International Journal of Mechanical and Materials Engineering (IJMME), Vol.6 (2011), No.

3, 391-399

Prediction of Burnishing Surface Integrity using Radial Basis Function

N.S.M. EL-Tayeb1 and S. Purushothaman2

1
Mechanical Engineering Department, Faculty of Engineering, University of Malaya, Kuala Lumpur, Malaysia-50603
2
Faculty of Engineering and Technology, Multimedia University, Jln Ayer Keroh Lama, Melaka, Malaysia-75450
Email: nabil.eltayeb@yahoo.com; nabil.eltayeb@um.edu.my

Received 12 November 2011, Accepted 24 December 2011

ABSTRACT
In this work, prediction of burnishing surface quality such as requires a high level of understanding to the
roughness (Ra) and Vickers hardness (HV) were achieved by phenomenon under investigation. This approach is
using supervised radial basis function (RBF). The process followed by many researchers (Ong et al. 2008,
state variables used were burnishing speed, feed, and depth. Moshkovich 2007, Man, 2006). It is commonly
RBF has achieved a minimum of 90.62 % of prediction and
proved to be convenient in terms of least computational
implemented where the analytical formulation of the
complexity and dealing with nonlinear data such as obtained cause and effect relationship between various factors is
in this work. complex and difficult to achieve. Nevertheless, the
strategy of experiment, observation, analysis, and
Keywords: Artificial Neural Network, Radial Basis conclusion is the most conventional approach which
Function, Burnishing process represents the corner stone of most scientific research
activities. However, the drawback of this approach is
1. INTRODUCTION that the conclusions are usually confined to the range
Burnishing is a plastic deformation process in which of the test conditions and cannot be generalized, i.e.
the force is applied to a workpiece surface by a hard conclusions have limited applications and this should
smooth ball or roller. The mechanism of burnishing be expected because of various types of test
operation is that asperities` peaks of the workpiece are equipments and techniques, different environment
compressed to fill in the valleys and thus conditions and other missing factors.
characteristics of surface under consideration are
The third approach is to use design of experiment (DoE)
changed (El-Tayeb et al.2007, Luca et al.2005, and El-
together with analysis of variance (ANOVA) to
Axir et al.2008). The characteristics of burnishing
develop an empirical model (regression model) based
surface depend upon controlling burnishing parameters
on some collected experimental data (inputs and
such as applied burnishing force (burnishing depth),
outputs). This approach is also implemented by several
burnishing speed, burnishing feed rate, and number of
researchers (Abu Shereehah 2008, El-Axir et al.2008
passes, geometry and material of burnishing tool, as
a&b, Shiou and Cheng 2008). This is a systematic
well as the material of burnishing surface. In the
approach which involves planning of experiments,
current work, the author’s intention is to implement a
collection and analysis of data with the aid of response
new approach to predict the burnishing process state
surface methodology (RSM) and occasionally used for
variables (Ra and HV). In this respect, there are
optimization (Hassan et al.1998). It is basically a
different approaches which can be adopted. The first
model formulation procedure to investigate how
approach is based on developing analytical model from
important factors affect the response of the experiment.
fundamental concepts (Hua et al.2008 and Shio and
It aims to a development of 1 st or 2nd order polynomial
Chen 2003) which requires a procedure to represent the
model that includes the parameters under consideration
response (output) as a function of the process
and their statistical significance. The fourth approach
parameters (inputs) (Korzynski 2007). In this approach,
uses unconventional methods, namely, artificial neural
although some parameters can be included and fairly
network (ANN), genetic algorithm (GA), particle
good prediction of the response is achieved, still a lot
swarm optimization (PSO) and Fuzzy logic (FL) (Esme
of other parameters cannot be included which usually
et al.2008 and El-Khabeery and El-Axir 2001). In this
contribute significantly to the state variables. In the
approach, an ANN simulates the way in which human
case of burnishing process, condition of burnishing tool,
brain processes information and makes decisions. This
deflection of workpiece, friction and heat generation
approach is also based on the collection of
between burnishing tool and workpiece are expected to
experimental data (inputs and outputs) for training and
contribute to the accuracy of the analysis. The second
testing. To the best of the author’s knowledge, rare
approach depends on experiments by studying the
applications of ANN to the burnishing process are
effect of one factor at a time and analysis the results
found in the literature. Therefore, the current work
(El-Tayeb et al.2006, 2009 and 2008). Experiments are
intends to implement the ANN approach for prediction
usually carried out to investigate the effect of most
of surface quality of ball burnishing brass surfaces.
important factors and the analysis of the results
391
The influence of ball burnishing parameters on surface
finish and surface hardness of brass was reported
earlier in several articles (El-Khabeery 2001, Hassan
2000, El-Axir 2003, Luo et al.2006, and El-Axir 2005)
using the second and third approaches. In the present
work, the authors considered three independent
burnishing parameters, namely burnishing speed,
burnishing depth (burnishing force), and burnishing
feed rate, and other parameters were kept constant by
using the fourth approach (ANN). Specifically, this
paper implements supervised radial basis function on
the collected experimental data to predict Ra and HV.
2. SCHEME OF INVESTIGATION
This work is planned in two main stages. In the first
stage, experimental preparation, measurements and
collection of results are conducted. In the second stage,
neural network with Radial Basis Function (RBF) is
implemented. Figure 1 shows the procedure of data
collection during experimentation, the training of ANN
algorithms and testing of ANN respectively.
3. EXPERIMENTAL SETUP
3.1 Work piece material
In the present work, the workpiece material was
selected from commercially available brass (C3605:
56-60% Pb, 3.5-4.5%Fe, 3.5-4.5% Fe, 0.50% Sn,
1.2%Zn, and Bal Al). The strength and relatively good
mechanical properties of the brass C3605 make this
material suitable in many industrial applications such
as bushes and valves etc. Some mechanical properties
of the C3605 brass material are given in Table 1. The
C3605 brass rod were turned and burnished by using
CNC turning machine. The CNC machine has a wide
range of parameter settings.

