0% found this document useful (0 votes)
39 views11 pages

Optimizing Overbreak Prediction Based On Geologica

Uploaded by

OSCAR SOLARI
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
39 views11 pages

Optimizing Overbreak Prediction Based On Geologica

Uploaded by

OSCAR SOLARI
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 11

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/270915619

Optimizing overbreak prediction based on geological parameters comparing


multiple regression analysis and artificial neural network

Article  in  Tunnelling and Underground Space Technology · September 2013


DOI: 10.1016/j.tust.2013.06.003

CITATIONS READS

33 2,776

2 authors:

Hyongdoo Jang E. Topal


Curtin University Curtin University
70 PUBLICATIONS   322 CITATIONS    94 PUBLICATIONS   2,964 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Uneven Break Optimisation View project

underground WSNs system View project

All content following this page was uploaded by Hyongdoo Jang on 29 March 2018.

The user has requested enhancement of the downloaded file.


This article appeared in a journal published by Elsevier. The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/authorsrights
Author's personal copy

Tunnelling and Underground Space Technology 38 (2013) 161–169

Contents lists available at SciVerse ScienceDirect

Tunnelling and Underground Space Technology


journal homepage: www.elsevier.com/locate/tust

Optimizing overbreak prediction based on geological parameters


comparing multiple regression analysis and artificial neural network
Hyongdoo Jang, Erkan Topal ⇑
Department of Mining Engineering, Western Australian School of Mines, The Curtin University, Kalgoorlie, WA, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Underground mining becomes more efficient due to the technological advancements of drilling and blast-
ing methods and the developing of highly productive mining methods that facilitate easier access to ore.
In the perspective of maximizing productivity in underground mining by drilling and blasting methods,
Keywords: overbreak control is an essential component. The causing factors of overbreak can simply divided as
Blasting blasting and geological parameters and all of the factors are nonlinearly correlated. In this paper, the
Overbreak blasting design of the tunnel was fixed as the standard blasting pattern and the research focus on effects
Underground mine
of geological parameters to the overbreak phenomenon. 49 sets of rock mass rating (RMR) and overbreak
Artificial neural network
Multiple regression analysis
data were applied to linear and nonlinear multiple regression analysis (LMRA and NMRA) and artificial
neural network (ANN) to predict overbreak as input and output parameters, respectively. The perfor-
mance of LMRA, NMRA, and optimized ANN models was evaluated by comparing coefficient correlations
(R2) and their values are 0.694, 0.704 and 0.945, respectively, which means that the relatively high level
of accuracy of the optimized ANN in comparison with LMRA and NMRA. The developed optimum over-
break predicting ANN model is suitable for establishing an overbreak warning and preventing system and
it will utilize as a foundation reference for a practical drift blasting reconciliation at mines for operation
improvements.
Crown Copyright Ó 2013 Published by Elsevier Ltd. All rights reserved.

1. Introduction reduce the possible failures on blasting operations. Geological fac-


tors, however, are unchangeable and they have a significant influ-
In underground mining, overbreak has long been recognized as ence on the overbreak phenomenon. In fact, if the rock is not strong
the principal cause of hazards and deterioration costs in mine enough to support itself, possibly no blasting techniques can stop
management and as such numerous related research projects have the occurrence of overbreak.
been conducted. Many research papers have been devoted to Notwithstanding the significance of overbreak in underground
clarifying the overbreak phenomenon, but they are still unable to mining and tunneling, it is known as an unpredictable phenome-
explain the exact occurrence process. According to former re- non. Acknowledging the importance of overbreak phenomenon, it
search, factors causing overbreak can be classified into two groups. is critical that it should be managed by proper systems. Predicting
Different terminology has been used by different researchers, but overbreak is the first step in developing an overbreak management
geological and blasting factors were the principal groups influenc- and blasting reconciliation system. Throughout the history of
ing overbreak (Mahtab et al., 1997; Mandal et al., 2008). mining and tunneling, there has been only one attempt to predict
Blasting parameters are changeable factors. Excavation condi- overbreak. Murthy et al., 2003 reported an overbreak prediction
tions affect the likelihood of overbreak and these influencing formula, comparing with the maximum charge per delay at Koyna
parameters are highly correlated with each other. Some particular Hydro-electric Project in India and the research used the peak
overbreak occurrences can be accounted for certain major influ- particle velocity (PPV) threshold levels for damage estimation.
encing parameters, but it cannot be generalized without interpret- However, the proposed overbreak prediction formula is not
ing other influencing factors. In modern underground blasting, reliable and cannot be generalized for further use. The reason is
techniques are reasonably organized with advanced final wall that the formula used only the maximum charge per delay, which
blasting methods such as, smooth blasting and presplitting. Along is only one factor of many others that need to be considered, and
with the advanced blasting methods, final wall customized the PPV is a site specific value that cannot be generally applied
explosives and computer base drill operating systems significantly to different sites.
In this paper, the blasting design of a tunnel was fixed as a
⇑ Corresponding author. Tel.: +61 8 9088 6156; fax: +61 8 9088 6151. standard blasting pattern and the study focuses on the effects of
E-mail address: [email protected] (E. Topal). geological parameters to the overbreak phenomenon. RMR param-

