BOULDER PREDICTION IN ROCK BLASTING USING ... developed neural network model is suitable for...

VOL. 12, NO. 1, JANUARY 2017 ISSN 1819-6608

ARPN Journal of Engineering and Applied Sciences ©2006-2017 Asian Research Publishing Network (ARPN). All rights reserved.

www.arpnjournals.com

47

BOULDER PREDICTION IN ROCK BLASTING USING ARTIFICIAL

NEURAL NETWORK

P. Y. Dhekne, Manoj Pradhan, Ravi K. Jade and Romil Mishra

Faculty, Department of Mining Engineering, National Institute of Technology, Raipur, India E-Mail: [email protected]

ABSTRACT

This paper explains the development of an artificial neural network (ANN) model for estimating the number of boulders resulting from the blasts in Limestone quarries. A number of ANN models are available for prediction of post-blast rock fragmentation distribution but none is available to predict boulder count. A data base of three hundred blasts was created for the development of the model. The records in the data base were generated from the blasts carried out in four Limestone quarries which have similar geotechnical set up. These quarries adopt different blast practice. The data base consists of blast design parameters, explosive type and the boulder count. 191records out of three hundred were used to train a two-hidden layer back-propagation neural network model to predict boulder count resulting from the blasts. The boulder count was considered to be a function of nine independent parameters. The ANN program code was developed inMatLabR2012a. The network was trained using Levenberg-Marquardt algorithm as it provides the highest stability and maximum learning speed. The network was extensively analyzed to assess its performance with different transfer functions and the number of hidden layers to estimate the optimum network architecture. 77 data sets were used to validate the results of the trained neural network model and the remaining 32 datasets were used for testing the developed model. Predictions of boulder count by the ANN model were compared with those using a statistical model developed in SPSS statistics 20.0. It was observed that the prediction capability of the trained neural network model was found to be strong and it provides an easy option to the field engineers to optimize the blast design so that the boulder-count is the minimum. Diversity of the blast data in the similar geotechnical set up is one of the most important aspects of the developed model. The developed neural network model is suitable for practical use at the Limestone quarries having similar geotechnical set up. Keywords: artificial neural network, multiple regression, blasting, rock fragmentation, boulder count.

INTRODUCTION

Efficiency of the unit operations of quarrying viz.

loading, transporting and crushing is the maximum with optimum rock fragmentation. Maximum efficiency minimizes the cost of production. Therefore optimum rock fragmentation is always one of the objectives in any production blasting. Rock fragmentation is said to be optimum if the rock needs no further treatment after the blast. A lot of research has already been conducted on the various aspects of the fragmentation with the sole objective of improving the same [1-2, 4-7, 17, 21-22, 24-31, 34-35, 38, 43,45-49, 51, 61-62, 65-71, 74-75, 78]. It is reported that more than 20 factors affect the blast results [19]. These factors can be grouped in four different categories: rock geotechnical parameters such as density, hardness, compressibility; explosive parameters such as density, velocity of detonation; technical parameters such as delay interval, primer strength and location and geometrical parameters such as burden, spacing, and stemming [3]. A number of empirical models are available to estimate post-blast rock fragmentation resulting from a blast [9-11, 39, 42, 64]. Concept of ANN has been applied to model various aspects of blast-induced ground vibrations [32, 37, 55, 76], air overpressure [72], fly rock [54,57,79], back break [15, 20, 36, 56, 73], powder factor [32-33, 44], estimation of blast geometry [13, 52, 59,77], estimation of an appropriate type of explosive [12], estimation of fragmentation [3, 15-16, 40-41, 50, 53, 58, 63, 80-81] etc. A critical analysis of the models indicates that the modeling using ANN is beneficial in rock blasting

where a large number of affecting variables and their complicated mutual dependence sometimes create difficulties in the application of empirical modeling. Outputs of the developed ANN based fragmentation models have been expressed either as the fragment size or as the sieve analysis. Practicing mining engineers not only make use of this form of the output to evaluate the fragmentation but at the same time they are also interested in knowing the boulder count so that they can plan the secondary breakage operations. Extensive literature survey has indicated that no ANN model is available to predict the boulder count. Therefore an ANN model has been developed to predict the boulder count. The data sets required for the development of the model have been generated from the Limestone quarries.

DESCRIPTION OF THE SITES The ANN model described in this paper has been

developed from the blast records generated from Baikunth, Hirmi, Sonadih and Rawan Limestone quarries. The quarries are situated in Raipur district of Chhattisgarh province of India and are located within a radius of 20 km. The quarries have an overburden which is generally hard Laterite and Clay having a thickness from 4 m to 6 m. The deposits consist of horizontal, thick bedded Stromatolitic Limestone of Raipur Group. Patches of Dolomitic Limestone, Argillaceous Limestone and Shale are found associated with the deposits. The deposits are flat and their thickness ranges from 30 m to 35 m. The deposits are having three sets of joints. The joints are open and are

mailto:[email protected]




48

having an aperture of 2 mm to 5 mm. The joint surface is rough. The geotechnical properties of the deposits are

summarized in Table-1.

Table-1. Geotechnical properties of limestone deposits (*Mean values).

Name of

the

quarry

Uniaxial *

compressive

strength (M Pa)

Density*

(g/cc)

Young’s *

Modulus

(G Pa)

Porosity*

(%)

Vertical *

spacing

between joints (m)

Horizontal*

spacing

between joints (m)

Baikunth 38 2.25 46 6 1.0 0.6

Rawan 43 2.38 49 5 1.5 0.9

Sonadih 45 2.35 48 7 1.4 0.8

Hirmi 44 2.40 50 5 2.0 1.0

A perusal of the properties indicates that the

deposits of the four quarries have a similar geotechnical set-up and can be considered as competent. The quarries are under active production stage and the winning is done using conventional drilling and blasting method. The holes are drilled in two to three rows on staggered pattern.

