International Journal of Machine Tools & Manufacture 43 (2003) 833–844

Predicting surface roughness in machining: a review

P.G. Benardos, G.-C. Vosniakos ∗
National Technical University of Athens, School of Mechanical Engineering, Manufacturing Technology Division, Iroon Polytexneiou 9,
157 80 Zografou, Athens, Greece

Received 20 December 2002

Abstract

The general manufacturing problem can be described as the achievement of a predefined product quality with given equipment,
cost and time constraints. Unfortunately, for some quality characteristics of a product such as surface roughness it is hard to ensure
that these requirements will be met. This paper aims at presenting the various methodologies and practices that are being employed
for the prediction of surface roughness.
The resulting benefits allow for the manufacturing process to become more productive and competitive and at the same time to
reduce any re-processing of the machined workpiece so as to satisfy the technical specifications. Each approach with its advantages
and disadvantages is outlined and the present and future trends are discussed. The approaches are classified into those based on
machining theory, experimental investigation, designed experiments and artificial intelligence (AI).
 2003 Elsevier Science Ltd. All rights reserved.

Keywords: Surface roughness; Surface roughness prediction; Machining; Review

1. Introduction of computer controlled machine tools, have brought up

There are two main practical problems that engineers
face in a manufacturing process. The first is to determine
the values of the process’ parameters that will yield the
desired product quality (meet technical specifications)
and the second is to maximize manufacturing system
performance using the available resources. The decisions
made by manufacturing engineers are based not only on
their experience and expertise but also on conventions
regarding the phenomena that take place during pro-
cessing. In the machining field, many of these phenom-
ena are highly complex and interact with a large number
of factors, thus preventing high process performance
from being attained. To overcome these problems, the
researchers propose models that try to simulate the con-
ditions during machining and establish cause and effect
relationships between various factors and desired pro-
duct characteristics. Furthermore, the technological
advances in the field, for instance the ever-growing use
∗
Corresponding author. Tel.: +30-1-772-1457; fax: +30-1-772-
1197.
E-mail address: vosniak@central.ntua.gr (G.-C. Vosniakos).
new issues to deal with, which further emphasize the
need for more precise predictive models.
Surface roughness is a widely used index of product
quality and in most cases a technical requirement for
mechanical products. Achieving the desired surface
quality is of great importance for the functional behavior
of a part. On the other hand, the process dependent nat-
ure of the surface roughness formation mechanism along
with the numerous uncontrollable factors that influence
pertinent phenomena, make almost impossible a straight-
forward solution. The most common strategy involves
the selection of conservative process parameters, which
neither guarantees the achievement of the desired surface
finish nor attains high metal removal rates.
The aim of this work is to present and discuss the
various methodologies and strategies that are adopted by
researchers in order to predict surface roughness. The
presentation of each approach along with its advantages
and disadvantages should help both the researchers and
practitioners by providing compact yet adequate infor-
mation so as to select the one that best suits their needs
and specific requirements. The paper focuses on turning
and milling since these are the most common cutting
processes and the majority of the published work applies

0890-6955/03/$ - see front matter  2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0890-6955(03)00059-2
834 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
to them. Surprisingly enough a similar review was not
found to be published before.
Surface roughness refers to deviation from the nomi-
nal surface of the third up to sixth order. Order of devi-
ation is defined in international standards [1]. First- and
second-order deviations refer to form, i.e. flatness, circu-
larity, etc. and to waviness, respectively, and are due
to machine tool errors, deformation of the workpiece,
erroneous setups and clamping, vibration and workpiece
material inhomogenities. Third- and fourth-order devi-
ations refer to periodic grooves, and to cracks and dilapi-
dations, which are connected to the shape and condition
of the cutting edges, chip formation and process kin-
ematics. Fifth- and sixth-order deviations refer to work-
piece material structure, which is connected to physical–
chemical mechanisms acting on a grain and lattice scale
(slip, diffusion, oxidation, residual stress, etc.). Different
order deviations are superimposed and form the surface
roughness profile, see Fig. 1.
2. Classification of approaches
The classification of the selected papers was not easy
due to two main reasons. First, there are many papers
that do not strictly follow a certain methodology in its
entirety, they rather select some of its basic principles
and combine them into a ‘new’ approach. Secondly,
there are many cases where researchers blend different
strategies into a single approach and therefore no single
classification would be entirely accurate.
Taking into account the above, four major categories
were created to classify the selected papers. These are:
(i) approaches that are based on machining theory to
develop analytical models and/or computer algorithms
to represent the machined surface; (ii) approaches that
Fig. 1. Surface form deviations (adapted from Ref. [1]).
examine the effects of various factors through the
execution of experiments and the analysis of the results;
(iii) approaches that use designed experiments; and (iv)
artificial intelligence (AI) approaches.
3. Machining theory based approach
This category includes approaches that place emphasis
on certain aspects from the theory of machining such as
process kinematics, cutting tool properties, chip forma-
tion mechanism etc. Computer-aided design (CAD)
methods and tools are utilized so as to achieve the goal
of building a model that will be able to simulate the
creation of the machined surface profile, thus visualizing
surface topography and assessing surface roughness.
In general, geometric model development forms the
basis of the approach through rigorous mathematical
equations. This model is then implemented by a com-
puter algorithm in order to handle the complex calcu-
lations. Also, there are some theoretical models that
relate surface roughness to cutting conditions such as the
feed rate [2]. These models are generally not accurate
so their improvement with the introduction of additional
parameters is examined by researchers.
In any case, experiments are conducted in order to
compare the predicted with the actual results. It must be
noted that despite the strong background of the afore-
mentioned theories, the phenomena that lead to the for-
mation of surface roughness are very complex and inter-
acting in nature so a comprehensive solution has not yet
been found.
The theoretical background used by the research
efforts in this category is considered a prerequisite for
anyone who is involved in machining studies and there-
fore no analytical description is presented in this paper.
However, work by Kaczmarek [3] can be used for refer-
ence purposes.
3.1. Machining theory studies
Grzesik [4] used the minimum undeformed chip thick-
ness to predict surface roughness in turning. The mol-
ecular–mechanical theory of friction (Kragelskii’s
theory) and the Hencky–Ilyushin theory of plasticity
were used to mathematically model the tribological
effects at the chip–cutting tool interface. The approach
was based on the assumption that the difference between
the theoretical and measured surface roughness values is
due to adhesion at the chip–cutting tool interface and
that the minimum undeformed chip thickness corre-
sponds to the transition from ploughing to microcutting.
