You are on page 1of 10

Construction and Building Materials 98 (2015) 276–285

Contents lists available at ScienceDirect

Construction and Building Materials


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

Adaptive neuro fuzzy prediction of deflection and cracking behavior


of NSM strengthened RC beams
Kh Mahfuz ud Darain a,b, Sahaboddin Shamshirband c, Mohd Zamin Jumaat a,⇑, M. Obaydullah a
a
Department of Civil Engineering, Faculty of Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia
b
Architecture Discipline, Science, Engineering and Technology School, Khulna University, 9208 Khulna, Bangladesh
c
Department of Computer System and Technology, Faculty of Computer Science and Information Technology, University of Malaya, 50603 Kuala Lumpur, Malaysia

h i g h l i g h t s

 Near Surface Mounted (NSM) strengthening for underperforming structures.


 A soft computing technique is employed for the serviceability prediction.
 Reinforced Concrete (RC) beams strengthened with the NSM technique.

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

Article history: The Near Surface Mounted (NSM) technique is a promising strengthening approach, though unpopular
Received 23 April 2015 due to reliable serviceability prediction options. In this paper, a soft computing technique is employed
Received in revised form 6 August 2015 to predict the deflection and cracking behavior of NSM strengthened RC beams. The Adaptive Neuro-
Accepted 10 August 2015
Fuzzy Inference System (ANFIS) was used to simulate the serviceability behavior of one control and six
NSM strengthened RC beams using variable bond lengths (1600, 1800 or 1900 mm) with different
NSM strengthening materials (steel and CFRP bars). The proposed ANFIS model employed variable load
Keywords:
and bond length as inputs while the output parameters were deflection and crack width of the steel
ANFIS
Serviceability
and CFRP bars. The ANFIS results were compared with the experimental results using the root-mean-
Structural strengthening square error (RMSE), coefficient of determination and Pearson coefficient. The results found that the
Deflection ANFIS approach showed an improvement in predictive accuracy and generalization capability in compar-
Crack width ison to the fuzzy approach. The highest level of accuracy with ANFIS was achieved when predicting CFRP
Soft computing bar crack width, where RMSE = 8.05E07. In contrast, the lowest level of accuracy was achieved when
predicting CFRP bar deflection, with RMSE = 0.045185.
Ó 2015 Published by Elsevier Ltd.

1. Introduction the moment curvature approach [11,12] to the more recently partial
interaction based moment rotation approach [13,14]. However,
NSM technique is a promising strengthening approach for these models cannot be directly applied to the NSM technique as
extending the design service life of underperforming structures its structural mechanics are quite unlike the EBR approach. When
[1–3]. This technique demonstrates better performance in enhanc- NSM reinforcement is inserted in the RC beam, a complex system
ing flexural and shear capacity as well as delaying the premature is developed due to the tension stiffening effect of the FRP–concrete
debonding problem of RC beams [4,5]. However, appropriate ser- interface in addition to the internal steel reinforcement-concrete
viceability (deflection, crack width, vibration of structure, etc.) pre- stiffening effect. As a result, the deflection behavior, crack width,
diction model is a major lack of this technique [6]. The researchers crack spacing and other serviceability parameters vary considerably
are updating previously solved Externally Bonded Reinforcement compared to the EBR strengthened beams.
(EBR) load–deflection model for making applicable for NSM tech- The promising applicability of the NSM system in real applica-
nique using the simple linear approach to the tri-linear response tion demands a prompt and straightforward, yet consistent and
(pre-cracking, cracking and post yielding) [7–10]. They also precise alternative system to predict the serviceability perfor-
explained the serviceability behavior of EBR from the view point of mance of the strengthened structures. The soft computing tech-
niques can be a better choice as they can efficiently and
⇑ Corresponding author. promptly predict any complicated problem if the operational data
E-mail address: zamin@um.edu.my (M.Z. Jumaat). are available. The neuro-fuzzy network or Adaptive Neuro-Fuzzy

http://dx.doi.org/10.1016/j.conbuildmat.2015.08.096
0950-0618/Ó 2015 Published by Elsevier Ltd.
K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285 277

