Sie sind auf Seite 1von 12

J Intell Manuf (2012) 23:639650

DOI 10.1007/s10845-010-0415-2
Support vector machines models for surface roughness prediction
in CNC turning of AISI 304 austenitic stainless steel
Ula s ayda s Sami Ekici
Received: 8 September 2009 / Accepted: 3 May 2010 / Published online: 13 May 2010
Springer Science+Business Media, LLC 2010
Abstract In the present investigation, three different type
of support vector machines (SVMs) tools such as least square
SVM (LS-SVM), Spider SVM and SVM-KM and an arti-
cial neural network (ANN) model were developed to estimate
the surface roughness values of AISI 304 austenitic stain-
less steel in CNC turning operation. In the development of
predictive models, turning parameters of cutting speed, feed
rate and depth of cut were considered as model variables.
For this purpose, a three-level full factorial design of experi-
ments (DOE) method was used to collect surface roughness
values. A feedforward neural network based on backpropa-
gation algorithm was a multilayered architecture made up of
15 hidden neurons placed between input and output layers.
The prediction results showed that the all used SVMs results
were better than ANN with high correlations between the
prediction and experimentally measured values.
Keywords Surface roughness Support vector machines
AISI 304 machining
Introduction
Turning is the primary operation in most of the production
processes in the industry. The turning operation produces the
components, which have critical features that require spe-
cic surface nish (Davim et al. 2008). Surface roughness is
U. ayda s (B)
Technical Education Faculty, Department of Manufacturing,
University of Firat, 23119 Elazig, Turkey
e-mail: ucaydas@rat.edu.tr
S. Ekici
Technical Education Faculty, Department of Electric,
University of Firat, 23119 Elazig, Turkey
a widely used index of product quality and in most cases a
technical requirement for mechanical products. Hence,
achieving the desired surface quality is of great importance
for the functional behaviour of a part (Benardos et al. 2003).
Surface roughness is the result of the process parameters such
as tool geometry (i.e. nose radius, edge geometry, rake angle,
tool tip radius, chamfer thick, etc.), cutting conditions (i.e.
feed rate, cutting speed, depth of cut, etc.) and workpiece
properties (Umbrello et al. 2008).
Nowadays, many papers have been published about the
prediction of surface quality in turning process. Researchers
have approached the problem from different points of view
and using different articial intelligence techniques. Durmu s
Karayel (Karayel 2008) presented a neural network model
for prediction and control of surface rougness in a computer
numerically controlled (CNC) lathe. A feedforward multi-
layered neural network was developed and the neural net-
work model was trained using the scaled conjugate gradient
algorithm(SCGA), which is a type of back-propagation. The
predicted surface roughness from the model was very close
to the values measured experimentally with a high correla-
tion coefcient. Additionally, the feed rate parameter was
found the most dominant factor on surface roughness, while
the effect of depth of cut was not regular and had a var-
iable character. Muthukrishnan and Davim (2009) studied
the surface roughness of Al-SiC (20p) composites in turn-
ing operation by using coarse grade polycrystalline diamond
(PCD) insert under different cutting conditions. A multilay-
ered perception ANNmodel has been constructed with back-
propagation algorithm using the input parameters of depth
of cut, cutting speed and feed. Additionally, the analysis
of variance (ANOVA) technique was adopted to study and
it was shown that the statistical inuence of feed rate on
surface roughness was the highest (51%), right after the
depth of cut (30%) and the cutting speed (12%). The result
123
640 J Intell Manuf (2012) 23:639650
of ANN model showed close matching between the model
output and directly measured surface roughness. Grzesik
and Brol (2003) made an assesment on surface quality of
turned, ground and honed specimens by using three differ-
ent approaches, i.e. statistical, fractal and ANNs. It has been
shown that the results of modeling techniques showed good
results. Surjya and Chakraborty (2005) developed a method-
ology for the prediction of surface roughness in turning using
back-propagation neural networks. Agood number of exper-
iments have been conducted on mild steel workpiece using
HSS as the cutting tool material. Optimum network archi-
tecture has been found out based on the mean square error
and convergence rate. Finally, the presented network model
was efcient to predict the surface roughness in an accept-
able error rate. Ba gc and I sk (2006) carried out orthogonal
cutting tests on unindirectional glassbre reinforced plastics
(GFRP), using cermet tools. Depth of cut, feed rate and cut-
ting speed parameters were varied. A statistical three level
full factorial experimental design technique was designed to
collect surface roughness data. AnANNandresponse surface
models were developed for surface roughness prediction.
Good agreement was observed between the predicted models
and experimental measurements. zel et al. (2007) investi-
gated the surface nishing and tool ank wear in turning of
AISI D2 steels (60HRC) using ceramic wiper (multi-radii)
design inserts. Surface roughness values as low as 0.18
0.20 m were achieved. Multiple linear regression models
and neural network models were developed for predicting
surface roughness and tool ank wear. The results showed
that neural network models were suitable for predicting the
outputs for a range of cutting conditions and could be utilized
in intelligent process planning for hard turning. Suresh et al.
(2002) developed a surface roughness prediction model for
machining mild steel, using response surface methodology
(RSM). A simple three level (3
4
) factorial design of exper-
iments was adopted for experimentation. A second order
mathematical model, in terms of machining parameters, was
developed for surface roughness prediction using RSM. The
factor effects of the individual process parameters have been
determined. An attempt also has been made to optimize
the surface roughness prediction model using Genetic Algo-
rithms (GA) to optimize the objective function. Minimum
and maximum surface rughness values and their respective
optimal machining conditions have also been determined by
GA program. Lee and Tarng (2001) used a computer vision
technique to inspect surface roughness of S45C steel under
a variation of turning operation. They used digital camera to
acquire a surface image of the workpiece and then the feature
of the surface image. Apolinomal network using a self-orga-
nizing adaptive modeling method was applied to contruct
the relationships between the features of the surface image
and the actual surface roughness. Roy (2006) developed a
GA trained fuzzy expert system for predict the surface nish
in ultra precison diamond turning of Al 6061/SiCp metal
matrix composite. Wong and Hamouda (2003) introduced
a new type of articial neuron in the design of neural net-
work for turning process, namely the product neuron, which
had multiplication instead of summation. The back-propa-
gation algorithm was used to train the network. Comparison
among neural network prediction, fuzzy logic predicton and
handbook data was made. Nalbant et al. (2009) presented an
experimental investigation of the effects of uncoated, PVD-
coated and CVD-coated cemented carbide inserts and cutting
parameters on surface roughness of AISI 1030 steel in CNC
