Noise PRediction

© All Rights Reserved

1 views

Noise PRediction

© All Rights Reserved

- PolSARpro_Software_EPottier.pdf
- Spam Svm Report Naveenb
- The Effects of Training Satisfaction Employee Be 2018 International Journal
- 100620
- INSIDER DATA THEFT PREVENTION USING BEHAVIOR PROFILING (
- Volume 10.1007978-1-4471-4884-5 Issue Chapter 1 2013 [Doi 10.1007_978!1!4471-4884-5_1] Bramer, Max -- [Undergraduate Topics in Computer Science] Principles of Data Mining Introduction to Data Mining
- JBiSE20090300008_39459927
- SVM KNN Classifier
- Genetic Algorithm with SRM SVM Classifier for Face Verification
- 00001870
- 10.1.1.114
- Abstract
- Sustainable Natural Resources Management
- SVM Lecture 2
- Huang Nakamori Wang 2005
- termal conditions.pdf
- 1-s2.0-S0039914014003038-main
- SVM Overview
- Forecasting Gas Production Rate.pdf
- CONTENT BASED IMAGE RETRIEVAL USING SVM ALGORITHM AND CONTOURLET TRANSFORM COEFFICIENTS DISTRIBUTION.

You are on page 1of 17

DOI 10.1007/s00366-015-0408-z

ORIGINAL ARTICLE

resulting from blasting

Danial Jahed Armaghani1 · Mahdi Hasanipanah2 · Edy Tonnizam Mohamad3

© Springer-Verlag London 2015

Abstract Blasting operations usually produce significant including coefficient of determination (R2), root mean

environmental problems which may cause severe damage square error and value account for were utilized to check

to the nearby areas. Air-overpressure (AOp) is one of the the performance capacity of the predictive methods. Con-

most important environmental impacts of blasting opera- sidering these performance indices and using simple rank-

tions which needs to be predicted and subsequently con- ing method, the best models were selected among all con-

trolled to minimize the potential risk of damage. This paper structed models. It was found that the ICA-ANN approach

presents three non-linear methods, namely empirical, arti- can provide higher performance capacity in predicting AOp

ficial neural network (ANN), and imperialist competitive compared to other predictive methods.

algorithm (ICA)-ANN to predict AOp induced by blasting

operations in Shur river dam, Iran. ICA as a global search Keywords Blasting · Air-overpressure · Artificial neural

population-based algorithm can be used to optimize the network · Imperialist competitive algorithm

weights and biases of the network connection for training

by ANN. In this study, 70 blasting operations were inves-

tigated and relevant blasting parameters were measured. 1 Introduction

The most influential parameters on AOp, namely maximum

charge per delay and the distance from the blast-face, were Blasting refers to the controlled use of explosives for the

considered as input parameters or predictors. Using the five purpose of breaking down, excavation, or removal of the

randomly selected datasets and considering the modeling rock mass. This is the most commonly used technique for

procedure of each method, 15 models were constructed rock fragmentation in civil and mining engineering appli-

for all predictive techniques. Several performance indices cations, e.g., quarry operations, road and dam construc-

tions, etc. However, blasting has a number of negative

side effects on the environment, such as air-overpressure

* Danial Jahed Armaghani (AOp), ground vibration, flyrock, dust and fumes, and so

danialarmaghani@yahoo.com on [1–5], especially if blasting operations are carried out

Mahdi Hasanipanah near to residential buildings, factories, etc. or they are not

hasanipanah.m@grad.kashanu.ac.ir designed appropriately [6, 7]. AOp resulting from blasting

Edy Tonnizam Mohamad is an undesirable side-effect of the use of explosives. AOp

edy@utm.my affects structures and can produce damage when quarrying,

1 which may result in conflict between the quarry manage-

Young Researchers and Elite Club, Qaemshahr Branch,

Islamic Azad University, Qaemshahr, Iran ment and those who are affected [8–10].

2 Several empirical equations have been developed to

Young Researchers and Elite Club, Qom Branch, Islamic

Azad University, Qom, Iran predict AOp induced by blasting operation [11–14]. As

3 a result, these methods are not accurate enough, whereas

Department of Geotechnics and Transportation, Faculty

of Civil Engineering, Universiti Teknologi Malaysia, 81310 prediction of AOp values with high degree of accuracy

UTM Skudai, Johor, Malaysia is essential to estimate the blasting safety area [14]. In

13

Engineering with Computers

addition, these equations need to be updated when new Generally, during blasting operations, AOp waves are

blasting parameters are available. Aside from empirical created from four main sources [11, 30]:

equations, the use of artificial intelligence (AI) techniques

such as artificial neural network (ANN) as a quick solution • Air pressure pulse (APP): the rock displacement at

for engineering problems has received attention recently in bench face as the blast progresses.

field of geotechnical engineering [15–19]. However, ANNs • Gas release pulse (GRP): generated by the gas escape

have a number of disadvantages: slow learning rate and through rock fractures.

getting trapped in local minima [20, 21]. The ANN perfor- • Rock pressure pulse (RPP): caused by vibration of

mance can be considerably improved by the use of optimi- ground.

zation algorithms such as imperialist competitive algorithm • Stemming release pulse (SRP): produced by the gas

(ICA) as reported by many scholars [5, 22, 23]. This study escape from blast-hole once the stemming is ejected.

is aimed to investigate the efficiency of the empirical and

intelligent systems, namely ANN and ICA-ANN in predict- Many parameters influence the AOp like maximum

ing AOp resulting from the blasting operation. charge per delay, burden and spacing, detonator accuracy,

stemming, charge depth, and direction of initiation, weak

strata, overcharging, atmospheric conditions, and condi-

2 Air‑overpressure tions arising from secondary blasting [11, 26, 31, 32].

Nevertheless, as mentioned by many authors [2, 11, 13]

The explosion occurs by the shock wave of chemical reac- the most influential parameters on AOp are maximum

tion when the reactive gases pressure reaches the sonic charge per delay and distance from the blast-face. Blasting-

velocity [24]. The gas pressure velocity quickly rises as induced AOp cannot be easily predicted since, in different

the explosive detonation occurs within the blast-hole. cases, the same blast design may generate different results.

Suddenly, surrounding rocks are loaded by the blast-hole

pressure, which produces a compressive holes pulse and

moves away quickly from the blast-hole. Mainly, the pres- 3 AOp prediction methods

sure in blasting is indicated by shock and gas mechanisms

[25, 26]. A large shock wave from explosive into the sur- Several empirical models have been proposed to predict

face produces AOp. Therefore, the AOp is a shock wave AOp using its influential parameters. The use of the cube-

which is refracted horizontally by density variations into root scaled distance (SD) factor is a common technique to

the atmosphere. AOp has two atmospheric pressure waves: predict AOp. The relationship between the SD and the two

an audible high frequency sound and sub-audible low fre- parameters, namely explosive charge weight per delay and

quency [26]. The lower boundary of detectable sound for distance from the blast-face, is formulated as follows:

human ear is 20 Hz and below than that it is unhearable.

However, sound of more than 20 Hz frequency may cause a SD = DW −0.33 (1)

concussion to the human ear [13]. According to Kuzu et al. where D denotes the distance (m or ft) and W is the explo-

[13], in terms of sound, AOp is measured based on Pascal sive charge weight (kg or lb), and SD is the scaled distance

(Pa) or Decibels (dB). factor (m kg−0.33 or ft lb−0.33). A site-specific AOp attenu-

In cases where the energy of AOp waves goes above the ation formula can be developed when statistical analy-

atmospheric pressure (194.1 dB), the surrounding struc- sis techniques (i.e., least squares regression analysis) are

tures may be, to some extent, damaged [27]. Explosion applied to the representative AOp data [33, 34]. The gener-

causes the average level and higher spectral frequencies alized predictor equation for the prediction of AOp is given

in AOp to be higher, whereas the AOp amplitude reduces as follows [11]:

by 6 dB for every doubling of distance from recipient to

AOp = H(SD)−β (2)

blast [28]. The attenuation range becomes smaller, −3.1 to

−10 dB, depending on the differences between the source in which AOp is measured in dB, and H and β are the site

spectra and propagation conditions. The AOp degree of factors. The site factor values, H and β, for some blasting

damage possibility for structures is 180 dB, general win- conditions are tabulated in Table 1.

dow breakage is 171 dB, and occasional window breakage Using both rock material and free air properties, some

is 151 dB [13]. According to Siskind et al. [11], as reported numerical models are introduced and linked to Autodyn

by the United States Bureau of Mines (USBM), a value 2D in the study conducted by Wu and Hao [35] to make

of 134 dB is recommended for AOp limitation. Therefore, a simulation of AOp and ground shock induced by sur-

many attempts have been made to control AOp values [13, face explosions. An empirical method was developed by