Inference System (ANFIS) [15] was used by researchers for model- specimen. Six RC specimens were strengthened using NSM Steel bars (3 in number)
and CFRP bars (3 in numbers) with variable bond length (1600 mm, 1800 mm and
ing [16,17], predictions [18–20] and control in various engineering
1900 mm). Consequently, a total of seven RC rectangular beams were considered in
systems [21,22]. It is a hybrid intelligent system that merges the this study with two testing variables. These variable parameters were the bonded
technique of the learning power of the artificial neural network length of NSM reinforcement, and the type of NSM reinforcement. The testing
(ANN) with the knowledge representation of fuzzy logic. matrix is shown in Table 1.
The concept of using artificial intelligence in structural strength-
ening is quite new. In this particular area of research, ANN has been 2.1.2. Specimens and materials
used more than the other artificial intelligence techniques [23], Rectangular 2.3 m long RC beams with 125-mm  250 mm cross-sectional
dimension were selected for the experimental study. The beams were designed as
such as fuzzy inference system (FIS), ANFIS and Genetic Algorithms
under reinforced beams (q = As/bd = 0.0085) with two 12 mm diameter deformed
(GA), etc. [24–27]. Flood et al. [24] studied the possibility of predict- bars as tensile reinforcement. Two 10 mm diameter deformed bars were used as
ing the externally reinforced RC beam’s deflection using the ANN hanger bars to hold the stirrups. Double-legged closed 8 mm diameter steel stirrups
model. Several researchers used the ANN model to predict the shear were used as shear reinforcement with spacing of 90 mm center to center to ensure
strength of the RC beam using FRP material [25] and steel plates that the flexure failure would prevail. Fig. 1 exhibits the features of the beam
arrangement where the dimensions were given in millimeter scale.
[26], and suggested some modification of the design equations.
Crushed stone (20 mm diameter) and natural river sand were used as coarse and fine
Nasrollahzadeh and Basiri [28] used FIS to predict the shear aggregate, respectively, for concrete casting. The compressive and flexural strength of
strength of 197 RC beams from the work of different researchers, hardened concrete were measured at 28 days based on the cube (100 mm 
which were reinforced with FRP. Darain et al. [29] demonstrated 100 mm  100 mm), cylinder (200 mm  100 mm diameter) and prism specimens
(500 mm  100 mm  100 mm). The compressive and flexural strengths of the con-
a simplified serviceability model using fuzzy logic to predict the
crete were determined according to BS EN (12390-3), ASTM (C39/C39M-14) and BS
deflection and crack width of NSM strengthened RC beams. ANFIS EN (12390-5) [33–35]. The material properties of concrete are described in Table 2.
is an improvement over ANN and FIS although it was rarely imple- The supplied manufacturer’s mechanical properties of deformed steel bars were
mented in this area for analysis and prediction of the strengthening checked in the laboratory to confirm whether they conform to the ASTM A615 [36]
efficacy of any structural component. specification. LaMaCo System Sdn Bhd. supplied the 12 mm diameter sand coated
carbon-epoxy pultruded FRP (CFRP) bars with a density of 1.65 g/mm2 for NSM
Park et al. [30] predicted the behavior of a FRP retrofitted circu-
strengthening. These CFRP bars showed linear elastic response up to ultimate fail-
lar column using the ANFIS model. Seven parameters of specimens ure. To adhere the NSM steel or CFRP bar with the concrete substrate, SikadurÒ 30
and FRP configuration were selected to predict the strength, strain, was used as the epoxy adhesive. According to the manufacturer’s results, the den-
and stiffness of the post-yielding modulus, which finally exhibited sity was 1.65 kg/l at 23 °C. The adhesive is composed of two components, a resin
and a hardener, which were blended together in a ratio of 3:1 until an even gray
good agreement with the constitutive model of other researchers.
color was attained. The details of the concrete, steel and CFRP material properties
Balasubramaniam et al. [31] studied different levels of corrosion are described in Table 2 and [37] SikadurÒ 30 properties are mentioned in Table 3.
damaged RC beams, which were strengthened with Glass FRP. Each of the strengthened beam specimens had a single groove
The ANFIS model was employed to assess the performance charac- (24 mm  24 mm) made using a diamond bladed concrete saw along the beam
teristics of those beams and the output of the model confirmed the length to accommodate the CFRP bar. A hammer and hand chisel were used to
remove the remaining concrete lugs of the groove, resulting in a rough surface
experimental results of the study. Jalal et al. [32] estimated the
inside of the groove that improves bonding. All the debris was then removed from
ultimate strength of a CFRP confined concrete cylinder from the groove using airbrushing pressure. An epoxy adhesive (SikadurÒ 30) was
reviewing several experimental studies and compared the perfor- applied into the groove to fill around 2/3 of the groove depth. The CFRP bar was
mance using ANN, multiple regression (MR), genetic programming then gently inserted into the groove, and pressed lightly to ensure proper epoxy
covering surrounding the bar. The groove’s outer surface was then levelled and left
(GP), ANFIS and several empirical models. The comparison demon-
for one week to achieve proper epoxy strength. The beams were then tested under
strated that the mean absolute percent errors (MAPEs) were lowest static loading using the four-point loading test.
(1.85%) for the ANFIS model amongst others. Suitable support condition was prepared where four point loading was applied
In the existing literature, the prediction model for the deflection using a 500 kN capacity closed-loop Instron universal testing machine (in Fig. 2).
and crack width of NSM strengthened RC beams is rare and almost Deflection measurement was performed with three vertical Linear Variable Differ-
ential Transducers (LVDT) which were placed at the midpoint, 250 mm away the
no work has been done on developing a prediction model of its ser-
center point (under the spreader beam load point) and 500 mm from the center
viceability behavior using ANFIS. To address this gap, the present of the RC beam. 5 mm strain gauges were affixed at the internal steel and NSM steel
study has an aim to develop an accurate yet rapid prediction model or CFRP bar to measure the strain during test. To measure concrete compressive
for the deflection and crack width of NSM strengthened RC beams strain, 30 mm strain gauges were fixed at the top of the concrete surface. Demec
discs were planted at the mid-span of the beams along the depth to measure trans-
using ANFIS. In this paper, two approaches were taken into consider-
verse strains. A Dino-Lite digital microscope was used to measure micro cracks on
ation. One control and six full sized NSM strengthened RC beams the surface of the beams.
were experimented and studied. Then, ANFIS was used to predict
the deflection and cracking behavior of NSM strengthened RC beams.
2.2. Adaptive neuro-fuzzy application

2. Material and methods 2.2.1. Neuro-fuzzy computing


Soft computing is an innovative approach in the construction of systems that
2.1. Experimental program are computationally intelligent, which possess humanlike expertise within a speci-
fic domain. These systems are supposed to adapt to changing environments, learn
2.1.1. Test specimens to do better and explain their decision-making process. It is usually more beneficial
In order to comprehend the proposed analysis approach, an experimental pro- to employ several computing methods in a synergistic way rather than building a
gram was carried out on NSM strengthened RC specimens. One RC beam was pre- system based exclusively on one technique. This is useful in confronting real-
pared without any strengthening material and was tested as the control world computing problems. The result of such synergistic use of computing tech-

Table 1
Test matrix of control and strengthened RC beam.

