You are on page 1of 18

The current issue and full text archive of this journal is available at

www.emeraldinsight.com/1741-038X.htm

Back-
Back-propagation artificial neural propagation
network approach for machining ANN approach
centre selection
315
Boppana V. Chowdary
Department of Mechanical and Manufacturing Engineering, Received June 2005
Faculty of Engineering, The University of the West Indies, St Augustine, Revised January 2006
Trinidad and Tobago, West Indies Accepted March 2006

Abstract
Purpose – Traditional machining centre selection methods may not guarantee a cost effective
solution. Properly trained back-propagation artificial neural network (BPANN) tend to select
reasonable machining centres when presented with machining parameters that they have never seen
before. The aim of this paper is to demonstrate the applicability of artificial neural networks (ANNs) to
machine centre selection problems.
Design/methodology/approach – A three-layer feedforward back-propagation supervised
training approach is selected to address the machining centre selection problem and demonstrated
its potential through an example. This is intended to help readers understand implications on
manufacturing system design and future research.
Findings – Very limited studies attempted the machining centre selection problem. Feedforward
ANN approach has been applied to a wide variety of manufacturing problems. Neural networks have
training capability to solve problems that are difficult for conventional computers or human beings.
The developed BPANN model has potential to solve the machine centre selection problem with notable
consistency and reasonable accuracy.
Practical implications – The BPANN model is an innovative approach fundamentally based on
artificial intelligence, which is not directly visible to the user, but is able to solve through a simpler and
supervised feedforward back-propagation training process. The model consists of an input layer,
a hidden layer and an output layer. The 18 neurons fixed in the input layer are same as the set of
machining centre parameters which are taken directly from the machine tool manufacturer’s
catalogues. Evidently the proposed three-layer ANN model has the capability of solving the machine
centre selection problem with three hidden neurons for threshold level of 0.9, noise level of 0.05 and
tolerance of 0.01.
Originality/value – The work size, weight, travel range, spindle speed range, horse power, feed,
accuracy, tool magazine and price are used as machining centre selection parameters. Machining
centres’ information in the form of 24 patterns along with the desired machining centres’ were used to
train and test the network.
Keywords Artificial intelligence, Neural nets, Machining centres, Training, Manufacturing systems
Paper type Research paper

Introduction
Selection of a proper machining centre is an important task in the design of a flexible Journal of Manufacturing Technology
manufacturing system, and thus, a crucial step for facilities planning. Thus, the Management
Vol. 18 No. 3, 2007
pp. 315-332
q Emerald Group Publishing Limited
The author wishes to thank the referees for their detailed comments and constructive criticisms 1741-038X
of the initial draft. DOI 10.1108/17410380710730648
JMTM importance of machining centre selection cannot be overlooked. Addition of a proper
18,3 machining centre in the production system can enhance the effectiveness of the
manufacturing process, provide effective utilization of manpower, increase production,
and improve system flexibility. However, with a wide range of machining centres
available today, selection of an appropriate machining centre alternative for a given
production scenario is not an easy task.
316 Decisions in design and selection of manufacturing systems are affected by the
ongoing development of new production facilities, practices and enhanced capabilities in
the market place and complex characteristics of the product. The machining centre
selection problem influences by operational capabilities in terms of various part design
features, performance measures set by the customer and then cost. The elicited
hierarchy (Wick, 1979; Varma and Kumar, 1989; Peter, 1990; Machinery and Production
Engineering, 1995) in selection of a machining centre is a suitable working length,
working diameter, achievable tolerance, number of axes, number of tool stations, price,
level of available power, speed, and automatic swarf disposal. Generally, selection
in terms of these multiple attributes cannot be easily attained with usual programming
tools. For this reason, the research studies in this field are progressively directed toward
the use of new approaches and methods developed in the AI world: knowledge-based
systems, fuzzy logic, inductive learning, neural networks (NN), and genetic algorithms.
This paper describes the applicability of artificial neural network (ANN) approach
to machining centre selection problem. This tool, in fact, introduces an innovative
approach fundamentally based on a knowledge not directly visible by the user, but
able to be stored through a simpler and more intuitive training process. The reasons for
using NN to solve the machining centre selection problem are: processing speed,
processing order, abundance and complexity, knowledge storage, and processing
control (Arad and El-Awawy, 1997). The paper proceeds as follows: a review of
literature is provided to identify the lack of research to demonstrate the ANN’s
applicability to solve the machining centre selection problem. The machining centre
selection problem is demonstrated in the following section. Then, the results and
discussion are presented. Finally, conclusions are given.

Review of literature
During the last two decades, a number of developments have been reported in
knowledge-based systems, fuzzy logic, inductive learning, NN and genetic algorithms
(Pham and Pham, 2000). ANN, one application of AI, has achieved considerable
success in recent years and opened a new dimension for scientific research and
industrial/business applications. Evidently this approach has been applied to several
areas in engineering and manufacturing. With the ready availability of high memory
and affordable computers now, there is a considerable potential for the extension
of ANN approach to solve challenging problems like machining centre selection.
The purpose of this section here is to describe the background principles of ANN and
to document some applications of ANN to manufacturing.

