ISSN: 2277-3754

ISO 9001:2008 Certified

International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 1, July 2012

Introduction to Artificial Neural Network

A.D.Dongare, R.R.Kharde, Amit D.Kachare
Abstract: - This paper presents an emergence of an Artificial knowledge. Artificial Neural Networks are modeled closely
Neural Network (ANN) as a tool for analysis of different following the brain and therefore a great deal of
parameters of a system. An Artificial Neural Network (ANN) is terminology is borrowed from neuroscience.
an information-processing paradigm that is inspired by the way
biological nervous systems such as brain, process information.
II. LITERATURE REVIEW
ANN consists of multiple layers of simple processing elements
called as neurons. The neuron performs two functions, namely,
A.O. Kurban investigated an artificial neural network are
collection of inputs & generation of an output. Use of ANN non-linear mapping systems with a structure loosely based
provides overview of the theory, learning rules, and applications on principles observed in the biological nervous systems. In
of the most important neural network models, definitions and greatly simplified terms from, a typical real neuron has a
style of Computation. The mathematical model of network branching dentritic tree that collects signals from many
throws the light on the concept of inputs, weights, summing other neurons in a limited area; a cell body that integrates
function, activation function & outputs. Then ANN helps to collected signals and generates a response signal (as well as
decide the type of learning for adjustment of weights with manages metabolic functions); and along branching axon
change in parameters. Finally the analysis of a system is
that distributes the response through contacts with dentritic
completed by ANN implementation & ANN training & prediction
quality.
trees of many other neurons. The response of each neuron is
a relatively simple non-linear function of its inputs and is
Keywords: Biological Inspiration, ANN Methodology, ANN largely determined by the strengths of the connections from
Implementation and Prediction. its inputs. In spite of the relative simplicity of the individual
units, systems containing many neurons can generate
I. INTRODUCTION complex and intersecting behaviors. In general terms, a NN
Many tasks involving intelligence or pattern recognition consists of large number of simple processors linked by
are extremely difficult to automate, but appear to be weighted connections. By analogy, the processing nodes
performed very easily by humans. For instance, humans may be called “neurons”. Each node output depends only on
recognize various objects and make sense out of the large the information that is locally available at the node, either
amount of visual information in their surroundings, stored internally or arriving via the weighted connections.
apparently requiring very little effort. It stands to reason Each unit receives inputs from many other nodes and
that computing systems that attempt similar tasks will profit transmits its output to other nodes. By itself, a single
enormously from understanding how humans perform these processing element is not very powerful; it generates a
tasks, and simulating these processes to the extent allowed scalar output with a single numerical value, which is a
by physical limitations. This necessitates the study and simple non-linear function of its inputs. The power of the
simulation of Neural Networks. The neural network of an system emerges from the combination of many units in an
human is part of its nervous system, containing a large appropriate way. A network is utilized different function by
number of interconnected neurons (nerve cells). “Neural” is varying the connection topology and the values of the
an adjective for neuron, and “Network” denotes a graph like connecting weights. Complex functions can be implemented
structure. Artificial Neural Network refers to computing by connecting the units together with appropriate weights. It
systems whose central theme is borrowed from the analogy has been shown that a sufficiently large network with an
of biological neural networks. Artificial Neural Networks appropriate structure and property chosen weights can
are also referred to as “Neural Nets”, artificial neural approximate with arbitrary accuracy any function satisfying
systems “parallel distributed processing systems” and certain broad constraints. [1] This model is a drastically
“connectionist systems”. For a computing system to be simplified approximation of real nervous systems. The
called by these pretty names, it is necessary for the system intent is to capture the major characteristics important in the
to have a labeled directed graph structure where nodes information processing functions of real networks without
perform some simple computations. From elementary graph varying too much about the physical constraints imposed by
theory we recall that a “Directed Graph” consists of a set of biology. Artificial NN are made up of simple, highly
“Nodes” (vertices) and a set of “Connections” interconnected processing units called neurons, each of
(edges/links/arcs) connecting pairs of nodes. In a neural which performs two functions, namely, aggregation of its
network, each node performs some simple computations, inputs from other neurons or the external environment and
and each connection conveys a signal from one node to generation of an output from the aggregated inputs.
another, labeled by a number called the “Connection Through this simple structure, neural networks have been
Strength” or “Weight” indicating the extent to which a shown to be able to approximate most continuous functions
signal is amplified or diminished by connection. This to any degree of accuracy, by choice of an appropriate
system is the alternative for human expertise and

