Sie sind auf Seite 1von 9

Applied Soft Computing 13 (2013) 15431551

Contents lists available at SciVerse ScienceDirect


Applied Soft Computing
j our nal homepage: www. el sevi er . com/ l ocat e/ asoc
Modeling and prediction of machining quality in CNC turning process
using intelligent hybrid decision making tools
C. Ahilan
a,1
, Somasundaram Kumanan
b,1
, N. Sivakumaran
c,1
, J. Edwin Raja Dhas
d,
a
Department of Mechanical Engineering, Oxford Engineering College, Tiruchirappalli, 620009 Tamil Nadu, India
b
Department of Production Engineering, National Institute of Technology Tiruchirappalli, 620015 Tamil Nadu, India
c
Department of Instrumentation and Control Engineering, National Institute of Technology Tiruchirappalli, 620015 Tamil Nadu, India
d
Department of Automobile Engineering, Noorul IslamCentre for Higher Education, Nagercoil 629180, Tamil Nadu, India
a r t i c l e i n f o
Article history:
Received 5 December 2011
Received in revised form19 January 2012
Accepted 13 March 2012
Available online 5 May 2012
Keywords:
CNC turning
Taguchi method
Neural network
Genetic algorithm
Particle swarmoptimization
Response Surface Methodology
Power consumption
Surface roughness
a b s t r a c t
Decision-making process in manufacturing environment is increasingly difcult due to the rapid changes
in design and demand of quality products. To make decision making process (selection of machin-
ing parameters) online, effective and efcient articial intelligent tools like neural networks are being
attempted. This paper proposes the development of neural network models for prediction of machining
parameters in CNC turning process. Experiments are designed based on Taguchis Design of Experiments
(DoE) andconducted with cutting speed, feed rate, depth of cut and nose radius as the process parameters
and surface roughness and power consumption as objectives. Results from experiments are used to train
the developed neuro based hybrid models. Among the developed models, performance of neural network
model trained with particle swarm optimization model is superior in terms of computational speed and
accuracy. Developed models are validated and reported. Signal-to-noise (S/N) ratios of responses are cal-
culated to identify the inuences of process parameters using analysis of variance (ANOVA) analysis. The
developed model can be used in automotive industries for deciding the machining parameters to attain
quality with minimum power consumption and hence maximum productivity.
2012 Elsevier B.V. All rights reserved.
1. Introduction
Turning is one of the important and widely used machining
processes in engineering industries. In turning, the cutting con-
ditions such as cutting speed, feed rate, depth of cut, features of
tools and work piece materials affects the process efciency and
performance characteristics [1,2]. Performance evaluation of CNC
turning is based on the performance characteristics like surface
roughness, material removal rate, tool wear, tool life, cutting force
and power consumption. Very few research attempts have been
donetoestimatethesignicanceof energyrequiredfor themachin-
ing process. Recent increase in energy demand and constraints in
supply of energy becomes a priority for the manufacturing indus-
try. Surface quality is an important performance to evaluate the
productivity of machine tools as well as machined components
[3]. Surface roughness is the critical quality indicator for machined
surfaces [4]. In todays manufacturing industry, special attention is
giventosurface nishandpower consumption. Usually, the desired

Corresponding author. Tel.: +91 04651 250566/366297;


fax: +91 04651 257266/250266.
E-mail address: edwinrajadhas@rediffmail.com(J. Edwin Raja Dhas).
1
Tel.: +91 0431 2503033/2503507; fax: +91 0431 2500133/2503502.
cutting parameters are determined based on experience or by use
of a handbook which does not guarantee optimal performance [5].
It is necessary to select the most appropriate machining settings
in order to improve cutting efciency, process at low cost and
produce high-quality products [68]. So it needs well suited pre-
dictive model for process studies. Researchers attempt different
modeling techniques to investigate the parameters of machining
process.
Quality control techniques such as factorial experiment [9],
design of experiment [10] and Taguchi method [11] have been
employed to predict surface roughness. Taguchi method adopts
a set of orthogonal arrays to investigate the effect of parameters
on specic quality characteristics to decide the optimumparame-
ter combination [12,13]. These kinds of arrays use a small number
of experimental runs to analyze the quality effects of parame-
ters as well as the optimum combination of parameters. Hence,
the Taguchi method has become a powerful design of experiment
method.
Response Surface Method (RSM) is a collection of mathemati-
cal and statistical method useful for the modeling and analysis of
engineering problems in which a response of interest is inuenced
by several variables and the objective is to optimize the response.
The rst step in RSMis to nd a suitable approximation for the true
functional relationship between response of interest Y and a set of
1568-4946/$ see front matter 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.asoc.2012.03.071
1544 C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551
controllable variables {X
1
, X
2
, . . ., X
n
}. Usually when the response
function is not known or non-linear, a general model is in the form:
Y = b
0
+
n