Artificial neural networks


The ANN is an information processing system comprising of a collection of simple
processing units, termed as nodes or neurons, connected together by links with varying
strengths or weights. Links are usually unidirectional, as in the case of biological
systems (Negnevitsky, 2005). The strengths of the links in the network are Back-
predetermined fixed values, which usually set through a supervised back-propagation propagation
training algorithm. Processing units have a state of activation, which varies with time.
Each processing unit has an output function, which maps infinite domain - the input, to ANN approach
a pre-specified range - the output.
In ANNs the propagation and the activation rules specify how the inputs to the
processing elements are combined to produce its activation level (Kronsjo and 317
Shumsheruddin, 1992). The activation of a node is the sum of the products of its inputs
and the weight of their links, which is given as in equation (1).
Ai ðt þ 1Þ ¼ SAj wij ð1Þ
Where Ai is the activity of the ith node and wij is the weight of the link from node j
to node i.
During training the learning rule specifies how the weights of the connections in the
network are to be adjusted. The weights are usually adjusted in a large number of
small steps. This can be done through the following equation:
Dwij ¼ hðT i 2 Ai ÞAj ð2Þ
Where wij is the change to be made to wij during one learning step and h is a learning
rate constant and Ti is the target value for the activity of the ith node.
The nodes of an ANN can be arranged in three or more layers. There may be an
arbitrary number of nodes in each layer. The first/input layer could receive data
normalized between 0 and 1. There are also connections from each node in the
second/hidden layer to all nodes in the third/output layer. The nodes in the second and
third layers have internal threshold associated with them. During normal operation,
the input patterns are used to set the activation level of the input nodes. Activity is then
propagated from the input nodes to the hidden nodes. Finally activity is propagated
from the hidden nodes to the output nodes. Propagation takes place according to the
equations (3) and (4).
neti ¼ aj wij ð3Þ
ai ¼ sðneti 2 ui Þ ð4Þ
Where neti is the net input to the ith node in the current layer, n is the number of nodes
in the previous layer and wij is the strength of the connection from the jth node, ai is the
activity of the ith node, ui is its threshold and s is sigmoid function:
sðaÞ ¼ 1=ð1 þ e 2a Þ ð5Þ
In ANNs, a sigmoid function is used to set the output value for any given neuron by
processing the sum of the weighted input values and any bias applied. The function
outputs a value that is close to zero for a low total input value and close to one for a high
input value. The slope of the function curve can be adjusted by including a threshold
value that can make the “step” between zero and one (Negnevitsky, 2005).
To train the network, the thresholds of the nodes and the strengths of the links are
initially set to small random values. Then the network is trained with a set of training
cases. Each case contains an input pattern and the corresponding desired output pattern.
JMTM The training for the complete set of training cases is continued to achieve good
18,3 performance. Then activity is propagated from the hidden layer to the output layer.
The actual activities of the output nodes are then compared with the target activities in
the training case and error will be computed. Then the weights of the links from the
hidden layer to the output layer are adjusted according to the following rule:

318 Dwij ¼ hdi aj ð6Þ


Where wij is the change to be made in the weight of the link from the jth node to the ith
node, h is a learning rate constant between zero and one, aj is the activity of the jth node
and di is the error in node “i ” in the output layer and can be obtained from equation (7).
di ¼ ai ð1 2 ai Þðt i 2 ai Þ ð7Þ
Where ai is the actual activity of the ith node and ti is the target activity for it.
Also the weights of the links from the input layer to the hidden layer are adjusted
according to the rule as given in equation (6). For the hidden layer the values of di are
calculated by propagating the error signals backwards from the output layer according
to the equation (8).
di ¼ ai ð1 2 ai ÞSk dk wik ð8Þ
Where dk is the value of di for the kth output node.
Properly trained back propagation networks tend to give reasonable answers when
presented with inputs that they have never seen. The next section deals with some
significant developments in ANN’s applications to manufacturing.

ANN applications to manufacturing


Although ANNs have been introduced for several decades, their use in manufacturing
area is quite recent and the applications to manufacturing problems are still very few. In
the field of process planning, some studies have been carried out to test the behaviour of
ANNs in solving operations sequencing problem. Study of Dong et al. (1995), illustrates
the potential of ANN approach to solve the sequencing problem. A feedforward neural
network was designed for each form feature in order to select the best sequence among a
set of previously classified sequences. Another study (LeTumelin et al., 1995) attempted
the detection of the appropriate sequence of operations for machining holes. Sakakura
and Inasaki (1992) described an application of ANN for selection of grinding parameters.
They developed a three-layer feedforward model to simulate the values of depth of cut
and feed to be used in grinding in order to obtain a given surface roughness. Li et al.
(1994) proposed an ANN model for selection of grinding wheel selection. The proposed
network has five input neurons that correspond to the “type of machine” the “work piece
material” the “hardness” the “surface roughness” and the “severity” of the grinding
operation; the output vector is formed by four elements, which represent the “abrasive
type” “grade” “grit size” and “bond”. It is remarkable that the network was trained using
a range of data directly taken from the manufacturer’s catalogue. Further applications of
ANN to manufacturing related problems are evidenced with selection of parameters of
cutting tools (Santochi and Dini, 1996), scheduling (Sabuncuoglu and Gurgun, 1996) and
lot sizing (Gaafar and Choueiki, 2000). The analysis of literature has revealed that the
ANN approach has been applied to a wide variety of manufacturing problems.
Very few attempts are aimed to solve machining centre selection problem. Some of Back-
the studies include Gopalakrishnan et al. (2004)’s object-oriented computer-based propagation
software program, which is the prototype of decision support system (DSS) framework.
They focused primarily on selection of vertical and horizontal machining centres with ANN approach
various options include high speed, productivity, machining complexity and
machining accuracy to facilitate and satisfy production requirements. Arslan et al.
(2004) also proposed a DSS model for selection of knowledge intensive machining 319
centres. Multi-criteria weighted average was used in their decision-making process to
rank the machines evaluated with respect to several criteria such as productivity,
flexibility, space, adaptability, precision, cost, reliability, safety and environment and
maintenance and service. It is evident that although DSS models have been developed
to address the machine tool selection problem, a specific AI-based NN approach
towards machining centre selection is needed since the NN can have capability to solve
problems that are difficult for human beings. This paper attempts to show the
potentiality of ANN to machining centre selection problem through an illustration.