Row to row delay with cord relays is maintained at 50 ms whereas the delay with shock tube initiation is 42 ms. Blasts are carried out either with Ammonium Nitrate- Fuel Oil mixture (ANFO) or with Site-Mixed Emulsion (SME) Explosive. The other details of blast practice are given in Table-2.

Table-2. Blast practice details.

Name of

the

quarry

Dia

(mm) Spacing

(m) Burden

(m) Depth

(m) Explosive

type Primer

Initiation

method

Secondary

breakage

method

Baikunth

115 5.0-6.0 3.0-4.0

8.0

ANFO Cartridge

booster/Cast booster

Cord relays

Secondary blasting/

Rock breaker 152 6.5-7.0 4.5-5.0 SME Cast booster

Shock tubes

Rawan 152 7.0 4.0 8.0 SME Cast booster Shock tubes

Rock breaker

Sonadih 100 4.0 3.0 9.0 SME Cast booster Shock tubes

Rock breaker

Hirmi

115 4.0 6.0

8.0

ANFO Cartridge booster/Cast

booster

Cord relays Rock

breaker 152 7.0 5.0 SME

Shock tubes

DATA COLLECTION

It has been established by the research [8] that the order of influence of various parameters for the fragmentation in competent rocks is: a) Explosive energy per unit volume of rock mass, i.e.

specific charge b) Explosive distribution within the rock mass c) Type of explosive d) Delay timing e) Joint system and its orientation with respect to blast

direction.

These findings have been the starting hypothesis for the selection of the input variables. Since the rock mass is competent; hence the specific charge has the most dominating role in deciding the quality of fragmentation. The specific charge is logically correlated with the number

of holes/row, number of rows, average depth, average spacing, average burden and total quantity of explosive fired in one round. If all these parameters are considered separately for the development for the model; then one can assess the effect of variation in individual parameter on the boulder count, if required. This is why the individual parameters were considered instead of specific charge. On the similar considerations the type of the explosive replaces the VoD and the density of the explosive. The explosive distribution is represented by the diameter of the blast hole and the stemming height. The geotechnical parameters and the delay practice were similar in the referred mines hence they have not been considered as an input variable. This makes the model site specific. In order to use the model in different geotechnical environment and/or with different delay practice and with different criterion for oversize fragments; the model requires training, validation and testing with new blast records.




49

This is not a major problem as most of the mines maintain the records which are required for the operationalizing the model. Thus the input variables are number of holes per row, number of rows, average spacing, average burden, average depth, diameter, average stemming, type of the explosive and the total charge. The target variable is the boulder count. The maximum feed size of the crushers of the quarries is 1 m. Therefore the fragments having the size more than 1 m are considered as boulders. More than three hundred blast records were generated and assessed for their technical feasibility for multiple regressions

modeling with SPSS. Subsequently, three hundred blast records were finalized for both statistical and ANN modeling. These three hundred records were divided in three groups. 191 records were used for training, 77 records for validation and 32 for testing of the ANN model. The records used for training and validation of ANN model were used for the development of statistical model whereas the 32 records used for testing of the ANN and statistical models were same. The Range of the data used for the development of the models is presented in Table-3.

Table-3. Range of the data used for the development of models.

Input variables

Value

Minimum Maximum Mean Standard

Deviation

Diameter, mm (D) 100 152 Not Applicable

Average burden, m (B) 2.80 4.65 3.73 0.61

Average spacing, m (S) 3.70 6.80 4.89 0.78

Average depth, m (H) 7.00 9.75 8.66 0.61

Average stemming, m (T) 2.60 4.30 3.33 0.39

Number of holes/row (N) 9 57 25 8.13

Number of rows (n) 2 3 Not Applicable

Type of the explosive ANFO and SME

Quantity of explosive fired per round, kg (Q)

1287 9567 4109 1897

Target variable

Value

Minimum Maximum Mean Standard Deviation

Boulder count (N’) 39 250 122 45.82

CONCEPT

ANN mimics the function of human brain. Since its inception, ANN has become popular and applicable to various fields of science and technology to solve complicated simulation problems. The ANN is capable of calculating arithmetic and logical functions, generalizing and transforming independent variables to the dependent variables, parallel computations, nonlinearity processing, handling noisy data, function approximation and pattern recognition [3]. ANN is trained using a set of real inputs and their corresponding outputs. For a better approximation, sufficient number of datasets is required. Performance of the trained model is checked with part of the available data known as testing datasets. To find out the best possible network, various topologies are constructed and tested. The process of model training-testing has to be continued until the optimum model with minimum error and maximum accuracy is achieved. A neural network has a layered structure, and each layer contains processing units or neurons. Input variables are placed in the input layer, whereas target variables are put in the output layer. The neurons in the hidden layers are the intermediate computation components (black box) of

the system. All of the layers are connected to each other by weighted connections. Each neuron is connected to the neurons in the subsequent layer. However, there is no connection between the neurons of the same layer. In the training process, the interconnections among the neurons are initially assigned specific weights. The network would be able to perform a function by adjusting the initial weights. In the process of ANN training, an initial arbitrary value (weight) is assigned to the connections and then to combine all of the weighted inputs and generate the neuron output the formula depicted in equation (1) is applied: O = ∑xiwi + b (1)

Where, xi is the input; wi is the connection weight and b is the bias.

