Sunteți pe pagina 1din 11

PREDICTION OF ULTIMATE SHEAR STRENGTH OF REINFORCED-

CONCRETE DEEP BEAMS USING NEURAL NETWORKS


By A. Sanad1 and M. P. Saka,2 Member, ASCE

ABSTRACT: This study explores the use of artificial neural networks in predicting the ultimate shear strength
of reinforced-concrete deep beams. One hundred eleven experimental data collected from the literature cover
the simple case of a simply supported beam with two point loads acting symmetrically with respect to the
centerline of the span. The data are arranged in a format such that 10 input parameters cover the geometrical
and material properties of the deep beam and the corresponding output value is the ultimate shear strength.
Among the available methods in the literature, the American Concrete Institute, strut-and-tie, and Mau-Hsu
methods were selected because of their accuracy and used to calculate the shear strength of each beam in the
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

set. Later, an artificial neural network is developed using two different software programs and the ultimate shear
strength of each beam is determined form these networks. It is found that the average ratio of actual and predicted
shear strength was 0.99 for the neural network, 2.08 for the American Concrete Institute method, 0.85 for the
strut-and-tie method, and 0.84 for the Mau-Hsu method. It is apparent that neural networks provide an efficient
alternative method in predicting the shear strength capacity of reinforced-concrete deep beams where several
equations exist, none of which produce an accurate result.

INTRODUCTION and Lu 1999). Among the methods available in the literature,


Reinforced-concrete deep beams are commonly used ele- the strut-and-tie model constitutes a viable approach that is
ments in a wide range of different types of structures, from adopted for design of reinforced-concrete deep beams by nu-
tall buildings to offshore gravity structures. The American merous codes such as those of the Comité Euro-International
Concrete Institute (ACI) code 318-95 (clause 10.7.1) (ACI du Béton-Fédération International de la Précontrainte (CEB-
1995) classifies the beam as a deep beam for flexure if the FIP), Paris, Cement and Concrete Association, London, and
clear-span/overall-depth ratio is <1.25 for simply supported Canadian Standards Association, Rexdale, Ont., Canada.
beams and 2.5 for continuous beams and as deep beams for This study explores the feasibility of using a multilayer
shear if the clear-span/effective-depth ratio is <5 for simply feed-forward artificial neural network to predict the ultimate
supported beams loaded on one face and supported on the shear strength of reinforced-concrete deep beams. A neural
opposite face so that compression struts can develop between network is an interconnected network of processing elements
loads and supports. that has the ability to be trained to map a given input into the
The structural behavior of deep beams differs from that of desired output. If the training data use the results of experi-
slender beams because of the small ratio between shear span ments carried out on reinforced-concrete beams, the neural
and depth. In contrast to slender beams, the response is char- network gives the shear strength capacity without making any
acterized by nonlinear strain distribution even in the elastic assumption about the behavior of the beam. Neural network
range. The strength of deep beams with a normal amount of modeling techniques have been successfully applied to differ-
longitudinal reinforcement is usually governed by shear, not ent areas of structural engineering. Some of these were on
flexure. A significant amount of load is carried to supports by structural analysis and design (Vanluchene and Roufei 1990;
a compression thrust joining the load and the reaction. This Hajela and Berke 1991; Arslan and Ince 1994; Chuang et al.
compression in the diagonal direction combined with the ten- 1998; Jenkins 1998; Waszczyszyn 1998) and some others were
sion along the beam bars constitutes the basis for the strut- on structural damage assessment (Wu et al. 1992; Buenfeld
and-tie model. This tied arch action is recognized as the force- and Hassanein 1998). The application of neural networks in
transferring mechanism of deep beams. The failure of a deep structural mechanics is described in Topping and Bahreinine-
beam may occur because of crushing of a compression strut jad (1997). In this paper, an artificial neural network is devel-
or loss of a beam bar anchorage (Subedi 1986, 1988; ASCE- oped for the prediction of the shear strength of reinforced-
ACI 1998; Shin et al. 1999; Tan and Lu 1999; Hwang et al. concrete deep beams and the results obtained are compared
2000). Unfortunately, no accurate theory exists for predicting with the experimental values and with those determined from
the ultimate shear strength of deep reinforced-concrete beams. the ACI code method, strut-and-tie method (Wen 1993), and
The great number of parameters that affect the beam strength Mau-Hsu method (Mau and Hsu 1989). It is shown that the
has led to a limited understanding of shear failure. These pa- performance of the neural network in predicting the shear
rameters include size effect; proportion, amount and arrange- strength is much more accurate than the methods considered.
ment of tensile, compressive, and web reinforcement; shear It is noticed that, although the average value of the predicted
span to depth ratio; properties of concrete and steel; anchorage strength to the actual strength ratio is around 0.48 in the ACI
of reinforcement into the supports; and size of bearing and code method, the same ratio is only 0.97 in the neural network.
loading areas (Asin and Walraven 1995; ASCE-ACI 1998; Tan
ULTIMATE SHEAR STRENGTH OF
1
Sr. Civ. Engr., Ministry of Housing, Municipal Affairs, P.O. Box 53, REINFORCED-CONCRETE DEEP BEAMS
Bahrain.
2
Prof., Civ. Engrg. Dept., Univ. of Bahrain, P.O. Box 32038, Isa Town, Accurate prediction of the ultimate strength of reinforced-
Bahrain (corresponding author). concrete deep beams is of prime importance for developing a
Note. Associate Editor: Julio Ramirez. Discussion open until Decem- reliable method for their design. For this purpose, many ex-
ber 1, 2001. To extend the closing date one month, a written request must periments have been carried out to investigate the behavior of
be filed with the ASCE Manager of Journals. The manuscript for this
paper was submitted for review and possible publication on July 8, 1999;
such beams. The data collected from these experiments con-
revised February 1, 2001. This paper is part of the Journal of Structural stitute the basis of some of the expressions available in the
Engineering, Vol. 127, No. 7, July, 2001. 䉷ASCE, ISSN 0733-9445/01/ literature for analytically predicting the ultimate shear strength
0007-0818–0828/$8.00 ⫹ $.50 per page. Paper No. 21339. of these beams (Ramakrishnan and Ananthanarayana 1968;
818 / JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001

