You are on page 1of 8

Advances in Engineering Software 40 (2009) 920–927

Contents lists available at ScienceDirect

Advances in Engineering Software


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

Predicting the compressive strength of mortars containing metakaolin


by artificial neural networks and fuzzy logic
Mustafa Sarıdemir *
Department of Civil Engineering, Niğde University, 51100 Niğde, Turkey

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

Article history: Artificial neural networks and fuzzy logic approaches have recently been used to model some of the
Received 23 September 2008 human activities in many areas of civil engineering applications. Especially from these systems in the
Received in revised form 24 November 2008 model experimental studies, very good results have been obtained. In this research, the models for pre-
Accepted 1 December 2008
dicting compressive strength of mortars containing metakaolin at the age of 3, 7, 28, 60 and 90 days have
Available online 22 January 2009
been developed in artificial neural networks and fuzzy logic. For purpose of building these models, train-
ing and testing using the available experimental results for 179 specimens produced with 46 different
Keywords:
mixture proportions were gathered from the technical literature. The data used in the multilayer feed-
Mortar
Metakaolin
forward neural networks and Sugeno-type fuzzy inference models are arranged in a format of five input
Compressive strength parameters that cover the age of specimen, metakaolin replacement ratio, water–binder ratio, superplast-
Neural network icizer and binder–sand ratio. According to these input parameters, in the multilayer feed-forward neural
Fuzzy logic networks and Sugeno-type fuzzy inference models, the compressive strength of mortars containing
metakaolin are predicted. The training and testing results in the multilayer feed-forward neural networks
and Sugeno-type fuzzy inference models have shown that neural networks and fuzzy logic systems have
strong potential for predicting compressive strength of mortars containing metakaolin.
Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction strength enhancement is due to a combination of the filler effect


and accelerated cement hydration. Subsequently, these effects
It is well known that pozzolana such as silica fume, fly ash and are enhanced by the pozzolanic reaction between MK and Ca(OH)2
granulated blast furnace slag are industrial by-products that pres- produced by the hydration cement [2]. Also, the replacement of
ent great variations of properties, which can sometimes compro- part of the cement by MK (up to 15%) reduces drying shrinkage
mise their use [1]. On the other hand, the utilization of calcined in comparison with concrete without MK [6,7].
clay in the form of metakaolin as a pozzolanic addition for mortar Several researches indicate that the presence of MK in mortar
and concrete has received considerable interest in recent years [2]. and concrete seems to increase the compressive strength (fc) as
These pozzolanic admixtures are used for reducing the cement compared to that of conventional mortar and concrete. These re-
content in mortar and concrete production [3]. Also, the use of poz- searches have demonstrated clearly that with intelligent use con-
zolanic materials such as silica fume, fly ash and metakaolin is nec- siderable improvement in fc, especially at the early stages of
essary for producing high performance concrete. These materials, curing, can be carried out. The increase of fc of mortar and concrete
when used as mineral admixtures in high performance concrete, with MK accounts for the increasing consumption of this admix-
can improve either or both the strength and durability properties ture in mortar and concrete. In the study of Vu et al. [3], the two
of the concrete [4,5]. groups of specimen series with 36 different mixes have been ob-
Metakaolin (MK) is a valuable pozzolanic material, it is a ther- tained. It has been designed that the binder–sand ratio are 1:2.75
mally activated alumino-silicate material obtained by calcining and 1:2 for the first group of 6 mixes and second group of 30 mixes,
kaolin clay within the temperature range of 700–850 °C [2,4,5]. respectively. Each group included mixes with 10%, 15%, 20%, 25%
The principal reasons for the use of clay-based pozzolans in mortar and 30% replacement by weight of cement with MK; and a control
and concrete have been materials availability and durability mix without any mineral admixture. At a 10%, 15% and 20% partial
enhancement. In addition, depending on the calcining temperature replacement of cement with MK, they have reported that the fc of
and clay type, it is also possible to obtain enhancement in strength, mortars increases at 7, 28, 60 and 90 days for water–binder ratios
particularly during the early stages of curing. The very early between 0.4 and 0.53. Furthermore, it is observed that at the 30%
replacement of cement with MK, the mortar fc at all levels of the
* Tel.: +90 388 225 2348. water–binder ratios generally shows a decline at 7, 28, 60 and 90
E-mail address: msdemir@nigde.edu.tr. days. In another paper, Parande et al. [5] have produced mortar

0965-9978/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.advengsoft.2008.12.008
M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927 921

