Composites: Part B 66 (2014) 162–173

Contents lists available at ScienceDirect

Composites: Part B
journal homepage: www.elsevier.com/locate/compositesb

Application of artificial intelligence techniques to predict

the performance of RC beams shear strengthened with NSM FRP rods.
Formulation of design equations
Ricardo Perera a,⇑, David Tarazona a, Antonio Ruiz b, Andrés Martín b
a
Department of Structural Mechanics, Technical University of Madrid, José Gutiérrez Abascal 2, 28006 Madrid, Spain
b
Department of Applied Mathematics, Technical University of Madrid, Alenza 4, 28003 Madrid, Spain

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

Article history: The use of near-surface mounted FRP reinforcement in reinforced concrete structures has seen a
Received 19 February 2014 considerable increase in recent years as a strengthening method. However, few simple models have been
Received in revised form 16 April 2014 published for predicting the shear capacity given by this technique. In this work, two different
Accepted 6 May 2014
approaches are proposed. Firstly, the use of neural networks as a means of predicting shear strength
Available online 23 May 2014
without the need to use complex models is developed. Secondly, with the purpose of generating a design
formula of simple application to evaluate the shear strength contribution provided by a near-surface
Keywords:
mounted system, a multi-objective optimization problem is solved. It results from considering simulta-
A. Polymer-matrix composites
B. Debonding
neously the experimental results of beams with and without NSM-FRP reinforcement. The performance
C. Computational modeling of both methodologies is compared with some experimental results. Sensitivity and parametric analyses
are performed too for further evaluation of the proposed models.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction compression parts of the cross section (shear), as they can be easily
pre-stressed; bars are protected by the concrete cover and so are
The use of externally bonded (EB) fibre-reinforced polymer less exposed to accidental and mechanical damage, fire, and
(FRP) composites has gained acceptance during the last two vandalism (flaws of FRP EBR) while the aesthetic of the strength-
decades in the construction engineering community, particularly ened structure is virtually unchanged. However, in spite of the
in the rehabilitation of reinforced concrete (RC) structures, as higher adherence surface available when compared to NSM-FRP,
demonstrated by the numerous research studies which have been the premature debonding problem, considered as a loss of bond
conducted and the different techniques applied in rehabilitation within the adhesive layer or at the FRP strip/adhesive or
projects [1,2]. Currently, to increase the shear resistance of RC adhesive/concrete interfaces, persists, which limits the efficiency
beams, FRP sheets are externally bonded (EB-FRP) and applied on of this kind of FRP strengthening system.
the external side surface of the beams to be strengthened with Despite the substantial amount of experimental work [7,11,13]
different configurations [3–6]. Of more recent application, the on this subject, little numerical work [7,14] has been developed
near-surface mounted FRP bar (NSM-FRP) method is another and, therefore, accurately calculating the shear capacity of FRP
technique successfully used to increase the shear resistance of RC shear strengthened RC beams remains a complex challenge that
beams. In the NSM method, FRP rods are embedded into grooves has not yet been fully resolved due to the numerous variables
intentionally prepared in the concrete cover of the side faces of involved in the procedure. Therefore, besides additional experi-
RC beams. Research has shown that a significant increase in the mental research, more numerical studies need to be carried out
shear resistance of RC beams is attainable using the NSM FRP to further clarify these aspects.
method [7–13]. NSM technology has some important advantages, Specific models [7,14] have been proposed to compute the
among which should be pointed out the greater ease with which shear capacity of a member strengthened in shear with NSM FRP
they can be anchored into adjacent members (bending) or reinforcement, and in particular the load at diagonal tension failure
of concrete. Considering debonding of the FRP bars as the most
important failure mode for this kind of strengthening (although
⇑ Corresponding author. Tel.: +34 913363278; fax: +34 913363004. tensile rupture of the NSM reinforcement is another possible
E-mail address: [email protected] (R. Perera).

http://dx.doi.org/10.1016/j.compositesb.2014.05.001
1359-8368/Ó 2014 Elsevier Ltd. All rights reserved.
 R. Perera et al. / Composites: Part B 66 (2014) 162–173 163

failure mode, it is rarely observed), numerical models incorporat- absence of other types of neural networks. MLPs consist of an input
ing local bond–slip relationships obtained from bond tests are used layer of source nodes, one or more hidden layers of computation
to evaluate the contribution of the NSM FRP rods. Another failure nodes or ‘neurons’ and an output layer. The number of nodes in
mode recently observed [13,15] consists of the progressive the input and the output layers depends on the number of input
detachment and outward expulsion of the concrete cover from and output variables respectively. The number of hidden layers
the underlying beam core. Such failure mode is related to the and the number of nodes in each hidden layer affects the general-
semiconical tensile fracture of the concrete surrounding each ization capability of the network. The interconnection topology
NSM strip and, in this respect, a new approach has been proposed defines the number of the types of arrangements. The remarkable
for modelling the contribution of NSM FRP strips for the shear computational characteristics of neural networks are their ability
strengthening of RC beams [16,17]. to learn functional relationships from training examples and to
The objective of this paper is to develop alternative methodolo- discover patterns and regularities in variables by simply presenting
gies to evaluate the capacity of FRP shear strengthened RC beams them with data.
by dealing with the problem from a different point of view by using The training of an MLP network involves finding values of the
artificial intelligence techniques. With this purpose two different connection weights which minimize an error function between
approaches will be presented: one concerned with the use of the actual network output and the corresponding target values in
artificial neural networks and the other based on the implementa- the training set. The most commonly used training algorithms
tion of an optimization approach developed jointly with the use of are based on back-propagation. The feed forward MLP neural net-
artificial neural networks (ANNs) and solved with genetic algo- work, used in this work, consisted of three layers: input, hidden
rithms (GAs). With these approaches some of the difficulties and output. The input layer had nodes representing the normalized
concerned regarding the numerical modelling can be overcome. features characteristics of the behaviour of strengthened beams.
As an alternative tool to conventional numerical techniques, neural Taking into account the high number of parameters which influ-
networks do not provide closed form solutions for modelling ence the shear failure mode, a suitable selection of the input vari-
problems but do, however, offer a complex and accurate solution ables or signature is essential to reach an optimum configuration of
based on a representative set of historical examples of the relation- the ANN and ensure the success of this approach, since the number
ship. Furthermore, they can adapt solutions over time to include of input parameters should also be chosen according to the number
new data. of training data. On the one hand, the number of parameters must
On the other hand, as a second proposal, an optimization be large enough to represent the system properly. However, on the
approach has also been developed to implement simple yet accu- other hand, a large number of input neurons in the ANN may
rate shear design equations for this kind of strengthening. This reduce the efficiency and accuracy of the training process.
approach is developed in a multi-objective framework by consider- In the ANNs, the activation functions of tan-sigmoid and linear
ing experimental results of RC beams with and without NSM-FRP. were used in the hidden and the output layers, respectively. The
Furthermore, the results obtained with the previous scheme based range of hidden layer nodes and the activation functions were
on ANNs are also used as a filter to choose the parameters to selected on the basis of training trials. The ANN was created,
include in the design equations. Genetic algorithms are used to trained and implemented using MATLAB with back-propagation
solve the optimization problem since they are especially suitable and the training algorithm of Levenberg–Marquardt. The ANN
for solving multi-objective problems when compared to standard was trained iteratively to minimize the performance function of
optimization methods. mean square error (MSE) between the network outputs and the
The key features of the two proposed procedures are outlined corresponding target values. In general, the more input–target
and their performance in predicting the capacity of NSM-FRP shear pairs employed in the training process, the higher the generaliza-
strengthened RC beams is evaluated by comparison with results tion ability of the training ANN.
from experimental tests and with predictions obtained using a For the configuration and learning of the NN, 69 experimental
simplified numerical model [14]. A sensitivity study of the predic- tests collected from DABASUM, the database for FRP-based shear
tions of both models for the input parameters is also carried out. strengthening of reinforced concrete beams developed by the
structural composites research group of Minho University (Portu-
gal) [23], were considered in detail. The RC beams included in
2. Prediction of the capacity of shear strengthened RC beams DABASUM were tested by Dias and Barros [15,24–26], Barros and
with NSM-FRP using ANNs Dias [11], De Lorenzis and Nanni [7], Rizzo and De Lorenzis [13]
and Kotynia [27]. To these data, 32 experimental tests carried
Artificial neural networks were not specifically developed for out at the Eduardo Torroja Institute for Construction Science, Spain
evaluating the capacity of civil engineering structures, but their [28] were added, resulting in the database reported in Appendix A.
pattern matching capability makes them very suitable to be The database includes 101 beams shear strengthened with NSM-
employed as a tool for this purpose. Due to their unique features FRP. The selected tests present a high diversity regarding the beam
they can be used to solve complex problems that cannot be han- geometric ratios, reinforcement ratios, material properties and,
dled by analytical approaches, even problems whose underlying therefore, failure loads. The 101 test beams were grouped
physical and mathematical models are not well-known, such as randomly into two sets: a training set containing 80% beams, and
those derived from the FRP strengthening of RC beams [18–22]. a validating set with 20% beams. To avoid the slow rate of learning
From this point of view, they might be suitable for determining near the end points the input and output data were scaled between
the capacity of RC beams strengthened with NSM-FRP shear the interval [1, 1].
reinforcement. The neural network was configured with only one output node
Artificial neural networks consist of a number of processing for the present data set corresponding to the beam capacity.
units or neurons interacting with each other via weighted However, as commented above, from the wide variety of parame-
connections to constitute a network, and are characterized by an ters affecting the behaviour of this strengthening technique, the
architecture or topology and a learning mechanism. Among the dif- suitable choice of the input parameters is critical for the success
ferent types of ANNs, multilayer perceptron (MLP) neural networks of the procedure. For this reason, the configuration and training
are quite popular and are used for the present work. In this paper, procedure of the NN was carried out in two steps. Initially, in a first
the terms ANN and MLP have been used interchangeably in the step, after an extensive bibliographic study, fourteen parameters
164 R. Perera et al. / Composites: Part B 66 (2014) 162–173

