Tool Wear Forecast Using Dominant Feature Identication of Acoustic

Emissions
Chee Khiang Pang
1
, Jun-Hong Zhou
2,3
, Zhao-Wei Zhong
3
, and Frank L. Lewis
4
AbstractIdentication and online prediction of lifetime
of cutting tools using cheap sensors is crucial to reduce
production costs and down-time in industrial machines. In this
paper, we use the acoustic emission from an embedded sensor
for computation of features and prediction of tool wear. A
reduced feature subset which is optimal in both estimation
and clustering least square errors is then selected using a
new Dominant Feature Identication (DFI) algorithm to reduce
signal processing and number of sensors required. Tool wear is
then predicted using an ARMAX model based on the reduced
features. Our experimental results on a ball nose cutter in a
high speed milling machine show a reduction in 16.83% in mean
relative error when compared to other methods proposed in the
literature.
Index TermsARMAX Model, Least Square Error (LSE),
Principal Component Analysis (PCA), Principal Feature Analy-
sis (PFA), Singular Value Decomposition (SVD), Tool Condition
Monitoring (TCM).
I. INTRODUCTION
I
N an era of intensive competition, the new challenges
faced by industrial manufacturing processes include max-
imizing productivity, ensuring high product quality, and
reducing the production time while minimizing the produc-
tion cost simultaneously. One of the causes of delay in
manufacturing processes is machine down time or failure of
the machining tools. As such, the detection and prediction
of tool wear becomes a vital role in manufacturing [1].
Currently, many methods exist for Tool Condition Mon-
itoring (TCM) which rely on the relationship between tool
conditions and measurable signals such as Acoustic Emis-
sions (AE) for detecting tool conditions. AE have been
extensively studied because of its non-intrusiveness, ease of
operation, and fast dynamic response [2] [3]. The frequency
range of AE is higher than that of machine vibration and
environmental noise [4], and AE sensors are small and can be
installed easily. The sampling process also does not interrupt
the machining operation [5], and using AE signals is less
expensive than using force signals.
The main goal of feature subset selection is to reduce the
number of features used in classication without compromis-
This work was supported in part by Singapore MOE AcRF Tier 1
Grant R-263-000-564-133, ARO Grant ARO W91NF-05-1-0314, NSF Grant
ECCS-0801330, and Alignment Tool (Singapore) Pte. Ltd.
1
C. K. Pang is with Department of Electrical and Computer Engineering,
National University of Singapore, Singapore [email protected]
2
J. -H. Zhou is with A*STAR Singapore Institute of Manufacturing
Technology, Singapore [email protected]
3
J. -H. Zhou and Z. -W. Zhong are with School of Mechanical
and Aerospace Engineering, Nanyang Technological University, Singapore
[email protected]
4
F. L. Lewis is with Automation & Robotics Research Institute, The
University of Texas at Arlington, Fort Worth, TX, USA [email protected]
ing on accuracy. Commonly used data compression schemes
include Principal Component Analysis (PCA) [6], Principal
Feature Analysis (PFA) [7], and Dominant Feature Identica-
tion (DFI) [8], which have been applied to pattern recognition
to identify key features from the original data. Both PCA
and PFA require computationally intensive calculations of
eigenvectors (and their corresponding eigenvalues) based
on the correlation or covariance matrix to derive a linear
transformation matrix for feature selection to compress the
raw data into lower dimensions. DFI on the other hand, uses
an inner product matrix of a lower dimension, and reduces
the Least Squares Error (LSE) induced when selecting the
principal components and clustering to identify the dominant
features. Compared with PCA and PFA, DFI is numerically
efcient, and reduces the complexity of feature selection
greatly while reducing the LSE during dominant feature
selection and clustering automatically.
In this paper, we use a new Dominant Feature Identi-
cation (DFI) [8] along with measurement of AE signals
to predict tool wear in a ball nose cutter. We present a
software decision tool for selecting the dominant features
that are most essential in predicting time series of tool wear
in industrial milling machines using an online, real-time, and
indirect approach, with data from installed AE sensors. With
the performance-related dominant features extracted using
the DFI methodology, Autoregressive Moving-Average with
eXogenous inputs (ARMAX) modeling [9] is utilized to pre-
dict the behavior of the performance degradation of machines
or cutters. The proposed methods were tested using a ball
nose cuter in an industrial high speed milling machine, and
comparisons were made with another technique for feature
selection in [7] using force sensors with a dynamometer with
recursive least squares in [8].
II. PRINCIPAL COMPONENTS AND DOMINANT FEATURES
In this section, the key results of Principal Component
Analysis (PCA) [6], Principal Feature Analysis (PFA) [7],
and Dominant Feature Identication (DFI) [8] will be pre-
sented. The major differences and merits of the approaches
will also be discussed.
A. Principal Components Analysis (PCA)
The SVD of a linear transformation X where X R
mn
of rank n < m is
X = UV
T
(1)
with U R
mn
and V R
nn
, such that U
T
U =
V
T
V = I
n
with I
n
R
nn
being an identity matrix of
2010 IEEE International Conference on Control Applications
Part of 2010 IEEE Multi-Conference on Systems and Control
Yokohama, Japan, September 8-10, 2010
978-1-4244-5363-4/10/$26.00 2010 IEEE 1063
dimension n. R
nn
is a diagonal matrix whose elements
are corresponding singular values (principal gains) arranged
in descending order, i.e., with = diag(
1
,
2
, ,
n
)
and
1