29]. Rodríguez et al. [31] to predict AOp generated by blasting

13

Engineering with Computers

blasting conditions

Siskind et al. [11] Quarry blasts, behind face 622 0.515

Quarry blasts, direction of initiation 19,010 1.12

Quarry blasts, front of face 22,182 0.966

Hopler [9] Confined blasts for AOp suppression 1906 1.1

Blasts with average burial of the charge 19,062 1.1

Hustrulid [12] Detonations in air 185,000 1.2

Kuzu et al. [13] Quarry blasts in competent rocks 261.54 0.706

Quarry blasts in weak rocks 1833.8 0.981

Overburden removal 21,014 1.404

Hajihassani et al. [14] Quarry blasts, front of face (distance of 300 m) 10,909 1.09

Quarry blasts, front of face (distance of 600 m) 959.48 0.45

works outside a tunnel. The performance of their method the AOp values predicted by SVM are much closer to the

was evaluated with several tests and it was shown capable actual values compared to empirical methods. Hajihassani

of being used under various conditions. Kuzu et al. [13] et al. [14] proposed an approach based on hybrid particle

introduced an empirical equation for predicting AOp. They swarm optimization (PSO) and ANN model for the predic-

considered two critical parameters of AOp, i.e., the weight tion of AOp induced by quarry blasting. For the evaluation

of explosive materials and the distance from the blast-face of the accuracy of the proposed PSO-based ANN model,

to monitoring point in their model. They analyzed 98 AOp the AOp values measured by this model were compared

readings recorded from the quarry blasting operations in to those obtained by the empirical formula and the results

different conditions and concluded that their proposed showed the applicability of the proposed PSO-based ANN

equation could predict AOp with a reasonable accuracy. approach to predict AOp. In order to predict AOp using two

Using the data recorded from two quarries, Segarra et al. intelligent systems, namely adaptive neuro-fuzzy infer-

[36] proposed a new AOp predictive equation. The AOp ence system (ANFIS) and ANN, 166 blasting operations

measurements and blasting data were extracted from a total were investigated by Jahed Armaghani et al. [39] and the

of 122 records of 40 rock blasting operations with low to most effective parameters on AOp were measured. At the

very low strength. After all, they achieved an empirical end of this study, they mentioned that when higher accu-

equation with 32 % accuracy. racy of AOp prediction is needed ANFIS model would be

In addition to the above-mentioned empirical methods, the proper alternative compared to ANN technique. Table 2

numerous studies have been developed to predict AOp shows some studies with their performances in predicting

using soft computing techniques. Khandelwal and Singh AOp using soft computing techniques.

[37] proposed an ANN model using distance from the blast-

face to monitoring point for the prediction of AOp. They

compared the performance of ANN with that of the multi- 4 Artificial neural network

variate regression analysis (MVRA) and the United States

Bureau of Mines (USBM) predictor and proved the supe- Artificial neural network (ANN), which is an information

riority of their proposed model over MVRA and USBM processing system, makes a simulation of the functions

in terms of estimation accuracy. Considering the distance and structure of human brain. This network is a highly

from the blast-face to the monitoring point and the maxi- interconnected structure consisting of a large number of

mum charge per delay, Mohamed [38] applied ANN and a simple processing elements (i.e., neurons). This network

fuzzy inference system (FIS) to the prediction of AOp. He is capable of recognizing similarities when they are pre-

made a comparison between the results obtained from the sented with new input parameters after appropriately pre-

predictive models and the values of regression analyses and dicting the output pattern. Neural network is applicable as

observed field data. The ANN and FIS models were shown an alternative for some statistical analysis techniques such

more accurate compared to the regression analysis. Using a as multivariable regression, auto correlation, trigonomet-

total of 75 datasets related to three mines in India and the ric, linear regression, and so on. Using three fundamental

technique of support vector machine (SVM), Khandelwal components, a network can be defined; the components are

and Kankar [2] attempted to predict AOp. The AOp values network architecture, transfer function, and learning law

obtained from the SVM technique were compared to those [42, 43]. These components are defined based on the given

of a generalized predictor equation. They concluded that problem(s).

13

Engineering with Computers

References Technique Input No. of dataset R2

Mohamed [38] ANN, FIS DI, C 162 2

RANN = 0.92

2 = 0.86

RFIS

Khandelwal and Kankar [2] SVM DI, C 75 R2 = 0.85

Tonnizam Mohamad et al. [40] ANN HD, S, B, N, D, ST, PF 38 R2 = 0.93

Hajihassani et al. [14] ANN-PSO HD, S, B, ST, PF, N, DI, C, RQD 62 R2 = 0.86

Hajihassani et al. [41] ANN-PSO BS, MC, HD, ST, SB, RQD, PF, NH, DI 88 R2 = 0.89

Jahed Armaghani et al. [39] ANN DI, MC 166 R2 = 0.83

S spacing, B burden, ST stemming, PF powder factor, SVM support vector machine, C maximum charge per delay, D hole diameter; HD hole

depth, N number of row, PSO particle swarm optimization, DI distance from the blast-face, RQD rock quality designation, SB subdrilling, NH

number of hole

Several algorithms have been designed to train the neu- some of the best countries are called imperialists while

ral networks, among which the back-propagation (BP) the remaining countries form the colonies of the selected

algorithm is known as the most robust technique [44–46]. imperialists. ICA comprises three main operators, namely

BP can solve predictive complex engineering problems; assimilation, revolution, and competition. During the

it makes BP so popular among all algorithms for training assimilation, colonies are attracted towards the imperialists;

ANN [47, 48]. Generally, the feed-forward back-propa- this is because of their enthusiasm for power, culture, and

gation neural network (BPNN) contains at least three lay- economy, which are possessed by imperialists. Through

ers, namely input, hidden, and output layers. In each layer, the revolution, some sudden changes occurred in the coun-

there are some elementary processing units (i.e., neurons) tries’ positions. Colonies, during assimilation and revolu-

each of which is linked with the next layer through weights. tion, have the potential to reach a better state compared to

Depending on the given problem, the number of hidden that of their corresponding imperialist and take control of

layers and their neurons (nodes) change. The number of the empire [50]. In competition, the imperialists strive for

input and output neurons is similar to variables of input and more colonies and the whole empires attempt to possess

output. For the purpose of differentiating between differ- the colonies of other empires. Depending on their power,

ent processing units, the values that are named biases are all imperialists have the chance to possess a minimum of

presented in transfer functions. The net output of a neuron one colony of the weakest empire. Therefore, during the

or layer to its real or actual output is planned by transfer competition operator, weak empires collapse gradually and

functions [45]. In training the network, the data process- more powerful empires enhance their power. This proce-

ing from the input layer to the next (hidden layer) contin- dure goes on until only the most powerful empire remains

ues until date reaches the last (output) layer (forward pass). in power and the others collapse.

Within this layer, a comparison is done between the output As explained before, ICA starts with producing a set

and the actual values. Then, the difference between output of random candidate solutions each of which is known as

and actual values is propagated back through the network a country. In an Nvar dimension optimization problem, a

(backward pass), which updates the biases of the individ- country is represented by a 1 × Nvar array (see Eq. 3).

ual neurons and the individual weights of the connections.

The process continues until the network error is converged

Country = P1 , P2 , P3 , . . . , PNvar , (3)

to a threshold that is defined by a corresponding function, where pi signifies various factors such as culture, economy,

which is generally the root mean squared error (RMSE). religion, and language, which are important in character-

izing a country. Consequently, the cost of each country

should be obtained through the evaluation of the function f

5 Imperialist competitive algorithm in variables P1 , P2 , P3 , . . . , PNvar ; therefore,

Imperialist competitive algorithm (ICA) is a global search

Costi = f countryi = f P1 , P2 , P3 , . . . , PNvar (4)

population-based algorithm that was developed by Atash- Subsequent to the generation of N countries (Ncountry), a

paz-Gargari and Lucas [49] to be used in optimization definite number of countries that have the lowest costs

problems. ICA gets started with a random initial popula- are chosen as the imperialists, i.e., Nimp and other coun-

tion or candidate solutions which call countries. However, tries (colonies) are represented by Ncol. The colonies are

13

Engineering with Computers

Fig. 1 Schematic of forming

initial imperialists and colonies

[49]

powers. For this reason, knowing all empires’ costs, the

normalized cost, is obtained as follows:

Cn = max{ci } − cn , (5)

i

maximum cost among imperialists, and Cn is the normal-

ized cost of the nth imperialist. The relatively low normal-

ized cost of each imperialist indicates its relative weakness.

Using the normalized cost of each empire, pn, the param-

eter of imperialist normalized power can be obtained by

Eq. 6. Based on this parameter, the colonies are distributed Fig. 2 Assimilation procedure in ICA [49]

among empires.

powerful imperialists (i.e., those with bigger stars) take the

possession of more colonies.