The neuron output is mapped to actual output by the function indicated in equation (2), y = f (O) = f (∑xiwi + b) (2)




50

DEVELOPMENT OF ANN MODEL The ANN model described in this paper was

developed using the syntax available in ANN tool box of MATLAB. A back-propagation neural network was selected due to its simplicity and uniform approximation of any continuous function. The number of neurons in the input layer was nine and the number of neurons in the output layer was one. The tool box provides a number of training functions and each is having its own advantages and disadvantages. The objective of the training function is to adjust the weights and the bias vectors so as to find the complicated relation between the input and output. The application of a training function depends on several factors, such as the complexity of the problem, the number of training samples, the structure of the network and error target. Levenberg–Marquardt (LM) algorithm is used for

training the network because it has good generalization ability and has the capability of providing good predictions. LM algorithm provides a numerical solution to the problem of minimizing a function over a space of parameters of the function. Not only the first derivative information but also the second derivative information of the target function is used in the LM algorithm. It can dynamically adjust the convergence direction of iteration according to the iteration result, so its convergence speed is very fast. Generally, Log-Sigmoid ((Logsig), Hyperbolic tangent Sigmoid (Tansig), Positive Linear (Poslin) and Linear (Purelin) transfer functions are used in back propagation neural network (BPNN) [73]. These four transfer functions are considered for optimization. The optimization is ascertained by regression coefficients (Table 4 and Figure-1).

Table-4. Regression coefficients obtained with various transfer functions.

S. No. Transfer function Regression coefficients

Training Validation Testing Overall

1 Poslin 0.69 0.82 0.57 0.69

2 Logsig 0.75 0.47 0.62 0.64

3 Purelin 0.96 0.98 0.94 0.96

4 Tansig 0.59 0.44 0.51 0.55

Note- 1, 2, 3 and 4 indicate the serial numbers of the transfer functions indicated in Table-4.

In the case of the multilayer network, which can

be used for fitting or pattern recognition, the number of hidden layers is not determined by the problem, since any number of hidden layers is possible. The standard procedure is to begin with a network with one hidden layer. If the performance of the network is not satisfactory, then more number of layers can be used in the network. It would be unusual to use more than two hidden layers. The training becomes more difficult when multiple hidden layers are used which leads to slow convergence. The numbers of neurons in the hidden layers are determined by

the complexities of the function that is being approximated or the decision boundaries that are being implemented. Unfortunately, we don’t normally know how complex the problem is until we try to train the network. The standard procedure is to begin with more neurons and then change it. The network was trained using different number of hidden layers and neurons. The regression coefficients were noted for each of the architectures. The architecture providing the maximum regression coefficient was selected for the prediction [23]. Asa result of optimization, a BPNN model (Figure-2) with one hidden layer with two neurons in it, Levenberg-Marquardt as training function and purelin function as transfer function is finalized.

Figure-2. Developed artificial neural network.

Testing of ANN model The model predicts the boulder count based on

the input variables of 32 records comprising of ANFO blasts in 152 and 115 mm diameter holes. And, SME blasts in 152 and 100 mm diameter holes and the predictions are compared with target variables of these 32 records. The predictive ability of the neural network is evident from Figures 3 and 4.

0

0.2

0.4

0.6

0.8

1

1.2

1 2 3 4

Reg

ress

ion c

oef

fici

ents

Figure-1. Regression coefficients obtained with various transfer functions

Training

Validation

Testing

Overall




51

Figure-3. Curves obtained in MATLAB during training, validation and testing of the data sets.

Figure-4. Regression lines obtained in MATLAB during training, validation and testing of the data set.




52

Sensitivity analysis Sensitivity analysis is carried out to assess the

influence on each of the parameter on the predictions of ANN. The results are expressed as the strength of relations between the input and variables. Cosine amplitude method (CAM) is used to identify the most sensitive factors affecting boulder count. In this method, all of the data pairs are expressed in common X-space. The data pairs used to construct a data array X defined in equation (3) is X = {X1, X2, X3,...Xn} (3)

Each of the elements in the data array X is a vector of lengths of n, as shown in equation (4): Xi = {Xi1, Xi2, Xi3,…Xin} (4)

Thus, each of the pairs is a point in n-dimensional space, where each point requires m coordinates for a full description. Each element of a relation, results in a pair wise comparison of two data pairs. The strength of the relation between the data pairs, and, is expressed by the membership function given in equation (5). 𝑟 = ∑ 𝑥𝑛= ∗𝑥√∑ 𝑥𝑛= ∗∑ 𝑥𝑛= (5)

The analysis is carried and the order of the

influence of various input variables on the predictions is depicted in Table-5 and Figure-5 shows the strength of relations of input variables and the order of influence of each of the input variable on the boulder count. Table-5. Order of the influence of various input variables

on the predictions and the order of influence.

S. No. Input variable Strength of

relation

1 No. of holes per row 0.82

2 No. of rows 0.48

3 Average spacing 0.72

4 Average burden 0.78

5 Average depth 0.92

6 Diameter 0.76

7 Average stemming 0.80

8 Type of the explosive 0.50

9 Total charge 0.81

Figure-5. Strength of relations of input variables (1, 2, 3etc indicate the serial number of variables shown

in Table-5).

Since the investigated rock mass is competent hence the variables pertaining to the specific charge viz. number of holes per row, number of rows, average spacing, average burden, average depth and total charge and those related to explosive distribution viz. average stemming and total charge are expected to show a major influence on the boulder count. Table 5 indicates that each of the concerned variables except the number of rows has an influence of more than 70% on the boulder count. A low value of the strength of relation is owing to the fact that the number of rows is either two or three and as such there is no substantial variation in this variable. As per the starting hypothesis; type of the explosive comes last in the order of the influence amongst the variables considered in this research. This fact is further corroborated by the findings and the type of the explosive has an influence of 50 % on the predictions. The influence is less because there was not much variation in the type of the explosive and only ANFO and SME were available in this research. The benefits of the developed ANN model are: a) The model uses a limited number of input variables.