i=1
b
i
X
i
+
n

i=1
n

j=1
b
ij
X
ij
+ (1)
where the expected response is Y and bs are the regression coef-
cients to be estimated. The least square technique is being used
to t a model equation containing the input variables by minimiz-
ing the residual error measured by the sum of square deviations
between the actual and estimated responses. The calculated coef-
cients or the model equations, however, need to be tested for
statistical signicance. RSM [14] is applied to nd a set of process
variables that produce the desired surface nish. RSM was devel-
oped to predict surface roughness end milling titaniumalloy using
uncoated tungsten carbide inserts [15].
ANOVAtest is performed using F-ratio. F ratio [16] measures the
signicance of the model under investigation with respect to the
variance of all the terms included in the error term at the desired
signicance level, .
Checks are made to determine whether the model actually
describes the experimental data by determining R
2
coefcient. The
adequacy of the model is investigatedby the examinationof residu-
als. The residuals, which are the difference between the respective,
observe responses and the predicted responses are examined using
the normal probability plots of the residuals and the plots of the
residuals versus the predicted response. If the model is adequate,
the points on the normal probability plots of the residuals should
form a straight line. On the other hand the plots of the residuals
versus the predicted response should be structure less, that is, they
should contain no obvious patterns.
Due to the inadequacy and inefciency of the linear regres-
sion models to explain the nonlinear properties existing between
machining parameters, intelligent systems such as Articial Neural
Network (ANN), fuzzy logic and expert systemhave been emerged.
ANN is a good technique used to handle problems of nonlin-
earity. ANNs trained with back propagation algorithm[17,18] have
been used to predict the weld bead geometry and penetration in
shielded metal-arc welding process. In real-world applications, the
back-propagation algorithm employed in ANN cannot guarantee
an optimal solution since it may converge to a set of sub-optimal
weights from which it cannot escape. Fuzzy control [19] systems
are effective to the uncertain, nonlinear as well as dynamic time-
varying processes control systems such as machining process. A
model is developed using fuzzy logic for the selectionof cutters and
machining conditions [20]. Optimization techniques like genetic
algorithm, particle swarm optimization and articial bee colony
algorithmareappliedtooptimizevarious manufacturingprocesses.
Prediction of surface roughness with genetic algorithm is pro-
posed [21,22]. The main disadvantage of GA is that it requires a
higher computational power. PSO is used for nding the optimal
parameters for submerged arc weld operations [23]. PSO follows a
collaborative population based search, which is based on the social
behavior of bird ocking and sh schooling. PSO is good in arriv-
ing at best solution by having memory of the previous experience.
Articial bee colony algorithmis employed to optimize multi pass
milling operation.
Eachintelligent technique has particular strengthandweakness
and this made it necessary for combining two or more intelli-
gent tools to improve the performance. The goal is to combine
a technique that has weakness in a particular function with a
technique that has strength in the same function [24]. Hybrid
algorithms are successfully applied for various manufacturing pro-
cesses [2527]. This paper addresses the development of hybrid
algorithms employing neural network embedded with genetic
algorithm and particle swarm optimization technique to predict
machining quality for a given set of process parameters. Finally
response surface models are generated to validate the adequacy of
the developed models. The developed technique promotes man-
ufacturers to develop unmanned factories to achieve the highest
level for automation.
2. Neural network models
The use of neural network models is vital in the modern manu-
facturing environment. Neural networks are dynamic systems that
consist of processing units called neurons with weighted connec-
tions to each other. Neural networks can learn, remember and
retrieve data. The signicant functions of neural network are tack-
ling non-linearity and mapping inputoutput information. The
different types of neural networks which are in practice are back
propagation neural network, counter propagation neural network
andradial basis functionneural network. Eachintelligent technique
has certain strengths and weaknesses and they cannot be applied
universally to every problem. This limitation is the central driv-
ing force behind the creation of intelligent hybrid systems where
two or more techniques are combined in a manner that overcomes
the limitations of individual techniques. The motivation for com-
bining different intelligent techniques is multiplicity of application
tasks, technique enhancement and realizing multifunctional tasks
[24]. Hence optimization techniques like GA and PSO algorithms
are employed in development of neural network models.
2.1. Back propagation neural network model
Back propagation neural network (BPNN) is a multiple layer
ANNwithinput layer, output layer andsomehiddenlayers between
the input and output layers. Its learning procedure is based on gra-
dient search with least mean squared optimality criteria. Once the
input data is fed to the nodes in the input layer (o
i
), this will be fed
to nodes (j) in the hidden layer through weighting factors (w
ji
).
The net input to node j:
net
j
=

i
w
ji
o
i
b
j
(2)
where b
j
is the bias over node j.
The output of the node j:
o
j
=
1
1 +e
net
j
(3)
Similarly the outputs fromnodes in the hidden layer are fed into
nodes in the output layer. This process is called the feed forward
stage. After feed forward, calculation output (o
pk
) can be obtained
from nodes in the output layer. In general, the output o
pk
will not
be the same as the desired known target t
pk
. Therefore the average
systemerror is:
E =
1
2p

k
(t
pk
o
pk
)
2
The error is then back propagated fromnodes in the output layer to
nodes in the hidden layer using gradient search method
p
w
kj
=
(E/w
kj
) =
k
o
j
delta value for output layer is:

k
= o
k
(1 o
k
)(t
k
o
k
)
delta value for hidden value is:

j
= o
j
(1 o
j
)

w
kj

k
C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551 1545
This process is called back propagation stage. After all examples are
trained the systemwill collect adjusted weights according to:
w
ji
=