C

n

pn = N (6)

As mentioned earlier, during the assimilation operator,

imp

C

i=1 i

colonies are attracted to their corresponding imperialists.

However, the power of an imperialist is dependent on the Figure 2 shows a schematic form of this procedure. As dis-

number of its colonies. Thus, for each imperialist, the ini- played in this figure, distance between an imperialist and

tial number of colonies, i.e., the Nth imperialist, is calcu- its colony is represented by d and the colony moves toward

lated using Eq. 7. its imperialist by x units, where x signifies a random num-

ber with uniform distribution (see Eq. 8).

N.C.n = round{pn .(Ncol )}, (7)

where N.C.n stands for the number of colonies for the nth

x ∼ U(0, β × d), (8)

imperialist. As shown in this equation, this is equal to the where β denotes a number close to 2 (many research-

nearest integer number that is obtained through the product ers have used the value of 2 for β [49]. In addition, Fig. 2

of the normalized cost of the nth imperialist and its colo- shows the deviation in the movement of a colony toward its

nies. The colonies are distributed amongst different empires corresponding imperialist (for enhancing the search prop-

using a random procedure and according to the empires’ erties of the algorithm). This deviation is modeled through

N.C. number. Subsequent to ascertaining the initial num- defining the parameter, θ parallel to the previous parameter

ber of imperialists and their colonies, the ICAs’ evolution- (i.e., x), θ is a random parameter with uniform distribution

ary loop gets started until the stop criterion is met. Figure 1 (see Eq. 9).

demonstrates a schematic view of the formation of initial

empires and their colonies. As shown in this figure, more θ ∼ U(−γ , γ ), (9)

13

Engineering with Computers

Fig. 3 Competition among

empires to attract the colony of

weakest empire [49]

where γ denotes a parameter whose values are determinant N.T.C.n = max{T.C.i } − T.C.n (11)

i

of the amount of deviation (search area) in the movement

of a colony. However, choosing θ equal to π/4 often leads

N.T.C.

to satisfactory results [49].

n

Ppn = N (12)

Nevertheless, as demonstrated in Eq. 10, in ICA, an imp

i=1 N.T.C.i

empire’s total power, i.e., the nth empire, is defined as the

summation of the nth empire’s power and a percentage of Based on the probability of attracting the targeted col-

the colonies’ power: onies for each empire, the above-mentioned colonies are

T.C.n = Cost imperialistn

allocated to the other empires. Then, the targeted colonies

are allocated to the empire with the highest relevant index

+ ξ mean Cost colonies of empiren , (10) in vector D. This competition procedure continues until

where T.C.n signifies the total cost of the nth empire and the weakest empire collapse (i.e., it loses all its colonies).

ξ represents a small number close to zero. As recommended Weak empires collapse one after another until only one

by Atashpaz-Gargari and Lucas [49], the value should be in empire remains, which possesses all countries as its colo-

the range of 0.002–0.1 for ξ. As implied by small values of nies. It is also noticeable that the algorithm can have other

ξ, the total cost of the nth empire roughly equals the total stopping criteria, for example, achieving a desirable value

cost of its imperialist. for the objective functions. Figure 4 presents flowchart of

Nevertheless, as previously stated, the gradual loss of the ICA algorithm.

the empire’s power results in the collapse of the empire. ICA has been widely utilized as optimization algorithm

Assuming an empire, i.e., the weakest empire, is about to optimize the complex problems. Kaveh and Talatahari

to collapse, there would be a competition among other [51] applied ICA to optimize the design of skeletal struc-

empires to attract the colonies of the weakest empire. tures. In the study conducted by Nazari-Shirkouhi et al.

Figure 3 shows this procedure. In this figure, the weak- [52], ICA was used to solve the integrated product mix-out-

est empire is represented as Empire 1, and Empires 2 to sourcing optimization problem. Application of ICA algo-

N strive to attract the weakest colony of Empire 1. For rithm in automated clustering of remote sensing images

the simulation of such a competition, first, the probability was presented by Karami and Shokouhi [53]. In addi-

of attracting the target colony in respect of the empire’s tion, this algorithm was applied to determine the optimum

power is determined through calculating the normalized weights and biases used in ANNs [22, 23, 54]. To predict

total power of the nth empire N.T.C.n (see Eq. 11). After- soil compaction indices, Taghavifar et al. [22] employed

wards, the probability of attracting a target colony for each conventional ANN and ICA–ANN techniques. They suc-

of the empires (Empires 2 to N) can be calculated through cessfully indicated that the network optimized by ICA can

Eq. 12. perform better compared to conventional ANN technique.

13

Engineering with Computers

man, Iran, in the adjacency of Sarchemeh copper mine and

Rafsanjan city, and at latitude of 30°1′48″ and longitude of

55°51′47″ (see Fig. 5). Also, this dam is the tallest asphaltic

concrete core dam in Iran. The type of bed rock in the dam

area is andesite and tuff. The site is situated in the area with

high seismicity where maximum design earthquake (MDE)

is 0.8 g [56]. The dam structure is separated to two parts

including main and saddle dam. The height of these dams

is 85.5 and 36.5 from the foundation for main and saddle,

respectively. Also, the crest length of the dam is 450 m.

Features of Shur river dam are listed in Table 3.

There are two mines in the vicinity of the Shur River

dam, namely main and second mines. In these mines, blast-

ing operations were used to produce fragmented rocks. In

these operations, ammonium nitrate fuel oil (ANFO) and

emulsion explosive were used as main explosive and prim-

ing, respectively. Since the mentioned mines are very close

to dam structure, AOp is one of the most important environ-

mental impacts of blasting operations in this region. There-

fore, it is essential to estimate AOp with the high degree

of accuracy. According to Tables 1 and 2, the most-used

input parameters in predicting AOp are maximum charge

per delay and distance from the blast-face. A total number

of 70 blasting operations were monitored at the dam site

and the values of AOp, maximum charge per delay and

distance from the blast-face were measured and used to

develop predictive models in this study. In this study, Mini-

mate Pulse produced by Instantel was utilized to measure

the AOp values. By using this equipment, AOp range from

88 dB (7.25 × 10−5 psi or 0.5 Pa) to 148 dB (0.0725 psi or

500 Pa) can be recorded. Properties of seismograph equip-

ment are shown in Table 4. In addition, summary of the

data used in the AOp modeling is listed in Table 5.

and intelligent techniques

Fig. 4 ICA flowchart [49]

namely empirical equation suggested by Siskind et al. [11],

ANN, and ICA-ANN, were applied in this study. In these

models, maximum charge per delay and distance from the

Furthermore, ICA was utilized to optimize ANN networks blast-face were considered as predictors or inputs. The fol-

in the studies carried out by Marto et al. [55] and Hajihas- lowing sections describe modeling procedure of the afore-

sani et al. [5] in predicting flyrock and ground vibration mentioned methods to predict AOp. Subsequently, to dem-

resulting from blasting, respectively. In both researches, onstrate the ability of these methods, the measured values

they showed the superiority of ICA-ANN technique in of AOp were compared with the corresponding predicted

comparison with the other used-methods. values of AOp.

13

Engineering with Computers

Property Description

Previously, several scholars suggested some empiri-

Type of dam Rockfill dam with asphalt concrete core cal equations for prediction of AOp using the suggested

Upstream side slope 1.75:1 (H:V) method of USBM (see Eq. 2). In this study, an attempt was

Downstream side slope 1.5:1 (H:V) made to develop an empirical equation for estimation AOp

Dam body volume 2,400,000 m3 using two predictors including maximum charge per delay

Asphalt concrete volume 16,000 m3 and distance from the blast-face. To do this, five differ-

Reservoir Volume 34,000,000 m3 ent datasets were selected randomly (to training and test-

Data of project From 2007 to 2012 ing datasets) from all 70 datasets for choosing the most

precise equation. A visual basic code was written to select

the random datasets through the randomizer function [57].

The idea behind using some data for testing is to check the

Table 4 Properties of minimate pulse performance capacity of the developed model. Swingler

Vibration monitoring Air-overpressure monitoring [58] suggested 20 % of whole datasets for testing purpose,

while this value was recommended as 25 % in the study

Range Up to 254 mm/s Linear range 88–148 dB

conducted by Looney [59]. Furthermore, Nelson and Ill-

Resolution 0.127 mm/s Linear resolution (500 Pa

ingworth [60] offered a range of 20 % to 30 % of whole

Peak)

datasets for testing purpose. Considering these recommen-

Accuracy ±5 % or 0.5 mm/sLinear accuracy 0.25 Pa

dations, 20 % (14 datasets) of whole datasets (70 datasets)

Frequency range2–250 Hz Linear frequency 2–250 Hz

response was selected randomly as testing datasets, whereas the

remaining 56 datasets were used for training the mod-

els. Using the selected datasets, five power equations (see

Eq. 2) were proposed to predict AOp as listed in Table 6.

