You are on page 1of 11

Applied Soft Computing 11 (2011) 26452655

Contents lists available at ScienceDirect

Applied Soft Computing


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

Prediction of lateral connement coefcient in reinforced


concrete columns using neural network simulation
Sema Noyan Alacal a , Bulent Akbas b , Bilge Doran a,
a
Department of Civil Engineering, Yldz Technical University, Davutpasa Campus, 34210 Esenler-Istanbul, Turkey
b
Department of Earthquake and Structural Engineering, Gebze Institute of Technology, 41400 Gebze-Kocaeli, Turkey

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

Article history: This paper presents an application of Neural Network (NN) simulation in civil engineering science. The
Received 27 April 2009 connement degree for conned concrete has been investigated by using a NN analysis as an alternative
Received in revised form 23 August 2010 approach. To accurately predict the behavior of a conned concrete, it is important to understand the
Accepted 26 October 2010
connement degree and its individual components. For the purpose of investigating connement effects,
Available online 3 November 2010
three empirical equations as a function of various parameters and an experimental work existing in the
literature were considered in this study. However, these analytical models are time consuming to use.
Keywords:
Therefore, there is still the need to develop simple but accurate method for determining the connement
Neural network
Connement coefcient
coefcient. In this context, the NN algorithm has been established, in order to validate these empirical
Volumetric ratio equations proposed for the connement coefcient. The approach adapted in this study was shown to
Reinforced concrete columns be capable of providing accurate estimates of lateral connement coefcient, Ks by using the six design
Ductility parameters. Finally, comparison with other empirical equations proposed for the lateral connement
coefcient illustrates the validity of the proposed algorithm.
2010 Elsevier B.V. All rights reserved.

1. Introduction under monotonic axial compression experimentally to examine


the impact of various parameters on the behavior of tied columns.
Design procedures for reinforced concrete (R/C) structures often The distribution of longitudinal steel, the column perimeter, the
require ductile behavior. The ductile behavior can be provided tie conguration, the amount of lateral connement and longi-
by conning the concrete with the lateral or spiral reinforce- tudinal column steel were the main variables. Also, Sheik and
ment. Lateral or spiral reinforcements in R/C columns play an Uzumeri [4] carried out a complete stressstrain curve for con-
important role in enhancing the strength and ductility. The mag- ned concrete based on the experimental results. In this model,
nitude of the increase in strength of the R/C columns can be they showed that the distribution of longitudinal steel and smaller
dened by a connement coefcient, Ks . For this purpose, differ- tie spacing result in higher strength and ductility of concrete. Saat-
ent behavior models are proposed to describe the connement cioglu and Razvi [5,6] proposed a stressstrain relationship for
degree of conned concrete according to various parameters by conned concrete. The relationship consists of a parabola for the
many researchers. Kent and Park [1] proposed a stressstrain curve ascending branch, and a linear portion for the descending branch.
for concrete conned by rectangular steel hoops. This curve is rep- The descending branch is dened as the strain corresponding to
resented by a second degree parabola up to maximum stress; a 85% of the peak stress. A constant residual strength has been
linear falling branch and a horizontal linear portion with stress con- assumed at 20% strength level beyond the descending branch. The
stant at 0.2 of the maximum stress. The slope of the linear falling parameters of the analytical model have been established from
branch was found to depend on the concrete cylinder strength, a large volume of experimental data, including poorly conned
the ratio of width of conned concrete to spacing of hoops and and well-conned concrete. The model has been compared with
the ratio of with of conned concrete to spacing of hoops and the a large number of column tested experimentally. Chung et al. [7]
ratio of volume of hoop steel to volume of concrete core. Park proposed a stressstrain relation of conned concrete from an
et al. [2], modied the stressstrain relation suggested by Kent empirical study of 65 columns. The strength of concrete, the vol-
and Park [1]. Sheik and Uzumeri [3] tested 24 short tied columns umetric ratio, the connement type of rectilinear ties, and the
distribution of longitudinal bars were the experimental parame-
ters.
Corresponding author. Tel.: +90 532 296 50 02; fax: +90 212 383 51 33. In order to validate the different behavior models proposed
E-mail address: doranbilge@gmail.com (B. Doran). by the researchers mentioned above, the Neural Network (NN)

1568-4946/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.asoc.2010.10.013
2646 S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655

Table 1
Experimental results for lateral connement coefcient adapted from Chung et al. [7].

Num. Specimen fc (MPa) m (mm) fyh (MPa) s (mm) s (mm) s Ks (exp.)

