You are on page 1of 7

International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

Contents lists available at ScienceDirect

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

Technical Note

Prediction of blast-induced ground motion in a copper mine


Radojica Lapevi a, Sran Kosti b,n, Radoje Pantovi c, Neboja Vasovi d
a

Department of Geotechnics, University of Belgrade Faculty of Mining and Geology, uina 7, 11000 Belgrade, Serbia
Department of Geology, University of Belgrade Faculty of Mining and Geology, uina 7, 11000 Belgrade, Serbia
Department of Mining Engineering, University of Belgrade Technical Faculty in Bor, Bor, Serbia
d
Department of Applied Mathematics, University of Belgrade Faculty of Mining and Geology, uina 7, 11000 Belgrade, Serbia
b
c

art ic l e i nf o
Article history:
Received 18 September 2013
Received in revised form
31 December 2013
Accepted 15 March 2014

1. Introduction
Drilling and blasting are commonly used rock excavation
techniques within the New Austrian Tunneling Method (NATM),
as a method of producing underground space by using all available
means to develop the maximum self-supporting capacity of the
rock itself to provide the stability of the underground opening
[1,2]. Even though Tunnel Boring Machines (TBMs) are now used
in many tunneling projects, most underground excavation in rock
is still performed using blasting. In the absence of an initial free
face, the solid blasting method is employed for rock excavation.
A greater proportion of annual tunnel advance is still achieved by
drilling and blasting [3]. The excavation of orebody T was also
performed using the NATM method with drilling and blasting,
since this technique has an unmatched degree of exibility
and can overcome the limitations of machine excavations. Unfortunately, blast-induced rock damage and overbreak in underground construction may result in increasing construction costs
and declining stability of the chamber. Considering this, it is of
great importance to properly design the blasting operations, in
order to avoid the possible occurrence of rock mass and support
damage and instability.
In practice, blast-induced ground motion is commonly
expressed by a peak particle velocity (PPV), estimated using various empirical ground motion attenuation relations [46]. These
equations are of great interest for engineers, since they enable
them to predict the maximum ground vibration depending on the
scaled distance [712]. However, considering the fact that a
number of parameters affect the blast induced ground vibrations,
empirical attenuation equations are sometimes not suitable for the
n

Corresponding author. Tel.: 381113219107.


E-mail address: srdjan.kostic@rgf.bg.ac.rs (S. Kosti).

http://dx.doi.org/10.1016/j.ijrmms.2014.03.002
1365-1609/& 2014 Elsevier Ltd. All rights reserved.

design of blasting patterns. In those cases, instead of these


conventional predictors, new techniques such as articial neural
networks (ANN) are being used. Khandelwal and Singh [13]
predicted the PPV by ANN, taking into consideration the distance
from the blast face to monitoring point and explosive charge per
delay. A few years later, the same authors developed a three-layer
feed-forward back-propagation neural network for predicting the
PPV and frequency and obtained a much higher coefcient of
determination in comparison to the conventional predictors [14].
Monjezi et al. [15] also developed a feed-forward back-propagation neural network model, with four input parameters, two
hidden layers and one output parameter (PPV). In this case, the
accuracy of prediction by using ANN was much higher (R2 0.95)
in comparison to the conventional predictors or mutlivariate
regression analysis (R2 0.380.80).
In this paper, we develop a PPV prediction model for the
specic case study. Even though there are already many ground
motion predictors, which could give a reasonable prediction of
PPV, there is a justied need for updating the existing models by
including PPV values of new recordings. This arises from the fact
that conventional predictors represent only approximate models,
which take into consideration a scaled distance as the most
important input unit, while the blast-induced ground motion
depends on a wide scale of different inuential parameters, such
as total charge, stemming, hole depth, physico-mechanical properties of rock mass and explosive characteristics [14]. The presented
analysis is done for the recordings of ground vibrations induced by
blasting at copper mine Bor in Serbia, during the excavation of
orebody . The blasting was performed at fourteen different
locations, with a total of 612 blast boreholes, and with maximum
1226 kg charge per delay. The ground vibrations were measured
at three monitoring points, placed at different distances from the
explosive charge.