by replacing cement with MK partially by weight at the ratios of


x1 1
5%, 10%, 15%, and 20%. Additionally, they have prepared these spec- wi1
imens using mixes with a water–cement ratio of 0.40 and the bin- x2 wi2
2 (net)j 1
Σwijoi
der–sand ratio of 1:3. They have reported that the fc increased with oj
wi3
the addition of MK till the end of 90 days. It is also observed that x3 Output
3 …
the addition of 15% of MK gives the best result when compared ... … Sum function Sigmoid activation
to other replacement levels. Besides, it is mentioned that the addi- ...
function
win
tion of more than 15% of MK with cement decreases the fc. Courard xn n
et al. [8] have produced specimens by considering the water–ce- Input
ment ratio of 0.5 and binder–sand ratio of 1:3 during their exper-
Fig. 1. The artificial neuron model.
imental study. Prepared specimens contain mixes with 10%, 15%,
20%, 25% and 30% replacement by weight of cement with MK;
and a control mix without any mineral admixture. They have ob-
Fig. 1 shows a typical neural network with input, sum function,
tained the maximum fc results at the mixes containing 15% MK
sigmoid activation function and output. The input to a neuron from
replacing cement by weight for 3, 7, 14 and 28 days. In the study
another neuron is obtained by multiplying the output of the con-
of Caldarone et al. [9] have produced concrete, with 5% and 10%
nected neuron by the synaptic strength of the connection between
MK, which showed increase strengths at ages up to 365 days.
them [29]. The weighted sums of the input components (net)j are
The authors reported that their MK-cement concretes exhibited
calculated by using the following equation
fc, which were slightly greater than silica fume–cement mixtures
at the same levels replacement by the pozzolans. Similar influences X
n

of MK on the fc of concrete have been reported by Wild et al. [10]. ðnetÞj ¼ wij oi þ b ð1Þ
i¼1
In the last years, artificial neural networks (ANN) and fuzzy lo-
gic (FL) approaches, a sub-field of intelligent systems, are being where (net)j is the weighted sum of the jth neuron for the input re-
widely used to solve a wide variety of problems in civil engineering ceived from the preceding layer with n neurons, wij is the weight
applications [11–26]. While ANN are low-level computational between the jth neuron in the preceding layer, oi is the output of
structures that perform well when dealing with raw data, FL deals the ith neuron in the preceding layer [15,16,24–26]. The quantity
with reasoning on a higher level, using linguistic information ac- b is called the bias and is used to model the threshold. The output
quired from domain experts. The above-mentioned capabilities signal of the neuron, denoted by oj in Fig. 1, is related to the network
make ANN and FL a very powerful tool to solve many civil engi- input (net)j via a transformation function called the activation func-
neering problems, particularly problems, where data may be com- tion [29]. The most common activation functions are ramp, sigmoid
plex or in an insufficient amount [12]. This purpose of the study is and Gaussian function. In general, for multilayer receptive models
to build models in ANN and FL systems to evaluate the effect of MK as the activation function (f (net)j) sigmoid function is used. The
on fc of mortars. For purpose of constructing these models, 46 dif- output of the jth neuron oj is calculated by using Eq. (2) with a sig-
ferent mixtures with 179 specimens of the 3, 7, 28, 60 and 90 days moid function as follows [15,16,24]:
fc results of mortars containing MK, used in training and testing for 1
the multilayer feed-forward ANN and Sugeno-type fuzzy inference oj ¼ f ðnetÞj ¼ ð2Þ
1 þ eaðnetÞj
models, were gathered from the technical literature [3,5,8]. In
training and testing of these models the age of specimen (AS), where a is constant used to control the slope of the semi-linear re-
metakaolin replacement ratio (MK%), water–binder (WB) ratio, gion. The sigmoid nonlinearity activates in every layer except in the
superplasticizer (SP%) and binder–sand (BS) ratio were entered as input layer [15,24,29]. The sigmoid function represented by Eq. (2)
input; while compressive strength (fc) values were used as output. gives outputs in (0, 1) [15,16,24–26].
These models were trained with 119 data of experimental results In recent years, ANN have been applied to many civil engineer-
and then remainders were used as only experimental input values ing problems with some degree of success. In civil engineering,
for testing and testing results similar to the experimental results neural networks have been applied to the detection of structural
were obtained. damage, structural system identification, modeling of material
behavior, structural optimization, structural control, ground water
monitoring, prediction of experimental studies, and concrete mix
2. Artificial neural networks proportions [14].

Artificial neural networks (ANN) are based on the present 2.1. Feed-forward neural networks
understanding of biological nervous system, although much of
the biological detail is neglected. ANN are enormously parallel sys- Feed-forward neural networks are the most popular and most
tems composed of many processing elements connected by links of widely used models in many practical applications. They are
variable weights [18,27]. Generally, ANN are made of an input known by many different names, such as multilayer perceptrons.
layer of neurons, sometimes referred to as nodes or processing In a feed-forward neural network, the artificial neurons are ar-
units, one or several hidden layer of neurons and output layer of ranged in layers, and all the neurons in each layer have connections
neurons. The neighboring layers are fully interconnected by to all the neurons in the next layer [13]. However, there is no con-
weight. The input layer neurons receive information from the out- nection between neurons of the same layer or the neurons which
side environment and transmit them to the neurons of the hidden are not in successive layers. In general, the feed-forward neural
layer without performing any calculation [17,28]. Layers between network consists of one input layer, one or two hidden layers
the input and output layers are called hidden layers and may con- and one output layer of neurons [29].
tain a large number of hidden processing units [14]. All problems, The input layer receives input information and passes it onto
which can be solved by a perceptron can be solved with only one the neurons of the hidden layer(s), which in turn pass the informa-
hidden layer, but it is sometimes more efficient to use two or three tion to the output layer. The output from the output layer is the
hidden layers. Finally, the output layer neurons produce the net- prediction of the net for the corresponding input supplied at the in-
work predictions to the outside world [17,28]. put nodes. Each neuron in the network behaves in the same way as
922 M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927

discussed in Eqs. (1) and (2). There is no reliable method for decid- rons until a satisfactory convergence is achieved for all the exam-
ing the number of neural units required for a particular problem. ples present in the training set [29]. The training process is
This is decided based on experience and a few trials are required successfully completed, when the iterative process has converged.
to determine the best configuration of the network [29]. In the The connection weights are captured from the trained network, in
studies, the multilayer feed-forward type of neural networks, as order to use them in the recall phase [12].
shown in Fig. 2, is considered. In a feed-forward network, the in-
puts and output variables are normalized within the range of 0–1. 2.3. Neural network model