The records of the required input and target variables

are normally maintained in every mine. Requirement

of a large number of records for training and

validation of the model is therefore not a problem.

b) The target variable i.e. boulder count helps to plan

the secondary breakage operations properly.

c) The delay practice and the geotechnical features can

be incorporated in the model for prediction of

boulder count. In such a case, the program can be

easily modified to incorporate the changes.

0.82

0.48

0.72 0.78

0.92

0.76 0.8

0.5

0.81

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 3 4 5 6 7 8 9

Str

ength

of

rela

tio

ns

Input variables




53

d) The model can be used to assess the blast designs for

minimizing the boulder count.

The limitations of the developed ANN model are: a) The predictions of the model are site specific and in

the present form; it can predict the boulder count

accurately in respect of only those mines which have

a similar rock mass and analogous delay practice.

b) The model has been developed with target variable as

boulder count which is; in this case; the fragments of

size more than 1 m. Skewed predictions may be

obtained, if the fragments of some other size are

considered as boulders.

c) MATLAB environment is necessary for the

development of this ANN model and subsequent

predictions. However, in the absence of the

MATLAB; ANN program code can be developed for

the prediction of the boulder count.

DEVELOPMENT OF STATISTICAL MODEL The statistical model has been developed to

validate the predictions by the developed ANN model

since no model is available for validation that could predict the boulder count. The statistical model has been developed in SPSS environment using multiple regression analysis. Multiple regression analysis is widely used for prediction of various results in a complex phenomenon like rock blasting. The regression analysis establishes a statistical relationship between independent variables and a dependent variable in the form of a statistical model which is fitted to a set of experimental data. This method of modeling has been used for prediction of rock fragmentation [3, 15-16, 40-41, 50, 53, 58, 63, 73, 80-81]. The regression analysis can also be used for estimation in the variation of the dependent variable w.r.t. predictors. This can be ensured by using probability distributions t test and F test. A t test is used to know whether calculated effects are statistically significant or not. Referring the F tables, the degree of significance of the computed F value can be determined. F value is used to indicate whether the developed model is significant or not [14, 60]. Since the type of the explosive is a nominal variable; it was required to develop two separate models for ANFO and SME blasts. The relevant coefficients for the model are presented in Tables 5 and 6. The developed statistical models are as depicted in equations (6) and (7): For ANFO blasts:

𝑁′ = − . 𝑄𝑁𝐻 + . 𝑆𝐷 + . 𝐵𝐷 + . 𝑛 − . 𝐻 + . 𝑇 + . 𝑄 − . (6)

For SME blasts: 𝑁′ = − . 𝑄𝑁𝐻 + . 𝑆𝐷 + . 𝐵𝐷 + . 𝑛 − . 𝐻 + . 𝑇 + . 𝑄 − . (7)

It is seen from the regression equations that in

case of ANFO and SME blasts; the independent variables of the multiple regression equation viz. total charge to the number of blast holes in a row and average depth of a blast hole are negatively correlated to the boulder count. This indicates that the boulder count depends upon the distribution of the explosive in a blast hole and a good distribution ensures fewer boulders. The boulder count is directly proportional to the average spacing to diameter ratio and average burden to diameter ratio which validates the basic blasting theory. The models further indicate that

the average height of stemming column is directly proportional to boulder count which is a very well-known principle. The boulder count increases with total charge owing to the reason that boulder count increases with the volume of blasted in-situ rock mass and more rock is displaced with the increase in the total charge. The boulder count however reduces with the average depth. Table-6 indicates the results of the significance tests carried out using t test. It indicates that all the variables except the average depth are statistically significant in the prediction of boulder count in case of both ANFO and SME blasts.




54

Table-6. Results of significance test for ANFO and SME blasts.

Variables ANFO Blasts SME Blasts

t Sig t Sig

(Constant) -8.855 0.000 -5.799 0.000

Total charge to the number of blast holes in a row and average depth of a blast hole

-2.914 0.004 -3.605 0.004

Average spacing to diameter 6.222 0.000 5.426 0.000

Average burden to diameter 4.085 0.000 4.297 0.000

Number of rows 3.194 0.002 7.907 0.002

Average depth of the blast holes -1.537 0.126 -1.900 0.060

Average height of stemming 7.729 0.000 3.677 0.000

Total charge 24.499 0.000 42.000 0.000

Testing of statistical model

The testing of the models indicates that each of the predictors has its tolerance greater than 0.1 and the variance inflation factor for each of the independent variables is less than 10. Further, the Durbin-Watson statistic for the models on ANFO and SME blasts is 1.913 and 1.836 respectively indicating the independence of the predictors. The residuals have also been found to have normal distribution and the P-P plots indicate the normal distribution of the residuals. It is therefore found that the proposed model has all the desirable properties as required for a statistical model. The statistical model developed for ANFO blasts indicates that the R is 0.979 and adjusted R2 is 0.957. It means that 95.7% of the variance in the boulder count can be explained by the independent variables considered in this regression whereas in case of

SME blasts the R is 0.984 whereas the adjusted R2 is

0.967. It means that 96.7% of the variance in the boulder count can be explained by the independent variables considered in this regression. The accuracy of the statistical model is corroborated by the F test. The test indicates that both the models in the present form are significant since p< 0.001. It validates the assumption of linearity between the dependent and independent variables. The predictions by ANN model and the statistical model are presented in Table 7 and Figures 6, 7 and 8 represent the an overview of prediction capabilities of the two models against the actual boulder count, scatter obtained in actual boulder count and predicted boulder count by statistical model and scatter obtained in the predictions by ANN model and Statistical model respectively.




55

Table-7. Prediction results of boulder count by ANN and statistical model.

