Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 MATERIAL AND MANUFACTURING TECHNOLOGY SERIES

ARTIFICIAL INTELLIGENCE IN
MANUFACTURING RESEARCH

No part of this digital document may be reproduced, stored in a retrieval system or transmitted in any form or
by any means. The publisher has taken reasonable care in the preparation of this digital document, but makes no
expressed or implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of information
contained herein. This digital document is sold with the clear understanding that the publisher is not engaged in
rendering legal, medical or any other professional services.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 MATERIAL AND MANUFACTURING
TECHNOLOGY SERIES
J. Paulo Davim, Editor

Tribology Research Advances

J. Paulo Davim
2009. ISBN: 978-1-60692-885-1

Drilling of Composite Materials

J. Paulo Davim (Editor)
2009. ISBN: 978-1-60741-163-5

Drilling of Composite Materials

J. Paulo Davim (Editor)
2009. ISBN: 978-1-60876-584-3 (Online book)

Artificial Intelligence in Manufacturing Research

J. Paulo Davim (Editor)
2010. ISBN: 978-1-60876-214-9
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 MATERIAL AND MANUFACTURING TECHNOLOGY SERIES

ARTIFICIAL INTELLIGENCE IN
MANUFACTURING RESEARCH

J. PAULO DAVIM
EDITOR
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Nova Science Publishers, Inc.

New York

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Copyright © 2010 by Nova Science Publishers, Inc.

All rights reserved. No part of this book may be reproduced, stored in a retrieval system or
transmitted in any form or by any means: electronic, electrostatic, magnetic, tape, mechanical
photocopying, recording or otherwise without the written permission of the Publisher.

For permission to use material from this book please contact us:
Telephone 631-231-7269; Fax 631-231-8175
Web Site: http://www.novapublishers.com

NOTICE TO THE READER

The Publisher has taken reasonable care in the preparation of this book, but makes no expressed or
implied warranty of any kind and assumes no responsibility for any errors or omissions. No
liability is assumed for incidental or consequential damages in connection with or arising out of
information contained in this book. The Publisher shall not be liable for any special,
consequential, or exemplary damages resulting, in whole or in part, from the readers’ use of, or
reliance upon, this material. Any parts of this book based on government reports are so indicated
and copyright is claimed for those parts to the extent applicable to compilations of such works.

Independent verification should be sought for any data, advice or recommendations contained in
this book. In addition, no responsibility is assumed by the publisher for any injury and/or damage
to persons or property arising from any methods, products, instructions, ideas or otherwise
contained in this publication.

This publication is designed to provide accurate and authoritative information with regard to the
subject matter covered herein. It is sold with the clear understanding that the Publisher is not
engaged in rendering legal or any other professional services. If legal or any other expert
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

assistance is required, the services of a competent person should be sought. FROM A

DECLARATION OF PARTICIPANTS JOINTLY ADOPTED BY A COMMITTEE OF THE
AMERICAN BAR ASSOCIATION AND A COMMITTEE OF PUBLISHERS.

LIBRARY OF CONGRESS CATALOGING-IN-PUBLICATION DATA

Artificial intelligence in manufacturing research / J. Paulo Davim.

p. cm.
Includes index.
ISBN 978-1-61761-564-1 (E-Book)
1. Manufacturing processes--Automation. 2. Manufacturing processes--Research. 3. Computer
integrated manufacturing systems. 4. Artificial intelligence. I. Davim, J. Paulo.
TS183.A78 2009
670.285'63--dc22
2009031035

Published by Nova Science Publishers, Inc. 

  New York

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 CONTENTS

Preface vii
Research and Review Studies
Chapter 1 Application of Neural Networks and Fuzzy Sets to Machining and
Metal Forming 1
U. S. Dixit
Chapter 2 Multi-Objective Optimization of Multi-Pass Milling Process
Parameters Using Artificial Bee Colony Algorithm 31
R. Venkata Rao and P. J. Pawar
Chapter 3 Optimization of Abrasive Flow Machining Process Parameters
Using Particle Swarm Optimization and Simulated Annealing
Algorithms 51
P. J. Pawar, R. Venkata Rao and J. P. Davim
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Chapter 4 Study of Effects of Process Parameters on Burr Height in Drilling of

AISI 316 Stainless Steel Using Artificial Neural Network Model 65
V. N. Gaitonde, S. R. Karnik and J. Paulo Davim
Chapter 5 Artificial Neural Network Modeling of Surface Quality
Characteristics in Abrasive Water Jet Machining of Trip Steel Sheet 79
N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos
Chapter 6 Multi-Objective Optimisation of Cutting Parameters for Drilling
Aluminium AA1050 101
Ramón Quiza and J. Paulo Davim
Chapter 7 Application of Fuzzy Logic in Manufacturing: A Study on
Modeling of Cutting Force in Turning GFRP Composites 111
K. Palanikumar and J. Paulo Davim
Chapter 8 Flank Wear Detection with AE Signal and FNN During Turning of
Al/15 Vol%Sic-MMC 129
Alakesh Manna

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 vi Contents

Chapter 9 Integration of Product Development Process Using STEP and PDM 141
S. Q. Xie and W. L. Chen
Index 175
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 PREFACE
Artificial intelligence is a subfield of computer science concerned with understanding the
nature of intelligence and constructing computer systems capable of intelligent action.
Artificial intelligence can be applied to all systems and manufacturing processes.
This book aims to provide the research and review studies on artificial intelligence in
manufacturing. The first chapter provide information on application of neural networks and
fuzzy sets to machining and metal forming. Chapter 2 discuss multi-objective optimization of
multi-pass milling process parameters using artificial bee colony algorithm. Chapter 3 is
focused on optimization of abrasive flow machining process parameters using particle swarm
optimization and simulated annealing algorithms. Chapter 4 discuss the effects of process
parameters on burr height in drilling of AISI 316 stainless steel using artificial neural network
model. Chapter 5 is focused on artificial neural network modeling of surface quality
characteristics in abrasive water jet machining of trip steel sheet. Subsequently, the chapter 6
deal with multi-objective optimization of cutting parameters for drilling aluminium AA1050.
The chapter 7 is dedicated application of fuzzy logic in manufacturing, a study on modeling
of cutting force in turning GFRP composites. The chapter 8 is dedicated on flank wear
detection with AE signal and FNN during turning of Al/15vol%SiC-MMC. Finally, the last
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

chapter of this research book is focused on integration of product development process using
STEP and PDM.
The present research book can be used as a text book for final undergraduate engineering
course (for example, mechanical, manufacturing, systems, etc) or as a subject on artificial
intelligence in manufacturing at the postgraduate level. Also, this book can serve as a useful
reference for academics, manufacturing and computational sciences researchers, mechanical,
systems and manufacturing engineers, professional in related industries with artificial
intelligence and manufacturing.
The Editor acknowledge their gratitude to Nova Publishers for this opportunity and for
their professional support. Finally, I would like to thank all the chapter authors for their
availability for this work.

J. Paulo Davim
Aveiro, Portugal
June 2009

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 RESEARCH AND REVIEW STUDIES
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 1

APPLICATION OF NEURAL NETWORKS AND FUZZY

SETS TO MACHINING AND METAL FORMING

U. S. Dixit∗
Department of Mechanical Engineering,
Indian Institute of Technology Guwahati 781 039, INDIA

ABSTRACT
Machining and metal forming are two major manufacturing processes. Neural
networks and fuzzy sets have been employed in the modeling, optimization and control of
these processes. This chapter describes how these soft computing tools can be used in the
areas of machining and metal forming starting from the brief background of the methods.
It is observed that neural networks are very effective tools of learning from the data.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Multilayer perceptron and radial basis functions neural networks have been extensively
used for this purpose. The advantages of fuzzy set based models are transparency of the
model unlike black box nature of neural networks, ease of representing the experts’
knowledge and ability to handle imprecise information.
In this chapter fuzzy set applications have been classified into three parts—
application of fuzzy set operations, application of fuzzy arithmetic and application of
fuzzy logic. The combination of fuzzy sets and neural networks is very effective and the
techniques may be used in conjunction with finite element model. Some suggestions for
future work have been provided.

1. INTRODUCTION
Of late neural networks and fuzzy sets are being applied in many areas of manufacturing.
These techniques are the constituents of artificial intelligence, as they try to emulate human
intelligence. In spite of development of a number of computational techniques and high speed
computers, there is no substitute for skilled manufacturing personnel. It is, therefore, natural
to focus the attention towards these techniques that resemble the thinking pattern of

∗
Corresponding author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 2 U. S. Dixit

experienced and skilled manufacturing personnel. Neural networks mimic the capability of a
human to learn from the data and fuzzy sets mimic the ability to make inferences based on
imprecise and linguistic information.
Fuzzy set and neural network based methods are parts of more general soft computing
methods. Soft computing methods are distinguished from the conventional hard computing
methods by the following characteristics:

• Soft computing methods can work with approximate data. Approximation occurs in
the form of uncertainty and imprecision. Uncertainty refers to statistical variation and
imprecision to the vagueness in the definition. As an example of uncertainty,
consider that the flow stress of a material may be approximated by a normal
distribution with a mean and a standard deviation. The more is the standard
deviation, the more is the uncertainty in the flow stress value. As an example of
imprecision, consider that the product A looks more beautiful than product B. One
can not quantify the beauty in a crisp manner. Thus, there is an imprecision in the
world ‘beautiful’. Based on uncertain and imprecise data, soft computing methods
can produce approximate but useful result. Although more often the results produced
by the soft computing techniques are not accurate, one will usually get an idea of the
extent of approximation.
• Soft computing methods can work with missing data. It is possible that in the
presence of missing data, soft computing model may be more approximate and/or
may not be available for the entire domain. However, usually, one will get an idea
about the extent of approximation or certainty factor of the result obtained from the
model.
• Soft computing methods have ability to learn from the data. One need not have a
generalized theory of the phenomenon. One analogy can be given for the learning
behavior of soft computing methods. Suppose one wants to learn driving a car. One
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

method can be to teach him/her the physics and engineering of car and prepare a list
of instruction for him/her to follow. Instructions can be of the form: “If you want to
increase the speed by 10% in top gear, you have to press the accelerator by 3 mm”.
Imagine how effective this method of learning is! Many a times, hard computing
methods have similar complexity. Another way of learning the car driving that is
akin to soft computing methods is as follows. Let the learner start driving the car.
He/she takes some actions and observes the result; then takes further action based on
the feedback. After trying for a number of times, he/she will learn the driving. Many
soft computing based methods learn from the data in the similar manner. A number
of pairs of input (action) and output (result) are presented to a soft computing tool to
train it, after which it becomes possible to predict the output not only for the training
data but for other cases also.
• Soft computing algorithms are simple and many a times quite general in nature, such
that the algorithm developed for one problem may be easily modified to make it
suitable for the other case. For example, the methods to train a driver and a cook will
be quite different if done in a hard computing way. However, doing in a soft
computing way requires the similar training method. Both the cook and the driver
have to learn from experience.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 3

The present chapter describes the application of fuzzy set theory and neural networks in
metal forming and machining. The main focus is on the application of these two major
constituents of the soft computing. Only a cursory description of the theory of these
techniques is provided in this chapter. The reader is encouraged to obtain the background of
these methods from other sources [1–4].

2. NEURAL NETWORK APPLICATIONS IN MACHINING

Neural networks map a point in a multi-dimensional input space to a point in a multi-
dimensional output space. Assuming x1, x2, x3,..........., xn are the input variables and o1, o2,
........, ok are the output variables, the mapping by a neural network can be represented as

oi = fi ( x1 , x2 , x3 ,........, xn ) , i = 1, 2,........., k (1)

Thus, any output is a function of input variables. A function fi is constructed by the neural
network based on the learning data. Usually, data will be supplied in the form of input-output
pair. The learning procedure utilizing the output (target) values is called supervised learning
procedure. Sometimes, the learning data may contain only inputs and no output values. Then,
also the neural network can learn using unsupervised learning procedure. The majority of
applications of neural network in machining make use of supervised learning procedure. In
the following subsection, two common types of neural networks are described.

2.1. Multi-Layer Perceptron and Radial Basis Function Neural Networks

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

For simplicity consider the case of two inputs and one output. Figure 1 shows a multi-
layer perceptron network architecture with two hidden layer. Each layer consists of a number
of neurons, which are basically the processing units. In this type of network, the weighted
outputs of the neurons of one layer become the inputs to the neurons of the next layer. Each
neuron of the input layers receives the value of one variable and emits it as it is to be used by
subsequent layer. Let uij be the weight associated with i-th input neuron and j-th neuron in the
first hidden layer. For the sake of simplicity, assume that each hidden layer consists of 3
neurons, although there is no restriction on the number of neurons in the hidden layers. If the
processing-function of the j-th neuron in the first hidden layer is denoted by dj, then the
outputs of the neurons of the hidden layers are given by

output of neuron 1=d1 ( u11 x1 + u21 x2 ) ,

output of neuron 2=d 2 ( u12 x1 + u22 x2 ) , (1)
output of neuron 3=d3 ( u13 x1 + u23 x2 ) .

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 4 U. S. Dixit

Figure 1. An MLP Neural Network.

The weighted outputs of hidden layer 1 reach hidden layer 2. Let vij be the weight
associated with i-th neuron in the first hidden layer and j-th neuron in the second hidden
layer. If the processing-function of the j-th neuron in the second hidden layer is denoted by ej,
then the outputs of the neurons of the hidden layers are given by

output of neuron 1=e1 ( v11d1 + v21d 2 + v31d3 ) ,

output of neuron 2=e2 ( v12 d1 + v22 d 2 + v32 d3 ) , (2)
output of neuron 3=e3 ( v13 d1 + v23 d 2 + v33 d3 ) .
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

The weighted outputs of hidden layer 2 reach the output layer. Let wij be the weight
associated with i-th neuron in the second hidden layer and j-th neuron in the output layer. If
the processing-function of the j-th neuron in the second hidden layer is denoted by fj, then the
outputs of the neurons of the hidden layers are given by

o1 =f1 ( w11e1 + w21e2 + w31e3 ) . (3)

In this architecture only two hidden layers are shown. In general, there can be more
hidden layers.
Note that if all the processing functions are linear, the output will be a linear combination
of input variables. However, if this were our aim, we would have used a multi-linear
regression procedure. In general, output may be a non-linear function of the input variables.
Hence, it is usual practice to have non-linear processing functions in the hidden layers. The
function in the output layer is taken linear and often of the following form:

f ( x) = x . (4)

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 5

In that case, Equation (3) is written as

o1 =w11e1 + w21e2 + w31e3 . (5)

It is clear that if all the processing functions are linear, the output will be a linear function
of input. This is seldom the case. Usually, the processing functions of hidden layer neurons
are nonlinear. By increasing the number of neurons in a hidden layer, it is possible to fit the
function that matches the supplied data. Hence, many researchers find it convenient to have
only one hidden layer.
In the above discussion, input to a neuron has been taken as the linear combination of the
outputs of the neurons of the previous layer, with no constant term. A constant term or bias
can be easily incorporated in it by having a neuron in each layer (except the output layer),
which always emits 1. The weight associated with that neuron will act as bias.
The processing (activation) functions of the hidden layer neurons are non-linear function,
which are bounded between 0 and 1 or between −1 and +1. One common function is the log
sigmoid function given by

1
f ( x) = , (6)
1 + exp(−cx)

which is bounded between 0 and 1 for −∞ < x < ∞ . The parameter c is a slope parameter.
Another common function is tan sigmoid given by

f ( x) = tanh(cx) , (7)

which is bounded between −1 and 1.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Having fixed the processing functions, the predictive capability of the neural network is
decided by the weights. The procedure to obtain the proper weights is as follows:

• Divide total learning data into training, testing (called cross-validation by some
authors) and validation (called testing by some authors) data. A typical ratio is 6:3:1.
• Determine the weights by any optimization algorithm, such that the error in
prediction is minimized for training data upto certain error goal. With the trained
neural network, the testing error is calculated. An optimum neural network provides
reasonable predictions for training as well as the testing data. If the training error is
less, but testing error is high, the network is called to over-fit the data. In that case,
the network is trained again (usually with relaxed error goal) and tested.
• The network architecture and weights that provide the tolerable amount of error in
training and testing data are frozen. The performance of the finalized neural network
is assessed on the validation data.

The most common algorithm for the training of MLP neural networks has been
backpropagation algorithm, in which the weights of the last layer are adjusted first and then

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 6 U. S. Dixit

the weights of other layers are adjusted in backward sequence. The optimization problem of
determining the weights poses challenging issues. There is a tendency to get trapped into
local minima. To avoid this problem, many non-traditional (particularly evolutionary)
optimization algorithms have been employed. The convergence speed is also a major concern.
Another common neural network is radial basis function (RBF) neural network. It has
only one hidden layer. A typical architecture consisting of 2 input, 3 hidden and one output
neurons is shown in Figure 2. No weights are associated between hidden layer and input
layer. The values of input variables reach the neurons of hidden layers as it is. With each
neuron in the hidden layer is associated a center, which is a fixed vector, whose size is equal
to the number of input variables. In fact, usually from the supplied learning data, some are
chosen to act as center. The processing functions of the neurons in the hidden layer are radial
basis functions, which map the Euclidean distance of an input vector from the center into a
real number. The weighted outputs of each neuron in the hidden layer reach the output
neuron, where it is summed. Thus, for the present neural network, the output is given by

3
o = ∑ wiφi ( x − ci 2 ), (8)
i =1

where x is the input vector, ci is the center of the i-th neuron in hidden layer, φi is the
corresponding radial basis function and wi is the corresponding weight. Specifically, for the
network shown in Figure 2, the output is given by

( )
3
o = ∑ wiφi ( x1 − ci1 ) 2 + ( x2 − ci 2 ) 2 , (9)
i =1

where ci1 and ci2 are the components of the center of i-th neuron. There are a number of
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

possible radial basis functions. Some of them are as follows:

( )
2 1/ 2
Multiquadrics: φi ( x ) = x - ci + si2 , (10)

( )
2 −1/ 2
Inverse Multiquadrics: φi ( x ) = x - ci + si2 , (11)
2
− si2 x-ci
Gaussians: φi ( x ) = e (12)
2
Thin Plate Splines : φi ( x ) = x - ci log x - ci . (13)

In Equations (10–12), si is called spread parameter that decides the zone of influence of a
neuron. The most commonly used radial basis function for the modeling of metal forming and
machining processes is Gaussian function.
From Equation (9) it is seen that once the centers and types of functions are fixed, the
output is a linear function of the weights. The weights can be obtained by multiple-regression
method. This procedure provides unique and fast solution.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 7

Figure 2. An RBF Neural Network.

2.2. Prediction of Machining Performance Using Neural Networks

The performance of a machining process is dependent on input variables and external

factors. One can carry out the prediction online as well as offline. Online prediction requires
the feedback from sensors. The important performance parameters that are modeled using
neural networks are cutting force, surface roughness, dimensional deviation, temperature, tool
life and tool wear. The sensors that are commonly employed for online prediction are
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

accelerometer, cutting force dynamometer, acoustic emission sensor and sound measuring
sensor. The network takes the cutting parameters like cutting speed, feed and depth of cut as
input and predicts the required performance parameter as output.
Although there have been a number of attempts to model the machining performance
parameters, there are a number of issues to be addressed. First challenge is to develop a neural
network model that requires less amount of learning data. The second challenge is outliers.
The experimental data often contains spurious data. Therefore, a data filtration scheme is
required. The third challenge is the statistical variation in the data. A number of replicate
experiments need to be carried out. The neural network model should be able to predict not
only the parameters, but their expected variations. There have been few attempts to predict
variations in the machining parameters. Kohli and Dixit [5] have used neural networks to
predict lower, most likely and upper estimates of surface roughness in a turning process.
Figure 3 shows their results for the case of dry turning of steel with coated carbide tools. It is
seen that in almost all the cases, experimental values fall in between lower and upper
estimates. There is a significant difference between upper and lower estimates indicating the
presence of large amount of statistical variation.
Compared to the estimation of cutting force and surface roughness, estimation of tool
wear is more difficult. Sick [6] reviewed 138 publications dealing with online and indirect

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 8 U. S. Dixit

tool monitoring in turning using neural networks and observed that a reliable tool wear
monitoring system is still far away from the goal.

Figure 3. Predicted values versus experimental values of dry turning of steel with coated carbide tool.
With permission from Kohli and Dixit [5]. Copyright [2005] Springer.

2.3. Optimization of Machining Performance Using Neural Networks

The optimization of machining process is an extensively researched area. The goals of

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

optimization problems have been minimum cost of machining, maximum production rate and
the weighted combinations of these two goals. These goals are subjected to various
constraints such as maximum surface roughness, maximum cutting force, maximum power
etc. Of late researchers have started solving multi-objective problems.
As a neural network model can provide performance of a machining operation as a
function of input parameters, it can be used in the optimization process. The objective and
constraint functions can be represented in the form of trained neural networks. Afterwards,
any suitable optimization method can be employed for solving the optimization problem. Cus
and Zuperl [7] have proposed a methodology of machining optimization with neural network.
They optimize a manufacturer’s implicit value function y that is a function of surface
roughness, cost of machining and production rate. Thus, the three goals are combined into
one depending on the manufacturer’s preference. A neural network is trained with cutting
speed, feed and depth of cut as input variables and manufacturer’s value function y as the
output variable. The training data is selected in such a manner, so that constraints are not
violated. The trained neural network is used as the objective function and optimization is
carried out by a suitable multi-variable optimization problem. Authors have provided an
example of CNC turning, in which empirical expressions have been used for surface
roughness, tool life, cutting force and cutting power. In place of empirical expressions, one
can use neural networks for prediction of machining performance parameters.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 9

Apart from offline optimization, online optimization of machining processes can also be
carried out using neural networks. Ko and Cho [8] have proposed an approach for the
adaptive optimization of face milling process that uses two neural networks— one for
predicting the tool wear and the other for predicting the performance parameters (cutting
force, cutting power and surface roughness). The first neural network used the feed force
signal for the prediction of length of the wear. The input variables for the second neural
network are cutting conditions and tool wear length. The process is optimized online using
exterior penalty function method to tackle the constraints. Liu and Wang [9] have proposed
neural network based control and optimization of milling process.
Neural network can be trained for producing the inverse relations too. One can develop a
network that predicts surface roughness as a function of cutting speed, feed and depth of cut.
Similarly, an inverse network can be developed to predict cutting speed, feed and depth of cut
for a prescribed surface roughness. This will not be difficult if the inverse problem has a
unique solution. One can suitably select the input variables and their ranges such that a unique
solution is obtained. Using this concept, Azouzi and Guillot [10] has proposed an inverse
process neurocontroller for optimizing the turning process. The force and vibrations sensors
are used to provide the feedback in the proposed neurocontroller.

3. NEURAL NETWORK APPLICATIONS IN METAL FORMING

In metal forming area, neural networks have been used for the prediction of performance
parameters and control. Most of the applications of neural networks are in the field of rolling,
as the rolling is one of the oldest, the most important and the most complex process. Multi-
layer perceptron network has found more application in the modeling and control of rolling
process than any other type of network. Pican et al. [11] used MLP neural network with
backpropagation algorithm for predicting the roll force in temper rolling. Temper rolling or
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

skin-pass rolling is a cold rolling process, in which the sheet metal is subjected to very light
reduction (0.5–4%) in thickness, in the presence of friction. The purpose of temper rolling
process is to provide a degree of surface hardening, restore temper, prevent stretcher strains
(Lueder’s bands), impart a desired finish, and to impart a desired degree of flatness. The roll
force in this process depends on a number of parameters. Pican et al. [11] have used the
following parameters as the input parameters:

• Number of passes: 0, 1 or 2
• Width of sheet: 600–1900 mm
• Thickness of the sheet: 0.3–3.0 mm
• Percentage reduction: 0.2–2.0 %
• Lubrication: no lubrication, lubricant 1, lubricant 2
• Back tension: 10–50 N
• Front tension: 10–70 N
• Sheet temperature: 0–60 °C
• Surface finish of work rolls: 21 different conditions
• Roll wear (basically in terms of the total length of sheet processed by the work rolls):
0–130 km

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 10 U. S. Dixit

• Diameter of work roll: 570–610 mm

• Flow stress (which is called as resistance by the authors): 120–430 MPa

The output layer just contains one neuron corresponding to roll force. The roll force
prediction ultimately helps in adjusting the initial gap between the rolls. The final network
contained two hidden layers— one with 48 neurons and other with 24 neurons. A total of
6221 training data and 1265 testing data were used for modeling. This data was collected
from the plant in about six months. The requirement of huge amount of data for neural
network modeling poses one of the challenges in its implementation.
Larkiola et al. [12] has predicted roll force by a combination of physics based model and
neural network model. In their approach, although the roll force was calculated based on
physics based model, the material and friction parameters in the model were obtained from
trained neural networks. The neural network for the computation of material parameter (2
material constants) consisted of 18 input neurons corresponding to alloying elements,
temperature of the last hot rolling pass, coiling temperature etc and was trained with 4500
training data. The final network contained one hidden layer with 10 neurons and provided
root mean square error of 4.9%. The friction parameter was obtained in an inverse manner, so
as to obtain the correct roll prediction by the physics based model. Thus, any error in the
estimation of material parameter or physics based model is compensated by the friction
parameter. The neural network for the computation of friction consisted of 10 input variables.
Cho et al. [13] used two different types of neural networks for the prediction of roll force
in cold rolling. One MLP neural network predicted the roll force directly and other MLP
neural network provided a corrective coefficient which has to be multiplied to the prediction
made by a physics based model. The authors observed that both the MLP neural network
models could improve the accuracy of prediction by 30–50% compared to the physics based
model. Prediction by neural network is probabilistic in nature. Therefore, in order to have
more reliability, the predictions of both neural networks and physics based model can be
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

considered together. One can take the weighted average of three predictions. Lee and Lee [14]
also proposed a model for the prediction of roll force in hot rolling that combines neural
networks with the physics based model. Xie et al. [15] has used neural network for predicting
the error in the mathematical model that predicts the coiling temperature. Recently, Geerdes
et al. [16] applied the neural networks in the prediction of temperature in a hot strip mill. The
authors observed that a combination of neural network model and physics based model
provides the best result, in which the neural network is used to correct the physics based
prediction.
Gorni [17] has described the application of neural networks in the modeling of hot plate
rolling process for following tasks:

• Predicting the thermal profile of slabs in the reheating furnace

• Detecting turn-up of the plate, i.e., detecting the excessive bowing up of the plate
during rolling
• Predicting the length of the discarded portion of the final rolled stock, when the plate
is rolled from cast slabs
• Pass scheduling i.e., planning the reductions to be achieved in different stands of a
tandem rolling mill.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 11

Author has pointed out the requirement of sufficient amount of data for training as one of
the challenges towards practical application of rolling process.
Gunasekera et al. [18] has developed a neural network model for cold rolling process.
The data required in the modeling was collected from a physics-based model. The authors
concentrated on reducing the error in prediction and training time of the neural network.
Among the physics based modeling techniques, finite element model (FEM) has been
found to be very effective in predicting various performance parameters of rolling. However,
FEM requires a large computational time and is not suitable for online prediction. A neural
network can be trained with FEM generated data offline. The trained neural network can be
used for predicting the performance parameter in subsequent applications, as the time
required for prediction by a neural network is very small. This possibility has been explored
by a number of researchers [19−21]. Among these, Dixit and Chandra [19] concentrated on
reducing the number of training and testing data for neural network modeling and prediction
the most likely, lower and upper estimates of roll force and roll torque.
Like in the rolling, neural networks can be employed to model other metal forming
processes. However, there is less number of papers describing the application of neural
networks in the other processes. Osakada and Yang [22] employed neural networks for the
process planning of cold forging. Raj et al. [23] has applied neural networks in the modeling
of hot extrusion and hot upsetting. Mori and Li [24] applied neural networks in the forging
process. The authors used two geometric parameters of the die and aspect ratio of the
workpiece as the input variables of the neural network model. The corresponding outputs
were the die filling state and the forging load. The training and testing data for the neural
network were obtained from FEM simulations. Kazan et al. [25] predicted spring-back in
wipe bending using neural networks.
There have been far less number of applications of RBF neural networks for the
prediction of performance parameters in rolling. Sbarbaro-Hofer [26] used RBF neural
networks in the control of the strip thickness in a steel rolling mill. Recently, Gudur and Dixit
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[27] used RBF neural network for predicting the roll force and roll torque in cold rolling
process. In their model, network had 5 input parameters. They used 55 training data and 30
testing data. The percentage root mean squared fractional errors in the roll force prediction
were 0.25% for training and 2.9% for testing. The corresponding values for the roll torque
were 0.77% and 3.5%.
The above description has highlighted the application of neural networks to metal
forming in the following ways: (1) use of neural network as an alternative computational tool
for the modeling of the process (2) use of neural network for correcting the predicted values
obtained by physics-based methods and (3) capturing the knowledge of physics-based model
with a view point of reducing the prediction time for online learning and/or optimization.
Gudur and Dixit [28] have presented one more application of neural network— its use as an
assistive tool to FEM model of the process. The general overview of the scheme for cold flat
rolling is shown in Figure 4. The authors have used a plane strain rigid-plastic FEM model
that provides the velocity components at each node alongwith the location of neutral point. A
radial basis function neural network (RBF NN) accepts these as input and gets trained. The
trained neural network is kept in a repository. Now, whenever a problem is to be solved,
initial velocity field and neutral point are obtained from this trained neural network. Further
iterations of FEM are carried out to obtain a more accurate solution and post-process the
results. It is observed that the reduction in the computational time is by a factor of order 10.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 12 U. S. Dixit

For a more refined mesh updated Lagrangian scheme, the procedure is expected to provide
more significant time saving.

Figure 4. A general overview of neural network assisted FEM model. With permission from Gudur and
Dixit [28]. (A label in the figure has been slightly modified). Copyright [2005] Elsevier.

4. FUZZY SET APPLICATIONS IN MACHINING

A fuzzy set is the set in which its members are allowed to have any positive membership
grade between 0 and 1. A membership grade denotes the strength of association of the
element within the set with 1 indicating the perfect association. To understand the concept of
fuzzy sets, consider the universe containing the first nine natural numbers from 1 to 9.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Suppose that one has to construct a fuzzy set F of “numbers close to 5” from these numbers.
There can be various solutions to this problem. Some of the solutions are as follows:

F = {0.2 / 3, 0.6 / 4, 1/ 5, 0.6 / 6, 0.2 / 7},

F = {0.3 / 3, 0.7 / 4, 1/ 5, 0.7 / 6, 0.3 / 7}, (14)
F = {0.1/ 2, 0.4 / 3, 0.9 / 4, 1/ 5, 0.9 / 6, 0.4 / 7, 0.1/ 8},

where 0.2/3 means that the membership grade of natural number 3 is 0.2. Note that in all the
above solutions, the natural number 5 is having the membership grade 1 as there is no doubt
that it is close to 5. On the other hand, there is subjectivity in the allocation of membership
grades to other numbers. This depends on how one perceives the word ‘close’. To find out
suitable values of membership grades is one of the most challenging tasks of fuzzy set theory.
It is natural that it needs some art in capturing the human thinking in the mathematical
framework.
Fuzzy sets have been extensively used in decision making by employing various
operations of fuzzy set theory. As a simple example of it, consider three cutting fluids L1, L2
and L3. Let A be the set of high specific heat and B the set of high thermal conductivity and

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 13

A = {0.7 / L1, 0.9 / L2, 0.8 / L3}, B = {0.8 / L1, 0.5 / L2, 0.6 / L3} (15)

As per expert opinion, a good cutting fluid should have high thermal conductivity and
high specific heat. In fuzzy set theory, the word ‘and’ can be replaced by an intersection
operator. Thus, the fuzzy set of high specific heat and high thermal conductivity is given by

A ∩ B = {0.7 / L1, 0.5 / L2, 0.6 / L3} . (16)

Note that here the following definition of A∩B has been used. The membership grade of
an element in the set A∩B is the minimum of its membership grades in A and B. Now,
examining the set A∩B (Eq. 16), it is seen that L1 has the highest membership grade and thus
is the most suitable cutting fluid. In the same way, the ‘or’ of English language can be
replaced by a union operation of the fuzzy set, which is usually defined as follows. The
membership grade of an element in set A∪B is the maximum of its membership in A and B.
A prominent branch of fuzzy set theory is the fuzzy arithmetic, which deals with the
fuzzy numbers. Fuzzy numbers are the generalization of interval number, where
corresponding to each membership grade, a different interval can be assigned. For example,
coefficient of friction in a machining process may be (0.7, 0.7), (0.6, 0.8) and (0.5, 0.9) at
membership grades of 1, 0.8 and 0.5 respectively. Many of the process variables are
imprecise and uncertain and thus are best represented by fuzzy numbers. When these
variables are put in a physics-based model, the result is obtained in the form of fuzzy
numbers, which is a more realistic picture. Note that apart from using fuzzy variables in a
physics model, the fuzziness can be introduced in the model itself, if the physics is uncertain.
For example, assume a physics-based model provides that the parameter y is related to
parameter x in the following way:
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

y = x2 . (17)

This can be converted to fuzzy model by using the following relation:

y = xa , (18)

where a is a fuzzy number, whose most likely value corresponding to the membership grade 1
is 2. Once the result is obtained in the form of fuzzy numbers, using this information in an
appropriate way is another important task.
Another important branch of fuzzy set theory is fuzzy logic, which uses IF-THEN rules
in which either the antecedent (IF part) or the consequent (THEN part) or both contain fuzzy
sets. The decision is taken by evaluating the fuzzy rules. In Mamdani fuzzy inference system,
both antecedent and consequent are fuzzy sets. Sugeno fuzzy model, also known as Takagi,
Sugeno and Kang (TSK) model contains fuzzy sets in antecedent and crisp function in the
consequent. A typical rule in Mamdani system is as follows:

If depth of cut is low and feed is low and cutting speed is high then surface roughness is
low.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 14 U. S. Dixit

The same rule in TSK system may look as follows:

If depth of cut is low and feed is low and cutting speed is high then surface roughness
−0.1
is 5 fv d.

In the following paragraphs, some applications of fuzzy set theory to machining will be
described. The applications have been classified into three groups— (i) Application of fuzzy
set operations, (ii) Applications of fuzzy arithmetic, (iii) Applications of fuzzy logic.

4.1. Applications of Fuzzy Set Operations

Fuzzy set operations are very helpful in taking decisions in the presence of conflicting
and incommensurable objectives. For example, consider that 5 different types of models of a
product have to be evaluated against two objectives. If a model has a membership grade μ1 in
one objective and μ2 in other objective, its overall membership grade can be defined as

μo = min( μ1 , μ2 ) . (19)

Here, the overall performance of the model is dependent on the most poorly performing
aspect. This type of strategy is called non-compensating strategy. An overall membership
grade based on a compensating strategy may be defined

μo = μ1 μ2 . (20)
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

A weighted combination of two strategies may also be taken. Thus, if membership grades
of a model in n different objectives are μ1 , μ 2 ,........., μ n , then overall membership grade
may be defined as

μo = (1 − α ) min( μ1 , μ2 , ,,,,, μn ) + α n μ1 μ2 ,,, μn , (21)

where α is a weight factor. If α=1, then Eq. (21) provides a pure compensating strategy and if
α=0, it provides a pure non-compensating strategy.
One can easily incorporate linguistic information in the decision process. For example, if
a component has a membership grade μ in a set of ‘good surface finish’, it can be assigned a
membership grade of μ in the set of ‘more or less good surface finish’. Thus, customers’
requirement available in the language form may be converted into mathematical form.
In the area of multi-objective machining optimization, the fuzzy set based decision
strategy may be very effective. There are a number of objectives to be satisfied in a multi-
objective machining optimization such as minimum cost of product, maximum production
rate, maximum profit rate and minimum surface roughness. In the fuzzy set based approach,
different solutions may be assigned membership grades in different objectives and then based
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 15

on the fuzzy set operations, the solution having the overall membership grade may be chosen.
Panchal et al. [29] have used this type of approach in the optimization of turning process.

4.2. Applications of Fuzzy Arithmetic

Fuzzy arithmetic is helpful when the values of variables are uncertain and imprecise.
Several researchers have used fuzzy arithmetic in machining optimization. Recently, Liu [30]
has developed a method that can obtain the optimum fuzzy cost of machining in a multi-pass
turning process, when the exponents and coefficients in the objective and constraint functions
are taken as fuzzy number. The optimum cost in the form of a fuzzy number provides more
information than the crisp value.
In developing a fuzzy arithmetic model, one may face difficulty in assigning suitable
fuzzy numbers to various variables. A simple way is to treat these variables as linear
triangular numbers. The most likely estimate of a variable can be assigned a membership
grade 1 and the lower and upper estimates of the variable can be assigned membership grades
of 0.5. With these three points a triangle can be constructed. The more sophisticated methods
can also be tried.

4.3. Applications of Fuzzy Logic

The fuzzy logic containing the IF-THEN fuzzy rules has been used in developing the
expert system and controller. Basically, a fuzzy logic based controller may be considered as
an expert system that can take the appropriate decision for a control action based on the
sensory feedback. Thus, software part of a controller is similar to software of an expert
system. Fuzzy logic has been widely used in modeling and control of machining problems.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 5 shows the block diagram of an expert system that can learn from shop floor data.
The system provides the necessary output that can be used for either process planning or
control. The expert system requires IF-THEN rules, which can either be supplied by an expert
or can be generated from the shop floor data. The shop floor data is limited in quantity. It
often becomes difficult to use that data for generating all the required rules. A good strategy
can be to train a neural network from the data, as neural networks are efficient in learning
from the data. The trained neural network can be used to generate lots of IF-THEN rule. The
strategy described in [31, 32] can be adopted for generating the rules. The generated rules can
be evaluated by an expert and expert can suitably modify them. Once the rule-base is
available, the inference module can predict the output for given set of input.
Abburi and Dixit [31] used this procedure for the surface roughness prediction in a
turning process. The input variables of their expert system are cutting speed v, feed f, depth of
cut d and acceleration of radial vibration of the tool-holder a. Figure 6 shows how the input
and output variables are fuzzified into a number of overlapping fuzyy subsets. In this type of
fuzzification, the value of a particular variable may be part of two fuzzy subsets. Accordingly,
for a given set of input variables, a number of rules get fired. The strength of different rules is
different. The combined action of the rules decides the output of the expert system based on
the standard procedure of fuzzy logic theory. The validation of the procedure with
experimental data is good. The similar procedure may be employed for modeling of other
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 16 U. S. Dixit

machining parameters such as cutting force, tool wear, and tool life. The expert system can
also be conveniently used for solving the inverse problems. For example, for a given surface
roughness, cutting speed, feed and depth of cut can be obtained. This type of expert system
also possesses a certain amount of extrapolation capability, unlike neural networks.

Figure 5. A block diagram of an expert system that has learning capability. With permission from
Abburi and Dixit [31]. Copyright [2006] Elsevier.

The fuzzy set based expert system developed in [31] has the following advantages:

• The system is able to learn from exemplars.

• Based on the input parameters, the system can predict the output quickly.
• The system is transparent. The user can see the rules and find the flaw in the system,
if any.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

• The system can solve the inverse problems quickly. The inverse problems usually
have multiple solutions, which the expert system can easily identify.
• A certain amount of extrapolation is possible, as the physics of the process usually do
not change abruptly.

In an early work, Fang and Jawahir has [33] has proposed a fuzzy logic based
methodology to assess total machining performance encompassing surface finish, tool-wear
rate, dimensional accuracy, cutting power and chip breakability. The authors quantified the
effect of major influencing factors on total machining performance by fuzzy set method and
developed a series of fuzzy set models to give quantitative assessments for the given set of
input conditions. Fang et al. [34] has presented a fuzzy set based methodology for predicting
the chip form and chip breakability in machining. A series of machining experiments were
conducted to establish a chip control database which is used as a standard to quantify the
variability of chip breaking due to effects varying process parameters.
In References [35, 36], process parameters are selected/optimized based on fuzzy logic.
Susanto et al. [37] applied fuzzy logic in tool wear monitoring of face milling process.
References [38, 39] deal with predicting burns in grinding process. References [40–42] apply
fuzzy logic to drilling process.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved. Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 17

Figure 6. Fuzzification of input and output variables for turning process..With permission from Abburi
and Dixit [31]. Copyright [2006] Elsevier.

5. FUZZY SET APPLICATIONS IN METAL FORMING

Fuzzy set theory has been applied in modeling and control of metal forming processes.
Amongst metal forming processes, rolling is one of the most investigated processes.
Therefore, there are a number of papers dealing with the application of fuzzy sets to it. In the
following sub-sections, the application of fuzzy sets to rolling process is described. The fuzzy
set theory can be applied to other metal forming processes in the similar way.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved. 18 U. S. Dixit

Figure 7. Membership functions of desired: (a) velocity (b) reduction and (c) power. With permission
from Dixit et al. [43]. Copyright [2002] Elsevier.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 19

5.1. Applications of Fuzzy Set Operations

An interesting application of fuzzy set operations has been provided in reference [43].
Here, the fuzzy set operations have been used for the conceptual design of a laboratory cold
rolling mill. The customer’s requirement of maximum mill speed, maximum possible
reductions and rolling power are represented as fuzzy sets as shown in Figure 7.
The customer will be very pleased to have a mill speed of more than 1m/s as indicated by
the constant membership grade of 1.0 after an outlet strip velocity of more than 1 m/s. As the
strip-velocity reduces below 1 m/s, the membership grade indicating the satisfaction level of
the customer linearly decreases to become 0 at 0 strip-velocity. In the same way, the
membership functions of desired reduction and power may be interpreted. Note that Figure 7
(c) represents a fuzzy set of low power. The membership grade in the fuzzy set of very low
power is obtained by squaring its membership grade in the fuzzy set of low power. Assuming
that the customer requires high reduction, high mill-speed and very low power, the overall
membership grade is obtained by the fuzzy intersection operation. Thus, the overall
membership grade μo is given by

μo = min ( μr , μv , μ p2 ) . (22)

Among the various possible solutions, the solution providing the maximum overall
membership grade is chosen.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 8. Membership functions of roll diameter from the viewpoints of low power and defect-free
rolling. With permission from Dixit et al. [43]. Copyright [2002] Elsevier.