In construction of these equations, results of maximum

Table 5 Summary of the data used in the modeling charge per delay and distance from the blast-face were used

Parameter Unit Symbol Category Min Max Mean to calculate the SD values. The statistical software package

of SPSS version 16 [61] was used to construct and analyze

Maximum charge kg C Input 135 1240 791.72 the empirical equations. As shown in Table 6, when con-

per delay

sidering only model development datasets, R2 of the pro-

Distance from the m DI Input 56 1541 366.88 posed empirical equations vary between 0.675 and 0.742.

blast-face

However, the R2 values are in the range of 0.703 and 0.899

Air-overpressure dB AOp Output 109.1 147.6 126.62

when testing datasets are taken into consideration. More

13

Engineering with Computers

Table 6 Proposed empirical equations for 5 randomly selected data- Table 7 Suggested equations for determining the number of neuron

sets in hidden layer [66]

2 AOp = 172.32 SD−0.092 0.742 3Ni Hush [62]

3 AOp = 169.32 SD−0.087 0.722 (Ni + N0)/2 Ripley [67]

4 AOp = 167.97 SD−0.085 0.687 2+N0 ×Ni +0.5N0 × N02 +Ni −3 Paola [68]

Ni +N0

5 AOp = 167.76 SD−0.083 0.675 2Ni/3 Wang [69]

√

Ni × N0 Masters [70]

2Ni Kaastra and Boyd [71]

information regarding selection of the best equation will be Kannellopoulas and Wilkinson [63]

given in results and discussion section. Ni number of input neuron, N0 number of output neuron

7.2 ANN modeling

technique to solve AOp problem. It should be noted that

In ANN modeling procedure of this study, the same data- only results of R2 were considered as performance criteria

sets performed in the analyses of empirical equation were to select the best model. Performance indices of all models

applied. As a first step of ANN modeling, all data should be with five hidden nodes for training and testing datasets are

normalized using the following equation: presented in Table 11. Levenberg–Marquardt (LM) learn-

ing algorithm was used in constructing ANN models. Study

Xnorm = (X − Xmin) / (Xmax − Xmin), (13) by Hagan and Menhaj [72] suggests the efficiency of this

where Xnorm represents the normalized value of the algorithm compared to other conventional gradient descent

measured parameter, Xmin is the minimum value of the techniques. More discussions regarding the selection of the

measured parameters, Xmax is the maximum value of the best ANN model to predict AOp will be given later.

measured parameters, and X is the measured value.

Performance of ANN models depends strongly on the 7.3 ICA‑ANN modeling

suggested architecture of the network [62, 63]. Therefore,

determination of the optimal architecture is required to As mentioned earlier, many optimization algorithms such

design an ANN model. The network architecture is defined as GA by Montana and Davis [73], PSO by Kennedy and

as the number of hidden layer(s) and the number of neurons Eberhart [74], artificial bee colony by Karaboga [75], and

in each hidden layer(s). As stated by Hecht-Nielsen [64] ant colony by Socha and Blum [76] have been proposed for

and Hornik et al. [65], any complex problem can be solved training ANN to obtain a set of weights and biases. These

using only one hidden layer. Hence, one hidden layer was algorithms can minimize the error values of the system as

chosen to build the ANN models in this study. In addition, stated by many researchers [22, 77–80]. In this objective,

determining neuron number(s) in the hidden layer is the ICA was applied for determining the optimum weights and

most critical task in the ANN architecture as stated by Son- biases of ANN model to minimize error values of the sys-

mez et al. [66]. Table 7 tabulates some equations related to tem. A hybrid of ICA-ANN model was employed to predict

the determination of number of neurons proposed by sev- AOp resulting from blasting operations at Shur river dam,

eral scholars. Based on this table and considering two neu- Iran. In order to have a perfect comparison between all pro-

rons in input layer (Ni) and one neuron in output layer (No), posed predictive models, the same randomly selected data-

the number of neurons which should be used in the hidden sets were utilized.

layer is in the range of 1–6.

In order to determine the optimum number of neurons 7.3.1 ICA parameter

in the hidden layer, using the same datasets of last part (5

randomly selected datasets), 30 ANN models were con- Before developing hybrid ICA-ANN model, some impor-

structed using one hidden layer and number of hidden tant parameters of ICA should be determined. As discussed

neurons of 1 to 6 as shown in Table 8. According to this earlier, the most effective factors on ICA algorithm include

table, considering average R2 values of both training and number of imperialism, number of country, number of dec-

testing datasets, model No. 5 with five hidden neurons out- ade, β, θ and ζ. Regarding β, θ, and ζ parameters, it seems

performs the other ANN models. Hence, five was selected that a review of the previous studies is needed. Atashpaz-

as number of hidden neurons in constructing ANN models. Gargari and Lucas [49] recommended β = 2 in the origi-

Therefore, structure of (2 × 5 × 1) was suggested for ANN nal study of ICA. Additionally, this value was successfully

13

Engineering with Computers

Table 8 R2 of trained ANN models to predict AOp using hidden neurons of 1–6

Model no. Nodes in hidden layers Network result

Iteration 1 Iteration 2 Iteration 3 Iteration 4 Iteration 5 Average

2 2 2 2 2

R R R R R R2

Train Test Train Test Train Test Train Test Train Test Train Test

1 1 0.806 0.823 0.819 0.733 0.802 0.796 0.797 0.815 0.811 0.801 0.807 0.794

2 2 0.819 0.787 0.799 0.844 0.794 0.836 0.821 0.785 0.813 0.826 0.809 0.816

3 3 0.887 0.728 0.840 0.861 0.814 0.918 0.843 0.863 0.845 0.633 0.846 0.801

4 4 0.879 0.669 0.876 0.794 0.865 0.810 0.847 0.803 0.799 0.890 0.853 0.793

5 5 0.862 0.835 0.879 0.882 0.879 0.827 0.859 0.910 0.858 0.885 0.867 0.868

6 6 0.884 0.780 0.877 0.679 0.846 0.785 0.870 0.880 0.840 0.802 0.863 0.785

Table 9 Effects of different Model no. Number of country Network result Ranking Total rank

numbers of countries in

predicting AOp Train Test Train Test

2 2 2

R RMSE R RMSE R RMSE R2 RMSE

2 50 0.760 0.101 0.745 0.100 4 8 7 4 23

3 75 0.776 0.091 0.767 0.098 5 12 8 5 30

4 100 0.735 0.101 0.842 0.096 3 8 11 6 28

5 150 0.797 0.097 0.770 0.075 11 9 9 10 39

6 200 0.785 0.096 0.683 0.112 8 10 5 2 25

7 250 0.797 0.097 0.664 0.094 11 9 2 7 29

8 300 0.780 0.096 0.738 0.107 7 10 6 3 26

9 350 0.789 0.095 0.673 0.112 9 11 3 2 25

10 400 0.760 0.103 0.840 0.069 4 7 10 11 32

11 450 0.804 0.095 0.907 0.048 12 11 12 12 47

12 500 0.795 0.097 0.677 0.092 10 9 4 8 31

applied by Hajihassani et al. [5], Taghavifar et al. [22], and [5], respectively, to predict flyrock and ground vibration

Ahmadi et al. [23]. Moreover, β > 1 was suggested in the induced by blasting. Therefore, in this study, 12 hybrid

study conducted by Niknam et al. [81]. Atashpaz-Gargari ICA-ANN models with various Ncountry were constructed

and Lucas [49] introduced θ = π/4 in the initial version of (see Table 9). In these models, to find the optimum Ncountry,

ICA and this value was successfully implemented by Marto number of decade (Ndecade) and number of imperialism

et al. [55] and Ebrahimi et al. [82] in different fields. In (Nimp) were set as 200 and 10, respectively. The obtained

the case of ζ, values of 0.1 by Atashpaz-Gargari and Lucas network results including RMSE and R2 for each model

[49], 0.05 by Taghavifar et al. [22], and 0.02 by Hajihassani were listed in Table 9. Furthermore, a simple ranking

et al. [5] have been suggested. Considering the above dis- method proposed by Zorlu et al. [57] was applied to obtain

cussion and using trial-and-error procedure method, values the optimum Ncountry. Note that, a model will be excel-

of 2, π/4 and 0.02 were selected for β, θ and ζ, respectively. lent if the R2 is one and RMSE is zero. Results of rank-

ing values for training and testing of each model as well

7.3.2 Number of country as their total rank values were calculated and presented

in Table 9. Based on total rank values, Ncountry = 450 can

In the previous researches, various Ncountry values were provide higher performance capacities of ICA-ANN model

utilized for solving the engineering problems. Value of compared to other number of countries. Therefore, value