Machining centre selection problem


With more and more manufacturing industries switching to computer-aided
manufacturing, the focus is now on the integration of processing stations with
computer control and high levels of automation. However, small and medium-sized
firms still essentially rely on an individual experience. For instance, a plant engineer is
usually responsible for facility layout, and the design of workstations. For many such
engineers, these tasks represent only a small portion of their numerous and diverse
responsibilities, so they may not develop the expertise that is needed today for solving
complex material processing problems. Traditionally, designers are faced with three
choices for selection of a right machining centre:
(1) relying on published literature such as books and articles and on personal
experiences;
(2) relying on the experience of machining centre suppliers; and
(3) appointing a consultant.

Designers relying on personal experience tend to select the machining centres with
which they are most familiar, however, the choice may not be the cost-effective
machining centre. Machining centre vendors have an inherent interest in selling their
machinery, so their recommendations might sometimes be biased. On the other hand,
consultants often charge substantial amounts for their recommendations and for the
time they spend on evaluations. Thus, these traditional machining centre selection
methods/options may not guarantee a cost-effective solution. The study of Gerrard
(1988) reveals that the role of engineering staff in authorization for final selection of
machining centre is 6 per cent, the rest belongs to upper and middle management. It is
evident that there is scope to apply ANN approach to machining centre selection
problem due to the following:
.
Availability of limited traditional tools to assist the facilities design engineer.
. Attempts have been made to apply ANN approach to several manufacturing
applications such as process planning, product design and analysis, process and
JMTM machine diagnosis, analysis of grinding operations and machine maintenance
18,3 analysis.
.
Time consuming and knowledge intensive iterative process.

The ANN model developed in this paper involves understanding of the key machining
centre options include job capacity, travel range, tool changer capacity, tool change
320 time, spindle speed and power, feed, accuracy, type of control system, and price.
The major steps of the proposed approach are:
(1) construction of a three-layer feedforward back-propagation artificial neural
network (BPANN) model;
(2) preparation of the training and testing patterns;
(3) supervised-learning and feedforward back-propagation training; and
(4) testing of the model output.

These steps are explained further in the following sections.

Construction of a three-layer feedforward BPANN model


The BPANN model used to select the machining centre consists of an input layer,
a hidden layer and an output layer, as shown in Figure 1. Each neuron or node in the

Work
Length Width Type of Control System Price

Input Layer …

Weights

Hidden Layer …

Weights

Figure 1. Output Layer


Feedforward three-layer
BPANN model for
machining centre selection
0 or 1 0 or 1 0 or 1 0 or 1
input layer contains known information. The number of neurons ( ¼ 18) in the input Back-
layer is same as the set of machining centre parameters identified for the selection of propagation
suitable machining centre as given in Table I. Lines represent weights connect the
input layer to the hidden layer. These weights are used to model the synaptic strength ANN approach
between neurons in the human brain, thus portraying the importance of one neuron’s
effect on the other of next layer. The hidden layer is also made-up of neurons, whose
number is fixed on the basis of trial and error method, which is explained in the 321
following sections. The output layer consists of four neurons only because four
machining centres are considered in the study for the purpose of illustration. An output
node is used to represent each of the closest possible machining centres. When the
network is working correctly, the output node corresponding to the current machining
centre will have the values as given in Table I. The activation of the neuron is
computed by applying a sigmoid function.

Preparation of the training and testing patterns


Training patterns refers to the input data coupled with its desired outputs were
prepared from the literature (Wick, 1979; Varma and Kumar, 1989; Peter, 1990;
Machinery and Production Engineering, 1995). Generally, the selected training
patterns should be sufficient enough to cover the chosen characteristics of a typical
machining centre. Data of four machining centres information in the form of 24 patterns
along with the desired machining centres were used for training the network and is
presented in Tables II and III. When testing the network, various input data sets are

Input neuron (Ni¼ 1-18) Machining centre parameter

1 Work length
2 Width
3 Thickness
4 Weight
5 Travel range in x-direction
6 Travel range in y-direction
7 Travel range in z-direction
8 Number of cutting tools
9 Maximum diameter of cutting tool
10 Maximum length of cutting tool
11 Tool change time
12 Spindle power
13 Spindle rpm
14 Feed rate
15 Accuracy (positional)
16 Accuracy (repeatability)
17 Type of control system
18 Price
Output neuron (Oi ¼ 1-4) Machining centre
Study code Type Table I.
1 MC01 Vertical CNC Machining centre
2 MC02 Horizontal CNC parameters, input and
3 MC03 Vertical CNC output neurons selected
4 MC04 Vertical CNC in the study
18,3

322

selection
Table II.
JMTM

for machining centre


Training patterns used
Travel range
Work capacity (inches) Tool changer capacity Spindle
Size Weight Tool diameter Tool length Tool change time
Pattern No. (inches) (pounds) x y z Number of tools (inches) (inches) (seconds) Hp Rpm

