A SURVEY OF ARTIFICIAL NEURAL NETWORKS APPLICATIONS IN WEAR AND MANUFACTURING PROCESSES

THE ANNALS OF UNIVERSITY “DUNĂREA DE JOS “ OF GALAŢI
FASCICLE VIII, 2004, ISSN 1221-4590
TRIBOLOGY
35
A SURVEY OF ARTIFICIAL NEURAL NETWORKS
APPLICATIONS IN WEAR AND MANUFACTURING PROCESSES
Minodora RIPA and Laurentiu FRANGU
University “Dunarea de Jos” of Galati, Romania
minodora.ripa@ugal.ro
ABSTRACT
The potential of the artificial neural networks in the field of wear and
manufacturing processes is presented in this paper. Their properties of learning and
nonlinear behavior make them useful to model complex nonlinear processes, better
than the analytical methods. The neural structures, considered appropriate for such
models, are presented. The applications found in the referenced papers mainly
consist of prediction and classification. They present some common points, specific
to the field: wear processes and particles, manufacturing processes, friction
parameters, faults in mechanical structures. The results obtained by the quoted
authors, in their interdisciplinary research are described, proving that neural
networks are an useful tool during the design stage as well as the running or
operation stage.
KEYWORDS: artificial neural networks, wear, manufacturing, prediction,
classification.
1. INTRODUCTION1
In wear and manufacturing processes, very
complex and highly nonlinear phenomena are
involved. Consequently, analytical models are
difficult or impossible to obtain. However, the
improvement of performances and reliability of
mechanical equipment and manufacturing tools
requires accurate modeling and prediction of this
phenomena. For this purpose, Artificial Neural
Networks (ANN) posses a number of properties for
modeling processes or systems: universal function
approximation capability, learning from experimental
data, tolerance to noisy or missing data, and good
generalization capability.
Two main functions of ANN are useful in
tribological applications:
− the continuous approximation of a multivariable
function, used for modeling of processes;
− classification, that is a discrete approximation
of functions, used for recognition of the operation
conditions of machinery.
The former function is usually obtainable by
feed-forward neural network (NN, called Multi Layer
Perceptron (MLP). The latter may be obtained using
self-organizing NN, as Kohonen and Adaptive
Resonance Theory (ART). MLP equally may be used
1
The paper represents an revised version of the paper presented at
NATO ASI “NIMIA-SC 2001” Crema, Italy.
for classification if adding a supplementary
discriminator of the output values. Both functions are
already applied in the field of tribology and
machinery and presented in several recent papers.
This work aims at evaluating the usefulness of
ANN in such applications, as it is reflected in some
relevant papers. It contains a first chapter as
introduction, a second chapter dedicated to the main
ANN structures, chapters three and four dealing,
respectively, with modeling and classification
applications and a conclusions part.
2. DESCRIPTION OF ARTIFICIAL
NEURAL NETWORK STRUCTURES
2.1. Multi Layer Perceptron
An example of a MLP type neural network is
presented in figure 1. Its basic structure is composed
of layers, namely the input layer, hidden layers and
output layer. The input layer accepts data from the
external world, the output layer generates outputs to
the external world. There may be one or more hidden
layers. Each layer consists of a number of nodes
(neurons, cells, processing elements). As shown in
figure 1, each pro-cessing element may have several
input paths but only one output. The inputs of a
neuron may come from the external world (in the
input layer) or from the outputs of other neurons (in

TRIBOLOGY
36
the hidden and output layer). Each neuron sums its
input signals, modified by the interconnection
weights. The sum, modified by an activation function
(frequently a sigmoidal function), is the output of the
neuron.
Fig. 1. Example of a neural network [17].
The information propagates from the input layer
to the output layer, through connections existing only
between elements in adjacent layers. In order to
prevent saturation of the activation function the data
used for train the NN are generally normalized.
The learning procedure starts with computing of
each output Ok, for a specified input vector, X. Then
the error between the computed output Ok and the
desired one, dk is used to modify the weights of the
neurons in order to decrease this error. The procedure
is repeatedly applied for all input vector sets, in order
to minimize the global approximation error of the
network. This is the back-propagation learning
algorithm. It has many versions, aiming at minimizing
the learning time and at a good convergence.