1 L8S5.5S10 19.6 8 550 6 100 0.007 1.06


2 L8S5.5E10 19.6 8 550 8 100 0.012 1.10
3 L8S5.5S5.5 19.6 8 550 6 55 0.012 1.26
4 L8S5.5E5.5 19.6 8 550 8 55 0.022 1.58
5 L8S5.5S3 19.6 8 550 6 30 0.022 1.76
6 L8S5.5E3 19.6 8 550 8 30 0.041 2.18
7 L12S5.5E3 20.1 12 550 8 30 0.042 2.10
8 L12S5.5E5.5 20.1 12 550 8 55 0.022 1.45
9 L12S5.5E10 20.1 12 550 8 100 0.012 1.03
10 L8C5.5F3 21.1 8 550 5 30 0.023 2.56
11 L8D5.5S5 21.1 8 550 6 50 0.023 2.21
12 L8D5.5S4 21.1 8 550 6 40 0.029 2.00
13 L12R5.5E7 21.1 12 550 8 70 0.040 2.60
14 L0S7S3 21.6 0 700 6 30 0.023 1.78
15 L4S7S3 21.6 4 700 6 30 0.023 1.84
16 L8S7S3 21.6 8 700 6 30 0.023 1.58
17 L8C7S4.5 21.6 8 700 6 45 0.023 1.78
18 L12S9E3 20.1 12 900 8 30 0.042 2.10
19 L12S9E5.5 20.1 12 900 8 55 0.022 1.40
20 L12S9E10 20.1 12 900 8 100 0.012 1.03
21 L8D13S4 24 8 1300 6 40 0.029 2.30
22 L8D13E4 24 8 1300 8 40 0.051 3.60
23 M8S5.5S10 38.3 8 550 6 100 0.007 1.06
24 M8S5.5E10 38.3 8 550 8 100 0.012 1.06
25 M8S5.5S5.5 38.3 8 550 6 55 0.012 1.25
26 M8S5.5E5.5 38.3 8 550 8 55 0.022 1.30
27 M8S5.5S3 38.3 8 550 6 30 0.022 1.41
28 M8S5.5E3 38.3 8 550 8 30 0.041 1.59
29 M12S5.5E3 40.2 12 550 8 30 0.042 1.60
30 M12S5.5E5.5 40.2 12 550 8 55 0.022 1.31
31 M12S5.5E10 40.2 12 550 8 100 0.012 1.13
32 M8C5.5F3 34.8 8 550 5 30 0.023 1.93
33 M8D5.5S5 34.8 8 550 6 50 0.023 1.96
34 M8D5.5S4 40.2 8 550 6 40 0.029 1.90
35 M12R5.5E7 34.8 12 550 8 70 0.040 2.16
36 M12S9E3 40.2 12 900 8 30 0.042 1.70
37 M12S9E5.5 40.2 12 900 8 55 0.022 1.35
38 M12S9E10 40.2 12 900 8 100 0.012 1.24
39 M8D13S4 40.2 8 1300 6 40 0.029 2.00
40 M8D13E4 40.2 8 1300 8 40 0.051 2.30
41 H8S5.5S10 50.5 8 550 6 100 0.007 0.93
42 H8S5.5E10 50.5 8 550 8 100 0.012 0.81
43 H8S5.5S5.5 50.5 8 550 6 55 0.012 1.06
44 H8S5.5E5.5 50.5 8 550 8 55 0.022 1.14
45 H8S5.5S3 50.5 8 550 6 30 0.022 1.24
46 H8S5.5E3 50.5 8 550 8 30 0.041 1.41
47 H12S5.5E3 50.5 12 550 8 30 0.042 1.40
48 H12S5.5E5.5 54.4 12 550 8 55 0.022 1.21
49 H12S5.5E10 54.4 12 550 8 100 0.012 1.05
50 H8C5.5F3 54.4 8 550 5 30 0.023 1.46
51 H8D5.5S5 56.4 8 550 6 50 0.023 1.57
52 H12R5.5E7 56.4 12 550 8 70 0.040 1.71
53 H0S7S3 56.4 0 700 6 30 0.023 1.35
54 H4S7S3 47.5 4 700 6 30 0.023 1.60
55 H8S7S3 47.5 8 700 6 30 0.023 1.27
56 H8C7S4.5 47.5 8 700 6 45 0.023 1.61
57 H12S9E3 47.5 12 900 8 30 0.042 1.33
58 H12S9E5.5 54.4 12 900 8 55 0.022 1.15
59 H12S9E10 54.4 12 900 8 100 0.012 1.03

algorithm has been established. The NN model was employed to data contains error or is incomplete; (2) NNs are able to adapt solu-
develop and test using the experimental work carried out by Chung tions over time and to compensate for changing circumstances; (3)
et al. [7]. NNs can evaluate theoretical, experimental, or empirical data based
Neural Network (NN) modeling has been widely used as an on good and reliable past experience or a combination of these [22].
alternative approach for establishing nonlinear empirical equa- In this study, a Neural Network (NN) model has been established
tions in engineering problems for the last two decades engineering to predict the lateral connement coefcient of R/C columns. For
[821]. this purpose, a NN model was constructed to estimate the lateral
Some of the major advantages of a NN model can be listed as: (1) connement coefcient (Ks ) in R/C rectangular columns. The multi-
NNs learn and generalize from examples and experience to produce layered feed-forward NN model was developed and tested with the
meaningful solutions to the problems even in cases where the input back-propagation algorithm to assess Ks .
S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655 2647

bcy

bcx
(a)
Actual
Average
Equivalent fl
k2fl

(b)
Fig. 1. Passive connement pressure: (a) variation of connement pressure with reinforcement arrangement; (b) equivalent uniform pressure.

