Documente Academic
Documente Profesional
Documente Cultură
Tunnel Blasting
ABSTRACT
The main purpose of this study was to apply the artificial neural network (ANN) models to determine
optimal powder factors and predict peak particle velocities, based on a series of observations and input
parameters. The used blasthole pattern was V-cut. The locations of blastholes and ignition patterns
remained almost constant. The input parameters were 14 geological conditions (i.e., dip, dip direction,
spacing, separation, and persistence of 2 discontinuity sets, tunnel orientation, rock strength, RQD, and
RM R) and 4 blasting conditions (i.e., distance between blasting points and measuring points, charge
weight per delay, drilling length, and blasting efficiency). Data for the ANN application in this study
were recently collected in a highway tunnel under construction in Korea. The main rock type in the site
was biotitic granite. An optimum ANN model was determined by training models with collected data.
The trained model was used to evaluate the prediction capability of the optimum model. It was shown
that the ANN model could predict the powder factor and the peak particle velocity depended upon the
selected input parameters. Particularly, the ANN model predicted more accurate peak particle velocities
than those obtained by the blasting vibration equation. M oreover, the more dominant factors among the
input parameters could be determined by analysis of the relative strength effect.
INTRODUCTION
The design of blasting pattern is important to determine the blasting efficiency, damage of the tunnel
wall, vibration and noise level caused by the blasting in tunnel blasting. The allowable peak particle
velocity with respect to any adjacent structures and the powder factor depend on a selection of the
optimum pattern. However, in Korean tunnel construction sites, blasting operations have been performed
on the basis of the experience of blasting engineers. Consequently, blasting engineers have given
different weighting for geological conditions that have an influence upon the powder factor and the
particle velocity.
The Powder Factor, which defines the weights of explosives used (in kilograms) per the broken rock
volume (in cubic meters), is inversely proportional to the cube root of the fracture frequency (Scott,
1996), and is proportion to the Blastability Index, that is obtained by summing the representative values
of the rock mass description, joint plane spacing, joint plane orientation, specific gravity influence and
hardness (Lilly, 1986). The charge weight per delay to control blasting vibrations which are generated
by blasting operation at under an allowable particle velocity, is estimated from equations of blasting
vibration which were obtained empirically. Systematic studies that consider synthetically geological and
blasting conditions for estimating the powder factors and predicting the peak particle velocity, however,
Data used for the ANN development in this study were collected directly at a highway construction site
in Korea. The main rock type was weathered biotitic granite. The blasthole pattern used at the site was
the V-cut and is shown in Fig. 1. Because data that were used during this study were collected in a
tunnel construction site that the locations of blastholes and ignition patterns were constant, the
mentioned blasting conditions were excluded as input parameters of ANN. Input parameters selected for
this study, were based on those previously identified. Fourteen input parameters were used to predict the
powder factors and 18 input parameters were used to predict the peak particle velocities generated by
tunnel blasting. Table 1 shows the input and output parameters. The input parameters are (A~F) to
predict the powder factor and (A~J) to predict the peak particle velocity.
For ANN training, 25 data sets were used to predict the powder factors in tunnel blasting, and 27 to
predict the peak particle velocities.
ANNs are classified as various types by the kind of learning algorithms, network topology and data they
accept, etc. Of these classifications, this study adopted a multi-layer perceptron using a back-
propagation learning algorithm, and a sigmoidal function as an activation function.
ANN has different learning efficiencies and prediction capacities, which depend upon the learning
constants, such as the learning rate and the momentum rate, the number of neurons in the input and
output layer and the architecture of the hidden layer. During the learning period of the ANN, these
parameters were changed to determine an optimal ANN model for predicting the powder factor and the
A prediction capacity of the selected ANN for powder factor and particle velocity in tunnel blasting was
assessed using sample data not used for ANN training, and the predicted velocities by ANN were
compared with the results of the equations for blasting vibration using the ANN training data.
Table 2 shows the data used to access the prediction capacity of an optimized ANN (architecture:18-5-5-
1, learning rate:0.95 and momentum rate:0.95) for predicting a powder factor in tunnel blasting. Table 3
shows the results of a comparison between the predicted powder factor obtained using the ANN model
and the measured values. The ANN model for predicting the powder factor showed that the RM S of the
3 3
predicted values was 0.0433 kg/m . The value was larger than the RM S (i.e., 0.000737 kg/m ) of the
errors obtained using the training data. This was due to an insufficient generalization of ANN because of
the small number of training data sets. However, the predicted values have a mean absolute error of 0.04
3
kg/m and a mean relative error of 4.72 % with respect to the measured values. The absolute and relative
error are defined as;
Absolute Error = Ti Oi (1)
T
Relative Error = i (2)
Oi
where T i and Oi represent the targeted output used in ANN learning and the output learned by ANN,
respectively.
