Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

Buckling Analysis of Cracked Composite aluminum column Using Neural

Network
A. Nezam Abadi1, M. Nasrollahi2*, I. Kalantari3, S. I. Miri Kerahroudi4
1
Assistant Prof, Department of Mechanical Engineering, Islamic Azad University, Arak Branch, Arak, Iran, a-nezamabadi@iau-arak.ac.ir.

Abstract: Acquiring buckling critical load by the means of complex equations which is expecting to be finally solved
by proving or eliminating various parameters via numerical or analytical methods is exclusively limiting our freedom to
alter problem‘s variables. We designed a neural network to model buckling of a cracked composite aluminum. Network
training data (inputs and outputs) is obtained from analytical calculations. To train the network, Levenberg-Marquardt
algorithm has been used. Then the network is tested using the crack angle, crack depth, fiber volume and crack place as
input variables. Hence a neural network can be used to analyze this type of buckling process and represent respective
inputs which determine problem condition in order to produce output network which is the same critical load. The
network manufactured for this process is a three-layer perceptron network (with a hidden layer) with an error back
propagation rule. To implement this network, MATLAB software has also been used. Researches show that only three
layers are needed; it has been verified that the three-layer perceptron with Sigmoid Output Function are comprehensive
approximations. These perceptron can be trained for the approximation of any kind of function in which the accuracy of
the applied approximation depends on the number of neurons in the hidden layer.

Key words: Buckling, Perceptron, Error Back Propagation, Neural Network, Composite Column, Levenberg-
Marquardt Algorithm

2
(C. Author.) Ph.D. Student in Mechanical Engineering Solid Design, Iran Aluminium Company (IRALCO), Arak, Iran, nasr.milad@gmail.com
3
M.Sc. Mechanical Engineering Solid Design, Iran Aluminium Company (IRALCO), Arak, Iran, iman1kalantary@gmail.com
4
Lecturer, Photography Department, Honar University, Qom, Iran, imanmiri@mail.com
 Proceedings of Iran International Aluminum Conference (IIAC2018)
April 24-25, 2018, Tehran, I.R. Iran
Introduction
In the general structure, artificial neural networks have
been inspired from the similar biology network [1]. In
the case that there is not required rule set for solving
the problem or knowing the phenomenon is so
complicated, it provoked scientists to invent the use of
an artificial intelligence system which has the ability of
human’s learning, creativity and flexibility [2]. A brief
investigation about mechanical fracture of composite
having one-sided cracks has been mentioned in the
paper written by Nikpour et.al. (1990) [3]. In this
regard, neural calculative methods were presented
which are fundamentally based on the gradual upgrade
of the system [4]. In recent decades, neural networks
have been considerably applied approximately in all of
sciences such as Aerospace, Automotive, Defense, Oil
and Gas, Robot, Electric and Electron. Neuron as the
smallest data processing unit in an artificial nerve
forms the basis of its function. Combining some
neurons forms a cell. Depending on its kind, a cell has
a specific duty in the network. How to connect neural
cells in different layers determines the network
structure which is called network architecture [5].
Researchers in Illinois University applied the neural
networks to analyze concrete pavement systems of the
airport and he could propose an analytical tool based
on the neural network in which extremely advanced
finite element solutions were drawn by a simple
method and simultaneously it runs several times faster
than the complex finite element programs. Roserio et
al [6] have researched the neural networks to detect
damages to composite layer sheets. They took
advantage of piezoelectric sensors and the neural
network to detect damages of composite layer sheets.
Sheidai et al [7] have assessed buckling and post-
buckling behavior of the columns using the neural
network. They could obtain desirable predictions for
buckling and post-buckling behavior of the columns
using Levenberg-Marquart algorithm.. Researches
show that only three layers are required. In the
reference [6], it is certified that 3-layer perceptron with
Sigmoid Output Function are comprehensive
approximations. These perceptron can be trained for
the approximation of any kind of function in which the
accuracy of the applied approximation depends on the
number of neurons in the hidden layer.
Designing neural network
Neural networks due to high velocity and satisfactory
accuracy to foresee the proper relationship between
input and output variables especially in solving
complex nonlinear problems have been considerably
regarded by the researchers and they aim to determine
passive function of the problem using lots of simple
calculative elements which have been connected via a
lot of connections [8]. At first, need for establishing
database of required data, preprocessing and
information distribution in order to analyze the
relationship between parameters is discussed. It is
transparent that the accuracy of the model will be
increasing by the number of dataset. The neural
network has three main stages. In the first step, input
and output parameters of the network are defined using
respective data certified by other related and reliable
papers. The second step includes network training
process which is happening in the hidden layers. The
network is learning to find the effect of input data on
output data with a particular order. After estimating
output data, in the last step, tentative data and the
network results are compared and based on the rate of
adaptability of these parameters, the accuracy is
calculated. It is obvious that the higher the accuracy of
the neural network is, the closer its output to the
tentative data and more acceptable usage is to predict
system performance [9].
Training data
Among training data, mostly a subset of them is set
aside to examine the network performance and they are
not used in training. After training the network, the
ability of the network to hold an intended map between
the input and output space of variables with
investigating its performance to respond these new
inputs were measured, and error rates were calculated.
This step is known as the verification step [8].
The number of the hidden layers
Multilayer perceptron can carry out a complicated
classification via applying the ample number of
perceptron layers in the network and desirable number
of perceptron in each layer. To process input data
which are analytical, more than two hidden layers are
not needed and the accuracy of the approximation is
controlled by the number of nerves in each layer not by
the number of the network layers [10]. At the end, it
can be concluded that applying two hidden layers is
not necessarily the best alternative and using more
layers can lead to the less using of the nerves in the
whole network [11]. So as to determine the structure of
the neural network for a specific usage, the number of
neurons in each layer and also the number of its hidden
layers must be specified. The number of the input and
output neurons of the network are easily determined
 Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

