Composite Structures 147 (2016) 4253

Contents lists available at ScienceDirect

Composite Structures
journal homepage: www.elsevier.com/locate/compstruct

A two-step approach for damage detection in laminated composite

structures using modal strain energy method and an improved
differential evolution algorithm
T. Vo-Duy, V. Ho-Huu, H. Dang-Trung, T. Nguyen-Thoi
Division of Computational Mathematics and Engineering, Institute for Computational Science, Ton Duc Thang University, Ho Chi Minh City, Viet Nam
Faculty of Civil Engineering, Ton Duc Thang University, Ho Chi Minh City, Viet Nam

a r t i c l e

i n f o

Article history:
Received 28 January 2016
Revised 3 March 2016
Accepted 9 March 2016
Available online 12 March 2016
Keywords:
Damage identification
Modal strain energy
Improved differential evolution algorithm
Composite structure
Mode shape error

a b s t r a c t
The paper presents a two-step approach based on modal strain energy method and an improved differential evolution algorithm for damage detection in laminated composite structures. First, the modal
strain energy based method is employed to identify a set of potential damaged elements. Then, the
improved differential evolution algorithm is utilized to minimize the function of mode shape error with
design variables relating to the extent of identified damaged elements. Here, the function of mode shape
error is defined by the shift between the mode shape of the damaged structure and that of the healthy
structure. The proposed approach is applied for a cross-ply (0
/90
/0
) laminated composite beam and
a cross-ply (0
/90
/0
) square laminated composite plate with multiple damaged elements. In addition,
the effect of noise on the accuracy of damage identification is also investigated. Numerical results show
that the proposed approach is effective for damage detection in laminated composite beam and plate
structures for both cases with noise and without noise.
 2016 Elsevier Ltd. All rights reserved.

1. Introduction
Due to outstanding properties, such as light weight, high stiffness and strength, etc., the composite materials have been used
widely and popularly in many different engineering disciplines
such as civil infrastructures, aerospace and automotive engineering. Damage in composite structures may significantly reduce their
stiffness and then lead to tragic consequences. As a result, the
development of reliable and efficient damage identification methods for composite structures is really necessary.
In the structural health monitoring (SHM) literature, vibrationbased damage detection methods are widely used for composite
structures. Reviews of these methods can be found in Refs. [14]
and their application for composite structures is reported in
[5,6]. In the categories of vibration-based damage detection methods, frequency change-based method [7,8], curvature mode shapebased method [9,10], modal strain energy based method [11,12],

Corresponding author at: Division of Computational Mathematics and Engineering, Institute for Computational Science, Ton Duc Thang University, Ho Chi
Minh City, Viet Nam.
E-mail addresses: voduytrung@tdt.edu.vn (T. Vo-Duy), hohuuvinh@tdt.edu.vn
(V. Ho-Huu), dangtrunghau@tdt.edu.vn (H. Dang-Trung), nguyenthoitrung@tdt.edu.
vn (T. Nguyen-Thoi).
http://dx.doi.org/10.1016/j.compstruct.2016.03.027
0263-8223/ 2016 Elsevier Ltd. All rights reserved.

flexibility based approach [13] and mode shape based method

[14] have been applied successfully for composite structures like
beams and plates. Besides, some other damage identification
methods, such as wavelet analysis method, optimization-based
method, guided Lamb wave method [15] also employed to identify
the damage location on composite structures. Although there were
amount of damage assessment methods based on vibration characteristics for composite structures, most of them were limited to
damage location only.
There have been some damage identification methods based on
the optimization algorithms which were used to identify the location and extent of damage [4,1619]. The main idea of these methods is to transform the damage identification problem into an
optimization problem in which the objective function is usually
defined by minimizing the difference between measured and analytical model characteristics of the structure and the design variables are the damage ratios of elements in the structure. Several
meta-heuristic optimization algorithms such as genetic algorithm
(GA), particle swarm optimization (PSO), artificial bee colony
(ABC) have been applied successfully for structural damage localization [16,20,21]. One important advantage of this approach is
that the complete set of model characteristics (e.g. modes or displacements) is not needed because the objective function involves
only the difference of main components of these vectors [22].

43

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

Another advantage of this approach is that they can determine

both location and extent of structural damage. However, the application of this approach for composite structures is still somewhat
limited. Besides, the implementation of optimization algorithms
for locating damage in structures still has some drawbacks that
need to be addressed. For example, (1) the accuracy of the optimization algorithms, (2) the expensive computational cost of the
optimization algorithms, (3) the invalid of the optimization algorithms for dealing with large design variables.
Among population based meta-heuristic algorithms (e.g. particle swarm optimization (PSO), artificial bee colony (ABC), and
cuckoo search (CS), etc.), the differential evolution (DE) algorithm,
firstly introduced by Storn and Price [23], has been proven to be
one of the most promising methods when it was evaluated for over
fifty different benchmark functions [24]. It also has been successfully applied and developed for numerous problems in various
fields such as communication [25], mechanical engineering
[2629], structural health monitoring [30,31], artificial neural
network training [32], and so forth. Nevertheless, similar to many
other meta-heuristic algorithms, the main limit of the DE also
concerns with the high computational cost [3335].
This paper hence makes an effort to fill in the above mentioned
research gaps by proposing a two-step approach for damage identification in laminated composite structures by means of using a
modal strain energy based method and a newly improved differential evolution algorithm. Firstly, the modal strain energy based
method referred to [36] is used to identify possible damage elements, and also to reduce the design variables of the optimization
problem for the second step. Secondly, an improved differential
evolution algorithm is utilized to assess the extent of damage
and also to reduce the false alarms of elements found in the first
step. The improvements on the DE are made in mutation and selection phases to reduce the computational cost of the algorithm. The
numerical examples consider a cross-ply (0
/90
/0
) laminated
composite beam and a cross-ply (0
/90
/0
) square laminated composite plate with multiple damaged elements. Moreover, the effect
of noise on the accuracy of the proposed procedure is also investigated. The damage identification results obtained by the proposed
approach are compared to those gained by the combination
between modal strain energy based method and the DE. The
numerical results show that regardless of the effect of noise, the
proposed approach can locate and quantify the extent of damages
efficiently, especially for the computational cost.