Therefore, it was believed that the trained ANN took changes of geological conditions into
appropriate considerations and had a satisfactory prediction capacity
The training data for ANN modeling was used to determine the equation for blasting vibration. Firstly,
in the case of using a root square scaled distance, a 50 % mean regression equation is eqn. (3) and a
95 % blasting vibration equation that uses in the blasting design and that considers the safety factor is
represented by eqn. (4).
1. 964
D (50 %) (R = 0.887) (3)
V = 780.0 1 / 2
W
The data presented in Table 4 was used to access the predictive capacity of the optimized ANN
(architecture:18-5-5-1, learning rate:0.95 and momentum rate:0.80) for predicting peak particle velocity
in tunnel blasting. Table 5 is the result of a comparison between the predicted values of the ANN model
and the measured peak particle velocities. The ANN model showed that the RM S of the errors of the
predicted values to the measured values was 0.0156 cm/sec. This value is larger than the RM S (i.e.,
0.00361 cm/sec) of the errors obtained using training data. However, the RM S of the values predicted by
the equation for blasting vibration with the training data was 0.150 cm/sec.
Fig. 2 represents the measured and the values predicted by ANN and by the equations for blasting
vibration, i.e., eqn. (3)-(6). It shows that the prediction values of the equations are 60 % more than the
measured values, which means that the equations dont consider the geological conditions sufficiently.
Therefore, the velocity prediction using ANN is more accurate than the method using the equations
for blasting vibration, i.e., the results of a regression analysis using measured data, which reflect only
local geological conditions, because the peak particle velocity is dependent upon geological conditions.
The relative strength effect (RSE) expresses the influent extent of different input values on one of the
output values. The larger the value of RSE becomes, the greater is the effect on the output, and the sign
of every RSE indicates how to the input affects the output. That is, if the output has no relation with the
input, the values are zero, and a positive action applies on the output when the values are positive, and a
negative action applies on the output when the values are negative. (Yang et al., 1998)
Fig. 3 shows the mean RSE values of each input parameter for the training data sets used for predicting
powder factors in tunnel blasting. In the figure, the length of the drawn line above each column indicates
the variance of the RSE values.
Hence, in the case of this tunnel, a hierarchy of the input parameters on the powder factors is obtained
as:
In case of the training data sets used to predict peak particle velocities in tunnel blasting, the mean RSE
values of each input parameter are shown in Fig. 4. In the figure, it is shown that the distance between
blasting points and measuring points (J) is the most dominant factor on the peak particle velocity among
the 16 input parameters considered. And although the secondary discontinuity spacing (C3) and dip (C2),
the major discontinuity dip (B2) and rock strength (D) in this order, have an impact on peak particle
velocity, the variance of the parameters is large. The tunnel orientation (A) and the RM R (F), however,
are less important factors than the other input parameters.
It is important to execute tunnel blasting within the range of allowable particle velocities and to
maximize the blasting efficiency. Therefore, it is necessary to investigate input parameters that have an
influence on the peak particle velocities generated by blasting vibration under different blasting
efficiency conditions. For the data from the good blasting efficiency group (data numbers: 1, 2, 5 and 6)
and the poor group (data numbers: 15, 19 and 20) among the training data for ANN, RSE analysis
results were examined. When the blasting efficiency, i.e., the ratio of the real to the designed advance is
more than 95 %, the data belongs to a good blasting efficiency group. On the other hand, when the
blasting efficiency is less than 80 %, the data belong to a poor blasting efficiency group. Fig. 5 shows
the RSE values of the good blasting efficiency group and Fig. 6 shows the values of the poor group.
The figures show that, independently of the blasting efficiency, the distance between blasting points
and the measuring (J) has greatest negative impact on the peak particle velocity among the considered
input parameters. This means that the greater the distance, the less the peak particle velocity. Both the
data of the good blasting efficiency group and the poor blasting efficiency group have similar curve
shapes, which are linked by the RSE values of each of the input parameters. However, the variance of
the RSE values in the good blasting efficiency group is lower than those of the poor blasting efficiency
group. This shows that variations of input parameter that has a large variance of the RSE value, have a
dominant impact on the velocity when the blasting efficiency is poor.
CONCLUSIONS
In this study, the geological and blasting conditions were directly measured in a tunnel construction and
were used to investigate the possibility of applying ANN to predict powder factors and peak particle
velocities. Parameters affecting the powder factor and the peak particle velocity in tunnel blasting were
also studied by adopting the RSE concept. The conclusions of this study are:
- The ANN model constructed to predict the powder factor, showed that the RM S of the errors of the
3
predicted values versus the measured values was 0.0433 kg/m . The value was larger than the RM S of
REFERENCES
Huang Yi and Stefan Wanstedt, 1998, The introduction of neural network system and its applications in
rock engineering, Engineering Geology 49, pp. 253-260.