2

n
> 0. X can be regarded as a
transformation from feature space R
n
into data space R
m
.
Note that X =

n
i=1

i
u
i
v
T
i
, where u
i
are the column
vectors of U and v
T
i
are the row vectors of V
T
, respectively.
Partition the SVD of X according to
X =

U
1
U
2

1
0
0
2

V
T
1
V
T
2

= U
1

1
V
T
1
+ U
2

2
V
T
2
(2)
with q < m as the desired number of singular values
to be retained in
1
for data space of dimension m. As
such,
2
contains the n q discarded singular values.
Obviously, U
1
R
mq
, U
2
R
m(nq)
,
1
R
qq
,
2

R
(nq)(nq)
, V
T
1
R
qn
, and V
T
2
R
(nq)n
.
Now the approximation

X to X is

X = U
1

1
V
T
1
. (3)
Then

X =

q
i=1

i
u
i
v
T
i
contains the columns u
i
of U
1
and
the rows v
T
i
of V
T
1
. The dominant singular values, i.e., the q
retained singular values, and their associated columns of U
are called principal components in PCA [6].
B. Principal Feature Analysis (PFA)
Principal Feature Analysis (PFA) is introduced in [7]. A
linear transformation Y is selected as
Y = U
T
1
X R
qn
(4)
so that x R
n
yields Y = Y x = U
T
1
Xx R
q
. Then
vectors z = Xx R
m
may be approximated in terms of
vectors Y R
q
according to
z = U
1
y. (5)
Combining with (2), we get an approximation of z R
m
in terms of vectors in the reduced space R
q
in (4) as
z = U
1
y = U
1
U
T
1
Xx
= (U
1
U
T
1
)(U
1

1
V
T
1
+ U
2

2
V
T
2
)x
= U
1

1
V
T
1
x =

Xx (6)
with

X depicted in (3).
PFA performs the data compression and dimension reduc-
tion to isolate the principal components based on the same
criteria as PCA, and proceeds to derive principal feature
components to arrive at a subset of the original feature vector
using the K-means clustering algorithm on columns of U
T
1
in (3) after singular value decomposition of the outer product
matrix XX
T
in [7]. This is computationally expensive in
general, since m n in most practical applications.
C. Dominant Feature Identication (DFI)
On the other hand, the Dominant Feature Identication
(DFI) methodology proposed in [8] uses the fact that
Y = U
T
1
X =
1
V
T
1
(7)
Therefore, DFI selects the best q columns of
1
V
T
1
using
K-means clustering methods for approximation using the
selected p dominant features within R
n
. This is in contrast
to [7].
Note that
X
T
X = V
2
V
T
(8)
with inner product matrix X
T
X an n n matrix with n
m. Therefore, V
1
and
1
can be computed using X
T
X. This
is far easier than using X
T
X as required in PCA and PFA.
More importantly, the procedure of rst selecting q princi-
pal components and then selecting p dominant features yields
the minimum overall approximation error in feature selection
and K-means clustering. In fact, the approximation error z
is given by
z = (X

