Composites Science and Technology 63 (2003) 539–548

www.elsevier.com/locate/compscitech

Multiple regression and neural networks analyses in

composites machining
J.T. Lina,*, D. Bhattacharyyaa, V. Kecmanb
a
Centre for Composites Research, Department of Mechanical Engineering, University of Auckland, Private Bag 92019, Auckland, New Zealand
b
Dynamics and Control Systems Research Group, Department of Mechanical Engineering, University of Auckland, Private Bag 92019,
Auckland, New Zealand

Received 29 June 2001; accepted 2 April 2002

Abstract
The machining forces-tool wear relationship of an aluminium metal matrix composite has been studied in this paper using mul-
tiple regression analysis (MRA) and generalised radial basis function (GRBF) neural network. The results show that using the
force-wear equation derived from MRA is a fairly accurate way of predicting the attainment of prescribed tool wear. However, the
use of a neural network analysis can further improve the accuracy of the tool wear prediction particularly when the functional
dependency is nonlinear. It is evident that the relationship derived from the feed force data is more accurate than that derived from
the cutting force.
# 2002 Elsevier Science Ltd. All rights reserved.
Keywords: A. Metal-matrix composites; B. Wear; C. Statistics; Neural networks

1. Introduction right time for tool change, continuous monitoring of the

Increasing quantities of metal matrix composites
(MMCs) are being used to replace conventional materi-
als in many applications, especially in the automobile
and recreational industries where the performance
requirements are getting more demanding. Among the
wide diversity of MMCs, the most popular types of
MMCs are aluminium alloys reinforced with ceramic
particles. These low cost composites provide higher
strength, stiffness and fatigue resistance [1,2], with a
minimal increase in density over the base alloy. The
superior mechanical properties achieved by the reinfor-
cements in MMCs, on the other hand, make them hard
to be machined to a required shape. Therefore, only
polycrystalline diamond (PCD) tools are often found to
be suitable for machining these MMCs for a reasonable
length of cutting time [3–5].
Due to the relatively high cost of PCD tools, they
need to be changed in time for resharpening to max-
imise the total useful tool life. In order to decide the
* Corresponding author. Tel.: +64-9-373-7599x4543; fax: +64-9-
373-7479.
E-mail address: rj.lin@auckland.ac.nz (J.T. Lin).
0266-3538/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved.
PII: S0266-3538(02)00232-4
tool wear needs to be done during the machining oper-
ation. The commonly used experimental method of
examining the tool wear using microscopy involves
interruptions in the cutting process. However, an indir-
ect way of monitoring the tool wear, in which a mea-
surable output might be used to indicate the extent of
tool wear without interrupting the machining process,
would be more suitable for practical applications. Such
outputs may be the cutting and feed forces that are
dependent on tool wear [6,7].
From previous research [3–7], it is now well known
that the primary wear mechanism in machining parti-
culate reinforced MMC is abrasion by the reinforcing
particles on the flank surface of the tool. Figs. 1 and 2
show the curves of the typical tool wear growth and
changes of machining forces for machining the DUR-
ALCAN1 (registered trademark of Alcan Aluminium
Ltd.) MMC (SiC particulate reinforced aluminium metal
matrix composite material, A359/SiC/20p) in three dif-
ferent cutting conditions. It is clear that the tool wear
and the two machining forces increase steadily with time
following the similar trends. The general Taylor’s equa-
tion showing the relationship between tool wear and
cutting conditions has been derived in an earlier paper
540 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548
Fig. 1. The comparison of curves of the typical tool wear growth and
changes of feed force in three different cutting conditions.
Fig. 2. The comparison of curves of the typical tool wear growth and
changes of cutting force in three different cutting conditions.
[7], which may be used to predict the tool life when
machining the SiC reinforced aluminium composites.
The similar trends of change in the tool wear and
machining forces suggest that the measurement of
machining forces during the machining process may be
used as an indirect way of monitoring tool wear if a
proper relationship can be found between these two
parameters.
In this paper the relationship between the tool wear
and the machining forces during machining DUR-
ALCAN material has been established using both mul-
tiple regression and artificial neural network analyses. A
general equation involving machining conditions (cut-
ting speed and feed rate), machining force (either feed
or/and cutting) and tool wear has been derived from