1 50 £ 33 £ 30 6,200 30 19 24 35 05.0 17.5 02 20 10-3,000


2 60 £ 43 £ 40 7,000 40 29 34 45 15.0 17.5 10 20 10-3,000
3 55 £ 38 £ 35 6,600 35 24 29 40 10.0 17.5 06 20 10-3,000
4 70 £ 53 £ 50 8,000 50 39 44 50 20.0 17.5 12 20 10-3,000
5 40 £ 23 £ 20 5,200 25 15 20 30 10.0 17.5 06 20 10-3,000
6 45 £ 28 £ 25 5,600 30 20 15 30 10.0 17.5 10 15 10-3,000
7 72 £ 50.5 £ 52 9,200 52 52 29 45 15.0 17.5 10 20 10-3,000
8 62 £ 40.5 £ 42 8,400 42 42 19 35 05.0 17.5 02 20 10-3,000
9 67 £ 45.5 £ 47 8,800 47 47 24 40 10.0 17.5 06 20 10-3,000
10 57 £ 35.5 £ 37 7,800 37 37 12 30 05.0 17.5 02 17 05-2,000
11 55 £ 40.5 £ 42 8,000 40 40 24 35 10.0 17.5 08 15 05-2,000
12 77 £ 60.5 £ 57 9,900 60 60 35 50 15.0 17.5 15 25 10-5,000
13 17 £ 29 £ 23 0500 29 17 07 29 06.0 12.0 14 02 60-2,000
14 07 £ 19 £ 13 0100 19 05 03 19 02.0 12.0 06 02 60-2,000
15 12 £ 24 £ 18 0300 24 12 05 24 04.0 12.0 10 02 60-2,000
16 15 £ 15 £ 12 0500 25 15 07 25 05.0 12.0 15 02 10-2,000
17 27 £ 20 £ 15 1,000 30 20 10 30 10.0 15.0 10 02 60-2,000
18 15 £ 20 £ 12 0550 24 14 07 30 05.0 12.0 15 02 10-2,000
19 31 £ 47 £ 31 3,400 45 30 26.6 29 13.5 15.0 14 7.5 20-4,000
20 21 £ 37 £ 21 2,600 35 20 16.6 19 03.5 15.0 06 7.5 20-4,000
21 26 £ 42 £ 26 3,000 40 25 21.6 24 08.5 15.0 10 7.5 20-4,000
22 25 £ 42 £ 25 3,000 35 30 25 25 13.0 12.0 15 10 20-4,000
23 20 £ 30 £ 20 2,600 35 20 20 20 05.0 12.0 10 10 20-4,000
24 31 £ 42 £ 31 3,500 40 40 25 30 10.0 12.0 10 10 20-4,000
Desired output neuron Desired
Accuracy (inches) values machining centre
Pattern No. Feed (ipm) Positional Repeatability Control system Price (US $) 01 02 03 04 Code

1 0-240 0.00015 0.00015 CNC 434,600 0 0 0 1 MC01


2 0-240 0.00025 0.00095 CNC 444,600 0 0 0 1 MC01
3 0-240 0.00020 0.00010 CNC 439,600 0 0 0 1 MC01
4 0-240 0.00030 0.00030 CNC 460,000 0 0 0 1 MC01
5 0-240 0.00010 0.00010 CNC 400,000 0 0 0 1 MC01
6 0-240 0.00015 0.00005 CNC 429,000 0 0 0 1 MC01
7 0-240 0.00025 0.00095 NC 483,000 0 0 1 0 MC02
8 0-240 0.00015 0.00015 NC 473,000 0 0 1 0 MC02
9 0-240 0.00020 0.00010 NC 478,000 0 0 1 0 MC02
10 0-140 0.00012 0.00012 NC 350,000 0 0 1 0 MC02
11 0-140 0.00015 0.00008 NC 408,000 0 0 1 0 MC02
12 0-500 0.00050 0.00095 NC 525,000 0 0 1 0 MC02
13 0.5-300 ^0.00010 ^0.00030 CNC 064,400 0 1 0 0 MC03
14 0.5-300 ^0.00010 ^0.00030 CNC 054,400 0 1 0 0 MC03
15 0.5-300 ^0.00010 ^0.00030 CNC 059,400 0 1 0 0 MC03
16 0.5-300 ^0.00010 ^0.00030 CNC 065,400 0 1 0 0 MC03
17 0.5-300 ^0.00010 ^0.00030 CNC 075,000 0 1 0 0 MC03
18 0.5-300 ^0.00010 ^0.00030 CNC 060,500 0 1 0 0 MC03
19 0.01-400 ^0.00010 ^0.00050 CNC 135,000 1 0 0 0 MC04
20 0.01-400 ^0.00010 ^0.00050 CNC 125,000 1 0 0 0 MC04
21 0.01-400 ^0.00010 ^0.00050 CNC 130,000 1 0 0 0 MC04
22 0.01-400 ^0.00005 ^0.00045 CNC 115,000 1 0 0 0 MC04
23 0.01-400 ^0.00005 ^0.00045 CNC 135,000 1 0 0 0 MC04
24 0.01-400 ^0.00005 ^0.00045 CNC 150,000 1 0 0 0 MC04
Notes: 01, 02, 03 and 04 – output neurons; MC01, MC02, MC03 and MC04 – machining centres
ANN approach
propagation

for machining centre


Back-

selection
Training patterns used
Table III.
323
JMTM applied to its input layer. The network generates the output. This output is used to
18,3 determine the network performance and no back-propagation is carried out. The
network’s output is compared with the desired outputs. These input data sets and the
desired outputs are collectively known as test patterns. The input values have been
coded with numerical values ranging between 0 and 1 in order to give a contribution to
the network independent from their real absolute values.
324
Supervised learning and back-propagation training
Supervised learning is a process that incorporates an external teacher and/or global
information. This process decides when to turn off the learning, how long and how often to
present each association for training, and it also supplies performance error information.
Supervised learning and back-propagation training algorithm proposed by Simpson
(1989) has been used in the study to adjust the weights of the network systematically such
that the error between the output and the corresponding desired output is minimum. Also
it maps new, never-before seen information (i.e. customer requirements for machining
centre) entering the input layer to the nearest machining centre at the output layer.
The actual training of the network is accomplished by “back-propagating” the error from
the output layer to the hidden layer and finally to the input layer. The error is simply the
difference between the desired output and the output calculated during training.