2.2. The back-propagation algorithm A multilayered feed-forward neural network with a back-prop-
agation algorithm was employed in the present study. The nonlin-
Neural networks learn the problems by samples like human ear sigmoid function was used in the hidden layer and the neuron
brain. Establishing the connections between neurons determines outputs at the output layer. The neural network architecture devel-
the structure of network; and training connections to obtain de- oped in this research has five neurons in the input layer and one
sired results determines the learning algorithm of network. There neurons in the output layer as demonstrated in Fig. 3. Two hidden
are several learning algorithm [30]. The feed-forward multilayer layers were used in the architecture of multilayer feed-forward
networks using a back-propagation training algorithm are em- neural network due to its minimum absolute percentage error val-
ployed in the present study. Back-propagation algorithm, as one ues for training and testing sets. In the first hidden layer six and in
of the most well-known training algorithms for the multilayer per- the second hidden layer five neurons were determined.
ceptron, is a gradient descent technique to minimize the error for a AS, MK%, WB, SP% and BS were entered as input; while fc values
particular training pattern in which it adjust the weights by a small were used as outputs in training and testing of the multilayer feed-
amount at a time [15,16,24]. The learning error for rth example is forward neural network model. Herein, 119 data of experiment re-
calculated by using the following equation sults were used for training, whereas 60 ones were employed for
testing. The limit values of input and output variables used in
1X
Er ¼ ðtj  oj Þ2 ð3Þ the multilayer feed-forward neural network model are listed in Ta-
2 j
ble 1. More detail regarding input and output variables can be ob-
tained from literature [3,5,8]. Learning rate and momentum rate
where tj is the output desired at neuron j and oj is the output pre-
values were determined and the model was trained through itera-
dicted at neuron j. As presented in Eqs. (1) and (2) the output oj is
tions. The values of parameters obtained in the multilayer feed-for-
a function of synaptic strength and outputs of the previous layer
ward neural network model are given in Table 2. The trained model
[29]. In the back-propagation phase, the error between the network
was only tested with the input values and the results found were
output and the desired output values is calculated using the so-
close to experiment results.
called generalized delta rule [31], and weights between neurons
are updated from the output layer to the input layer by using the
following equation [12] 3. Fuzzy logic inference system

wij ðm þ 1Þ ¼ wij ðmÞ þ gðdj oj Þ þ bwij ðtÞ ð4Þ The fuzzy logic (FL) concept provides a natural way of dealing
where, dj is the error signal at a neuron j, oj is the output of neuron j, with problems where the source of imprecision is an absence of
m is the number of iteration, and g, b are called learning rate and sharply defined criteria rather than the presence of random vari-
momentum rate, respectively. dj in Eq. (4) can be calculated using ables. The fuzzy approach considers cases where linguistic uncer-
the partial derivative of the error function Er in the output layer tainties play some role in the control mechanism of the
and other layer, respectively, by Eqs. (5) and (6) [12,29] phenomena concerned [32]. Thus, generally speaking, such ap-
proaches are always needed. One of the methods of solving such
dj ¼ oj ðt j  oj Þð1  oj Þ ð5Þ approaches is the fuzzy logic inference system [19,21–26]. Fuzzy
X
dj ¼ oj ð1  oj Þ dk wkj ð6Þ inference systems (FIS) are powerful tools for the simulation of
k nonlinear behaviors with the help of FL and linguistic fuzzy rules
[22]. FIS are a rule based system comprise of three conceptual com-
where, the kth layer means the upper layer of the jth layer [12]. The
ponents. These are: (1) a rule base, includes fuzzy If–then rules, (2)
above operations are repeated for each example and for all the neu-
a data base, defines the membership function and (3) an inference
system, combines the fuzzy rules and produces the system results
Feed-forward [33]. There are various FIS methodologies, such as Mamdani-type
and Sugeno-type [34–36]. The fuzzy modeling or fuzzy identifica-

N11
AS N21
N12
Inputs Outputs MK% N22
N13
WB N23 fc
Output N14
layer SP% N24 Output
N15 layer
Input N25
BS
layer Hidden N16
Input 2. Hidden
layer
layer 1. Hidden layer
Back-propagation layer

Fig. 2. The architecture of a multilayer neural network. Fig. 3. The system used in the multilayer neural network model for fc.
M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927 923

Table 1 2
Oi = μAi ( x )
3
Oi = wi
4
Oi = wi O = w z
5
i i i
The input and output quantities used in the models. wi
= μAi ( x).μBi ( y ) = x y 6
A1 w1 + w2 Oi = wi z i
Input variables Data used in training and testing the model
x w1 w1
i

Minimum Maximum A2 Π N z1 w1 z1

Age of specimen, day 3 90


Σ z(x,y)
B1 Π N z2 w2 z2 Layer 6
Metakaolin,% 0 30 y w2 w2
Water–binder ratio 0.40 0.53 Layer 3 Layer 5
Layer 1 B2 Layer 4
Superplasticizer,% 0 1.3
Binder–sand ratio 0.33 0.50
Layer 2

Output variable Fig. 5. Corresponding adaptive neuro-fuzzy inference system architecture.