2.2. Kohonen self-organizing neural
network
The network consists of output units, typically
arranged in a two-dimensional plane, with weights
between each unit and input units. When an input
vector is fed to the network, only one output unit,
which has the best-matching weight. i.e. the weight
vector is the closest to the input vector, is selected as
a 'winner'. After learning, units in the network have
modified weights such that neighboring units have
similar weight vectors. Hence similar inputs are
linked with winner units that are located close to each
other, while winner units for different inputs are
located far away in the network. Thus the feature map
is created and inputs to the network are automatically
classified on the map. The advantage of the feature
map is that the weight of each output unit directly
shows a corresponding input vector itself.
The main property of the Kohonen network is
the unsu-pervised learning. That permits to divide the
input vectors set in clusters without prior knowledge
about their similarities.
Fig. 2. The configuration of ART network [29].
2.3. Adaptive Resonance Theory Network
(ART)
The ART network is shown schematically in
figure 2. It has two layers: the first is the
input/comparison layer and the second is the
output/recognition layer. These layers are connected
together with extensive use of feedback from the
output layer to the input layer along with the feed
forward connections. Associated with each
connection, the ART network has feed forward
weights (wjis) from the input layer to the output layer
and feed back weights (tijs), from the output layer to
the input layer. Between the input and output layers
there is also a reset circuit. This circuit is actually
responsible for comparing the evaluated Euclidean
distance (making use of the current inputs and the
most recent weights) to a vigilance threshold that
determines whether the pattern under consideration
pertains to one of the already generated clusters or a
new cluster must be created.
Although the structure is different with respect
to the Kohonen network, their main functions are
similar: unsupervised clustering of the input space,
based on experimental data. They are useful in pattern
classification applications.

TRIBOLOGY
37
3. MODELING AND PREDICTION
The MLP network is suitable for modeling of the
processes. It offers a continuous approximation of a
multivariable function, that is not analytically
obtainable, but it is properly described by the
experimental data set. Usually, the purpose of the
model is the one-step ahead prediction, i.e. to
determine the value of the output function at the
moment t+1, knowing the present and the previous
values of the output and inputs (t is integer). This
ability is used in applications such as tool wear or
lubricant wear prediction. There are also applications
where the current time is not one of the variables,
such as the model of the pressure distribution in a
rectangular gas bearing.
3.1. Off-line modeling
The paper [17], Jones and co-authors present an
ANN that models a simple mechanical system
involving friction and wear, according to three
standard experiments: rub-shoe, pin-on-disk and four
balls models. The input variables are: speed, load,
viscosity, sliding distance, friction coefficient and
temperature. The output functions are the wear
volume (µm3
) and the wear rate (m3
/m).
The purpose of this model is to predict the
lifetime of the friction elements for given operating
conditions (even without carrying on a real
experiment), or to perform accelerated-life testing, on
more complex mechanical systems. A feed-forward
NN is trained on the basis of experimental data to
output the current wear volume or the current wear
rate. On different data set the output yielded by the
NN is compared to the recorded one, in order to
validate the model (fig. 3).
The focus of this work was the evaluation of the
performance of different network dimensions [17],
including recurrent network that filters the input data.
Fig. 3. Comparison between actual and
predicted data [13].
As a general result the models prove to be good
approximations of the wear volume or rate.
Papers [18] and [13] present other related results
of the same research group.
A similar problem is solved by Gimondo and
co-authors [14]. The prediction of the linear wear
(mm) is performed by a MLP network, the considered
inputs being: load, velocity, relative humidity and
sliding distance.
The papers [33], [34], [39] proposed by another
group of authors, deal with the same problem of the
wear volume prediction, but in fretting experiments.
The main changes concern the dimension of the
network, including the number of the relevant inputs
(fig. 4). Two interesting aspects presented by Zhang
and co-authors [39] have to be mentioned: the friction
coefficient is no longer considered as fixed input with
previously known value, but as a variable, which has
Fig. 4. Construction of ANN [39].

