You are on page 1of 13

ASSESSING THE RELIABILITY OF RC CODE PREDICTIONS THROUGH THE USE OF

ARTIFICIAL NEURAL NETWORKS


Ahmad A, Demitrios M. Cotsovos, Heriot-Watt University, Edinburgh, UK
Nikos D. Lagaros, National Technical University of Athens, Athens, GR
Abstract:
A large number of tests have been conducted to date in order to investigate in detail the behaviour of simply supported
reinforced concrete (RC) beam specimens under static loading. The test data obtained forms the basis for the
development of the physical models currently adopted by the codes of practice (e.g. ACI, EC2 and JSCE) for the
ultimate limit-state (ULS) design of RC members. However, a comparison of the code predictions concerning important
aspects of RC structural response at the ULS with their experimental counterparts often exhibits significant differences.
This raises concerns regarding the validity of the fundamental assumptions adopted by the various codes concerning the
mechanics underlying RC structural response at the ULS as well as the effectiveness of the proposed design solutions in
safeguarding the often stringent structural performance requirements (dictated by the codes of practice) usually
associated with strength and ductility. Present work forms the initial step of a more general study aiming to assess the
reliability of the RC design codes by comparing their predictions with their counterparts established: (i) experimentally,
(ii) numerically via nonlinear finite element analysis (iii) through the use of Artificial Neural Networks (ANNs) and (iv)
by employing an alternative physical model adopted by the Compressive Force Path (CFP) method which provides
design solutions considerably different to those of the available design codes. In the present study ANNs are employed
to directly assess the available test data and to objectively quantify the effect of important design parameters on RC
structural response. The training and validation of the ANNs is achieved through the use of databases formed by the test
data mentioned above. The input parameters are selected on the basis of the physical models employed for describing the
mechanisms underlying RC structural response at the ULS and are associated with the specimen design details. The
comparative study reveals that for certain values of input parameters the predictions of the CFP method and the ANNs
correlate closer to the relevant experimental data than their counterparts obtained by the codes. This highlights the
urgent need to re-assess current codes for RC design and the underlying assumptions upon which they are based.
Keywords: Reliability, Design Codes, Ultimate limit state, RC beams, Shear-span ratio, Artificial Neural Network, CFP

NOMENCLATURE ULS assume that an RC member, after crack-


𝑎𝑣 : shear span formation, behaves essentially as a truss, with
b, bw : beam width concrete and reinforcement bars acting in
d : effective depth compression forming struts and the reinforcement
x : depth of the compressive zone bars acting in tension forming ties [Fig.1b]. On the
As : area of tensile reinforcement basis of the above physical models a sectional
Av : area of transverse reinforcement approach is usually adopted when assessing load-
fc : uniaxial compressive strength of bearing capacity. This process is limited to the
concrete “critical” regions [lcr, Fig.1b] of the RC member
fy : longitudinal reinforcement yield stress which are characterised by high values of internal
fyw : transverse reinforcement yield stress actions (i.e. shear forces, bending moments)
[Fig.1a]. At these regions the flexural and shear
Vc , VRd,c : shear resistance of concrete alone
capacity are assessed independently. In the case of
Vs : shear resistance of the transverse
flexural capacity a generally good agreement is
reinforcement
observed between the predicted and
V : shear resistance at failure
experimentally established values, whereas in the
M : bending moment at failure
case of shear capacity the above values often
1. INTRODUCTION exhibit significant differences. Such differences
In practice, the ultimate limit state (ULS) design can be attributed to the empirical nature of the
of reinforced concrete (RC) members is carried formulae employed for predicting shear capacity,
out through the available design codes (e.g. the derivation of which is based on the regression
ACI[1], EC2[2] and JSCE [3]).The physical analysis of available test data which is
models (i.e. truss-analogy, strut & tie models) [4- characterised by considerable scatter.
6] adopted by the majority of available codes for The Compressive Force Path (CFP) method
describing RC structural response approaching the [7]adopts a physical model for describing RC

306
structural behaviour at the ULS which is production of powerful and computationally
fundamentally different to those adopted by the efficient numerical tools capable of successfully
codes of practice. Its predictions have been replacing time-consuming NLFEA procedures in
validated for a range of RC structural forms under structural analysis. Such tools are lately being
static and dynamic loads [7]. It assumes that RC successfully employed to predict RC structural
members, after crack-formation, behave as arch- response at the ULS [12-24].
like structures connected at the points of contra-
flexure (in the case statically indeterminate
elements, Fig.1a). This type of behaviour is
enforced by the reinforcement provided. It should
be noted that the areas of the RC structure
considered as “critical” by the CFP method
[locations 1-7 in Fig.1c] are different to the critical
regions (lcr) considered by the codes [Fig.1c].
Furthermore, it is interesting to point out, that
unlike current RC design codes (which adopt a
sectional approach), the CFP method considers the
structural elements as a whole, accounting for the
geometry, boundary conditions and the true
(triaxial) stress-state that develops within concrete
[7]. Due to the above differences, the CFP method
can lead to different predictions of structural
behaviour and design solutions compared to those
of the codes without compromising the
requirements for strength and ductility.
This detailed study of RC structural response and
the validation of the predictions of the previous
design methods are achieved either experimentally
Figure 1: (a) indeterminate RC beam specimen and its
or numerically. To date a large number of tests
bending moment and shear force diagrams at
have been conducted on simply supported RC the ULS, b) TA Model, c) CFP Model
beam specimens subjected to concentrated static
loads applied at specific points along their span. Present work forms the initial step of a more
This data forms the basis for the development of a general study aiming to assess the reliability/
number of databases [8,9] describing specific accuracy of the predictions of the RC design codes
aspects of RC structural response at the ULS (e.g. and their underlying assumptions concerning the
load-bearing capacity, modes of failure). Such mechanics associated with RC structural response
databases can be further enriched (to account for a at the ULS. This is achieved by comparing the
wider range of input parameters) through the use various code-predictions concerning load-bearing
of nonlinear finite element analysis (NLFEA) capacity and mode-of-failure with their
packages which have been developed to predict counterparts established: (i) experimentally, (ii)
RC structural response [10, 11]. numerically, (ii) by appropriately trained ANNs
and (iv) by employing the CFP method. The latter
In addition to the (more traditional) methods methods forms the basis for ongoing research
discussed earlier, Soft Computing (SC) methods assessing the reliability of the existing RC design
can be used to form powerful computation tools methods and the ability of NNs and alternative
which are being increasingly used in structural assessment methods (i.e. the CFP method) of
analysis. Unlike NLFEA procedures, such being implemented in practical structural analysis
methods are based on heuristic approaches which for accurately predicting RC structural response
do not strictly adhere to the principles of even for complex load scenarios (e.g. associated
theoretical mechanics. Neural Networks (NNs) with fire, high-rate loading or environmental
have been frequently employed as the basis for the actions) on short term or long term response. In
development of SC methods resulting in the
307
the present article, although the comparative study - The area of the compressive zone has a
is limited to simply-supported RC beams under significant effect on shear capacity.
static loading, it is shown that for certain values of - In the case of cyclic (e.g. earthquake) loading
the parameters considered, the predictions of the the inclined struts of the TA cannot form due to
CFP method and the ANNs correlate closer to the the densely spaced inclined intersecting cracks
relevant experimental data than their counterparts on the web of the RC element
established by the design codes. This raises
questions concerning the validity of the
assumptions adopted by the various physical
models adopted by the majority of design codes.
The latter observation raises further uncertainty
concerning the ability of such models to form the
basis for the development of more advanced
assessment methods that can realistically predict
RC structural response under more complex
loading scenarios associated with high Figure 2: Truss analogy model
temperatures, fire, impact, blast as well as The latter assumptions are incompatible with the
environmental actions. TA and lead to a fundamentally different physical
2. RC DESIGN CODES model for describing RC structural behaviour at
the ULS. In the case of simply-supported RC
The majority of existing RC design codes are beams subjected to transverse point loads applied
adopt the truss-analogy (TA) for describing the at certain location along the element span the latter
behaviour of RC members at the ULS. Flexural model takes the form presented in Fig.3.Failure is
capacity is owed to the resultant of the considered to occur due to the development of
compressive and tensile stresses acting normal to transverse tensile stresses at specific locations
the cross-section of the element (see Fig.2). In the along the path followed by the compressive force.
case of shear current design philosophy assumes These locations are dependent on the value of the
that the internal shear component acting on an RC shear span-to-depth ratio (av/d). The manner in
structural element is transferred by a combination which av/d affects the beam load-carrying capacity
of: (i) tension acting on the stirrups, (ii) direct (expressed as Mu/Mf) is indicated in Fig.4[7] in
shear transfer by the uncracked portion of the RC which four distinct types of structural element
structural element’s and (iii) through the cracked behaviour are identified.
portion of the element’s cross-section by means of
aggregate interlock and dowel action. On the basis
of the previous assumptions the shear-transfer
mechanism adopted by the codes can be observed Location 2