20

R. Lapevi et al. / International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

2. Geological setting and rock excavation method


Bor mine deposit is one of the largest copper deposits in
Europe, with production of almost 5.7 million mt of ore in the
rst six months of 2011 [16]. Several orebodies have been mined at
this site, but currently mining is concentrated in orebody T,
a relatively small orebody with 200,000 mt of ore, but rich in
copper (grading 5%-plus). Production from orebody T was
10,000 mt/month in 2011 [16].

Orebody T is located in the central part of Bor copper mine,


at relatively great depth, between 470 and 520 m below the
surface. It is 60 m long, 40 m wide and 4550 m high [16]. From
the geological point of view, wider area is mainly built of the
andesites, which are, in general, hydrothermally altered, due to
chloritization and kaolinization. Some smaller parts of andesites
are silicated. Tectonically, orebody T is located between two
large fault zones with eastwest direction, as shown in Fig. 1.
During 2011, detailed geological and geotechnical analysis
showed that exploitation of this ore body with conventional
mining methods would be associated with higher risk, cost and
potential losses of very rich ore. Therefore, it was suggested that
further excavation should be done with complete ensuring of
stability, using the NATM method. In that way, an underground
chamber of great dimensions was being formed. The ore excavation started from the upper zone, with the approach to the
excavated area through the spiral ramp and upper transportation
ramp. The excavated ore was transported through the vertical
shaft to the transportation horizon. The excavation was performed
in horizontal layers, about 5 m high, and was conducted from top
to bottom in phases (Fig. 2).
Every excavation phase consisted of blasting and successive
installation of the support system, with grouted anchors of different lengths (9 m, 15 m, 20 m and 25 m), connected with vertical
ribs and horizontal beams of reinforced concrete, inbetween
which a 25 cm thick shotcrete MBB35 is projected, with reinforcing mesh R503 on both sides of the shotcrete. The excavated
underground chamber is one of the deepest and biggest chambers
in this part of Europe, approximately 50 m high, 40 m long, 50 m
wide, and 470520 m below the surface (Fig. 3).

3. Blasting and eld measuring


Fig. 1. Geological cross-section of the orebody T: 1 pelite and tuff; 2 weakly
silicated andesite; 3 silicated andesite; 4 kaolinized andesite; 5 chloritized
andesite; 6 orebody T; 7 orebody F; and 8 old tailings. Red lines denote
large fault zones. (For interpretation of the references to color in this gure legend,
the reader is referred to the web version of this article.)

Field work consisted of (a) blasting, performed at fourteen


different locations inside the underground chamber and
(b) recording of blast-induced ground motion, at three monitoring
stations (Fig. 4). Every blasting series consisted of 2082 horizontal boreholes, with two to fteen boreholes per blasting interval,

Fig. 2. (a) Middle phase of the excavation. The excavated area is approached through the spiral and upper transportation ramps. The excavated ore is transported through the
vertical shaft. (b) Excavation phases of orebody T. Numbers denote phases of the excavation. Red line denotes the contour of orebody, while the black line denotes the
excavation contour. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of this article.)

R. Lapevi et al. / International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

drilled into the side wall of the underground chamber. Maximal


amount of explosives per interval was 1226 kg.
The ground vibrations were registered with 3 accelerometers
of D110-T type, with dynamic range of 120 dB, and sensitivity
threshold of 101000 Hz. The accelerometer S1 was installed about
12 m below the chamber (at  133 m depth), while accelerometers
S2 and S3 were installed in the chamber (at  121.5 m and
121.2 m depth, correspondingly). All three accelerometers were

21

placed on metal platforms, tighten to rock and secured with


special metal protection from the ying rock remnants (Fig. 5).
This step was necessary, since the stability of the recording station
needed to be secured, in order to obtain the valid data for the
analysis. We are aware of the fact that this acquisition system
would affect the quality of the results, since the accelerometers
were not directly coupled to rock. However, we assume that the
acquired data were recorded precisely enough for this analysis,
considering the large registered values of peak particle velocities.

