1-s2.0-s1674775513001157-main.pdf

Journal of Rock Mechanics and Geotechnical Engineering 6 (2014) 6776

Journal of Rock Mechanics and Geotechnical Engineering

Journal of Rock Mechanics andEngineering

journa l homepage: www.rockge

An ANN cedGol-E-G

Mahdi Saa Department o nb Department o ty of Ac Faculty of Eng

a r t i c l

Article history:Received 14 AReceived in re13 November Accepted 18 November 2013

Keywords:Blast-induced ground vibrationEmpirical predictorsArticial neural network (ANN)Multiple linea

of thess as t an aGol-E

propagation multi-layer perceptron (MLP) was used and trained with LevenbergMarquardt algorithm.To construct ANN models, the maximum charge per delay, distance from blasting face to monitoring point,stemming and hole depth were taken as inputs, whereas peak particle velocity (PPV) was considered asan output parameter. A database consisting of 69 data sets recorded at strategic and vulnerable locationsof GEG iron mine was used to train and test the generalization capability of ANN models. Coefcient of

1. Introdu

Blastingmining andvibration hastructures ation and assblasting acttion, vital cpropagation

CorresponE-mail add

(M. KhandelwPeer review uAcademy of Sc

ELSEVIER

1674-7755 2Sciences. Prodhttp://dx.doi.or regression determination (R2) and mean square error (MSE) were chosen as the indicators of the performance ofthe networks. A network with architecture 4-11-5-1 and R2 of 0.957 and MSE of 0.000722 was foundto be optimum. To demonstrate the supremacy of ANN approach, the same 69 data sets were used forthe prediction of PPV with four common empirical models as well as multiple linear regression (MLR)analysis. The results revealed that the proposed ANN approach performs better than empirical and MLRmodels.

2013 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting byElsevier B.V. All rights reserved.

ction

is one of the most pervasive excavation methods in civil engineering projects. Since blast-induced grounds an inevitable impact on rock mass as well as nearbynd human beings, the prediction of blast-induced vibra-essment of its effects must be performed prior to actualivities. In order to control the harm of blasting vibra-onsideration should be taken into the generation and

mechanism of blast-induced vibration. While any of

ding author. Tel.: +91 2942471379.resses: [email protected], [email protected]).nder responsibility of Institute of Rock and Soil Mechanics, Chineseiences.

Production and hosting by Elsevier

013 Institute of Rock and Soil Mechanics, Chinese Academy ofuction and hosting by Elsevier B.V. All rights reserved.rg/10.1016/j.jrmge.2013.11.001

three kinematic descriptors (displacement, velocity and accelera-tion) could be employed to describe ground motion, among these,peak particle velocity (PPV) is the most preferable (ISRM, 1992).All the empirical predictors are based on two parameters, i.e. themaximum charge per delay and the distance from the blastingface to the monitoring point. It is well-known that the intensityof PPV is closely associated with physico-mechanical parametersof rock mass, blast design as well as explosive (Khandelwal andSingh, 2009; Khandelwal et al., 2011), which are underestimatedor overestimated while using available conventional predictors. Itis important to understand the inter-relation among these param-eters, to use appropriate blasting pattern, and to evaluate thedetrimental impact of blasting. Moreover, empirical models arenot well suitable for predicting any other important parame-ters such as frequency, air over pressure, y rocks, etc., whichare equally important and critical for safe, smooth and environ-mentally friendly excavation of rock mass for mining and civilengineering projects (Monjezi et al., 2006; Khandelwal and Singh,2009).

To understand the complicated nature of ground vibration andto overcome limits of empirical models, articial neural network(ANN) which has the ability to address the complicated problemscan be implemented. ANN is a branch of the articial intelligencescience and has been developed rapidly since 1980s. This method-based approach to predict blast-induohar iron ore mine, Iran

adata, Manoj Khandelwalb,, M. Monjezi c

f Mining Engineering, Shahid Bahonar University of Kerman, Kerman 76196-37147, Iraf Mining Engineering, College of Technology & Engineering, Maharana Pratap Universiineering, Tarbiat Modares University, Tehran, Iran

e i n f o

ugust 2013vised form2013

a b s t r a c t

Blast-induced ground vibration is onecause substantial damage to rock maan attempt has been made to presenblast-induced ground vibration of the Geotechnical

otech.org

ground vibration of

griculture & Technology, Udaipur 313001, India

inevitable outcomes of blasting in mining projects and maywell as nearby structures and human beings. In this paper,pplication of articial neural network (ANN) to predict the-Gohar (GEG) iron mine, Iran. A four-layer feed-forward back

68 M. Saadat et al. / Journal of Rock Mechanics and Geotechnical Engineering 6 (2014) 6776

Table 1Blasting parameters of GEG iron ore mine.