Consequently, an existing model for predicting the
roughness of a turned surface was improved and the dif-
ference between the measurements and predicted results
was markedly reduced.
 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
In Ref. [5], a surface topography simulation model
was established to simulate the surface finish profile gen-
erated after a turning operation with known vibration
characteristics. The model incorporated the effects of
tool geometry, cutting parameters and the relative
motion between the cutting tool and the workpiece on
the surface finish profile, which was ‘decomposed’ into
three directions, namely, the radial, tangential and axial
direction. The vibration frequency ratio (FR), which was
defined as the ratio of vibration frequency (in Hz) over
spindle rotational speed (in rps), was thought to influ-
ence the period of the surface waviness along the axial
direction. It was also found that the effects of the radial
direction vibrations on the surface roughness measures
were much more significant than those of either the tan-
gential direction vibrations or the axial direction
vibrations as would be expected.
The study of Baek et al. [6] presented a surface rough-
ness model for face-milling operations considering the
profile and the runout error of each insert in the cutter
body. It was stated that because of manufacturing errors
in making the cutters, axial (affecting the depth of cut)
and radial (affecting the surface roughness) runout errors
exist. The feed rate was also taken into account so as to
formulate a geometric model. After the model validation
with experimental cutting data, the material removal rate
was maximized through optimization of the feed rate
with the surface roughness as a constraint by means of
a bisection optimization algorithm.
The computer modeling technique developed in Ref.
[7] could represent the spectrum of components of sur-
face topography ranging over shape, waviness and
roughness in a way suitable for generating macro- and
micro-level finishing position commands. In order to do
this, a three-dimensional (3D) filter was created in four
steps. First, surface shape filtering was used to ‘remove’
the surface shape from inspection data resulting in a
native surface shape. Surface topography analysis by 3D
motif filter method was used in the next step to obtain
surface roughness and waviness motif elements. Surface
shape error could then be calculated as the deviation of
the neutral surface shape and the design surface shape
within each surface waviness element. Then the sculp-
tured surface shape was added back to obtain the regen-
erated surface roughness, surface waviness and surface
shape error. Finally, a surface roughness model was
obtained by using B-spline curve fitting of the regener-
ated roughness data within each roughness element, and
similarly the surface waviness model and surface shape
error model could be obtained by the B-spline fitting of
the regenerated waviness data and regenerated surface
shape error data within each waviness motif element.
The entire model was based on the 3D motif elements
and the necessary combination rules that were created.
Ehmann and Hong [8] introduced a new method to
represent the surface generation process, which they
835
called ‘surface-shaping system’. Their system basically
consisted of two parts, one that modeled the machine
tool kinematics and another that modeled the cutting tool
geometry. For the latter, specific interest was given in
the area of the cutting edge that was described as the
intersection of the tool’s face and flank surfaces along
with the respective angles. In general, the surface-shap-
ing system could account for spindle runout and machine
vibrations while additional research for the estimation of
cutting forces was still underway. In the work the system
was applied for the simulation of the 3D topography of
a peripherally milled surface.
Kim and Chu [9] stated that the surface roughness
could be determined by the maximum height of the
effective scallop including the effects of cutter marks
and conventional scallops. Through a texture superpo-
sition procedure, 3D surface texture, according to given
cutting conditions and cutter types, could be formed. The
runout effect (classified as geometric runout caused by
the eccentricity of the cutter axis and the irregularity of
the cutting edges and as dynamic runout caused by
vibration, chatter and the tool deflection) was included to
make the predicted surface closer to the actual machined
surface. The generalized cutter description was created
by introducing the fillet radius as a variable and covered
the cases of a ball end mill, a filleted end mill and a flat
end mill. The method besides providing the machined
surface texture can be used for optimal cutter path gener-
ation.
The work of Lee et al. [10] emphasizes on surface
roughness and profile in high-speed end milling. A
method for simulating the machined surface was
presented using the acceleration signal instead of the cut-
ting forces. The argument provided was that the
vibration, which is caused by the high speed of the
spindle deteriorates the geometric accuracy of the
machined surface. A geometric end milling model was
used for modeling the end mill offset and tilt angle. The
computer algorithm was developed in terms of cutting
conditions, cutter and workpiece geometry, and runout
parameters to determine the angular position of the end
mill. The coordinates of the flute end positions were cal-
culated using the geometric model of the end milling
process and the peak frequency components of the accel-
eration signal. From the generated surface data, surface
roughness could be calculated and the profile plotted.
3.2. Remarks
The conclusion that can be drawn is that these, theor-
etical for the most part, studies simulate the cutting pro-
cess in terms of kinematics and cutting tool properties.
Additional parameters such as vibrations are included in
an attempt to more accurately depict the phenomenon
and the obtained results can be characterized as fairly
good. The drawback of the approach is that a lot of other
836 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
factors that contribute to the roughness formation mech-
anism are not considered, for example wear and deflec-
tion of the cutting tool or certain thermal phenomena.
The integration of these factors to the already existing
models is estimated to increase their accuracy, especially
in cases of finishing procedures where their influence
is greater.
4. Experimental investigation approach
The experimental approach may be thought of as the
most ‘obvious’ method: experiments with the factors that
are considered to be the most important are conducted
and the obtained results are used to investigate the effect
of each factor as well as the influencing mechanism on
the observed quality characteristic. Regression analysis
is often employed in order to build models based on the
experimental data. The researcher’s intuition and insight
play a great role in this approach but a high understand-
ing of the examined phenomenon is also necessary for
the experiment to yield any meaningful results. The
experimental approach is mainly adopted in cases where
there can be no analytical formulation of the cause and
effect relationships between the various factors.
4.1. Experimental studies
The relationship between tool life, surface roughness
and vibration was examined in Ref. [11]. The variables
that were considered included the cutting speed, feed
rate, depth of cut, tool nose radius, tool overhang,
approach angle, workpiece length and workpiece diam-
eter and the accelerations in both radial and feed direc-
tions. The acceleration signals were fed to an FFT ana-
lyzer that produced ASCII files. The experimental data
were analyzed to produce regression analysis models.