J. Struct. Eng., 2001, 127(7): 818-828


Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

FIG. 1. Basic Parameters for Shear Strength Prediction of Simply Supported Deep Beam: (a) Deep Beam; (b) Cross Section

Smith and Vantsiotis 1982; Rogowsky et al. 1986; Mau and in such a way that 10 basic parameters are listed as input
Hsu 1989; Tan and Lu 1999; Hwang et al. 2000). values and the shear strength is included as the corresponding
output value. The complete list of the data is given in the
Experimental Data Appendix, where the name and the source of each specimen
are referenced (Sanad 1997). The data collected contain the
The experimental data collected from the literature cover ranges, which vary from 0.95 to 5.40 for effective-span/depth
the shear strength of the specimens, which are simply sup- ratios, from 1.57 to 47 for effective-depth/breadth ratios, and
ported and subjected to two point loads acting symmetrically from 0.23 to 3.00 for shear-span/effective-depth ratios.
with respect to the centerline of the span. This case provides
a larger amount of data than other cases do, which is essential Numerical Methods
for better training of a network. During the collection of the
data, specimens that do not have shear-related failures have Several methods exist in the literature for the prediction of
been excluded from the training set. The basic parameters that the shear strength of reinforced-concrete deep beams. The
control the shear strength of deep beams, based on previous most prominent among these, which are selected and used in
research works, are shown in Fig. 1. These parameters are this study for the comparison of the results from the neural
listed below: network, are outlined in the following.

• Effective span of beam (L) ACI Code Method


• Breadth of beam (bw) The recommended formula in the code consists of a set of
• Effective depth of beam (d ) empirical rules based on a large amount of test data. The for-
• Shear span (a) mula uses the concept of shear friction theory and assumes
• Cylinder compressive strength of concrete ( f ⬘)
c that, at failure, the web steel is at yield. The shear friction
• Yield strength of horizontal steel ( fyh) theory assumes that the concrete may crack in an unfavorable
• Yield strength of vertical steel ( fyv) manner or that slip may occur along a predetermined plane of
• Reinforcement ratio of horizontal tensile steel (␳h) weakness such as an existing or potential crack, interface be-
• Reinforcement ratio of total horizontal steel (␳ht) tween dissimilar materials, or interface between two concrete
• Reinforcement ratio of transverse steel (␳ v) cast at different times. Reinforcement in such a case should
be designed to cross the potential or actual crack or shear plane
In addition to those listed above, there are two more param- to prevent direct shear failure. The formula given in section
eters that are also critical in deep beam behavior. These are 11.8 of the ACI code for the nominal shear strength of rein-
anchorage of longitudinal steel into supports and size of bear- forced-concrete deep beams with web reinforcement is
ing and loading areas. However, the values of these parameters
were not available in most of the experimental data collected Vn = ␾ ⫻ (Vc ⫹ Vs) (1)
from the literature. Hence they are excluded from the input
in which the shear strength is computed by
data set used in training the artificial neural network in this
study.
The experimental data include 111 deep beam results, which
are taken from the tests carried out by Mau and Hsu (1989),
Vc = 冋3.5 ⫺ 2.5
Mu
Vu d
册冉
1.9兹f ⬘c ⫹ 2,500␳w
Mu

Vu d
bw d (2)