with respect to specification of the respective problem
so that the number of input and output neurons is equal
to the number of input and output variables of the
problem, receptively [12]. Speaking of the number of
the hidden layers, it is theoretically proven that it is
enough to only have a hidden layer with nonlinear
activity functions to create any kind of nonlinear map
with any desirable accuracy between input and output
space of variables [13]. Nevertheless, determining the
number of the neurons of a hidden layer is not as easy
as the determining the number of the neurons of input
and output layers.
column’s manufacturer material were optional. In fact,
the main purpose of column analysis is to display the
changes rate of the critical load compared to four
aforementioned factors but not material.
Critical load changes in ratio with each of factors after
column analysis are as the graphs which have been
brought in the following. These graphs all have been
obtained from numerical analytical calculations [15].
2D Graph 2
1.0

a=0.4 b=0.5 8=45

0.8
In most of the cases, there is not a specific method to
obtain the precise number of the hidden neurons expect 0.6

of training the networks with different hidden neurons

Pcr
0.4
and then estimating the general error (Error in the
verification step) of each. If a network does not 0.2

converge well, it might need more hidden neurons and

0.0
if a network converges well, less hidden neurons can
be tested and according to the whole performance of 0.0 0.2 0.4 0.6 0.8 1.0

system, it can be agreed over its structure. The main Fibr volume

idea is that as much as possible, the least number of

hidden neurons is used and also the error of the Fig 2: Critical load changes in ratio with fiber
verification step is acceptable, because each hidden volume
As it is observed, the critical load quantity is linearly
neuron increases calculative load and causes more
increasing with an increase in the quantity of
complexity.
composite fibers.
Ultimately, as a total conclusion, it can be uttered that
2D Graph 3
determining the number of the neurons of the hidden
layer is empirical and requires trial and error [14].

Architecture and total form of the designed neural Pcr for a=0.4 8=45 v=0.41
network is according to the following fig (1). 446.008

446.006
Pcr

446.004

Fig 1: Schematic design of Neural Network.

