. 83 pp.
[34] Jaiswal A, Shrivastava BK. Numerical simulation of coal pillar strength. Int J [50] Mark C, Molinda GM. Rating coal mine roof strength from exploratory drill
Rock Mech Min Sci 2009;46(4):779–88. core. In: Ozdemir L, Hanna K, Haramy KY, Peng S., editors. Proceedings of the
[35] Xiao T, Jianbiao B, Lei X, Xuebin Z. Characteristics of stress distribution in
15th international conference on ground control in mining. Golden, Colo:
floor strata and control of roadway stability under coal pillars. Min Sci Tech
Colorado School of Mines; 1996. p. 415–28.
2011;21(2):243–7.
[51] Bieniawski ZT. Engineering rock mass classifications. New York: Wiley; 1989.
[36] Singh N, Porter I, Hematian J. Finite element analysis of three-way roadway
[52] Mark C, Chase FE, Molinda GM. Design of longwall gate entry systems using
junctions in longwall mining. Int J Coal Geol 2001;45(2-3):115–25.
roof classification. New technology for longwall ground control. In: Proceed-
[37] Islam R, Hayashi D, Kamruzzaman ABM. Finite element modeling of stress
ings of the U.S. Bureau of Mines Technology Transfer Seminar. Pittsburgh, PA:
distributions and problems for multi-slice longwall mining in Bangladesh,
U.S. Department of the Interior, Bureau of Mines; 1994. p. 5–17.
with special reference to the Barapukuria coal mine. Int J Coal Geol 2009;
[53] Bieniawski ZT. A method revisited: coal pillar strength formula based on field
78(2):91–109.
[38] Lianchong L, Tianhong Y, Zhengzhao L, Wancheng Z, Chunan T. Numerical investigations. In: Proceedings of the workshop on coal pillar mechanics and
investigation of groundwater outbursts near faults in underground coal design. Pittsburgh, PA: U.S. Department of the Interior, Bureau of Mines,
mines. Int J Coal Geol 2011;85(3-4):276–88. IC 9315; 1992. p. 158–65.
[39] Jing L. A review of techniques, advances and outstanding issues in numerical [54] Singh AK, Singh R, Maiti J, Kumar R, Mandal PK. Assessment of mining
modeling for rock mechanics and rock engineering. Int J Rock Mech Min Sci induced stress development over coal pillars during depillaring. Int J Rock
2003;40(3):283–353. Mech Min Sci 2011;48(5):805–18.
[40] Lawrence W. A method for the design of longwall gateroad roof support. Int J [55] Jimenez R, Recio D. A linear classifier for probabilistic prediction of squeezing
Rock Mech Min Sci 2009;46(4):789–95. conditions in Himalayan tunnels. Eng Geol 2011;121(3-4):101–9.
[41] Hoek E, Marinos PG. Tunnelling in overstressed rock. In: Vrkljan I, editor. [58] Mitchell TM. Machine learning. Chapter generative and discriminative
Rock Engineering in Difficult Rock Conditions—Soft Rocks and Karst. classifiers: Naive Bayes and Logistic Regression; 2005.
London: Taylor & Francis Group; 2010. p. 49–60. [56] Jordan M. Linear classification: Chapter 7 of an introduction to probabilistic
[42] Singh R, Mandal PK, Singh AK, Kuman R, Sinha A. Coal pillar extraction at graphical models (unpublished manuscript). Department of Statistics,
deep cover: with special reference to Indian coalfields. Int J Coal Geol University of California, Berkeley; 2003.
2011;86(2-3):276–88. [57] Duda R, Hart E, Stork D. Pattern classification. 2nd ed. New York: Wiley;
[43] Singh R, Singh AK, Maiti J, Mandal PK, Singh R, Kumar R. An observational 2000.
approach for assessment of dynamic loading during underground coal pillar [59] Wasserman L. All of statics: a concise course in statistical inference.
extraction. Int J Rock Mech Min Sci 2011;48(5):794–804. New York: Springer; 2004.
[44] Mark C. Pillar design methods for longwall mining. U.S. Department of the [60] Ihaka R, Gentleman RR. A language for data analysis and graphics. J Comp
Interior. U.S. Bureau of Mines, IC 9247; 1990. 53 pp. Graph Stat 1996;5:299–314.

You might also like