CNC Turning Turning Selection of

Machine Operation Burnishing Parameters

Measurements, Burnishing
Ra and HV Operation

(a) Schematic flow of experimental work

Read input and Train using Store final

target outputs RBF weights

(b) Training RBF

Read input Process with Compare outputs

pattern final weight with threshold

(c) Testing RBF

Fig.1 Stages of experimental work and data analysis

3.2 Burnishing operation After

The burnishing setup is shown schematically in Fig.2a
Burnishing
Burnishing Speed

and a typical arrangement is shown in Fig.2b. Turned part

Tailstock
Table 1 Mechanical properties of the brass rods C3605.

Properties Values
Workpiece
Tensile Strength, t (MPa) 338 - 469 Ball Diameter
12, 14, 16 mm
Yield strength, y (MPa) 124 - 310

53
Burnishing
Elongation at Break (%)
Force/Depth
Modulus of Elasticity, E (GPa) 97
(a) Schematic drawing showing all burnishing features

392
 This burnishing tool is based on using four standard
Kerosene
and commercially available carbon chromium steel
balls (used in ball bearings) having a Rockwell
hardness of HRC 63 and arithmetic surface roughness
average (Ra) of 0.012. A ball adapter can
accommodate the main burnishing ball (8mm dia.)
supported by three other small balls (6mm dia.) to
prevent sticking the burnishing ball during burnishing
process. Furthermore, the burnishing tool is designed in
Unburnished Burnished a way that the main ball is free to rotate by the friction
effect due to rotation of the workpiece. As shown in
Fig.3b, the ball adapter contained a lubrication channel
(b) Typical setup of burnishing operation
to lubricate the burnishing ball continuously during
Fig.2 Burnishing operation burnishing operation. When burnishing force is applied
to the workpiece, an axial sliding distance of the guide
rod is detected by a dial gage. This axial distance is
Initially C3605 material was received in the form of converted into burnishing force by using a linear
cylindrical rods (125mm length x 37mm diameter). equation determined by a calibration process
The cutting and burnishing tools were clamped
simultaneously onto the turret of the CNC turning
7
machine and thus turning and burnishing operations
were completed one after another without separations
process. The preparation of burnishing specimens
starts by turning the rods into circular bars of 24mm
diameter. This turning process is followed by setting
the desired burnishing conditions (burnishing
parameters) and applying burnishing process. The
burnishing parameters considered are given in Table 2.
During burnishing operation, Kerosene was (1) Shank, (2) Helical compression spring, (3) Ball adapter, (4)
continuously supplied into the burnishing zone, Fig. 2b, Carbon chromium steel ball, 12mm,(5) A dial gauge, (6) A spring
as a coolant fluid to avoid heat generation or adhesion guide, and (7) Supporting balls.
(a) Assembled sectional view of ball burnishing tool.
problems. After finishing burnishing process, the
burnished cylindrical bar was cut into small discs of
24mm diameter and 10mm thickness for subsequent
tests and assessments.

