Vous êtes sur la page 1sur 15

644

A neural network model for predicting maximum shear capacity of concrete beams without transverse reinforcement

Ayman Ahmed Seleemah

Abstract: Different relationships have been proposed by codes and researchers for predicting the shear capacity of members without transverse reinforcement. In this paper, the applicability of the artificial neural network (ANN) tech - nique as an analytical alternative to existing methods for predicting this shear capacity is investigated using a critically reviewed and agreed upon database of experimental work that serves as a basis of comparison and (or) assessment of existing and new relationships. Both ANN and eight different codes and researcher’s predictions of the shear capacity of the specimens of the database were compared. The ANN predictions are much superior to those of any of the cur - rent available relationships.

Key words: artificial neural networks, shear capacity, transverse reinforcement, beams.

Résumé : Les codes et les chercheurs ont proposé diverses relations pour prédire la résistance au cisaillement de mem - brures sans armature transversale. Le présent article étudie l’applicabilité de la technique des réseaux neuronaux artifi - ciels comme alternative analytique aux méthodes existantes de prédiction de la résistance au cisaillement. Nous avons utilisé pour cela une base de données du travail expérimental acceptée et examinée scrupuleusement pour servir de base de comparaison et-ou d’évaluation des relations existantes et nouvelles. Nous avons comparé les prévisions émises par les réseaux neuronaux artificiels, huit différents codes et des chercheurs concernant la résistance au cisaillement d’échantillons de la base de données. Nous avons trouvé que les prévisions des réseaux neuronaux artificiels sont gran- dement supérieures à celles de toute autre relation actuellement disponible.

Mots clés : réseaux neuronaux artificiels, résistance au cisaillement, armature transversale, poutres.

[Traduit par la Rédaction] Seleemah 657

Introduction

Civil and structural engineers attempt to improve the anal - ysis, design, and control of the behavior of structural sys - tems. The behavior of structural systems, however, is complex and often governed by both known and unknown multiple variables, with their interrelationship generally un - known, nonlinear, and sometimes very complicated. The tra - ditional approach used in most research in modeling starts with an assumed form of an empirical or analytical equation and is followed by a regression analysis using experimental data to determine unknown coefficients such that the equa - tion will fit the data. In the last two decades, researchers explored the potential of artificial neural networks (ANNs) as an analytical alterna - tive to conventional techniques, which are often limited by strict assumptions of normality, linearity, homogeneity, vari -

Received 22 April 2004. Revision accepted 4 January 2005. Published on the NRC Research Press Web site at http://cjce.nrc.ca on 3 August 2005.

A.A. Seleemah. Department of Civil Engineering, Benha High Institute of Technology, Benha, Egypt (e-mail:

seleemah55@yahoo.com).

Written discussion of this article is welcomed and will be received by the Editor until 31 December 2005.

able independence, etc. Researchers found ANNs particu- larly useful for function approximation and mapping prob- lems, which are tolerant of some imprecision and have a considerable amount of experimental data available. In a strict mathematical sense, ANNs do not provide closed-form solutions for modeling problems but offer a complex and ac - curate solution based on a representative set of historical ex - amples of the relationship. Advantages of ANNs include the ability to learn and generalize from examples, produce meaningful solutions to problems even when input data con - tain errors or are incomplete, adapt solutions over time to compensate for changing circumstances, process information rapidly, and transfer readily between computing systems (Flood and Kartam 1994). This paper focuses on the prediction of the ultimate shear strength of reinforced concrete beams. This topic is still an active area of research because of the complexity of the shear transfer mechanism and the large number of affecting parameters. While many efforts have been conducted to un - derstand the shear behavior of reinforced concrete beams and (or) to derive equations for estimating such shear capac - ity, some researchers explored the application of ANNs for such predictions. For example, Sanad and Saka (2001) applied ANNs to predict the ultimate shear capacity of rein - forced concrete deep beams; Mansour et al. (2004) success - fully used ANNs to predict the shear capacity of reinforced concrete beams with transverse reinforcement; and Oreta

Seleemah

(2004) applied ANNs to a set of 155 experimental tests to simulate the size effect on the shear strength of reinforced concrete beams without transverse reinforcement. In the cur - rent paper, one of the largest and most confident databases is utilized to investigate the applicability of the ANN tech - nique to predict the concrete contribution to shear resistance of beams. Moreover, a comparison of both ANN and differ - ent codes and researcher’s predictions for this important structural criteria is conducted.

Shear capacity of beams without transverse reinforcement

Different relationships have been proposed by researchers all over the world for predicting the shear capacity of beams without transverse reinforcement. The importance of these relationships is that they represent the concrete contribution to the overall shear resistance of a reinforced concrete beam with transverse reinforcement, which is a common design case.