Explosive type Blast hole pattern Bench height (m) Hole diameter (m) Rows per blast Holes per row

ANFO 03

is capable oa process, wSimilar to epensive andANN to pretwo parameitoring poinndings wiat an opencusing rock,of ANN andanalysis (Khseveral ANNmore inputMonjezi et Siahbisheh the distancming and hresults withBy sensitiviing face is thparameter oa model to (2013) propShur River quency bas2010, 2012and found vods. The idevibration inful functionempirical m(MLR) analy

2. Site des

The GEGin Kerman a point appShiraz and has six anomopen-pit mSanandajSclassied intop, bottomand-blast mthe rock typof blast-inding design pof GEG min

3. Empiric

In order (SD) laws aare linear rat various cal blast-in

ious inednd an PPn m

coef(The predngefo

ltiple

ltipleelatindene moted ated fn (S

+ 1 Y is t0 is coefe prndep

ANN

133

H is coe

of PPPPVs

rvie

is neuuncti

andnsistsing f thest cotechn

ulti-l

netimpicula

mornt roStaggered 15 0.2

f extracting the relation between inputs and outputs ofithout the physics being explicitly provided to them.

mpirical models, this method is computationally inex- easy to implement. Khandelwal and Singh (2007) useddict PPV at a magnesite mine in India. They consideredters, i.e. the distance from the blasting face to the mon-t and explosive charge per delay, and compared theirth commonly used predictors. In their other researchesast mine, they studied blast vibration and frequency

blast design and explosive parameters with the help compared their results with multivariate regressionandelwal and Singh, 2006). Mohammad (2009) used

models on Assiut limestone and concluded that using data can improve the capability of ANN to predict PPV.al. (2011) developed an ANN model to predict PPV atproject in Iran, using the maximum charge per delay,e from the blasting face to the monitoring point, stem-ole depth as input parameters and compared their

empirical models and multivariate regression analysis.ty analysis, they found that the distance from the blast-e most effective and the stemming is the least effectiven the PPV. Dehghani and Ataee-pour (2011) developedpredict PPV using dimensional analysis. Monjezi et al.osed an ANN-based solution for prediction of PPV at

dam, Iran. Other researchers predicted PPV and/or fre-ed on ANN models in different projects (Amnieh et al.,; Alvarez-Vigil et al., 2012; Mohamadnejad et al., 2012)ery superior results compared to conventional meth-a of the present study is to predict blast-induced ground

Gol-E-Gohar (GEG) iron ore mine based on the power- approximation tool, ANN. The results of both ANN andodels were compared with multiple linear regressionsis to nd the applicability of each method.

cription and data set

iron ore mine is located in 60 km southwest of SirjanProvince of Islamic Republic of Iran. The mine lies atroximately equidistant from the cities of Bandar Abas,Kerman, at an altitude of 1750 m above sea level. GEGalies out of which, the rst one is under extraction by

ethod. This mine is situated on the northeast margin ofirjan tectonic-metamorphic belt. Iron ores at GEG are

three types based on their chemical characterization,, and oxide magnetic. The deposit is excavated by drill-ethod. Because of the complex discontinuity existence,e variations and the water bearing beds, the evaluationuced ground vibration is critically important. The blast-arameters of the GEG are listed in Table 1. A photographe is shown in Fig. 1.

al methods for predicting PPV

to control the harm of blasting vibration, scaled distance

by vardetermscale abetweebetweehigherUSBM better and La

4. Mu

Mulinear rindepethat thpredicconducequatio

Y = 0wheretors, is the cwith thinput-iused in

PPV =

whereThe

valuesdicted

5. Ove

ANNhumanpute finputstem coprocesment othe mo(MLP)