40 was suggested by Ahmadi et al. [23] in predicting oil of 450 (model No. 11) was selected as Ncountry in modeling

flow rate of the reservoir. Ncountry = 56 and Ncountry = 135 procedure of ICA-ANN. For more information regarding

were introduced by Marto el al. [55] and Hajihassani et al. simple ranking method refer to Zorlu et al. [57]. It should

13

Engineering with Computers

the network performance of

ICA-ANN

be noted that, in the modeling procedure of the ICA-ANN training and testing datasets are shown in Table 10. Here,

technique, 80 % of whole datasets was assigned to train the a simple ranking procedure proposed by Zorlu et al. [57]

system, whereas, the remaining 20 % was considered as was used to select the best number of imperialists. Based

testing datasets. on total rank values presented in Table 10, Nimp = 20

(model No. 4) shows better performance of the network

7.3.3 Number of decade compared to other Nimp. Therefore, value of 20 was cho-

sen as optimum Nimp in modeling of ICA-ANN for pre-

Determination of Ndecade is the next step of ICA-ANN mod- diction of AOp.

eling procedure. A practical way to determine the optimum

Ndecade is to compare the results of the analysis of differ- 7.3.5 Network architecture

ent Ndecade. Therefore, another parametric study was per-

formed to identify the effect of Ndecade on the network’s In the last step of ICA-ANN modelling, using the sug-

performance. In these analyses, a fixed Ndecade = 1000 gested ANN structure (2 × 5 × 1), same randomly selected

was utilized. Such analyses were conducted to monitor datasets, and obtained ICA parameters from the previous

the changes in the cost function or RMSE for each dec- steps, five ICA-ANN models were trained. In addition,

ade. Results of parametric study on Ndecade for Ncountry of similar to two other predictive methods, the testing datasets

25 to 500 can be seen in Fig. 6. Based on this figure, sig- were also used in each train. The relevant results of training

nificant changes are observed in the first decades, whereas and testing datasets of ICA-ANN models as well as their

the changes were moderate up to number of decade equal discussion will be given in the following section.

to 800. After the mentioned Ndecade, there are no significant

changes in the RMSE results for all numbers of countries.

Hence, the optimum Ndecade was set to be 800 in ICA-ANN 8 Results and discussion

modeling. It is worth mentioning that in determining num-

ber of decades, the other obtained ICA parameters from In this research, three non-linear techniques, namely power

previous steps were utilized. empirical equation, ANN, and ICA-ANN, were developed

to predict AOp obtained from Shur river dam, Iran. During

7.3.4 Number of imperialists the modeling process of this study, all 70 datasets were ran-

domly selected to 5 different datasets (training and testing)

Knowing the optimum number of countries and decades, for development of the non-linear models. Some perfor-

to obtain the optimum number of imperialists, another mance indices including R2, value account for (VAF) and

sensitivity analyses are needed. For this purpose, Nimp RMSE were computed to check the capacity performance

values ranging from 5 to 65 were utilized to determine of all predictive models:

the best Nimp. In this step, the achieved parameters from N ′ 2

previous steps were utilized. Performance indices, i.e., 2 i=1 y − y

R = 1 − N

R2 and RMSE of different number of imperialists for i=1 (y − ỹ)

2 (14)

13

Engineering with Computers

Table 10 Effects of different Model no. Number of country Network result Ranking Total Rank

numbers of imperialists in

predicting AOp Train Test Train Test

2 2 2

R RMSE R RMSE R RMSE R2 RMSE

2 10 0.816 0.091 0.544 0.117 12 11 3 8 34

3 15 0.759 0.095 0.766 0.117 3 8 11 8 30

4 20 0.739 0.097 0.894 0.084 2 7 13 13 35

5 25 0.795 0.097 0.658 0.094 6 7 6 12 31

6 30 0.800 0.091 0.696 0.119 8 11 8 7 34

7 35 0.807 0.093 0.606 0.115 9 10 5 9 33

8 40 0.810 0.091 0.595 0.120 11 11 4 6 32

9 45 0.778 0.097 0.797 0.095 4 7 12 11 34

10 50 0.799 0.088 0.709 0.128 7 12 9 4 32

11 55 0.808 0.094 0.365 0.111 10 9 1 10 30

12 60 0.799 0.097 0.659 0.094 7 7 7 12 33

13 65 0.779 0.097 0.736 0.095 5 7 10 11 33

VAF = [1 − ] × 100 diction of UCS. Additionally, this method was success-

var(y)

(15) fully implemented in the study conducted by Yagiz et al.

[83]. They applied this method to select the best ANN

N

and non-linear multiple regression models for predic-

1

RMSE = (y − y′ )2 , (16) tion of PR values of Queens tunnel, USA. A ranking

N

i=1

value was calculated and assigned for each training and

where y, y′ and ỹ are the measured, predicted, and testing dataset separately (see Table 11). It is worth not-

mean of the y values, respectively, N is the total number ing that the RMES results of ANN and ICA-ANN tech-

of data and P is the number of predictors. Theoretically, niques were obtained for normalized datasets, whereas

the model will be excellent if R2 is one, VAF is 100 these values were achieved in modeling of empirical

(%), and RMSE is zero. Results of models performance models using the original (not normalized) datasets.

indices (R2, RMSE and VAF) for all randomly selected This is due to limitation to make a power relationship

datasets based on training and testing are presented in for AOp prediction. Total ranking of training and test-

Table 11. High performances of the training datasets ing datasets for three predictive models is shown in

indicate that the learning processes of the predictive Table 12. According to this table, model No. 1 exhib-

models are successful if those of the testing datasets ited the best performance of AOp prediction for empiri-

reveal that the models’ generalization ability ia satisfac- cal method, while models No. 2 and 3 yielded the best

tory. Based on Table 11, since, the obtained results of results in predicting AOp for ANN and ICA-ANN tech-

models’ performance indices are very similar, select- niques, respectively. When considering both training

ing the best models is too difficult. Sometimes, R2 of and testing datasets, the prediction performances of the

the training dataset of a predictive model may be higher ICA -ANN models are higher than those of empirical

while that of the testing dataset for the same model may and ANN models. The selected power equation (model

be lower. For instance, according to the obtained results No. 1) is shown in the following equation:

of ANN model, R2 values of 0.879 and 0.859 were

achieved for train 3 and train 4, respectively, which AOp = 167.51 SD−0.083 (17)

indicate higher performance capacity of train 3 in train- The graphs of predicted AOp using empirical, ANN, and

ing stage while these values were obtained 0.827 and ICA-ANN techniques against the measured AOp for train-

0.910 for test 3 and test 4, respectively, which express ing and testing datasets are shown in Figs. 7, 8, 9, respec-

higher performance capacity of test 4 in testing stage. tively. As shown in these figures, the ICA-ANN model can

To overcome this difficulty, Zorlu et al. [57] introduced perform better in the prediction of AOp in comparison to

a simple ranking method to choose the best predictive other predictive models. Based on these figures, R2 equal

models. They used this procedure in selecting the best to 0.961 for testing dataset suggests the superiority of the

13

Engineering with Computers

Table 11 Performance indices of each model and their rank values for all predictive approaches

Method Model R2 RMSE VAF Rating for R2 Rating for RMSE Rating for VAF Rank value

Train 2 0.742 5.216 74.160 5 5 5 15

Train 3 0.722 5.501 72.225 4 2 4 10

Train 4 0.687 5.339 68.675 2 3 2 7

Train 5 0.675 5.339 67.510 1 3 1 5

Test 1 0.886 4.221 85.833 4 3 4 11

Test 2 0.703 4.841 61.482 1 1 1 3

Test 3 0.793 3.167 79.235 2 5 2 9

Test 4 0.867 4.191 85.390 3 4 3 10

Test 5 0.899 4.272 89.621 5 2 5 12

ANN Train 1 0.862 0.075 86.163 4 4 3 11

Train 2 0.879 0.075 87.850 5 4 4 13

Train 3 0.879 0.076 87.854 5 3 5 13

Train 4 0.859 0.075 85.906 3 4 2 9

Train 5 0.858 0.074 85.750 2 5 1 8

Test 1 0.835 0.093 82.116 2 1 2 5

Test 2 0.882 0.060 87.964 3 5 3 11

Test 3 0.827 0.082 65.485 1 3 1 5

Test 4 0.910 0.080 88.798 5 4 5 14

Test 5 0.885 0.084 88.423 4 2 4 10

ICA-ANN Train 1 0.971 0.037 97.075 4 4 5 13

Train 2 0.962 0.038 96.172 1 3 1 5

Train 3 0.972 0.033 97.040 5 5 4 14

Train 4 0.963 0.045 96.220 2 1 2 5

Train 5 0.964 0.041 96.437 3 2 3 7

Test 1 0.952 0.039 95.132 1 2 1 5

Test 2 0.968 0.038 96.794 4 3 4 11

Test 3 0.961 0.049 96.059 3 1 3 7

Test 4 0.969 0.025 96.825 5 5 5 15

Test 5 0.960 0.037 95.298 2 4 2 8

ICA-ANN technique in predicting AOp, while these val- Table 12 Results of total rank for all predictive techniques obtained

ues are 0.886 and 0.882 for empirical and ANN models, from five randomly selected datasets

respectively. It is worth mentioning that the performance Method Model Total rank

capacity of the ICA-ANN model is higher than the perfor-

Empirical 1 21

mance capacities of the other techniques implemented by

2 18

previous scholars (see Table 2).

