You are on page 1of 17

ABSTRACT

This is a study regarding stress strain relationship of different types of


concretes. A tool called Artificial Neural Network (ANN) is used extensively
for the study. ANN is actually a mimic of human neural system, where a
number of neurons form a nervous system. ANN provided by deep learning
toolbox of MATLAB is used for studying. ANN is trained with approximately
1700 stress-strain input data’s and corresponding theoretical model and grade as
output. After training ANN it is capable of giving theoretical model and grade
for any experimental stress-strain data’s we provide. Depending upon How
much data points were used in training of ANN the accuracy of result will vary,
The more number of training points ,more accurate the result.

1
INTRODUCTION

What is ANN?

ANNs are artificial adaptive systems that are inspired by the functioning
processes of the human brain. They are systems that are able to modify their
internal structure in relation to a function objective. They are formed by
mimicking human nervous system. As like human nervous system is comprised
of many neurons that connects and forms a network, ANN is also a network of
artificial neurons called perceptron. They have high ability in pattern
recognition, hence very useful in studying data and creating difficult data’s.

Fig. 1 Structure of a single neuron

Application of ANN

The most typical problem that an ANN can deal with can be expressed as
follows: given N variables, about which it is easy to gather data, and M
variables, which differ from the first and about which it is difficult and costly to
gather data, assess whether it is possible to predict the values of the M variables
on the basis of the N variables. When the M variables occur subsequently in
time to the N variables, the problem is described as a prediction problem; when
the M variables depend on some sort of static and/or dynamic factor, the
2
problem is described as one of recognition and/or discrimination and/or
extraction of fundamental traits. To correctly apply an ANN to this type of
problem, we need to run a validation protocol. We must start with a good
sample of cases, in each of which the N variables (known) and the M variables
(to be discovered) are both known and reliable. When the M variables occur
subsequently in time to the N variables, the problem is described as a prediction
problem; when the M variables depend on some sort of static and/or dynamic
factor, the problem is described as one of recognition and/or discrimination
and/or extraction of fundamental traits. To correctly apply an ANN to this type
of problem, we need to run a validation protocol. We must start with a good
sample of cases, in each of which the N variables (known) and the M variables
(to be discovered) are both known and reliable.

Architectures of ANN
 Single layer feed forward architecture
 Multi-layer feed forward architecture
 Recurrent network architecture

Single layer feed forward neural network

 A layer of n neurons constitutes a single layer feed forward neural


network
 This is so called because, it contains a single layer of artificial neurons.
 In a single layer neural network, the inputs x1, x2,…………Xm are
connected to the layers of neurons through the weight matrix W.
 The output of any kth neuron can be determined as follows.

( )
m
Ok =f k ∑ ( W ik X i) + θk
i=1

θ k denotes the threshold value of the Kth neuron.

3
Fig. 2 Lay out of single layer feed forward neural network

Multi-layer feed forward architecture

 Made up of multiple layers.


 Thus architecture of this class besides processing an input and an output
layer also have one or more intermediary layers called hidden layers.
 The hidden layers aid in performing useful intermediary computation
before directing the input to the out put layer.
 If there are “l” number of neurons in the input layer, “m” number of
neurons in the hidden layer and “n” number of neurons in the out put
layer, It can be designated as l-m-n network.

4
Fig. 3 Multi-layer feed forward architecture

Recurrent network architecture

 The network differs from feedforward network architectures in the sense


that there is at least one feedback loop.
 There could also be neurons with self-feedback links, that is, the output of
a neuron is fed back into itself as input.
 Depending on different type of feedback loops, several recurrent neural
networks are known such as Hopfield network, Boltzmann machine
network etc..

5
Fig. 4 Recurrent network architecture

Advantages of ANN

 ANN’s exhibits mapping capabilities, that is, they can map input patterns
to their associated output pattern.
 The ANN’s posses the capability to generalize. This is the power to apply
in application where exact mathematical model to problem are not
possible.
 The ANN’s are robust system and fault tolerant. They can therefore,
recall full patterns from incomplete, partial or noisy patterns.
 The ANN’s can process information in parallel, at high speed and in a
distributed manner.