2. Damage locating using modal strain energy based method

The modal parameters of a structure, such as frequencies, mode
shapes, and damping ratios are related to the damage in the structure. Therefore, the change of these modal parameters between
damaged and healthy structures was used to locate damage. In
addition, the change of flexibility matrix, mode shape curvature
and modal strain energy established by these modal parameters
is also considered as efficient indicator of damage. In the previous
studies [1,37], authors proved that the change of the modal strain
energy is better than the change of others in locating damage.
In this paper, the modal strain energy based method [36] using
modal strain energy change ratio (MSRCR) as damage indicator is
applied to identify the location of damage on laminated composite
structures. The brief review of the MSECR formulation is presented
as follows.
For the ith mode shape of the damaged and healthy structures,
the modal strain energy (MSE) for both structures at the jth
element can be calculated by
T

MSEhij Uhi Kj Uhi ;

MSEdij Udi Kj Udi

where MSEij is the strain energy; Kj is the stiffness matrix of the jth
element of the healthy structure; Ui is the ith mode shape vector
and the superscripts h and d denote the healthy and damaged
states, respectively. The superscript T denotes the vector transpose.
As pointed out in [36], Eq. (1) shows that when damage occurs in
elements of a system, the MSE will change little in the undamaged
elements, but there will be a larger change in the damaged elements. As a result, the modal strain energy change ratio (MSECR)
was proposed to locate damaged elements. This indicator is
expressed by

MSECRj

m
1X
MSECRij
m i1 MSECRmax
i

where

MSECRij

jMSEdij  MSEhij j

MSEhij

and

MSECRmax
maxfMSECRik g
i

3. Damage severity assessment

One type of damage identification methods is to transform the
damage localization into an optimization problem in which the
objective function is defined as the difference between the modal
parameter of the damaged and healthy structures. The variables
of the problem are defined as the damage severity of elements.
Many kinds of objective functions have been proposed. Some typical objective functions can be mentioned as the frequency error
[20], the MDLAC coefficient error (mode shape) [38] or the flexibility matrix error [39].
3.1. Objective function based on the mode shape error
In this paper, an objective function f based on the change of
mode shape is defined as follows

f x



nm  d
X
Ui  Uhi x
 d
;
U 
i1

x x1 ; . . . ; xn 2 0; 1n

where x is the design variable vector corresponding with the extent

of damage of n elements; Ui is the ith mode shape vector; nm is the
number of modes considered; k  k is the Euclidean norm and the
superscripts h and d denote the healthy and damaged states,
respectively.
3.2. Differential evolution algorithm
The differential evolution (DE) algorithm [23], designed to dealing with nonlinear, non-differentiable and multimodal objective
function for continuous optimization problems, has been demonstrated effectiveness and robustness by many researches for both
benchmark and real-world problems [34]. In the DE, a few parameters are used and make easy implement for users. The basic procedure of the algorithm is briefly summarized via four phases as
follows.
3.2.1. Initialization
Initially, a population with NP individuals is created by randomly sampling from the search space. Each individual is a vector
containing D design variables xi = {xi,1, xi,2, xi,3,. . ., xi,D} and is produced as

xi;j xlj rand0; 1  xuj  xlj i 1; 2; :::; NP;

j 1; 2; :::; D

44

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

where xlj and xuj are the lower and upper bounds of xj , respectively;
rand[0, 1] is a uniformly distributed random number in [0, 1]; NP is
the size of population; and D is the number of design variables.
3.2.2. Mutation
Secondly, each individual called the target vector xi in the population is used to generate a mutant vector vi by mutation operations. Four popular mutation operations are often used in the DE
as follows

-rand=1 :

v i xr

F  xr2  xr3

-rand=2 :

v i xr

F  xr2  xr3 F  xr4  xr5

-best=1 :

vi xbest F  xr

 xr 2

-best=2 :

vi xbest F  xr

 xr2 F  xr3  xr4

9
10

where integers r1, r2, r3, r4, r5 are randomly selected from {1, 2,. . .,
NP} such that r1 r2 r3 r4 r5 i; the scale factor F is randomly chosen within [0, 1]; and xbest is the best individual (i.e.
the individual having the smallest value of the objective function)
in the current population.
After mutation, if the boundary constraints of the jth components vij of mutant vector vi are violated, it will be reflected back
to allowable region as follows

v ij

8 l
l
>
< 2xj  v ij if v ij < xj
u
2xj  v ij if v ij > xuj
>
:
v ij otherwise

11

delta jf mean =f best  1j

3.2.3. Crossover
Next, each target vector xi generates a trial vector ui by replacing some elements of the vector xi by some elements of the mutant
vector vi using binomial crossover operation

uij

v ij

if rand0; 1 6 CR or j jrand

xij

otherwise

12

where i e {1, 2,. . ., NP}; j e {1, 2,. . ., D}; jrand is an integer selected
from 1 to D; and CR e [0, 1] is the crossover control parameter.
3.2.4. Selection
Finally, the trial vector ui is compared to the target vector xi.
The better one with lower objective function value will be chosen
for the next generation

xi

ui if f ui 6 f xi
xi otherwise

3.3.1. Modification of the mutation phase

In the population based optimization algorithms, the balance
between two aspects (i.e. global exploration and local exploitation)
is great important [34]. The standard differential evolution (DE)
algorithm, however, does not keep this balance in its procedure
because it uses only one mutation operator which is only good
on one side [41,42]. More specifically, the rand/1 and rand/2 operators are good at global search but limited at local search, whereas
the best/1 and best/2 operators are good at local search but bad at
global search. Thus, in this study, to effectively equalize the global
and local search abilities of the DE, these four popular mutation
operators will be adaptively employed for the mutation phase during search process. In each generation, two of them are chosen following an adaptive scheme based on the absolute deviation
(denoted as delta) between the best and the mean of objective
function in the previous generation. In particular, in the sooner
generations, when the value of the delta is still big, two mutation
operators rand/1 and rand/2 are randomly selected with the
rate of 0.7/0.3 for exploring the domain containing the global optimum. After many generations of the global search process, when
the diversity of the population is gradually stable and the delta is
small, two mutation operators best/1 and best/2 are randomly
picked with the rate of 0.7/0.3 for accelerating the convergence
rate of the algorithm. Note that, the delta here is considered as
the adaptive control parameter that is compared with a predicted
small threshold to decide which one of mutation operators rand/1
and rand/2 or best/1 and best/2 to be used. More specifically,
the value of delta is defined by

13

3.3. Improved differential evolution algorithm

In the DE, the parameters such as mutant factor F and crossover
control parameter CR and trial vector generation strategies have
significant influence on its performance [40]. Commonly, to
acquire the most satisfactory optimization performance for a particular problem, they are often defined using a trial-and-error
search. Obviously, it may lead to a huge amount of the computational cost. In addition, a rigorous selection mechanism in the
selection phase may make the DE computationally expensive.
Therefore, to overcome these limitations, the paper introduces
two modifications in two phases, mutation and selection. In the
mutation phase, an adaptive mutation scheme with multimutation operators is presented, while in the selection phase, a
new selection technique is introduced. Their details are presented
in the following subsections.