2. Assessment of the experimental and analytical works (1) Park et al. [2], proposed the modied stressstrain relation in
conned concrete and observed a new relation for passive con-
2.1. Experimental work for the lateral connement coefcient nement controlled by Ks . However, the location of longitudinal
reinforcement and lateral reinforcement arrangements were
Chung et al. [7] proposed a stressstrain relation of conned not considered. An empirical equation can be written as
concrete from an empirical study of 65 columns and investigated
s fyh
the connement effects of concrete columns by lateral ties. Test Ks = 1 + (1)
columns have 200 mm square cross-section, 600 mm high, 17 mm fc
concrete cover thickness, and 12 mm longitudinal reinforced diam- In this equation, s can be expressed by
eter. In this experimental work, L, M, and H denote the compressive
strength of concrete of 20, 39, and 54 MPa, respectively, and P rep- Ash s
s = (2)
resents the plain concrete column without any bars. In this study, sbc dc
Chung et al.s [7] experimental data are used to employ the NN where bc and dc are the width and depth of the conned core
model. Some necessary data set are chosen from their experiments measured to outside of ties, respectively, and s is the center to
and given in Table 1. Here, the compressive strength of concrete center spacing of ties. s is the total length of the lateral ties.
fc , number of longitudinal bars m, yield strength of lateral tie fyh , (2) Sheikh and Uzumeri [4] accomplished an experimental study
diameter of lateral ties h , spacing between lateral ties s, and vol- on the stressstrain relation and connement, in their study,
umetric ratio of lateral ties s , and nally the lateral connement
coefcient Ks are given.
Input Layer
X1
2.2. Analytical proposals for the lateral connement coefcient
Hidden Layer
X2 1
The magnitude of the increase in strength of the R/C columns
can be dened by a connement coefcient. In the literature, dif-
ferent analytical models for the connement degree of conned
2
concrete as a function of various parameters are proposed by many Output Layer
researchers [2,4,6]. These empirical equations for lateral conne-
ment coefcient, Ks are given below. i
Xi Y

Table 2 m
Design parameters.

Design Denition Range


Xn
X1 fc 19.656.4 (MPa)
X2 m 812 N/A
X3 fyh 5501300 (MPa) Bias Node 2
X4 h 68 mm
X5 s 30100 mm Bias Node 1
X6 s 0.0070.051 N/A
Y Ks 0.813.60 N/A Fig. 2. Typical neural network model.
2648 S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655

Table 3
Cases used in constructing the NN model.

Case no. (X1 ) (X2 ) (X3 ) (X4 ) (X5 ) (X6 ) (y)


fc m fyh h s s Ks

1 19.6 8 550 6 100 0.007 1.06


2 19.6 8 550 8 100 0.012 1.10
3 54.4 8 550 5 30 0.023 1.46
4 19.6 8 550 8 55 0.022 1.58
5 19.6 8 550 6 30 0.022 1.76
6 19.6 8 550 8 30 0.041 2.18
7 20.1 12 550 8 30 0.042 2.10
8 20.1 12 550 8 55 0.022 1.45
9 20.1 12 550 8 100 0.012 1.03
10 21.1 8 550 5 30 0.023 2.56
11 21.1 8 550 6 50 0.023 2.21
12 21.1 8 550 6 40 0.029 2.00
13 21.1 12 550 8 70 0.040 2.60
14 21.6 0 700 6 30 0.023 1.78
15 21.6 4 700 6 30 0.023 1.84
16 21.6 8 700 6 30 0.023 1.58
17 21.6 8 700 6 45 0.023 1.78
18 20.1 12 900 8 30 0.042 2.10
19 20.1 12 900 8 55 0.022 1.40
20 38.3 8 550 8 30 0.041 1.59
21 20.1 12 900 8 100 0.012 1.03
22 24 8 1300 6 40 0.029 2.30
23 24 8 1300 8 40 0.051 3.60
24 38.3 8 550 6 100 0.007 1.06
25 38.3 8 550 8 100 0.012 1.06
26 34.8 8 550 5 30 0.023 1.93
27 34.8 8 550 6 50 0.023 1.96
28 40.2 8 550 6 40 0.029 1.90
29 34.8 12 550 8 70 0.040 2.16
30 40.2 12 900 8 30 0.042 1.70
31 40.2 12 900 8 55 0.022 1.35
32 40.2 12 900 8 100 0.012 1.24
33 40.2 8 1300 6 40 0.029 2.00
34 40.2 8 1300 8 40 0.051 2.30
35 50.5 8 550 6 100 0.007 0.93
36 50.5 8 550 8 100 0.012 0.81
37 50.5 8 550 6 55 0.012 1.06
38 50.5 8 550 8 55 0.022 1.14
39 54.4 12 900 6 55 0.022 1.15
40 54.4 12 550 8 100 0.012 1.05
41 56.4 12 550 8 70 0.040 1.71
42 56.4 0 700 6 30 0.023 1.35
43 40.2 12 550 8 100 0.012 1.13
44 54.4 12 900 6 100 0.012 1.03
45 47.5 4 700 6 30 0.023 1.60
46 47.5 8 700 6 30 0.023 1.27
47 47.5 8 700 6 45 0.023 1.61
48 47.5 12 900 6 30 0.042 1.33