X)x =

Xx = U
2

2
V
T
2
x (9)
and an estimate for z R
m
taking into account both q < n
retained singular values and p < n features is given by

z = U
1
y. (10)
This implies that the procedure of rst selecting q PCs
and then selecting p dominant features yields the minimum
overall approximation error given by
z

z = (z z) + ( z

z) = U
2

2
V
T
2
x + U
1

Y x (11)
where

Y = Y

Y and

Y is the approximation of Y by
replacing the vectors in Y with the vectors closest to the
centroids after K-means clustering. For note that
z
T
( z

z) = (U
2

2
V
T
2
x)U
1

Y x
= x
T
V
2

2
U
T
2
U
1

Y x = 0, (12)
i.e., the error in neglecting n q singular values and the
clustering error are orthogonal. This means that there is no
better way of selecting the dominant features other than the
DFI methodology proposed. Interested readers are referred
to [8] for more in-depth discussions, and the details of the
mathematical proofs are omitted here for brevity.
III. PREDICTION OF TOOL WEAR USING DOMININANT
FEATURES
In this section, we predict the tool wear using the dominant
features selected using DFI and conventional static models
or proposed dynamic Auto-Regressive Moving Average with
eXogenous (ARMAX) models.
Dene

f
1
f
2
f
n

as the matrix of features

computed from the signal measured by the AE sensor. Each
feature f
i
is a time signal of length m samples. Select p
dominant features using DFI as in the previous section. This
selects p columns in X as a reduced matrix

X R
mp
.
1064
A. Disccussion of Tool Wear Prediction Models and Cutting
Conditions
In current literature, tool wear is usually predicted with
standard static linear prediction models, e.g., in [10] [11].
The simplest and most commonly used is the linear Multiple
Regression Model (MRM) of the form
y(k) =
T
k

(k) (13)
where
T
k
R
p
is the vector containing the measured
dominant features at time k m (or a row vector in

X
at time k) and

(k) is the vector of unknown coefcients to

be regressed for time series forecast of tool wear y(k). The
coefcients in

(k) are estimated by conducting experiments

under one cutting condition. However in the real production,
cutting conditions (spindle speed, feed rate, and depth of
cut, etc.) are varied according to the production requirement.
The model that was built using one cutting condition may
not represent other cutting conditions.
To overcome these difculties, we propose to use dynamic
ARMAX model instead of the static MRM modeling in
our study for linear regression that depends on current
and past values of the features, as well as also on past
predicted values of tool wear. This gives the prediction model
more information about the existing feature and tool wear
conditions.
B. ARMAX Model Based on DFI
We now desire to predict the real tool wear y(k) at time k
in terms of the identied p dominant features using an
ARMAX model with Extended Least Squares (ELS). The
general ARMAX model has the following structure [9]
A(z
1
) y(k) = B(z
1
)u(k n
u
) + C(z
1
)(k) (14)
where y(k) is the predicted tool wear, u(k) =

u
1
u
2
u
p

T
is a column vector of measurements
from the p dominant features (or a row vector in

X at time k),
and n
u
is the input delay. z
1
is the unit backward shift
operator and (k) is the estimation error. A, B, and C are
polynomials of in ascending powers of delays as
A(z
1
) =1 + a
1
z
1
+ a
2
z
2
+ + a
na
z
na
B(z
1
) =

b
11
+ b
12
z
1
+ + b
1nb
z
nb1+1
b
21
+ b
22
z
1
+ + b
2nb
z
nb2+1
.
.
.
b
p1
+ b
p2
z
1
+ + b
pnb
z
nbp+1

T
C(z
1
) =1 + c
1
z
1
+ c
2
z
2
+ + c
nc
z
nc
(15)
where n
a
, n
pb
, and n
c
as the orders of A, B, and C,
respectively.
If there is no input delay, i.e., n
u
= 0, we can write the
ARMAX model from (14) in time series and the standard
Least Squares (LS) problem can now be reformulated into the
Extended LS (ELS). The combination of the DFI algorithm
with the usage of a dynamic ARMAX model using ELS is
called the new DFI methodology.
The coefcients of the ARMAX model are updated with
the real time captured sensor data and estimated using the
common Recursive LS algorithm. The model is constructed
dynamically with the latest cutting information to reect
the real cutting conditions more accurately. As such, the
problem of building different models under different cutting
conditions, or the inaccuracies when using one model for
different cutting conditions, can be solved by this approach.
IV. INDUSTRIAL TEST BED
An application related to milling machining tool condition
is selected for the experiment. Tool condition is an important
factor in the milling machining process. Tool wear and tool
failure may result in a loss in surface nish and dimensional
accuracy of the nished parts, and even possible damage to
the work piece and machine. An image of tool wear is shown
in Fig. 1.
Fig. 1. Flank wear at cutting edge.
A. Experimental Setup
In our experiment, we used a r oders TEC vertical milling
machine as our test bed, and a ball nose cutter was se-
lected for our testing. The cutting process was performed
by predened procedures. After each cutting process, tool
snapshots were taken to measure the amount of tool wear.
An LECIA MZ12.5 high performance stereomicroscope was
used to measure the tool wear of the cutting tool.
For our application, an 8152B211 Piezotron