TRIBOLOGY
38
to be modeled by the network. The authors consider
that studying the relevance of various parameters of
the tribological behavior of the system will help to
understand the relation between tribological
proprieties and material parameters in fibber
composites. This is a challenging start to using NN
for knowledge extraction.
The authors of the paper [21] make use of a MLP
to model the tool wear in a face milling process. The
input vector contains forces, feed rate, eccentricity
and the work piece geometry. The output is the
average flank wear. Different number of neurons in
the hidden layer is tested but the results seem to be
very close one to each other (very small error). A
specific element for the cutting process is that the
cutting force itself evolves along with the tool wear.
Not only the cutting force influences the tool wear but
also an inverse influence exists.
The influence of the cutting force on the tool is
separately evaluated outside the network. In order to
validate the neural model, its output is compared to
previously identified analytical models; the neural
model proves to be better.
Das and co-authors [9] apply the same type of
ANN to model the turning process. The input vector
contains three forces, cutting velocity and feed
(mm/rev). The output is the average flank wear of the
cutting tool. The model is closed to the real process,
excepting for large feed rates, when the error
increases due to the unstable associated phenomena.
Bilgesu and co-authors [6] present a similar
problem, the prediction of tool wear and life, under
varying operating conditions. The tool is a three-cone
bit for the drilling process in oil industry and the input
variables are: weight on bit, rotational speed, pump
rate, formation hardness, bit type and torque.
In papers [2, 3], Ao and co-authors used a MLP
to predict some parameters of the surface topography,
after a specified wear time: RMS, skewness and
kurtosis. These parameters further specify the
statistical properties of the worn surface, which is
subject of a surface reconstruction.
Benardos and Vosniakos [4] used a MLP
network to predict the roughness in CNC face milling.
The considered variables were: the depth of cut, the
feed rate per tooth, the cutting speed, the engagement
and wear of the cutting tool, the use of cutting fluid
and the three components of the cutting force. The
same method is applied by Őzel and Karpat [25, 26]
for roughness and tool wear prediction.
Supplementary, the authors compare the neural
network prediction ability to that of a regression-
based model. Another paper that compares the
prediction ability of neural networks and regression
models for surface roughness is [11].
Mainsah and Ndumu [23] considered a less used
source of information: the 3D-roughness
measurement. A MLP network processed this
information, in order to discriminate the different
surface types produced by manufacturing processes
(honing, shot-blasting, grounding). Moreover, the
output of their network allows the recognition of the
worn surfaces.
The paper of Chien and Chou [8] deals also with
the roughness prediction. The inputs to the MLP neu-
ral network are cutting speed, feed and depth of cut,
whereas the outputs are final surface roughness of the
workpiece, the cutting force and the tool life. This NN
provides a reliable model of the manufacturing proc-
ess. It should be noticed the use of the model in an op-
timisation procedure, which chooses the suitable cut-
ting parameters, to obtain an optimum of the cutting
regime, under constraints related to the final rough-
ness.
Finally, a neural model application that does not
imply prediction but space modeling is presented by
Karkoub and Elkamel [19]. The studied process is the
pressure distribution in a rectangular gas bearing. The
inputs of the network are: a dimensionless normalized
distance (along the width of the bearing), the supply
pressure, the gap width between the two plates and a
form factor of the plates. The first two variables are
more relevant for the pressure value (correlation coef-
ficient 0.8) than the last two (correlation coefficient
0.05). The same network models the load-carrying
capacity of the gas bearing, more accurately than the
theoretical model derived by Kassab from the Navier-
Stokes model. This is due to the highly nonlinearity of
the process, difficult to catch in a simple analytical
model.
3.2. On-line models
The papers discussed above aim at off-line pre-
dicting of different parameters, describing the manu-
facturing processes or wear processes.
On the contrary, the roughness prediction and
manufacturing quality evaluation based on acoustical
emission (AE) and vibrations represent a distinct class
of applications. In these applications, some on-line
measured data contribute to monitoring and diagnosis
of the processes.