Stability and system optimality


As the proposed BPANN has one hidden layer and the number of neurons in the hidden
layer is initially set at 10. This number is arbitrarily fixed since there are no set rules to
decide how many neurons should be used in a hidden layer (Shanker et al., 1996). However,
there are a minimum number of hidden neurons that a network should have. A network
with fewer hidden neurons will not train. This minimum number of hidden neurons is
considered a factor for system optimality, because in the case of hardware implementation
of the network, a smaller number of neurons would mean less hardware.
The back-propagation algorithm selected in this study is a gradient descent type
(Simpson, 1989), in which the network weights are moved along the negative of the
gradient of the performance function in an iterative manner. The weights involved in
the network are adjusted in each of the iterations so as to reduce the error along a
descent direction. In doing so, two parameters, called learning rate (h) and momentum
factor (mf), are introduced in the literature (Yu and Chen, 1997) for controlling the size
of weight adjustment along the descent direction and for dampening oscillations of the
iterations. These two parameters have to be empirically chosen in the conventional
back-propagation algorithm. In general, the “mf” should be less than unity to stabilize
the behaviour of the algorithm. The selection of learning rate is also more arbitrary
due to the fact that the error surface usually consists of many flat and steep regions
and behaves quite differently from application to application (Yu and Chen, 1997). The
learning rate is set at 1.0 for the default network. A large learning rate is helpful for
accelerated learning, when the weight search crosses a plateau. It can be set in between
0.0 and 4.0 (Chaudhuri and Sequeira, 1992). A value 0.0 means that the weights will not
change during training. As a result, the network will not train. A value 4.0 provides the
network with large changes in weight values, which would mean fast training time.
Therefore, an efficient back-propagation algorithm should be capable of dynamically
varying its learning rate and “mf”.
Stability Back-
The initial network with 18 input neurons, ten hidden neurons and four output neurons propagation
is called the default network. The tolerance used for the default network is 0.10. For
example, if a vector input (in terms of customer specifications like job size, weight, ANN approach
travel range, feed, and speed range) corresponds to a desired output of 0 0 0 1
(machining centre1) and the network calculated an output of 0 0 0 (0.9 to 1.1), the
training information is considered learnt. An output below 0.9 and above 1.1 means 325
the network has not yet learnt that particular training information. Reality is that the
network has to learn all the training data before it is considered trained. Thus, training
tolerance must be considered in training the data, which translates to the minimum
error allowed. There is a minimum training tolerance below which a network will fail
to learn or converge. Also adding of the input noise to the input vectors causes the
network for variations in training time. However, it helps the network to generalize and
resulting network is a better predictor. For the default network, the input noise is set
at 0.0.

System optimality
Once the default NN is trained, its parameters need to be optimized to yield an
optimum performance. The changes made in the network parameters such hidden
neurons, input noise, learning rate and tolerance and the corresponding results
obtained are shown in Table IV. The first step towards system optimality is changing
the number of hidden neurons. Changing the number of hidden neurons subsequently
changes the number of total weights. Hence, they have to be reinitialized and the new
network has to be trained. If the new network is trained and the hidden neurons are
further reduced then the network needs to be retrained with the reduced number of
hidden neurons. In this way an initial number of 10 in the default network is reduced to
three hidden neurons. Next the input noise and the back-propagation learning rate are
varied and experimentation continued. During the investigation, the optimal network
was found at the noise level of 0.05 and minimum tolerance of 0.07.

Testing of the model output


The flow chart for machining centre selection shown in Figure 2 has been used while
developing the BPANN model in “C” language. The trained network for machining
centre selection has been simulated and tested for its validity. While testing the neural

Hidden neurons Input noise Tolerance Learning rate Iterations

10 default network 0.00 0.10 1 98


05 0.00 0.10 1 81
04 0.00 0.10 1 78
03 0.00 0.10 1 95
02 0.00 0.10 1 Does not train
01 0.00 0.10 1 Does not train
03 0.00 0.07 2 385
03 0.05 0.09 2 2,161
03 0.05 0.08 2 1,230 Table IV.
03 optimum network 0.05 0.07 2 1,043 BPANN model optimality
03 0.05 0.06 2 Does not train results
JMTM Training Patterns
18,3
Network Training and
Validation
Increase Network
326 Complexity

(Desired Output
N
Increase – Actual Output) ≤
Training Patterns Specified Accuracy

Yes

Feed the Test Patterns and


Validation

No (Desired Output
Yes
Figure 2. – Actual Output) ≤ Stop
Flow chart for machining Specified Accuracy
centre selection using
BPANN approach

network, various patterns (Table II), have been applied to the network. The output
generated from the neural network model is compared to the desired output. The
network output and the desired output against the respective pattern numbers are listed.