turning operation and its prediction using ANNs. The ANN
results and experimental values were compared by statisti-
cal error analyzing method. The absolute fraction of vari-
ance value was obtained as 0.99985 within an acceptable
accuracy. Paiva et al. (2007) presented a hybrid approach,
combining RSM and principal component analysis (PCA) to
optimize multiple correlated responses in a turning process.
The mixed ceramic tool life, processing cost per piece, cut-
ting time, the total turning cycle time, surface roughness and
material removal rate were the main outputs under investiga-
tion. It was shown that the multiresponse optimization was
achieved at a cutting speed of 238 m/ min, with a feed rate
of 0.08 mm/rev and a depth of cut of 0.32 mm settings in
turning of AISI 52100 hardened steel. An attempt has also
been made by Gaitonde et al. (2009) to analyze the effects
of depth of cut and machining time on machinability aspects
such as machining force, power, specic cutting force, sur-
face roughness and tool wear using second order mathemat-
ical models during turning of high chromium AISI D2 cold
work tool steel with CC650, CC650WG and GC6050WH
ceramic inserts. From parametric analysis, it was revealed
that the CC650WG wiper insert performed better with ref-
erence to surface roughness and tool wear. Basheer et al.
(2008) presented an experimental work on the analysis of
machined surface quality on Al/SiCp composites leading to
an ANN-based model to predict the surface roughness. The
model showed high coefcient of correlation of R = 0.977
and mean absolute error of 10.4% between the actual and
predicted values. The best surface quality was obtained at
lowest value of feed rate, the smaller particle size and the
largest tool-nose radius parameter combinations. Ezugwu
et al. (2005) used a three layered feed forward backprop-
agation neural network for the analysis and prediction of
the tangential force, axial force, spindle motor power con-
sumption, machined surface roughness, average ank wear,
maximum ank wear and nose wear during high-speed turn-
ing of nickel-based, Inconel 718 alloy using CNMG 120412
inserts. Correlation coefcients between the model predic-
tions and experimental values were ranging from 0.6595 for
cutting force to 0.9976 for nose wear prediction. Al-Ahmari
(2007) compared the performance of multiple linear regres-
sion analysis (RA), RSMand computational neural networks
123
J Intell Manuf (2012) 23:639650 641
(CNN) models in prediction of tool life, cutting force and
surface roughness in turning austenitic AISI 302 stainless
steels. Descriptive statistics and hypothesis tests were con-
ducted to study and it was found that the CNN models were
better than RA and RSM models. zel and Karpat (2005)
utilized neural network modeling to predict surface rough-
ness and tool ank wear over the machining time for veriety
of cutting conditions in nish hard turning of hardened AISI
H-13 steels using cubic boron nitride (CBN) tools. Regres-
sion models were also developed in order to capture process
specic parameters. Predictive ANN models were found to
be capable of better predictions for machining outputs than
regression model.
As can be seen from the literature, articial intelligence
tools like GA, ANN, regression modeling and fuzzy logic are
found to be extremely used in modeling of surface nishing
of turned parts. Although a broad literature survey has been
conducted in turning of different materials under different
turning parameters, little attention has been paid to turning of
AISI 304 austenitic stainless steels. Additionally, the predict-
ing results of used techniques seem to have some disadvan-
tages, i.e. low convergence rate, obvious over tting, and
especially poor generalisation when few samples are avail-
able (Shi andGindy2007). Inthis context, this studybecomes
necessary to model the surface roughness during machining
AISI 304 material by using support vector machines (SVM),
aneffective articial iltelligence tool, different fromthe novel
literature. SVM based on statistical learning theory is a new
achievement in the eld of data-driven modeling and imple-
mented succesfully in classication, regression and function
estimation (Kwok 1999; Cao and Tay 2003).
Austenitic stainless steel machining
The AISI 300 series of austenitic stainless steels represent
the largest group of steels in use of total (OSullivan and
Cotterell 2002). However, their machinability is more dif-
cult than the other alloy steels due to the reasons such
as having low heat conductivity, built-up edge (BUE) ten-
dency and high work hardening. Many attemps have been
made to improve the machinability of austenitic stainless
steels. Xavior and Adithan (2009) used three types of cut-
ting uids, namely coconut oil, emulsion and neat cutting
oil (immiscible with water) to determine their effects on
tool wear and surface roughness during turning of AISI
304 with carbide tools. Tekiner and Ye silyurt (2004) deter-
mined the best cutting parameters in the AISI 304 austen-
itic stainless steel machining by considering the acoustic
emission during the cutting process. Tool ank wear, BUE,
chip formations, surface roughness of machined parts and
machine power consumptions were investigated and used
for determine the optimal cutting speed and feed rate. Ciftci
(2006) presented an experimental work in dry turning of
austenitic stainless steels (AISI 304 and AISI 316) using
CVD multilayer coated cemented carbide tools. The inu-
ences of cutting speed, cutting tool coating top layer and
workpiece material were investigated on the machined
surface roughness and the cutting forces. The tool wear
characteristics were also examined under scanning electron
microscope (SEM). Korkut et al. (2004) investigated the
effect of cutting speed on tool wear and surface roughness in
turning of AISI 304 austenitic stainless steel. Correlation was
made between the tool wear/surface roughness and the chips
obtained at the cutting speeds of 120, 150 and 180 m/ min.
Apart from the introduction, it is possible to nd no
attempt on modeling the process parameters in turning of
AISI 304 austenitic stainless steels from the machinability
literature of these steels. The objective of this paper is to use
the SVM for estimate surface roughness in turning AISI 304
austenitic stainless steel. Three different toolboxes namely,
least square SVM(LS-SVM), Spider and SVM-KMbased on
clustering by k-means were used in modeling surface rough-
ness. Additionally, an ANN model has also been used to
model surface roughness to compare the results with SVMs.
Experimental setup and cutting conditions
Turning experiments were conducted in dry machining con-
ditions on a JOHNFORD TC-35 lathe machine with Fanuc
18-TCNCcontrol andprogrammabletailstockandmaximum
spindle speed of 3.500 rpmand a 15 kWdrive motor. Figure 1
shows the CNClathe machine where the turning is operated.
Cutting tools and workpiece materials
The cutting tools used were commercial grade cemented
carbide inserts produced by Kennametal with the geometry
of CNMG 120408. The inserts were recommended for
Fig. 1 Orthogonal cutting of AISI 304 austenitic stainless steel
123
642 J Intell Manuf (2012) 23:639650
Table 1 Chemical composition of AISI 304 austenitic stainless steel
Elements Composition (%wt.)
C 0.0452
Mn 1.78
Si 0.32
S 0.019
P 0.0397
Cr 18.35
Ni 8.9
Mo 0.456
Nb 0.0312
Cu 0.768
Ti 0.0001
V 0.06
Fe Balance
Table 2 Mechanical and physical properties of tested material
Density (1000 kg/m
3
) 8
Poissons ratio 0.28
Elastic modulus (GPa) 194
Tensile strength ( MPa) 515
Yield strength ( MPa) 206
Elongation (%) 41
Hardness (HRB) 87
Thermal expansion (10
6
/

