Geotech Geol Eng (2012) 30:665–684

DOI 10.1007/s10706-012-9496-3

ORIGINAL PAPER

New Prediction Models for Mean Particle Size in Rock Blast

Fragmentation
P. H. S. W. Kulatilake • T. Hudaverdi •

Qiong Wu

Received: 29 April 2011 / Accepted: 25 January 2012 / Published online: 10 February 2012
 Springer Science+Business Media B.V. 2012

Abstract The paper refers the reader to a blast data obtained similarity groups. The blast data that were
base developed in a previous study. The data base not used for training were used to validate the trained
consists of blast design parameters, explosive param- neural network models. For the same two similarity
eters, modulus of elasticity and in situ block size. A groups, multivariate regression models were also
hierarchical cluster analysis was used to separate the developed to predict mean particle size. Capability
blast data into two different groups of similarity based of the developed neural network models as well as
on the intact rock stiffness. The group memberships multivariate regression models was determined by
were confirmed by the discriminant analysis. A part of comparing predictions with measured mean particle
this blast data was used to train a single-hidden layer size values and predictions based on one of the most
back propagation neural network model to predict applied fragmentation prediction models appearing in
mean particle size resulting from blast fragmentation the blasting literature. Prediction capability of the
for each of the obtained similarity groups. The mean trained neural network models as well as multivariate
particle size was considered to be a function of seven regression models was found to be strong and better
independent parameters. An extensive analysis was than the existing most applied fragmentation predic-
performed to estimate the optimum value for the tion model. Diversity of the blasts data used is one of
number of units for the hidden layer for each of the the most important aspects of the developed models.

P. H. S. W. Kulatilake (&) Keywords Rock mass  Blast fragmentation 

Geological Engineering Program, Department Cluster analysis  Discriminant analysis 
of Materials Science and Engineering, Neural networks  Multivariate regression analysis
University of Arizona, Tucson, AZ 85721, USA
e-mail: [email protected]

T. Hudaverdi
Department of Mining Engineering, Istanbul Technical
University, Maslak, 34469 Istanbul, Turkey
1 Introduction
T. Hudaverdi  Q. Wu
Geological Engineering Program, University of Arizona, Control of the particle size distribution of a muckpile
Tucson, AZ 85721, USA after blasting is always an important subject for mining
industry. Blasting has a significant impact on down-
Q. Wu
Faculty of Engineering, China University of Geosciences, stream processes of mining such as loading, crushing
Wuhan 430074, China and grinding. Improvement of blasting results provides

123
666 Geotech Geol Eng (2012) 30:665–684

increase in loader and excavator productivity due to charge (kg Anfo/m3) are explosive parameters. All
increased diggability capacity, and increased bucket these parameters are also controllable. The third group
and truck fill factors. Suitable and uniform particle size consists of rock mass structure parameters. Number of
distribution results increase in crusher and mill discontinuity sets, orientation, size, spacing and
throughput and decrease in energy consumption in intensity distributions of each discontinuity set belong
size reduction process. Mckenzie (1966) found, in the to the third group. Physical and mechanical properties
studies at Quebec Cartier Mines, that the efficiency of of the intact rock and discontinuities belong to the
all the subsystems of mining is dependent on the fourth group. Density, dynamic compressive strength,
fragmentation (Chakraborty et al. 2004). Today, dynamic tensile strength, shear strength, dynamic
researchers suggest ‘mine to mill’ blasting approach elastic properties, hardness, durability, mineral com-
that is defined as optimization of the blast design to position and grain size of intact rock, and strength,
maximize the overall profitability rather than individ- deformability, roughness and infilling material prop-
ual operations (Kanchibotla et al. 1999; Grundstrom erties of discontinuities belong to the fourth group.
et al. 2001). Additionally, uniform particle size distri- The parameters of the third and fourth groups are
bution also eliminates the need of the secondary uncontrollable.
blasting of the big boulders. The parameters of the aforementioned 4 groups
Several studies have been conducted on blastability should be considered together to explain fragmentation
and prediction of fragmentation. The term blastability process. Because a large number of parameters influ-
refers to the ease with which a rock mass can be ence fragmentation distribution, it is obvious that the
fragmented by blasting and is closely related to fragmentation process is extremely complex and thus it
fragmentation. The parameters that determine frag- is an extremely challenging task to develop models to
mentation by blasting may be divided into four groups: predict fragmentation distribution. Therefore, even
(a) Blast design parameters; (b) Explosive parameters; though some of the fragmentation prediction models
(c) Rock mass structure parameters; (d) Intact rock and that appear in the literature have contributed to
discontinuity physical and mechanical properties. improving the state-of-the-art on the subject, none of
Burden, spacing between boreholes, bench height, them include all the important parameters. In some of
drill-hole diameter, hole length, charge depth, stem the available prediction models crude, highly simpli-
height, subdrilling, drilling pattern (square or stag- fied or inappropriate procedures have been used in
gered), hole inclination (vertical or inclined), blasting estimating rock mass fracture geometry parameters.
direction and blasting sequence (instantaneous or Inappropriate distributions have been used to represent
delayed) are all blast design parameters. All these joint orientation. Corrections for sampling biases have
parameters are controllable. Figure 1 shows most of not been applied in modeling joint size, joint orienta-
the blast design parameters used in a bench blast. The tion and joint intensity. Estimation of fracture spacing
diameter of the drill hole is the most important has been described in a highly vague manner. It is
parameter for any blast design. It influences the important to note that spacing of a fracture set changes
selection of all other parameters. The hole is generally with the direction and the correct spacing is obtained in
drilled slightly below the floor level to obtain a clean the direction perpendicular to the fracture plane. In
breakage. This total length of the hole is known as hole some of the blast fragmentation papers, RQD is used as
length. The extra length of the hole below the floor or a parameter. It is important to note that RQD changes
the grade level is called the sub-drilling. The part of with the direction and thus many values within a wide
the drill hole at the top which is not filled with range exist for RQD for the same rock mass. In situ
explosives is known as stemming height. Some inert block size estimation has not been done in a compre-
material, such as drill cuttings, sand, crushed stone, hensive manner. Therefore, it is important to use better
etc., are used as stemming to contain the explosive and accurate procedures in estimating rock mass
gases in the hole for a slightly longer time to increase fracture geometry parameters in developing rock blast
rock fracturing. The second group consists of explo- fragmentation data bases in the future. Such quality
sive parameters. Explosive type (Anfo, water gel, data bases should then be used to improve the existing
emulsion or dynamite), its density (changes between models or to develop new models to predict rock blast
0.80 and 1.60 g/cm3), strength, resistivity and specific fragmentation distribution.

123
Geotech Geol Eng (2012) 30:665–684 667

Fig. 1 Blast design parameter terminology (Ash 1973)

Because the blast fragmentation distribution depends areas as material sciences (Li et al. 2006), voice
on many parameters, and the process is highly recognition, loan-risk assessment, stock market anal-
complex due to the heterogeneity and anisotropy of a ysis, box office revenue forecasting (Zhang et al.
discontinuous rock mass system, it is impossible to 2009) and military target discrimination. In geosci-
derive an equation for fragmentation distribution ences and geo-engineering, neural networks have been
purely from theoretical and mechanistic reasoning. applied in rock mechanics and rock engineering
In such situations, empirical approaches are used (Zhang et al. 1991; Ghaboussi 1992; Lee and Sterling
incorporating case history data along with statistical 1992), soil engineering (Kung et al. 2007), well-log
based procedures in developing prediction equations and well-test interpretation (Rogers et al. 1992; Al-
for complex geotechnical processes. Multivariate Kaabl and Lee 1993), seismic and satellite image
regression analysis has been used to develop frag- processing (de Groot 1993; Penn et al. 1993),
mentation prediction models (Chakraborty et al. groundwater characterization and remediation (Rizzo
2004; Hudaverdi et al. 2011). However, capturing of and Doughery 1994; Rogers and Dowla 1994),
high non-linearity incorporating many parameters is earthquake intensity prediction (Tung et al. 1994),
a difficult task even with multivariate regression oil reservoir prediction (Yu et al. 2008) and conduc-
analysis. tive fracture identification (Thomas and La Pointe
Due to its excellent ability of non-linear pattern 1995). Neural network approach (NNA) is highly
recognition, generalization, self-organization and self- suitable for systems with highly non-linear complex
learning, the Artificial Neural Network Approach relations between input and output parameters that are
(ANNA) has been proved to be of widespread utility in difficult to develop through physical reasoning and
engineering and is steadily advancing into diverse mathematical modeling. Linking between the rock