3.3 Burnishing Tool

A ball burnishing tool was designed and fabricated to
carry out the burnishing operation during the current
work, Fig.3 (El-Tayeb 2007). Figure 3a shows a
schematic drawing of the burnishing tool.

Table 2 Burnishing parameters. (b) Typical ball burnishing tool

Burnishing Parameters Values Fig.3. Ball burnishing tool (El-Tayeb 2007)

Burnishing Speed (rpm)
Burnishing depth (mm)
Burnishing Force (N)
Ball Diameter (mm)
Burnishing Feed Rate
(mm/rev)
Lubricant
150, 250, 350, 450 3.4 Surface roughness and hardness measurements
0.05, 0.1, 0.15, 0.2 Surface roughness and micro hardness measurements
were carried out before and after burnishing operation.
109, 187, 221, 246
Surface roughness, Ra, was measured by using Mahr
8mm Perthometer S2 Profilometer. Meanwhile micro
0.03, 0.06, 0.09, hardness (Vickers’s hardness, HV) measurements were
0.12 done using Vickers` hardness tester (Model 450VA).
The average of more than three measurements is
Kerosene reported in the current work. The experimental data
collected is given in Table 3

393
 Table 3 Experimental data of burnishing

Measured Measured 38 250 0.10 0.09 0.395 163.7

B. Speed B. Depth B. Feed
Pattern Roughness Hardness
(rpm) (mm) (mm/rev) 39 250 0.15 0.09 0.609 165.8
(Ra), m (HV)
1.25* 140.0* 40 250 0.20 0.09 1.310 177.7
1 150 0.05 0.03 1.009 145.9 41 350 0.05 0.09 0.933 145.2
2 150 0.10 0.03 0.333 150.9 42 350 0.10 0.09 0.562 153.8
3 150 0.15 0.03 0.523 154.3 43 350 0.15 0.09 0.640 164.9
4 150 0.20 0.03 0.629 165.3 44 350 0.20 0.09 0.715 182.3
5 250 0.05 0.03 0.996 184.4 45 450 0.05 0.09 1.014 154.2
6 250 0.10 0.03 0.946 191.7 46 450 0.10 0.09 1.120 165.4
7 250 0.15 0.03 1.188 193.7 47 450 0.15 0.09 1.520 175.1
8 250 0.20 0.03 1.395 207.9 48 450 0.20 0.09 1.450 184.9
9 350 0.05 0.03 0.727 181.5 1.200* 140.0*
10 350 0.10 0.03 0.996 175.8 49 150 0.05 0.12 0.933 191.9
11 350 0.15 0.03 1.318 185.2 50 150 0.10 0.12 0.562 192.9
12 350 0.20 0.03 1.006 192.3 51 150 0.15 0.12 0.640 186.7
13 450 0.05 0.03 0.521 156.4 52 150 0.20 0.12 0.715 198.5
14 450 0.10 0.03 0.994 169.4 53 250 0.05 0.12 0.631 153.5
15 450 0.15 0.03 0.841 181.8 54 250 0.10 0.12 0.395 157.5
16 450 0.20 0.03 1.579 183.2 55 250 0.15 0.12 0.609 165.9
1.285* 140.0* 56 250 0.20 0.12 1.310 168.1
17 150 0.05 0.06 0.637 179.4 57 350 0.05 0.12 1.014 196.0
18 150 0.10 0.06 0.269 186.4 58 350 0.10 0.12 1.120 192.4
19 150 0.15 0.06 0.491 190.6 59 350 0.15 0.12 1.520 194.5
20 150 0.20 0.06 0.830 205.3 60 350 0.20 0.12 1.450 185.5
21 250 0.05 0.06 0.987 149.8 61 450 0.05 0.12 1.111 189.2
22 250 0.10 0.06 0.947 166.0 62 450 0.10 0.12 1.294 194.7
23 250 0.15 0.06 0.758 169.8 63 450 0.15 0.12 1.916 192.8
24 250 0.20 0.06 0.923 186.8 64 450 0.20 0.12 1.735 203.3
25 350 0.05 0.06 0.616 170.8 * Un-burnished
26 350 0.10 0.06 0.901 182.5
4. ARTIFICIAL NEURAL NETWORK