C) 17.4
Thermal conductivity (W/m-k) 15.8
Electric resistivity (10
9
-m) 720
machining austenitic stainless steels by Kennametal. The
inserts were clamped mechanically on a rigit tool holder.
As far as possible, the tests were carried out in accordance
with ISO 3685. The workpiece material was AISI 304 au-
stenitic stainless steel. This steel nds its applications in air
craft ttings, aerospace componenets, such as shipbuilding,
shafts, valves, special screws, crygenic vessels and compo-
nents for severe chemical environments (Xavior and Adithan
2009). The workpiece specimens were 120 mm long and
40 mm in diameter. Table 1 shows the chemical composition
of the machined workpiece material while the physical and
mechanical properties are listed in Table 2.
Turning parameters and experimental design
The design of experiments (DOE) method is an important
step in selection of cutting parameters and their levels within
an effective and suitable experimental run. The correct choise
of the experimental design has a large inuence on the accu-
racy and the construction cost of the approximations. The 27
different combination of turning parameters were determined
Table 3 Turning parameters and their factor levels
Symbol Turning parameter Level 1 Level 2 Level 3
A Cutting speed (m/ min) 30 60 90
B Feed rate ( mm/rev) 0.15 0.25 0.35
C Depth of cut ( mm) 0.5 1 1.5
Table 4 Experimental results
Exp.
No
Feed rate
( mm/dev)
Cutting
speed
(m/dak)
Depth of
cut ( mm)
Surface
roughness
(m)
1 0.10 60 0.5 3.48
2 0.10 60 1.0 2.54
3 0.10 60 1.5 3.01
4 0.10 120 0.5 2.89
5 0.10 120 1 2.15
6 0.10 120 1.5 2.66
7 0.10 180 0.5 1.12
8 0.10 180 1 1.35
9 0.10 180 1.5 0.94
10 0.10 60 0.5 3.27
11 0.10 60 1 3.86
12 0.10 60 1.5 2.65
13 0.10 120 0.5 2.99
14 0.10 120 1 2.34
15 0.10 120 1.5 1.63
16 0.10 180 0.5 1.21
17 0.10 180 1 0.86
18 0.10 180 1.5 1.45
19 0.10 60 0.5 6.21
20 0.10 60 1 4.83
21 0.10 60 1.5 5.36
22 0.10 120 0.5 3.17
23 0.10 120 1 2.69
24 0.10 120 1.5 2.31
25 0.10 180 0.5 2.45
26 0.10 180 1 2.35
27 0.10 180 1.5 2.97
and tested by using a three-level full factorial experimental
design method to determine surface roughness of samples. In
this study, three independent variables, such as cutting speed,
feed rate and depth of cut were used as turning parameters.
The machining parameters and their factor levels are shown
in Table 3. Surface roughness measurements were carried out
on the machined surfaces using a Mitutoyo SJ-211 portable
device. Three measurements were made on the each surface
averaged for an accurate reading. Table 4 shows the experi-
mental results.
123
J Intell Manuf (2012) 23:639650 643
Support vector machines
SVM is a very useful soft computing method based on
statistical learning theory presented by Vladimir Vapnik
(Salat and Osowski 2004; Vapnik 1998). SVM try to
construct a separating hyperplane maximizing the margin
between two data sets according to their classes which
have been previously mapped to a high dimensional space
(Thukaram 2005). To determine the margin, two parallel
hyperplanes are constructed on each side of the separating
hyperplane. An optimal separation is achieved by the hyper-
plane which has the largest distance to the neighboring data
points of the both classes, because the larger the margin the
better the generalization error of the classier is obtained.
SVM has been used in many pattern recognition and regres-
sion estimation problems and has been applied to the prob-
lems of dependency estimation, forecasting, and constructing
intelligent machines (Ekici et al. 2009).
SVM for classication problems
SVMs have the potential to handle very large feature spaces,
because the training of SVM is carried out so that the dimen-
sion of classied vectors does not have as a distinct inuence
on the performance of SVM as it has on the performance
of conventional classiers. That is why it is noticed to be
especially efcient in large classication problems (Vapnik
1998). Also, SVM-based classiers are claimed to have good
generalization properties compared to conventional classi-
ers, because in training the SVM classier, the so-called
structural misclassication risk is to be minimized, whereas
traditional classiers are usually trained so that the empiri-
cal risk is minimized (Dash et al. 2007). The mathematical
expression of the SVM classier can be dened as below.
Let n-dimensional input x
i
(i =1, 2, ..., M), M be the
number of samples belong to class-1 or class-2, and asso-
ciated labels be y
i
=1 for class-1 and y
i
= 1 for class-
2 respectively. For linearly separable data a hyperplane
f (x) =0 which separates the data can be determined as,
f (x) = w
T
x +b =
n