multiple regression analysis (MRA), and the same data
have been analysed using generalised radial basis func-
tion neural network. The accuracy of predicting tool
wear using both methods will also be compared.
2. Background of neural network
The use of neural networks (NNs) represents a new
methodology in many different applications including
tool condition monitoring for machining of traditional
enginering materials [8–22]. It is a promising field of
research in predicting experimental trends and has
become increasingly popular in the last few years as
they can often solve problems much faster compared to
other approaches with the additional ability to learn.
Generally, a neural network means a network of many
simple processors (‘‘units’’) operating in parallel, each
possibly having a small amount of local memory. The
units are connected by communication channels (‘‘con-
nections’’) which usually carry numeric (as opposed to
symbolic) data, encoded by one of various ways [23].
One of the best known examples of a biological neural
network is the human brain. It has the most complex
and powerful structure which, by learning and training,
controls human behaviour toward responding any
problem encountered in every-day life. As for the artifi-
cial neural networks (ANN), they have been developed
to try to emulate this biological network for the purpose
of learning the solution to a physical problem from a
given set of examples.
Among the many neural networks that have been
developed, the most popular neural networks are
known respectively as the multilayer perceptron (MLP)
and the generalised radial basis function (GRBF) net-
work [8,9]. These networks currently form the basis for
the majority of practical applications. In this paper,
however, the GRBF network is chosen, since it provides
several advantages over the MLP network and has been
proven successfully in different applications [9,19–22].
The GRBF network is based on the simple intuitive idea
that an arbitrary function y(x) can be approximated as
the linear superposition of a set of localised basis func-
tions jj(x). A network diagram, as shown in Fig. 3, can
represent the structure of the GRBF network. The
architecture of a GRBF network consists of three lay-
ers—input, output and hidden layers. The hidden layer,
situated between the input and output layer, is where
the basis functions operate to intervene between the
 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548
Fig. 3. Architecture of a high dimensional input and output GRBF
network [9], that maps c-dimensional input vectors into d-dimensional
output, ie. It represents a Rc!Rd function.
external input and the network output. For most applica-
tions the basis functions are commonly chosen to be the
Gaussian. Each of these functions Gk computes a localised
function of the input vector. The connecting lines between
the inputs and the Gaussian functions represent (a) the
elements of the vector
k that describes the location of the
centre (in input space) and (b) the elements of the vector
 k that describes the width or standard deviation of the
Gaussian function (in input space). The lines connecting
the Gaussian functions to the outputs represent the
weights w of the neural network which are comparable to
the parameters in the case of curve fitting with simple
polynomials. The outputs are then obtained as linear
combinations of the values of the Gaussian functions.
3. Experimental procedure
The type of MMC studied in this paper consists of
A359 aluminium (Si: 9.23
9.31, Fe: 0.08
0.15, Cu:
0.01
0.03, Mn: < 0.02, Mg: 0.55
0.61, Zn: < 0.01, Ti:
< 0.02, Sr: 0.013
0.015, Al: REM, in wt.%) reinforced
with 20 vol.% of silicon carbide (SiC) particles of 12.8-
mm mean diameter [4] and is processed by ingot metal-
lurgy with the ceramic particles stirred into the melt.
The specimens came in as round bars with diameters
around 175 mm, which allowed the machining experi-
ments to be conducted at high linear cutting speeds.
The cylindrical bars were turned on an infinitely vari-
able speed control lathe with polycrystalline diamond
(PCD) inserts, TPG322 COMPAX 1500 (Appendix A),
at three different speeds (300–700 m/min) with three
different feed rates (0.1–0.4 mm/rev) and a selected
constant depth of cut of 0.5 mm. No lubricant or cool-
ing fluid was used during machining. Table 1 summaries
the details of the experimental conditions. The use of
PCD inserts in this study is due to its excellent perfor-
mance compared to ceramic tools when machining alu-
minium MMCs [4,6,7].
541
Table 1
Experimental details
Workpiece material Direct-Chill Cast A359/SiC/20p
(175 mm400 mm long diameter billets)
Machine Dean Smith & Grace 1609 centre lathe
Tool insert TPG322 COMPAX 1500 (PCD)
Tool holder TARP-16-3hR175.2-2525-16
(Sandvik Coromat)
Back rake angle: 6
Cutting parameters Cutting speed: 300, 500 and 700 m/min
Feed rate: 0.1, 0.2 and 0.4 mm/rev
Depth of cut: 0.5 mm