123
668 Geotech Geol Eng (2012) 30:665–684

blast mean fragment size and the blast design param- Rosin–Rammler equation seems to be the most
eters, explosive parameters, rock mass structure popular one.
parameters, and intact rock and discontinuity physical It was experienced by many that the rock mass
and mechanical properties is a very complex, non categories defined by Kuznetsov (1973) are very wide
linear process. Therefore, NNA will be highly suitable and need more precision (Chakraborty et al. 2004).
to relate the mean fragment size to the aforementioned Cunningham (1983, 1987) modified the Kuznetsov’s
blast related parameters belonging to the four groups. equation to estimate the mean fragment size and used
Application of NNA to predict rock blast mean the Rosin–Rammler distribution to describe the entire
fragmentation size was dealt with in a previous paper size distribution. The uniformity exponent of Rosin–
(Kulatilake et al. 2010). A summarize account of it is Rammler distribution was estimated as a function of
given in this paper. Needed future research to improve blast design parameters. Rock factor ‘‘A’’ in Kuznet-
the currently existing models are discussed in the sov’s equation was estimated incorporating Lilly’s
paper. blasting index, BI (1986). The final equation suggested
by Cunningham, known as Kuz-Ram model, can be
given as follows:
2 Literature Review
X50 ¼ A  ðV=QÞ0:8  Q0:167  ðE=115Þ0:633 ð3Þ
A previous paper (Kulatilake et al. 2010) has covered where
the literature on the topic to an extensive level by
A ¼ 0:06  BI ð4Þ
referring to the following papers: Ghosh et al. (1990),
Mojtabai et al. (1990), Ouchterlony et al. (1990), and
Chakraborty et al. (1994), Pal Roy (1995),Hagan
BI ¼ 0:5  ðRMD þ JPS þ JPO þ RDI þ SÞ ð5Þ
(1995), Aler et al. (1996), Ozcelik (1998), Jhanwar
et al. (2000), Castro et al. (1998), Latham and Lu In Eq. 3, E is relative weight strength of explosive
(1999), Hamdi and Du Mouza (2005), Hall and (Anfo = 100) and V = BSH where B = burden (m),
Brunton (2002), Latham et al. (2003), Sanchidrian S = blast hole spacing (m) and H = bench height
et al. (2007), Gheibie et al. (2009), and Rustan (1998). (m). In Eq. 5: RMD is rock mass description (powdery
Kuznetsov (1973) has suggested the following empir- or friable = 10, blocky = 20 and massive = 50); JPS
ical equation to predict the mean fragmentation size is joint plane spacing (close \ 0.1 m = 10, 0.1
resulting from rock blasting: - 1.0 = 20, [ 1.0 = 50); JPO is joint plane orienta-
tion (horizontal = 10, dip out face = 20, strike nor-
X50 ¼ AðV=QÞ0:8 Q0:167 ð1Þ
mal to face = 30, dip into face = 40) and RDI is rock
In Eq. 1: X50 is the mean fragment size (cm); ‘A’ is a density influence equal to 25d-50, where d is density
rock factor (7 for medium rock, 10 for hard highly and S is rock strength, equal to 0.05 UCS, where UCS
fissured rock, and 13 for hard weakly fissured rocks); is uniaxial compressive strength. Even though a few
V is the rock volume (m3); Q is the mass of explosive other equations such as SveDefo’s fragmentation
per blast hole (kg). Kuznetsov also has suggested to model (Hjelmberg 1983) and Kou and Rustan’s model
use Rosin–Rammler equation (Rosin and Rammler (1993) are also available in the literature to estimate
1933) given below to estimate the complete fragmen- mean fragmentation size, Kuz-Ram model seems to be
tation distribution resulting from rock blasting: the most popular one.
Research at the JKMRC, Australia and elsewhere
Y ¼ expðX=Xc Þr ð2Þ
has demonstrated that the Kuz-Ram model underes-
In Eq. 2, Y = Proportion of the material larger than X, timates the contribution of fines in the fragment size
Xc = characteristic size = X50 and r = uniformity distribution. Hall and Brunton (2002) claim that the
exponent. Even though Schumann Distribution JKMRC model provides better prediction than Kuz-
(Schuhmann 1959) and Swebrec equation (Nie and Ram model due to improved estimation of the fines to
Rustan 1987) are also suggested in the literature to intermediate size (\ 100 mm) of the fragmentation
predict the complete fragmentation distribution, distribution (Chakraborty et al. 2004). The JKMRC

123
Geotech Geol Eng (2012) 30:665–684 669

model calculates the coarse and fines distributions too large, confined gases may cause ground vibra-
independently. JKMRC uses Kuz-Ram model to tions and back-break. The particle size of the
calculate the course fraction. muckpile may be coarser than expected under such
a situation. The ratio of burden to hole diameter (B/
D) is one of the most important parameters. Ash
3 Used Blast Database and Scope of Study (1973) suggested the ratio of burden to hole
diameter (B/D) as 30 for average conditions. The
In a previous study conducted by the second and first B/D ratio is equal to 25 for low density explosives
authors of this paper (Hudaverdi et al. 2011), many such as Anfo. For the used data, the mean B/D ratio
blasts performed in different parts of the world and is 27.21 with a standard deviation of 4.77. In this
reported in the literature were carefully analyzed and study, the ratio of the bench height to burden (H/B)
put together to create a blast data base to develop is used instead of the ratio of hole length to burden
fragmentation prediction models. A total of 109 blasts (L/B) used by Ash. The ratio of bench height to
were used in this study. Ninety-seven blasts were used burden indicates the stiffness of the rock beam
for model development and twelve blasts were used under blast induced stress. Hustrulid (1999) indi-
for model validation. For details of this blast data base, cated that the H/B ratio is 1.6 or more for most of
the reader is referred to Hudaverdi et al. (2011). the open-pit operations. The mean H/B ratio of the
Five main blast design parameters are used in the data used is 3.44 with a standard deviation of 1.64.
developed multivariate and neural network models. Because the data base was large and diverse, it
They are the Burden (B, m), Spacing (S, m), Bench turned out to be a difficult assignment to find common
height (H, m), Stemming (T, m) and Hole diameter intact rock and rock mass parameters for all the
(D, m). Several blasting researchers have considered selected blast data to use in developing fragmentation
blast design parameters as ratios. In the conducted distribution models. On the other hand, it was possible
studies by the authors of this paper, the blast design to find in situ block size for all the blasts in the data
parameters of all the blast data are also used as base. Therefore, in situ block size which is accepted as
ratios. The ratio of bench height to drilled burden one of the key parameters of the fragmentation process
(H/B), ratio of spacing to burden (S/B), ratio of was used to represent rock mass structure in the data
burden to hole diameter (B/D) and ratio of stem- base. With respect to intact rock, the modulus of
ming to burden (T/B) are the blast design param- elasticity turned out to be the most common parameter
eters used. All blasts in the database were performed available for all the blasts and was used to represent
using Anfo. Therefore, there was no need to use any intact rock properties in the data base. Thus seven
parameter related to explosive type. The Powder parameters were used to establish fragmentation
factor (Pf) has been considered as an explosive prediction models based on multivariate analysis and
parameter. The ratio of spacing to burden is NNA and incorporating the blast design parameters,
determined based on energy coverage of the bench. powder factor, modulus of elasticity (E, GPa) and in
For square pattern, S/B ratio is 1. The mean S/B situ block size (XB, m). Table 1 shows the descriptive
ratio of the used blast data is 1.20. Generally, the statistics of the parameters that were used to develop
ratio of stemming to burden applied is around 1. For multivariate analysis and neural network based frag-
the used data, the mean T/B ratio is 1.27 with a mentation prediction models.
standard deviation of 0.69. Low T/B ratio may cause The cluster analysis was performed on this data to
premature release of explosive gases and result in separate the blast data into two different similarity
fly-rock and inefficient fragmentation. Conversely, groups. The main difference between the two groups
excessive stemming length means low specific was found to be the modulus of elasticity value. The
charge and may cause large boulders. Most of the data belonging to the two groups are given in Tables 1
blast design calculations start with burden determi- and 2 in Hudaverdi et al. (2011), respectively. The
nation. If the burden is too small, detonation gases mean elastic modulus values are 51.14 and 17.99 for
escape to the atmosphere. Escape of the detonation Groups 1 and 2, respectively. Group memberships
gases cause noise and airblast. That means less were then analyzed and confirmed by the discriminant
energy is used for fragmentation. If the burden is analysis. A part of the blast data was used to train