S. No. Explosive used Actual boulder

count

Predicted

boulder count by

ANN

Predicted

boulder count by

statistical model

1 ANFO 94 125 93

2 ANFO 99 119 98

3 ANFO 148 137 140

4 ANFO 128 131 120

5 ANFO 113 123 110

6 ANFO 110 120 109

7 ANFO 144 134 138

8 ANFO 135 139 130

9 ANFO 139 132 138

10 SME 70 71 72

11 SME 61 47 51

12 SME 88 102 90

13 SME 39 23 33

14 SME 45 24 48

15 SME 66 50 62

16 SME 91 89 86

17 SME 94 100 99

18 SME 84 89 89

19 ANFO 126 121 106

20 ANFO 129 120 109

21 ANFO 90 98 89

22 ANFO 51 53 49

23 ANFO 64 82 71

24 ANFO 60 68 92

25 ANFO 88 106 90

26 ANFO 64 75 82

27 SME 130 110 133

28 SME 152 131 146

29 SME 100 83 111

30 SME 112 102 119

31 SME 107 95 114

32 SME 135 111 131




56

Figure-6. An overview of prediction ability of the two models.

Figure-7. Scatter obtained between the actual boulder count and the predicted boulder count by ANN.

0

20

40

60

80

100

120

140

160

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

Bo

uld

er c

ount

Blast number

Actual boulder count

Predicted boulder count byANN

Predicted boulder count bystatistical model

0

20

40

60

80

100

120

140

160

0 20 40 60 80 100 120 140 160

Pre

dic

ted

bo

uld

er c

ount

by A

NN

Actual boulder count

R² = 0.8184




57

Figure-8. Scatter obtained between the predictions by ANN and statistical model.

RESULTS

Rock blasting is multi-faceted phenomenon, the results of which are affected are affected by the interaction of a number of uncontrollable parameters. It can therefore be safely said that an accuracy of more than 80 % in the predictions of boulder count is sufficient. It is observed that the correlation coefficient between the actual boulder count and predicted boulder count by ANN is more than 0.9; which can be considered as satisfactory. It was also found that in case of an ANN model; an accuracy of more than 80 % was achieved in 24 predictions out of 32. The accuracy was less than 80% but was around 75 % in four predictions out of 32 cases. It is difficult to state the exact cause of unexpected variation in the predicted boulder count in the remaining four cases as blasting is a complex phenomenon and the results can be affected because of unexpected changes in uncontrollable parameters. The boulder count may vary because of the unforeseen presence of subsurface cracks/fissures, due to unexpected localized changes in the elastic properties/joint properties, etc. The predictions by ANN and the statistical model are in close agreement. The coefficient of correlation between the actual boulder count and the predicted boulder count by the statistical model exceeds 0.9 and the value of the same between predictions by ANN and the statistical model is around 0.9. The test of significance indicated that in case of ANFO and SME blasts; all the variables except the average depth added statistically significantly to the prediction of boulder count with p< 0.05. The statistical model has been developed to ascertain the accuracy of predictions by the ANN model which has nine input variables. The selection of these variables is based on the well-established research findings [8] and the presentation

of full model is the best method to reflect the range of predictors investigated [18]. Therefore the average depth which is the statistically insignificant predictor has been retained in the proposed statistical models. The accuracy of the statistical model is corroborated by the F test. The test indicates that both the models in the present form are significant since p< 0.001.

CONCLUSIONS

The artificial neural network seems to be a competence measure to predict rock fragmentation. In this study, a new ANN model was established to predict boulder count in rock blasting. The new model is a BPNN with two hidden layers with two neurons in each of the hidden layer. The model uses LM training algorithm and purelin transfer function. The model indicated an overall high correlation coefficient of 0.96. The correlation coefficient between actual and predicted number of boulder count was more than 0.90. The performance of the model was validated using a statistical model developed in SPSS which also showed a correlation coefficient around 0.9. A perusal of the results indicated that majority of the predictions by two models were in close agreement. It can therefore be concluded that the developed ANN model has satisfactory predictive abilities. The sensitivity analysis of the ANN model indicates that the spacing, burden and stemming have substantial influence on the boulder count. Further, the tests of significance indicate that the explosive distribution, height of stemming column and the ratios of spacing and burden to diameter are statistically significant in case of both ANFO and SME blast models. Due attention is therefore required to be paid to maintain the

0

20

40

60

80

100

120

140

160

0 20 40 60 80 100 120 140 160

Pre

dic

ted

bo

uld

er c

ou

nt

by

sta

tist

ica

l m

od

el

Predicted boulder count by ANN

R² = 0.7847




58

accuracy in drilling and charging the blast holes so that the boulder count can be minimized in the referred quarries.

ACKNOWLEDGEMENT

The authors are thankful to the management of Rawan, Hirmi, Sonadih and Baikunth Limestone Quarries from where the blast records were generated. Without their unflinching support; the study could not be completed.

REFERENCES

[1] Aler J., Mouza J.D. and Arnould M. 1996.

Measurement of the fragmentation efficiency of rock

mass blasting and its mining applications, Int J of

Rock Mech and Min Sci and Geo Abstracts. 3: 125-

139.

[2] Ash R. 1963. The mechanics of the rock breakage

(Part 1), Pit and Quarry. 56(2): 98-100.

[3] Bahrami A., Monjezi M., Goshtasbi K. and

Ghazvinian A. 2011. Prediction of rock fragmentation

due to blasting using Artificial Neural Network,

Engineering with Computers. 27: 177-181.

[4] Bhandari S. 1997. Rock Fragmentation by Blasting

Engineering In: Rock Blasting Operations, A. A.

Balkema/Rotterdam/Brookfield.

[5] Castro J.T., Liste A. V. and Gonzalez A.S. 1998.