4. Data set
A total of forty-two blast vibration records were used for
development of an ANN model, from which twenty-one data sets
were used for training and the rest for the validation and testing of
the neural network. Different blast parameters collected from the
site are PPV (mm/s), total charge (kg), maximum charge per delay
(kg), charge per hole (kg), delay time (s) and distance between the
shot point and monitoring station (m).

5. Prediction of peak particle velocity (PPV)


Fig. 3. Excavated underground chamber with installed support system.

5.1. Prediction of PPV using conventional predictors

Fig. 4. Distribution scheme of blasting series and measuring stations bottom of


underground chamber. Numbers 114 denote the distribution of blasting shots,
while S1, S2 and S3 stand for the position of measuring stations. Chamber bottom
has irregular shape, indicated by different heights of particular points, denoted as
negative depth below the sea level. Station S1, located in the lower chamber, is
shown in the same level with other points, due to simplicity, but with the clearly
lower depth (  133 m).

In order to justify the development of a new PPV prediction model


by using the ANN approach, rst we turn to common empirical
attenuation equations, which represent prediction models for PPV as
a function of scaled distance [17]. Various conventional predictors
proposed by different researchers are given in Table 1 [712]. These
equations are developed on the basis of the assumption that the total
energy of the ground motion generated by blasting varies directly
with the weight of detonated explosives per delay and the distance
from the blasting shot point.
The site constants were determined from the multiple regression analysis of the forty-two recordings (Table 2).
The relationship between measured and predicted PPV by conventional predictor equations is given in Fig. 6. As can be seen, in case
of using conventional predictors for estimating PPV, the coefcient
of determination (R2) varies between 0.13 (LangeforsKihlstrom) and
0.31 (General predictor), which could be explained due to uncontrollable underground physical conditions and their effect on fragmentation mechanism [18]. On the other hand, regarding the past
researches on this subject, the values of R2 above 0.7 indicate that the
measurement data could be used for PPV prediction by deploying
the aforementioned equations [19,20], which is not the case in
present study.

Fig. 5. D110-T type accelerometers at the recording sites S1 (a), S2 (b) and S3 (c).

22

R. Lapevi et al. / International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

5.2. Prediction of PPV using articial neural network approach (ANN)


Results of the analysis from the previous section indicate that
conventional methods did not give accurate prediction of PPV.
Table 1
Conventional predictors.a
Conventional predictor

Equation
p
v KR= Q max   B
q
v K Q max =R2=3 B

Duvall and Petkof (USBM) [7]


Langefors and Kihlstrom [8]

v KR  B Q max A
p
v KR= 3 Q max   B
p
v n KR= Q max   1

General predictor (Davies et al. [9])


Ambraseys and Hendron [10]
CMRI (Pal Roy [12])

a
v is the peak particle velocity (PPV) in mm/s, Qmax is the maximum charge per
delay, in kg, R is the distance between the blasting source and vibration monitoring
point, in meters, and K, B, A and n are site constants.

Table 2
Calculated values of site constants.
Equation

Duvall and Petkof (USBM) [7]


Langefors and Kihlstrom [8]
General predictor (Davies et al. [9])
Ambraseys and Hendron [10]
CMRI (Pal Roy [12])