AE sensor
from Kistler was used for tool condition monitoring in
the experiment. Due to its small form factor, it can be
easily mounted near the source of emission for optimal
measurement of the AEs. During the machining process, the
AE signal is measured by the integrated impedance converter
which measures the AE signal from 100 kHz to 900 kHz.
The experimental setup with both the dynamometer (force
sensor) and AE sensor deployed in a TEC vertical milling
machine is shown in Fig 2. A 6-mm ball nose tungsten
carbide cutter was chosen to machine a Titanium Ti6Al4V
work piece. The AE sensors position was chosen to be near
the cutting path in order to efciently capture the dynamic
changes of the cutting process. The mounting position does
not change cutters static and dynamic stiffness, and add
additional constraints on the milling process.
1065
Fig. 2. Experimental setup.
Our experiment was carried out with a new cutter until the
tool wear became too large, i.e, broken down and damaged.
The evolution of measured tool wear during the experimental
stage is shown in Fig. 3.
Fig. 3. Stages in evolution of tool wear.
In our study, the signals were rstly truncated to the de-
sired lengths to facilitate signal processing. Next, noise in the
raw signals was ltered by the wavelet de-noising method,
which consists of decomposition, identication of threshold
detail coefcient, and reconstruction. The processed noise-
free signal can then be reconstructed via the inverse wavelet
transform of the threshold wavelet coefcients. A third level
decomposition was also applied to the signals.
B. Selection of Dominant Features
The characteristics of the AE signal that can be used for
TCM have been studied by researchers in various aspects,
e.g., spectral, statistical, and time series analysis, etc., to
analyze tool wear. As such, these sixteen features from
the methodologies found to be effective in TCM in current
literature are summarized in Table I and form the scope of
the feature subset selection.
With the AE features and force features obtained, the DFI
methodology described earlier and in [8] are used to obtain
a feature subset which is minimum in overall approximation
and K-means clustering error.
TABLE I
FEATURES AND NOMENCLATURE
No Feature Notation
1 Skewness AEskew
2 Kurtosis AEkts
3 Crest-factor AEcrest
4 Peak AEmax
5 Total amplitude of AE AEa
6 Mean of RMS AErms
7 Average AE AEca
8 Standard deviation AEstd
9 Mean of band power AEmb
10 Standard deviation of band power AEstdb
11 Delta (change in signal) AEdlt
12 Absolute deviation AEad
13 Rind down count AEc
14 Rise Time AEr
15 Area under curve AEca
16 Duration AEd
V. EXPERIMENTAL RESULTS
In the experiment, 1250 time points of measured force and
AE sensor data were captured under the following machine
settings: spindle speed 7400 rpm, feed rate 474 mm/min, and
depth of cut 0.1 mm. During the data measurement phase,
the tool was removed from the chuck and the tool wear was
measured by hand periodically. Specically, the ank wear of
each individual tooth of the cutting tool was measured with
an LECIA MZ12.5 high performance stereomicroscope. This
yields the baseline actual tool wear plot shown in Figs. 58.
The measured force and AE data were detrended, i.e., the
mean value was subtracted and normalized. Based on these
data, the sixteen features in TablesI were computed as func-
tions of time. This yields two sets of n = 16 feature vectors
(each of which is a function of time and has m = 1250 data
points) and X in (1) that has n = 16 columns and m = 1250
rows. Next, the PFA and DFI of feature selection were
performed. The selected dominant features were then used
as the input to build non-dynamic and dynamic ARMAX
tool wear prediction models. The prediction accuracies with
different approaches were analyzed and the results are shown
in this section.
The resulting singular values are plotted in Fig. 4. It can
be seen that fourth and subsequent singular values for both
force and AE signals are very small as compared to the rst
three singular values. As such, it appears that retaining the
rst three singular values would be sufcient to capture the
relevant trend in tool wear prediction for this experiment.
However, it is not clear which of the sixteen features are
required for tool wear prediction. We therefore use the DFI
methodology to select the most important features.
A. Comparison of Standard Non-Dynamic Prediction Mod-
els with Dynamic ARMAX Model
In current literature, tool wear is usually predicted with