j =1
w
j
x
j
+b = 0 (1)
where w is an n-dimensional vector and b is a scalar. The
vector w and the scalar b determine the position of the sep-
arating hyperplane. This separating hyperplane satises the
constraints f (x
i
) 1 if y
i
= 1 and f (x
i
) 1 if y
i
= 1
and this results in
y
i
f (x
i
)=y
i
_
w
T
x
i
+b
_
+1 for i = 1, 2, . . . , M (2)
The separating hyperplane that creates the maximum dis-
tance between the plane and the nearest data is called as
the optimal separating hyperplane as shown in Fig. 2. The
Class-2
m
Class-1
w
1 b x w = +
1 b x w = +
0 b x w = +
w
2
m =
H
H1
H2
SV
SV
SV
W
+
Fig. 2 Optimal separating hyper plane
geometrical margin is found to be w
2
(Dash et al. 2007).
Considering noise with the slack variables
i
and error
penalty C, the optimal hyperplane can be found by solving
the following convex quadratic optimization problem;
minimize
1
2
w
2
+C

M
i =1

i
(3)
subject to
y
i
_
w
T
x
i
+b
_
1
i
for i = 1, 2, . . . , M

i
0 for all i
(4)
where
i
is the distance between the margin and the exam-
ples x
i
lying on the wrong side of the margin. Applying the
Kuhn-Tucker conditions into the equivalent Lagrange dual
problem;
maximize
W () =
M

i =1

i

1
2
M

i,k=0

k
y
i
y
k
x
T
i
x
k
(5)
subjects to
M

i =1
y
i

i
= 0, C
i
0, i = 1, 2, . . . , M (6)
The number of variables in the dual problem gives the num-
ber of training data. Denoting the optimal solution of the
dual problem with

and w

the equality conditions in (2)


holds for the training input-output pair (x
i
, y
i
) only if the
associated

= 0. In this case, the training example x


i
is
called as support vector (SV). The number of SVs is con-
siderably lower than the number of training samples making
SVM computationally very efcient (Dash et al. 2007). The
value of the optimal bias b

is found from the geometry,


b

=
1
2

SVs
y
i

i
_
s
T
1
x
i
+s
T
2
x
i
_
(7)
123
644 J Intell Manuf (2012) 23:639650
where s
1
and s
2
are the arbitrary SVs for class-1 and
class-2 respectively. Only the samples associated with SVs
are summed because the other elements of the optimal
Lagrange multiplier

are equal to zero. The nal decision


function is given by
f (x) =

SVs

i
y
i
x
T
i
x +b

(8)
The unknown data sample x is then classied as,
x
_
Class-1, if f (x) 0
Class-2, otherwise
(9)
The nonlinear classication problems can also be solved by
using SVM applying a kernel function. The classied data is
mapped onto a high-dimensional feature space where the lin-
ear classication is possible. Using a nonlinear vector func-
tion,
(x) =
1
(x),
2
(x), . . . ,
m
(x), m n (10)
to map the n-dimensional input vector x into the
m-dimensional feature space, the linear decision function in
dual form is given by,
f (x) =

SVs

i
y
i

T
(x
i
) (x) (11)
Notice that in (11) as well as (4), the inner products are used.
A function that returns a dot product of the feature space
mapping of the original data points is called a kernel func-
tion K (x, z) =
T
(x)(x). The learning in the feature space
does not require the inner products where a kernel function
is applied. Using a kernel function, the decision function can
be written as,
f (x) =

SVs

i
y
i
K (x
i
, x) (12)
There are different kernel functions used in the literature.
Mercers theorem states that any symmetric positive-defi-
nite matrix can be regard as a kernel matrix. In this paper,
Gaussian radial basis kernel function which gives the best
results is selected. The radial basis kernel function is dened
as,
K (x, z) = exp
_

|x z|
2
2
2
_
(13)
where is the width of the Gaussian function and it will
be called as kernel parameter in this paper. The detailed
information about the SVM can be found in Vapnik (1998)
and Vojtech and Hlavac (2004).
SVM for regression problems
Although SVM is used for classication problems, it is
also an excellent tool for regression problems (Thukaram
2005; Smola et al. 1998; Hearst 1998). A version of a SVM
for regression was proposed in 1996 by Vladimir Vapnik,
Harris Drucker, Chris Burges, Linda Kaufman and Alex
Smola (Drucker et al. 1997). This method is called support
vector regression (SVR). The model produced by support
vector classication (as described above) only depends on
a subset of the training data, because the cost function for
building the model does not care about training points that
lie beyond the margin. Analogously, the model produced by
support vector regression only depends on a subset of the
trainingdata, because the cost functionfor buildingthe model
ignores any training data that are close (within a threshold )
to the model prediction.
In SVM regression, the learning task is transformed to the
minimization of the error function, dened through the so
called -insensitive loss function which controls the accu-
racy of the regressor. The variables measure the cost of the
errors on the training points. These are zero for all points that
are inside the band.
L

(d, y(x)) ;
L

(d, y(x)) = {|d y(x)| , for |d y(x)|


0, for |d y(x)| <
(14)
where is the assumed accuracy, d is the destination, x is the
input vector which includes cutting speeds, feed rates and
depths of cut, and y(x) is the actual output of the network
(surface roughness). The actual output of the SVM network
is dened by
y(x) =
K

j =0

j
(x) = w
T
(x) (15)
where w = [w
0
, w
1
, . . . w
k
]
T
is the weight vector and
(x) = [
0
(x),
1
(x), . . . ,
k
(x)]
T
the basis function vec-
tor. The learning task is dened as the minimization problem
of the error function E (Salat and Osowski 2004).
E =
1
p
p

i =1
L

(d
i
, y (x
i
)) (16)
at the upper bound on the weight vector, w
2
< C
0
, where
C
0
is a user-specied constant and p is the number of learning
data pairs (x
i
, d
i
). Introducing the slack variables
i
and

i
,
the learning problem can be redened as the minimization of
the cost function

_
w, ,

_
= C
_
p

i =1
_

i
+

i
_
_
+
1
2
w
T
w (17)
at the following functional and boundary constraints:
d
i
w
T
(x
i
) +
i
w
T
(x
i
) d
i
+