Side cutting edge angle: 0
Dry (no lubricant or cooling fluid applied)
Tool dynamometer Kistler 3-D force dynamometer
(Type 9441)
The tool flank wear, VBmax as shown in Fig. 4, was
measured at predetermined intervals using optical
microscope. Both the cutting and feed forces were
monitored at every pass using a piezoelectric 3-D force
dynamometer (Kistler 9441).
4. Data analysis
For both multiple regression and neural network
analyses, feed force and cutting force data were used to
derive the generalised force-tool wear equations. The
independent variables used were cutting speed, feed rate
and feed or/and cutting force, and the dependent vari-
able was the tool flank wear.
Basically the neural network analysis was divided into
three parts [24]: (1) data preprocessing; (2) hidden layer
training; and (3) output layer training and neural net-
work testing. The data preprocessing was to restructure
the measured data which could facilitate the following
training and testing procedures. After data preproces-
sing, the hidden layer training was to optimise both the
Gaussian functions’ centres and the standard devia-
tions. In optimising the Gaussian functions’ centres, the
number of the Gaussian functions (k) needed to be
firstly determined by a supervised training process and
then followed the determination of the locations of the
centres in the input space. Once the locations of the
Fig. 4. Criteria based on features of tool flank wear.
542 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548
Gaussian functions were determined, the standard
deviations, which affect the degree of overlapping of the
Gaussian functions, could be calculated and adjusted by
the fine-tune parameter kks. This parameter also had to
be determined by supervising the training process.
Lastly, the performance of the GRBF neural network
model [9] was evaluated through the output layer train-
ing and neural network testing.
The optimum number of Gaussian functions (k) and
the optimum standard deviation fine-tune parameter
(kks) were obtained by repeating the whole hidden and
output layer trainings as well as the neural network
testing until the performance criterion (a low value of
the overall mean testing error) was achieved. The final
performance of the GRBF model with the optimised
parameters for the tool wear-machining force relation-
ship was verified using the validation data set that was
totally independent from the data used for hidden layer
and output layer trainings.
5. Results and discussion
5.1. Multiple regression analysis (MRA)
In order to find the relationship between the tool wear
and the two machining forces, firstly, a linear regression
analysis has been carried out individually on the
experimental data for each different cutting condition.
In this part of analysis, two forms of data fit, a simple
linear form and power law, have been implemented. The
results are shown in Figs. 5 and 6 for feed force and
cutting force respectively. From these individual linear
regression results, it appears that the tool wear-feed
force equation can describe the experimental behaviour
more accurately than the equation relating tool wear
and cutting force. In most cases of the individual ana-
lysis, the power law relationship has been more accurate
Fig. 5. Typical regression result of the tool wear against feed force in
the cutting conditions of 300 m/min and 0.1 mm/rev.
than the linear equation in matching the experimental
data; that is evident from the value of the determination
coefficient R2. In general, though, the accuracy of both
linear form and power law is good. Therefore, with any
one particular cutting condition, machining force-tool
wear relationship can be derived in either of the two
forms to be used to predict the tool wear with sufficient
accuracy.
However, for practical machining operations, differ-
ent cutting conditions have to be employed to satisfy
different requirements. Since the cutting speed and the
feed rate are the most important parameters as far as
the machinability is concerned, while deriving the gen-
eral equations over a range of cutting conditions both of
them have been taken into consideration.
As there is more than one independent variable
involved in the analysis, the technique of multiple
regression has been carried out using all the experi-
mental data obtained from different machining condi-
tions applied in this study. Using feed force, cutting force
and both forces, the general equations for tool wear are
subsequently obtained by Eqs. (1)–(3), respectively:
W ¼ 102:439  V0:149  f 0:197  Fx1:15 ;
ð1Þ
R 2 ¼ 0:929;
and
W ¼ 106:162  V 0:188  f 1:226  Fz2:017
ð2Þ
R 2 ¼ 0:887;
and
W ¼ 102:719  V 0:119  f 0:273  Fx1:046  Fz0:159 ;
ð3Þ
R 2 ¼ 0:884
Fig. 6. Typical regression results of the tool wear against cutting force
in the cutting conditions of 300 m/min and 0.1 mm/rev.
 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548 543