p
w
ji
Updating of weights is done according to
w
ji
(n +1) = w
ji
(n) +w
ji
(4)
2.2. Neural network model trained with genetic algorithm
Genetic algorithm is widely used to solve optimization prob-
lems. The standard genetic algorithmproceeds as follows: aninitial
population of individuals (that is, a set of solutions for the given
problemrepresented by chromosomes) is generated at randomor
heuristically. The data is normalized in the range 0.10.9 by using
Eq. (4)
x
norm
= 0.1 +0.8

x
i
x
min
x
max
x
min

(5)
The steps used in genetic algorithmare (a) initializing the pop-
ulation randomly, (b) cross over with probability and (c) mutation
with probability. In each generation, the individuals in the current
population are decoded and evaluated according to some prede-
ned quality criterion, referred as the tness function. To form a
new population for next generation, the individuals are selected
accordingtotheir tness. The selectedbest populationwill undergo
crossover and mutationoperationto produce newoffsprings. Then
some or all population are replaced with newly created offspring
based on their tness. This action is motivated by a hope that the
newpopulationwill bebetter thantheoldone. This is repeateduntil
some condition (for example number of populations or improve-
ment of the best solution) is satised. If GA has been designed well,
the population will converge to an optimal solution to the problem.
2.3. Neural network model trained with particle swarm
optimization
In this proposed model, back propagation algorithm of neural
network is replaced by PSO algorithm. Here the randomly gener-
ated weights are assigned in each link of neural network. In particle
swarm optimization algorithm, PBest is the location of the best
solution of a particle has achieved so far. GBest is the location of
the best solution that any neighbor of a particle has achieved so
far. Initially random numbers are generated for each particle and
these randomvalues are considered as PBest and present weights.
Velocity is calculated using Eq. (6) and added with the present
weight in each link of neural network. For each particle, the newly
calculated weights are compared with the PBest weights and the
minimumerror produced weights are stored in PBest. Initial veloc-
ity V is assumed to be 1 and GBest is the weights of minimumerror
produced particle. Newweight is calculated as in Eq. (7).
Velocity[] = wVelocity[] +C
1
rand
1
(PBest[] present[])
+C
2
xrand
2
(GBest[] present[]) (6)
Present[] = Present[] +Velocity[] (7)
where C
1
and C
2
are two positive constants named learning factors.
rand
1
and rand
2
are two randomfunctions in the range [0, 1]. w is
an inertia weight to control the impact of the previous history of
velocities on the current velocity. The operator w plays the role of
balancing the global search and the local search; and was proposed
to decrease linearly with time from a value of 1.40.5. As such,
global search starts with a large weight and then decreases with
Fig. 1. Schematic diagramof CNC set-up for turning operation.
time to favor local search over global search. When the number of
iterations is equal to the total number of particles, goal is compared
with the error produced by the GBest weights. If the error produced
by the GBest weights are less than or equal to the goal, weights in
the GBest are used for testing and prediction. Otherwise weights of
minimumerror are stored in GBest and the iterations are repeated
until goal reached.
3. Data acquisition
CNC (FANUC control) lathe with7.5kWspindle power and max-
imumspindle speed of 4500rpmis used to performthe machining
operation. Schematic diagram of the experimental set-up used is
shown in Fig. 1. FLUKE 43B power quality analyzer to measure
the power consumption (W) of cutting process is connected to
the power supply of CNC turning center as shown in Fig. 2. Power
consumption is measured for each setting of machining operation
and idle running operation. Taylor-Hobson Talysurf as is used to
measure surface roughness. AISI 304 stainless steel of 70 HRC in
the form of round bars with 50mm diameter and 200mm cut-
ting lengthare usedfor machining. Photographof the experimental
setup is shown in Fig. 3. The percentage of chemical composition of
work material is 0.08 C, 1820 Cr, 2 Mn, 810.5 Ni, 0.045 P, 0.03 S, 1
Fig. 2. Photograph of the Fluke 43b power quality analyzer.
1546 C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551
Fig. 3. Photograph of CNC experimental set-up.
Si andremainingFe. It is mainlyusedfor chemical equipments, food
processing equipment, dairy equipment, textile dyeing equipment,
cryogenic vessels and hospital surgical equipment. Standard car-
bide tool inserts CNMG120404, CNMG120408 and CNMG120412