in Fig.2. As a result shear failure is either


associated with the failure of the vertical ties or
the inclined struts of the TA model representing
the RC structural component considered at ULS.
To determine shear capacity each code employs its
Figure 3: The CFP physical model
own empirical formulae (see Table 6), the
derivation of which is based on the regression Type I behaviour is characterized by a flexural
analysis of available test data. mode of failure preceded by longitudinal splitting
of the concrete in the compressive zone of the
3. THE CFP METHOD
beam. This occurs when concrete strength in the
The fundamental assumptions upon which the compressive zone is exhausted due to the
CFP method is based are: development of transverse tensile stresses induced
- Concrete behavior is brittle and characterized by volume dilation of concrete in the adjacent
by triaxiality regions which include primary flexural cracks. It
- The contribution of aggregate interlock and should be noted that due to the internal actions
dowel action to shear capacity is negligible developing before the loss of bond (τ) on a portion

308
of RC beam between two consecutive cracks location 2 (Fig.3) due to the loss of bond between
(concrete tooth) a complex triaxial state of stress the longitudinal reinforcement and the surrounding
develops within the compressive region (see Fig.5) concrete. Figure 6 indicates a portion of the beam
which differs considerably to that exhibited under between two cross-sections defined by consecutive
uniaxial compression [25]. This allows the cracks (concrete tooth), accompanied with the
maximum stresses developing within the internal forces which develop at these cross-
compressive zone of the beam to attain values sections before and after the loss of bond. Based
approximately equal to 1.5 times fc [7, 25]. on the equations of equilibrium the bending
moment and shear forces are given by eqs(3) & (4).
M=Fs∙z (3)
Type I 𝑎𝑣 /𝑑 > 5
Type II 2.5 < 𝑎𝑣 ⁄𝑑 < 5 V=dM/dx=dFs/dx∙z+Fs∙(dz/dx) (4)
Type III 1 < 𝑎𝑣 ⁄𝑑 < 2.5
Type IV 𝑎𝑣 ⁄𝑑 < 1

Figure 4: Effect of av/d on the load-carrying capacity


(Mu/Mf) of RC beams.
Type II behavior is characterized by a brittle
mode of failure usually caused by tensile stresses
Figure 6: Internal forces developing on a concrete
developing either in the region of change of the tooth (a) before and (b) after the loss of bond.
CFP direction (location 1, see Figs 3 & 5) or in the
region of the cross-section where the maximum Loss of bond can lead to an extension of cracking
bending moment combines with the shear force and, hence, to a reduction of the depth of the
(location 2,see Fig.3 and Fig.5). The transverse compressive zone (see Fig.6b)
stress resultant at location 1 is considered 𝐹𝑐 ∙ (𝑥 − 𝑥1 )/2 = 𝑉 ∙ 𝑥/2 (5)
numerically equal to the acting shear force, and,
its effect is considered to spread over a distance of This reduction leads to an increase of the intensity
d, on either side of location 1. The value of the of the compressive stress field thus leading to
tensile force that can be sustained at this location dilation of the volume of concrete in the
is determined by eq.(1) [7]. compressive zone, which in turn causes the
development of transverse tensile stresses (σt, see
𝑇𝐼𝐼,1 = 0.5 ∙ 𝑏 ∙ 𝑑 ∙ 𝑓𝑡 (1) Figs 5&6) in the adjacent regions Eq.(6).
|σ𝑡 | = 𝑓𝑐 ⁄[5(𝑓𝑦𝑘 𝐴𝑠 ⁄𝑉 − 1)] (6)
By considering these transverse tensile stresses
and the ensuing complex triaxial stress state it is
possible to express the shear force (VII,2) that can
be sustained at locations 2 form Eq. (7) [7].
Figure 5: Internal action developing along the CFP of 𝑉𝐼𝐼,2 = 𝐹𝑐 ∙ [1 − 1/(1 + 5 ∙ 𝑓𝑡 /𝑓𝑐 )] (7)
a beam exhibiting type II behaviour
If the developing shear force is higher than the
If the developing shear force is higher than the value provided by Eq. (6), stirrups are needed. To
value provided by Eq. (1), stirrups are uniformly calculate the amount of stirrups required it is
placed over a length d on both sides of location 1. necessary to calculate the vertical and the
The amount of stirrups required is provided by Eq. horizontal stress resultants in the region between
(2) and their spacing should not exceed 0.5d. (i) the area where the inclined and the horizontal
𝑉 = 𝐴𝑠𝑤,𝐼𝐼1 𝑓𝑦𝑤 (2d/s+1) (2) portion of the compressive path meet (region 1,
Fig.5) and (ii) the point at which the load is
Further to location 1, transverse tensile stresses applied:
within the compressive zone may also develop at
T𝐼𝐼,2𝑣 = σ𝑡 (𝑎𝑣 − 2.5𝑑)𝑏⁄2 (8)
309
T𝐼𝐼,2ℎ = σ𝑡 (𝑎𝑣 − 2.5𝑑)𝑥⁄2 (9) On the other hand, the load-carrying capacity (PD)
corresponding to the strength of the inclined leg of
Based on Eqs (8) & (9), the amount of stirrups
the ‘frame’ will be equal to the vertical component
required is obtained from Eqs (10) & (11):
of the load (FD) that can be carried by this leg. As
A𝑠𝑤,𝐼𝐼2𝑣 = T𝐼𝐼,2𝑣 ⁄𝑓𝑦𝑤 (10) indicated in Fig.8, FD is easily calculated by taking
the depth of the leg equal to av/3 as recommended
A𝑠𝑤,𝐼𝐼2ℎ = T𝐼𝐼,2ℎ ⁄𝑓𝑦𝑤 (11)
in [7].
Type III behavior is characterized by a brittle PD = FD z/ (z2+av2)1/2 (17)
mode of failure caused by the deep penetration of
the inclined, closest to the support, crack into the Where; FD = (av/3) b fc (18)
compressive zone of the beam. This crack reduces
the strength of the uncracked concrete in the Therefore, the load-carrying capacity of a beam in
compressive zone on the region where the inclined the case of type IV behavior will be
and the horizontal compressive path of the model Pu = min(Pf, PD) (19)
meet (region 1, Fig.3), which causes a reduction
on the flexural capacity of the beam. Based on the
internal action presented in Fig.7 a measure of the
maximum shear force that the concrete alone can
carry on this region is provided by Eq.(12).