Unfortunately, these proposed relationships are usually empirical and designed to fit a limited set of shear test re - sults that is most familiar to the researcher. Moreover, the relationships differ quite considerably in their selected pa- rameters because there is no generally accepted model for the load transfer and the ultimate capacity of members with- out transverse reinforcement. This fact was discussed in the two state-of-the-art reports by the Joint ASCE–ACI Com- mittee 445 on Shear and Torsion (1998) and the CEB (1997). Furthermore, the test data used for comparing pro- posals with tests differ very much from one researcher to an- other in the amount and quality of selected data. Therefore, it is essential to collect a unique and critically reviewed database of experimental work that would serve as

a basis of comparison and (or) assessment of assumed rela -

tionships. Moreover, it is important also to explore other methods for predicting the shear capacity that would give re - sults superior to those suggested relationships. In this paper, the capability of ANNs as an analytical alternative to con- ventional techniques for predicting the ultimate shear capac - ity of concrete beams is investigated.

Shear database

A databank established by Reineck (1999) and checked by

a group of experts was used to verify the empirical equation

for members without transverse reinforcement proposed for the German standard DIN 1045-1 (DIN 2001). Kuchma (2000) conducted a wide-range collection of test data on members both with and without transverse reinforcement. Recently, both authors have merged their databanks into a collection shear databank (CSDB) that contains mainly beams with a rectangular cross section (933 beams, 231 of which have a shear span to beam depth ratio of less than

2.40).

A subcommittee was then established within the Joint ASCE–ACI Committee 445 on Shear and Torsion to come to

a consensus on criteria that must be satisfied if a specific test result is to be used for assessing the capability of empirical design rules or expressions derived from models for struc -

645

Table 1. Ranges of parameters in the database.

b w (mm)

76.2–1000.0

d (mm)

110–2000

a / d

2.41–8.03

d / b w

0.25–7.20

M / V d

1.41–7.03

f 1c (MPa)

12.6–110.9

f 1t (MPa)

1.27–5.29

f y (MPa)

275.9–1779.3

ρ l (%)

0.138–6.635

Max. aggregate size (mm)

6.35–38.00

Note: M, maximum bending moment; f 1t , uniaxial con- crete tensile strength; ρ l , percentage of main steel. All other terms as defined in the text.

tural behavior. A sanctioned set of criteria was developed for accepting a test result into an evaluation-level database. Examples of these criteria include a minimum compressive strength, a minimum overall height and width, and a check against flexural failure. A databank of members that satisfy these agreed-upon criteria was established and called the evaluation shear databank (ESDB) (Reineck et al. 2003). This databank contained 439 shear tests collected from 64 references. All beams in this database have a rectangular cross section, do not contain shear reinforcement, and were subjected to point loads. Extensive discussions and review on the ESDB were then conducted by the American Con- crete Institute (ACI) Subcommittee 445-F on Beam Shear, which led to some changes in the database, and a total of 41 tests were excluded for several reasons. The revised version (398 tests) was intended to serve as a basis for any code changes. This database is utilized in this study to evaluate and dem- onstrate the capability of the ANN technique for predicting the shear capacity of members without transverse reinforce - ment. Moreover, it was also used to evaluate eight different codes and formulas that exist for predicting such shear ca - pacity, and several comparisons of the predictions of these methods and those of the ANNs are conducted. It should be pointed out that the aforementioned database covers a very wide range of beam depth, beam breadth, shear span to beam depth ratio, maximum aggregate size used in concrete and its tensile and compressive strengths, main reinforcement percentage, and yield stress. Table 1 summarizes the ranges of the parameters in the database.

Existing models for shear capacity prediction

Eight different major codes and methods of estimation of the shear capacity of reinforced concrete beams without shear reinforcement were collected and are as follows: ACI Committee 318 (1999), BSI (1990), ECS (1992), NZS (1995), Zsutty (1971), Collins and Kuchma (1999), Niwa et al. (1986), and Reineck (1991). The expressions utilized by each of these methods are summarized in Table 2. These methods were applied to the whole database (398 speci - mens) to achieve the prediction of each method for the shear capacity of each specimen.

© 2005 NRC Canada

646

Can. J. Civ. Eng. Vol. 32, 2005

Table 2. Summary of the expressions to calculate the concrete contribution to shear resistance.

Reference

Equation

 

Remarks

ACI Committee 318 1999

BSI 1990

V

c

V

c

=

=

( F
(
F

/)6

bd

cw

0.27 b d

w

γ

m

(

100 A / bd

s

w

13

/

)(

f

cu

)

/

13

 

f cu 40.0 MPa

ECS 1992

NZS 1995

V rd1 = [ τ Rd k (1.2 + 40 ρ l )] b w d

V

c

=+( 0.07

10 ρ )

l

′

f bd

cw

 

k = (1.6 – d /1000) 1.0; ρ l = A s / b w d 0.02 a / d 2

Zsutty 1971

1 / 3

b d

a / d 2.5

Collins and Kuchma 1999

Niwa et al. 1986

V

u

V

c

V

=

=

=

2.2 ( f

ρ

cl

da /)

w

245 c 1275 + S E
245
c
1275 + S
E

f bd

w

where S

0.2

bd

(

f

c

13

/

)(

100

A bd

/

sw

)