Compressive strength, MPa 15.45 71.20
O2i ¼ lAi ðxÞ; O2i ¼ lBi ðyÞ; i ¼ 1; 2 ð7Þ

Table 2 where, lAi and lBi are the appropriate membership functions for Ai
The values of parameters used in the multilayer neural network model. and Bi fuzzy sets, respectively. Many various membership functions
Parameters ANN
such as trapezoidal, triangular, Gaussian function, etc. can be ap-
plied to determine the membership grades [33]. In this study, the
Number of input layer neurons 5
Number of hidden layer 2
generated triangular membership function was calculated as in
Number of first hidden layer neurons 6 the following equation
Number of second hidden layer neurons 5 8 9
Number of output layer neuron 1 >
> 0; x 6 ai >
>
>
> xai >
>
Momentum rate 0.9 <
b a
; ai 6 x 6 bi =
Learning rate 0.7 O2i ¼ lAi ðxÞ ¼ i
ci x
i
ð8Þ
Error after learning 0.00006
>
>
> ci bi
; bi 6 x 6 c i >
>
>
>
: >
;
Learning cycle 15.000 0; ci 6 x
where, ai, bi and ci are the membership function’ parameters set that
tion, first explored systematically by Sugeno et al. [35,36], has change the shape of membership function from 1 to 0. These
found numerous practical applications in control, prediction and parameters are referred to as the premise parameters [33,38].
FIS [20–23]. For simplicity, it was assumed that the FIS had two in- Layer 3: This layer is the rule layer. Each neuron in this layer
puts, x and y, and one output, z. For the zero-order Sugeno-type corresponds to a single Sugeno-type fuzzy rule. A rule neuron re-
inference system, typical two rules can be expressed [24–26,33]: ceives inputs from the respective fuzzification neurons and calcu-
lates the firing strength of the rule it represents. The outputs of
R1 : If x is A1 and y is B1 ; then z1 ¼ p1 x þ q1 y þ r1 third layer, called as firing strengths O3i , are the products of the cor-
R2 : If x is A2 and y is B2 ; then z2 ¼ p2 x þ q2 y þ r2 responding degrees obtaining from layer 2, called as w in the fol-
lowing equation [25,33]
where, x and y are the crisp inputs to the node i, Ai and Bi are the
linguistic labels (low, medium, high, etc.) described by convenient O3i ¼ wi ¼ lAi ðxÞlBi ðyÞ; i ¼ 1; 2 ð9Þ
membership functions and pi, qi and ri are the consequence param-
Layer 4: This layer is the normalization layer. Each neuron in
eters (i = 1 or 2) [25,33]. If z1 and z2 are constants instead of linear
this layer receives inputs from all neurons in the rule layer, and cal-
equations, then we have zero-order Sugeno-type fuzzy model
culates the normalized firing strength of a given rule. The output of
[35–37].
each neuron in this layer is the ratio of the ith rule’s firing strength
Figs. 4 and 5 demonstrate the fuzzy reasoning mechanism and
to the sum of all rules’ firing strength, and is calculated as in the
the corresponding adaptive neuro-fuzzy inference system (ANFIS),
following equation
respectively [24–26,37]. The scheme of ANFIS consists of five layers
(Fig. 5) [25,33].The design procedure for each layer in the premise wi
O4i ¼ wi ¼ P ; i ¼ 1; 2 ð10Þ
and the consequence of ANFIS is as follows: i wt
Layer 1: In this layer, the input signals are distributed to the
Layer 5: This layer is the defuzzification layer. Each neuron in
neurons in the next layer directly [25,38].
this layer is connected to the respective normalization neuron,
Layer 2: This layer is the fuzzification layer. Each neuron in this
and also receives initial inputs, x and y. In this layer, the contribu-
layer produces membership grades of the crisp inputs which be-
tion of ith rule’s towards the total output or the model output and/
long to each of convenient fuzzy sets by using the membership
or the function defined Eq. (11) is calculated
functions. Each neuron’s output O2i is calculated as in Eq. (7)
[25,33] O5i ¼ wi zi ¼ wi ðpi x þ qi y þ r i Þ; i ¼ 1; 2 ð11Þ

Minimum or Product
μ ( x) μ ( y)
B1 [min(ai,bi) or ai,xbi]
a1 A1
b1 w1 z1 = p1 x + q1 y + r1
R1
X Y
y w1 z1 + w2 z 2
x Z=
μ (x) μ ( y) w1 + w2
A2 B2
R2 b2
w2 z 2 = p2 x + q2 y + r2
a2
X Y
x y
Fig. 4. Inference system of zero-order Sugeno-type model with two fuzzy rules.
924 M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927

AS2 AS3 AS4 AS5


1 AS1

M. degree
0.5

0
10 20 30 40 50 60 70 80 90
Age of specimen, day

1 MK1 MK2 MK3 MK4


1 WB1 WB2 WB3 WB4
M. degree

M. degree
0.5 0.5

0 0
0 5 10 15 20 25 30 0.4 0.42 0.44 0.46 0.48 0.5 0.52
Metakaolin, % Water-binder ratio