Sl. no. Notation Description Groove size (mm) Total bonded length of strengthened bars 2L (mm)
1 CB Control RC beam – –
2 N1.6F NSM strengthened RC beam with 12 mm CFRP bar 24 mm  24 mm 1600
3 N1.6S NSM strengthened RC beam with 12 mm steel bar 1600
4 N1.8F NSM strengthened RC beam with 12 mm CFRP bar 1800
5 N1.8S NSM strengthened RC beam with 12 mm steel bar 1800
6 N1.9F NSM strengthened RC beam with 12 mm CFRP bar 1900
7 N1.9S NSM strengthened RC beam with 12 mm steel bar 1900
278 K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285

P/2 P/2

900 500 900


2-10 mm 2-10 mm

250
8 mm @ 90 mm c/c.
NSM steel/CFRP bar
150 2-12 mm 2000 150
2300

24 mm
NSM steel
/CFRP bar

250 mm 24 mm Epoxy

125 mm

Fig. 1. Reinforcement and strengthening detail of RC beam.

Table 2
Properties of concrete, internal steel reinforcement, NSM steel and CFRP bar.

Material Compressive strength Flexure strength Yield stress Ultimate strength Elastic modulus Elongation
(MPa) (MPa) (MPa) (MPa) (GPa) (%)
Concrete 43.24 (cube) 5.01 – – 30.1 –
35.63 (cylinder)
RC beam steel – u 12 mm – – 400 480 200 27
NSM steel – u 12 mm – – 520 587 200 20
Steel – u 10 mm – – 529 578 200 21
Steel – u 8 mm – – 380 450 200 29
CFRP – u 12 mm – – – 2400 165 –

human knowledge and implement decision-making and differentiation. The combi-


Table 3 nation and integration of these two complementary methodologies produce a novel
Properties of SikadurÒ 30 [37]. discipline called neuro-fuzzy computing.
Strength 15 °C (MPa) 35 °C (MPa)
2.2.2. Adaptive Neuro-Fuzzy Inference System
Compressive strength 70–80 85–95
The ANFIS (Adaptive Neuro-Fuzzy Inference System) is a class of adaptive net-
Tensile strength 14–17 16–19
works functionally equivalent to the fuzzy inference systems. Simulations were run
Shear strength 24–27 26–31
in MATLAB toolbox and the results were observed on the corresponding output
blocks. In this study, the fuzzy inference system used has two inputs, x and y and
niques is the construction of complementary hybrid intelligent systems. The epit- one output f. In this study, the first-order Sugeno fuzzy model, with two fuzzy if-
ome of designing and constructing intelligent systems of this kind is neuro-fuzzy then rules was used as follows (Eqs. (1) and (2)):
computing: firstly, neural networks recognize patterns and adapt to cope with
evolving environments; and, secondly, fuzzy inference systems, which include Rule 1 : if x is A and y is C then f 1 ¼ p1 x þ q1 y þ r1 ð1Þ

Fig. 2. Instrumentation and testing arrangement of the prepared specimen.