In reference [43], the fuzzy set theory has also been applied for deciding the roll
diameter. Large diameter rolls provide better rigidity, better cooling, lesser value of the
minimum coefficient of friction, a greater value of maximum possible reduction and lesser
tendency of central bursting and split end defects in the rolled product. On the other hand,

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 20 U. S. Dixit

small rolls provide lesser rolling power and lesser spread. Thus, there are conflicting
objectives in the selection of proper roll diameter. The proper roll diameter will maximize the
overall objective, which is measured by the overall membership grade. For example, Figure 8
shows the membership functions of roll diameter from the viewpoints of low power and
defect-free process. The intersection of two curves is the point that maximizes the overall
membership grades in these two objectives. In the same way, other objectives may be
considered. The art lies in choosing the proper shape functions.
Dixit and Dixit [44] applied fuzzy set operations in the scheduling of tandem mills. The
conflicting objectives of minimum power and maximum reliability were optimized based on
the fuzzy set theory. In this work, the material parameter and friction are considered as fuzzy
number.

5.2. Applications of Fuzzy Arithmetic

Metal forming problems have been extensively modeled using finite element method. A
finite element package requires the data about the material properties and friction coefficient.
Often these values are imprecisely known and can therefore be represented as fuzzy numbers.
Dixit and Dixit (1996) considered the material and friction parameters as fuzzy number
and obtained the values of roll force, roll torque and roll pressure distribution from a finite
element model. The material is assumed to strain harden in the following way:

n
⎛ ε eq ⎞
σ y = ( σ y )0 ⎜ 1 + ⎟ , (23)
⎝ b ⎠

where (σ ) is the yield stress of the raw material, ε eq is the equivalent strain and b and n
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

y 0

are material handling constants. The parameters ( σ ) , b and n are taken as fuzzy. Figure 9
y 0

shows the membership functions of ( σ ) , b and n for a steel and coefficient friction. Based
y 0

on these parameters, variation of roll force with reduction for some membership grades is
shown in Figure 10. Reference [11] mentioned in the figure pertains to the work of Shida and
Awazuhara [46].

5.3. Applications of Fuzzy Logic

Fuzzy logic has been widely used for controlling the various processes including rolling.
Recently, Gudur and Dixit [47] employed it for the modeling of roll force and roll torque. For
this purpose, a Sugeno fuzzy model is used and the methodology of reference [48] is adopted.
The representation of IF-THEN rules in the TS fuzzy model is given as

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 21

Ri : IF x1 is A1 (i ) and x2 is A2 (i ) .......and xk is Ak (i ),
(24)
THEN yi = ai0 + ai1 x1 + ai2 x2 + ........ + aik xk .

where Ri is the ith rule, vector x = [ x1 , x2 ,....xk ] represents the set of k input variables like
j
reduction, roll radius etc and A (i ) is the fuzzy subset corresponding to the jth input
j
variable. Note that depending on the rule, A (i ) can be low (L), medium (MD) or high (H).
The output of the ith IF-THEN rule is yi , which is assumed to be a linear function of the input
variables. The coefficients associated with the input variables in the linear functions are
j
denoted as ai , where subscript denotes the rule number and superscript denotes the
associated input variable. The aggregated output of the Sugeno fuzzy model, ŷ is expressed
by a weighted average of the rule consequents:

NR

∑w y i i
yˆ = i =1
NR
, (25)
∑w
i =1
i

where wi is the strength of the ith fuzzy rule and N R is the total number of rules. The
strength of the rule is taken as the product of the membership grades of the input values of the
variables in the corresponding fuzzy subsets.
j
The coefficients ai of Eq. (24) can be obtained by minimizing the maximum
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

approximation error ( λ ) between the actual output and the fuzzy model output over the given
input-output datasets, which can be obtained experimentally or from FEM model. For this,
the following linear programming (LP) model is solved:

Minimize λ ,
subject to,
o p − yˆ p ≤ λ , p = 1, 2,........m,
(26)
−o p + yˆ p ≤ λ , p = 1, 2,........m,
λ ≥ 0,
aij ≥ 0, i = 1, 2,.......N R , j = 1, 2,.....k .

where the subscript p represents the pth input-output pattern and m is the total number of
patterns in the dataset. The actual output of the pth data is o p . The maximum approximation
j
error is termed as λ . The design variables of the problem are ai and λ .

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved. 22 U. S. Dixit

Figure 9. Membership functions of (a) (σ ) y 0 (b) b (c) n (d) coefficient of friction. With permission
from Dixit and Dixit [45]. Copyright [1996] Elsevier.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 23

Figure 10. Variation of roll force with reduction at different membership grades. With permission from
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Dixit and Dixit [45]. Copyright [1996] Elsevier.

The method described above has been extended to the prediction of lower and upper
estimates of the roll force and the roll torque based on the model described in [48]. The
maximum approximation error for computing the coefficients for lower and upper prediction
are termed as λ1 and λ2 respectively. The coefficients for lower and upper estimates are
obtained by solving the following LP models:

Minimize λ1 ,
subject to,
o p − yˆ p ≤ λ1 , p = 1, 2,........m,
(27)
o p − yˆ p ≥ 0, p = 1, 2,........m,
λ1 ≥ 0,
ai ≥ 0,
j
i = 1, 2,.......N R , j = 1, 2.....k .

and

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 24 U. S. Dixit

Minimize λ2 ,
subject to,
−o p + yˆ p ≤ λ2 , p = 1, 2,........m,
(28)
o p − yˆ p ≤ 0, p = 1, 2,........m,
λ2 ≥ 0,
ai ≥ 0,
j
i = 1, 2,.......N R , j = 1, 2,.....k .

Gudur and Dixit [47] have also proposed a method to remove the outliers from the
supplied data for learning.
Figure 11 shows the fuzzification of 5 input variables used in [47]. The input variables
are ratio of roll radius to inlet thickness (R/h1), coefficient of friction (f), percentage reduction
(r), and hardening parameters (b and n). The learning data from the fuzzy model is generated
from a radial basis function neural network (NN) model.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 11. Fuzzification of input variables for a typical rolling process. With permission from Gudur
and Dixit [2009]. Copyright (2009) Springer.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 25

Figure 12. Lower, most likely and upper estimates of roll force. With permission from Gudur and Dixit
[2009]. Copyright (2009) Springer.

The NN model in turn was trained by an FEM model. The lower, most likely and upper
estimate of roll force is shown in Figure 12.
The fuzzy logic can be applied in the online control of gauge thickness of the sheet in the
rolling process. It can also be used for controlling the flatness. Recently, Li and Janabi-Sharifi
[49] has proposed a fuzzy logic based strategy for inter-stand tension control in a hot strip
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

rolling mill. The control of tension is essential for maintaining the good quality of the
product. The tension is controlled by regulating the speed of work rolls. Authors have used
armature current of drive rolls as an indicator of the forward tension. Some other interesting
papers describing the fuzzy set applications to other metal forming processes are [50–53].

CONCLUSION
This chapter has highlighted the applications of neural networks and fuzzy sets to
machining and metal forming. Both these techniques have been applied in modeling,
optimization and control of the processes. Each of these techniques can be used separately or
together. For example Abburi and Dixit [31] used neural networks to learn from shop floor
data; the neural network predicted data in turn were used to develop a fuzzy expert system.
The hybrids of these methods have also been used, for example an adaptive-network-based
fuzzy inference system (ANFIS) is a new soft computing tool that takes inspiration from
fuzzy sets as well as neural networks. It uses the fuzzy rules of Sugeno type. The parameter of
the membership functions and output functions are adaptively changed based on the input

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 26 U. S. Dixit

data. The training procedure is similar to neural networks and it uses a layered architecture
similar to MLP network.
The neural network and fuzzy sets can also be used in conjunction with finite element
model. Finite element model has established itself as a robust tool in modeling, however it
takes a lot of computational time and does not provide a close form solution. Neural network,
fuzzy sets and finite element method in combination can provide a fast, transparent and robust
model.
For more successful applications of neural network and fuzzy sets in the machining and
metal forming area, the following issues need the attention of the researchers:

• These methods require a lot of data for learning. Although a number of methods have
been proposed to minimize the data required for learning, continuing effort is needed
in this direction. The soft computing based methods should have the capability to
learn from limited, imprecise and missing data.
• The robust data filtration algorithms should be developed to deal with the outliers in
the data. The data may contain statistical variation. The soft computing based model
should be able to predict the output in the form of a statistical variable. At present,
barring a few exceptions, most of the models predict the mean value of the
parameter.
• Fuzzy set theory is an excellent tool for incorporating the linguistic information in
the model. However, construction of a suitable membership function for a variable is
an art. There is a need to develop more systematic methods for it.
• Compatible hardware system like sensors, actuators etc should also be developed to
take the maximum advantage of the soft computing methods.

REFERENCES
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[1] Haykin, S. (2004), Neural Networks— A Comprehensive Foundation, Prentice-Hall of

India, New Delhi.
[2] Klier, G.J. and Folger, T.A. (1993), Fuzzy Sets, Uncertainty and Information, Prentice-
Hall of India, New Delhi.
[3] Bojadzieve, G. and Bojadzieve, M. (1995), Fuzzy Sets, Fuzzy Logic Applications,
Advances in Fuzzy Systems— Application and Theory, Vol. 5, World Scientific,
Singapore.
[4] Dixit, P.M. and Dixit, U.S. (2008), Modeling of Metal Forming and Machining— By
Finite Element and Soft Computing Methods, Springer-Verlag, London.
[5] Kohli, A. and Dixit, U.S. (2005), A neural-network-based methodology for the
prediction of surface roughness in a turning process, International Journal of Advanced
Manufacturing Technology, Vol. 25, pp. 118−129.
[6] Sick, B. (2002), On-line and indirect tool wear monitoring in turning with artificial
neural networks: a review of more than a decade of research, Mechanical Systems and
Signal Processing, Vol. 16, pp. 487−546.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 27

[7] Cus, F. and Zuperl, U. (2006), Approach to optimization of cutting conditions by using
artificial neural networks, Journal of Materials Processing Technology, Vol. 173, pp.
281−290.
[8] Ko, T.J. and Cho, D.W. (1998), Adaptive optimization of face milling operations using
neural networks, ASME Journal of Manufacturing Science and Engineering, Vol. 120,
pp. 443−451.
[9] Liu, Y. and Wang, C. (1999), Neural network based adaptive control and optimization
in the milling process, International Journal of Advanced Manufacturing Technology,
Vol. 15, pp. 791−795.
[10] Azouzi, R. and Guillot, M. (1998), On-Line Optimization of the Turning Process Using
an Inverse Process Neurocontroller, ASME Journal of Manufacturing Science and
Engineering, Vol. 120, pp. 101−108.
[11] Pican, N., Alexandre, F. and Bresson, P. (1996), Artificial neural networks for the
presetting of a steel temper mill, IEEE Transactions: Expert, Vol. 11, pp. 22−27.
[12] Larkiola, J., Myllykoski, P., Nylander, J. and Korhonen, A.S. (1996), Prediction of
rolling force in cold rolling by using physical models and neural computing, Journal of
Materials Processing Technology, Vol. 60, pp. 381−386.
[13] Cho, S., Cho Y. and Yoon, S. (1997), Reliable roll force prediction in cold mill using
multiple neural networks, IEEE Transactions: Neural Networks, Vol. 8, pp. 874−882.
[14] Lee, D. and Lee, Y. (2002), Application of neural-network for improving accuracy of
roll force model in hot-rolling mill, Control Engineering Practice, Vol. 10, pp.
473−478.
[15] Xie, H.B., Jiang, Z.Y., Liu, X.H. and Wang, G.D. and Tieu, A.K. (2006), Prediction of
coiling temperature on run-out table of hot strip mill using data mining, Journal of
Materials Processing Technology, Vol. 177, pp. 121−125.
[16] Geerdes, W.M., Alvarado, M. Á.T., Cabreara-Rios, M. And Cavazos, A. (2008), An
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

application of physics-based and artificial neural networks-based hybrid temperature

prediction schemes in a hot strip mill, ASME Journal of Manufacturing Science and
Engineering, Vol. 130, 014501, pp. 1−5.
[17] Gorni, A.A. (1997), The application of neural networks in the modelling of plate rolling
processes, JOM-e, Vol. 49, no. 4, April, electronic document.
[18] Gunasekera, J.S., Jia, Z., Malas, J.C. and Rabelo, L. (1998), Development of a neural
network model for a cold rolling process, Engineering Applications of Artificial
Intelligence, Vol. 11, pp. 597−603.
[19] Dixit, U.S. and Chandra, S. (2003), A neural network based methodology for the
prediction of roll force and roll torque in fuzzy form for cold flat rolling process,
International Journal of Advanced Manufacturing Technology, Vol. 22, pp. 883−889.
[20] Yang, Y.Y., Linkens, D.A., Talamantes-Silva, J. and Howard, I.C. (2003), Roll force
and torque prediction using neural network and finite element modelling, Transactions
of the Iron and Steel Institute of Japan, Vol. 43, pp. 1957−1966.
[21] Yang, Y.Y., Linkens, D.A., Talamantes-Silva, J. (2004), Roll load prediction— data
collection, analysis and neural network modelling, Journal of Materials Processing
Technology, Vol. 152, pp. 304−315.
[22] Osakada, A. and Yang G.B. (1991), Neural networks for process planning of cold
forging, CIRP Annals, Vol. 40, pp. 243−246.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 28 U. S. Dixit

[23] Raj, K.H., Sharma, R.S., Srivastava, S. and Patvardhan, C. (2000), Modelling of
manufacturing processes with ANNs for intelligent manufacturing, International
Journal of Machine Tools and Manufacture, Vol. 40, pp. 851−868.
[24] Mori, T. and Li, S. (2005), Determination of design parameters in metal forging process
by using FEM and neural network, Proceedings of the 10th India-Japan Joint Seminar,
February 21−26, 2005, IIT Kanpur.
[25] Kazan, R., Firat, M. and Tiryaki, A.E. (2009), Prediction of springback in wipe bending
process of sheet metal using neural network, Materials and Design, Vol. 30, pp.
418−423.
[26] Sbarbaro-Hofer, D., Neumerkel, D. and Hunt, K. (1993), Neural control of a steel
rolling mill, IEEE Transactions: Control Systems, Vol. 13, pp. 69−75.
[27] Gudur, P.P. and Dixit, U.S. (2009), An application of fuzzy inference for studying the
dependency of roll force and roll torque on process parameters in cold flat rolling,
International Journal of Advanced Manufacturing Technology, Vol. 42, pp.41–52.
[28] Gudur, P.P. and Dixit, U.S. (2008), A neural network-assisted finite element analysis of
cold flat rolling, Engineering Applications of Artificial Intelligence, Vol. 21, pp. 43–52.
[29] Panchal, J.H., Khanna, R. and Dixit, U.S. (2000), Optimization of turning process using
a neuro-fuzzy controller, Proceedings of Sixteenth National Convention of Mechanical
Engineers and All India Seminar on Future Trends in Mechanical Engineering,
Research and Development, Roorkee, September 29-30, 2000.
[30] Liu, S.-T. (2006), Optimization of a machining economics model with fuzzy exponents
and coefficients, International Journal of Production Research, Vol. 44, pp. 3083–
3104.
[31] Abburi, N.R. and Dixit, U.S. (2006), A knowledge-based system for the prediction of
surface roughness in turning process, Robotics and CIM, Vol. 22, pp. 363–372.
[32] Chen, J.C. and Black, J.T. (1997), A fuzzy-nets in-process (FNIP) systems for tool-
breakage monitoring in end-milling operations, International Journal of Machine Tools
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

and Manufacture, Vol. 37, pp. 783-800.

[33] Fang, X.D. and Jawahir, I.S. (1994), Predicting total machining performance in finish
turning using integrated fuzzy-set models of the machinability parameters,
International Journal of Production Research, Vol. 32, pp. 833–849.
[34] Fang, X.D., Fie, J. and Jawahir, I.S. (1996), A hybrid algorithm for predicting chip
form/chip breakability in machining, International Journal of Machine Tools and
Manufacture, Vol. 36, pp. 1093–1107.
[35] Hashmi K., El Baradie, M.A. and Ryan, M. (1999), Fuzzy-logic based intelligent
selection of machining parameters, Journal of Materials Processing Technology, Vol.
94, pp. 94−111.
[36] Iqbal, A., He, N., Li, L. and Dar, N.U. (2007), A fuzzy expert system for optimizing
parameters and predicting performance measures in hard-milling process, Expert
Systems with Applications Vol. 32, pp. 1020−1027.
[37] Susanto, V. and Chen, J.C. (2003), Fuzzy logic based in-process tool wear monitoring
system in face milling operations, International Journal of Advanced Manufacturing
Technology Vol. 3, pp. 186−192.
[38] Ali, Y.M. and Zhang, L.C. (2004) A fuzzy model for predicting burns in surface
grinding of steel, Journal of Material Processing Technology, Vol. 44, pp. 563−571.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Neural Networks and Fuzzy Sets to Machining and Metal Forming 29

[39] Liu, Q., Chen, X. and Gindy, N (2005) Fuzzy pattern recognition of AE signals for
grinding burn, International Journal of Machine Tools and Manufacture, Vol. 45, pp.
811−818.
[40] Biglari, F.R. and Fang, X.D. (1995) Real-time fuzzy logic control for maximizing the
tool life of small–diameter drills, Fuzzy sets and systems, Vol. 72, pp. 91−101.
[41] Chung B.M. and Tomizuka M. (2001) Fuzzy logic modeling and control for drilling of
composite laminates, 10th IEEE International Conference on Fuzzy Systems,
Melbourne, pp. 509−512.
[42] Hashmi K., Graham I.D. and Mills B. (2000), Fuzzy logic based data selection for the
drilling process, Journal of Materials Processing Technology, Vol.108, pp. 55−61.
[43] Dixit, U.S., Robi, P.S. and Sarma, D.K. (2002), A systematic procedure for the design
of a cold rolling mill, Journal of Materials Processing Technology, Vol. 121, pp.
69−76.
[44] Dixit, U.S. and Dixit, P.M. (2000), Application of fuzzy set theory in scheduling of
tandem rolling mills, ASME Journal of Manufacturing Science and Engineering, Vol.
122, pp. 494−500.
[45] Dixit, U.S. and Dixit, P.M. (1996), A finite element analysis of flat rolling and
application of fuzzy set theory. International Journal of Machine Tools and
Manufacture, Vol. 36, pp. 947−969.
[46] Shida, S. and Awazuhara, H. (1973), Rolling load and torque in cold rolling, Journal of
the Japan Society for Technology of Plasticity, Vol. 14, pp. 267−278.
[47] Gudur, P.P. and Dixit, U.S. (2009), An application of fuzzy inference for studying the
dependency of roll force and roll torque on process variables in cold flat rolling,
International Journal of Advanced Manufacturing Technology, Vol. 42, pp. 41−52.
[48] Skrjanc, I., Blazic, S. and Agamennoni, O. (2005), Interval fuzzy model identification
using l∞-norm, IEEE Transactions: Fuzzy Systems, Vol. 13, pp. 561−568.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[49] Li, G. and Janabi-Sharifi, F. (2009) Fuzzy looperless tension control for hot strip
rolling, Fuzzy Sets and Systems, Vol.160, pp. 521−536.
[50] Ong, S.K, De Vin, L.J., Nee, A.Y.C. and Kals, H.J.J. (1997), Fuzzy set theory applied
to bend sequencing for sheet metal bending, Journal of Materials Processing
Technology, Vol. 69, pp. 29–36.
[51] Osakada, K., Yang, G. and Mori, K. (1993), Determination of Optimum Forming Path
in Three-Roll Bending by Combination of Fuzzy Reasoning and Finite Element
Simulation, CIRP Annals— Manufacturing Technology, Vol. 42, pp. 291–294.
[52] Hsiang, S.H. and Lin Y.W. (2008), Application of fuzzy theory to predict deformation
behaviors of magnesium alloy sheets under hot extrusion, Journal of Materials
Processing Technology, Vol. 201, pp.138–144.
[53] Shiraishi, M., Nikawa, M., Kubota, T. and Goto (2007), Y., Prediction of curvatures of
parts extruded through rotary die using fuzzy inference, Journal of Materials
Processing Technology, Vol.187-188, pp. 702–705.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 2

MULTI-OBJECTIVE OPTIMIZATION OF MULTI-PASS

MILLING PROCESS PARAMETERS USING ARTIFICIAL
BEE COLONY ALGORITHM

R. Venkata Rao1∗ and P. J. Pawar2

1
Department of Mechanical Engineering,
S.V. National Institute of Technology,
Ichchanath, Surat, Gujarat – 395 007, INDIA
2
Department of Production Engineering,
K.K. Wagh Institute of Engineering Education and Research
Nashik – 422 003, Maharashtra, INDIA
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

ABSTRACT
The effective optimization of machining process parameters affects dramatically the
cost and production time of machined components as well as the quality of the final
products. This chapter presents the details of multi-objective optimization of a multi-pass
milling operation. The two objectives considered are maximization of production rate and
minimization of production cost subjected to various constraints of arbor strength, arbor
deflection, cutting power, and surface roughness.
Various cutting strategies are considered to determine the optimal process
parameters like the number of passes, depth of cut for each pass, cutting speed, and feed
per tooth. The optimization is carried out using a recently developed non-traditional
optimization algorithm namely, artificial bee colony (ABC) algorithm. An example is
presented and solved to illustrate the effectiveness of the algorithm. The results show that
the artificial bee colony algorithm can be effectively used in the multi-objective
optimization of multi-pass milling process.

∗
Corresponding author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 32 R. Venkata Rao and P. J. Pawar

INTRODUCTION
In today’s manufacturing environment many large industries have attempted to introduce
the flexible manufacturing system (FMS) as a strategy to adapt to the ever-changing
competitive market requirement. The flexible manufacturing system involves highly
automated and computer controlled machines. Due to high capital and machining costs, there
is an economic need to operate these machines as efficiently as possible in order to obtain the
required pay back. The success of any machining operation depends on the selection of
machining process parameters. Proper selection of process parameters play a significant role
to ensure quality of product, to reduce the machining cost, to increase productivity in
computer controlled machining processes and to assist in computer aided process planning. A
human process planner selects machining process parameters using his own experience or
from the handbooks. But these parameters do not give optimal result. Various optimization
strategies and algorithms ranging from elementary numerical search methods to more
systematic approaches employing non-traditional techniques for optimization of process
parameters in case of single pass milling operation had been reported in the literature.
However, as multi-pass operations are often preferred to single pass operations for economic
reasons, recent efforts have been directed towards determination of optimal machining
conditions for multi-pass operations.
Traditionally, mathematical programming techniques like linear programming, method of
feasible direction, dynamic programming and geometric programming had been used to solve
optimization problems in milling. However, these traditional methods of optimization do not
fare well over a broad spectrum of problem domains. Moreover, traditional techniques may
not be robust. Numerous constraints and multiple passes make machining optimization
problems complicated and hence these techniques are not ideal for solving such problems as
they tend to obtain a local optimal solution.
Considering the drawbacks of traditional optimization techniques, attempts are being
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

made to optimize the machining problem using evolutionary optimization techniques.

Evolutionary computation consists of a variety of methods including optimization paradigms
that are based on evolution mechanisms such as biological genetics and natural selection.
These methods use the fitness information instead of the functional derivatives making them
more robust and effective. These methods thus avoid the problem of getting trapped in local
optima and enable to obtain a global (or nearly global) optimum solution. Efforts are
continuing to use more recent optimization algorithms, which are more powerful, robust and
able to provide accurate solution. Artificial bee colony (ABC) algorithm developed by
Karaboga [1] and Karaboga and Basturk [2, 3] is one of the recent algorithms and no effort
has been yet made for optimization of process parameters of any of the machining processes
by using this algorithm. This chapter presents an application of artificial bee colony algorithm
(ABC) for multi-objective optimization of process parameters of multi-pass milling operation.
Milling is the machining process in which the metal is removed by a rotating multiple
tooth cutter. Milling operation can be performed in a single pass or in multi-passes. Multi-
pass operations are often preferred to single pass operations for economic reasons and are
generally used to machine stocks that cannot be removed in a single pass. In case of face
milling operation a significant amount of stock material is to be removed and hence work
piece is required to be machined using multi-pass operations. Determination of the optimal

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 33

cutting parameters like number of passes, depth of cut for each pass, speed, and feed is
considered crucial of multi-pass milling operation. Various investigators had used
optimization techniques, both traditional and non-traditional, for optimization of multi-pass
milling operation in the past.
Shin and Joo [4] used dynamic programming optimization method for milling process
parameter optimization. Dynamic programming can solve both continuous and discrete
variables and yield a global optimal solution. However, if the optimization problem involves
a large amount of independent parameters with a wide range of values such as the cutting
parameters in milling operation, the use of dynamic programming is limited. Wang [5] used a
neural network based approach to optimize milling process parameters. However,
optimization by using neural networks may often ends in local minima or fails to converge on
a desired result.
Tolouei-Rad and Bidhendi [6] used the method of feasible direction and considered
maximization of profit rate as an objective function in milling operation. The feasible solution
denotes the local minimum of the problem. However, this local minimum need not be the
global one unless the problem is convex programming problem. Optimization model
developed in their work was non-convex.
Sonmez et al. [7] studied multi-pass milling operation based on the maximum production
rate criterion and used an algorithm adopted from the study of Agapiou [8] which was
proposed for the multi-pass turning operations. The authors had developed a strategy to
determine the optimum cutting parameters for multi-pass milling based on the maximum
production rate criterion. The optimum number of passes was determined via dynamic
programming, and the optimal values of the cutting conditions were found based on the
objective function developed for the typified criterion by using geometric programming.
Although the results showed significant improvement over handbook recommendations, the
optimization techniques used in their work either tend to result in local minima or take a long
time to converge on a reasonable result.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Shunmugam et al. [9] used genetic algorithm (GA) for milling process parameter
optimization with total production cost as the objective function. The authors had divided the
cost into two separate minimization problems of roughing and finishing operations. GA was
used for finding the optimum number of rough passes and allocation of the total stock in each
of the rough and the finish passes to achieve the minimum total production cost. However,
although GA has advantages over the traditional techniques, it has the drawbacks of lack of
ability to perform good local search and premature convergence.
During the past decade, different optimization methods had been integrated to improve
performance of algorithms and to reach the global optimum results. In order to optimize the
machining parameters, the evolutionary methods had been modified or hybridized by using
other optimization techniques. Liu and Wang [10] modified the genetic algorithm by defining
and changing the operating domain and used for optimization of milling parameters. The
results and the convergence speed of their approach were better than that of genetic
algorithm. Wang et al. [11] proposed a new hybrid approach, named genetic simulated
annealing (GSA), based on genetic algorithm and simulated annealing to find optimal
machining parameters in milling operations. In their approach, genetic algorithm and
simulated annealing were combined. They pointed out that the results obtained were found to
be better than those of genetic algorithm and geometric programming.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 34 R. Venkata Rao and P. J. Pawar

Baskar et al. [12] considered a specific case in milling operation and solved the same by
using three different non-traditional optimization techniques comprising a global search
algorithm (genetic algorithm), local search algorithm (local hill climbing) and memetic
algorithm. Significant improvement was obtained with the above techniques in comparison
with the results given by handbook and the method of feasible direction. This was because of
the limitations of enumerative search followed by the direct search method.
Onwubolu [13] proposed a new optimization technique based on Tribes for determination
of the cutting parameters in multi-pass milling operations such as plain milling and face
milling by simultaneously considering multi-pass rough machining and finish machining. The
optimum milling parameters were determined by the maximum production rate criterion
subjected to several practical technological constraints. The results obtained confirmed the
well known fact that multi-pass approach is preferred to single-pass approach if the total
depth of cut to be removed in milling exceeds the maximum allowable depth of cut. Although
the results obtained in his work using tribes showed significant improvement over other
traditional and non-traditional algorithms, but the results are not valid as some of the
constraints in the solution obtained are violated.
Yildiz [14] developed a new hybrid optimization approach by hybridizing the immune
algorithm with hill climbing local search algorithm to maximize the total profit rate in milling
operations. The results showed that the hybrid approach was more effective to optimize the
cutting parameters for milling operations than genetic algorithm, the feasible direction
method, and the handbook recommendations. Zarei et al. [15] presented a harmony search
(HS) algorithm to determine the optimum cutting parameters for multi-pass face-milling.
Total production cost was considered as the objective function.
It is also revealed from the literature that few efforts has been made for multi-objective
optimization of some machining processes like drilling [16], grinding [17, 18] and turning
[19]. However, no effort has been yet made for multi-objective optimization of multi-pass
milling process.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

The literature related to optimization of milling process is thus mainly concerned with
single objective optimization only (considering either production cost or production time),
with single pass or multi-pass operation. This chapter therefore provides an approach to
develop a methodology for multi-objective optimization of multi-pass milling process with
optimization using artificial bee colony algorithm. The two objectives considered are (1)
maximization of production rate (i.e. minimization of production time) and (2) minimization
of production cost. Although a low production time would mean a low production cost, it
should be realized that machining parameters giving minimum production time would not be
identical to those giving minimum production cost. Hence it is necessary to determine the
compromise best solution which would ultimately lead to highest benefit in terms of both cost
and productivity. Various practical constraints considered are, arbor strength, arbor
deflection, cutting power, and surface roughness. Feed per tooth, cutting speed and depth of
cut are considered as process parameters. The upper and lower bounds of the process
parameters are also included in the study.
The next section presents a multi-objective optimization model of multi-pass milling
process.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 35

MULTI-OBJECTIVE OPTIMIZATION MODEL OF MULTI-PASS

MILLING PROCESS
The objective functions and the constraints are formulated as discussed below.

Objective Functions

Two objectives considered are:

(1) Maximization of production rate ( i.e. minimization of production time)

(2) Minimization of production cost

Maximization of Production Rate

Production rate is maximized by minimizing the production time. Equation (1) gives the
production time for multi-pass milling operation [7].

⎛1 ⎞ e ⎛ uv ⎞ ⎛ nv ⎞
⎜ −1 ⎟ v ⎜ −1 ⎟ rv ⎜ −1 ⎟ q v
Np
πDL Td πLV ⎝m ⎠
am ⎝m ⎠
f m ⎝m ⎠
a z λ m
TPr = TP + TL + N pTa + ∑ + i i zi
⎛ bv ⎞
r s
(1)
f zi z1000Vi
× (B B B B )
i =1 1 ⎜ −1 ⎟ 1
m ⎝m ⎠
1000C D v m h p t
m

where, TP = machine preparation time; TL = Loading and unloading time; Np = Number of

passes; Ta = Process adjusting and quick return time; D = cutter diameter; L= length of cut; Td
= Tool changing time; fz = feed per tooth; z = Number of teeth; V = cutting speed; a = depth
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

of cut, ar = width of the cut, Bm, Bk, Bp, Bt = correction coefficients, m, ev, uv, rv, nv, qv, bv, =
exponents, Cv = process constant, λs = cutting inclination angle.
Also, let,

TPr = T1 + T2 (2)

where,

Ts
T1 (min) = + TL + N pTa (3)
Nb

⎛1 ⎞ ⎛ uv ⎞ r ⎛ nv ⎞
⎜ −1 ⎟ ev ⎜ −1 ⎟ v ⎜ −1 ⎟ qv
πDL Td πLVi λ
Np ⎝m ⎠ ⎝m ⎠ m ⎝m ⎠
a m
f a z m
T2 (min) = ∑ + i zi r s
f zi z1000Vi ⎛ bv ⎞ (4)
× (B B B B )
i =1 1 ⎜ −1 ⎟ 1
⎝m ⎠
1000Cv m D m h p t
m

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 36 R. Venkata Rao and P. J. Pawar

Minimization of Production Cost

Equation (5) gives the production cost in multi-pass milling operation [9].

Np
U t = U f + ∑ U ri + A4 (5)
i =1

where, A4 = Cost of tool preparation, = ko Tp; ko = Overhead cost,

Cost per pass is given by (U)

tm ⎛ zT ⎞
U ($) = kotm + (kt z ) + kotm ⎜⎜ d ⎟⎟ + ko ( Lh1 + h2 ) (6)
TR ⎝ TR ⎠

πDL
Machining time = tm = (7)
1000 zf zV

kt = cost of cutting edge;

⎛ bv ⎞

× (Bm Bh B f Bt )
1 ⎜⎜ m ⎟⎟ 1
⎝ ⎠
Cv D m m
TR= tool replacement life = ⎛ nv ⎞
(8)
⎛1⎞
⎜ ⎟ ev ⎛ uv ⎞ rv ⎜⎜ ⎟⎟ qv
V ⎝m⎠
a m
fz ⎜ m⎟
⎝ ⎠ ar m ⎝ m
z ⎠
λs m

h1= tool returns time; h2 = Tool advance/return time;

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Constraints

Following four constraints are considered in this optimization model:

Arbor Strength
The arbor is subjected to torsion from the action of resistance to cutting. Therefore, the
selected values of process parameters should ensure that the arbor is safe from strength point
of view as specified by equation (9).

Fs - Fc ≥ 0 (9)

b e u
where, Fc = mean peripheral cutting force = C zp ar zD z a z f z z (10)
Czp=, process constants, bz, ez, and uz are exponents.

Permissible force for arbor strength (Kg) = Fs

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 37

0.1kb d a3
Fs = (11)
0.08 La + 0.65 ((0.25L ) a
2
+ (0.5αD ) 2 )
where, kb = permissible bending strength of arbor; da = arbor diameter; La = arbor length
between supports; α = kb / (1.3 kt); kt = permissible torsional strength of arbor.

Arbor Deflection
The selected values of process parameters should be checked for arbor deflection as
given by equation (12).

Fd - Fc ≥0 (12)

4 Eed a4
where, permissible force for arbor deflection (Kg) = Fd = (13)
L3a
where, E = modulus of elasticity of arbor material; e = permissible value of arbor deflection.

For rough milling operation, e=0.2 mm and for finish milling operation, e=0.05 mm.

Power
Power required for the cutting operation should not exceed the effective power
transmitted to cutting point by the machine tool. This is ensured by equation (14).

FcV
Pc − ≥0 (14)
6120
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

where, Pc = Cutting power (KW) = Pm * η

Pm = nominal motor power, η=overall efficiency.

Surface Roughness Constraint

For roughing operation the surface roughness value should not exceed the specified limit
of 25 µm as given by equation (15).

0.0321 f z2
25 × 10− 3 − ≥0 (15)
re

where, re = nose radius = 0.999 mm

For finishing operation the surface roughness value should not exceed the specified limit
of 2.5 µm as given by equation (16).

0.0321 f z2
2.5 × 10− 3 − ≥0 (16)
re

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 38 R. Venkata Rao and P. J. Pawar

Process Variables
The three process variables considered are depth of cut (a), feed per tooth (fz) and cutting
speed (V). The bounds for these variables are given by equations below.

1≤ a ≤ 4 (mm) (17)
0.1 ≤ fz ≤ 0.6 (mm/tooth) (18)
50 ≤ V ≤ 300 (m/min) (19)

The next section describes the artificial bee colony algorithm as an optimization
technique.

ARTIFICIAL BEE COLONY ALGORITHM

A branch of nature inspired algorithms, called swarm intelligence, is focused on insect
behavior in order to develop some meta-heuristics which can mimic insect's problem solution
abilities. Interaction between insects contributes to the collective intelligence of the social
insect colonies.
These communication systems between insects have been adapted to scientific problems
for optimization. The foraging behavior, learning, memorizing and information sharing
characteristics of honeybees have recently been one of the most interesting research areas in
swarm intelligence. Artificial bee colony (ABC) algorithm is developed to model the
intelligent behaviors of honeybee swarms .The honeybee swarms consists of two essential
components (i.e. food sources and foragers) and defines two leading modes of the behavior
(i.e. recruitment to a nectar source and abandonment of a source).
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Food Sources

The value of a food source depends on different parameters such as its proximity to the
nest, richness of energy and ease of extracting this energy. For the simplicity, the
‘‘profitability’’ of a food source can be represented with a single quantity.

Foragers

Foragers can be unemployed, employed or experienced.

Unemployed Foragers

If it is assumed that a bee has no knowledge about the food sources in the search field,
bee initializes its search as an unemployed forager. There are two possibilities for an
unemployed forager:

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 .

Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 39

• Scout Bee: If the bee starts searching spontaneously without any knowledge, it will
be a scout bee. The percentage of scout bees varies from 5% to 30% according to the
information into the nest. The mean number of scouts averaged over conditions is
about 10%.
• Recruit: If the unemployed forager attends to a waggle dance done by some other
bee, the bee will start searching by using the knowledge from waggle dance.

Employed Foragers

When the recruit bee finds and exploits the food source, it becomes an employed forager
and memorizes the location of the food source. After the employed foraging bee loads a
portion of nectar from the food source, it returns to the hive and unloads the nectar to the food
area in the hive. There are three possible options related to residual amount of nectar for the
foraging bee. If the nectar amount decreased to a low level or exhausted, foraging bee
abandons the food source and become an unemployed bee. If there are still sufficient amount
of nectar in the food source, it can continue to forage without sharing the food source
information with the nest mates or it can go to the dance area to perform waggle dance for
informing the nest mates about the same food source. The probability values for these options
highly related to the quality of the food source.

Experienced Foragers

These types of foragers use their historical memories for the location and quality of food
sources. This type of forages can be an inspector, which controls the recent status of food
source already discovered. It can also be a reactivated forager by using the information from
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

waggle dance. It tries to explore the same food source discovered by self if there are some
other bees confirm the quality of same food source. It can also be scout bee to search new
patches if the whole food source is exhausted. It can also be a recruit bee, which is searching
a new food source declared in dancing area by another employed bee.
Communication among bees related to the quality of food sources occurs in the dancing
area. The related dance is called waggle dance. Since information about all the current rich
sources is available to an onlooker on the dance floor, she probably could watch numerous
dances and choose to employ herself at the most profitable source. There is a greater
probability of onlookers choosing more profitable sources since more information is
circulating about the more profit able sources. Employed foragers share their information
with a probability, which is proportional to the profitability of the food source, and the
sharing of this information through waggle dancing is longer in duration. Hence, the
recruitment is proportional to profitability of a food source. The detailed steps of artificial bee
colony algorithm are explained in section 5.
The nest section provides an example of multi-objective multi-pass milling process to
demonstrate and validate the application of the artificial bee colony algorithm.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 40 R. Venkata Rao and P. J. Pawar

EXAMPLE
Now, the following experimental data are considered to demonstrate and validate the
application of artificial bee colony algorithm for the optimization of process parameters of the
multi-pass milling operation. The data considered for the example are as given below:

Motor power (Pm) = 5.5 kW, efficiency, η = 0.7

Arbor diameter, da = 24 mm, arbor length between supports, La = 210 mm
Permissible bending stress of arbor, kb: 100 MPa
Permissible torsional stress of arbor, kt: 85 MPa
Modulus of elasticity of arbor material, E = 200 GPa,
Tool material: HSS, tool diameter, D = 120 mm, number of teeth, z = 8,
Material: structural carbon steel
Tensile strength: 750 MPa, Brinell hardness number = 150
Length of cut, L = 200 mm, width of cut, ar = 50 mm, depth of cut, a = 5 mm
Machine preparation time, Tp = 0.75 min.
Loading and unloading time of one work piece, TL = 1.5 min.
Tool change time, Td = 5 min.
Process adjusting and quick return time, Ta = 0.1 (min/part)
Cutting inclination λs= 30°
Overhead cost, ko = 0.5 ($/min), Cost of cutting edge, kt = 2.5 ($/cut edge)
Tool returns time, h1 =7×10−4 min/mm, Tool advance time, h2 = 0.3 min
Constants: Bm=1, Bk=1, Bp=0.8, Bt=0.8, m=0.33, ev=0.3, uv=0.4, rv=0.1, nv=0.1, qv=0,
Cv=35.4, bv=0.45, Czp = 68.2, bz = -0.86, ez = 0.86, and uz = 0.72.

Now various steps of ABC algorithms are applied as described below:

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Step 1: Parameter Selection

As discussed in the description of ABC algorithm, food source represents a possible

solution to the problem of multi-objective optimization with minimization of production time
and minimization of cost as considered in this chapter. Number of initial solutions (i.e. the
number of food sources) considered for the given problem are five. The value of each food
source depends on the fitness value of the objective function given by equations (1) and (5).
For every food source there is only one employed bee (employed forager). In other
words, the number of employed bees is equal to number of food sources. Thus, the number of
employed bees is considered to be five. The unemployed forager can be scout or an onlooker
bee. The number of onlooker bees must be greater than the number of employed bees. The
effect of number of onlooker bees on the convergence of solution is shown in Figure 1. It is
observed that as the number of onlooker bees (n0) and hence the population size increases, the
algorithm perform better in terms of convergence rate. However, after a sufficient value of
number of onlooker bees, any increment in the value does not improve the performance of the
algorithm.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 41

Figure 1. Effect of number of onlooker bees (no) on convergence rate for rough milling.

For the problem considered in this chapter, number of onlooker bees is considered to be
eleven, which can provide an acceptable convergence speed for search. The colony size is the
sum of number of employed bees and number of onlooker bees. Hence the colony size is
sixteen. Number of scout bees is usually 5-30% of the colony size. The number of scout bee is
taken as 5% of the colony size i.e. one. The parameters of optimization thus selected are
summarized as below:
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

• Number of employed bees=5

• Number of onlookers bees= 11
• Number of scout bees=1
• Maximum number of iterations= 150

Step 2: Calculate the Nectar Amount of Each Food Source

The employed bees are moved to the food sources and the nectar amount of these food
sources is evaluated based on their fitness value as defined by the objective function given by
equation (1) and (5) subjected to constraints given by equations (9),(12), (14), (15) and (16).