123
670 Geotech Geol Eng (2012) 30:665–684

Table 1 Descriptive statistics of the input parameters used to develop fragmentation prediction models
Minimum Maximum Mean Coefficient. of variation

S/B 1.00 1.75 1.20 0.091

H/B 1.33 6.82 3.44 0.477
B/D 17.98 39.47 27.21 0.175
T/B 0.50 4.67 1.27 0.542
Pf (kg/m3) 0.22 1.26 0.53 0.449
XB (m) 0.02 2.35 1.17 0.409
E (GPa) 9.57 60.00 30.74 0.576
S spacing, B burden, H hole depth, D drill hole diameter, T stemming height, Pf powder factor, XB mean block size, E elastic modulus

Table 2 Summary of discriminant analysis

Function Eigenvalue % of Variance Cumulative % Canonical correlation coefficient

1 18.095 100.0 100.0 0.973

2 3
neural network models for each of the obtained x11 ... x1j ... x1m
similarity groups. The blast data that were not used 6... ... ... ... ... 7
6 7
for training were used to validate the trained neural X¼6
6 xi1 ... xij ... xim 7
7 ð6Þ
network models. Also, multivariate regression analy- 4... ... ... ... ... 5
sis was performed for each of the obtained similarity xn1 ... xnj ... xnm
groups to develop prediction models for mean particle
In matrix X, each row represents an object and each
size.
column shows data for a different variable (Kaufman
and Rousseeuw 1990). For the conducted study, a data
matrix of 97 blast data was formed. In the matrix, the
4 Application of Multivariate Analysis objects are the blasts and the measurable variables are
the 7 blast design and rock mass parameters.
The cluster analysis which is also called segmentation In the data matrix, usually a standardization process
analysis or taxonomy analysis is used to create is applied to weigh each measurable variable equally
relatively homogeneous groups of variables or cases. and to remove the effects of different units of
The cluster analysis identifies a set of groups that measurement across the different variables. Each
minimize within group variation and maximize variable in the data matrix, xij, was standardized by
between-group variation. Several clustering tech- subtracting the column mean, xj , and dividing by the
niques are available: the hierarchical clustering, Kmin column standard deviation, dxj , as shown in Eq. (7):
clustering, two step clustering and fuzzy clustering xij  xj
(Kaufman and Rousseeuw 1990). The hierarchical Zij ¼ ð7Þ
dxj
clustering technique is the most common clustering
technique that is applied in earth science. This The above operation is also called z-score standard-
technique creates relatively homogeneous groups of ization and it produces the Z matrix given in Eq. 8.
cases or objects (the blasts in the conducted research) 2 3
z11 . . . z1j . . . z1m
based on selected variables or characteristics. Each 6... ... ... ... ... 7
object is identified as a separate cluster and then the 6 7
Z¼6 7
6 zi1 . . . zij . . . zim 7 ð8Þ
clusters are combined until only one is left. 4... ... ... ... ... 5
As the first step of the hierarchical cluster analysis, zn1 . . . znj . . . znm
a data matrix (proximity matrix) is formed. If there are
‘n’ objects having ‘m’ measurable variables, ‘n 9 m’ Several hierarchical clustering methods exist: the
data matrix, X, is formed as shown in Eq. 6: median clustering, Ward’s method, nearest neighbor,

123
Geotech Geol Eng (2012) 30:665–684 671

furthest neighbor, average linkage. The cluster method level between the objects ‘a’ and ‘b’. Next, the objects
defines the rules for cluster objects. The basic criterion are arranged into a hierarchy so that objects with the
for any hierarchical clustering is the distance. The highest mutual similarity are placed together to form
similarity between objects is determined based on the clusters. Then the groups having closest similarity to
distance between each other. A small distance indi- other groups are connected together until all of the
cates the two objects are similar, whereas a large objects are placed into a hierarchical tree diagram
distance indicates dissimilarity. The objects that are known as a Dendrogram (Kulatilake et al. 2007).
similar should belong to the same cluster, and objects In this study, the average clustering technique was
that are dissimilar should belong to different clusters. used to form the clusters. The linkage function that
The distance (similarity) between two objects is some determines the distance between two clusters is
function of their measurable variables. computed as the distance between average values of
The distance between clusters may be computed the two clusters (Everitt 1993). Figure 2 shows the
using several functions: the Euclidian distance; Pear- basic logic behind the average clustering technique. In
son correlation distance; Minkowski distance; Block Fig. 2, the clusters X and Y contain the blasts Rc6 and
distance; and Chebychev distance (SPSS 2008). The Mg6, and Mg1, Mg2, Mg3, respectively. ‘dn’ is the
Pearson correlation distance (Pd) was used in this distance between individual blasts. The distance
study to determine the distance between clusters. The between the clusters X and Y (Dxy) is defined as the
Pearson correlation coefficient was calculated using average of the distances between all pairs of individual
z- score values given in matrix Z. The Pearson blasts: Dxy = (d1 ? d2 ? d3 ? d4 ? d5 ? d6)/6.
correlation coefficient between any two series of The dendrogram resulting from the performed
numbers (vectors) za = {za1, … zaj, … zam} and hierarchical cluster analysis (SPSS 2008) is shown in
zb = {zb1, … zbj, …, zbm} is defined as: Fig. 3. The dendrogram shows the relative size of the
m    calculated distance coefficients at which the blasts and
1X zajza zbjzb
r za zb ¼ ð9Þ clusters were combined. The blasts that have smaller
m j¼1 dza d zb distance coefficients are combined. The X axis shows
the blasts. The Y axis shows the rescaled version of the
where zaj indicates the z-score of jth variable for object
calculated distance. The blasts with low distance
‘a’; zbj indicates the z-score of jth variable for object
coefficient (high similarity) are close together.
‘b’; ‘m’ is the number of measured variables on each
As seen in the Dendrogram, the blasts are divided
object. za is the average of the values in vector za; zb is
into two main clusters (groups) between rescaled
the average of the values in vector zb; dza is the
distances of 22 and 25. The first group that appears on
standard deviation of the values in vector za and dzb is
the right side of the dendrogram graphic is given in
the standard deviation of the values in vector zb and
Table 1 in Hudaverdi et al. (2011). The second group
rza zb is the correlation coefficient between the vectors
of blasts that appears on the left side of the dendrogram
za and zb. In this study, objects ‘a’ and ‘b’ are the blasts
is given in Table 2 in Hudaverdi et al. (2011). Group 1
(for example, blasts Rc6 and Mg1). The vectors
contains 35 blasts and Group 2 contains 62 blasts.
za = {za1, … zaj, … zam} and zb = {zb1, … zbj, …,
zbm} are the z-scores of the blast design and rock mass
parameters that belong to blasts ‘a’ and ‘b’, respec-
tively. The Pearson correlation distance was computed
as Pd = 1 - r and varies between 0 (when correlation
coefficient is ?1) and 2 (when correlation coefficient
is -1). A small distance indicates the two objects are
similar, whereas a large distance indicates dissimilar-
ity (Garson 2009).
Computation of a similarity measurement using the
pearson distance between all possible pairs of objects
produces an n 9 n symmetrical matrix M. Each Fig. 2 Distances between blasts belonging to two clusters to
coefficient mza zb in the matrix indicates the similarity calculate average distance between the two clusters

123
672 Geotech Geol Eng (2012) 30:665–684

Fig. 3 Dendrogram of the hierarchical classification for the examined blasts

Fig. 4 Mean vector variables for the two groups

The mean variable vectors of the groups may be investigate the differences between Groups 1 and 2 more
examined to understand which parameters are effec- precisely. The discriminant function used to classify the
tive on the occurrence of two different groups. Figure 4 blasts can be presented based on the unstandardized
shows the mean variable vectors of Groups 1 and 2 that discriminant function coefficients as SPSS (2008):
are shown in Tables 1 and 2 of Hudaverdi et al. (2011),
respectively. As seen in Fig. 4, the main difference L ¼ 4:467ðS=BÞ  0:551ðH=BÞ  0:123ðB=DÞ
between the two groups is the modulus of elasticity þ 1:642ðT=BÞ  3:005ðPfÞ þ 0:309ðXB Þ
value. The mean elastic modulus values are 51.14 and þ 0:208ðEÞ þ 3:577 ð10Þ
17.22 for Groups 1 and 2, respectively. Also, Pf and
H/B values have some differences between the two The discriminant score is the value resulting from apply-
groups. The mean Pf values for Groups 1 and 2 are ing the discriminant function formula (Eq. 10) to the
respectively, 0.41 and 0.60. The mean H/B values for data belonging to a given blast. The histograms shown in
Groups 1 and 2 are respectively, 2.44 and 3.85. Fig. 5 display the discriminant scores for the blasts of
The discriminant analysis technique was applied to the Groups 1 and 2. The sharp difference between the
make sure that the blasts are grouped (clustered) discriminant scores of the two groups indicates strong
correctly. Also, the discriminant analysis enables one to discrimination of the discriminant function and the