5.1. M

MLPers of sA parttwo ordifferere developed by various eld investigators. SD lawsegression, in a loglog plane between PPVs recordeddistances during a blast. Table 2 shows the empiri-duced ground vibration predictor equations proposed

work is calloutput sidecalled hiddration of th27 1020

researchers. The values of site constants K and B are by plotting the graph between PPV and SD on loglogre shown in Table 2. Fig. 2 shows the loglog plotsV and different SDs. Fig. 3 illustrates the relationshipeasured and predicted PPVs by various SD laws. Thecient of determination for AmbreseysHenderson andUnited States Bureau of Mines) predictors indicates theiction capability over Bureau of Indian Standard (BIS)rsKihlstrom predictors.

linear regression analysis

linear regression (MLR) is a method used to model theonship between a dependent variable and one or moret variables. MLR is based on least squares, which meansdel is t such that the sum of squares of differences ofnd measured values are minimized. An MLR has beenor the prediction of PPV. MLR is given by the followingcheaffer et al., 2011):

X1 + + PXP + e (1)he predicted variable, Xi (i = 1, 2, . . ., P) are the predic-alled intercept (coordinate at origin), i (i = 1, 2, . . ., P)cient on the ith predictor and e is the error associatededictor. MLR model was developed based on the sameendent variables and output-dependent variables as

model. This resulted in the following equation:

.8090 0.0002Qmax0.1103D + 11.7270H + 4.3550S(2)

the hole depth and S is the stemming.fcient of determination for predicted and measuredV is 0.276. Fig. 4 shows the plot of measured and pre-

by MLR model.

w of articial neural network

a form of articial intelligence which is based on theronal system. ANN can be used to learn and com-ons for which the analytical relationships between

outputs are unknown. An ANN is a computing sys-ing of highly interconnected set of simple informationelements called neurons or perceptrons. The arrange-se neurons determines the ANN architecture. One ofmmonly implemented ANNs is multi-layer perceptronique.

ayer perceptron network

works are feed-forward networks having several lay-le computing elements, called neurons or perceptrons.r network could contain one or more layers in whiche perceptrons can be combined. The layers of MLP haveles. The interfacing layer at the input side of the net-

ed sensory layer (or input layer in common); the one at

is referred as output layer. All intermediate layers areen layers (Sethi and Jain, 1991). Based on the congu-e connections between neurons of different layers, the

M. Saadat et al. / Journal of Rock Mechanics and Geotechnical Engineering 6 (2014) 6776 69

number of A2012). Feedof ANNs, inthrough all2007). MLPtine called (a)

(b)

Fig. 1. (a) Location of GEG mine in Iran map and (b) the

NN architectures can be obtained (Raai and Moosavi,-forward neural networks (FFNNs) are a special kind

which the inputs are received and simply forwarded the next layers to obtain the outputs (Engelbrecht,

employs an iterative gradient-based optimization rou-back-propagation (BP) learning technique, a kind of

FFNN modeof BP is thattage is thatand Hoft, 19are continutraining and picture of GEG mine.

l (Rumelhart et al., 1986; Leondes, 1998). The advantage it is simple and easy to understand, but the disadvan-

convergent speed is slow and BP is not so robust (Lin94). The mathematical functions implemented by MLPous and differentiable, which signicantly simplies

error analysis. The perceptron is the basic structural


(a) USBM model. (b) Langefor s-Kihls trom model.

y = 121.0x1.25

1

10

100

1000

0.1 1 10 100

PPV

(mm

/s)

USBM scal ed dist anc e D/W 1/2

y = 0.442x1.637

1

10

100

1000

1 10 100

PPV

(mm

/s)

LangeforsK ihlstrom sc aled distan ce W1/2/D1/3

(c) Am bra seys -Hendron model .

y = 812 .0x1.29

1

10

100

1000

1 10 100

PPV

(mm

/s)

Ambrasey s-Hendron scaled dist anc e D/W1/3

(d) BI S model.

y = 0.44 2x0.818

1

10

100

1000

1 10 100 1000

PPV

(mm

/s)

Bureau of indi an stand ard s scaled di stanc e W/D2/3

Fig. 2. Loglog plots between PPV and scaled distance for various models (W is the charge per delay, kg; D is the distance between blasting face to vibration monitoringpoint, m).

(a) USBM predictor.