Ghani and Choudhury [12] followed a similar
approach in which the vibration signals were used to
monitor tool wear and to verify the correlation between
tool wear progression and surface roughness during turn-
ing. The experiments were conducted on nodular cast
iron with ceramic tool, something that lead to very short
tool life (approximately 1.5 min).
Jang et al. [13] focused on the development of an on-
line roughness measuring technique by studying the
effects of cutting vibration during hard turning. It was
assumed that the average surface roughness is the result
of the superposition of a theoretical profile (process
kinematics) and an oscillatory profile (process
dynamics). An inductance type displacement sensor was
used to measure the relative movement between cutting
tool and workpiece. The results showed that surface
roughness along the workpiece had specific frequency
components that were determined by feed marks, in the
lower frequency range, and that were closely related to
the natural frequencies of spindle–workpiece system, in
the high frequency range.
Acoustic emission analysis was employed in Ref. [14]
to predict surface quality. Acoustic emission (AE) is
defined as the class of phenomena whereby transient
elastic waves are generated by the rapid release of
energy from localized sources within a material. In the
case of turning such sources can be found in the primary
(due to chip formation), secondary (due to friction
between cutting tool and chip) and tertiary (due to fric-
tion between cutting tool flank and workpiece) cutting
zones. Instead of using the RMS value of the AE meas-
ured signals, a new quantity called AERMS20 was intro-
duced in the paper and correlated with surface rough-
ness.
The potential for use of cryogenic cooling in turning
and particularly its effects on tool wear, dimensional
accuracy and surface finish was examined in Ref. [15].
For that purpose, a comparison between liquid nitrogen
jet on the cutting tool and dry machining was performed.
The results showed that cryogenic cooling improved the
surface finish by reducing the deterioration of the auxili-
ary cutting edge caused by abrasion, chipping and built-
up edge formation.
A series of experiments that aimed at determining the
cutting speed for built-up-edge formation in turning was
conducted in Ref. [16] and based on the experimental
data two mathematical models which correlate surface
roughness with cutting parameters were established. The
first one included the effects of the cutting tool’s nose
radius, feed rate, cutting speed and depth of cut while
the second added the effect of the material’s hardness
on surface finish. The most important terms were found
to be the feed rate, tool’s nose radius and cutting speed,
in that order.
A detailed investigation was carried out in Ref. [17]
concerning the effects of the cutting edge geometry and
workpiece hardness on the surface finish and cutting
forces in the finish hard turning of steel. The experi-
mental study included accurate characterization of the
edge geometry so as to measure edge features, exper-
imentation on a rigid machine tool with good spindle
accuracy to minimize vibration and analysis of variance
to establish the statistical significance of the parameters.
The findings revealed the significant effects of the edge
hone radius and the workpiece hardness to surface
roughness and cutting forces.
Baptista and Antune Simoes [18] concentrated on ana-
lyzing the effects of 3 and 5 axis milling of sculptured
surfaces on the surface finish, thus introducing para-
meters such as stepover and feed direction. In the case
of 5 axis milling a ball nose mill was compared to an
end mill, the results indicating that lower surface rough-
ness can be obtained by using the end mill inclined in
the feed direction.
A different approach was conducted in Ref. [19]
 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
where an ultrasonic system was developed for in-process
monitoring and control of surface roughness. An ultra-
sonic sensor connected to a PC, produced a pulse which
was reflected by the surface of the workpiece and meas-
ured the amplitude of the returned signal. The system
once calibrated with data from a stylus profilometer
could produce the actual roughness values. The main
advantage was that it was not affected by cutting fluids
and chips as is the case of other in-process schemes.
Additionally, a control scheme that aims at maintaining
constant surface roughness throughout tool degradation
was demonstrated.
The argument presented in Ref. [20] stated that the
length of cut, number of teeth, cutting forces and the
engagement of the cutting tool’s tooth with the work-
piece are all affected by the relative position of the cut-
ting tool and workpiece and consequently tool life, tool
wear and surface roughness are influenced by it.
The relation between vibrations and surface gener-
ation was investigated in Ref. [21] for slab milling oper-
ations. Since the periodical microgroove structure that is
generated by sequential engagement of the cutter teeth
is the primary parameter of surface roughness, it was
suggested that in a dynamic process, the different surface
profiles would depend on the amplitude, frequency and
phase of the relative motion between the workpiece and
the cutting tool. The separate contribution of each of the
cutter’s teeth to the resulting surface profile was verified
by conducting experiments with known vibrational
characteristics.
4.2. Remarks
This is the most conventional approach adopted. The
experiment–observation–conclusion strategy is the cor-
nerstone of every scientific research activity. Its advan-
tage lies in the fact that it is not difficult to implement
and that, depending on the level of understanding of the
participating phenomena, it can produce very good
results. On the other hand, the obtained conclusions have
little or no general applicability. It must be pointed out
that it is very easy for an experiment not to produce the
expected results because there are too many factors to
be considered regarding not only the examined phenom-
enon but also the equipment that is used.
5. Designed experiments approach
The reason why designed experiments were classified
under a different category from the previous approaches
is because they constitute a systematic method concern-
ing the planning of experiments, collection and analysis
of data with near-optimum use of available resources.
The response surface methodology (RSM) and Tagu-
chi techniques for design of experiments (DoE) seem to
837
be the most wide-spread methodologies for the surface
roughness prediction problem and are therefore
described in principle in the following paragraphs.
5.1. Response surface methodology overview
In response surface methodology, the factors that are
considered as most important are used to build a poly-
nomial model in which the independent variable is the
experiment’s response. In order to find the global mini-
mum of the response, experiments that ‘prune’ the
response surface are designed and conducted and the
gradient of the response surface is used along with a
steepest ascent algorithm as follows [22].
5.1.1. Stage 1
1. Select the factors to be investigated.
2. Design and run a two-level factorial experiment in a
localized region of the response surface.
3. Compute the estimates of the effects and thereby cal-
culate the coefficients of the linear model:
Y ⫽ b0 ⫹ b1X1 ⫹ b2X2 ⫹ % ⫹ bnXn.
4. Select a reference factor to be used as a guide in
determining the appropriate steps along the direction
of each factor in order to continue moving along the
path of steepest ascent.