1 50.5 8 550 6 30 0.022 1.24


2 38.3 8 550 6 55 0.012 1.25
3 38.3 8 550 8 55 0.022 1.30
4 38.3 8 550 6 30 0.022 1.41
5 54.4 12 550 8 55 0.022 1.21
6 40.2 12 550 8 30 0.042 1.60
7 50.5 8 550 8 30 0.041 1.41
8 50.5 12 550 8 30 0.042 1.40
9 40.2 12 550 8 55 0.022 1.31
10 19.6 8 550 6 55 0.012 1.26
11 56.4 8 550 6 50 0.023 1.57

parameter Ks was affected by the gain in the concrete strength where Ack is the area of conned core concrete, Ast is the area
due to connement provided by rectilinear reinforcement. of total longitudinal reinforcement, m is the number of the lon-
Besides, they considered the effect of rectilinear reinforcement gitudinal steels, and ci is the net distance between longitudinal
as a product of the square root of volumetric ratio of steel. From steels bounded by perimeter. In Eq. (3), volumetric ratio of lat-
a regression analysis, Ks can be written as eral ties s is expressed by

  2   Ash s
b2c mci2 s s = (5)
Ks = 1 + 1 1 s fyh (3) sbc dc
140Pocc 5.5b2c 2bc

where bc and dc are the width and depth of the conned core
Pocc = 0.85fc (Ack Ast ) (4) measured to center of ties, respectively.
S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655 2649

Table 4 (3) Saatcioglu and Razvi [6] reported in their study that the passive
Combinations of design parameters used in estimating Ks .
lateral pressure generated by laterally expanding concrete and
Combo no. (X1 ) (X2 ) (X3 ) (X4 ) (X5 ) (X6 ) restraining transverse reinforcement in square and rectangular
fc m fyh h s s columns was not always uniform. Therefore, the connement

1 model proposed was based on the computation of equivalent

2

uniform pressure that gives the same effect as the nonuniform
3 connement pressures that develop in columns with different

4
reinforcement arrangements as shown in Fig. 1. Equation for Ks
5
can be written as
6

7
6.7 0.17
8 Ks = 1 + (f1e ) f1e (6)

9 fc

10

11
In Eq. (6), the equivalent uniform pressure f1e can be dened as
12
13

Ash fyh sin
14 f1e = k2 f1 = k2 (7)
sbc

Fig. 3. Error on the estimated Ks vs. actual Ks for the eleven testing samples.
2650 S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655

Fig. 3. (Continued )

where f1 is the average lateral pressure, is the angle between feedforward radial basis function network, etc. Feedforward neural
the transverse reinforcement and bc , and coefcient k2 dened networks are simple networks consisting of an input layer, hidden
to reduce the average pressure is obtained from Eq. (8). layer/layers and an output layer through which the information

 b  1  moves only in one direction [23]. In this paper, feedforward NN
bc c modeling is adapted. Fig. 2 shows a typical three-layer feedforward
k2 = 0.26 1 (8)
s ci f1 NN with n input nodes, m hidden nodes and one output. The input
nodes represent the data presented to the NN, whereas the output
3. Neural network design nodes represent the produced NN output. The hidden layer (Fig. 2)
functions as the interface.
NNs consists of simple mathematical structures handling highly A group of processing elements (PEs) exists in a typical NN,
non-linear problems easily. There are many NN types such as which is dened by computing the sum of their weighted inputs,
feedforward NN, self-organizing maps, recurrent neural networks, subtracting its threshold from the sum, and transferring these
results by a function as follows [23]:

Table 5

n
ui = wij xj i
NNs performance.
(9)
Combo no. EPEED (training) (%) EPEED (testing) (%) WE (%)
j=1
1 19.64 12.28 15.96
2 19.41 15.47 17.44 where ui represents the output of a PE, wij represents the synap-
3 11.79 21.22 16.51 tic weights associated with PE i, xj represents the input signal,  i
4 8.46 15.93 12.20 represents the threshold value of the PE, and () presents the
5 19.08 14.27 16.68
6 12.43 25.82 19.13
transformation (or activation) function [23]. The most common
7 9.28 16.91 13.10 form of activation function used in the construction of a NN is the
8 11.71 22.09 16.90 hyperbolic tangent function. This function generates output values
9 7.71 10.38 9.05 between 1 and 1 as given below [24]:
10 7.35 16.03 11.69
11 10.32 22.95 16.64 f (xi ) = tanh(xi ) (10)
12 7.67 12.52 10.10
13 7.24 6.49 6.87 where is used for controlling the slope of the function. There is no
14 6.34 4.07 5.21
specic rule as to how to determine the number of PEs required. For
S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655 2651

many practical engineering problems, a single hidden layer with an 3.1. Modeling the neural network
optimum number of PEs is considered to be sufcient [22]. Super-
vised learning NN algorithms, for example back propagation neural The accuracy of a NN model is signicantly affected by the
networks train slowly and require many training data. However, it input variables. The six design parameters for the input layer were
is the most commonly used NN for the analysis of structural and selected to evaluate the Ks (output data) in R/C columns. The param-