Results and discussion


In order to validate the capability of the trained neural net-work, 24 sets of test data are
used (Tables V and VI). Based on the results some critical observations can be drawn.
Input neurons were set at an approximate activity levels in between 0 and 1 through
normalization procedure to indicate the customer desired machining centre
specifications. Four output neurons were used to select the possible or approximate
machining centre in this study. When the model is generating accurate output, the
output neuron corresponding to the correct machining centre will have an activity of 1.0
and all others will have any value in between 0 and 1.0, however, it depends on the noise
level added to the input data and the set threshold levels. For two different combinations
of noise and threshold levels the results were obtained and presented in Tables VII and
VIII. Among the 24 test patterns, except in three situations (i.e. in pattern number 3, 9,
and 17) for threshold level of 0.9, mf of 0.9, learning rate of 2, tolerance of 0.01 and without
any noise, the generated network output is coinciding with the desired output
(Table VIII). In all the three cases, the desired machining centre is MC01. This shows that
the BPANN has to be validated on a set of data, which has not been used in the training
process, particularly in the case of pattern numbers 3, 9 and 17. If the trained network
Travel range
Work capacity (inches) Tool changer capacity Spindle
Pattern Weight Number of Tool diameter Tool length Tool change time
No. Size (inches) (pounds) x y z tools (inches) (inches) (seconds) Hp Rpm

1 62 £ 40.5 £ 42 8400 42 42 19 35 5 12.5 2 15 10-2,000


2 72 £ 50.5 £ 52 9,200 52 52 29 45 15 22.5 10 25 10-4,000
3 50 £ 33 £ 30 6,200 30 19 24 35 5 12.5 2 15 10-2,000
4 65 £ 50 £ 45 7,900 50 39 44 55 22 16.5 14 20 10-3,000
5 40 £ 22 £ 19 5,150 27 14 21 28 09 17.5 06 20 10-3,000
6 42 £ 26 £ 24 5,400 29 19.5 14 33 11 15.5 09 17 10-3,000
7 73 £ 45 £ 51 9,000 49 52 26 42 14 20.0 9 26 10-4,000
8 17 £ 29 £ 23 500 29 17 7 29 6 12.0 14 3 14-2,400
9 60 £ 43 £ 40 7,000 40 29 34 45 15 22.5 10 25 10-4,000
10 55 £ 33.5 £ 35 7,770 37 40 13 30 05 18.5 02 17 5-2,000
11 54 £ 39.5 £ 41 7,940 38 38 25 32 11 17.5 10 13 5-2,000
12 75 £ 60 £ 56 9,820 58 58 33 55 16 17 13 27 10-5,000
13 21 £ 37 £ 21 2,600 35 20 16.6 19 3.5 10 6 5 35-3,900
14 7 £ 19 £ 13 100 19 5 3 19 2 7.0 6 6 60-1,900
15 31 £ 47 £ 31 3,400 45 30 26.6 29 13.5 20 14 14 60-2,000
16 15 £ 14.5 £ 11 490 23 16 10 25 05 11 16 02 10-2,000
17 26 £ 21 £ 15.5 0955 30 20 10 33 09 15.5 11 02 60-2,000
18 14.5 £ 20.5 £ 11.5 0550 24.5 14.5 08 27 05 12 16 02 10-2,000
19 52 £ 40 £ 37 6,800 36 26 31 42 11.0 15.5 7 7.5 80-2,000
20 18 £ 28 £ 22 400 29 16 8 25 6 11 15 20 20-4,000
21 29 £ 46 £ 30 3,450 44 29 27 26 13 18 13 20 25-4,000
22 24 £ 42 £ 23.5 2,915 36.5 33 27.5 26 11 10 16 09 20-4,000
23 19 £ 29.5 £ 19 2,500 37 20 20 19 05 11 11 10 20-4,000
24 30 £ 41 £ 30.5 3,400 44 40 26.5 32 11 13 09 08 20-4,000
ANN approach

User selected
propagation

specifications/test data
Back-

Table V.
327
JMTM
Pattern No. Feed (ipm) Accuracy (inches) Control system Price (US $)
18,3
1 0-200 0.00022 0.00011 CNC 440,960
2 0-240 0.00025 0.00095 CNC 65,000
3 0-230 0.00020 0.00010 CNC 140,000
4 0-200 0.00022 0.00011 CNC 445,000
328 5 0-240 0.00025 0.00095 CNC 405,000
6 0-230 0.00020 0.00010 CNC 425,000
7 0-220 0.000215 0.00020 NC 482,000
8 0-210 0.00026 0.00090 NC 475,000
9 0-240 0.00020 0.00010 NC 65,000
10 0-220 0.000215 0.00020 NC 360,000
11 0-210 0.00026 0.00090 NC 408,000
12 0-240 0.00020 0.00010 NC 500,000
13 0.5-300 ^0.0010 ^0.0003 CNC 445,000
14 0.5-290 ^0.0010 ^0.0003 CNC 124,000
15 0.5-250 ^0.0010 ^0.0003 CNC 130,000
16 0.5-300 ^0.0010 ^0.0003 CNC 650,000
17 0.5-290 ^0.0010 ^0.0003 CNC 745,000
18 0.5-250 ^0.0010 ^0.0003 CNC 060,400
19 0.01-380 ^0.00024 ^0.00014 CNC 438,000
20 0.01-400 ^0.00026 ^0.0009 CNC 475,000
21 0.5-300 ^0.00010 ^0.0003 CNC 058,500
Table VI. 22 0.01-380 ^0.00024 ^0.00014 CNC 115,000
User selected 23 0.01-400 ^0.00026 ^0.0009 CNC 132,000
specifications/test data 24 0.5-300 ^0.00010 ^0.0003 CNC 150,000

performs poorly on the validating set, it is assumed that there is some important
information in the validating data set, which the network was unable to learn. The
patterns with large deviations against their desired output are then added to the training
set and the network is retrained using the new set of training patterns. This process of
training and validation is repeated until the performance of the trained network on the
validating set is acceptable.