189
 ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 1, July 2012
number of neuron units (Kurban and Yildirim, 2003; from the combination of many units in an appropriate way.
Yildirim and Uzmay, 2003). [2] [1] The ANN does not really solve the problem in a strictly
mathematical sense, but it demonstrates information
processing characteristics that give an approximate solution
III. BIOLOGICAL INSPIRATION to a given problem. The ANNs have been widely used in
Human brain is made up of a network of neurons that are complex non linear function mapping, image processing,
coupled with receptors and effectors. Receptors are called pattern recognition & classification & so on. Feed-forward
“dendrites” and effectors are called “axons”.[3] Fig. 1 networks are common type of neural networks. A feed
shows that the dendrites collects the signals from many forward network comprises an input layer, where the inputs
other neurons in a limited area; a cell body or soma that of the problem are received, hidden layers, where the
integrates collected signals & generates a response signal & relationship between the inputs & outputs are determined &
along branching axon that distributes the response through represented by synaptic weights, & an output layer which
contacts with dendrite trees of many other neurons. [4] emits the outputs of the problem. The neural feed forward
network is modeled with three basic elements: a) A set of
synapses characterized by synaptic weights, b) An adder or
linear combiner for summing the input signals. c) An
activation function for limiting the amplitude of the output
of neuron to some finite value. The input of the activation
function can be increased by using a bias term. Here, we
have made use of a certain ANN architecture known as the
multi-layer-feed-forward neural network or Multi Layer
Perceptron (MLP)[5] .

Fig. 1 Biological Neuron

IV. ANN METHODOLOGY

ANNs are basically massive parallel computational
models that imitate the function of human brain. An ANN
consists of large number of simple processors linked by
weighted connections. By analogy, the processing nodes
may be called “neurons”. Each node output depends only on
the information that is locally available at the node, either
stored internally or arriving via the weighted connections.
Each unit receives inputs from many other nodes transmits
its output to yet another nodes. By itself, a single processing
element is not very powerful; it generates a scalar output Fig.2 Style of Neural Computation
with single numerical value, which is a simple non-linear
function of its inputs. The power of the system emerges
In above Fig.2 an input is presented to the neural network artificial neural networks, the designer chooses the network
and a corresponding desired or target response set at the topology, the performance function, the learning rule, and
output (when this is the case the training is called the criterion to stop the training phase, but the system
supervised). An error is composed from the difference automatically adjusts the parameters. So, it is difficult to
between the desired response and the system output. This bring a priori information into the design, and when the
error information is fed back to the system and adjusts the system does not work properly it is also hard to
system parameters in a systematic fashion (the learning incrementally refine the solution. But ANN-based solutions
rule). The process is repeated until the performance is are extremely efficient in terms of development time and
acceptable. It is clear from this description that the resources, and in many difficult problems artificial neural
performance hinges heavily on the data. If one does not networks provide performance that is difficult to match with
have data that cover a significant portion of the operating other technologies. Denker 10 years ago said that "artificial
conditions or if they are noisy, then neural network neural networks are the second best way to implement a
technology is probably not the right solution. On the other solution" motivated by the simplicity of their design and
hand, if there is plenty of data and the problem is poorly because of their universality, only shadowed by the
understood to derive an approximate model, then neural traditional design obtained by studying the physics of the
network technology is a good choice. This operating problem. At present, artificial neural networks are emerging
procedure should be contrasted with the traditional as the technology of choice for many applications, such as
engineering design, made of exhaustive subsystem pattern recognition, prediction, system identification, and
specifications and intercommunication protocols. In control.[6]