where W is tool wear in mm, V is cutting speed in m/ 5.2. Neural network analysis
min, f is feed rate in mm/rev, Fx̂ and Fẑ are feed force
and cutting force respectively in Newton (N). When Both feed force data set and cutting force data set
compared with the experimental data obtained for spe- have been pre-processed to the required format for the
cific cutting conditions, the value generated by the gen- training procedure. After the repeated training and
eral equations can predict the trend with sufficient testing procedure, the final GRBF model configurations
accuracy even though, as expected, it is not as accurate for machining forces-tool wear relationships can be
as those predicted by the equation derived using data obtained. When reapplying these optimised parameters
from that particular cutting condition. However, it is into the GRBF neural network model, the prediction of
interesting to note that when a 10% error zone is put tool wear can be made. Figs. 9 and 10 show the exam-
in, the band covers most of the experimental data, as ples of experimental data comparisons with the two
shown in Figs. 7 and 8. Overall speaking, the derived predictions from multiple regression and neural net-
general equations can be used as reasonably reliable work analyses under the two extreme cutting conditions
guides for deciding the extent of tool wear in machining studied in this paper. In the case of feed force, it can be
the DURALCAN1 MMC (A359/SiC/20p) under var- observed that when the force level is low (less than
ious cutting conditions using PCD tools. about 70 N), the relationship between tool wear and
Judging from the value of R2 (the square of multiple feed force follows almost a linear pattern, so do the
correlation coefficient), it is also evident that even in the predictions from both multiple regression and neural
general equation, using only feed force gives the best
indication for the extent of tool wear. The use of both
feed and cutting forces together renders worse results
while the worst comes from using only cutting force
data. The reason for this may be attributed to the
greater fluctuation of the cutting force compared to the
feed force. This is caused by irregular high-speed
impacts of the reinforcing particles onto the tool surface
in the cutting direction, which results in the magnitude
of the cutting force to fluctuate during the machining
process. On the other hand, the impact of the particles
onto the tool is much less in feed direction due to the
much smaller magnitude of feed rate compared to that
of the cutting speed. Therefore, even though the tool
goes into the material in the feed direction with a similar
probability of encountering the reinforcing particles, the
fluctuation for feed force is not so severe as that for the
cutting force. Fig. 8. Typical general regression result of the tool wear against cut-
ting force in the cutting conditions of 300 m/min and 0.1 mm/rev.