For instance, Risbood, Dixit and Sahasrabudhe
[27] included in their prediction system designed for
turning processes, the feedback provided by the radial
component of cutting force and the acceleration of ra-
dial vibration of the tool holder. The NN learns to
predict on-line the dimensional deviation, then it al-
lows the proposed system to control the surface finish.
Chen and Savage [7] propose a similar method,
that evaluates in real-time the roughness (Ra
parameter) in milling processes, based on vibration
signals (the sensor is an accelerometer). The predicted
variable is Ra, under multiple cutting conditions.
Moreover, the NN is able to classify the surface
roughness, in fuzzy terms, such as a human operator
should perform. Other applications using the AE data
for on-line tool monitoring are reviewed by Li [20].
Yao and co-authors [38] present a new hybrid
model for tool condition monitoring and optimal tool

TRIBOLOGY
39
management in end milling operation. To facilitate
effective tool replacement at a proper state or time,
the model includes a wavelet fuzzy NEURAL
NETWORK, processing the AE signal, and a fuzzy
classifier of tool wear state, based on the detected
cutting parameters supported by the cutting database.
A method for monitoring the grinding processes
is presented by Liu and co-authors [22]. On-line
detection and prediction of grinding quality is based
on a wavelet NN and fuzzy back propagation learning
algorithm.
More about indirect and on-line tool wear
monitoring in turning with artificial neural networks
is presented by Bernhard [5].
Tansel and co-authors [30, 31] studied the
estimation of the tool wear in micro machining. Paper
[30] concerns analytical estimations of the cutting
forces. A very interesting application, based on the
neural model, is developed in the paper [31]. The
authors propose a periodic inspection device, to
evaluate the tool condition, in order to preserve the
quality of the manufacturing. The inspection is
executed as it follows: at the specified sampling
periods, the tool moves from the workpiece to a test
piece and cuts a slot, then comes back to the
workpiece (fig. 5). During the test, the cutting forces
are measured and they become input data to the neural
model previously learned. The output of the network
(Neural-Network-based Periodic Tool Inspector-
N2
PTI) is the estimation of the tool wear. It permits to
decide when to bring a new tool, being in a better
condition. This is a typical diagnose application.
Fig. 5. Operation of the N2
PTI [31].
An original contribution of the authors is the
comparison between more variants of network
implementation. Concerning the structure, they used
the classical MLP, with backpropagation training rule,
as well as a probabilistic neural network (PNN). This
kind of network (described in paper [31]) has four
layers: input, pattern, summation and output. One
neuron is assigned to the pattern layer for each
training case. After training, pattern and summation
layer neurons create the output. For each class there is
one summation neuron. Training is very fast;
however, the size of the network depends on the
number of training cases. If there are many training
cases, a large network will be established. In the paper
[31] the basic version of PNN is used, having a single
scaling parameter. The authors compare the
advantages of the MLP to the PNN. They establish,
based on experimental data, that backpropagation
learning is 10 times longer than the training of PNN,
but the average estimation error is 2 to 5 times better.
Another contribution of the authors is the comparison
between three encoding methods, that is the selection
of the parameters fed to the network input. The
concerned methods are:
- force-variation-based encoding (the inputs are the
differences between the maximum and minimum
values of the forces, in each period);
- segmental-average-based encoding (10 samples
per period are presented to the network, for each
cutting force);
- wavelet-transformation-based encoding (16
norma-lized parameters, provided by the wavelet
transforma-tion, are presented to the network).
The study of the authors found the wavelet
transformation as better estimating the forces and the
wear, but requiring more time and costly resources.
At the other end, the force-variation-based encoding
is simply to implement, but provides the worst
estimate.
Two other papers deal with neural models,
having to predict variables similar to wear. Paper [28]
presents an application concerning the degradation of
the physical properties of the lubrication oil, during its
lifetime. Properties such as viscosity, flash point,
water content, insoluble rating, are sampled during the
life of the oil. The MLP network is trained to
associate the age of the oil (as output) to its properties
(inputs). The result is the ability of the network to
predict the moment when each property of the oil
reaches a certain value, considered unacceptable
(called by the authors the "rejection time").