0886-7798/$ - see front matter Crown Copyright Ó 2013 Published by Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.tust.2013.06.003
Author's personal copy

162 H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169

eters were collected through 49 blasting sections as geological data holes, firing sequences, cut design, blasting hole deviation, explo-
and overbreak data were individually investigated. Various meth- sive characteristics, charge concentrations, coupling ratio, powder
ods have been applied in engineering for overbreak prediction. In factors, blast-induced shock wave and energy levels, etc. All blast-
this study, linear and nonlinear multiple regression analysis (LMRA ing parameters affect overbreak in complex mutual correlation
and NMRA) and artificial neural network (ANN) are used to predict within just a few milliseconds. Actually, to obtain a smooth
potential overbreak. The geological data sets are put as input fracture plane without any wall damage, proper blasting design
parameters and encountered overbreak results are used as output and accurate drilling should take precedence and influence the
parameters to ANN models and simultaneously to LMRA and selection of other blasting parameters. With these large influences
NMRA. Consequently, the optimum overbreak predicting model on overbreak and the flexibility to manipulate blasting parameters,
is selected by comparing measured and predicted overbreaks and many research projects have been undertaken to understand the
the correlation coefficient (R2) of each proposed model. management of overbreak and to unveil the influences of blasting
parameters on the overbreak phenomenon. For example, Rustan
2. Characteristics of overbreak (1998) conducted model and field blasting tests to define the
optimal delay time between contour holes in lock blasting by com-
Overbreak is a surplus excavated area of rock beyond the theo- paring simultaneous and micro-sequential initiation systems. Both
retical contour in an excavation, and it can occur in any kind of systems have advantages and disadvantages, but simultaneous ini-
underground excavation method. It is known to be inevitable when tiation system was found to be superior in minimizing overbreak.
drilling and blasting method, and it is affected by the majority of As a result, field tests showed the maximum radial crack length
conditions of excavation. Although excavation by drilling and into the remaining rock mass of simultaneous initiation systems
blasting is known to be an inexpensive method, it becomes exorbi- created a 1.3–9.0 times less than the micro-sequential initiation
tant when safety of structures is endangered. Overbreak of rock be- system which had only 1 ms firing delay between contour holes.
yond the designed periphery of a tunnel is one major factor that
puts excavated spaces at risk and significantly affects operational 2.1.2. Geological parameters
and management efficiency. Overbreak jeopardizes both workers Geological parameters are fixed factors and most of them, such
and equipment in the underground excavations and increases dilu- as the strength of rock mass, discontinuous characteristics, water
tion of the ore in mine operations. In addition, it adversely affects conditions, stress conditions, and the topography of the surround-
mine management by creating unproductive works such as dilu- ing area, have significant influence on the overbreak phenomenon.
tion, requirements for additional supports and their removal, all Many research projects have been conducted to understand the
of which adds to production costs. For example, Chakraborty relationship between geological factors and overbreak. For in-
et al. (1996) demonstrated the adverse effects of overbreak during stance, Hagan (1992) emphasized the importance of pre-existing
the Koyna Hydro-electric project in India that was excavated by joints and beddings on in situ rock. According to his paper, frac-
drilling and blasting methods. The area was predominantly cov- tures in the rock tend to dominate the nature of the blast-induced
ered with basalt and volcanic breccia with dense joint sets that fracture pattern and it usually influences overbreak more than the
caused 7–12% of overbreak of total excavation. Undesired excava- mechanical and physical properties of the rock.
tions in the form of overbreak increased the total construction cost Among the geological factors, the orientation of discontinuity is
by about 19%. Fig. 1 illustrated the typical overbreak and under- one of the major factors influencing overbreak phenomenon. Hoek
break in a tunnel after blasting. and Brown (1980) reported that a discontinuity plane having strike
parallel to the tunnel axis is considered to have an unfavorable ef-
2.1. Factors influencing to overbreak fect on overbreak. Generally, less overbreaks and under-breaks are
observed where the strike of the discontinuity is nearly perpendic-
Overbreak is known as an unavoidable phenomenon, and the ular to the tunnel axis and contrastively greater when they are
causing factors for overbreak have significant mutual correlation. nearly parallel. In detail of other orientations, drives with dip are
As mentioned earlier, the causing factors of overbreak can split into more favorable than drive against dip where the strike of disconti-
two categories – blasting and geological parameters. nuity is perpendicular to tunnel axis and fair and very unfavorable
for dip has angle of 20–45° and 45–90°, respectively, when the
2.1.1. Blasting parameters strike is parallel to tunnel axis (Bieniawski, 1973).
As blasting parameters are able to be modified, overbreak can Although the perpendicular orientation of discontinuity strikes
be managed by varying these parameters. The components of against tunnel axis has been considered to have advantageous dri-
blasting parameters include blasting geometry, subdrilling, guide vage conditions than parallel, it still has significant influences to