are used for machining.
To perform the experimental design, three levels of machining
parameters cutting speed in m/min, A (100.125.150) feed rate in
mm/rev, B (0.05, 0.1, 0.15), depth of cut in mm, C (0.20, 0.35, 0.50)
and nose radius in mm, D (0.4, 0.8, 1.2). An appropriate orthogo-
nal array for the experiments must have the degrees of freedom
greater than or at least equal to those for the process parame-
ters. In this study, an L
27
(3
4
) orthogonal array is used because it
has 26 degrees of freedom more than the 8 degrees of freedom
in the machining parameters. Experimental combinations of the
machining parameters and the observations of power consump-
tion and surface roughness are shown in Table 1. Responses from
experimentation are used to develop the proposed neural network
models.
Table 1
Experimental design and their responses.
Trial
numbers
A B C D Power (W) Surface
roughness
(m)
1 1 1 1 1 213 2.04
2 1 1 2 2 320 1.74
3 1 1 3 3 332 2.02
4 1 2 1 2 283 1.25
5 1 2 2 3 340 1.1
6 1 2 3 1 393 1.02
7 1 3 1 3 275 1.5
8 1 3 2 1 350 1.12
9 1 3 3 2 620 1.35
10 2 1 1 2 392 1.82
11 2 1 2 3 438 1.52
12 2 1 3 1 441 1.78
13 2 2 1 3 391 1.04
14 2 2 2 1 570 0.84
15 2 2 3 2 668 1.02
16 2 3 1 1 394 1.16
17 2 3 2 2 617 1.26
18 2 3 3 3 760 1.48
19 3 1 1 3 448 2.02
20 3 1 2 1 516 1.54
21 3 1 3 2 585 1.94
22 3 2 1 1 476 1.08
23 3 2 2 2 625 1.16
24 3 2 3 3 765 1.42
25 3 3 1 2 528 1.46
26 3 3 2 3 706 1.38
27 3 3 3 1 873 1.64
Input
layer
Hidden
layers
Output
layer
Surface
roughness
Power
Consumption
Cutting Speed
Feed rate
Depth of cut
Nose radius
Fig. 4. Developed structure of BPNN model.
4. Development of proposed neural network models
To handle multiplicity of application tasks and realizing multi-
functionality different neural network models are proposed and
developed.
4.1. Development of proposed NNBPN model
Topography of the proposed NNBPN model to predict sur-
face roughness and power consumption is shown in Fig. 4.
It is feed forward back propagation network trained with
LevenbergMarquardt back propagation algorithm. Experimental
data is usedfor trainingandtestingthedevelopedbackpropagation
neural network model. The learning function is gradient descent
algorithmwith momentumweight and bias learning function. The
number of hidden layers and neurons are determined through a
trial and error method, in order to accommodate the converged
error. The structure of the developed neural network is 4962 (4
neurons in the input layer, 9 neurons in 1st hidden layer and 6 neu-
rons in 2nd hidden layer and 2 neurons in the output layer). With
a learning rate of 0.57 and a momentum term of 0.9, the network
is trained for 10,000 iterations. Error between the desired and the
actual outputs is less than 0.001 at the end of training process.
4.2. Development of proposed NNGA model
The proposed method to predict machining using NNGA model
is depicted in Fig. 5. The topology of a neural network model as
shown in Fig. 4 is dened and will remain xed after the initial-
ization. The learning function is gradient descent algorithm with
momentum weight and bias learning function. The transfer func-
tion and error criteria (mean square error) are xed. Training is
done by genetic algorithmsearch. In this application each string or
chromosome in the population represents the weight and bias val-
ues of the network. The initial population is generated randomly.
By selecting suitable parameters like selection criteria, probabil-
ity of crossover, probability of mutation, initial population etc., to
the GA, high efciency and performance is achieved. The objective
function is minimization of the mean square error. Fitness function
consideredis the minimumof the meansquare error andcomputed
by recalling the network. After getting the tness values of all chro-
mosomes, they are ranked based on the best tness values. For the
production of offspring for next generation half of the best-ranked
population is selected. This half population undergo crossover with
crossover probability (0.9). This againwill be mutated to give a new
offspring, withmutationprobability(0.08), whichis combinedwith
selected best population to form a new population for the next
C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551 1547
Results Results Network
Testing
Network
Prediction
Network Weight
Matrix
Ranking
Network with
optimal weight set
Data from
Experiments
Generation of
optimized
weights
Cross over
Mutation
Evaluation of
objective
function
Fig. 5. NNGA model for prediction of machining quality.
1
-0.3
--0.7
X1
X2
1.0
3
2
1
2
1
-0.9
1.4
-1.9
-0.4
1.4
-0.6
-1.4
0.5 Parent 1
0.4
3
4
2
2
2
X3
X4
Y1
Y2
0.9
0.7
1.5
0.6
-0.8
-0.4
-0.4
1.6
1.7
2.0
2.0
2.0
-1.0
1.0
1.2
2.0
2.0
2.0
0.6
2.0
1.0
1.3
1.5
0.7
0.9
1.0
2.0