Step 3: Determine the Probabilities by Using the Nectar Amount

If the nectar amount of a food source “θi” is Fi, then the probability (Pi) of choosing this
food source by an onlooker bee is expressed as:

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 42 R. Venkata Rao and P. J. Pawar

∑ (1/ f k ) −1
Pi = k =1
(20)
fi

where “S” is the number of food sources.

Step 4: Calculate the Number of Onlooker Bees, Which Will Be Sent to Food
Sources

Based on the probabilities calculated in step 3, the number (N) of onlookers bees sent to
food source “θi” is calculated as:

N = Pi * m. (21)

where, ‘m’ is the total number of onlooker bees.

Step 5: Calculate the Fitness Value of Each Onlooker Bee

After watching the dances of employed bees, an onlooker bee goes to the region of food
source “θi” by the probability given by equation (20). The position of the selected neighbor
food source is calculated as the shown in equation (22).
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

θi (c+1) = θi ( c) ± φi (c) (22)

where, “c” is number of generation. φi (c) is a randomly produced step to find a food source
with a more nectar around ‘θi'. φi (c) is calculated by taking the difference of the same parts of
θi(c) and θk(c) (“k” is a randomly produced index) food positions. If the nectar amount
Fi(c+1) at θi(c+1) is higher than at θi(c), then the bees go to the hive and share information
with others and the position θi(c) of the food source is changed to θi(c+1) otherwise θi(c) is
kept as it is. If the position ‘θi' of the food source “i” cannot be improved through the
predetermined number of trials, then that food source ‘θi' is abandoned by its employed bee
and then the bee becomes a scout. The scout starts searching new food source, and after
finding the new source, the new position is accepted as ‘θi'.

Step 6: Evaluate the Best Solution

Position of the best onlooker bee is identified for each food source. The global best of the
honeybee swarm in each generation is obtained and it may replace the global best at previous
generation if it has better fitness value.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 43

Step 7: Update the Scout Bee

The worst employed bees, as many as the number of scout bees in the population, are
respectively compared with the scout solutions. If the scout solution is better than employed
solution, employed solution is replaced with scout solution. Else employed solution is
transferred to the next generation without any change.
Using ABC algorithm, various feasible cutting strategies are adopted to determine the
optimum number of passes required (i.e. number of roughing and finishing passes) and depth
of cut for each pass.
The results of optimization are described below:

1. Determination of optimum process parameters of finish milling considering only

production rate as the objective.
Table 1 shows the results of optimization of finish milling operation considering only
production rate as the objective. It may be observed that all the constraints are satisfied. For
finish milling, values of depth of cut higher than 1.5 are not considered as for these values the
deflection constraint is violated for any combination of feed and speed in specified range.

Table 1. The results of optimization of finish milling operation considering only

production rate as the objective

a fz V T2 SC DC PC RC
1 0.213 92.727 0.711 64.0751 0.0751 1.6390 0.001044

1.5 0.131 100.189 1.071 64.2630 0.2630 1.4448 0.001949

a: depth of cut (mm); fz: feed per tooth (mm/tooth); V: cutting speed (m/min);
SC, DC, PC and RC are the values of left hand side of equations (9), (12), (14) and (16) related to
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

strength, deflection, power and roughness constraints respectively.

2. Determination of optimum process parameters of finish milling considering only

production cost as the objective
Table 2 shows the results of optimization of finish milling operation considering only
production cost as the objective.

Table 2. The results of optimization of finish milling operation considering only cost as
the objective

a fz V U SC DC PC RC
1 0.212 50 1.199 64.5688 0.5687 2.6618 0.00106
1.5 0.130 50 1.639 65.0649 1.0649 2.6658 0.00196

3. Multi-objective optimization of finish milling

The combined objective function of finish milling (ZF) is formulated as given by equation
(23).

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 44 R. Venkata Rao and P. J. Pawar

T2 U
Z F = w1 × + w2 × (23)
d1 d2

where, w1 and w2 are the weights assigned to the objectives of production rate and production
cost, respectively. The values of weights can be calculated by using the analytic hierarchy
process [20]. However, in the present example, equal weights are assumed. d1 and d2 are the
minimum values of production time and production cost respectively for a given depth of cut,
obtained in step 1 and step 2 above. Table 3 shows the results of multi-objective optimization
for finish milling operation for a=1 mm {d1=0.711; d2=1.199} and for a=1.5 mm{d1=1.071;
d2=1.639}.

Table 3. The results of multi-objective optimization for finish milling operation

a fz V T2 U ZF SC DC PC RC
1 0.211 57.08 0.869 1.311 1.156 65.063 1.063 2.498 0.0010

1.5 0.131 59.58 1.332 1.821 1.174 64.263 0.263 2.431 0.0019

4. Determination of optimum process parameters of rough milling considering only

production rate as the objective
Table 4 shows the results of optimization of rough milling operation considering only
production rate as the objective. It may be observed that all the constraints are satisfied. For
rough milling, values of depth of cut higher than 2.5 are not considered as for these values the
strength constraint is violated for any combination of feed and speed in specified range.

Table 4. The results of optimization of rough milling operation considering only

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

production rate as the objective

a fz V T2 SC DC PC RC

1 0.353 50.00 0.608 0.0607 374.061 2.1348 0.0210

1.5 0.217 81.28 0.794 0.0552 374.055 1.0617 0.0234

2 0.154 85.91 1.064 0.3008 374.301 0.9063 0.0242

2.5 0.118 89.34 1.335 0.2609 374.261 0.7881 0.0245

a: depth of cut (mm); fz: feed per tooth (mm/tooth); V: cutting speed (m/min);
SC, DC, PC and RC are the values of left hand side of equations (9), (12), (14) and (15) related to
strength, deflection, power and roughness constraints respectively.

5. Determination of optimum process parameters of rough milling considering only

production cost as the objective
Table 5 shows the results of optimization of rough milling operation considering only
production cost as the objective.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 45

Table 5. The results of optimization of rough milling operation considering only cost as
the objective

a fz V U SC DC PC RC

1 0.352 50.00 1.083 0.489 374.489 2.138 0.021

1.5 0.217 50.00 1.431 0.402 374.403 2.137 0.023
2 0.154 50.00 1.771 0.300 374.301 2.136 0.024
2.5 0.117 50.00 2.109 1.542 375.542 2.146 0.024

6. Multi-objective optimization of rough milling

Now, the combined objective function for rough milling is formulated as given by
equation (24) which is similar to equation (23).

T2 U
Z R = w1 × + w2 × (24)
d1 d2

In case of rough milling, Table 6 shows the results of multi-objective optimization for
rough milling operation for a=1mm {d1=0.608; d2=1.083}, for a=1.5mm{d1=0.794;
d2=1.431}; for a=2mm: {d1=1.064; d2=1.771} and for a=2.5mm: {d1=1.335; d2=2.109}.

Table 6. The results of multi-objective optimization for rough milling operation

a fz V T2 U ZR SC DC PC R
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

1 0.352 50.00 0.610 1.084 1.000 0.489 374.489 2.1380 0.0210

1.5 0.217 50.00 0.965 1.432 1.106 0.402 374.402 2.1376 0.0234

2 0.154 52.91 1.288 1.852 1.127 0.300 374.300 2.0370 0.0242

2.5 0.117 55.41 1.620 2.284 1.145 1.542 375.542 1.9620 0.0245

7. Formulation of equivalent objective function (Zeq) for finish milling and rough milling
operations
The equivalent objective function Zeq is written as:

Zeq = (T2)eq + (U)eq (25)

where (T2)eq = T2/ (T2min)a=1; (U)eq = U/(Umin)a=1

The values of equivalent objective function for various depths of cut in finish milling
operation are given in Table 7 and that for rough milling operation are given in Table 8.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 46 R. Venkata Rao and P. J. Pawar

Table 7. Equivalent objective function for finish milling

a fz V (T2)eq (U)eq (Zeq)finish

1 0.211 57.08 1.222 1.092 2.314
1.5 0.131 59.58 1.873 1.517 3.390

Table 8. Equivalent objective function for rough milling

a fz V (T2)eq (U)eq (Zeq)finish

1 0.352 50.00 1.000 1.000 2.000
1.5 0.217 50.00 1.587 1.321 2.908
2 0.154 52.91 2.118 1.708 3.826
2.5 0.117 55.41 2.664 2.107 4.771

8. Determination of total equivalent objective function

The total equivalent objective function (Zeq)total is then obtained for various possible
combinations of depth of cut in finish and rough milling operation, to cut a total depth of cut
of 5 mm.

(Zeq)total = (Zeq)finish+(Zeq)rough. (26)

9. Evaluation of best solution

Table 9 shows various possible combinations of depth of cut in finish and rough milling
operation. As shown in Table 9, the best solution is the one having minimum value of total
equivalent objective function. It is observed from Table 9, that from production rate point of
view, combination 1 {i.e. 1, 1, 1, 1, 1} is better, whereas, from cost point of view,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

combination 4 is better {i.e. 1, 2, 2}. The compromise best solution is 4 having minimum
equivalent value (Zeq) of 9.966.

Table 9. Possible combinations to cut a total depth of 5 mm

S.N. afinish arough arough arough arough Zeq (T2)total Utotal

1 1 1 1 1 1 10.314 3.309 5.647
2 1 1.5 1.5 1 10.130 3.409 5.259
3 1 2 1 1 10.140 3.377 5.331
4 1 2 2 9.966 3.445 5.015
5 1 2.5 1.5 9.993 3.454 5.027
6 1.5 2 1.5 10.124 3.585 5.105

Application of artificial bee colony (ABC) algorithm to solve the considered multi-
objective multi-pass milling process leads to the following optimum solution.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 47

• Total number of passes required = 3.

• Number of rough cutting passes = 2 with depth of cut in each pass = 2 mm.
• Number of finish pass = 1 with depth of cut in finishing pass = 1 mm.
• Feed per tooth for a = 2 mm in roughing operation = 0.154 mm/tooth
• Cutting speed for a = 2 mm in roughing operation = 52.91 m/min
• Feed per tooth for a = 1mm in finishing operation = 0.211 mm/tooth
• Cutting speed for a = 1 mm in finishing operation = 57.08 m/min
• Total production time = Tpr = T1 + T2 = 2.550 + 3.445 = 5.995 min
• Total production cost = U + A4 = 5.015 + 0.375 = $ 5.39

Figure 2. Convergence of ABC algorithm for finish milling.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 3. Convergence of ABC algorithm for rough milling

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 48 R. Venkata Rao and P. J. Pawar

As shown in Figures 2 and 3, the convergence rate of artificial bee colony algorithm is
very high and it requires little iteration for convergence to the optimal solution.

CONCLUSION
Multi-pass operations are often preferred to single pass operations for economic reasons.
Determination of optimal process parameters such as the number of passes, depth of cut for
each pass, speed, and feed is considered crucial in multi-pass machining. The effective
optimization of the process parameters affects dramatically the cost and production time of
machined components as well as the quality of the final products. In this chapter, multi-
objective optimization aspects of multi-pass milling operation are considered. The two
objectives considered are minimization of production time (i.e. maximization of production
rate) and minimization of production cost subjected to the various practical constraints such
as arbor strength, arbor deflection, cutting power, and surface roughness. Process parameters
considered are feed per tooth, cutting speed, and depth of cut with their upper and lower
bound values.
The performance of artificial bee colony (ABC) algorithm is studied in terms of
convergence rate and accuracy of the solution. The algorithm is applied to obtain the
optimum process parameter values for various selected cutting strategies. The optimum
strategy is selected based on the compromise best of production time and production cost.
The convergence rate of artificial bee colony algorithm is very high and it requires only few
iterations for convergence to the optimal solution. The artificial bee colony algorithm can also
be easily modified to suit optimization of process parameters of other machining processes
such as grinding, turning, drilling, etc.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

REFERENCES
[1] Karaboga, D. (2005). An idea based on honey bee swarm for numerical optimization.
Technical Report-TR06’. Computer Engineering Department, Erciyes University.
[2] Karaboga, D, and Basturk, B. (2007). A powerful and efficient algorithm for numerical
function optimization: artificial bee colony (ABC) algorithm. Journal of Global
Optimization, 39(3), 459-471.
[3] Karaboga, D., and Basturk B. (2008). On the performance of artificial bee colony
(ABC) algorithm. Applied Soft Computing, 8, 687-697.
[4] Shin, Y.C, and Joo, Y.S. (1992). Optimization of machining conditions with practical
constraints. International Journal of Production Research, 30, 2907-2919.
[5] Wang, J. (1993). Multiple objective optimization of machining operations based on
neural networks. International Journal of Advanced Manufacturing Technology, 8, 235-
243.
[6] Tolouei-Rad, M., and Bidhendi I.M. (1997). On the optimization of machining
parameters for milling operations. International Journal of Machine Tools and
Manufacture, 37, 1–16.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimization of Multi-Pass Milling Process Parameters.. 49

[7] Sonmez, A.I., Baykasoglu, A., Dereli, T., and Filiz, I.H. (1999). Dynamic optimization
of multi-pass milling operations via geometric programming. International Journal of
Machine Tools and Manufacture. 39, 297–32.
[8] Agapiou, J.S. (1992). The optimization of machining operations based on a combined
criterion – Part 2: Multi-pass operations. Journal of Engineering for Industry, 114, 508–
513.
[9] Shunmugam, M.S., and Reddy. S.V.B. (2000). Narendran AA. Selection of optimal
conditions in multi-pass face-milling using a genetic algorithm. International Journal of
Machine Tools and Manufacture, 40, 401–414.
[10] Lui, Y.M., and Wang. C.J., (1999). A modified genetic algorithm based optimization of
milling parameters. International Journal of Advanced Manufacturing Technology, 15,
796-809.
[11] Wang, Z.G., Rahman M., Wong, Y.S., and Sun, J. (2005). Optimization of multi-pass
milling using parallel genetic algorithm and parallel genetic simulated annealing.
International Journal of Machine Tools and Manufacture, 45, 1726–1734.
[12] Baskar, N., Asokan, P., Saravanan, R., and Prabhaharan, G. (2006). Selection of
optimal machining parameters for multi-tool milling operations using a memetic
algorithm. Journal of Material Processing Technology, 174, 239-249.
[13] Onwubolu, G.C. (2006). Performance-based optimization of multi-pass face milling
operations using Tribes. International Journal of Machine Tools and Manufacture, 46,
717–727.
[14] Yildiz, A.R., (2009). A novel hybrid immune algorithm for optimization of machining
parameters in milling operations. Robotics and Computer-Integrated Manufacturing,
25(2), 261-270.
[15] Zarei, O, Fesanghary, M., Farshi, B., Saffar, R.J., and Razfar, M.R.
(2009).Optimization of multi-pass face-milling via harmony search algorithm. Journal
of Materials Processing Technology, 209, 2386-2392.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[16] Davim, J.P., and Antonio, C.A.C. (2001). Optimization of cutting conditions in
machining of aluminium matrix composites using a numerical and experimental model.
Journal of Material Processing Technology, 112, 78–82.
[17] Baskar, N., Saravanan, R., Asokan, P., and Prabhaharan, G. (2004). Ant colony
algorithm approach for multi-objective optimization of surface grinding operations.
International Journal of Advanced Manufacturing Technology, 23, 311–317.
[18] Saravanan, R., and Sachithanandam, M. (2001). Genetic algorithm (GA) for
multivariable surface grinding process optimization using a multi-objective function
model. International Journal of Advanced Manufacturing Technology, 17(5), 330–338.
[19] Abburi, N.R. and Dixit, U.S. (2007). Multi-objective optimization of multipass turning
processes. International Journal of Advanced Manufacturing Technology, 32, 902–910.
[20] Rao, R.V. (2007). Decision making in the manufacturing environment using graph
theory and fuzzy multiple attribute decision making methods. London, Springer.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 3

OPTIMIZATION OF ABRASIVE FLOW MACHINING

PROCESS PARAMETERS USING PARTICLE SWARM
OPTIMIZATION AND SIMULATED ANNEALING
ALGORITHMS

P. J. Pawar1, R. Venkata Rao∗1 and J. P. Davim2

1
Department of Mechanical Engineering,
S.V. National Institute of Technology, Ichchanath, Surat,
Gujarat – 395 007, INDIA
2
Department of Mechanical Engineering, University of Aveiro,
Campus Santiago, 3810-193 Aveiro, PORTUGAL
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

ABSTRACT
With the development of industry manufacturing technology, fine surface finish is in
high demand in a wide spectrum of industrial applications. Abrasive flow machining
(AFM) is an advanced finishing process capable of producing excellent surface finish of
the order of few nanometers. AFM process is used to deburr, polish, radius, and remove
recast layers of critical components in aerospace, automotive, electronic, and die-making
industries. The AFM process is becoming popular due to its ability to give repeatable and
consistent results. However, as the surface finish requirement increases, the operational
cost of these processes increases exponentially. Quality, cost, time and efficiency of these
processes can be improved significantly by choosing the optimum values of their process
parameters.
This paper presents two advanced optimization algorithms known as particle swarm
optimization (PSO) and simulated annealing (SA) to find the optimal combination of
process parameters of AFM process. The results of the proposed algorithms are compared
with the previously published results.

∗
Corresponding author, e-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 52 P. J. Pawar, R. Venkata Rao and J. P. Davim

INTRODUCTION
Finishing operations represent a critical and expensive phase of overall manufacturing
processes. The most labor intensive and uncontrollable area in the manufacturing of precision
parts involves final finishing operations, which frequently demand as much as 15% of the
total manufacturing cost. The dimensional and alignment accuracy and quality of surface
finish are taken care of by finishing processes such as grinding, lapping, honing and super-
finishing (i.e. traditional methods of finishing). But the applications of these traditional
abrasive-finishing processes are limited to the production of work pieces of basic forms such
as flat, cylindrical, etc. These finishing processes are being pushed to their limits of
performance especially in components of hard materials, and complicated shapes. Also, the
traditional fine finishing operations employ a rigid tool that subjects the work piece to
substantial normal stresses, which may cause micro-cracks resulting in reduced strength and
reliability of the machined part. Hence, there is a need to develop new finishing processes
with wider bounds of application areas, better quality performance, higher productivity, and
automatic operation. Abrasive flow machining (AFM) is such an advanced finishing process.
AFM is an advanced finishing process that can be used to deburr, radius, polish, remove
recast layer, and to produce compressive residual stresses. This process was developed by the
Extrude Hone Corporation, USA in 1960s as a method to deburr and polish difficult-to-reach
surfaces and edges by flowing abrasive laden polymer with special rheological properties.
However, as the use of the AFM process is a costly affair, the optimum selection of process
parameters of these processes is essential for efficient and economic utilization of process
capabilities. From the review of past literature, it is understood that very few efforts have
been made for parameter optimization of AFM process. Thus, efforts must be extended by
using advanced optimization algorithms which are more powerful, robust, and able to provide
accurate solution. This paper is intended to apply two such optimization algorithms known as
particle swarm optimization (PSO) and simulated annealing (SA) for optimization of process
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

parameters of AFM process.

The next section presents optimization aspects of AFM process.

OPTIMIZATION ASPECTS OF ABRASIVE FLOW MACHINING PROCESS

AFM process removes small quantity of material by flowing a semisolid abrasive laden
compound called ‘media’ (abrasive particles uniformly suspended in viscous chemical
compound) through or across the surfaces of the work piece to be finished. As shown in
Figure 1 [1], two vertically opposed cylinders extrude media back and forth through passages
formed by the work piece and tooling. The machining action compares to a grinding or
lapping operation as the media gently and uniformly abrades the surfaces and/or edges. The
media acts as a ‘self deformable stone’ having protruding abrasive particles acting as cutting
tools. The media is composed of semisolid carrier (e.g. polyborosilixane) and abrasive grains.
The abrasive action during AFM depends on the extrusion pressure, flow volume and media
flow speed determined by the machine setting in relation to media type, passage area, and
media formulation which includes media viscosity, and abrasive type and size.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Optimization of Abrasive Flow Machining Process Parameters… 53

Figure 1. Schematic diagram of AFM process [1].

AFM process is gaining widespread attention due to its ability to produce consistent and
predictable results. Removing stress raisers at sharp corners by producing controlled radii on
edges can substantially improve thermal and mechanical fatigue strength of highly stressed
components. Additional benefits over traditional finishing processes include a substantial
time saving and better control with regard to the accuracy and squareness of the bearing
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

surfaces. The process can deburr holes as small as 0.2 mm and radius edges from 0.025 - 1.5
mm. Tolerances can be held to ±5 μm [2,3].

REVIEW OF LITERATURE ON AFM PROCESS

AFM process is relatively a new process and much information is not available in the
literature that deals with optimization of the process parameters. The relationship between
process parameters and performance characteristics are not known completely. The abrasion
ability of abrasive media is governed by many factors, especially by grain size, abrasive
concentration, extrusion pressure, and hardness of work piece material. In order to analyze the
influence of such parameters and other AFM conditions upon material removal and surface
roughness of the machined surface, experimental investigations were carried out by many
researchers [1-4]. A stochastic modeling and analysis technique called data dependent system
(DDS) was used by Williams and Rajurkar [4] to study the surface roughness profiles before
and after AFM. Rajeshwar et al. [5] presented a simulation model to determine the
characteristics of media flow during machining. The finite difference method was chosen for
obtaining the solution.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 54 P. J. Pawar, R. Venkata Rao and J. P. Davim

Williams [6] used acoustic emission signals to analyze the mechanism of surface
generation in AFM and compared them to the acoustic emission signals of grinding process.
Jain and Adsul [1] studied the effects of different process parameters, such as number of
cycles, concentration of abrasive, abrasive mesh size, and media flow speed, on material
removal and surface finish in terms of polynomial equations.
Jain et al. [7] developed a back propagation neural network for modeling of AFM
process. The inputs were media flow speed (v), percentage concentration of the abrasives (c),
abrasive mesh size (d), and number of cycles (n). The outputs were material removal rate
(MRR) and surface roughness (Ra). Based on the results of simulation, the possibility of using
the neural network model for surface quality and material removal rate prediction for AFM
process was confirmed. This model could be used to study the AFM process by examining
the effects of the input process parameters on the performance of AFM process. Jain et al. [8]
developed a model for the flow of AFM media through cylindrical work piece and solved the
same by using finite element method (FEM). The model was shown to predict the radial
stresses at the work piece surface with reasonable accuracy. The normal stress, so obtained
from the flow model, was used for the estimation of material removal rate and surface
roughness.
Centrifugal force assisted abrasive flow machining (CFAAFM) process has recently been
tried as a hybrid machining process with the aim towards the performance improvement of
AFM process by applying centrifugal force on the abrasive laden media with a rotating
centrifugal force generating rod introduced in the work piece passage. Walia et al. [9]
explored the application of centrifugal force for the productivity enhancement of the process.
The authors had reported that centrifugal force enhances the material removal rate and
improves the scatter of surface roughness value in AFM.
Ali-Tavoli et al. [10] proposed an approach using group method of data handling
(GMDH)-type neural networks and genetic algorithms for modeling the effects of number of
cycles and abrasive concentration on both material removal rate and surface roughness, using
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

some experimentally obtained training and testing data for brass and aluminum in AFM
process. However, the approach had not considered other important process input parameters
such as media flow speed, abrasive mesh size, etc. Also, the approach is computationally
more complex. Furthermore, genetic algorithms (GA) provide a near optimal solution for a
complex problem having large number of parameters and constraints. This is mainly due to
difficulty in determination of optimum controlling parameters.
Jain et al. [11] presented the details of process parameters optimization of AFM and
magnetic abrasive finishing (MAF) processes using real-coded genetic algorithms. The
authors described a surface roughness model that was developed to form the objective
function for the optimization of AFM process. However, the results of GA presented by the
authors are erroneous and ineffective.

OPTIMIZATION MODEL OF AFM PROCESS

The optimization model for abrasive flow machining process is formulated in the present
work based on the analysis given by Jain et al. [11]. The optimization model is formulated as
shown below:

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Optimization of Abrasive Flow Machining Process Parameters… 55

Objective function:

Minimize (Ra)Ns=Ra0[1 – 226814{(Kafm/HwRa0) (Ph/d) ((1+ηa(ApmcLs/Aw)cdNs)1/2–1)}2]2

(1)

where, (Ra)Ns = Final surface value after Ns number of strokes; Rao = Initial surface
roughness, Hw = Brinell hardness of work piece, Ph = Extrusion pressure, Apmc = Cross-
sectional area of piston of medium-containing cylinder, Ls = Stroke length of piston, Aw =
Cross-sectional area of work piece, ηa = Proportion of abrasives effectively participating in
machining, c = Volumetric concentration of abrasive particles, d = Mesh size, and Kafm =
Proportionality constant relating normal radial stress acting on the abrasive grain.
Constraint: The constraint is to ensure that the final surface roughness value after AFM
process is smaller than the initial surface roughness value. This is mathematically expressed
as below:

1 – 226814{(Kafm/HwRa0) (Ph/d) ((1 + ηa (ApmcLs/Aw)cdNs)1/2 – 1) }2 ≥ 0.0 (2)

Decision variables (i.e. process parameters):

Four decision parameters were considered by Jain et al. [11] in the optimization problem
and these were, concentration of abrasives by volume (c), abrasive mesh size (d), number of
strokes (Ns), and extrusion pressure (Ph). The parameter bounds for these 4 decision
parameters were as shown below.

0.05 ≤ c ≤ 0.5, 8 ≤ d ≤ 1000, 1 ≤ Ns ≤ 100, and 0.7 ≤ Ph ≤ 25 MPa

The optimization model of AFM process is solved using PSO and SA algorithms which
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

are described in the next two sections.

PARTICLE SWARM OPTIMIZATION

Particle swarm optimization (PSO) is an evolutionary computation technique developed
by Kennedy and Eberhart [12]. It exhibits common evolutionary computation attributes
including initialization with a population of random solutions and searching for optima by
updating generations. Potential solutions, called particles, are then "flown" through the
problem space by following the current optimum particles. The updates of the particles are
accomplished as per the following equations.

Vi+1= w * Vi + c1*r1*(pBesti – Xi) + c2*r2*(gBesti – Xi) (3)

Xi+1 = Xi + Vi+1 (4)

Equation (3) calculates a new velocity (Vi+1) for each particle (potential solution) based
on its previous velocity, the best location it has achieved (‘pBest’) so far, and the global best

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 56 P. J. Pawar, R. Venkata Rao and J. P. Davim

location (‘gBest’), the population has achieved. Equation (4) updates individual particle’s
position (Xi) in solution hyperspace. The two random numbers ‘r1’ and ‘r2’ in equation (3) are
independently generated in the range [0, 1].
The acceleration constants ‘cl’ and ‘c2’ in equation (3) represent the weighting of the
stochastic acceleration terms that pull each particle towards ‘pBest’ and ‘gBest’ positions. ‘c1’
represents the confidence the particle has in itself (cognitive parameter) and ‘c2’ represents
the confidence the particle has in swarm (social parameter). Thus, adjustment of these
constants changes the amount of tension in the system. Low values of the constants allow
particles to roam far from target regions before being tugged back, while high values result in
abrupt movement toward, or past through target regions [13]. The inertia weight ‘w’ plays an
important role in the PSO convergence behavior since it is employed to control the
exploration abilities of the swarm. The large inertia weights allow wide velocity updates
allowing to globally explore the design space while small inertia weights concentrate the
velocity updates to nearby regions of the design space. The optimum use of the inertia weight
“w” provides improved performance in a number of applications.
Unlike genetic algorithm, PSO algorithm does not need complex encoding and decoding
process and special genetic operator. PSO takes real number as a particle in the aspect of
representation solution and the particles update themselves with internal velocity. In this
algorithm, the evolution looks only for the best solution and all particles tend to converge to
the best solution. In the implementation process, particles randomly generated at the
beginning or generated by internal velocity during the evolutionary process usually violate the
system constraints resulting in infeasible particles. Therefore, the handling of system
constraints, particularly nonlinear equation constraints, and the measurement and evaluation
of infeasible particles is very important. To cope with constrained problems with evolutionary
computation, various approaches such as rejection of infeasible individuals, repair of
infeasible individuals, replacement of individuals by their repaired versions, and penalty
function methods can be adopted. Among them, the penalty function method is particularly
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

promising as evidenced by recent developments [13] and the same is adapted in the present
work.

SIMULATED ANNEALING
Simulated annealing is a probabilistic hill climbing soft computing algorithm. A
probabilistic algorithm is one in which the objective function is evaluated at sample of points
chosen randomly from the feasible region. A hill climbing technique is one, which allows for
an increase in the value of objective function in a controlled manner. The methodology of
simulated annealing algorithm is described below.
Let a feasible configuration (k) is a point in an allowable region, T is a controlling
parameter and the cost C(k) is the quantity to be minimized. The neighborhood of a
configuration is a set of predefined feasible points from which next configuration is chosen. If
T is a quantity analogous to the temperature in annealing of solids, then the manner in which
the temperature is to be decremented and the number of moves required is to be decided. The
cost function is said to be in the state of equilibrium at temperature T when probability of
being in the configuration k is:

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Optimization of Abrasive Flow Machining Process Parameters… 57

e −C (k ) / T
Pr {config . = k } = (5)
Z (T )

where Z (T ) =
∑ e −C (k ) / T , (6)
j

If ‘i’ is the current configuration with cost C(i) then using the Metropolis algorithm [14],
we can say that the probability of accepting ‘j’ as next configuration depends on the
difference in the function value at these two points or on ΔC = C ( j ) − C (i ) and is
calculated using the Boltzman probability distribution:

⎧⎪ 1 if ΔC ≤ 0⎫⎪
Pr{new= j | current= i} = ⎨ ⎬ (7)
⎪⎩e−ΔC / T otherwise⎪⎭

Thus the probability of accepting new configuration is either 1 or e − ΔC / T depending

upon the sign of ΔC .
The nest section provides the application of proposed PSO and SA algorithms.

APPLICATION OF PSO AND SA ALGORITHMS TO AFM PROCESS

PARAMETERS OPTIMIZATION
The optimization of process parameters of AFM process using the proposed PSO and SA
algorithms is discussed below based on the model given by Jain et al. [11].
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Optimization Using PSO Algorithm

To ensure the convergence of PSO algorithm, the condition specified by equation (8)
must be satisfied [15].

max (|λ1|, |λ2|) <1 (8)

where, λ1 and λ2 are the eigen values given by equations (13) and (14).

λ1 = (1+ w - φ1 - φ2 + γ) / 2 (9)

λ2 = (1+ w - φ1 - φ2 - γ) / 2 (10)

and γ = [(1+ w - φ1 - φ2)2 – 4w)]1/2 (11)

φ1= r1*c1 and φ2= r2*c2

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 58 P. J. Pawar, R. Venkata Rao and J. P. Davim

Considering the feasible range for the value of ‘φ1 + φ2’ as 0 to 4 and that for ‘w’ as 0 to
1, it can be observed that for convergent trajectories the relation given by equation (12) must
be satisfied.

w > 0.5 (φ1 + φ2 ) –1 (12)

Now, in the present study the following values of ‘w’, ‘c1’ and ‘c2’ are used.

• Inertia weight factor (w) = 0.65

• Acceleration coefficients: c1 = 1.65 and c2 = 1.75

Considering the extreme possibility of random number as ‘r1’=0.95 and ‘r2’=0.95, the
right hand term in equation (16) is 0.5*(0.95*1.65 + 0.95*1.75) –1 = 0.61, which is less than
0.65 thus satisfies the equation (16). Hence, the values of ‘w’, ‘c1’ and ‘c2’ selected in the
present work are appropriate for convergence of the algorithm.
In the present work values of the constants shown in equations (1) and (2) are considered
as Apmc = 5026.5 mm2, Aw = 484 mm2, Ls = 80 mm, Hw = 2000 MPa, ηa = 0.2, Kafm = 0.4, and
Rao= 3.0 μm.. For the selected values of coefficients ‘c1’, ‘c2’ and inertia weight ‘w’ as
discussed above, the convergence of the PSO algorithm is shown in Figure 2.
The optimum process parameter values obtained by using PSO algorithm are given as:

c = 0.32 (i.e. 32%), d = 1000, Ph = 15.96 N/mm2, Ns = 73.10 (˜73)

Optimum value of final surface roughness = 0.000026 μm

Value of surface finish improvement constraint (given by equation (2)) = 0.00293

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0.25

0.2
Ra (microns)

0.15

0.1

0.05

0
0 5 10 15 20 25 30 35 40 45
Generation no.

Figure 2. Convergence of PSO algorithm for AFM process.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Optimization of Abrasive Flow Machining Process Parameters… 59

The results obtained by using PSO algorithm showed that the final roughness value is
almost equal to zero (i.e. 0.000026 μm) without violating the surface roughness constraint.
Thus, the results obtained by PSO algorithm provide large improvement in the solution. Table
1 shows comparison of the results of optimization using GA, PSO, and SA algorithms.
Validity of the model and the optimality of the above mentioned solution could be
confirmed from the Figures 3 to 4. As shown in Figure 3, the surface roughness initially
decreases up to a certain value and then increases with increase in volumetric concentration of
abrasives. This nature is confirmed experimentally by Jain and Adsul [1]. Also the surface
finish improvement constraint continuously decreases with increase in concentration of
abrasives. For the optimization problem under consideration, the surface finish decreases up
to value of ‘c’= 0.32 (i.e.32%) after which surface roughness increases as shown in Figure 3.
Also the constraint value is violated for any value of ‘c’ more than 0.32. Thus the selection of
optimum value of ‘c’ = 0.32 is appropriate.

Table 1. Comparison of optimization results of AFM process

Optimization c d Ns Ph Ra Constraint
technique value
GA [11] 0.196 851 23.3 0.81 0.1041a 0.18628a

GA 0.196 851 23.3 0.81 2.996b 0.999b

PSO 0.32 1000 73.10 15.96 0.000026 0.00293

SA 0.217 622.3 73.69 15.20 0.000105 0.00705

a: Values wrongly calculated by Jain et al. [11].
b: corrected values.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0 0.1 0.2 0.3 0.4 0.5 0.6

2.5 c
Ra (microns) & constraint value

1.5

1 Ra (microns)
0.5 Constraint

-0.5

-1

Figure 3. Variation of surface roughness and constraint with volumetric concentration of abrasive
particle ‘c’.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 60 P. J. Pawar, R. Venkata Rao and J. P. Davim

0 150 300 450 600 750 900 1050

300 d

Ra (microns) & constraint value

250

200

150 Ra (microns)
Constraint
100

50

-50

Figure 4. Variation of surface roughness and constraint with abrasive mesh size ‘d’.

Figure 4 shows the variation of surface roughness and constraint with abrasive mesh size.
The surface roughness decreases with increase in ‘d’ following the law of diminishing return.
Also the constraint is violated for lower values of ‘d’. Hence the upper bound value of ‘d’ =
1000 is selected as an optimum value.
Figure 5 shows variation of surface roughness and constraint with number of strokes. As
shown in Figure 5, the surface roughness initially decreases up to a certain value and then
increases with increase in number of strokes. Also the surface finish improvement constraint
continuously decreases with increase in concentration of abrasives. For the optimization
problem under consideration, the surface roughness decreases up to value of ‘Ns’= 73.10 after
which surface roughness increases as shown in Figure 5. The constraint is also getting
violated after this value of ‘Ns’. Hence the optimum value of number of strokes ‘Ns’ = 73.10
is selected.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 6 shows variation of surface roughness and constraint with extrusion pressure. As
shown in Figure 6, minimum value of surface roughness occurs for extrusion pressure ‘Ph’ =
15.96 without violating constraint surface roughness improvement constraint. Hence the
optimum value of ‘Ph’=15.96 MPa is selected.
It is also observed from the numerical results that very high surface finish of 0.000026 μ
is achieved by selecting optimum parameters as provided by PSO. Although this value is not
practically attainable as we cannot machine the material below its atomic size, but it can be
ensured that with the given set of parameters, maximum possible and attainable value of
surface finish will be obtained, as the values of suggested optimum parameters lies within
their specified bounds and also the set of optimum parameters satisfies the constraint on
surface roughness improvement.
Keeping in view of the above discussion, one should consider the limiting but practically
infeasible case of surface improvement constraint value being equal to zero. The theory of
constrained optimization mentions that a feasible solution (lying in the feasible region) that
optimizes the objective function and gives smaller value of the constraint is likely to be an
optimum solution. Therefore, those values of the decision variables lying in the feasible
region and minimizing the final surface roughness value and surface finish improvement
constraint should be the optimum solution.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Optimization of Abrasive Flow Machining Process Parameters… 61

0 15 30 45 60 75 90 105
2.5 Ns

Ra (microns) & constraint value

2

1.5

Ra (microns)
1
Constraint

0.5

-0.5

Figure 5. Variation of surface roughness and constraint with number of strokes ‘Ns’.

0 3 6 9 12 15 18 21 24 27
7 Ph (MPa)
Ra (microns) & constraint value

6
5
4
Ra
3 (microns)
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

2 Constraint
1
0
-1
-2

Figure 6. Variation of surface roughness and constraint with extrusion pressure (Ph).

Table 2 presents the results of optimization of process parameters of AFM for various
parameter values of w, c1 and c2. It can be observed that the feasible values of surface
roughness and the surface finish improvement constraint are 0.0785 and 0.1607 respectively
corresponding to w = 0.45, c1 = 1.55 and c2 = 1.45 and the corresponding feasible values for
c, d, Ns, and Ph are 0.27, 500, 55 and 13 respectively. These feasible values are better than
those presented by Jain et al.[11] who suggested the feasible values of surface roughness and
the surface finish improvement constraint as are 0.1041 and 0.18628 respectively. Thus PSO
results of optimization are proved better than the GA results given by Jain et al. [11].

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 62 P. J. Pawar, R. Venkata Rao and J. P. Davim

Table 2. Results of optimization of AFM process for different values of w, c1, c2 of PSO

S.No w c1 c2 c d Ns Ph Ra Constraint

1 0.20 1.15 1.15 0.302 476.38 48.78 12.08 0.180 0.245

2 0.25 1.10 1.10 0.294 416.17 48.58 12.39 0.0426 0.119

3 0.30 1 1 0.305 433.47 53.07 12.36 0.0064 0.0462

4 0.45 1.55 1.45 0.27 500 55 13 0.0785 0.1607

5 0.65 1.65 1.75 0.32 1000 73.1 15.96 0.000026 0.0029

6 0.85 1.95 2 0.50 486 33.2 12.92 0.00362 0.0347

7 0.95 1.90 1.9 0.30 1000 30.71 25 0.0041 0.037

c = Concentration of abrasives by volume, d = Abrasive mesh size, Ns = Number of strokes,
Ph=Extrusion pressure, and
Ra = Surface roughness value.

0.0035

0.003

0.0025
Ra (microns)

0.002

0.0015
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0.001

0.0005

0
0 10 20 30 40 50 60 70 80
No.of generations

Figure 7. Convergence of SA algorithm for optimization of AFM process.

Optimization Using SA Algorithm

Now using the simulated annealing technique, the objective function considering
maximization of the difference between the initial and final roughness value is written as:

Min. Z =Z1 - Penalty * Z2 (13)

where, Z = Combined objective function.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Optimization of Abrasive Flow Machining Process Parameters… 63

Z1 = Objective function given by equation (14)

Z2 = Constraint given by equation (15)

For the present case the penalty is considered as 1 if Z2 <0; else penalty = 0.

The initial temperature is obtained by calculating the average of the function values at a
boundary points. The convergence of the SA algorithm for optimization of AFM process
parameters is shown in Figure 7.
As shown in Table 1, the results of genetic algorithm [11] are erroneous and the corrected
results are ineffective. The results obtained by using SA algorithm are thus better than the
results obtained by using GA. However, the results obtained by using PSO algorithm are
better than the results provided by both GA and SA.

CONCLUSIONS
Abrasive flow machining (AFM) process is an economic and effective advanced
finishing process which is capable of providing excellent surface finish on difficult to
approach regions on a wide range of components. With AFM process, it has become possible
to substitute various time-consuming deburring and polishing operations that had often led to
non-reproducible results. In the present work, optimization aspects of process parameters of
AFM process are considered using particle swarm optimization (PSO) and simulated
annealing (SA) algorithms. The objective considered is, minimization of surface roughness
subjected to a constraint and parameter bounds.
The comparison of the performance of various algorithms in terms of accuracy of
solution and rate of convergence reveals that PSO algorithm outperformed the other
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

algorithms such as genetic algorithm and simulated annealing. The results obtained by using
PSO algorithm showed significant improvement over other optimization techniques such as
genetic algorithm (GA) and simulated annealing (SA). On the other hand, the results of SA
are better than the results of GA. Compared to other advanced optimization methods, few
trials are required to predict the best and worst operating parameters of PSO algorithm and
the algorithm requires only 30 to 40 iterations for convergence to the optimal solution. Unlike
genetic algorithm, PSO algorithm does not need complex encoding and decoding process and
special genetic operator. The proposed PSO and SA algorithms can be easily modified to suit
single or multi-objective optimization of process parameters of other advanced finishing
processes such as magnetic abrasive finishing, magnetic float polishing, ion beam machining,
etc.