5. Select a few experimental conditions along the path
of steepest ascent and run trials to determine if the
response continues to increase. If the response ceases
to increase, a new path should be generated.
6. If a new path is needed, design and run a new two-
level factorial experiment. All previous steps are
repeated until no substantial improvement in the
response is obtained. At this point stage 2 is conduc-
ted.
5.1.2. Stage 2
앫 Design and run a three-level factorial experiment in
the region where the path of the steepest ascent yields
no substantial improvement in the response.
앫 Compute the coefficients of the model:
Y ⫽ b0 ⫹ b1X1 ⫹ b2X2 ⫹ % ⫹ b11X21 ⫹ b22X22 ⫹ %
⫹ b12X1X2 ⫹ % ⫹ bn⫺1,nXn⫺1Xn.
앫 Using the above model, determine the nature of the
stationery point of the response surface. The station-
ery point is one where the gradient vanishes.
The sequential nature of RSM allows the experimenter
to learn about the process or system under study as the
investigation proceeds. This ensures that over the course
of the RSM application the experimenter will learn: (i)
how much replication is necessary; (ii) the location of
the region of the optimum; (iii) the type of approximat-
838 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
ing function required; (iv) the proper choice of experi-
mental designs; and (v) whether or not transformations
on the responses or any of the process variables are
required [23].
5.2. Taguchi techniques for DoE overview
DoE dictates a series of steps to follow for the experi-
ment to yield an improved understanding of product or
process performance [24].
Planning phase:
1. State the problem.
2. State the objectives of the experiment.
3. Select the quality characteristics and the measure-
ment systems.
4. Select the factors that may influence the quality
characteristics.
5. Select levels for the factors.
6. Select the appropriate Taguchi fractional matrices or
orthogonal arrays (OAs).
7. Select the interactions that may influence the qual-
ity characteristic.
8. Assign factors to OAs and locating interactions.
Execution phase:
9 Conduct the experiment repetitions as described by
the OAs.
Analysis phase:
10 Analyze the experimental results, e.g. using analysis
of variance (ANOVA).
11 Conduct a confirmation experiment.
There are three types of OAs, dealing with two-level
factors, three-level factors and mixed-level factors. The
selection of the appropriate OA is based on the following
criteria: the number of factors and interactions of inter-
est, the number of levels for the factors of interest and
the desired experimental resolution or cost limitations.
The first two determine the smallest OA that it is poss-
ible to use, while the third gives the possibility to con-
duct a larger experiment with higher resolution. Resol-
ution can vary from 1 (lowest) to 4 (highest) and it
indicates the clarity with which each individual effect of
factors and interactions may be evaluated in an experi-
ment.
In order to assign the various factors to an OA’s col-
umns, the following mathematical property should be
taken into account. If one factor is assigned to any parti-
cular column and a second factor to any other particular
column, a specific third column will automatically have
the interaction of those factors assigned to it. The pattern
of which columns will be interaction columns is known
for all of the OAs and it is visualized through the interac-
tion tables and linear graphs.
5.3. Designed experiments studies
The objective of Davim [25] was to establish a corre-
lation between cutting velocity, feed and depth of cut
with the surface roughness in turning. For that purpose,
a plan of experiments, based on Taguchi techniques, was
designed and executed. The results showed that the cut-
ting velocity had the greater influence, followed by the
feed and that the error achieved was smaller than that
of a geometric theoretical model.
An effort to predict surface roughness in turning of
high-strength steel based on RSM was made in Ref. [26].
The adequacy of the developed model results was not
very good but the conclusion was that the effect of feed
is much more pronounced than the effects of cutting
speed and depth of cut on the surface roughness.
Feng and Wang [27] included a total of six para-
meters, namely the workpiece hardness, feed, tool point
angle, depth of cut, spindle speed and cutting time to
build a model for finish turning operations. Hypothesis
testing established the adequacy of the model, while its
performance was deemed satisfactory. Feed was also
identified as the most important factor along with cut-
ting time.
RSM was also used in Ref. [28] for assessing the
influence of the past technological processes of the
workpiece material on the surface roughness of the
machined surfaces. The findings indicate which machin-
ing parameters have to be used for different combi-
nations of cutting tool–workpiece material for the
achievement of the desired surface roughness. Further
work, regarding the influence of workpiece material
properties, cutting parameters and TiN (PVD) hard coat-
ing on the surface roughness of fine-turned workpieces
by using Taguchi’s signal-to-noise response tables and
interaction graphs was also conducted by the same
researchers [29].
An extensive work was reported in Ref. [30] concern-
ing the role of cutting tool vibrations on surface rough-
ness during dry turning operations. Six parameters were
taken into account, including the workpiece and cutting
tool length and a full factorial design was adopted. The
analysis of variance and interaction analysis of the
experimental data revealed that the feed rate and tool
nose radius are the most influential parameters and that
the best surface roughness was achieved with low feed
rate, large tool nose radius and high cutting speed. Fur-
thermore, the amplitude of vibration measured at the
tool’s natural frequency and the variation of this natural
frequency indicated that the chip variation acted as a
dynamic force that excited tool vibration.
Two other implementations of the RSM can be found
in Refs. [31,32] where a surface roughness model is
 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
developed for end milling of 190 BHN steel and inconel
718. First- and second-order models were constructed
along with contour plots that more easily enable the
selection of the proper combination of cutting speed and
feed to increase the metal removal rate without sacrific-
ing surface quality.
The purpose of the work carried out in Ref. [33] was
to study the influence of tool geometry and cutting con-
ditions on the machined surface quality and to build a
model that would be able to predict it. Investigation of
the above factors in relation to the residual stresses was
also carried out. The innovation of the work was that
RSM and Taguchi’s method were combined to develop
the model.
RSM was also successfully applied for surface rough-
ness modeling of difficult to machine materials as the
EN32 (a semi-free cutting carbon casehardening steel)
[34]. A first-order model covering the speed range 30–
35 m/min and a second-order model covering the speed
range 24–38 m/min were built in terms of cutting speed,
feed rate and axial depth of cut under dry cutting con-
ditions.
5.4. Remarks
Although the common goal of the techniques investi-
gated is to organize the experimental procedure and the
necessary data processing, each follows a different path.