/

13

E

=

35

Sa /( + 16 )

x

1.4

S x 0.9 d

Reineck 1991

c

w

d 1000

(/

)

1

/ 4

0.75 +

a / d

V u = (0.4 b w df ct + V du )/(1 + 0.54 λ ); f ct = 0.246( f c ) 2/3 ;

E s = 2 000 000 MPa

V du / b w df c = 1.33 ρ 8/9 /( f c ) 2/3 ; λ = f c d / E s ρw u …, w u = 0.09

mm

Note: E s , modulus of elasticity of steel; f ct , axial tensile strength of concrete; k , constant relating to section depth and curtailment; S E , crack spacing parameter; S x , crack spacing; V c , nominal shear strength; V du , dowel force; V rd1 , shear resistance; V u , ultimate shear force; w u , limiting crack width; γ m , partial safety factor for strength; λ , dimension-free value for the crack width, which determines the friction capac - ity. All other terms as defined in the text.

It should be mentioned that the original values of the ba-

sic design shear strength τ Rd of the Eurocode EC2 (ECS 1992) is given by the code in a tabular form in terms of the characteristic cylinder strength f ck (in MPa). Appropriate conversions were conducted to relate this factor to the uniax- ial concrete compressive strength f 1c . Moreover, for the Brit-

ish Standard BS5400-4 (BSI 1990), the characteristic concrete cube strength, f cu , was also converted in terms of f 1c . Generally speaking, appropriate conversions for the unit system and (or) the compressive strength were conducted whenever such conversions were essential. The equations most frequently used in these conversions are as follows:

[1]

[2]

f

ck

f

1c

=

f

c

′− 1.60

= 0.95

f

c

[3]

where f c is the cylinder strength, and f c,cube is the cube strength of concrete.

f

1c

= 0.75

f

c,cube

Neural network based modeling

A neural network is a nonlinear dynamic system consist -

ing of a large number of highly interconnected processing units, called artificial neurons. Its architecture and opera - tions are inspired by the biological structure of neurons and the internal operation of the human brain. The main compu - tational characteristics of neural networks are their ability to learn functional relationships from examples and to discover patterns and regularities in data through self-organization. In this study, the multilayer feed-forward neural networks are utilized because they are very suitable for modeling the non - linear mapping type of problems. Figure 1 shows a typical

Fig. 1. Typical architecture of multilayer ANN.

a typical Fig. 1. Typical architecture of multilayer ANN. architecture of multilayer feed-forward neural networks with

architecture of multilayer feed-forward neural networks with an input layer, an output layer, and one hidden layer. As shown in Fig. 1, the artificial neurons are arranged in layers, and all the neurons in each layer have connections to all the neurons in the next layer. Associated with each con - nection between these artificial neurons, a weight value ( w i ) is defined to represent the connection weight. The connec -

© 2005 NRC Canada

Seleemah

tions between the neurons are individually weighed so that the total of i inputs ( X i ) to the single neuron is

[4]

where b is the connection weight associated with a bias node that has an input value of 1.0. The weights may be positive or negative such that some inputs will be excitatory and others will be inhibitory. This input passes through an activation function to produce the values of Y i of the hidden layer(s) or O i of the output layer. The activation function may have many forms. The most fa - miliar and effective form in our case is the sigmoid function, defined as

[5] output = 1/{[1 + exp[– α(input)]}

where α is a constant that typically varies between 0.01 and

1.00.

Signals are received at the input layer, pass through the hidden layers, and reach the output layer, producing the out - put of the network. The learning process primarily involves the determination of connection weight and bias matrices and the pattern of connections. It is through the presentation of examples, or training cases, and application of the learn - ing rule that the neural network obtains the relationship em- bedded in the data. In applications of multilayer feed-forward neural networks to engineering modeling, there are four important issues to be addressed: (i) representation of the problem in which the composition of the input and output layers must be deter- mined, (ii) determination of the configuration of hidden layer(s) through trial and error process, (iii) training of the ANN, and (iv) generalization verification of the trained ANN. The representation scheme depends on the nature of the problem and the intended use of the neural network model. Experience shows that there is no unique solution for repre- sentation schemes; different neural networks can produce similar results with the same set of training data.

input =

i

wX

i

i +

b

Input and output layers of ANN In this study, the neural networks were designed to have an input layer that consists of six input nodes representing the most important parameters that affect the shear capacity of reinforced concrete beams. Based on careful study of re - cent approaches for the shear phenomena in concrete mem - bers, it was decided to design the input layer to consist of ( i ) beam depth ( d ), ( ii ) beam breadth ( b w ), ( iii ) ratio of shear span ( a ) to beam depth ( a / d ), ( iv ) uniaxial concrete compres - sive strength ( f 1c ), ( v ) percentage of main reinforcement (ρ l ), and ( vi ) steel yield stress ( f y ). The output layer consisted of one node representing the ultimate shear capacity of the beam.