R = 0.4461

0

50

100

150

200

250

4003002001000Pre

dict

ed PP

V by

USB

M (m

m/s)

Measured PPV (mm/s)(b) Langefors-Kihlstrom predictor.

R = 0.1397

01020304050607080

4003002001000

Pred

icte

d PP

V by

Lan

gefo

rs

Kih

lstro

m (m

m/s)

Measured PPV (mm/s)

(c) AmbraseysHendron predictor.

R = 0.4341

0

50

100

150

200

250

4003002001000

Pred

icte

d PP

V by

Am

bras

eys

Hen

dron

(mm/

s)

Measured PPV (mm/s)(d) Bureau Indian Standards predictor.

R = 0.1397

0

10

20

30

40

506070

80

4003002001000

Pred

icte

d PP

V by

Bur

eau

of In

dian

St

anda

rd (m

m/s)

Measured PPV (mm/s)

Fig. 3. Graphs between measured and predicted PPVs by various predictors.


Table 2Different empirical predictors.

Name Equation K B

USBM(Duvall and Fogleson, 1962)

v = K(D/

Qmax)B

121.0 1.25

LangeforsKihlstrom(Langefors and Kihlstrom, 1963)

v = K(

Qmax/D2/3)B

0.442 1.637

AmbraseysHendron(Ambraseys and Hendron, 1968)

v = K(D/Qmax1/3)B

812.0 1.29

Bureau of Indian Standard(Indian Standard Institute, 1973)

v = K(Qmax/D2/3)B 0.442 0.818

Note: v is the PPV (mm/s) and Qmax is the maximum charge per delay (kg).

R = 0.2769

Pred

icte

d PP

V by

MLR

m

ode

l (mm

/s)

Fig. 4.

element of fThe perceptceptron areweighted intion f (Blackas follows:

y = f(

Ni=1

where xi is input, b is t

5.2. Networ

NetworkThe traininstages (Sum

(1) initializ(2) feed-for(3) back-pr(4) updatin

Fig. 5. Str

igmoi

ious imation problems. LevenbergMarquardt appears to be the

method for training moderate-sized FFNNs (up to severald weights). This algorithm has high accuracy and conver--60-40-20

020406080

100120140160

4003002001000Measu red PP V (mm/s)

Graph between measured and predicted PPVs by MLR model.

eed-forward multi-layer perceptron (FFMLP) networks.ron with n inputs is shown in Fig. 5. The inputs to a per-

weighted with an appropriate weight (w). The sum ofputs and the bias (b) form the input for transfer func-well and Chen, 2009). Transfer function can be written

)

Fig. 6. S

Varapproxfastesthundrewixi + b (3)

the ith input, wi is the weight associated with the ithhe bias and f is the transfer function of the perceptron.

k training

training is essential before interpreting new results.g algorithm of back propagation (BP) involves fourathi and Paneerselvam, 2010):

ation of weights,ward,opagation of errors, andg of weights and biases.

ucture of an elementary perceptron (Blackwell and Chen, 2009).

gence speesoftware, bein function aMATLAB enstudy, this a

The behaand weightlayer, wherthe output lare taken ttansigmoiused in BP (is dened a

f = 11 + eex

where ex is

Fig. 7. Linear gg

a =t an[s ig(n

a =log10 [si

)]

(n)]

d transfer functions used for hidden layers (MathWorks Inc., 2009).

training algorithms have been developed for functiond. It also has an efcient implementation in MATLABcause that the solution of the matrix equation is a built-nd that its attribute becomes even more pronounced invironment (MathWorks Inc., 2009). Thus, in the presentlgorithm is used for training the network.vior of ANN mainly depends on both transfer functions

s. The output of transfer function is passed to the outpute it is multiplied by the connection weights betweenayer and hidden layer, and again a number of productso generate the output for the network. Logsigmoid,d and linear transfer functions are the most commonlyFigs. 6 and 7). The logarithmic sigmoid function (logsig)s (MathWorks Inc., 2009)

(4)

the weighted sum of the inputs for a processing unit.

a=pur eli n(n)

transfer functions used for output layer (MathWorks Inc., 2009).


Table 3The range of variables used to train the network.