123
Geotech Geol Eng (2012) 30:665–684 673

Fig. 5 The discriminant scores of the blasts belonging to the two groups

accuracy of the classification of the blasts into two function and the groups. The canonical correlation is
groups. If there is a hesitation about group membership used to state to what extent the discrimination function is
of a prospective blast, it is possible to assign the blast to useful in determining group differences. For the conducted
Group 1 or 2 by the discriminant function. study, a canonical correlation of 0.973 was obtained. This
A summary of the discriminant analysis is shown in indicates a highly successful discrimination.
Table 2. The second column in Table 2 shows the Table 3 shows the tests of equality of group means.
eigenvalue. Only one eigenvalue exists for each The tests of equality of group means measure each
discriminant function. The eigenvalue, also called independent variable’s potential before the model is
the characteristic root of each discriminant function, developed. Each test displays the results of a one-way
reflects the ratio of importance of the dimensions analysis of variance (ANOVA) for the independent
which classify objects (blasts) of the dependent variable variable using the grouping variable as the factor. If the
(Garson 2009). The larger the eigenvalue, higher the significance value (which is obtained through the F
variance in the dependent variable explained by statistic, and the two degrees of freedom, df1 and df2,
the discriminant function. The third column displays values) is larger than 0.10, it indicates that the
the percent of variance explained by each function. The parameter is not effective on occurrence of group
fourth column is the cumulative percent of variance (SPSS 2008). Accordingly, the S/B has no effect on
explained. Thus a more general version of Table 2 can occurrence of the groups. This finding agrees with the
be used to understand the relative importance of results appearing on Fig. 4. The effect of a parameter
discriminant functions if more than one function exists.
Since we sorted the blasts into only two groups, only
Table 3 Results of tests of equality of group means
one discriminant function exists for the conducted
study. Thus, the canonical correlation coefficient is the Wilks’ lambda F df1 df2 Sig.
only important indicator in Table 2 for the conducted S/B 0.988 0.97 1 88 0.280
discriminant analysis. The canonical correlation is a
H/B 0.826 16.55 1 88 0.000
measure of the association between the groups formed
B/D 0.965 2.51 1 88 0.068
by the dependent variable and the given discriminant
T/B 0.972 3.69 1 88 0.103
function. When the canonical correlation coefficient is
Pf 0.841 15.17 1 88 0.000
zero, there is no relation between the groups and the
XB 0.932 10.59 1 88 0.010
function. When the canonical correlation is close to 1,
E 0.161 540.77 1 88 0.000
there is a high correlation between the discriminant

123
674 Geotech Geol Eng (2012) 30:665–684

on the occurrence of the groups indicates only the

discrimination ability of the selected parameter in
cluster and discriminant analysis. It does not indicate S/B

the effect of the parameter on the fragmentation H/B

process. The discrimination ability of a parameter is B/D
highly related to its variance. For example, in our
T/B X50
analysis, the range of the S/B parameter is relatively
tight. On the other hand, the S/B is an important Pf

parameter for blasting and site engineers are very XB

meticulous in selecting the value for the S/B parameter. E
Wilks’ lambda is another measure of a variable’s
potential. Wilks’ lambda is a test statistic used in
multivariate analysis of variance. It tests whether there
are differences between the means of the groups of the
Input Layer Hidden Layer Output Layer
subjects on a combination of dependent variables (Everitt (7 neurons) (N neurons) (1 neuron)
and Dunn 1991; Huberty and Olejnik 2006). Smaller
values indicate the variable is better at discriminating Fig. 6 The structure of a 7-N-1 BP neural network
between groups. Wilks’ lambda test results show that the
parameter that has a dominant effect on the occurrence of et al. 1995; Maier and Dandy 1998). Figure 6 shows the
the two groups is the modulus of elasticity. BP network configuration used in this study assuming
the optimum number of hidden units as N. Section 5.2
deals with estimation of N in great detail.
5 Application of Artificial Neural Network Kulatilake et al. (2010) covered in detail the
Approach (ANNA) equations associated with the information transfer
between the input layer and the hidden layer as well as
5.1 Setting Up and Training of ANNA between the hidden layer and the output layer. In
addition, the same paper covered in detail the equa-
The back-propagation (BP) network, a multilayer feed- tions associated with the training of the network.
forward ANNA, is perhaps the most popular network As stated before, the blasting data have been divided
architecture today as it contains the highlights of the into two groups by the value of elastic modulus. To
neural network theory, simple in structure and clear in increase the prediction precision, BP neural network
mathematical meaning. It has been proved that any was applied separately to each group. For group 1,
continuous function can be uniformly approximated by thirty-five sets of data given in Table 1 were used to
BP network model with only one hidden layer (Cybenko train the network and the five sets of data given in
1989). So a single-hidden layer BP network is used in Table 4 were used to predict and validate the network.
this paper to predict the mean particle size of rock frag- For group 2, sixty-two sets of data given in Table 2 were
mentation resulting from blasting. As stated previously, used to train the network and the seven sets of data given
the mean particle size X50 is considered to be a function in Table 5 were used to predict and validate the network.
of seven independent parameters. Consequently, the As their orders of magnitude are different, before
parameters S/B, H/B, B/D, T/B, Pf, XB and E are used as running the neural networks, the original data were
inputs and X50 as the output in the BP network model. In normalized using Eq. 11 given below:
the literature different opinions are expressed with
xi  xmin
respect to designing the neural network structure with yi ¼ ð11Þ
xmax  xmin
respect to the number of nodes and the weights to obtain
accurate performance from a trained network for a given In Eq. 11, x is the vector before normalization; y is the
number of training samples. This aspect was discussed vector after normalization; xi and yi are respectively
in detail in Kulatilake et al. (2010). Research conducted the element of vector x and vector y; xmax and xmin are
in the past has shown that the number of hidden units has respectively, the maximum and minimum element of
a great impact on the ANNA prediction results (John vector x.

123
Geotech Geol Eng (2012) 30:665–684 675

Table 4 Prediction results of the 8 simulations for Group 1 (for N = 9)

Blast no. X50 (m) X50R (m) X50K (m) X50N (BP Neural Network) (m)
1 2 3 4 5 6 7 8 l d

En13 0.47 0.39 0.44 0.39 0.42 0.42 0.42 0.42 0.42 0.42 0.42 0.42 0.0255
RU7 0.64 0.51 0.65 0.96 0.96 0.38 0.23 0.38 0.96 0.96 0.24 0.63 0.5571
Mg8 0.44 0.40 0.39 0.37 0.37 0.37 0.37 0.37 0.37 0.37 0.37 0.37 0.0000
Mg9 0.25 0.24 0.30 0.26 0.27 0.27 0.27 0.30 0.28 0.28 0.28 0.28 0.0430
Rc1 0.46 0.52 0.72 0.48 0.31 0.41 0.47 0.48 0.36 0.47 0.49 0.43 0.1544
X50 measured mean particle size (m), X50K mean particle size based on Kuznetsov’s equation (m), X50N predicted mean particle size
based on neural network model (m), X50R mean particle size based on developed regression model (m)

Table 5 Prediction results of the 8 simulations for Group 2 (for N = 7)

Blast no. X50 (m) X50R (m) X50K (m) X50N (BP neural network) (m)
1 2 3 4 5 6 7 8 l d