1.0
1.6
2.0
1.0
2.0
1.0
1.8
1.0
1.7
2
1
-0.8
--0.7
X1
X2
-0.7
3
2
1
2
1
-0.3
1.4
-1.4
-0.4
1.4
-0.6
-1.7
1.9
Parent 2
1.4
3
4
2
2
X3
X4
Y1
Y2
0.3
1.7
0.5
0.9
-0.4
-0.2
-1.4
-0.6
1.7
2.0
2.0
2.0
-1.0
1.0
1.2
2.0
2.0
2.0
0.6
1.0
1.3
1.6
1.5
0.7
0.9
1.0
2.0
1.0
1.6
2.0
1.0
2.0
1.0
1.8
1.0
1.7
Fig. 6. Typical neural network model coded with weights.
For each particle, calculate the output of neural network and error
Find out the minimum error produced particle
Store the corresponding weights in gBest
Find velocity of i
th
particle and add velocity to the present weights to get
new weights
Error(new weights)
< Error(pBest)
Yes
Velocity (next particle) = Velocity(current particle)
No
Iteration = Iteration + 1, i = i+1
Is Iteration
= No. of
Particles
No
Yes
gBest = weights(minimum error
produced particle)
Is goal
reached?
No
Yes
Iteration = 0, i = 0
weights in gBest are
used for testing and
prediction
Stop
Find out random values and stored as pBest and present
weights, Assign Iteration = 0 and i = 0
Store the newly
generated weights in
pBest
Start
Fig. 7. Flowchart for training the neural network using PSO algorithm.
generation. Typical neural network model of 452 architecture
with weights is depicted in Fig. 6.
Where X
1
, X
2
, . . . are inputs and Y
1
, Y
2
, . . . are outputs. This
will be continued till the stopping criteria are reached. The stop-
ping criteria for this network are the number of generations. Neural
network coded with optimal weight set chromosome produced by
genetic algorithmis forwarded to predict the quality of weld.
4.3. Development of proposed NNPSO model
In the development of NNPSO model, back propagation algo-
rithm of neural network is replaced by PSO algorithm. The data
required for training and testing the NNPSO model is taken from
Table 1. PSO algorithm used to train the neural network is shown
in Fig. 7. The total number of population size is 40. The dimension
size is 4 and the value of inertia weight is 0.51.4. Number of itera-
tions allowed is equal to total number of particles. Value of velocity
factors (C
1
and C
2
) is taken as 1.4.
5. Validation of the developed models
The developed models are validated from conrmatory tests.
Table 2shows the outcomes fromconrmatory tests. Table 3shows
the accuracy of the values predicted from the developed mod-
els. The percentage of error predicted by the developed models
are calculated by [{Observed value predicted value}/predicted
value] 100.
Time required for training the developed network models are
compared in terms of number of epochs as shown in Figs. 8 and 9.
The congurations of the computing machine used are Intel Pen-
tiumIV1.8GHz processor, 512MB RAMand 80GB Hard Disk Drive.
1548 C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551
Table 2
Results fromconrmatory experiment.
Test no. Cutting speed
(m/s)
Feed rate
(mm/rev)
Depth of cut
(mm)
Nose radius
(mm)
Observed values
Power
consumption
(W)
Surface
roughness
(m)
1 100 0.1 0.2 0.4 245 1.14
2 125 0.15 0.5 0.8 815 1.35
3 150 0.05 0.35 1.2 526 1.92
Table 3
Results fromthe developed models in terms of percentage error.
Percentage of error BPNN model for Percentage of error NNGA model for Percentage of error NNPSO model for
Power consumption Surface
roughness
Power
consumption
Surface
roughness
Power
consumption
Surface
roughness
2.94 3.63 1.606 3.06 1.6 2.63
1.74 9.39 1.49 4.92 0.617 2.87
1.68 4.43 0.94 2.673 0.766 2.12
5.1. Analysis of CNC turning parameters on power consumption
and surface roughness
Power consumption and surface roughness have lower-the-
better criterion. The S/N ratios of responses are computed using
(8) and (9). The main effects plot for power consumption is
shown in Fig. 10. Higher the difference between the minimum
and maximumS/N ratios in each factor is, higher the effect on the
power consumption. It clearly shows that cutting speed (A) was a
0
50
100
150
200
250
300
350
400
450
8 7 6 5 4 3 2 1
Trial Nos (Surface roughness)
E
p
o
c
h
s
BPNN NNGA NNPSO
Fig. 8. Training time required by the developed models for surface roughness pre-
diction.
0
50
100
150
200
250
300
350
400
450
1 2 3 4 5 6 7 8
Trial Nos (Power consumption)
E
p
o
c
h
s
BPNN NNGA NNPSO
Fig. 9. Training time required by the developed models for power consumption
prediction.
dominant parameter for power consumption followed by depth of
cut (C) and feed rate (B). Nose radius (D) has a little effect on power
consumption. Contribution of parameters on power consumption
is identied using ANOVA and presented in Table 4. Because of the
lower-the-better characteristics, highest S/N ratio in each factor
was desirable to obtain minimum power consumption. The S/N
ratio
ij
for the ith performance characteristic in the jth experiment
is expressed in Eq. (8).