Conclusions and further research


Selecting a suitable machining centre for a given set of machining centre features is
a complex problem since it can be affected by many factors. Identifying the
correlation between selecting an appropriate machining centre in accordance with
its varying parameters is the key problem to be resolved. After reviewing the
literature, the work size, weight, travel range, spindle speed range, horse power,
feed, accuracy, tool magazine and price are chosen as machining centre selection
parameters to test the potential of the BPANN model. Data of four machining
centres’ information in the form of 24 patterns along with the desired machining
centres’ were used to achieve the optimum neural network parameters. Results show
that the developed ANN model has the capability of solving the machine centre
selection problem with three hidden neurons for threshold level of 0.9, noise level of
0.05 and tolerance of 0.01. When testing the network, 24 input data sets are used.
Except in three data sets, the network displayed a robust performance for threshold
Back-
Suggested machining centre
Pattern No. Network output Desired output Network output Desired output propagation
1 0 0 1 0 0 0 1 0 MC02 MC02
ANN approach
2 0 0 1 0 0 0 1 0 MC02 MC02
3 0 0 0 1 0 0 0 1 MC01 MC01
4 0 0 0 1 0 0 0 1 MC01 MC01 329
5 0 0 1 0 0 0 1 0 MC02 MC02
6 0 0 1 0 0 0 1 0 MC02 MC02
7 0 1 0 0 0 1 0 0 MC03 MC03
8 0 1 0 0 0 1 0 0 MC03 MC03
9 0 0 0 1 0 0 0 1 MC01 MC01
10 0 0 0 1 0 0 0 1 MC01 MC01
11 1 0 0 0 1 0 0 0 MC04 MC04
12 1 0 0 0 1 0 0 0 MC04 MC04
13 0 1 0 0 0 1 0 0 MC03 MC03
14 0 1 0 0 0 1 0 0 MC03 MC03
15 1 0 0 0 1 0 0 0 MC04 MC04
16 1 0 0 0 1 0 0 0 MC04 MC04
17 0 0 0 1 0 0 0 1 MC01 MC01 Table VII.
18 0 0 1 0 0 0 1 0 MC02 MC02 Results generated for
19 0 0 0 1 0 0 0 1 MC01 MC01 BPANN model for
20 0 1 0 0 0 1 0 0 MC03 MC03 threshold level of 0.9 and
21 0 0 1 0 0 0 1 0 MC02 MC02 noise level of 0.05
22 0 1 0 0 0 1 0 0 MC03 MC03 (mf ¼ 0.9,
23 1 0 0 0 1 0 0 0 MC04 MC04 Tolerance ¼ 0.01, and
24 1 0 0 0 1 0 0 0 MC04 MC04 learning rate ¼ 2)

level of 0.9, mf of 0.9, learning rate of 2, tolerance of 0.01 and without adding any
noise to it (Table VIII). This outcome tells that the NN needs to be validated with
some additional machining centre information, which has not been used in the
earlier training process. Thus, the study has shown that a properly developed ANN
model provides a valid alternative for solving the machining centre selection
problem.
The BPANN model demonstrated in this paper is an innovative approach
fundamentally based on AI which is not directly visible to the user, but able to solve
through a simpler and supervised feedforward back-propagation training process.
This was carried out through training examples directly taken from the machine tool
manufacturer’s catalogues. The advantages of this approach, in comparison with other
programming methods, are:
.
there is no need to know the explicit function for selecting a suitable machining
centre for a user set specifications; and
.
capable of accomplishing the selection of machining centres in a short computing
time.

The potential of ANN as a machining centre selection tool is usually based on a large
sample size. The generalization capabilities of ANNs are highly dependent on the
number of patterns in the training set. Hence, there is scope to train and test
the BPANN model based on further larger sample size in the future. Also it could be
JMTM
Suggested machining centre
18,3 Pattern No. Network output Desired output Network output Desired output

1 0 0 1 0 0 0 1 0 MC02 MC02
2 0 0 1 0 0 0 1 0 MC02 MC02
3 0 1 0 0 0 0 0 1 MC03 MC01
330 4 0 0 0 1 0 0 0 1 MC01 MC01
5 0 0 1 0 0 0 1 0 MC02 MC02
6 0 0 1 0 0 0 1 0 MC02 MC02
7 0 1 0 0 0 1 0 0 MC03 MC03
8 0 1 0 0 0 1 0 0 MC03 MC03
9 0 1 0 0 0 0 0 1 MC03 MC01
10 0 0 0 1 0 0 0 1 MC01 MC01
11 1 0 0 0 1 0 0 0 MC04 MC04
12 1 0 0 0 1 0 0 0 MC04 MC04
13 0 1 0 0 0 1 0 0 MC03 MC03
14 0 1 0 0 0 1 0 0 MC03 MC03
15 1 0 0 0 1 0 0 0 MC04 MC04
16 1 0 0 0 1 0 0 0 MC04 MC04
Table VIII. 17 0 1 0 0 0 0 0 1 MC03 MC01
Results generated for 18 0 0 1 0 0 0 1 0 MC02 MC02
BPANN model for 19 0 0 0 1 0 0 0 1 MC01 MC01
threshold level of 0.9 and 20 0 1 0 0 0 1 0 0 MC03 MC03
noise level of 0.0 21 0 0 1 0 0 0 1 0 MC02 MC02
(mf ¼ 0.9, 22 0 1 0 0 0 1 0 0 MC03 MC03
tolerance ¼ 0.01, and 23 1 0 0 0 1 0 0 0 MC04 MC04
learning rate ¼ 2) 24 1 0 0 0 1 0 0 0 MC04 MC04

interesting to the researchers to compare the performance of ANN approach with other
association rules or decision tree models (e.g. classification and regression tree)
especially to examine whether ANN approach has any superiority in solving machine
centre selection problem.
With respect to the other modelling tools for machining centre selection, IF-THEN
rules often used in knowledge-based expert systems have the remarkable advantage of
representing the knowledge with simple and independent structures. Continuing
research is now being directed towards the development of knowledge-based neural
system to demonstrate the potential of machining centre selection using several rules
such as flexibility, quality and productivity.