PROCEDURE
6
1. Creating data for training ANN
Our aim is to establish a relationship between stress strain characteristics and
model and grade of a concrete. For this ANN should be trained with many stress
strain data points corresponding to different grade of concrete and different
theoretical model. Four of the models used for the study is explained below.

a) KENT AND PARK MODEL

Fig. 5 Stress strain relationship for kent and park model

70

60 M15
50
M20
M25
40
M30
30 M33
Stress (MPa)

20

10

0
-10 0 10 20 30 40 50 60
-10

-20

-30 Strain

Fig.6 Kent and park model

b) POPOVICS MODEL
7
Fig. 7 Stress strain relationship for popovics model

70

60

50 M20
M25
40 M30
Stress (MPa)

M35
30 M40
M45
20 M50
m55
10
M60
0
0 10 20 30 40 50 60
Strain

Fig. 8 Popovics model

8
c) HOGNESTAD MODEL

Fig.9 Stress strain relationship for hognestad model

80

70

60
M20
50 M25
Stress (MPa)

M30
40 M33
M35
30 M40
M45
20 M50
m55
10 M60

0
0 10 20 30 40 50 60
Strain

Fig 10. Hognestad model

9
d) RUSCH MODEL

Fig. 11 Stress strain relationship for Rusch model

70

60

50

20
40 M25
Stress (MPa)

M30
M35
30 M40
M45
M50
20 M55
M60
10

0
0 0.0005 0.001 0.0015 0.002 0.0025 0.003 0.0035 0.004

Strain

Fig. 12 Rusch model

10
2) Training of ANN

From four models and different grade of concretes approximately 1700 data
points are made, now we should train the ANN with these data’s
 Training input data and out put data is uploaded in the work space of
MATLAB in the form of a matrix

 These matrixes are transposed, because ANN understand input data and
corresponding output data in that order.

 A neural network having sufficient “n” number of neurons in hidden


layer is created using the code “ net=feedforwardnet(n); ”

 Now the code for training the ANN is entered-

net=train(net,traininput,traintarget);
3) Testing data

 The data which is required to be tested is uploaded into workspace of


ANN

 Code for testing - y=net(testinput); is entered.

 Now test out put is stored into Y.

11
RESULTS

Training of ANN

Fig. 13 Training of ANN

Fig.14 Results of training

12
TESTING OF DATA

Fib.15 Test input provided

13
Fig.16 Test out put obtained from ANN

14
CONCLUSION

ANN is obviously a powerful tool for data studies. But in case of this study we
have not obtained expected result using ANN. This may be attributed to less
number of training data we have provided, or may be for this kind of study
(identifying model and grade of concrete from stress strain characteristics) using
ANN is not appropriate. There are many areas especially in image recognition
where ANN is dominating. How ever by conducting this experiment we were
able to understand the basic working of ANN and how to solve a problem using
ANN.

15
REFERENCE

 Enzo Grossia and Massimo Buscema,(2007). “Introduction to artificial


neural networks.” Journal of soft computing, European Journal of
Gastroenterology & Hepatology.

 Kevin Gurney University of Sheffield,(1997). “An introduction to neural


networks”, UCL Press Limited 11 New Fetter Lane London EC4P 4EE

 NPTEL :: Computer Science and Engineering - NOC:Introduction to Soft


Computing

16
APPENDIX

Computer programs used in ANN

traininput=traininput’; - Transposing the matrix


traintarget=traintarget’;
testinput=testinput’;
net=train(net,traininput,traintarget); - Training of ANN
net=feedforwardnet(10); - creating hidden layer having 10 neurons
y=net(testinput); - Finding out put from ANN
testoutput=y’;

17

You might also like