civil engineering problems due to its versatile and robust technique eters are X1 : compressive strength of unconned concrete fc (MPa),
and are capable of solving predictive problems [24]. Supervised X2 : number of longitudinal bars (m), X3 : yield strength of lateral
learning algorithms with static back-propagation neural network tie, fyh (MPa), X4 : diameter of lateral ties, h (mm), X5 : spacing
are selected for the purpose of this study [19]. The NN model in between lateral ties, s (mm), X6 : volumetric ratio of lateral ties, s .
this study was developed in three phases: the modeling, the train- Based on the previous experimental results, these six parameters
ing, and the testing phases. The analysis of data, the identication are assumed to be the predominant Ks drivers of this study, and the
of RC column and section parameters, and the internal rules were ranges of data for the selected design variables are given in Table 2.
considered in the modeling phase. The preparation of the data and A total of 59 cases were used for the NN model (Table 3). The Ks
the adaptation of the learning law for the training were performed and the input variables from the 59 cases were divided into two sets.
during the training phase. And the prediction accuracy of the model One set was used for the training of the NN (rst 48 cases in Table 3),
was evaluated at the testing phase, i.e. the comparison of the actual and the other was used for validating the performance of the trained
lateral connement coefcients (Ks ) and the estimated lateral con- network (testing). For testing purposes, 20% of the data (11 cases)
nement coefcients (Ks ). were selected at random order for the testing set for each training

Fig. 4. Sensitivity analyses results.


2652 S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655

Fig. 4. (Continued )

cycle (cases below the dotted line in Table 3). Normalization of the where n is the number of samples to be evaluated in the training
data set including training and testing was made for the modeling phase (n = 48 for this study), xi is the model output related to the
of the NN. For normalization, the data set were multiplied by nor- sample i (i = 1, 2,. . ., n), and X(i) is the target output, i.e. the esti-
malization coefcients: amplitude and offset. The normalization mated Ks . Training is stopped when the MSE remained unchanged
coefcients were computed based on the minimum and maximum for a given number of epochs. An epoch number of 1000 was found
values found among all of the data set [24]. One hidden layer was to be sufcient for the training phase.
used in the NN model. In the NN model, the momentum coefcient
of 0.7 for the hidden and output layers performed very well. The 3.3. The testing phase
step size in learning rate was selected as 1.0 for the hidden and 0.1
for the output layers. Testing set gives an idea about the performance of the NN. The
testing is performed with the best weights obtained during the
training. The NNs performance in this study was measured by using
3.2. The training phase
the Ks percentage error (PEKs ) formula as follows:
The NN analyses have been carried out by Neuro Solutions [24]. x(i) X(i)
To investigate which combination of the six design parameters esti- PEKs = x100% (12)
X(i)
mates the Ks , a total of 14 combinations were used (Table 4). The
NN models in this study were created using an input layer of 3-, 4-, To evaluate the entire NNs overall performance, weighted error
5-, and 6-six interconnected PEs corresponding to the 3-, 4-, 5-, and (WE) was dened as follows [26]:
6-six input parameters, respectively, and one PE corresponding to
an output layer selected as the target. The error between the NN WE(%) = 0.5(Average PEKs for Training Set)
output and target output is processed back through the network +0.5(Average PEKs for Testing Set) (13)
(backward pass) adjusting the individual weights [21]. During the
learning, a gradual reduction of error between the model output
and the target output occurs and the error is minimized so as to Table 5 summarizes the performance of the NN models for each
minimize the sum of squared errors [23]. The mean square error Combo. PEKs for the testing set in each NN model is given in Fig. 3.
(MSE) is dened as [25]: The highest error in the testing set was observed for Combo no. 6
and the smallest error was for Combo 14 (Fig. 3). Average PEKs for

n 2 the 11 testing cases for Combo 14 was calculated as 4.07%, while it
(x
i=1 i
X(i))
was 6.34% for the training set (48 cases). Thus, the WE was found
MSE = (11)
n to be 5.21% (Table 5).
S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655 2653

Table 6
Comparison of the experimental and analytical results.

Num. Specimen Ks,prop . Ks,exp /Ks,prop .

(1) (2) (3) (4) (5) (1) (2) (3) (4) (5)

