You are on page 1of 7

Tunnelling and Underground Space Technology 27 (2012) 23–29

Contents lists available at ScienceDirect

Tunnelling and Underground Space Technology


journal homepage: www.elsevier.com/locate/tust

ANN and ANFIS performance prediction models for hydraulic impact hammers
Melih Iphar ⇑
Department of Mining Engineering, Eskisehir Osmangazi University, Meselik, Eskisehir, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: Hydraulic impact hammers are mechanical excavators that can be used in tunneling projects economi-
Received 7 February 2011 cally under geologic conditions suitable for rock breakage by indentation. However, there is relatively less
Received in revised form 23 May 2011 published material in the literature in relation to predicting the performance of that equipment employ-
Accepted 15 June 2011
ing rock properties and machine parameters. In tunnel excavation projects, there is often a need for accu-
Available online 14 July 2011
rate prediction the performance of such machinery. The poor prediction of machine performance can lead
to very costly contractual claims. In this study, the application of soft computing methods for data anal-
Keywords:
ysis called artificial neural network (ANN) and adaptive neuro-fuzzy inference system (ANFIS) to predict the
Hydraulic impact hammers
Artificial neural networks
net breaking rate of an impact hammer is demonstrated. The prediction capabilities offered by ANN and
Adaptive neuro-fuzzy interference system ANFIS were shown by using field data of obtained from metro tunnel project in Istanbul, Turkey. For this
Schmidt hammer purpose, two prediction models based on ANN and ANFIS were developed and the results obtained from
Rock quality designation those models were then compared to those of multiple regression-based predictions. Various statistical
performance indexes were used to compare the performance of those prediction models. The results sug-
gest that the proposed ANFIS-based prediction model outperforms both ANN model and the classical
multiple regression-based prediction model, and thus can be used to produce a more accurate and
reliable estimate of impact hammer performance from Schmidt hammer rebound hardness (SHRH)
and rock quality designation (RQD) values obtained from the field tests.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction with the advent of more powerful equipment and their relatively
low capital costs, companies in the tunneling business are now
Compared to drilling and blasting method, mechanization of increasingly considering them for excavation. They can be
underground excavation provides an important potential for faster mounted on different types of excavators, and break the rock
development, reduced costs, smaller capital expenditure, improved mainly by impact forces applied by a piston or hammer hitting
strata control and a safer working environment (Robbins, 2000). an adapter at high speeds. The first impact hammers were pneu-
However, despite these advantages offered by mechanization, dril- matic with low impact energy, approximately 1000 J/blow. Today,
ling and blasting has wide application where the excavation however, due to the technological improvements, large hydraulic
lengths are relatively short and/or where the rock formations are hammers that can deliver around 10–20 kJ/blow are available
too hard and abrasive for economical cutting. Any decision to exca- (Bilgin et al., 2002; Tuncdemir, 2008). A detailed survey about
vate by mechanical means or drill and blast should consider factors the operating parameters of hydraulic impact hammers available
such as advance rates, costs, problems of installation, feasibility in the market has been carried out by Tuncdemir (2008). Perhaps
and the ability to negotiate adverse geological conditions (Terezo- due to their presently limited application in tunnel excavation,
poulos, 1987). little quantitative information is available in literature on the
Roadheaders and tunnel boring machines (TBMs) are predomi- performance prediction of hydraulic impact hammers. The perfor-
nantly used for mechanical excavations in tunnel drivage opera- mance prediction of this equipment is generally based on the man-
tions. Apart from roadheaders and TBMs, hydraulic impact ufacturer’s and the contractor’s experience. Recently, however,
hammers are increasingly being employed in tunneling projects. some attempts have been made to correlate intact and mass
In tunnel excavation, hydraulic impact hammers are generally properties of rocks to net breaking rate (NBR) of hydraulic impact
employed to allow economic excavation of relatively fractured or hammers by using classical multiple regression analysis. Using the
altered and geologically disturbed rock formations. They have been data collected from metro tunnels in Istanbul, Bilgin et al. (2002)
in use for about thirty years primarily for demolition. However, stated that Schmidt hammer rebound hardness (SHRH) values gave
statistically significant correlations with the NBR for a Montabert
⇑ Tel.: +90 222 239 37 50; fax: +90 222 239 36 13. BRH 625 type hydraulic hammer, when the rock formations were
E-mail address: miphar@gmail.com grouped based on their rock quality designation (RQD) values.