Fig. 9. Comparison between experimental data and predictions from

multiple regression analysis and neural network analysis for the tool
Fig. 7. Typical general regression result of the tool wear against feed wear against feed force in the cutting conditions of 700 m/min and 0.4
force in the cutting conditions of 300 m/min and 0.1 mm/rev. mm/rev.
544 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548
Fig. 10. Comparison between experimental data and predictions from
multiple regression analysis and neural network analysis for the tool
wear against cutting force in the cutting conditions of 700 m/min and
0.4 mm/rev.
network model. Also the difference between the two
predictions can hardly be distinguished. However, when
the high force level is approached, the relationship
between tool wear and feed force becomes nonlinear.
Then the predictions from these two separate approa-
ches highlight their differences. It becomes clear that the
prediction from neural network model follows the
experimental data much more closely than that from
multiple regression analysis. Furthermore, due to the
wide range of machining conditions employed in this
study, the continuous 3-D models based on the derived
tool wear-machining force relationships from both ana-
lyses can be developed, one of the typical examples is
Fig. 11. Examples of 3-D models for comparison between experimental data and predictions from multiple regression analysis and neural network
analysis. Cutting speed is 700 m/min, while the changing parameters are feed rate and feed force.
shown in Fig. 11. The discrete experimental data have
also been superimposed to indicate how well these
models fit the experimental data. As mentioned earlier,
the difference between the multiple regression and
neural network lies mainly in the nonlinear regions
(higher force levels). Comparing the plots in Fig. 11(a)
and (b) as well as Fig. 12(a) and (b), it is evident that
MRA model overpredicts some of the data points
around the high tool wear region while the NN model
depicts the empirical trend reasonably well.
These different but close results produced from the
two methods can be easily explained. As a matter of
fact, the multiple regression analysis, which is based on
linear regression technique, can be regarded as a special
case of neural networks [22]. Instead of using a non-
linear function as the activation function, in linear
regression analysis, a linear activation function is used.
Hence the GRBF neural network employed in this
paper, which uses a nonlinear Gaussian function, is able
to make more accurate predictions in the nonlinear
region of the machining forces-tool wear relationship. A
very similar scenario can also be observed from the
results produced by the cutting force data set. Judging
from the value of the mean testing error produced
during the training processes of both data sets (0.3724
and 0.5474 for feed force and cutting force respec-
tively), it is again clear that the wear prediction from
feed force is more accurate than that from cutting
force. However, if both feed and cutting force data are
put into the neural network training process, an inter-
esting but unexpected result is noticed in terms of the
value of the mean test error. Instead of producing a
 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548 545