(a) (b)
Figure 8: (a) Internal actions and (b) cracking on
a beam exhibiting type IV behaviour
Figure 7: Internal actions developing in a RC beam 4. EXPERIMENTAL DATABASES
exhibiting type III behaviour.
𝑀𝐼𝐼𝐼 In the present study two databases are formed; the
𝑉𝐼𝐼𝐼 = (12) first contains data obtained from 352 simply
𝑎𝑣
where: supported RC beam specimens without stirrups
(2.5𝑑) and a second, form 190 simply supported RC
(2.5𝑑) (𝑀𝑓 −𝑀𝐼𝐼 )(2.5𝑑−𝑎𝑣 )
𝑀𝐼𝐼𝐼 = 𝑀𝐼𝐼 + (13) beam specimens with stirrups. This data includes
(1.5𝑑)
(2.5𝑑) parameters associated with the design details of
𝑀𝐼𝐼 = 2.5𝑑𝑉𝐼𝐼,1 (14)
the RC specimens as well as certain aspects of RC
The stirrups required are provided by Eq.(15) and structural response (load-bearing capacity and
are distributed within the shear span with a mode of failure) exhibited at the ULS. All
spacing smaller than 0.5d: specimens are simply supported and subjected to
A𝑠𝑤,𝐼𝐼𝐼 = 2(𝑀𝑓 − 𝑀𝐼𝐼𝐼 )⁄(𝑎𝑣 𝑓𝑦𝑤 ) (15) concentrated loads applied at certain locations
along their span. These databases form the basis of
Type IV behavior can be characterized by two the comparative study mentioned earlier. Table 1
modes of failure either due to the failure of the provides the statistical information concerning the
horizontal element of the CFP model or due to variation of certain parameters associated with the
failure of the uncracked end portion of the beam specimens considered.
(inclined leg of the ‘frame’ of the CFP model) in
compression. From the moment equilibrium of the 5. ARTIFICIAL NEURAL NETWORKS
free body in Fig.8a the flexural capacity (Mf) can ANNs mimics the biological NNs in the central
be easily calculated and consequently the nervous system and the brain of animals and
associated load-carrying capacity (Pf) can be humans (Fig.9). They are used to estimate
determined form Eq.(16). or approximate functions that depend on a large
𝑃𝑓 = 𝑀𝑓 /𝑎𝑣 (16) number of input parameter the effect of which is
not clearly established or quantified. ANNs have

310
the ability to learn, generalize, categorize and The outputs of the activation function form the
predict values due to their adaptive nature and input value for the neurons of the next layer (see
their ability to remember information introduced Fig.9). The weights are initially randomly
to them during their training. They consist of a assigned and their final values are obtained
number of layers. Each layer consists of a system through the training process and use of the
of interconnected "neurons". Each link (forming available data.
between two neurons) has a specific weight. These In the present study NNs are used to predict the
weights are multiplied by the input values load-bearing capacity of simply-supported RC
generated by the neurons. The values obtained beam specimens. Figure 10 describes the process
from all neurons of a specific layer during the through which this is achieved. Based on past
latter process are then transferred through the links experience [26-28] multilayer feed-forward NNs
and summed with the bias (see Fig.9). This latter (MLFNNs) (see Fig. 11) are considered to be the
sum is then introduced into a predefined activation most appropriate for the type of problem presently
function representing the relationship between studied. The later type of NN typically consists of
layers and is described by Eq.(20). an input layer, one or more hidden layers and an
𝑂 = 𝑓(∑𝑥𝑖 𝑤𝑖 + 𝑏) (20) output layer. The number of hidden layers depends
on the architecture of the NNs adopted. Each layer
where O is the output from neural network,
has its own predefined activation function which
xi are the input values,
affects the overall performance of the ANN.
wi are the weight coefficients and
Signals are generated by the neurons and travel
b is the bias value.
from left to right through the path described earlier
Table 1: Statistics analysis of the parameters included (see Fig. 11).
in the databases
Calibration
units Min Max Mean St.Dev+ COV* Database of ANN
Proposed
ANN Models
(a) Beams without stirrups Model
b mm 50 500 174.45 80.45 0.46
d mm 70 681 263.13 102.16 0.38 Normalization Erorr
ANN Type
of Paramters Function
av/d 2.5 5.49 3.385 0.73 0.21
ρl % 0.5 4.68 1.99 0.88 0.44
Selection of
fy MPa 303 555 413.10 61.90 0.14 ANN Optimize
Input
Architecture ANN Model
fc MPa 12.2 79.3 36.91 16.38 0.44 Paramters
Vu KN 7 281 65.24 46.55 0.71
(b) Beams with stirrups Figure 10: Flow Chart of ANN developing
b Mm 100 457 211.64 72.96 0.35 In the present investigation logistic and hyperbolic
d mm 113 925 382.12 172.28 0.46
activation functions are used between the input
av/d 1 6.98 3.36 1.48 0.45 and middle layer and hyperbolic activation
ρl % 0.18 4.65 2.34 0.95 0.41
function are employed between the middle and
fy MPa 250 910 417.95 91.86 0.22
output layer of the NN. The resulting error can be
fc MPa 13.8 108.7 41.82 21.53 0.52
ρw % 0.06 1.23 0.38 0.25 0.66
calculated by the following Eq. (21);
Vu KN 7 876 273.83 216.1 0.79 1
E(w) = 2 ∑i [T 2 − O2 ] (21)
*Coefficient Of Variance
+
Standard Deviation
where T and O are the target (defined in the
Dendrites
Terminal Branches database) and output (predicted by the NN) value
of Axon
x1 respectively
w1
x2
w2 To minimize the error obtained from Eq.(21) the
x3 w3
back-propagation technique (DETA RULE) [29,
S 30] is employed. This process is carried out from
Axon right to left of the NN (see Fig. 11).and makes use
wn of the information included in the databases. Its
xn
aim is to correct/adjust the values of the weights
Figure 9 Artificial Neural Networks (of the links connecting the neurons) which are