Input Output

Charge weight per delay equivalent to ANFOQ (kg)

Distant from the blastpoint D (m)

StemmingS (m)

Hole depthH (m)

Peak particle velocityPPV (mm/s)

160631,573 401092 37 1418 2.2360

The tangent sigmoid function (tansig) is dened as follows(MathWorks Inc., 2009):

f = eex eex

eex + eex (5)

5.3. Monitoring the validation and performance of ANN

Typically neural networks adapt satisfactorily to training usu-ally the data but the test performance does not provide signicantresults. The unfavorable generalizing property is caused by theovertting of neural network parameters to the training data. Over-tting occurs when the neural network begins to memorize thetraining set instead of learning them, and consequently loses the

ability to generalize (Raai and Moosavi, 2012). As the training pro-cess continues, the error on the training set decreases while onthe validation set, it decreases initially and subsequently increasesafter some training period. Early stopping is an effective remedyfor resolving this issue. In early stopping, weights are initialized tovery small values. Part of the data sets is used for training the net-work and the other part is used for monitoring the validation error.Training is stopped when the validation error begins to increase.

Early stopping in its basic form is rather inefcient, as it is verysensitive to the initial condition of the network and only part ofthe data is used for training the model. These problems can easilybe alleviated by using committee of early stopping networks, withdifferent partitioning of the data to training, test and validation setsfor each network. MATLAB random selection of data sets is also

t

ith

Apply ANN for prediction of output (PPV)

?

20% for validation of the network

Divide data sets into three categories

Define the architecture of the network (inputs, outputs, number of neurons in first hidden layer, number of neurons in second hidden layer, and transfer functions)

Choose blast data sets

20% for test of the network 60% for training the network No Is there any significanrelationship betweentargets and outputs?

Simulate the network wchosen blast data sets

Yes

No Is the network response satisfactoryYes

END

Fig. 8. Flow chart for ANN method.


Table 4Results of different examined ANN models with one hidden layer.

Number of neurons Transfer function R R2 MSE

Training Validation Test Overall

25 tansig 0.99 0.72 0.82 0.93 0.77 0.004logsig 0.99 0.69 0.88 0.88 0.65 0.011

23 tansig 0.9 0.75 0.93 0.82 0.42 0.0163logsig 0.97 0.62 0.9 0.92 0.45 0.0158

20 tansig 0.99 0.94 0.88 0.93 0.89 0.00199logsig 0.99 0.61 0.97 0.85 0.135 0.0359

18 tansig 0.99 0.83 0.82 0.96 0.48 0.0158logsig 0.97 0.67 0.99 0.96 0.86 0.00347

15 tansig 0.99 0.65 0.93 0.96 0.86 0.00354logsig 0.97 0.83 0.94 0.96 0.45 0.0155

12 tansig 0.99 0.83 0.8 0.97 0.58 0.022logsig 0.99 0.58 0.95 0.97 0.64 0.0143

Table 5Results of different examined ANN models with two hidden layers.

Number of neurons Transfer functions R R2 MSE

Training Validation Test Overall

105 tansigtansig 0.99 0.89 0.95 0.98 0.77 0.00825115 tansigtansig 0.99 0.97 0.89 0.97 0.93 0.00139115 tansiglogsig 0.99 0.92 0.96 0.93 0.95 0.000722125 logsiglogsig 0.99 0.94 0.95 0.96 0.94 0.00647125 126 126 127

helpful in tprocedure smight havetrained on tmost appro(R2) and meThe perform

MSE = 1N

Ni=

where Ti, Ooutput and

elop

epara

d-ford funutpu

n. In r pretansiglogsig 0.97 0.94 tansigtansig 0.98 0.97 logsigtansig 0.99 0.85 tansigtansig 0.99 0.91

his remedy. This random selection might stop trainingo that the trained ANNs with different validation sets

different generalization errors even if ANNs would behe same training set (Chen and Wang, 2007). Finally, thepriate model based on its correlation of determinationan square error (MSE) will be chosen as the ANN model.ance function, MSE, can be calculated as follows:

6. Dev

6.1. Pr

Feesigmoiin the oimatiotool fo1

(Ti Oi) (6)

i and N represent the measured output, the predictedthe number of input-output data sets, respectively.

effective paThese parampropagationwhen inputment, pre-p

Fig. 9. Proposed ANN architecture0.99 0.98 0.94 0.001250.98 0.93 0.9 0.0005630.98 0.97 0.94 0.001380.95 0.94 0.69 0.00829

ment of the ANN-based prediction

tion of data for ANN models

ward back propagation articial neural network withctions in the hidden layers and linear transfer functiont layer are appropriate structures for function approx-this study, ANN was used as a function approximationdiction of PPV from four variables, which are the most

rameters in predicting PPV (Monjezi et al., 2011).eters are listed in Table 3. LevenbergMarquardt back

algorithm gives the most convincing performances and targets are scaled. In order to meet this require-rocess of the data is needed. The performance of this

.