Fig. 1. Overbreak and under-break in tunnels conventionally excavated.


Author's personal copy

H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169 163

overbreak phenomenon. Chakraborty et al. (1994) conducted tall. The tunnel was schemed through the cretaceous period sedi-
scaled blasting model tests and field observation to assess the ef- mentary rock area. The dominant gritty sandstone was found
fects of perpendicular joint orientations via tunnel axis and rock throughout the tunnel and different joint patterns and weathered
mass qualities on overbreak. In field observation of inclines under rates of in situ rock allowed to collect various geological data sets.
sedimentary rocks, massive overbreak was noted where the strike Greenish1 and dark gray shales were occasionally appeared as the
of joint planes was normal to the tunnel axis. More overbreaks order of 10–50 cm thick stratum. The stratification bedding plains
were observed at the side wall than crown when the joint is gently were dominant through the tunnel and often two vertical joint sets
dipped. On the other hand, the crown got more damaged than the sometimes associated with some random joints were observed. The
side wall when the joints were presented with the steep dip. The strikes of bedding planes were mostly perpendicular to the tunnel
author commented on this phenomenon that the shock waves axes and gently dipped. The average bedding spaces were between
were likely transmitted in the lateral directions of joint planes 10 and 60 cm, and in most of cases, the bedding planes were tightly
but were restricted in the forward direction and relatively similar healed or slightly altered. However, the conditions of bedding and
results were verified by the laboratory scaled blasting tests. On joints were turned poor where these two are intersected. The stereo-
the observed inclines, the rock mass conditions were determined graphic projection of joints and the location of Bigye tunnel are illus-
by Rock Mass Quality (Q) (Barton, 1974) and the value was turned trated in Figs. 2 and 3, respectively.
out between 0.82 and 24. On the comparison Q-value with over- The general blueprints of tunnel supporting types and blasting
break, the approximate lengths of overbreak were smaller than designs were subdivided into six patterns based on the RMR rates.
0.6 m when the Q-value is between 12 and 24. Contrariwise, up Powder factors of each pattern varied from 0.723 kg/m3 to
to 2 m of conspicuously massive overbreaks was found out when 0.914 kg/m3, but the perimeter hole was defined as 0.55 kg/m3.
the Q-value is less than 12. These observations clearly indicate Also, the burden and spacings of the perimeter holes were deter-
the strong dependence between rock mass quantity & joint orien- mined as 750 mm and 600 mm, respectively. The borehole diame-
tation and tunnel blast performance. Moreover, characteristics of ter is 45 mm and 17 mm of Finex; one of the manufactured
discontinuous in the rock mass are acknowledged as the key fac- emulsion base explosive for perimeter holes was charged to perim-
tors of overbreak and crack propagations. The related literature eter holes. Details of perimeter blasting designs are tabulated in
on this area is listed in Fordyee et al. (1993), Fourney and Dick Table 1.
(1995), Mandal et al. (2008), and Simha (1996). As input data for the overbreak predicting models, RMR criteria,
such as unconfined strength of rock (Sc), rock quality designation
(RQD), spacing of joints (Js), state of joints (Ja), spatial orientation
2.2. Overbreak prevention methods
of joints (Jo), and ground water condition (Jw), were investigated
before blasting. Additionally, the depths of overbreak were ob-
Over the past 60 years, many researchers have presented
served after blasting as output data. The Sc, RQD, and Js results of
advancing blasting techniques and methods for the prevention of
RMR investigation throughout the entire tunnel on average were
overbreak. One cautious blasting method ‘‘smooth blasting theory’’
80 MPa, 68, and 400 mm, respectively. The gently dipped sedimen-
was developed at Nitroglycerin AB’s Detonic Research Laboratory
tary bedding was acknowledged to have noticeable discontinuities
in Stockholm by Lundgorg and reported by Langefors (1953). It is
and the adjustment for the joint orientation value rated to the or-
recognized as one of the effective methods in preventing overbreak
der of 5 through the entire length of the tunnel. Hence, the rate of
and has been applied to various mines and construction sites. The
Jo of the order of 5 was summed up to calculate RMR which is
smooth blasting method is highly efficient when used for fissure-
counted as one of input parameter. The average RMR rate calcu-
less and hard rock. Nevertheless, it is still hard to control the over-
lated at 57.53, which indicates that it is in the moderate rock range.
break and blast-induced damages if the in situ rock is not strong
In order to simplify the overbreak predicting processing, RMR rates
enough to support itself and if it contains any discontinuities.
were used to proposed models. Details of investigated input and
Mandal et al. (2008) accounted for modified smooth blasting
output parameters are arranged in Table 2.
patterns for minimizing damages to the peripheral area and also
mentioned that geo-technical properties and in situ stress condi-
tions of rock mass should be considered for minimizing the magni- 4. Development of overbreak prediction methods
tude of overbreak. The research advises a long delay time between
buffer and perimeter row to avoid shearing effect of periphery rock MRA and multiple layer perceptron (MLP) ANNs were applied to
mass and the burden of perimeter hole should not be more than find an overbreak predicting model using the collected data sets.
0.2 times of the depth of holes to minimize the duration of gaseous MRA is a statistical analysis method for interpreting the variance
energy impact to the perimeter wall. The conclusion mentioned of dependent variable by using given independent variables and
that the magnitude of vibration should not be the criteria to specify it has been widely used for prediction and forecasting. ANN is a
the damage level. branch of artificial intelligence (AI). ANN does an impression of
Recently, Iverson and Hustrulid (2011) reported a new concept the way which a human brain works in processes such as studying,
for perimeter control blasting, which suggests that a proper design memorizing, and reasoning and one of the fascinating abilities of
of the buffer raw of the blast holes is the key to successful perimeter ANN is a prediction by performing massive parallel computations.
control. In his study, the criteria for burden and spacing of perime-
ter holes are decided by calculating the damage radius of buffer
4.1. Multiple regression analysis to predict overbreak
holes. Notwithstanding many efforts, it is fact that if the rock is
weak, possibly no blasting technique can create a smoother, more
In order to find the overbreak predicting model, MRA was em-
solid face than that permitted by the inherent rock formation.
ployed by using collected RMR criteria and overbreak data sets as
independent and dependent variable, respectively. The purpose
3. Site description and data collection of MRA is to identify RMR criteria that explain variations in the
overbreak. In this paper, both LMRA and NMRA were adopted
In 2010, data were collected at Bigye tunnel, which is located in
the Gumi city, South Korea. The tunnel was constructed as part of 1
For interpretation of color in Fig. 3, the reader is referred to the web version of
the local road improvements plan and is 12.08 m wide and 8.40 m this article.
Author's personal copy