Jain and co-authors [16] present a neural
network approach to modeling the abrasive flow
machining process. This is a manufacturing process,
involving the same phenomena and variables as those
presented so far (for typical wear processes).
Consequently, the models are similar. The authors
trained a MLP to model the material removal rate and
the quality of the obtained surface (surface finish).
The four inputs are the flow speed, the abrasive
concentration, the abrasive mesh size and the number
of abrasive cycles. The integral of the material
removal rate provides the material loss, so the effects
may be predicted on-line, in order to control the
abrasive flow machining process. The idea of using
the model for on-line control is important for
fabrication processes, which need good prediction of
the critical parameters or variables. The same network
may be used to study the possible effects of the

TRIBOLOGY
40
process, prior to performing the real experiment. The
validation of the neural model proved that better
results are obtained, with respect to previously
identified multivariable regression models.
A remarkable work is presented in the Ph.D.
thesis of Flepp [12]. The purpose is to on-line
classifying the operating conditions of axial face seals
of large pumps (normal and abnormal). Three neural
networks were trained to process the acoustic
emission data and their classification abilities are
compared. The three neural network structures are:
MLP, Self Organizing Map and Time-Delay Neural
Network.
4. RECOGNITION OF FAULTS IN
BEARINGS, MACHINERY AND IN
MATERIAL STRUCTURE
Classification is the oldest application of the
neural networks. This ability is useful in a mechanical
environment, mainly to recognize the faults, as part of
a diagnose effort. As presented in chapter II, the
networks containing neurons with threshold activation
function in the output layer are suitable for
classification problems. However, linear or sigmoidal
activation functions are also suitable for this purpose,
if the output value is considered as a probability
density function, subject of a further threshold
discrimination. The training of the network may be
supervised or unsupervised (clustering). As a
consequence, both MLP and self-organizing maps
may be used in a classification problem.
Aiordachioaie and co-authors [1] present the
ability of a Kohonen network to classify 4 types of
bearing faults and their combinations. They are: outer
bearing race defect, inner bearing race defect, ball
defect and train defect. The source of information is
the vibration transducer. At each sampling moment, a
narrow horizon Fourier analysis provides the spectral
components, which are fed as inputs of the network.
This one recognizes the bearing vibration signatures.
The training is supervised, requiring short time.
Vyas and Satishkumar [35] introduce an
application of MLP to fault classification of a rotor-
bearing system. The inputs still are spectral
components of the vibration signals, whereas the
meaning of the outputs of the MLP is that of
probability density function of the fault signatures.
The outputs are used to discriminate four faults (rotor
with mass unbalance, rotor with bearing cap loose,
rotor with misalignment and play in spider coupling).
The network does not offer a quantification of the
fault, once it is recognized, but manages to recognize
even combinations of faults. The authors mention that
the influence of the training parameters (in a back-
propagation procedure) on the classification abilities
have to be investigated further.
Paper [29], published by Subrahmanyam and
Sujatha, contains a very complex and rigorous
analysis of the classification of faults, localized in ball
bearings. It uses both MLP and self-organizing
network (ART2) to recognize two main faults. The
process information, provided by piezo-electric
accelerometers, is subject of the Fourier analysis,
which computes the spectrum components. These are
then compressed in 8 significant descriptors, fed as
inputs of the network. The meaning of the output,
normalized between 0.1 and 0.9, is that of a scalar
description of the possible defects of the bearing. The
value 0.1 corresponds to a normal bearing, 0.6 to a
ball defect and 0.9 to an outer race defect. The MLP is
trained in a supervised procedure, whereas the ART
network performs a clustering process. Both networks
performed a 100% reliable recognition of the defect
bearings (on the presented data sets). MLP
distinguished the possible states of the defect
bearings, for diagnose purposes, with a rate of success
of 95%. The ART2 network was less accurate in
recognizing different defects, but it was 100 times
faster in training.