1 L8S5.5S10 1.20 1.31 0.9 0.97 0.9 0.97 0.97 0.97 0.72 0.97
2 L8S5.5E10 1.34 1.341 1.24 1.16 1.11 0.99 0.99 0.99 0.95 0.99
3 L8S5.5S5.5 1.34 1.59 1.31 1.30 1.28 0.99 0.99 0.99 0.97 0.99
4 L8S5.5E5.5 1.59 1.79 1.40 1.54 1.48 1.06 1.06 1.06 1.02 1.06
5 L8S5.5S3 1.62 1.95 1.951 1.72 1.031 1.03 1.03 1.03 1.02 1.03
6 L8S5.5E3 2.07 2.28 1.65 2.05 0.990 0.99 0.99 0.99 1.06 0.99
7 L12S5.5E3 2.04 1.03 1.64 1.98 2.04 1.03 1.03 1.03 1.06 1.03
8 L12S5.5E5.5 1.57 0.9 0.9 0.9 0.98 0.98 0.98 0.98 0.97 0.98
9 L12S5.5E10 1.33 1.46 1.23 1.12 1.13 0.91 0.91 0.91 0.92 0.91
10 L8C5.5F3 1.60 1.89 2.58 1.96 2.32 1.10 1.10 1.10 1.31 1.10
11 L8D5.5S5 1.57 1.76 1.52 1.88 2.07 1.07 1.07 1.07 1.18 1.07
12 L8D5.5S4 1.73 1.93 1.62 0.88 2.26 0.88 0.88 0.88 0.88 0.88
13 L12R5.5E7 1.86 1.88 0.970 2.33 0.97 0.97 0.97 0.97 1.12 0.97
14 L0S7S3 1.051 2.30 1.051 a 12.30 1.05 1.05 1.05 a 1.05
15 L4S7S3 1.71 1.58 1.821 1.82 1.95 0.94 0.94 0.94 1.01 0.94
16 L8S7S3 1.71 1.97 10.91 1.73 1.75 0.91 0.91 0.91 0.91 0.91
17 L8C7S4.5 1.71 1.87 2.52 0.8 0.89 0.89 0.89 0.89 0.95 0.89
18 L12S9E3 2.70 2.83 1.78 2.11 2.12 0.99 0.99 0.99 1.00 0.99
19 L12S9E5.5 1.94 2.13 1.47 1.9 1.45 0.96 0.96 0.96 0.88 0.96
20 L12S9E10 1.54 0.9 0.9 1.59 1.05 0.98 0.98 0.98 0.86 0.98
21 L8D13S4 2.52 2.25 1.78 2.46 2.31 1.00 1.00 1.00 0.94 1.00
22 L8D13E4 3.65 2.70 2.00 2.63 3.23 1.11 1.11 1.11 1.37 1.11
23 M8S5.5S10 1.10 1.16 1.10 1.03 1.03 1.03 1.03 1.03 0.86 1.03
24 M8S5.5E10 1.17 1.21 1.12 0.97 0.96 1.11 1.11 1.11 1.09 1.11
25 M8S5.5S5.5 1.17 1.30 1.16 1.09 1.13 1.11 1.11 1.11 1.15 1.11
26 M8S5.5E5.5 1.30 1.40 1.0 1.30 1.27 1.02 1.02 1.02 1.00 1.02
27 M8S5.5S3 1.32 1.49 1.26 1.44 1.38 1.02 1.02 1.02 0.98 1.02
28 M8S5.5E3 1.55 1.65 1.33 1.72 1.63 1.52 1.52 1.52 0.93 1.52
29 M12S5.5E3 1.52 1.72 1.32 1.65 1.56 1.03 1.03 1.03 0.97 1.03
30 M12S5.5E5.5 1.29 1.44 1.29 1.24 1.32 0.99 0.99 0.99 1.05 0.99
31 M12S5.5E10 1.16 1.23 1.12 0.93 1.08 1.05 1.05 1.05 1.21 1.05
32 M8C5.5F3 1.36 1.54 1.96 1.71 2.00 0.97 0.97 0.97 1.13 0.97
33 M8D5.5S5 1.35 1.46 1.351 1.64 1.74 1.13 1.13 1.13 1.19 1.13
34 M8D5.5S4 1.38 1.49 1.33 1.92 1.72 1.11 1.11 1.11 0.99 1.11
35 M12R5.5E7 1.52 1.53 1.36 2.04 2.17 0.99 0.99 0.99 1.06 0.99
36 M12S9E3 1.85 1.92 1.39 1.76 1.59 1.07 1.07 1.07 0.97 1.07
37 M12S9E5.5 1.47 1.57 1.24 1.32 1.29 1.05 1.05 1.05 1.02 1.05
38 M12S9E10 1.27 1.29 1.14 1.00 0.98 1.26 1.26 1.26 1.25 1.26
39 M8D13S4 1.91 1.75 1.47 2.15 1.94 1.03 1.03 1.03 0.93 1.03
40 M8D13E4 2.58 2.01 1.60 2.30 2.53 0.91 0.91 0.91 1.00 0.91
41 H8S5.5S10 1.08 1.12 1.07 0.94 0.99 0.94 0.94 0.94 0.81 0.94
42 H8S5.5E10 1.13 1.16 1.09 0.91 0.85 0.85 0.85 0.85 0.89 0.85
43 H8S5.5S5.5 1.13 1.23 1.12 1.001 1.06 1.00 1.00 1.00 1.05 1.00
44 H8S5.5E5.5 1.23 1.231 1.15 0.930 1.22 0.93 0.93 0.93 0.95 0.93
45 H8S5.5S3 1.24 1.37 1.20 1.34 1.28 0.97 0.97 0.97 0.93 0.97
46 H8S5.5E3 1.41 1.50 1.25 0.970 1.45 0.97 0.97 0.97 0.88 0.97
47 H12S5.5E3 1.41 1.57 1.25 1.55 1.44 0.98 0.98 0.98 0.90 0.98
48 H12S5.5E5.5 1.21 1.33 1.14 0.95 1.28 0.95 0.95 0.95 1.05 0.95
49 H12S5.5E10 1.12 1.17 0.95 0.86 1.10 0.95 0.95 0.95 1.22 0.95
50 H8C5.5F3 1.23 1.35 0.961 1.52 1.14 1.14 1.14 1.14 0.96 1.14
51 H8D5.5S5 1.21 1.29 1.19 1.45 1.38 1.14 1.14 1.14 1.08 1.14
52 H12R5.5E7 1.32 1.33 1.22 1.0 1.0 1.07 1.07 1.07 0.95 1.07
53 H0S7S3 1.27 1.50 1.50 a 1.46 0.93 0.93 0.93 a 0.93
54 H4S7S3 1.32 1.26 1.23 1.48 1.48 1.08 1.08 1.08 1.08 1.08
55 H8S7S3 1.32 1.44 1.23 1.41 1.32 0.96 0.96 0.96 0.90 0.96
56 H8C7S4.5 1.32 1.40 1.69 1.53 1.45 1.11 1.11 1.11 1.05 1.11
57 H12S9E3 1.72 1.78 1.33 1.68 1.47 0.91 0.91 0.91 0.79 0.91
58 H12S9E5.5 1.35 1.42 1.17 1.22 0.89 0.89 0.89 0.89 0.94 0.89
59 H12S9E10 1.120 1.22 1.11 0.92 1.094 0.94 0.94 0.94 1.12 0.94