14

where f best is the objective function value of the best individual and
f mean is the mean objective function value of the whole population.
The adaptive mutation scheme is briefly described as follows
if (delta > threshold)
if rand[0, 1] > 0.3

v i xr

F  xr2  xr3

15

F  xr2  xr3 xr4  xr5

16

else

v i xr

end
else (delta 6 threshold)
if rand[0, 1] > 0.3

vi xbest F  xr

 xr 2

17

 xr2 F  xr3  xr4

18

else

vi xbest F  xr
end
end

where the scale factors F is randomly generated in the interval [0.4,

1] during searching progress instead of being fixed as in the original
DE. This aims to create the variety of searching directions for the
both cases of the delta (delta > threshold and delta 6 threshold) and
threshold is the criterion value which is chosen based on the stopping criterion of the algorithm and is discussed in more detail in
Section 3.3.3.
3.3.2. Modification of the selection phase
From selection mechanism in the standard DE, it can be recognized that some good information of unselected individuals can be
neglected. This is because an unselected individual is worse than

45

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

its target individual in the pair, but it can be still better than other
individuals in the entire population. As a result, the DE may lead to
slow convergence. In fact, the good individuals often contain useful
characteristics for searching progress. Therefore, to keep good
information for the next generation, the elitist selection technique
introduced in [29,43] is employed for the selection phase instead of
the basic selection of the DE.
3.3.3. The proposed improved differential evolution algorithm
By integrating two above-mentioned modifications into the DE,
the improved differential evolution (IDE) algorithm is summarized
as follows
Algorithm 1: The proposed algorithm (IDE)
1: Generate the initial population
2: Evaluate the fitness for each individual in the population
3: while (delta > tolerance and MaxIter is not reached) do
4: for i = 1 to NP do
5: jrand = randint(1, D)
6: CR = rand[0.7, 1]
7: F = rand[0.4, 1]
8:
for j = 1 to D do
9:
if rand[0,1] < CR or j = jrand then
10:
if delta > threshold then
11:
if rand > 0.3 then % using rand/1
12:
Select randomly r1 r2 r3 i,
"i2{1, 2,. . ., NP}
13:
ui;j xr1;j F  xr2;j  xr3;j
14:
else
% using rand/2
15:
Select randomly r1 r2 r3 r4 r5 i,
"i2{1, 2,. . ., NP}
16:
ui;j xr1;j F  xr2;j  xr3;j F  xr4;j  xr5;j
17:
end if
18:
else delta 6 threshold
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:

if rand > 0.3 then % using best/1

Select randomly r1 r2 best i,
"i2{1, 2,. . ., NP}
ui;j xbest;j F  xr1;j  xr2;j
else
% using best/2
Select randomly r1 r2 r3 r4 best i,
"i2{1, 2,. . ., NP}
ui;j xbest;j F  xr1;j  xr2;j F  xr3;j  xr4;j
end if
end if
else
ui;j xi;j
end if
end for
Evaluate the trial vector ui
end for
Do selection phase following Section 4.2
Define f best ; f mean
delta jf mean =f best  1j
end while

where tolerance is the allowed error; MaxIter is the maximum

number of iterations; and randint(1, D) is a function which returns
a uniformly distributed random integer number between 1 and D.
According to Algorithm 1, the IDE will finish the searching
progress either when the delta is less than or equal to the tolerance
or when the maximum number of iterations (MaxIter) is achieved.
It can be recognized that the threshold is chosen based on the
tolerance and obviously, it must be bigger than the tolerance.

Moreover, it is important to note that choosing the value of threshold will directly effect on the global and local search capabilities of
the IDE. If the value of threshold is too big compared to the value of
the tolerance, the IDE will have a priority in global searching; otherwise, if it is too small, the IDE will have a priority in local searching.
Thus, to set an adequate value for the threshold, it should be based
on the features of a particular problem. For instance, for highly
nonlinear and complex problems, the value of threshold should
be small (104 or 105 for threshold and 106 for tolerance). In contrast, for small and simple problems, the value of threshold should
be large (102 or 103 for threshold and 106 for tolerance). In this
work, we found that the threshold of 104 and tolerance of 106 are
good values which are capable of efficiently equalizing between
exploration and exploitation capacities.
4. Numerical examples
The effectiveness and robustness of the proposed procedure are
verified on a symmetric cross-ply (0
/90
/0
) cantilevered beam
and a symmetric cross-ply (0
/90
/0
) clamped square plate. Some
damage scenarios with multiple damage locations in the structures
are investigated. The damage in these structures is simulated by
reducing the stiffness of selected elements, i.e. Ked 1  aKeh ,
where Ked and Keh are the stiffness matrices corresponding to damaged and healthy structures of the eth element and a (0 6 a 6 1) is
the damage ratio of the eth element. In addition, the influence of
noise on the accuracy of the proposed procedure is also considered.
More specifically, the noise on the mode shape is added as follows
[44]

Unoise
1 g2rand0; 1  1jUij j
ij

19

Table 1
The parameters of the cross-fly (0
/90
/0
) cantilevered composite beam.
Parameters/unit

Value

Length (L)/m
Thickness (h)/m
Width (b)/m
Youngs modulus (E1)/GPa
Youngs modulus (E2)/GPa
Mass density (q)/kg/m3
Poisson ratio (v12)
Shear modulus (G12)/GPa
Shear modulus (G23)/GPa
Shear modulus (G13)/GPa

0.1
0.01
0.01
40
1
1000
0.25
0.6 E2
0.5 E2
0.6 E2

Fig. 1. Element numbering of a cross-fly (0
/90
/0
) cantilevered composite beam.