were chosen as input parameters: (a) Geometry of the beam trained network configurations. If for one specific configuration
(rectangular or T section), (b) ratio of the NSM-FRP transversal one or more parameters present an anomalous value when
reinforcement qf, (c) concrete average compressive strength fcm, compared to the rest of the trained networks, a lower mark is
(d) shear span/depth ratio of the member a/d, (e) distance between assigned to that configuration.
FRP bars sf, (f) elastic modulus of the FRP reinforcement Ef, (g) By a combination of the marks assigned to each particular
design yielding stress of the shear steel reinforcement fsy, (h) angle network for the three criteria, the most suitable network was
between the FRP bars and the longitudinal axis of the member af, chosen. According to this, finally, the values of the network param-
(i) ratio of the steel shear reinforcement qsw, (j) ratio of the steel eters considered by this study were as follows: Number of input
longitudinal reinforcement qsl, (k) height of the beam section h, parameters: 9 (parameters (a)–(i)); number of output parameters:
(l) length of the beam L, (m) FRP ultimate strain efu and (n) width 1; number of hidden layers: 1; number of hidden neurons: 11.
of the beam cross section bw.
A first filter on the input parameters was performed training the 3. Proposal for a new shear design equation
network several times by randomly varying the training samples
and applying the Garson index to every case. Through operations As demonstrated in the results shown in further sections, ANNs
between the weight matrices generated in two consecutive layers predict capacity values closer to experimental results than existing
of the ANN, the Garson index identifies the relative importance of conventional models. However, from an engineering design point
all input variables regarding the output variable [29]. This method- of view, a more practical approach based on the use of simple
ology is not useful if the index changes a lot from one trained design equations would be recommendable to estimate the capac-
network to another. However, when the index associated with ity of NSM-FRP shear strengthened beams. To date, except for one
one parameter remains low for a high number of networks it recently published paper [30], there are no works published in this
means that this parameter might be considered irrelevant and, sense for this technology. For this reason, one of the purposes of
therefore, removable as input parameter. the present study is the proposal of design equations to calculate
After this first step, three parameters, beam length, FRP the shear capacity of RC beams NSM-FRP strengthened in shear.
ultimate strain and beam width, were removed, with eleven input Based on the observed behaviour of the analyses carried out with
parameters remaining, (a)–(k), as input parameters. the neural networks, an optimization approach is proposed to
For the available data sample, eleven parameters are close to develop simple yet accurate shear design equations for NSM-FRP
the maximum recommendable. Therefore, in a second step the shear strengthened RC beams. The optimization approach is devel-
NN was configured and trained again by varying the number of oped in a multiobjective framework by considering RC beams with
input parameters from seven to eleven. To do this, the first seven and without NSM-FRP reinforcement and is solved with genetic
parameters, (a)–(g), considered to be more relevant, were kept as algorithms (GAs), since they are especially suitable for solving
input parameters in all possible configurations and the training multiobjective problems when compared to standard optimization
was performed considering all the possible combinations of the methods [31].
other four parameters, ranging from zero to four. Furthermore,
for each combination of input parameters, the number of hidden 3.1. Contribution of NSM-FRP rods to shear capacity
nodes was varied between 4 and 13 and all the possible configura-
tions were trained. From this study, it is difficult to choose the best The design shear strength of beams strengthened with NSM-
configuration from all those tested. To do so, a posterior multi- FRP can be computed by adding a third term to the basic equation
criterion analysis was carried out considering three different given by ACI code [32], to account for the contribution of the FRP
objectives to be evaluated at each network. Firstly, the mean reinforcement
square error between the trained network output and the V Rd ¼ V c þ V s þ V f ð1Þ
corresponding target values in the training and validation sets is
estimated for each network. The lower the MSE value, the higher where Vc is the contribution of concrete, Vs is the contribution of the
the mark given to the evaluated network. steel stirrups and Vf is the contribution of the FRP.
The second criterion involves a parametric study whose pur- The shear strength of the concrete Vc for a cracked section
pose is to evaluate how each trained network is able to represent depends on the dowel action of the longitudinal reinforcement,
the effect of each particular input parameter on the beam capacity. the diagonal tensile strength of the uncracked part of the concrete
These studies are carried out by simply varying each of the input and the aggregate interlocking effect; Vc may be calculated accord-
parameters and keeping the others constant. Through parametric ing to the provisions in existing RC design codes using the expres-
studies, the performance of the different ANNs in simulating the sions for shear strength provided by concrete without web
physical behaviour of the shear strengthening of an RC beam can reinforcement which are based on empirical data. One of the
be verified. Neural network predictions consistent with the exper- easiest ways to determine the concrete contribution is based on
imental tests would yield qualitative evidence of the ability of the the dimensions of the member and the compressive strength of
ANN to simulate the physical phenomenon. However, due to the the concrete, fcm:
limited number of experimental tests available, it is not always 1=n
V c ¼ bw dðfcm Þ ð2Þ
possible to evaluate if the NN is able to predict the experimental
trend. For this reason, besides a comparison with experimental where bw is the web width, d is the distance from the extreme
results, the possible anomalies relative to the behaviour of the compression fibre of the cross-section to the centroid of the longi-
network for some particular parameter and comparison among tudinal reinforcement and n is usually equal to 2 or 3 depending
predictions of the different networks are considered in order to on the code chosen.
evaluate the quality of each network. A mark is given for each The contribution from steel stirrups can be expressed by using
network and each parameter as a result of this parametric the truss theory as follows
evaluation.
V s ¼ ðAs =sÞfy z cot h ð3Þ
Finally, the last criterion involves the use of the Garson index,
used previously to filter the number of input parameters from 14 where As is the area of the transverse stirrups at spacing s, fy is the
to 11. In this case, the Garson indices associated with the input yield strength of the shear reinforcement, z is the flexural lever arm
parameters of the different networks are compared among all the and h is the local inclination of the compression strut.
 R. Perera et al. / Composites: Part B 66 (2014) 162–173 165