0886-7798/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tust.2011.06.004
24 M. Iphar / Tunnelling and Underground Space Technology 27 (2012) 23–29

Recently, Kucuk et al. (2011) have suggested a method to predict of inferential statistics. Therefore, it is important that these
the performance of impact hammers based on rock and machine assumptions should be evaluated for each tested analysis. How-
parameters using soft computing techniques such as adaptive neu- ever, in practice, it is extremely time consuming or nearly impos-
ro-fuzzy inference system (ANFIS). Aksoy et al. (2011) developed sible to check assumptions and make individual decisions about
empirical equations to predict the instantaneous breaking rate of how to address potential violations in each case (Wager et al.,
impact hammers considering the data obtained from different 2005). Considering the limitations of the LS method, as an alterna-
locations in various rock masses. Block punch index and geological tive, this paper aims to demonstrate the possible use of data anal-
strength index for rock mass or material properties and power of ysis tools called artificial neural network (ANN) and adaptive neuro-
impact hammer were used as independent variables in the deter- fuzzy inference system (ANFIS) for the performance prediction of
mination of rock mass excavability depending on the excavation hydraulic impact hammers.
type such as slope and tunnel excavation.
In any excavation project, performance prediction of mechani-
2. Artificial neural network analysis
cal excavators is of great concern to contractors from the point
view of planning and cost estimation before the commencement
Artificial neural networks (ANN) represent a technology that is
of the project. Factors affecting the excavation performance of tun-
rooted in many disciplines: neurosciences, mathematics, statistics,
neling machines can be briefly summarized as: (i) intact rock prop-
physics, computer science and engineering (Haykin, 1999). A
erties (UCS, tensile strength, hardness, abrasiveness, elasticity,
broader definition of a practical neural network is that it is a collec-
plasticity, brittleness), (ii) rock mass properties (bedding planes,
tion of interconnected neurons that incrementally learn from their
joints, fractures, fault zones), (iii) machine parameters (available
environment (data) to capture essential linear and nonlinear
power, machine weight, thrust and torque), (iv) cutting mode em-
trends in complex data, so that it provides reliable predictions
ployed (sumping, traversing, undercutting, overcutting), (v) cutting
for new situations containing even noisy and partial information.
tools (drag bits, roller cutters), (vi) cutting geometry (spacing of the
Neurons are the basic computing units that perform local data pro-
cutters, cutting depth, rake angle, angle of attack), (vii) operational
cessing inside a network. These neurons form massively parallel
factors (gradient, water inflow, haulage capacity, support require-
networks, whose function is determined by the network structure
ments), (viii) size and shape of the opening, and (ix) operator skills.
(i.e., how neurons are organized and linked to each other), the con-
As can be seen from the above mentioned information, perfor-
nection strengths between neurons, and the processing performed
mance prediction of mechanical excavators is a highly complicated
at neurons (Samarasinghe, 2007).
task. Nevertheless, a number of empirical performance prediction
The computing process in a neuron is idealized in a model neu-
methods have been developed in the past for hydraulic impact