Training of the network In a multilayer feed-forward neural network, training re - fers to the iterative process involving the presentation of training data to the network, the invocation of learning rules to modify the connection weights, and, usually, the evolution of the network architecture, such that the knowledge embed - ded in the training data is appropriately captured by the weight structure of the network. During the training phase,

647

the training data consist of input and associated output pairs representing the problem that we want the network to learn.

Training and testing patterns An important factor that can significantly influence the ability of a network to learn and generalize is the number of patterns in the training set. Although it increases the time re - quired to train a network, increasing the number of training patterns provides more information about the shape of the solution surface and thus increases the potential level of ac - curacy that can be achieved by the network. Another impor - tant condition is that the training data should be well distributed within the problem domain. Carpenter and Barthelemy (1994) stated that a necessary condition for obtaining a unique approximation is to have the number of training pairs equal to or greater than the number of weights and biases associated with the network. In our case this number ranges between nine and 161 for networks with one and 20 nodes in a single hidden layer, re - spectively. It also ranges between 11 and 191 for networks with two hidden layers having one and 10 neurons per hid - den layer, respectively. Oreta and Kawashima (2003) stated that the minimum number (NT) of training data pairs, for ANNs with bias terms at the input and hidden layers, should be NT = J ( I + 1) + K ( J + 1), where I is the number of input nodes, J is the number of hidden layer nodes, and K is the number of output nodes. Some ANN experts suggest that NT should be in the range 10( I + J + K ), which yields a number between 80 and 270. Applying all suggestions to our case, we reach a number of training data pairs between nine and 270. Carpenter and Hoffman (1995) suggested that an ANN model that is 20%–50% overdetermined tends to produce reasonably good approximations. Compromising these sug- gestions and having in mind that we have a total of 398 data patterns, it was decided to use 50% of the data (199 speci- mens) for training and save the other 50% for testing or vali - dation. Training data were first selected randomly, and then they were checked to make sure that they satisfy a good dis - tribution within the problem domain.

Back-propagation learning algorithm In this study, the training phase of ANNs is implemented by using the back-propagation learning algorithm. A back- propagation network typically starts out with a random set of weights. The network adjusts its weights each time it sees an input–output pair. Each pair requires two stages: a forward pass and a backward pass. The forward pass involves pre - senting a sample input to the network and letting activations flow until they reach the output layer. During the backward pass, the network’s actual output (from the forward pass) is compared with the target output and error estimates are com - puted for the output units. The weights connected to the out - put units can be adjusted to reduce those errors. We can then use the error estimates of the output units to derive error es - timates for the units in the hidden layers. Lastly, errors are propagated back to the connections stemming from the input units. The back-propagation algorithm updates its weights incrementally, after seeing each input–output pair. After it has seen all the input–output pairs (and adjusted its weights

© 2005 NRC Canada

648

Can. J. Civ. Eng. Vol. 32, 2005

Fig. 2. Training and testing root mean square (RMS) error trace: ( a ) ANN 6-12-1; ( b ) ANN 6-4-4-1.

(RMS) error trace: ( a ) ANN 6-12-1; ( b ) ANN 6-4-4-1. many times), it

many times), it is said that one epoch has been completed. Training a back-propagation network usually requires many thousands of epochs. An error criteria for the network output is usually chosen and the maximum number of iterations is set to provide a condition for terminating the learning process. The perfor - mance of ANNs can be monitored by monitoring the train - ing error with respect to the number of iterations. If the network “learns,” the error will approach a minimum value. After the training phase, the ANNs can be tested for the other set of patterns, which the network has never seen, where the final values of the weights obtained in the training phase are used. No weight modification is involved in the testing phase.

Normalization of input and output data Neural networks usually provide improved performance when the data lie within the range (0, 1). Because a sigmoid function is used, a slow rate of learning occurs near the end points of the sigmoid function. To avoid this, all input values and associated outputs are transformed to values ranging from 0.1 to 0.9 rather than from 0.0 to 1.0. The following equation is used:

[6] XT = 0.1 + 0.8[( X X min )/( X max X min )]

where XT is the normalized value, X is the original value, X max is the maximum value of the attribute or output, and X min is the minimum value of the attribute or output.

Network validation and error analysis After an ANN model has been trained, validation of the network on the test patterns should be undertaken. The usual practice for ANN model validation is to evaluate the net - work performance measure using a selected error metric based on both training and testing data. The evaluation and validation of an ANN prediction model can be done by us - ing common error metrics such as the root mean squared er - ror (RMS error). The definition for RMS error is given in the following equation:

[7]

RMS =

n m

 

∑∑

(O ij

2

Y )/ nm

ij

  ∑∑ ( O ij − 2 Y )/ nm ij

i

=

1

j =

1

where n is the number of patterns in the validation set, m is the number of components in the output vector, O is the out - put of a single neuron, and Y is the target output for the sin - gle neuron.