190
 ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 1, July 2012
Table 1. Terminology of Neuron network shown in Fig. 4 is a 3-2-3-2 feed forward network;
Biological Terminology ANN Terminology it contains three nodes in the input layer (layer 0), two
nodes in the first hidden layer (layer 1), three nodes in the
Neuron Node/Unit/Cell/Neurode
second hidden layer (layer 2), and two nodes in the output
Synapse Connection/Edge/Link layer (layer 3). These networks, generally with no more
than four such layers, are among the most common neural
Synaptic Efficiency Connection Strength/Weight nets in use, so much so that some users identify the phrase
“neural networks” to mean only feed forward networks.
Firing Frequency Node Output Conceptually, nodes in successively higher layers abstract
successively higher level features from preceding layers. In
A. Mathematical Model the literature on neural networks, the term “feed forward”
When creating a functional model of the biological has been used sometimes to refer to layered or acrylic
neuron, there are three basic components of importance. networks. [8]
First, the synapses of the neuron are modeled as weights.
The strength of the connection between an input and a
neuron is noted by the value of the weight. Negative weight
values reflect inhibitory connections, while positive values
designate excitatory connections [Haykin]. The next two
components model the actual activity within the neuron cell.
An adder sums up all the inputs modified by their respective
weights. This activity is referred to as linear combination.
Finally, an activation function controls the amplitude of the
output of the neuron. An acceptable range of output is Fig. 4 Feed Forward Networks
usually between 0 and 1, or -1 and 1. Mathematically, this
process is described in the figure, C. Neural Learning
It is reasonable to conjecture that neurons in an animal’s
brain are “hard wired.” It is equally obvious that animals,
especially the higher order animals, learn as they grow.
How does this learning occur? What are possible
mathematical models of learning? In this section, we
summarize some of the basic theories of biological learning
and their adaptations for artificial neural networks. In
artificial neural networks, learning refers to the method of
modifying the weights of connections between the nodes of
a specified network. Learning is the process by which the
random-valued parameters (Weights and bias) of a neural
network are adapted through a continuous process of
simulation by the environment in which network is
embedded. Learning rate is defined as the rate at which
network gets adapted. Type of learning is determined by the
Fig. 3 Mathematical Model manner in which parameter change takes place. Learning
may be categorized as supervised learning, unsupervised
From this model the interval activity of the neuron can be learning and reinforced learning. In Supervised learning, a
shown to be, teacher is available to indicate whether a system is
performing correctly, or to indicate a desired response, or to
validate the acceptability of a system’s responses, or to
indicate the amount of error in system performance. This is
(1) in contrast with unsupervised learning, where no teacher is
available and learning must rely on guidance obtained
heuristically by the system examining different sample data
The output of the neuron, Yk, would therefore be the
or the environment. Learning is similar to training i.e. one
outcome of some activation function on the value of Vk. [7]
has to learn something which is analogous to one has to be
B. Feed Forward Networks
This is a subclass of acrylic networks in which a trained. A neural network has to be configured such that the
connection is allowed from a node in layer i only to nodes application of a set of inputs produces (either 'direct' or via
in layer i+1 as shown in Fig.4. These networks are a relaxation process) the desired set of outputs. Various
succinctly described by a sequence of numbers indicating methods to set the strengths of the connections exist. One
the number of nodes in each layer. For instance, the way is to set the weights explicitly, using a priori