The focus on this section is on the contribution from the tests or assumptions. From the previous considerations, the follow-
composite. The problem is thus to find an expression for Vf. This ing equation is adopted:
is, however, not so easy considering the high number of parame-  C7 !C 8
ters involved. In the following, an approach to compute the contri- Af fcm
V f ¼ C 6 Ef bw d sin af ðcot af þ cot C 5 45Þ ð8Þ
bution of NSM-FRP rods to the shear capacity of an RC beam is b w sf qf Ef
presented. In the proposal, the predictions obtained with the
neural network model of Section 2 are used as starting point. Fur- where af is the angle for FRP reinforcement fibre direction in
thermore, it is assumed that the law of superposition from Eq. (1) is relation to the beam’s longitudinal axis, Ef is the FRP Young’s mod-
valid for NSM-FRP strengthened beams. ulus, Af is the FRP cross section, sf is the spacing between adjacent
strips and qf = (Af/bwsf) is the FRP reinforcement ratio. The term
(fcm/qfEf) obeys the idea of introducing in Eq. (8) a design effective
3.2. Multi-objective NSM-FRP shear design equations
strain representative of the debonding of FRP bars; for this reason,
the same relevant parameters used in [35] for failure by debonding
The proposed approach to obtain shear design equations is
in externally bonded FRP reinforcement have been considered in
developed as an optimization problem to define in a combined
the expression. All the parameters contributing in Eq. (8), directly
way equations for the shear design of RC beams with or without
or indirectly, were included as input parameters in the NN
NSM-FRP shear reinforcement and whose objective will be to
approach. C6 to C8 are unknown coefficients to be determined
minimize the difference between the measured shear strength of
during the optimization procedure.
reinforced concrete beams with and without NSM-FRP reinforce-
The solution of the design problem will require searching for an
ment and their calculated values. To perform this, the problem will
optimum set of coefficients, C1 to C8, by simultaneously minimiz-
be developed in a multi-objective framework by simultaneously
ing the objective functions in Eqs. (4) and (5).
optimizing the equation to predict the shear capacity of reinforced
concrete beams without NSM-FRP shear reinforcement and the
equation to predict the shear capacity of NSM-FRP shear strength- 3.3. Multi-objective evolutionary algorithm
ened reinforced concrete beams. In both cases, the objective func-
tions will be constructed as a measure of how well the model’s The presence of two objectives in the problem, in principle,
predicted output agrees with the experimentally measured data gives rise to a set of optimal solutions, known as Pareto-optimal
and have an expression as follows: solutions, instead of a single optimal solution. In the absence of
any further information, one of these Pareto-optimal solutions
F 1 ¼ V exp RC  V predRC ¼ V exp RC  ðV c þ V s Þpred ð4Þ cannot be said to be better than the other. This demands a user
to find as many Pareto-optimal solutions as possible.
F 2 ¼ V exp RCþFRP  V predRCþFRP ¼ V exp RCþFRP  ðV c þ V s þ V f Þpred ð5Þ Pareto-based approaches constitute the main motivation for
using evolutionary algorithms, such as genetic algorithms (GAs),
where Vexp RC and VpredRC are the measured and the model-predicted to solve multi-objective optimization problems since, unlike classi-
shear capacities, respectively, in the case of beams without NSM- cal optimization methods, in a single run of the algorithm several
FRP shear reinforcement, and Vexp RC+FRP and VpredRC+FRP represent points of the Pareto-optimal set are found. Multiobjective GAs will
the same when NSM-FRP shear reinforcement is added. provide equations capable of estimating the shear strength values
To define the objective functions in Eqs. (4) and (5) and, there- of NSM-FRP shear strengthened RC beams as closely as possible to
fore, implement the optimization problem, expressions for Vc,pred, the experimentally measured ones.
Vs,pred and Vf,pred should be developed. The equations should be Genetic algorithms have been well accepted as fairly effective in
the simplest and most accurate possible and should consider the dealing with multi-objective optimization problems due to their
main basic design parameters on the shear strength. To this end, features, such as multi-point, direct, and parallel searching. Due
Eqs. (2) and (3) might be taken as the starting point for Vc,pred to their potential for solving multi-objective optimization
and Vs,pred resulting in the following expressions: problems, a wide variety of algorithms has been proposed in the
literature [36].
C
C 1 ðfcm Þ 3 bw d The Non-dominated Sorting Genetic Algorithm II (NSGA-II) [37]
Vc ¼ C2
ð6Þ
ða=dÞ is one of the most prominent multi-objective evolutionary
algorithms and has been used in this paper. This algorithm is a
V s ¼ C 4 ðAs =sÞfy z cotðC 5 45Þ ð7Þ modified version of the NSGA [38] algorithm, which can be consid-
ered as belonging to the second generation of multi-objective
where C1, C2, C3, C4 and C5 are unknown coefficients to be deter- evolutionary algorithms. This revised version is more efficient
mined. Although not considered usually in the design equations, (computationally speaking) and incorporates a better sorting algo-
the shear span/depth ratio (a/d) has been included in Eq. (6) since, rithm that keeps diversity without specifying any additional
as some studies have demonstrated [33], it has a high influence on parameters or elitism, and no sharing parameter needs to be
the shear capacity. The form of Eq. (7) is well accepted and is similar chosen a priori.
in the different design equations; however, the main difference NSGA-II is based on several layers of classifications of the
among them is in the choice of the crack inclination, h. For this rea- individuals. Once the population is initialized and before
son, a coefficient, C5, has been included to consider this aspect. selection is performed, the population is ranked on the basis of
The lack of design recommendations dealing with the NSM-FRP non-domination (Fig. 1). The first front being completely non-
technique causes difficulties to develop an expression for the FRP dominated individuals in the current population and the second
contribution. The equation used in [34] for external shear strength- front being dominated by the individuals in the first front only
ening was adopted as starting point and some modifications based and so on. Each individual in each front is assigned rank (fitness)
on the predictions obtained with the NN model proposed here and values or is based on the front to which it belongs. Individuals in
the studies carried out in [14,16,17] were incorporated into it. Fur- the first front are given a fitness value of 1 and individuals in
thermore, for an easy application from a structural engineering the second are assigned a fitness value of 2 and so on.
point of view, it would be suitable to avoid the introduction of In addition to a fitness value, a new parameter called crowding
parameters in the equation whose definition requires previous distance is calculated for each individual (Fig. 1). The crowding
166 R. Perera et al. / Composites: Part B 66 (2014) 162–173

Fig. 1. NSGA-II multi-objective algorithm.