K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285 279

Rule 2 : if x is B and y is D then f 2 ¼ p2 x þ q2 y þ r 2 ð2Þ Layer 3: Each node in the 3rd layer is a fixed node. The ith node calculates the
proportion of the firing strength of the ith rule to the sum of the firing strength of all
where parameters fpi ; qi ; r i g are referred to as premise parameters. In other words rules (Eq. (7)):
these parameters are the user defined parameters which needs to be optimized by
ANFIS training procedure. The reasoning mechanism for this Sugeno model is illus- wi
trated in Fig. 3(a); while Fig. 3(b) shows the corresponding equivalent ANFIS archi- O3;i ¼ wi ¼ ; i ¼ 1; 2: ð7Þ
w1 þ w2
tecture. Nodes of the same layer have similar functions. The output of the ith node in
layer l is denoted as Ol;i . For the sake of convenience, the outputs of this layer are called normalized fir-
Layer 1: Each node i in this first layer is an adaptive node with a node function ing strengths.
(Eqs. (3) and (4)) Layer 4: Each node I in this 4th layer is an adaptive node with a node function
(Eq. (8))
Ol;i ¼ lAi ðxÞ; for i ¼ 1; 2; or ð3Þ
O4;i ¼ wi f i ¼ wi ðpi x þ qi y þ ri Þ; ð8Þ
Ol;i ¼ lBi2 ðyÞ; for i ¼ 3; 4 ð4Þ
where wi is a normalized firing strength from the 3rd layer and fpi ; qi ; r i g is this
where x or y is the input variable to node i and Ai or Bi2 is an associated linguistic node’s parameters set. In this layer, variables are referred to as consequent parame-
label (such as ‘small’ or ‘large’). In other words, Ol;i is the membership grade of a ters. In other words, these parameters now present optimal values after ANFIS train-
fuzzy set A and B ð¼ A1 ; A2 ; B1 ; A2 Þ. It stipulates the extent to which the specified ing procedure.
input x or y satisfies the quantifier A. In this instance, the membership function Layer 5: In this 5th layer, the single node is a fixed node. The fixed node calcu-
can be any suitable parameterized membership function. The generalized bell func- lates the total output as the summation of all the signals that are incoming (Eq. (9)):
tion is used here (Eq. (5)) as it has the best abilities for the generalization of nonlin-
X P
ear parameters: wi f
O5;i ¼ wi f i ¼ Pi ð9Þ
i i wi
1
lA ðxÞ ¼  2bi ð5Þ
xci The parameters in the ANFIS architecture were identified by applying the
1þ ai
hybrid learning algorithms. In the forward pass of this algorithm, functional signals
go forward until Layer 4. Consequent parameters are identified by the least squares
where fai ; bi ; ci g are the parameters for the membership function. These membership estimate. In the backward pass, the error rates propagate backwards. Premise
parameters are updated during ANFIS training procedure. The bell-shaped function parameters are updated by the gradient descent.
varies accordingly as the values of the variables change, therefore manifesting differ-
ent types of membership functions for fuzzy set A. Variables in the first layer are
called premise variables.
Layer 2: Each node in the 2nd layer is a fixed node and its output is the conse- 2.2.3. Input variables
quence of all signals which are incoming (Eq. (6)): In this study, the load and bonded length were used for generating the ANFIS
models. The experimental data are given in Table 4. There are total 36 measure-
O2;i ¼ wi ¼ lAi ðxÞlBi ðyÞ; i ¼ 1; 2 ð6Þ ments. 70% data of each group was used for training and 30% for testing of the mod-
els since it was shown as the most suitable data division for training and testing. In
Every node output represents the firing strength of a rule. It other words, the this case all groups are included in the training of the models and in later in testing
layer determines the strength of the rules. as well. It is suitable to have more training than testing data.

Fig. 3. (a) A two-input first order Sugeno fuzzy model with two rules; (b) equivalent ANFIS architecture.
280 K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285

2.3. Evaluation of model performances remarkably. In Table 5, there is a comparison of the NSM strength-
ened RC beams of different bond lengths with the control beam. In
The following statistical indicators were selected for calculation of the ANFIS
accuracy.
this table, the first crack load, first crack deflection, ultimate crack
(1) Root-mean-square error (RMSE) and ultimate crack deflection are analyzed with the percentile dif-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ference. From the Table 5, it is visible that the 1st crack and ulti-
Pn 2
i¼1 ðP i  Oi Þ mate load of NSM strengthened beams were increased
RMSE ¼ ð10Þ
n considerably compared to the control specimen. At the first crack
(2) Pearson correlation coefficient (r) load, the deflection of the strengthened beam seemed to be more
P   Pn   Pn  than that of the control beam. This phenomenon was changed
n ni¼1 Oi  Pi  i¼1 Oi  i¼1 P i
r ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 P   P ð11Þ afterwards and the ultimate deflection was less than the control
 P  2  P 2 
n ni¼1 O2i  n
i¼1 Oi  n ni¼1 P2i  n
i¼1 P i specimen.
The failure pattern of the strengthened beams is depicted in
(3) Coefficient of determination (R2)
Figs. 4 and 5. From the figures, it is clearly visible that the CFRP
hP i2
n strengthened beams failed due to premature debonding failure.
i¼1 ðOi  Oi Þ  ðP i  P i Þ
2
R ¼P   P   ð12Þ However, the NSM steel beam overcomes the premature failure
n
Oi  Oi  ni¼1 P i  P i
i¼1 and their failure mode was concrete crushing after yielding the
where Oi = ANFIS value, Pi = measurement values, and n = the total number of test internal steel. The bond length of NSM reinforcement for the
data. While some controversy regarding the validity of this error metrics does exist, N1.8F and N1.8S beams was 1800 mm whereas 1900 mm bond
it pertains to comparison of forecasting methods across different data sets [38,39], length was selected for the N1.9F and N1.9S specimens. The bond
which falls beyond the scope of this article. efficiency was superior in the case of the steel bar as NSM rein-
forcement and the failure pattern confirmed the composite action
3. Results of the NSM reinforcement and RC beam.
Fig. 6 shows the load–deflection behavior of the control and
3.1. Experimental results NSM strengthened RC beam. As shown in the figure, beyond the
1st cracking of the concrete the load–deflection relationship is
Depending upon the type of strengthening material (CFRP and the same for both the specimens. After the elastic limit, the embed-
steel) and bond length (1600 mm, 1800 mm and 1900 mm), the ded NSM reinforcement contributed a stiffer response and showed
load–deflection behavior was influenced considerably. In all cases, a pronounced effect in enhancing the flexural response of the
the ultimate capacity of the strengthened RC beam was increased strengthened beam. The NSM steel reinforced beam demonstrated
a nearly bilinear response characteristic of the RC beam in under
reinforced conditions. With the exception of the N1.6S beam, the
Table 4
Experimental data.
other two steel strengthened beams failed in flexure with the for-
mation of a wider flexural crack at midspan, which further led to
Input data Steel bar output FRP bar output concrete crushing at the top of the midspan. The CFRP strength-
Load Length of NSM Deflection Crack Deflection Crack ened beams demonstrated a trilinear response with a sharp fall
bar width width after the failure of the beams. All the beams failed due to concrete
0 1600 0 0 0 0 cover separation. The slope of the load–deflection curve of the NSM
0 1800 0 0 0 0 steel strengthened beam was steeper compared to the NSM CFRP
0 1900 0 0 0 0
specimens, which also indicated the higher stiffness of these
10 1600 0.434 0 0 0
10 1800 0.761 0 0.349 0 beams.
10 1900 0.1 0 0.078 0 In Fig. 7, the load vs. crack width is demonstrated for the control
15 1600 0.762 0.042 0.2125 0.103 and strengthened beam. After strengthening, the crack width of all
15 1800 1.091 0.063 0.761 0.06 the strengthened beams was less than that of the control specimen.
15 1900 0.314 0 0.344 0.047
20 1600 1.089 0.088 0.631 0.16
It is evident that the bond length of the NSM reinforcing bar signif-
20 1800 1.518 0.106 1.305 0.136 icantly influences the crack width of the tested specimens. The
20 1900 0.496 0 0.795 0.079 beam with a bond length of 1900 mm showed less crack width
30 1600 1.682 0.15 1.759 0.206 compared to the specimens with other bond lengths (1600 mm
30 1800 2.44 0.215 2.585 0.213
and 1800 mm). Interestingly, the steel NSM bar performed better
30 1900 1.173 0.11 1.838 0.2
40 1600 2.314 0.24 2.897 0.315 compared to the CFRP strengthening bar. Even the N1.9S showed
40 1800 3.434 0.231 3.773 0.304 the minimum deflection (8.5 mm) and the least crack width
40 1900 2.062 0.115 2.91 0.242 (0.3 mm) compared to the N1.9F beam, which had a deflection of
50 1600 3.003 0.282 5.7875 0.385 9.88 mm with 0.44 mm crack width across the same load point
50 1800 4.409 0.247 4.996 0.383
50 1900 2.946 0.192 4.055 0.268
(100 kN).
60 1600 4.179 0.313 7.7125 0.45
60 1800 5.393 0.296 6.126 0.448
60 1900 3.81 0.209 5.175 0.312 3.2. ANFIS model analysis
70 1600 5.389 0.472 10.1375 0.493
70 1800 6.541 0.351 7.634 0.488
70 1900 4.635 0.245 6.349 0.354 Initially, the ANFIS network was trained with measured data by
80 1600 6.859 0.547 12 0.554 the previously presented experimental procedures. Three bell-
80 1800 7.857 0.567 8.701 0.542 shaped membership functions were used to fuzzify the ANFIS
80 1900 5.735 0.267 7.498 0.397 inputs. After the training process, the ANFIS networks were tested
90 1600 8.44 0.61 14.1375 0.57
90 1800 9.288 0.637 10.112 0.56
to determine the deflection and crack width of the steel and CFRP
90 1900 7.036 0.286 8.614 0.421 bars. According to the experiments, the input parameters were
100 1600 0 0 16.925 0.597 applied load (F) and bonded length (L). The final outputs – deflec-
100 1800 0 0 11.368 0.578 tion for steel bar strengthened beams (DS), crack width for steel
100 1900 8.563 0.304 9.879 0.439
bar strengthened beams (WS), deflection for CFRP bar strength-
K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285 281