311
initially randomly selected, so that the output Table 2a: Values of r expressing the effect between
values of the NN realistically agree with target parameters considered in the database for the case of
values defined in the database. The NN then beams without stirrups
b d av/d ρl fy.l fc Vu
employs the adjusted weights in order to obtain b 1.00
more accurate predictions which in turn result in d 0.45 1.00
av/d -0.10 -0.01 1.00
smaller errors. This iterative process is repeated ρl -0.12 -0.10 0.20 1.00
until the errors become acceptable. fy.l 0.23 0.00 -0.06 -0.03 1.00
fc 0.17 0.01 -0.07 0.30 0.41 1.00
Vu 0.77 0.66 -0.13 0.21 0.27 0.38 1.00

Table 2b: Values of r expressing the effect between


parameters considered in the database for the case of
beams with stirrups
b d av/d ρl ρw fy.l fc Vu
b 1.00
d 0.85 1.00
av/d 0.03 -0.12 1.00
ρl -0.17 -0.18 -0.10 1.00
ρw -0.42 -0.48 -0.32 0.29 1.00
fy.l 0.08 0.06 0.43 -0.48 -0.24 1.00
fc 0.08 -0.09 0.33 0.31 0.12 0.23 1.00
Vu 0.41 0.45 -0.60 0.23 0.19 -0.48 -0.13 1.00

Table 3: Architecture of ANN Models


Beams without Stirrups
No of
Sr. Function Out-
Figure 11: Multi-layer Feed Forward NN (MLFNN) Model Input Hidden
No of Hidden put
Neurons
5.1. SELECTION OF INPUT PARAMETERS 1 ST-4-H-1 b,d,av/d,fc Sigmoid 5-10 [Vu]
2 TT-4-H-1 b,d, av/d,fc tanghn 5-10 [Vu]
The relationship between two parameters 3 ST-5-H-1 b,d, av/d,fc,ρl Sigmoid 6-12 [Vu]
4 TT-5-H-1 b,d, av/d,fc,ρl tanghn 6-12 [Vu]
(considered in the database) is expressed through
5 ST-6-H-1 b,d, av/d,fc,ρl,fy Sigmoid 7-14 [Vu]
the Pearson's correlation coefficient(r) described 6 TT-6-H-1 b,d, av/d,fc,ρl,fy tanghn 7-14 [Vu]
by Eq.(22) [31]. 7 ST-4-H-H-1 b,d,av/d,fc Sigmoid 5-10 [Vu]
8 TT-4-H-H-1 b,d, av/d,fc tanghn 5-10 [Vu]
𝒏𝜮𝒙𝒚−(𝜮𝒙)(𝜮𝒚)
𝒓= (22) 9 ST-5-H-H-1 b,d, av/d,fc,ρl Sigmoid 6-12 [Vu]
√𝒏(𝜮𝒙𝟐 )−(𝜮𝒙)𝟐 √𝒏(𝜮𝒚𝟐 )−(𝜮𝒚)𝟐 10 TT-5-H-H-1 b,d, av/d,fc,ρl tanghn 6-12 [Vu]
11 ST-6-H-H-1 b,d, av/d,fc,ρl,fy Sigmoid 7-14 [Vu]
12 TT-6-H-H-1 b,d, av/d,fc,ρl,fy tanghn 7-14 [Vu]
where: r is Pearson's correlation coefficient,
Beams with Stirrups
n total number of samples, 1 ST-4-H-1 b,d,av/d,fc Sigmoid 5-10 [Vu]
x, y are two variables 2 TT-4-H-1 b,d,av/d,fc tanghn 5-10 [Vu]
3 ST-5-H-1 b,d,av/d,fc,ρl Sigmoid 6-12 [Vu]
Tables 2a&b describe the correlation between the 4 TT-5-H-1 b,d,av/d,fc,ρl tanghn 6-12 [Vu]
parameters considered by the ANN employed. The 5 ST-6-H-1 b,d,av/d,fc,ρl,ρw Sigmoid 7-14 [Vu]
linear correlation coefficient (r) can obtain values 6 TT-6-H-1 b,d,av/d,fc,ρl,ρw tanghn 7-14 [Vu]
7 ST-7-H-1 b,d,av/d,fc,ρl,ρw,fy Sigmoid 8-16 [Vu]
between 1 and -1 (1>r>-1). Negative values of r 8 TT-7-H-1 b,d,av/d,fc,ρl,ρw,fy tanghn 8-16 [Vu]
are considered to result in an inverse affect 9 ST-4-H-H-1 b,d,av/d,fc Sigmoid 5-10 [Vu]
between the values of the two associated 10 TT-4-H-H-1 b,d,av/d,fc tanghn 5-10 [Vu]
11 ST-5-H-H-1 b,d,av/d,fc,ρl Sigmoid 6-12 [Vu]
parameters (i.e. an increase in the value of one
12 TT-5-H-H-1 b,d,av/d,fc,ρl tanghn 6-12 [Vu]
parameter will result in the decrease in the value 13 ST-6-H-H-1 b,d,av/d,fc,ρl,ρw Sigmoid 7-14 [Vu]
of the other). Positive values of r have a similar 14 TT-6-H-H-1 b,d,av/d,fc,ρl,ρw tanghn 7-14 [Vu]
effect on the values of the associated parameters 15 ST-7-H-H-1 b,d,av/d,fc,ρl,ρw,fy Sigmoid 8-16 [Vu]
16 TT-7-H-H-1 b,d,av/d,fc,ρl,ρw,fy tanghn 8-16 [Vu]
(i.e. an increase in the value of one parameter will
result in an increase of the other) [32]. Higher 5.2. ANN ARCHITECTURE
values of |r| result in a more pronounced effect.
After selecting the appropriate combination of
Based the above, different combinations of input
input parameters the architecture of the ANN is
parameters are selected for the case of the beams
determined. This includes the selection of the type
without and with stirrups as shown in Table 3.
of activation function of each layer, the number of