The RSM is mainly a model formulation procedure to
investigate how important factors affect the response of
an experiment and leads to the development of first- and
second-order polynomial models that include the para-
meters under consideration and their statistical signifi-
cance. These models are used to create contour plots that
can be more practically utilized to draw conclusions
compared to using a polynomial function. On the other
hand, the Taguchi DoE is more of a factor-screening pro-
cedure to determine the significance of each factor, that
is, it identifies the most influential parameters and the
values that produce the desired output without for-
mulating any kind of model. Nevertheless, it must be
pointed out that because of their generality and strong
statistical background, certain tasks of these method-
ologies can be isolated and applied to a wide range of
engineering problems where the size of the search space
must be reduced.
6. Artificial intelligence approach
AI is implemented in engineering problems through
the development of artificial neural network (ANN)
models, genetic algorithms (GAs), fuzzy logic and
expert systems. Simulating the way in which human
beings process information and make decisions, in the
839
surface roughness prediction problem published work, is
based mainly on ANNs and GAs.
6.1. Artificial neural networks overview
An ANN is an information processing system that dis-
plays similar behavior to that of its biological analog. It
is essentially a mathematical model that mimics the
human reasoning and neurobiology and that is based on
the following assumptions [35]:
앫 Information processing occurs in a number of simple
elements called neurons.
앫 Signals are transmitted between neurons over connec-
tion links.
앫 Each connection link has an associated weight that
multiplies the signal transmitted.
앫 Each neuron applies an activation function to the
incoming signal to determine its output signal (Fig.
2).
ANNs are mostly used for pattern recognition, pattern
association and classification, constrained optimization
and systems modeling with applications ranging from
simple signal processing to medical diagnosis.
The two main characteristics of an ANN are: (i) the
pattern of arrangement of the neurons, namely the archi-
tecture of the network, which generally dictates what
type of problems can be dealt with; (ii) the method of
determining the weights of the connections, either using
a training algorithm generalizing the relation of input to
output vectors-examples (supervised training) or infer-
ring classifications that are inherent to the data and gen-
erating an exemplar vector for each class that is created
(unsupervised training).
The most common type of ANNs that have been used
in the literature examined is the feedforward ANNs
trained with some variation of the backpropagation
algorithm. The typical architecture of these ANNs can
Fig. 2. Neuron model.
840 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
be seen in Fig. 3. The neurons are arranged in layers,
i.e. the input layer, the hidden layer and the output layer,
from left to right. The input layer is used to present the
data in the ANN and the output to produce the ANN’s
response.
6.2. Neuro-fuzzy systems overview
Fuzzy logic is based on the concept of fuzzy sets. A
fuzzy set is a set without a crisp, clearly defined bound-
ary that is, it can contain elements with a partial degree
of membership (usually between 0 and 1). A more strict
definition is that a fuzzy set m of X is a function from
the set X to the unit interval, i.e. m:X→[0,1]. The func-
tion that defines how each element is mapped to the
degree of membership is called the membership func-
tion (MF).
If–then rules are used to formulate the ‘logic’ by cre-
ating statements that contain knowledge. The typical
form of such a rule is: If x is A then y is B, where x is
called the antecedent, y the consequent and A and B are
linguistic values defined by fuzzy sets. The outputs of
all rules used are aggregated in a single fuzzy set, which
is later defuzzified so as to obtain a single number.
In a neuro-fuzzy system ANNs are used to determine
parameters of fuzzy systems thereby creating or improv-
ing a fuzzy system automatically. The implementation
of a neuro-fuzzy system can be made so as the two sys-
tems work independently of each other (cooperative
neuro-fuzzy system) or within a homogeneous architec-
ture (hybrid neuro-fuzzy system). These systems must
not be confused with fuzzy neural networks, which are
essentially ANNs that use fuzzy methods to learn faster
or perform better. For a more complete description of
neuro-fuzzy systems please refer to Ref. [36].
Fig. 3. Supervised learning procedure of a typical feedforward net-
work.
6.3. Genetic algorithms overview
GAs are search algorithms for optimization, based on
the Darwinian theory of evolution. The power of these
algorithms is derived from a very simple heuristic
assumption that the best solution will be found in the
regions of solution space containing high proportion of
good solution, and that these regions can be identified
by judicious and robust sampling of the solution space.
The mechanics of a GA are simple and involve the
coding of solution states in chromosomes as a series of
binary elements (0 and 1). A set (i.e. population) of can-
didate solution states (i.e. chromosomes) is generated
and evaluated. A fitness function is used to evaluate each
of the solutions in the population. The chromosomes
encoding the better solutions are broken apart and
recombined through the use of genetic operators in suc-
cession to get a new solution (i.e. offspring) that is gen-
erally better or more fit, in one generation or iteration.
These operators are essentially mathematical models of
genetic operations that take place in the human body.
The simplest form of GA involves three types of oper-
ators: selection (copying of the strings into a ‘mating
pool’, in proportion to their fitness values), crossover
(swapping parent strings partially, causing offspring to
be generated) and mutation (occasional random alter-
ation with a small probability of the value of a string
position, in binary strings, this simply means changing
0 to 1 or vice versa, see Fig. 4.
The genetic operators produce various and different
effects according to their individual or combined use,
for instance too-weak selection will result in too-slow
evolution. Also, without mutation every string in the
population might come to have a one at the first bit pos-
ition and there would then be no way to obtain a string
beginning with a zero [37].
A GA is typically iterated for 50–500 or more gener-
ations. For the exact description of how a GA works
refer to Ref. [37].
Fig. 4. Mutation and crossover genetic operators.
 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
6.4. Artificial intelligence studies
A sensor fusion technique was presented in Ref. [38]
to evaluate the surface roughness and dimensional devi-
ation during turning. The systematic method for the
selection of the candidate sensors determined the aver-
age effect of each in the performance of the measuring
system. The sensors that affected it the most were fused
using ANN modeling. The results showed that the mod-
els created with the above technique were more accurate
than regression analysis models that were developed for
comparison purposes.
Sensor fusion incorporating ANNs is also described
in Ref. [39]. Capacitive, inductive and fiber optic sensors
were used so as to cover an as wide as possible range
of application by detecting features that cannot be sensed
by a single sensor type. The RMS value of the three
sensors along with the type of manufacturing process
(face turning, milling, electro-discharge machining and
grinding), all of which were coded in binary format,
were used to train a 17 × 20 × 15 ANN and the results
obtained could be characterized as fairly good.