Network topology Since there is no direct and precise way of determining the most appropriate number of hidden layers and number of neurons to include in each hidden layer for a problem, a trial and error procedure is typically used to approach the best network topology for a particular problem. In this paper, several network topologies were examined. These included networks with one hidden layer containing 1, 2, 4, 6, 8, 10, 12, 14, 16, 18, and 20 neurons in the single hidden layer; and networks with two hidden layers containing 1, 2, 4, 5, 6, 8, and 10 neurons in each hidden layer. The target network would be the smallest sized network that not only produces a minimum error for the training pattern but also gives a gen- eralized solution that performs well for the testing pattern.

Results and discussion

Error trace Figure 2 shows the RMS error trace with the number of it- erations for networks with a single hidden layer (ANN 6-12- 1) and those with two hidden layers (ANN 6-4-4-1). Clearly, the RMS training error decreases with an increase in the number of iterations, whereas the RMS testing error de - creases to a minimum and then begins to increase again. During the first part of the training, performance on the training set improves as the network adjusts its weights through back-propagation. Performance on the test set (dataset that the network is not allowed to learn on) also im - proves. After a while, the weights shift around, looking for a path for further improvement. When this path is found, per - formance on the training set improves but performance on the test set gets worse due to the fact that the network begins to “memorize” the individual input–output pairs rather than settling for weights that generally describe the mapping for all cases. Therefore, it is not always beneficial to increase the maximum number of allowed iterations. In this study, this number was set to 40 000 iterations, which was fairly suitable for most of the studied cases.

Achieving best network

The performance of different networks in terms of RMS

error of both training and testing patterns is shown in Fig. 3. Increasing the number of neurons causes the RMS training error to decrease. In other words, networks with a large number of neurons in the hidden layers achieve good learn - ing in terms of RMS training error. On the other hand, they give worse results in terms of RMS testing error. This indi - cates that too much power of the network (large number of

© 2005 NRC Canada

Seleemah

649

Fig. 3. RMS error for different network topologies: ( a ) networks with a single hidden layer; ( b ) networks with two hidden layers.

hidden layer; ( b ) networks with two hidden layers. Fig. 4. Statistical errors for all

Fig. 4. Statistical errors for all patterns (398 specimens): ( a ) networks with a single hidden layer; ( b ) networks with two hidden layers.

hidden layer; ( b ) networks with two hidden layers. neurons in the hidden layers) allows

neurons in the hidden layers) allows the network to “memo- rize” the training pattern rather than extract the noticeable features that map the input data to the output data. To judge which network performs better, a statistical error of each network was calculated for all patterns, that is, 398 pattern sets. The results are shown in Fig. 4, in which the av- erage error, median, standard deviation (SD), and standard variation (SV) are shown for each of the tested networks. For the networks with a single hidden layer, the networks with six and 10 neurons in the hidden layer are better than all other networks. For the networks with two hidden layers, the one with six neurons in each hidden layer can be consid - ered best. These three networks were selected and allowed to run 200 000 more iterations to reach a total of 240 000 itera - tions. The maximum, minimum, average, SD, and SV of the ratio of experimental to model-predicted shear capacity ( γ mod = V exp / V pred ) for the three networks with 40 000 itera - tions and their counterparts with 240 000 iterations are shown in Fig. 5. Note that the networks with 240 000 itera - tions are denoted ANN#. Clearly the ANN 6-6-1 network with 40 000 iterations was superior to all other networks. It is concluded that too much increase in the number of itera - tions, although consuming a large amount of computer run - ning time, did not improve the performance of the networks at all. A comparison of the experimental and predicted shear ca - pacity by the aforementioned six networks for all 398 data patterns is shown in Fig. 6. The line of equality and the lines of plus or minus 10% error are also plotted to facilitate visu - alization and judgment of the results. The network that gives results closer to the equality line is of course better. ANN 6- 6-1 is obviously the best network and has the greatest num -

6-1 is obviously the best network and has the greatest num - Fig. 5. Statistical data

Fig. 5. Statistical data for the ratio between experimental and predicted shear capacity ( γ mod = V exp / V pred ) for different net- works.

e x p / V p r e d ) for different net- works. ber of

ber of predictions lying between the plus and minus 10% error lines. Figure 7 shows a histogram of the ratio between experimental and predicted shear capacity for ANN 6-6-1. The histogram is slightly skewed to the left-hand side and has its maximum occurring at γ mod = 0.96.