REFERENCES
[1] Jain, V.K., Adsul, S.G. (2000). Experimental investigation into abrasive flow
machining. International Journal of Machine Tools and Manufacture, 40, 1003-1021.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 64 P. J. Pawar, R. Venkata Rao and J. P. Davim

[2] Rhoades, L.J. (1987). Abrasive flow machining with not-so-silly putty. Metal
Finishing, July 27-29.
[3] Rhoades, L.J. (1991). Abrasive flow machining: a case study. Journal of Materials
Processing Technology, 28, 107-116.
[4] Williams, R.E., Rajurkar, K.P. (1992). Stochastic modeling and analysis of abrasive
flow machining. Transactions of ASME, Journal of Engineering for Industry, 114, 74-
81.
[5] Rajeshwar, G., Kozak, J., Rajurkar, K.P. (1994). Modeling and computer simulation of
medium flow in abrasive flow machining process. Proceedings of the International
Mechanical Engineering Congress and Exposition, Chicago, 965-971.
[6] Williams, R.E. (1998). Acoustic emission characteristics of abrasive flow machining.
Transactions of ASME, Journal of Manufacturing Science and Engineering, 120, 264-
271.
[7] Jain, R.K., Jain, V.K., Kalra, P.K. (1999). Modeling of abrasive flow machining
process: a neural network approach. Wear, 231, 242-248.
[8] Jain, R.K., Jain, V.K., Dixit, P.M. (1999). Modeling of material removal and surface
roughness in abrasive flow machining process. International Journal of Machine Tools
and Manufacture, 39, 1903-1923.
[9] Walia, R.S., Shan, H.S., Kumar, P. (2006). Multi-response optimization of CFAAFM
process through Taguchi method and utility concept. Materials and Manufacturing
Processes, 21, 907-914.
[10] Ali-Tavoli, M., Zadeh, N.N., Khakhali, A., Mehran, M. (2006). Multi-objective
optimization of abrasive flow machining processes using polynomial neural networks
and genetic algorithms. Machining Science and Technology, 10, 1-20.
[11] Jain, N.K., Jain, V.K., and Jha, S. (2007). Parametric optimization of advanced fine-
finishing processes. International Journal of Advanced Manufacturing Technology, 34,
1191-1213.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[12] Kennedy, J., and Eberhart, R. (1995). Particle swarm optimization. Proceedings of
IEEE, International Conference on Neural Networks, 4, 1942-1948.
[13] Dong, Y., Tang, J., Xu, B., and Wang, D. (2005). An application of swarm optimization
to nonlinear programming. Computer and Mathematics with Applications, 49, 1655-
1668.
[14] Metropolis, N., Rosenbluth, A., Rosenbluth, M., Teller, A., and Teller, E. (1953).
Equation of state calculations by fast computing machines. Journal of Chemical
Physics, 21, 1087-1092.
[15] Bergh, F., and Engelbrecht, A.P. (2006). A study of particle swarm optimization
particle trajectories. Information Sciences, 176, 937-971.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 4

STUDY OF EFFECTS OF PROCESS PARAMETERS ON

BURR HEIGHT IN DRILLING OF AISI 316 STAINLESS
STEEL USING ARTIFICIAL NEURAL NETWORK
MODEL

V. N. Gaitonde1,a, S. R. Karnikb and J. Paulo Davimc

a
Department of Industrial and Production Engineering,
B.V.B. College of Engineering and Technology,
Hubli – 580 031, INDIA
b
Department of Electrical and Electronics Engineering,
B.V.B. College of Engineering and Technology,
Hubli – 580 031, INDIA
c
Department of Mechanical Engineering,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

University of Aveiro, Campus

Santiago, 3810-193 Aveiro, PORTUGAL

ABSTRACT
Burr size at the exit of holes in drilling degrades the precision of products and must
be removed in subsequent deburring processes to allow the components to meet specified
tolerances. Therefore understanding the drilling burr formation and its process parameters
is essential for predicting the burr size, which in turn provides useful information to
minimize the burr formation. An artificial neural network (ANN) based model has been
developed to study the effect of process parameters such as cutting speed, feed, drill
diameter, point angle and lip clearance angle on burr height during drilling of AISI 316
stainless steel. A multilayer feed forward ANN; trained using error back-propagation
training algorithm (EBPTA) has been employed. The input-output patterns required for
ANN training are obtained from the drilling experiments planned through the Box-
Behnken design. The simulation results demonstrate the effectiveness of ANN based
model to analyze the effects of drilling process parameters on burr height.

1
Corresponding author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 66 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

Keywords: Drillig; AISI 316 stainless steel; Burr height; Box-Behnken design; Artificial
neural network.

1. INTRODUCTION
The burr is a plastically deformed material generated on the edge of the drilled
component. These burrs cause several problems for the quality and functionability of the
products. There are no tools available for deburring if the exit burr is formed inside the cavity.
Hence, special tools must be used, which in turn increases the deburring cost. Therefore, it is
essential to understand the factors affecting the burr formation at the exit of holes in drilling
so as to reduce the burr at the manufacturing stage. This necessitates accurate models to
analyze the effects of drilling process parameters on burr height.
The finite element models, empirical drilling charts and the response surface
methodology (RSM) based models were employed in the past to study the influence of
various parameters on burr formation during drilling of several workpiece materials [1, 2].
The AISI 316 stainless steel work material finds broad applications in chemical industries,
aircraft designs and in the manufacture of medical apparatus due to high corrosive resistance
properties. But, the burr formation in AISI 316 stainless steel drilling is a greater concern due
to high ductility of the work material. Hence, an attempt has been made in this work to
analyze the interaction effects of process parameters such as cutting speed, feed, drill
diameter, point angle and lip clearance angle on exit burr height during drilling. A multi-layer
feed forward artificial neural network (ANN) has been employed to capture the relationship
between the burr height and the affecting process parameters. The experiments required for
ANN training were planned as per Box-Behnken design (BBD) during drilling of AISI 316
stainless steel workpieces using HSS twist drills.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

2. ARTIFICIAL NEURAL NETWORK MODELING

The artificial neural network (ANN) is a flexible modeling tool with an aptitude to learn
the mapping between input and output process parameters [3]. The purpose of ANN
development is to imitate human brain so as to implement the various functions such as
association, self-organization and generalization. The ANNs are parallel computer models of
processes and the mechanisms, which constitute biological nerve systems. The ANNs are
attractive in view of their high execution speed and modest computer hardware requirements
in addition to an adaptive nature and capability in solving complex and non-linear problems.
The ANN is made up of neurons connected via links. The information is processed within a
neuron and is propagated to other neurons through the links connecting neurons.
In this investigation, a multi-layer feed forward ANN using error back propagation
training algorithm (EBPTA) has been adapted. The EBPTA is a supervised learning algorithm
based on generalized delta rule [3], which requires a set of inputs and desired outputs, known
as training patterns. The EBPTA uses a gradient search technique, which updates the synaptic
weights during learning stage in such a way that mean square error between the actual output
pattern and the desired output pattern of network is minimized. The multilayer feed forward

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Study of Effects of Process Parameters on Burr Height… 67

ANN consists of neurons divided into input layer, hidden layers and the output layer. The net
activation input for ith neuron is given by [3]:

n
neti = ∑ wij x j (1)
j =1

where, wij = weight of link connecting neuron i to j; xj = the output of jth neuron . For an
unipolar sigmoid transfer function, the ouput of ith neuron is given as [3]:

1
oi = (2)
1 + eη neti

where, η is the scaling factor. The training algorithm used here is based on the weight updates
so as to minimize the sum of squared error for K number of neurons in the output layer and is
given by [3]:

1 K
E= ∑ ( d k , p − ok , p )
2
(3)
2 k = 1

where, dk,p = desired output for pth pattern. The synaptic weights of the links are updated as
[3]:
w ji( n +1 ) = w ji( n ) + α δ pj o pi + β Δw ji( n ) (4)

where, n is the learning step, α is the learning rate and β is the momentum constant. The error
term δpj in Eq. 4 is given by [3]:
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

For output layer: δ pk = ( d kp − okp )( 1 − okp ) ; k=1, . . .K (5)

For hidden layer: δ pj = o pj ( 1 − o pj ) ∑ δ pk wkj ; j =1,. . .J (6)

where, J is the number of neurons in the hidden layer.

The steps involved in ANN training using EBPTA are mentioned below:

(1) The network weights are initialized to small random values.

(2) The input/desired output pairs are presented one by one, updating the weights each
time.
(3) The mean square error (MSE) due to all outputs and NP number of patterns is
1 NP K
computed as, MSE = ∑ ∑ ( d kp − o kp )
2
(7)
NP p =1 k =1
(4) If (MSE < specified tolerance) or (epochs > (epochs)max)
Then stop.
Else, goto step 2.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 68 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

3. EXPERIMENTAL DETAILS
The database required for the development of ANN is obtained through the drilling
experiments. As, the classical experimental design methods are too complex, an experimental
layout plan based on Box-Behnken design (BBD) [4] has been selected to perform the
experiments. In the current study, cutting speed, feed, drill diameter, point angle and lip
clearance angle are selected as the process parameters. Three levels for each of the process
parameters are identified. Usually, in industries during the manufacture of valves for the
chemical plants, the drilling of stainless steel material is performed using HSS twist drills
with the cutting speed in the range 8 - 16 m/min. The feed in the range 0.04 - 0.12 mm/rev is
generally preferred with higher drill diameters in order to avoid excessive temperature rise
during machining. The range of point angle for drilling of stainless steel was selected based
on the investigations carried out by Stein [1]. The levels of lip clearance angle were identified
through the preliminary experiments. Accordingly, the ranges of the process parameters were
selected in the current study. The experimental parameters and their levels identified are
presented in Table 1. Table 2 illustrates the experimental layout plan as per BBD, which
consists of 46 sets of process parameter combinations.
A three-axis ‘YCM-V116B’ CNC vertical machining center (Make: Yeong Chin
Machinery Industries Co., Taiwan) was used to conduct the drilling experiments. The
machining center is equipped with a maximum feed of 5000 mm/min and a variable spindle
speed from 45 - 4000 rpm with a 15 kW drive motor. The maximum table travel along X –
axis is 1100 mm, along Z – axis is 630 mm and the maximum saddle travel along Y – axis is
600 mm. The drilling fixture was used to clamp the specimens on a flat surface and the fixture
was mounted in the vise on the table of machining center.
The AISI 316 stainless workpieces of 320 mm × 60 mm × 25 mm were used for all the
drilling experiments. The chemical composition and mechanical properties of work material
are listed in Table 3. The workpieces were polished on exit surface before drilling for the
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

preparation of burr height measurements. The HSS parallel shank stub series twist drills
(Make: Addison and Co. Ltd., India) confirming to IS: 5100/DIN: 1897/BS: 328/ ISO
specifications were utilized for drilling experiments. The required point angle and lip
clearance angles were ground as per BBD. ‘Cut60EP’ water-soluble oil was used as coolant
through out the experimentation.

Table 1. Experimental parameters and their levels

Levels
Parameter Notation Unit
1 2 3
Cutting speed v m/min 8 12 16
Feed f mm/rev 0.04 0.08 0.12
Drill diameter d mm 16 22 28
Point angle θ degree 118 126 134
Lip clearance angle ψ degree 8 9 10

The burr height was measured on ‘RPP-400’ toolmakers’ microscope (Make: Sicherun-
Gen Versehen, Germany) with a resolution of 1 micron at 30 × magnification. To measure the
burr height, the focus was put on top of burr and then on exit surface. The burr height is the
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Study of Effects of Process Parameters on Burr Height… 69

distance between two foci. It was observed that the burrs were more or less uniform and
hence the burr height values were recorded at four equally spaced locations around the
circumference and the average reading was taken as the process response. The experimental
layout plan along with the observed values of burr height is given in Table 2.

4. RESULTS AND DISCUSSION

4.1. ANN Training

The experimental results as shown in Table 2 were utilized for the development of ANN
model. The training of ANN for 46 input-output cominations has been performed using neural
network tool box avilable in 'MATLAB’ software [4].

Table 2. Experimental layout plan and measured values of burr height

Trial no. Process parameter settings Burr height (mm)

v f d θ ψ
(m/min) (mm/rev) (mm) (degree) (degree)
1 8 0.08 22 126 8 0.056
2 16 0.08 22 126 8 0.276
3 12 0.12 28 126 9 0.752
4 12 0.04 28 126 9 0.132
5 12 0.08 22 134 10 0.341
6 12 0.08 22 126 9 0.842
7 16 0.04 22 126 9 0.818
8 12 0.08 22 126 9 0.842
9 12 0.08 28 134 9 0.016
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

10 12 0.08 22 118 8 0.009

11 8 0.08 22 134 9 0.022
12 12 0.04 22 118 9 0.025
13 16 0.08 16 126 9 1.113
14 16 0.08 22 134 9 0.578
15 12 0.08 28 126 8 0.010
16 12 0.08 22 134 8 0.022
17 12 0.08 22 126 9 0.842
18 12 0.04 22 126 8 0.147
19 12 0.12 22 134 9 0.370
20 12 0.08 16 126 10 1.179
21 12 0.12 22 118 9 0.331
22 16 0.12 22 126 9 1.322
23 12 0.04 16 126 9 1.227
24 12 0.08 22 126 9 0.842
25 12 0.08 16 134 9 0.181
26 8 0.08 16 126 9 0.610
27 12 0.12 16 126 9 1.024

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 70 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

Table 2. (Continued)

Process parameter settings Burr height (mm)

Trial no. v f d θ ψ
(m/min) (mm/rev) (mm) (degree) (degree)
28 12 0.08 16 118 9 0.419
29 12 0.12 22 126 10 0.895
30 12 0.08 22 126 9 0.842
31 8 0.08 22 118 9 0.012
32 12 0.04 22 126 10 1.114
33 12 0.12 22 126 8 0.755
34 12 0.08 16 126 8 0.354
35 8 0.08 22 126 10 0.371
36 8 0.08 28 126 9 0.007
37 8 0.04 22 126 9 0.595
38 16 0.08 28 126 9 0.376
39 12 0.08 22 126 9 0.842
40 12 0.08 28 118 9 0.009
41 12 0.08 28 126 10 0.219
42 16 0.08 22 118 9 0.014
43 8 0.12 22 126 9 0.541
44 12 0.08 22 118 10 0.013
45 12 0.04 22 134 9 0.377
46 16 0.08 22 126 10 1.072

Table 3. Chemical composition and mechanical properties of AISI 316 stainless steel

Chemical 0.08 C, 2.5 Mo, 2.0 Mn, 1.0 Si, 0.03 S, 0.045 P, 12.0 Ni, 17.0 Cr,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

composition (wt %) 65.0 Fe

Tensile strength: 620 MPa
Mechanical Yield strength: 415 MPa
properties Hardness: 190 BHN
Modulus of elasticity: 193 GPa

All the inputs and the desired outputs were normalized using the following equation:

2( X − X min )
X normal = −1 (8)
( X max − X min )

where, Xmin = minimum value in the matrix of pattern for X; Xmax = maximum value in the
matrix of pattern for X. This normalization maps all the inputs and desired outputs between –1
and +1. The simulated multi-layer feed forward ANN architecture consists of 5 neurons in the
input layer (corresponding to five input parameters, v, f, d, θ and ψ ), 1 neuron in the output
layer (corresponding to one output parameter). The ANN training simulation was performed
using a variable learning rate training procedure “traingdx” of MATLAB NN toolbox [5]. A
trial and error procedure has been employed to optimize the number neurons in the hidden
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Study of Effects of Process Parameters on Burr Height… 71

layer. Appropriate learning rate parameters for faster convergence and the momentum factor
for increasing the rate of learning have been employed. The training has been continued until
the mean squared error (MSE) reaches 10-5 or 5000 epochs. After successful training, one
hidden layer with 11 neurons in the ANN structure with a learning rate of 0.3 and a
momentum constant of 0.5 were found to be suitable for burr height model.

4.2. ANN Testing

The ANN was initially tested with 46 input training patterns. For each input training
pattern, the ANN predicted burr height value was compared with the respective experimental
measured value and were found to be very close for each of the training patterns. For the
validation purpose, the drilling experiments were performed for 12 new combinations of input
process parameters, which do not belong to the training data set. The experimental conditions
along with the measured burr height values are summarized in Table 4. The comparison of the
ANN predicted and the measured values of burr height for the validation data set is depicted
in Figure 1. It can be seen that the ANN predicted values follow almost the same trend as that
of the measured values, which validates the developed ANN burr height model.

Table 4. Experimental conditions and measured values of burr height for validation
data

Burr height
Levels of process parameters
Test trial (mm)
no. v f d θ ψ
(m/min) (mm/rev) (mm) (degree) (degree)
1 8 0.04 22 134 10 0.781
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

2 8 0.04 28 134 10 0.053

3 8 0.04 16 134 10 1.112
4 8 0.04 22 134 10 0.773
5 12 0.12 28 126 10 0.612
6 12 0.12 22 134 8 0.187
7 12 0.12 22 134 10 0.681
8 16 0.12 22 126 8 0.813
9 16 0.12 22 126 10 1.834
10 16 0.12 16 126 10 1.221
11 16 0.04 16 126 9 1.323
12 16 0.04 16 134 9 0.421

4.3. Parametric Analysis of Process Parameters on Burr Height

The developed ANN model was utilized to study the interaction effects of selected
process parameters on burr height. To analyze the interaction effects, the 3D response surface
plots were generated considering two parameters at a time, while the other parameters are
held constant at their respective center levels. These interaction plots are presented in Figures
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 72 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

2-11. It is observed from Figure 2 that for a given cutting speed, the burr height increases
with increase in feed. A similar trend is observed with the cutting speed variations for a
specified feed. However, it is to be noted that, the burr height is highly sensitive at higher
values of cutting speed and feed. Moreover, minimum burr height results with lower values of
cutting speed and feed combination. From the interaction plot of cutting speed and drill
diameter (Figure 3), it is clearly evident that lower cutting speed is necessary for all drill
diameter values in the range 16 -28 mm in order to minimize the burr height.

1.8
Experimental
1.6
ANN predicted
1.4
Burr height (mm)

1.2

0.8

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8 9 10 11 12
Test trial no.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 1. Comparison of the experimental and the ANN predicted values of burr height for validation
data set.

1.5
Burr height (mm)

0.5

0
0.12
0.1 16
0.08 14
12
0.06 10
0.04 8
Feed (mm/rev) Cutting speed (m/min)

Figure 2. Interaction effect of cutting speed and feed on burr height.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Study of Effects of Process Parameters on Burr Height… 73

1.5
Burr height (mm)
1

0.5

0
28
26 16
24
22 14
20 12
18 10
16 8
Drill diameter (mm) Cutting speed (m/min)

Figure 3. Interaction effect of cutting speed and drill diameter on burr height.

1
Burr height (mm)

0.5
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0
134
130 16
126 14
122 12
10
118 8
Point angle (degree) Cutting speed (m/min)

Figure 4. Interaction effect of cutting speed and point angle on burr height.

Figure 4 illustrates the interaction effects of cutting speed and point angle on burr height.
It is seen from this figure that the burr height is less sensitive to cutting speed variations
irrespective of the point angle. On the other hand, maximum burr height results when the
point angle is in the range 122-130 degree for all values of cutting speed.
The interaction effects of cutting speed and lip clearance angle on burr height (Figure 5)
suggests that the lower values of lip clearance angle and cutting speed are necessary for
minimizing the burr height. In other words, the burr height increases either with the increase
in cutting speed or with the increase in lip clearance angle.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 74 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

1.5

Burr height (mm) 1

0.5

0
10
9.5 16
9 14
8.5 12
10
8 8
Lip clearance angle (degree) Cutting speed (m/min)

Figure 5. Interaction effect of cutting speed and lip clearance angle on burr height.

1.5
Burr height (mm)

0.5
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0
28
26 0.12
24
22 0.1
20 0.08
18 0.06
16 0.04
Drill diameter (mm) Feed (mm/rev)

Figure 6. Interaction effect of feed and drill diameter on burr height.

As seen from Figure 6, the burr height is almost insensitive to feed variations for drill
diameters up to 22 mm. However, beyond 22 mm drill diameter, lower feed value is
beneficial for minimizing the burr height. It can be observed from Figure 7 that for any
selected feed value, lower point angle is necessary to minimize the burr height. Moreover, for
a specified feed, initially the burr height sharply increases with the point angle up to 126
degree. The interaction behavior of burr height with feed and lip clearance angle (Figure 8) is
almost same as that with lip clearance angle and cutting speed (Figure 5).

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Study of Effects of Process Parameters on Burr Height… 75

1
Burr height (mm)

0.5

0
134
130 0.12
126 0.1
122 0.08
0.06
118 0.04
Point angle (degree) Feed (mm/rev)

Figure 7. Interaction effect of feed and point angle on burr height.

1.5
Burr height (mm)

0.5
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0
10
0.12
9 0.1
0.08
0.06
8 0.04
Lip clearance angle (degree) Feed (mm/rev)

Figure 8. Interaction effect of feed and lip clearance angle on burr height.

It is clear from Figure 9 that, lower point angle is necessary to minimize the burr height
for all drill diameters in the range 16-28 mm. Further, it is also observed that low lip
clearance angle is essential as shown in Figure 10. Figure 11 exhibits the interaction effect of
point angle and lip clearance angle. For any given lip clearance angle, the burr height shraply
increases with point angle up to 126 degree and then shows decreasing tendency for further
increase in point angle. Therefore, lower point angle is found to be suitable for minimizing
the burr height.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 76 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

Burr height (mm)

0.5

0
134
130 28
26
126 24
22
122 20
118 18
16
Point angle (degree) Drill diameter (mm)

Figure 9. Interaction effect of drill diameter and point angle on burr height.

1.5
Burr height (mm)

0.5

0
10
28
26
9 24
22
20
18
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

8 16
Lip clearance angle (degree) Drill diameter (mm)

Figure 10. Interaction effect of drill diameter and lip clearance angle on burr height.

1
Burr height (mm)

0.5

0
10
134
9 130
126
122
8 118
Lip clearance angle (degree) Point angle (degree)

Figure 11. Interaction effect of point angle and lip clearance angle on burr height.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Study of Effects of Process Parameters on Burr Height… 77

From the above discussion, it is clearly evident that the ANN model is very useful in
analyzing the effects of drilling process parameters on burr height. With the developed model,
it is possible to predict the burr height for a given combination of selected process parameters.
Moreover, the developed ANN model is very useful for the selection of appropriate control
parameters, which result in minimum burr height. The principal advantage of ANN modeling
is that it has the ability to capture any degree of non-linearity that exists between the process
parameters and the response with good generalization ability.

CONCLUSION
The multilayer feed forward artificial neural network (ANN) architecture; trained using
error back propagation algorithm (EBPTA) was developed to analyze the effects of process
parameters on burr height during drilling of AISI 316 stainless material. The input-output
patterns required for ANN training were obtained through drilling experiments planned as per
Box-Behnken design (BBD). A good agreement was observed between the ANN predicted
and the experimental values of burr height. The simulated results demonstrated the
effectiveness of the ability of ANN model in realizing a non-linear relationship between the
burr height and the identified drilling process parameters. The following conclusions are
drawn from the analysis:

• There exists a highly non-linear relationship between the burr height and the selected
drilling process parameters.
• The burr height increases with increase in feed and is highly sensitive at higher
values of cutting speed and feed. The minimum burr height results with lower values
of cutting speed and feed combination.
• For minimizing the burr height, lower values of cutting speed are necessary for all
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

the selected drill diameters. Further, lower point angle is necessary for minimizing
the burr height for any given feed and for all drill diameters in the range 16-28 mm.
• The burr height increases either with increase in cutting speed or with increase in lip
clearance angle. The lower feed value is beneficial for minimizing the burr height
beyond 22 mm drill diameter.
• The maximum burr height exists when the point angle is in the range 122-130 degree
for all values of cutting speed.

REFERENCES
[1] Stein, J. M. (1997), “The Burrs from Drilling: An Introduction to Drilling Burr
Technology”, Burr Technology Information Series TM, USA.
[2] Gaitonde, V. N., Karnik, S. R., Achyutha, B. T., Siddeswarappa, B. and Davim, J. P.
(2009), “Predicting burr size in drilling of AISI 316L stainless steel using response
surface analysis”, International Journal of Materials and Product Technology, 35(1/2),
228-245.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 78 V. N. Gaitonde, S. R. Karnik and J. Paulo Davim

[3] Schalkoff, R. B. (1997), “Artificial Neural Networks”, McGraw-Hill International

Edition.
[4] Montgomery, D. C. (2004), “Design and Analysis of Experiments”, 5th edition, John
Wiley and Sons, New York.
[5] Math Works Incorporation (2005), “MATLAB User Manual Version 7.1 R14”, Math
Works Incorporation, Natick, MA.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 5

ARTIFICIAL NEURAL NETWORK MODELING OF

SURFACE QUALITY CHARACTERISTICS IN ABRASIVE
WATER JET MACHINING OF TRIP STEEL SHEET

N. M. Vaxevanidis∗1, A. Markopoulos1 and G. Petropoulos2

1
Department of Mechanical Engineering Educators,
School of Pedagogical and Technological Education (ASPETE),
141 21 N. Heraklion, Athens, GREECE
2
Department of Mechanical and Industrial Engineering,
University of Thessaly, Pedion Areos, 38334 Volos, GREECE

ABSTRACT
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

In this chapter the application of Artificial Neural Networks (ANNs) for the
modeling of surface quality characteristics (kerf geometry and surface roughness) in
Abrasive Water Jet Machining (AWJM) of transformation induced plasticity (TRIP) steel
sheet is discussed. For the development of the models the neural network toolbox of
Matlab is used and all networks are feed-forward models trained with the Levenberg-
Marquardt algorithm. Models constructed possess three inputs, namely nozzle diameter,
stand-off distance and travel speed, one hidden layer with five neurons and two outputs
(kerf and Ra). Training of the models was performed with data from an extensive series of
statistically designed experiments concerning AWJM of two TRIP type steel sheets. The
reported results indicate that the proposed ANN model can satisfactorily predict the
surface roughness and the mean kerf in AWJM; moreover, it can be considered as
valuable tools for process planning in workshop.

INTRODUCTION
In abrasive water jet machining (AWJM), high velocity water containing abrasive
particles is used to cut different materials ranging from soft, ductile to hard and brittle

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 80 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

materials. AWJM technology nowadays is considered to be one of the most developing

advanced non-traditional methods used in industry for material processing with the distinct
advantages of no thermal distortion, high machining versatility, high flexibility and small
cutting forces. Because of these capabilities, it can be applied for machining materials with
higher performance and be more cost-effective than traditional and some non-traditional
machining processes. AWJM is widely used in the processing of materials such as titanium,
steel, brass, aluminium, stone, Inconel, any kind of glass and composites [1, 2].
The intensity and the efficiency of the machining process depend on several AWJM
process parameters. They are classified as hydraulic, abrasive, work material and cutting
parameters [3, 4]. Surface roughness, which is used to determine and to evaluate the surface
quality of a product, together with the kerf geometry formed in a through cut [5], are the two
major quality attributes of an AWJMed product which were evaluated during present
research. For the estimation and/or the prediction of surface roughness in machining
processes various methodologies have been adopted; for an overview see [6]. These different
approaches are based on machining theory and/or process modeling, experimental
investigation, design of experiments [7] and intelligent techniques. The present trend of
research in the field of manufacturing favours intelligent techniques, probably due to the
enhanced computing power. They consider the machining conditions and the information
stored in the experimental data to develop the models. The various intelligent techniques
attempted by researchers are (artificial) neural network (ANN), fuzzy logic, neuro-fuzzy
approach and genetic algorithms [8, 9]. Following the succesful implementation of ANNs for
Ra prediction due to electro-discharge machining (EDM); see Refs [9, 10], in this chapter the
application of Artificial Neural network (ANN) for the modeling of surface quality
characteristics (kerf geometry and surface roughness) in Abrasive Water Jet Machining
(AWJM) of TRIP steel sheet is presented.
In brief, the structure of this chapter reads as follows: In the subsequent two sections the
main features of AWJM and ANNs are, respectively, summarized. Then, the experimental
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

features (materials, measuring techniques and the structure of ANNs selected) are outlined.
Next, the application and the suitability of the proposed modeling are presented and discussed
(Results and Discussion) and the representative literature follows.

ABRASIVE WATER JET MACHINING

The so-called non conventional (or non-traditional or advanced) machining processes
have been emerged from the need in modern manufacturing practice for a more frequent use
of harder, tougher or stronger workpiece materials, which are much more difficult to machine
with traditional methods. Reference is made to all kinds of high strength thermal resistant
alloys, to various kinds of carbides, fiber-reinforced composite materials, ceramic materials,
various modern composite tool materials etc [1, 11].
These advanced machining processes become still more important when one considers
precision and ultra-precision machining. In a number of processes, material is removed even
in the form of atoms or molecules individually or in groups. These processes are based on the

∗
Corresponding author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 81

direct application of energy for material removal by mechanical erosion, thermal erosion or
electro-chemical/chemical dissolution [12].
In general, advanced non conventional machining processes are classified into four main
groups, namely:

• Mechanical machining processes

• Thermal machining processes
• Electro-chemical and chemical machining processes
• Hybrid machining processes

Among these processes abrasive water jet machining (AWJM) is regarded as the fastest
growing major machine tool process in the world [13]. AWJM is a technique for cutting or
separating materials by means of a high-velocity slurry jet, formed as a result of injecting
abrasive particles to a waterjet ejected by an orifice; see Figure 1. Note, that pure water jet
cutting was introduced around 1970 and the abrasive water jet cutting in 1983.
The main advantages of the AWJM process are being able to cut versatile geometries and
its ability to cut both ductile materials like aluminum, brass, steel and titanium and brittle
materials like glass, stone and ceramics without any influence on their microstructure. The
water pressure, combined with the water flow rate defines the ability of the AWJM, since
both factors define the maximum available power of the abrasive-water mixture [11, 14].
Note that the physical principles and the phenomena involved, the process parameters
and the capabilities and applications of AWJM as well as recent advances and the limitations
of the process can be found in a number of excellent reference works, see for example [1, 4,
15]; therefore only some main features are summarized below.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 1. Schematic illustration of AWJM.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 82 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

The abrasive water jet-cutting process is characterized by a large number of operational

parameters which determine the efficiency, economy and quality of the entire process. In
general, the parameters in AWJM can be divided into four categories [4]:

1. Hydraulic parameters
– Pump pressure (p)
– Water-orifice diameter
– Water flow rate
2. Mixing and acceleration parameters
– Focus diameter
– Focus length
3. Cutting parameters
– Traverse rate
– Number of passes (np)
– Stand off distance (x)
– Impact angle
4. Abrasive parameters
– Abrasive mass flow-rate
– Abrasive particle diameter
– Abrasive particle size distribution
– Abrasive particle shape
– Abrasive particle hardness

Material removal in AWJM of metals is comprised of microchip formation, plowing and

rubbing, all of which occur through shear deformation. The predominance of any one of these
three mechanisms is depended on the abrasive attack angle and the mechanical properties of
the target material; for an overview see [2, 15]. At low cutting depths, abrasives impinge on
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

the target material at shallow attack angles, hence promoting “cutting wear”. The
impingement angle was believed to become more obtuse with larger cutting depths, thereby
inducing a change of removal mechanisms to “deformation wear”. Material removal within
the deformation wear region occurs through a cyclic cutting action and commonly results in a
wavy surface texture which serves as the division between the cutting wear and deformation
wear regions.
The pressure at which a water jet operates is about 400 MPa, which is sufficient to
produce a jet velocity of 900 m/s. Such a high-velocity jet is able to cut materials such as
ceramics, composites, rocks, metals etc. [1]. The AWJM process can easily cut both
electrically non-conductive and conductive, and difficult-to-machine materials. This process
does not produce dust, thermal defects, and fire hazards. Recycling of water and abrasives is
possible to some extent. It is a good process for shaping and cutting of composite materials,
and creates almost no delamination [4, 12].
The kerf geometry of a through cut generated by abrasive waterjets may be described as
in Figures 2(a) and 2(b). It is characterized by a small rounded corner at the top edge due to
the plastic deformation of material caused by jet bombardment. As the kerf is wider at the top
than at the bottom due to the decrease in water pressure, a taper is produced. In addition, the
plastically deformed material rolls over at the bottom of the kerf forming burrs at the jet exit
when cutting ductile materials.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 83

Figure 2. Geometry of a typical cut in AWJM; (a) a schematic representation and (b) macrograph of a
cut sheet.

The primary interests in sheet steel processing are the kerf shape (kerf width and kerf
taper) and kerf quality (cut surface roughness) as well as burrs which may be formed at the jet
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

exit [16]. These characteristics were considered in the present study; see next section
“experimental”.

ARTIFICIAL NEURAL NETWORK

Artificial Neural Networks Overview

Two main and important features of neural networks are their architecture, i.e., the way
that the network is structured, and the algorithm used for its training. After the appropriate
training, the selected network has the ability to interconnect one value of output to given
input. These two features of neural networks along with some techniques used for the
improvement of their performance are briefly presented below. Note that the origin, the
development and the mathematical details for implementing the ANNs can be found in a
number of excellent reference works, see for example Refs [17-19]; therefore they are not
discussed here.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 84 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

Neural Network’s Architecture

Artificial neural networks are mathematical representations of the human brain function.
The “core” element of a neural network is the neuron. Neurons are connected to each other
with a set of links, called synapses and each synapse is described by a synaptic weight.
Neurons are placed in layers and each layer’s neurons operate in parallel. The first layer is the
input layer. The activity of input units represents the non-processed information that entered
the network; at that layer neurons do not perform any computations. The hidden layers follow
the input layer and the activity of each hidden unit is determined from the activity of the input
units and the weights at the connections of input and hidden units. A network can have many
or none hidden layers and their role is to improve the network’s performance. The existence
of these layers at the network becomes more necessary as the number of input neurons grows.
The last layer is the output layer. The behavior of output units depends upon the activity of
the hidden units and the weights between hidden units and output units. The output of the
layer is the output of the whole network; output layer neurons, in contrast to input layers,
perform calculations.
There are two types of neural networks: the feed-forward and the recurrent ones. Feed-
forward neural networks allow the signals to travel in only one direction: from input to
output, i.e. the output signal of a neuron is the input of the neurons of the following layer and
never the opposite. The inputs of the first layer are considered the input signals of the whole
network and the output of the network is the output signals of last layer’s neurons. On the
contrary, recurrent networks include feedback loops allowing signals to travel forward and/or
backward [18]. Feed-forward neural networks are characterized by simple structure and easy
mathematical description [19]; therefore they were selected for the modeling of surface
roughness in the present chapter.
In general, there is not a standard algorithm for calculating the proper number of hidden
layers and neurons. For relatively simple systems, as the present case, a trial-and-error
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

approach is usually applied in order to determine which architecture is optimal for a problem.
Networks that have more than one hidden layers have the ability to perform more complicated
calculations. However, for most applications, one hidden layer is enough, while for more
complicated applications the simulation usually takes place using two hidden layers [20]. The
existence of more than necessary hidden layers complicates the network, resulting in a low
speed of convergence during training and large error during operation. Therefore, the
architecture of a neural network always depends upon the specific situation examined and
must not be more complex than needed [10, 17].

Neural Network’s Training

Once the number of layers and the number of units in each layer are selected the
network's weights must be set in order to minimize the prediction error of the network; this is
the role of the training algorithms. The historical cases that were gathered are used to
automatically adjust the weights in order to minimize this error. The error of a particular
configuration of the network can be determined by running all the training cases through the
network and comparing the actual output generated with the desired or target outputs. The
differences are combined together by an error function resulting the network’s error. Usually
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 85

the mean square error (MSE) of the network’s response to a vector p, is calculated, according
to the equation:

1 l
∑ ( d p ,i − o p ,i )
2
Ep =
2 i =1

In the preceding equation op,i are the values of the output vector which occur for the input
vector p and dp,j the values of the desirable response corresponding to p. The procedure is
repeated until MSE becomes zero. Each time that the program passes through all pairs of
training vectors an epoch is completed; training usually ends after reaching a great number of
epochs.
One of the frequently used training algorithms is the back-propagation (BP) algorithm. It
is usually applied in feed-forward networks with one or more hidden layers [10, 21, 22]. The
input values vectors and the corresponding desirable output values vectors are used for the
training of the network until a function is approached which relates the input vectors with the
particular output vectors. When the value of the mean square error is calculated, it is
propagated to the back in order to minimize the error with the appropriate modification of the
weights.
Another important parameter of the neural network models is their ability to generalize,
i.e. the ability of neural networks to provide logic responses for input values that were not
included in the training. Correctly trained back-propagation networks are able to perform
generalization; this ability provides the opportunity of training the network using a
representative set of input - desirable output values pairs.

Improvement Techniques
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

When an algorithm is applied to the network random values are given to the weight
factors. The convergence speed and the reliability of the network depend upon the initial
values of weights; thus different results may be observed during the application of the same
algorithm to the network. There are only a few elements that can guide the user for the
selection of the proper values. A wrong choice may result to small convergence speed or even
to network’s paralysis, where training stops. Furthermore, due to the fact that the algorithm
searches for the minimum error, the network may be stabilized at a local minimum instead of
the total minimum. As a result, most of the times, incorrect response values of the network
are produced. To overcome these problems variations of the most used algorithms have been
proposed; for further information on this topic Refs. [18, 19] may be consulted. Worth
mentioning, also, that a very common and simple technique used for overcoming problems of
this type is the repetition of the algorithm many times and the use of different initial values of
the weight factors.
One of the problems that occur during the training of neural networks is over-fitting
which undermines their generalization ability. The error appears to be very small at the set of
the training vectors, however, when new data are imported to the network the error may
become extremely large. This phenomenon is attributed to fact that the network memorized
the training examples and did not learn to generalize under the new situations. The
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 86 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

generalization ability of a network is assured when the number of training data is quite greater
than the number of network’s parameters. However, when the network is large the relations
between input and output become rather complicated. Hence, a network should not be larger
than needed to solve the given problem. Note, also, that two improvement techniques were
applied during modeling, namely, normalization of the used data and the early stopping
technique. Both these techniques and their application to the particular problem are briefly
discussed in section: Experimental/ANN modeling.
Note that in the area of machining, neural networks have been used for the prediction of
cutting forces, surface roughness, dimensional deviation, tool wear and tool life; see for an
overview [23]. To the authors’ knowledge, applications of ANNs to AWJM are quite limited;
see Refs. [22, 24, 25].

EXPERIMENTAL
Abrasive Water Jet Machining

Machining was performed on a SIELMAN HELLENIC HYDROJET industrial AWJM

system. In each specimen three slots of 3 cm in length were cut on both sides; each slot
corresponds to different machining conditions. After processing, each specimen was
separated in order to allow roughness measurements to be performed on the machined
surface. The specimens were processed with three different diameters of the nozzle (nozzle
diameter), three different distance values between the nozzle and the sheet steel (stand off
distance) and three different cutting speeds (travel speed). Concisely, the design variables for
AWJM modeling are given in the following table (Table 1). Worth mentioning that the high
value of stand off distance is not usually applicable but was considered to establish a wide
range of variation.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Table 1. Design variables of the machining process.

Nozzle diameter (mm) Stand-off distance (mm) Travel speed (mm/min)

0.95 20 200
1.2 64 300
1.5 96 400

Materials

TRIP (Transformation Induced Plasticity) multi-phase steels belong to a new generation

of steel grades exhibiting an enhanced combination of strength and ductility, with extensive
applications in automotive and aerospace industry.
The TRIP steels tested are designated TRIP 800 HR-FH and TRIP 700 CR-FH. Their
chemical composition is summarized as follows.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 87

• TRIP 800 HR-FH: 0.22 % C, 0.04 % Si, 1.66 % Mn, 1.49 % Al, 0.012 % P, 0.001 %
S, 0.027% Cr, 0.003 %Mo, 0.02 %Ni and 0.027 %Cu.
• TRIP 700 CR-FH: 0.21 % C, 1.53 % Si, 1.78 % Mn, 0.04 % Al, 0.019 % P, 0.003 %
S, 0.026% Cr, 0.001 % Mo and 0.027 % Cu.

Specimens of both materials are of square form (10x10 cm2) but differ in thickness,
hardness and processing method; see Table 2.

Table 2. Properties of materials tested

Material TRIP 800 HR-FH TRIP 700 CR-FH

Thickness 1.25 mm 0.9 mm

Hardness 20 HRC 23 HRC

Processing method Hot rolling Cold rolling

Measuring Techniques

The roughness measurements were carried out on twenty-seven specimen surfaces using
Surtronic 3+ stylus profilometer supported by the commercial software Talyprof®. The cut-
off selected was 0.8 mm and the measurements were undertaken in the direction of the cut.
The parameter values appear as averages of 5 measurements on each surface at the medium
area of the cut. Although, there is serious surface inhomogeneity between the entrance and the
exit areas of the beam, these values are considered as representative for the quality control of
the AWJMachined surfaces. In Figure 3 a typical profile is shown, indicative of the striated
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

pattern generated by AWJM; see also [26].

Figure 3. Typical roughness profile of AWJMed surface.

As it is illustrated in Figure 2(b) the kerf is of tapered form and to evaluate this
characteristic, the semi-sum of the upper area width and the lower area width was employed.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 88 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

ANN Modeling

Two different neural networks models were constructed for each steel type, namely TRIP
800 and TRIP 700. Each model can predict “mean kerf diameter” (Kerf) and “mean surface
roughness” (Ra). These models were constructed using Matlab® with the neural network
toolbox (version R2008a) [27].
All networks are feed-forward models trained with the Levenberg-Marquardt algorithm.
They have three inputs, namely nozzle diameter, stand-off distance and travel speed, one
hidden layer with five neurons and two outputs (kerf and Ra). The transfer function in the
hidden layer is the hyperbolic tangent sigmoid function. The architecture of the selected
(optimized) network is presented in Figure 4.
For each model the early stopping technique is applied. In this method the existing data
are divided into three subsets. The first subset consists of the training vectors, which are used
to calculate the gradient and to form the weight factors and the bias. The second subset is the
validation group. The error in this group is observed during training and like training group
normally decreases during the initial phase of training. However, when the network begins to
adjust the data more than needed, the error in that group raises and when this increase is
continued for a certain number of repetitions, training stops. The third subset is the test group
and its error is not used during training. It is used to compare the different models and
algorithms. The number of data as well as which data are included in every group is randomly
selected by the program.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 4. Optimized neural network architecture.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 89

Furthermore, Principal Component Analysis is used in the research outlined in the

present chapter. In some situations, the dimension of the input vector is large, but the
components of the vectors are highly correlated (redundant). It is useful in this situation to
reduce the dimension of the input vectors. An effective procedure for performing this
operation is principal component analysis. This technique has three effects: it orthogonalizes
the components of the input vectors (so that they are uncorrelated with each other), it orders
the resulting orthogonal components (principal components) so that those with the largest
variation come first, and it eliminates those components that contribute the least to the
variation in the data set. In the present analysis, firstly the input vectors are normalized, so
that they have zero mean and unity variance. Then the principal components that contribute
less than 1% to the total variation in the data set are eliminated.