Table 2
First eight frequencies of the healthy (undamaged) and damaged cases of the cross-fly
(0
/90
/0
) cantilevered composite beam.
Mode

Intact (FOBT) [45]

Intact (present)

Damaged (present)

1
2
3
4
5
6
7
8

5.479

5.478
22.315
46.656
70.534
81.654
95.151
120.047
146.039

5.287
21.358
45.205
65.486
78.128
91.516
115.300
140.890

46

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

where Unoise
is the jth component of the ith mode shape vector conij
taminated by noise and g is the noise level. In the following examples, 3% noise level is chosen. Parameters of the IDE and DE for all
tests are listed in Table 3. To obtain the statistical result of damage
severity evaluation, five independent runs are performed for both
optimization algorithms. The codes of the finite element analysis
for the laminated composite beam and plate, the IDE and DE are
written in Matlab.
4.1. A three cross-ply (0
/90
/0
) cantilevered beam
4.1.1. Model of damage
In this example, the symmetric cross-fly (0
/90
/0
) cantilevered beam previously analyzed by Khdeir and Reddy [45] is
carried out. The material properties of the beam are given in
Table 1. The beam is divided into 16 equal elements, as depicted
in Fig. 1. A local damage is assumed by a reduction of 30% and

50% in stiffness of the 4th and 10th elements, respectively. The first
eight normalized frequencies of the healthy and damaged beams
are listed in Table 2 in comparison with those obtained by Khdeir
and Reddy.
4.1.2. Identification of damage sites using modal strain energy based
method
The MSECR for all elements of the beam is calculated for cases
without noise and with noise. In the case of noise-free, the first
two modes are used to calculate the MSECR while in the case of
considering measurement noise, various numbers of modes (6, 7,
and 8) are utilized to compute the MSECR. The MSECR for both
cases are depicted in Fig. 2.
It can be seen that for the case of noise-free, the two actual
damaged elements are located exactly; however, for the case of
considering measurement noise, besides the two actual damaged
elements, some false alarm elements which have large MSECR
but they are not the actual damaged elements (e.g. the 11th and
16th elements) are also identified. It should also be mentioned that
for the case of considering measurement noise, the usage of different number of modes influents on the damage assessment. For
example, the usage of six modes leads to three false alarm elements (e.g. the 11th, 13th and 16th elements) while the use of
seven and eight modes leads to two false alarm elements (the
11th and 16th elements).
4.1.3. Quantification of damage severities using the IDE and DE
In this section, the damage severity assessment at the damaged
elements is implemented using the IDE and DE. The assessment
carries out two cases: noise-free and considering measurement

Fig. 2. MSECR for all elements of the cross-fly (0
/90
/0
) cantilevered composite
beam.

Table 3
Parameters of the IDE and DE.
Parameter

DE

IDE

Size of population (NP)

Mutant factor (F)
Crossover control parameter (CR)
Stop criterion (tolerance)
Maximum iteration (MaxIter)
Selection scheme
Mutation operator
Adaptive parameter (threshold)

20
0.8
0.9
106
5000
Binomial
rand/1
Invalid

20
rand[0.4, 1]
rand[0.7, 1]
106
5000
Elitist
rand/1; best/1; rand/2; best/2
102

Fig. 3. Convergence histories of the IDE and DE in the case of noise-free.

Table 4
The statistical results of damage severity assessment and the average number of structural analyses in the case of noise-free for the cross-fly (0
/90
/0
) cantilevered composite
beam.
Method

DE
x1

Bst. f
Avg. value
Wst. f
Std. f
Avg. NSA

0.3000
0.3000
0.3000
3176

IDE
x2
0.500
0.500
0.500

f
07

1.3725  10
1.9886  1007
2.8733  1007
7.4637  1008

x1

x2

0.3000
0.3000
0.3000

0.500
0.500
0.500

2.3657  1007
5.1209  1007
9.6922  1007
2.7468  1007

1024

x1 = damage ratio of element 4; x2 = damage ratio of element 10; f = objective function; Bst. f = best value with respect to f; Avg. value = average value with respect to both x1,
x2 and f; Wst. f = worst value with respect to f; Std. f = standard deviation with respect to f; Avg. NSA = average number of structural analyses.

47

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

Table 5
The statistical results of the damage severity assessment and the average number of structural analyses in the case of considering noise for the cross-fly (0
/90
/0
) cantilevered
composite beam.
Method

Bst. f
Avg. value
Wst. f
Std. f
Avg. NSA

DE

IDE

x1

x2

x3

x4

x1

x2

x3

x4

0.3008
0.3007
0.3008

0.4976
0.4976
0.4975

0.0018
0.0019
0.0018

5.8083  1007
8.4020  1007
1.5211  1006

0.1111
0.1111
0.1111
1.8791  1007

0.3008
0.3007
0.3007

0.4975
0.4976
0.4975

0.0017
0.0018
0.0020

9.8067  1008
1.6419  1006
4.4038  1006

0.1111
0.1111
0.1111
4.2471  1007

8056

2536

x1 = damage ratio of element 4; x2 = damage ratio of element 10; x3 = damage ratio of element 11; x4 = damage ratio of element 16; f = objective function; Bst. f = best value
with respect to f; Avg. value = average value with respect to x1, x2, x3, x4 and f; Wst. f = worst value with respect to f; Std. f = standard deviation with respect to f; Avg.
NSA = average number of structural analyses.

noise. The objective function, as given in Section 3.1, is evaluated

using the first four and eight modes corresponding to the cases
without noise and with measurement noise. The design variables
are denoted by x = (x1, x2, . . ., xi, . . ., xn) (i = 1, . . ., n) are the damage
extents of n suspected elements including actual damaged elements and false alarm elements. Particularly, for the case of
noise-free, x = (x1, x2) in which x1 and x2 are the damage extents
of the 4th and 10th elements, respectively while for the case of

noise 3%, x = (x1, x2, x3, x4) in which x1, x2, x3, and x4 are the damage
extents of the 4th, 10th, 11th and 16th elements, respectively.
4.1.3.1. Case of noise free. The statistical results including best,
average and worst solutions, standard deviation, and the average
number of structural analyses of damage quantification using the
IDE and DE are provided in Table 4. It can be observed that the
damage ratios of the 4th and 10th elements are correctly identified
for both optimization algorithms. The best (smallest) and worst
(largest) value of the objective function are very small and the

Table 6
The parameters of the cross-ply (0
/90
/0
) square composite plate.
Parameter/unit

Value

The length of a side (a)/m

The thickness (t)/m
Youngs Modulus (E1)/N/m2
Youngs Modulus (E2)/N/m2
Poisson ratio v12
Poisson ratio v21

1
0.1
40
1
0.25
0.00625

Table 7
Two damage scenarios in the cross-ply (0
/90
/0
) square composite plate.
Scenario 1

Scenario 2

Element No.