Site constants
K

114.2
97.46
210.46
124.8
83.09

0.08
 0.01
0.14
0.1

 0.077

90.95

As a next step, we develop a neural network model, by using


the same approach as in [21] with total charge, maximum charge
per delay, distance from monitoring station to blasting shot,
charge per hole and delay times as input parameters, whereas
PPV was considered as the single output parameter (Table 3).
Stemming and hole depth were the same for all the boreholes
(0.5 m and 3 m, correspondingly), so these parameters were not
analyzed.
In the present study, in order to create an adequate ANN model for
prediction of PPV, a three-layer articial neural network is used with
the back-propagation training rule optimized by BroydenFletcher
GoldfarbShannon (BFGS) algorithm and with sigmoid activation
function. The mathematical summary of the back-propagation learning algorithm is given in [22]. The back-propagation learning algorithm for multilayer networks performs a gradient descent in weight
space to search for a minimum of some cost function. A general
drawback of gradient-based numerical optimization methods is their
slow convergence [23,24]. In learning problems, in particular, one
typically starts a long way from the solution, and spends most of the
time oscillating in weight space, because the gradient is sharp in some
directions, but shallow in others. Consequently, the learning parameters tend to be selected in an ad-hoc manner, according to the
particular problem and the current performance of the network.
We apply the BFGS algorithm, which is considered one of the best
of the quasi-Newton's techniques, that uses a local quadratic approximation of the error function, like Newton's method, but employs an
approximation of the inverse of the Hessian matrix to update the
weights, thus getting the lowest computational cost. The BFGS
algorithm is error tolerant, yields good solutions and converges in a
small number of iterations [25]. The computational advantage of BFGS

Fig. 6. Measured PPV vs. predicted PPV by conventional predictors: (a) USBM, (b) LangeforsKihlstrom, (c) General predictor, (d) AmbraseysHendron, and (e) CMRI. It is
clear that each of the predictor gives rather low coefcient of determination, in the range R2 0.130.31.

R. Lapevi et al. / International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

Table 3
Inputoutput parameters for the ANN training and their range.
Type of data

Parameter

Range

Inputs

Total charge (kg)


Maximum charge per delay, Qmax (kg)
Distance from blasting shot (m)
Charge per hole (kg)
Delay time (ms)
Peak particle velocity (mm/s)

40140
1226
846.7a
1.82
34500
281873

Output

a
Distance from the shot point to the monitoring stations was determined as
the minimal distance through rock along the edge of the chamber.

exception. Mohamadnejad et al. [29] examined even smaller


number of data (37) using support vector machine algorithm
and regression neural network, obtaining rather high prediction
accuracy (R2 0.92). Moreover, Monjezi et al. [21] developed
a four-layer feed-forward back-propagation neural network, using
only twenty data sets. In this case, high prediction accuracy was
also obtained (R2 0.927).
In order to utilize the most sensitive part of neuron and since
output neuron being sigmoid can only give output between 0 and
1, scaling of the output parameter was necessary, and was
performed in the following way:
scaled value

Table 4
Proposed articial neural network with various number of hidden nodes.
No. of hidden
nodes
1

10

12

15

20

25

30

Data set

Coefcient of
determination (R2)

Mean squared
error (MSE)

Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing
Training
Validation
Testing

0.645
0.404
0.624
0.582
0.712
0.484
0.684
0.604
0.658
0.717
0.392
0.659
0.724
0.511
0.648
0.789
0.539
0.635
0.769
0.714
0.694
0.702
0.501
0.916
0.692
0.610
0.672
0.994
0.599
0.650

0.012
0.025
0.047
0.024
0.017
0.024
0.010
0.056
0.022
0.017
0.042
0.015
0.015
0.028
0.019
0.008
0.035
0.036
0.007
0.033
0.021
0.001
0.077
0.008
0.023
0.027
0.054
0.001
0.013
0.352

especially holds for small to moderate sized problems [26], which is


the case in the present analysis.
After analyzing several cases of networks with various numbers
of hidden nodes, the most precise model for PPV prediction was
obtained by a neural network with one hidden layer and twenty
hidden nodes (Table 4).
In this case, we chose a feed-forward back-propagation neural
network with logistic activation function and BFGS training algorithm, as it was already used in [14,15,27]. The total data set
comprising forty-two points has been divided as follows: 50% of
the data for training (twenty-one recordings), 20% for validation
(eight) and the remaining 30% for testing (thirteen).
We are aware of the fact that the analysis of this relatively small
data set could lead to ambiguous results and interpretations.
Common approach usually considers more than a 100 PPV
recordings, in order to obtain reliable results [14,15,28], enabling,
in that way, a training of network with larger dataset, which is a
crucial step towards the ANN model with high prediction accuracy.
However, regarding the application of ANN approach for prediction of blasting vibration, the analysis of small data sets is not an