ron shown in Fig. 1, in which signals are received, accumulated, or
hammers (Bilgin et al., 1996, 1997, 2002; Aksoy, 2009).
summed (R) in the cell body and processed further [f (R)] to pro-
Bilgin proposed the following equation to predict the perfor-
duce an output.
mance of hydraulic impact hammer considering the data gathered
The other neurons that receive this output signal (and the out-
from tunnel drivages of Istanbul metro (Bilgin et al., 1996, 1997):
put signals from other neurons) in turn process the information lo-
IBR ¼ 4:24PðRMCIÞ0:567 ð1Þ cally and pass the output signal to other neurons until the process
is completed and a concept is generated or reviewed, or an action is
 2=3 taken.
RQD The type of neural network used in this study is multi layer per-
RMCI ¼ rc ð2Þ
100 ceptron (MLP) as shown in Fig. 2. The MLP is the most popular and
widely used nonlinear neural network for solving many practical
where IBR is instantaneous or net breaking rate (m3/h), P, cutting
problems in applied sciences such as engineering (Samarasinghe,
power of the hydraulic hammer (HP), RMCI, rock mass cuttability
2007). In Fig. 2, x1, xn are input variables comprising the input
index (MPa), rc, uniaxial compressive strength (MPa), RQD, rock
layer. The first set of arrows represent the weights (or input-hid-
quality designation (%).
den neuron connections) that link this layer to the hidden middle
Aksoy (2009) suggested the following equation to calculate the
layer, consisting of one or many hidden neurons so called because
net excavation (NE, m3/h) of impact hammer by using excavator
they are not exposed to the external environment (data), as are the
power (P, HP) and weak rock breakability index (WRBI):
input and output neurons. Hidden neurons sum the corresponding
NE ¼ 15:423P0:057 WRBI0:229 ð3Þ weighted inputs as denoted by R in Fig. 2. Each hidden neurons
passes its weighted sum through a nonlinear transfer function, de-
  noted by r. The outputs of the hidden neurons are fed through the
WRBI ¼ 265BPI0:25 GSI0:2 =20 ð4Þ
second set of weights (hidden-output neuron connections) into the
output neuron(s), which assemble the outputs by computing the
where BPI is the block punch strength index (MPa), GSI, the geolog-
weighted sum and passing it through a linear or nonlinear func-
ical strength index.
tion. The output of these neurons makes up the network output,
Almost common to most of the empirical approaches men-
which is usually a single output in prediction.
tioned above is the use of classical regression analysis techniques
This structure is trained to learn by repeated exposure to exam-
for establishing performance prediction equations. Regression
ples (input–output data) until the network produces the correct
analysis is an important statistical tool that is routinely applied
output. Learning involves incrementally changing the connection
in most sciences, and out of many possible regression techniques,
the least squares (LS) method has been generally adopted because
of tradition and ease of computation. However, despite to its wide-
spread application, the LS method is based on three fundamental
assumptions that are often disregarded in practise: (a) errors are
Σ f (Σ)
independent from one another, (b) errors are normally distributed,
and (c) error variance is constant across levels of the predicted val-
ues. Violations of these assumptions can produce both false posi-
tive and false negative results and undermine the interpretability Fig. 1. A simple model neuron.
M. Iphar / Tunnelling and Underground Space Technology 27 (2012) 23–29 25

