Вы находитесь на странице: 1из 5

Rock Engineering and Rock Mechanics: Structures in and on

Rock Masses Alejano, Perucho, Olalla & Jimnez (Eds)

2014 Taylor & Francis Group, London, 978-1-138-00149-7

Artificial Neural Network (ANN) based model for predicting

of overall strength of Volcanic Bimrock
H. Sonmez, A. Coskun, M. Ercanoglu, D. Turer & K.E. Kasapoglu
Department of Geological Engineering, Applied Geology Division, Hacettepe University, Beytepe, Ankara, Turkey

C. Tunusluoglu
Department of Geological Engineering, Applied Geology Division, Canakkale Onsekizmart Universitesi,
anakkale, Turkey

ABSTRACT: The uniaxial compressive strength of rock material (UCS) is one of the fundamental input
parameters for engineering applications to be constructed on/in rock masses such as deep slopes, tunnels and
dams. However, preparation of the high quality cores for laboratory studies is generally difficult for some types of
rock such as laminated and/or fragmented rock material. To overcome this difficulty empirical prediction models
were developed by considering some input parameters. Geological mixtures composed of rock blocks surrounded
by weak matrix material are known as Block-In-Matrix-Rock (Bimrock) in literature. Agglomerate is a special
type of Bimrock, which is composed of andesite fragments surrounded by tuff matrix and it is an example of
Volcanic Bimrock. Preparation of core samples for experimental studies from agglomerate is problematic due
to the strength contrast between andesite rock fragments and tuff matrix. To overcome these difficulties, some
prediction tools have been studied by regression analyses in the literature. In this study, Artificial Neural Network
(ANN) as a prediction tool was used to construct a model for prediction of overall UCS of Volcanic Bimrock.
While Volumetric Block Proportion (VBP), Volumetric Block Count (VBC) and fractal dimensions (1 and 2
dimensional) were selected as input parameters, normalized overall uniaxal strength of agglomerate to uniaxal
compressive strength of tuff matrix is output parameter. Fractal geometry has been used as popular method
to define irregular shapes as a quantity in literature. The boundary strength between an-desite fragments and
tuff matrix is also sensitive to fragment shape and surface roughness of andesite fragments. Therefore fractal
dimensions were selected as input parameters to incorporate this effect on boundary strength. While previously
developed computer code FRACRUN was used to determine average fractal dimension of andesite fragments in
agglomerate cores, previously developed computer code ANNES was used for ANN based model construction.
In addition, similar to Volumetric Joint Count (Jv) which is widely used in rock mass characterization, Volumetric
Block Count (VBC) was defined as another input parameter for determination of Bimrock UCS considering
some of studies about performed in literature. The highest prediction performance was obtained from the model
which considers Volumetric Block Proportion (VBP), Volumetric Block Count (VBC) and 1D fractal dimension
as inputs.


Preparation of cores from Bimrocks is extraordinarily

difficult due the strength contrast between blocks and
weak matrix. Bimrocks are divided mainly into two
subgroups namely welded and unwelded Bimrocks
(Altinsoy 2006, Sonmez & Tunusluoglu 2010). While
the strength between matrix and blocks is almost
equal to the strength of matrix for welded Bimrocks,
the strength between matrix and blocks is less than
strength of matrix for unwelded Bimrocks. In this
study, Ankara agglomerate which is a kind of welded
Volcanic Bimrock was considered as the study rock
material (Fig. 1). While Volumetric Block Proportion (VBP), Volumetric Block Count (VBC), fractal
dimensions (1 and 2 dimensional) were selected as
input parameters, normalized overall uniaxal strength
of agglomerate to uniaxal compressive strength of tuff

Uniaxial compressive strength of rock material is a

crucial input parameter for engineering designs to
be constructed in/on rock masses. The uniaxial compressive strength of rock material is determined by
conventional laboratory tests employed on high quality core samples. However, for some rocks such as
laminated and fragmented rock materials, preparation of high quality cores is almost impossible. To
overcome this difficulty some empirical prediction
models for Ankara agglomerate studied in literature
(Gokceoglu 2002, Sonmez et al. 2004, Sonmez et al
2006a) Mixtures of rocks composed of geotechnically
significant strong blocks, within a bonded weak matrix
of finer texture defined as Bimrock by Medley (1994).