1 SP1 SP2 SP3


1 BS1 BS2 BS3
M. degree

M. degree
0.5 0.5

0 0
0 0.2 0.4 0.6 0.8 1 1.2 0.35 0.4 0.45 0.5
Superplasticizer, % Binder-sand ratio

Fig. 6. Membership functions of input variables for fc.

where wi is the ith neuron’s output from the previous layer. pi, qi output layer backward to the input nodes. This learning rule is ex-
and ri are the parameters set in the consequence function and also actly the same as the back-propagation learning rule used in the
the coefficients of linear combination in Sugeno-type fuzzy infer- common feed-forward neural networks [25,37]. From the ANFIS
ence system [33,38]. architecture in Fig. 5, it is observed that given the values of premise
Layer 6: This layer is represented by a single summation neuron. parameters, the overall output z can be expressed as a linear com-
This neuron calculated the sum of outputs of all defuzzification bination, a hybrid-learning rule is employed here, which combines
neurons. Each rule’s fuzzy results are transformed into a crisp out- the gradient descent and the least-squares method to find a feasi-
put in this layer by defuzzification process using the following ble set of antecedent and consequent parameters [25,37,39].
equations [25,33]
3.1. Fuzzy inference system model
w1 ðx; yÞz1 ðx; yÞ þ w2 ðx; yÞz2 ðx; yÞ w1 z1 þ w2 z2
Zðx; yÞ ¼ ¼ ð12Þ
w1 ðx; yÞ þ w2 ðx; yÞ w1 þ w2
X P In this study, totally 179 data experiment results were used in
w i z i
O6i ¼ zðx; yÞ ¼ wi zi ¼ Pi ð13Þ the processes of Sugeno-type fuzzy inference model in FL system.
i i wi Of these, 119 data experiment results were used for training and
60 data for testing of the model. The limit values of input and out-
Therefore, an adaptive network is presented in Fig. 5, which is
put variables used in Sugeno-type fuzzy inference model are listed
functionally equivalent to a FIS in Fig. 4 [24–26,37].
in Table 1. In the rule base, fuzzy variables were connected with
The basic learning rule of ANFIS is the back-propagation gradi-
‘‘prod” operators and the implication of the each rule was calcu-
ent descent, which calculates error signals recursively from the

75 75
70 ANN training results 70 ANN testing results
Predicted compressive strength, MPa
Predicted compressive strength, MPa

65 FL training results 65 FL testing results


60 60
55 55 FL testing
FL training
50 50 y = 0.9964x - 0.5842
y = 0.9966x + 0.1823
45 45 R2 = 0.9659
R2 = 0.9993
40 40
35 35 ANN testing
ANN training y = 0.9964x + 0.0946
30 30
y = 0.9923x + 0.4666 R2 = 0.9738
25 25
R2 = 0.9887
20 20
15 15
10 10
10 15 20 25 30 35 40 45 50 55 60 65 70 75 10 15 20 25 30 35 40 45 50 55 60 65 70 75
Experimental compressive strength, MPa Experimental compressive strength, MPa
Fig. 7. Comparison of fc experimental results with training results of models. Fig. 8. Comparison of fc experimental results with testing results of models.
M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927 925

Table 3
Testing data sets for comparison of experimental results with testing results predicted from models.