Subedi et al. (1986), Kong et al. (1995), ‘‘Discussion’’ (1995), which should not be greater than 6兹f ⬘b c w d. The upper bound
Kang et al. (1995), and Ramakrishnan and Ananthanarayana on the term within the first bracket is 2.5. Furthermore the
(1968). It is apparent that each set of experimental data is in shear strength Vn for the deep beam should not be greater than
a different format. After a thorough study of the tables and 8兹f ⬘b
c w d when ln /d < 2. When ln /d is between 2 and 5, the
diagrams given in the above references, the data are rearranged following expression is to be used for Vn :
JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001 / 819

J. Struct. Eng., 2001, 127(7): 818-828


Vn =
2
3 冉10 ⫹
d冊
ln
兹f ⬘b
c wd (3)
equations in the strut-and-tie model. The expression is dimen-
sionless and contains four variables that express the horizontal
and vertical reinforcement ratios, concrete strength, and shear-
Note that (2) and (3) are in inches and pounds whereas this span ratio. The constants in the formula were calibrated using
paper is in SI units. Appropriate unit conversion is carried out the experimental data available in the literature. The expres-
when the equation is used in the computation of the shear sion for the shear strength of the deep beam is as follows:
strength of deep beams.
Vn = v bw d (7)

冋 冉 冊 冉 冊册
The shear strength Vs is computed by
ln ln v 1
1⫹ 11 ⫺ = [K(wh ⫹ 0.03)
Av d Av h d f ⬘c 2
Vs = ⫹ fy d (4)
s 12 s2 12 ⫹ 兹K 2(wh ⫹ 0.03)2 ⫹ 4(wh ⫹ 0.03)(wv ⫹ 0.03)] ⱕ 0.3 (8)
with the following limitations:
where Vn = nominal shear strength of the deep beam; ␾ =
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

shear strength reduction factor, taken as 0.85; Vc and Vn = wh = ␳h fy /f ⬘c ⱕ 0.26; wv = ␳v fy /f ⬘c ⱕ 0.12 (9a,b)
shear strengths provided by concrete and shear reinforcements, K = 2d/h, for 0 ⱕ a/h ⱕ 0.5 (10a)

冋冉 冊册
respectively; Vu and Mu = factored shear force and moment at
the critical section; d = effective depth of the section; bw = d h 4 2a
K= ⫺ , for 0.5 < a/h ⱕ 2 (10b)
width of the section; ln = clear span of concrete; f ⬘c = cylinder h a 3 3h
compressive strength of concrete; ␳w = ratio of flexural tensile
reinforcement; Av = area of shear reinforcement perpendicular K = 0, for a/h > 2 (10c)
to flexural tension reinforcement within a distance s; Av h = where d and bw = effective depth and width of the beam; f ⬘c =
area of shear reinforcement parallel to flexural reinforcement cylinder compressive strength of concrete; h = total depth of
within a distance s2 ; and fy = yield strength of shear reinforce- the beam; a = shear span; K = ratio of the effective compres-
ment.
Validation of the ACI formula was carried out by a number
of researchers. It was reported in Rogowski et al. (1986) that
there is little agreement between the experimental values of
shear strength and those predicted by the ACI code formula.
In general, test-to-code ratios were conservative for simply
supported beams with a mean-test/code ratio of 1.88. It is ob-
served in Sanad (1997) that, for high-strength concrete deep
beams within the ranges of compressive strengths of 43–96
MPa, shear-span-to-depth ratio of 0.22–1.5, and slenderness
ratio (h/bw) of 4–50, the ACI method was overwhelmingly
conservative (almost 2) in all cases. Mau and Hsu (1989) have
carried out a number of experiments on deep beams and com-
pared test results with those of the ACI method. The mean
value of computed shear strength to the experimental one was
0.507. It is clear that the ACI formula predicts the shear
strength of deep beams on the conservative side, with a very
large safety margin.

Strut-and-Tie Method
This method was proposed by Wen (1993) based on a strut-
and-tie model. The proposed formula is
Vu = 1.8 ft bw d (5)
ft = 6.96兹f ⬘[1
c ⫹ m(␳k sin2␪ ⫹ ␳v cos2␪)] (6)
where Vu = shear strength of the deep beam; ft = allowable
tensile strength of concrete; bw and d = width and effective
depth of the beam; f c⬘ = cylinder compressive strength of con- FIG. 2. Artificial Neural Network: (a) Neural Network; (b) Processing
crete; m = modular ratio of steel reinforcement to concrete; ␳h Element
and ␳v = steel ratios of horizontal and vertical steel reinforce-
ment; and ␪ = angle with the horizontal made by the line
joining the edges of the loading plates and supports. TABLE 1. Ranges of Parameters in Database
It was reported by Wen that the formula is fairly accurate,
with an actual/predicted ratio of 0.83, in computing the shear Input parameters Range
strength of deep beams with a shear span to depth ratio <1.04. L/d 0.95–5.4
But as this ratio becomes >1.29, the validity of the equation d/bw 2.83–47
declines rapidly because the behavior of the beam starts to a/d 0.22–2.16
deviate from that of the deep beam. f c⬘ 12.5–76
fy h 250–600
fy v 0–460
Mau-Hsu Method ␳h 0.05–1.94
␳ht 0.14–2.95
Mau and Hsu (1989) derived an explicit expression for the ␳v 0–2.45
shear strength of deep beams, using the three equilibrium
820 / JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001