3 19

To demonstrate capability of the developed models in

predicting AOp, several empirical models including Siskind 4 17

et al. [11], Hopler [9], and Kuzu et al. [13] from all pre- 5 17

sented models in Table 1 were selected to predict AOp val- ANN 1 16

ues. Predicted AOp values using Siskind et al. [11], Hop- 2 24

ler [9], and Kuzu et al. [13] models, respectively, against 3 18

the all 70 AOp values measured in the site, are shown in 4 23

Figs. 10, 11, 12. R2 values of 0.667, 0.655, and 0.689 for 5 18

Siskind et al. [11], Hopler [9], and Kuzu et al. [13] mod- ICA-ANN 1 18

els, respectively, indicate lower prediction capacities of the 2 16

empirical models, while these values are obtained as 0.728, 3 21

0.880, and 0.971 for power equation, ANN and ICA-ANN 4 20

models, respectively. 5 15

13

Engineering with Computers

Fig. 7 R2 of measured and predicted AOp values for training and Fig. 10 Predicted values of AOp using Siskind et al. [11] model

testing datasets using empirical model against monitored AOp

Fig. 8 R2 of measured and predicted AOp values for training and Fig. 11 Predicted values of AOp using Hopler [9] model against

testing datasets using ANN technique monitored AOp

9 Conclusions

blasting using empirical, ANN and ICA-ANN methods. This

was accomplished using the blasting data obtained from 70

blasting operations at Shur river dam, Iran. The most influen-

tial parameters on AOp, namely maximum charge per delay

and the distance from the blast-face, were considered as input

parameters or predictor, whereas the values of measured AOp

were set as the output parameter. Using the five randomly

selected datasets and considering the modeling procedure of

each model, 15 models were constructed for all predictive

techniques. Considering some model performance indices

Fig. 9 R2 of measured and predicted AOp values for training and

such as R2, RMSE and VAF and using simple ranking method

testing datasets using ICA-ANN technique proposed by Zorlu et al. [57], the best models were selected

13

Engineering with Computers

Society of Explosives Engineers

10. Sawmliana C, Roy PP, Singh RK, Singh TN (2007) Blast

induced air overpressure and its prediction using artificial neural

network. Mining Technology 116(2):41–48

11. Siskind DE, Stachura VJ, Stagg MS, Koop JW, 1980. In: Siskind

DE, editor. Structure response and damage produced by airblast

from surface mining. United States Bureau of Mines

12. Hustrulid WA, 1999. Blasting principles for open pit mining:

general design concepts. Balkema.

13. Kuzu C, Fisne A, Ercelebi SG (2009) Operational and geological

parameters in the assessing blast induced airblast-overpressure in

quarries. Appl Acoust 70:404–411

14. Hajihassani M, Jahed Armaghani D, Sohaei H, Mohamad ET,

Marto A (2014) Prediction of airblast-overpressure induced by

blasting using a hybrid artificial neural network and particle

swarm optimization. Appl Acoust 80:57–67

Fig. 12 Predicted values of AOp using Kuzu et al. [13] model against

15. Tonnizam Mohamad T, Jahed Armaghani DJ, Noorani SA, Saad R,

monitored AOp

Alavi Nezhad SV (2012) Prediction of flyrock in boulder blasting

using artificial neural network. Elect J Goetech Eng 17:2585–2595

16. Jahed Armaghani D, Tonnizam Mohamad E, Momeni E, Naraya-

among all created models. The results indicated that the ICA- nasamy MS, Mohd Amin MF (2014) An adaptive neuro-fuzzy

inference system for predicting unconfined compressive strength

ANN approach outperforms the other predictive methods. and Young’s modulus: a study on Main Range granite. Bull Eng

The R2 equal to 0.961 for testing dataset suggests the supe- Geol Environ. doi:10.1007/s10064-014-0687-4

riority of the ICA-ANN technique in predicting AOp, while 17. Monjezi M, Hasanipanah M, Khandelwal M (2013) Evaluation and

these values were obtained as 0.886 and 0.882 for empirical prediction of blast-induced ground vibration at Shur River Dam, Iran,

by artificial neural network. Neural Comput Appl 22:1637–1643

and ANN methods, respectively. Although all proposed mod- 18. Momeni E, Nazir R, Jahed Armaghani D, Maizir H (2014) Pre-

els in this study are applicable for prediction of AOp, they can diction of pile bearing capacity using a hybrid genetic algorithm-

be used depending on the conditions. When high accuracy based ANN. Measurement 57:122–131

of AOp prediction is required, the ICA-ANN model would 19. Trivedi R, Singh TN, Raina AK (2015) Prediction of blast-induced

flyrock in Indian limestone mines using neural networks. Journal

be the proper alternative as it can optimize the weights and of Rock Mechanics and Geotechnical Engineering 6(5):447–454

biases of the network connection to train by ANN. 20. Wang XG, Tang Z, Tamura H, Ishii M, Sun WD (2004) An

improved backpropagation algorithm to avoid the local minima

problem. Neurocomputing 56:455–460

21. Adhikari R, Agrawal RK (2011) Effectiveness of PSO based

References neural network for seasonal time series forecasting. Indian Inter-

national Conference on Artificial Intelligence (IICAI). Tumkur,

1. Shirani Faradonbeh R, Monjezi M, Jahed Armaghani D (2015) India, pp 232–244

Genetic programing and non-linear multiple regression tech- 22. Taghavifar H, Mardani A, Taghavifar L (2013) A hybridized arti-

niques to predict backbreak in blasting operation. Eng Comput. ficial neural network and imperialist competitive algorithm opti-

doi:10.1007/s00366-015-0404-3 mization approach for prediction of soil compaction in soil bin

2. Khandelwal M, Kankar PK (2011) Prediction of blast-induced facility. Measurement 46(8):2288–2299

air overpressure using support vector machine. Arabian J Geosci 23. Ahmadi MA, Ebadi M, Shokrollahi A, Majidi SMJ (2013)

4:427–433 Evolving artificial neural network and imperialist competitive

3. Khandelwal M, Kumar DL, Yellishetty M (2011) Application of algorithm for prediction oil flow rate of the reservoir. Applied

soft computing to predict blast-induced ground vibration. Eng Soft Computing 13(2):1085–1098

Comput 27(2):117–125 24. Baker WE, Cox PA, Kulesz JJ, Strehlow RA, Westine PS, 1983.

4. Jahed Armaghani D, Hajihassani M, Mohamad ET, Marto A, Explosion hazards and evaluation. Elsevier Science.

Noorani SA (2014) Blasting-induced flyrock and ground vibra- 25. Roy PP (2005) Rock blasting effects and operations. A.A.

tion prediction through an expert artificial neural network based Balkema, India

on particle swarm optimization. Arab J Geosci 7:5383–5396 26. Bhandari S (1997) Engineering rock blasting operations. A.A.