27 350 0.15 0.06 0.602 188.4 Artificial neural networks are computing elements,
28 350 0.20 0.06 1.036 200.3 which are based on the structure and function of the
29 450 0.05 0.06 1.093 170.4
biological neurons. These networks have nodes or
neurons, which are described by difference or
30 450 0.10 0.06 0.948 182.8 differential equations. The nodes are interconnected
31 450 0.15 0.06 1.401 191.3 layer-wise or intra-connected among themselves. Each
32 450 0.20 0.06 1.341 node in the successive layer receives the inner product
189.0
of synaptic weights with the outputs of the nodes in the
1.216* 140.0* previous layer. The inner product is called the
33 150 0.05 0.09 0.412 146.3 activation value. When the vectors are binary or bipolar,
34 150 0.10 0.09 0.288 148.6
hard-limiting non-linearity is used. When the vectors
are analog, a squashed function is used. Some of the
35 150 0.15 0.09 0.480 161.5 squashed functions are sigmoid (0 to 1), tanh (-1 to +1),
36 150 0.20 0.09 0.621 168.4 Gaussian, logarithmic and exponential.
37 250 0.05 0.09 0.631 156.1

394
4.1 Radial Basis function (RBF) more than one output is considered, a combination
Figure 4 shows the flow of information from input criterion has to be considered. The total number of
layer through RBF layer and finally to the output layer. patterns decided for each class after selection for
The concept of distance measure is used to associate training and testing are given in Table 4.
the input and output pattern values. Radial Basis
Functions are capable of producing approximations to Table 4. Distribution of patterns and their classification range
an unknown function ‘f’ from a set of input data (Patterns used for Training and Testing RBF)
abscissa. The approximation is produced by passing an
input point through a set of basis functions, each of Serial Number Class Ra Range of HV
which contains one of the RBF centres, multiplying the
result of each function by a coefficient and then 1 I {50}(≤1.225) {27}(≤175)
summing them linearly. The data have been trained 2 II {14}(>1.225) {37}(>175)
using radial basis function (exponential function). The
Ra 1, 2, 3, 5, 6, 7, 9, 10, 12, 13, 14, 15, 17,
RBF learns all 40 training patterns during training (class I): 18, 19, 20, 21, 22, 23, 24, 25, 26, 29, 30,
process. 31, 32, 33, 34, 35, 36, 37, 38, 39, 41, 42,
43, 44, 45, 46, 49, 53, 54, 55, 57, 58, 59,
Hidden 60, 61, 62.
Layers
Input Output
Inputs Layer Layer
Ra 8, 11, 16, 27, 28, 40, 47, 48, 50, 51, 52,
(class II): 56, 63, 64
Roughness,
B. Speed
Ra HV 1, 2, 3, 4, 13, 14, 21, 22, 23, 25, 29, 33,
B. Depth
(class I): 34, 35, 36, 37, 38, 39, 41, 42, 43, 45, 46,
53, 54, 55, 56,
Hardness,
B. Feed rate
HV HV 5, 6, 7, 8, 9, 10, 11, 12, 15, 16, 17, 18,
(class II): 19, 20, 24, 26, 27, 28, 30, 31, 32, 40, 44,
47, 48, 49, 50, 51, 52, 57, 58, 59, 60, 61,
62, 63, 64.