Table 1
1 w01 1 Measured SHRH, RQD and recorded NBR values (Bilgin et al., 2002).

x1 1 Tunnel chainage (m) SHRH RQD (%) Recorded NBR (m3/h)


Σ σ Istanbul metro line 1-Levent
15 + 636 41 66 32.00
x2 15 + 661 58 52 22.85
Σ y 15 + 663 41 59 30.00
15 + 670 48 60 18.20
wn1 15 + 691 51 40 17.70
Σ σ 15 + 762 41 37 22.85
xn wnm m 15 + 769 47 41 26.60
Istanbul metro line 2-Levent
Fig. 2. Configuration of a multilayer neural network. 15 + 687 43 9 17.70
15 + 695 37 20 18.50
15 + 704 58 49 10.60
15 + 706 54 49 10.60
strengths (weights) until the network learns to produce the correct
15 + 709 45 55 22.20
output. The final weights are the optimized parameters of the 15 + 710 50 9 17.70
network. 15 + 722 32 2 20.00
The procedure used to perform the learning process is called the 15 + 736 45 4 16.55
learning algorithm, the function of which is to modify the synaptic 15 + 785 47 32 20.66
15 + 790 49 2 15.00
weights of the network in an orderly fashion to attain a desired de-
sign objective (Haykin, 1999). The Levenberg–Marquardt algo- Istanbul metro line 1-Taksim
15 + 264 51 90 4.27
rithm which appears to be the fastest method for training
15 + 260 61 30 3.94
moderate-sized feedforward neural networks (MathWorks, 2010) 15 + 252 54 66 2.94
was used in this study. 15 + 247 56 54 2.77
In the literature, it was suggested that one hidden layer may be
sufficient for most problems (Caydas and Hascalik, 2008; Lee et al.,
2003; Tiryaki, 2008, 2009; Yilmaz et al., 2002; Yagiz et al., 2009;
Singh, 2004; Suwansawat and Einstein, 2006; Javadi, 2006; Son-
mez et al., 2006; Jonak and Gajewski, 2006; Gajewski and Jonak,
2006). Therefore, one hidden layer was preferred in this study.
However, the number of neurons is the most critical task in the
ANN structure (Sonmez et al., 2006). The heuristics proposed for
this purpose were summarized in the studies of Sonmez et al.
(2006), and Sonmez and Gokceoglu (2008). It is evident that the
performance of ANN architecture increases when the number of
neurons used in hidden layer(s) increases. However, minimum
number of hidden layer(s) and minimum number of neurons as
possible should be selected by considering acceptable performance
criterion (Sonmez and Gokceoglu, 2008). Also, a large number of
hidden layer nodes have large number of associated undetermined
parameters, and if the number of training pairs is small, the net-
work will then tend to memorize rather than generalize (Tiryaki,
2008). So, the number of hidden neurons was determined as 5. In
this way, an ANN model for NBR prediction of hydraulic impact
hammer from SHRH and RQD values was developed by employing
MATLAB software. When using the function ‘‘train’’ in MATLAB
software, by default, the data is randomly divided so that 60% of
the samples are assigned to the training set, 20% to the validation
set, and 20% to the test set (MathWorks, 2010). For training, 80%
out of the raw data in a paper previously published by Bilgin Fig. 3. Predicted NBR values by ANN model versus recorded values.
et al. (2002) were randomly selected and used (Table 1). In that
study, the field data that was collected from metro tunnel drivages
in Istanbul, where a Montabert BRH 625 type impact hammer at- NBR ¼ 0:7482  SRH þ 0:032905  RQD þ 51:4945 ðm3 =hÞ ð5Þ
tached to a hydraulic excavator has been used. The rocks excavated
in that project are mainly fine-grained, laminated, fractured and Using Eq. (5); NBR values were predicted from SHRH and RQD val-
interbedded siltstone, sandstone and mudstone. The Schmidt ham- ues, and plotted against observed NBR values (Fig. 4).
mer (N-type) rebound hardness tests were performed on tunnel
faces by following the test procedure proposed by Poole and Farm- 3. ANFIS method and algorithm
er (1980). The NBR of the impact hammer was recorded for each
experimental tunnel face along with the RQD values, as described ANFIS, being developed by Jang (Jang, 1993), is a soft computing
by Brown (1981). technique which incorporates the concept of fuzzy logic into neu-
Using the ANN model; the NBR values have been predicted by ral networks, and has been widely used in many applications of
using RQD and SHRH values, and plotted against recorded NBR val- engineering science as well as the earth sciences (Grima et al.,
ues for both training and testing data (Fig. 3). 2000; Grima, 2000; Gokceoglu et al., 2004; Singh et al., 2005; Iphar
ANN model then compared with a multiple regression model. et al., 2008; Yilmaz and Yuksek, 2009; Yilmaz, 2009; Kucuk et al.,
The obtained multiple regression equation is as follows: 2011).
26 M. Iphar / Tunnelling and Underground Space Technology 27 (2012) 23–29

Fig. 6. The architecture of ANFIS (Jang et al., 1997).