191
 ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 1, July 2012
knowledge. Another way is to 'train' the neural network by 3 .Reinforced Learning
feeding it teaching patterns and letting it change its weights Reinforcement Learning is type of learning may be
according to some learning rule. We can categorize the considered as an intermediate form of the above two types
learning situations in two distinct sorts. These are of learning. Here the learning machine does some action on
1. Supervised Learning the environment and gets a feedback response from the
Supervised learning or Associative learning in which the environment. The learning system grades its action good
network is trained by providing it with input and matching (rewarding) or bad (punishable) based on the environmental
output patterns. These input-output pairs can be provided by response and accordingly adjusts its parameters. Generally,
an external teacher, or by the system which contains the parameter adjustment is continued until an equilibrium state
neural network (self-supervised). Example: An occurs, following which there will be no more changes
archaeologist discovers a human skeleton and has to in its parameters. The self organizing neural learning may
determine whether it belonged to man or woman. In doing be categorized under this type of learning. [7]
this, the archaeologist is guided by many past examples of D. Back Propagation Network
male and female skeletons. Examination of these past The back propagation algorithm (Rumelhart and
examples (called the training set) allows the archaeologist McClelland, 1986) is used in layered feed-forward ANNs.
to learn about the distinctions between male and female This means that the artificial neurons are organized in
skeletons. This learning process is an example of supervised layers, and send their signals “forward”, and then the errors
learning, and the result of learning process can be applied to are propagated backwards. The network receives inputs by
determine whether the newly discovered skeleton belongs to neurons in the input layer, and the output of the network is
man or woman. given by the neurons on an output layer. There may be one
or more intermediate hidden layers. The back propagation
algorithm uses supervised learning, which means that we
provide the algorithm with examples of the inputs and
outputs we want the network to compute, and then the error
(difference between actual and expected results) is
calculated. The idea of the back propagation algorithm is to
reduce this error, until the ANN learns the training data.
The training begins with random weights, and the goal is to
adjust them so that the error will be minimal. [9] Back
propagation network has gained importance due to the
Fig. 5 Supervised Learning shortcomings of other available networks. The network is a
multi layer network (multi layer perception) that contains at
least one hidden layer in addition to input and output layers.
2 .Unsupervised Learning
Unsupervised learning or Self-organization in which an Number of hidden layers & numbers of neurons in each
(output) unit is trained to respond to clusters of pattern hidden layer is to be fixed based on application, the
within the input. In this paradigm the system is supposed to complexity of the problem and the number of inputs and
discover statistically salient features of the input population. outputs. Use of non-linear log-sigmoid transfer function
Unlike the supervised learning paradigm, there is no a priori enables the network to simulate non-linearity in practical
set of categories into which the patterns are to be classified; systems. Due to this numerous advantages, back
rather the system must develop its own representation of the propagation network is chosen for present work. [3]
input stimuli. Example: In a different situation, the Implementation of back propagation model consists of two
archaeologist has to determine whether a set of skeleton phases. First phase is known as training while the second
fragments belong to the same dinosaur species or need to be phase is called Testing. Training, in back propagation is
based on gradient decent rule that tends to adjust weights
differentiated into different species. For this task, no
and reduce system error in the network. Input layer has
previous data may be available to clearly identify the
species for each skeleton fragment. The archaeologist has to neurons equal in number to that of the inputs. Similarly,
determine whether the skeletons (that can be reconstructed output layer neurons are same in the number as number of
from the fragments) are sufficiently similar to belong to the outputs. Number of hidden layer neurons is deciding by trial
same species, or if the differences between these skeletons and error method using the experimental data. [10]
are large enough to warrant grouping them into different E.ANN Development & Implementation
In this work, both ANN implementation & training is
species. This is an unsupervised learning process, which
developed, using the neural network toolbox of Mat Lab.
involves estimating the magnitudes of differences between
Different ANNs are build rather than using one large ANN
the skeletons. One archaeologist may believe the skeletons
including all the output variables. This strategy allowed for
belong to different species, while another may disagree, and
better adjustment of the ANN for each specific problem,
there is no absolute criterion to determine who is correct.
including the optimization of the architecture for each
output.

192
 ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 1, July 2012
F. ANN Training & Prediction quality back propagation network. The network is capable of
One of the most relevant aspects of a neural network is predicting the parameters by experimental system. The
its ability to generalize, that is, to predict cases that are not network has parallel structure and fast learning capacity.
included in the training set. One of the problems that occur The collected experimental data such as speed, load, &
during neural network training is called over fitting. The values of pressure distribution etc. are also employed as
error on the training set is driven to a very small value, but training and testing data for an artificial neural network.
when new data is presented to the network, the error is The neural network is a feed forward three layered network.
large. The network has memorized the training examples, Quick propagation algorithm is used to update the weight of
but it has not learned to generalize to new situations. One the network during the training. The ANN has a superior
method for improving network generalization is to use a performance to follow the desired results of the system and
network that is just large enough to provide an adequate fit. is employed to analyze such systems parameters in practical
The larger network you use the more complex functions the applications.
network can create. There are two other methods for
VII. ACKNOWLEDGEMENT
improving generalization that are implemented in Mat Lab
The authors would like to acknowledge & thanks to Dr.
Neural Network Toolbox software: regularization & early
Y.R. Kharde, Principal, and Shree. Saibaba Institute of
stopping. The typical performance function used for
Engineering Research & Allied Sciences, Rahata, Prof. S.B.
training feed forward neural networks is the mean sum of
Belkar, Asso. Prof., P.R.E.C. Loni & Prof. R.R. Navthar,
squares of the network errors,
Asstt. Prof., P.D.V.V.P. COE Ahmednagar for their
immense help in this work.
(2)