164 H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169

Fig. 2. Stereographic projection of joints by DIPS.

Fig. 3. Location of Bigye tunnel, Gumi, Gyeongbuk province, South Korea.

Table 1
Details of perimeter design.

Hole diameter Space Burden Powder factor


Blasting design 45 mm 600 mm 750 mm 0.55 kg/m3
Type Diameter Velocity Gas
Explosive Emulsion 17 mm 4000 m/s 640 L/kg

and RMR criteria and overbreak results were set as independent dependence and moderate correlations appeared to Js (0.608)
and dependent variables, respectively. Pearson’s correlation values and Jw (0.500).
of RMR, RQD, and Sc calculated from 0.822, 0.765, and 0.608, When multiple regression analysis is performed, the multicollin-
which means that the data sets have a strong linear inverse earity problem has to be controlled. It occurs when the correlation
Author's personal copy

H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169 165

Table 2
Details of 49 sets of input and output parameters for overbreak predicting models.

Type of data Name Average Std. deviation Max. value Min. value
Input Sc 8.16 2.33 12.00 2.00
RQD 13.63 3.93 20.00 3.00
Js 10.12 2.73 15.00 5.00
Ja 17.06 4.87 23.00 7.00
Jw 8.55 1.64 10.00 6.00
RMR 57.53 13.47 77.00 29.00
Output Overbreak (cm) 13.53 12.28 50.00 0.00