distance is a measure of how close an individual is to its found, the parameter values are obtained through the inverse of
neighbours. A large average crowding distance will result in better the decoded values. The GA search space has been limited for each
diversity in the population. Parents are selected from the popula- parameter separately by observing those equations provided by
tion by using binary tournament selection based on the rank and commonly used shear design standards and recommendations
crowding distance. An individual is selected if the rank is less than and after carrying out some previous numerical tests. Specifically,
the other or if the crowding distance is greater than the other. The for the parameters C1 to C5 in Eqs. (6) and (7) the following range of
selected population generates offspring from crossover and values has been considered, it being based on previous studies
mutation operators. [33]: C1 = [0, 1]; C2 = [0, 1]; C3 = [0.1, 1]; C4 = [0, 2]; C5 = [0.5, 1.3].
The population with the current population and current For the choice of the parameters C6 to C8 involved in the expression
offspring is sorted again based on non-domination and only the of Vf, a previous iterative study was carried out testing 168
best N individuals are selected, where N is the population size. different combinations of interval ranges for the three parameters
The selection is based on rank and on the crowding distance in considered jointly with the purpose of determining the optimal
the last front. Unlike NSGA, an external memory is not required range for each interval. As a criterion for the choice, the ratios Vexp
and elitism is introduced by combining the best parents with the RC/VpredRC and Vexp RC+FRP/VpredRC+FRP were used. The optimal choice
best offspring obtained. corresponds to those intervals for which both ratios are the closest
to one but without overtaking the unit value in order to be
3.4. Solution of the problem conservative in the predictions. Finally, the following ranges were
adopted: C6 = [0, 3]; C7 = [1.5, 2]; C8 = [1, 1.5].
To solve the optimization problem, two different databases Taking into account the stochastic nature of GAs, 40 indepen-
were considered. The first collects the experimental results of 40 dent runs were performed per GA and test problem in order to
RC beams without FRP reinforcement included in [39], and found decrease the influence of random effects, taking as parameter
mainly from test results compiled by Cladera [40] and Bhatt and values the average values of all the runs. With these conditions
Kader [41]. The tests include the results of simply supported RC each problem was solved 10 times to adopt as final parameter
beams subjected mostly to a concentrated load at the midspan at values the best ones obtained from the 10 times, which are the
various shear span-to-depth ratios, strengths of concrete and steel, following: C1 = 0.784; C2 = 0.469; C3 = 0.735; C4 = 1.816;
steel ratios and geometrical sizes with the purpose of evaluating C5 = 0.983; C6 = 1.735; C7 = 1.549; C8 = 1.119
the method when applied to a different range of values. All the
specimens failed in shear. Some results included at DABASUM
4. Results
[23] belonging to specimens without NSM-FRP reinforcement used
for control in test campaigns of RC beams with NSM-FRP rods have
The predictions of the two proposed models in Sections 2 and 3,
also been incorporated into the database, finally resulting in 55
respectively, are compared with the test data, Vexp, given in Appen-
samples.
dix A. The results are also shown in the table of Appendix A.
The second database includes the 70 experimental results of
beams shear strengthened with NSM-FRP collected from DABA-
SUM [23] plus the 31 experimental results carried out at the
Eduardo Torroja Institute, Spain [28] both used for the training
and validation of the ANNs presented in Section 2 (Appendix A).
The beam parameters available from the experimental database
are the load configuration, the geometry of the beam, the mechan-
ical properties and configuration of concrete and internal
reinforcement, and the geometry, configuration and mechanical
properties of the external reinforcement.
In all the works performed in this study, the following parame-
ters were chosen for the application of the proposed multiobjective
genetic algorithm: (a) size of population = 100; (b) crossover prob-
ability = 0.6; (c) mutation probability = 0.03; and (d) maximum
number of generations = 100. These values are typical in GAs and
were chosen after some previous numerical tests. Each parameter,
C1 to C8, was coded into a 6-bit binary number and, then, the eight
parameters were translated into a chromosome of 48 bits length. A
higher number of bits would involve more precision in order to
estimate the parameters. Once the optimal chromosome has been Fig. 2a. Familiar data.
 R. Perera et al. / Composites: Part B 66 (2014) 162–173 167

Fig. 3a. Neural networks: Total set of data.

Fig. 2b. Unfamiliar data.

Fig. 3b. Multi-objective proposal: Total set of data.

Figs. 2a and 2b show the predictions of the two proposed
models as compared to the experimental values for both familiar
and unfamiliar data of the model. These last, not included in the
development of the model, were used as validation set. In both
approaches, the same validation set was used. The predictions lie
above or below the target line, i.e., the line where the predicted
value is equal to the experimental value. The nearer the points
gather around the diagonal line, the better the predicted values.
From a comparison point of view, the linear regression slope and
the linear correlation coefficient for the proposed models are
shown in Table 1. As was to be expected, in the case of using ANNs, Fig. 4a. Neural networks: Unfamiliar data.
better correlations are obtained for the training set although in
both cases, the low scatter of data around the diagonal line con-
firms the efficiency of the ANN as predictor of the shear stress
capacity. However, in the multi-objective approach the perfor-
mance for both familiar and unfamiliar data is practically similar,
which confirms the ability of this proposal to be used as design
equation. Comparisons of the shear strength predictions made by
both methods for the whole of the data sets are shown in
Figs. 3a and 3b. As expected, the neural network approach outper-
forms the multi-objective proposal. However, when only validation
data are considered (Figs. 4a and 4b) the average performance of Fig. 4b. Multi-objective proposal: Unfamiliar data.

the multi-objective design equation is better than that derived

from the neural network approach.
The statistical performance of the proposed models is shown in confirm the goodness of fit obtained with the proposed design
Table 2. The comparison is based on the average standard deviation equation obtained with the multi-objective approach.
(SD) and coefficient of variation (COV) of the ratio Vexp/Vpred. With Comparisons have also been carried out, when possible, with
the same purpose, the average absolute error (AAE) calculated for n theoretical predictions given by the numerical model proposed in
specimens as [14]. In this model, the FRP contribution to the shear capacity is
evaluated assuming that the failure mode is by debonding of the
1X n
jV exp  V pred j NSM reinforcement taking place in each single NSM bar separately.
AAE ¼ 100 ð9Þ
n i¼1 V exp The shear force resisted by the FRP is computed as the sum of the
forces resisted by the FRP rods intersected by a shear crack which
is also shown in Table 2. In general terms, it is clear that, statisti- will depend on the embedment lengths of all the bars crossed by
cally, the predictions with both procedures are very good and the crack and the bond strength. Then, the value of Vf

Table 1
Linear regression slopes and coefficients of correlation of all the predictions performed.

Familiar data set Unfamiliar data set Total data set

Neural Multi-objective design Neural Multi-objective design Neural Multi-objective design
network equation network equation network equation
Linear regression slope 0.98 0.8478 0.94 0.8674 0.97 0.8473
Linear correlation 0.9834 0.9138 0.9662 0.9028 0.9798 0.9115
coefficient
168 R. Perera et al. / Composites: Part B 66 (2014) 162–173

Table 2
Average, SD, COV and AAE of Vexp/Vpred for the predicted values.

Neural networks Multi-objective design equations

Familiar data set Unfamiliar data set Total data set Familiar data set Unfamiliar data set Total data set
Average 1.01 1 1.01 0.98 0.98 0.98
SD 0.07 0.08 0.07 0.13 0.15 0.14
COV % 6.50 7.59 6.72 13.58 15.51 13.91
AAE % 3.56 6.06 4.08 11.28 13.24 11.69