Fig. 12. Examples of 3-D models for comparison between experimental data and predictions from multiple regression analysis and neural network
analysis. Cutting speed is 700 m/min, while the changing parameters are cutting force and feed force.

worse result than that using only feed force data, as was
the case in multiple regression analysis, the outcome
(the mean test error of 0.3264) is actually better than
using either force data set. This could be due to the
neural network analysis being able to map data between
higher dimensional spaces.
5.3. Statistical comparison of both methods
In this paper we have defined and used some statis-
tical terms (Appendix B) to monitor the performance of
both analyses with different scenarios. Table 2 compares
the results between NN and classical non-linear multiple
regression analysis in terms of these three statistical
measures. The results again confirm the superiority of
using the feed force data over cutting force data.
Although both sets of force data can be used together to
produced better outcome than that of using cutting
force, its accuracy can only just match that obtained
from using feed force data. It clearly indicates that the
best result can be obtained using the feed force alone and
it renders that by measuring only feed force the tool wear
monitoring for machining particulate reinforced MMC
materials can be accurately and effectively achieved.

Table 2
Comparison of statistical results for the GRBF model and multiple non-linear regression analysis

GRBF model Multiple non-linear regression

2
rms cov (%) r rms cov (%) r2

(a) Independent data set from the pre-processed data set

Feed force data 0.017 8.19 0.9935 0.024 11.30% 0.9876
Cutting force data 0.018 9.05 0.9844 0.025 12.31% 0.9710
Feed and cutting forces data 0.020 9.97 0.9907 0.024 12.05% 0.9861

(b) Pre-processed data set

Feed force data 0.015 7.39 0.9949 0.024 11.86% 0.9868
Cutting force data 0.021 10.50% 0.9890 0.031 15.42% 0.9763
Feed and cutting forces data 0.015 7.62% 0.9946 0.023 11.29% 0.9882

(c) Original data set

Feed force data 0.023 11.83% 0.9882 0.029 15.14% 0.9791
Cutting force data 0.039 20.72% 0.9627 0.042 22.08% 0.9560
Feed and cutting forces data 0.025 13.23% 0.9842 0.032 17.02% 0.9739
546 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548

The comparisons of the predicted tool wear against
the measured tool wear using both methods are shown
in Fig. 13. A perfect prediction would see all the plotted
points sitting on the 45 line (the solid line), the accu-
racy of each method can be easily compared by the clo-
seness of the data clusters to this line. The 10% lines
(the dotted lines) have also been put into the diagrams
to facilitate the comparison.
One of the advantages of using neural network ana-
lysis is its ability of automatic noise filtering during the
model training procedure; good results can be obtained
without any preprocessing of the recorded data pattern.
However, due to the discrete nature of data recording
during the machining operation and the inevitable min-
ute changes in the operational conditions during data
measurement and recording, proper data preprocessing
can significantly improve the prediction ability of the
trained model.
As shown in the comparisons made above, both
analyses can make descent tool wear prediction using