The work of Chien and Chou [40] can be divided in
two parts. The first part deals with the building of three
predictive models, using multilayer functional-link net-
works, for surface roughness, cutting force and tool life,
respectively. The second part focused on finding the
optimum cutting conditions by combining surface rough-
ness and tool life network with a genetic algorithm. In
that way, the cutting conditions that maximized the
metal removal rate were obtained, under the constraints
of surface roughness and cutting tool life.
Suresh et al. [41] also adopted a two stage approach
towards optimizing for surface roughness. First, experi-
mental results were used to build two mathematical mod-
els for surface roughness by a regression method accord-
ing to RSM. Secondly, the second-order mathematical
model was taken as an objective function and was optim-
ized with a GA to obtain the machining conditions for
a desired surface finish.
Polynomial networks were considered in Ref. [42] to
construct the relationships between the cutting para-
meters (cutting speed, feed rate, depth of cut) and cutting
performance (tool life, surface roughness and cutting
force). In that way, a machining database could be con-
structed for turning operations. It was stated that a com-
parison between polynomial networks and backpropag-
ation networks has shown that the former have higher
prediction accuracy and fewer internal connections.
Additionally, the best network structure was determined
by using an algorithm for synthesis of polynomial net-
works (ASPN). The principle of the ASPN criterion is
to select a network as accurate as required but as less
complex as possible, too.
Li et al. [43] developed a hybrid machining model
that integrated analytical models and neural network
841
models for predicting all of the machining characteristic
factors. The analytical component was based on the pre-
dictive machining theory proposed by Oxley [44] and
served as a predictor for the cutting forces, temperature
in the cutting region and chip geometry. It also served
as a pre-processor of the neural network model that pre-
dicted the tool wear, surface roughness and chip break-
ability, which could not be dealt with in a completely
analytical manner.
Another approach that used a criterion for determining
a network’s architecture automatically can be found in
Ref. [45]. The aim was to develop a prediction model
prior to the implementation of the actual machining pro-
cess to determine certain cutting conditions (cutting
speed, feed rate and depth of cut) in order to obtain a
desired surface roughness value and cutting force value.
Furthermore, using the obtained cutting force, the cutting
power and optimal metal removal rate could be calcu-
lated next. The abductive networks that were created
using the predicted square error (PSE) criterion perfor-
med more accurately than the respective regression
analysis models.
In Ref. [46] an approach that could evaluate the influ-
ence of machine tool characteristics on cutting processes
using adaptive prediction was presented. The network
for predicting surface roughness had as inputs the cutting
speed, the affinity between cutting tool and workpiece,
the chip discontinuity (evaluated by the chip strain), the
built-up-edge formation (evaluated by average tempera-
ture around the cutting edge), the width of flank wear
and the theoretical roughness considering tool wear.
ANN modeling was also used along with designed
experiments by Benardos and Vosniakos [47] in face
milling. The final model was created in terms of feed
rate per tooth, depth of cut, engagement of the cutting
tool, use of cutting fluid and the component of the cut-
ting force along the direction of the feed. The results of
this study showed that ANNs can be extremely accurate
even when they are used for complex problems such as
finishing of Al alloys. Furthermore, the use of designed
experiments had as a result the systematic screening of
the ANN topologies with a view to optimizing the final
topology used.
In the research of Tsai et al. [48] an accelerometer
was employed as an in-process sensor. After collecting
vibration data from the machine tool–workpiece system
during an end milling process in a CNC vertical machin-
ing center, a backpropagation ANN was trained and
tested not only to achieve the goal of in-process surface
roughness recognition, but also to increase production
rate and product quality. In addition, a statistical multiple
regression model was also developed and compared with
the proposed ANN model. The parameters included were
the spindle speed, feed rate, depth of cut and the
vibration average per revolution (VARP). It was found
that the ANN models behaved better under all situations.
842 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
A new technique, an in-process surface recognition
system that used the fuzzy-nets and a sensor-testing sys-
tem to measure surface roughness in end milling oper-
ations was presented in Ref. [49]. The training procedure
consisted of five steps: (a) division of the input space
into fuzzy regions; (b) generation of fuzzy rules from
given pairs through experiments; (c) avoidance of con-
flicting rules; (d) development of a combined fuzzy rules
base; and (e) determination of a mapping system based
on the fuzzy rules base. Then the fuzzy rules and mem-
bership functions would be generated and adjusted to
meet the requirements of the control parameters of the
milling system. The technique was initially used for one
type of material and cutting tool but it was later
expanded by Chen and Savage [50] by adding a new set
of parameters including tool diameter and workpiece
material.
An adaptive neuro-fuzzy inference system (ANFIS)
and computer vision were used to predict surface rough-
ness in turning in Ref. [51]. The computer vision system,
comprising a digital camera connected to a PC and the
appropriate light sources, provided surface images that
were analyzed to calculate the arithmetic average of gray
levels (number of shades of gray). This information as
well as the cutting parameters were given, for a total of
four inputs, to the ANFIS and the roughness value could
then be obtained.
6.5. Remarks
In spite of the fact that these methodologies were
developed decades ago, recent advances in the field, the
ANN training algorithms for example, as well as the
rapid increase of available computing power have
brought revived interest among researchers. It is obvious
that the approaches that have been described in this sec-
tion can produce very good results and simultaneously
offer the possibility for on-line monitoring and/or control
of the process.
The main advantages that ANNs have in contrast to
classic programming is that they can manage noisy or
incomplete data with ease, there is no need to explicitly
formulate the problem, the solution algorithm or to write
code and that the process of information is distributed
over the neurons which operate in parallel, therefore
resulting in increased computational power in contrast
to the sequential operation of today’s computers. The
most obvious drawback of the ANNs is that there is no
guarantee for their resulting performance in an appli-
cation.
The simplicity of operation and efficiency with multi-
criterion optimization problems are the two main attrac-
tions of the GA approach. Since there are ready-for-use
GAs available, it is not necessary to write code for a
selected application from scratch. The only prerequisite
is the coding of the data in order to form the chromo-
somes and the formulation of the appropriate fitness
function. However, the main disadvantage of a GA is
that it is very demanding on computational power since
a run can take days to be completed on a typical PC plat-
form.