parametric analysis results are generally cases that would be

expected from a structural engineering viewpoint. The tendencies
Table 3
Comparison of the two proposed approaches with the Rizzo and De Lorenzis’ model evaluated with both methods do not differ a lot except for the most
[14] for the eight specimens tested by Dias and Barros [24]. sensitive parameters in which a higher disparity occurs. Table 5
shows the regression coefficients for every relevant input variable
Neural Multi-objective design Rizzo and De Lorenzis’
network equation model and for the two proposals presented in this work. These coeffi-
cients indicate the increment of output for every unit increment
Average 1.02 1.02 1.08
SD 0.05 0.03 0.16
of the considered input relative to its standard deviation. Positive
COV % 4.61 3.11 14.89 coefficients indicate increments of the shear strength when the
AAE % 2.86 2.44 13.29 input variable increases, while the opposite happens for negative
coefficients. The higher the absolute values of the coefficients,
corresponding to the most unfavourable crack position and crack is the higher the variation of the shear strength. From Table 5, it is
calculated. To this value, the concrete and steel contributions are observed that the input parameters that mostly affect the shear
added to determine the shear capacity. For concrete contribution, strength are, in decreasing order, the concrete average compressive
values computed using the Model Code [42] and ACI 318 [32] were strength fcm, the FRP reinforcement ratio qf, and the ratio of the
checked. Better predictions, when compared to experimental steel shear reinforcement qsw. However, the distance between
results, were obtained by using ACI code and, for this reason, only FRP bars sf has a high influence when analyzed with the
these results are shown here. The steel stirrups contribution is multi-objective design equation by comparison with the neural
evaluated considering the number of steel stirrups intersected by network approach. In general, for most of the parameters,
the shear crack and the crack opening if the bars have not yielded. regression coefficients are higher for the multi-objective approach
However, with this model, the calculation of Vf requires the than for the NN approach and are also more consistent with the
availability of bond tests to estimate the bond strength. Due to this, predictions carried out in [30] which might reveal the higher
only when these bond tests were available for the strengthened effectiveness of the multi-objective proposal approach when
beams of the experimental database, could numerical predictions applied to a high number of specimens beyond the training
with this model be carried out. Eight experimental specimens domain, as happens in the parametric study.
taken from [24] were used for comparison. The predictions are
shown in Table 3. The two proposed approaches work better than 6. Statistical considerations of the design equation
the Rizzo and De Lorenzis’ model. Furthermore, in this case, the
predictions are slightly better when the proposed design equation In a further step, the reliability of the proposed design equation
is used. should be evaluated considering the uncertainty of the studied
phenomenon with the purpose of reaching the levels of safety
5. Sensitivity and parametric analyses required for design purposes, i.e. to defining its characteristic
value. For this, the design by testing approach originally proposed
For further verification of both proposed models, a parametric in [43] and applied in [44,45] can be used.
analysis was performed in this study. This analysis might be used A new variable or error function d that represents the model
to evaluate the input parameters that mostly affect the result. error for all available test data is defined as follows
For this study, an average value was computed for all the input V pred ¼ V pred;det ðfcm ; bw ; d; a; As ; s; fy ; Ef ; Af ; af ; sf Þd ð10Þ
parameters in both proposals from all the experimental tests.
Further, a range of variation was chosen for each parameter to be where Vpred,det is the deterministic relationship defined with the
studied also taking as a reference the values of the experimental algorithm proposed in Section 3 and evaluated in Section 4 and d
beams (Table 4). At each analysis, all the input parameters were is a random variable assumed to be normally distributed with unit
set to the average value while the chosen parameter to be studied mean. Eq. (10) represents the probabilistic strength model for RC
was varied into the range of values specified in Table 4. Fig. 5 beams shear strengthened with NSM FRP rods
presents the tendency of the shear strength predictions to the The error function is defined for each ith experimental test of
variations of various design parameters, af, fcm, a/d, Ef, sf, qf and qsw. the available database as follows
As shown in this figure, the shear strength of the concrete beam V exp;i
continuously increases due to increasing the parameters of fcm, qf di ¼ ð11Þ
V pred;det ðfcm;i ; bw;i ; di ; ai ; As;i ; si ; fy;i ; Ef ;i ; Af ;i ; af ;i ; sf ;i Þ
and qsw and decreases with increases in af, Ef, sf and a/d. The

Table 4
Descriptive statistics of the variables used in the sensitivity study .

Parameter qf fcm (MPa) a/d sf (mm) Ef (MPa) af qsw (mm)

Mean 0.0022 34.79 2.8 178 163,000 65 0.25
Standard deviation 0.0029 18 0.48 0.0974 28136.5 15.14 0.076
Range 0.0005–0.0112 18.60–59.40 2.2–3.3 45–365 104,000–174,000 45–90 0.19–0.36
 R. Perera et al. / Composites: Part B 66 (2014) 162–173 169

Fig. 5. Shear strength parametric analysis.

where Vexp,i and Vpred,det(fcm,i, bw,i, di, ai, As,i, si, fy,i, Ef,i, Af,i, af,i, sf,i) are The following expression is defined for the 5% percentile
the experimental and predicted shear strength values, respectively, (characteristic) value of the shear strength defined in Eq. (10)
for the ith test.
From Eq. (11), the mean value, variance and CoV of the error
function are estimated as V pred;k ¼ V pred;det  1:64½VarðV pred Þ0:5 ð13Þ

X
n
1 X n
d ¼ 1 di s2d ¼ ðdi  dÞ
2
CoV ¼
sd
ð12Þ where the variance of Vpred is evaluated from the variables chosen to
n i¼1
n  1 i¼1 d
be uncertain.
170 R. Perera et al. / Composites: Part B 66 (2014) 162–173

Table 5 For the coefficients of variation values of 0.05 and 0.2 have been
Regression coefficients for the contribution of each variable to the shear strength. assumed for Ef and fcm, respectively. The coefficient of variation of
Parameter MO design equation Neural network the variable d is estimated from the statistical distribution of the
qf 50.42 12.37
experimental data applying Eq. (12)
fcm 73.11 17.94 Fig. 6a shows the comparison based on the ratio Vexp/Vpred
a/d 5.23 11.68 between the design values obtained for all beams in the experi-
sf 23.83 7.38 mental database by applying the multi-objective approach
Ef 2.06 14.6
proposed in Section 3 and the corresponding values obtained by
af 8.75 8.74
qsw 22.72 18.8 using the proposed design-by-testing procedure followed in this
paper (Eq. (13)). Although, as shown in Section 4, the shear
strength calibrated by solving the multi-objective approach gives
a good approximation of the experimental results, it cannot be
used for design purposes due to its non-conservative nature.
However, the points representing the predictions obtained with
Eq. (13), which cover the randomness and uncertainty of the
phenomenon, show an acceptable level of safety and reliability
for a design procedure. Furthermore, a level of safety might be eas-
ily increased by using percentiles lower than 0.05. As an example,
Fig. 6b also shows values corresponding to the 0.005 percentil.

7. Conclusions

Two new methods have been proposed for estimating the shear
strength of NSM-FRP RC beams. Firstly, an artificial neural network
has been developed to predict this shear strength using a database
of experimental results. Different criteria have been used to choose
the most suitable network. Secondly, an alternative approach has
been implemented based on solving a multi-objective optimization
problem and whose purpose is to provide a user-friendly design
equation rather than complex models. This approach is formulated
by simultaneously minimizing the difference between the
measured shear strength of reinforced concrete beams with and
without NSM-FRP reinforcement and their calculated values with
proposed equations based on design equations and previous
experiences.
It was observed that the performance of both models is satisfac-
tory to a high degree of accuracy when compared to the experi-
mental results available in the literature. Their validity is
guaranteed by the effectiveness and robustness for assessing the
shear strength beyond the training domain. Both models are
expected to be very useful for the evaluation of the shear capacity
Fig. 6. Characteristic value of the shear strength. of RC members strengthened with NSM-FRP in pre-design stages.
The proposed formulations were subsequently employed to
carry out some parametric studies with the purpose of identifying
the more relevant parameters affecting the shear strength of this
For simplicity, if we consider the parameters Ef and fcm as
kind of strengthened beams. The results of the sensitivity analysis
normal random variables statistically independent from one
indicate that fcm, qf, qsw and sf exert a dominant influence on the
another, the variance of Vpred is given by
variations of the shear resistance of RC beams strengthened with
NSM-FRP.
VarðV pred Þ ¼ C 2Ef  VarðEf Þ þ C 2fcm  Varðfcm Þ þ C 2d  VarðdÞ ð14Þ Finally, a design-by-testing procedure was developed to state a
simplified design formula considering the randomness and
where the coefficients are the values of the partial derivatives of uncertainty involved with the studied phenomenon. By using this
Vpred with respect to the basic random variables evaluated from formula, the safety levels of the design procedure can be
Eq. (10) and computed at the mean values of the basic variables; controlled.
Cd is the value of the partial derivative ofVpred with respect to the
error function
  Acknowledgement
@V pred
C Ef ¼
@Ef Ef ;m ;fcm;m ;dm ¼1 The writers acknowledge the support for the work reported in
  this paper from the Spanish Ministry of Science and Innovation
@V pred (Project BIA2010-20234-C03-01).
C fcm ¼ ð15Þ
@fcm Ef ;m ;fcm;m ;dm ¼1
  Appendix A
@V pred
Cd ¼
@d Ef ;m ;fcm;m ;dm ¼1
See Table A1.
 R. Perera et al. / Composites: Part B 66 (2014) 162–173 171

Table A1
Geometrical and mechanical data of the experimental RC beams.