ij
= 10log(L
ij
) (8)
For lower-the-better performance characteristic, L
ij
is expressed
in Eq. (9).
L
ij
=
1
n
n

k=1
y
2
ijk
(9)
where k is number of tests, y
ijk
is experimental value of the ith
performance characteristic in the jth experiment at the kth tests.
S/N ratio of four factors main effects plot for surface roughness
is shown in Fig. 11. The contribution of the parameters on surface
roughness is identied using ANOVA and presented in Table 5.
150 125 100
-50
-52
-54
-56
0.15 0.10 0.05
0.50 0.35 0.20
-50
-52
-54
-56
1.2 0.8 0.4
A
M
e
a
n
B
C D
Main Effects Plot for S/N ratio-Power Consumption
Data Means
Fig. 10. Parameters (A, B, C and D) main effect plot for power consumption.
C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551 1549
Table 4
Analysis of variance analysis for power consumption.
Factor DOF SS MS F value % contribution
A 2 123.18 61.59 86.75 50.23
B 2 32.11 16.05 22.61 13.09
C 2 72.73 36.36 51.21 29.66
D 2 4.5 2.25 3.17 1.84
Error 18 12.69 0.71 5.18
Total 26 245.21 100.00
Table 5
Analysis of variance analysis for surface roughness.
Factor DOF SS MS F value % contribution
A 2 7.55 3.77 7.69 6.41
B 2 87.74 43.87 89.53 74.45
C 2 9.14 4.57 9.33 7.76
D 2 4.67 2.33 4.76 3.96
Error 18 8.75 0.49 7.42
Total 26 117.85 100.00
Table 6
R
2
value of response surface models for power consumption and surface roughness.
Response surface regression model R
2
value for power consumption R
2
value for surface roughness
Linear 89.09% 32.60%
Linear +square 91.31% 92.51%
Linear +interaction 96.41% 36.84%
Full quadratic (linear +square +interaction) 98.63% 96.75%
5.2. Development of response surface models
Response surface model have linear model, quadratic model,
interaction model and full quadratic model and are as follows (for
four input parameters):
Linear equation:
Y = b
0
+b
1
X
1
+b
2
X
2
+b
3
X
3
+b
4
X
4
(10)
150 125 100
-1
-2
-3
-4
-5
0.15 0.10 0.05
0.50 0.35 0.20
-1
-2
-3
-4
-5
1.2 0.8 0.4
A
M
e
a
n
B
C D
Main Effects Plot for S/N ratio-Surface Roughness
Data Means
Fig. 11. Parameters (A, B, C and D) main effect plot for surface roughness.
Quadratic equation:
Y = b
0
+b
1
X
1
+b
2
X
2
+b
3
X
3
+b
4
X
4
+b
5
(X
1
)
2
+b
6
(X
2
)
2
+b
7
(X
3
)
2
+b
8
(X
4
)
2
(11)
Interaction equation:
Y = b
0
+b
1
X
1
+b
2
X
2
+b
3
X
3
+b
4
X
4
+b
5
X
1
X
2
+b
6
X
1
X
3
+b
7
X
1
X
4
+b
8
X
2
X
3
+b
9
X
2
X
4
+b
10
X
3
X
4
(12)
Second order equation (Full model)
Y = b
0
+b
1
X
1
+b
2
X
2
+b
3
X
3
+b
4
X
4
+b
5
(X
1
)
2
+b
6
(X
2
)
2
+b
7
(X
3
)
2
+b
8
(X
4
)
2
+b
9
X
1
X
2
+b
10
X
1
X
3
+b
11
X
1
X
4
+b
12
X
2
X
3
+b
13
X
2
X
4
+b
14
X
3
X
4
(13)
R
2
values for power consumption model is tabulated in Table 6 and
the desired full quadratic model is in Eq. (14).
Power consumption = 1192.36 +19.1028 X
1
1913.93 X
2
513.16 X
3
+450.454 X
4
0.0617 X
2
1
4755.56 X
2
2
313.58 X
2
3
206.597 X
2
4
+10.5422 X
1
X
2
+4.2193 X
1
X
3
1.1089 X
1
X
4
+8714.07 X
2
X
3
+118.889 X
2
X
4
+110.37 X
3
X
4
(14)
where X
1
is cutting speed, X
2
is feed rate, X
3
is depth of cut and X
4
is nose radius.
The normal probability plot of the residual for power consump-
tion is shown in Fig. 12. R
2
values for surface roughness model is
tabulated in Table 6 and full quadratic model is given in Eq. (15).
Surface roughness = 9.80674 0.07608 X
1
52.8926 X
2
9.17185 X
3
+0.33241 X
4
+0.00026 X
2
1
+198.00 X
2
2
1550 C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551
Table 7
Validation test results of developed RS models for power consumption and surface roughness.
RS model value Error in %
Power consumption (W) Surface roughness (m) Power consumption (W) Surface roughness (m)
227 1.18 7.35 3.51
793 1.4 2.70 3.70
509 1.77 3.23 7.81
+9.18519 X
2
3
0.10417 X
2
4
+0.05404 X
1
X
2
+0.01849 X
1
X
3
0.00002 X
1
X
4
+5.45185 X
2
X
3
0.13333 X
2
X
4
+0.0037 X
3
X
4
(15)
The normal probability plot of the residuals for surface rough-
ness is shown in Fig. 13. The developed model for power
consumption and surface roughness is validated by conducting
experiments (new set of input conditions) and their results are
presented in Table 7.
75 50 25 0 -25 -50
99
95
90
80
70
60
50
40
30
20
10
5
1
Residual
P
e
r
c
e
n
t
Normal Probability Plot
(response is Power consumption)
Fig. 12. Developed response surface models normal probability plot for power
consumption.
0.15 0.10 0.05 0.00 -0.05 -0.10 -0.15
99
95
90
80
70
60
50
40
30
20
10
5
1
Residual
P
e
r
c
e
n
t
Normal Probability Plot
(response is Surface roughness)
Fig. 13. Developed response surface models normal probability plot for surface
roughness.
6. Results
Experiments are conducted in CNC machine using Taguchis
principles. Results fromexperimentation are used to train and test
the developed neural network models. The performance of the
developed hybrid neural network models is measured in terms of
computational accuracy and speed. Conrmatory tests are done to
validate this approach. FromTable 6 the predictive error computed
using NNPSO is less compared with other developed models. The
reason is that NNPSO model searches solution in the search space
different from other neural network models. It maintains a inter-
nal memory to store the Gbest and Pbest solutions. Each individual