Damage ratio

Element No.

Damage ratio

37
47

0.40
0.60

33
57
74

0.15
0.20
0.25

Fig. 4. Convergence histories of the IDE and DE in the case of considering noise.

Fig. 5. (a) A cross-ply (0
/90
/0
) square composite plate; (b) element numbering of the plate.

48

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

solutions (damage ratios) at these values are the same as actual

damage ratios. Here, it should be noted that because the values
of the objective function are very close (e.g. 0.299999793564568
and
0.299999687349585
for
the
4th
element
and
0.499999901387455 and 0.499999831222834 for the 10th element) and all computed entries are rounded to four decimal places,
the damage ratios are the same for different values of the objective
function. Also from Table 4, it can be seen that the IDE requires
much less structural analyses than the DE. In addition, the standard

Table 8
First eight frequencies of the intact (undamaged) and damaged cases of the cross-ply
(0
/90
/0
) square composite plate.
Mode

Intact
[47]

Intact
[46]

Intact
(present)

Damaged case
1 (present)

Damaged case
2 (present)

1
2
3
4
5
6
7
8

7.5622
10.2890
14.3270
14.8030
16.0780
19.3810
20.2000
21.8920

7.4108
10.3930
13.9130
15.4290
15.8060
19.5720
21.4890
21.6230

7.4444
10.4378
13.9895
15.5196
15.8873
19.6861
21.6269
21.8126

7.3770
10.3767
13.8495
15.3784
15.8133
19.5522
21.0219
21.6020

7.4216
10.4002
13.9511
15.4577
15.8409
19.6224
21.4823
21.7258

deviation of the objective function is very small for both

algorithms.
The convergence histories of the best run yielding the smallest
value of the objective function of the IDE and DE are shown in
Fig. 3. It is easy to see that the IDE converges much faster than
the DE. The IDE stops searching process around 60 iterations
(1200 structural analyses) while the DE stops searching around
160 iterations (3200 structural analyses).
4.1.3.2. Case of noise level 3% for mode shape. Unlike the previous
case, due to the effect of noise, two new design variables (damage
ratios of elements 11 and 16) are included in this case. The statistical results including best, average, and worst solutions, standard
deviation, and the average number of structural analyses of the IDE
and DE are given in Table 5. As can be seen from the table, the damage ratios of the 4th and 10th elements agree well with the actual
damage ratios while the damage ratios of the 11th and 16th elements are nearly to zero for both optimization algorithms. This
helps eliminate the 11th and 16th elements from the suspected
elements. It can also be realized that the values of the objective
function are larger than those in the case of noise-free. Again, the
IDE uses much smaller the average number of structural analyses
than the DE and the standard deviation of the objective function

Fig. 6. MSECR for all elements of the cross-ply (0
/90
/0
) square composite plate of scenario 1. (a) noise-free, using 2 modes; (b) noise 3%, using 4 modes; (c) noise 3%, using
5 modes; (d) noise 3%, using 6 modes.

49

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

is very small for both algorithms. The convergence histories of the

best run of the IDE and DE are compared in Fig. 4. Obviously,
the IDE has much faster convergence than the DE. Particularly,
the IDE stops searching process around 148 iterations (2960 structural analyses) while the DE stops searching process around 380
iterations (7600 structural analyses).
4.2. A three cross-ply (0
/90
/0
) clamped composite plate
4.2.1. Model of damage
The geometry and material parameters of the laminated composite plate are given in Fig. 5 and Table 6 respectively. The free
vibration behavior of the plate was previously analyzed by many
researchers such as Ferreira et al. [46] using a high order collocation method and Phung-Van et al. [47] using an edge-based
smoothed discrete shear gap method (ES-DSG3). In this example,
the free vibration behavior of the plate is analysed by using finite
element method with mesh 10x10 of 9-node quadratic quadrilateral element. Two damage scenarios consisting of two and three
damaged elements are investigated. The details of damaged elements and their damage ratios of these scenarios are summarized
in Table 7. The first eight frequencies of the laminated composite
plate in the healthy and damaged states are presented in Table 8.
It can be recognized that for the healthy case, the frequencies are

in good agreement with the corresponding results obtained by

[46] and [47] and for the damaged cases, the frequencies are
decreased but the change of frequencies between damaged and
healthy cases is not significant.
4.2.2. Identification of damage sites using modal strain energy based
method
For both scenarios of damage, the MSRCR is approximated by
using the first two modes for the case of noise-free and the first
four, five and six modes for the case of considering measurement
noise. The MSECR for all elements of scenario 1 and scenario 2 is
respectively depicted in Figs. 6 and 7.
Fig. 6 indicates that for the case of noise-free, the 35th and 47th
elements are the suspected elements which are the same as actual
damaged elements and for the case of considering measurement
noise, the damage identification results depend on the number of
modes for approximating the MSECR. The usage of lower number
of modes (four modes) yields the largest MSEC at the 35th and
47th elements but it also leads to the high value of the MSCE of
the 46th element which can be seen as the false alarm element.
Whereas, the usage of larger number of modes (five and six modes)
can eliminate the false alarm element.
Fig. 7 illustrates that three damaged elements are located
exactly for the case of noise free without any false alarm elements.

Fig. 7. MSECR for all elements of the cross-ply (0
/90
/0
) square composite plate of scenario 2. (a) Noise-free, using 2 modes; (b) noise 3%, using 4 modes; (c) noise 3%, using
5 modes; (d) noise 3%, using 6 modes.

50

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

Similar to scenario 1, the number of modes for approximating the

MSECR influents on the damage identification result for the case of
considering measurement noise.
4.2.3. Quantification of damage severities using the IDE and DE
This section aims to estimate the severity of damage found in
the previous section for both scenarios with and without noise.
For both cases of noise-free and considering measurement noise,
the design variables for scenario 1 are the damage ratios of the
37th and 47th elements while those for scenario 2 are the damage
ratios of the 33th, 57th and 74th elements. The numbers of modes
used to approximate the objective function for both damage scenarios are taken to be four and six corresponding to the cases without noise and with measurement noise.