RESULTS AND DISCUSSION

Based on the design variables for AWJM modeling presented in Table 1, the process
parameters and the corresponding measured values for Ra and kerf are tabulated in Tables 3
and 4 for TRIP 800 and TRIP 700 steel grades respectively. In the same Tables the results
from the proposed ANN as well as the deviation (“Dev.”) between measured and calculated
by ANN values are presented.

Table 3. Results for TRIP 800 steel grade

Stand off Kerf

Nozzle Travel Ra by
distance Kerf (by Ra
diameter speed Dev. (%) ANN Dev. (%)
(mm) (mm) ANN) (μm)
(mm) (mm/min) (μm)
(mm)
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0.95 20 200 0.915834 0.9337 1.95079 5 5.1112 2.224

0.95 20 300 0.872053 0.9506 9.007136 5.3 5.4966 3.709434
0.95 20 400 0.857991 1.0182 18.67257 5.9 5.7055 3.29661
0.95 64 200 1.113319 0.8841 20.5888 6 6.0626 1.043333
0.95 64 300 1.0768 0.9828 8.729569 5.7 5.7161 0.282456
0.95 64 400 0.96584 1.0425 7.937132 6.7 6.0615 9.529851
0.95 96 200 1.164883 1.1374 2.359293 6.1 6.0564 0.714754
0.95 96 300 0.952434 1.0155 6.621561 6.3 6.3097 0.153968
0.95 96 400 0.924122 0.9993 8.135073 6.6 6.5694 0.463636
1.2 20 200 1.18478 1.1435 3.484191 6.1 6.0777 0.365574
1.2 20 300 1.175643 1.1385 3.159377 6.3 6.2883 0.185714
1.2 20 400 1.048306 1.1523 9.920195 6.2 6.1874 0.203226
1.2 64 200 1.285587 1.2588 2.08364 6.8 6.6491 2.219118
1.2 64 300 1.322628 1.2069 8.749853 6.7 6.6936 0.095522
1.2 64 400 1.150741 1.1895 3.368178 6.6 6.5893 0.162121
1.2 96 200 1.447989 1.5168 4.752177 6.85 6.902 0.759124
1.2 96 300 1.434771 1.4156 1.336171 7.05 6.9841 0.934752

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 90 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

Table 3. (Continued)

Stand off Kerf

Nozzle Travel Ra by
distance Kerf (by Ra
diameter speed Dev. (%) ANN Dev. (%)
(mm) (mm) ANN) (μm)
(mm) (mm/min) (μm)
(mm)
1.2 96 400 1.26562 1.2695 0.306569 7 7.0369 0.527143
1.5 20 200 1.353131 1.2437 8.087244 6.1 6.1126 0.206557
1.5 20 300 1.224977 1.1939 2.536946 5.7 5.8553 2.724561
1.5 20 400 1.384728 1.1655 15.83185 5.75 6.3982 11.27304
1.5 64 200 1.501308 1.5067 0.359153 6.7 6.7411 0.613433
1.5 64 300 1.34832 1.421 5.390412 6.85 6.8475 0.036496
1.5 64 400 1.616713 1.3643 15.61273 6.75 6.5312 3.241481
1.5 96 200 1.58843 1.5798 0.543304 7.05 6.9039 2.07234
1.5 96 300 1.560049 1.5604 0.022499 6.8 6.9268 1.864706
1.5 96 400 1.516709 1.5101 0.435746 6.9 6.9479 0.694203

Table 4. Results for TRIP 700 steel grade

Kerf
Nozzle Stand off Travel Ra by
Kerf (by Ra
diameter distance speed Dev. (%) ANN Dev. (%)
(mm) ANN) (μm)
(mm) (mm) (mm/min) (μm)
(mm)
0.95 20 200 0.9780385 0.9994 2.184116 4.5 4.6297 2.882222
0.95 20 300 1.0265984 0.9859 3.964393 4.1 4.794 16.92683
0.95 20 400 0.9612056 1.0554 9.79961 5.2 5.692 9.461538
0.95 64 200 1.11525 1.1088 0.578346 6.25 5.9997 4.0048
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0.95 64 300 1.1554065 1.1524 0.260211 6.15 6.0786 1.160976

0.95 64 400 1.088564 1.0386 4.5899 5.95 6.0815 2.210084
0.95 96 200 1.163348 1.2351 6.167716 6.45 6.6784 3.541085
0.95 96 300 1.1400259 1.2733 11.69044 6.5 6.6293 1.989231
0.95 96 400 1.0820467 1.0371 4.15386 7.1 7.0151 1.195775
1.2 20 200 1.350591 1.3847 2.525487 6.2 6.1749 0.404839
1.2 20 300 1.329437 1.4459 8.760325 6.1 6.1591 0.968852
1.2 20 400 1.3352925 1.3235 0.88314 6.3 6.2844 0.247619
1.2 64 200 1.4216195 1.4469 1.778289 6.45 6.678 3.534884
1.2 64 300 1.4231235 1.5675 10.14504 7.3 7.6855 5.280822
1.2 64 400 1.3685895 1.451 6.021565 8.3 7.7036 7.185542
1.2 96 200 1.610029 1.5636 2.883737 7.9 7.8212 0.997468
1.2 96 300 1.5572765 1.604 3.000334 8.8 7.9298 9.888636
1.2 96 400 1.4473185 1.4209 1.825341 8.8 8.4547 3.923864
1.5 20 200 1.651102 1.6122 2.356123 7 6.9934 0.094286
1.5 20 300 1.464466 1.7267 17.90646 7 7.2331 3.33
1.5 20 400 1.739387 1.6887 2.914073 7 7.1597 2.281429
1.5 64 200 1.8041005 1.7254 4.362312 7.35 7.2777 0.983673

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 91

Table 4. (Continued)

Kerf
Nozzle Stand off Travel Ra by
Kerf (by Ra
diameter distance speed Dev. (%) ANN Dev. (%)
(mm) ANN) (μm)
(mm) (mm) (mm/min) (μm)
(mm)
1.5 64 300 1.8291155 1.7522 4.205065 8.1 7.8708 2.82963
1.5 64 400 1.791895 1.7174 4.157331 8.65 8.6593 0.107514
1.5 96 200 1.8016555 1.7028 5.486926 9.05 7.8651 13.09282
1.5 96 300 1.710008 1.7094 0.035555 8.2 8.3314 1.602439
1.5 96 400 1.730972 1.6332 5.648387 8.3 9.215 11.0241

Modeling of TRIP 800 Steel Grade

The MSE of training of the selected ANN was about 0.0153 and its training took 11
epochs to complete. The performance of the model is clearly indicated in Figure 5. The best
validation performance is 0.015296 at epoch 5. From this figure is also evident that validation
and testing group MSEs are higher than that of the training group, as expected; see also [10].
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 5. Results of the neural network training for AWJM of TRIP 800 grade.

For the evaluation of the generalization ability (quality) of the trained neural networks
linear fits between the outputs of the model and the experimental data, for all the
measurements, without discrimination to which group they belong, was performed. The graph
of the linear fit is presented in Figure 6; note that “Target” represent the experimental results
and “Output” the output values of the model. The best linear fit function is calculated as
being: Output=0.81Target+0.22, while the correlation coefficient R is equal to 0.89883 for
Kerf and Output=0.85Target+0.94, while the correlation coefficient R is equal to 0.93022 for
Ra (with R=1 meaning that the best linear fit is achieved and the A=T curve match perfectly).
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 92 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

Figure 6. Correlation between experimental data and neural network output.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 7. Regression analysis for the whole ANN model.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved. Artificial Neural Network Modeling of Surface Quality Characteristics… 93

Figure 8. Experimental data and ANN results for Ra and kerf in AWJM of TRIP 800 steel grade.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 94 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

A regression analysis was also performed for the whole model, see Figure 7. In the
graphs of this figure, training, validation and test data group are depicted, first separately and
then, all-in-one group.
Results in Figures 5-7 indicates that the proposed ANN can very satisfactorily predict the
output data required. Furthermore, in Figure 8, 3-D plots of the centre-line average surface
roughness (Ra) and the kerf versus the stand off distance and the travel speed are presented
for all three nozzle diameters used. In these figures both the experimental data and the neural
network outputs are presented. It is evident for all six plots that the experimental and the
calculated values exhibit small discrepancies, indicating once more the reliability of the
neural network constructed.

Modeling of TRIP 700 Steel Grade

Process parameters, measured values and the results from the proposed ANN applied in
AWJM of TRIP 700 grade as well as the deviation (“Dev.”) between measured and calculated
by ANN values, are summarized in Table 4.
Results obtained from ANN modeling are qualitatively almost identical with the ones
concerning TRIP 700 steel grade. The MSE of training of the selected ANN was about 0.271
and its training took 10 epochs to complete. The performance of the model is clearly indicated
in the next Figure 9.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 9. Results of the neural network training for AWJM of TRIP 700 grade.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 95

The best validation performance is 0.27091 at epoch 4. From this figure is again evident
that validation and testing group MSEs are higher than that of the training group, as expected.
The graph of the linear fit is presented in Figure 10; note that “Target” represent the
experimental results and “Output” the output values of the model. The best linear fit function
is calculated as being: Output=0.88Target+0.17, while the correlation coefficient R is equal to
0.9533 for Kerf and Output=0.82Target+1.2, while the correlation coefficient R is equal to
0.94728 for Ra (with R=1 meaning that the best linear fit is achieved and the A=T curve
match perfectly).
Results in Figures 9 and 10 indicate that the proposed ANN can very satisfactorily predict
the output data required. Furthermore, in Figure 11, 3-D plots of the centre-line average
surface roughness (Ra) and the kerf versus the stand off distance and the travel speed are
presented for all three nozzle diameters used. In these figures both the experimental data and
the neural network outputs are presented. It is evident for all six plots that the experimental
and the calculated values exhibit small discrepancies; a behaviour which was already reported
for TRIP 800 steel grade.
In general, ANN models for both TRIP 800 and TRIP 700 steel grades were proven to
perform well for AWJM, giving reliable predictions and providing thus a possible way to
avoid time- and money-consuming experiments.
Note also that modeling of AWJM by ANN could be combined with a Knowledge Based
System (KBS) based on both rules and Object-Oriented structures such as Semantic Networks
(SN) and Frames, which maintain, process and relate taxonomies of products and materials,
their attributes and properties, processes’ parameters, quality control specifications and
manufacturability rules; see preliminary results on the topic reported in Ref. [28].
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 10. Correlation between experimental data and neural network output for TRIP 700 steel grade.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved. 96 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

Figure 11. Experimental data and ANN results for Ra and kerf in AWJM of TRIP 700 steel grade.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 97

CONCLUSION
Abrasive waterjet machining (AWJM) is a technique for cutting or separating materials
by means of a high-velocity slurry jet, formed as a result of injecting abrasive particles to a
waterjet ejected by a nozzle. The main advantages of the AWJM are being able to cut
versatile geometries and its ability to cut both ductile materials like aluminium, brass, steel
and titanium and brittle materials like glass, stone and ceramics without any influence on their
microstructure. In sheet steel processing with AWJM the primary output characteristics, as far
as the quality of the cut is concerned are the kerf shape and the cut surface roughness. The
average kerf width and the centre-line average surface roughness (Ra) were considered in the
present study. For the prediction of these features in AWJM two different ANNs were
developed using the neural network toolbox of Matlab®. Models constructed possess three
inputs, namely nozzle diameter, stand off distance and travel speed, one hidden layer with
five neurons and two outputs (kerf and Ra). Training of the models was performed with data
from an extensive series of statistically designed experiments concerning AWJM of two TRIP
(transformation induced plasticity) type steel sheets.
The reported results indicate that the proposed ANNs models can satisfactorily predict
the surface roughness and the mean kerf in AWJM; moreover, they can be considered as
valuable tools for the process planning and moreover, provide a possible way to avoid time-
and money-consuming experiments.

REFERENCES
[1] Jain, VK. Advanced (Non-traditional) Machining Processes (ch. 11). In Davim, JP,
editor. Machining - Fundamentals and Recent Advances. Springer; 2008; 299-327.
[2] Arola, D; Ramulu, M. Material removal in abrasive waterjet machining of metals
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Surface integrity and texture. Wear, 1997 210, 50-58.

[3] Hashish, M. Optimization factors in abrasive waterjet machining. Journal of
Engineering for. Industry (Tranactions of the. ASME) 1991 113, 29–37.
[4] Momber AW; Kovacevic, R. Principles of Abrasive Water Jet Machining, Springer,
1997.
[5] Vassilakis, S; Petropoulos, G; Sokovic, M; Vaxevanidis, NM; Ntziantzias, I. Macro-
geometric and Surface Roughness Quality of TRIP Steel Sheets Processed by Abrasive
Waterjet Machining. Proc. 9th International Conference on Management of Innovative
Technologies-MIT 2007, 8th - 10th October, 2007, Fiesa, Slovenia, pp. 196-203.
[6] Benardos, PG; Vosniakos, GC. Predicting surface roughness in machining: a review.
International Journal of Machine Tools and Manufacture, 2003 43, 833-844.
[7] Petropoulos, G; Mata, F; Davim, JP. Statistical study of surface roughness in turning of
peek composites. Materials and Design, 2008 29, 218-223.
[8] Jesuthanam, CP; Kumanan, S; Asokan, P. Surface roughness prediction using hybrid
neural networks. Machining Science and Technology, 2007 11, 271-286.
[9] Markopoulos, A; Vaxevanidis, NM; Petropoulos, G; Manolakos, DE. Artificial neural
network modeling of surface finish in Electro-discharge machining of tool steels.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 98 N. M. Vaxevanidis, A. Markopoulos and G. Petropoulos

Proceedings of ESDA2006: 8th Biennial ASME Conference on Engineering Systems

Design and Analysis, July 4-7, 2006, Torino, Italy. (paper ESDA2006-95609).
[10] Markopoulos, A; Manolakos, DE; Vaxevanidis, NM. Artificial Neural Network Models
for the Prediction of Surface Roughness in Electrical Discharge Machining. Journal of
Intelligent Manufacturing, 2008 19, 283-292.
[11] Snoeys, R; Staelens, F; Dekeyser, W. Current Trends in Non-Conventional Material
Removal Processes. Annals of the CIRP, 1986 35, 467-480.
[12] Mc Geough JA. Advanced Methods of Machining, Chapman and Hall, London, 1988.
[13] Ma, C; Deam, RT. A correlation for predicting the kerf profile from abrasive water jet
cutting. Experimental Thermal and Fluid Science, 2006 30, 337-343.
[14] Hoogstrate, AM; van Luttervelt, CA. Opportunities in Abrasive Water-Jet Machining.
Annals of the CIRP, 1997 46, 697-714.
[15] Kovacevic, R., Hashish, M., Mohan, R., Ramulu, M., Kim, T.J., Geskin, E.S. State of
the art of research and development in abrasive waterjet machining. Journal of
Manufacturing Science and Engineering (Transactions of the ASME), 1997 119, 776-
785.
[16] Wang, J; Wong, WCK. A study of abrasive waterjet cutting of metallic coated sheet
steels. International Journal of Machine Tools and Manufacture, 1999 39, 855–870.
[17] Davalo, E; Naim, P; Rawsthorne, A. Neural networks, Macmillan Education, London,
1991.
[18] Fausset, LV. Fundamentals of neural networks: Architectures, algorithms and
applications, Prentice Hall, New Jersey, 1994.
[19] Haykin, S. Neural networks, a comprehensive foundation, Prentice Hall, New Jersey,
1999.
[20] Feng, C-XJ; Yu, Z-G; Kusiak, A. Selection and validation of predictive regression and
neural network models based on designed experiments. IIIE Transactions, 2006 38, 13-
23.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[21] Tsai, K-M; Wang, PJ. Predictions on surface finish in electrical discharge machining
based upon neural network models. International Journal of Machine Tools and
Manufacture, 2001 41, 1385-1403.
[22] Caydas, U; Hascalık, A. A study on surface roughness in abrasive waterjet machining
process using artificial neural networks and regression analysis method. Journal of
Materials Processing Technology, 2008 202, 574–582.
[23] Deb, S; Dixit, US. Intelligent Machining: Computational Methods and Optimization,
(ch. 12). In Davim, JP, editor. Machining - Fundamentals and Recent Advances.
Springer; 2008; 329- 358.
[24] Parikh PJ; Lam, SS. Parameter estimation for abrasive water jet machining process
using neural networks. International Journal of Advanced Manufacturing Technology,
2009 40, 497–502.
[25] Yang, L; Song, J; Hu, B. Neural network parametric modelling of abrasive waterjet
cutting quality. International Journal of Abrasive Technology, 2007 1, 198-207.
[26] Petropoulos, G; Tsolakis, N; Vaxevanidis, N; Antoniadis, A. Topographic description
of abrasive waterjet machined surfaces. Proc. 2nd European Conference on Tribology -
ECOTRIB 2009, 7-10 June 2009, Pisa, Italy, pp. 309-314.
[27] Demuth, H; Beale, M. Neural networks toolbox for use with Matlab, User’s guide.
2001.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Artificial Neural Network Modeling of Surface Quality Characteristics… 99

[28] Karbadakis, NA; Mekras, ND; Kontovazenitis, PV; Petropoulos, GP; Vaxevanidis, NM.
Implementation of a knowledge based system for modeling non-conventional
machining processes. Nonconventional Technologies Review, 2007 1, 43-48.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 6

MULTI-OBJECTIVE OPTIMISATION OF CUTTING

PARAMETERS FOR DRILLING ALUMINIUM AA1050

Ramón Quizaa and J. Paulo Davim∗b

a
Department of Mechanical Engineering, University of Matanzas,
Autopista a Varadero, km 3 1/2, Matanzas 44740, CUBA
b
Department of Mechanical Engineering,
University of Aveiro Campus Santiago,
3810-193 Aveiro, PORTUGAL

ABSTRACT
This paper presents a multi-objective approach to the optimisation of drilling process
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

of aluminium AA1050, considering two important, conflicting objectives: material

removal rate and obtained surface roughness. From experimental data were developed
statistical models for the significant variables involved in the cutting process. To carry
out the optimisation process, two methods were used: the numerical method GRG2 and a
multi-objective genetic algorithm. Both methods were compared taking into account their
respective advantages and drawbacks. The paper remarks the conveniences of the a
posteriori approach in the optimisation and decision-making process.

NOMENCLATURE
d Hole diameter [mm];
FF Feed force [N];
f Feed [mm/rev];
M Material removal rate [mm3/s];
MT Cutting torque [N·m];
n Spindle speed [rpm];
P Cutting power [kW];

∗
Corresponding Author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 102 Ramón Quiza and J. Paulo Davim

PMOT Motor power [kW];

Ra Surface roughness [μm];
V Cutting speed [m/min];
η Transmission efficiency;

1. INTRODUCTION
In general, machining aluminium alloys required cutting fluids because the tendency of
the chip to stick to the rake face of the cutting tool. However, the appropriate selection of the
cutting fluid and lubrication technique is a hard task because the great number of parameters
involved. Recently Mendes et al. [1] investigated the performance of cutting fluids when
machining aluminium alloys. Subsequently, Davim et al [2] presented an experimental study
on drill of aluminium (AA1050) under dry, minimum quantity of lubricant and flood-
lubricated conditions. The results of the tests show it is possible to obtain similar
performances to flood-lubricated conditions by using minimum quantity of lubricant.
Optimal selection of cutting parameters is a key issue in any machining process.
However, it remains been a difficult work [3] in spite of much work has been done in this
field and several optimisation approach have been proposed [4].
Although the common practice in cutting parameters optimisation is considering the
whole problem as single-objective, the complex nature of the cutting processes actually
makes the problem multi-objective [5]. Some multi-objective approaches have been reported
in cutting parameters optimisation [6-8] but mainly they use a priori techniques, where the
decision maker combines the different objectives into a scalar cost function. This actually
simplifies the multi-objective problem, turning it into a single-objective one [9].
On the other hand, in the a posteriori techniques, the decision-maker is presented with a
set of non-dominated optimal candidate solutions and chooses from that set. These solutions
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

are optimal in the wide sense that no other solution in the search space is superior to them
when all optimisation objectives are simultaneously considered [10]. They are also known as
Pareto-optimal solutions. The main advantage of this approach is to allow selecting the most
proper solution under different conditions, without running again the optimisation process.
In obtaining the Pareto-optimal solutions, classical optimisation methods (weighted sum
methods, goal programming, min–max methods, etc.) are not efficient, because they cannot
find multiple solutions in a single run, thereby requiring them to be applied as many times as
the number of desired Pareto-optimal solutions [11].
On the contrary, studies on evolutionary algorithms have shown that these methods can
be efficiently used to eliminate most of the above-mentioned difficulties of classical
methods [12]. Nevertheless, there is not any certainty that evolutionary methods can obtain
actual optimal values, for each specific problem [13].
Quiza and co-workers [14] carried out an a posteriori multi-objective optimisation for
drilling laminate composite materials, by using a genetic algorithm, considering productivity
and surface quality as optimisation targets. Outputs show the advantages of the a posteriori
approach, but they are not compared with classical (numerical) methods.
In this paper an aluminium AA1050 drilling process is modelled and optimised,
considering two objectives: productivity (material removal rate) and surface quality

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimisation of Cutting Parameters for Drilling … 103

(roughness). The a posteriori approach is used and two methods, numeric and evolutive, are
employed to find the Pareto-optimal solutions. Outcomes form both methods are compared.

2. EXPERIMENTAL STUDY
In order to achieve the objective of this experimental work, Aluminium AA1050
(aluminium with max. 0.4%Fe, 0.25%Si, 0.05% Cu, 0.05%Mg, 0.05%Mn, 0.05%V, 0.05%
Zn and 0.03%Ti) was tested. A machining center “VCE500 MIKRON” with 11kW spindle
power and maximum spindle speed of 7500 rpm were used to perform the cut holes (diameter
5 mm) in 15 mm thick aluminium discs under flood-lubricated conditions.. A helical K10
drill (R415.5-0500-30) was manufactured according DIN6537 by Sandvik®. The drill was a
point angle 140º, 10% cobalt grade and 28 mm of flute length. A Kistler® piezoelectric
dynamometer 9272 with a load amplifier was used to acquire the cutting torque, MT; and the
feed force, FF. Data acquisitions were made through piezoelectric dynamometer by interface
RS-232 to load amplifier and PC using the appropriate software Dynoware Kistler®.
The surface roughness was evaluated (Ra according to ISO 4287/1) with a Hommeltester
T1000 profilometer. A full factorial design was carried out for the two independent variables:
cutting speed, V, and feed, f. Experimental data is shown in Table 1.

Table 1. Experimental data

V f n MT FF Ra
No.
[m/min] [rev/min] [rpm] [N·m] [N] [μm]
1 60 0.15 3820 1.02 833 1.80
2 60 0.20 3820 1.27 1113 2.09
3 60 0.25 3820 1.54 1373 2.13
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

4 75 0.15 4775 1.00 897 1.78

5 75 0.20 4775 1.22 1154 1.93
6 75 0.25 4775 1.45 1367 2.00
7 90 0.15 5730 0.98 826 1.64
8 90 0.20 5730 1.18 1110 1.87
9 90 0.25 5730 1.32 1206 1.75

3. MODELLING
Models were obtained from the experimental data by using multiple regression analyses.
For every studied variable several models were tried and analysed. They were compared by
considering their correlation coefficients and the variance of the model. Furthermore, the
influence of each independent variable was analysed by means of the corresponding t-Student
tests, and the lack of autocorrelation was indicated by the outcomes of the Durbin-Watson
(DW) tests. Finally, the most convenient model was selected for each dependent variable.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 104 Ramón Quiza and J. Paulo Davim

For the surface roughness, the most suitable model is the quadratic form:

Ra = −0.29 − 45.33 f 2 − 0.04347 fV + 23.59 f (1)

The R-squared statistic for the above-mentioned model is 95.6 % and the mean absolute
error is 0.0276. Since the p-value in the ANOVA table (see Table 2) is less than 0.01, there is
a statistically significant relationship between the variables at the 99% confidence level. The
DW statistic has a value of 2.504 and an associated p-value of 0.2209, therefore there is no
there is no indication of serial autocorrelation in the residuals. In Figure 1 is shown the
graphical representation of the adjusted model for the surface roughness.

The most convenient model for the cutting torque is:

9.758 f 0.7053
MT = (2)
V 0.2186
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 1. Graphical representation of surface roughness model.

Table 2. ANOVA for the surface roughness model

Source Sum of Squares D.F. Mean Square F-Ratio P-Value

Model 0.2046 3 0.0682 36.31 0.0008
Residual 0.0094 5 0.0019 --- ---
Total (Corr.) 0.2140 8 --- --- ---

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimisation of Cutting Parameters for Drilling … 105

Table 3. ANOVA for the cutting torque model

Source Sum of Squares D.F. Mean Square F-Ratio P-Value

Model 0.2076 2 0.1038 165.3 0.0000
Residual 0.0038 6 0.0006 --- ---
Total (Corr.) 0.2114 8 --- --- ---

This model has an R-squared statistic of 98.2 % and a mean absolute error of 0.0165. As
can be seen in Table 3, the P-value of the ANOVA is less than 0.01, so there is a statistically
significant relationship between the variables at the 99% confidence level. The DW test
showed a value of 0.0868, with an associated p-value of 0.060. Since this value is greater than
0.05 there is no indication of serial autocorrelation in the residuals. A graphical representation
of the cutting torque model is shown in Figure 2.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 2. Graphical representation of cutting torque model.

For the feed force, was adjusted the following model:

FF = 171.0 + 5501 f − 11.57 fV (3)

For this model, the R-square statistic is 94.7 % and the mean absolute error 38.29. In
Table 4 is shown the ANOVA. Since the p-value is less than 0.01 there is a statistically
significant relationship between the variables at the 99% confidence level. Because the value
of the DW statistic is 1.450, with an associated p-value of 0.114, there is no indication of
serial autocorrelation in the residuals. In Figure 3 is shown the graphical representation of the
feed force model.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 106 Ramón Quiza and J. Paulo Davim

Table 4. ANOVA for the feed force model

Source Sum of Squares D.F. Mean Square F-Ratio P-Value

Model 329551 2 164775 54.28 0.0001
Residual 18213 6 3035 --- ---
Total (Corr.) 347764 8 --- --- ---
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 3. Graphical representation of feed force model.

4. OPTIMISATION
4.1. Formulation of the Optimisation Problem

In the analysed optimisation problem, as decision variables were considered the cutting
parameters: feed and speed. Two conflicting objectives were simultaneously optimised. The
first is the material removal rate, M, which can be computed as:

M = 250dfV (4)

This parameter allows evaluating the productivity of the drilling process and is inversely
proportional to the machining time. The second objective is the surface roughness, which can
be computed by the adjusted model (Eq. 1). It characterises the surface quality of the
elaborate hole. As can be noted, the first objective must be maximised since the second one

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimisation of Cutting Parameters for Drilling … 107

must minimised. Some constraints must be considering in the optimisation process. In the
first place, the machining parameters must be into the experimental ranges, in order to
guarantee the validity of the previously obtained models.

f MIN ≤ f ≤ f MAX (5a)

VMIN ≤ V ≤ VMAX (5b)

Moreover, the cutting torque and the feed force must be less or equal to the maximum
values allowed by the machine tool.

M T ≤ M T − MAX (6a)
FF ≤ FF − MAX (6b)

Furthermore, the cutting power:

2π M T n
P= ≤ η PMOT
6 × 104 (7)

must be satisfied by the machine tool motor, considering the losses in the transmission.
In this problem the following values was considered (see Table 5):

Table 5. Optimisation parameters for the considered problem

Parameter Value
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Hole diameter 5 mm
Maximum allowed cutting torque 1.25 N·m
Maximum allowed feed force 1200 N
Motor power 5.6 kW
Transmission efficiency 75 %

4.2. Optimisation by Numerical Method

In order to obtain a Pareto’s front by numerical method, one objective is transformed into
constraint. Several values are successively assigned to this objective and the other one is
optimised once for every value established for the first one. In this problem, surface
roughness was transformed into constraint and material removal rate was kept as objective.
To solve the optimisation problem the Generalized Reduced Gradient (GRG2) algorithm was
used. In Table 6, there are shown the outcomes of the optimisation process.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 108 Ramón Quiza and J. Paulo Davim

Table 6. Outcomes of the optimisation by using the GRG2 method

V f MT FF n P Ra M
No.
[m/min] [rev/min] [N·m] [N] [rpm] [kW] [μm] [mm3/s]
1A 90.0 0.219 1.25 844 5730 0.750 1.85 49264
2A 90.0 0.185 1.11 885 5730 0.666 1.80 41706
3A 90.0 0.171 1.05 934 5730 0.630 1.75 38508
4A 90.0 0.160 1.00 996 5730 0.601 1.70 36088
5A 90.0 0.151 0.96 1148 5730 0.577 1.65 34059

4.3. Optimisation by Using Genetic Algorithm

The other considered approach carries out the optimisation process by using a micro
genetic algorithm. This genetic algorithm was originally proposed by Coello and Toscano
[15] and it was used by Quiza and co-workers to solve optimisation problems in machining
processes [14, 16].

Table 7. Outcomes of the optimisation by using the micro-GA

V f MT FF n P Ra M
No.
[m/min] [rev/min] [N·m] [N] [rpm] [kW] [μm] [mm3/s]
1B 89.0 0.217 1.25 1141 5665 0.739 1.86 48275
2B 89.2 0.213 1.23 1123 5676 0.730 1.85 47480
3B 89.3 0.204 1.19 1082 5685 0.709 1.84 45546
4B 89.8 0.198 1.16 1055 5716 0.697 1.83 44446
5B 89.3 0.191 1.14 1024 5682 0.677 1.82 42618
6B 89.2 0.184 1.11 993 5677 0.659 1.80 41017
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

7B 89.8 0.180 1.09 974 5715 0.652 1.79 40398

8B 90.0 0.176 1.07 956 5729 0.643 1.77 39599
9B 89.7 0.174 1.06 948 5709 0.636 1.76 39009
10B 89.7 0.168 1.04 921 5713 0.621 1.74 37690
11B 89.2 0.164 1.02 904 5679 0.607 1.72 36574
12B 88.9 0.160 1.00 887 5657 0.595 1.71 35542
13B 89.2 0.158 0.99 877 5679 0.592 1.69 35236
14B 89.2 0.155 0.98 864 5676 0.584 1.68 34549
15B 89.1 0.153 0.97 855 5675 0.578 1.67 34094
16B 88.7 0.150 0.96 842 5649 0.568 1.65 33275

For the considered application, the algorithm’s parameters were selected as follow: static
population size: 500, dynamic population size: 20; chromosome length: 64 (32 for each
decision variable), maximum epochs count: 100; maximum evolutionary periods count: 50;
mutation likelihood: 0.01 and maximum Paretian size: 25. After carrying out the optimisation
process, there were obtained 10 near-optimal solutions, which are shown in Table 7.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Multi-Objective Optimisation of Cutting Parameters for Drilling … 109

5. DISCUSSION
To compare outcomes from both approaches, a Pareto's fronts graph was constructed (see
Figure 4). In this graph, non-dominated solutions are arranged in a coordinated plane whose
axes are the optimisation objectives.
As can be seen, the solutions contained in the front obtained by the numerical method are
better than contained in the front obtained by micro genetic algorithm. The term “better”
means that for every solution belong to the numerical obtained front there is a point in the
genetic algorithm obtained front that dominates it. Moreover, the genetic front has a higher
density of points, allowing selecting the most proper solution from a larger set. Naturally,
executing new runs of the optimisation process must increase the amount of points in the
numerical front. Nevertheless, this must be done by try-and-error and can take a great amount
of time and personal effort.
The main drawback of the genetic algorithm approach is its elevated computational cost,
determined by the high amount of evaluations of the objective functions that must be
executed. However, this fact is widely paid back by a practically total elimination of human
interaction in the optimisation process. It should be noted that the try-and-error process, used
in the numerical approach, is not only highly time consuming but also very tedious and
ineffective from the human user point of view.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 4. Paretos’s fronts.

CONCLUSION
As conclusions of the work can be remarked the convenience of the use of evolutionary
algorithms in multi-objective optimisation of cutting processes, instead its high computational
cost. Finally, the advantage of the a posteriori approach in the cutting parameters
optimisation must be remarked. From the Pareto’s front graph a flexible decision-making can
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 110 Ramón Quiza and J. Paulo Davim

be done, considering different technological requirement in the elaborated holes. By using a

pre-elaborated Pareto’s front, the selection of the most convenient cutting parameters, for an
established surface roughness can be done without carry out the optimisation process once
again. For example, if the only important criterion is productivity, the extreme point 1B must
be chosen. This point has a poor surface quality but corresponds to the highest material
removal rate. On the contrary, if the surface quality must be keep as low as possible, the other
extreme point, 16B, is the most appropriated instead its low productivity. In an intermediate
case, where a hole must be drilled with a surface roughness of Ra 1,75 μm, point 10B must be
selected.

ACKNOWLEDGMENTS
The authors acknowledge to the MSc A. Festas and P. Reis as well as the Mechanical
Engineers C. Peixoto and R. Gomes for their participation in the experimental work carried in
University of Aveiro.

REFERENCES
[1] Mendes, O. C.; Ávila, R. F.; Abrão, A. M.; Reis, P.; Davim, J. P. Ind. Lubric. Trib.
2006, 58 (5), 260-268.
[2] Davim, J. P.; Sreejith, P. S.; Gomes, R.; Peixoto, C. Proc. IMechE Part B: J. Eng.
Manuf. 2006, 220, 1605-1611.
[3] Jain, R. K.; Jain, V. K. J. Mater. Process. Technol. 2000, 108 (1), 62-67.
[4] Mukherjee, I.; Ray, P. K. Comput. Ind. Eng. 2006, 50 (1), 15-34.
[5] Yih-Fong, T.; Fu-Chen, C. Int. J. Machin. Machinab. Mat. 2006, 1 (1), 76-93.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

[6] Lee, B. Y.; Tarng, Y. S.; Lii, H. R. J. Mater. Process. Technol. 2000, 105 (1), 1-6.
[7] Zuperl, U.; Cus, F. Robot. Comput. Int. Manuf. 2003, 19 (1-2), 189-199.
[8] Cus, F.; Balic, J. Robot. Comput. Int. Manuf. 2003, 19 (1-2), 113-121.
[9] Van Veldhuizen, D. A.; Lamont, G. B. Evol. Comput., 2000, 8 (2), 125-147.
[10] Abbass, H. A.; Sarker, R. A. In: Congress on Evolutionary Computation, Piscataway,
NJ (U.S.A.), 2001; pp 971-978.
[11] Coello, C. A. IEEE Comput. Intell. Mag. 2006, 1 (1), 28-36.
[12] Soodamani, R.; Liu, Z. Q. Int. J. Approx. Reason. 2000, 23(2), 85-109.
[13] Zitzler, E.; Laumanns, M.; Bleuler, S. In: Metaheuristics for Multiobjective
Optimisation; Gandibleux, X.; Sevaux, M.; Sörensen, K.; T'kindt, V. Eds.; Springer-
Verlag: Berlín (Alemania), 2004; pp 3-37.
[14] Quiza, R.; Reis, P.; Davim, J. P. Compos. Sci. Technol. 2006, 66 (15), 3083-3088.
[15] Coello, C. A.; Toscano, G. In: First International Conference on Evolutionary Multi-
Criterion Optimization. Zitzler, E. Ed.; Springer-Verlag: New York (U.S.A.), 2001;
pp 126-140.
[16] Quiza, R.; Rivas, M.; Alfonso, E. Eng. Appl. Artif. Intell. 2006, 19 (2), 127-133.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 7

APPLICATION OF FUZZY LOGIC IN

MANUFACTURING: A STUDY ON MODELING OF
CUTTING FORCE IN TURNING GFRP COMPOSITES

K. Palanikumar∗ and J. Paulo Davim

Sri Sairam Institute of Technology,
Sai Leo Nagar, Chennai – 600 044, INDIA.
Department of Mechanical Engineering, University of Aveiro
Campus Santiago, 3810-193 Aveiro, PORTUGAL

ABSTRACT
Artificial intelligence (AI) tools are playing an important role in manufacturing
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

engineering. They are used for modeling, control and optimization of parameters in
manufacturing. Fuzzy logic is one of the important tools in artificial intelligence and is
used in varieties of manufacturing applications. This chapter discusses the application of
fuzzy logic for modeling process parameters for cutting force in turning of GFRP
composites. Taguchi method in design of experiments principle is used for conducting
the experiments. An L27 (313) orthogonal array is used for the investigation. The cutting
parameters selected were: cutting speed, feed and depth of cut. Fuzzy rules were
developed for correlating the cutting parameters with cutting force in turning of GFRP
composites. The model predicted values and measured values are fairly close to each
other. The confirmation test results proved the fact that the developed models are
effectively representing the cutting force in turning GFRP composites.

1. INTRODUCTION
Artificial intelligence (AI) is the study of ideas that enables computers to be intelligent.
AI systems have a significant impact on the automation of engineering manufacture. This will
provide intelligent interfaces to sophisticated engineering analysis automate the use of

∗
Corresponding author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 112 K. Palanikumar and J. Paulo Davim

manufacturing data base and provide engineering assistance to handle varieties of engineering
tasks. AI is used for modeling, control and optimization of parameters in manufacturing.
Fuzzy logic is one of the important tool in artificial intelligence, which can be used for
modeling, analysis and optimization of manufacturing processes. The fuzzy set theory
initiated by Zadeh [1] is a mathematical theory of inexact reasoning that allows modeling of
the reasoning process in human linguistic terms. This theory proved to be an effective means
for dealing with objectives that are linguistically specified. Linguistic term such as low,
medium and high may be defined by fuzzy sets [2]. Fuzzy logic has been applied successfully
for many manufacturing processes especially in machining processes. AI techniques are
insensitive to the noise produced in the machining data due to variations in parameters such
as material properties, temperature, and cutting geometry, the same are being used in
modeling of machining processes in the present day research. Suleyman Yaldiz et al. [3] have
evaluated the cutting force in turning. They have compared the experimental results obtained
by designed dynamometer to fuzzy model for predicting cutting force. Surface roughness
prediction for turning was carried out by Yue Jiao et al. [4]. They have used Fuzzy adaptive
network for modeling the machining process Fang [5] has used fuzzy logic diagnosis for
machining condition monitoring of finish turning process. Palanikumar et al. [6] have used
fuzzy logic for optimization of machining parameters in machining GFRP composites. This
method is highly useful when the process is complex and uncertain in nature. The works
carried out by the researchers indicated that fuzzy logic can be applied for manufacturing
processes. In the present chapter, application of fuzzy logic for the modeling of cutting force
in turning GFRP composites is introduced. Fuzzy rule based modeling technique is used for
the present investigation.
Glass fiber reinforced composite finds many applications including construction
industries, marine, electrical, chemical, pharmaceutical, transportation, sport goods
manufacturing industries etc. Composites materials exhibit good resistance to corrosion and
wear. The tailorability of composites for specific applications has been one of its greater
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

advantages and also one of the more perplexing challenges to adopting are alternative to
metallic materials in many situations. Even though GFRP parts may be produced by molding
processes, they require further machining to facilitate dimensional control for easy assembly
and control of surface quality for functional aspects. Machining of FRP parts includes
turning, drilling, milling, grinding and the like. The machining of fiber reinforced composite
materials are different from the traditional engineering materials. Fibre delamination,
fuzziness, debonding of fiber and matrix materials, excessive wear on the cutting tool are
some of the problems which encountered during the machining of FRPs [7]. Over the last
several years, the phenomenon of machining of FRP composites has been analyzed by
researchers. Everstine and Rogers [8] proposed an analytical theory of machining FRPs. In
their classical study, they prepared a theory related to plane deformation of incompressible
composites reinforced by strong parallel fibers.
In machining process, cutting force plays an important role in deciding the surface and
power requirement for cutting. Generally the machining force increases with the feed rate and
decreases with the cutting velocity [9]. The machining force evaluation in cutting GFRP
composites has been carried out by Paulo Davim et al. [10,11]. They have analyzed the
cutting force in drilling and milling operation using cemented carbide drill bits and end mill
cutter. The influence of cutting force in turning GFRP composites is carried out by Sang-Ook

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 113

et al. [12]. It was found that the single crystal diamond tool was performed better than both
poly crystal diamond and cubic boron nitride tools in turning of GFRP composites.
Eventhough many attempts have been made to analyze the cutting force in turning of
GFRP composites, there is no systematic study has been reported. The present chapter
investigates the effect of different machining parameters on cutting force in machining of
GFRP composites. The experiments are conducted on a lathe using Poly Crystalline Diamond
(PCD) cutting tools. The experiments are conducted based on L27 orthogonal array with
prefixing cutting parameters. The machining parameters considered for the experiments are
cutting speed, feed, and depth of cut. Fuzzy rule based modeling technique is used for
modeling the machining parameters in machining of GFRP composites. The results indicate
that fuzzy logic modeling technique can be successfully applied for the prediction of cutting
force in turning of GFRP composites.