Figure 1. View from Ankara Agglomerate (Sonmez &

Tunuslugolu 2010).

matrix is obtained as the output parameter. The artificial neural network (ANN) was used as a prediction
tool for construction of the models. Prediction performances of the generated models were compared in
terms of fractal dimensions (1 and 2 dimensional).



In this study, the database established by Sonmez &

Tunusluoglu (2010) was considered. Establishment of
database was performed on total of 70 core specimen
having 6 cm diameter and height between 2 to 2.5. The
top and the bottom circular surfaces and rectangular
cylindrical side of the samples were scanned for image
analysis. Then unit weight () and uniaxial compressive strength (UCS) of each core were determined by
laboratory tests as suggested by ISRM (2007).
As it can be followed from the literature, fractal
dimension (D) is a statistical quantity that gives an
indication of how completely a fractal appears to fill
space, as one zooms down to finer and finer scales (Das
2011). In literature, fractal geometry has been used as
popular tool to define particularly irregular shapes by
the theory of fractal developed by Mandelbrot (1967)
(Hyslip & Vallejo 1997, Kruhl & Nega 1996, Bagde
et al. 2002, Gulbin & Evangulova 2003, Pardini 2003,
Kolay & Kayabali 2006, Hamdi 2008, Zorlu 2009,
Sezer 2009). In this study, square gridcell (box) count
method for 2D and segment (line) count method for
1D were followed in the algorithm of FRACRUN.
FRACRUN has the capability of determining fractal dimensions of many closed polygons on a single
image, with a click on the start button. The calculations
in 1D and 2D procedure for a single shape followed
by FRACRUN are summarized in Figure 2.
While, the relation between overall strength of
Bimrock and volumetric block proportion (VBP) has
been widely investigated in the literature, some studies
were focused on the characterization of Bimrocks such
as determination of volumetric block proportion, block
size distribution and their quantification (Lindquist

Figure 2. Schematically illustration for determination of 2D

(box-count) and 1D (segment-line count) fractal dimensions
of a single block (Sonmez & Tunusoglu 2010).

1994, Lindquist & Goodman 1994, Medley 2002). In

this study, by considering the studies performed on
the quantification of blocks in Bimrocks, volumetric
block count (VBC) was defined as additional input
parameter together with VBP. Because the amount of
the weakest component may depends not only on VBP
but also on VBC. In counting of the number of block in
a volume, the engineering volume should be defined.
For this study, engineering volume was considered as
the volume of cores. Therefore the number of block
counted for each core was used as VBC in number of
block/cubic unit. The VBC values may include some
errors due to counting of blocks on 2D scanned images.
Considering the simple regression relations between
UCS and input parameters, two multi inputs ANN
models were investigated. While VBP, VBC and 1D
were used as multi input parameters for the first model,


Figure 4. The ANN architecture considered in this study.

Figure 3. A criteria for termination of training and selection

of optimum network architecture (Basheer & Hajmeer, 2000).