4.2.3.1. Case of noise free. The statistical results and the average
number of structural analyses of damage severity evaluation for
two damage scenarios are respectively presented in Tables 9 and
10. It can be observed that the true damage ratios are correctly
identified. Similar to the example of the beam, the IDE requires
much less structural analyses than the DE. In addition, the standard
deviation of both optimization algorithms is very small. Fig. 8 illustrates the convergence histories of the IDE and DE for both damage
scenarios. From the figure, it is easy to see that the IDE converges
considerably faster than the DE.
4.2.3.2. Case of noise level 3% for mode shape. Tables 11 and 12 provide the statistical results in term of best, average, and worst solutions, standard deviation, and the average number of structural

Table 9
The statistical results of damage severity assessment and the average number of structural analyses for scenario 1 of the cross-ply (0
/90
/0
) square composite plate in the case of
noise-free.
Method

DE

Bst. f
Avg. value
Wst. f
Std. f
Avg. NSA

IDE

x1

x2

x1

x2

0.40
0.40
0.40

0.60
0.60
0.60

6.2740  1008
1.8519  1007
2.8811  1007
8.8870  1008

0.40
0.40
0.40

0.60
0.60
0.60

2.0043  1007
4.3163  1007
1.0967  1006
3.7665  1007

7548

2660

x1 = damage ratio of element 37; x2 = damage ratio of element 47; f = objective function; Bst. f = best value with respect to f; Avg. value = average value with respect to both
x1, x2 and f; Wst. f = worst value with respect to f; Std. f = standard deviation with respect to f; Avg. NSA = average number of structural analyses.

Table 10
The statistical results of damage severity assessment and the average number of structural analyses for scenario 2 of the cross-ply (0
/90
/0
) square composite plate in the case of
noise-free.
Method

Bst. f
Avg. value
Wst. f
Std. f
Avg. NSA

DE

IDE

x1

x2

x3

x1

x2

x3

0.15
0.15
0.15

0.20
0.20
0.20

0.25
0.25
0.25

2.6988  1007
6.3076  1007
1.1374  1006
3.3291  1007

0.15
0.15
0.15

0.20
0.20
0.20

0.25
0.25
0.25

9.6003  1008
4.5083  1007
6.0394  1007
2.0442  1007

8872

2520

x1 = damage ratio of element 33; x2 = damage ratio of element 57; x3 = damage ratio of element 74; f = objective function; Bst. f = best value with respect to f; Avg.
value = average value with respect to both x1, x2, x3 and f; Wst. f = worst value with respect to f; Std. f = standard deviation with respect to f; Avg. NSA = average number of
structural analyses.

Fig. 8. Convergence histories of the IDE and DE in the case of noise-free. (a) Scenario 1; (b) scenario 2.

51

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

Table 11
The statistical results of damage severity assessment and the average number of structural analyses for scenario 1 of the cross-ply (0
/90
/0
) square composite plate in the case of
considering noise.
Method

DE

Bst. f
Avg. value
Wst. f
Std. f
Avg. NSA

IDE

x1

x2

x1

x2

0.4054
0.4055
0.4058

0.5957
0.5956
0.5956

0.1035
0.1035
0.1035
7.5396  1008

0.4054
0.4053
0.4051

0.5956
0.5957
0.5959

0.1035
0.1035
0.1035
1.2968  1007

7644

2688

x1 = damage ratio of element 37; x2 = damage ratio of element 47; f = objective function; Bst. f = best value with respect to f; Avg. value = average value with respect to both x1,
x2 and f; Wst. f = worst value with respect to f; Std. f = standard deviation with respect to f; Avg. NSA = average number of structural analyses.

Table 12
The statistical results of damage severity assessment and the average number of structural analyses for scenario 2 of the cross-ply (0
/90
/0
) square composite plate in the case of
considering noise.
Method

Bst. f
Avg. value
Wst. f
Std. f
Avg. NSA

DE

IDE

x1

x2

x3

x1

x2

x3

0.1456
0.1457
0.1463

0.1975
0.1976
0.1975

0.2511
0.2511
0.2510

0.1056
0.1056
0.1056
6.4839  1008

0.1461
0.1458
0.1460

0.1980
0.1975
0.1971

0.2511
0.2511
0.2512

0.1056
0.1056
0.1056
4.7573  1008

13,956

3464

x1 = damage ratio of element 33; x2 = damage ratio of element 57; x3 = damage ratio of element 74; f = objective function; Bst. f = best value with respect to f; Avg.
value = average value with respect to both x1, x2, x3 and f; Wst. f = worst value with respect to f; Std. f = standard deviation with respect to f; Avg. NSA = average number of
structural analyses.

Fig. 9. Convergence histories of the IDE and DE in the case of considering noise. (a) Scenario 1; (b) scenario 2.

analyses for both damage scenarios, respectively. As shown in

Table 11, the IDE and DE can determine the extent of damaged elements with very small error ratios. Particularly, the relative error
between obtained average damage ratios and the actual damage
ratios of the 37th and 47th elements by the DE are 1.38% and
0.73%, respectively while those by the IDE are 1.32% and 0.72%,
respectively. Similar to scenario 1, the damage ratios, obtained
by the IDE and DE, are also identified with very small error ratios
in scenario 2, as illustrated in Table 12. Again, the IDE requires
much less structural analyses than the DE. In addition, the standard
deviation of the IDE and DE is also very small.
The comparison of convergence histories of the IDE and DE for
two damage scenarios is shown in Fig. 9. It is clearly seen that
the convergence speed of the IDE is significantly higher than that
of the DE.