among the independent variables is strong. Hence, the standard er- found and results are tabulated in Table 3. The high values of R2
rors of the coefficients are increasing, which leads to an erroneous are derived when Sc, Log (RQD), Quadratic (Js), Cubic (Ja), Quadratic
conclusion of multiple regression analysis. Variance Inflation Factor (Jw), and Cubic (RMR). The optimum overbreak prediction equation
(VIF) and Tolerance are commonly used to verify the multicollinear- was derived by combining only Quadratic (Js) and Inverse (Ja). R2 of
ity problem. VIF measures how much the variance of the estimated the optimized model is 0.708. The values of the independent vari-
coefficients increases over the case of no correlation among the ables are less than 0.05, Durbin–Watson value is 1.902 and the VIF
independent variables. If two independent variables are not corre- is 1.579, which means that the model is unaffected by autocorrela-
lated, then all the VIFs will be equal to 1. Generally, if the value of tion, multicollinearity, or heteroskedasticity problems. Variables of
VIF is over five, the model demonstrates strong multicollinearity Sc, RQD, Jw, and RMR are eliminated due to the same reason of
problem. Tolerance is an inverse number of VIF, and if the tolerance LMRA results. The final nonlinear model is mathematized at:
value is less than 0.1, it is acknowledged that there is multicolline-
OB ¼ 3:457  0:072½Quadratic ðJ s Þ þ 272:58½Inverse ðJ a Þ ð3Þ
arity problem associated with the model. The simple solution of the
multicollinearity problem is the elimination of the suspicious Comparisons between measured and predicted overbreak by
variable. LMRA and NMRA are illustrated in Fig. 4a and b, respectively.

4.1.1. Linear multiple regression analysis 4.2. Overview of artificial neural network
LMRA is employed to generate a mathematical relationship that
describes variations on the overbreak against RMR criteria. Before ANN is one of AI methods and it is computation inference model
processing LMRA, the tolerance of Ja was calculated to be 0.000, simply mimicked by biological structure of the human brain. ANN
which means that the variable Ja had the highest possibility of mul- has the ability to generalize a solution from the pattern presented
ticollinearity when all the variables were considered. Hence, Ja was to it during training. Once ANN is trained with sufficient numbers
eliminated and LMRA model (Eq. (1)) is the obtained using SPSS of sample data sets, a new input of relatively similar pattern, pre-
software. dictions can be done on the basis of the previous learning. ANN
consists of group of neuron layers: input, hidden, and output.
OB ¼ 54:825 þ 0:114½Sc  þ 0:338½RQD  0:910½J s  þ 0:883½J w  The artificial neuron, simply called neuron, is a simple processing
 0:785½RMR ð1Þ element and it is complicatedly interconnected with each other.
In a similar way with a biological brain, neuron receives signals
where OB is the overbreak expressed in centimeter and Sc, RQD,
from other neurons which can be intensified and weakened by
Js, Jw and RMR are in RMR rates and the correlation coefficient (R2)
proper activation functions and the connection weights modulate
of Eq. (1) is calculated as 0.706. However, the VIF values of RQD
input signals as synapsis does in brain. Limited learning ability of
and RMR were obtained as 22.33 and 81.80, respectively, with
ANN can achieve by adjusting the weights with chosen learning
the tolerance values for Sc, Js, and Jw were 0.169, 0.154, and
algorithm.
0.146, respectively. It clearly indicates that all the variables in Eq.
(1) have the multicollinearity problem. Hence, instead of entering
4.2.1. Learning algorithms of ANN
all the variables into the model, the stepwise method was used.
The learning situation of ANN can be classified with unsuper-
In this method, possible combinations of variable were considered
vised and supervised learning, and this research will focus on the
and the combination or variables were removed if any of these
back propagation algorithm which is one of the powerful super-
have inappropriate value of multicollinearity problem. The up-
vised learning algorithms. The back propagation algorithm is firstly
dated overbreak predicting model was derived as following
introduced by Bryson and Ho (1969), and it has been attracting
equation:
attention on the middle of the 1980s. Since the perceptron (Rosen-
OB ¼ 53:607  1:210½J a   1:919½J s  ð2Þ blatt, 1958), an initial learning algorithm of ANN, can only solve
2
The correlation coefficient (R ) is 0.694 which is less than Eq.
(1), but the model does not have any autocorrelation, multicollin- Table 3
earity, or heteroskedasticity problems. Summary of nonlinear models of independent variables.

Independent variables R2 of equation


4.1.2. Nonlinear multiple regression analysis Linear Logarithmic Inverse Quadratic Cubic
Many problems in engineering are heavily complex when large
Sc 0.370 0.340 0.268 0.370 0.370
numbers of variables are present. Regarding LMRA, it is limited in RQD 0.585 0.601 0.514 0.600 0.600
its ability to derive a proper mathematical relationship. As earlier Js 0.571 0.636 0.652 0.656 0.656
mentioned, the independent and dependent variables have nonlin- Ja 0.596 0.625 0.639 0.610 0.683
ear relationship so NMRA was applied to improve the overbreak Jw 0.250 0.253 0.256 0.258 0.258
RMR 0.676 0.697 0.703 0.689 0.727
prediction. Before employing NMRA, the optimum nonlinear model
of each independent variable against dependent variable was Dependent variable: Overbreak.
Author's personal copy