23

max value unscaled value


max value  min value

In this way, numerical values of the analyzed parameter were


normalized in the range of [0,1]. The resulting neural network
model with scaled values for training, validation and testing set is
shown in Fig. 7(a)(c), while the same model with scaled, actually
measured values is given in Fig. 7(d). Rather high coefcient of
determination (R2 0.916) demonstrates good performance of the
proposed network.

6. Evaluation of models performance


If we compare the values of PPV predicted by different methods
(conventional predictors and ANN), it is clear that prediction by
ANN is closer to the measured PPV, while conventional predictors
give weaker prediction (Fig. 8). Performances of the developed
predictor models were evaluated using different standard statistical error criteria given in Table 5 [21]. Calculated statistical errors
are given in Table 6. It is clear that ANN has the lowest values
of Mean Absolute Percentage Error (MAPE), Variance Absolute
Relative Error (VARE) and MEDian Absolute Error (MEDAE), while
it has the highest value of Variance Account For (VAF), in comparison to conventional predictors, which conrms the best
predictive power of the suggested ANN model.

7. Sensitivity analysis
Sensitivity analysis represents a method that enables us to
determine the relative strength of effect (RSE) for each input unit
on the nal value of PPV [14,21]. In this case, it was carried out by
the hierarchical analysis [30], where the RSE parameter is determined in the following way:
RSEki C :::W jn k Gek W jn  1 jn Gejn W jn  2 jn  1 Gejn  1 W jn  3 jn  2 Gejn  2 :::W ij1 Gej1
jn jn  1

j1

2
where C is normalized constant which controls the maximum
absolute values of RSEki, W is a connected weight, and G(ek) exp
( ek)/(1 exp(  ek))2 denotes the hidden units in the n, n 1,
n  2,,1 hidden layers [30].
Global sensitivity analysis, which was carried out for all the
input parameters, indicated that the distance from blasting shot
point and delay time have the strongest impact on the PPV value
(Fig. 9), which compares well with the previous research on this
topic [14,15,21].

8. Conclusions
We developed an articial neural network model for PPV
prediction, on the basis of the recorded ground vibrations induced
by blasting at fourteen different locations, during the excavation of
orebody T in copper mine Bor in Serbia. The recording was

24

R. Lapevi et al. / International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

Fig. 7. The comparison of the predicted and measured values of PPV for training (a), validation (b) and testing (c) set (scaled values); (d) the same comparison for the
unscaled values (testing dataset).

Fig. 8. Comparison of recorded and predicted PPV by using different predictors.


Abbreviations AH, GP and LK stand for AmbraseysHendron, General Predictor and
LangeforsKihlstrom, respectively. It is clear that ANN gives more accurate prediction in comparison to the conventional predicting techniques.

performed at three different monitoring stations, with a total


dataset of forty-two measurements. Even though the analyzed
dataset is relatively small from the perspective of ANN modeling, it
represents valuable experimental datum that has never been