NBR (Fig. 7). For this purpose, the database with 21 SHRH, RQD
and NBR values given in Table 1were used to construct and train
the fuzzy inference system (FIS) whose membership function
parameters are tuned (adjusted). The training data set (80% of
whole data) was randomly chosen from the data given in Table
1, and the remaining data were used as test data. Several different
Fig. 4. Predicted NBR values by multiple regression analysis versus recorded values. models were constructed to obtain satisfying results by changing
the training options in ANFIS algorithm. The initial fuzzy models
were constructed using the grid partition algorithm. The member-
ANFIS can simulate and analyze the mapping relation between ship functions of the inputs were tuned using the hybrid method,
the input and output data through a hybrid learning to determine which consists of back-propagation for the parameters associated
the optimal distribution of membership function. It is mainly based with the input membership function and the least square estima-
on the fuzzy ‘‘if–then’’ rules from the Takagi and Sugeno type (Jang tion for the parameters associated with the output membership
et al., 1997) as shown in Fig. 5. It involves a premise part and a con- functions. After the training process, 25 ‘‘if–then’’ fuzzy rules were
sequent part. The equivalent ANFIS architecture of the type from obtained in the form of following example:
Takagi and Sugeno is shown in Fig. 6. It comprises five layers in this
inference system and each layer involves several nodes, which are Rule 1: if SHRH is ‘‘Low’’ and RQD is ‘‘Low’’ then NBR is
described by the node function. f1 = p1  SHRH + q1  RQD + r1.
The output signals from nodes in the previous layers will be ac-
cepted as the input signals in the present layer. After manipulation where p1, q1, r1 are the consequent parameters of the first rule in the
by the node function in the present layer, the output will be served rule base.
as input signals for the next layer. Here; square nodes, named Each input variables was represented by five membership func-
adaptive nodes, are adopted to demonstrate that the parameter tions in the constructed ANFIS model, and the type of the member-
sets in these nodes are adjustable. Whereas, circle nodes, named ship functions (Eq. (6)) used for the inputs was the bell
fixed nodes, are adopted to demonstrate that the parameter sets membership function which is reported as a popular method for
are fixed in the system. The ANFIS procedure was described and specifying fuzzy sets because of their smoothness and concise
explained in detail by Jang (1993). notation (Jang, 1993).
In this study, geotechnical and machine performance data col-
lected previously during Istanbul metro tunnel excavation (Bilgin 1
f ðx; a; b; cÞ ¼  2b ð6Þ
et al., 2002) were evaluated by constructing ANFIS-based predic- 1 þ xc
a

tion model. A neuro-fuzzy model whose input parameters are the
SHRH and RQD, and output parameter being the NBR of the where ‘‘a’’, ‘‘b’’ and ‘‘c’’ are the three fitting parameters of the bell
hydraulic impact hammer was developed for the prediction of function. Each of these parameters has a physical meaning: ‘‘a’’ is

Premise Part Consequent Part

µ A1 µ B1
w1
f1=p1x+q1y+r1
w1 f 1 + w2 f 2
f =
X Y w1 + w2
µ A2 µ B2
f2=p2x+q2y+r2 = w 1 f1 + w 2 f 2
w2

X Y
x y

Fig. 5. A two-input first-order Sugeno fuzzy model with two rules (Jang et al., 1997).
M. Iphar / Tunnelling and Underground Space Technology 27 (2012) 23–29 27

inputmf rule outputmf Output

. .
. .
. .
SHRH . .
. .
. .
. .
. .
. .
. .
. .
. . NBR
. .
. .
. .
. .
. .
. .
. .
. .
RQD . .

Fig. 7. ANFIS architecture of the constructed fuzzy model.

the half width, ‘‘b’’ (together with ‘‘a’’) controls the slopes at the 4. Model results and performances
crossover points where membership function value is 0.5 and ‘‘c’’
determines the center of the corresponding membership function. For the performance prediction of hydraulic impact hammer,
The input membership functions in the constructed ANFIS model the NBR values were predicted from SHRH and RQD by using pres-
after training are given in Fig. 8. These membership functions en- ently developed ANN, ANFIS models and also by multiple regres-
able user to express input variables (SHRH and RQD) in linguisti- sion equation (Eq. (5)), and the obtained results were given in
cally, such as ‘‘Very Low’’, ‘‘Low’’, ‘‘Moderate’’, ‘‘High’’ and ‘‘Very High’’. Table 2. Also; in Figs. 3, 4 and 9, plots of the predicted NBR values
The results obtained with this data analysis technique were also by each method versus the field recorded values are illustrated, for
compared to those of classical least squares method currently used multiple regression analysis, ANN and ANFIS models, respectively.
in practice and those of ANN based prediction model. In order to check and compare the prediction performances of
the ANN, ANFIS and multiple regression based models, the variance
account for VAF (Eq. (7)) and the root mean square error RMSE (Eq.
(a) 1
VL L M H VH (8)) performance indexes were used:

0.8
Membership Degree

 
ðv ar ðyi  y^ i ÞÞ
VAF ¼ 1  100% ð7Þ
0.6 v arðyi Þ

0.4
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 XN
0.2 RMSE ¼ ðy  y ^i Þ2 ð8Þ
N i¼1 i