Authors Label Shape bw h fcm qsl fsy,w qsw Ef sf qf af a/ Vexp VNN VMO
(T, R) (mm) (mm) (MPa) (MPa) (mm) (GPa) (mm) d (kN) (kN) (kN)
De Lorenzis and Nanni [7] B90-7 T 152.4 406 31.0 0.024 0 0.00 105 178 0.0052 90 3.0 115.2 126.42 139.00
B90-5 T 152.4 406 31.0 0.024 0 0.00 105 127 0.0073 90 3.0 127.7 114.91 152.78
B45-7 T 152.4 406 31.0 0.024 0 0.00 105 178 0.0052 45 3.0 165.5 165.63 173.90
B45-5 T 152.4 406 31.0 0.024 0 0.00 105 127 0.0073 45 3.0 177.9 175.99 193.14
Barros and Dias [11] A10_VL R 150 300 49.2 0.008 0 0.00 150 200 0.0009 90 2.2 79.3 79.23 113.41
A10_IL R 150 300 49.2 0.008 0 0.00 150 300 0.0006 45 2.2 79.0 80.60 122.21
A10_VL1 R 150 300 49.2 0.008 0 0.00 150 200 0.0009 90 2.2 79.3 79.23 113.41
A12_VL R 150 300 49.2 0.011 0 0.00 150 100 0.0019 90 2.2 117.6 117.78 131.14
A12_IL R 150 300 49.2 0.011 0 0.00 150 150 0.0012 45 2.2 131.2 131.21 143.00
B10_VL R 150 150 56.2 0.017 0 0.00 150 100 0.0019 90 2.4 65.6 64.10 65.47
B12_VL R 150 150 56.2 0.025 0 0.00 150 50 0.0037 90 2.4 69.6 69.37 77.96
Dias and Barros [24] 2S-4LV-B T 180 400 39.7 0.028 542 0.19 171 180 0.0008 90 2.5 202.4 204.33 207.69
2S-7LV-B T 180 400 39.7 0.028 542 0.19 171 114 0.0013 90 2.5 224.5 225.40 220.55
2S-10LV-B T 180 400 39.7 0.028 542 0.19 171 80 0.0018 90 2.5 238.5 238.55 232.41
2S-4LI45-B T 180 400 39.7 0.028 542 0.19 171 275 0.0005 45 2.5 235.7 235.76 217.35
2S-7LI45-B T 180 400 39.7 0.028 542 0.19 171 157 0.0009 45 2.5 253.0 249.97 236.13
2S-10LI45-B T 180 400 39.7 0.028 542 0.19 171 110 0.0013 45 2.5 267.9 268.15 250.62
2S-4LI60-B T 180 400 39.7 0.028 542 0.19 171 243 0.0006 60 2.5 231.8 229.91 218.89
2S-6LI60-B T 180 400 39.7 0.028 542 0.19 171 162 0.0009 60 2.5 236.6 243.59 232.31
2S-9LI60-B T 180 400 39.7 0.028 542 0.19 171 108 0.0014 60 2.5 247.6 251.84 248.28
4S-4LV-B T 180 400 39.7 0.028 542 0.32 171 180 0.0008 90 2.5 254.7 252.16 253.79
4S-4LVa-B T 180 400 39.7 0.028 542 0.32 171 180 0.0008 90 2.5 263.5 252.16 253.79
4S-7LV-B T 180 400 39.7 0.028 542 0.32 171 114 0.0013 90 2.5 256.4 256.79 266.65
4S-4LI45-B T 180 400 39.7 0.028 542 0.32 171 275 0.0005 45 2.5 265.5 264.98 263.45
4S-7LI45-B T 180 400 39.7 0.028 542 0.32 171 157 0.0009 45 2.5 286.9 258.13 282.23
4S-4LI60-B T 180 400 39.7 0.028 542 0.32 171 243 0.0006 60 2.5 266.3 265.47 264.99
4S-6LI60-B T 180 400 39.7 0.028 542 0.32 171 162 0.0009 60 2.5 274.6 257.17 278.40
Rizzo and De Lorenzis [13] NB90-73-a R 200 210 29.3 0.044 665.3 0.36 146 73 0.0069 90 3.0 176.4 162.35 164.83
NB90-73-b R 200 210 29.3 0.044 665.3 0.36 146 73 0.0069 90 3.0 148.6 162.35 164.83
NB90-45-b R 200 210 29.3 0.044 665.3 0.36 146 45 0.0112 90 3.0 150.8 148.99 178.07
NB45-146-a R 200 210 29.3 0.044 665.3 0.36 146 146 0.0034 45 3.0 161.3 160.22 166.89
NB45-73-a R 200 210 29.3 0.044 665.3 0.36 146 73 0.0069 45 3.0 150.2 149.01 187.48
NS90-73-a R 200 210 29.3 0.044 665.3 0.36 122 73 0.0044 90 3.0 172.7 126.40 155.76
NS45-146-a R 200 210 29.3 0.044 665.3 0.36 122 146 0.0022 45 3.0 154.9 155.27 157.49
Kotynia [27] BI-2/3B T 150 360 39.0 0.026 340 0.29 163 360 0.0013 45 2.5 176.6 171.04 188.82
BI-3/5A T 150 360 38.5 0.026 340 0.29 163 210 0.0023 45 2.5 212.6 212.91 206.22
Eduardo Torroja Institute B90-3a R 200 350 22.8 0.020 400 0.25 165 230 0.0022 90 3.0 103.8 112.07 149.97
[28] B90-3b R 200 350 26.0 0.020 400 0.25 165 230 0.0022 90 3.0 114.6 114.73 161.97
B90-6a R 200 350 26.7 0.020 400 0.25 165 115 0.0044 90 3.0 160.1 182.96 183.80
B90-6b R 200 350 24.1 0.020 400 0.25 165 115 0.0044 90 3.0 164.9 161.31 171.91
B45-3a R 200 350 29.1 0.020 400 0.25 165 230 0.0022 45 3.0 177.7 155.91 197.83
B45-3b R 200 350 23.9 0.020 400 0.25 165 230 0.0022 45 3.0 183.5 179.34 173.46
B45-6a R 200 350 23.0 0.020 400 0.25 165 115 0.0044 45 3.0 173.2 179.99 191.89
B45-6b R 200 350 28.5 0.020 400 0.25 165 115 0.0044 45 3.0 232.7 212.86 223.85
L90-3a R 200 350 22.8 0.020 400 0.25 165 230 0.0016 90 3.0 111.1 116.02 144.45
L90-3b R 200 350 26.0 0.020 400 0.25 165 230 0.0016 90 3.0 124.6 138.49 155.58
L90-6a R 200 350 26.7 0.020 400 0.25 165 115 0.0033 90 3.0 177.2 188.24 174.94
L90-6b R 200 350 24.1 0.020 400 0.25 165 115 0.0033 90 3.0 137.1 165.95 164.00
L45-3a R 200 350 29.1 0.020 400 0.25 165 230 0.0016 45 3.0 172.8 173.93 187.71
L45-3b R 200 350 23.9 0.020 400 0.25 165 230 0.0016 45 3.0 209.0 185.63 165.34
L45-6a R 200 350 23.0 0.020 400 0.25 165 115 0.0033 45 3.0 180.9 185.67 181.43
L45-6b R 200 350 28.5 0.020 400 0.25 165 115 0.0033 45 3.0 226.0 222.91 210.54
B90-3a R 200 350 22.8 0.020 400 0.25 165 230 0.0022 90 3.0 131.2 112.07 149.97
B90-3b R 200 350 26.0 0.020 400 0.25 165 230 0.0022 90 3.0 120.6 114.73 161.97
B90-6a R 200 350 26.7 0.020 400 0.25 165 115 0.0044 90 3.0 181.0 182.96 183.80
B90-6b R 200 350 24.1 0.020 400 0.25 165 115 0.0044 90 3.0 162.0 161.31 171.91
B45-3a R 200 350 29.1 0.020 400 0.25 165 230 0.0022 45 3.0 133.6 155.91 197.83
B45-3b R 200 350 23.9 0.020 400 0.25 165 230 0.0022 45 3.0 194.9 179.34 173.46
B45-6a R 200 350 23.0 0.020 400 0.25 165 115 0.0044 45 3.0 188.8 179.99 191.89
B45-6b R 200 350 28.5 0.020 400 0.25 165 115 0.0044 45 3.0 193.1 212.86 223.85
L90-3a R 200 350 22.8 0.020 400 0.25 165 230 0.0016 90 3.0 123.1 116.02 144.45
L90-3b R 200 350 26.0 0.020 400 0.25 165 230 0.0016 90 3.0 138.8 138.49 155.58
L90-6a R 200 350 26.7 0.020 400 0.25 165 115 0.0033 90 3.0 198.8 188.24 174.94
L90-6b R 200 350 24.1 0.020 400 0.25 165 115 0.0033 90 3.0 157.6 165.95 164.00
L45-3a R 200 350 29.1 0.020 400 0.25 165 230 0.0016 45 3.0 174.3 173.93 187.71
L45-3b R 200 350 23.9 0.020 400 0.25 165 230 0.0016 45 3.0 203.4 185.63 165.34
L45-6a R 200 350 23.0 0.020 400 0.25 165 115 0.0033 45 3.0 186.7 185.67 181.43
L45-6b R 200 350 28.5 0.020 400 0.25 165 115 0.0033 45 3.0 216.0 222.91 210.54
Dias and Barros [15] 2S-3LV-A T 180 400 31.1 0.029 533 0.19 167 267 0.0006 90 2.5 189.6 190.41 172.94
2S-5LV-A T 180 400 31.1 0.029 533 0.19 167 160 0.0010 90 2.5 214.2 211.06 182.54