446.002

Modeling buckling behavior of the composite

column by analytical method 446.000
0.0 0.2 0.4 0.6 0.8 1.0 1.2
In the buckling process, it aims to acquire critical load
the place for crak
P Fig 3: Critical load changes in ratio with crack
quantity cr =0. In this column due to having crack and
fibers, critical load quantity depends on four factors of place
crack height, crack depth, fiber volume and fiber angle
In above fig, the exact quantity of the load is slightly
with Cartesian coordinate system. Thus, to design the
varying from place to place. Because its unit in based
neural network, we have 4 inputs which impact on
on kilo newton, this trivial difference cannot be
single output of the problem. To analyze this column,
detected but is considered in drawing the graph.
no specific material has been sought and all of
mechanical features which are attributed to the Pcr
Horizontal vector also doesn’t reflect =0. This is
 Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

concluded from the fig (3) that the critical load Due to the factors affecting on the critical load quantity
quantity doesn’t much depend on crack place. which have been investigated in above figs, network
design will be examined in the following.
2D Graph 4

1200

1000 Network design

Pcr for a=0.4 b=0.5 V=0.41 After writing respective program for network
800
manufacturing, 148 quadric inputs are represented for
600 training the network. In this section, the network
P cr

accuracy is enhancing by changing in the number of

400
200
0
0 10 20
Angle Fibrs
Fig 4: Critical load changes in ratio with fiber angle
Here there is not any specific trend between rise in
fiber angle and critical load quantity. That’s why,
detecting the relationship between fiber angle and
critical load quantity is so hard for the network and as
we will see in the future, the highest network error
occurs in this part in which fiber angle is changing.
neurons of the hidden layer, the number of training
courses and or the amount of final layer. This action is
a pragmatic one which is carried out to reach the
optimal manner. As it was aforementioned, more
30 40 50
number of neurons in the hidden layer doesn’t
necessarily mean a rise in network accuracy and an
escalation in the number of training courses also leads
to excessive adjustment of the network. In fact, to
design a proper network for any kind of function, a sort
of tradeoff between various factors must be achieved.
Training data
20 data which have been proposed to be trained by 148
network training data are detailed as the table (1)

2D Graph 5 Table 1: Training data for designing a neural

network
261.0012
Inputs Output
Data No

V b Θ a
Pcr
261.0010

(Fiber (crack (Fiber (crack

261.0008 Volume) Place) Angle) depth) (load)
261.0006 P cr for V=0.41 b=0.5 8=45 1 0 0.5 45.38 0.4 0.0012
Pcr

2 0.2 0.5 45.38 0.4 0.0222

261.0004
3 0.5 0.5 45.38 0.4 0.0536
261.0002 4 0.9 0.5 45.38 0.4 0.0975
5 0.41 0 45.38 0.4 0.0446
261.0000
6 0.41 0.2 45.38 0.4 0.0446
260.9998
7 0.41 0.5 45.38 0.4 0.0446
0.0 0.2 0.4 0.6 0.8 1.0 8 0.41 0.9 45.38 0.4 0.0446
Depth fo crak 9 0.41 0.5 0 0.4 0.0027
Fig 5: Critical load changes in ratio with crack 10 0.41 0.5 5 0.4 0.0367
depth 11 0.41 0.5 10 0.4 0.0026
12 0.41 0.5 15 0.4 0.0030
In above fig like fig (4) also, there is not a slight 13 0.41 0.5 20 0.4 0.0228
14 0.41 0.5 30 0.4 0.0453
difference between critical loads. The important
15 0.41 0.5 40 0.4 0.0046
conclusion that has been drawn from this section is that 16 0.41 0.5 45 0.4 0.0114
critical load quantity is reducing faster if crack depth 17 0.41 0.5 45.38 0 0.0026
passes the half of the whole column and before that, 18 0.41 0.5 45.38 0.2 0.0026
drop in critical load quantity is not tangible. This 19 0.41 0.5 45.38 0.5 0.0026
section that means passing the half of depth ratio 0.5 20 0.41 0.5 45.38 0.9` 0.0026

must be considered in fracture mechanic. It is quoteworthy that the amount of crack place and
depth has been normalized and it is calculated in ratio
with the entire column. For example, b= 0.4 indicates
 Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

that the crack is far from the top of the column as Hence, to gain an optimal network, the correlation
much as 0.4 of the column height. between output data from the neural network with
target data was evaluated.
Load quantity after multiplying by 10 is based on
Meganewton. This unit was chosen because of input
smallness, creating coordination into the network
between inputs and outputs and also increasing
network accuracy and its better convergence.