35 40 45 50 55 60 ^i is the
where var symbolizes the variance, yi is the measured value, y
SHRH predicted value, and N is the number of samples.
The interpretation of the above performance indexes are as fol-
lows: the higher the VAF, the better the model performs. For exam-
(b) 1
VL L M H VH
ple, a VAF of 100% means that the measured output has been
predicted exactly. VAF = 0 means that the model performs as
0.8 poorly as a predictor using simply the mean value of the data.
Membership Degree

The lower the RMSE, the better the model performs (Grima and
0.6 Babuska, 1999; Gokceoglu, 2002; Gokceoglu and Zorlu, 2004). Con-
trary to VAF, RMSE also accounts for a bias in the model, i.e. an off-
0.4
set between the measured and predicted data.
When the VAF and RMSE performance indexes are considered
for each predictive model (Table 3), it’s clear that the developed
0.2
ANFIS model gives a better prediction performance than that of
the ANN and classical multiple regression model. Also, the ANFIS
0 model has a lower standard error of estimate and a higher correla-
tion coefficient (r) than those of the ANN and classical multiple
10 20 30 40 50 60 70 80 90
regression model. Therefore, it can be said that ANFIS is the best
RQD
prediction tool for the estimation of net breaking rate of hydraulic
Fig. 8. Input membership functions for ANFIS model (a) Input 1: SHRH (b) Input 2: impact hammers from Schmidt hammer rebound hardness and
RQD. rock quality designation values for this study.
28 M. Iphar / Tunnelling and Underground Space Technology 27 (2012) 23–29

Table 2
Predicted NBR values by multiple regression, ANN and ANFIS-based models.

Tunnel chainage (m) Recorded NBR Predicted NBR by the Predicted NBR by Predicted NBR by
(m3/h) regression model ANN-based model ANFIS-based model
(Eq. (5)) (m3/h) (m3/h) (m3/h)
Istanbul metro line 1-Levent
15 + 636 32.00 22.99 31.97 32.00
15 + 661 22.85 9.81 8.79 22.85
15 + 663 30.00 22.76 28.38 29.45
15 + 670 18.20 17.56 16.44 18.20
15 + 691 17.70 14.65 18.95 17.70
15 + 762 22.85 22.04 18.46 18.77
15 + 769 26.60 17.68 27.27 26.60
Istanbul metro line 2-Levent
15 + 687 17.70 19.62 17.77 17.70
15 + 695 18.50 24.47 17.29 10.33
15 + 704 10.60 9.71 8.82 10.60
15 + 706 10.60 12.70 9.95 10.60
15 + 709 22.20 19.64 29.21 29.22
15 + 710 17.70 14.38 18.83 17.70
15 + 722 20.00 27.62 20.03 20.00
15 + 736 16.55 17.96 17.74 9.95
15 + 785 20.66 17.38 24.16 20.66
15 + 790 15.00 14.90 16.21 15.00
Istanbul metro line 1-Taksim
15 + 264 4.27 16.30 6.84 4.27
15 + 260 3.94 6.84 7.69 3.94
15 + 252 2.94 13.26 8.65 2.94
15 + 247 2.77 11.37 6.64 2.77

5. Conclusions

In metro tunneling projects, the accurate prediction of excava-


tion performance of mechanical excavators is of great concern to
contractors with respect to both contractual claims arising from
poor equipment selection, and to project budgetary and time lim-
itations. Today, as an alternative to drilling and blasting, mechan-
ical excavation is increasingly being employed to raise production
rate and minimize costs. Hence, there is a continuous demand in
the industry to develop accurate means of machine performance
prediction.
In this study; ANN, ANFIS and multiple regression models were
developed using SHRH and RQD for the prediction of the perfor-
mance of a hydraulic impact hammer. The proposed ANFIS-based
model has performed best when VAF (88.14), RMSE (2.90), standard
error of estimation (2.83) and correlation coefficient (0.95) are con-
sidered. However, the proposed ANN and ANFIS models can be im-
proved by adding machine related parameters of hydraulic impact
hammer as new inputs to the models. Besides, some other poten-
tially important variables such as the orientation of the hammer
which may be defined as linguistically can also be included in the
predictive models. This parameter may affect the net breaking rate
of the hydraulic impact hammer since free face plays an important
role in favor of net breaking rate as stated by Tuncdemir (2008). So,
Fig. 9. Predicted NBR values by ANFIS model versus recorded values. for further studies, the linguistic definitions such as hammer orien-
tation can easily be considered in a fuzzy model since fuzzy set the-
ory can cope with the linguistic terms and complex systems.