7. Conclusions and discussion
The current work presented a review of the different
approaches that are used for predicting the surface
roughness and certain remarks concerning each approach
can be found in the respective sections. As is evident
from the referenced papers, in recent years there has
been a great deal of research activity in the field and the
results that have been produced are good. The trend that
is formed encourages more automated systems building
for on-line monitoring, measuring or control and is
mainly driven by the fact that the processes themselves
have been automated to a great extent. All the method-
ologies that are presented here can exhibit advantages
and disadvantages when compared to one another, but
given this trend the most promising seem to be the theor-
etical and the AI approaches.
A comparison of these two approaches reveals that AI
models take into consideration the particularities of the
equipment used and the real machining phenomena,
information that is stored in the experimental data used
to develop the models. On the other hand, the theoretical
approach is based on conventions and idealizations,
which are responsible for errors and limitations. Surpris-
ingly enough, a combined effort that would involve both
AI and analytical modeling so as to validate, refine or
correct the theoretical models was not found in the litera-
ture.
Other advantages of the AI approach are that the mod-
els created seem to be the most realistic and accurate,
they probably exhibit the highest level of integration
with computers and that this approach can be used in
conjunction with other more conventional techniques.
With these facts taken into consideration, it can be con-
cluded that there are not so many efforts as would have
been expected. The same applies to the existing number
of hybrid AI research approaches, such as the neuro-
fuzzy systems. The advantages that they offer
(knowledge representation in the form of if-then rules,
ANN assisted parameter determination) should be more
than enough to encourage researchers to adopt these
techniques, yet this has not been the case.
Optimization of cutting conditions for a certain sur-
face roughness is another field that has not received too
much attention. GAs and other optimization algorithms
could be ideally used in conjunction with the developed
models for the prediction of surface roughness but as is
evident from the above, very few similar approaches
have been found.
 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
Fig. 5. Fishbone diagram with the parameters that affect surface roughness.
It must also be noted that despite the fact that accurate
models have been developed there are still issues to be
dealt with. Certain cases like high accuracy machining,
where surface roughness is of great importance, are still
under investigation and factors such as the cutting tool’s
deflection or the thermal conditions must be introduced
to future models for a more realistic depiction of surface
roughness creation. The integration of the existing mod-
els to a more general advisory system, which could be
used by a machine tool operator for example, could be
another very useful and practical application.
Finally, the set of parameters that are thought to
influence surface roughness and thus have been investi-
gated by the researchers, is diagrammatically displayed
in Fig. 5.
References
[1] DIN4760, Form Deviations; Concepts; Classification System,
Deutches Institut Fuer Normung, e.V., 1982.
[2] G. Boothroyd, W.A. Knight, Fundamentals of Machining and
Machine Tools, Marcel Dekker, New York, 1988.
[3] J. Kaczmarek, Principles of Machining by Cutting, Abrasion and
Erosion, Peter Peregrinus, London, 1983.
[4] W. Grzesik, A revised model for predicting surface roughness in
turning, Wear 194 (1996) 143–148.
[5] S.C. Lin, M.F. Chang, A study on the effects of vibrations on
the surface finish using a surface topography simulation model
for turning, International Journal of Machine Tools and Manufac-
ture 38 (1998) 763–782.
[6] D.K. Baek, T.J. Ko, H.S. Kim, Optimization of feedrate in a face
843
milling operation using a surface roughness model, International
Journal of Machine Tools and Manufacture 41 (2001) 451–462.
[7] C.-C.A. Chen, W.-C. Liu, N.A. Duffie, A surface topography
model for automated surface finishing, International Journal of
Machine Tools and Manufacture 38 (1998) 543–550.
[8] K.F. Ehmann, M.S. Hong, A generalized model of the surface
generation process in metal cutting, CIRP Annals 43 (1994)
483–486.
[9] B.H. Kim, C.N. Chu, Texture prediction of milled surfaces using
texture superposition method, Computer Aided Design 31 (1999)
485–494.
[10] K.Y. Lee, M.C. Kang, Y.H. Jeong, D.W. Lee, J.S. Kim, Simul-
ation of the surface roughness and profile in high speed end mill-
ing, Journal of Materials Processing Technology 113 (2001)
410–415.
[11] O.B. Abouelatta, J. Madl, Surface roughness prediction based on
cutting parameters and tool vibrations in turning operations, Jour-
nal of Materials Processing Technology 118 (2001) 269–277.
[12] A.K. Ghani, I.A. Choudhury, Study of tool life, surface roughness
and vibration in machining nodular cast iron with ceramic tool,
Journal of Materials Processing Technology 127 (2002) 17–22.
[13] D.Y. Jang, Y.G. Choi, H.G. Kim, A. Hsiao, Study of the corre-
lation between surface roughness and cutting vibrations to
develop an on-line roughness measuring technique in hard turn-
ing, International Journal of Machine Tools and Manufacture 36
(1996) 453–464.
[14] C. Beggan, M. Woulfe, P. Young, G. Byrne, Using acoustic emis-
sion to predict surface quality, International Journal of Advanced
Manufacturing Technology 15 (1999) 737–742.
[15] N.R. Dhar, S. Paul, A.B. Chattopadhyay, The influence of cryo-
genic cooling on tool wear, dimensional accuracy and surface
finish in turning AISI 1040 and E4340C steels, Wear 249 (2001)
932–942.
[16] P. Munoz-Escalona, Z. Cassier, Influence of the critical cutting
speed on the surface finish of turned steel, Wear 218 (1998)
103–109.
844 P.G. Benardos, G.-C. Vosniakos / International Journal of Machine Tools & Manufacture 43 (2003) 833–844
[17] J.D. Thiele, S.N. Melkote, Effect of cutting edge geometry and
workpiece hardness on surface generation in the finish hard turn-
ing of AISI 52100 steel, Journal of Materials Processing Tech-
nology 94 (1999) 216–226.
[18] R. Baptista, J.F. Antune Simoes, Three and five axis milling of
sculptured surfaces, Journal of Materials Processing Technology
103 (2000) 398–403.
[19] S.A. Coker, Y.C. Shin, In-process control of surface roughness
due to tool wear using a new ultrasonic system, International
Journal of Machine Tools and Manufacture 36 (1996) 411–422.