Blasting index for exploitation of aggregates, 7th Mine

Planning and Equipment Selection Symposium,

Calgary, Canada. pp. 165-168.

[6] Chakraborty A. K., Jethwa J.L., and Paithankar A. G.

1990. Effects of joint orientation and rock mass

quality on tunnel blasting, Engineering Geology. 37:

247-262.

[7] Chakraborty A.K., Raina A. K., Ramlu M.,

Choudhury P.B., Haldar A., Sahu P. and

Bandopadhyay C. 2004. Parametric study to develop

guidelines for blast fragmentation improvement in

jointed and massive formations, Engineering

Geology. 73: 105-116.

[8] Chiappetta R.F. 1998. Choosing the right delay timing

for the blasting application, optimization and

maintainingfield control, 8th High-Tech Seminar on

State-of-the Art, Blasting Technology,

Instrumentation and Explosives Applications,

Allentown PA, USA. pp. 215-254.

[9] Chung S.H. and Katsabanis P.D. 2000. Fragmentation

prediction using improved engineering formulas, Int J

for Blasting and Fragmentation. 4: 198-207.

[10] Cunningham C.V.B. 1983. The Kuz-Ram Model for

prediction of fragmentation from blasting, 1st

International Symposium on Rock Fragmentation by

Blasting, Luleå University of Technology, Luleå,

Sweden. pp. 439-453.

[11] Cunningham C.V.B. 1987. Fragmentation estimations

and Kuz-Ram Model-Four years on, 2nd International

Symposium on Rock Fragmentation by Blasting,

Keystone, Colorado. pp. 475–487.

[12] Dhekne P.Y., Pradhan M. and Jade R. 2014.

Assessment of performance of explosives by

estimating the number of oversize boulders using

ANN Model, The Indian Mining and Engineering

Journal. 53: 8-13.

[13] Dhekne P.Y., Pradhan M. and Jade R.K. 2015.

Assessment of the effect of blast hole diameter on the

number of oversize boulders using ANN Model,

Journal of the Institution of Engineers (India) Series

D, DOI 10.1007/s40033-015-0083-7.

[14] Drape N.R. and Smith Jr H. 1981. Applied Regression

Analysis, 2nd ed, John Wiley and Sons Inc, New

York, USA.

[15] Ebrahimi E., Monjezi M., Khalesi M.R. and

Armaghani D.J. 2015. Prediction and optimization of

backbreak and rock fragmentation using an Artificial

Neural Network and Bee Colony Algorithm, Bulle

Engineering geo and Environ, DOI 10,1007/s10064-

015-0720-2.

[16] Enayatollahi I., Aghajani B.A. and Asa A. 2014.

Comparison between Neural Network and multiple

regression analysis to predict rock fragmentation in

open pit mines, Technical Note, Rock Mechanics and

Rock Engineering. 47: 799-807.

[17] Esen S., Onederra I., and Bilgin H.A. 2003.

Modelling the size of crushing zone around a blast

hole, Int. J of Rock Mech and Min Sci. 40: 485-495.

[18] Forstmeier W. and Schielzeth H. 2011. Cryptic

multiple hypotheses testing in linear models:

Overestimated effect sizes and the winner's curse,

Behavioural Ecology and Sociobiology. 65: 47-55.




59

[19] Gama D. D. 1983. Use of comminution theory to

predict fragmentation of jointed rock masses

subjected to blasting, 1st International Symposium on

Rock Fragmentation by Blasting, Luleå University of

Technology, Luleå, Sweden. pp. 565-579.

[20] Gate W., Ortiz L. and Florez R. 2005. Analysis of

rock fall and blasting back break problems, American

Rock Mechanics Conference, Vol 4, Anchorage,

Alaska. pp. 1290-1298.

[21] Ghosh A., Daemen J.J.K., and Vanzyl D. 1990.

Fractal based approach to determine the effect of

discontinuities on blast fragmentation, 31st U.S.

Symposium on Rock Mechanics: Contributions and

Challenges, Golden, Colorado. pp. 905-912.

[22] Grundstrom C., Kanchibotla S., Jankovic A. and

Thornton D.M. 2001. Blast fragmentation for

maximizing the SAG mill throughput at Porgera

Goldmine, 27th Annual Conference on Explosives and

Blasting Technique, Orlando, Florida. pp. 383-399.

[23] Hagan M.T., Demuth H.B., Beale M.H. and De Jesus

O. Neural Network Design 2nd

edition,http://hagan.okstate.edu/NNDesign.pdf.

[24] Hagan T.N. 1995. The effect of rock properties on the

design and results of tunnel blasts, J of Rock Mech

and Tunnelling Tech. 1(1): 25-39.

[25] Hall J. and Brunton I. 2002. Critical comparison of

Kruttschnitt Mineral Research Center (JKMRC) blast

fragmentation models, Fragblast: Int J for Blasting

and Fragmentation. 6: 207-220.

[26] Hamdi E. and Mouza J.D. 2005. A methodology for

rock mass characterization and classification to

improve blast results, Int J of Rock Mech and Min

Sci. 42: 177-194.

[27] Hamdi E., Mouza J.D. and Fleurisson J.A. 2001.

Evaluation of the part of blasting energy used for rock

mass fragmentation, Fragblast: The Int J of Blasting

and fragmentation. 5: 180-193.

[28] Hjelmberg H. 1983. Some ideas on how to improve

calculations of the fragment size distribution in bench

blasting, 1st International Symposium on Rock

Fragmentation by Blasting, Luleå University of

Technology, Luleå, Sweden. pp. 469-494.

[29] Hustrulid W. 1999. The Fragmentation System

Concept, In: Blasting Principles for Open Pit Mining,

A.A. Balkema, Rotterdam.

[30] Jhanwar J.C., Jethwa J.L. and Reddy A.H. 2000.