Network implementation
Two different programs in order to implement neural
network for relevant buckling problem have been
written using MATLAB software which both have
been brought in the index.

Here, we compare two programs and the comparison

of their outputs with desirable outputs will be taken
into the consideration. As it was mentioned before, in
both programs, there is a three-layer perceptron
network (with a hidden layer) with an error back
propagation. In fact, back propagation rule is an old
and common rule in designing perceptron networks.
Perception is usually used in engineering and technical Fig 6: Desirable and real output graph at the
applications, and other kinds of neural networks such expense of the optimal neural network
as Kohonen and Hopfield are employed in the contexts
specified for the computer and electricity fields. In above chart, the stretch line represents the real
output and the spots represent the desirable output. As
Owing to the constant number of neurons of the first it is seen, these two lines have a relatively good
and third layers, only neurons of the hidden layer can similarity and this issue reflects the relatively good
be changed. This work has been done and one network responsiveness of the network.
with different numbers of hidden neurons has been
tested, and obtained results have been compared. X value in above graph is regression and is calculated
based on the following formula.
As it is observed, at first, desirable input and output
)1
  T   o 
N
matrices are inserted into the network. Because of this N   Ti oi  
N N
i 1 i i 1 i
fact that our inputs are 148 quadric data, matrix P is
r i 1

N  T     T 
N
also a 4*146 matrix. It means that the matrix has 4 2 N
i i 1 i
rows and 146 columns and each column indicates one i 1

input data and because in each section, single output of

the matrix t is a line matrix with 150 columns, each
column in matrix t represents desirable output
proportionated to the same column in the matrix p.
Then, we create our desirable network using newff
command. Newff command builds a feed-forward
perceptron network. In this program, Tangent Sigmoid
Stimulation Function has been defined for the hidden
layer and for the output layer, linear function has been
defined. Opting for a linear function for the output
layer is a common matter in designing neural network.
Hence, to reach an optimal network, for all of designed
models, different quantities of Root Mean Square Error
(RMSE), Mean Absolute Error (MAE), Coefficient of
correlation(R2), Mean Absolute Percentage Error
(MAPE) have been calculated using below equations
[37,38] and compared in each step. Results are
obtained as the following graphs. Obtained results
have been depicted in fig 2 which have been compared
with each other.
 Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

1 N
1 N M (2 MSE
RMSE 
N
 (O
i 1
i  Ti ) 2 MSE    (T k i  O k i )
mn K 1 i 1
N

1 N
(T  Oi )  (O i  Oi ) 2 20000

MAPE    i  100 R  1
2 i 1
N
18000
N i 1 Ti
 (O )
i 1
i
2 16000
Col 6 vs Col 4
14000

Mean square Error

12000

In the relation (2), parameter m indicates output 10000

neurons and N is the number of training examples. 8000

6000

The best performance of the tested networks and their 4000

2000
proposed answers have been expressed in table (2)
0

Table 2: Calculated values for cracked composite

50 -60- 70- 80

7- 20- 35- 35

8 -26- 18- 11

8- 22- 39- 30

7- 25- 43- 46

5 -20 -35 35

9 -27- 21 -13

6- 25 -15- 10

8- 27- 20- 12

25- 37 -14- 33

25- 35- 15- 30

25- 34- 15- 31

10- 25- 39 -35

10- 25- 42- 40

8- 15- 25- 11
column according to the best results
NETWORK The Network geomatry
STRUCTURE R MSE RMSE µ 
(INPUT NEURONS) Fig 7: Mean Square Error based on inputs
5 -20 -35 35 0.99218 554.347 23.5446 2.3272 23.5099