Fig. 13. Diagrams of predicted tool wear against measured tool wear for the comparison of neural network and multiple regression analyses with
pre-processed data patterns, (a) cutting force data, (b) feed force data, (c) cutting and feed force data.
 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548
Fig. 13 (continued).
machining force data. Although the results produced
by multiple regression analysis have been quite satis-
factory and the difference between these two methods of
The use of GRBF neural network can further
analysis is marginal in some regions, the neural network
model has been proved to be able to further improve the
performance of multiple regression analysis.
6. Conclusions
From the results discussed above, the following con-
cluding remarks can be made:
By measuring the machining forces, tool wear
Due to the superior results achieved by using
can be monitored without interrupting the
machining operation. Hence the decision making
for tool change becomes less time consuming and
maximises the total effective tool life.
General equations for monitoring the tool wear
have been derived from the feed and cutting for-
ces data. For any single cutting condition, either
a power law or a linear equation may be used to
describe the progress of tool wear with similar
accuracy. A general power law equation for a
wide spectrum of cutting conditions is also
obtainable.
Owing to the differences in impact velocities,
the fluctuation observed while measuring the
cutting force is more severe than that for the
feed force measurement. Therefore, in all cases
the feed-force based equation gives higher
547
accuracy than that obtained from the cutting
force.
improve the accuracy of tool wear prediction,
especially within the region where the change of
tool wear exhibits a nonlinear relationship with
the machining forces. However, in the linear
range, both methods perform the prediction with
very similar accuracy. The reason for this can be
attributed to the fact that the regression analysis
is a special case of neural network method, which
employs a linear activation function as the basis
function.
feed force data only, it can be asserted that tool
wear monitoring for machining particulate rein-
forced MMC materials can be accurately and
effectively achieved by measuring only the feed
force.
Appendix A
TPG is an ISO code; T=Triangle; P=Clearance
angle (11 degrees); G=Tolerance on thickness and
inscribed circle ( 0.025 mm).
322 is an American code relating to the dimensions of
the insert: 3=edge length in 10ths of an inch; 2=thick-
ness in 16ths of an inch; 2=nose radius in 64ths of an
inch. This equals to the ISO code 160308: 16=edge
length (16 mm); 03=thickness (3 mm); 08=nose radius
(0.8 mm).
548 J.T. Lin et al. / Composites Science and Technology 63 (2003) 539–548
Compax 1500 refers to the grade of polyscrystalline
diamond supplied by GE; ie the particle size of the dia-
mond used in the matrix: 1500=coarse; 1300=medium;
1600=fine. The coarse grade is usually chosen for wear
resistance, the fine grade for better surface finish on
items such as alloy wheels.
Appendix B
Those terms defined and used in this paper to monitor
the performance of both Multiple Regression and
Neural Networks analyses are the root-mean-square
error (rms), coefficient of variation (cov) and absolute
fraction of variance (r2), which are as follows:
1. rms: root-mean-square error defined as
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP
u n  2
u y  tmea;m
tm¼1 pre;m
rms ¼
n [12] Kuo RJ, Cohen PH. Manufacturing process control through
2. cov: coefficient of variation in per cent defined as
rms
cov ¼   100
tmea 
3. r2: absolute fraction of variance (value of 1
denotes perfect prediction) defined as
n 
P 2
ypre;m  tmea;m
r2 ¼ 1  m¼1 n  2
P [17] Zawada-Tomkiewicz A. Classifying the wear of turning tools
tmea;m
m¼1
where n is the number of data patterns, ypre,m, tmea,m
indicate the predicted value and the measured value
respectively of one data point m, and tmea is the mean
value of all measured data points that were used for the
NN training.
References
[1] Bhattacharyya D, Bowis ME, Gregory JT. The influence of alumina
microsphere reinforcement on the mechanical behaviour and weld-
ability of a 6061 aluminium metal matrix composite. In: Srivatsan
TS, Bowden DM, editors. Machining of composite materials.
Materials Park, (OH, USA): ASM International; 1992. p. 49–56.
[2] Couper MJ, Xia K. Development of microsphere reinforced
metal matrix composites. In: Hansen N et al., editors. Proceedings
12th Riso Int. Symposium on Materials Science. Denmark: Roskilde,
1991, p. 291–8.
[3] Lane C. Machinability of aluminium composite as a function of
matrix alloy and heat treatment. In: Srivatsan TS, Bowden DM,
editors. Machining of composite materials. Materials Park (OH,
USA): ASM International; 1992. p. 3–16.
[4] Lane C. The effect of different reinforcements on PCD tool life
for aluminium composites. In: Srivatsan TS, Bowden DM, edi-
tors. Machining of composite materials. Materials Park, (OH,
USA): ASM International; 1992. p. 17–27.
[5] Tomac N, Tønnessen K. Machinability of particulate aluminium
matrix composites. Annals of the CIRP 1992;41(1):55–8.
[6] Cronjäger L, Meister D. Machining of fibre and particle-rein-
forced aluminium. Annals of the CIRP 1992;41(1):63–6.
[7] Lin JT, Bhattacharyya D, Lane C. Machinability of a silicon
carbide reinforced aluminium metal matrix composite. Wear
1995;181-183:883–8.
[8] Bishop CM. Neural networks and their applications. Review of
Scientific Instruments 1994;65(6):1803–32.
[9] Kecman V. Learning and soft computing. Cambridge (MA,
USA): The MIT Press; 2000.
[10] Das S, Bandyopadhyay PP, Chattopadhyay AB. Neural-net-
works-based tool wear monitoring in turning medium carbon
steel using a coated carbide tool. Journal of Materials Processing
Technology 1997;63(1–3):187–92.
[11] Kuo RJ, Cohen PH. Intelligent tool wear estimation system
through artificial neural networks and fuzzy modelling. Artificial
Intelligence in Engineering 1998;12:229–42.
ðA1Þ
integration of neural networks and fuzzy model. Fuzzy Sets and
Systems 1998;98:15–31.
[13] Ghasempoor A, Moore TN, Jeswiet J. On-line wear estimation
using neural networks. Proceedings of the Institution of
Mechanical Engineers, Part B: Journal of Engineering Manu-
facture 1998;212(B2):105–12.
ðA2Þ [14] Liu Q, Altintas Y. On-line monitoring of tool wear in turning
using a neural network. International Journal of Machine Tools
and Manufacture 1999;39(3):489–504.
[15] Liu Q, Altintas Y. On-line monitoring of flank wear in turning with
multilayered feed-forward neural network. International Journal
of Machine Tools and Manufacture 1999;39(12):1945–59.
[16] Dimla Sr DE, Lister PM. On-line metal cutting tool condition
monitoring II: tool-state classification using multi-layer percep-
tron neural networks. International Journal of Machine Tools
and Manufacture 2000;40(5):769–81.
ðA3Þ
with neural networks. Journal of Materials Processing Technol-
ogy 2001;109(3):300–4.
[18] Sick B. Tool wear estimation in turning with process-specific pre-
processing and time-delay neural networks. International Journal
of Smart Engineering System Design 2001;3(1):1–14.
[19] Elanayar SVT, Shin YC. Design and implementation of tool wear
monitoring with radial basis function neural networks. In: Pro-
ceedings of the 1995 American Control Conference. Part 3 (of 6),
1995. p. 1722–26.
[20] Elanayar SVT, Shin YC. Robust tool wear estimation via radial basis
function neural networks. American Society of Mechanical Engineers,
Production Engineering Division (Publication) PED: 1992. p. 37–51.
[21] Kuo RJ, Cohen PH. Multi-senor integration for on-line tool wear
estimation through radial basis function networks and fuzzy
neural network. Neural Networks 1999;12:155–370.
[22] Shah FF. Radial basis function approach to financial time series
modelling. Master thesis, The University of Auckland, 1998.
[23] Sarle WS ed., Neural network FAQ: Part 1 of 7—Introduction,
periodic posting to the Usenet newsgroup comp.ai.neural-nets.
Available from: ftp://ftp.sas.com/pub/neural/FAQ.html, 1997.
[24] Swider DJ, Kecman V, Browne MW, Bansal PK. Steady-state
modelling of vapour-compression liquid chillers with neural net-
works. Internal report. The University of Auckland, 1998.