(continued on next page)

172 R. Perera et al. / Composites: Part B 66 (2014) 162–173

Table A1 (continued)

Authors Label Shape bw h fcm qsl fsy,w qsw Ef sf qf af a/ Vexp VNN VMO
(T, R) (mm) (mm) (MPa) (MPa) (mm) (GPa) (mm) d (kN) (kN) (kN)
2S-8LV-A T 180 400 31.1 0.029 533 0.19 167 100 0.0016 90 2.5 237.6 222.86 193.43
2S-3LI45-A T 180 400 31.1 0.029 533 0.19 167 367 0.0006 45 2.5 196.8 203.24 181.41
2S-5LI45-A T 180 400 31.1 0.029 533 0.19 167 220 0.0010 45 2.5 230.4 230.44 193.09
2S-8LI45-A T 180 400 31.1 0.029 533 0.19 167 138 0.0016 45 2.5 229.2 230.59 206.23
2S-3LI60-A T 180 400 31.1 0.029 533 0.19 167 325 0.0006 60 2.5 224.4 225.64 182.42
2S-5LI60-A T 180 400 31.1 0.029 533 0.19 167 195 0.0009 60 2.5 235.2 240.14 194.34
2S-7LI60-A T 180 400 31.1 0.029 533 0.19 167 139 0.0013 60 2.5 243.6 249.87 203.80
Dias and Barros [25] 2S-7LV-C T 180 400 18.6 0.028 539 0.19 174 114 0.0013 90 2.5 164.2 163.34 144.62
2S-4LI45-C T 180 400 18.6 0.028 539 0.19 174 275 0.0008 45 2.5 169.8 155.55 143.25
2S-7LI45-C T 180 400 18.6 0.028 539 0.19 174 157 0.0013 45 2.5 183.9 164.39 151.27
2S-4LI60-C T 180 400 18.6 0.028 539 0.19 174 243 0.0007 60 2.5 169.0 168.92 143.91
2S-6LI60-C T 180 400 18.6 0.028 539 0.19 174 162 0.0011 60 2.5 178.6 182.47 149.64
4S-7LV-C T 180 400 18.6 0.028 539 0.32 174 114 0.0013 90 2.5 189.1 190.63 190.22
4S-4LI45-C T 180 400 18.6 0.028 539 0.32 174 275 0.0008 45 2.5 208.3 200.16 188.85
4S-7LI45-C T 180 400 18.6 0.028 539 0.32 174 157 0.0013 45 2.5 213.8 199.026 196.87
4S-4LI60-C T 180 400 18.6 0.028 539 0.32 174 243 0.0007 60 2.5 207.4 223.53 189.51
4S-6LI60-C T 180 400 18.6 0.028 539 0.32 174 162 0.0011 60 2.5 217.4 214.90 195.24
Dias and Barros [26] 3S-6LV-D T 180 400 59.4 0.031 539 0.19 174 180 0.0008 90 3.3 193.5 193.62 254.79
3S-10LV-D T 180 400 59.4 0.031 539 0.19 174 114 0.0013 90 3.3 245.9 245.95 274.92
3S-5LI45-D T 180 400 59.4 0.031 539 0.19 174 275 0.0008 45 3.3 246.1 246.37 269.92
3S-9LI45-D T 180 400 59.4 0.031 539 0.19 174 157 0.0013 45 3.3 281.8 280.90 299.32
3S-5LI60-D T 180 400 59.4 0.031 539 0.19 174 243 0.0007 60 3.3 249.0 251.09 272.33
3S-8LI60-D T 180 400 59.4 0.031 539 0.19 174 162 0.0011 60 3.3 292.3 273.38 293.34
3S-5LI45F1- T 180 400 59.4 0.031 539 0.19 174 275 0.0008 45 3.3 265.7 246.37 269.92
D
3S-5LI45F2- T 180 400 59.4 0.031 539 0.19 174 275 0.0008 45 3.3 245.3 246.37 269.92
D
5S-5LI45-D T 180 400 59.4 0.031 539 0.29 174 275 0.0008 45 3.3 279.8 298.80 305.75
5S-9LI45-D T 180 400 59.4 0.031 539 0.29 174 157 0.0013 45 3.3 313.8 315.25 335.15
5S-5LI60-D T 180 400 59.4 0.031 539 0.29 174 243 0.0007 60 3.3 278.2 300.47 308.16
5S-8LI60-D T 180 400 59.4 0.031 539 0.29 174 162 0.0011 60 3.3 327.3 313.17 329.17
5S-5LI45F-D T 180 400 59.4 0.031 539 0.29 174 275 0.0008 45 3.3 306.0 298.80 305.75
5S-5LI60F-D T 180 400 59.4 0.031 539 0.29 174 243 0.0007 60 3.3 277.4 300.47 308.16