5. Conclusions
In this work, a two-step approach based on modal strain energy
based method and an improved differential evolution (IDE) algorithm is presented for damage identification of laminated composite structures. First, the modal strain energy based method is used
to identify damage locations, and then a new improved differential
evolution algorithm is employed to quantify the extent of these
damages. The IDE is proposed by modifying the mutation phase
and selection phase of the standard differential evolution (DE)
algorithm. For the mutation phase, multiple mutation operators
are adaptively utilized for keeping the trade-off between global
exploration and local exploitation of the optimization algorithm.
For the selection phase, the elitist scheme is used to replace the
traditional section scheme of the DE. The numerical examples for

52

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

a cross-ply (0
/90
/0
) composited beam and the three-layer

square laminated composite plate are performed to show the effectiveness and robustness of the proposed approach. In these examples, some damage scenarios including two and three damaged
elements are simulated by reducing the stiffness of these elements.
Also, the effect of noise is investigated. The numerical results
demonstrate that the modal strain energy based method is effective for locating damage, regardless the effect of noise. The IDE
and DE can quantify damage extents correctly not only in case of
noise-free but also in case of noise level of 3% in mode shape. In
addition, the IDE always requires much less structural analyses
than the DE. This shows the practice and robustness of the IDE in
damage severity assessment for the laminated composite structures with a large number of variables.
Acknowledgement
This research is funded by Vietnam National Foundation for
Science and Technology Development (NAFOSTED) under Grant
number 107.99-2014.11.
References
[1] Doebling SW, Farrar CR, Prime MB, Shevitz DW. Damage identification and
health monitoring of structural and mechanical systems from changes in their
vibration characteristics: a literature review. Los Alamos National Laboratory;
1996.
[2] Salawu OS. Detection of structural damage through changes in frequency: a
review. Eng Struct 1997;19:71823. http://dx.doi.org/10.1016/S0141-0296
(96)00149-6.
[3] Carden EP. Vibration based condition monitoring: a review. Struct Heal Monit
2004;3:35577. http://dx.doi.org/10.1177/1475921704047500.
[4] Fan Wei, Qiao Pizhong. Vibration-based damage identification methods: a
review and comparative study. Struct Heal Monit 2011;10:83111. http://dx.
doi.org/10.1177/1475921710365419.
[5] Montalvao D, Maia NMM, Ribeiro AMR, Montalvao D, Maia NMM, Ribeiro AMR.
A review of vibration-based structural health monitoring with special
emphasis on composite materials. Shock Vib Dig 2006;38:295324. http://
dx.doi.org/10.1177/0583102406065898.
[6] Zou Y, Tong L, Steven GP. Vibration-based model-dependent damage
(delamination) identification and health monitoring for composite structuresa
review. J Sound Vib 2000;230:35778. doi: 10.1006/jsvi.1999.2624.
[7] Cawley P, Adams RD. A vibration technique for non-destructive testing of fibre
composite structures. J Compos Mater 1979;13:16175. http://dx.doi.org/
10.1177/002199837901300207.
[8] Kessler SS, Spearing SM, Atalla MJ, Cesnik CES, Soutis C. Damage detection in
composite materials using frequency response methods. Compos Part B Eng
2002;33:8795. http://dx.doi.org/10.1016/S1359-8368(01)00050-6.
[9] Hamey CS, Lestari W, Qiao P, Song G. Experimental damage identification of
carbon/epoxy composite beams using curvature mode shapes. Struct Heal
Monit 2004;3:33353. http://dx.doi.org/10.1177/1475921704047502.
[10] Lestari, Qiao W, Hanagud P, Sathya. Curvature mode shape-based damage
assessment of carbon/epoxy composite beams. J Intell Mater Syst Struct
2007;18:189208. http://dx.doi.org/10.1177/1045389X06064355.
[11] Kumar M, Shenoi RA, Cox SJ. Experimental validation of modal strain energies
based damage identification method for a composite sandwich beam. Compos
Sci
Technol
2009;69:163543.
http://dx.doi.org/10.1016/
j.compscitech.2009.03.019.
[12] Hu H, Wang B-T, Lee C-H, Su J-S. Damage detection of surface cracks in
composite laminates using modal analysis and strain energy method. Compos
Struct 2006;74:399405. http://dx.doi.org/10.1016/j.compstruct.2005.04.020.
[13] Vo-Duy T, Nguyen-Minh N, Dang-Trung H, Tran-Viet A, Nguyen-Thoi T.
Damage assessment of laminated composite beam structures using damage
locating vector (DLV) method. Front Struct Civ Eng 2015;9:45765. http://dx.
doi.org/10.1007/s11709-015-0303-0.
[14] Moreno-Garca P, Arajo dos Santos JV, Lopes H. A new technique to optimize
the use of mode shape derivatives to localize damage in laminated composite
plates.
Compos
Struct
2014;108:54854.
http://dx.doi.org/10.1016/
j.compstruct.2013.09.050.
[15] Su Z, Ye L, Lu Y. Guided lamb waves for identification of damage in composite
structures: a review. J Sound Vib 2006;295:75380. http://dx.doi.org/10.1016/
j.jsv.2006.01.020.
[16] Seyedpoor SM. A two stage method for structural damage detection using a
modal strain energy based index and particle swarm optimization. Int J Non
Linear
Mech
2012;47:18.
http://dx.doi.org/10.1016/j.ijnonlinmec.
2011.07.011.
[17] Nanthakumar SS, Lahmer T, Rabczuk T. Detection of flaws in piezoelectric
structures using extended FEM. Int J Numer Methods Eng 2013;96:37389.
http://dx.doi.org/10.1002/nme.4565.