Mr12 0.20 0.16 0.24 0.12 0.32 0.12 0.17 0.22 0.12 0.22 0.12 0.18 0.4093
Db10 0.35 0.16 0.08 0.17 0.28 0.16 0.17 0.20 0.74 0.20 0.74 0.33 0.7621
Sm8 0.18 0.19 0.35 0.19 0.19 0.19 0.19 0.19 0.19 0.19 0.19 0.19 0.0000
Oz8 0.23 0.17 0.20 0.18 0.14 0.14 0.12 0.13 0.15 0.13 0.15 0.14 0.1268
Oz9 0.17 0.16 0.23 0.19 0.19 0.18 0.22 0.19 0.17 0.19 0.17 0.19 0.0853
Ad23 0.21 0.18 0.11 0.22 0.19 0.23 0.17 0.22 0.22 0.22 0.22 0.21 0.0919
Ad24 0.20 0.14 0.13 0.19 0.21 0.19 0.17 0.25 0.23 0.25 0.23 0.22 0.1332
X50 measured mean particle size (m), X50K mean particle size based on Kuznetsov’s equation (m), X50N predicted mean particle size
based on neural network model (m), X50K mean particle size based on developed regression model (m)

Several algorithms are available in the literature to to minimize the objective function value given by
train a neural network. Each of them has its advantages Eq. 12 and thus to arrive at an optimized trained
and disadvantages. For a given problem, it is difficult to network. In a previous paper (Kulatilake et al. 2010),
say which one works best. It depends on several factors, the best training method was decided by trying
such as the complexity of the problem, the number of different methods and observing the performance of
training samples, the structure of the network, error each method on a plot between mean square error
target and so on. The information flows through the (mse) value and number of training cycles. The
network from the input layer to the output layer via the training of the network was stopped after it has been
hidden layer. The objective of the training is to adjust the trained for many cycles to reach a stable mse value.
weights and thresholds that exist between the input layer Four training methods were used to train the same
and the hidden layer, and the hidden layer and the output selected network. The LM algorithm showed the
layer to develop and estimate a complicated non-linear highest stability among the four training algorithms.
function between the output and input variables. The Also it reached the global minimum with the lowest
objective function given in Eq. 12 is used to obtain an number of training cycles. Therefore, only the LM
optimized trained network. algorithm was used for further modeling work with
neural networks.
1X T
mse ¼ ðy  Ct Þ2 ð12Þ
T t¼1 t
5.2 Procedure to Estimate Number of Units
In Eq. 12, yt is the expected output and Ct is the for the Hidden Layer
calculated output; T is the number of data sets used in
the training sample. The weights and thresholds are Choosing an appropriate number for the units in the
adjusted using the gradient decreased learning method hidden layer is not a straightforward task (Maier and

123
676 Geotech Geol Eng (2012) 30:665–684

 
Dandy 1998). The number of input parameters, errorij ¼ eij  rij  ð14bÞ
number of output parameters, number of data sets
available and the characteristics of the functional In Eq. 14b, eij denotes the prediction result of the ith
relation between the output and the input parameters network under a certain N for the jth blast number, rij
may affect the optimum number for the units in the denotes the corresponding actual value for the same
hidden layer. At present, the authors are not aware of blast number. In Eq. 14a, J is the number of blast data
any accepted procedure or formula available to used for prediction for a certain group. The correlation
determine the aforesaid optimum number. This opti- coefficient between the predicted value and the
mum number may even change with different run measured value for the aforementioned J blast data
(simulation) numbers for the same problem. Two was used as the second parameter to evaluate the
empirical formulae available in the literature were accuracy of each ith network under a certain N value.
used in Kulatilake et al. (2010) to estimate the In evaluating the accuracy, several random simula-
optimum number for the hidden layer units. tions were performed for each ith network under a
Based on Kolmogorov’s theorem, Hecht-Nelson certain N value.
(1987) has suggested that 2n ? 1 (where n is the
number of input parameters) should be used as the 5.3 Results, Prediction and Validation
upper bound for the number of hidden units for a one-
hidden-layer back-propagation network. Because in For group 1, five blasts were used for the prediction
our study n = 7, the number of hidden units for both and validation. Note that under each N value, 8
Groups1 and 2 should be B15 according to Hecht- simulations were made. As an example, the prediction
Nelson’s suggestion. According to the second empir- obtained for each blast under N = 9 for each of the
ical formula (Ge and Sun 2007), the number of hidden simulations made is shown in Table 4. En13 blast has
units, N, should satisfy the following inequality: the same values of S/B, H/B, B/D, T/B, Pf, XB and E as
for En4 blast. Therefore, the prediction result of X50
X
n for En13 blast is almost the same as the value for En4
CiN [ k ð13aÞ blast. For RU7 & RU1, all the blasting parameter
i¼0
values are the same apart from the value for T/B. That
where has led to a large variation of the predicted value with
respect to the simulation number. Table 4 also
N!
CiN ¼ ð13bÞ provides the predicted mean, l, and coefficient of
i!ðN  iÞ!
variation, d, obtained for each blast from the 8
In inequality Eq. 13a, n is the number of input simulations.
parameters and k is the number of data sets used. The RMSE values and the coefficient of variations
Note that If i [ N, CiN ¼ 0. Application of inequality obtained for Group 1 for different N values are given
Eq. 13a to group 1 (n = 7, k = 35) and group 2 in Table 6. High correlation coefficient values indicate
(n = 7, k = 62) results in N C 6 for both groups. predictions close to the measured values. The consis-
Therefore, use of the aforementioned two empirical tency of the correlation coefficient values shows high
criteria results in 6 B N B 15 for both Groups 1 and 2. homogeneity of the Group 1 samples. N = 9 has
Accuracy of the network was considered to deter- resulted in the lowest RMSE and the highest correla-
mine the optimum value for N. To evaluate the tion coefficient. That means for group 1, N = 9 is the
accuracy of the network for each N, two parameters optimum value. Table 4 shows a comparison between
were used. The Root Mean Square Error, RMSE, was neural network predictions, measured values and
used as the first parameter and it was defined by the predictions based on the Kuznetsov’s equation. All
following equation: the blast data were examined carefully and the rock
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi factor ‘A’ was estimated for each blast to apply the
PJ
j¼1 errorij
2 Kuznetsov’s equation. For all 5 blasts, neural network
RMSEi ¼ ð14aÞ predictions are close to the measured values. This can
J
be also seen from the regression analysis results given
where in Fig. 7a. For 4 out of the 5 blasts, predictions based

123
Geotech Geol Eng (2012) 30:665–684 677

Table 6 Prediction results of mean particle size from ANNA for Group 1 (for N = 6–15)
Blast no En13 RU7 Mg8 Mg9 Rc1 Correlation RMSE
coefficient (with X50)

X50 (m) 0.47 0.64 0.44 0.25 0.46 1.00

N=6
l 0.42 0.86 0.37 0.27 0.40 0.90 0.1093
d 0.0169 0.1595 0.0079 0.0479 0.3480
N=7
l 0.42 0.56 0.37 0.27 0.54 0.88 0.0642
d 0.0275 0.5837 0.0127 0.0410 0.4018
N=8
l 0.42 0.74 0.37 0.28 0.51 0.93 0.0645
d 0.0000 0.4460 0.0056 0.0698 0.3098
N=9
l 0.42 0.63 0.37 0.28 0.43 0.96 0.0429
d 0.0284 0.5553 0.0071 0.0275 0.1549
N = 10
l 0.42 0.64 0.37 0.27 0.58 0.87 0.0666
d 0.0003 0.5480 0.0031 0.0249 0.2648
N = 11
l 0.42 0.78 0.37 0.28 0.60 0.89 0.0974
d 0.0146 0.4193 0.0036 0.0272 0.1477
N = 12
l 0.42 0.61 0.37 0.28 0.39 0.94 0.0531
d 0.0000 0.5691 0.0069 0.0584 0.2199
N = 13
l 0.42 0.67 0.37 0.28 0.46 0.95 0.0431
d 0.0000 0.5472 0.0065 0.0488 0.3262
N = 14
l 0.42 0.60 0.37 0.27 0.55 0.77 0.0592
d 0.0000 0.5611 0.0019 0.0420 0.1954
N = 15
l 0.42 0.74 0.37 0.26 0.54 0.93 0.0691
d 0.0000 0.4175 0.0000 0.0602 0.2556
X50 measured mean particle size (m)

on Kuznetsov’s equation are close to the measured coefficient of variation, d, obtained for each blast from
values. This can be also seen from the regression the 8 simulations. The RMSE values and the coeffi-
analysis results given in Fig. 7b. Note that Group 1 cient of variations obtained for Group 2 for different N
blast data come from hard rocks that have high elastic values are given in Table 7. The results show high
modulus values. fluctuation of correlation coefficient values for Group
For Group 2, seven blasts were used for the 2 data. This shows that the homogeneity of Group 2 is
prediction and validation. Note that under each N weaker than that of Group 1. N = 7 has resulted in the
value, 8 simulations were made. As an example, the lowest RMSE value and the highest correlation
prediction obtained for each blast under N = 7 for coefficient. That means for Group 2, N = 7 is the
each of the simulations made is shown in Table 5. The optimum value. Table 5 shows a comparison between
same table also provides the predicted mean, l, and neural network predictions, measured values and