Table 5
Summary of experimental test results.

Beam ID Pcr (kN) % Pcr Dcr (mm) % Dcr Pu (kN) % Pu Du (mm) % Du Failure mode
CB 10.6 – 0.3 64.4 – 24.7 FF
N1.9F 15.0 41.5 0.3 0 133.2 106.8 19.2 22 CCS
N1.9S 23.0 117.0 0.7 133 103.8 61.2 12.4 50 FF
N1.8F 14.0 32.1 0.6 100 130.8 103.1 18.6 25 CCS
N1.8S 15.5 46.2 1.0 233 99.6 54.7 16.8 32 FF
N1.6F 17.5 65.1 0.4 33 109.5 70.0 13.6 45 CCS
N1.6S 14.4 35.5 0.7 133 94.1 46.1 9.4 62 CCS

Where Pcr = first crack load; % Pcr = percent increase in first crack load over the control beam; Pu = ultimate load; % Pu = percent increase in ultimate load over the control
beam; Dcr = deflection at 1st crack; Du = mid-span deflection at ultimate load, % Dcr = percent increase of deflection at 1st crack compare to control beam; % Du = percent
increase of deflection at ultimate load compare to control beam; FF = flexural failure (concrete crushing after steel yielding), CCS = concrete cover separation.

Fig. 4. Failure behavior of (a) N1.8F and (b) N1.8S beams.

Fig. 5. Failure behavior of (a) N1.9F and (b) N1.9S beams.