i
,

i
0
(18)
123
J Intell Manuf (2012) 23:639650 645
The two variables and C are free parameters that control
the dimension of the approximating function. Both must be
selected by the user (Salat and Osowski 2004; Burges 1998;
Zhang 2004). The optimization problem is solved by the
introduction of the Lagrangian function and Lagrange mul-
tipliers
i
,

i
(i = 1, 2, ....., p) responsible for functional
constraints. The minimization of the Lagrangian function has
been transformed to the so-called dual problem,
max
p

i =1
d
i
_

i
_

i =1
_

i
+

i
_

1
2
p

i =1
p

j =1
_

i
_
_

j
_
K
_
x
i
, x
j
_
(19)
at the constraints
p

i =1
_

i
_
0
0
i
C,
0

i
C
(20)
where K
_
x
i
, x
j
_
=
T
(x
i
)
_
x
j
_
is a inner-product kernel
dened in accordance with Mercers theorem on the basis of
the learning data set. Note that the dual problem belongs to
the quadratic programming optimization tasks with respect
to the Lagrange multipliers. The solution of it is relatively
easy and leads to the global minimum. After solving the dual
problem, the optimum solution for the vector w is given by
w =
N
s

i =1
_

i 0

0i
_
(x
i
) (21)
where N
s
= K is the number of so-called support vectors
(equal to the number of nonzero Lagrange multipliers). The
network output signal y(x) can be expressed through the
Lagrange multipliers and kernel function K (x, x
i
) (Salat and
Osowski 2004);
y(x) =
Ns

i =1
_

i 0

0i
_
K (x, x
i
) +
0
(22)
and does not need to know the explicit form of the nonlin-
ear basis function (x). The most known kernel functions
used in practice are radial (Gaussian), polynomial, spline,
or sigmoidal functions (Salat and Osowski 2004). The most
important is the choice of coefcients (kernel parameter)
and C for obtaining the optimum network. The constant C is
the weight, determining the balance between the complex-
ity of the network, characterized by the weight vector w and
the error of approximation, measured by the slack variables

i
(i = 1, 2, . . . , p). For the normalized input signals, the
value of C is usually adjusted as much bigger than 1. In this
paper, these parameters will be adjusted by testing a lot of
different values to be found the best results.
Articial neural networks
ANNs are very efcient on adaptation and learning and for
this reason they are used as modeling tools in a number of
applications. An ANN is made of three types of layers: an
input layer which accepts the input variables. Hidden layers
has some number of neurons, and an output layer made of
one neuron that in the case examined herein gives the sur-
face roughness Ra. Hidden and output layers are composed
of some number of neurons that perform a specic nonlinear
function. The neurons of one layer are interconnected to the
neurons of the pre and after layers through weighted links.
Each neuron of the hidden and output layers is offset by a
threshold value (Sharma et al. 2008).
The mathematical model of an articial neurons behavior
is the simplication of the biological brain neuron as shown
in Fig. 3. Various inputs x(n) to the network multiplied by
weights w(n) are sent to a neuron. Performing accumulation
and threshold, the neuron sums the weighted inputs, passes
the result through a non-linear transfer function and provides
an output Yi (Zhang et al. 2007; Chang et al. 2010)
Y
i
= f
_
n1

i =0
w
i
x
i

_
(23)
where the inputs of x
i
in this study corresponds to cutting
speed, feed rate, and depth of cut; is the internal threshold
or offset of a neuron; and f is the non-linear transfer function.
The most commonly used f is dened by the sigmoid logistic
function as:
f (x) =
1
1 +e
x
(24)
the weights are dynamically updated using the BP algorithm.
The difference between the target output and actual output
(learning error) for a sample p is (Tosun and Ozler 2002)
E
p
=
1
2
K

k=1
_
d
pk
o
pk
_
2
(25)
where d
pk
and o
pk
are the desired and calculated output for
kth output, respectively. K denotes the number of neuron in
output of network. The average error for whole system is
obtained by (ayda s and Hasalk 2008):

1
1
n
i
i i
x w
X
0
X
1
X
2
W
1
W
2
W
n-1
Input
Output Transfer
function
Fig. 3 The behavior of an articial neuron
123
646 J Intell Manuf (2012) 23:639650
E
p
=
1
2
P

p=1
K

k=1
_
d
pk
o
pk
_
2
(26)
where P is the total number of instances. For the purpose
of minimizing E
p
, the weights of the inter-connections are
adjusted during the training procedure until the expected
error is achieved. To adjust the weights of the networks, the
process starts at the output neuron and works backward to the
hidden layer. The weights in BP based on the delta learning
rule can be expressed as follows:
w
new
= w
old
+w (27)
w =
E
p
w
out
j
(28)
where out
j
is the j th neuron output, is the learning rate
parameter controlling stability and rate of convergence of
the network, which is a constant between 0 and 1. Once the
weights of all the links of the network are decided, the deci-
sion mechanism is then developed.
Estimation of surface roughness using SVM regression
and ANN
To estimate the surface roughness, 27 experiments were per-
formed rstly. In each experiment; cutting speed, feed rate
and depth of cut were changed and the surface roughness was
recorded. The cutting speed, feed rate and depth of cut were
selected as 30-60-90, 0.15-0.25-0.35 and 0.5-1-1.15 respec-
tively. Three parameters including cutting speed, feed rate
and depth of cut were selected as the inputs (x
i
) of SVM
while the surface roughness was the output (y
i
).
All of the computer implementations were performed by
using MATLAB software (Demuth et al. 2006). In the lit-
erature, there are several SVM toolboxes for MATLAB.
To evaluate the performances of different SVM toolboxes,
least square SVM (LS-SVM) (Suykens et al. 2002),
Spider (Weston et al. 2006) and SVM-KM(Canu et al. 2005)
based on clustering by k-means were experimented. Gauss-
ian radial basis function which was used most often in the
literature was selected as the kernel function. As mentioned
above, the free parameters and C for the kernel function
should be adjusted by users to obtain the best SVM outputs.
Grid search is the simplest way to determine the values for
parameters and C. By setting the upper and lower bounds
(search interval) and the jumping interval in the search, we
can nd sets of values for parameters and C which lead to
the highest accuracy rate of classication in this interval. In
this study, search intervals were selected as [0 10] for kernel
parameter and [0 1000] for the regularization parameter C.
The jumping intervals were adjusted as 0.1 and 10 for and
C respectively.
Fig. 4 Grid search for LS-SVM
1 2 3 4 5 6 7 8 9
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Number of Experiment
S
u
r
f
a
c
e