5. Hajihassani M, Jahed Armaghani D, Marto A, Tonnizam Mohamad E Balkema, Netherlands

(2014) Ground vibration prediction in quarry blasting through an 27. Glasstone S, Dolan PJ. The effects of nuclear weapons. Washington

artificial neural network optimized by imperialist competitive algo- (D.C.): US Department of Defense and Energy Research; 1977.

rithm. Eng. Geol. Environ, Bull. doi:10.1007/s10064-014-0657-x 28. Stachura VJ, Siskind DE, Kopp JW (1984) Airheast and ground

6. Kuzu C (2008) The importance of site-specific characters in vibration generation and propagation from contour mine blast-

prediction models for blast-induced ground vibrations. Soil ing. U.S. Dept. of the Interior, Bureau of Mines

Dynamic Earthq Eng 28:405–414 29. Rodríguez R, Lombardía C, Torno S (2010) Prediction of the air

7. Singh TN, Dontha LK, Bhardwaj V (2008) Study into blast wave due to blasting inside tunnels: approximation to a ‘phono-

vibration and frequency using ANFIS and MVRA. Mining Tech- metric curve’. Tunn Undergr Sp Technol 25:483–489

nology 117(3):116–121 30. Wiss JF, Linehan PW (1978) Control of vibration and blast noise

8. Konya CJ, Walter EJ (1990) Surface blast design. Prentice Hall, from surface coal mining. Wiss, Janney, Elstner and Associates

Englewood Cliffs Inc, Northbrook, IL (USA)

13

Engineering with Computers

31. Rodríguez R, Toraño J, Menéndez M (2007) Prediction of the 51. Kaveh A, Talatahari S (2010) Optimum design of skeletal struc-

airblast wave effects near a tunnel advanced by drilling and tures using imperialist competitive algorithm. Computers and

blasting. Tunn Undergr Sp Technol 22:241–251 Structures 88(21–22):1220–1229

32. Dowding CH. Construction vibrations. In: Dowding, editor;

52. Nazari-Shirkouhi S, Eivazy H, Ghodsi R, Rezaie K, Atashpaz-

2000. p. 204–207. Gargari E (2010) Solving the integrated product mix-outsourcing

33. Rosenthal MF, Morlock GL. Blasting guidance manual, office of problem using the imperialist competitive algorithm. Expert Sys-

surface mining reclamation and enforcement. US Department of tems with Applications 37(12):7615–7626

the Interior; 1987. 53. Karami S, Shokouhi SB (2012) Application of imperialist com-

34. Cengiz K (2008) The importance of site-specific characters in petitive algorithm for automated classification of remote sensing

prediction models for blast-induced ground vibrations. Soil Dyn images. International Journal of Computer Theory and Engineer-

Earthquake Eng 28:405–414 ing 4(2):137–143

35. Wu C, Hao H (2005) Modelling of simultaneous ground shock 54. Ahmadi MA (2011) Prediction of asphaltene precipitation using

and air blast pressure on nearby structures from surface explo- artificial neural network optimized by imperialist competitive

sions. Int J Impact Eng 31:699–717 algorithm. Journal of Petroleum Exploration and Production

36. Segarra P, Domingo JF, López LM, Sanchidrián JA, Ortega MF Technology 1(2–4):99–106

(2010) Prediction of near field overpressure from quarry blast- 55. Marto A, Hajihassani M, Jahed Armaghani D, Tonnizam Mohamad

ing. Appl Acoust 71:1169–1176 E, Makhtar AM (2014) A novel approach for blast-induced flyrock

37. Khandelwal M, Singh TN (2005) Prediction of blast induced prediction based on imperialist competitive algorithm and artificial

air overpressure in opencast mine. Noise Vib Control Worldw neural network. Sci. World. J, Article ID 643715

36:7–16 56. Report Design (2008) Shur River Dam. Australian Tailing Con-

38. Mohamed MT (2011) Performance of fuzzy logic and artificial sultants, ATC

neural network in prediction of ground and air vibrations. Int J 57. Zorlu K, Gokceoglu C, Ocakoglu F, Nefeslioglu HA, Acikalin S

Rock Mech Min Sci 48:845–851 (2008) Prediction of uniaxial compressive strength of sandstones

39. Jahed Armaghani D, Hajihassani M, Monjezi M, Mohamad ET, using petrography-based models. Eng. Geol. 96(3):141–158

Marto A, Moghaddam MR (2015) Application of two intelligent 58. Swingler K (1996) Applying Neural Networks: A Practical

systems in predicting environmental impacts of quarry blasting. Guide. Academic Press, New York

Arab J Geosci. doi:10.1007/s12517-015-1908-2 59. Looney CG (1996) Advances in feed-forward neural networks:

40. Tonnizam Mohamad E, Hajihassani M, Jahed Armaghani D,

demystifying knowledge acquiring black boxes. IEEE Transac-

Marto A. Simulation of blasting-induced air overpressure by tions on Knowledge and Data Engineering 8(2):211–226

means of artificial neural networks. Int Rev Modell Simulations 60. Nelson M, Illingworth WT (1990) A Practical Guide to Neural

2012;5:2501–6. Nets. Addison- Wesley, Reading MA

41. Hajihassani M, Jahed Armaghani D, Monjezi M, Mohamad

61. SPSS Inc. (2007) SPSS for Windows (Version 16.0). Chicago:

ET, Marto A (2015) Blast-induced air and ground vibra- SPSS Inc.

tion prediction: a particle swarm optimization-based artifi- 62. Hush DR (1989) Classification with neural networks: a perfor-

cial neural network approach. Environ Earth Sci. doi:10.1007/ mance analysis. Proceedings of the IEEE International Confer-

s12665-015-4274-1 ence on Systems Engineering. Dayton, OH, USA, pp 277–280

42. Simpson PK (1990) Artificial neural system—foundation, para- 63. Kanellopoulas I, Wilkinson GG (1997) Strategies and best prac-

digm, application and implementations. Pergamon Press, NewYork tice for neural network image classification. International Journal

43. Kosko B (1994) Neural networks and fuzzy systems: a dynamical of Remote Sensing 18:711–725

systems approach to machine intelligence. Prentice-Hall, New Delhi 64. Hecht-Nielsen R (1987) Kolmogorov’s mapping neural network

44. Khandelwal M, Singh TN (2009) Prediction of blast-induced existence theorem. Proceedings of the First IEEE International

ground vibration using artificial neural network. Interna- Conference on Neural Networks. San Diego, CA, USA, pp 11–14

tional Journal of Rock Mechanics and Mining Sciences 65. Hornik K, Stinchcombe M, White H (1989) Multilayer feedfor-

46(7):1214–1222 ward networks are universal Approximators. Neural Networks

45. Bahrami A, Monjezi M, Goshtasbi K, Ghazvinian A (2011) Pre- 2:359–366

diction of rock fragmentation due to blasting using artificial neu- 66. Sonmez H, Gokceoglu C, Nefeslioglu HA, Kayabasi A (2006)

ral network. Eng Comput 27(2):177–181 Estimation of rock modulus: for intact rocks with an artificial

46. Jahed Armaghani D, Momeni E, Alavi Nezhad Khalil Abad SV, neural network and for rock masses with a new empirical equa-

Khandelwal M (2015) Feasibility of ANFIS model for predic- tion. Int. J. Rock Mech. Min. Sci. 43:224–235

tion of ground vibrations resulting from quarry blasting. Environ 67. Ripley, B.D., 1993. Statistical aspects of neural networks. In:

Earth Sci. DOI 10.1007/s12665-015-4305-y Barndoff- Neilsen OE, Jensen JL, Kendall WS, editors. Net-

47. Maulenkamp F, Grima MA (1999) Application of neural net- works and chaos-statistical and probabilistic aspects. London:

works for the prediction of the unconfined compressive strength Chapman & Hall, pp. 40-123.

(UCS) from Equotip hardness. Int J Rock Mech Min Sci 68. Paola, J.D., 1994. Neural network classification of multispectral

36:29–39 imagery. MSc thesis, The University of Arizona, USA.

48. Momeni E, Armaghani DJ, Hajihassani M, Amin MFM (2015) 69. Wang, C., 1994. A theory of generalization in learning machines

Prediction of uniaxial compressive strength of rock samples with neural application. PhD thesis, The University of Pennsyl-

using hybrid particle swarm optimization-based artificial neural vania, USA.

networks. Measurement 60:50–63 70. Masters T (1994) Practical neural network recipes in C++. Aca-

49. Atashpaz-Gargari E, Lucas C 2007. Imperialist competitive algo- demic Press, Boston MA

rithm: an algorithm for optimization inspired by imperialistic 71. Kaastra I, Boyd M (1996) Designing a neural network for fore-

competition. In: IEEE Congr Evol Comput, pp 4661–4667 casting financial and economic time series. Neurocomputing

50. Atashpaz-Gargari E, Hashemzadeh F, Rajabioun R, Lucas C

10:215–236

(2008) Colonial competitive algorithm, a novel approach for 72. Hagan MT, Menhaj MB (1994) Training feed forward networks

PID controller design in MIMO distillation column process. Int J with the Marquardt algorithm. IEEE Transactions on Neural Net-

Intell Comput Cybern 1:337–355 works 5(6):861–867

13

Engineering with Computers

73. Montana DJ, Davis L (1989) Training feedforward neural net- 79. Monjezi M, Khoshalan HA, Varjani AY (2012) Prediction of

works using genetic algorithms. IJCAI 89:762–767 flyrock and backbreak in open pit blasting operation: a neuro-

74. Kennedy, J., and Eberhart, R. (1995). Particle Swarm Optimiza- genetic approach. Arabian Journal of Geosciences 5(3):441–448

tion. IEEE International Conference on Neural Networks. Perth, 80. Tonnizam Mohamad, E., Jahed Armaghani, D., Momeni, E., Alavi

Australia. 1942-1948. Nezhad Khalil Abad, S.V., 2014. Prediction of the unconfined

75. Karaboga, D. (2005). An idea based on honey bee swarm for compressive strength of soft rocks: a PSO-based ANN approach.