Mean 1.03 0.97 1.12 1.01 1.02


Standard deviation 0.20 0.20 0.22 0.13 0.10
Coefcient of variation 0.19 0.20 0.20 0.13 0.10

(1) Kent and Park; (2) Sheikh and Uzumeri; (3) Saatcioglu and Razvi;(4) Eq. (14); (5) NN.
a
Not considered (m = 0, see Table 1).

Sensitivity analysis was also carried out to investigate the effect training is disabled [24]. Sensitivity analyses results related to dif-
that each network input was having on the network output. This ferent combinations dened in Table 4 are shown in Fig. 4. During
gives the user the option of removing the insignicant channels the analysis, the network weights are not affected, because the net-
from the network reducing the size of the network. This would work training is disabled. The inputs to the network are shifted
reduce the complexity and the training times. During the sensitivity slightly and the corresponding change in the output is reported as
analysis, the network weights are not affected, because the network a percentage summing to 100% in total [24]. In this study, for the 3-
2654 S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655

parameter Combos (Fig. 4ad), the most effective parameters were design parameters. The results are considered to be encourag-
found to be fc (X1 ), h (X4 ), and s (X6 ). For the highest performance ing for further research of expanded data sets. By setting up
Combo (Combo 14), the input parameter s (X6 ) was found to be some random variations in the design parameters, it is possible
the most effective parameter on predicting the lateral connement to estimate the Ks for new or existing R/C rectangular columns.
coefcient (42.84%) (Fig. 4n), while the most insignicant param- It should be noted that neural networks learn from examples
eter was found to be the input parameter fyh (X3 ) (3.14%). In a real of which performance strongly depends on the quality and the
case, fyh is expected to have signicant effect on Ks . The reason it quantity of examples. The more examples there are, the less the
had such a little effect in this study might be due to only four dif- prediction error is. Thus, to study modeling and prediction meth-
ferent fyh values (550 MPa, 500 MPa, 900 MPa, and 1300 MPa) were ods, and construct an accurate prediction model of Ks , there is a
used in constructing the NN model. If there were more fyh values need for an experimental research in which full-scale data of R/C
included in the NN model, the effect of on the Ks would proba- rectangular columns with various congurations are considered.
bly be higher. s (X6 ) was also the most effective parameter in any This type of a broad range of experimental data would also be
Combo when included. The input parameters s (X1 ), m (X2 ), h helpful for evaluating the sensitivity analyses results.
(X4 ), and s (X5 ) had the effect of 11.61%, 12.21%, 15.79%, and 14.41%, 3. The accuracy of the connement coefcient depends strongly
respectively, on the lateral connement coefcient for Combo on the data provided to the NN as well as the number of input
14 (Fig. 4n). variables. For cases when there is not enough data for testing,
NN models with 4 or less variables estimate the connement
3.4. Comments on results coefcient with higher errors. For such highly nonlinear prob-
lems, errors less than 10% for both training and testing phases
Data from 11 cases were used for testing purposes in this study can be considered to be good enough in structural engineering
out of a total 48 cases. The results for the 6-parameter Combo problems.
(Combo 14) showed 94.79% of average accuracy. The highest per-
formance observed in 6-parameter Combo (Combo 14) shows that References
even the small attributes provided by the other parameters such
[1] D.C. Kent, R. Park, Flexural members with conned concrete, Journal of the
as fyh (X3 ) may enhance the NNs prediction capability, i.e. the
Structural Division, ASCE 97 (7) (1971) 19691990.
more the number of design parameters, the higher the accuracy. [2] R. Park, M.J. Priestley, W.D. Gill, Ductility of square-conned concrete columns,
MSE was measured as 0.0069 and 0.0024 at the end of the train- Proceedings ASCE 108 (4) (1982) 929950.
ing and testing phases, respectively. The results were also tested [3] S.A. Sheikh, S.M. Uzumeri, Strength and ductility of tied concrete columns,
Journal of the Structural Division, ASCE 108 (12) (1980) 27032722.
through Hegazy and Ayeds method [26]. In their study, Hegazy [4] S.A. Sheikh, S.M. Uzumeri, Analytical model for concrete connement in tied
and Ayed [26] developed a simple NN simulation in a spread- columns, Journal of the Structural Division, ASCE 106 (5) (1982) 10791101.
sheet format for cost-estimating of highway projects. They used [5] M. Saatcioglu, S.R. Razvi, Strength and ductility of conned concrete, Journal of
the Structural Engineering, ASCE 118 (6) (1992) 15901607.
an optimization technique used by Excel Solver which can be used [6] M. Saatcioglu, S.R. Razvi, Connement model for high-strength concrete, Jour-
as an optimization of nonlinear problems. In this study, the same nal of the Structural Engineering, ASCE 125 (3) (1999) 281289.