R
o
u
g
h
n
e
s
s
Estimated
Experimental
Fig. 5 Comparison of experimental and estimated results of LS-SVM
The data set obtained from the experiments was divided
the two groups for training and testing of SVM. One data
set obtained from 18 experiments was used for training and
the rest of them (9 experiment) which are not introduced to
the network before were used for testing of SVM. The grid
search results of LS-SVM are shown in Fig. 4.
As can be shown in Fig. 4, the better results are obtained
while > 0.5 and C > 120. After grid search, the optimum
parameters are found as = 7.7 and C = 670 when the sum
of average error is equal to 0.2997. The comparison of real
and estimated surface roughness results which are obtained
from the best SVM structure constructed by the optimum
free parameters are shown in Fig. 5.
The grid search results of Spider are shown in Fig. 6. The
better results are obtained from the values of > 0.5 and
C > 40. The minimum sum of average error is 0.2748 by
adjusting > 3.9 and C > 900. The comparison of the
experimental and estimated results is illustrated in Fig. 7.
Finally, a grid search implementation was performed for
the SVM-KM. The grid search results for SVM-KMare illus-
trated in Fig. 8. As can be shown in Fig. 8, better test results
are obtained by adjusting > 0.4 and C > 10. The best
results are obtained from the testing when the free param-
eters are selected as = 0.5 and C = 30. By using these
parameter values, the sumof average error is equal to 0.2832.
123
J Intell Manuf (2012) 23:639650 647
Fig. 6 Grid search for spider
1 2 3 4 5 6 7 8 9
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Number of Experiment
S
u
r
f
a
c
e

R
o
u
g
h
n
e
s
s
Estimated
Experimental
Fig. 7 Comparison of experimental and estimated results of spider
Fig. 8 Grid search for SVM-KM
The comparison results of the experimental and estimated
data belonging to surface roughness are shown in Fig. 9.
All of the grid search applications show that the selec-
tion of the optimum parameters is very important for the
performance of SVM. It can be said that the selection of ker-
nel parameter as 0.4 and C > 1 is very effective to be
obtained the best performance of SVM. By adjusting the opti-
mal free parameters, LS-SVM, Spider and SVM-KM give
the similar results which have the acceptable error tolerance.
To investigate the performance of articial neural networks
(ANN), the same data set is used as inputs of ANN. A multi
layered feed forward neural network with the back-propaga-
tion algorithm is employed. The multi layered network has
1 2 3 4 5 6 7 8 9
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Number of Experiment
S
u
r
f
a
c
e

R
o
u
g
h
n
e
s
s
Estimated
Experimental
Fig. 9 Comparison of experimental and estimated results of SVM-KM
0 5 10 15 20 25
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Number of hidden layer neurons
A
v
e
r
a
g
e

e
r
r
o
r
Fig. 10 Performance of ANNs with different hidden layer neurons
one input, one hidden and one output layer. The network
architecture has 18 inputs, and 1 output neurons. Number
of hidden layer neurons is varied from 1 to 25. The perfor-
mances of these networks which have different number of
neurons in hidden layers are shown in Fig. 10. As shown in
Fig. 10, the best test results are obtained from the network
which has 15 hidden neurons in the hidden layer. Hyperbolic
tangent sigmoid function is selected as the transfer function
for input and hidden layer. Linear transfer function is used in
the output layer. The network is trained by using the Leven-
berg-Marquardt training function. After 1000 epochs, train-
ing of the network is stopped and the testing results of this
network are shown in Fig. 11.
It is observed from the Fig. 11 that results of ANN are
worse than the results of SVMs. On the other hand, the train-
ing time of ANN is very long (122.93 s) whereas the time
needed for the training of SVM is very short (only a few sec-
onds). The average errors and elapsed times for each SVMs
and ANNs are shown in Table 5.
The testing results of all SVMs and ANNs are compared
with the experimental results as can be shown in Table 6. To
123
648 J Intell Manuf (2012) 23:639650
1 2 3 4 5 6 7 8 9
0
1
2
3
4
5
6
Number of Experiment
S
u
r
f
a
c
e