Influence of air-deck blasting on fragmentation in

jointed rocks in an open-pit Manganese mine,

Engineering Geology. 57: 13-29.

[31] Jimeno C.L., Jimeno E.L. and Carcedo F.J.A. 1995.

Controllable Parameters of Blasting, In: Drilling and

blasting of Rocks, Balkema A. A., Rotterdam.

[32] Jong Y. and Lee C. 2002. Application of Neural

Networks to prediction of powder factor and peak

particle velocity in tunnel blasting, 28thAnnual

Conference on Explosives and Blasting Technique,

Las Vegas, Nevada. 2: 67-76.

[33] Jong Y.H. and Lee C.I. 2004. Influence of geological

conditions on the powder factor for tunnel blasting,

Int J of Rock Mech and Min Sci. 41: 533-538.

[34] Kanchibotla S.S., Valery W., and Morrell S. 1999.

Modeling fines in blast fragmentation and its impact

on crushing and grinding, EXPLO 99 Conference,

Kalgoorlie, Australia. pp. 137-144.

[35] Karami A. and Afluni Z.S. 2013. Sizing of rock

fragmentation modeling due to bench blasting using

adaptive neuro-fuzzy inference system (ANFIS), Int. J

of Min Sci and Tech. 23: 809-813.

[36] Karami A. R., Mansouri H., Farsangi M.A.E. and

Nezamabadi H. 2006. Backbreak prediction due to

bench blasting: An Artificial Neural Network

approach, J of Mine, Metals and Fuels. 54: 418-420.

[37] Khandelwal M. and Singh T.N. 2006. Prediction of

blast induced ground vibrations and frequency in

opencast mine: A Neural Network Approach, J of

Sound Vibration. 289: 711-725.

[38] Kojovic T., Michaux S., and McKenzie C. 1995.

Impact of blast fragmentation on crushing and

screening operations in quarrying, EXPLO 95

Conference, AusIMM, Brisbane, Australia. pp. 427-

436.

[39] Kou S. and Rustan A. 1993. Computerised design and

result prediction of bench blasting, 4th International

Symposium on Rock fragmentation by Blasting,

Vienna, Austria. pp. 263-271.




60

[40] Kulatilake P.H.S.W., Hudaverdi T. and Qiong W.

2012. New prediction models for mean particle size in

rock blast fragmentation, Geotech and Geological

Engineering, 30: 665-684.

[41] Kulatilake P.H.S.W., Qiong W., Hudaverdi T. and

Kuzu C. 2010. Mean particle size prediction in rock

blast fragmentation using Neural Networks,

Engineering Geology. 114: 298-311.

[42] Kuznetsov V. M. 1973. Mean diameter of fragments

formed by blasting rock, Soviet Mining Science. 9(2):

144-148.

[43] Latham J.P. and Lu P. 1999. Development of an

assessment system for the blastability of rock masses,

Int J of Rock Mech and Min Sc. 36: 41-55.

[44] Leu S.S., Lin S.F., Chen C.K. and Wang S.W. 1998.

Analysis of powder factors for tunnel blasting using

Neural Networks, Fragblast: Int J for Blasting and

Fragmentation, 2: 433-448.

[45] Lilly P.A. 1986. An empirical method of assessing

rock mass blastability, Large Open Pit Conference,

IMM, Australia. pp. 89-92.

[46] Mckenzie S. 1966. Cost of explosives-Do you

evaluate it properly? Mining Congress Journal. 52 (5):

32-41.

[47] Mishnaevsky Jr L.L. and Schmauder S. 1996.

Analysis of rock fragmentation with the use of the

theory of fuzzy sets, ISRM International Symposium

(Eurock '96), Torino, Italy. 735-740.

[48] Mishra A. K., Kumar A., and Nizami A. A. 2008.

Digital blasting technology, Conference on Emerging

Trends in Mining and Allied Industries, NIT

Rourkela. pp. 80-92.

[49] Mishra A.K. 2008. Safety engineering in mine

blasting operations, Course Material on Safe Mining

Methods, Design and Technology, IIT, Kharagpur.

[50] Moghadam M.J., Farsangi M.A.E., Mansouri H. and

Nezamabadi H. 2006. Muck -pile fragmentation

prediction using Artificial Neural Networks, J of

Mines, Metals and Fuels. 54: 421-423.

[51] Mojtabai N., Farmer I. W. and Savely J.P. 1990.

Optimisation of rock fragmentation in bench blasting,

31st U.S. symposium on Rock Mechanics

Contributions and Challenges, Golden, Colorado. pp.

897-904.

[52] Monjezi M, Singh T.N., Khandelwal M., Sinha S.,

Singh V. and Hosseini I. 2006. Prediction and

analysis of blast parameters using Artificial Neural

Network, Noise Vibration Worldwide. 37: 8-16.

[53] Monjezi M., Amiri H., Farrokhi A. and Goshtasbi K.

2010. Prediction of rock fragmentation due to blasting

in Sarcheshmeh Copper mine using Artificial Neural

Networks, Geotech and Geological Engineering. 28:

423-430.

[54] Monjezi M., Dehghan H. and Samimi N.F. 2007.

Application of TOPSIS method in controlling fly rock

in blasting operations, 7th International Science

Conference SGEM, Sofia, Bulgaria. pp. 41-49.

[55] Monjezi M., Ghafurikalajah M. and Bahrami A. 2011.

Prediction of blast-induced ground vibration using

Artificial Neural Networks, Tunnelling and

Underground Space Technology. 26: 46-50.

[56] Monjezi M., Hashemi S.M., Hashemi R., Johari M.V.

and Khandelwal M. 2014. Artificial Neural Network

as a tool for back break prediction, Geotech and Geo

Engineering. 32: 21-30.