Number in the ( ) indicates the threshold

X1
W 1i Number in the {} indicates total number of patterns
considered in each class for each target variable
X2
W 2i ∑ W ij Xi
Inputs Output Yj

X3 W 3i Out of these patterns, the number of patterns used for

training the network is 40 patterns where the remaining
Activation patterns (24) are used for testing the classification
Weights Function performance of the network. The patterns for training
the network were selected in away to represent the
Fig.4 Configuration of ANN showing details of relationship entire population of the data. The selection of patterns
between input and output of one neuron
is done by:
nf 2
5. IMPLEMENTATION OF ANN
j 1 x ij xj
5.1 Normalizing the data Ei2 2 . (2)
The patterns are normalized so that the values of the i
features are in the range of 0 to 1. By this process, the And
computational complexity is reduced. The
nf 2
normalization of the patterns is done by using
2 j 1 x ij xj
following equations: i . (3)
L
xi Where (Ei2) is the total variance of a pattern, (nf) is the
xi . (1) number of features, (xj) is the mean for each feature,
xmax and ( i2) is the variance. The value of Ei2 is found for
each pattern. Patterns with maximum Ei2 are chosen
Where (xi) is the value of a feature, and (xmax) is the from each class for training the network. For each
maximum value of the feature. pattern target output is defined, so that during training,
supervised algorithms can be used and during testing,
5.2 Training and testing data the required output can be obtained by giving inputs.
Based on the classification range of the outputs, two The classification range is given in Table 4 and also the
classes (I and II) are decided. If only one output is training pattern numbers corresponding to Table 3.
considered, the range of classification is simple. If

395
5.3 Implementation of RBF (b) % Decrease in Roughness vs Burnishing depth for different burnishing
Following are the steps of implementing BRF speeds. (at 0.06mm/rev burnishing feed )
120
1. Decide number of centres. A centre is one of the At 150rpm

Decreasing in Ra, %
training patterns or an average of few training 80
patterns. At 250rpm

40
2. Find the distance between a pattern and a centre At 350rpm
and hence for all centres. The number of centres is 0
At 450rpm
based on the number of nodes in the hidden layer 0.00 0.05 0.10 0.15 0.20 0.25
-40
3. Apply exponential (-x) activation and hence RBF Burnishing depth, mm
is obtained.
4. Add a bias of ‘1’ to the RBF and hence obtain an
RBF matrix (3 x 5). The value 3 represents three (c) % Decrease in Roughness vs Burnishing depth for different burnishing
speeds. (at 0.09mm/rev burnishing feed )
nodes in the input layer and 5 represent 4 RBF and
120
‘1’ bias in the hidden layer. At 150rpm

Decreasing in Ra, %
80
5. Find the inverse of the RBF matrix and process At 250rpm
with the target values to obtain the final weight 40
At 350rpm

0
6. RESULTS AND DISCUSSIONS 0.00 0.05 0.10 0.15 0.20 0.25
At 450rpm
Figures 5&6 show decreasing and increasing (%) in -40
roughnes and hardness of burnishing surfaces at Burnishing depth, mm
different burnishing parameters as measured within the
course of this work. The figures are presented for 150- (d) % Decrease in Roughness vs Burnishing depth for different burnishing
speeds. (at 0.12mm/rev burnishing feed )
450 rpm for different feed rates 0.03-0.12mm/rev. The
effects of burnishing depth on roughness at different 80 At 150rpm
Decreasing in Ra, %

burnishing speeds and different feed rates are shown in

40 At 250rpm
Fig.5. Fig.5a shows that the maximum percantage of
reduction achieved in Ra is 200% at lower burnishing 0 At 350rpm
speeds (150 and 250rpm). However at higher 0.00 0.05 0.10 0.15 0.20 0.25
-40 At 450rpm
burnishing speed (350 and 450 rpm), the % reduction
in the reoughness becomes less, i.e by about 50%. -80
Burnishing depth, mm