166 H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169

Fig. 4. Comparison between measured and predicted overbreak using LMRA (a) and NMRA (b).

simple problems in pattern classification, the back propagation 4.2.2. Artificial neural network application to mining industry
algorithm within a multilayer ANN can approximate any nonlinear Historically, mining industry has close connection with the AI.
function. In 1979, a knowledge base expert system, one of offshoot tech-
The perceptron learning algorithm is based on McCulloch and niques of AI, earned great reputation with developing a com-
Pitts’s neuron model, and it is a single layer neural network that puter-based consultation system for mineral exploration,
consists of adjustable synapse weights and a hard limit training PROSPECTOR, and it had been successfully identified the location
function. For example, there is a set of learning samples consisting of a molybdenum deposit in Washington State, US (Campbell
of input vectors (x1,2,. . .,n) and a desire output (target) (Yd). The per- et al., 1982; Hart et al., 1978).
ceptron learning algorithm starts with randomly setting weights In recent years, the ANN model is more frequently used for
for the connections of inputs and a neuron. After picking one of in- developing high tech computer technologies and its abilities have
put vector, neuron will activate with developed significantly. Many mining researchers have been
applying ANN to the various problems that cannot solved with fun-
" # damental statistical analyses such as predicting fragmentations,
Xn
YðpÞ ¼ step xiðpÞ wiðpÞ  h ð4Þ back breaks, ground vibrations, strength of mineral admixture
i¼1 concretes, fly rocks, effectiveness of blast wall barriers, and even
optimizing blasting parameters in certain condition; however,
where p and Y(p) present number of steps and output of p steps,
none of researches have not been done on predicting overbreak
respectively. wi(p) is weights at p steps and h presents an initial
with ANN. The prediction ability of ANN has been testified by com-
threshold. If the output of p steps (Y(p)) is different with the target
paring results with other statistical methods and it acknowledged
(Yd), all the connection weights modify with
as superior model than any other prediction method. The ANN re-
lated with mining literatures are list as Atici (2011), Kulatilake
wi ðp þ 1Þ ¼ wi ðpÞ þ Dwi ð5Þ
et al. (2010), Monjezi et al. (2010), Monjezi et al. (2011a), Monjezi
where Dwi(p) presents an updated weight value (according to: and Dehghani (2008), Monjezi et al. (2011b), Remennikov and Rose
Dwi(p) = a(Y(p)  Yd)), wi(p + 1) is a new weight for step 2, and a (2007), and Zhang and Li (2011).
is indicated learning rate. Then, the second input (xi) activated with
the new weight (wi(p + 1)) and it repeats until the output con- 4.2.3. Define the proper learning algorithm and ANN architectures for
verges to the target (Yd). overbreak prediction
The back propagation learning algorithm with multilayer ANN When applying ANN, defining optimum number of hidden
similarly processes with perceptron and Fig. 5 shows a typical layers, the number of neurons in layers and preventing over-fitting
architecture of the feed-forward back propagation ANN assuming are crucial confronting problems. The performance of ANN is
the optimum number of input data as a, neurons in a hidden layer widely affected by functional relations between input and output
as b and hidden layers as c. parameters, the architecture of model, and the learning algorithm.
Unlikely from the perceptron, normally the sigmoid function is Few empirical formulas were suggested by researchers, such as
adopted as a transfer function and outputs of hidden layer are used Hecht-Nielsen (1987) and Kaastra and Boyd (1996), but a heuristic
as an updated input for the next layer. Through hidden layers, in- approach is applied to define the optimum architecture of ANN
put data are processed until it reaches to the output layer which is because the proper numbers may even alter with different simula-
named as the forward pass. In output layer, the output is compared tions in a same problem. It is well known fact that most of over-fit-
with the target, measured value, and the error differences are pro- ting problem can be solved with applying more than two layers
cessed back through the network by updating all weights of con- and the performance of ANN are not evidently improves with more
nections and biases of individual neurons, the backward pass. than two layers.
Those forward and backward pass processes are repeated until Neural Networks version No. 4.3.8 of Alspace (O’Neill et al.,
the network error converges to an initially setting minimum error. 2011) was implemented to build MLP and the basic architecture
Author's personal copy

H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169 167

Fig. 5. Typical architecture of the feed-forward back propagation ANN.