7 -20 – 35- 35 0.97943 1438.0815 37.92 -1.34 38.0288

1.05
7- 25- 43- 46 0.99089 650.1165 25.4974 0.56065 25.579

8- 22- 39- 30 0.9897 710.4605 26.45 1.6929 26.6922

1.00

10 -25 -39 -35 0.83417 14503.466 120.4303 15.191 119.879

0.95
10- 25- 42- 40 0.97545 1544.1747 39.294 4.7507 39.144
Regrasion

Col 6 vs Col 5
25 -37- 14- 33 0.99983 11.4833 3.3887 0.19935 3.3975 0.90

6 -25- 15- 10 0.99796 145.755 12.0729 -0.45012 12.106

0.85
8 -15- 25 -11 0.99542 323.667 17.9949 1.035 18.027

8- 26 -18 -11 0.97023 2086.778 45.6813 3.2924 45.7193 0.80

8- 27 -20- 12 0.9983 114.933 10.72 -0.156 10.756

0.75
50 -60- 70- 80

7- 20- 35- 35

8 -26- 18- 11

8- 22- 39- 30

7- 25- 43- 46

5 -20 -35 35

9 -27- 21 -13

6- 25 -15- 10

8- 27- 20- 12

25- 37 -14- 33

25- 35- 15- 30

25- 34- 15- 31

10- 25- 39 -35

10- 25- 42- 40

8- 15- 25- 11
9 -27- 21- 13 0.9922 548.249 23.41 0.4146 23.49

130.855
50- 60 -70- 80 0.7964 17978.24 134.083 -31.184
1

25 -35- 15 -30 0.99995 3.7371 1.9339 -0.0085009 1.94

The Network Geometry
25 -34- 15- 31 1 0.16021 0.40026 -0.0003453 0.40199

Fig 8: Comparing regression for the neural network

In the fig 6, Mean Square Error obtained form training
15 neural networks with the 3-layer network structure It is witnessed that Mean Square Error and also
and with altercation in the neurons of the hidden layer regression are reaching desirable values and the best
has been propounded. Also, fig 7 has displayed the condition between above networks can be easily
total regression obtained from these networks. By specified. Regression lines have been drawn in order to
examining them, we arrived at this conclusion that a display the rate of overlapping of obtained quantities.
network having (25-34-15-31) neurons seem suitable With respect to fig 8, what expected from the selected
for this dissertation. network is that the quantities obtained from the
software that is obtained from finite elements have
adapted to the quantities obtained from the neural
network. The most important advantage is lack of need
for any kind of complex and time-consuming input
file. Similar to finite elements methods, this method
also provides a dramatic reduction in the required time
for calculating each analysis.
 Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

2D Graph 2

250

DLtta IN ANN=-20
200 DLttaIN EXP=20

DLtta IN ANN=15
DLtta IN exp=15

Pcr (Mpa)
150

100

50

0
4 5 6 7 8 9 10 11

1/ Volom af fibres
Fig 10: Comparing critical load obtained from the

neural network using numerical methods for two fiber

angles of 15 and 20 degrees

Conclusion
It can be concluded from the example used to draw a
neural network in this dissertation that if there is
enough data in an engineering problem, a neural
network can be drawn for that problem. So, after
passing a training course and being ensure of the
network performance, for the new data, instead of
using analytical calculations and or using other
analytical programs, it gives us intended output. The
results obtained from this research can be summarized
as the following:
1. The type of proposed network in order to estimate
buckling behavior of the cracked composite
column is a 3-layer neural network with
Reinforcement Learning Algorithm and
behavioral results are very consistent with
numerical works.
Fig 9: Comparing graphs for the selected networks 2. 2-Designing the neural network can be applied to
(25-34-15-31) model different buckling of the cracked
composite sheets with various directions and
After training the neural network, desirable output can various fabrics in different material modes.
be received from the network by introducing new 3. Taking advantage of hyperbolic functions as a
output vector. Input vector can include each of possible transfer function in the neural network in order to
variables of the inputs or a proportion of each input express nonlinear behavior of the buckling of the
and buckling critical load is obtained in output. For column articulates close-to-reality answers.
verification, we compare 2 different fiber angles of 20 4. As it was observed in some of the cases, the
and 15 degrees by the numerical method and the neural expansion of input neurons causes escalation in
network method. As it was seen, the answers calculation error and we must always look for the
converged towards each other with good most optimal possible mode.
approximation.
 Proceedings of Iran International Aluminum Conference (IIAC2018)