2D were considered as input parameters instead of 1D

for the second model.
ANN has been one of the attractive prediction
tool used in geo-engineering applications due to the
high performance on the modeling of nonlinear multivariate problems (Sonmez et al. 2006b). The backpropagation artificial neural network among the other
methods has been widely considered for empirical
predication models in geo-engineering applications
(Goh et al. 1995, Shi et al. 1998, Neaupane & Achet
2004, Lee et al. 2003, Gomez & Kavzoglu 2005,
Ermini et al. 2005, Yesilnacar & Topal 2005, Sonmez
et al. 2006b).
Although the complexity of BMNN architecture
increases by addition of hidden layers, the simplest
structure having sufficient and applicable prediction
capacity is preferred to avoid overlearning. Each layer
including input(s) and output(s) layers consist of neurons (nodes), and the neurons are joined by weighted
links. The final weights and thresholds of activation
for decreasing the error between observed and computed outputs under a sufficient level defined by user
is set by training phase of ANN algorithm (Sonmez
et al. 2006b). The sigmoid transfer function is considered as an activation function in this study, because it
is commonly preferred in application of predication
purposes. The maximum number of training cycles
(epoch) for back and forward stages is limited by
10000, together with the minimum threshold for root
mean square error (RMSE) as 0.001.
Another important point of the ANN application
is avoiding of overlearning, otherwise ANN may
lose its generalization capacity (Fig. 3). While the
learning rate was used as 0.1, the momentum coefficient was set to 0.95, considering recommendation
and application used in literature. For solution of
the most problems by ANN one hidden layer may
be sufficient (Rumelhart et al. 1986, Hect-Neilsen
1987, Lippmann 1987, Basheer 2000). By considering

Figure 5. The graph of root mean square error (RMSE)

versus number of cycles (epochs) for two models.

the recommendation mentioned in the literature, one

hidden layer was preferred in this study.
Overlearning problem may be observed when large
number of neurons preferred in hidden layers. For this
propose some recommendation is available in literature (Hecht-Nielsen, 1987, Hush 1989, Ripley 1993,
Wang 1994, Masters 1994, Kaastra & Boyd 1996,
Kannellopoulas & Wilkinson 1997). In this study,
three neurons were used by considering number of
inputs and output (Fig. 4).
Whole data in the database were normalized
between 0 and 1. Then the database composed of 70
data set was divided into 55 data set for training (80%
of data) and 15 data set for testing (20% of data set).
After ANNES was run for both first and second models, the graph for number of training cycle versus root
mean square error for both training and testing data
sets were drawn as in Figure 5.


Preparation of cores for laboratory test from Bimrocks

is extraordinarily difficult due the strength contrast
between blocks and weak matrix. The artificial neural
network (ANN) was used for construction of the UCS
prediction models. The main important findings of the
study can be summarized as follows.
The use of Volumetric Block Count (VBC) together
with Volumetric Block Proportion (VBP) exhibits high
importance on the prediction of overall strength of
1D (line) fractal dimension seems to be more sensitive than 2D fractal as input parameter on prediction of
overall strength of Bimrock. Because 1D (line) fractal
dimension may define roughness of andesite blocks
(grain) better than 2D fractal dimension. Similar to
the study carried by Kolay & Kayabali (2006), higher
1D fractal dimension is obtained for rougher grains.
The contact strength between matrix and blocks may
be expected to be higher with the increase of surface
roughness of the blocks as emphasized in literature.

The database used in this study was established during
TUBITAK Project (The Scientific and Technological
Research Council of Turkey, Project No: 108Y002).
Altinsoy H. 2006. Matriks iinde blok ieren kayalarn
makaslama dayanmn belirlenmesi iin fiziksel model
esasl bir arastrma. MSc thesis, Hacettepe University (in
Aycan C., Sonmez H., Kasapoglu K.E., Dinc O. & Tunusluoglu C., 2010. Determination of Uniaxial Compressive
Strength of Ankara Agglomerate Considering Fractal
Geometry of Blocks. Geophysical Research Abstracts,
Vol. 12, EGU2010-5899, EGU General Assembly 2010.
Bagde M.N., Raina A.K., Chakraborty A.K. & Jethwa J.L.,
2002. Rock mass characterization by fractal dimension.
Engineering Geology 63, 141155.
Baheer I., 2000. Selection of methodology for modeling hysteresis behavior of soils using neural networks. J Comput
aided Civil Infrastruct Engg 2000; 5(6): 445463.
Basheer I.A. & Hajmeer, M., 2000. Artificial neural networks: fundamentals, computing, design, and application.
J of Microbiological Methods 2000; 43: 331
Das S. 2011. Functional Fractional Calculus. 2nd edition.
Springer-Verlag Berlin Heidelberg.
Ermini L. Catani, F. & Casagli N. 2005. Artificial Neural
Networks applied to landslide susceptibility assessment.
Geomorphology 2005; 66: 327343.
Goh A.T.C., Wong, K.S. & Broms, B.B. 1995. Estimation of
lateral wall movements in braced excavations using neural
networks. Can Geotech J 1995; 32: 10591064.
Gokceoglu C. 2002. A fuzzy triangular chart to predict the
uniaxial compressive strength of the Ankara agglomerates from their petrographic composition. Engineering
Geology, 66 (12), 3951.
Gomez H. & Kavzoglu T. 2005. Assessment of shallow landslide susceptibility using artificial neural networks in
Jabonosa River Basin, Venezuela. Engineering Geology
2005; 78: 1127.