A different approach of a diagnosis problem is
presented by Dey and Stori [10]. The purpose is the
classification of the root causes of the variations of
the sequential machining operations and to provide a
probabilistic confidence level of the diagnosis. The
information is provided by multiple sensors and
processed by a Bayesian neural network. This is a
different kind of network, which evaluates conditional
probabilities and whose structure is that of an acyclic
graph.
The last four papers have a common element:
they deal with the classification of the wear
conditions. Umeda and co-authors [32] indicate a
MLP network as a proper classifier of the sliding
conditions, in a pin-on-disk sliding experiment. The
information considered relevant is the shape of the
wear particles, studied at the microscope and
compressed in 4 descriptors that are the network
inputs. The 5 outputs of the network are normalized
(0 to 1). Three of them represent the type of lubricant,
whereas the remaining two represent the low- or high-
load conditions of the process. After a supervised
training, the network is able to recognize the
mentioned sliding conditions. In order to make use of
image information (copying the human abilities, of
image processing and pattern recognition), the same
paper proposes a self-organizing network to classify
the surface images. The inputs are two textural
parameters, extracted from the image of the wear
particles (rather distinct to the worn surfaces than
similar). After the unsupervised training, a Kohonen
network is able to recognize the above mentioned
sliding conditions.
Paper [24], published by Myshkin and co-authors,
also deals with the wear debris classification, in order
to detect the phase of the wear process of the
lubricated sliding surfaces. The shape of the particles,
studied by microscopy, is described by Fourier
descriptors, compressed in 16 histogram parameters,

TRIBOLOGY
41
fed as inputs to the MLP network. The single output
of the network is able to point to the wear stage
(initial or final), by classifying the shape of the wear
debris.
In papers [36] and [37], Xu and co-authors present
MLP networks that classify wear particles, in terms of
wear mode. The authors created powerful software for
managing the classification possibilities of the
networks. The inputs are Fourier descriptors in one
example, to classify the shape of the particles, or
elements of the co-occurrence matrix of the gray
image (the texture), in other example, to discriminate
the worn surfaces in "smooth" and "rough". Other
examples of MLP networks that classify wear
particles, based on image analysis and Fourier
descriptors, appear in Hundal’s paper [15].
5. CONCLUSIONS
The purpose of this study was to assess the
potential of using neural networks to predict the
performance and life of mechanical systems. The
mentioned papers, as well as many others, not listed
here, succeeded to model tribological processes,
making use of ANN. In general, their conclusions
indicate ANN as a good modeling method, due to the
learning, generalization and nonlinear behavior
properties.
The main functions performed by the ANNs are
prediction (model) and classification. The purpose of
the prediction may be the diagnosis (prediction of the
lifetime), accelerated life-time testing, on-line control
of manufacturing processes that involve wear and
prediction of the main properties of the mechanical
systems, during the conceptual design stage. The
classification is useful for diagnose purposes:
recognition of the conditions of operation and
recognition of the faults. Usually, the result will have
a cost- and timesaving effect. Different technologies,
variables and sources of information are subject of the
modeling process. Some of the ANN inputs require
more previous processing, like Fourier analysis,
filtering and image processing.
The most popular ANNs are MLP - for
prediction and classification, and self-organizing
maps - Kohonen, ART, for classification. Their
properties and the diversity of the difficult mechanical
processes suggest that new applications of the ANN
in tribology will appear in the near future.
REFERENCES
1. Aiordachioaie D., R., Teodorescu R., Puscasu G., 1995, Fault
Detection in Electrical Machines with Neural Networks in Proc.
ELECTROMOTION'95 Symposium, Cluj-Napoca, Romania.
2. Ao Y., Wang Q. J., Penny C., 2001, Simulating the Worn Sur-
face in a Wear Process in Proc. World Tribology Congress, Vienna,
2001 (paper on CD).
3. Ao Y., Wang Q.J., Chen P., 2002, Simulating the worn surface
in a wear process, Wear, vol. 252 (1-2), p. 37-47.
4. Benardos P.G. and Vosniakos G.C., 2002, Prediction of sur-
face roughness in CNC face milling using neural networks and Ta-
guchi's design of experiments, Robotics and Computer-Integr.