in the population tries to emulate the Gbest and Pbest solutions in
the memory by updating two PSOequations. But NNGAmodel iter-
atively searches for several good individuals in the population, and
make the population to emulate the best solutions found in that
generation through reproduction, crossover and mutation opera-
tors. Hence it requires substantial computational time to perform
decision making whereas back propagation algorithmtraining net-
work may converge to a set of sub-optimal weights. Hence the
effectiveness of NNPSO model in nding the true global optimal
solution is competent than the other neural network models. The
developedneural networkmodel withPSOpredicts theoutput with
an accuracy of error is hardly less than 2%.
Results in Figs. 8 and 9 indicate that the developed ANN model
trained with PSO needs minimum number of epochs and hence
computational timerequiredis less. Experimental datafromTable1
are used for analyzing the inuence of process parameters on
responses (power consumption and surface roughness) and build-
ingtheresponsesurface(RS) model of CNCturningprocess. Effect of
CNCturning parameters onpower consumptionandsurface rough-
ness are evaluated using ANOVA and calculation of S/N ratio. For
power consumption (Fig. 10 and Table 4) cutting speed is the most
critical factor followed by depth of cut, feed rate and nose radius
whereas feed rate (Fig. 11 and Table 5) has a signicant effect
Fig. 14. Percentage error of power consumption model and surface roughness
model.
C. Ahilan et al. / Applied Soft Computing 13 (2013) 15431551 1551
on surface roughness of CNC turning process. Results of ANOVA
(Tables 4 and 5) and comparisons of experimental data represent
that the response surface model for power consumption and sur-
face roughness are well suitedwithexperimental results witha 95%
condence interval.
Normal probability plot of residual for power consumption
(Fig. 12) and surface roughness (Fig. 13) reveals that the residu-
als fall on a straight line implying that the errors are distributed
normally and they have no obvious pattern and unusual struc-
ture. From this it is identied that the developed models are
adequate. Fig. 14shows that the maximumabsolute error andaver-
age absolute error of power consumption response surface models
prediction is 10.90% and 3.24%. Similarly, for surface roughness
model the maximumabsolute error is 9.76% and average absolute
error is 3.72%. Results reveal that the developedmodels are reliable.
Adequacies of the prediction models are validated.
7. Conclusions
Growth of an industry depends on effective and efcient deci-
sion making process. Hence development of intelligent decision
making tools needs attention. In this work, different hybrid deci-
sionmakingtools suchas NNBPN, NNGAandNNPSOalgorithmhave
been studied for the optimization of machining parameters in CNC
turning operations. Optimization procedures based on the above
approaches have been developed and successfully implemented.
Signicant improvement is obtained with the above techniques
when compared to the results by handbook, charts and tables. This
is due to the limitations of enumerative search followed by the
direct search method. In the use of non-conventional techniques,
NNPSOalways yields better results. So it is suggested to use NNPSO
for solving problemof turning operations. Results obtained in this
work are intended for use by numerical control or manually oper-
ated machines.
References
[1] M.C. Shaw, Metal Cutting Principles, Oxford University Press, NewYork, 1984.
[2] G. Boothroyd, W.A. Knight, Fundamentals of Machining and Machine Tools,
Marcel-Dekker, NewYork, 1989.
[3] P.G. Benardos, G.C. Vosniakos, Predicting surface roughness in machining: a
review, International Journal of Machine Tool and Manufacturing 43 (2003)
833844.
[4] I.P. Arbizu, C.J.L. Perez, Surface roughness prediction by factorial design of
experiments in turning processes, Journal of Materials Processing Technology
143144 (2003) 390396.
[5] M.P. Groover, Fundamentals of Modern ManufacturingMaterials Processing
and Systems, Prentice-Hall, Englewood Cliffs, 1996.
[6] J. Kopac, S. Sali, Tool wear monitoring during the turning process, Journal of
Materials Processing Technology 113 (2001) 312316.
[7] K. Ihsan, K. Mustafa, C. Ibrahim, S. Ulvi, Determination of optimum cutting
parameters during machining of AISI 304 austenitic stainless steel, Materials
and Design 25 (2004) 303305.
[8] M.A. Xavior, M. Adithan, Determining the inuence of cutting uids on tool
wear andsurface roughness duringturningof AISI 304austenitic stainless steel,
Journal of Materials Processing Technology 209 (2009) 900909.
[9] I.A. Chouhury, M.A. El Baradie, Surface roughness inthe turning of highstrength
steel by factorial design of experiments, Journal of Materials Processing Tech-
nology 67 (1997) 5561.
[10] J.P. Davim, Anoteonthedeterminationof optimal cuttingconditions for surface
nish obtained in turning using design of experiments, Journal of Materials
Processing Technology 116 (2001) 305308.
[11] D.I. Lalwani, N.K. Mehta, P.K. Jain, Experimental investigations of cutting
parameters inuence on cutting forces and surface roughness in nish hard