ened beams (DF) and crack width for CFRP bar strengthened beams Consequently, it follows that the prediction results are in very good
(WF) of ANFIS – were verified using MATLAB ANFIS Toolbox. agreement with the measured values for the ANFIS method. This
Fig. 8 shows the ANFIS decision surfaces for steel bar deflection, observation can be confirmed with very high value for the coeffi-
FRP bar deflection, steel bar crack width and FRP bar crack width, cient of determination. The number of either overestimated or
respectively. Fig. 8(a) shows that the maximum steel bar deflection underestimated values produced is limited. Consequently, it is
occurs for a load of 70 kN and for the 1700 mm long NSM bar. In obvious that the predicted values enjoy high level precision.
contrast, Fig. 8(b) shows that the maximum FRP bar deflection The performance of the ANFIS model in estimating the deflec-
occurs for the load of 100 kN and for the 1600 mm long NSM bar. tion and crack width were evaluated according to the statistical
‘‘Fig. 8(c) and (d) show that the crack width of steel and CFRP criteria, such as RMSE, coefficient of determination R2 and Pearson
bar is getting bigger with the increment of applied load. Almost coefficient r.
the same trend is noticed for widening the crack width with In order to demonstrate the merits of the proposed ANFIS
increasing load for the variable bond length of NSM steel or CFRP approach on a more definite and tangible basis, the prediction
bar”. accuracy of the ANFIS model was compared with the prediction
The predicted values using the ANFIS model are shown in Fig. 9 accuracy of the fuzzy method [40], genetic programming (GP)
in the form of scatter plots. In this section, the performance of the [41] and artificial neural network (ANN) method [42], which were
steel and CFRP in terms of deflection and crack width in the ANFIS used as a benchmark. The efficiency of the ANFIS model over clas-
predictive model are reported. It can be seen that most of the sical methods is listed in Table 6. Conventional error statistical
points fall along the diagonal line for the ANFIS prediction models. indicators, RMSE, r and R2, were used for comparison. Table 6 sum-
282 K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285

Load (kN)

Deflection (mm)
Fig. 6. Load–deflection curve of the control and NSM strengthened RC beams.

Fig. 7. Load vs. crack width diagram of the control and NSM strengthened RC beams.

marizes the prediction accuracy results for the test datasets since results obtained with the benchmark models. The main parameters
the training error is not a credible indicator for the prediction used for ANN model are:
potential of a particular model. The ANFIS model outperformed
the Fuzzy models according to the results in Table 6. The ANFIS  first layer: 2 nodes (5 inputs),
model provides significantly better results than the benchmark  hidden layers: 3, 6, 10 (three nodes in the second layer, 6 nodes
models. The one reason for the higher ANFIS accuracy could be in the third layer and 10 nodes in the fourth layer),
related to the adaptability of the ANFIS network since ANFIS found  output layer: 1 node (1 output),
optimal values for the membership functions. Another reason  number of iteration: 1000,
could be in respect of the membership function type, since, in  activation function: continuous log-sigmoid function.
the ANFIS network, bell-shaped membership functions are used,
and, in fuzzy logic, triangular membership functions are used. In The following main parameters employed in GP modeling are
On the basis of the comparison of the RMSE analysis with Fuzzy, selected by trial and error procedure:
it may be concluded that the proposed ANFIS outperformed the
K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285 283

Steel Steel
bar bar
defle crack
ction width
(mm) (mm)

Length of NSM Load (kN) Length of NSM Load (kN)


steel bar (mm) steel bar (mm)

(a) (c)

CFRP CFRP
bar bar
deflec crack
tion width
(mm) (mm)

Length of NSM Load (kN) Length of NSM


CFRP bar (mm) Load (kN)
CFRP bar (mm)

(b) (d)
Fig. 8. ANFIS decision surfaces for: (a) steel bar deflection, (b) CFRP bar deflection, (c) steel bar crack width, and (d) CFRP bar crack width.

 population size: 1024, bar showed better bonding capacity and ductile behavior,
 function set: þ; ; ; ; pffi; x2 ; ln; e x ; a x which, eventually, prevented the premature failure unlike
 head size: 5–9, the CFRP NSM material, even though the steel’s strength
 chromosomes: 20–40, capacity was lower than the CFRP NSM specimens. The crack
 number of genes: 3–4, width pattern of the steel NSM bars showed stiffer behavior
 mutation rate: 95.66, than that of the CFRP bars. Irrespective of the strengthening
 crossover rate: 34.44, materials (CFRP or steel), the 1900 mm bond length showed
 inversion rate: 110.33. better efficiency than the other lengths (1600 mm and
1800 mm).
4. Conclusion (2) The study indicated that the predicted deflection and crack
width closely conform to the experimental results, which
This paper investigated the feasibility of using the adaptive affirms the performance of the ANFIS model. This model
neuro-fuzzy technique to predict the serviceability behavior of demonstrated the superior predictive accuracy and capabil-
NSM strengthened RC beams. The variable load and bond length ity of generalization over the fuzzy logic approach and arti-
of NSM steel and CFRP bar were used as input parameters to ficial neural network (ANN) and genetic programming (GP)
observe the deflection and crack width of the tested RC beam. approaches as well as benchmark models. The highest level
The simulated output from ANFIS was compared with the experi- of accuracy with ANFIS was achieved for the CFRP bar crack
mental data and classical fuzzy technique. The conclusions are as width prediction RMSE = 8.05E07. In contrast, the lowest
follows: level of accuracy was achieved for the FRP bar deflection pre-
diction RMSE = 0.045185. The main advantages of the ANFIS
(1) The experimental test revealed that the NSM strengthened model are computationally efficient, very adaptable with
RC beam is superior to the unstrengthen control specimen optimization and adaptive techniques. The proposed ANFIS
in terms of strength and stiffness. The chosen test parame- model can be used for estimating the deflection and crack
ters (bond length and NSM strengthening material) signifi- width of NSM strengthened steel or CFRP material, which
cantly influenced the deflection pattern, cracking behavior will eventually save valuable time and the expense of rigor-
and failure mode of the tested specimens. The steel NSM ous relevant experimental tests.
284 K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285

Fig. 9. Scatter plots of predicted values using ANFIS models for (a) deflection prediction of steel/CFRP bar; (b) crack width prediction of CFRP/steel bar.