analyzed in such a manner for any excavated object in Serbia, as


far as the authors are aware.
The conducted analysis showed that, by using conventional equations, a small coefcient of determination is obtained (R2 0.130.31),
indicating that these predictors may not be used for the purpose of
underground blasting, probably due to large recorded values of PPV
and complex surface geometry, which could signicantly affect the
seismic wave propagation. Moreover, structural-tectonic conditions,
such as the large fault zones, or the existence of different joint systems,
could determine the main propagation direction, which all makes the
application of conventional predictors even more difcult.
On the other hand, by applying ANN, we developed a prediction model with satisfying accuracy (R2 0.916). We applied the
BFGS learning algorithm, which is error tolerant and converges in
a small number of iterations. Another advantage of such an
approach is that the selection of hidden layers and number of
neurons in those layers demands no specic theorems, and it is
usually obtained by trial and error. Further analysis showed that
ANN has the lowest values of statistical error parameters MAPE,
VARE, and MEDAE, while it has the highest value of VAF, in
comparison to conventional predictors.
As for the inuence of input parameters on PPV, global
sensitivity analysis showed that the distance from the blasting
shot point and delay time have the strongest impact on the nal
value of PPV in comparison to the other parameters (maximum
charge per delay, total charge and charge per hole).

R. Lapevi et al. / International Journal of Rock Mechanics & Mining Sciences 69 (2014) 1925

Table 5
Statistical error parameters used for models' evaluation.
Statistical parameter
Mean absolute
percentage error
Variance absolute
relative error
Median absolute error
Variance account for
a

25

References
a

Equation
i

h


MAPE 1n  ni 1 ti t i xi   100



2  





V ARE 1n  ni 1 ti ti xi   meant i ti xi 
 100
MEDAE mediant i  xi
h
i
t i  xi
V AF 1  varvart
 100
i

ti represents measured value of PPV, while xi denotes predicted value of PPV.

Table 6
Statistical errors of different models for predicting PPV.
Model

MAPE

VARE

MEDAE

VAF

Duvall and Petkof (USBM) [7]


Langefors and Kihlstrom [8]
General predictor (Davies et al. [9])
Ambraseys and Hendron [10]
CMRI (Pal Roy [12])
ANN

64.29
63.18
72.00
64.46
71.70
16.38

57.07
56.50
63.52
57.19
62.78
16.07

314.22
345.88
353.57
326.00
292.95
110.885

68.42
70.19
68.05
68.57
59.28
91.17

Fig. 9. Relative strength of effect (RSE) of each input parameter on the recorded
value of PPV, as a result of global sensitivity analysis.

However, even though the results of the analysis are satisfactory and encouraging (regarding the predictive power of ANN in
a rst place) there are still certain questions that remain open. Is
the developed model only valid for the investigated area (T
orebody, particularly), or could it be used in a general case of blast
induced vibrations during the underground excavation? Also,
could the developed model be improved by analyzing a larger
dataset? Only in that way, by broadening the presented research,
would the prediction power of the suggested model be fully
evaluated.
Acknowledgments
This research was supported by the Ministry of Education,
Science and Technological Development of the Republic of Serbia
(Grant nos. 36009, 176016 and 171017).

[1] Wittke W, Pierau B, Erichsen C. New austrian tunneling method (NATM)