Manuf., vol. 18 (5-6), p. 343-354.
5. Bernhard S., 2002, On-line and indirect tool wear monitoring in
turning with artificial neural networks: A review of more than a
decade of research, Mech. Syst. and Signal Processing, vol. 16 (4),
p.487-546.
6. Bilgesu H.I. et al., A New Approach to Predict Bit Life Based
on Tooth or Bearing Failures, Proc. SPE Eastern Reg. Tech. Conf.
and Exhibition, Pittsburgh, P.A., p. 253-257.
7. Chen J.C. and Savage M., 2001, A fuzzy-net-based multilevel
in-process surface roughness recognition system in milling opera-
tions, Int. J. of Adv. Manuf. Tech., vol.17 (9), p.670-676.
8. Chien W.-T. and Chou C.-Y., 2001, The predictive model for
machinability of 304 stainless steel, J. Of Mater. Processing Tech.,
vol. 118 (1-3), p. 442-447.
9. Das, S. et al., 1996, Evaluation of wear of turning carbide inserts
using neural networks, Int. J. Mach. Tools Manufact., vol. 36, no. 7,
p. 789-797.
10. Dey, S., Stori, J.A., 2005*, A Bayesian network approach to
root cause diagnosis of process variation, Int. J. Mach. Tools &
Manufact., vol. 45, p. 75-91 (*available online 2 nov. 2004).
11. Feng Chang-Xue and Wang Xian-Feng, 2003, Surface
roughness predictive modeling: Neural networks versus regression,
IIE Transactions (Inst. of Ind. Eng.), 35 (1), p. 11-27.
12. Flepp B., 1999, Wear Diagnosis of Mechanical Seals with
Neural Network, Ph.D. Thesis, ETH, Zurich.
13. Fusaro, R.L., 1998, Feasibility of Using Neural Network Models to Ac-
celerate the Testing of Mechanical Systems, NASA Research Center.
http://www.lerc.nasa.gov/Other_Groups/RT1997/5000/5930fusaro.
htm (23.08.2004).
14. Gimondo, P. et al., 2001, Neural Networks for Predicting Tri-
bological Experimental Results in Proc. World Tribology Congress,
Vienna, 2001 (paper on CD).
15. Hundal H.S. et al., 1997, Particle shape characterization using
image analysis and neural networks, Powder Technology, vol. 91,
no. 3, p. 217-227.
16. Jain R.K., Jain V.K., Kalra P.K., 1999, Modelling of abra-
sive flow machining process: a neural network approach, Wear, vol.
231, no. 2, p. 242-248.
17. Jones S.P., Jansen R., Fusaro R.L., 1997, Preliminary inves-
tigation of neural network techniques to predict tribological proper-
ties, Tribology Transactions, vol. 40, no. 2, p. 312-320.
18. Jones S.P., Jansen R., Fusaro R.L., 1998, Tribological Applica-
tion of ANN, GRNN NASA. www.grc.nasa.gov/WWW/spacemech/
neural.html (23.08.2004).
19. Karkoub M., Elkamel A., 1997, Modelling pressure distribu-
tion in a rectangular gas bearing using neural networks, Tribol. Int.,
vol.30, nr. 2, p. 139-150.
20. Li X., 2002, A brief review: Acoustic emission method for tool
wear monitoring during turning, Int. J. Mach. Tools & Manufact.,
vol. 42 (2), p. 157-165.
21. Lin, S.C., R.J. Lin, 1996, Tool wear monitoring in face mill-
ing using force signals, Wear, vol. 198, no.1-2, pp. 136-142.
22. Liu G.J., Gong Y.D., Wang W.S., 2004, Study on On-Line
Monitoring Method for Grinding Quality, Key Eng. Materials, vol.
259-260, p. 287-292.
23. Mainsah E., Ndumu D.T., 1998, Neural Network Applica-
tions in Surface Topography, Int. J. Mach. Tools Manufact., vol. 38
(5-6), p. 591–598.