[20] A.E. Diniz, J.C. Filho, Influence of the relative positions of tool
and workpiece on tool life, tool wear and surface finish in the
face milling process, Wear 232 (1999) 67–75.
[21] U. Heisel, Vibrations and surface generation in slab milling,
CIRP Annals 43 (1994) 337–340.
[22] A. Garcia-Diaz, D.T. Phillips, Principles of Experimental Design
and Analysis, Chapman and Hall, London, 1995.
[23] R.H. Myers, D.C. Montgomery, Response Surface Methodology,
Process and Product Optimization Using Designed Experiments,
Wiley, New York, 1995.
[24] P.J. Ross, Taguchi Techniques for Quality Engineering: Loss
Function, Orthogonal Experiments, Parameter and Tolerance
Design, McGraw-Hill, New York, 1996.
[25] J.P. Davim, A note on the determination of optimal cutting con-
ditions for surface finish obtained in turning using design of
experiments, Journal of Materials Processing Technology 116
(2001) 305–308.
[26] I.A. Choudhury, M.A. El-Baradie, Surface roughness in the turn-
ing of high-strength steel by factorial design of experiments, Jour-
nal of Materials Processing Technology 67 (1997) 55–61.
[27] C.X.J. Feng, X. Wang, Development of empirical models for sur-
face roughness prediction in finish turning, International Journal
of Advanced Manufacturing Technology 20 (2002) 348–356.
[28] J. Kopac, M. Bahor, Interaction of the technological history of a
workpiece material and the machining parameters on the desired
quality of the surface roughness of a product, Journal of Materials
Processing Technology 92–93 (1999) 381–387.
[29] J. Kopac, M. Bahor, M. Sokovic, Optimal machining parameters
for achieving the desired surface roughness in fine turning of cold
pre-formed steel workpieces, International Journal of Machine
Tools and Manufacture 42 (2002) 707–716.
[30] M. Thomas, Y. Beauchamp, A.Y. Youssef, J. Masounave, Effect
of tool vibrations on surface roughness during lathe dry turning
process, Computers in Industrial Engineering 31 (1996) 637–644.
[31] M. Alauddin, M.A. El-Baradie, M.S.J. Hashmi, Computer-aided
analysis of a surface roughness model for end milling, Journal
of Materials Processing Technology 55 (1995) 123–127.
[32] M. Alauddin, M.A. El-Baradie, M.S.J. Hashmi, Optimization of
surface finish in end milling inconel 718, Journal of Materials
Processing Technology 56 (1996) 54–65.
[33] K.H. Fuh, C.F. Wu, A proposed statistical model for surface qual-
ity prediction in end milling of Al alloy, International Journal of
Machine Tools and Manufacture 35 (1995) 1187–1200.
[34] A. Mansour, H. Abdalla, Surface roughness model for end mill-
ing: a semi-free cutting carbon casehardening steel (EN32) in dry
condition, Journal of Materials Processing Technology 124
(2002) 183–191.
[35] L. Fausett, Fundamentals of Neural Networks, Prentice-Hall
International Editions, Englewood Cliffs, NJ, 1994.
[36] D. Nauck, F. Klawonn, R. Kruse, Foundations of Neuro-Fuzzy
Systems, Wiley, New York, 1997.
[37] M. Mitchell, An Introduction to Genetic Algorithms, MIT Press,
Cambridge, MA, 1999.
[38] R. Azouzi, M. Guillot, On-line prediction of surface finish and
dimensional deviation in turning using neural network based
sensor fusion, International Journal of Machine Tools and Manu-
facture 37 (1997) 1201–1217.
[39] S. Varghese, V. Radhakrishnan, A multi sensor approach to in-
process monitoring of surface roughness, Journal of Materials
Processing Technology 44 (1994) 353–362.
[40] W.T. Chien, C.Y. Chou, The predictive model for machinability
of 304 stainless steel, Journal of Materials Processing Technology
118 (2001) 442–447.
[41] P.V.S. Suresh, P. Venkateswara Rao, S.G. Deshmukh, A genetic
algorithmic approach for optimization of surface roughness pre-
diction model, International Journal of Machine Tools and Manu-
facture 42 (2002) 675–680.
[42] B.Y. Lee, Y.S. Tarng, H.R. Lii, An investigation of modeling of
the machining database in turning operations, Journal of
Materials Processing Technology 105 (2000) 1–6.
[43] X.P. Li, K. Iynkaran, A.Y.C. Nee, A hybrid machining simulator
based on predictive theory and neural network modeling, Journal
of Materials Processing Technology 89–90 (1999) 224–230.
[44] P.L.B. Oxley, Mechanics of Machining: An Analytical Approach
to Assessing Machinablity, Ellis Horwood, Chichester, UK, 1989.
[45] W.S. Lin, B.Y. Lee, C.L. Wu, Modeling the surface roughness
and cutting force for turning, Journal of Materials Processing
Technology 108 (2001) 286–293.
[46] T. Matsumura, H. Sekiguchi, E. Usui, An evaluation approach of
machine tool characteristics with adaptive prediction, Journal of
Materials Processing Technology 62 (1996) 440–447.
[47] P.G. Benardos, G.C. Vosniakos, Prediction of surface roughness
in CNC face milling using neural networks and Taguchi’s design
of experiments, Robotics and Computer Integrated Manufacturing
18 (2002) 343–354.
[48] Y.H. Tsai, J.C. Chen, S.J. Lou, An in-process surface recognition
system based on neural networks in end milling cutting oper-
ations, International Journal of Machine Tools and Manufacture
39 (1999) 583–605.
[49] S.J. Lou, J.C. Chen, In-process surface recognition of a CNC
milling machine using the fuzzy nets method, Computers in
Industrial Engineering 33 (1997) 401–404.
[50] J.C. Chen, M. Savage, Fuzzy-net-based multilevel in-process sur-
face roughness recognition system in milling operations, Inter-
national Journal of Advanced Manufacturing Technology 17
(2001) 670–676.
[51] S.-Y. Ho, K.-C. Lee, S.-S. Chen, S.-J. Ho, Accurate modeling
and prediction of surface roughness by computer vision in turning
operations using an adaptive neuro-fuzzy inference system, Inter-
national Journal of machine Tools and Manufacture 42 (2002)
1441–1446.