J. Struct. Eng., 2001, 127(7): 818-828


sive stress in the transverse direction to the effective shear Other Methods
stress in the shear element; and wh and wv = reinforcement There are a number of other methods in the literature sug-
indices in the horizontal and vertical directions, respectively. gested for the prediction of the shear strength of reinforced-
Mau and Hsu stated that their formula is to be used within concrete deep beams. Some of these are outlined in the fol-
the following limitations: lowing.
The CEB-FIP model code (1990) uses the strut-and-tie
• Span/effective-depth ratio (L/d) should be between 0.95 method (ASCE-ACI 1998). The code allows the use of the
and 3.3 lower-bound approach of the theory of plasticity. The com-
• Reinforcement ratio of horizontal web steel (␳v h) should pressive stress in the strut is limited by the value calculated
be between 0 and 0.0091 using the effectiveness factor, which is a function of the char-
• Reinforcement ratio of vertical steel (␳v) should be be- acteristic cylinder strength. The required minimum amount for
tween 0.0018 and 0.0245 the shear reinforcement in the webs of beams is expressed in
• Reinforcement ratio of compression steel (␳c) should be terms of a mechanical reinforcing ratio rather than a geomet-
between 0 and 0.0092 rical value minimum ␳v .
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

• Concrete cylinder compression strength ( f ⬘)


c should be The Construction Industry Research and Information As-
close to 21 MPa sociation (CIRIA) method (1984) determines the shear
strength of deep beams with a clear-span-to-depth ratio of 0–
TABLE 2. Ultimate Shear Strength Values for Deep Beam Specimens 0.7. The method requires the depth of flexural cracks and the
Obtained by Neural Networks
angle these cracks make with the steel bars as input informa-
Propagator Software Predict Software tion. Such information was not available in the experimental
data collected.
Specimen Error Error
number (%) Vpred /Vexp (%) Vpred /Vexp
There are other methods reported in the literature, such as
the formula of Ramakrishnan and Ananthanarayana (1968) and
3 ⫺24.5 0.75 ⫺9.1 0.91 Zsutty’s equation (1971). None of the above methods were
7 ⫺23.0 0.77 ⫺1.6 0.98
23 ⫺13.2 0.87 0.2 1.00 selected in the comparative study because either they require
36 ⫺4.4 0.96 10.7 1.11 additional information that is not available in the experimental
45 ⫺19.6 0.80 2.1 1.02 data collected or they neglect the effect of one or more of the
80 ⫺16.1 0.84 0.5 1.01 selected input parameters that are used in the training of the
86 17.4 1.17 20.6 1.21 neural network.
93 ⫺5.1 0.95 0.6 1.01
95 ⫺15.0 0.85 ⫺0.4 1.00 NEURAL NETWORK MODELING
99 ⫺9.2 0.91 7.4 1.07
Average ⫺11.27 0.887 3.10 1.032
An artificial neural network is a computational tool that at-
tempts to simulate the architecture and internal operational fea-

FIG. 3. Comparison of Shear Strengths of Deep Beams Obtained by Various Methods: (a) Strut-and-Tie Method; (b) ACI Code Method; (c) Mau-Hsu
Method; (d) Neural Network

JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001 / 821

J. Struct. Eng., 2001, 127(7): 818-828


tures of the human brain and nervous system. It consists of a 1997). A review of papers on the same topic can be found in
number of processing elements, which represent neurons that Waszczyszyn (1996). These references reveal the fact that a
are connected to each other to constitute a network. Each pro- multilayer feed-forward neural network model is the most
cessing element can have many inputs, but it sends only one widely used network for its efficient generalization capabili-
output. Similar to the real behavior of a neuron, if the sum of ties. A typical sample of such a neural network is shown in
inputs to a given processing element exceeds a set threshold, Fig. 2(a) and consists of an input layer, output layer, and one
it fires an input to other neighboring processing elements. hidden layer.
Weights are applied to input signals, which are the strength of The training process of a multilayer feed-forward neural
the connection between processing elements. The values of network consists of an iterative process in which the values
these weights are determined by using a number of training of the connection weights are determined (Hertz et al. 1991).
input and output values. The neural network gains its problem- Among the available training methods, back-propagation is the
solving ability through this process. An excellent description most successful and widely used in neural network applica-
of the basic anatomy of different types of neural networks is tions (Kaveh and Khalegi 1998). In this method, the input is
given in Rumelhart et al. (1986) and Hertz et al. (1991). propagated from the input layer through the hidden layers to
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