Table 6
Performance statistics of the ANFIS and fuzzy models of steel and FRP bar estimation based upon different statistical indicators.

ANFIS model Fuzzy inference system (FIS)


RMSE R2 r RMSE R2 r
Steel bar deflection 0.002249 1 1 0.0525 0.9831 0.9915
FRP bar deflection 0.045185 1 0.999997 0.0672 0.9962 0.9980
Steel bar crack width 5.8E07 1 1 0.0343 0.9945 0.9972
FRP bar crack width 8.05E07 1 1 0.0278 0.9922 0.9960
GP model ANN model
Steel bar deflection 0.0649 0.9763 0.9734 0.0785 0.9637 0.9615
FRP bar deflection 0.0785 0.9845 0.9832 0.0892 0.9767 0.9680
Steel bar crack width 0.0456 0.9842 0.9742 0.0647 0.9646 0.9672
FRP bar crack width 0.0453 0.9811 0.9797 0.0577 0.9627 0.9660

[2] L. Hollaway, A review of the present and future utilisation of FRP composites in
the civil infrastructure with reference to their important in-service properties,
Acknowledgement Constr. Build. Mater. 24 (12) (2010) 2419–2445.
[3] M.R. Coelho, J.M. Sena-Cruz, L.A. Neves, A review on the bond behavior of FRP
NSM systems in concrete, Constr. Build. Mater. 93 (2015) 1157–1169.
The authors gratefully acknowledge the support given by [4] L. De Lorenzis, J.G. Teng, Near-surface mounted FRP reinforcement: an
University of Malaya (UM), Malaysia for funding the study through emerging technique for strengthening structures, Compos. Part B Eng. 38 (2)
the High Impact Research Grant UM.C/HIR/MOHE/ENG/36 – (2007) 119–143.
[5] R. El-Hacha, S.H. Rizkalla, Near-surface-mounted fiber-reinforced polymer
‘‘Strengthening Structural Elements for Load and Fatigue” reinforcements for flexural strengthening of concrete structures, ACI Struct. J.
(D000036-16001). 101 (5) (2004) 717–726.
[6] M.M. Rahman, M.Z. Jumaat, M.A. Rahman, I.M. Qeshta, Innovative hybrid
bonding method for strengthening reinforced concrete beam in flexure, Constr.
References Build. Mater. 79 (2015) 370–378.
[7] M.T. El-Mihilmy, J.W. Tedesco, Deflection of reinforced concrete beams
[1] J. Teng, J. Chen, S.T. Smith, L. Lam, Behaviour and strength of FRP-strengthened strengthened with fiber-reinforced polymer (FRP) plates, ACI Struct. J. 97 (5)
RC structures: a state-of-the-art review, Proc. ICE Struct. Build. 156 (1) (2003) (2000).
51–62.
K.M.u. Darain et al. / Construction and Building Materials 98 (2015) 276–285 285