Further increase in the burnishing feed rate, showed

less extent of reduction in surface roughnes as shown Fig.5 Decreasing % of roughness vs burnishing depth at
different burnishing speeds and feed rates
in Figs.5b-d, especially at higher burnishing speeds and
burnishing depth. Thus increase both burnishing speed
Similarly, the maximum % improvement in the
and burnishing depth have negative impact on
hardness of burnishing surfaces is occurred at lower
improvement of surface roughnes especially at higher
feed rate, Fig.6a, i.e. 50%. At lower feed rates (0.03,
feed rate, Fig.5d. Burnishing speed 450 rpm produced
0.06, and 0.9 mm/rev), Fig.6a-c indicates that
rougher surface compared to unburnshed onse as
burnishing depth has a substantial contribution in
shown in Fig.5c & d.
increasing the hardness of burnishing surface. However,
by increasing the feed rate up to 0.12mm/rev, Fig.6d,
(a) % Decrease in Roughness vs Burnishing depth for different burnishing insignificant increase of hardness is achieved.
speeds. (at 0.03mm/rev burnishing feed )
220 (a) % Increase in HV vs Burnishing depth for different burnishing speeds
180 At 150rpm (at 0.03mm/rev burnishing feed)
Decreasing in Ra, %

50
140 At 150rpm
Increasing in HV, %

At 250rpm 40
100 At 250rpm
30
60 At 350rpm
20 At 350rpm
20
-20 At 450rpm 10
At 450rpm
0.00 0.05 0.10 0.15 0.20 0.25 0
-60 0.00 0.05 0.10 0.15 0.20 0.25

Burnishing depth, mm Burnishing depth, mm

396
 (b) % Increase in HV vs Burnishing depth for different burnishing speeds (d) % Increase in HV vs Burnishing depth for different burnishing speeds
(at 0.06mm/rev burnishing feed) (at 0.12mm/rev burnishing feed)
50 50

Increasing in HV, %
At 150rpm At 150rpm
Increasing in HV, %

40 40

30 At 250rpm 30 At 250rpm

20 20
At 350rpm At 350rpm
10 10
At 450rpm 0 At 450rpm
0
0.00 0.05 0.10 0.15 0.20 0.25
0.00 0.05 0.10 0.15 0.20 0.25
Burnishing depth, mm Burnishing depth, mm

(c) % Increase in HV vs Burnishing depth for different burnishing speeds

(at 0.09mm/rev burnishing feed)
Fig.6 Increasing % of hardness vs burnishing depth at
50
different burnishing speeds and feed rates
Increasing in HV, %

At 150rpm
40

30 At 250rpm Thus, the results shown in Figs 5&6 indicate that lower
20
At 350rpm values of burnishing feed rate and speed are favorate in
10

0 At 450rpm
improving the surface roughness and hardness.Table 5
0.00 0.05 0.10 0.15 0.20 0.25 lists the predicted values of Ra and HV by using ANN,
Burnishing depth, mm
i.e. RBF.