2. EXPERIMENTAL DETAILS
The work piece material used for the present investigation is glass fiber reinforced
polymer composite pipes consisting of E-glass fibers and polyester resin. The pipes are
manufactured through filament winding process. The fiber volume fraction for the composites
is 55%. The machining operation considered is turning and the machine used for the
machining is Bharat make all geared lathe of model Nagmati 175 with a spindle speed range
of 54 - 1200 rpm and power of 2.25 kW. The cutting tool used for the machining is PCD tool
which is shown in Figure 1. The experiments for the investigation are conducted by using
Taguchi’s orthogonal array [13], the suitable array selected for the investigation was L27,
which needs 27 runs and has 26 DOFs. It can conduct three levels of parameters. To check
the degrees of freedom (DOFs) in the experimental design, for the three-level test, the three
main factors take 6 (3*(3-1)) DOFs. The DOF for three second-order interactions is 12
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

(3*((3-1)*(3-1))) and the total DOFs required is 18. As per Taguchi’s experimental design
method, the total DOFs of selected orthogonal array must be greater than or equal to the total
DOFs required for the experiment and hence L27 orthogonal array has been selected [14].

Figure 1. Tool holder and cutting tool used for the investigation.

The cutting parameters selected for the present investigation are cutting speed (V), depth
of cut (d) and feed (f). The cutting parameters used and their levels considered are presented
in Table 1. The three level L27 orthogonal array is shown in Table 2, where the numbers 1, 2,
and 3 stand for the levels of the factors. This array specifies 27 experimental run and have 13
columns. The assignment of cutting parameters and interactions to columns is presented in
Table 2.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 114 K. Palanikumar and J. Paulo Davim

Table 1. Experimental factors and levels

Factor Notation Unit Factor levels

1 2 3
Cutting speed V m/min 82 126 170
Depth of cut d mm 0.5 1.0 1.5
Feed f mm/rev 0.096 0.143 0.191

Table 2. Experimental Layout Using L27 Orthogonal Array

Column numbers
1 2 3 4 5 6 7 8 9 10 11 12 13
Trial No. V d V*d V*d f V*f V*f d*f -- -- d*f -- --
1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 2 2 2 2 2 2 2 2 2
3 1 1 1 1 3 3 3 3 3 3 3 3 3
4 1 2 2 2 1 1 1 2 2 2 3 3 3
5 1 2 2 2 2 2 2 3 3 3 1 1 1
6 1 2 2 2 3 3 3 1 1 1 2 2 2
7 1 3 3 3 1 1 1 3 3 3 2 2 2
8 1 3 3 3 2 2 2 1 1 1 3 3 3
9 1 3 3 3 3 3 3 2 2 2 1 1 1
10 2 1 2 3 1 2 3 1 2 3 1 2 3
11 2 1 2 3 2 3 1 2 3 1 2 3 1
12 2 1 2 3 3 1 2 3 1 2 3 1 2
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

13 2 2 3 1 1 2 3 2 3 1 3 1 2
14 2 2 3 1 2 3 1 3 1 2 1 2 3
15 2 2 3 1 3 1 2 1 2 3 2 3 1
16 2 3 1 2 1 2 3 3 1 2 2 3 1
17 2 3 1 2 2 3 1 1 2 3 3 1 2
18 2 3 1 2 3 1 2 2 3 1 1 2 3
19 3 1 3 2 1 3 2 1 3 2 1 3 2
20 3 1 3 2 2 1 3 2 1 3 2 1 3
21 3 1 3 2 3 2 1 3 2 1 3 2 1
22 3 2 1 3 1 3 2 2 1 3 3 2 1
23 3 2 1 3 2 1 3 3 2 1 1 3 2
24 3 2 1 3 3 2 1 1 3 2 2 1 3
25 3 3 2 1 1 3 2 3 2 1 2 1 3
26 3 3 2 1 2 1 3 1 3 2 3 2 1
27 3 3 2 1 3 2 1 2 1 3 1 3 2

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 115

The cutting force developed during machining is measured by using a Kistler quartz 3-
component dynamometer type 9257B. The arrangement used for measuring cutting force is
given in Figure 2.

Figure 2. Schematic Diagram of Experimental Setup.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 3. Typical cutting force observed in machining of GFRP composites.

The experiments are repeated for three times and the average values are used for further
analysis. The dynamometer measures the active cutting force regardless of its application
point. Both the average value of the force and the dynamic force increase may be measured.
The usable frequency range depends mainly on the resonance frequency of the entire
measuring rig. The force to be measured is introduced via a top plate and distributed between

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 116 K. Palanikumar and J. Paulo Davim

four 3-component force sensors arranged between the base and top plates. Each of the sensors
has three pairs of quartz plates, one sensitive to pressure in the z-direction and the other two
to shear in the x and y directions respectively. The measurement is virtually without
displacement. The dynamometer is connected to a 3-channel charge amplifier type 5807A
through a connecting cable type 1687B5 which in turn is connected to the PC by a 37-pin
cable from the A/D board. The dynamometer is calibrated for the cutting force in the range
from 0 to 1000 N. The typical cutting force observed from the Kistler dynamometer is
presented in Figure 3.
The experimental condition used and the experimental results are presented in Table 3.

Table 3. Experimental results

S.No Cutting speed (V), Depth of cut (d), Feed rate (f), Cutting force(Fc),
m/min mm mm/rev. N
1 82 0.5 0.096 178.5
2 82 0.5 0.143 201.2
3 82 0.5 0.191 231.5
4 82 1 0.096 192
5 82 1 0.143 207.8
6 82 1 0.191 242.4
7 82 1.5 0.096 212.5
8 82 1.5 0.143 213
9 82 1.5 0.191 251.4
10 126 0.5 0.096 171.2
11 126 0.5 0.143 190.8
12 126 0.5 0.191 225
13 126 1 0.096 187.9
14 126 1 0.143 201.7
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

15 126 1 0.191 236.6

16 126 1.5 0.096 172.3
17 126 1.5 0.143 212.3
18 126 1.5 0.191 241.8
19 170 0.5 0.096 173.5
20 170 0.5 0.143 192.6
21 170 0.5 0.191 221.6
22 170 1 0.096 179.7
23 170 1 0.143 203.6
24 170 1 0.191 234.6
25 170 1.5 0.096 191.6
26 170 1.5 0.143 208.4
27 170 1.5 0.191 241.5

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 117

3. FUZZY RULE BASED MODEL

Fuzzy logic is based on imprecision, specifically, the way people make decisions on
imprecise and non-numerical information. Fuzzy logic mimics the way humans make
decisions using linguistic reasoning. Fuzzy logic modeling is based on mathematical theory
combining multi-valued logic, probability theory and Artificial Intelligence methods and can
be used to tackle complex problems. Fuzzy modeling is based on fuzzy set theory and the
linguistic statements expressed mathematically which corresponds to the analysis of a human
expert. Fuzzy systems base their decisions on inputs and outputs in the form of linguistic
variables. The variables are tested with IF-THEN rules, which produce one or more responses
depending on which rules are asserted. The response of each rule is weighed according to the
degree of membership of its inputs and the centroid of the responses is calculated to generate
the appropriate output [15 -17].

3.1. Fuzzy Set Theory

The fuzzy set theory, introduced by Zadeh in 1965 in his seminar paper entitled ‘Fuzzy
Sets’ to deal with domains of pattern recognition, communication of information and
abstraction [15]. A fuzzy set as the name implies, is a set without a crisp boundary, that is, the
transition from “belong to a set” to “not belong to a set” is gradual and this smooth transition
is characterized by membership functions that give fuzzy sets flexibility in modeling
commonly used linguistic expressions such as “ the water is hot ” or the “ the temperature is
high ” [18].
If X is a collection of objects denoted generally by x, then a fuzzy set A in X is defined as
a set of ordered pairs:
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

A = {( x, μ A ( x )) | x ∈ X }, (1)

where μ A (x) is called the membership function for the fuzzy set A. This function maps
each element of X to a membership grade or value between 0 and 1 and ( x, μ A ( x )) is a
singleton. Usually X is referred to as the universe of discourse or simply the universe and it
may consist of discrete objects or continuous space [18]. Another common way of
representing a fuzzy set is

A= U μ A ( xi ) / xi (2)
xi ∈ X

Here the fuzzy set A is the collection or union of all μ A ( x i ) / x i .

If set A is a crisp set, it is determined by the membership function that precisely identifies
the boundaries of the set (a,b) on the universe of discourse. If it is a fuzzy set, it is determined
by the membership function that shows the distribution of degrees of relevance across the
universe of discourse [19].

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 118 K. Palanikumar and J. Paulo Davim

Let X denote a universal set. Then the membership function μ A by which a fuzzy set A
is usually defined has the form:

μ A : X → [0,1] (3)
For example, we can define a possible membership function for the fuzzy set of real
numbers close to 0 as follows.

1
μ A ( x) = (4)
1 + 10 x 2

Given a crisp universal set X, let P ( X ) denote the set of all fuzzy subsets of X and let

P k ( X ) be defined recursively by the equation

P k ( X ) = P ( P k −1 ( X )) (5)

for all integers k ≥ 2 . Then fuzzy sets of level k are formally defined by membership
functions of the form

μ A : P k −1 ( X ) → [0,1] (6)

A fuzzy set A is defined on universal set that is finite and countable is written as

n
A = ∑ μ i / xi
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

(7)
i =1
A = ∫ μ A ( x) / x (8)
X

3.2. The Fuzzy Inference System

The fuzzy inference system is a popular computing framework based on the concepts of
fuzzy set theory, fuzzy IF – THEN rules, and fuzzy reasoning. It has been used in a wide
variety of fields, such as automatic control, data classification, decision analysis, expert
system, and time series prediction, robotics and pattern recognition. It is a knowledge based
system that converts linguistic variables into its knowledgebase using fuzzy rules and
membership functions. Because of its multidisciplinary nature, the fuzzy inference system is
known by numerous other names, such as fuzzy – rule based system, fuzzy expert system,
fuzzy model, fuzzy associative memory, fuzzy logic controller and simply fuzzy system. The
basic structure of a fuzzy inference system consists of three conceptual components: a rule
base, which contains a selection of fuzzy rules; a data base (or dictionary), which defines the
membership functions used in the fuzzy rules; and a reasoning mechanism, which performs
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 119

the inference procedure upon rules and given facts to derive a reasonable output or conclusion
[14-20]. The concept of fuzzy reasoning for three-input-one-output fuzzy logic unit is
described as follows. The fuzzy rule base consists of a group of IF- THEN statements with
three inputs, x1, x2 and x 3 and one output y, i.e.

Rule 1: if x1 is A1 and x2 is B1 and x3 is C1 then y is D1 else

Rule 2: if x1 is A2 and x2 is B2 and x3 is C2 then y is D2 else
……………………………………………………………..
……………………………………………………………..
Rule n: if x1 is An and x2 is Bn and x3 is Cn then y is Dn. (9)

Ai, Bi, Ci and Di are fuzzy subsets defined by the corresponding membership functions,
i.e., μ Ai , μ Bi μ Ci and μ Di . 27 fuzzy rules were developed. By taking the max-min
compositional operation, the fuzzy reasoning of these rules yields a fuzzy output. Suppose
that x1, x2 and x3 are the three input values of the fuzzy logic unit, the membership function of
the output of fuzzy reasoning can be expressed as [20]:

μDo ( y) = [μ A1 (x1 ) ∧ μB1 (x2 ) ∧ μC1 (x3 ) ∧ μD1 ( y) ∨...μ An (x1 ) ∧ μBn (x2 ) ∧ μCn (x3 ) ∧ μDn ( y)]
(10)

where ∧ is the minimum operation and ∨ is the maximum operation. Finally, a

defuzzification method is used to transform the fuzzy output into a non-fuzzy value y0 . The
membership functions can be of different forms like triangular, trapezoidal, Gaussian,
sigmoid, etc. The triangular shaped membership function for input is specified by three
parameters {a, b, c} as follows:
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

⎧0, x ≤ a.
⎪ x−a a ≤ x ≤ b.
⎪ ,
⎪b − a
triangle (x; a, b, c) = ⎨ (11)
⎪c − x , b ≤ x ≤ c.
⎪c−b
⎪
⎩ 0, c ≤ x.

By using min and max, an alternate expression for the proceeding equation is

⎛ ⎛ x−a c− x⎞ ⎞
trimf ( x; a, b, c) = max⎜⎜ min⎜ , ⎟,0 ⎟⎟ (12)
⎝ ⎝b−a c−b⎠ ⎠

where a, b, c stands for the triangular fuzzy triplet and it determines the x coordinates of the
three corners of the underlying triangular membership function. For example, the functions
considered for cutting speed is LOW, MEDIUM and HIGH as shown in Figure 4. Similarly
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 120 K. Palanikumar and J. Paulo Davim

for other input parameters, the membership functions have been fixed and are shown in
Figures 5 and 6. The output responses of the fuzzy process can be viewed only in fuzzy
values and they have to be defuzzified. The membership function used for the output response
cutting force is presented in Figure 7.

LOW MEDIUM HIGH

1

0.8
Degree of membership

0.6

0.4

0.2

90 100 110 120 130 140 150 160 170

Cutting speed, m/min

Figure 4. Membership function for cutting speed.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

LOW MEDIUM HIGH

1

0.8
Degree of membership

0.6

0.4

0.2

0.1 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19
Feed, mm/rev.

Figure 5. Membership function for feed.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 121

LOW MEDIUM HIGH

1

0.8

Degree of membership
0.6

0.4

0.2

0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
Depth of cut, mm

Figure 6. Membership function for depth of cut.

LOWEST LOW MED HIGH HIGHEST

1
LOWER LOW-MED HI-MED HIGHER

0.8
Degree of membership
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

0.6

0.4

0.2

180 190 200 210 220 230 240 250

Cutting force, N

Figure 7. Membership function for output response cutting force in machining GFRP composites.

The linguistic forms used for the input parameters and output response cutting force in
machining of GFRP composites with respect to experiment number is presented in Table 4.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 122 K. Palanikumar and J. Paulo Davim

Table 4. Linguistic form used for cutting parameters and output response cutting force

Cutting speed (V), Depth of cut (d), Feed rate (f), Cutting force (Fc),
S.No m/min mm mm/rev. N
1 LOW LOW LOW LOWEST
2 LOW LOW MEDIUM LOW-MED
3 LOW LOW HIGH HIGH
4 LOW MEDIUM LOW LOW
5 LOW MEDIUM MEDIUM MEDIUM
6 LOW MEDIUM HIGH HIGHER
7 LOW HIGH LOW MEDIUM
8 LOW HIGH MEDIUM MEDIUM
9 LOW HIGH HIGH HIGHEST
10 MEDIUM LOW LOW LOWEST
11 MEDIUM LOW MEDIUM LOW
12 MEDIUM LOW HIGH HIGH
13 MEDIUM MEDIUM LOW LOWER
14 MEDIUM MEDIUM MEDIUM LOW-MED
15 MEDIUM MEDIUM HIGH HIGHER
16 MEDIUM HIGH LOW LOWEST
17 MEDIUM HIGH MEDIUM MEDIUM
18 MEDIUM HIGH HIGH HIGHER
19 HIGH LOW LOW LOWEST
20 HIGH LOW MEDIUM LOW
21 HIGH LOW HIGH HIGH-MED
22 HIGH MEDIUM LOW LOWEST
23 HIGH MEDIUM MEDIUM LOW-MED
24 HIGH MEDIUM HIGH HIGHER
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

25 HIGH HIGH LOW LOW

26 HIGH HIGH MEDIUM MEDIUM
27 HIGH HIGH HIGH HIGHER

In the present study, the centroid defuzzification method has been selected, as it produces
the centre of area of the possibility distribution of the inferenced output and is a more
frequently used defuzzification method calculating the centroid of the area under the
membership function

y0 =
∑ yμ D ( y) 0
(13)
∑ μ D ( y) 0

The non-fuzzy value y0 gives the output value in numerical form. For example the value
of cutting force at a cutting condition of cutting speed of 126 m/min, depth of cut of 1mm and
feed of 0.143 mm/rev., is obtained as 201 N. Matlab® fuzzy logic tool box is used for this
calculation and calculated result is shown in Figure 8. In the above example single rule is

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 123

applied. Also for different input values multiple rules can be applied. In the example
Mamdani max-min approach is used as an inference engine.

Figure 8. Cutting force observed using fuzzy logic.

4. RESULTS AND DISCUSSIONS

The study on cutting force in turning of GFRP composites is one of the important areas of
research. To maintain the proper surface and reducing the power requirement for cutting
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

minimization of cutting force is required. The minimization of cutting force is the serious task
because fibre reinforced composite materials is made of two different materials like the hard
fibers and soft matrix materials. The property of the materials may not be uniform throughout.
Machining of glass fiber reinforced composite materials poses big problems for the
manufacturing engineers. Fibre pull-out, delamination, matrix crazing, high tool wear are
some of the problems encountered while machining GFRPs. The cutting parameters such as
cutting speed, depth of cut and feed are playing a vital role in deciding the cutting force in
turning of GFRP composites. Also the cutting force on the machining of composites depends
on material properties, orientation of fibers in the matrix, type of weave, bond strength, etc.
For analyzing the cutting force in turning of GFRP composites, mathematical relations
are required. In this investigation, Fuzzy rule based modeling technique is used for the
modeling of machining parameters in machining of GFRP composites with respect to the
cutting force. Fuzzy logic models are one of the important tools in the artificial intelligence
and can be effectively used for the modeling of manufacturing processes such as turning.
Figure 9 shows the correlation between the cutting force observed during machining and
fuzzy logic model. The adequacy of the developed model is verified through R-Sq (R2) value,
which is presented in the figure. The quantity R-Sq is called as coefficient of determination
and is used to judge the adequacy of the models developed. In the present case the coefficient

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 124 K. Palanikumar and J. Paulo Davim

of determination is 0.9848 which shows high correlation that exists between the fuzzy logic
model developed and experimental results and hence fuzzy logic can be effectively used for
the prediction of cutting force in turning of GFRP composites. Further the comparison of
results obtained from the experiment and fuzzy model is plotted in Figure 10, in which all the
points are coincide with each other and hence fuzzy logic can be effectively utilized for the
prediction of cutting force in machining of GFRP composites.

300
280
Predicted cutting force, N

260
R2 = 0.9848
240
220
200
180
160
140
120
100
100 150 200 250 300

Experimental cutting force, N

Figure 9. Correlation graph for cutting force.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

260
240
220
Cutting force, N

200
180
160
Experimental results
140
Predicted results by
120 Fuzzy logic model

100
0 10 20 30
Experiment number

Figure 10. Comparison of experimental result and fuzzy output for cutting force.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 125

The effect of cutting parameters such as cutting speed, depth of cut and feed can be
analyzed through three-dimensional surface plots. Surface plots are helping to visualize the
response surface. This plot represents the functional relationship between the response and
the experimental factors. The response surface plots also help to visualize how the response
reacts to changes in the experimental factors. The surface plot shows only two factors at a
time, and the extra factor should be kept at a constant level.
Figure 11 shows the surface plot of cutting force with variables cutting speed and depth
of cut while keeping the feed at constant middle level. The figure indicates that the increase
of cutting speed reduce the cutting force in turning of GFRP composites. The reason being, at
high cutting speed, the removal of glass fibers in the GFRP composites is easy and it
produces low cutting force. The increase in depth of cut increases the cutting force in turning
of GFRP composites. The increase in cutting force increases the load on the cutting tool and it
leads to the increased cutting force.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 11. Surface plot of cutting speed and depth of cut.

Figure 12. Surface plot of depth of cut and feed.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 126 K. Palanikumar and J. Paulo Davim

Figure 13. Surface plot of cutting speed and feed.

The surface plot of varying the variables depth of cut and feed is depicted in Figure 12.
The figure indicates that the increase of feed increases the cutting force abruptly. The reason
being the increase of feed rate increases the tooth load on the cutting tool and also produces
chatter or vibration in cutting of GFRP composites which in turn increases the cutting force.
The effect of cutting speed and feed on cutting force is presented in Figure 13. The figures
indicated that the increase of feed and depth of cut increases the cutting force in turning of
GFRP composites. The magnitude of cutting force variation is high while increasing the feed
compared to the depth of cut. From the graphs, one can infer that high cutting speed, low
depth of cut and low feed are preferred for cutting the GFRP composites.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Fuzzy output
300
Experimental result
250 Verification test
Cutting force, N

200

150

100

50

0
1 2 3 4
Trial number

Figure 14. Verification test result for cutting force.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Application of Fuzzy Logic in Manufacturing 127

For verification of fuzzy logic model and experimental results, verification tests were
carried out. The verification tests were carried out as per the corresponding experimental
numbers 1, 7, 15 and 26. The verification test results for cutting force in turning are presented
in Figure 14. From the figures it can be asserted that the variation between experimental
results and the model are within the limit and are very close to each other and hence fuzzy
logic model can be effectively used for the prediction of cutting force parameters in turning of
GFRP composites.

CONCLUSIONS
Experiments are conducted for modeling the parameters which affect the cutting force in
turning of GFRP composites. Taguchi’s experimental design has been used for
experimentation. Fuzzy rule based modeling technique is employed for modeling the
machining parameters. Based on the experimental results and data analysis, the following
conclusions can be drawn.

1. Fuzzy rule based modeling technique is used for predicting the cutting force in
turning of GFRP composites.
2. The model for cutting force is well correlated with experimental results. The results
indicated that artificial intelligence assisted tool can be well suited for the prediction
of cutting force in turning GFRP composites.
3. In machining of GFRP composites, the increase in cutting speed reduces the cutting
force, whereas the increase of feed increases the cutting force.
4. In machining of GFRP composites feed is the main parameter which affects the
cutting force in machining of GFRP composites.
5. The confirmation experiment reveals that the developed model can be effectively
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

used for predicting the cutting force in turning of GFRP composites.

6. The artificial intelligence tool used for the manufacturing process is simple and can
be used as an on-line monitoring tool, if proper equipments are used.

REFERENCES
[1] Zadeh L.A. (1965), Fuzzy sets, Information and control, 8:338-353.
[2] Hasmi K., Graham I.D and Mills B, (2000), Fuzzy logic based data selection for the
drilling process, Journal of Materials Processing Technology 108: 55-61.
[3] Suleyman Yaldiz, Faruk Unsacar and Haci Saglam (2006), Comparison of experimental
results obtained by designed dynamometer to fuzzy model for predicting cutting forces
in turning, Materials and Design, 27: 1139-1147.
[4] Yue Jiao, Shuting Lei, Pei Z.J, and Lee E.S, (2004), Fuzzy adaptive networks in
machining process modeling: surface roughness prediction for turning operations,
International Journal of Machine Tools and Manufacture, 44:1643-1651.
[5] Fang, X.D (1995). Expert system support fuzzy diagnosis of finishturning process
states. International Journal of Machine Tools and Manufacture 35 (6): 913–924.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 128 K. Palanikumar and J. Paulo Davim

[6] Palanikumar, K.; Karunamoorthy, L.; Karthikeyan, R.; Latha, B (2006). Optimization
of machining parameters in turning GFRP composites using a carbide (K10) tool based
on the Taguchi method with fuzzy logics. Metals and Materials 12/6: 483–491.
[7] Santhanakrishnan, G., Krishnamoorthy, R. and Malhotra, S. K. (1989). High Speed
Steel Tool Wear Studies in Machining of Glass Fiber-reinforced Plastics, Wear, 132:
327–336.
[8] Evestine, G. C. and Rogers, T. G. (1971). A Theory of Machining of Fiber-reinforced
Materials, J. Composite Materials, 5: 94–106.
[9] Arno kopf, StefenFeistritzer, Klaus Udier (2006). Diamond coated cutting tools for
machining of non-ferrous metals and fiber reinforced polymers, International Journal
of Refreactory metals and hard materials, 24: 354-359.
[10] Paulo Davim J. Pedro Reis. Conceicao A.C (2004/). Experimental study of drilling
glass fiber reinforced plastics (GFRP) manufactured by hand lay up, Computer Science
and Technology, 64(2): 289 – 297.
[11] Paulo Davim J. Pedro Reis. Conceicao Antonio C, (2004). A study on milling of glass
fiber reinforced plastics manufactured by hand-lay up using statistical analysis
(ANOVA)’, Composite Structures, 64: 493-500.
[12] Sang-Ook An. Eun-Sang Lee Sang-Lai Noh, (1997), A study on the cutting
characteristics of glass fibre reinforced plastics with respect to toll materials and
geometries’ Journal of Materials Processing Technology, 68: 60-67.
[13] Ross, P. J. (1989). Orthogonal Array Selection and Utilization, Taguchi Techniques for
Quality Engineering, McGraw-Hill, New York.
[14] Palanikumar, K, (2006). Cutting parameters optimization for surface roughness in
machining of GFRP composites using Taguchi’s method, Journal of Reinforced plastics
and composites, 25(16): 1739-1751.
[15] Nandi A.K. and Pratihar D.K, (2004). An expert system based on FBFN using a GA to
predict surface finish in ultra-precision turning”, Journal of Materials Processing
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Technology, 155-157: 1150-1156.

[16] Oguzhan Yilmaz, Omer Eyercioglu, Nabil N.Z. Gindy (2006), A user-friendly fuzzy-
based system for the selection of elsctro discharge machining process parameters,
Journal of Materials Processing Technology , 172: 363-371.
[17] Palanikumar. K (2009), Surface roughness model for machining glass fiber reinforced
plastics by PCD tool using fuzzy logics, Journal of Reinforced Plastics and
Composites, Jun 2009; DOI. 0731684408092009v1.
[18] Jang, J.S.r. Sun C.T, Mizutani E., Neuro – Fuzzy and soft computing – A
Computational approach to learning and machine intelligence, Pearson Education,
2005.
[19] Riza C. Berkan, Sheldon L. .Trubatch , Fuzzy systems design principles –Building
fuzzy IF-THEN Rule bases, IEEE Press, Standard Publishers Distributors, 2000.
[20] Lin J.L., Wang K.S., Yan B.H., Tarng Y.S. (2000), Optimization of the electrical
discharge machining process based on the Taguchi method with fuzzy logic, J. Mater
Process. Technol. 102; 48-55.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 8

FLANK WEAR DETECTION WITH AE SIGNAL AND

FNN DURING TURNING OF AL/15 VOL%SIC-MMC

Alakesh Manna∗
Department of Mechanical Engineering, Punjab Engineering College,
Deemed University, Chandigarh- 160012, INDIA

ABSTRACT
This paper describe predictive machining approach with fuzzy neural network (FNN)
modeling of the cutting tool flank wear in order to estimate the performance of CCGX
09 T3 04 Al-H 10 insert during turning of Al/15 vol%SiC-MMC. In the present work a
new approach for cutting tool wear detection with cutting conditions estimated wear
through acoustic emission (AE) signal is presented. The measured tool wear and
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

estimated tool wear by conditions monitoring and detected signals are compared and
graphically analyzed. Investigated results prove that the new method of FNN is reliable
and appropriate to control and monitor the cutting tool wear.

1. INTRODUCTION
In modern high-tech industries, manufacturing of various complex components
machining processes plays an important role which involving metal cutting from a significant
part of most manufacturing task. To ensure the require degree of reliability in the final
product, there is a need to maintain high degree of quality control. Automation in the metal
cutting process is reliable methods of monitoring the cutting tool wear and tool failure.
Sensors and automatic control can monitor the variation of cutting parameters during metal
cutting. To fulfill the object several sensing technique have been developed in the recent year
for estimating and automatic control the variation of cutting parameter during metal cutting.
Some of these techniques have provided quiet useful under laboratory conditions; few of
them have been discussed in a review article (Tlusty and Andrews, 1983). One of the

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 130 Alakesh Manna

important techniques is the acoustic emission can be used to measure the cutting tool wear.
Very beginning of tool failure is also indicated by some of these parameters as sudden
changes in their characteristics the sensor is expected to be highly sensitive to the parameters,
it is to monitor and it should be at the same time sensitive to the others parameters of the
metal cutting process as well as to any varying conditions not under process control. It is
essential that a sensor is used for monitoring the progressive tool wear in machining, it should
have minimum sensitivity to process parameter such as cutting speed, feed and depth of cut,
On the other hand sensor whose output is to be used for monitoring cutting speed and also an
input in automatic control of the machining process should not be affected by the condition of
tool wear. Acoustic emission (AE) is an important method can meet the ideal requirement to
monitor the cutting tool wear during metal cutting.
The frequency distribution pattern of force signals is influenced by cutting conditions
(Rao and Shin, 1999). Not only frequency distribution pattern of force signal but also AErms,
the total energy of forces are influenced by the cutting conditions, tool wear, chip fracture,
cutting edge deterioration (Chung Choo and Saini 2000; Usui and Hirata, 1978, Usui et. al.
1978). The AE can be used effectively to detect the tool condition as the frequency of the AE
signal is much higher than that of machine vibrations and other noises which eliminates the
interference in cutting (Iwata and Moriwaki, 1997; Dornfeld, 1990). In turning operation,
with in the range of input variables e.g. cutting speed 20-35 m/min, feed rate 0.075 – 0.125
mm/rev and depth of cut 0.4-0.8 mm the neural network comes ahead of the design of
experiment in nearness of the predictions of the experimental values of flank wear
(Chaudhury and Bartarya, 2003). Elanayar and Shin (1995) proposed a model, which
approximates flank and crater wear propagation and their effects on cutting force by using
radial basis function neural networks. The generic approximation capabilities of radial basis
function neural networks are used to identify a model and a state estimator is designed based
on this identified model. A wide range of tool monitoring techniques utilizing neural
networks has been reviewed by Dimla et al. (1997). They concluded that neural networks are
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

adequate for tool condition monitoring. They also pointed out the confusion in the
interpretation of TCM techniques in literature as on-line or off-line systems. Ghasempoor et.
al. (1999) proposed a tool wear classification and continuous monitoring neural network
system for turning by employing recurrent neural network design. In the study of Li et al.
(1999), neural network models have also been integrated with analytical models such as
Oxley’s theory to form a hybrid machining model for the prediction of tool wear and work-
piece surface roughness. Neural networks are used to predict difficult-to-model machining
characteristic factors. Liu and Altintas (1999) derived an expression to calculate flank wear in
terms of cutting force ratio and other machining parameters. The calculated flank wear, force
ratio, feed rate and cutting speed are used as an input to a neural network to predict the flank
wear in the next step. Chaudhury, S.K; et. al. (1999) used optoelectronic sensor and back
propagation of the neural network to predict the tool wear. Very recently an experimental
investigation was carried out to predict the ceramic tool wear during machining of D2 AISI
steel using neural network by Quiza, Ramon; et. al. (2008) and they concluded that the neural
network model has shown better capability to make accurate predictions of tool wear under
the conditioned studied.

∗
Corresponding author, mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Flank Wear Detection with AE Signal and FNN During Turning 131

The different wear on a cutting tool may occur during turning and it is really be very
difficult to avoid in practice. However, the flank wear on the cutting tool during turning has
the greatest influence on the quality of the machined surface and work-piece dimensional
accuracy. Keeping in view, the present research strives to monitor the cutting tool flank wear
using AE sensor and based on Fuzzy Neural Net work (FNN) classification.

2. EXPERIMENTAL CONDITION
Work-piece material: Al/15vol%SiC-MMC is selected as work-piece materials for
experimental investigation. Table-1 shows the chemical composition of work-piece used for
experimentation. Table 2 represents the parameters and their levels considered for
experimental investigation.

Table 1. Chemical composition of Al/SiC-MMC used for experimentation

Types of Density; %SiC %Si %Mg %Fe %Cu %Mn %Zn %Ti Al
MMC APS
LM6Mg 2.67g-cm-3 15.0 12.0 0.50 0.12 0.17 0.11 0.10 0.10 Rem.
15SiC 23μm

Table 2. Physical and Mechanical Properties of Al / SiC – MMC

Material Density Tensile Yield % 0f Hardness Modulus of

(gm/cm3) Strength Strength Elongation (BHN) Elasticity
(Mpa) (Mpa ) (Gpa)
LM6 Mg 15 2.67 140 120 1 - 1.5 115 78
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

SiCP

Machine Tool; Cutting Tools and Measuring Equipment Used

(i) Machine Tool: Combination Turret Lathe(motor h.p. 7.5, Word Co. England)

Table 3. Details of cutting tool used for experimentation

Cutting Cutting tool Tool material Rake Clearance Nose Cutting Cutting
tool specification and grade angle angle radius edge fluid
angle used
T-Max-U CCGX 09 T3 Uncoated
Positive 04 Al-H 10 tungsten 50 70 0.4 800 Not used
rhombic carbide (WC) mm
insert (HW-K10)

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 132 Alakesh Manna

Table 4. Cutting parameters and their levels

Sl. No. Machining parameters Level

1 3
1 A: Cutting speed, m/min. 40 100 160
2 B: Feed, mm/rev. 0.16 0.32 0.48
3 C: Depth of cut, mm. 0.50 0.75 1.00
(ii) Tool holders Specification: SCLCR 20 20 K 09
(iii) Mitutoyo Shop Microscope with 30X magnification and resolution 1 μm

2.1. Experimental Setup

AE sensor is mounted on the tool shank and the signal output (Irms and Vrms, Sensor
output) is amplified and converted into the digital form by means of an A/D converter
connected in between the PC and Low pass filter. Figure 1 shows the scheme of the
experimental setup utilized for experimentation.
As the measure of the AE signal the RMS value of the AE is directly related to the
average power dissipation involved in the process, the AE count clearly relates to the
occurrence of the discreet events like microscopic slips and fractures in the metals during
metal cutting. The AE sensing clearly reveals the information about the hardness and the
dislocation densities in the material and therefore has potential to be useful in the
identification of individual types of tool wear.
The signal from AE the sensor is first filtered and amplified in a sufficiently wide band to
encompass most of the AE spectrum and remove any unwanted noise, especially at low
frequency. Assume that AE signal arises mainly from primary contact region between the tool
and the chip formed as well as due to the flank wear of the cutting tool and analysis the
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

estimated data.

Turned face

Work-piece

CCGX 09 T3 04 Al-H 10, T-Max-U

SCLCR 20 20 K 09 Uncoated WC (HW-K 10)

Tool Holder

Monitor

A/D converter PC

Low pass filter

AE Sensor

Pre-amplifier Band pass filter

Figure 1. Scheme of the experimental setup.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Flank Wear Detection with AE Signal and FNN During Turning 133

3. TOOL WEAR DETECTION WITH FNN MODEL

The flow diagram of the user interface is shown in figure 2. The basic function of the
developed user interface are promoting the user to enter the necessary information such as
turning parameters setting conditions and tool wear size and as well as display the estimated
wear to the user. Considering 0.30 mm flank wear is tool life (Bhattacharya, 1984) of the
cutting tool insert, flank wear from 0.00 to 0.35 mm was categorized in to seven categories
with 0.05 mm increment.

Start

Preparing weight
for Neural Network

Input Turning
Parameter as required

Spindle and feed

current signal

Estimation of flank
wear
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Showing estimated
results

Yes Tools wear

monitoring

End

Figure 2. Flow chart for on-line FNN model.

In this study, the flank wear width was categories into Category-A, Category-B,
Category-C, Category-D, Category-E, Category-F and Category-G according to the width of
the cutting tool flank wear during turning of Al/SiC-MMC using CCGX 09 T3 04 Al-H 10
insert and by utilizing SCLCR 20 20 K 09 cutting tool holder. Table-5 shows the tool wear
and their category.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 134 Alakesh Manna

Table 5. Cutting tool flank wear categorization

Category A B C D E F G
of Flank
wear width
Flank wear 00-0.05 0.05-0.10 0.10-0.15 0.15-0.20 0.20-0.25 0.25-0.30 0.30-0.35
width(mm)

Considering 0.30 mm flank wear is tool life an attempt has been made to develop the
model for the current signal. The effect of the cutting parameters e.g. cutting speed (V,
m/min), feed rate (f, mm/rev) and depth of cut (d, mm) on the spindle and feed current signal
can be expressed in the form of the exponents of the equations as follows

I spindle = C1.Vn1.fn2.dn3 Eqn.1

I feed = C2.Vm1.fm2.dm3 Eqn.2

where, I spindle = spindle current (IS, A), I feed = feed current (IF, A), C1 and C2 are the constant
depends on the properties of the work materials and cutting tool and as well as insert
geometry; n1,n2,n3 and m1,m2,m3 are the exponential for spindle peak current and feed current
respectively. From the above predicted exponents of the equations, it is observed that a unit
increase in either the cutting speed, feed rate or depth of cut would have exponential effects
of n1, n2, n3 and m1, m2, m3 on the average spindle and feed current signal respectively.

The logarithmic values of the spindle current (IS, A ) can be expressed as follows

IS1 n10 n11 n12 n13

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

IS2 n20 n21 n22 n23 1

IS3 n30 n31 n32 n33 log V
IS4 = n40 n41 n42 n43 x log f
IS5 n50 n51 n52 n53 log d
IS6 n60 n61 n62 n63
IS7 n70 n71 n72 n73
Eqn. 3

The logarithmic values of the feed current (IF , A ) can be expressed as follows

IF1 m10 m11 m12 m13

IF2 m20 m21 m22 m23 1
IF3 m30 m31 m32 m33 log V
IF4 = m40 m41 m42 m43 x log f
IF5 m50 m51 m52 m53 log d
IF6 m60 m61 m62 m63
IF7 m70 m71 m72 m73
Eqn. 4
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Flank Wear Detection with AE Signal and FNN During Turning 135

where, IS1, IS2 ….. . IS7 and IF1, IF2. . . . . .IF7 are the algorithmic values of the spindle peak
current and feed current respectively. The input variables for this analysis are 1, log V, log f
and log d for turning parameters cutting speed (V, m/min), feed rate (f, mm/rev) and depth of
cut (d, mm) respectively. Output variables are IS1, IS2 ….. . IS7 and IF1, IF2. . . . . .IF7 for spindle
peak current and feed current respectively. Utilizing the above input-output relational matrix
the weights of the neural network can be calculated by regression analysis. The matrix
equation 3 between the spindle current (IS, A) and turning parameters, and matrix equation 4
between feed current (IF, A) and turning parameters can be expressed by the neural network
topology is shown in figure 3.
The FNN model can be expressed in four separate heads such as fuzzy logic based tool
wear classification, input normalization, neural network base tool wear estimation and fuzzy
logic based adjustment of tool wear. As earlier mentioned that the spindle current signal and
feed current signal models equation 3 and 4 at the various cutting tool wear can be explained
with the help of neural network structure is shown in figure 3. The measure spindle current
and feed current are considered as real value and estimated spindle currents for IS1, IS2….IS7
and feed current for IF1, IF2….IF7 considered as mid value of a bunch of various tool wear
classifications. The measure values (i.e. real values) are compared with the estimated values
by utilizing fuzzy logic based classification, and the membership degree of the tool wear can
be calculated by utilizing the fuzzy curve of membership degree is shown in figure 4. The
same method is utilized for both IS and IF. Here, three input variables each of which is
classified into seven fuzzy sets and an output variables is also classified into seven fuzzy sets
and based on the experimental work, forty-nine rules are set. These rules are classified into
seven groups corresponding to seven categories of flank wear.

O u tp u t L a y e r
IS 1
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

IS 2
IS 3
IS 4 S p in d le P e a k
In p u t L a y e r C u rre n t
1 IS 5
IS 6
lo g V IS 7

lo g f
IF1
lo g d
IF2
IF3
IF4 F e e d c u rre n t
o u tp u t
IF5
IF6
IF7

Figure 3. Spindle current signal and feed current signal with Neural Network.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 136 Alakesh Manna

1 A B C D E F G
Membership
degree

0.0 IS1 IS2 IS3 IS4 IS5 IS6 IS7 Current, A

Figure 4. Fuzzy logic based curve for classify the membership degree of the tool wear.

4. RESULTS AND DISCUSSIONS

AE signal generated during metal cutting mostly originates in the plastic deformation
zones on the work piece and related to the associated stress relaxation, in which stress energy
in the materials gets dissipated as thermal energy and micro-elastic pulses. A large number of
pulses are emitted by sources randomly distributed over the plastic deformation zone. The AE
signal detected from the process appears as a continuous noise. Figure 5 depicts the schematic
appearance of the generated AE record. Here, 1 to 21 represents the time (s) from 0 s to 1.5 s
with equal increment of 0.05 s. During experiment it was found that Vrms increased
monotonically with increase the individual setting value of all three turning parameters such
as cutting speed, feed rate and depth of cut. The parameter AE-mode measures the value of
the voltage amplitude corresponding to the peak in the amplitude distribution of the digitized
AE signal envelope.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 5. AE signal at cutting speed (Vc) = 100 m/min, feed rate (f) = 0.32 mm/rev and depth of cut (d)
= 0.75 mm.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Flank Wear Detection with AE Signal and FNN During Turning 137

Time (s)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
0.4

0.3

0.2

0.1
AE (V)

-0.1

-0.2

-0.3

-0.4 AE Signal Average

Poly. (AE Signal) 5 per. Mov. Avg. (AE Signal)
-0.5

Figure 6. shows Polynomial trend and Moving Average trend of AE signal.