[57] Monjezi M., Mehrdanesh A., Malek A., and

Khandelwal M. 2013. Evaluation of effects of blast

design parameters on flyrock using Artificial Neural

Networks, Neural Computing and Applications. 23:

349-356.

[58] Monjezi M., Mohamadi H.A., Barati B. and

Khandelwal M. 2014. Application of soft computing

in predicting rock fragmentation to reduce

environmental blasting side effects, Arab Journal of

Geoscience. 7: 505-511.

[59] Monjezi M., Yazdian A., and Hesami S.M. 2006. Use

of back propagation Neural Network to estimate

burden in tunnel blasting, J of Mines, Metals and

Fuels. 12: 424-429.

[60] Montgomery D.C., Peck E.A. and Vining G.G. 2006.

Introduction to Linear Regression Analysis, John

Wiley & Sons Inc, New Jersey, USA.

[61] Morin M.A. and Ficarazzo F. 2006. Monte Carlo

simulation as a tool to predict blasting fragmentation

based on the Kuz–Ram model, Computers and

Geosciences. 32: 352-359.

[62] Nie S. L. and Rustan A. 1987. Techniques and

procedures in analysing fragmentation after




61

blasting by photographic method, 2nd International

Symposium on Rock Fragmentation by Blasting,

Keystone, Colorado. pp. 36-47.

[63] Oraee K. and Asi B. 2006. Prediction of rock

fragmentation in open pit mines using Neural

Network Analysis, 15th International Symposium on

Mine Planning and Equipment Selection (MPES),

Torino, Italy. pp. 19-24.

[64] Ouchterlony F. 2003. Influence of blasting on size

distribution and properties of muckpile fragments: A

state-of-the-art review, MinFo Project P2000-10,

Swebrec, Luleå University of Technology, Sweden.

pp. 21-48.

[65] Ouchterlony F., Niklasson B. and Abrahamsson S.

1990. Fragmentation monitoring of production blasts

at Mrica, International Symposium on Rock

Fragmentation by Blasting, Frag Blast 3, Brisbane,

Australia. pp. 283-289.

[66] Ozcelik Y. 1998. Effect of discontinuities on fragment

size distribution in open-pit blasting: A case study,

Trans of the Institution of Mining and Metallurgy.

107: 146-150.

[67] Pal R. P. 1995. Breakage assessment through cluster

analysis of joint set orientations of exposed benches

of opencast mines, Geotechnical and Geological


[68] Rai P. and Baghel S.S. 2004. Investigation of firing

patterns on fragmentation in an Indian opencast

limestone mine, Quarry Management Journal, UK. pp.

33-37.

[69] Raina A. K., Ramulu M., Choudhary P.B., and

Chakraborty A.K. 2002. Fragmentation prediction in

different rock masses characterized by drilling index,

7th International Symposium on Rock fragmentation

by Blasting, Met Industry Press, Beijing, China. pp.

117-121.

[70] Rustan P.A. 1998. Automatic image processing and

analysis of rock fragmentation- Comparison of

systems and new guidelines for testing the systems,

Int J for Blasting and Fragmentation, Fragblast. 2: 15-

23.

[71] Sanchidrian J.A., Segarra P. and Lopez L.M. 2007.

Energy components in rock blasting, Int J of Rock

Mech and Min Sci. 44: 130-147.

[72] Sawmliana C., Pal R.P., Singh R. K. and Singh T.N.

2007. Blast induced air overpressure and its prediction

using Artificial Neural Network, Mining Technology,

Trans, Inst, Min, Metall, A. 116: 41-48.

[73] Sayadi A., M Monjezi., Talebi N. and Khandelwal M.

2013. A comparative study on the application of

various Artificial Neural Networks to simultaneous

prediction of rock fragmentation and back break, J of

Rock Mech and Geotech Engineering. 5: 318-324.

[74] Schuhmann Jr R. 1959. Energy input and size

distribution in comminution, Am, Inst, Min, Metall,

ANNME. 214: 22-25.

[75] Shao P., Xu Z. W., Zhang H.Q. and He Y. N. 2009.

Evolution of blast-induced rock damage and

fragmentation prediction, Procedia Earth and

Planetary Science, International Conference on

Mining Science & Technology (ICMST2009),

Xuzhou, China. 1: 585-591.

[76] Singh T. N. 1995. A study of blast induced ground

vibration at Dharapani Magnesite Mine, Pitthoragarh

Himalaya, India, 3rd International Symposium on

Headwater Control, New Delhi, India. pp. 183-188.

[77] Tawadrou A.S. and Katsabanis P.D. 2005. Prediction

of surface blast patterns in Limestone quarries using

Artificial Neural Networks, Fragblast: Int J for

Blasting and Fragmentation, 9: 233-242.

[78] Thote N. R. and Singh D.P. 2000. Effect of air-

decking on fragmentation: A few case studies of

Indian mining, Explosive & Blasting Technique,

Balkema. pp. 265-270.

[79] Trivedi R., Singh T. N. and Raina A.K. 2014.

Prediction of blast induced fly rock in Indian

Limestone mines using Neural Networks, J of Rock

Mech and Geotech Engineering. 6: 447-454.

[80] Xiu S. Z., Dan H., Jian Z. and Shu Z. 2013. Combined

ANN prediction model for rock fragmentation

distribution due to blasting, J of Information and

Computational Science. 10: 3511-3518.

[81] Zhu H.B. and Wu L. 2005. Application of Gray

correlation analysis and Artificial Neural Network in

rock mass blasting, J of Coal Science and


BOULDER PREDICTION IN ROCK BLASTING USING ... developed neural network model is suitable for...

Documents

Transcript of BOULDER PREDICTION IN ROCK BLASTING USING ... developed neural network model is suitable for...