Comparison with different models The ANN 6-6-1 is employed to predict the shear capacity of all 398 beam specimens, together with the eight different codes and methods listed in Table 2. The results are pre - sented in Fig. 8, which shows, in separate plots, comparisons between experimental and predicted shear capacity by each of the methods. It is clear from Fig. 8 that the ACI Commit - tee 318 (1999), BSI (1990), ECS (1992), Collins and

© 2005 NRC Canada

650

Can. J. Civ. Eng. Vol. 32, 2005

Fig. 6. Comparison of experimental and predicted shear capacity for different networks. The broken lines represent 10% error on either side of the line of equality.

represent 10% error on either side of the line of equality. Fig. 7. Histogram of the

Fig. 7. Histogram of the ratio of experimental to predicted shear capacity for ANN 6-6-1.

of experimental to predicted shear capacity for ANN 6-6-1. Kuchma (1999), and Niwa et al. (1986)

Kuchma (1999), and Niwa et al. (1986) methods give wide variations on either side of the equality line. This means that these methods underestimate the shear capacity for some specimens and overestimate the shear capacity for other specimens. The NZS (1995) and Zsutty (1971) methods overestimate the shear capacity. The Reineck (1991) method is the only conservative method, in which most of the pre - dictions are underestimates. In contrast, the predictions of the ANN are extremely impressive, with most results lying on or very close to the equality line. This accuracy suggests that the most critical variables controlling the shear capacity of concrete beams without transverse reinforcement are the six that have been used as input data to the ANN model.

The ratio of the experimental to predicted shear capacity by specific model ( γ mod = V exp / V pred ) can be interpreted as the additional factor of safety implied by this method, since no strength reduction factor was applied during the calcula- tion of shear capacity by any of the methods. For economic considerations, this additional safety factor should not be very large because there is a concrete strength reduction fac- tor, γ m , that is applied during any design process. The ratio γ mod also should not be much lower than unity, since this means that the method overestimates the shear capacity of the beam and may lead to unsafe design. A comparison of statistical calculations of the ratio of the experimental to predicted shear capacity for all models ( γ mod ) and for all data patterns was conducted and the re - sults are shown in Fig. 9. As observed earlier, the method of Reineck (1991) gives the most conservative results, with a maximum ratio of 4.40, a minimum ratio of 0.87, and an av- erage ratio of 1.43. The ANN method gives the best results, with a maximum ratio of 1.26, a minimum ratio of 0.74, and an average ratio of 0.94. All other methods have a maximum ratio lying between 2.00 and 3.30, a minimum ratio between 0.32 and 0.58, and an average ratio between 0.87 and 1.47. Both the SD and SV are minimum for the ANN method. To describe the influence of dominant parameters on the predictions of each model, the ratio of actual to predicted shear capacity ( γ mod ) is plotted versus the primary parame - ters in Figs. 10–13. A common feature of all the plots is the very large scatter in the predictions by different models, in which γ mod ranged from 0.33 to 4.41. Figure 10 shows the plot of γ mod versus the ratio of shear span to beam depth ( a / d ). Most tests were conducted for low a / d ratios, for example, 81% of all tests (323 of 398 tests)

© 2005 NRC Canada

Seleemah

651

Fig. 8. Comparison of experimental and predicted shear capacity for 398 specimens. The broken lines represent 10% error on either side of the line of equality.

represent 10% error on either side of the line of equality. Fig. 9. Statistical data for

Fig. 9. Statistical data for the ratio between experimental and predicted shear capacity for different shear proposals.

and predicted shear capacity for different shear proposals. were carried out for beams with an a

were carried out for beams with an a / d ratio of less than or equal to 4.0. A large scatter is noticed in the predictions made by all methods for a / d ratios between 2.4 and 4.0. For the ACI Committee 318 (1999), BSI (1990), ECS (1992), NZS (1995), Collins and Kuchma (1999), and Reineck (1991) models, the additional safety factor, γ mod , increases with a decrease in the a / d ratio, indicating a beneficial influ - ence due to direct load transfer that is not captured by these models. The results obtained from the neural network indi - cate consistent accuracy in all ranges of the a / d ratio, indi - cating that the shear phenomena were well captured by this network. Variation of the additional safety factor, γ mod , with the concrete uniaxial compressive strength ( f 1c ) for the models considered in this study is shown in Fig. 11. Most of the tests in the database (73%) were carried out for normal- strength concrete (NSC) having f 1c less than 45.0 MPa (289 of 398 tests). The scatter of the predictions by the ACI Com - mittee 318 (1999), BSI (1990), and Collins and Kuchma

© 2005 NRC Canada

652

Can. J. Civ. Eng. Vol. 32, 2005

Fig. 10. Ratio of experimental to predicted shear capacity versus the ratio a / d for different shear proposals.

of experimental to predicted shear capacity versus the ratio a / d for different shear proposals.

© 2005 NRC Canada

Seleemah

653

Fig. 11. Ratio of experimental to predicted shear capacity versus uniaxial concrete compressive strength ( f 1c ) for different shear proposals.

capacity versus uniaxial concrete compressive strength ( f 1 c ) for different shear proposals. ©

© 2005 NRC Canada

654

Can. J. Civ. Eng. Vol. 32, 2005

Fig. 12. Ratio of experimental to predicted shear capacity versus percentage of main steel (ρ 1 = A s / b w d ) for different shear proposals.

versus percentage of main steel ( ρ 1 = A s / b w d )

© 2005 NRC Canada

Seleemah

655

Fig. 13. Ratio of experimental to predicted shear capacity versus beam depth ( d ) for different shear proposals.

of experimental to predicted shear capacity versus beam depth ( d ) for different shear proposals.