Fig. 6. MSE variance by different learning algorithm on hidden layers.

was defined as one input layer with 6 neurons, two hidden layers
with 15 neurons each, and one output layer with one neuron
(6–15–15–1). In order to determine the proper leaning algorithm,
sigmoid (SIG) and hyperbolic tangent (TANH) functions were alter-
natively applied to two hidden layers and linear (LIN) function was
commonly applied to the output layer. The performance of ANN is
evaluated with the sum of mean squares error (MSE) on the train-
ing data at 1000 iterations which were enough to observe the con-
vergence of MSE. 49 data sets of input and output data were
applied to proposed MLP, and 20% of the data were randomly clas-
sified as a test set and rest of 80% of data were defined as a training
set. The four different types of ANN were operated 10 times each
and trimmed means, an average value that discarding a maximum
and a minimum value, were calculated to examine the central ten-
dency of each model. The lowest trimmed mean of MSE is appeared
when TANH function is applied to both first and second hidden
layer, and the results are illustrated in Fig. 6.
In order to optimize the architecture of ANN, a heuristic ap-
proach is used to choose appropriate number of neurons in two
hidden layers and the optimum architecture of the overbreak pre-
dicting ANN model was decided at 6–13–5–1 and the model
reached MSE value of 0.1 at 17,690 iterations.
Comparison results between measured and predicted overbreak
by the optimum ANN model are illustrated in Fig. 7, and the corre- Fig. 7. Comparison between measured and predicted overbreak using the opti-
lation coefficient (R2) of optimized model was 0.945. mized ANN model.
Author's personal copy

168 H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169

Fig. 8. Comparison of measured and predicted overbreak of LMRA, NMRA, and the optimum ANN.

5. Results and discussion optimized ANN – were evaluated by comparing their coefficient
correlations (R2). Their values are 0.694, 0.708, and 0.945, respec-
The purpose of this study is to develop an optimum overbreak tively. This shows the relatively high level of accuracy of the opti-
predicting model based on provided geological data. Multiple mized ANN in comparison with LMRA and NMRA.
regression analysis models and the multiple layers ANN models The optimized overbreak predicting ANN can effectively predict
were used to find the best description model for predicting over- potential overbreak by simply adopting RMR criteria. It can also be
break. The performance of LMRA, NMRA, and the optimized ANN used as an overbreak warning and preventing system. Even though
model was evaluated by comparing their correlation coefficients the proposed first version of optimized overbreak predicting ANN
(R2). Their values are 0.694, 0.708, and 0.945, respectively. Conse- model would be reliably performed at limited conditions of sedi-
quentially, the optimum ANN model, architecture of 6–13–5–1, mentary rock area, it can become a generalized model by retraining
provided a better result than the linear and nonlinear multiple with additional overbreak and geological data sets.
regression analysis models. The graphical comparisons of mea-
sured and predicted overbreak for LMRA, NMRA, and the optimum
ANN model are shown in Fig. 8. References
The causing factors of overbreak are complex and mutually cor-
Atici, U., 2011. Prediction of the strength of mineral admixture concrete using
related, so it is important to consider all parameters together. multivariable regression analysis and an artificial neural network. Expert
Some of parameters, however, such as Sc, RQD, Jw, were eliminated Systems with Applications 38, 9609–9618.
when processing LMRA and NMRA because of statistical problems. Barton, N., 1974. Engineering classification of rock masses for the design of tunnel
support. Rock Mechanics 6, 189.
The developed optimum ANN model can interpret 94.5% of over- Bieniawski, Z., 1973. Engineering classification of jointed rock masses. Civil
break phenomenon without disregarding any of dependent Engineer in South Africa 15.
variables. Bryson, A.H., Ho, Y.C., 1969. Applied Optimal Control. Blaisdell.
Campbell, A.N., Hollister, V.F., Duda, R.O., Hart, P.E., 1982. Recognition of a hidden
Considering that the RMR survey is a fundamental routine task mineral deposit by an artificial intelligence program. Science 217, 927–929.
as part of the face mapping process before blasting, the optimized Chakraborty, A.K., Jethwa, J.L., Paithankar, A.G., 1994. Assessing the effects of joint
ANN model can practically be used as a prior warning system for orientation and rock mass quality on fragmentation and overbreak in tunnel
blasting. Tunnelling and Underground Space Technology 9, 471–482.
potential overbreak. This allows the experienced blasting engineer Chakraborty, A.K., Murthy, V.M.S.R., Jethwa, J.L., 1996. Blasting problems in
to minimize overbreak by altering the base blasting scheme. While underground constructions through deccan trap formation: some experiences
the optimized overbreak predicting ANN model cannot cover all of at Koyna Hydro-electric project, stage IV. Tunnelling and Underground Space
Technology 11, 311–324.
the rock condition, the model can generalize by adopting and
Fordyee, D.L., Fourney, W.L., Dick, R.D., Wang, X.J., 1993. Effects of joints on stress
retraining with other data sets. wave transmission. In: 4th International Symposium on Rock Fragmentation,
pp. 211–219.
Fourney, W.L., Dick, R.D., 1995. The utilization of explosive loading as a
6. Conclusion nondestructive evaluation tool in geologic materials. International Journal of
Solids and Structures 32, 2511–2522.
Hagan, T.N., 1992. Safe and cost-efficient drilling and blasting for tunnels, caverns,
The main objective of this study is to build an optimized over- shafts and raises in India. In: Proc. of a Workshop on Blasting Technology for
break predicting model based on various geological data. LMRA Civil Engineering Projects, New Delhi, India, pp. 16–18.
and NMRA and multilayer ANNs were applied to find the most Hart, P., Duda, R., Einaudi, M., 1978. PROSPECTOR—a computer-based consultation
system for mineral exploration. Mathematical Geology 10, 589–610.
accurate model for predicting overbreak. The blasting design of
Hecht-Nielsen, R., 1987. Kolmogorov’s mapping neural network existence theorem.
the perimeter area of the tunnel was fixed as standard scheme, In: The 1st IEEE International Conference on Neural Networks, San Diego, CA,
and RMR criteria and overbreak data were investigated on each USA, pp. 11–14.
face before and after blasting. Hoek, E., Brown, E.T., 1980. Underground Excavations in Rock.
Iverson, S.R., Hustrulid, W.A, 2011. Design concept for perimeter control blasting in
The performance of the proposed overbreak predicting models drifting operations. In: 45th U.S. Rock Mechanics/Geomechanics Symposium.
that used different methods – namely LMRA, NMRA, and the American Rock Mechanics Association, San Francisco, California.
Author's personal copy