standard non-dynamic linear prediction models, e.gs., in [10]
[11]. The simplest and most commonly used is the linear
Multiple Regression Model (MRM) in (13). For non-intrusive
1066
Fig. 4. Plot of principal components versus singular values.
and online prediction, the commonly used RLS algorithm
was also used for our experiment.
MRMs using RLS were identied to predict the baseline
measured tool wear using all of the original sixteen force and
AE features as described in the previous section. A Mean
Relative Error (MRE) of 7.15% for all sixteen force features
and 10.53% for all sixteen AE features are observed as shown
in Fig. 5. A better prediction performance is obtained using
the force MRM when compared to the AE MRM. The MRE
of the AE MRM is above 10%, and is unacceptable in TCM
applications.
Fig. 5. RLS using all sixteen force and AE features
Using ELS techniques, dynamic ARMAX models in (14)
were identied to predict the baseline measured tool wear
using all sixteen original force and AE features. The ac-
tual measured tool wear and the predicted tool wear using
ARMAX models as functions of time are shown in Fig. 6.
Clearly, the predictions are good. An MRE of 1.13% using
force features and 3.22% using AE features are observed
in Fig. 6, and represent the best possible prediction of tool
wear using these two sets of sixteen features. Both force
and AE with dynamic ARMAX models using ELS perform
better than those using the same features with non-dynamic
MRM using RLS. We also observed that using AE with
ARMAX and ELS obtained better performance than using
force with MRM and RLS. This justies the proposed usage
of AE sensors in TCM, as effective tool wear prediction
can be realized with improvement in system identication
techniques.
Here, we consider a single cutting condition to demon-
Fig. 6. ARMAX models using all sixteen force and AE features.
TABLE II
COMPARISION OF MODEL ACCURACIES
Model type MRE(%)
ARMAX using ELS with DFI 7.19
MRM using RLS with DFI 17.52
MRM using RLS with PFA 24.02
strate the effectiveness of our combination of dominant fea-
ture identication and dynamic ARMAX modeling for tool
wear prediction for simplicity but without loss of generality.
B. Comparison of proposed ARMAX Model using ELS with
DFI, MRM using RLS with DFI, and MRM using RLS with
Principal Feature Analysis (PFA)
It is desired to use fewer features to predict the tool wear.
Next, we use the DFI method [8] and PFA method [7]
in current literature as the feature selection techniques to
reduce the number of features in the feature space and select
the important features. The selected features obtained using
both DFI and PFA methods are then used as inputs, and
the ARMAX models using ELS and MRMs using RLS
techniques are used to build the prediction models to forecast
tool wear. To showcase the effectiveness of our proposed new
DFI methodology using AE sensors and ARMAX models
with ELS, only AE features are used in this comparison for
simplicity but without loss of generality.
Here, four dominant features (namely kurtosis, count, area,
and duration) were selected using DFI, and four dominant
features (namely kurtosis, delta, count, and duration) were
selected by using PFA. Using ELS and RLS to compute
the ARMAX model and MRM, respectively, mathematical
models were built to predict tool wear as shown in Fig. 7.
The actual measured tool wear, tool wear prediction using
ARMAX model with ELS and DFI (new DFI), tool wear
prediction using MRM with RLS and DFI, and tool wear pre-
diction using MRM with RLS and PFA, with all predictions
using the best four features with DFI and PFA, are shown
in Fig. 7. The MREs of the different models are shown in
Table II. A reduction in 16.83% in MRE is observed when
the proposed new DFI methodology is used as compared to
other methods proposed in the literature.
1067
Fig. 7. Examples of tool wear prediction using four dominant AE features
and three principal components.
C. Comparison of Tool Wear Prediction Using AE Measure-
ments and Force Measurements
Next, we compare the tool wear prediction when using
force and AE sensors using four dominant features selected
by DFI only. An ARMAX model with ELS is built with the
four selected AE dominant features (new DFI), and an MRM
model is built with the four selected force dominant features.
The actual measured tool wear and prediction results from
the constructed models are shown in Fig. 8. An MRE
of 7.19% when using the four AE features and 13.56% when