© 2005 NRC Canada

656

(1999) methods for the specimens with high-strength con - crete (HSC) having f 1c greater than 45.0 MPa is very appar - ent. The results of the NZS (1995), Zsutty (1971), and Niwa et al. (1986) methods are mostly unaffected by the variation of f 1c , indicating better recognition of the effect of f 1c on the shear capacity of beams without shear reinforcement. Among all of the models shown, the results obtained from the neural network are the most consistent, with values close to unity for a wide variation in concrete uniaxial compres - sive strength. Figure 12 shows a plot of γ mod versus the longitudinal re - inforcement percentage ρ l (= A s / b w d , where A s is the area of longitudinal reinforcement). Fifty-five percent of all tests (219 of 398 tests) were conducted for a relatively high rein - forcement ratio of ρ 1 > 2%. Only 14.6% of the tests (58 of

398 tests) were conducted for ρ 1 < 1%. The results obtained

using the ACI Committee 318 (1999), BSI (1990), ECS (1992), Zsutty (1971), and Collins and Kuchma (1999) methods show a general increase in the model’s additional safety factor, γ mod , with the increase in ρ 1 indicating a pro - nounced beneficial effect of the reinforcement ratio that is not properly captured by any of these methods. Some of these methods, namely those of ECS and Zsutty, included the reinforcement ratio in their equations, but it seems that the beneficial effect of reinforcement exceeds what was in- cluded in the relationships. The models of Niwa et al. (1986) and Reineck (1991) are nearly unaffected by the change in the reinforcement ratio, indicating a reasonable capturing of the effect of reinforcement ratio. The performance of the neural network model is superior to that of all other models. It not only is unaffected by the change in reinforcement ra- tio, but also gives values of γ mod near or equal to unity.

Lastly, the variation of the additional safety factor, γ mod ,

with beam depth, d , is shown in Fig. 13; 87% of all the spec- imens (346 of 398 specimens) had a beam depth of less than

500 mm. Once again, most methods yield large scatter in the

results, especially for specimens with beam depths in the range 140–300 mm. The additional safety factor, γ mod , pre -

dicted using the ACI Committee 318 (1999), BSI (1990), NZS (1995), Zsutty (1971), and Collins and Kuchma (1999) models decreases with an increase in depth, indicating im - proper capturing of the effect of depth on the overall shear capacity of beams. In contradiction to all other methods, the predictions of the ANN are very consistent and very close to unity.

Summary and conclusions

Different relationships have been proposed by researchers all over the world for predicting the shear capacity of con - crete members without transverse reinforcement. Unfortu - nately, these proposed relationships are usually empirical and designed to fit a limited set of shear test results that is most familiar to the researcher. Accordingly, the relation - ships differ quite considerably in their selected parameters, since there is no generally accepted model for the load trans - fer and ultimate shear capacity of members without trans - verse reinforcement. A unique and critically reviewed database of experimental work conducted on beams without transverse reinforcement was collected from 64 different references. This database

Can. J. Civ. Eng. Vol. 32, 2005

contained a total of 398 tests and was intended, by the Joint ASCE–ACI Committee 445 on Shear and Torsion (1998) and the ACI Subcommittee 445-F on Beam Shear, to serve as a basis for any code changes regarding the shear phenom - ena. The database covered a very wide range of beam parameters, including their dimensions, concrete strengths, reinforcement ratios, and shear span to beam depth ratios. This database was utilized in this study to evaluate the fol - lowing eight different existing codes and formulas for pre - dicting the shear capacity of members without transverse reinforcement: ACI Committee 318 (1999), BSI (1990), ECS (1992), NZS (1995), Zsutty (1971), Collins and Kuchma (1999), Niwa et al. (1986), and Reineck (1991). Moreover, the database was used to evaluate and demon - strate the capability of the feed-forward back-propagation artificial neural networks (ANNs) for predicting such shear capacity. Based on careful study of recent approaches for the shear phenomena in concrete members, it was decided to design the ANN to have six nodes in the input layer containing data regarding beam depth ( d ), beam breadth ( b w ), shear span to beam depth ratio ( a / d ), uniaxial concrete compressive strength ( f 1c ), percentage of main reinforcement (ρ 1 ), and steel yield stress ( f y ). The output layer consisted of one node representing the ultimate shear capacity of the beam. Several network topologies were examined. These in- cluded networks with one hidden layer containing 1, 2, 4, 6, 8, 10, 12, 14, 16, 18, and 20 neurons in the single hidden layer; and networks with two hidden layers containing 1, 2, 4, 5, 6, 8, and 10 neurons in each hidden layer. All the net- works were trained on 199 shear tests, and the performance of the networks on both 199 training datasets and 199 testing datasets was compared; the network with the best prediction was selected for comparison with the eight models. In comparing the predictions of the ANN with those from the eight models, it was found that the results obtained from the ANN were the most accurate, giving values of maximum shear capacity very close to the experimental values. More - over, the results obtained using the ANN were very consis - tent and covered a very wide range of variation of any of the input parameters. This accuracy suggests that the most criti - cal variables controlling the shear capacity of concrete beams without transverse reinforcement are the six variables used as input data to the ANN model.