numerical optimization (Vol. 200). Technical report-tr06, Erci- Bull. Eng. Geol. Environ. doi:10.1007/s10064-014-0638-0

yes university, engineering faculty, computer engineering 81. Niknam T, Taherian Fard E, Pourjafarian N, Rousta A (2011) An

department. efficient hybrid algorithm based on modified imperialist com-

76. Socha K, Blum C (2007) An ant colony optimization algorithm petitive algorithm and K-means for data clustering. Engineering

for continuous optimization: application to feed-forward neural Applications of Artificial Intelligence 24(2):306–317

network training. Neural Comput Appl 16:235–247 82. Ebrahimi E, Mollazade K, Babaei S (2014) Toward an automatic

77. Karaboga, D., Akay, B., Ozturk, C. (2007). Artificial bee colony wheat purity measuring device: A machine vision-based neural

(ABC) optimization algorithm for training feed-forward neural networks-assisted imperialist competitive algorithm approach.

networks. In Modeling decisions for artificial intelligence (pp. Measurement 55:196–205

318-329). Springer Berlin Heidelberg. 83. Yagiz S, Gokceoglu C, Sezer E, Iplikci S (2009) Application of

78. Sivagaminathan RK, Ramakrishnan S (2007) A hybrid approach two non-linear prediction tools to the estimation of tunnel boring

for feature subset selection using neural networks and ant colony machine performance. Eng. Appl. Artif. Intel. 22(4):808–814

optimization. Expert systems with applications 33(1):49–60

13

- PolSARpro_Software_EPottier.pdfUploaded byaqzeeshan78
- Spam Svm Report NaveenbUploaded byAzizul Huq
- The Effects of Training Satisfaction Employee Be 2018 International JournalUploaded bySyahirah Senin
- 100620Uploaded byvol2no6
- INSIDER DATA THEFT PREVENTION USING BEHAVIOR PROFILING (Uploaded byIJIERT-International Journal of Innovations in Engineering Research and Technology
- Volume 10.1007978-1-4471-4884-5 Issue Chapter 1 2013 [Doi 10.1007_978!1!4471-4884-5_1] Bramer, Max -- [Undergraduate Topics in Computer Science] Principles of Data Mining Introduction to Data MiningUploaded byRegina Yunisa
- JBiSE20090300008_39459927Uploaded byNguyen Huyen
- SVM KNN ClassifierUploaded byMunaf Rashid
- 00001870Uploaded byhajarooooo
- Genetic Algorithm with SRM SVM Classifier for Face VerificationUploaded byAnonymous Gl4IRRjzN
- 10.1.1.114Uploaded byHossein Yadollahtabar
- AbstractUploaded byYasmine TA
- Sustainable Natural Resources ManagementUploaded byAlexandra Scarlet
- SVM Lecture 2Uploaded byHeenal Mehta
- Huang Nakamori Wang 2005Uploaded byjoanaguirre
- termal conditions.pdfUploaded byKhriiz Hg
- 1-s2.0-S0039914014003038-mainUploaded byAndreea Marin
- SVM OverviewUploaded bywebdaxter
- Forecasting Gas Production Rate.pdfUploaded byadeeyo
- CONTENT BASED IMAGE RETRIEVAL USING SVM ALGORITHM AND CONTOURLET TRANSFORM COEFFICIENTS DISTRIBUTION.Uploaded byIJAR Journal
- Survey on Feature Selection and Dimensionality Reduction TechniquesUploaded byAnonymous kw8Yrp0R5r
- A Machine Learning Introductory Tutorial with Examples _ Toptal.pdfUploaded byGuilherme Germano
- Movie Buzz: Forecasting Movie’s SuccessUploaded byEditor IJRITCC
- Dong Zhao; Deyi Xue -- Parametric Design With Neural Network Relationships and Fuzzy Relationships Considering UncertUploaded bybaconkho
- Yesidmauricioospinadavila.2009.PDFUploaded byAntonio Andia Mirano
- Soft Computing Healthcare Intelligence through Multisensory Activity RecognitionUploaded byEditor IJTSRD
- Machine Learning_part 1Uploaded bycjon
- CAT 3 NNFLC with ans.docxUploaded bySHAIK ROSHAN
- Phd Msc Positions Synchromedia 2019Uploaded byJunaid Akram
- DTSpaper110915.pdfUploaded byManohar Lal

- Necessary and Sufficient Observability Conditions for Bearings Only TmaUploaded byRishav Kanth
- A 105502Uploaded byRishav Kanth
- 02 Transient HarmonicUploaded byRishav Kanth
- Gen_Contact Localization and Motion Analysis in the OceanUploaded byRishav Kanth
- Active_PAssive_Sensor Data Fusion Using Kalman FilterUploaded byRishav Kanth
- Measures of Nonlinearity for EstimationUploaded byRishav Kanth
- PRediction of Airblast Overpressure Induced by Blasting Using a HybridUploaded byRishav Kanth
- Attenuation of Blast Waves when detonating explosives inside barriers.pdfUploaded byRishav Kanth
- Refed and Validated Noise Contour SystemUploaded byRishav Kanth
- An ANN-based approach to predict blast-induced ground vibration of Gol-E-Gohar iron ore mine, Iran.pdfUploaded byRishav Kanth
- Target Motion Analysis With Passive Data FusionUploaded byRishav Kanth
- sensors-13-01151Uploaded byThuyhoi Dinh
- Target Motion AnalysisUploaded byrodrigosilva1973
- Leg-By-leg Bearings-Only Target Motion Analysis Without Observer ManoueverUploaded byRishav Kanth
- Constrained Bearings OnlyUploaded byRishav Kanth
- 0ENTR08032Uploaded byRishav Kanth
- a119z6Ap12Uploaded byRishav Kanth
- Witte KindUploaded bySaeed Javdani
- Links to Linkages LS Dyna PapersUploaded byRishav Kanth
- Define LoadUploaded byRishav Kanth
- Courant–Friedrichs–Lewy condition paper.pdfUploaded byRishav Kanth
- Huang LstcUploaded byRishav Kanth
- 20ssgUploaded byRishav Kanth
- 17dscsdcUploaded byRishav Kanth
- 16aasasUploaded byRishav Kanth
- 15fdsdUploaded byRishav Kanth
- 14dsdsUploaded byRishav Kanth
- 11sxdUploaded byRishav Kanth

- 80010454Uploaded byselcuksl
- ATMega16 AVR Microcontroller LCD Digital ClockUploaded byAdrianMartinezMendez
- Nut PlatesUploaded byJoe Komsun
- Meteorology JeppesenUploaded byАртём Огородиков
- torsional constantUploaded bykushaljp
- Psychic Merits and FlawsUploaded byRaMinah
- EPSTEIN & MARTIN - Qualitative Aproach to Empirical Legal Research.pdfUploaded byDouglas Zaidan
- Preturi Ventiloconvectoare Perete BrizaUploaded bystarsplesh
- Smart Data Cloud Security AllianceUploaded byfewdisc
- U.S. Pat. 8,637,753, Inventor Dean Zelinsky, Z-Glide Neck, 2014.Uploaded byDuane Blake
- Arrays in Python.pptUploaded byAman Singh Raghuwanshi
- Altivar 312 Drive_ATV312HU40N4Uploaded byPaul Ramos Carcausto
- Application of image processing in seed technology: A surveyUploaded bycuterohit16
- LG Balancing Urban& Rural SegmentUploaded byRiyas Noorudeen
- Vacuum Breaker VB14 VB21-Technical InformationUploaded byCarlos
- Wiki.developerforce.com-Using Apex Managed Sharing to Create Custom Record Sharing Logic Developer Force ComUploaded bymastaflamasta
- Impact of RFID in RetailUploaded bybhar4tp
- Price Sound System Lt. 14Uploaded byCahyo
- ISO 26000 Mgt BriefingUploaded byroelens
- PHY-TEST-EM05(3-3)Uploaded byphysicspalanichamy
- gsb pdfUploaded bysachin s
- Nouveau Document TexteUploaded byAnonymous 8erOuK4i
- power-steering-system-flushing-general-procedures.pdfUploaded byMichael Hernandez
- genre analysisUploaded byapi-252218826
- Graphs of y=a sin b(x - c) + d and y=a cos b(x - c) + dUploaded byFrances Bless
- 00 - Technology Pioneers Rev Aug 2010Uploaded byAlexandre Uffler
- Uniaxial Deformation and Orientation OfUploaded byBabic Ugljesa
- Maxon Formelsammlung eUploaded byAmante Bandido
- Concentrator Lab (1)Uploaded byLiven WenChuan Liu
- Smart Quill2Uploaded bysuchu138522

## Much more than documents.

Discover everything Scribd has to offer, including books and audiobooks from major publishers.

Cancel anytime.