123
678 Geotech Geol Eng (2012) 30:665–684

Fig. 7 Predicted mean particle size (m) versus measured mean particle size (m): a based on neural network models; b based on
Kuznetsov’s equation; c based on developed regression models

123
Geotech Geol Eng (2012) 30:665–684 679

Table 7 Prediction results of mean particle size from ANNA for Group 2 (for N = 6–15)
Blast no Mr12 Db10 Sm8 Oz8 Oz9 Ad23 Ad24 Correlation RMSE
coefficient (with X50)

X50 (m) 0.20 0.35 0.18 0.23 0.17 0.21 0.20 1.00
N=6
l 0.19 0.19 0.18 0.14 0.17 0.20 0.19 0.11 0.0834
d 0.5159 0.5700 0.1365 0.1059 0.1287 0.2082 0.1551
N=7
l 0.18 0.33 0.19 0.14 0.19 0.21 0.22 0.81 0.0425
d 0.4093 0.7621 0.0000 0.1268 0.0853 0.0919 0.1332
N=8
l 0.28 0.30 0.19 0.14 0.21 0.25 0.20 0.49 0.0640
d 0.8467 0.8615 0.0005 0.0322 0.3374 0.3733 0.0956
N=9
l 0.41 0.43 0.19 0.14 0.20 0.29 0.20 0.59 0.1136
d 0.6653 0.6185 0.0000 0.0495 0.0832 0.6729 0.0409
N = 10
l 0.33 0.33 0.19 0.14 0.19 0.26 0.20 0.52 0.0767
d 0.7884 0.7237 0.0000 0.0089 0.3196 0.3535 0.1177
N = 11
l 0.31 0.35 0.17 0.13 0.16 0.19 0.19 0.68 0.0671
d 0.7188 0.7518 0.1878 0.0681 0.1654 0.2405 0.2480
N = 12
l 0.31 0.45 0.19 0.14 0.18 0.29 0.22 0.78 0.0865
d 0.7063 0.6182 0.0000 0.0060 0.1346 0.4057 0.1344
N = 13
l 0.39 0.36 0.19 0.14 0.17 0.23 0.20 0.49 0.0951
d 0.7322 0.6801 0.0000 0.0050 0.1009 0.1552 0.0909
N = 14
l 0.39 0.38 0.18 0.14 0.16 0.21 0.19 0.57 0.0955
d 0.6019 0.7096 0.1361 0.0714 0.1654 0.2076 0.1746
N = 15
l 0.21 0.30 0.18 0.14 0.17 0.20 0.19 0.79 0.0484
d 0.6902 0.6597 0.1365 0.0545 0.1518 0.1937 0.1562
X50 measured mean particle size (m)

predictions based on the Kuznetsov’s equation. For all Figure 7a shows the linear regression analysis
7 blasts, neural network predictions are close to the performed between the predictions obtained from the
measured values. This can be also seen from the neural network models developed for Groups 1 and 2
regression analysis results given in Fig. 7a. Only for and the measured mean particle size. Figure 7b shows
about 50% of the blasts, predictions based on the linear regression analysis performed between the
Kuznetsov’s equation are close to the measured predictions based on Kuznetsov’s equation for Groups
values. This can be also seen from the regression 1 and 2 and the measured mean particle size. In
analysis results given in Fig. 7b. Note that group 2 Fig. 7a, the prediction line has an intercept close to
blast data come from rocks that have relatively low zero and a slope close to 1.0 with a R2 value of 0.9407
elastic modulus values. (which indicates a strong regression fit). These results

123
680 Geotech Geol Eng (2012) 30:665–684

Table 8 Regression statistics obtained for Eq. 15

R R2 Adjusted R2 Standard error Observations

Model summary
0.841 0.708 0.632 0.0916 35
Sum of squares df Mean square F Significance

Analysis of variance (ANOVA)

Regression 0.551 7 0.079 9.356 0.000
Residual 0.227 27 0.008
Total 0.778 34

indicate that the matching between the neural network X50 ¼ 208ðS=BÞ2:788 ðH=BÞ0:112 ðB=DÞ0:027
predictions and the measured values is very strong. In
Fig. 7b, even though the prediction line has an  ðT=BÞ0:321 ðPfÞ0:360 ðXB Þ0:233 ðEÞ1:802
intercept close to zero and a slope close to 1.0, the ð15Þ
R2 value is only 0.5697 (which indicates only a
R, the multiple correlation coefficient, is the linear
moderate level regression fit). In addition, the 95%
correlation between the observed and model-predicted
confidence band in Fig. 7a is much narrower than that
values of the dependent variable. Its large value (close
in Fig. 7b. These results clearly show that the neural
to 1) indicates a strong relation. R2, the coefficient of
network predictions are better than the predictions
determination, is the squared value of the multiple
based on Kuznetsov’s equation.
correlation coefficient. R2 is the percent of variance in
the dependent variable explained collectively by all of
the independent variables. R2 value close to 1 also
6 Prediction of Mean Particle Size Based indicates importance of regression. The regression row
on Multivariate Regression Analysis in Tables 8 and 9 provide information about the
variation accounted by the regression model. The
The multiple regression analysis (Draper and Smith residual row displays information about the variation
1981) was applied to develop a prediction equation for that is not explained by the regression model (Draper
each group. The dependent variable of the multiple and Smith 1981; Montgomery et al. 2006). For
regression analysis is the mean particle size (x50R) and example, the sum of squares values given in Table 8
the independent variables are the all blast design show that over seventy percent of the variance in the
parameters, elastic modulus and in situ block size. mean particle size (x50R) is explained by the regression
Equation 15 given below was developed for Group model. The F test is applied to test the significance of
1 that has high Young’s modulus values. Table 8 the regression model. If the significance value of the F
shows the obtained regression statistics. statistic is less than 0.05, it means that the variation

Table 9 Regression statistics obtained for Eq. 16

R R2 Adjusted R2 Standard error Observations

Model summary
0.859 0.739 0.705 0.1119 62
Sum of squares df Mean square F Significance

Analysis of variance (ANOVA)

Regression 2.178 7 0.311 22.808 0.000
Residual 0.770 54 0.014
Total 2.948 61

123
Geotech Geol Eng (2012) 30:665–684 681

explained by the model is not due to chance. In other This will result in less fragmentability and higher
words, the null hypothesis of no linear relationship of mean particle size.
x50R to the 7 independent variables is rejected. Table 8 Equations 15 and 16 were applied respectively, to
shows a significance value of very close to zero based the 5 blasts shown in Table 4 and the 7 blasts shown in
on the F and the degrees of freedom (df) value Table 5 to predict mean particle size based on the
calculated. That indicates the importance of the developed regression equations. The values obtained
developed regression equation for Group 1. are shown in Tables 4 and 5, respectively. For all 5
The equation given below was developed for Group blasts belonging to Group 1, the regression based
2 that has low elastic modulus values. Table 9 shows predictions are close to the measured values. For the 7
the regression statistics obtained for Eq. 16. Again a blasts belonging to Group 2, apart from DB10, for the
significance value of very close to zero was obtained rest, the regression based predictions are close to the
under ANOVA results. All these values indicate that measured values. Figure 7c shows the regression
the regression is important and strong for Group 2. analysis performed between the predictions based on
the regression equation and the measured mean
X50 ¼ 0:60ðS=BÞ0:547 ðH=BÞ0:535 ðB=DÞ0:427 particle size. Even though the intercept of the predic-
 ðT=BÞ0:101 ðPfÞ0:115 ðXB Þ0:434 ðEÞ1:202 tion line is almost zero, the slope (equal to 0.86) is
ð16Þ slightly off from 1.0. However, the R2 value of 0.82
indicates a strong regression fit and the 95% confi-
In Eqs. (15) and (16), the exponents obtained for S/B, dence band is much tighter than the one appears in
B/D and XB are positive. That means, the mean Fig. 7b. Comparison of Fig. 7b, c shows that the
particle size increases with increasing S/B, B/D and regression based predictions have better reliability
XB. The exponents obtained for T/B and PF are than the predictions based on Kuznetsov’s equation.
negative. That means, the mean particle size decreases Comparison of Fig. 7a, c shows that the neural
with increasing T/B and PF. These results can be network predictions are better than the predictions
reasoned out easily intuitively. For both rock groups, based on developed multivariate regression models.
the obtained exponent value for S/B is high in the
regression models. It indicates that S/B is an important
parameter for mean particle size prediction models. 7 Discussion
For the high modulus rock group, it has turned out to
be the most important parameter. For the low modulus Note that even though both the multivariate regression
rock group, it has turned out to be the second most models and neural network models are non-linear
important parameter The coefficients associated with models, the neural network models can be considered
the modulus of elasticity are negative in Eqs. (15) and as more advanced non-linear models than multivariate
(16). This means increase of the elastic modulus regression models. It is important to note that neural
results in decrease of the mean particle size. The network results do not provide a unique answer. The
modulus of elasticity is an indicator of rock stiffness. results depend on the factors such as network geom-
In the developed models, if the stiffness of rock etry, internal parameters of the learning algorithm and
increases the fragmentability of rock increases. The the simulation number. The deviation associated with
energy transmission velocity increases with increasing the simulation number can be reduced by computing
Young’s modulus. This in turn increases the frag- the mean value coming out of several simulations as
mentability and results in lowering the mean particle done in this paper. For engineering and science
size. In the developed model, H/B is positively problems, it is a difficult task to find large data bases
correlated to mean particle size. This result may be that are suitable to use in neural network modeling.
explained as follows. H/B can be increased by Therefore, as shown in this paper, attempts should be
increasing the H value and keeping the B value made to find the optimum network geometry and the
constant. The energy transmitted volume can be best learning algorithm to obtain the best possible
expected to increase with increasing H. If the same results for problems having a limited number of data.
total energy is used under increasing H, it will result in Best learning algorithms can be obtained as shown in
decreasing the energy transmitted per unit volume. the paper through numerical experimentation to