using the four force features are observed. The ARMAX
model with four AE dominant features performs better than
MRM model with four force dominant features [8], and
further justies the use of AE sensors with DFI and dynamic
ARMAX models using ELS for effective TCM.
Fig. 8. Examples of prediction using four dominant features and three
principal components.
The feature selection of DFI was performed ofine in this
study. In the online monitoring, only the selected features
were extracted in order to save the computational effort and
also for fast data processing. The ARMAX model was then
updated dynamically online with the new feature points. This
approach is different from the normal tool-wear monitoring
system in which static tool reference models [10] [11] are
necessary. Since the nature of the cutting conditions is always
changing in the real production, building a generic static
model suitable for all cutting conditions is very challenging
and almost impossible. On the other hand, different static
tool-wear models are built for different cutting conditions
when a generic model is lacking which is very time con-
suming. The approach in this paper updates the dynamic
ARMAX model on-line with new data points makes it
suitable for different cutting conditions, as no static models
are needed. The dynamic ARMAX model that is updated
on-line also gives a more accurate tool wear prediction.
VI. CONCLUSION
In this paper, the Dominant Feature Identication (DFI)
methodology using Singular Value Decomposition (SVD) of
collected tool wear data from Acoustic Emission (AE) and
force sensors is used to construct models for prediction of
times series of deterioration of an industrial cutting tool.
Our experimental results using AE sensors show signicant
reduction in both Mean Squares Errors (MSEs) and Mean
Relative Error (MRE) when a dynamic ARMAX model with
Extended Least Squares (ELS) technique is employed, which
is promising in replacing force sensors and conventional
non-dynamic models for effective online Tool Condition
Monitoring (TCM) and tool wear prediction. Our future
works include improving signal-to-noise ratio of the AE
signals, as well as their corresponding signal processing and
spectral analysis.
ACKNOWLEDGEMENTS
The authors would like to thank B. S. Lim, X. Li, and S.
J. Pua from A*STAR SIMTech, Singapore, for their help in
the experiments.
REFERENCES
[1] C. L. Jiaa and D. A. Dornfeld, A Self-Organizing Approach to the
Prediction and Detection of Tool Wear, ISA Transactions, Vol. 37,
No. 4, pp. 239255, September 1998.
[2] X. Li and X. Yao, Multi-Scale Statistical Process Monitoring in
Machining, IEEE Trans. Ind. Electron., Vol. 52, No. 3, pp. 924927,
June 2005.
[3] J. Sun, G. S. Hong, M. Rahman, and Y. S. Wong, Identication
of Feature Set for Effective Tool Condition Monitoring by Acoustic
Emission Sensing, International Journal of Production Research, Vol.
42, No. 5, pp. 901918, March 2004.
[4] T. W. Liao , Q. Zou, L. Mann, and M. E. Zodhi, Online PCBN Tool
Failure Monitoring System Based on Acoustic Emission Signatures,
IEE Proceedings: Science, Measurement and Technology, Vol. 142,
No. 5, pp. 404410, September 1995.
[5] J. -S. Kim, M. -C. Kang, B. -J. Ryu, and Y. -K. Ji, Development of
an On-Line Tool-Life Monitoring System Using Acoustic Emission
Signals in Gear Shaping, International Journal of Machine Tools &
Manufacture, Vol. 39, No. 11, pp. 17611777, November 1999.
[6] I. T. Jolliffe, Principal Component Analysis, Springer-Verlag, New-
York, 1986.
[7] I. Cohen, Q. Tian, X. S. Zhou, and T. S. Huang, Feature Selection
Using Principal Feature Analysis, in Proc. of the 15th International
Conference on Information Processing, Rochester, NY, USA, Septem-
ber 22-25, 2002.
[8] J. -H. Zhou, C. K. Pang, F. L. Lewis, and Z. -W. Zhong, Intelligent
Diagnosis and Prognosis of Tool Wear Using Dominant Feature
Identication, IEEE Trans.Ind. Informat., Vol. 5, No. 4, pp. 454464,
November 2009.
[9] L. Ljung, System Identication-Theory For the User, 2
nd
Edition, PTR
Prentice Hall, Upper Saddle River, N.J., 1999.
[10] P. Palanisamy, I. Rajendran, and S. Shanmugasundaram, Prediction
of Tool Wear Using Regression and ANN Models in End-Milling
Operation, The International Journal of Advanced Manufacturing
Technology, Vol. 37, Nos. 12, pp. 2941, April 2008.
[11] Z. Chen and X. M. Zhang, Monitoring of Tool Wear Using Feature
Vector Selection and Linear Regression, Lecture Notes in Computer
Science, Advances in Natural Computation, Vol. 3611, pp. 16, 2005.
1068