Data used in models construction Compressive strength, MPa


AS, day MK, % WB SP, % BS Experimental ANN FL Ref.
3 5 0.50 0.00 0.33 34.40 31.74 35.09 [8]
3 15 0.50 0.00 0.33 31.70 33.42 33.27 [8]
3 5 0.40 0.00 0.33 31.50 26.07 29.31 [5]
3 15 0.40 0.00 0.33 29.00 29.56 30.78 [5]
7 10 0.50 0.00 0.33 48.60 46.05 43.23 [8]
7 10 0.40 0.00 0.33 43.50 43.49 39.58 [5]
7 10 0.48 0.00 0.36 18.99 16.44 16.96 [3]
7 25 0.52 0.00 0.36 16.30 16.74 15.21 [3]
7 10 0.53 0.00 0.50 25.90 26.50 25.24 [3]
7 25 0.53 0.00 0.50 27.50 27.37 28.29 [3]
7 10 0.50 0.00 0.50 30.40 30.45 28.73 [3]
7 25 0.50 0.00 0.50 29.10 29.51 29.00 [3]
7 10 0.47 0.00 0.50 31.90 32.84 30.06 [3]
7 25 0.47 0.00 0.50 30.20 31.26 30.52 [3]
7 0 0.44 0.50 0.50 35.40 37.10 32.87 [3]
7 25 0.44 0.50 0.50 33.80 35.42 34.88 [3]
7 15 0.40 1.30 0.50 41.90 41.37 44.29 [3]
7 25 0.40 1.30 0.50 39.60 40.17 40.16 [3]
28 5 0.50 0.00 0.33 65.10 64.08 62.57 [8]
28 15 0.50 0.00 0.33 71.20 69.29 63.98 [8]
28 0 0.40 0.00 0.33 40.00 47.40 36.03 [5]
28 15 0.40 0.00 0.33 60.00 59.52 59.79 [5]
28 15 0.50 0.00 0.36 26.24 22.70 27.96 [3]
28 25 0.52 0.00 0.36 23.25 23.88 19.99 [3]
28 10 0.53 0.00 0.50 38.50 38.11 38.83 [3]
28 25 0.53 0.00 0.50 38.10 38.16 37.89 [3]
28 10 0.50 0.00 0.50 41.30 40.41 40.43 [3]
28 25 0.50 0.00 0.50 39.90 40.35 40.34 [3]
28 10 0.47 0.00 0.50 43.60 41.42 42.95 [3]
28 25 0.47 0.00 0.50 42.80 42.54 41.30 [3]
28 10 0.44 0.50 0.50 45.90 44.47 45.79 [3]
28 25 0.44 0.50 0.50 44.70 43.63 42.95 [3]
28 10 0.40 1.30 0.50 47.20 47.05 46.50 [3]
28 25 0.40 1.30 0.50 47.70 46.26 45.76 [3]
60 10 0.48 0.00 0.36 29.20 30.02 27.74 [3]
60 25 0.52 0.00 0.36 28.45 29.20 26.87 [3]
60 10 0.53 0.00 0.50 43.80 43.41 42.49 [3]
60 25 0.53 0.00 0.50 42.50 41.55 41.71 [3]
60 10 0.50 0.00 0.50 46.00 44.89 44.27 [3]
60 25 0.50 0.00 0.50 43.60 43.13 43.33 [3]
60 10 0.47 0.00 0.50 46.90 45.92 43.43 [3]
60 25 0.47 0.00 0.50 43.90 44.94 45.03 [3]
60 10 0.44 0.50 0.50 48.50 48.27 46.88 [3]
60 25 0.44 0.50 0.50 47.30 46.56 46.03 [3]
60 10 0.40 1.30 0.50 50.90 51.40 50.20 [3]
60 25 0.40 1.30 0.50 51.30 49.84 49.46 [3]
90 5 0.40 0.00 0.33 52.00 56.54 55.70 [5]
90 20 0.40 0.00 0.33 65.00 66.58 70.03 [5]
90 10 0.48 0.00 0.36 31.23 31.50 28.58 [3]
90 25 0.52 0.00 0.36 32.28 31.83 28.65 [3]
90 10 0.53 0.00 0.50 44.50 46.80 45.77 [3]
90 25 0.53 0.00 0.50 43.40 44.43 43.44 [3]
90 10 0.50 0.00 0.50 46.70 47.84 48.45 [3]
90 25 0.50 0.00 0.50 45.30 45.54 44.92 [3]
90 10 0.47 0.00 0.50 47.40 48.57 47.98 [3]
90 25 0.47 0.00 0.50 45.60 47.19 45.78 [3]
90 10 0.44 0.50 0.50 51.20 50.76 52.17 [3]
90 25 0.44 0.50 0.50 50.50 49.03 48.68 [3]
90 10 0.40 1.30 0.50 54.20 54.06 55.96 [3]
90 25 0.40 1.30 0.50 53.50 52.40 52.11 [3]

lated using ‘‘wtaver” (weighted average) defuzzification method.


The membership functions of the training data set for the input
variables of fc was determined of the triangular type and premise Table 4
parameter sub-spaces were determined by using clustering of the The fc statistical values of proposed models.
training data set. In this study, triangular membership functions,
Statistical parameters ANN FL
for the input variables AS, MK%, WB, SP% and BS were fuzzyfied
Training set Testing set Training set Testing set
by dividing them into 5, 4, 4, 3 and 3 partitions, respectively. The
membership function plots of input variables used in the training RMS 1.1344 1.7928 0.2887 2.1809
R2 0.9887 0.9738 0.9993 0.9659
of fc is shown in Fig. 6. Obviously, all of these fuzzy sets are com-
MAPE 2.0577 3.2288 0.2872 4.2169
plete and consistent on their domains and of course they are con-
926 M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927