Figure 6. Cross correlation between predicted and measured UCS values of Bimrock cores for two models.

RMSE values of testing data tends to increase after

2100 cycles (epoch) and 1900 for the first (inputs:
VBC, VBP, 1D) and the second (inputs: VBC, VBP,
2D), respectively. Therefore, above these thresholds,
generalization of the models decreases. By using the
weights and thresholds of activation at 2100th and
1900th epochs for the two models predicted values of
training and testing data were generated by ANNES.
Then comparison of the two models was made by cross
correlation graph (Figure 6).
As it can be followed from Figure 6, the first model
which consider 1D (segment-line) fractal dimension
exhibit slightly better prediction capacity than the second model which considers 2D (box-counting) fractal
dimension. Because 1D fractal dimension have better
capability for defining roughness of the blocks.


In this study, Ankara agglomerate as a kind of volcanic bimrock was selected for the study material.


Gulbin Y.L. & Evangulova E.B. 2003. Morphometry of

quartz aggregates in granites: fractal images referring to
nucleation and growth processes. Mathematical Geology
35 (7), 819833
Hamdi E. 2008. A fractal description of simulated 3D discontinuity networks. Rock Mechanics and Rock Engineering
41, 587599.
Hecht-Nielsen R. 1987. Kolmogorovs mapping neural network existence theorem. Poc of the First IEEE International Conference on Neural Networks, San Diego CA,
USA 1987; 1114.
Hush D.R. 1989. Classification with neural networks: a performance analysis. Proc. Of the IEEE Int Conf on systems
Engg, Dayton Ohia USA 1989; 277-280.
Hyslip J.P. & Vallejo, L.E., 1997. Fractals analysis of the
roughness and size distribution of granular materials.
Engineering Geology 48, 231244.
ISRM 2007. Ulusay, R.; Hudson, J.A., eds. The Blue Book
The Complete ISRM Suggested Methods for Rock Characterization, Testing and Monitoring: 19742006. Ankara:
ISRM & ISRM Turkish National Group. p. 628. ISBN
Kaastra I. & Boyd M. 1996. Designing a neural network for
forecasting financial and economic time series, Neurocomputing 1996; 10(3): 215236.
Kanellopoulas I. & Wilkinson G.G. 1997. Strategies and best
practice for neural network image classification, International Journal of Remote Sensing 1997; 18: 711725.
Kolay E. & Kayabali K. 2006. Investigation of the effect
of aggregate shape and surface roughness on the slake
durability index using the fractal dimension approach.
Engineering Geology 86, 271294.
Kruhl J.H. & Nega M. 1996. The fractal shape of sutured
quartz grain boundaries: application as a geothermometer.
Geologische Rundschau 85, 3843.
Lee S., Ryu J.H., Lee M.J. & Won J.S. 2003. Use of an artificial neural network for analysis of the susceptibility to
landslides at Boun, Korea. Environmental Geology 2003;
44: 820-833.
Lindquist E. S. 1994. The Strength and Deformation Properties of Melange, Ph.D. Dissertation, Department of Civil
Engineering, University of California at Berkeley.
Lindquist E.S. & Goodman R.E. 1994. Strength and deformation properties of a physical model mlange, Proceedings,
1st North American Rock Mechanics Symposium, Austin,
TX, May 1994.
Lippmann R.P. 1987. An introduction to computing with
neural nets. IEEE ASSP Mag 1987; 4: 422.
Mandelbrot B. 1967. How Long Is the Coast of Britain?
Statistical Self-Similarity and Fractional Dimension,
Science, New Series, Vol. 156, No. 3775. (May 5, 1967),
pp. 636638.
Masters T. 1994. Practical Neural Network Recipes in C++.
Academic Press, Boston MA 1994.
Medley E.W. (1994) Engineering characterization of
mlanges and similar block-in-matrix rocks (bimrocks) PhD Dissertation, Dept. Civil Engineering, Univ.
California at Berkeley.