Table 3 References
Results of the statistical performance analysis.

Model VAF RMSE Standard error of Correlation Aksoy, C.O., 2009. Performance prediction of impact hammers by block punch index
(%) estimation coefficient, r for weak rock masses. International Journal of Rock Mechanics and Mining
Sciences 46 (8), 1383–1388.
Multiple 39.75 6.40 6.73 0.63 Aksoy, C.O., Ozacar, V., Demirel, N., Ozer, S.C., Safak, S., 2011. Determination of
regression instantaneous breaking rate by geological strength index, block punch index
model and power of impact hammer for various rock mass conditions. Tunnelling and
ANN based model 74.68 4.16 4.33 0.87 Underground Space Technology 26 (4), 534–540.
ANFIS based 88.14 2.90 2.83 0.95 Bilgin, N., Yazici, S., Eskikaya, S., 1996. A model to predict the performance of
model roadheaders and impact hammers in tunnel drivages. In: Proceedings of the
Eurock ’96, Balkema, pp. 715–720.
M. Iphar / Tunnelling and Underground Space Technology 27 (2012) 23–29 29

Bilgin, N., Kuzu, C., Eskikaya, S., 1997. Cutting performance of rock hammers and Mathworks, 2010. Matlab Manual, 1984–2010. The MathWorks, Inc.
roadheaders in Istanbul metro drivages. In: Proceedings of the Word Tunnel Poole, R.W., Farmer, I.W., 1980. Consistency and repeatability of Schmidt hammer
Congress’97, Tunnels for People, Balkema, pp. 455–460. rebound data during field-testing. International Journal of Rock Mechanics,
Bilgin, N., Dincer, T., Copur, H., 2002. The performance prediction of impact Mining Sciences and Geomechanical Abstracts 17, 167–171.
hammers from Schmidt hammer rebound values in Istanbul metro tunnel Robbins, R.J., 2000. Mechanization of underground mining: a quick look backward
drivages. Tunnelling and Underground Space Technology 17, 237–247. and forward. International Journal of Rock Mechanics and Mining Sciences 37,
Brown, E.T., 1981. Rock Characterization Testing and Monitoring. Pergamon Press, p. 413–421.
211. Samarasinghe, S., 2007. Neural Networks for Applied Sciences and Engineering:
Caydas, U., Hascalik, A., 2008. A study on surface roughness in abrasive waterjet From Fundamentals to Complex Pattern Recognition. Auerbach Publications
machining process using artificial neural networks and regression analysis Taylor & Francis Group, p. 570.
method. Journal of Materials Processing Technology 202, 574–582. Singh, T.N., 2004. Artificial neural network approach for prediction and control of
Gajewski, J., Jonak, J., 2006. Utilisation of neural networks to identify the status of ground vibrations in mines. Mining Technology (Transactions of the Institution
the cutting tool point. Tunnelling and Underground Space Technology 21, 180– of Mining and Metallurgy Section A) 113, A251–A256.
184. Singh, T.N., Verma, A.K., Singh, V., Sahu, A., 2005. Slake durability study of shaly rock
Gokceoglu, C., 2002. A fuzzy triangular chart to predict the uniaxial compressive and its predictions. Environmental Geology 47, 246–253.
strength of Ankara agglomerates from their petrographic composition. Sonmez, H., Gokceoglu, C., 2008. Discussion on the paper by H. Gullu and E. Ercelebi
Engineering Geology 66, 39–51. ‘‘A neural network approach for attenuation relationships: An application using
Gokceoglu, C., Zorlu, K., 2004. A fuzzy model to predict the uniaxial compressive strong ground motion data from Turkey (in press)’’. Engineering Geology 97,
strength and the modulus of elasticity of a problematic rock. Engineering 91–93.
Applications of Artificial Intelligence 17 (1), 61–72. Sonmez, H., Gokceoglu, C., Nefeslioglu, H.A., Kayabasi, A., 2006. Estimation of rock
Gokceoglu, C., Yesilnacar, E., Sonmez, H., Kayabasi, A., 2004. A neuro-fuzzy model modulus: For intact rocks with an artificial neural network and for rock masses