Table 5 Predicted Ra and HV by RBF over 24 tested patterns

Pattern B. Speed B. depth B. Feed Measured Measured Predicted Predicted

(rpm) (mm) (mm/rev) Ra ( m) Hardness (Ra) by (HV) by
(HV) RBF RBF
2 150 0.1 0.03 0.333 150.9 0.6837 167.2451
4 150 0.2 0.03 0.629 165.3 0.7435 180.9562
6 250 0.1 0.03 0.946 191.7 0.7419 168.5599
11 350 0.15 0.03 1.318 185.2 1.0316 178.7787
12 350 0.2 0.03 1.006 192.3 1.2005 185.6389
16 450 0.2 0.03 1.579 183.2 1.5242 191.2632
18 150 0.1 0.06 0.269 186.4 0.6874 167.3657
19 150 0.15 0.06 0.491 190.6 0.7071 173.9312
20 150 0.2 0.06 0.83 205.3 0.7472 181.0644
21 250 0.05 0.06 0.987 149.8 0.6792 162.8382
26 350 0.1 0.06 0.901 182.5 0.8839 172.5807
27 350 0.15 0.06 0.602 188.4 1.0351 178.889
31 450 0.15 0.06 1.401 191.3 1.303 184.8423
32 450 0.2 0.06 1.341 189 1.5272 191.3622
33 150 0.05 0.09 0.412 146.3 0.6999 161.8431
36 150 0.2 0.09 0.621 168.4 0.7584 181.3887
37 250 0.05 0.09 0.631 156.1 0.6905 163.2116
43 350 0.15 0.09 0.64 164.9 1.0454 179.2192
47 450 0.15 0.09 1.52 175.1 1.3126 185.1564
48 450 0.2 0.09 1.45 184.9 1.5363 191.6588
50 150 0.1 0.12 0.562 192.9 0.7175 168.3269
52 150 0.2 0.12 0.715 198.5 0.7769 181.9273
58 350 0.1 0.12 1.12 192.4 0.9125 173.5044
63 450 0.15 0.12 1.916 192.8 1.3286 185.678
397
For more clarity, the two process state variables Ra and
HV, measured and predicted, are plotted in Fig.7 for 24
tested patterns listed in Table 5. The figure compares
the closeness between measured values of Ra and HV
(Table 5, columns 5&6) with the predicted values by
RBF (Table 5, columns. 7&8).
Fig. 7 Target and predicted Ra and HV versus 24 patterns
6.1 Performance of RBF in predicting Ra and HV
The maximum prediction percentage during testing
depends upon the type of data and number of centres.
Table 6 summaries the prediction performance of RBF.
Table 6 RBF Performance
Items
Configuration of ANN (RBF) 3x5x1
% Classification Total test patterns = 40 90.62 (58)*
Iterations 1
Computational effort for one pattern 620 (Equation 4)
Total Computational effort for all the 24800
training patterns until convergence
* Number in the parenthesis is the total number of
patterns predicted
TCE for RBF = 2nc2 np +inv(nc2) np nc2 . (4)
Where; TCE: Total computational effort, nc :Number
of centers+1(bias), np : Number of training patterns inv
is inbuilt function inverse of a matrix
398
6. CONCLUSIONS
Highest reduction in Ra (200%) was achieved at
lower burnishing speeds. Increasing both
burnishing speed and depth showed a negative
impact on improvement of surface roughnes
especially at higher feed rate. 50 % improvement
in the hardness of burnishing surfaces was
achieved at lower feed rate. Furthermore, at lower
feed rates burnishing depth has a substantial
contribution in increasing the hardness of
burnishing surface.
A radial basis function algorithm was successfully
used to predict the surface roughness and hardness
for the burnishing brass surfaces. RBF has proved
to be the best when dealing with nonlinear data (i.e.
Ra and HV). In implementing RBF the number of
arithmetic operation is substantially small
compared with other training algorithms such as
back propagation algorithm.
REFERENCES
El-Tayeb N.S.M., K.O. Low, P.V. Brevern. 2007.
Influence of roller burnishing contact width and
burnishing orientation on surface quality and
tribological behaviour of Aluminium 6061. J. Mater.
Process. Technol, 186: 272–278.
luca L., S. Neagu-Ventzel and I. Marinescu. 2005.
Effects of working parameters on surface finish in
ball-burnishing of hardened steels, Pre. Eng. 29:
253-256.
El-Axir M.H., O.M. Othman and A.M. Abodiena. 2008.
Improvements in out-of-roundness and