Medley E. 2002. Estimating block size distributions of

mlanges and similar block-in-matrix rocks (bimrocks),
Proceedings, 5th North American Rock Mechanics
Symposium, University of Toronto, Toronto, Canada,
July 2002, pp. 599606.
Neaupane K.M. & Achet S.H. 2004. Use of backpropagation
neural network for landslide monitoring: a case study in
the higher Himalaya. Engineering Geology 2004; 74: 213
Pardini G. 2003. Fractal scaling of surface roughness in artificially weathered smectite rich soil regoliths. Geoderma
117, 157167.
Ripley B.D. 1993. Statistical aspects of neural networks. In:
Barndoff-Neilsen OE, Jensen JL, Kendall WS (Eds.), Networks and Chaos-Statistical and Probabilistic Aspects,
Chapman & Hall, London 1993; 40123.
Rumelhart D.E. & Hinton G.E. 1986. Williams RJ. Learning internal representation by error propogation. In:
Rumelhart DE, McClelland JL (Eds.), Parallel Distributed
Processing 1986; 1: 318362.
Sezer E. 2009. A computer program for fractal dimension (FRACEK) with application on type of mass
movement characterization. Computers and Geosciences
Shi J., Ortiago J.A.R. & Bai, J. 1998. Modular neural networks for predicting settlements during tunnelling. J
Geotech Geo-Env Engg ASCE 1998; 124(5): 389394
Sonmez H. & Tunusluoglu C. 2010. Eklemli Kaya
Ktleleri ve BIMROCKlar iin Birlestirilmis Jeomekanik
Snflama Siteminin ve Genellestirilmis Ampirik
Yaklasmn Gelistirilmesi. TUBITAK Proje No. 108Y002.
(in Turksih)
Sonmez H., Tuncay E. & Gokceoglu C. 2004. Models to predict the uniaxial compressive strength and the modulus
of elasticity for Ankara Agglomerate. Int. J. Rock Mech.
Min. Sci., 41 (5), 717729.
Sonmez H., Gokceoglu C., Medley E.W., Tuncay E. &
Nefeslioglu H.A. 2006a. Estimating the uniaxial
compressive strength of a volcanic bimrock. Int. J. Rock
Mech. Min. Sci., 43 (4), 554561.
Sonmez H., Gokceoglu C., Kayabasi A. & Nefeslioglu, H.A.
2006b. Estimation of rock modulus: for intact rocks with
an artificial neural network and for rock masses with a
new empirical equation, Int. J. Rock Mech. Min. Sci., 2006
43(2), 224235.
Wang C. 1994. A theory of generalization in learning
machines with neural application. PhD Thesis, The Univ
of Pennsylvania, USA 1994.
Yesilnacar E.K. & Topal T. 2005. Landslide Susceptibility Mapping: comparison between logistic regression
and neural networks in a medium scale study, Hendek
region TURKEY), No. 2512. Engineering Geology,
79, (2005), pp. 251266.
Zorlu K. 2008. Description of the weathering states of building stones by fractal geometry and fuzzy inference system
in the Olba ancient city (Southern Turkey). Engineering
Geology 101 (2008), 124133.