123
682 Geotech Geol Eng (2012) 30:665–684

minimize the mse between the predicted value and the 2010) to train a single-hidden layer back propagation
expected value and to maximize the training speed and neural network model to predict mean particle size
the stability of the calculated mse with number of resulting from blast fragmentation for each of the
training cycles. There is no universally accepted obtained similarity groups. The mean particle size was
theoretical basis for choosing the network geometry. considered to be a function of seven independent
Therefore, in practical use, it should be obtained parameters. It turned out to be a difficult assignment to
through numerical experimentation as shown in the find common intact rock and rock mass parameters for
paper to minimize the RMSE obtained between the all the selected blast data to use in developing
prediction and the measured value. This will increase fragmentation distribution models. On the other hand,
the workload when using the neural network approach. it was possible to find in situ block size for all the
The learning and memory ability of a neural network blasts in the data base. Therefore, in situ block size was
depend on the training samples used. Therefore, if new used to represent rock mass structure in the developed
data become available, to obtain accurate predictions, models. With respect to intact rock, the modulus of
the network has to be rebuilt again from the very elasticity turned out to be the most common parameter
beginning. available for all the blasts and was used to represent
Researchers use different procedures in estimating intact rock properties in the developed models. It was
in situ block size. A wide variation is possible for the possible to incorporate most of the important blast
determination technique of the in situ block size. In the design parameters in the developed models.
future, attempts should be made to provide uniformity Capability of the developed neural network models
in estimating the in situ block size to increase was determined by comparing neural network predic-
accuracy. At present, the developed models incorpo- tions with measured mean particle size and the
rate elastic modulus to represent the intact rock. In the predictions based on one of the most applied frag-
future, attempts should be made to determine addi- mentation prediction models appearing in the blasting
tional intact rock parameters such as uniaxial com- literature. Prediction capability of the trained neural
pressive strength, tensile strength, alteration and network models was found to be strong and better than
density, and discontinuity geometry parameters of the most applied fragmentation prediction model. For
the rock mass that would be subjected to blasting. It the same two similarity groups, multivariate regres-
would be interesting to study whether the similarity sion models were also developed to predict mean
groups emerging from the results of cluster and particle size. The prediction capability of the multi-
discriminant analyses change due to incorporation of variate regression models was also found to be strong
additional parameters in the modeling procedure. This and better than the most applied fragmentation
paper has looked into only the mean particle size prediction model. The prediction capability of the
resulting from blasting. In a future research, it is neural network models seems to be superior to that of
important to extend this research to cover the particle multivariate regression models for the used data. No
size distribution. other study reported in the literature has used a large
data base as that used in this study. Therefore, the
diversity of the blasts data base is one of the strongest
8 Summary and Conclusions features of the developed models. The variety of the
blasts is also an important element that increases the
In a previous paper (Hudaverdi et al. 2011), many versatility and reliability of the developed models.
blasts performed in different parts of the world and The developed neural network models as well as
reported in the literature were put together to create a multivariate regression models are not complex and
blast data base to develop fragmentation distribution are suitable for practical use at mines. As a result of
models. A hierarchical cluster analysis was used to this study, two different neural network models and
separate the blasts data into two different groups of two different multivariate regression models were
similarity based on the intact rock stiffness. The group developed to predict mean particle size resulting from
memberships obtained from cluster analysis was blasting. This provides an opportunity to use a
confirmed by a discriminant analysis. A part of this different prediction model in accordance with the
blast data was used in another study (Kulatilake et al. value of modulus of elasticity of intact rock.

123
Geotech Geol Eng (2012) 30:665–684 683

Application of the developed prediction models to Garson GD (2009) Statnotes: topics in multivariate analysis.
new blasts will test the reliability of them. Attempts http://www2.chass.ncsu.edu/garson/pa765/statnote.htm. 24
Feb 2009
should be made to enlarge the blast database that will Ge ZX, Sun ZQ (2007) Neural network theory and MATLAB
be used to develop fragmentation prediction models R2007 application. Publishing House of Electronics
presented in this study. Neural network and multivar- Industry, Beijing, pp 108–122, 48–50
iate statistical modeling procedures used in this paper Ghaboussi J (1992) Potential applications of neuro-biological
computational models in Geotechnical engineering. In:
have shown the capability of developing new frag- Proceedings of the 4th international symposium on
mentation prediction models. numerical models in Geotech, Swansea, UK, pp 543–555
Gheibie S, Aghababaei H, Hoseinie SH, Pourrahimian Y (2009)
Acknowledgments This study was partially supported by the Modified Kuz–Ram fragmentation model and its use at the
Research Fund of the Istanbul Technical University (project Sungun Copper Mine. Int J Rock Mech Min Sci 46:967–973
name: ‘the investigation of environmentally friendly blast Ghosh A, Daemen JJK, Vanzyl D (1990) Fractal based approach
designs for improvement of fragmentation in Istanbul region to determine the effect of discontinuities on blast frag-
quarries’). The authors are grateful to the Research Fund of the mentation. In: Proceedings of the 31st U.S. symposium on
Istanbul Technical University for their financial support. rock mechanics, Balkema, Rotterdam, pp 905–912
Grundstrom C, Kanchibotla S, Jankovic A, Thornton DM (2001)
Blast fragmentation for maximizing the SAG mill
throughput at Porgera Goldmine. In: Proceedings of the
References 27th annual conference on explosives and blasting tech-
nique, Orlando, Florida, pp 383–399
Aler J, Du Mouza J, Arnould M (1996) Measurement of the Hagan TN (1995) The effect of rock properties on the design and
fragmentation efficiency of rock mass blasting and its results of tunnel blasts. J Rock Mech Tunnel Tech 1(1):25–39
mining applications. Int J Rock Mech Min Sci Geomech Hall J, Brunton I (2002) Critical comparison of Kruttschnitt
Abstr 33:125–139 Mineral Research Center (JKMRC) blast fragmentation
Al-Kaabl AU, Lee WJ (1993) Using artificial neural nets to models. Fragblast 6(2):207–220, Swets and Zeitlinger, The
identify the well-test interpretation model. SPE Form Eval Netherlands
8:233–240 Hamdi E, Du Mouza J (2005) A methodology for rock mass
Ash RL (1973) The influence of geological discontinuities on characterization and classification to improve blast results.
rock blasting. Ph.D. Thesis, University of Minnesota, 289 p Int J Rock Mech Min Sci 42:177–194
Castro JT, Liste AV, Gonzalez AS (1998) Blasting index for Hecht-Nelson R (1987) Kolmogorov’s mapping neural network
exploitation of aggregates. In: Singhal RK (ed) Proceed- existence theorem. In: Proceedings of the 1st IEEE annual
ings of the 7th mine planning and equipment selection international conference on neural networks. IEEE Press,
symposium, Oct, 6–9, 1998, Calgary, pp 165–168 San Diego, Piscataway, NJ, pp III: 11–14
Chakraborty AK, Jethwa JL, Paithankar AG (1994) Effects of Hjelmberg H (1983) Some ideas on how to improve calculations
joint orientation and rock mass quality on tunnel blasting. of the fragment size distribution in bench blasting. In:
Engg Geol 37:247–262 Proceedings of the 1st ınternational symposium on rock
Chakraborty AK, Raina AK, Ramulu M, Choudhury PB, Haldar fragmentation by Blasting, Lulea University Technology
A, Sahu P, Bandopadhyay C (2004) Parametric study to Lulea, Sweden, pp 469–494
develop guidelines for blast fragmentation improvement in Huberty CJ, Olejnik S (2006) Applied MANOVA and Dis-
jointed and massive formations. Engg Geol 73:105–116 criminant Analysis, 2nd edn. Wiley, New Jersey
Cunningham CVB (1983) The KuzRam model for prediction of Hudaverdi T, Kulatilake PHSW, Kuzu C (2011) Prediction of
fragmentation from blasting. In: Holmberg R, Rustan A blast fragmentation using multivariate analysis procedures.
(eds) Proceedings of 1st international symposium on rock Int J Num Anal Meth Geomech 35(12):1318–1333
fragmentation by Blasting, Aug, 22–26, 1983, Lulea, Hustrulid W (1999) Blasting principles for open pit mining.
pp 439–453 A. A. Balkema, Rotterdam
Cunningham CVB (1987) Fragmentation estimations and SPSS Inc (2008) SPSS Base 16.0 User’s Guide
KuzRam model–four years on. In: Proceedings of 2nd Jhanwar JC, Jethwa JL, Reddy AH (2000) Influence of air-deck
symposium on rock fragmentation by Blasting, Keystone, blasting on fragmentation in jointed rocks in an open-pit
Colorado, pp 475–487 manganese mine. Engg Geol 57:13–29
Cybenko G (1989) Approximation by superpositions of a sig- John FCK, Lim BS, Lennie ENL (1995) Optimal design of
moidal function. Math Control Syst Signal 2:303–314 neural networks using the Taguchi method. Neurocom-
De Groot PFM (1993) Reservoir characterization from 3-D puting 225–245
seismic data using artificial neural networks and stochastic Kanchibotla SS, Valery W, Morrell S (1999) Modeling fines in
modeling techniques. AAPG Bull 77:1617–1618 blast fragmentation and its impact on crushing and grind-
Draper NR, Smith H Jr (1981) Applied regression analysis, 2nd ing. In: Proceedings of the Explo-99 Conference, Kal-
edn. Wiley, New York, NY goorlie, pp 137–144
Everitt BS (1993) Cluster analysis. Edward Arnold, London Kaufman L, Rousseeuw PJ (1990) Finding groups in data: an
Everitt BS, Dunn G (1991) Applied multivariate data analysis. introduction to cluster analysis. Wiley-Interscience Publi-
Edward Arnold, London cation, Wiley, New York