References [15] Dias SJE, Barros JAO. Shear strengthening of T cross section reinforced concrete
beams by near-surface mounted technique. J Compos Constr
2008;12(3):300–11.
[1] Bank LC. Composites for construction. Wiley; 2006.
[16] Bianco V, Barros JAO, Monti G. Three dimensional mechanical model for
[2] Balaguru P, Nanni A, Giancaspro J. FRP composites for reinforced and
simulating the NSM FRP strips shear strength contribution to RC beams. Eng
prestressed concrete structures. New York and London: Taylor & Francis; 2009.
Struct 2009;31(4):815–26.
[3] Triantafillou TC, Antonopoulos CP. Design of concrete flexural members
strengthened in shear with FRP. J Compos Constr 2000;4(4):198–205. [17] Bianco V, Barros JAO, Monti G. New approach for modeling the contribution of
[4] Täljsten B, Elfgren L. Strengthening concrete beams for shear using CFRP- NSM FRP strips for shear strengthening of RC beams. J Compos Constr ASCE
materials: evaluation of different application methods. Compos Part B – Eng 2010;14(1):36–48.
2000;31(2):87–96. [18] Perera R, Arteaga A, De Diego A. Artificial intelligence techniques for prediction
[5] Chaallal O, Nollet MJ, Perraton D. Strengthening of reinforced concrete beams of the capacity of RC beams strengthened in shear with external FRP
with externally bonded fibre-reinforced-plastic plates: design guidelines for reinforcement. Compos Struct 2010;92(5):1169–75.
shear and flexure. Can J Civ Eng 1998;25(4):692–704. [19] Naderpour H, Kheyroddin A, Ghodrati Amiri G. Prediction of FRP-confined
[6] Perera R, Vique J, Arteaga A, De Diego A. Shear capacity of reinforced concrete compressive strength of concrete using artificial neural networks. Compos
members strengthened in shear with FRP by using strut-and-tie models and Struct 2010;92(12):2817–29.
genetic algorithms. Compos Part B – Eng 2009;40:714–26. [20] Perera R, Barchín M, Arteaga A, De Diego A. Prediction of the ultimate strength
[7] De Lorenzis L, Nanni A. Shear-strengthening of reinforced concrete beams with of reinforced concrete beams FRP-strengthened in shear using neural
near-surface mounted fiber-reinforced polymer rods. ACI Struct J networks. Compos Part B – Eng 2010;41(4):287–98.
2001;98(1):60–8. [21] Tanarslan HM, Secer M, Kumanlioglu A. An approach for estimating the
[8] Täljsten B, Carolin A, Nordin H. Concrete structures strengthened with capacity of RC beams strengthened in shear with FRP reinforcements using
near-surface mounted reinforcement of CFRP. Adv Struct Eng 2003;6(3): artificial neural networks. Constr Build Mater 2012;30:556–68.
201–13. [22] Jalal M, Ramezanianpour AA. Strength enhancement modeling of concrete
[9] Barros JAO, Ferreira DRSM, Fortes AS, Dias SJE. Assessing the effectiveness of cylinders confined with CFRP composites using artificial neural networks.
embedding CFRP laminates in the near surface for structural strengthening. Compos Part B – Eng 2012;43(8):2990–3000.
Constr Build Mater 2006;20(7):478–91. [23] DABASUM: database for FRP-based shear strengthening of reinforced concrete
[10] Teng JG, De Lorenzis L, Wang B, Rong L, Wong TN, Lam L. Debonding failures of beams. Minho University (Portugal). <http://dabasum.civil.uminho.pt/>.
RC beams strengthened with near-surface mounted CFRP strips. J Compos [24] Dias SJE, Barros JAO. Performance of reinforced concrete T beams strengthened
Constr ASCE 2006;10(2):92–105. in shear with NSM CFRP laminates. Eng Struct 2010;32(2):373–84.
[11] Barros JAO, Dias SJE. Near surface mounted CFRP laminates for shear [25] Dias SJE, Barros JAO. Shear strengthening of RC T-section beams with low
strengthening of concrete beams. Cem Concr Compos 2006;28:276–92. strength concrete using NSM CFRP laminates. Cem Concr Compos
[12] De Lorenzis L, Teng JG. Near-surface mounted FRP reinforcement: an emerging 2011;33(2):334–45.
technique for strengthening structures. Compos Part B – Eng [26] Dias SJE, Barros JAO. NSM shear strengthening technique with CFRP laminates
2007;38(2):119–43. applied in high-strength concrete beams with or without pre-cracking.
[13] Rizzo A, De Lorenzis L. Behavior and capacity of RC beams strengthened Compos Part B – Eng 2012;43(2):290–301.
in shear with NSM FRP reinforcement. Constr Build Mater 2009;23(4): [27] Kotynia R. Shear strengthening of RC beams with NSM CFRP laminates. In: Proc
1555–67. FRPRCS-8 international symposium on fiber reinforced polymer reinforcement
[14] Rizzo A, De Lorenzis L. Modeling of debonding failure for RC beams for concrete structures, Patras, CD; 2007.
strengthened in shear with NSM FRP reinforcement. Constr Build Mater [28] Eduardo Torroja Institute for Construction Science (CSIC). Internal report;
2009;23(4):1568–77. 2013.
 R. Perera et al. / Composites: Part B 66 (2014) 162–173 173

[29] Garson GD. Interpreting neural-network connection strengths. AI Expert [36] Abraham A, Jain L, Goldberg R. Evolutionary multiobjective optimization:
1991;6(4):47–51. theoretical advances and applications. London: Springer-Verlag; 2005.
[30] Bianco B, Monti G, Barros JAO. Design formula to evaluate the NSM FRP strips [37] Deb K, Pratap A, Agarwal S, Meyarivan T. A fast and elitist multi-objective
shear strength contribution to a RC beam. Compos Part B – Eng genetic algorithm: NSGA-II. IEEE Trans Evol Comput 2002;6(2):181–97.
2014;56:960–71. [38] Srinivas N, Deb K. Multiobjective optimization using nondominated sorting in
[31] Coello CA. Recent trends in evolutionary multiobjective optimization. In: genetic algorithms. Evol Comput 1994;2(3):221–48.
Abraham, Jain, Goldberg, editors. Evolutionary multiobjective optimization: [39] Perera R, Vique J. Strut-and-tie modelling of reinforced concrete beams using
theoretical advances and applications. London: Springer-Verlag; 2005. genetic algorithms optimization. Constr Build Mater 2009;23:2914–25.
[32] American Concrete Institute. ACI building code requirements for structural [40] Cladera A. Shear design of reinforced high strength concrete beams. PhD thesis
concrete and commentary. ACI 318-05; 2005. published by ACHE (Spanish Concrete Association), ISBN 84-89670-2; 2003.
[33] Kani MW, Huggins MW, Wittkopp RR. Kani on shear in reinforced [41] Bhatt P, Kader MA. Prediction of shear strength of reinforced concrete beams
concrete. Toronto: Department of Civil Engineering, University of Toronto by nonlinear finite element analysis. Comput Struct 1998;68:139–55.
Press; 1979. [42] CEB-FIP Model Code 90. Bulletin d’Information No. 213/214, final version.
[34] Perera R, Ruiz A. Design equations for reinforced concrete members [43] European Committee for Standardization. EN 1990 – Eurocode – base of
strengthened in shear with external FRP reinforcement formulated in an structural design; 2002.
evolutionary multi-objective framework. Compos Part B – Eng [44] Monti G, Alessandri S, Santini S. A procedure for the statistical determination
2012;43(2):488–96. of capacity models. Constr Build Mater 2009;23:1487–94.
[35] fib – Bulletin 14. Externally bonded FRP reinforcement for RC structures. [45] Bilotta A, Faella C, Martinelli E, Nigro E. Design by testing procedure for
Technical report by Task Group 9.3 FRP (fiber reinforced polymer) intermediate debonding in EBR FRP strengthened RC beams. Eng Struct
reinforcement for concrete structures. Féderation Internationale du Béton – 2013;46:147–54.
fib; 2001. 130p.