312
hidden layers and the number of neurons in each experimental data. To avoid the over-fitting and
hidden layers. Due to the absence of relevant enhance the performance of the NN, the latter
guidelines, the architecture of the NN depends on normalised data is initially divided into three sub-
the type of problem considered and it is sets: each of these sub-sets is employed for
determined through a trial-and-error process. training, validation and testing purposes.
Figure 12 shows the terminology used in the Matlab [34] is used to develop the NN models and
present study for identifying the characteristics of to randomly divide the database into the
different NN configurations considered in Table 3 previously mentioned sub-sets: 60% of the
in order to identify the most optimal architecture. database is used for training, 20% for validation
Number of Input
and another 20% for testing purposes. Each NN
Activation
Function between
Parameters model is trained for 100 epochs (iterations
Input and Hidden
Layers
including the full MLFFBP process). This process
S-T-4-H-1 is stopped when one of the following conditions is
Number of met: (i) a maximum of 6 validation failures are
Activation Function Single Hidden layer-
Output
Parameters
exhibited (validation failure is exhibited when the
between Hidden and
Output Layer
containing different
number of neurons
NN performance during each iteration fails to
improve or remains constant), (ii) the value of
Activation Number of Input performance goal is 0 (expressing the difference
Function Parameters
between Input between the target and output values) and (iii) the
and Hidden minimum performance gradient (related to the
Layers S-T-4-H-H-1 Number of
Output rate at which the weights are adjusted through the
Parameters
MLFFBP process) becomes 10-10. These
Activation Function
between Hidden and
Double Hidden layer-
containing different
conditions are proposed by the Levenberg-
Output Layer number of neurons Marquardt back-propagation method which is
currently adopted [34].
Figure 12: Terminology describing the structure of
the NNs Models

5.3. NORMALIZATION OF DATABASE


The performance of an ANN depends upon the
quality of the database. The normalization of the
parameters considered in the database has
significant impact on the ANN procedure. (a)
Considering that the various input combinations
are associated with different units the
normalization process allows their conversion to
unit less parameters. To avoid problems associated
with low learning rates of the ANN [20]it is better
to normalize the values of the parameters between (b)
an appropriate upper and lower limit value of the Figure 13: r-value for optimized NN models for the
subject parameter. In this work all parameters case of beams (a) without and (b) with
associated with all beam specimens are normalised stirrups.
between [0.1, 0.9] instead of [0, 1], by using A total of 84 different NN models were created for
following Eq. (23). the case of the RC beams without stirrups and 102
0.8 (0.9−0.8xmax ) NN models for the case of the RC beams with
X=x + (23)
max −xmin xmax −xmin stirrup (see Table 3). From each case, the NN
5.4. CALIBRATION OF ANN MODEL model exhibiting the maximum value of r between
the output and target values (Fig.13) is selected as
The calibration of the NN is achieved through the shown in Table 4.
multi-layer free forward back-propagation process
(MLFFBP) [26-28,33] and the use of the available

313
Table 4: Error Value for Training, validation and the second hidden layer has a significant effect on
Testing for NN models the efficiency of the NN.
Beams without Stirrups
ANN Training Validation Testing All 5.6. FINAL WEIGHTS AND BIAS VALUES
Sr.
Models 60% 20% 20% 100%
1 ST-4-9-1 0.955 0.947 0.932 0.948 Tables 5a&b describe the weights and bias
2 TT-4-10-1 0.969 0.957 0.933 0.956 properties of the selected optimize ANN for both
3 ST-5-7-1 0.991 0.968 0.977 0.984 the cases. By employing the properties defined in
4 TT-5-9-1 0.984 0.979 0.957 0.977
5 ST-6-9-1 0.986 0.984 0.976 0.983
section 5.2 and using the weights and bias values
6 TT-6-10-1 0.983 0.972 0.977 0.98 between different layers of ANN, the prediction
7 ST-4-8-8-1 0.966 0.932 0.932 0.95 can be obtain for unknown output values.
8 TT-4-8-8-1 0.96 0.958 0.939 0.956
9 ST-5-8-8-1 0.986 0.98 0.949 0.981 Table 5a: Weights and Bias Values for beams without
10 TT-5-9-9-1 0.991 0.964 0.972 0.98 Stirrups
11 ST-6-10-10-1 0.989 0.984 0.975 0.985 iw{1,1} b{1,1} b{2,1} b{3,1}
12 TT-6-9-9-1 0.992 0.982 0.972 0.984 -3.632 2.6675 5.0517 -8.368 0.5693 5.339 1.6317 0.388
Beams with Stirrups 3.1865 -6.608 5.7105 -2.857 2.4465 -5.466 -3.797
1 ST-4-10-1 0.984 0.953 0.974 0.975
7.9237 0.8532 -2.707 -3.446 -3.198 0.027 2.3929
2 TT-4-10-1 0.978 0.942 0.948 0.964
-4.956 -6.033 4.2813 0.9676 -1.33 5.674 -1.1
3 ST-5-8-1 0.985 0.967 0.971 0.979
7.2659 -4.715 1.8235 -0.608 5.1597 -3.997 0.8024
4 TT-5-6-1 0.981 0.967 0.976 0.977
4.6836 -6.396 -6.03 -3.582 1.8808 5.493 1.3801
5 ST-6-11-1 0.996 0.976 0.977 0.986
6 TT-6-7-1 0.992 0.977 0.98 0.984 2.2947 -10.61 1.2904 -0.667 1.2709 5.332 1.524
7 ST-7-13-1 0.992 0.96 0.952 0.975 -5.39 -2.503 -6.037 -3.591 1.2125 5.912 -1.076
8 TT-7-13-1 0.991 0.966 0.979 0.984 lw{2,1}
9 ST-4-7-7-1 0.974 0.966 0.975 0.973 0.9545 -0.796 0.6828 -2.263 0.7074 0.168 -0.431 0.7311
10 TT-4-10-10-1 0.981 0.972 0.957 0.975 1.0827 0.4859 0.2999 0.3316 1.8101 -3.319 0.7244 -0.380
11 ST-5-10-10-1 0.986 0.98 0.974 0.98 -1.652 -0.109 -0.617 -1.49 -1.009 -0.331 1.3319 -1.055
12 TT-5-9-9-1 0.988 0.975 0.974 0.983 -0.307 1.4808 0.8582 -0.675 2.6663 -1.317 1.2959 -1.157
13 ST-6-10-10-1 0.991 0.983 0.968 0.986 0.6028 0.4372 -3.063 0.2497 -0.335 1.481 -1.25 -2.065
14 TT-6-12-12-1 0.992 0.975 0.979 0.985 0.8101 -0.891 -1.694 -0.385 -0.048 1.225 1.5152 -2.139
15 ST-7-12-12-1 0.996 0.98 0.97 0.987 1.186 -1.348 -1.883 0.7109 -0.876 -1.845 1.3778 0.9419
16 TT-7-11-11-1 0.994 0.978 0.986 0.989 -0.56 -0.208 0.2223 -2.232 -1.075 1.211 -1.847 1.7009
lw{3,2}
5.5. OPTIMIZE ANN MODEL
1.6723 0.6711 1.0028 0.0934 0.094 0.438 0.2436 0.6218
ANNs exhibiting higher value of r are considered  b{1,1}=Bias values for first hidden layer
to be the most optimal. Form Table 4, for the case  b{2,1}=Bias values for second hidden layer
of RC beams without stirrups, ST-5-8-8-1  b{3,1}=Bias values for output layer
(employing a combination of input parameters: b,  iw{1,1}=Weights values for Input Layers
d, av/d, fc, ρl) is the most optimum NN model. No  lw{2,1}=Weights values for first hidden layer
 lw{3,2}=Weights values second hidden layer
significant effect is observed on NN-performance
when employing only single hidden layer. The
importance of the database in defining the
architecture of NNs becomes evident when
comparing NN ST-5-7-1 to other NNs developed
in the past [13-24]. Form this comparison it is
evident that the use of different databases can lead
to differences in the structure (architecture) of the
NN which however, does not necessarily result in
more accurate predictions. Changes in the
architecture of the NN may also occur when
extending the database upon which the training
process of the NNs is based. For the case of the
RC beams with stirrups, ST-7-12-12-1 (employing
a combination of input parameters: b, d, av/d, fc, ρl,
ρw, fy) is the most optimum ANN model. It should Figure 14: Weights and Bias values for ANN without
also be also pointed out that in this case the use of stirrups