123
684 Geotech Geol Eng (2012) 30:665–684

Kou S, Rustan A (1993) Computerized design and result pre- McKenzie C (ed) International symposium on rock frag-
diction of bench blasting. In: Proceedings of the 4th inter- mentation by Blasting, FragBlast, Aug 26–31, 1990, Bris-
national symposium on rock fragmentation by blasting, bane, Australia, pp 283–289
Vienna, pp 263–271 Ozcelik Y (1998) Effect of discontinuities on fragment size
Kulatilake PHSW, Park J, Balasingam P, Mckenna SA (2007) distribution in open-pit blasting—a case study. Trans Inst
Hierarchical probabilistic regionalization of volcanism for Min Metall Sect A Min Indust 107:146–150
Sengan Region Japan. Geotech Geol Eng 25(1):79–109 Pal Roy P (1995) Breakage assessment through cluster analysis
Kulatilake PHSW, Wu Q, Hudaverdi T, Kuzu C (2010) Mean of joint set orientations of exposed benches of opencast
particle size prediction in rock blast fragmentation using mines. Geotech Geol Eng 13:79–92
neural networks. Eng Geol 114:298–311 Penn BS, Gordon AJ, Wendlandt RF (1993) Using neural net-
Kung TC, Hsiao CL, Schuster M, Juang CH (2007) A neural works to locate edges and linear features in satellite ima-
network approach to estimating excavation-induced wall ges. Comput Geosci 19:1545–1565
deflection in soft clays. Comput Geotech 34:385–396 Rizzo DM, Doughery DE (1994) Characterization of aquifer
Kuznetsov VM (1973) Mean diameter of fragments formed by properties using artificial neural networks; neural kriging.
blasting rock. Soviet Min Sci 9(2):144–148 Water Resour Res 30:483–497
Latham JP, Lu P (1999) Development of an assessment system Rogers LL, Dowla FU (1994) Optimization of groundwater
for the blastability of rock masses. Int J Rock Mech Min Sci remediation using artificial neural networks with parallel
Geomech Abstr 36:41–55 solute transport modeling. Water Resour Res 30:457–481
Latham JP, Kemeny J, Maerz N, Noy M, Schleifer J, Tose S Rogers SJ, Fang JH, Karr CL, Stanley DA (1992) Determination
(2003) A blind comparison between results of four image of lithology from well logs using a neural network. AAPG
analysis systems using a photo-library of piles of sieved Bull 76:731–739
fragments. Fragblast 7(2):105–132 Rosin P, Rammler E (1933) The laws governing the fineness of
Lee C, Sterling R (1992) Identifying probable failure modes for powdered coal. J Inst Fuel 7:29–36
underground openings using a neural network. Int J Rock Rustan PA (1998) Automatic image processing and analysis of
Mech Min Sci 29(1):49–67 rock fragmentation–comparison of systems and new
Li Q, Yu JY, Mu BC, Sun XD (2006) BP neural network pre- guidelines for testing the systems. Int J Blast Fragm Frag-
diction of the mechanical properties of porous NiTi shape blast, 15–23, Balkema, Rotterdam
memory alloy prepared by thermal explosion reaction. Sanchidrian JA, Segarra P, Lopez LM (2007) Energy compo-
Matl Sci Eng 419:214–217 nents in rock blasting. Int J Rock Mech Min Sci 44:130–147
Lilly PA (1986) An empirical method of assessing rock mass Schuhmann RJ (1959) Energy input and size distribution in
blastability. In: Proceedings of the large open pit confer- comminution. Trans Am Min Metall AIME Trans 214:22–25
ence, IMM, Australia, pp 89–92 Thomas AL, La Pointe PR (1995) Conductive fracture identi-
Maier HR, Dandy GC (1998) The effect of internal parameters fication using neural networks. In: Proceedings of the 36th
and geometry on the performance of back-propagation US symposium on rock mechanics, Balkema, Rotterdam,
neural networks: an empirical study. Environ Model Softw pp 627–632
13:193–209 Tung ATY, Wong FS, Dong W (1994) Prediction of the spatial
Mckenzie AS (1966) Cost of explosives—do you evaluate it distribution of the modified Mercalli intensity using neural
properly? Min Congr J 52(5):32–41 networks. Earthq Eng Struct Dyn 23:49–62
Mojtabai N, Farmer IW, Savely JP (1990) Optimisation of rock Yu SW, Zhu KJ, Diao FQ (2008) A dynamic all parameters
fragmentation in bench blasting. In: Proceedings of the 31st adaptive BP neural networks model and its application on
US symposium on rock mechanics, Balkema, Rotterdam, oil reservoir prediction. Appl Math Comput 195:66–75
pp 897–901 Zhang Q, Song JR, Nie XY (1991) The application of neural
Montgomery DC, Peck EA, Vining GG (2006) Introduction to network to rock mechanics and rock engineering. Int J
linear regression analysis. Wiley, New Jersey Rock Mech Min Sci 28:535–540
Nie SL, Rustan A (1987) Techniques and procedures in analyzing Zhang L, Luo JH, Yang SY (2009) Forecasting box office rev-
fragmentation after blasting by photographic method. In: enue of movies with BP neural network. Expert Syst Appl
Proceedings of the 2nd international symposium on rock 36:6580–6587
fragmentation by Blasting, Keystone, Colorado, pp 36–47
Ouchterlony F, Niklasson B, Abrahamsson S (1990) Fragmen-
tation monitoring of production blasts at Mrica. In:

123