Output 0. 98*Target+0 .04 4

Traini ng: R=0 .9945 5

0.25Data

Valida tion: R=0.9253 6

Fit

Y=T

Data

Target

0.0 0.2 0.4 0.6 0.8 1.0

1.0

All: R=0.9334 7

algorithm wso that therelationshipparameters

Xnew = X Xma

where Xnewis the maxifor all data.data in the

6.2. Networprediction

A comma set of 69 dmine. Theseabout 60% fnetwork (14(14 data set

MATLABTraining davector; whthe accuractraining prOutput 0. 99* Targe t+0 .0009 8

0.00 0.05 0.10 0.15 0.20 0.25

Target

Fit

Y=T

0.35

Fit

Y=T

Data

Test: R=0.9661 21.0

0.20

0.15

0.10

0.05

0.00

Ou

tpu

t

0.8

0.6

0.4

0.2

0.0

Ou

tpu

t

0.30

0.25

0.20

0.15Ou

tpu

t

0.8

0.6

0.4Ou

tpu

t Output 0. 85*Target+0 .01 4

0.0 0.1 0.2 0.3

Target

0.0

0.10

0.05

0.00

0.2

0.0

Fig. 10. Coefcient of correlation for the training, testing, vali

ill be high when the input and target are normalizedy fall approximately in the range [0, 1]. The following

is implemented for normalizing the input and target (Sivanandam et al., 2006):

Xminx Xmin

(7)

is the normalized value, X is the original value, Xmaxmum value for all data and Xmin is the minimum value

The normalized values are used as input and targetednetwork.

k architecture and validation of ANN-based

ittee of early stopping networks was built by employingata points recorded in vulnerable part of GEG iron ore

data sets are randomly divided into three categories,or training the network (41 data sets), 20% for test of the

data sets) and 20% for validation of training procedures). The ow chart for ANN method is illustrated in Fig. 8.

software was used for implementation of ANN.ta were used to obtain the weight matrix and biasile test and validation data were chosen to monitory and validation of the network. After stopping, theocedure and measurement of the accuracy of the

network pto reach thmodel. Valiperform thedeterminatMSE were uThe resultslayers are s

0

50

100

150

200

250

300

350

0

Pred

icte

d PP

V by

A

NN

( mm

/s)

Fig. 11. Fit

Y=T

DataOutput 0. 98*Target+0 .009 9

0.2 0.4 0.6 0.8 1.0

Target

dation and overall data sets.

erformance, 69 data sets, were simulated in ordere nal outputs and to test the accuracy of the ANNdation check is carried out to show that ANN models

nonlinear regression analysis properly. Coefcient ofion (R2) (between predicted and measured PPVs) andsed to compare the performance of each ANN model.

for different ANN models with one and two hiddenhown in Tables 4 and 5. These results revealed that an

R = 0.9577

100 200 300 400Measured PP V (mm /s)

Graph between measured and predicted PPVs by ANN model.


PPVs

Table 6R2 and RMSE o

Model

ANN AmbraseysLangeforsKBureau of InUSBM MLR

Note: RMSE is

ANN with ahidden layeand purelinas the mostcorrelation training, tes

Fig. 11 shby ANN mobetween mproposed mto measureand predict

7. Results

Fig. 12 sPPVs by ANfour input pthe measurmates or ovbased on emdence. MLRthe maximuthe monitoTable 6 showcan be easil

8. Conclus

Using baof hidden larithm, MSE respectivelyping netwoparts of GEGbeen foundlayer with 1output layeput parame

on is moR an

maxe tocal meters

effecLR mamen to LR antentised. ealedrceerablirica

strune sm

t of

auth

ncesFig. 12. Comparison of measured and predicted

f PPV by various models.