The application of neural networks in civil and structural the output layer. The error between the predicted and expected
engineering covers a range of diverse areas (Kartam et al. output values is then back-propagated from the output layer to
the input layer in which the connection weights are modified.
TABLE 3. Performance of Methods Selected for Predicting Shear This process is repeated until the error is minimized.
Strength of 10 Deep Beams Two different neural network software programs were used
Predicted Strength/Actual Strength to develop an artificial neural network by which the ultimate
shear strength of a reinforced-concrete deep beam can be pre-
Specimen ACI Truss Mau-Hsu Neural
dicted. These were professional versions of Propagator (Prop-
number method method method network
agator 1996) and Predict (Predict 1997). One hundred one of
3 0.33 0.89 0.97 0.91 the experimental data given in the Appendix were used to train
7 0.52 0.84 0.94 0.98
the network, and the remaining 10 were employed in testing
23 0.47 1.04 1.10 1.00
36 0.49 1.26 0.95 1.11 the network. Before feeding the data into the network, some
45 0.32 1.03 0.97 1.02 of the basic input parameters selected earlier are converted into
80 0.44 0.76 1.67 1.01 the following nondimensional form:
86 0.67 2.14 2.04 1.21
93 0.36 0.86 1.48 1.01
95 0.23 1.62 1.30 1.00 • Effective-span/effective-depth ratio (L/d )
99 0.31 1.91 1.45 1.07 • Effective-depth/breadth ratio (d/bw)
Average 0.41 1.24 1.29 1.03
• Shear-span/effective-depth ratio (a/d )
• Cylinder compressive strength of concrete ( f ⬘)
c

FIG. 4. Variation of Actual Strength to Predicted Strength with L/d Ratio: (a) ACI Code Method; (b) Strut-and-Tie Method; (c) Mau-Hsu Method;
(d) Neural Network

822 / JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001

J. Struct. Eng., 2001, 127(7): 818-828


• Yield strength of longitudinal steel ( f yh) until their best combination was obtained. After a number of
• Yield strength of transverse steel ( f yv) trials, the best net was determined, which has the following
• Reinforcement ratio of horizontal tensile steel (␳h) features:
• Reinforcement ratio of total horizontal steel (␳ht)
• Reinforcement ratio of transverse steel (␳v) • The net consists of four layers: input layer of 10 nodes,
two hidden layers of 6 and 5 nodes, and output layer of
The output was selected as the ultimate shear strength 1 node.
(V/bd ) of the deep beam. The ranges of these parameters in • The linear transfer function is used for the proceeding
the experimental database used are given in Table 1. elements in the input layer, and the sigmoidal transfer
When a nonlinear transfer function such as sigmoid or tanh function is used in others.
are used in a network, it becomes necessary to scale the orig- • A learning rate of 0.4 and momentum factor of 0.2 pro-
inal data before training the network. The reason for scaling vide the best combination.
is that the nonlinear transfer functions work best over a limited
range. For a sigmoid function, which was used in this study, The results obtained using this net for 10 specimens re-
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

the output numbers are between 0 and 1. Hence, an appropriate served for testing the network are given in Table 2. The net-
scaling conversion (Sanad 1997) is applied to the values of work does not yield values within the desired tolerance limits.
the parameters so that they range within 0 and 1. However, these results are found to be better than those ob-
tained from the ultimate shear strength formulas suggested in
Propagator Software the ACI, strut-and-tie, and Mau-Hsu methods.
The Propagator neural network software version 1.0 has the Predict Software
facility of scaling the inputs before feeding them into the net-
work. The program requires a training file, validation file, and The professional version of this software is fully automated,
testing file. It displays the best training and validation errors and it selects subparameters from a general list of input pa-
and the cycle in which the best weights for training are ob- rameters. The data can be entered by MS Excel, which pro-
tained. Several networks were created by varying the number vides easy input. The software requires the selection of certain
of layers, number of nodes in each layer, learning rate, and parameters prior to its use for training, which are listed below:
momentum factor. The number of layers and number of nodes
in each layer were first fixed, and the learning rate and mo- • Input variable selection—This section uses a genetic al-
mentum factor were varied from 0 to 1 by increments of 0.1 gorithm to determine a good subset of the full set of input

FIG. 5. Variation of Actual Strength to Predicted Strength with a/d Ratio: (a) Strut-and-Tie Method; (b) ACI Code Method; (c) Mau-Hsu Method;
(d) Neural Network

JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001 / 823

J. Struct. Eng., 2001, 127(7): 818-828


variables. Genetic algorithms are a recent addition to nu- combination that is used in the network for predicting the
merical search techniques for finding the minimum of a shear strength of deep beams was very noisy data, compre-
constrained or unconstrained function. This new class of hensive data transformation, moderate variable selection, and
algorithms is developed by imitating living beings. The exhaustive network search of the aforementioned parameters.
structure of the genetic algorithm is based on the principle The input and hidden layers have seven processing elements
of survival of the fittest (Goldberg 1989). The parameters while the output layer has only one. The ultimate shear
required to be selected in this section are variable selec- strength values obtained form this network for 10 deep beam
tion model, convergence, population size factor, and re- specimens reserved for testing the network are given in Table
production factors. 2. It is clear that the network developed from the Predict soft-
• Network parameters—This section decides the architec- ware performs extremely well within an average error of 3.1%
ture of the network, such as the number of processing for out-of-training data.
elements in input, hidden, and output layers.
• Learning parameters—This software supports two learn- ANALYSIS OF RESULTS
ing rules, adoptive gradient and Kalman filter.
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

The neural network developed using Predict software is em-


The main variables to be selected before creating the net- ployed to predict the shear capacity of deep beam specimens
work in Predict software are collected under the following listed in the Appendix, together with three methods, namely,
titles. Each title has a number of options available for selec- ACI, strut-and-tie, and Mau-Hsu. The results obtained are
tion. These are shown in Fig. 3. It is clear from the results that the ACI
method is very conservative in predicting the actual strength
• Noise—Noisy, moderate, very noisy whereas the Mau-Hsu method underestimates the actual
• Data transformation—Moderate, comprehensive superfi- strength beyond about 300 kN. The strut-and-tie method pro-
cial, scale only duces wide variations on either side of the desired ratio of 1.
• Variable selection—Moderate, comprehensive, exhaus- The average ratio of actual strength to predicted strength of
tive, superficial all specimens is 2.07 in the ACI method, 0.85 in the strut-and-
• Network search—Moderate, comprehensive, exhaustive, tie method, 0.84 in the Mau-Hsu method, and 0.99 in the neu-
superficial ral network. Table 3 shows the same ratio for only 10 beams,
which were not used in training but were reserved for testing
Several combinations of the above parameters were tested the network. The average ratio of actual strength to predicted
until the best results were obtained from the network. The final strength in this set of data is 2.44 for the ACI method, 0.81

FIG. 6. Variation of Actual Strength with f ⬘c : (a) Neural Network; (b) Strut-and-Tie Method; (c) Mau-Hsu Method; (d) ACI Code Method

824 / JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001

J. Struct. Eng., 2001, 127(7): 818-828


TABLE 4. Shear Strength Values Obtained by Various Methods for methods. The actual and predicted strengths of the beam are
Specimen 12 listed in Table 4.
Actual strength Predicted strength Error The results presented above clearly demonstrate the accu-
Method (N) (N) (%) Vpred /Vexp racy and efficiency of the trained neural network method over
ACI 172,716 100,836.7 ⫺41.6 0.58
the other three methods.
Truss 172,716 153,758.3 ⫺11.0 0.89
Mau-Hsu 172,716 163,967.3 ⫺5.1 0.95 CONCLUSIONS
Propagator 172,716 162,118.6 ⫺6.1 0.94
Predict 172,716 176,051.2 1.9 1.02 A comparative study was carried out for the use of multi-
layer feed-forward neural networks in predicting the ultimate
shear strength of simply supported deep beams subjected to
for the strut-and-tie method, 0.78 for the Mau-Hsu method, two point loads acting symmetrically with respect to the cen-
and 0.97 for the neural network. These values clearly show terline of the span. It is found that the strength values obtained
that the neural network performs much better than the methods from the artificial neural network are much more accurate than
selected in this study. those determined from ACI code, strut-and-tie, or Mau-Hsu
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