vex. Moreover, training continued for over 10 epochs and process type fuzzy inference models are suitable and predict the 3, 7, 28,
terminated by the observation of the stability in error reduction. 60 and 90 days fc values very close to the experimental values.
Thus, 720 rules are obtained for the fc as in the following:
Ri: If (AS is ASmfj) and (MK% is MK%mfk) and (WB is WBmfp) and 5. Conclusions
(SP% is SP%mfm) and (BS is BSmfn) Then (fc is fcmfi), i = 1, . . . , 720;
j = 1, . . . , 5; k, p = 1, . . . , 4; m, n = 1, . . . , 3. In this study, artificial neural networks and fuzzy logic systems
were used for the prediction the 3, 7, 28, 60 and 90 days compres-
4. Results and discussion sive strength values of mortars containing metakaolin. In the mod-
el developed in artificial neural networks system, a multilayered
In the present study, three norms were used to comparative feed-forward neural network with a back-propagation algorithm
evaluation of the performance of the multilayer feed-forward neu- was used. In the multilayer feed-forward neural network model,
ral network and Sugeno-type fuzzy inference models. These norms two hidden layers were selected. In the first hidden layer 6 neurons
are root-mean-squared (RMS) error, absolute fraction of variance and in the second hidden layer 5 neurons were determined. In the
(R2) and mean absolute percentage error (MAPE) between model’s model developed in fuzzy logic system, Sugeno-type fuzzy infer-
results and experimental results which are according to the follow- ence system was used. These models were trained with input
ing equations , respectively [24–26,38] and output data. Using only the input data in trained models the
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3, 7, 28, 60 and 90 days compressive strength values of mortars
u n
u1 X containing metakaolin were found. The compressive strength val-
RMS ¼ t ðti  oi Þ2 ð14Þ
n i¼1 ues predicted from testing, for the multilayer feed-forward neural
P P P network and Sugeno-type fuzzy inference models, are very close to
ðn t i oi  ti oi Þ2 the experimental results. The statistical parameter values of RMS,
R2 ¼  P P  P P  ð15Þ
n t 2i  ð ti Þ2 n o2i  ð oi Þ2 R2 and MAPE have shown obviously this situation.
Pn  As a result, compressive strength values of mortars containing
1 i¼1 jt i  oi j metakaolin can be predicted in the multilayer feed-forward neural
MAPE ¼ Pn x100 ð16Þ
n i¼1 t i network and Sugeno-type fuzzy inference models in a quite short
period of time with tiny error rates. The conclusions have demon-
where t is the target value, o is the output value, n is the pattern.
strated that artificial neural networks and fuzzy logic systems are
In the present study, fc of mortars containing MK was predicted
practicable methods for predicting compressive strength values
using the multilayer feed-forward neural network and Sugeno-
of mortars containing metakaolin.
type fuzzy inference models. In the training and testing of the mul-
tilayer feed-forward neural network and Sugeno-type fuzzy infer-
References
ence models, various experimental data from three different
sources [3,5,8] are used. All results, obtained from experimental [1] Gleize PJP, Cyr M, Escadeillas G. Effects of metakaolin on autogenous shrinkage
studies [3,5,8] and predicted by using the training and testing re- of cement pastes. Cement Concrete Compos 2007;29(2):80–7.
sults of he multilayer feed-forward neural network and Sugeno- [2] Sabir BB, Wild S, Bai J. Metakaolin and calcined clays as pozzolans for concrete:
a review. Cement Concrete Compos 2001;23(6):441–54.
type fuzzy inference models, for 3, 7, 28, 60 and 90 days fc were gi- [3] Vu DD, Stroeven P, Bui VB. Strength and durability aspects of calcined kaolin-
ven in Figs. 7 and 8. The linear least square fit line, its equation and blended Portland cement mortar and concrete. Cement Concrete Compos
the R2 values were shown in these figures for the training and test- 2001;23(6):471–8.
[4] Poon CS, Kou SC, Lam L. Compressive strength, chloride diffusivity and pore
ing data. Also, inputs values and experimental results with testing
structure of high performance metakaolin and silica fume concrete. Constr
results obtained from the multilayer feed-forward neural network Build Mater 2006;20(10):858–65.
and Sugeno-type fuzzy inference models were given in Table 3. As [5] Parande AK, Babu BR, Karthik MA, Deepak Kumaar KK, Palaniswamy N. Study
seen in Figs. 7 and 8, the values obtained from the training and on strength and corrosion performance for steel embedded in metakaolin
blended concrete/mortar. Constr Build Mater 2008;22(3):127–34.
testing in the multilayer feed-forward neural network and Su- [6] Zhang MH, Malhotra VM. Characteristics of a thermally activated alumino-
geno-type fuzzy inference models are very close to the experimen- silicate pozzolanic material and its use in concrete. Cement Concrete Res
tal results. The results of the multilayer feed-forward neural 1995;25(8):1713–25.
[7] Brooks JJ, Megat Johari MA. Effect of metakaolin on creep and shrinkage of
network and Sugeno-type fuzzy inference models demonstrate concrete. Cement Concrete Compos 2001;23(6):495–502.
that the ANN and FL systems can be successfully applied to estab- [8] Courard L, Darimont A, Schouterden M, Ferauche F, Willem X, Degeimbre R.
lish accurate and reliable prediction models. Durability of mortars modified with metakaolin. Cement Concrete Res
2003;33(9):1473–9.
As seen in Figs. 7 and 8, good results were obtained from the [9] Caldarone MA, Gruber KA, Burg RG. High reactivity metakaolin: a new
multilayer feed-forward neural network and Sugeno-type fuzzy generation mineral admixture. Concrete Int 1994(November):37–40.
inference models. The statistical parameter values of RMS, R2 and [10] Wild S, Kahib JM, Jones A. Relative strength, pozzolanic activity and cement
hydration in superplasticised metakaolin concrete. Cement Concrete Res
MAPE showed obviously this situation. The statistical values of 1996;26(10):1537–44.
RMS, R2 and MAPE including all the station for both training and [11] Bai J, Wild S, Ware JA, Sabir BB. Using neural networks to predict workability of
testing, are given in Table 4. While the statistical values of RMS, concrete incorporating metakaolin and fly ash. Adv Eng Softw 2003;34(11–
12):663–9.
R2 and MAPE from training in the multilayer feed-forward neural
[12] Ince R. Prediction of fracture parameters of concrete by artificial neural
network model was found as 1.1344, 0.9887 and 2.0577%, respec- networks. Eng Fract Mech 2004;71(15):2143–59.
tively, these values were found in testing as 1.7928, 0.9738 and [13] Adhikary BB, Mutsuyoshi H. Prediction of shear strength of steel fiber RC
3.2288%, respectively. Similarly, while the statistical values of beams using neural networks. Constr Build Mater 2006;20(9):801–11.
[14] Kewalramani AM, Gupta R. Concrete compressive strength prediction using
RMS, and MAPE from training in the Sugeno-type fuzzy inference ultrasonic pulse velocity through artificial neural networks. Auto Constr
model was found as 0.2887, 0.9993 and 0.2872%, respectively, 2006;15(15):374–9.
these values were found in testing as 2.1809, 0.9659 and [15] Pala M, Özbay E, Öztasß A, Yüce MI. Appraisal of long-term effects of fly ash and
silica fume on compressive strength of concrete by neural networks. Constr
4.2169%, respectively. The best value of R2 is 0.9993 for training Build Mater 2007;21(2):384–94.
set in the Sugeno-type fuzzy inference model. The minimum value [16] Topçu IB,_ Sarıdemir M. Prediction of properties of waste AAC aggregate
of R2 is 0.9659 for testing set in the Sugeno-type fuzzy inference concrete using artificial neural networks. Comput Mater Sci
2007;41(1):117–25.
model. All of the statistical values in Table 4 show that the pro- [17] Demir F. Prediction of elastic modulus of normal and high strength concrete by
posed the multilayer feed-forward neural network and Sugeno- artificial neural network. Constr Build Mater 2008;22(7):1428–35.
M. Sarıdemir / Advances in Engineering Software 40 (2009) 920–927 927