[18] Nanthakumar SS, Lahmer T, Rabczuk T. Detection of multiple flaws in

piezoelectric structures using XFEM and level sets. Comput Methods Appl
Mech Eng 2014;275:98112. http://dx.doi.org/10.1016/j.cma.2014.03.001.
[19] Nanthakumar SS, Lahmer T, Zhuang X, Zi G, Rabczuk T. Detection of material
interfaces using a regularized level set method in piezoelectric structures.
Inverse Probl Sci Eng 2016;24:15376. http://dx.doi.org/10.1080/
17415977.2015.1017485.
[20] Friswell MI, Penny JET, Garvey SD. A combined genetic and eigensensitivity
algorithm for the location of damage in structures. Comput Struct
1998;69:54756.
[21] Yang Z-B, Chen X-F, Xie Y, Miao H-H, Gao J-J, Qi K-Z. Hybrid two-step method
of damage detection for plate-like structures. Struct Control Heal Monit 2015.
http://dx.doi.org/10.1002/stc.1769. n/a n/a.
[22] Moslem K, Nafaspour R. Structural damage detection by genetic algorithms.
AIAA J 2002;40:1395401.
[23] Storn R, Price K. Differential evolution a simple and efficient heuristic for
global optimization over continuous spaces. J Glob Optim 1997;11:34159.
http://dx.doi.org/10.1023/A:1008202821328.
[24] Civicioglu P, Besdok E. A conceptual comparison of the Cuckoo-search, particle
swarm optimization, differential evolution and artificial bee colony
algorithms. Artif Intell Rev 2013;39:31546. http://dx.doi.org/10.1007/
s10462-011-9276-0.
[25] Storn R. On the usage of differential evolution for function optimization. In:
Fuzzy Inf Process Soc 1996 NAFIPS, 1996 Bienn Conf North Am 1996:51923.
http://dx.doi.org/10.1109/NAFIPS.1996.534789.
[26] Ho-Huu V, Nguyen-Thoi T, Nguyen-Thoi MH, Le-Anh L. An improved
constrained differential evolution using discrete variables (D-ICDE) for
layout optimization of truss structures. Expert Syst Appl 2015;42:705769.
http://dx.doi.org/10.1016/j.eswa.2015.04.072.
[27] Le-Anh L, Nguyen-Thoi T, Ho-Huu V, Dang-Trung H, Bui-Xuan T. Static and
frequency optimization of folded laminated composite plates using an
adjusted differential evolution algorithm and a smoothed triangular plate
element. Compos Struct 2015;127:38294. http://dx.doi.org/10.1016/
j.compstruct.2015.02.069.
[28] Ho-Huu V, Nguyen-Thoi T, Le-Anh L, Nguyen-Trang T. An effective reliabilitybased improved constrained differential evolution for reliability-based design
optimization of truss structures. Adv Eng Software 2016;92:4856. http://dx.
doi.org/10.1016/j.advengsoft.2015.11.001.
[29] Ho-Huu V, Nguyen-Thoi T, Vo-Duy T, Nguyen-Trang T. An adaptive elitist
differential evolution for truss optimization with discrete variables. Comput
Struct 2016;165:5975. http://dx.doi.org/10.1016/j.compstruc.2015.11.014.
[30] Jena PK, Thatoi DN, Parhi DR. Differential evolution: an inverse approach for
crack detection. Adv Acoust Vib 2013;2013:110.
[31] Tang H, Xue S, Fan C. Differential evolution strategy for structural system
identification.
Comput
Struct
2008;86.
http://dx.doi.org/10.1016/
j.compstruc.2008.05.001.
[32] Donate JP, Li X, Snchez GG, Miguel AS. Time series forecasting by evolving
artificial neural networks with genetic algorithms, differential evolution and
estimation of distribution algorithm. Neural Comput Appl 2011;22:1120.
http://dx.doi.org/10.1007/s00521-011-0741-0.
[33] Fan H-Y, Lampinen J. A trigonometric mutation operation to differential
evolution. J Glob Optim 2003;27:10529. http://dx.doi.org/10.1023/
A:1024653025686.
[34] Das S, Abraham A, Chakraborty UK, Konar A. Differential evolution using a
neighborhood-based mutation operator. Evol Comput IEEE Trans
2009;13:52653. http://dx.doi.org/10.1109/TEVC.2008.2009457.
[35] Elsayed SM, Sarker RA, Essam DL. An improved self-adaptive differential
evolution algorithm for optimization problems. Ind Inf IEEE Trans
2013;9:8999. http://dx.doi.org/10.1109/TII.2012.2198658.
[36] Shi ZY, Law SS, Zhang LM. Structural damage localization from modal strain
energy change. J Sound Vib 1998;218:82544. http://dx.doi.org/10.1006/
jsvi.1998.1878.
[37] Gunes B, Gunes O. Structural health monitoring and damage assessment Part I:
a critical review of approaches and methods. Int J Phys Sci 2013;8:1694702.
http://dx.doi.org/10.5897/IJPS11.1327.
[38] Koh BH, Dyke SJ. Structural health monitoring for flexible bridge structures
using correlation and sensitivity of modal data. Comput Struct
2007;85:11730. http://dx.doi.org/10.1016/j.compstruc.2006.09.005.
[39] Chen X, Yu L. Flexibility-based objective functions for constrained
optimization problems on structural damage detection. Adv Mater Res
2011;186:3837. http://dx.doi.org/10.4028/www.scientific.net/AMR.186.383.
[40] Wang Y, Cai Z, Zhang Q. Differential evolution with composite trial vector
generation strategies and control parameters. IEEE Trans Evol Comput
2011;15:5566. http://dx.doi.org/10.1109/TEVC.2010.2087271.
[41] Qin AK, Huang VL, Suganthan PN. Differential evolution algorithm with
strategy adaptation for global numerical optimization. Evol Comput IEEE Trans
2009;13:398417. http://dx.doi.org/10.1109/TEVC.2008.927706.
[42] Jia G, Wang Y, Cai Z, Jin Y. An improved (muy + lamda)-constrained differential
evolution for constrained optimization. Inf Sci (Ny) 2013;222:30222. http://
dx.doi.org/10.1016/j.ins.2012.01.017.
[43] Padhye N, Bhardawaj P, Deb K. Improving differential evolution through a
unified approach. J Glob Optim 2013;55:77199. http://dx.doi.org/10.1007/
s10898-012-9897-0.
[44] Chan T, Samali B, Wang F, Chan T, Thambiratnam D, Tan A, et al. Correlationbased damage detection for complicated truss bridges using multi-layer

T. Vo-Duy et al. / Composite Structures 147 (2016) 4253

genetic algorithm. Adv Struct Eng 2012;15:693706. http://dx.doi.org/
10.1260/1369-4332.15.5.693.
[45] Khdeir AA, Reddy JN. Free vibration of cross-ply laminated beams with
arbitrary boundary conditions. Int J Eng Sci 1994;32:197180. http://dx.doi.
org/10.1016/0020-7225(94)90093-0.
[46] Ferreira JM, Castro LMS, Bertoluzza S. A high order collocation method for the
static and vibration analysis of composite plates using a first-order theory.

53

Compos
Struct
2009;89:42432.
http://dx.doi.org/10.1016/
j.compstruct.2008.09.006.
[47] Phung-Van P, Thai CH, Nguyen-Thoi T, Nguyen-Xuan H. Static and free
vibration analyses of composite and sandwich plates by an edge-based
smoothed discrete shear gap method (ES-DSG3) using triangular elements
based on layerwise theory. Compos Part B Eng 2014;60:22738. http://dx.doi.
org/10.1016/j.compositesb.2013.12.044.