The variation of the ratio of actual strength to predicted methods. Although the average value of the ratio of actual
strength with the L/d ratio is shown in Fig. 4. A large variation strength to predicted strength was 2.07 for all deep beams in
in the accuracy is noticed in the predictions made by the ACI the ACI method, 0.85 in the strut-and-tie method, and 0.84 in
method for L/d ratios, particularly between 3 and 3.5. The the Mau-Hsu method, it was only 0.99 in the neural network.
strength predictions determined by the strut-and-tie and Mau- These average ratios change to 2.44, 0.81, 0.78, and 0.97,
Hsu methods seem to be unaffected by the variation of the respectively, when the methods are employed for 10 beams
L/d ratio. Results obtained from the neural network indicate that are not used in training the network. These results clearly
consistent accuracy in all ranges of the L/d ratios. indicate the accuracy of the neural network in predicting the
The variation of the ratio of actual strength to predicted shear capacity of deep beams. Furthermore comparison has
strength with the L/d ratio in four methods is shown in Fig. revealed that although the ACI, strut-and-tie, and Mau-Hsu
5. Once again, the ACI method yields large variations. The methods were affected with the variations of L/d and a/d ratios
strut-and-tie and Mau-Hsu methods perform better compared and the compressive strength of concrete, neural network per-
to the ACI method. The neural network results are not affected formance was unaffected by these variations.
by the variations in the a/d ratio. It should be pointed out that the data collected from
The ratio of actual strength to predicted strength is plotted the literature have not covered the ranges of L/d > 3.5, 1.25
against the cylinder compressive strength of the concrete f ⬘c for < a/d < 1.4, and a/d > 1.75. It is therefore not possible
the methods considered in this study. The graphs are shown to state that the network performs correctly for all ranges.
in Fig. 6. It is apparent from the graphs that the ACI method However, this discrepancy can easily be eliminated if all
overestimates the actual strength excessively for beams made ranges are included in the training data. Furthermore, in this
out of concrete having a compressive strength <40 N/mm2. study, simply supported deep beams under two concentrated
The strut-and-tie and Mau-Hsu methods are mostly unaffected loads are considered. Beams under distributed loads or contin-
by the variation of f c⬘ except in some isolated cases. Among uous deep beams are not covered. However, from the results
these, the results obtained from the neural network is the con- obtained, it can be concluded that, with proper training, neural
sistent one, having values close to 1 for a wide variation of networks can provide an excellent alternative for predicting
compressive strengths of concrete. the strength capacities of structural members, particularly in
Finally, the deep beam specimen No. 12 is selected from cases where no accurate single expression exists for the pre-
the database and its strength is determined using different diction.

JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001 / 825

J. Struct. Eng., 2001, 127(7): 818-828


APPENDIX

TABLE 5. Experimental Data for Ultimate Shear Strength of Deep Beams


Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

826 / JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001

J. Struct. Eng., 2001, 127(7): 818-828


TABLE 5. (Continued )
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

Note: Specimens chosen for testing neural network have been highlighted.

JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001 / 827

J. Struct. Eng., 2001, 127(7): 818-828


REFERENCES reinforced concrete deep beams in current practice.’’ The Struct. Engr.,
London, 53(4), 173–180.
American Concrete Institute (ACI). (1995). ‘‘Building code requirements Mau, S. T., and Hsu, C. T. (1989). ‘‘Formula for shear strength of deep
for reinforced concrete.’’ ACI 318-95, Detroit. beams.’’ ACI Struct. J., 86(8), 516–523.
Arslan, A., and Ince, R. (1994). ‘‘Neural network based design of edge Predict professional. (1997). Neural Ware Inc., Pittsburgh.
supported reinforced concrete slabs.’’ Artificial intelligence and object Propagator: version 1.0. (1996). A. R. D. Environmental Inc., Md.
oriented approaches for structural engineering, B. H. V. Topping, ed., Ramakrishnan, V., and Ananthanarayana, Y. (1968). ‘‘Ultimate strength
Civil-Comp Press, Edinburgh, U.K., 91–97. of deep beams in shear.’’ ACI Struct. J., 65(2), 87–98.
ASCE-American Concrete Institute (ACI) 445 on Shear and Torsion. Rumelhart, D. E., McClelland, J. L., and Parallel Distributed Processing
(1998). ‘‘Recent approaches to shear design of structural concrete.’’ J. Research Group. (1986). Parallel distributed processing explorations
Struct. Engrg., ASCE, 124(12), 1375–1417. in the microstructure of cognition, 1st Edition, Vol. 1, MIT Press, Cam-
Asin, M., and Walraven, J. (1995). ‘‘Numerical analysis of reinforced bridge, Mass.
concrete continuous deep beams.’’ Heron, 40(2), 163–178. Rogowsky, D. M., MacGregor, J. G., and Ong, Y. (1986). ‘‘Tests on
Buenfeld, N. R., and Hassanein, N. M. (1998). ‘‘Predicting the life of reinforced concrete deep beams.’’ ACI Struct. J., 83(4), 614–623.
concrete structures using neural network.’’ Proc., Instn. Civ. Engrs., Sanad, A. (1997). ‘‘Design of reinforced concrete members using neural
Struct. and Build., London, 128(February), 38–48. networks.’’ MSc thesis, University of Bahrain, Isa Town, Bahrain.
Chuang, P. H., Goh, A. T. C., and Wu, X. (1998). ‘‘Modeling the capacity
Downloaded from ascelibrary.org by Arba Minch University on 05/14/20. Copyright ASCE. For personal use only; all rights reserved.