[18] Altun F, Kisßi Ö, Aydin K. Predicting the compressive strength of steel fiber [29] Mukherjee A, Biswas SN. Artificial neural networks prediction of mechanical
added lightweight concrete using neural network. Comput Mater Sci behavior of concrete at high temperature. Nucl Eng Des 1997;178(1):1–11.
2008;42(2):259–65. [30] Tosun N, Özler L. A study of tool life in hot machining using artificial neural
[19] S
ß en Z. Fuzzy algorithm for estimation of solar irradiation from sunshine networks and regression analysis method. J Mater Process Technol
duration. Sol Energy 1998;63(1):39–49. 2002;124(1–2):99–104.
[20] Muthukumaran P, Demirli K, Stiharu I, Bhat RB. Boundary conditioning for [31] Rumelhart DE, Hinton GE, William RJ. Learning internal representations by
structural tuning using fuzzy logic approach. Comput Struct error propagations. In: Rumelhart DE, McClelland JL, editors. Proceedings
2000;74(5):547–57. parallel distributed processing foundations, vol. 1. Cambridge: MIT Press;
[21] Demir F. A new way of prediction elastic modulus of normal and high strength 1986.
concrete-fuzzy logic. Cement Concrete Res 2005;35(8):1531–8. [32] Demir F, Korkmaz KA. Prediction of lower and upper bounds of elastic modulus
_
[22] Inan G, Göktepe AB, Ramyar K, Sezer A. Prediction of sulfate expansion of PC of high strength concrete. Constr Build Mater 2008;22(7):1385–93.
mortar using adaptive neuro-fuzzy methodology. Build Environ [33] Firat M, Güngör M. River flow estimation using adaptive neuro fuzzy inference
2007;42(7):1264–9. system. Math Comput Simulat 2007;75(3–4):87–96.
_ Sarıdemir M. Prediction of rubberized concrete properties using
[23] Topçu IB, [34] Mamdani EH, Assilian S. An experiment in linguistic synthesis with a fuzzy
artificial neural network and fuzzy logic. Constr Build Mater logic controller. Int J Man-Machine Stud 1975;7:1–13.
2008;22(4):532–40. [35] Takagi T, Sugeno M. Fuzzy identification of systems and its applications to
[24] Topçu IB,_ Sarıdemir M. Prediction of compressive strength of concrete modeling and control. IEEE Trans Syst Man Cyb 1985;15:116–32.
containing fly ash using artificial neural network and fuzzy logic. Comput [36] Sugeno M, Kang GT. Structure identification of fuzzy model. Fuzzy Sets Syst
Mater Sci 2008;41(3):305–11. Man Cyb 1993;23:665–85.
[25] Topçu IB,_ Sarıdemir M. Prediction of mechanical properties of recycled [37] Akbulut S, Hasiloğlu AS, Pamukcu S. Data generation for shear modulus and
aggregate concretes containing silica fume using artificial neural networks damping ratio in reinforced sands using adaptive neuro-fuzzy inference
and fuzzy logic. Comput Mater Sci 2008;41(1):74–82. system. Soil Dyn Earthquake Eng 2004;24(11):805–14.
_ Sarıdemir M. Prediction of rubberized mortar properties using
[26] Topçu IB, [38] Fazel Zarandi MH, Türksen IB, Sobhani J, Ramezanianpour AA. Fuzzy
artificial neural network and fuzzy logic. J Mater Process Technol 2008;199(1– polynomial neural networks for approximation of the compressive strength
3):108–18. of concrete. Appl Soft Comput 2008;8(11):488–98.
[27] Lippman R. An introduction to computing with neural nets. IEEE ASSP Mag [39] Jang JSR, Sun CT. Neuro-fuzzy modeling and control. Proc IEEE
1987;4:4–22. 1995;83:378–405.
[28] Mansour MY, Dicleli M, Lee JY, Zhang J. Predicting the shear strength of
reinforced concrete beams using artificial neural network. Eng Struct
2004;26(6):781–99.

You might also like