[8] H. Charkas, H.A. Rasheed, H. Melhem, Rigorous procedure for calculating [25] R. Perera, M. Barchín, A. Arteaga, A.D. Diego, Prediction of the ultimate strength
deflections of fiber-reinforced polymer-strengthened reinforced concrete of reinforced concrete beams FRP-strengthened in shear using neural
beams, ACI Struct. J. 100 (4) (2003). networks, Compos. B Eng. 41 (4) (2010) 287–298.
[9] W.-W. Wang, J.-G. Dai, K.A. Harries, Performance evaluation of RC beams [26] B.B. Adhikary, H. Mutsuyoshi, Artificial neural networks for the prediction of
strengthened with an externally bonded FRP system under simulated vehicle shear capacity of steel plate strengthened RC beams, Constr. Build. Mater. 18
loads, J. Bridge Eng. 18 (1) (2011) 76–82. (6) (2004) 409–417.
[10] H.A. Rasheed, H. Charkas, H. Melhem, Simplified nonlinear analysis of [27] N. Boely, R.M. Botez, G. Kouba, Identification of a non-linear F/A-18 model by
strengthened concrete beams based on a rigorous approach, J. Struct. Eng. the use of fuzzy logic and neural network methods, Proc. Inst. Mech. Eng. Part
130 (7) (2004) 1087–1096. G: J. Aerosp. Eng. 225 (5) (2011) 559–574.
[11] H.A. Rasheed, R. Nayal, H. Melhem, Response prediction of concrete beams [28] K. Nasrollahzadeh, M.M. Basiri, Prediction of shear strength of FRP reinforced
reinforced with FRP bars, Compos. Struct. 65 (2) (2004) 193–204. concrete beams using fuzzy inference system, Expert Syst. Appl. 41 (4) (2014)
[12] P.H. Bischoff, Reevaluation of deflection prediction for concrete beams 1006–1020.
reinforced with steel and fiber reinforced polymer bars, J. Struct. Eng. 131 [29] K.M.U. Darain, M.Z. Jumaat, M.A. Hossain, M.A. Hosen, M. Obaydullah, M.N.
(5) (2005) 752–767. Huda, et al., Automated serviceability prediction of NSM strengthened
[13] P. Visintin, D.J. Oehlers, R. Muhamad, C. Wu, Partial-interaction short term structure using a fuzzy logic expert system, Expert Syst. Appl. 42 (1) (2015)
serviceability deflection of RC beams, Eng. Struct. 56 (2013) 993–1006. 376–389.
[14] A.A. Shukri, K.M.U. Darain, M.Z. Jumaat, The tension-stiffening contribution of [30] T.-W. Park, U.-J. Na, S.-J. Kwon, Prediction of ultimate strength and strain of
NSM CFRP to the behavior of strengthened RC beams, Materials 8 (7) (2015) concrete columns retrofitted by FRP using adaptive neuro-fuzzy inference
4131–4146. system, J. Korea Concr. Inst. 22 (1) (2010) 19–27.
[15] J.-S. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE Trans. [31] V. Balasubramaniam, P. Raghunath, K. Suguna, An adaptive neuro-fuzzy
Syst. Man Cybern. 23 (3) (1993) 665–685. inference system based modeling for corrosion-damaged reinforced HSC
[16] A. Talei, L.H.C. Chua, C. Quek, A novel application of a neuro-fuzzy beams strengthened with external glass fibre reinforced polymer laminates, J.
computational technique in event-based rainfall–runoff modeling, Expert Comput. Sci. 8 (6) (2012) 879.
Syst. Appl. 37 (12) (2010) 7456–7468. [32] M. Jalal, A.A. Ramezanianpour, A.R. Pouladkhan, P. Tedro, Application of
[17] D. Petković, N.D. Pavlović, Ž. Ćojbašić, N.T. Pavlović, Adaptive neuro fuzzy genetic programming (GP) and ANFIS for strength enhancement modeling of
estimation of underactuated robotic gripper contact forces, Expert Syst. Appl. CFRP-retrofitted concrete cylinders, Neural Comput. Appl. 23 (2) (2013) 455–
40 (1) (2013) 281–286. 470.
[18] C. Wu, K. Chau, Prediction of rainfall time series using modular soft computing [33] EN B. 12390-3: 2009, Testing hardened concrete. Making and curing
methods, Eng. Appl. Artif. Intell. 26 (3) (2013) 997–1007. specimens for strength tests 2009, ISBN, 940137696.
[19] S. Shamshirband, D. Petković, N.B. Anuar, M.L. Mat Kiah, S. Akib, A. Gani, et al., [34] ASTM, Standard Test Method for Compressive Strength of Cylindrical Concrete
Sensorless estimation of wind speed by adaptive neuro-fuzzy methodology, Specimens. ASTM C39/C39M-14, American Society for Testing and Materials
Int. J. Electr. Power Energy Syst. 62 (2014) 490–495. (ASTM) International, USA, 2014.
[20] H. Basser, S. Shamshirband, H. Karami, D. Petković, S. Akib, A. Jahangirzadeh, [35] EN B. 12390-5: 2009. Flexural Strength of Test Specimens Testing Hardened
Adaptive neuro-fuzzy selection of the optimal parameters of protective spur Concrete, 2009, 1–14.
dike, Nat. Hazards (2014) 1–12. [36] A615M-14 AA, Standard Specification for Deformed and Plain Carbon-Steel
[21] T.L. Grigorie, R.M. Botez, A.V. Popov, M. Mamou, Y. Mébarki, A hybrid fuzzy Bars for Concrete Reinforcement, ASTM International, 2014.
logic proportional-integral-derivative and conventional on-off controller for [37] SikadurÒ-30, Product Data Sheet-Adhesive for bonding reinforcement. Edition:
morphing wing actuation using shape memory alloy-Part 1: Morphing system 2014-01_1. Retrieved June 02, 2014, from: <http://mys.sika.com/en/solutions_
mechanisms and controller architecture design, Aeronaut. J. 116 (1179) (2012) products/02/02a013/02a013sa06/02a013sa06100/02a013sa06105.html>.
433. [38] J.S. Armstrong, R. Fildes, Correspondence on the selection of error measures for
[22] T.L. Grigorie, R.M. Botez, A.V. Popov, M. Mamou, Y. Mébarki, A hybrid fuzzy comparisons among forecasting methods, J. Forecast. 14 (1) (1995) 67–71.
logic proportional-integral-derivative and conventional on-off controller for [39] J.S. Armstrong, F. Collopy, Error measures for generalizing about forecasting
morphing wing actuation using shape memory alloy, Part 2: Controller methods: empirical comparisons, Int. J. Forecast. 8 (1) (1992) 69–80.
implementation and validation, Aeronaut. J. 116 (1179) (2012) 451–465. [40] L.A. Zadeh, Fuzzy Logic, Neural Networks and Soft Computing. One-page
[23] A.H. Al-Rahmani, H.A. Rasheed, Y. Najjar, A combined soft computing- Course Announcement of CS 294-4, University of California, Berkley, 1992.
mechanics approach to inversely predict damage in bridges, Proc. Comput. [41] J.R. Koza, Genetic Programming: On the Programming of Computers by Means
Sci. 8 (2012) 461–466. of Natural Selection, The MIT Press, Cambridge, MA, 1992. 840 p..
[24] I. Flood, L. Muszynski, S. Nandy, Rapid analysis of externally reinforced [42] G.-B. Huang, L. Chen, C.K. Siew, Universal approximation using incremental
concrete beams using neural networks, Comput. Struct. 79 (17) (2001) 1553– constructive feed forward networks with random hidden nodes, IEEE Trans.
1559. Neural Networks 17 (2006) 879–892.

You might also like