Shin, S.-W., Lee, K.-S., Moon, J.-I., and Ghosh, S. K. (1999). ‘‘Shear
of pin-ended slender reinforced concrete columns using neural net- strength of reinforced high-strength concrete beams with shear span-
works.’’ J. Struct. Engrg., ASCE, 124(7), 830–838. to-depth ratios between 1.5 and 2.5.’’ ACI Struct. J., 96(4), 549–556.
Comité Euro-International du Béton-Fédération International de la Pré- Smith, K. N., and Vantsiotis, A. C. (1982). ‘‘Shear strength of deep
contrainte (CEB-FIP). (1990). Model code for concrete structures, Ce- beams.’’ ACI Struct. J., 79(3), 201–213.
ment and Concrete Association, London. Subedi, N. K. (1988). ‘‘Reinforced concrete deep beams—A method of
Construction Industry Research and Information Association (CIRIA). analysis.’’ Proc., Instn. Civ. Engrs., London, Part 2, 128, 12–25.
(1984). Guide 2: The design of deep beams in reinforced concrete, Ove Subedi, N. K., Vardi, A. E., and Kubata, N. (1986). ‘‘Reinforced concrete
Arup and Partners, London.
deep beams—Some test results.’’ Mag. of Concrete Res., London,
‘‘Discussion of ‘The design of reinforced concrete deep beams in current
38(137), 206–219.
practice,’ by F.-K. Kong, P. J. Robins, and G. R. Sharp.’’ (1975). The
Tan, K. H., and Lu, H. Y. (1999). ‘‘Shear behavior of large reinforced
Struct. Engr., London, 53(7), 300–301.
concrete deep beams and code comparisons.’’ ACI Struct. J., 96(5),
Goldberg, D. (1989). Genetic algorithms in search, optimization and ma-
836–845.
chine learning, Addison-Wesley, Reading, Mass.
Hajela, P., and Berke, L. (1991). ‘‘Neuro-biological computational models Topping, B. H. V., and Bahreininejad, A. (1997). Neural computing for
in structural analysis and design.’’ J. Comp. and Struct., 41(4), 657– structural mechanics, Saxe-Coburgh Publications, Edinburgh, U.K.
667. Vanluchene, D., and Roufei, S. (1990). ‘‘Neural networks in structural
Hertz, J., Krogh, A., and Palmer, R. G. (1991). Introduction to the theory engineering.’’ J. Microcomputers in Civ. Engrg., 5(3), 207–215.
of neural computing, Addison-Wesley, Redwood City, Calif. Waszczyszyn, Z. (1996). ‘‘Some recent and current problems of neuro-
Hwang, S.-J., Lu, W.-Y., and Lee, H.-J. (2000). ‘‘Shear strength prediction computing in civil and structural engineering.’’ Advances in computa-
for deep beams.’’ ACI Struct. J., 97(3), 367–376. tional structures technology, B. H. V. Topping, ed., Civil-Comp Press,
Jenkins, W. M. (1998). ‘‘Structural re-analysis by neural network.’’ Ad- Edinburgh, U.K., 43–58.
vances in Engineering Computational Technology, B. H. V. Topping, Waszczyszyn, Z. (1998). ‘‘Some new results in application of back prop-
ed., Civil-Comp Press, Edinburgh, U.K., 229–237. agation neural networks in structural and civil engineering.’’ Advances
Kang, H. T., Kong, F.-K., Susanto, T., and Lingwei, G. (1995). ‘‘High in computational structures technology, B. H. V. Topping, ed., Civil-
strength concrete deep beams with effective span and shear span var- Comp Press, Edinburgh, U.K., 173–187.
iations.’’ ACI Struct. J., 92(4), 395–405. Wen, B. S. (1993). ‘‘Strut-and-tie model for shear behavior in deep beams
Kartam, N., Flood, I., and Garrett, J. H. (1997). Artificial neural networks and pile caps failing in diagonal splitting.’’ ACI Struct. J., 90(4), 356–
for civil engineers: Fundamentals and applications, ASCE, New York. 363.
Kaveh, A., and Khalegi, A. (1998). ‘‘Prediction of strength for concrete Wu, X., Ghaboussi, J., and Garrett, J. H. (1992). ‘‘Use of neural networks
specimens using artificial neural networks.’’ Advances in computational in detection of structural damage.’’ J. Comp. and Struct., 42(4), 649–
structures technology, B. H. V. Topping, ed., Civil-Comp Press, Ed- 659.
inburgh, U.K., 165–171. Zsutty, T. (1971). ‘‘Shear strength prediction for separate categories of
Kong, F.-K., Robins, P. J., and Sharp, G. R. (1975). ‘‘The design of simple beam tests.’’ ACI J. Proc., 68(2), 138–143.

828 / JOURNAL OF STRUCTURAL ENGINEERING / JULY 2001

J. Struct. Eng., 2001, 127(7): 818-828

S-ar putea să vă placă și