It is important to extract the feature of the signals during turning to ensure the reliability
of the tool monitoring system. In monitoring the cutting tool flank wear AE signal monitored
the complicated information during turning.
Figure 6 shows the detail AE signal generated during turning through various graphical
representations such as average, polynomial trend and moving average trend of the AE signal.
Polynomial equation given below represents the polynomial trend of the cutting tool wear AE
signal during turning.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

y = 18.208x6 - 73.261x5 + 108.62x4 - 72.722x3 + 21.302x2 - 2.0708x + 0.0293 ---- Eqn.5

where, ‘y’ represents the cutting tool wear AE(V) signal, ‘x’ represents the time in second.
Figure 6 also shows the 5 % moving average trend of the generated AE signal during turning.
Average AE signal generated is slightly above the zero value and as shown in the figure 6.
The behavior of the AE signal due to the appearance of the cutting tool wear can be predicted
from the above polynomial relation. The developed equation 5 also can be utilized to predict
the AE signal at any instant by putting the time (s) of continuous cutting. From the
polynomial trend of the AE signal and mathematical relation it is clear that AE signal
increases rapidly after 1.2 s of the turning operation was started. Hence, it is concluded that at
the very beginning the AE signal was very small as the cutting tool was fresh and sharp edge,
but after certain time (s) the AE signal increases as the cutting tool wear appears and
increases with time phenomenon. During experiment it was observed that the generated AE
signal increases with increase of cutting time.
According to the Taguchi method, L27 (313) orthogonal array, turning parameters and their
levels (table 4) were utilized and 27 x 3 = 81 experiments were performed i.e. repeating the
each set of experiment three times and taking average flank wear to obtain the real values of
cutting tool flank wear. The membership degree of the cutting tool wear under different tool

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 138 Alakesh Manna

wear classification is estimated and fused using fuzzy logic and plotted in the graph is shown
in figure-7. The figure also shows the actual and estimated flank wear width (mm) against
each of the tool wear classification. From the figures 5, 6 and 7, it is clear that the above
described method is suitable to estimate the cutting tool wear during turning.

0.4
TW1 Estimated
Average Flank Wear width (mm)

0.35 TW1 Actual

TW2 Estimated
0.3
TW2 Actual
0.25 TW3 Estimated
TW3 Actual
0.2

0.15

0.1

0.05

0
A = 00-0.05 B = 0.05 - 0.10 C = 0.10-0.15 D = 0.15-0.20 E = 0.20-0.25 F = 0.25-0.30 G = 0.30-0.35

Tool Wear Classification

Figure 7. Cutting tool wear classification with comparison between estimated and actual wear.

CONCLUSIONS
The study on AE during turning has lead to a better understanding of the process and AE
signal parameters such as Vrms ,ISi ( i = 1,2.3-------7) and IFi ( i = 1,2,…..7) is to be used in tool
wear monitoring. The AE signal can be successfully employed to detect the tool failure or
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

breakage during turning of Al/SiC-MMC. AE signal levels are very sensitive to the cutting
tool wear, also strongly depends on the cutting speed and depth of cut. At the beginning the
AE signal was very small as the cutting tool was fresh and sharp edge, but after certain time
(s) the AE signal increases as the cutting tool wear appears and increases with time
phenomenon. The use of artificial intelligence technique such as FNN to offer a lot of
promise, as the method work well even when exact sensor modules are not available.

REFERENCES
Tlusty,J.; Andrews,G.C (1983) “A Critical review of Sensors for Unmanned Machining” Ann
CIRP 32, pp. 563-572.
Rao, B.C; Shin, Y.C.(1999) “A comprehensive dynamic cutting force model for chatter
prediction in turning” Int. Journal of Machine Tools and Manufacture, vol.39,pp.1631-
1654.
Chungehoo, C; Saini,D.(2000) “The total energy and the total entropy of forces signals-new
parameters for monitoring oblique turning operations” Int. Journal of Machine Tools and
Manufacture, vol.40,pp.1879-1897.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Flank Wear Detection with AE Signal and FNN During Turning 139

Usui,E; Hirota,A.(1978) “Analytical prediction of three dimensional cutting process part

2:chip formation and cutting force with conventional single point tool” ASME Journal of
Engineering for Industry, vol.100(2), pp.229-235.
Usi,E; Hirota,A; and Masuko,M (1978) “Analytical prediction of three dimensional cutting
process part-1: basic cutting model and energy approach” ASME Journal of Engineering
for Industry, vol.100(2), pp.222-228.
Iwata,K.; Moriwaki,T(1977) “ An application of acoustic emission to in-process sensing of
tool wear” Annals of the CIRP; vol.26, pp.21-26.
Donfeld,D.A.(1990) “Newral network sensor fusion for tool condition monitoring” Annals of
the CIRP; vol.39, pp.101-105.
Chaudhury,S.K; Bartarya,G.(2003) “Role of temperature and surface fenish in predicting tool
wear using neural network and design of experiments” Int. Journal of Machine Tools and
Manufacture, vol.43,pp.747-753.
Elanayar, S; Shin,Y.C (1995) “Robust tool wear estimation with radial basis function neural
networks”, ASME Journal of Dynamic Systems, Measurement and Control, vol.117,
pp.459–467.
Dimla, D.E; Lister,P.M; Leighton,N, (1997) “Neural network solutions to the tool condition
monitoring problem in metal cutting—a review critical review of methods”, International
Journal of Machine Tools and Manufacture, vol. 39,pp. 1219–1241.
Ghasempoor,A; Jeswiet,J; Moore,T.N, (1999) “Real time implementation of on-line tool
condition monitoring in turning”, International Journal of Machine Tools and
Manufacture, vol. 39 ,pp. 1883–1902.
Li, X.P; Iynkaran,K; Nee,A.Y.C.,(1999) “A hybrid machining simulator based on predictive
machining theory and neural network modeling”, Journal of Material Processing
technology, vol.89/90 , pp. 224–230.
Liu,Q; Altintas,Y, (1999) “On-line monitoring of flank wear in turning with multilayered
feed-forward neural network”, International Journal of Machine Tools and
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Manufacture,vol. 39, pp. 1945–1959.

Chaudhury, S.K.; Jain,V.K; and Rama Rao;Ch.V.V.(1999) “On-line monitoring of tool wear
in turning using a neural network”, International Journal of Machine Tools and
Manufacture, vol. 39, pp. 489–504.
Quiza, R.; Figueira, L.; Davim, J.P. (2008) “ Comparing statistical models and artificial
neural networks on predicting the tool wear in hard machining D2 AISI steel”
International Journal of Advanced Manufacturing Technology, vol.37, No.7-8, pp.641-
648.
Bhattacharya, A.(1984), Metal Cutting: Theory and Practice, Central Book Publisher,
Calcutta, India.
Haykin,S., (1994) Neural network: A Comprehensive Foundation, Macmillian, New York.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 In: Artificial Intelligence in Manufacturing Research ISBN 978-1-60876-214-9
Editor: J. Paulo Davim © 2010 Nova Science Publishers, Inc.

Chapter 9

INTEGRATION OF PRODUCT DEVELOPMENT

PROCESS USING STEP AND PDM

S. Q. Xie∗ and W. L. Chen

Department of Mechanical Engineering, University of Auckland,
20 Symonds Street, Auckland, NEW ZEALAND

ABSTRACT
In today’s highly competitive global economic environment, enterprises solely
replying on traditional technologies find themselves struggle to meet the demands of the
international market that has become more and more dynamical and customized oriented.
Integration plays an important role in enterprise strategy, and its benefits are enormous.
The conventional integration technologies and methods have limitations for supporting
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

information exchange and sharing in various stages of product development. This has led
to problems such as information loss, data format incompatibility, one-way and static
integration. In this chapter, an integrated platform based on the STandard for the
Exchange of Product data (STEP) and Product Data Management (PDM) is proposed to
overcome these problems. This platform is designed as a multi-layer reconfigurable
architecture. Each level is devoted to a specific task and is designed to bring the interface
with the integrated platform from different level of abstraction and functionality. Two
key issues are discussed in detail in the chapter. One issue is the product modeling which
is solved through a proposed Generic Product Model Data (GPMF). The GPMF consists
of an EXPRESS data model (EDM), a STEP-based modeling environment, a ‘five-phase’
modeling method, and three EDM data exchange and sharing methods. The other key
issue is process modeling, which is overcome by a proposed workflow modeling method.
The workflow modeling environment is comprised of one engine, three models and one
application module. Case studies are carried out to validate the proposed integrated
platform, and the results show that the integrated platform is compatible, comprehensive
and flexible, and is able to support product data exchange and sharing in a dynamical and
bidirectional manner.

∗
Corresponding author, E-mail: [email protected]
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 142 S. Q. Xie and W. L. Chen

1. INTRODUCTION
In today’s highly competitive global economic environment, the demand for high quality
products manufactured at lower costs with shorter development times has forced most of the
manufacturing industries to consider various new product design, manufacturing, and
management strategies [1]. One of the strategies is to combine the strength of many
companies for the development of products, especially for high value added customized
products. Normally it has become very a common practice if a product is designed in a place,
while manufactured and assembled in other places. This situation can be characterized by the
following buzzwords: globalization, customer satisfaction, parallelization, agility, virtual
enterprise, total quality management [1]. This strategy requires manufacturing companies,
especially One-of-a-Kind Production (OKP) companies, to work together closely. Special
critical problems for these OKP companies include high customization, ‘Once’ successful
approach, loose or fatter production, and complicated product data and information flow [2-
4]. The exchange and sharing of product and product related information has become more
and more important for them and their partners. Thus, there is an urgent need for better
process management and more integration within more decentralized and modular individual
OKP companies as well as among companies cooperating on collaborative projects.
With the rapid advancement in information technology, computer-aided solutions such as
Computer-Aided Design (CAD), Computer-Aided Process Planning (CAPP), Computer-
Aided Manufacturing (CAM), Enterprise Resource Planning (ERP), have been widely used in
many enterprises. These solutions or systems are, however not necessarily connected in a
seamless way, thus many “islands of automation” are formed in a company that may result in
communication error, rework and duplication. These “islands of automation” need to be
conquered to better suit the collaborative and distributed environment. The fundamental
requirements for a new OKP information system are identified as enterprise integration, open
and dynamical structure, supporting cooperation and collaboration [3, 4]. It is evident that
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

many OKP companies are struggling with integration strategy at different levels other than
the underlying technology. In some cases, where integration does not exist among these
computer-aided solutions, promising product technologies may come to a sudden halt against
these barriers[5]. Benefits of integration to manufacturing companies has been identified[6],
such as enabling a functional unit of an enterprise to communicate easily with other relevant
functional units, accurate data transfer among the manufacturing plant, and/or subcontracting
facilities; and faster response to required changes and increase flexibility towards introduction
of new products.
To thrive in the competitive market, OKP companies must restructure their processes (by
administrative, technical, or support processes), improve the way they manage these
processes, and find a way to make them more integrated. Considerable research effort has
been placed in this area. However, so far, the integration is still limited. This chapter proposes
an integrated platform based on STEP and PDM. The platform serves as a substrate for
integrating software tools and systems for OKP companies such as CAD, CAPP, CAM and
ERP. The goal of the research is the creation of a technological infrastructure for a new, open
integrated platform, which can adapt to continuous changes in OKP product development. A
number of recent developments are discussed. They include the integration based on STEP

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 143

and integration based on PDM. A prototype version of the platform has been successfully
developed and demonstrated with case studies.

2. LITERATURE REVIEW
For manufacturing companies to survive and thrive in the current competitive global
customer-driven market environment, technologies for supporting integration play a
significant role to improve their performances. The potential benefits of the integration have
attracted great research interest in this area.

2.1. Integration Based on STEP

STEP[7], formally known as ISO 10303, is an international standard designed to enable

the exchange of product data between heterogeneous computer systems used throughout a
product life cycle. STEP provides both broadly useful data modeling methods and data
models focused on specific industrial uses. Considerable effort has been placed on STEP-
based integration over the years.
Li, et al developed a feature based parametric product modeling system, which employed
a product model based on the STEP and managed by an object-oriented database[8]. This
system was suitable for applications in a computer integrated manufacturing environment.
Gu, et al developed a STEP-based generic product modeling system, which was designed and
implemented according to the generic resources of STEP[9]. The system can therefore be
used to integrate manufacturing activities, such as process planning and inspection planning
in concurrent engineering environment. Usher et al presented a STEP-based object-oriented
product model based on STEP AP 224[10]. This model was proposed for supporting CAPP
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

analysis. Chin et al presented a STEP-based part information model for process planning
purpose[11]. Their models included a process planning information model and a production
resource information model. Song et al presented a STEP-based die and product integrated
information model (DPIIM), in which integrated resources of STEP were utilized to model
six EXPRESS schemas[12]. These models could support the concurrent development of
stamp and die products. Zha et al presented a product data exchange using PDES/STEP-based
assembly model for the concurrent integrated design and assembly planning[13]. Shaharoun
et al utilized STEP to describe geometric data of a particular plastic product[14]. The
geometrical descriptions of the product were transferred into a CAD system to assist the
design and machining of a suitable mold for the plastic product. Cai et al proposed a method
to build self-defined APs for all kind of machine parts based on the STEP[15]. They
implemented this method to develop two APs for presenting the geometric data model in the
cone gear product for final driver of automobile driving axle system. Zhao et al delineated an
object-oriented feature-based aero-engine blade product modeling system[16]. In this
modeling system, the design platform utilized STEP to standardize data modeling and to
support the information transmission from design platform to analyzing system. Jasnoch et al
developed a collaborative virtual prototyping environment to integrate existing CAD

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 144 S. Q. Xie and W. L. Chen

systems[17]. The underlying product model of this environment was a STEP-based integrated
product model.
It can be concluded that STEP-based modeling method has become the core of
integration to organize product data in the standardized representation, which greatly
enhances the capability of data exchanging and sharing in the integrated manufacturing
environment.

2.2. Integration Based on PDM

Another integration method is based on PDM. PDM systems are designed to control
product-related data and associated workflow processes. As an integration tool connecting
many different areas, PDM manages product data throughout the enterprise, ensuring that the
right information is available to the right person at the right time and in the right format. In
this way, PDM improves communication and cooperation among diverse groups, and forms
the basis for organizations to restructure their product development processes and institute
initiatives such as concurrent engineering and collaborative product development. Research
work has been placed on PDM-based integration during the past few years.
The product data exchange between heterogeneous CAD and PDM systems is a crucial
issue for the integration of product development systems. Oh et al introduced a UML-based
mapping methodology for the product data models to exchange the product structure data
between a CAD system and a PDM system[18]. Dong et al proposed three modes of
integration of CAD/PDM, the three modes are encapsulation, interface, and seamless
integration and discussed the contents and data which were involved in the integration
process[19]. Eynard et al presented a UML approach for the specifications of a PDM system
implementation. The chosen object-oriented approach and the used UML diagrams for the
modeling and integration of product, process, and resource data was detailed for a turboprop
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

aircraft project[20]. The Bill of Material (BOM) was put forward to solve the information
integration between the CAD system and PDM system by Qiu et al[21]. They studied the
features of the information flow integration between the 2-and 3-dimensional CAD system
and the PDM system, illustrated the feasibility of integration based on the BOM. Wu et al
developed a Web-based PDM system using Web services component technology for data
exchange and sharing in one company and among cooperative enterprises[22]. Chan et al
proposed a concurrent control model for PDM that could cater for version management and
product architecture [23]. Meng et al built an integrated platform based on PDM system.
Information integration and application integration were achieved based on process-driven.
The data can be accessed each other through using the API and integration tools provided by
PDM system [24].
The research work of application system integration based on PDM is another focus,
especially the integration between PDM and ERP. Hou et al proposed a methodology for the
PDM/ERP/CAD integration. The UML-based methodology was applied for the product data
integration model. The PDM structure model diagram and the class model diagram as well as
workflow method were introduced. By workflow method, the data publican is realized by the
trigger[25]. Qiao et al presented an information conversion and integration method between
PDM systems and ERP systems. The information transformation and data synchronization
between the two systems were addressed by developing a conversion mechanism, with which
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 145

an engineering BOM can be converted to manufacturing BOM via process planning BOM, so
as to achieve information semantic unification between the two systems[26].

3. REMAINING PROBLEMS
The conventional integration technologies and methods have significantly enhanced the
performance of integration among various application systems. For example, geometrics
models are utilized to model product geometric information, and they can support CAD
systems to exchange and share product information. However, owing to the new challenges
from the manufacturing environment and information technology, there are some limitations
for conventional technologies:

(1) Point-to-point or one-to-one integration. The conventional integration technologies

are normally used to model products for supporting the integration of one or two
systems such as a CAD or CAPP system or special application system. They can’t be
directly used in the integration of systems that are employed in other stages of
product development, such as computer aided manufacture or computer aided
engineering (CAE). Hence, more work should be carried out to establish seamless
product data exchange and sharing across the entire product development processes.
For instance, the CAD based geometric product model does not define all the
necessary information for the downstream manufacturing processes. The
representation format of product data can impede the fluent information flow
between systems, and lead to the high development costs as a result of unnecessary
costly rework and redesign[27]. With the new emerging software system, the point-
to-point integration has to be upgraded.
(2) Problem of cooperation: many complex products are usually developed by
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

combining the strength of several manufacturing companies. Hence, data exchange

and sharing between these companies should be very efficient and effective. Most
companies structure their products using different modeling methods. Hence, it has
become an issue for them to cooperate with each other in support of the development
of a particular product. Normally, an extra data conversion process should be carried
out. This is very inefficient. Sometimes, conflicts about the model structures may
lead to loss of the information which cannot be converted[27]. New modeling
methods have to be developed to build up the high compatibility product model.
(3) Static and one-way integration. Most conventional integration technologies have
normally been used to integrate static product data and normally only support what
we call “One-way integration”. Due to lack of dynamical and bidirectional
integration mechanism, product related information may not be complete, correct and
up-to-date. As the product development is a very complex process involved many
stages and organizations, product information is normally produced dynamically,
which might be changed or modified many times according to the customer
requirements during the product lifecycle. Downstream changes related to a product
are also common, hence product information must be transferred bidirectionally and
dynamically among different product development stages and applications. A

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 146 S. Q. Xie and W. L. Chen

suitable bidirectional and dynamical integration method is required to support the

cooperation and collaboration among distributed teamwork.
(4) Limited expansibility and opening. Conventional integration technologies are usually
focused on a special issue, for example, traditional dedicated manufacturing
software/hardware systems have been designed to produce specific products. When
the need arose to quickly produce customized products, the systems are not up to the
job and quickly became obsolete. The systems are also standalone and there are no
open standard interfaces and mechanisms for them to communicate or integrate with
other systems employed by partners in the global environment. This has caused a
series of problems such as product data management problems, software
management problems and scheduling problems[28]. Hence, there is a strong need to
develop systems that are reconfigurable and extensible for the integration to keep up
with the continuous changes in the global manufacturing environment and
information technologies. The integration systems should be able to dynamically
integrate new subsystems or tools into the environment or remove existing systems
from the systems without influencing the basic.

4. AN INTEGRATED PLATFORM BASED ON STEP AND PDM

This section proposes an integrated platform based on STEP and PDM that aims to
provide an open and extensible infrastructure for product data exchanging and sharing in the
current manufacturing environment. In terms of the problems mentioned above, the research
work aims mainly focused on product data integration and integration framework. The goal of
the research work is to propose a generic integrated platform, which is reconfigurable and
supports efficient product information exchange and sharing dynamically and bidirectionally.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

4.1. Integrated Platform Architecture

An integrated platform can be characterized by a set of methods and accessible

mechanisms able to support and assist in data and application integration tasks. Due to the
complexity and diversity of data and application, the platform is proposed to meet the
following fundamentals.

(1) Open and expandable. Information technologies used in various OKP companies are
very different, with the emerging of new technology, there are always various
applications and data, which are needed to be integrated. In addition, different
industries may require different type information for the product development and
manufacturing. The integrated platform should be open and expandable to satisfy the
new emerging application integration requirements, not for some particular
applications or for company current applications only.
(2) Dynamical and bidirectional integration. This is also very important for the
integrated platform. As product development process normally involves activities in
various stages, product related information is often transferred from one person to

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 147

another person, from one place to another place, so product information is very
active, the data exchange and sharing needs to be dynamical and bidirectional.
(3) Reconfigurable and flexible. Most OKP companies are no longer geographically
confined; they are run in virtual environments in which related manufacturing
operations may be widely distributed geographically. The size of OKP companies
also differs significantly. The integrated platform needs to be reconfigurable for
various OKP companies. The function components of integrated platform must be
plug-in and take-out and the system architecture should be reconfigurable and
flexible.

A multi-layer, reconfigurable integrated platform architecture is proposed based on above

fundamentals, as shown in Figure.1. The architecture can be described by several levels. Each
level is devoted to a specific task and is designed to bring the interface with the integrated
platform from different level of abstraction and functionality. The main goal of such
architecture is to facilitate the application’s integration task, offering different level of access
and integration with the platform, providing an integrated and concurrent environment with
which employees of company can share product development software tools and data
dynamically, and work concurrently for the development of OKP products.
Figure 1 shows the proposed architecture that comprises five main layers; each one
corresponds to a specific task.

(1) Support layer provides the necessary computer net environment that supports the
system running. It includes a computer network, a Database Management System
(DBMS), Operation Systems (OS), desk PCs and Servers.
(2) Database layer is the core database for the integrated platform, which stores the
company’s product design data, manufacturing data, industry resource data and
knowledge database. Product design data can be product design models, drawings,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

product structure, part and other product design related information. Manufacturing
data includes manufacturing process information, machining operations, cutting
characteristics, tool requirements, NC codes etc. Industry resource data includes the
resources such as material, machine tool, fixture database, tool which helps to guide
the engineer to design or manufacture product. Knowledge database is a set of
industry and company product design and manufacturing knowledge, such as the
process design rules, the tool selection rules, the cutting parameter selection rules etc.
(3) Platform layer is comprised of product modeling, process modeling, enterprise
application integration, which is the core and important layer. Product modeling is to
provide environment and methods for modeling various types of customized product
for supporting efficient data exchange and sharing. Process modeling is used to
provide modeling environment and methods for modeling the product development
process, such as the new product development workflow, the engineering changing
workflow etc. This helps to promote the product data moving and sharing among
team members and different departments.
(4) Application integration is to provide standard methods and open interfaces for
integrating various applications into the platform. These applications might be used
in different product development stages, such as CAD in design stage, CAPP in
process planning stage, CAM in manufacturing stage.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 148 S. Q. Xie and W. L. Chen

(5) Function layer is the functional implement layer for purpose of product data
management. Based on the database layer and platform layer, the function modules
in this layer can read/write, present, modify, compare, and transfer all the product
data stored in the database of integrated platform. For the purpose of dynamical data
exchange and sharing, a workflow management module and a project management
module are also provided. The organization and authorization management module is
to ensure the safety of data access in the product development process. Integration
interface module is a set of diverse application interfaces.
(6) Application layer is to provide industrial solution for a given company or industry,
such as motorcar, tools, mould, electronics etc. Each layer of the integrated platform
will be deployed flexibly according to the requirements of a particular company.
(7) As product data is required to be efficiently exchanged and shared in a product
lifecycle, the integrated platform must integrate the more often used software tools in
different stage of product lifecycle. These software tools can be classified by two
groups, one is tools software, such as CAD, CAPP, CAM etc, the other is
management software, such as ERP, CRM (Customer Relationship Management),
SCM (Supply Chain Management), and MES (Manufacturing Execution System).
(8) To develop this proposed integrated platform successfully, there are two key issues
need to be solved. First is about the product modeling. The goal of integrated
platform is to establish seamless product data exchange and sharing across the entire
product development processes and among different cooperative companies, the
implementation foundation is the generic product model based on international
standard, which would be accepted globally. This is the focus task and challenge for
product modeling. The other key issue is about process modeling. As the integrated
platform is aiming to dynamical integration distinguished from the traditional
method, how to dynamically integrate the product data and related applications is
also very important.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

4.2. Generic Product Modeling Framework

Product modeling technology is the key and indispensable technologies for supporting
efficient data exchanging and sharing in the proposed integrated platform architecture. This
section introduces the generic product modeling framework (GPMF) that attempts to provide
an infrastructure for modeling various types of customized products. The output of the GPMF
is a set of data models defined to model a product at different stages of its development
processes. Figure 2 shows the structure of the GPMF developed based on STEP. It consists of
four functional components including an EXPRESS data model - EDM, a STEP-based
modeling environment, a ‘five-phase’ modeling method, and three EDM data exchange and
sharing methods.
The EXPRESS data model (EDM) is the core of the proposed GPMF. It defines a
complete product data structure and uses the standardized data format. There are eleven
defined EXPRESS schemas defined and STEP AP 203 included in the EDM. Each EXPRESS
schema utilizes either STEP resources or STEP-based compatible resources defined to model
a particular type of product information.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,

Tools Management
Application
Software software
integration layer 1
2
a1 b1
5
6 integration
a2 b2
3 7
a3 b3
4 8
a4 GND b4
0

Motorcar Tools Electrical equipment Mould Electronics Others

Function
CAD ERP
layer
Product structure Product configuration Object Workflow
management management management management

Change Organisation/ Integration

Project management
management authorization management interface
CAPP CRM

Platform
layer
Enterprise Application
Product modelling Process modelling
Integration
CAM SCM

Database
layer

CAE Product Knowledge MES

Product Industry
manufacturing database
design data resource data
data

Others
Support Others
layer
Computer Net DBMS OS PC/SERVER
PC

Figure 1. Integrated platform logical architecture.

 Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,

Modularization Five-Phase
Modeling Method
Basic modeling objects Completion of Constraints EDM Data
Exchange &
Sharing Methods
Relationships &Attributes Model Integration

DBMS
(SQL-Server,
Product General EXPRESS Data Oracle)
Information Module Model (EDM)
Product Product
Data Product Geometric Model
Data Module
Resource Module SDAI
Product Manufacturing (ST-Developer,
Data Module VC++)

Application Protocol 203

STEP-Based Neutral File
Modelling Environment (STEP Part 21
File)
Generic Resources
Newly Defined
Modeling Language Resources
EXPRESS/EXPRESS-G

Figure 2. Structure of STEP-based GPMF.

 Integration of Product Development Process Using STEP and PDM 151

The STEP-based modeling environment is developed for the GPMF. Within the
environment, a modeling language-EXPRESS and its graphical representation method
EXPRESS-G are used to model product structure. STEP generic resources are utilized to
model product information defined by STEP. STEP AP 203 is used to model product
geometric information, and new modeling resources are defined for modeling product
information that is not covered in STEP.
The ‘five-phase’ modeling method is proposed to develop the EDM. The method defines
a formal approach to logically organize all the tasks for building up the EDM in the modeling
processes.
Three EDM data exchange and sharing methods are used in the GPMF. As shown in
Figure 2, product data is exchanged and shared through either neutral physical files(STEP
Part 21 file), or Standard Data Access Interface(SDAI), or Database Management
Systems(DBMS). The product models defined within the GPMF are exchanged or shared
using one of the methods. These three methods are easily integrated into any application
software environment, which makes it easy to implement the product models defined by the
GPMF in applications.

4.2.1. STEP-Based Modeling Environment

The STEP-based modeling environment in the GPMF is to facilitate data sharing and
exchange through well defined STEP-compliant product data models. The EXPRESS data
modules and the modeling methods are used to develop STEP-compliant data models for the
integration of product development systems. A number of methods and resources are
developed for establishing such an environment.

(1) Modeling Language: EXPRESS and EXPRESS-G

EXPRESS modeling language consists of language elements that can allow unambiguous
data definition and specification of constraints on the data defined, and is one part of STEP
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

defined in Part 11[29]. The resources in STEP including generic resources and APs, are
normally represented as EXPRESS schema. In the GPMF, EXPRESS modeling language is
utilized to develop the EDM, which represents the structure of product data. The data model
in the EDM is represented by one or more schemas, which “group together the modeling
objects with related meaning and purpose”[30]. The most important EXPRESS language
element is the ENTITY data type, which defines the objects of interest in the domain being
modeled. The ENTITY is characterized by its attributes and constraints. EXPRESS language
also supports various kinds of data types, including simple types, aggregations types, and
constructed types.

(2) Generic Resources and STEP AP 203

The generic resources are directly utilized to define the basic data objects in an EDM. For
example, entities defined in product_definition_schema in STEP Part 41 [31] are utilized to
present the product general information in the EDM; STEP Part 45[32] is used to define the
EXPRESS schema that represents the data structure for material information. The generic
resources are defined in STEP. The use of the generic resources enhances the compatibility of
the GPMF. The STEP AP 203[33] is utilized to model product geometric information in the
EDM. It is employed by different CAD systems as the data model to structure product

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 152 S. Q. Xie and W. L. Chen

geometric data. Thus, the product geometric data module of EDM can be integrated with
CAD systems using STEP AP 203.

(3) New Modeling Resources

STEP has already defined many modeling resources to support the product modeling.
However, it is still under development. There are still not sufficient resources defined to
present information of different types in product development processes. Hence, two types of
new modeling resources are defined and utilized in the GPMF as the supplementation of
STEP-defined modeling resources.
One type of new resources is called hybrid modeling resources. They are modified from
the STEP generic resources. For instance, the document_schema in Part 41[31] is modified to
generate a data model for the document information. The method_definition_schema and
process_property_schema in Part 49[34] are modified to define the data model in the EDM
for the process information. New STEP-compliant modeling resources are defined by the
authors to model the product data which are not covered in STEP [27].

4.2.2. ‘Five-Phase’ Modeling Methodology

The ‘five-phase’ modeling method is proposed for the development of the EDM. The
method is to standardize the modeling of different types of product information. It is
composed of following five phases:

(1) Phase One: Modularization

The main tasks in this phase include: (a) product modeling objective analysis, (b)
classifying the product data, (c) modularizing the EDM. In the modularization phase, the
EDM is divided into four modules including a product general information module, a product
geometric data module and a product manufacturing data module, and a resources module
that is developed to present the sharable basic modeling objects extracted from the other
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

modules.

(2) Phase Two: Basic Modeling Objects

This phase defines the basic modeling objects and the general structure of each EDM
module. The phase starts by analyzing the structure of the defined EDM modules. The
fundamental elements of the modules are identified and defined as the basic modeling objects.
After defining the basic modeling objects, the way to structure them is analyzed and applied
in the EDM. For example, in the product manufacturing data module, product assembling
information is normally defined by four basic modeling objects including an assembly
product object, a product component object, a subassembly component object, and an object
called connector is defined to represent the connections between these objects. These objects
are named as: assembly_product, part, subassembly, and connector respectively.

(3) Phase Three: Relationships and Attributes

This phase refines and enriches the basic modeling objects defined in the second phase by
adding relationships to the defined objects. The tasks are included in this phase: (a) definition
of attributes and relationships between entities; (b) enhancement of the defined entities; (c)
definition of new entities; and (d) correctness checking. This phase must be continued until

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 153

the EDM has reached the desired level of details for representing the content of objects and
their relationships with other objects.

(4) Phase Four: Constraints Modeling

This phase models the constraints of an EDM. It has the three tasks: (a) definition of
constraints of objects and their relationships, (b) addition of global constraints to the model,
and (c) model error- checking. Constraints modeling defines the objects and their
relationships based on requirements. A complete EDM module is developed after this phase.
For example, one local constraint named WR1 is defined in the part. This constraint specifies
the range of the level_in_assembly_hierarchy attribute, which defines that the part must be in
the second or under the second level of the assembly model tree. The error-checking task
detects possible grammar errors, missing information, and conflicts of constraints.

(5) Phase Five: Model Integration

This phase integrates the modules defined after the four steps to form an EDM. The main
tasks involved in this phase are to check how data is represented in each module, define the
inputs and outputs the four modules, and evaluate whether EDM is complete, minimally
redundant, unambiguous, and error free.

4.2.3. EDM Data Exchange and Sharing Methods

To support data exchanging and sharing, it needs to be mapped to a model that is
accessible through commercial software tools. There are three EDM data exchange and
sharing methods used in the GPMF. These three methods are: (1) implementation via neutral
file, (2) implementation via SDAI, and (3) implementation via DBMS. Figure 3 shows the
different level method which may have different implementation complexity and difficulty
Level.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Applications

+ DBMS

Complex SDAI

Neutral Format
-

- Difficult +
Figure 3. Data exchange and sharing methods.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 154 S. Q. Xie and W. L. Chen

EDM-defined data structures can be used to support various stages of a product

development process, e.g. CAD, CAPP, CAM, etc. These systems can be integrated via the
three implementation methods for EXPRESS data model. An integrated product development
platform can be then developed. Product data can be made available by using API and data is
presented as EDM-defined structures. Different application systems can access the product
database to read and write product data. The data structure defined in the EDM should be
mapped into the product database. This structure information can support the possible DBMS
applications to implement this level method.

4.2.4. EXPRESS Data Model

Figure 4 shows the structure of the EDM. It consists of four modules. The product
general information module, product geometric data module and product manufacturing data
module are defined and based on the results of classifying product data in the first modeling
phase. The resources module is developed by grouping the sharable basic modeling objects to
support the development of other modules.
The product general information module represents the product data, which are not
directly related with the product manufacturing, such as the product identity, product
property, and the relationship between products. In this module, a product_definition_schema
is defined to support modeling this aspect of the product data. The product geometric data
module supports modeling product geometric data, such as shape information and dimension
information. This module is a key to integrate different computer aided systems. In the
product geometric data module, STEP AP 203 is directly utilized for representing and
exchanging product 3D geometric information. The product manufacturing data module is the
core of the proposed EDM, and consists of nine EXPRESSS schemas to support modeling
different manufacturing data.

(1) A supplier_informaton_schema is defined to model the supplier information.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Manufacturing companies need information about suppliers and their products to

arrange the manufacturing processes.
(2) A manufacturing_facility_schema is defined to model the facility information. The
manufacturing facilities, such as machines, tools and fixtures are directly involved in
the product manufacturing processes. They are the key resources determining the
product manufacturing processes and influencing the product quality.
(3) A product_document_schema is defined for documents aspect. The documents are
defined as the industry standards and documented criteria that regulate and guide to
manufacture of a product.
(4) A bill_of_material_information_schema is defined for representing Bill of Material
(BOM) information. The BOM, which can be called part list as well, is a list of
product components or resources for manufacturing or assembling this product. The
BOM is utilized to determine the product cost and used as an effective tracking
source during the manufacturing and assembling processes.
(5) A material_information_schema is defined to represent the material information. The
material is the basic element for a product and it directly influences the selection of
proper manufacturing facility and manufacturing processes.
(6) A process_planning_information_schema is defined for modeling process data. The
manufacturing process information is the detailed description about the actual
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 155

manufacturing activities. It is essential to consider this aspect information in design

stage, which leads to optimize the manufacturing processes.
(7) An assembly_information_schema is defined for these two aspects product data.
Assembly product is one of the most important types of products. To develop an
assembly product, the components information and the assembling method are
required.
(8) An inspection_information_schema is defined to model inspection information. The
inspection processes are essential parts of product development processes to control
product quality. The inspection results can assist to indicate the problems of a
product and its production processes.
(9) The cost_information_schema is defined for the cost information occurred in product
development processes. The inclusion of cost information is critical for any
engineering or manufacturing organization. Managing cost information can help the
firm to increase its own competitive ability.

The resources module in Figure 4 defines basic modeling objects that are shared by the
other modules. All these basic modeling objects are grouped into the supporting_schema. The
resources in the supporting_schema are represented as EXPRESS entity, the constructed
TYPE, and FUNTION. Through the EXPRESS schema interface, these resources in this
schema are utilized by other schemas to structure an effective and efficient data model
representation. The four modules of EDM are developed by applying the STEP generic
resources including Part 41, Part 45, and Part 49, STEP AP 203 as well as the new defined
STEP compatible modeling resources.

4.3. Process Modeling and Integration

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Product development process is dynamical, which involves many sub tasks and
participants. Hence, as result of the process, product data will be filled in gradually in
different life cycle stages, flowed among different organizations and different software
applications. The accuracy and timeliness of product data is very important during the
process, which may influence the efficiency of team collaboration. Process modeling
technology is the key and indispensable technology for supporting dynamical data
exchanging and sharing in the proposed integrated platform architecture. This section
introduces a process modeling method based on workflow modeling that attempts to provide
an open infrastructure for modeling various product development process.

4.3.1. Workflow Modeling Principle

The Workflow modeling principle is described in Figure 5. The modeling environment is
comprised of one engine, three models and one application module. Workflow engine is the
core part of the modeling environment, which is responsible for driving workflow by
interpreting workflow models, calling appropriate application system, handling related
product data. Three models are organization model, workflow model and product data model.
Organization model defines WHO will participate in the product development process.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,

Support Other Modules

supplier_informaton_schema
manufacturing_facility_schema
product_document_schema
bill_of_material_information_schema
material_information_schema
process_planning_information_schema
assembly_information_schema
Schema Product_Definition inspection_information_schema
Level STEP-AP203 STEP-AP203
_Schema cost_information_schema

Module Product General Product Geometric Product Manufacturing Resource

Level Information Module Data Module Data Module Module

EXPRESS Data Model(EDM)

Figure 4. Structure of the EDM.

 Integration of Product Development Process Using STEP and PDM 157

Workflow editor

Refer Refer
Define

Organization Model Workflow Model Product Data Model

Explained Apply
Refer Apply

Workflow Engine Application System

Maintain Use Apply

Update

Workflow Workflow Use Application

Control Data Data Data

Figure 5. Workflow modeling principle figure.

This is only a logical definition, they can be totally geographically distributed

organizations, teams and roles. Product data model defines WHAT to be input and output of
the process; they are the data source driven by every workflow node. Application systems are
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

registered and bound with different data type. When a type of data is dealt with, the bound
application system will be called automatically to do it. For example, when a CAD file is
generated by Pro/Engineering and needs to be revised, the workflow engine will help to call
Pro/Engineering to open the file automatically.
Workflow model is made up of workflow templates. A workflow template is composed
of a series of tasks, which may have time sequence. Task type can be classified into
compound task and task step. Task step is the minimal execution task unit which can not be
decomposed any more. Compound task is made up of task step or tasks. Figure 6 is a sketch
map of workflow template. A, C, D are both task step, B is a compound task. The arrow in the
figure indicates task or task step moving direction. These tasks or task step can be executed in
serial or parallel, which are driven by the workflow engine. Task B have one input arrow, two
output arrows. That means, task B receives information comes from task step A, deals with
the information and judges the result according to the predefined rules, if the result meets the
rules, task goes forward to task step C, if not, goes back to task step A.
Workflow editor is the core module for creating different workflow model according to
the integration requirements. A workflow template is established by the workflow editor,
which includes the tasks, the executive roles of these tasks and the input and output product
data.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 158 S. Q. Xie and W. L. Chen

Figure 6. Workflow template.

4.3.2. Task Correlation and Status Definition

A task is divided into a series of serial or parallel sub task according to its complexity.
There are different time sequence relationships among these subtasks, which are defined as
task correlation. Task correlation is defined as four relationships: end-begin, begin-begin,
end-end, begin-end.

• End-begin: task B can only begin when task A is ended.

• Begin-begin: task B can only begin when task A starts
• End-end: task B can only end when task A is ended
• Begin-end: task B can only end when task A starts.

As product development process is long and constantly repeated, the related task status is
often changed according to inputs and outputs. Task status is shifted by workflow engine
automatically, which includes seven types: initialization, wait, work, pause, termination,
finish, storage. Figure 7. shows the change of task status.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

• Initialization: status before a task starts.

• Wait: task is started, but cannot be changed to work status due to the restriction of
task correlation. Wait status will be changed to work status automatically when task
satisfies task correlation.
• Work: task has been start and in work status.
• Pause: task status is changed from work to pause. It can return to work status by
activation.
• Termination: task is terminated. It cannot be restored after its termination.
• Finish: the status after task has been finished and signed.
• Storage: the status after task has been finished or been terminated. Task cannot be
restored after its storage.

4.3.3. Task Step Model

Task step model is the minimal execution unit in the proposed workflow model. It is
responsible for interpreting and handling the input data and requirements and exporting
current task’s result. A task step model is proposed as Figure 8.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 159

Satisfy task
Work Sign
correlation

Wait Finish
Start
Activation

Start Pause Archive

Cancel
Pause

Initialization

Storage
Termination Archive

Figure 7. Task status changing in workflow.

S2 I4
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

A B

I1 O1
I2 O2

I3 O3

S1 I5
Figure 8. Task step model.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 160 S. Q. Xie and W. L. Chen

I1~I5 are the inputs of task step model, O1~O3 are the outputs of task step. S1~S2 are its
own attributes. Their definitions are as following:

• I1: the input product data or files that are needed to be dealt with by the task step.
• I2: the time requirement for the task step. It defines the expectative start time and end
time.
• I3: the people who will execute the task step. The people can be one or more real
persons; it also can be an organization, not a real person.
• I4: authorization definition. This shows what kind of authorization to the input data
and files the people may have in this task step. Normally, the authorization activities
may include browsing, modifying only, marking/modifying and marking only.
• I5: signature mode. This mode defines in what situation can pass the current task
step. There are four signature modes: anyone pass, all pass, half pass and two third
pass.
• O1: the output product data or data files of task step.
• O2: signature opinion. After finishing current task step, the people must give opinion
about current task step: agree, disagree or abstain.
• O3: the next task step or compound task after current task is finished.
• S1: task step status. The seven types of status are introduced in Section 4.3.2.
• S2: task step correlation with other task step. The four correlation relationships are
introduced in Section 4.3.2.