314
The values inside the {,} represent their matrix experimentally. The predictions of the design
positions as shown in Fig.14. The latter figure codes concerning shear capacity appear to
provides a representation of the weights and bias generally underestimate – often considerably – the
values in the form of matrix. The size of the experimentally established values.
matrix depends on the number of input neurons, Table 7 provides statistical information concerning
number of hidden layers, the number of neuron in the predictions of shear capacity obtained
the hidden layers and the number of neuron in experimentally, through the use of the trained NNs
output layer. The green sign shows the positive and the various assessment methods presently
value and red sign shows the negative value. The considered. Similar information is provided in
size of rectangle is linked to the magnitude of the Table 8 for the ratio of the experimentally
associated values. established shear capacity VTest and its predicted
Table 5b: Weights and Bias Values for beams with counterparts VPred obtained either from the ANN
Stirrups or the design codes. The mean value provides an
b{1, b{2, b{3,
iw{1,1} 1} 1} 1} indication concerning how safe or unsafe the
-1.2 -2.6 2.7 -3.2 -3.2 -3.4 6.6 8.19 -5.6 1.4
-5.1 4.8 -5.1 0.8 4.0 -2.2 2.4 -1.3 1.2 predictions are whereas the Coefficient of
-8.8
5.0 0.6
0.7
3.1
-0.4
-6.2
-2.6
-4.7
1.1
-0.5
6.4
5.4
4.3
-2.7
0.2
6.6
-3.2
Variance (CV) and Standard Deviation (St. Dev.)
7.1 0.4 1.6 -1.4 0.4 5.1 3.5 -8.8 0.5 provides information concerning the accuracy of
-3.6
2.9 -2.9
1.3
6.1
7.1
-4.7
4.9
-3.8
2.8
-0.7
-5.7
-4.9
-2.6
3.5
-3.6
-0.3
2.4
the predictions and the scatter.
4.3 4.8 5.2 -0.8 -4.7 -0.4 -5.1 0.5 -0.1
5.1 -2.8 -1.9 -2.1 1.3 -6.9 2.2 5.0 1.8 Table 6. RC Design Code
0.1 6.6 1.1 5.3 -0.3 -5.0 3.4 -6.5 4.6 Code Formula Limitations
0.9 0.1 -0.6 1.0 5.7 3.1 6.8 -5.8 -6.5 𝑉𝑛 =𝑉𝑐 +𝑉𝑠 𝑉𝑐 ≤ 0.29√𝑓𝑐 𝑏𝑑
318-11

𝐴𝑠 𝑉
ACI-

-0.1 -3.6 -3.0 -4.8 3.8 -0.8 -6.8 10.3 -0.1


𝑉𝑐 = (0.16(𝑓c)1/2 + 17 ) 𝑏𝑑
lw{2,1}
𝑀𝑏
0.1 1.2 -1.4 2.1 0.7 -0.3 2.8 -3.3 -0.8 3.1 -2.4 0.8 𝐴𝑠𝑤 𝑓𝑦𝑤𝑑
𝑉𝑠, =
0.9 -1.8 0.0 -2.0 1.7 -2.7 -3.2 -0.8 -1.4 3.5 -0.5 -3.8 𝑠
-3.6 0.8 -4.0 -4.4 0.3 -3.7 -6.7 2.9 -3.5 1.2 2.3 -0.8 𝑉𝑛 =𝑉𝑅𝑑,𝑐 +𝑉𝑅𝑑,𝑠 𝑉𝑅𝑑,c
1 3
3.2 -2.3 0.8 0.7 1.5 -0.4 -0.9 1.5 -0.8 -2.3 -2.8 2.5
VRd,c = α1 0.18 k(100ρl fck ) ⁄3 bwd ≥ 0.035√𝑓𝑐𝑘 k ⁄2 𝑏𝑑
-3.6 -2.8 1.5 0.3 0.9 -1.3 -0.4 2.7 -0.6 2.0 -3.1 0.7
-3.7 0.7 1.8 0.9 -1.4 3.1 -2.7 0.9 -0.4 1.4 1.4 -2.7
For av/d> 2, α1 = 1
-0.4 -1.7 1.6 0.5 -3.0 -1.3 0.3 -3.8 2.1 -3.1 1.9 0.2 For av/d<2α1 = v
2α 𝐴𝑠
𝜌𝑙 = < 0.02
EC2

d
0.1 -4.1 0.6 -1.1 -2.3 2.7 -1.4 2.9 2.2 1.8 0.8 -1.3
Av fyv (0.9d)Cotθ 𝑏𝑑
-1.8 0.4 -2.6 -2.9 -0.6 1.0 3.1 2.6 -2.3 -0.3 -2.1 -2.1 VRd,s =
3.2 0.1 -2.7 -0.7 -3.2 2.3 1.3 -0.0 -1.0 -2.4 1.3 -2.5 s
-0.4 -2.1 -3.0 3.4 0.7 1.7 -0.0 2.1 1.0 0.1 -2.4 3.5
200
-2.8 2.1 -1.3 -3.3 1.9 -3.5 2.6 0.7 -1.3 -2.9 1.2 0.7 𝑘 = (1 + √ )<2
lw{3,2}
𝑑
0.2 -1.7 2.7 0.1 -0.2 -0.1 -0.3 01 -0.1 -1.0 -0.1 2.2
Vn =Vc +Vs
1
Vc = 0.9βd βp βn (fc ′) ⁄3 bwd
6. COMPARATIVE STUDY where βd = (
39.4
)
1⁄
4
≤ 1.5 ,
JSCE