approach was adapted in this study to make a comparison of the [7] H.S. Chung, K.H. Yang, Y.H. Lee, H.C. Eun, Stressstrain curve of laterally con-
ned concrete, Engineering Structures 24 (2002) 11531163.
results obtained from the NN model developed in this study. The [8] I. Flood, A neural network approach to the sequencing of construction tasks,
WE was found to be 4.90% with an average PEKs of 4.53% and in: Proceedings of the Sixth International Symposium on Automation and
5.26% for training and testing cases, respectively, using Hegazy Robotics in Construction, Construction Industry Institute, Austin, Texas, 1989,
pp. 204211.
and Ayeds method [26]. A nonlinear regression analysis was also [9] L. Ren, Z. Zhao, An optimal neural network and concrete strength modeling,
carried and the following nonlinear equation with a coefcient Advances in Engineering Software 33 (2002) 117130.
correlation of 0.84 was obtained by using a commercial statistical [10] S. Lai, M. Serra, Concrete strength prediction by means of neural network,
Construction of Building Materials 11 (2) (1997) 9398.
software [27].
[11] A. Cladera, A.B. Mari, Shear design procedure for reinforced normal and high-
strength concrete beams using articial neural networks. Part I. Beams without
Ks = 41.1fc
(0.263) (0.129)
m(0.071) fyh h (0.813) s(0.209) h (0.591) (14) stirrups, Engineering Structures 26 (2004) 917926.
[12] S. Saadata, M.N. Nooria, G.D. Bucknera, T. Furukawab, Y. Suzukic, Structural
Eq. (14) yielded an average PEKs of 27.58%. The experimental and health monitoring and damage detection using an intelligent parameter vary-
analytical values obtained from suggested empirical equations and ing (IPV) technique, International Journal of Non-Linear Mechanics 39 (2004)
16871697.
from Eq. (14) and the NN model for lateral connement coef- [13] M. Tehranizadeh, M. Sa, Application of articial intelligence for construction
cient, Ks are compared by means of statistical analysis. Mean values, of design spectra, Engineering Structures 26 (2004) 707720.
standard deviations and coefcients of variation of the ratio of [14] M.E. Williams, M.I. Hoit, Bridge pier live load analysis using neural networks,
Advances in Engineering Software 35 (2004) 645652.
the experimental to analytical values, are calculated in Table 6. As [15] F. Amini, M.R. Tavassoli, Optimal active structural control force, number and
shown in Table 6, it is observed that the NN model constructed in placement of controllers, Engineering Structures 27 (2005) 13061316.
this study yielded better results compared to the proposed empir- [16] M. Papadrakakis, N.D. Lagaros, V. Plevris, Design optimization of steel structures
considering uncertainties, Engineering Structures 27 (2005) 14081418.
ical equations and Eq. (14).
[17] Z. Zhao, Steel columns under rea neural network based strength model,
Advances in Engineering Software 37 (2) (2004) 97105.
4. Conclusions [18] T. Kerh, S.B. Ting, Neural network estimation of ground peak acceleration at
stations along Taiwan high-speed rail system, Engineering Applications of Arti-
cial Intelligence 18 (7) (2005) 857866.
In this study, the NN model was achieved for developing and [19] W.T. Yeung, J.W. Smith, Damage detection in bridges using neural networks for
testing the prediction of connement coefcient in R/C rectangular pattern recognition of vibration signatures, Engineering Structures 27 (2005)
685698.
columns. The data of 48 cases were used to train the NN. The testing [20] J.J. Lee, W.L. Lee, J.H. Yi, C.B. Yun, H.Y. Jung, Neural networks-based damage
of the NN was done by the data of 11 testing cases. The following detection for bridges considering errors in baseline nite element models,
conclusions can be drawn based on the results of the analyses: Journal of Sound and Vibration 280 (2005) 555578.
[21] A.F. Ashour, M.A. Alqedra, Concrete breakout strength of single anchors in
tension using neural networks, Advances in Engineering Software 36 (2005)
1. Volumetric ratio of lateral ties was found to be the most effective 8797.
parameter on lateral connement coefcient. [22] M.Y. Raq, G. Bugmann, D.J. Easterbrook, Neural network design for engineer-
ing applications, Computers and Structures 79 (2001) 15411552.
2. The approach adapted in this study was shown to be capable [23] S. Haykin, Neural Networks: A Comprehensive Foundation, Macmillan College
of providing the best accurate estimates of Ks by using the six Publishing Company, Inc., NJ, 1994.
S.N. Alacal et al. / Applied Soft Computing 11 (2011) 26452655 2655

[24] Neuro Solutions. Neurodimension, Inc., Version 4.24, 2003. [26] T. Hegazy, A. Ayed, Neural network model for parametric cost estimation of
[25] H.M. Gunaydin, S.Z. Dogan, A neural network approach for early cost estimation highway projects, Journal of Construction and Engineering and Management
of structural systems of buildings, International Journal of Project Management 124 (3) (1998) 210218.
22 (2004) 595602. [27] IBM SPSS Statistics 19, Trial Software, 2010.

You might also like