24. Myshkin N.K. et al., 1997, Classification of wear debris using
a neural network, Wear, vol. 203-204, p. 658-662.
25. Őzel T. and Karpat Y., 2003, Prediction of Surface Rough-
ness and Tool Wear in Finish Dry Hard Turning Using Back
Propagation Neural Network, Proc. 17-th Int. Conf. on Production
Research, Blacksburg, VA., paper on CD.
26. Őzel, T. and Karpat, Y., 2005, Predictive modeling of surface
roughness and tool wear in hard turning using regression and neural
networks, Int. J. Mach. Tools & Manufact., vol. 45, p. 467-479
(available online 2 nov. 2004).
27. Risbood K.A., Dixit U.S., Sahasrabudhe A.D., 2003, Predic-
tion of surface roughness and dimensional deviation by measuring

TRIBOLOGY
42
cutting forces and vibrations in turning process, J. of Mater. Proc-
essing Tech, vol.132 (1-3), p. 203-214.
28. Sinha A.N., Mukherjee P.S., De A., 2000, Assessment of use-
ful life of lubricants using artificial neural network, Ind. Lubric.
and Tribology, vol. 52, no. 3, p. 105-109.
29. Subrahmanyam M. and Sujatha C., 1997, Using neural net-
works for the diagnosis of localized defects in ball bearings, Tribol.
Int., vol. 30, no. 10, p. 739-752.
30. Tansel, I.N. et al., 2000, Tool wear estimation in micro-
machining. - Part I: tool usage-cutting force relationship, Int. J.
Mach. Tools Manufact., vol. 40, no. 4, p. 599-608.
31. Tansel, I.N. et al., 2000, Tool wear estimation in micro-
machining. - Part II: neural-network-based periodic inspector for
non-metals, Int. J. Mach. Tools Manufact., vol. 40, no. 4, p. 609-
620.
32. Umeda, J. Sugimura and Y. Yamamoto, 1998, Characteriza-
tion of wear particles and their relations with sliding conditions,
Wear, vol. 216, no. 2, p. 220-228.
33. Velten, K, R. Reinicke, and K. Friedrich, 2000, Wear vol-
ume prediction with artificial neural networks, Tribol. Int., vol. 33,
no. 10, p. 731-736.
34. Velten, K., R. Reinicke, and K. Friedrich, 2000, Neuronale
Netze zur Vorhersage und Analyse des Verschleissverhaltens po-
lymerer Gleitwerkstoffe, Mat.-wiss. u. Werkstofftech., vol. 31, p.
712-714.
35. Vyas, N.S. and D. Satishkumar, 2001, Artificial neural net-
work design for fault identification in a rotor-bearing system,
Mechanism and Machine Theory, vol. 36, no. 2, p.157-175.
36. Xu, K. and A.R. Luxmoore, 1997, An integrated system for
automatic wear particle analysis, Wear, vol. 208, no. 1-2, p. 184-
193.
37. Xu K., A.R. Luxmoore and F. Deravi, 1997, Comparison of
Shape Features for the Classification of Wear Particles, Eng. Ap-
plic. of Artificial Intelligence, vol. 10, no. 5, p.4 85-493.
38. Yao Y.X. et al., 2004, A Hybrid Model for Tool Condition
Monitoring and Optimal Tool Management, Material Science Fo-
rum, vol. 471-472, p. 865-870.
39. Zhang Z., et al., 2001, Wear Prediction of Polymer Compos-
ites using Artificial Neural Networks, in Proc. CMSE/1, p. 203-206.

A SURVEY OF ARTIFICIAL NEURAL NETWORKS APPLICATIONS IN WEAR AND MANUFACTURING PROCESSES

More Related Content

Similar to A SURVEY OF ARTIFICIAL NEURAL NETWORKS APPLICATIONS IN WEAR AND MANUFACTURING PROCESSES (20)

More from Monique Carr (20)

Recently uploaded (20)

A SURVEY OF ARTIFICIAL NEURAL NETWORKS APPLICATIONS IN WEAR AND MANUFACTURING PROCESSES