d
1⁄
A comparative study between the predicted values βp = (100ρl ) 3 ≤ 1.5
0≤ βn ≤2.0
of shear capacity obtained from the NN, the design Av fyv (0.87d)
codes and their experimentally established Vs =
s
Note: Factors in all design codes such as load factors, strength reduction factors and
counterparts (obtained from the databases) is material reduction factors are not considered in this investigation because all these
presented in Figs 15(a&b) for the case of RC factors are associated with the design philosophy. Similarly application on the limit on
concrete compressive strength is also not applicable in this investigation.
beams without and with stirrups respectively. The
formulae adopted by the RC design codes (i.e. From the latter tables both the NN and the CFP
ACI, EC2 and JSCE) [1-3] for predicting shear method appear to be capable of providing
capacity are summarized in Table 6. predictions at least comparable to their
counterparts established experimentally and
The predictions provided by the NN correlate through the use of the design codes. Furthermore,
closest with the experimental data as the relevant it is worth mentioning that in spite being based on
points in Figs 16(a&b) are close to the diagonal assumptions incompatible with the physical
line of the graph. The predictions obtained from models adopted by the codes, the CFP method is
the CFP method, especially for the case of the RC capable of providing predictions of shear capacity
beams with stirrups, exhibit good agreement with closer to the test data compared to those of the
the test data. However for some specimens the codes. The same applies, in the case of the NNs
CFP method values overestimates or which although are based on heuristic approaches
underestimates the shear-capacity determined

315
which do not strictly adhere to the principles of The graphs, presented in Figs 16a&b, present the
theoretical mechanics, they appear to be capable normal (Gaussian) distribution of the ratios
of providing predictions which provide a closer fit VTest/VPred calculated for each RC beam specimen.
to the available test data compared to the For the case without stirrups, the predictions
predictions of the available design codes. obtained from the NNs (expressed as VTest/VNN)
exhibit the best Gaussian distribution characterised
by the lowest ST. DEV (see Table 7). Among the
RC design codes the predictions of EC2
(VTest/VEC2) are the closest to those obtained by the
NN (VTest/VNN). The predictions obtained from the
other design methods, including CFP, are
characterised by larger scatter. For the case of RC
beams with stirrups the predictions of the NN
(VTest/VNN) once again exhibit the best Gaussian
distribution, however in this case it is the CFP
method (instead of EC2) that provides the closest
predictions (VTest/VCFP)to those of NN (VTest/VNN).
(a)
The predictions obtained from the other design
methods are generally more conservative but at the
same time are characterised by considerably more
scatter.

(b)
Figure 15: Comparison of experimentally established
and predicted values of shear-capacity for
the case of the RC beams (a) without and
(b) with stirrups.
(a)
Table 7: Statistical analysis concerning Vu
Beams without Stirrups
(KN) Min Max Mean St.Dev. CV
Test 7.00 281.00 65.24 46.55 0.71
NN 12.00 251.50 62.37 43.03 0.69
CFP 5.73 376.87 75.25 62.52 0.83
ACI 3.64 288.28 52.44 44.33 0.85
EC2 4.89 266.92 65.13 47.63 0.73
JSCE 5.19 304.09 66.04 47.95 0.73
Beams with Stirrups
(KN) Min Max Mean St.Dev. CV
Test 7.00 876.00 273.82 216.09 0.79
NN 42.40 826.60 274.92 216.44 0.79
CFP 11.12 944.92 307.62 212.69 0.69
ACI 34.69 1060.45 199.64 153.87 0.77 (b)
EC2-45 33.49 1065.37 205.76 148.18 0.72 Figure 16: Normal distribution of VTest/VPred for the
EC2 68.12 2110.06 329.00 259.61 0.79 case of the RC beams (a) without and (b)
JSCE 39.52 944.36 191.09 140.28 0.73 with stirrups

316
Table 8: Statistical analysis concerning VTest/ VPred - In spite being based on assumptions
Beams without web reinforcement incompatible with the physical models adopted
Min Max Mean St.Dev CV% by the current RC design codes, the CFP
Vtest/VNN 0.58 1.53 1.03 0.15 0.14 method is capable of providing predictions of
Vtest /VCFP 0.42 2.17 0.94 0.22 0.23 shear capacity closer to the test data compared
Vtest /VACI 0.47 2.77 1.35 0.33 0.25 to those of the codes.
Vtest /VEC2 0.52 1.85 1.02 0.19 0.18
Vtest /VJSCE 0.33 2.23 1.02 0.27 0.26 - NNs, although being based on heuristic
Beams with web reinforcement approaches which do not strictly adhere to the
Vtest/VNN 0.13 2.26 1.01 0.23 0.23 principles of theoretical mechanics, appear to
Vtest /VCFP 0.23 2.32 0.87 0.31 0.36 be capable of providing predictions which
Vtest /VACI 0.17 4.36 1.43 0.92 0.64 provide a closer fit to the available test data
Vtest /VEC2-45 0.21 4.21 1.36 0.91 0.67 compared to those of the available design codes
Vtest /VEC2 0.10 2.66 0.84 0.52 0.62 and the CFP method.
Vtest /VJSCE 0.16 5.64 1.56 1.19 0.76
- The predictions obtained by the codes appear to
Table 8 presents the results obtained from the generally underestimate shear capacity and are
statistical analysis with respect to the ratio of characterized by considerable scatter.
VTest/VPred, based on this analysis the predictions of
the ANN essentially coincide with the - The latter observations raises further
experimentally established values (VTest) as the uncertainties concerning the ability of physical
mean value of VTest/VNN is 1.03 and 1.01 (≈1) when models adopted by the RC design codes to form
considering the beam specimens without and with the basis for the development of more advanced
stirrups respectively. For the case of the beams assessment methods that can realistically
without stirrups the mean values of VTest/VEC2 and predict RC structural response under more
VTest/VJSCE are 1.02 whereas the mean value of complex loading scenarios associated with high
VTest/VCFP is 0.94 which is 9.10 % less than the temperatures, fire, impact and blast or even
mean values of VTest/VNN. It should be noted that in environmental actions. This highlights the
terms of standard deviation the predictions urgent need to re-assess current codes for RC
obtained from the NN are associated with the design.
lowest value and the prediction of the ACI and - It should be noted that the present study is
CFP with the biggest. For the case of the RC currently being extended to include additional
beams with stirrups the mean value of VTest/VCFP test data in order to further validate the
and VTest/VEC2 are 0.87 and 0.84, which are the conclusion mentioned above. Furthermore, the
closest value of VTest/VPred amongst the other various conclusions are also being corroborated
design methods to the mean value of VTest/VNN. through the use on NLFEA.
Furthermore, it is important to notice that standard
deviation characterising the predictions of the NN REFERENCES
is the lowest followed by that associated with the 1. Building Code Requirements for Structural Concrete (ACI
predictions of the CFP. The standard deviation 318-08) and Commentary. 2008: American Concrete
characterising the predictions of the other codes is Institute 38800 Country Club Drive Farmington Hills, MI
48331 U.S.A.
considerable higher. 2. EN 1992-1-1 (2004) (English): Eurocode 2: Design of
concrete structures - Part 1-1: General rules and rules for
7. CONCLUSIONS buildings. 2004: Avenue Marnix 17, B-1000 Brussels.
The objective of the work described in the present 3. JSCE Guidelines for Concrete No.15-Standard
Specifications For Concrete Structures - 2007. 2007:
study is to compare the predictions concerning
Yotsuya 1-chome, Shinjuku-ku, Tokyo 160-0004,
shear capacity and mode of failure established (i) JAPAN.
experimentally, (ii) through the use of the CFP 4. Martin, B.T., Sanders, D.H., Verification and
method (iii) the available design codes and (iv) by Implementation of Strut-and-Tie Model in LRFD Bridge
employing ANNs. Design Specifications. 2007: 301 Manchester Road,
Poughkeepsie, NY 12603.
The most important conclusions derived from this 5. Williams, C., Deschenes, D., Bayrak, O., Strut-and-Tie
comparative study are: Model Design Examples for Bridges: Final Report. 2012:

317
Center for Transportation ResearchThe University of 20. Oreta, A.W.C., Simulating size effect on shear strength of
Texas at Austin1616 Guadalupe, Suite 4.202Austin, TX RC beams without stirrups using neural networks.
78701. Engineering Structures, 2004. 26(5): p.681-691.
6. Li, B., Ngoc Tran, C.T., Reinforced concrete beam 21. Afrifa, R.O., Asamoah, M.A., Ansah, E.O., Artificial
analysis supplementing concrete contribution in truss Neural Network Model for Low Strength RC Beam Shear
models. Engineering Structures, 2008. 30(11): p.3285- Capacity. Journal of Science and Technology (Ghana),
3294. 2012. 32(2): p.119-132.
7. Kotsovos, M.D., Compressive Force-Path Method-Unified 22. Shuraim, A.B., A novel approach for evaluating the
Ultimate Limit-State Design of Concrete Structures. concrete shear strength in reinforced concrete beams.
Springer Cham Heidelberg New York Dordrecht London. Latin American Journal of Solids and Structures, 2014.11:
8. Panagiotakos, T.B., Fardis, M.N., Deformations of p 093-112.
Reinforced Concrete Members at Yieldingand Ultimate. 23. Reddy, S.L., Rao, R.N., Rao, T.D., Shear Resistance of
ACI Structural Journal, 2001. 98(2): p.135-148.. High Strength Concrete Beams Without Shear
9. Reineck, K.H., Kuchma, D.A., Kim, K.S., Marx, S., reinforcement.International Journal Of Civil And
Shear Database for Reinforced Concrete Members Structural Engineering, 2010. 1(1) :p.101-11.
withoutShear Reinforcement. ACI Structural Journal, 24. Wei, W.W., Gong, J.X., Che, Y., Shear strength
2003. 100(2): p.240-249. prediction for reinforced concrete beams without stirrups.
10. Kotsovos, M.D., Pavlovic, M.N. "Structural concrete: Magazine of Concrete Research, 2011. 63(6): p. 433-440.
finite-element analysis for limit-state design." (1995). 25. Kotsovos, G.M., Kotsovos, M.D., Criteria for structural
11. Kotsovos, M.D., Spiliopoulos, K.V., Modeling of crack failure of RC beams without transverse reinforcement.
closure for finite-element analysis of structural concrete, The Structural Engineer,, 2008. 86: p. 55-56
Computers & Structures, 1998.69: p.383–398. 26. Svozil, D., KvasniEk,V., Pospichal, J., Introduction to
12. Abdalla, J.A., Elsanosi, A., Abdelwahab, A., Modeling multi-layer feed-forward neural networks. Chemometrics
and simulation of shear resistance of R/C beams using and Intelligent Laboratory Systems, 1997. 39: p. 43-62.
artificial neural network. Journal of the Franklin Institute, 27. Taner,M.T., Neural Networks And Computation Of Neural
2007. 344(5): p.741-756. Network Weights And Biases By The Generalized Delta
13. Amani, J., Moeini, R., Prediction of shear strength of Rule And Back-Propagation Of Errors. Rock Solid
reinforced concrete beams using adaptive neuro-fuzzy Images,” Houston, Texas 1995: p 1-11.
inference system and artificial neural network. Scientia 28. Moreira, M., Fiesler, E., Neural Networks with Adaptive
Iranica, 2012. 19(2): p.242-248. Learning Rate and Momentum Terms. 1995, Idiap: Nstitut
14. Cladera, A., Marí, A.R., Shear design procedure for Dalle Molle D’intelligence Artificielle Perceptive Case
reinforced normal and high-strength concrete beams Postale 609 - 1920 Martigny - Valais - Suisse.
using artificial neural networks. Part I: beams without 29. LeCun, Y. A., Bottou, L., Orr, G. B., & Müller, K. R.
stirrups. Engineering Structures, 2004. 26(7): p. 917-926. Efficient backprop. In Neural networks: Tricks of the
15. Cladera, A., Marí, A.R., Shear design procedure for trade. Springer Berlin Heidelberg, 2012: pp. 9-48.
reinforced normal and high-strength concrete beams 30. Wilson, D.R., Martinez, T.R., The general inefficiency of
using artificial neural networks. Part II: beams with batch training for gradient descent learning. Neural
stirrups. Engineering Structures, 2004. 26(7): p. 927-936. Networks, 2003. 16(10): p. 1429-1451.
16. Jung, S., Kim, K.S., Knowledge-based prediction of shear 31. Moore, D.S., McCabe, G.P., Introduction to the practice
strength of concrete beams without shear reinforcement. of statistics. . 1989: The Registrar General’s Decennial
Engineering Structures, 2008. 30(6): p. 1515-1525. Supplement for England and Wales.
17. Yang, K.H., Ashour, A.F., Song, J.K., Shear Capacity of 32. Taylor, R., Interpretation of the correlation coefficient: a
Reinforced Concrete Beams Using Neural Network. basic review. Journal of diagnostic medical sonography ,
International Journal of Concrete Structures and Materials, 1990.6(1): p. 35-39.
2007. 1(1): p. 63-73. 33. Castillo, E., Berdinas, B.G., Romero, O.F., Betanzos,
18. Mansour, M.Y., Dicleli, M., Lee, J.Y., Zhang, J., a A.A., A Very Fast Learning Method for Neural Networks
Predicting the shear strength of reinforced concrete Based on Sensitivity Analysis. Journal ofMachine
beams using artificial neural networks. Engineering Learning Research, 2006. 7: p. 1159-1182.
Structures, 2004. 26(6): p. 781-799. 34. Beale, M.H., Hagan, M.T., Demuth,H.B., Neural Network
19. Iruansi, O., Guadagnini, M., Pilakoutas, K., Neocleous, Toolbox™-User's Guide, in R2015a. 2015: The
K., Predicting the Shear Strength of RC Beams without MathWorks, Inc.3 Apple Hill DriveNatick, MA 01760-
Stirrups Using Bayesian Neural Network, in 4th 2098.
International Workshop on Reliable Engineering
Computing (REC 2010). 2010.

318

You might also like