References
Arad, B.S. and El-Awawy, A. (1997), “On fault training of feedforward neural networks”, Neural
Networks, Vol. 10 No. 3, pp. 539-53.
Arslan, M.C., Çatay, B. and Budak, E. (2004), “A decision support system for machine tool
selection”, Journal of Manufacturing Technology, Vol. 15 No. 1, pp. 101-9.
Chaudhuri, S.P. and Sequeira, C. (1992), “Neural networks for automatic target recognition”,
working paper, Electrical Engineering Department, University of Lowell, Lowell, MA,
pp. 1-10.
Dong, J.X., Tang, X.Q. and Wang, S.C. (1995), “Inference mechanism for CAPP tool based on Back-
artificial neural network”, Proceedings of 1st Congress on Intelligent Manufacturing,
Puerto Rico, pp. 994-1002.
propagation
Gaafar, L.K. and Choueiki, M.H. (2000), “A neural network model for solving the lot-sizing ANN approach
problem”, Omega, Vol. 28, pp. 175-84.
Gerrard, W. (1988), “Selection procedures adopted by industry for introducing new machine
tools”, in Worthington, B. (Ed.), Advances in Manufacturing Technology, 331
Proceedings, Fourth National Conference on Production Research, Kogan Page, London,
pp. 525-31.
Gopalakrishnan, B., Yoshii, T. and Dappili, S.M. (2004), “Decision support system for machining
centre selection”, Journal of Manufacturing Technology, Vol. 15 No. 2, pp. 114-54.
Kronsjo, L. and Shumsheruddin, D. (1992), Advances in Parallel Algorithms, Wiley, New York,
NY, pp. 66-85.
LeTumelin, C., Garro, O. and Charpentier, P. (1995), “Generating process plans using neural
networks”, Proceedings of Second International Workshop on Learning in Intelligent
Manufacturing Systems, Budapest.
Li, Y., Mills, B., Moruzzi, J.L. and Rowe, W.B. (1994), “Grinding wheel selection using a neural
network”, Proceedings of 10th National Manufacturing Research Conference
Loughborough, pp. 597-601.
Machinery and Production Engineering (1995), “Machining centres - technology update”,
Machinery and Production Engineering, Vol. 153, pp. 45-6.
Negnevitsky, M. (2005), Artificial Intelligence – A Guide to Intelligent Systems, 2nd ed.,
Addison-Wesley/Pearson Education Limited, Harlow.
Peter, K. (1990), “Machining centres”, Machinery and Production Engineering, Vol. 148, pp. 81-90.
Pham, D.T. and Pham, P.T.N. (2000), Computational Intelligence for Manufacturing,
Computational Intelligence in Manufacturing Handbook, CRC Press, New York, NY.
Sabuncuoglu, I. and Gurgun, B. (1996), “A neural network model for scheduling problems”,
European Journal of Operational Research, Vol. 93, pp. 288-99.
Sakakura, M. and Inasaki, I. (1992), “A neural network approach to the decision-making for
grinding operations”, Ann. CIRP, Vol. 41 No. 1, pp. 3453-6.
Santochi, M. and Dini, G. (1996), “Use of neural networks in automated selection of technological
parameters of cutting tools”, Computer Integrated Manufacturing Systems, Vol. 9 No. 3,
pp. 137-48.
Shanker, M., Hu, M. and Hung, M. (1996), “Effect of data standardization on neural
network training”, Omega International Journal of Management Science, Vol. 24 No. 4,
pp. 385-97.
Simpson, P.K. (1989), Artificial Neural Systems Paradigms and their Applications and
Implementations, Pergamon Press, London.
Varma, A. and Kumar, S. (1989), “Knowledgeable expert system for flexile manufacturing”,
Proceedings of the 4th International Conference on CAD, CAM, Robotics and Factories of
the Future, Tata McGraw-Hill, New Delhi, pp. 603-18.
Wick, C. (1979), “Machining center update”, Manufacturing Engineering, September, pp. 76-94.
Yu, X.-H. and Chen, G.-U. (1997), “Efficient back propagation learning using optimal learning
rate and momentum”, Neural Networks, Vol. 10 No. 3, pp. 517-27.
JMTM Further reading
18,3 Leshno, M., Lin, V., Pinkus, A. and Schocken, S. (1993), “Multilayer feedforward neural
networks with a non-polynomial activation function can approximate any function”,
Neural Networks, Vol. 6, pp. 861-7.
Sprecher, D. (1993), “A universal mapping for kolmogorov’s superposition theorem”, Neural
Networks, Vol. 6, pp. 1089-94.
332 White, H. (1990), “Connectionist nonparametric regression: multi-layer feedforward networks can
learn arbitrary mappings”, Neural Networks, Vol. 3, pp. 535-49.

Corresponding author
Boppana V. Chowdary can be contacted at: chowdary@eng.uwi.tt

To purchase reprints of this article please e-mail: reprints@emeraldinsight.com


Or visit our web site for further details: www.emeraldinsight.com/reprints

You might also like