R2 RMSE

0.957 8.796Hendron 0.434 25.74ihlstrom 0.139 11.7dian Standard 0.139 11.66

0.446 22.610.276 20.85

roots of mean square error.

rchitecture 4-11-5-1, tansig transfer function in its rstr, logsig transfer function in its second hidden layer

transfer function in its output layer, can be selected tting predictor of PPV (Fig. 9). Fig. 10 illustrates thecoefcient (R) for the proposed ANN model includingt, validation and overall data.ows the graph between measured and predicted PPVs

del. The graph demonstrates that the value of R2 is 0.95easured and predicted PPVs. It is clearly shown that theodel has the capability to predict the PPV very close

d PPV values, and that the accuracy between measureded PPVs is acceptable.

and discussion

hows a comparison between measured and predicted

vibratiof ANNand ML

Theing facempiriparamtion asfrom Mtor parsolutioThe Mlike powere uels revLongfoconsidof empnearbythe mi

Conic

The

Refere

N, MLR and different predictor equations. Consideringarameters, ANN model predicted PPV is very close toed data. Prediction by empirical equations underesti-erestimates the value of PPV. Therefore, any predictionpirical equations requires more consideration and pru-

results show that the relationships between PPV andm charge per delay, distance from the blasting face to

ring point, stemming, and hole depth cannot be linear.s R2 and RMSE for all models. The authenticity of ANN

y established in comparison with other predictors.

ions

ck propagation ANN model with an optimum numberyer neurons and LevenbergMarquardt training algo-and R2 for prediction of PPV were 0.000722 and 0.95,. For building ANN models, a committee of early stop-rks was trained by 69 data sets recorded at vulnerable

iron ore mine in Iran. Optimum ANN architecture has to have four neurons in the input layer, two hidden1 and 5 neurons, respectively, and one neuron in ther. Considering the complexity between inputs and out-ters, the application of ANN technique in the ground

Alvarez-Vigil Adicting blaneural net2012;55:1

Ambraseys NRin enginee

Amnieh HB, tion in S2010;48(3

Amnieh HB, Speak parti2012;50(9

Blackwell WJ, Massachus

Bureau of Indiato undergr

Chen K, WangDehghani H, A

in a blastiSciences 2

Duvall WI, Fogblasting vi

Engelbrecht APInc; 2007.

ISRM. SuggestRock Mech

Khandelwal Mto predict2011;27(2 by various models.

proven to be promising. To ascertain the authenticitydel, the results were compared with empirical modelsalysis.imum charge per delay and the distance from the blast-

the monitoring point are the major parameters forodels. Whereas in MLR and ANN models, two more

, i.e. stemming and hole depth, are taken into considera-tive variables to predict PPV. The poor results obtainedodel revealed that the relationship between predic-

ters and PPV must not be linear. The capability of ANNnd nonlinear relationships is a testament to this fact.alysis could possibly be improved if other techniquesal or even exponential non-linear multiple regressionThe results obtained from traditional empirical mod-

that the Bureau of Indian Standard overestimates butKihlstrom underestimates predicted PPV values. Thise uctuation in prediction of PPV shows that any usel predictors without validation causes damage to thectures and imposes nancial penalties for hindrance toooth working.

interest

ors declare no conict of interest.

E, Gonzales-Nicieza C, Lopez Gayarre F, Alvarez-Fernandez MI. Pre-sting propagation velocity and vibration frequency using articialwork. International Journal of Rock Mechanics and Mining Sciences0816., Hendron AJ. Dynamic behavior of rock masses. In: Rock mechanicsring practices. London: Wiley; 1968. p. 2037.Mozdianfard MR, Siamaki A. Predicting of blasting vibra-archeshmeh copper mine by neural network. Safety Science):31925.iamaki A, Soltani S. Design of blasting pattern in proportion to thecle velocity (PPV): articial neural networks approach. Safety Science):19136.Chen FW. Neural networks in atmospheric remote sensing. MA, USA:

etts Institute of Technology; 2009. p. 7880.n Standard (BIS). Criteria for safety and design of structures subjectedound blast. ISI Bull; IS-6922; 1973.