It is possible to improve generalization, if you modify the REFERENCES

[1] A.O. Kurban, “Analysis of shafts surface pressures using
performance function by adding a term that consists of the
neural network”, Industrial Lubrication and Tribology,
mean of the sum of the squares of the network weights & (2004), Volume 56, No. 4, Page no. 217-225.
biases,
[2] Fazil Canbulut, Cem Sinanoglu and Sahin Yildirim, “Neural
msereg = λmse +(1-λ)msw, (3) network analysis of leakage oil quantity in the design of
partially hydrostatic slipper bearings”, Industrial Lubrication
Where λ is the performance ratio, & and Tribology, (2004), Volume 56, No. 4, Page no. 231-243
[3] Dr. R. R. Srikanth. “Application of ANN in Condition
Monitoring: A Case Study”, (Conference proceeding
(4) “Condition Monitoring of Mechanical System”) (2009),
Gitam University, Vishakhapattanam, Page no. 31-44.
Using this performance function causes the network to have
[4] R.R.Navthar & Dr. N.V. Halegowda, “Analysis of oil film
smaller weights & biases, & this force the network response thickness in Hydrodynamic Journal Bearing using Artificial
to be smoother & less likely to over fit. Once the different Neural Networks”, Ciit International Journal of Artificial
stages of the training process & the ANN structure had been Intelligent System & Machine Learning (Nov. 2011), Volume
determined, & before the optimization procedure is 3, No.12, Page no. -762-766
developed, it is important to estimate the ANN prediction
[5] M. Ananda Rao & J. Srinivas, “Dynamics of rotors supported
qualities. There is excellent agreement of predicted values on fluid film bearings using neural networks”, Centre for
& expected values. This close agreement shows that the System Dynamics & Condition Monitoring & Diagnostic
ANN can be used in the data analysis, of theoretical work to Studies, Page no.149-155
generate the missing data in the theoretical program. The
[6] Ghorbanian, M. Ahmadi, R. Soltani, “Design predictive tool
results of model ANN are compared with the hydrodynamic
& optimization of journal bearing using neural network
simulation data. [6] model & multi objective genetic algorithm”, Scientia Iranica
B (2011), Volume 18, No.5 Page no.-1095-1105
V. ANN FOR HYDRODYNAMIC JOURNAL
BEARING [7] “Neural networks: A requirement for intelligent systems”
The Artificial Neural Network can be used for prediction [8] Kishan Mehrotra, Chilkuri K. Mohan, Sanjay Ranka,
of pressure distribution in hydrodynamic journal bearing “Elements of Artificial Neural Networks”, Penram
which can be further used for stability analysis of International Publishing (India), (1997), Volume 1, Page no.-
hydrodynamic journal bearing. [11]. 1-41.
[9] Carlos Gershenson, ”Artificial Neural Network for
Beginners”
VI. CONCLUSION
As the ANN is an emerging technology it can be used for [10] R.R.Navthar & Dr. N.V. Halegowda, “Applications of
data analysis in applications such as pattern recognition, Artificial Neural Network in Pressure Distribution Analysis
prediction, system identification & control. From above of Hydrodynamic Journal Bearings”, ASME Digital Library,
e-books, International Conference on Computer Technology
theories it can be seen that ANN is a radial basis function

193
 ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 1, July 2012
& Development (2011), (Third ICCTD 2011) , Page no. -
717-722
[11] R.R.Navthar & Dr. N.V. Halegowda, “Pressure Distribution
Analysis of Hydrodynamic Journal Bearings using Artificial
Neural Network”, ASME Digital Library, e-books,
International Conference on Computer & Automation
Engineering, (Fourth ICCAE 2012), Page no. -153-161

AUTHOR BIOGRAPHY

Prof.A.D. Dongare
M.E. (Design Engg), Ph.D.(App)
P.R.E.C. Loni Rahata Ahmednagar.
Area of Research: Design and Tribology.
Professional Membership: IE (I).

Prof.R.R. Kharde
M.E. (Tribology), Head, Dept. of
Mechanical Engineering.
P.R.E.C. Loni Rahata Ahmednagar.
Area of Research: Design and Tribology.
Professional Membership: IE (I).

A.D. Kachare

B.E. (Mechanical), P.G. Student.

P.R.E.C. Loni Rahata Ahmednagar.
Area of Research: Design.
Professional Membership: IE (I).

194