for modulus of deformation of jointed rock masses. Computers and Geotechnics with a new empirical equation. International Journal of Rock Mechanics and
31 (5), 375–383. Mining Sciences 43, 224–235.
Grima, M.A., 2000. Neuro-Fuzzy Modelling in Engineering Geology. A.A. Balkema, Suwansawat, S., Einstein, H.H., 2006. Artificial neural networks for predicting the
Rotterdam, p. 244. maximum surface settlement caused by EPB shield tunneling. Tunnelling and
Grima, M.A., Babuska, R., 1999. Fuzzy model for the prediction of unconfined Underground Space Technology 21, 133–150.
compressive strength of rock samples. International Journal of Rock Mechanics Terezopoulos, N.G., 1987. Influence of geotechnical environments on mine tunnel
and Mining Science 36 (3), 339–349. drivage performance. In: Szwilski, A.B., Richards, M.J. (Eds.), Proceedings,
Grima, M.A., Bruines, P.A., Verhoef, P.N.W., 2000. Modeling tunnel boring machine Underground Mining Methods and Technology. Elsevier, Amsterdam, pp.
performance by neuro-fuzzy methods. Tunnelling and Underground Space 139–155.
Technology 15 (3), 260–269. Tiryaki, B., 2008. Application of artificial neural networks for predicting the
Haykin, S., 1999. Neural Networks: A Comprehensive Foundation. Prentice Hall cuttability of rocks by drag tools. Tunnelling and Underground Space
International, Inc., p. 841. Technology 23, 273–280.
Iphar, M., Yavuz, M., Ak, H., 2008. Prediction of ground vibrations resulting from the Tiryaki, B., 2009. Estimating rock cuttability using regression trees and artificial
blasting operations in an open-pit mine by adaptive neuro-fuzzy inference neural networks. Rock Mechanics and Rock Engineering 42, 939–946.
system. Environmental Geology 56, 97–107. Tuncdemir, H., 2008. Impact hammer applications in Istanbul metro tunnels.
Jang, R.J.S., 1993. Anfis: adaptive-network-based fuzzy inference system. IEEE Tunnelling and Underground Space Technology 23, 264–272.
Transactions on Systems Man and Cybernetics 23, 665–685. Wager, T.D., Keller, M.C., Lacey, S.C., Jonides, J., 2005. Increased sensitivity
Jang, R.J.S., Sun, C.T., Mizutani, E., 1997. Neuro-Fuzzy and Soft Computing. Prentice- nueroimaging analyses using robust regression. NeuroImage 26, 99–113.
Hall, Inc., p. 614. Yagiz, S., Gokceoglu, C., Sezer, E., Iplikci, S., 2009. Application of two non-linear
Javadi, A.A., 2006. Estimation of air losses in compressed air tunneling using neural prediction tools to the estimation of tunnel boring machine performance.
network. Tunnelling and Underground Space Technology 21, 9–20. Engineering Applications of Artificial Intelligence 22, 808–814.
Jonak, J., Gajewski, J., 2006. Identifying the cutting tool type used in excavations Yilmaz, I., 2009. Landslide susceptibility mapping using frequency ratio, logistic
using neural networks. Tunnelling and Underground Space Technology 21, regression, artificial neural networks and their comparison: a case study from
185–189. Kat landslides (Tokat–Turkey). Computers and Geosciences 35, 1125–1138.
Kucuk, K., Aksoy, C.O., Basarir, H., Onargan, T., Genis, M., Ozacar, V., 2011. Prediction Yilmaz, I., Yuksek, G., 2009. Prediction of the strength and elasticity modulus of
of the performance of impact hammer by adaptive neuro-fuzzy inference gypsum using multiple regression, ANN, and ANFIS models. International
system modeling. Tunnelling and Underground Space Technology 26, 38–45. Journal of Rock Mechanics and Mining Sciences 46, 803–810.
Lee, S.J., Lee, S.R., Kim, Y.S., 2003. An approach to estimate unsaturated shear Yilmaz, S., Demircioglu, C., Akin, S., 2002. Application of artificial neural networks to
strength using artificial neural network and hyperbolic formulation. Computers optimum bit selection. Computers and Geosciences 28, 261–269.
and Geotechnics 30, 489–503.

You might also like