Lilly P., 1986, An empirical method of assessing rock mass blastability, Large Open Pit M ining Conf.,
Newman, pp. 89-92.
M eulenkamp F. and Alvarez Grima M ., 1999, Application of neural networks for the prediction of the
unconfined compressive (UCS) from Equotip hardness, Int. J. Rock M ech. M in. Sci. 30, pp. 207-222.
Scott A., 1996, 'Blastability' and Blast Design, 5th Int. Symp. on Rock Fragmentation by Blasting,
M ontreal, pp. 27-36.
Sou-Sen Leu, Sheng-Feng Lin, Ching-Kuang Chen and Shih-Wen Wang, 1998, Analysis of powder
factors for tunnel blasting using neural networks, FLAGBLAST 2, pp. 433-448.
Yang Y. and Zhang Q., 1997, A Hierarchical Analysis for Rock Engineering Using Artificial Neural
Networks, Rock M ech. Rock Eng., 30(4), pp. 207-222.
Yang Y. and Zhang Q., 1998, The Application of Neural Networks to Rock Engineering System (RES),
Int. J. Rock M ech. M in. Sci. 35(6), pp. 727-745.
Table 2. Testing data sets for accessing an optimum ANN to predict the powder factor.
Sample Number S-11 S-12 S-13 S-14
Tunnel orientation 090 090 090 090
M ajor dip direction 259 086 081 257
discontinuity dip 82 86 88 87
spacing 2 2 3 2
separation 2 2 2 2
persistence 2 2 2 3
Secondary dip direction 346 005 357 340
discontinuity dip 77 73 77 78
spacing 1 2 2 1
separation 3 3 2 2
persistence 3 3 3 3
Rock strength 3 4 5 2
RQD 1 2 2 2
RM R 38 52 39 44
Table 4. Testing data sets for accessing an optimum ANN to predict the peak particle velocity.
Sample Number S-21 S-22 S-23 S-24
Tunnel orientation 090 090 090 090
M ajor dip direction 086 086 086 081
discontinuity dip 86 86 86 88
spacing 2 2 2 3
separation 2 2 2 2
persistence 2 2 2 2
Secondary dip direction 005 005 005 357
discontinuity dip 73 73 73 77
spacing 2 2 2 2
separation 3 3 3 2
persistence 3 3 3 3
Rock strength 4 4 4 5
RQD 2 2 2 2
RM R 52 52 52 56
Drilling length (m) 1.70 1.70 1.70 2.20
Blasting efficiency (%) 88.2 88.2 88.2 81.8
Charge weight per delay (kg) 2.56 2.56 2.56 3.20
Distance (m) 70.0 80.5 93.5 93.9
Table 5. Results of a comparison between the predicted values obtained using the ANN model and the
measured peak particle velocity.
Sample Peak Particle Velocity (cm/sec) Error
Number M easured value Value predicted by Absolute Error Relative Error
ANN (cm/sec) (%)
S-21 0.28 0.265 0.015 5.36
S-22 0.20 0.193 0.007 3.50
S-23 0.14 0.166 -0.026 -18.57
S-24 0.20 0.195 0.005 2.50
0.8
0.6
0.4
0.2
0.0
S-21 S-22 S-23 S-24
Sa mp le N um be r
Fig. 2. Comparison of the measured and the predicted peak particle velocities.
1.0
0.8
0.6
0.4
0.2
R.S.E.
0.0
A B1 B2 B3 B4 B5 C1 C2 C3 C4 C5 D E F
-0.2
-0.4
-0.6
-0.8
-1.0
INPU T PARAME TER
Fig. 3. M ean RSE values of each input parameter for the training data sets (powder factor).
0.8
0.6
0.4
0.2
R.S.E.
0.0
A B1 B2 B3 B4 B5 C1 C2 C3 C4 C5 D E F G H I J
-0.2
-0.4
-0.6
-0.8
-1.0
INPU T PARAME TER
Fig. 4. M ean RSE values of each input parameter for the training data sets (peak particle velocity).
1.0
0.8
0.6
0.4
0.2
1
2
R.S.E.
0.0 5
A B1 B2 B3 B4 B5 C1 C2 C3 C4 C5 D E F G H I J
6
-0.2
-0.4
-0.6
-0.8
-1.0
INPU T PARAMET ER
1.0
0.8
0.6
0.4
0.2
15
0.0 19
R.S.E.
A B1 B2 B3 B4 B5 C1 C2 C3 C4 C5 D E F G H I J 20
-0.2
-0.4
-0.6
-0.8
-1.0
INPU T PARAMET ER