microhardness of inner surfaces by internal ball
burnishing process, J. Mater. Process. Technol. 196:
120-128
Hua W., Y.J. Man and B. Liu. 2008. Contact force
studies of a burnishing slider, Tri. Int. 41: 60-66
Shiou, F.J. and Chen, C.H. 2003. Freeform surface
finish of plastic injection mold by using ball-
burnishing process, J. Mater. Process. Technol., 140
(1-3): 248-254.
Korzynski M. 2007. Modeling and experimental
validation of the force-surface roughness relation
for smoothing burnishing with a spherical tool, Int.
J. Mach. Tools Manuf. 471: 1956-1964.
El-Tayeb N.S.M., K.O. Low, P.V. Brevern. 2006.
Enhancement of Surface Quality and Tribological
Properties using Ball Burnishing Process,
Machining Science and Technology (JMST) 12(2):
234-248.
El-Tayeb N.S.M., K.O. Low, P.V. Brevern. 2009. On
the Surface and Tribological Characteristics of
Burnished Cylindrical Al-6061. Tribo. Int., 42(2):
320-326.
El-Tayeb N.S.M., K.O. Low, P.V. Brevern. 2008. The
influence of roller burnishing process on hardness
and roughness of cylindrical polymer surfaces,
 IMechE 2008, Proc. IMechE Vol. 222 Part J: J. Eng
Tribol.: 947-955.
Ong Low K.., N.S.M. El-Tayeb, P.V. Brevern. 2008.
Experimental study on the sliding friction and wear
characteristics of roller Burnished polymer. Intern.
Polym. Process. 23(3): 292-299
Moshkovich A., V. Perfilyev, K. Yutujyan and L.
Rapoport. 2007. Friction and wear of solid lubricant
films deposited by different types of burnishing,
Wear 263: 1324-1327
Man Y.J., S.k. Yu and B. Liu.2006. Characterization
and formation mechanism understanding of
asperities to be burnished, J. M.M. Mater. 303: 101-
105.
Shereehah Tareq A. Abu. 2008. Developing and
investigating of elastic ball burnishing tool, Int. J.
Adv. Manuf. Technol. 36: 270-279
El-Axir M.H., O.M. Othman and A.M. Abodiena. 2008.
Study on the inner surface finishing of aluminum
alloy 2014 by ball burnishing process, J. Mater
Process Technol, 202: 435-442
El-Axir M.H., O.M. Othman and A.M. Abodiena. 2008.
Improvements in out-of-roundness and
microhardness of inner surfaces by internal ball
burnishing process, J Mater Process Technol. 196:
120-128
Shiou F.J. and C.H. Cheng. 2008. Ultra-precision
surface finish of NAK80 mould tool steel using
sequential ball burnishing and ball polishing
processes, J. Mater Process Technol 20: 554-559
399
Hassan A.M., Al-Jalil, H.F., Ebied. A.A. 1998.
Burnishing force and number of ball passes for the
optimum surface finish of brass components. J.
Mater Process Technol 83: 176-179
Esme Ugur, Aysun Sagbas, Funda Kahraman, M.
Kemal Kulekci. 2008. Use of artificial neural
networks in ball burnishing process for the
prediction of surface roughness of AA7075
aluminum alloy. Materiali in tehnologije /
Materials and technology, MTAEC9 42(5): 215-
219
El-Khabeery M.M. and M.H. El-Axir. 2001.
Experimental techniques for studying the effects of
milling roller-burnishing parameters on surface
integrity, Int. J. Mach. Tools Manuf. 41(12): 1705–
1719
Hassan A.M. and A.M.S. Momani. 2000. Further
improvements in some properties of shot peened
components using the burnishing process, Int. J.
Mach. Tools Manuf. 40(12): 1775–1786.
El-Axir M.H. and M.M. El-Khabeery. 2003. Influence
of orthogonal burnishing parameters on surface
characteristics for various materials, J. Mater.
Process. Technol. 132: 82–89.
Luo H.Y., J.Y. Liu, L.J. Wang and Q.P. Zhong. 2006.
Study of the mechanism of the burnishing process
with cylindrical polycrystalline diamond tools, J
Mater Process Technol, 180: 9-16.
El-Axir M. H. and A. A. Ibrahim. 2005. Some surface
characteristics due to center rest ball burnishing, J.
Mater Process Technol. 167: 47-53.