4.4. Integration Scenario

Figure 9 shows the tasks involved in the integration of tools throughout product
development life cycle, from design to fabrication and assembly. STEP and workflow are
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

used within the cycle. The design phase results in STEP data and includes the geometry of the
part, which is defined in AP203 or AP214. The data file is then transferred to macro process
planning which is a high-level process plan for a machined part. The process plan information
for both numerical controlled and manually operated applications, and associated product
definition data is defined in AP 240. The macro process planning defines the sequence of
machine process, each of them identifies the machine tool, machine setup, clamping
positions, a list of machine operations, and a list of machining features that are eligible to be
machined per process. Micro process planning comes after macro process planning, which is
closely related to a CNC machine. This is done by AP238[35] or ISO 14649. Micro process
planning defines the process information for a specific class of machine tools, such as turning,
milling, drilling, reaming or tapping. The feature sets are used during the computer-aided
manufacturing phase. Based on this, machine independent STEP-NC files are generated that
are executed by CNC machine tools and results in the machined parts. The manufactured part
is transferred to assembly line, and then the end product will be produced.
The figure helps to understand the activities in a whole product development chain. The
organizations can be in one company or geographically distributed companies. The product
development activities are driven by workflows automatically with related product data.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,

Organizations Activities Applications

Requirements

A STEP AP203 or AP214 CAD

Design

STEP AP 240 STEP AP240 & AP238 STEP NC AP238

B Macro Micro CAPP/CAM

NC
Process Process
Programming
Planning planning

Machine
C mechanical
parts CNC
Fixture

Fixture parts library Raw part

D
Design,Manufacture Material
CAD/ERP
fixture Preparation Machined parts

Fixture

E Product Assemble ERP

Figure 9. The integration scenario in product development life cycle.

 Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,

Workflow Model

App1 App2

App1 App2 App3 App4

App3 App4

Product data based on STEP

current way new way based on integrated platform

Figure 10 The different way of action in product development life cycle.

 Integration of Product Development Process Using STEP and PDM 163

In different phases, different applications may be used in product development, such as

CAD used in design phase, CAPP/CAM used in process planning phase, ERP used in
manufacturing phase. This accelerates product designing and manufacturing as a whole.
The difference between current way and integration way is shown in Figure 10. As in
current way, the integration between applications such as App1~4 is one-to-one, there are no
close relationships among them. Product data exchange and sharing interface may have to be
customized to a special application and the data flow is manual. While using the new way
based on integrated platform, the data exchange and sharing is via database based on STEP,
every application is interoperable with the common product database without one-to-one. The
action of exchange and sharing is driven by workflow engine automatically. This improves
greatly the efficiency and quality during product development process.

5. CASE STUDY
A prototype system names Teamwork Solutions has been developed to validate and
demonstrate the feasibility and the compatibility of the proposed integrated platform. Main
software tools used to develop the system include: Microsoft Visual C++ 2005 which is
employed to develop framework and functional modules; ST-Developer 12, which translates
data between STEP information models and integrated platform; Microsoft SQL Server 2005
which is utilized to construct the system’s repositories. One interface of the prototype system
is shown in Figure 11.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 11. Main control center interface of prototype system.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 164 S. Q. Xie and W. L. Chen

To illustrate the related functionality of the prototype system, the following main tests
will be carried out:

(1) Create a product development workflow and start it.

(2) Design a product using Pro/Engineering tool.
(3) Product information integrated into system’s database.
(4) Proofread the designed product.
(5) Return to design stage and revise the product.

The case studies are focused on the information and activities integration only using
partial stage of product development life cycle; the other product development activities are
very similar in principle.

(1) Scenario 1: Workflow modeling

An authorized person, e.g. product manager, can create company’s own special new
product development workflow by using workflow modeling module of system. The example
of workflow created by prototype system is shown in Figure 12. A new designed product will
go through several phases which are showed as task icons, each phase stands for a special
task for product development, such as part designing, proofreading, verifying, process
planning, standardization and confirmation. The arrow indicates the flow direction of product
data and task. Every task node may have different signature options and authorization to the
input data which can be defined through user interface shown in Figure 13.
After modeling the workflow, the product manager can start the workflow, and specify
the executor of each phase. The first task is to design a product. When the executor of this
task logins system, he will receive the task information.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 12. Workflow modeling example.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 165

Figure 13. Workflow task node definition.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 14. Transmission shaft product model designed in Pro/Engineering.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 166 S. Q. Xie and W. L. Chen

(2) Scenario 2: Part Design

The aim of this phase is to design a product using a CAD tool. The CAD tool can be
registered in the prototype system in advance, which will be called when current task is
executed. Figure 14 shows a transmission shaft product designed by Pro/Engineering. This
product includes nine parts; its hierarchy relational structure is shown at the left side of the
figure, which has two levels.

(3) Scenario 3: Product Modeling and Integration

After finishing design the product, the result will be stored in system’s database in order
to be exchanged and shared by other applications and other organizations. The product data
will be modeled based on above mentioned methods before it is stored in database. The
geometric product data of the transmission shaft product are modeled into a Part 21 file
through an AP203 output interface defined in Pro/Engineering.
Figure 15 shows the windows display results of the transmission shaft when this Part 21
file format product model is loaded in ST-Viewer in the ST-Developer toolkit. It is apparent
that the geometric product data in the product model can be read by systems to regenerate the
3D geometric models which has STEP AP 203 interface to load/output product geometric
data. The screenshots of the transmission shaft product model in both Pro/Engineering and
ST-Viewer are the same, which means that the geometric product data of the product model
can be exchanged and shared by the above two application systems.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 15. Transmission shaft product model loaded in ST-Viewer.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 167

Part 21 file is the important output of design phase, it is the foundation of further product
data exchange and sharing in different product development stages. The Part 21 file of
transmission shaft is shown in Figure 16.

Figure 16. Part 21 file for geometric data of transmission shaft product.

This Part 21 file can be imported by other application systems to achieve the product
geometric information. The prototype system has proposed a special interface to read directly
Part 21 file based on STEP AP 203. The developed user interface is shown in Figure 17.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 17. STEP file read interface.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 168 S. Q. Xie and W. L. Chen

After reading into the Part 21 file, the system deals with the data and stores them into its
own database. Through another view, the transmission shaft product structure and 3D model
can be browsed, as shown in Figure 18. It can be found that screenshots of the transmission
shaft product 3D model and product structure in both Pro/Engineering, ST-Viewer and
prototype system are the same, which further verifies that the product data model can be
exchanged and shared among application systems when they are based on STEP.

Figure 18. Browse product 3D model in system.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

(4) Scenario 4: Proofread the Designed Product

As the designer finishes his task and submit the result of his task, the workflow will
automatically move to the next predefined task: proofread. The executor maybe another
person. He will receive the task when he logins in the system. As the design result has been
stored in the database and flow automatically to current task executor, the executor can then
browse the 3D model, check the structure of product etc as he likes under the given
authorization. He may then give his opinion about the designed result: agree, disagree or
abstain, attached with his reason and suggestion. In this example, we assume the given
opinion is “disagree”, so the workflow has to be back to the front phase: part design. The
interface of this phase is shown in Figure 19.

(5) Scenario 5: Return to Design and Revise the Product

When the design result is returned back, the product designer has to revise the product
according to proofreader’s suggestions. This example illustrates the bidirectional integration
between CAD and prototype system. The prototype system can import and access product

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 169

data from CAD system, which is called “positive integration”. The CAD system can also
access database of the prototype system, which is called “reverse integration”. When a
designer logins the system, he will receive his task of revising his former designed product.
Driven by the given workflow model of the prototype system, Pro/Engineering may be called
and the former designed product is opened, see Figure 20.
The screenshots shows the same 3D product model and the same product structure as
before without data loss. The transmission shaft product model data is transferred to
Pro/Engineering positively and automatically by prototype system. The change is found at the
menu of Pro/Engineering, which is added a new menu named “PDM”. This menu is made up
of several communication functions, which can be used to access database of prototype
system under the system’s authorization. These functions include “connect PDM”,
“disconnect PDM”, “Check in”, “Object information”, etc. With the help of reverse
integration, designer can search, compare, copy or edit product database, execute own task
under CAD environment. Due to the interface of reverse integration is developed based on
COM/DCOM/COBRA standard, it can also be used to integrate with other applications such
as CAPP, CAM etc.
When designer has finished revising the product, he can check in his product file and
submit his current task. Then the task will go forward to the next driven automatically by
workflow engine until they are finished. All the product data will be transferred dynamically
and bidirectionally among tasks and all will be stored in system’s database together with its
related process.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Figure 19. Interface of task step-proofread.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 170 S. Q. Xie and W. L. Chen

Figure 20. Reverse integration in Pro/Engineering.

Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

CONCLUSION
In the advent of information technology and in the globalization of society, integration
has become a vital issue for businesses to grow and improve efficiency. The need to exchange
semantically rich information has also grown. In this chapter, an integrated platform based on
STEP and PDM is proposed to integrate tools and systems in order to achieve high
information exchange and sharing performance improvements for OKP companies. Two key
issues are discussed in detail in the chapter. One is the product modeling which is solved
through the proposed GPMF; the other is process modeling which is achieved by the
workflow modeling. Case studies have been carried out to validate the proposed integrated
platform. From this research work, the following conclusions are drawn:

(1) The proposed integrated platform is compatible for integration of different systems
and software tools. This has been validated in our case studies.
(2) The proposed modeling methods enable the integration of various application
systems bidirectional through common product database and the proposed integrated
platform.
(3) The product data flow can be driven forward or back by the proposed workflow
model dynamically and automatically.
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 171

(4) The dynamical integration capability of the proposed modeling scheme eliminates
the need for human interpretation, errors and rework in product development life
cycle.

The developed prototype system illustrates that the basic mechanisms of the integrated
platform are feasible; however, future research work is still required to further improve the
proposed platform.

(1) As information technology develops very quickly, new systems and tools will be
developed to support product development processes, further research work is
required to improve the proposed integrated platform in order to integrate these new
tools and systems.
(2) Develop new product modeling methods for the integrated platform. An
implementation method needs to be developed to map between extensible mark-up
language (XML) and EXPRESS, which is defined in STEP Part 28[36].
(3) As STEP itself is still at its development stage, further research work should keep up
with the new pace of STEP to achieve better performance, such as STEP-NC etc.

REFERENCES
[1] F.B.Vernadat, Enterprise modeling and integration : principles and applications. 1996,
London ; New York :: Chapman and Hall.
[2] Tu, Y., Xie SQ,Zhou ZD. An information modeling framework to concurrent product
design and manufacturing. in Proceedings of IAM’2001. 2001. American University in
Dubai, U.A.E.
[3] Tu, Y.L., S.Q. Xie, and J.J. Kam, Rapid one-of-a-kind production. International
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Journal of Advanced Manufacturing Technology, 2006. 29(5): p. 499-510.

[4] Xie, S.Q. and Y.L. Tu, Rapid one-of-a-kind product development. International
Journal of Advanced Manufacturing Technology, 2006. 27(5-6): p. 421-430.
[5] Xu, X., Integrating advanced computer-aided design, manufacturing, and numerical
control: principles and implementations. 2009, New York: Hershey.
[6] Nagalingam, S.V. and G.C.I. Lin, CIM-still the solution for manufacturing industry.
Robotics and Computer-Integrated Manufacturing, 2008. 24(3): p. 332-344.
[7] J.Kemmere, S., STEP:The Grand Experience. NIST Special Publication 939. 1999,
Washington: U.S.Government Printing Office.
[8] Li, H., et al. Feature-based, parametric modeling system for CAD/CAPP/CAM
integrated system. in Proceedings of the IEEE International Conference on Industrial
Technology. 1996.
[9] Gu, P. and K. Chan, Product modelling using STEP. Computer-Aided Design, 1995.
27(3): p. 163-179.
[10] Usher, J.M., A STEP-based object-oriented product model for process planning.
Computers and Industrial Engineering, 1996. 31(1-2): p. 185-188.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 172 S. Q. Xie and W. L. Chen

[11] Chin, K.S., Y. Zhao, and C.K. Mok, STEP-based multiview integrated product
modelling for concurrent engineering. International Journal of Advanced
Manufacturing Technology, 2002. 20(12): p. 896-906.
[12] Song, Y., X. Chu, and F. Cai, Real-time concurrent product and process design system
for mechanical parts. High Technology Letters, 1999. 5(1): p. 74-80.
[13] Zha, X.F. and H. Du, A PDES/STEP-based model and system for concurrent integrated
design and assembly planning. CAD Computer Aided Design, 2002. 34(14): p. 1087-
1110.
[14] Shaharoun, A.M., J. Ab Razak, and M.R. Alam, A STEP-based geometrical
representation as part of product data model of a plastics part. Journal of Materials
Processing Technology, 1998. 76(1-3): p. 115-119.
[15] Cai, C.T., et al., Design method of application protocol of the machine parts based on
STEP. Jisuanji Jicheng Zhizao Xitong/Computer Integrated Manufacturing Systems,
CIMS, 2002. 8(11): p. 892-895.
[16] Zhao, W. and M. Ma, Feature modeling for aeroengine blades according to STEP.
Beijing Hangkong Hangtian Daxue Xuebao/Journal of Beijing University of
Aeronautics and Astronautics, 1999. 25(5): p. 535-538.
[17] Jasnoch, U. and S. Haas, A collaborative environment based on distributed object-
oriented databases. Computers in Industry, 1996. 29(1-2): p. 51-61.
[18] Oh, Y., S.-h. Han, and H. Suh, Mapping product structures between CAD and PDM
systems using UML. Computer-Aided Design, 2001. 33(7): p. 521-529.
[19] Dong, J.H., et al., Research on CAD/PDM integration. Jixie Kexue Yu
Jishu/Mechanical Science and Technology, 2001. 20(2): p. 288-290.
[20] Eynard, B., et al., UML based specifications of PDM product structure and workflow.
Computers in Industry, 2004. 55(3): p. 301-316.
[21] Qiu, J.X., X.W. Wang, and X.C. Fan, Information integration between computer aided
design system and product data management system in collaborative product design.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Jilin Daxue Xuebao (Gongxueban)/Journal of Jilin University (Engineering and

Technology Edition), 2005. 35(5): p. 505-510.
[22] Wu, J.W., et al., Management strategy of product data in distributed collaborative
design environment. Zhejiang Daxue Xuebao (Gongxue Ban)/Journal of Zhejiang
University (Engineering Science), 2005. 39(10).
[23] Chan, E. and K.M. Yu, A concurrency control model for PDM systems. Computers in
Industry, 2007. 58(8-9): p. 823-831.
[24] Meng, X.J., et al. Research on integration platform based on PDM for networked
manufacturing. in IEEM 2007: 2007 IEEE International Conference on Industrial
Engineering and Engineering Management. 2007.
[25] Hou, J., et al. Integration of the CAD/PDM/ERP system based on collaborative design.
in Proceedings - ISECS International Colloquium on Computing, Communication,
Control, and Management, CCCM 2008. 2008.
[26] Qiao, L. and Y. Zhang, Implementation approach to information integration between
PDM and ERP systems. Beijing Hangkong Hangtian Daxue Xuebao/Journal of Beijing
University of Aeronautics and Astronautics, 2008. 34(5): p. 587-591.
[27] Xie, S.Q., W.Z. Yang, and Y.L. Tu, Towards a generic product modelling framework.
International Journal of Production Research, 2008. 46(8): p. 2229-2254.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Integration of Product Development Process Using STEP and PDM 173

[28] Xie, S.Q., X. Xu, and Y.L. Tu, A reconfigurable platform in support of one-of-a-kind
product development. International Journal of Production Research, 2005. 43(9): p.
1889-1910.
[29] Industrial Automation Systems and Integration: Product Data Representation and
Exchange: Part 11: Description Methods: The EXPRESS Language Reference Manual,
Reference Number: ISO 10303-11:1994(E), 1994.
[30] Kahn, H., et al., A generic framework for transforming EXPRESS information models.
CAD Computer Aided Design, 2001. 33(7): p. 501-510.
[31] Industrial automation systems and integration: Product data representation and
exchange: Part 41: Integrated generic resource: Fundamentals of product description
and support, Reference number: ISO 10303-41:2000(E), 2000.
[32] Industrial Automation Systems and Integration: Product Data Representation and
Exchange: Part 45: Integrated Generic Resource: Materials, Reference Number: ISO
10303-45:1998(E), 1998.
[33] Industrial automation systems and integration: Product data representation and
exchange: Part 203: Application protocol: Configuration controlled 3D designs of
mechanical parts and assemblies, Reference number: ISO 10303-203:1994(E), 1994.
[34] Industrial Automation Systems and Integration: Product Data Representation and
Exchange: Part 49: Integrated generic resources: Process structure and properties,
Reference number: ISO 10303-49:1998(E), First edition, Switzerland., 1998.
[35] Industrial automation systems and integration: Product data representation and
exchange: Part 238: Application protocol:Application interpreted model for
computerized numberical controlers. Reference number:ISO 10303-238, 2007.
[36] Industrial automation systems and integration: Product data representation and
exchange: Part 28: Implementation methods: XML representations of EXPRESS
schemas and data, Reference number: ISO 10303-28:2003(E), 2003.
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 INDEX

application, vii, 1, 3, 9, 10, 11, 17, 19, 27, 28, 29,

32, 39, 40, 52, 54, 57, 64, 79, 80, 81, 85, 86,
A 108, 111, 112, 115, 139, 141, 144, 145, 146,
147, 148, 151, 154, 155, 157, 163, 166, 167,
Aβ, 13 168, 170, 172
ABC, 31, 32, 38, 40, 43, 46, 47, 48 aptitude, 66
academics, vii arithmetic, 1, 13, 14, 15
acceleration, 15, 56, 82 artificial intelligence, vii, 1, 111, 112, 123, 127,
accelerator, 2 138
accuracy, 10, 16, 27, 48, 52, 53, 54, 63, 131, 155 aspect ratio, 11
acoustic, 7, 54, 129, 130, 139 assignment, 113
acoustic emission (AE) , 7, 54, 129, 130, 139 atoms, 80
acquisitions, 103 autocorrelation, 103, 104, 105
activation, 5, 67, 158 automation, 111, 142, 173
actual output, 21, 66, 84 availability, vii
actuators, 26
adaptive control, 27
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

adjustment, 56, 135 B

administrative, 142
aerospace, 51, 86 back, 11, 32, 52, 54, 56, 65, 66, 77, 85, 109, 130,
AFM, 51, 52, 53, 54, 55, 57, 58, 59, 61, 62, 63 157, 168, 170
agility, 142 barriers, 142
algorithm, vii, 2, 5, 9, 28, 31, 32, 33, 34, 38, 39, behavior, 2, 38, 56, 74, 84, 137
40, 43, 46, 47, 48, 49, 56, 57, 58, 59, 62, 63, bending, 11, 28, 29, 37, 40
65, 66, 67, 77, 79, 83, 84, 85, 88, 101, 102, benefits, 53, 141, 143
107, 108, 109 bias, 5, 88
alloys, 80, 102 bounds, 34, 38, 52, 55, 60, 63
alternative, 11, 112 brain, 66, 84
aluminium, vii, 49, 80, 97, 101, 102, 103 brass, 54, 80, 81, 97
aluminium alloys, 102 broad spectrum, 32
aluminum, 54, 81 browsing, 160
amplitude, 136 burn, 29
analytical models, 130 burns, 16, 28
ANN, 65, 66, 67, 68, 69, 70, 71, 72, 77, 79, 80,
86, 88, 89, 90, 91, 92, 93, 94, 95, 96
annealing, vii, 33, 49, 51, 52, 56, 62, 63
C
ANOVA, 104, 105, 106, 128
C++, 163
API, 144, 154
APP, 148, 154, 171

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 176 Index

CAD, 142, 143, 144, 145, 147, 148, 151, 154, convex, 33
157, 163, 166, 168, 169, 171, 172, 173 cooling, 19
CAE, 145 correlation, 91, 95, 98, 103, 123, 158, 160
CAM, 142, 147, 148, 154, 163, 169, 171 correlation coefficient, 91, 95, 103
carbide, 7, 8, 112, 128, 131 corrosion, 112
carbides, 80 corrosive, 66
carbon, 40 cost-effective, 80
carrier, 52 costs, 32, 142, 145
case study, 64 CRM, 148
cast, 10 cross-validation, 5
categorization, 134 cubic boron nitride, 113
ceramic, 80, 130 customers, 14
ceramics, 81, 82, 97 cutting fluids, 12, 102
chemical composition, 68, 86, 131 cutting force, vii, 7, 8, 9, 16, 36, 80, 86, 111, 112,
chromosome, 108 113, 115, 120, 121, 122, 123, 124, 125, 126,
classical, 68, 102, 112 127, 130, 138, 139
classification, 118, 130, 131, 135, 138 cutting tools, 52, 113, 128
close relationships, 163 cycles, 54
Co, 68, 131
cobalt, 103
codes, 147 D
collaboration, 142, 146, 155
data analysis, 127
communication, 38, 117, 142, 144, 169
data collection, 27
communication systems, 38
data mining, 27
compatibility, 145, 151, 163
data set, 71, 72, 89
complexity, 2, 146, 153, 158
data structure, 148, 151, 154
components, 6, 11, 31, 38, 48, 51, 52, 53, 63, 65,
data transfer, 142
89, 118, 129, 147, 148, 154, 155
database, 16, 68, 143, 147, 148, 154, 163, 164,
composites, vii, 49, 80, 82, 97, 111, 112, 113,
166, 168, 169, 170
115, 121, 123, 124, 125, 126, 127, 128
decision making, 12, 49
composition, 68, 70, 86, 131
decision-making process, 101
computation, 10, 32, 55, 56
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

decisions, 14, 117

computer science, vii
decoding, 56, 63
computer systems, vii, 143
defects, 19, 82
computing, 1, 2, 3, 23, 25, 26, 27, 56, 64, 80,
definition, 2, 13, 151, 152, 153, 154, 157, 160,
118, 128
165
concentration, 53, 54, 55, 59, 60
deformation, 29, 82, 112, 136
concurrency, 172
degrees of freedom, 113
concurrent engineering, 143, 144, 172
density, 109
conductive, 82
dependent variable, 103
conductivity, 12, 13
derivatives, 32
confidence, 56, 104, 105
detection, vii, 129
configuration, 56, 57, 84
deviation, 2, 7, 86, 89, 94
confusion, 130
discourse, 117
constraints, 8, 9, 31, 32, 34, 35, 36, 41, 43, 44,
discrete variable, 33
48, 54, 56, 107, 151, 153
discrimination, 91
construction, 26, 112
discs, 103
control, 1, 9, 11, 15, 16, 17, 25, 27, 28, 29, 53,
dislocation, 132
56, 77, 87, 95, 111, 112, 118, 127, 129, 144,
displacement, 116
155, 163, 171, 172
distribution, 2, 20, 57, 82, 117, 122, 130, 136
convergence, 6, 33, 40, 41, 48, 56, 57, 58, 63, 71,
diversity, 146
84, 85
division, 82
conversion, 144, 145
ductility, 66, 86
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Index 177

duplication, 142 flatness, 9, 25

duration, 39 flexibility, 80, 117, 142
dust, 82 flexible manufacturing, 32
float, 63
flood, 102, 103
E flow, vii, 2, 51, 52, 53, 54, 63, 64, 81, 82, 133,
142, 144, 145, 163, 164, 168, 170
economics, 28
flow rate, 81, 82
education, 31, 79, 98, 128
fluid, 13, 102, 131
elasticity, 37, 40, 70
food, 38, 39, 40, 41, 42
election, 147
fracture, 130
e-mail, 51
fractures, 132
emission, 7, 54, 64, 129, 130, 139
freedom, 113
employees, 147
frequency distribution, 130
encapsulation, 144
friction, 9, 10, 13, 19, 20, 22, 24
encoding, 56, 63
function values, 63
energy, 38, 81, 130, 136, 138, 139
functional aspects, 112
enterprise, 141, 142, 144, 147
fusion, 139
entropy, 138
fuzzy logic, vii, 1, 13, 14, 15, 16, 25, 29, 80, 111,
environment, 32, 49, 141, 142, 143, 144, 145,
112, 113, 118, 119, 122, 123, 127, 128, 135,
146, 147, 148, 151, 155, 169, 172
138
equilibrium, 56
fuzzy set theory, 17, 26, 29
erosion, 81
fuzzy sets, vii, 1, 12, 13, 17, 19, 25, 26, 29, 112,
estimating, 129
117, 118, 127, 135
estimator, 130
evolution, 32, 56
evolutionary process, 56 G
execution, 66, 157, 158
experimental condition, 71, 116 gauge, 25
experimental design, 68, 113, 127 Gaussian, 6, 119
expert system, 28 generalization, 13, 66, 77, 85, 91
extrapolation, 16 generation, 42, 43, 54, 86
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

extrusion, 11, 29, 52, 53, 55, 60, 61 genetic algorithms, 54, 64, 80
genetics, 32
geometric programming, 32, 33, 49
F glass, 80, 81, 97, 113, 123, 125, 128
globalization, 142, 170
fabrication, 160
goals, 8
factorial, 103
grades, 12, 13, 14, 15, 20, 21, 23, 86, 89, 95
failure, 129, 138
grain, 53, 55
fatigue, 53
grains, 52
feedback, 2, 7, 9, 15, 84
graph, 49, 91, 95, 109, 124, 138
FEM, 11, 12, 21, 25, 28, 54
grouping, 154
ferrous metals, 128
groups, 14, 80, 81, 135, 144, 148
fiber, 80, 112, 113, 123, 128
GSA, 33
fibers, 112, 113, 123, 125
Gujarat, 31, 51
filament, 113
filtration, 7, 26
finite element method, 20, 26, 54 H
fire, 82
fire hazard, 82 handling, 20, 54, 56, 155, 158
fitness, 32, 40, 41, 42 hardening, 9, 24
flank, vii, 129, 130, 131, 132, 133, 134, 135, 137, hardness, 40, 53, 55, 82, 87, 132
139 harm, 34, 49

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 178 Index

harmony, 34, 49 interface, 103, 133, 141, 144, 147, 148, 155, 163,
hazards, 82 164, 166, 167, 168, 169
heat, 12, 13 interference, 130
height, vii, 65, 66, 68, 69, 70, 71, 72, 73, 74, 75, interval, 13
76, 77 ion beam, 63
heterogeneous, 143, 144 islands of automation, 142
high-level, 160 ISO, 68, 103, 143, 160, 173
high-tech, 129 iteration, 48
hip, 130
hip fracture, 130
honey, 48 J
human, 1, 12, 32, 66, 84, 109, 112, 117, 171
judge, 123
human brain, 66, 84
judges, 157
humans, 117
Jun, 128
hybrid, 27, 28, 33, 34, 49, 54, 97, 130, 139, 152
hybrids, 25
hyperbolic, 88 L

L1, 12, 13
I
L2, 12
labor, 52
identification, 29, 132
Lagrangian, 12
identity, 154
language, 13, 14, 151, 171
implementation, 10, 56, 80, 139, 144, 148, 153,
law, 60
154, 171
layered architecture, 26
inclusion, 155
learning, 1, 2, 3, 5, 6, 7, 11, 15, 16, 24, 26, 38,
incompatibility, 141
66, 67, 70, 128
incompressible, 112
learning behavior, 2
independent variable, 103
life cycle, 143, 155, 160, 161, 162, 164, 171
indication, 104, 105
lifecycle, 145, 148
industrial, 51, 86, 143, 148
likelihood, 108
industrial application, 51
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

limitations, 34, 81, 141, 145

industry, 51, 80, 86, 147, 148, 154, 171
linear, 4, 5, 6, 15, 21, 32, 66, 77, 91, 95
inertia, 56, 58
linear function, 4, 5, 6, 21
inferences, 2
linear programming, 21, 32
information exchange, 141, 146, 170
linear regression, 4
information sharing, 38
linguistic, 2, 14, 26, 112, 117, 118, 121
information technology, 142, 145, 170, 171
linguistic information, 2, 14, 26
infrastructure, 142, 146, 148, 155
linguistically, 112
inhomogeneity, 87
links, 66, 67, 84
insects, 38
location, 11, 39, 55
inspection, 143, 155
losses, 107
inspiration, 25
low power, 19, 20
instruction, 2
lubrication, 9, 102
integration, vii, 141, 142, 143, 144, 145, 146,
147, 148, 151, 157, 160, 161, 163, 164, 168,
169, 170, 171, 172, 173 M
integrity, 97
intelligence, vii, 1, 38, 111, 123, 127, 128, 138 machines, 32, 64, 154
interaction, 66, 71, 73, 74, 75, 109 magnesium, 29
interaction effect, 66, 71, 73, 75 magnetic, 54, 63
interaction effects, 66, 71, 73 management, 142, 144, 146, 148, 172
interactions, 113 manufacturer, 8

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Index 179

manufacturing, vii, 1, 28, 32, 49, 51, 52, 66, 80, neural network, vii, 1, 2, 3, 5, 6, 7, 8, 9, 10, 11,
111, 112, 123, 127, 129, 142, 143, 144, 145, 12, 15, 16, 24, 25, 26, 27, 28, 33, 48, 54, 64,
146, 147, 152, 154, 155, 160, 163, 171, 172 65, 66, 69, 77, 79, 80, 83, 84, 85, 86, 88, 91,
manufacturing companies, 142, 143, 145 92, 94, 95, 97, 98, 129, 130, 135, 139
mapping, 3, 66, 144 Neural Network Model, v, 65, 66, 79, 98
market, 32, 141, 142, 143 neural networks, vii, 1, 3, 5, 7, 8, 9, 10, 11, 15,
mathematical programming, 32 16, 25, 26, 27, 33, 48, 54, 64, 83, 84, 85, 86,
matrix, 49, 70, 112, 123, 135 88, 91, 97, 98, 130, 139
MDH, 54 neurons, 3, 4, 5, 6, 10, 66, 67, 70, 79, 84, 88, 97
measurement, 56, 116 next generation, 43
measures, 28, 115, 136 Ni, 70, 87
mechanical properties, 68, 70, 82 NIST, 171
MED, 122 nitride, 113
media, 52, 53, 54 noise, 112, 132, 136
membership, 12, 13, 14, 15, 19, 20, 21, 23, 25, non-ferrous metal, 128
26, 117, 118, 119, 122, 135, 136, 137 non-linearity, 77
memorizing, 38 normal, 2, 52, 54, 55
memory, 118 normal distribution, 2
MES, 148 normalization, 70, 86, 135
metals, 82, 97, 128, 132
Metropolis algorithm, 57
microchip, 82 O
microscope, 68
obsolete, 146
microstructure, 81, 97
oil, 68
mining, 27
online, 7, 9, 11, 25, 127, 130, 133, 139
modeling, vii, 1, 6, 9, 10, 11, 15, 17, 20, 25, 26,
online learning, 11
29, 53, 54, 64, 66, 77, 79, 80, 84, 86, 89, 94,
operator, 13, 56, 63
95, 97, 99, 111, 112, 113, 117, 123, 127, 129,
optimization, vii, 1, 5, 6, 8, 9, 11, 14, 15, 25, 27,
139, 141, 143, 144, 145, 147, 148, 151, 152,
31, 32, 33, 34, 36, 38, 40, 41, 43, 44, 45, 48,
153, 154, 155, 157, 164, 170, 171, 172
49, 51, 52, 53, 54, 55, 57, 59, 60, 61, 62, 63,
models, 1, 10, 14, 16, 23, 26, 27, 28, 66, 79, 80,
64, 111, 112, 128
85, 88, 95, 97, 98, 101, 103, 107, 111, 123,
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

optimization method, 8, 33, 63

130, 135, 139, 141, 143, 144, 145, 147, 148,
optoelectronic, 130
151, 153, 155, 163, 166, 173
orientation, 123
modules, 138, 148, 151, 152, 153, 154, 155, 163
outliers, 7, 24, 26
modulus, 37
mold, 143
molecules, 80 P
momentum, 67, 71
money, 95, 97 parallelization, 142
movement, 56 paralysis, 85
multidisciplinary, 118 parameter, 5, 6, 7, 10, 11, 13, 20, 25, 26, 33, 48,
multiple regression, 103 52, 55, 56, 58, 61, 63, 68, 69, 70, 85, 87, 106,
multiple regression analyses, 103 127, 129, 136, 147
mutation, 108 Pareto, 102, 103, 107, 109
particle shape, 82
particle swarm optimization (PSO), vii, 51, 52,
N
55, 63, 64
particles, 52, 55, 56, 79, 81, 97
nanometers, 51
pattern recognition, 29, 117, 118
natural, 1, 12, 32
PCs, 147
natural selection, 32
penalty, 9, 56, 63
nerve, 66
pharmaceutical, 112

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 180 Index

physics, 2, 10, 11, 13, 16, 27

piezoelectric, 103
Q
planning, 10, 11, 15, 27, 32, 79, 97, 143, 145,
quality control, 87, 95, 129
147, 154, 160, 163, 164, 171, 172
quartz, 115, 116
plants, 68
plastic, 11, 82, 136, 143
plastic deformation, 82, 136 R
plasticity, 79, 97
plastics, 128, 172 Radial Basis Function, 3
play, 32, 143 radius, 21, 24, 37, 51, 52, 53, 131
plug-in, 147 random, 55, 56, 58, 67, 85
point-to-point, 145 random numbers, 56
polyester, 113 range, 33, 43, 44, 56, 58, 63, 68, 72, 73, 75, 77,
polymer, 52, 113 86, 113, 115, 130, 153
polymers, 128 rat, 82
polynomial, 54, 64, 137 raw material, 20
poor, 110 RBF, 6, 7, 11
population, 40, 43, 55, 56, 108 reading, 69, 168
population size, 40, 108 real numbers, 118
power, 8, 9, 16, 18, 19, 20, 31, 34, 37, 40, 43, 44, reasoning, 112, 117, 118, 119
48, 80, 81, 101, 102, 103, 107, 112, 113, 123, recognition, 29, 117, 118
132 regenerate, 166
predicting performance, 28 regression, 4, 6, 94, 98, 103, 135
prediction, 5, 7, 8, 9, 10, 11, 15, 23, 26, 27, 28, regression analysis, 94, 98, 135
54, 80, 84, 86, 97, 112, 113, 118, 124, 127, regression method, 6
130, 138, 139 rejection, 56
preference, 8 relationship, 53, 66, 77, 104, 105, 125, 154
press, 2 relationships, 152, 153, 158, 160, 163
pressure, 20, 52, 53, 55, 60, 61, 62, 81, 82, 116 relaxation, 136
principal component analysis, 89 relevance, 117
probability, 39, 41, 42, 56, 57, 117 reliability, 10, 20, 52, 85, 94, 129, 137
probability distribution, 57
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

repair, 56
probability theory, 117 repetitions, 88
problem space, 55 research and development, 28, 98
process control, 130 residuals, 104, 105
product design, 142, 147, 163, 166, 168, 171, 172 resin, 113
product life cycle, 143 resistance, 10, 36, 66, 112
production, 8, 14, 31, 33, 34, 35, 36, 40, 43, 44, resolution, 68, 132
46, 47, 48, 52, 142, 143, 155, 171 resources, 143, 147, 148, 151, 152, 154, 155, 173
productivity, 32, 34, 52, 54, 102, 106, 110 response surface methodology (RSM), 66
profit, 14, 33, 34, 39 returns, 36, 39, 40
profitability, 38, 39 rheological properties, 52
program, 85, 88 rigidity, 19
programming, 21, 32, 33, 49, 64, 102 robotics, 28, 49, 118, 171
propagation, 54, 65, 66, 77, 85, 130 rolling, 9, 10, 11, 17, 19, 20, 24, 25, 27, 28, 29,
property, 123, 152, 154 87
protocol, 172, 173 roughness, 7, 8, 9, 13, 14, 15, 26, 28, 31, 34, 37,
prototype, 143, 163, 164, 166, 167, 168, 169, 171 43, 44, 48, 53, 54, 55, 58, 59, 60, 61, 62, 63,
prototyping, 143 64, 79, 80, 83, 84, 86, 87, 88, 94, 95, 97, 98,
pulses, 136 101, 102, 103, 104, 106, 107, 110, 112, 127,
pure water, 81 128, 130
P-value, 105 roughness measurements, 86, 87
R-squared, 104, 105

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 Index 181

standards, 154
S statistical analysis, 128
steel, vii, 7, 8, 11, 20, 27, 28, 40, 65, 66, 68, 70,
safety, 148
77, 79, 80, 81, 83, 86, 88, 89, 90, 93, 94, 95,
sample, 56
96, 97, 130, 139
satisfaction, 19, 142
stochastic, 53, 56, 64
scalar, 102
stochastic model, 53
scaling, 67
stock, 10, 32, 33
scatter, 54
storage, 158
scheduling, 10, 20, 29, 146
strain, 11, 20
schema, 148, 151, 152, 154, 155
strains, 9
schemas, 143, 148, 151, 154, 155, 173
strategies, 14, 31, 32, 43, 48, 142
search, 32, 33, 34, 38, 39, 41, 49, 66, 102, 169
strength, 12, 15, 21, 31, 34, 36, 37, 40, 43, 44, 48,
searches, 85
52, 53, 70, 80, 86, 123, 142, 145
searching, 39, 42, 55
stress, 2, 10, 20, 40, 53, 54, 55, 136
selecting, 60, 102, 109
strokes, 55, 60, 61, 62
self-organization, 66
subjectivity, 12
semantic, 145
subtasks, 158
sensing, 129, 132, 139
suppliers, 154
sensitivity, 130
surface roughness, 7, 8, 9, 13, 14, 15, 26, 28, 31,
sensors, 7, 9, 26, 116
34, 37, 48, 53, 54, 55, 58, 59, 60, 61, 63, 64,
sequencing, 29
79, 80, 83, 84, 86, 88, 94, 95, 97, 98, 101, 103,
series, 16, 68, 79, 97, 118, 146, 157, 158
104, 106, 107, 110, 127, 128, 130
services, 144
swarm, vii, 38, 42, 48, 51, 52, 55, 56, 63, 64
set theory, 3, 12, 13, 14, 17, 19, 20, 26, 29, 112,
swarm intelligence, 38
117, 118
swarms, 38
shape, 20, 82, 83, 97, 154
synapse, 84
shaping, 82
synapses, 84
sharing, 38, 39, 141, 142, 144, 145, 146, 147,
synchronization, 144
148, 151, 153, 155, 163, 167, 170
systems, vii, 28, 29, 38, 66, 84, 111, 117, 128,
shear, 82, 116
130, 142, 143, 144, 145, 146, 151, 154, 157,
shear deformation, 82
166, 167, 168, 170, 171, 172, 173
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

sigmoid, 5, 67, 88, 119

sign, 57
signals, 29, 54, 84, 129, 130, 137, 138 T
simulation, 53, 54, 64, 65, 70, 84
simulations, 11 targets, 102
skin, 9 team members, 147
software, 15, 69, 87, 103, 142, 145, 146, 147, technology, 51, 80, 139, 142, 144, 145, 146, 148,
148, 151, 153, 155, 163, 170 155, 170, 171
specific heat, 12, 13 teeth, 35, 40
spectrum, 32, 51, 132 temperature, 7, 9, 10, 27, 56, 63, 68, 112, 117,
speed, 1, 2, 6, 7, 8, 9, 13, 14, 15, 19, 25, 31, 33, 139
34, 35, 38, 41, 43, 44, 47, 48, 52, 54, 65, 66, tension, 9, 25, 29, 56
68, 72, 73, 74, 77, 79, 84, 85, 86, 88, 89, 90, test data, 94
91, 94, 95, 97, 101, 102, 103, 106, 111, 113, thermal energy, 136
114, 116, 119, 120, 122, 123, 125, 126, 127, thinking, 1, 12
130, 132, 134, 135, 136, 138 three-dimensional, 125
spindle, 68, 103, 113, 134, 135 time consuming, 109
SQL, 163 time series, 118
stages, 141, 145, 146, 147, 148, 154, 155, 167 titanium, 80, 81, 97
stainless steel, vii, 65, 66, 68, 70, 77 tolerance, 67
standard deviation, 2 topology, 135
standardization, 164
Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,
 182 Index

torque, 11, 20, 23, 27, 28, 29, 101, 103, 104, 105, variability, 16
107 variables, 3, 4, 6, 7, 8, 9, 10, 11, 13, 15, 17, 21,
total energy, 130, 138 24, 29, 33, 38, 55, 60, 86, 89, 101, 103, 104,
total product, 33 105, 106, 117, 118, 125, 126, 130, 135
tracking, 154 variance, 89, 103
training, 2, 5, 8, 10, 11, 26, 54, 65, 66, 67, 69, 70, variation, 2, 7, 20, 26, 60, 86, 89, 126, 127, 129
71, 77, 83, 84, 85, 88, 91, 94, 95 vector, 6, 21, 85, 89
transfer, 67, 88, 142, 148 velocity, 11, 18, 19, 55, 56, 79, 81, 82, 97, 112
transformation, 79, 97, 144 versatility, 80
transition, 117 vibration, 15, 126
transmission, 107, 143, 166, 167, 168, 169 virtual enterprise, 142
transparency, 1 viscosity, 52
transparent, 16, 26
transportation, 112
travel, 68, 79, 84, 86, 88, 94, 95, 97 W
trial, 70, 71, 84
water, vii, 68, 79, 81, 82, 98, 117
trial and error, 70
water-soluble, 68
tribes, 34
wear, vii, 7, 9, 16, 26, 28, 82, 86, 112, 123, 129,
tungsten, 131
130, 131, 132, 133, 134, 135, 136, 137, 138,
tungsten carbide, 131
139
WHO, 155
U windows, 166
workers, 102, 108
UML, 144, 172 workflow, 141, 144, 147, 148, 155, 157, 158,
uncertainty, 2 159, 160, 163, 164, 168, 169, 170, 172
undergraduate, vii
unification, 145
uniform, 69, 123 X
universe, 12, 117
XML, 171, 173
updating, 55, 67
Copyright © 2009. Nova Science Publishers, Incorporated. All rights reserved.

V Y

yield, 20, 33
validation, 5, 15, 71, 72, 88, 91, 94, 95, 98
validity, 107
values, 3, 6, 7, 8, 11, 12, 15, 20, 21, 33, 36, 37, Z
39, 43, 44, 45, 48, 51, 56, 57, 58, 59, 60, 61,
62, 63, 67, 69, 71, 72, 73, 77, 85, 86, 87, 89, Zn, 103, 131
91, 94, 95, 102, 107, 111, 115, 119, 120, 123,
130, 134, 135, 137

Artificial Intelligence in Manufacturing Research, Nova Science Publishers, Incorporated, 2009. ProQuest Ebook Central,