L. Trends in neural computation. Rotterdam: Springer; 2007. p. 234.taee-pour M. Development of a model to predict peak particle velocityng operation. International Journal of Rock Mechanics and Mining011;48(1):518.leson DE. Review of criteria for estimating damage to residences frombration. USBM-RI, 5968; 1962.. Computational intelligence. Chichester, England: John Wiley & Sons,p. 27.ed method for blast vibration monitoring. International Journal ofanics and Mining & Geomechanics Abstracts 1992;29(2):1456., Kumar DL, Yellishetty M. Application of soft computing

blast-induced ground vibration. Engineering with Computers):11725.


Khandelwal M, Singh TN. Evaluation of blast-induced ground vibration predictors.Soil Mechanics and Earthquake Engineering 2007;27(2):11625.

Khandelwal M, Singh TN. Prediction of blast induced ground vibrations and fre-quency in opencast mine: a neural network approach. Journal of Sound andVibration 2006;289(4/5):71125.

Khandelwal M, Singh TN. Prediction of blast-induced ground vibration using arti-cial neural network. International Journal of Rock Mechanics and MiningSciences 2009;46(7):121422.

Langefors U, Kihlstrom B. The modern technique of rock blasting. New York: Wiley;1963.

Leondes CT. Neural network systems techniques and applications. Salt Lake City,USA: Academic Press; 1998. p. 87.

Lin BR, Hoft RG. Neural networks and fuzzy logic in power electronics. ControlEngineering Practice 1994;2(1):11321.

MathWorks Inc. Neural networks toolbox for use with MATLAB, Users Guide Version7.8. MA, USA: The MathWorks Inc; 2009.

Mohamadnejad M, Gholami R, Ataei M. Comparison of intelligence science tech-niques and empirical methods for prediction of blasting vibrations. Tunnellingand Underground Space Technology 2012;28:23844.

Mohammad MT. Articial neural network for prediction and control of blastingvibration in Assiut (Egypt) limestone quarry. International Journal of RockMechanics and Mining Sciences 2009;46(2):42631.

Monjezi M, Ghafurikalajahi M, Bahrami A. Prediction of blast-induced groundvibration using articial neural networks. Tunnelling and Underground SpaceTechnology 2011;26:4650.

Monjezi M, Hasanipanah M, Khandelwal M. Evaluation and prediction ofblast-induced ground vibration at Shur River Dam, Iran, by arti-cial neural network. Neural Computing and Applications 2013;22(7/8):163743.

Monjezi M, Singh TN, Khandelwal M, Sinha S, Singh V, Hosseini I. Prediction andanalysis of blast parameters using articial neural network. Noise & VibrationWorldwide 2006;37(5):816.

Raai H, Moosavi M. An approximate ANN-based solution for convergence of linedcircular tunnels in elasto-plastic rock masses with anisotropic stresses. Tun-nelling and Underground Space Technology 2012;27(1):529.

Rumelhart DE, Hinton GE, Williams RJ. Learning internal representations by errorpropagation. In: Rumelhart DE, McClelland JL, editors. Parallel distributed pro-cessing: explorations in the microstructure of cognition, foundation. Cambridge:MIT Press; 1986. p. 31862.

Scheaffer R, Mulekar M, McClav J. Probability and statistics for engineers. 5th ed.Boston, USA: Brooks/Cole; 2011. p. 599.

Sethi IK, Jain AK. Articial neural network and statistical pattern recognition. USA:Elsevier; 1991. p. 77.

Sivanandam S, Sumathi S, Deepa SN. Introduction to neural network using Matlab6.0. New Delhi: Tata McGraw-Hill; 2006. p. 533.

Sumathi S, Paneerselvam S. Computational intelligence paradigms: theory &applications using MATLAB. 1st ed. Boca Raton, FL, USA: CRC Press; 2010.p. 856.

An ANN-based approach to predict blast-induced ground vibration of Gol-E-Gohar iron ore mine, Iran1 Introduction2 Site description and data set3 Empirical methods for predicting PPV4 Multiple linear regression analysis5 Overview of artificial neural network5.1 Multi-layer perceptron network5.2 Network training5.3 Monitoring the validation and performance of ANN

6 Development of the ANN-based prediction6.1 Preparation of data for ANN models6.2 Network architecture and validation of ANN-based prediction

7 Results and discussion8 ConclusionsConflict of interestReferences

1-s2.0-s1674775513001157-main.pdf

Documents