R
o
u
g
h
n
e
s
s
Estimated
Experimental
Fig. 11 Comparison of experimental and estimated results of ANN
Table 5 Elapsed time and average errors for SVMs and ANN
LS-SVM Spider SVM-KM ANN
Average error 0.2997 0.2748 0.2832 0.3275
Elapsed time (s) 0.45 0.42 0.25 122.93
compare the results clearly, a post-process operation is per-
formed by using postreg command of MATLAB. Postreg
command post-processes the network training set by per-
forming a linear regression between each element of the net-
work response and the corresponding target. Where R is the
regression value (R = 1 means perfect correlation) (Sharma
et al. 2008). The post-regression results are illustrated in the
Fig. 12. The post-regression results show that the R values
of the SVMs are better than the R value of the ANN.
Table 6 Comparison of SVMs and ANN results with experimental
ndings
Experimental LS-SVM Spider SVM-KM ANN
2.5400 2.7519 3.2399 2.9430 2.9435
2.1500 2.2457 2.5437 2.6151 2.7495
1.3500 1.1037 1.0326 1.0177 1.0359
3.8600 3.5411 3.8731 3.6891 2.9905
2.3400 2.2062 2.6080 2.5215 2.2527
0.8600 1.6884 1.5769 1.5380 1.3245
4.8300 4.7266 4.8229 4.8105 5.7824
2.6900 2.1497 2.7047 2.4034 3.0646
2.3500 2.5687 2.3941 2.3714 2.7639
Statistical paired t test
In order to validate the modeling results of SVM and ANN
with the experimental results, the statistical paired t test was
employed. SPSS package programwas used for this purpose.
The results of test are summarized in Table 7. As can be seen
from this table, all pairs are significance at 0.001 levels with
high correlation coefcients. The most correlated pair was
pair 2 (spider SVM). Pairs 3 (SVM-KM), 1 (LSSVM) and 4
(ANN) were followed.
Concluding remarks
In this study, the experimental observations were incorpo-
rated into the SVMs models for orthogonal cutting of AISI
0 1 2 3 4 5
0
2
4
6
LS-SVM
Y
=
(
0
.
8
4
)
T
+
(
0
.
4
2
)
Outputs vs. Targets (T), R=0.94496
0 1 2 3 4 5
0
2
4
6
Spider
Y
=
(
0
.
9
1
)
T
+
(
0
.
4
4
)
Outputs vs. Targets (T), R=0.95715
0 1 2 3 4 5
0
2
4
6
SVM-KM
Y
=
(
0
.
8
9
)
T
+
(
0
.
3
9
)
Outputs vs. Targets, R=0.95602
0 1 2 3 4 5
0
2
4
6
ANN
Y
=
(
1
)
T
+
(
0
.
1
3
)
Outputs vs. Targets (T), R=0.91451
data points
best linear fit
Y=T
data points
best linear fit
Y=T
data points
best linear fit
Y=T
data points
best linear fit
Y=T
Fig. 12 Post-regression results
123
J Intell Manuf (2012) 23:639650 649
Table 7 Paired samples t -test
N Correlation Mean t Sig.
Pair 1 EXP & LSSVM 9 .945 .0013333 .010 .000
Pair 2 EXP & SPIDER 9 .957 .2028778 1.754 .000
Pair 3 EXP & SVMKM 9 .956 .1044111 .888 .000
Pair 4 EXP & ANN 9 .915 .2152778 1.179 .001
304 austenitic stainles steels. Three different SVMs mod-
els were developed to predict surface roughness after the
turning process. Additionally, a multi layered feedforward
backpropagation ANN model was also used for the same
purpose. Spider SVM was found the most predictive model
with R of 0.95715, while the SVM-KM (R = 0.95602) was
the second accuracy model and the LS-SVM was the least
(R = 0.94496). All the used SVM predictions were better
than ANN results (R = 0.91451). The ANN model involves
more computationally time than SVMss.
References
Al-Ahmari, A. M. A. (2007). Predictive machinability models for a
selected hard material in turning operations. Journal of Materials
Processing Technology, 190, 305311.
Ba gc, E., & I sk, B. (2006). Investigation of surface roughness in
turning unidirectional GFRP composites by using RS methodol-
ogy and ANN. International Journal of Advanced Manufacturing
Technology, 31, 1017.
Basheer, A. C., Dabade, U. A., Joshi, S. S., Bhanuprasad, V. V., & Gad-
re, V. M. (2008). Modeling of surface roughness in precison
machining of metal matrix composites using ANN. Journal of
Materials Processing Technology, 197, 439444.
Benardos, P. G., & Vosniakos, G. C. (2003). Predicting surface rough-
ness in machining: A review. International Journal of Machine
Tools and Manufacture, 43, 833844.
Burges, C. J. C. (1998). A tutorial on support vector machines for pat-
tern recognition. Data Mining and Knowledge Discovery, 2, 121
167.
Canu, S., Grandvalet, Y., Guigue, V., & Rakotomamonjy, A. (2005).
SVM and Kernel Methods Matlab Toolbox, Perception Systmes
et Information, INSA de Rouen, Rouen, France.
Cao, L. J., & Tay, F. E. H. (2003). Support vector machine with
adaptive parameters in nancial time series forecasting. IEEE
Transactions on Neural Networks, 14, 15061518.
ayda s, U., & Hasalk, A. (2008). A study on surface roughness
in abrasive waterjet machining process using articial neural
networks and regression analysis method. Journal of Materials
Processing Technology, 202, 574582.
Chang, P. C., Lin, J. J., & Dzan, W. Y. (2010). Forecasting of manu-
facturing cost in mobile phone products by case-based reasoning
and articial neural network models. Journal of Intelligent Man-
ufacturing. doi:10.1007/s10845-010-0390-7.
Ciftci, I. (2006). Machining of austenitic stainless steels using CVD
multi-layer coated cemented carbide tools. Tribology Interna-
tional, 39, 565569.
Dash, P. K., Samantaray, S. R., & Ganapati, P. (2007). Fault classica-
tion and section identication of an advanced series-compensated
transmission line using support vector machine. IEEE Transac-
tions on Power Delivery, 22, 6773.
Davim, J. P., Gaitonde, V. N., & Karnik, S. R. (2008). Investiga-
tion into the effect of cutting conditions on surface roughness
in turning of free machining steel by ANNmodels. Journal of
Materials Processing Technology, 205, 1623.
Demuth, H., Beale, M., & Hagan, M. (2006). Neural network toolbox
5, users guide (pp. 916). MA: The Mathworks, Inc.
Drucker, H., Burges, C. J. C., Kaufman, L., Smola, A., & Vapnik, V.
(1997). Support vector regression machines. Advances in Neural
Information Processing Systems 9, NIPS 1996, 155161, MIT
Press.
Ekici, S., Yldrm, S., & Poyraz, M. (2009). A pattern recogni-
tion application for distance protection. Journal of The Fac-
ulty of Engineering and Architecture of Gazi University 24,1,
5156
Ezugwu, E. O., Fadare, D. A., Bonney, J., Da Silva, R. B., & Sales, W.
F. (2005). Modeling the correlation between cutting and process
parameters in high-speed machining of Inconel 718 alloy using an