turning of MDN250 steel, Journal of Materials Processing Technology 206
(2008) 167179.
[12] D.C. Montgomery, Design and Analysis of Experiments, 4th ed., Wiley, New
York, 1997.
[13] M. Nalbant, H. Gokkaya, G. Sur, Application of Taguchi method in the opti-
mization of cutting parameters for surface roughness in turning, Materials and
Design 28 (2007) 13791385.
[14] N. Mohanasundararaju, R. Sivasubramanian, N. Alagumurthi, Optimisation
of work roll grinding using Response Surface Methodology and evolution-
ary algorithm, International Journal of Manufacturing Research 3 (2) (2008)
236251.
[15] T.L. Ginta, A.K.M. Nurul Amin, H.C.D. Mohd Radzi, Development of surface
roughness models in end milling titanium alloy Ti6Al4V using uncoated
tungsten carbide inserts, European Journal of Scientic Research 28 (4) (2009)
542551.
[16] R.A. Fisher, Statistical Method for Research Worker, Oliver & Boyd, London,
1925.
[17] A.M. Zain, H. Haron, S. Sharif, Prediction of surface roughness in the end milling
machining using Articial Neural Network, Expert Systems with Applications
37 (2) (2010) 17551768.
[18] K. Sundara Murthy, I. Rajendran, Astudy on optimisation of cutting parameters
and prediction of surface roughness in end milling of aluminium under MQL
machining, International Journal of Machining and Machinability of Materials
7 (2010) 112128.
[19] L.A. Zadeh, Fuzzy sets, Information and Control 8 (1965) 338353.
[20] J.C Chen, M. Savage, Fuzzy-net-based multilevel in-process surface roughness
recognition system in milling operations, International Journal of Advanced
Manufacturing Technology 17 (1995) 670676.
[21] M. Brezocnik, M. Kovacic, M. Ficko, Predictionof surface roughness withgenetic
programming, Journal of Materials Processing Technology 157158 (2004)
2836.
[22] S.K.N. Reddy, V.P. Rao, Selection of optimumtool geometry and cutting condi-
tions using a surface roughness prediction model for end milling, International
Journal of Advanced Manufacturing Technology 26 (2005) 12021210.
[23] J. Edwin Raja Dhas, S. Kumanan, Optimization of parameters of submerged arc
weld using non conventional techniques, Applied Soft Computing 11 (2011)
51985204.
[24] S. Goonatilake, S. Khebbal, Intelligent Hybrid System, John Wiley & Sons, New
York, 1995.
[25] J. Rezaeian, N. Javadian, R. Tavakkoli-Moghaddam, F. Jolai, A hybrid approach
based on the genetic algorithm and neural network to design an incremental
cellular manufacturing system, Applied Soft Computing 11 (2011) 41954202.
[26] N. Nameer, El-Emam, H. Riadh, Al-Rabeh, An intelligent computing technique
for uid owproblems using hybrid adaptive neural network and genetic algo-
rithm, Applied Soft Computing 11 (2011) 32833296.
[27] R. Fabricio, Bittencout, E. Luis, E. Zarate, Hybrid structure based on previous
knowledge and GA to search the ideal neurons quantity for the hidden layer of
MLPapplication in the cold rolling process, Applied Soft Computing 11 (2011)
24602471.
Dr. C. Ahilan is Professor and Head in the Department of Mechanical Engineering
Oxford Engineering College, Tiruchirappalli, 620009 Tamil Nadu, India. He received
his BE in Mechanical engineering fromthe PSG College of Technology, India in 1997
and MTech in manufacturing technology fromthe National Institute of Technology
Tiruchirappalli in 2007. He has worked as a Research Assistant in the Department
of Production Engineering, National Institute of Technology Tiruchirappalli. He has
published 5 papers in international conferences, and 4 international journals. His
research area includes intelligent industrial energy management systems.
Dr. Somasundaram Kumanan is a Professor and Dean in the Department of Pro-
duction Engineering at National Institute of Technology, Tiruchirappalli, India. He
obtained his Doctorate Degree in Manufacturing Management from Indian Insti-
tute of Technology, Madras, India. He has published 50 papers in international
conferences and 100 international journals. His research interests are Intelligent
Manufacturing Systems, Modeling, Simulation and Optimization of Manufacturing
Systems.
Dr. N. Sivakumaran is an Assistant Professor in the Department of Instrumen-
tation and Controls Engineering, National Institute of Technology Tiruchirappalli,
India. He received his BE in electronics and instrumentation from Bharathidhasan
University, India, ME in process control and instrumentation engineering from
Annamalai University, India and PhD in the area of process control and instrumen-
tation fromNational Institute of Technology Tiruchirappalli. He has published more
than 22 papers in national/international conferences, 13 international journals and
3 national journals. His teaching and research interests are in the areas of Process
control and Control systems.
Dr. J. Edwin Raja Dhas is a Professor and Head in the Department of Automo-
bile Engineering at Noorul IslamUniversity, India He obtained his Doctoral Degree
fromNational Institute of Technology, Tiruchirappalli, India. He has published more
than 22 papers in national/international conferences and 15 international/national
journals. His research interests are Non traditional simulation and optimization of
Manufacturing Systems.

Das könnte Ihnen auch gefallen