Acknowledgment

The author would like to thank Professor Amr W. Sadek of Cairo University and the Kuwait Institute for Scientific Research (KISR) for his valuable suggestions regarding de - termination of the most appropriate number of training and testing patterns.

References

ACI Committee 318. 1999. Building code requirements for struc - tural concrete (ACI318-99) and commentary (318R-99). Ameri - can Concrete Institute, Farmington Hills, Mich. BSI. 1990. Steel, concrete and composite bridges — Part 4: Code of practice for design of concrete bridges. British Standard BS5400-4, British Standards Institution, London, U.K.

© 2005 NRC Canada

Seleemah

Carpenter, W.C., and Barthelemy, J. 1994. Common misconcep - tions about neural networks as approximators. ASCE Journal of Computing in Civil Engineering, 8 (3): 345–358. Carpenter, W.C., and Hoffman, M.E. 1995. Training backpropaga - tion neural networks. AI Expert, 10 (3): 30–33. CEB. 1997. Concrete tension and size effects: utilisation of con - crete tension in structural concrete design and relevance of size effect. Bulletin 237. Comité Euro-International du Béton (CEB), Lausanne, Switzerland. 258 pp. Collins, M.P., and Kuchma, D.A. 1999. How safe are our large, lightly reinforced concrete beams, slabs, and footings. ACI Structural Journal, 96 (4): 482–490. DIN. 2001. Plain concrete, reinforced and prestressed concrete structures — Part 1: Design and construction (DIN 1045-1). Deutsches Institut für Normung e.V. (DIN), Berlin, Germany. ECS. 1992. Eurocode 2: design of concrete structures — Part 1:

General rules and rules for buildings (DD ENV 1992-1-1, BSI). European Committee for Standardization (ECS), Brussels, Bel - gium. 253 pp. Flood, I., and Kartam, N. 1994. Neural networks in civil engineer - ing. I: principles and understanding. ASCE Journal of Com - puting in Civil Engineering, 8 (2): 131–148. Joint ASCE–ACI Committee 445 on Shear and Torsion. 1998. Re - cent approaches to shear design of structural concrete. ASCE Journal of Structural Engineering, 124 (12): 1375–1417. Kuchma, D.A. 2000. Shear Data Bank. Available from http://www. ce.uiuc. edu/kuchma/sheardatabank [cited December 2003]. Mansour, M.Y., Dicleli, M., Lee, J.Y., and Zhang, J. 2004. Pre- dicting the shear strength of reinforced concrete beams using ar- tificial neural networks. Engineering Structures, 26 (6): 781–799.

657

Niwa, J., Yamada, K., Yokozawa, K., and Okamura, M. 1986. Re- evaluation of the equation for shear strength of RC beams with - out web reinforcement. Proceedings of the Japan Society of Civil Engineering, 5 (372): 1986–1988. [In Japanese.] NZS. 1995. New Zealand standard code of practice for the design of concrete structures (NZS 3101). Standard Association of New Zealand (NZS), Wellington, New Zealand. Oreta, A.C. 2004. Simulating size effect on shear strength of RC beams without stirrups using neural networks. Engineering Structures, 26 (5): 681–691. Oreta, A.C., and Kawashima, K. 2003. Neural network modeling of confined compressive strength and strain of circular concrete col- umns. ASCE Journal of Structural Engineering, 129(4): 554–561. Reineck, K.H. 1991. Ultimate shear force of structural concrete members without transverse reinforcement derived from a me - chanical model. ACI Structural Journal, 88 (5): 592–602. Reineck, K.H. 1999. Shear design of members without transverse reinforcement in DIN 1045-1. Final research report for DIBt, Project IV 1-5-876/98. Chapter 3 and Appendixes. University of Leipzig, Aachen, and University of Stuttgart, TU Munchen, Ger - many. [In German.] Reineck, K.H., Kuchma, D., Kim, K.S., and Marx, S. 2003. Shear database for reinforced concrete members without shear rein - forcement. ACI Structural Journal, 100 (2): 240–249. Sanad, A., and Saka, M.P. 2001. Prediction of ultimate shear strength of reinforced concrete deep beams using neural networks. ASCE Journal of Structural Engineering, 127(7): 818–827. Zsutty, T.C. 1971. Shear strength prediction for separate categories of simple beams tests. ACI Journal, 68 (2): 138–143.

© 2005 NRC Canada