stability analysis and design. Geotech Eng Res Pract, WBI-PRINT 2006;5
(Aachen).
[2] Surhone LM, Timpledon MT, Marseken SF. New Austrian tunneling method.
Vienna: VDM Publishing; 2010.
[3] Girmscheid G, Schexnayder C. Drill and blast tunneling practices. Pract Period
Struct Des Constr 2002;7:12533.
[4] Henrych J. The dynamics of explosion and its use. New York: Elsevier; 1979.
[5] Wiss JF. Construction vibrations, state-of-the-art. J Geotech Eng ASCE
1981;107:16781.
[6] Dowding CH. Construction vibrations. 2nd ed. New Jersey: Prentice-Hall;
2000.
[7] Duvall WI, Petkof B. Spherical propagation of explosion of generated strain
pulses in rocks. USBM, Rep Invest 1959;5483:212.
[8] Langefors U, Kihlstrom B. The modern techniques of rock blasting. 3rd ed.
New York: Wiley; 1963.
[9] Davies B, Farmer IW, Attewell PB. Ground vibrations from shallow sub-surface
blasts. Engineer (London) 1964:5539.
[10] Ambraseys NR, Hendron AJ. Dynamic behavior of rock masses: rock mechanics
in engineering practices. In: Stagg KG, Zienkiewicz OC, editors. Rock
mechanics in engineering practice. New York: Wiley; 1968.
[11] Ghosh A, Daemen JK. A simple new blast vibration predictor, In: Proceedings
of the 24th US rock mechanics symposium; 1983. p. 15161.
[12] Pal Roy P. Putting ground vibration predictors into practice. Colliery Guardian
1993;241:6370.
[13] Khandelwal M, Singh TN. Evaluation of blast induced ground vibration
predictors. Soil Dyn Earthq Eng 2007;27:11625.
[14] Khandelwal M, Singh TN. Prediction of blast-induced ground vibration using
articial neural network. Int J Rock Mech Min 2009;46:121422.
[15] Monjezi M, Ghafurikalajahi M, Bahrami A. Prediction of blast-induced ground
vibration using articial neural networks. Tunn Undergr Sp Technol 2011;26:
4650.
[16] Stojadinovi S, iki M, Pantovi R. RTB Bor: the comeback of Serbian copper.
Eng Min J 2011;212:1027.
[17] Kuzu C. The importance of site-specic characters in prediction models for
blast-induced ground vibrations. Soil Dyn Earthq Eng 2008;28:40514.
[18] Erarslan K, Uysal O, Arpaz E, Cebi MA. Reply to the comments by Tarkan Erdik
on Barrier holes and trench application to reduce blast induced vibration in
Seyitomer coal mine. Environ Earth Sci 2010;61:10956.
[19] Kahriman A, Ozer U, Aksoy M, Kradogan A, Tuncer G. Environmental impacts
of bench blasting at Hisarcik Boron open pit mine in Turkey. Environ Geol
2006;50:101523.
[20] Iphar M, Yavuz M, Ak H. Reply to the comment on Prediction of ground
vibrations resulting from the blasting operations in an open pit mine by
adaptive neuro-fuzzy inference system by Tarkan Erdik. Environ Earth Sci
2009;59:4736.
[21] Monjezi M, Hasanipanah M, Khandelwal M. Evaluation and prediction of
blast-induced ground vibration at Shur River Dam, Iran, by articial neural
network. Neural Comput Appl 2013;22:163743.
[22] Pao YH. Adaptive pattern recognition and neural network. Reading. MA:
Addison-Wesley; 1989.
[23] Becker S, Le Cun Y. Improving the convergence of back-propagation learning
with second-order methods. In: Touretzky D, Hinton G, Sejnowski T, editors.
Proceedings of the 1988 connectionist models summer school; 1988. p. 2937.
[24] Castillo E, Berdinas-Guijarro B, Fontenia-Romero O, Alonso-Betanzos A. A very
fast learning method for neural networks based on sensitivity analysis. J Mach
Learn Res 2006;7:115982.
[25] Dennis JE, Schnabel RB. Numerical methods for unconstrained optimization
and nonlinear equations. Philadelphia: Society for Industrial and Applied
Mathematics; 1987.
[26] Watrous RL. Learning algorithms for connectionist networks: applied gradient
methods of nonlinear optimization. In: Proceedings of the 1st international
conference on neural networks, vol. 2; 1987. p. 61927.
[27] Khandelwal M, Singh TN. Prediction of blast induced ground vibrations and
frequency in opencast mine: a neural network approach. J Sound Vib
2006;289:71125.
[28] Dehghani H, Ataee-pour M. Development of a model to predict peak particle
velocity in a blasting operation. Int J Rock Mech Min Sci 2011;48:518.
[29] Mohamadnejad M, Gholami R, Ataei M. Comparison of intelligence science
techniques and empirical methods for prediction of blasting vibrations. Tunn
Undergr Sp Technol 2012;28:23844.
[30] Yang Y, Zang O. A hierarchical analysis for rock engineering using articial
neural networks. Rock Mech Rock Eng 1997;30:20722.

You might also like