articial neural network. International Journal of Machine Tools
and Manufacture, 45, 13751385.
Gaitonde, V. N., Karnik, S. R., Figueira, L., & Davim, J. P. (2009)
Machinability investigations in hard turning of AISI D2 cold
work tool steel with conventional and wiper ceramic inserts.
International Journal of Refractory Metals and Hard Materials
(in press).
Grzesik, W., & Brol, S. (2003). Hybrid approach to surface rough-
ness evaluation in multistage machining processes. Journal of
Materials Processing Technology, 134, 265272.
Hearst, M. A. (1998). Support vector machines. IEEE Intelligent
Systems 1821.
Karayel, D. (2008). Prediction and control of surface roughness in
CNC lathe using articial neural network. Journal of Materials
Processing Technology.
Korkut, I., Kasap, M., Ciftci, I., & Seker, U. (2004). Determination
of optimum cutting parameters during machining of AISI 304
austenitic stainless steel. Materials and Design, 25, 303305.
Kwok, J. (1999). Moderating the outputs of support vector machine
classier. IEEETransactions on Neural Networks, 10, 10181031.
Lee, B. Y., & Tarng, Y.S. (2001). Surface roughness inspection by
computer vision in turning operations. International Journal of
Machine Tools and Manufacture, 41, 12511263.
Muthukrishnan, N., & Davim, J. P. (2009). Optimization of machin-
ing parameters of Al/SiC-MMC with ANOVA and ANN anal-
ysis. Journal of Materials Processing Technology, 209, 225232.
Nalbant, M., Gkkaya, H., Tokta s, I., & Sur, G. (2009). The exper-
imental investigation of the effects of uncoated, PVD- and
CVD-coated cemented carbide inserts and cutting parameters
on surface roughness in CNC turning and its prediction using
articial neural network. Robotics and Computer Integrated Man-
ufacturing, 25, 211223.
OSullivan, D., & Cotterell, M. (2002). Machinability of austenitic
stainless steel SS 303. Journal of Materials Processing Technol-
ogy, 124, 153159.
zel, T., & Karpat, Y. (2005). Predictive modeling of surface rough-
ness and tool wear in hard turning using regression and neural
network. International Journal of Machine Tools and Manufac-
ture, 45, 467479.
zel, T., Karpat, Y., Figueira, L., & Davim, J. P. (2007). Modeling
of surface nish and tool ank wear in turning of AISI D2
steel with ceramic wiper inserts. Journal of Materials Processing
Technology, 189, 192198.
Paiva, A. P., Ferreira, J. R., & Balestrassi, P. P. (2007). A multi-
variate hybrid approach applied to AISI 52100 hardened steel
turning optimization. Journal of Materials Processing Technol-
ogy, 189, 2635.
123
650 J Intell Manuf (2012) 23:639650
Roy, S. S. (2006). Design of genetic-fuzzy expert system for predicting
surface nish in ultra-precison diamond turning of metal matrix
composite. Journal of Materials ProcessingTechnology, 173, 337
344.
Salat, R., & Osowski, S. (2004). Accurate fault location in the power
transmission line using support vector machine approach. IEEE
Transactions on Power Systems, 19, 879886.
Sharma, V. S., Dhiman, S., Schgal, R., &Sharma, S. K. (2008). Estima-
tion of cutting forces and surface roughness for hard turning using
neural networks. Journal of Intelligent Manufacturing, 19, 473
483.
Shi, D., & Gindy, N. N. (2007). Tool wear predictive model based
on least squares support vector machines. Mechanical Systems
and Signal Processing, 21, 17991814.
Smola, A. J., & Schlkopf, B. (1998). A tutorial on support vec-
tor regression. Technical Report NC2-TR-1998-030, ESPRIT
Working Group in Neural and Computational Learning.
Suresh, P. V. S., Rao, P. V., & Deshmukh, S. G. (2002). A genetic
algorithmic approach for optimization of surface roughness pre-
diction model. International Journal of Machine Tools and Man-
ufacture, 42, 675680.
Surjya, K. P., & Chakraborty, D. (2005). Surface roughness prediction
in turning using articial neural network. Neural Computing and
Applications, 14, 319324.
Suykens, J. A. K., Van Gestel, T., De Brabanter, J., De Moor, B., &
Vandewalle, J. (2002). Least squares support vector machines.
World Scientic, Singapore, (ISBN 981-238-151-1).
Tekiner, Z., & Ye silyurt, S. (2004). Investigation of the cutting param-
eters depending on process sound during turning of AISI 304
austenitic stainless steel. Materials and Design, 25, 507513.
Thukaram, D., Khincha, H. P., & Vijaynarasimha, H. P. (2005). Arti-
cial neural network and support vector machine approach for
locating faults in radial distribution systems. IEEE Transactions
on Power Delivery, 20, 710721.
Tosun, N., & Ozler, L. (2002). A study of tool life in hot machin-
ing using articial neural networks and regression analysis
method. Journal of Materials Processing Technology, 124, 99
104.
Umbrello, D., Ambrogio, G., Filice, L., & Shivpuri, R. (2008). A
hybrid nite element methodArticial neural network approach
for predicting residual stress and the optimal cutting conditions
during hard turning of AISI 52100 bearing steel. Materials and
Design, 29, 873883.
Vapnik, V. (1998). The support vector method of function estima-
tion. In J. Suykens, J. Vandewalle (Eds.), Nonlinear modeling:
Advanced black-ox techniques (pp. 5586). Dordrecht: Kluwer
Vojtech, F., & Hlavac, V. (2004). Statistical pattern recognition tool-
box for MATLAB (SPRTOOL). Users Guide; http://cmp.felk.
cvut.cz/cmp/cmpsoftware.html.
Weston, J., Elisseeff, A., Bakr, G., & Sinz, F. (2006). The spider
software package. http://www.kyb.tuebingen.mpg.de/bs/people/
spider.
Wong, S. V., & Hamouda, A. M. S. (2003). Machinability data rep-
resentation with articial neural network. Journal of Materials
Processing Technology, 138, 538544.
Xavior, M. A., & Adithan, M. (2009). Determining the inuance of
cutting uids on tool wear and surface roughness during turn-
ing of AISI 304 austenitic stainless steel. Journal of Materials
Processing Technology, 209, 900909.
Zhang, J. H. R. (2004). A new algorithm of improving fault loca-
tion based on SVM. Eighth IEE International Conference on
Developments in Power System Protection, 1, 204207.
Zhang, J. Z., Chen, J. C., & Kirby, E. D. (2007). The development of an
in-process surface roughness adaptive control system in turning
operations. Journal of Intelligent Manufacturing, 18, 301311.
123

Das könnte Ihnen auch gefallen