April 24-25, 2018, Tehran, I.R. Iran

References; [11] M.T. Hagan, H.B. Demuth and M. Beale, Neural

network design, Boston: PWS Publishing com.,
[1] N. Kartam, I. Flood, J. Garrett, Artificial neural 1996, p.161.
networks for civil engineers: fundamentals and [12] J.A. Freeman and D.M. Skapura, Neural
applications, ASCE, New York, 1997, p. 216. Networks: algorithms, application, and
[2] J. McCarthy, What is Artificial Intelligence, programming techniques, Addision-Wesely,
Stanford University, 2004, p.15. Reading. MA, 1992, p. 45.
[3] K. Nikpour, Buckling Of Cracked Composite [13] J.G. Polhill, M.K. Weir, “An approach to
Columns, Int. J. Solids Structures, V. 26, No. 12, Guaranteeing Generalization in Neural
1990, pp. 1371-1386. Networks”, Neural Networks, 2001, pp.1035-
[4] P. Ramchandra Baviskar, V.B. Tungikar, 1048.
Multiple Cracks Assessment using Natural [14] J.F. Briceno, H.EI- Mounayri and S.
Frequency Measurement and Prediction of Crack Mukhopadhyay, Selecting and artificial neural
Properties by Artificial Neural Network, network for efficient modeling and accurate
International Journal of Advanced Science and simulation of the milling process”, International
Technology Vol. 54, 2013. Journal of Machine Tools & Manufacture 42
[5] M. Y. Rafiq, G. Bugmann, D.J. Easterbrook, (2002) pp.663-674.
Neural network design for engineering [15] A. Kumar, V. Kumar, Numerical Analysis of
application, Computers & Structures 79(17), Crack Effect on Natural Frequency of Cantilever
2001, p.p. 1541–1552. Composite Beam, International Journal of
http://dx.doi.org/10.1016/S0045-7949(01)00039- Engineering Research (IJOER) [Vol-1, Issue-3,
6 June- 2015].
[6] L. Roserio , U. Ramos, R. Leal, Neural Networks
in Damage Detection of Composite Laminated
Plates, Proceedings of the 6th WSEAS Int. Conf.
on Neural Networks, Lisbon, Portugal, June 16-
18, 2005 (p.p. 115-119).
[7] M.R. Sheidaei, R. Bahrami Nejad, Evaluating
Buckling Analysis of Columns Using Neural
Network, to be published.
[8] K. M. Saridakis, A.C. Chasalevris, C.A.
Papadopoulos, A.J. Dentsoras, Applying neural
networks, genetic algorithms and fuzzy logic for
the identification of cracks in shafts by using
coupled response measurements, Computers and
Structures 86 (11–12), 2008, p.p. 1318–1338.
http://dx.doi.org/10.1016/j.compstruc.2007.08.00
4
[9] M. Y. Rafiq, G. Bugmann, D.J. Easterbrook,
Neural network design for engineering
application, Computers & Structures 79(17),
2001, p.p. 1541–1552.
http://dx.doi.org/10.1016/S0045-7949(01)00039-
6
[10] J. Ghaboussi, J.H. Garrett, X. Wu, Knowledge-
based modeling of material behavior with neural
networks, Journal of Engineering Mechanics
ASCE 117(1), 1991, p.p. 132–153.
http://dx.doi.org/10.1061/(ASCE)0733-
9399(1991)117:
1(132)