H. Jang, E. Topal / Tunnelling and Underground Space Technology 38 (2013) 161–169 169

Kaastra, L., Boyd, M., 1996. Designing a neural network for forecasting financial and Monjezi, M., Ghafurikalajahi, M., Bahrami, A., 2011b. Prediction of blast-induced
economic time series. Neurocomputing, 215–236. ground vibration using artificial neural networks. Tunnelling and Underground
Kulatilake, P.H.S.W., Qiong, W., Hudaverdi, T., Kuzu, C., 2010. Mean particle size Space Technology 26, 46–50.
prediction in rock blast fragmentation using neural networks. Engineering Murthy, V.M.S.R., Dey, K., Raitani, R., 2003. Prediction of over-break in underground
Geology 114, 298–311. tunnel blasting a case study. Journal of Canadian Tunneling Canadien, 109–115.
Langefors, U., 1953. Slätsprängning [Smooth blasting]. Jernkontorets Annaler O’Neill, K., Sueda, S., Tung, L., Santos, J., Arksey, N., Porter, K., Knoll, B., 2011. In:
Almqvist and Wiksells Bocktryckeri AB. Alspace, Neural Networks, 4.3.8 ed. The University of British Columbia,
Mahtab, M.A., Rossler, K., Kalamaras, G.S., Grasso, P., 1997. Assessment of geological Vancouver.
overbreak for tunnel design and contractual claims. International Journal of Remennikov, A.M., Rose, T.A., 2007. Predicting the effectiveness of blast wall
Rock Mechanics and Mining Sciences 34, 185.e181–185.e113. barriers using neural networks. International Journal of Impact Engineering 34,
Mandal, S., Singh, M., Dasgupta, S., 2008. Theoretical concept to understand plan 1907–1923.
and design smooth blasting pattern. Geotechnical and Geological Engineering Rosenblatt, F., 1958. The perceptron: a probabilistic model for information storage
26, 399–416. and organization in the brain. Psychological Review 65, 386–408.
Monjezi, M., Dehghani, H., 2008. Evaluation of effect of blasting pattern parameters Rustan, A.P., 1998. Micro-sequential contour blasting – how does it influence the
on back break using neural networks. International Journal of Rock Mechanics surrounding rock mass. Engineering Geology 49, 303–313.
and Mining Sciences 45, 1446–1453. Simha, K., 1996. Effect of open joint on stress wave propagation. In: 5th
Monjezi, M., Ahmadi, M., Sheikhan, M., Bahrami, A., Salimi, A.R., 2010. Predicting International Symposium on Rock Fragmentation by Blasting.
blast-induced ground vibration using various types of neural networks. Soil Zhang, C.L., Li, X.P., 2011. Study on optimization of blasting parameters and its effect
Dynamics and Earthquake Engineering 30, 1233–1236. on anchoring rock beam of a underground workshop. Applied Mechanics and
Monjezi, M., Bahrami, A., Varjani, A., Sayadi, A., 2011a. Prediction and controlling of Materials 90, 2045–2052.
flyrock in blasting operation using artificial neural network. Arabian Journal of
Geosciences 4, 421–425.

View publication stats

You might also like