SCIENCE CHINA

Technological Sciences
Science China Press and Springer-Verlag Berlin Heidelberg 2011 tech.scichina.com www.springerlink.com

*Corresponding author (email: [email protected], [email protected], [email protected])
RESEARCH PAPER
December 2011 Vol.54 No.12: 31193129
doi: 10.1007/s11431-011-4595-6
On-line chatter detection using servo motor current signal
in turning
LIU HongQi
1*
, CHEN QingHai
1*
, LI Bin
1,2*
, MAO XinYong
1
, MAO KuanMin
2

& PENG FangYu
1

1
National NC System Engineering Research Center, Huazhong University of Science &Technology, Wuhan 430074, China;
2
State Key Laboratory of Digital Manufacturing Equipment & Technology, Huazhong University of Science &Technology,
Wuhan 430074, China
Received August 24, 2011; accepted September 16, 2011; published online November 5, 2011

Chatter often poses limiting factors on the achievable productivity and is very harmful to machining processes. In order to
avoid effectively the harm of cutting chatter, a method of cutting state monitoring based on feed motor current signal is pro-
posed for chatter identification before it has been fully developed. A new data analysis technique, the empirical mode decom-
position (EMD), is used to decompose motor current signal into many intrinsic mode functions (IMF). Some IMFs energy and
kurtosis regularly change during the development of the chatter. These IMFs can reflect subtle mutations in current signal.
Therefore, the energy index and kurtosis index are used for chatter detection based on those IMFs. Acceleration signal of tool
as reference is used to compare with the results from current signal. A support vector machine (SVM) is designed for pattern
classification based on the feature vector constituted by energy index and kurtosis index. The intelligent chatter detection sys-
tem composed of the feature extraction and the SVM has an accuracy rate of above 95% for the identification of cutting state
after being trained by experimental data. The results show that it is feasible to monitor and predict the emergence of chatter
behavior in machining by using motor current signal.
chatter detection, current signal, empirical mode decomposition (EMD), support vector machine (SVM)

Citation: Liu H Q, Chen Q H, Li B, et al. On-line chatter detection using servo motor current signal in turning. Sci China Tech Sci, 2011, 54: 31193129,
doi: 10.1007/s11431-011-4595-6



1 Introduction
Chatter is a self-excited vibration that occurs during ma-
chining operations and becomes a common limitation to
productivity and part quality. Chatter occurrence also has
several negative effects, such as excessive noise, machine
tool damage, disproportionate tool wear, damage of the
tooling structure and the spindle bearing, and waste of ma-
terials and energy. For these reasons, chatter avoidance is a
topic of enormous interest for many researchers who have
focused on the solutions to the problem of chatter. A review
of the literature regarding the chatter avoidance finds that
there are two main types of strategies: using the lobbing
effect and changing the system behavior. The first strategy
aims to avoid chatter by situating the machining process in
the stable zone of the stable lobes diagram (SLD) [1, 2], and
the second strategy is to enlarge the stable zone of SLD by
expanding the stability frontier or simply changing the sys-
tem behavior [3, 4]. It is found that appropriate control
strategy may be designed to avoid the emergence of cutting
chatter by adopting controllable stiffness, which relies on

3120 Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12
the on-line detection of cutting condition. Therefore, on-line
chatter detection and identification are the base of the chat-
ter avoidance. Combined with tool path optimization [5, 6]
and tool positioning strategy [7, 8], chatter can be avoided,
the surface quality and the processing efficiency can be im-
proved effectively.
For chatter detection, the sensors mostly applied are dis-
placement, acceleration, microphones, AE and plate dyna-
mometers. In relation to the signal processing, several ana-
lytical techniques are mainly applied to feature extraction,
including time series modeling, spectral analysis [9] and
time-frequency analysis. In addition, several smart pattern
classifiers are introduced so far. Table 1 illustrates some
researches which have focused on chatter detection. The
table is organized to indicate the machining process, meas-
ured physical quantities, sensors setup, signal processing
techniques, chatter classification criteria and the reference.
According to Table 1 and lots of literature related to the
chatter detection, these mostly used sensors are difficult to
set up. Moreover, some sensors setup may affect the ma-
chining operations, and they are expensive. Devillez et al.
[10] detected tool vibration with eddy current sensors, but
they must design a customized tool holder first, that is too
inconvenient. Most of the chatter detection methods tabu-
lated in Table 1 can only be used in laboratory, and be dif-
ficult to use in practical production. Hence, in order that the
on-line chatter detection technique can be applied to manu-
facturing production, the selected physical quantities should
be easily measured and be effectively used to detect chatter.
Nowadays, along with the servo system performance
unceasing enhancement, the response speed and sensitivity
improve continuously. So drive motor current can reflect
the cutting state of machining processes. Many researches
have been carried out on cutting force [11] tool-wear and
tool-breakage detection by using current signal, but this
paper focuses on detecting chatter occurrence during ma-
chining operations by adopting drive motor current. The
advantage of chatter detection based on motor current is that
it is easy to acquire current signal, the sensors are cheap and
convenient to set up, and it can avoid the interference from
the cutting environment. So on-line chatter detection based
on current signal is conducive to extent of spreading and
application of this technique. Soliman and Ismail [15] used
ratio of the second moments of the high- and low-frequency
components of the spindle current to detect chatter in mill-
ing. This method might work well for low spindle speeds.
However, because spindle of miller has small inertia, it is
mainly suitable for milling.
Because spindle of the heavy duty NC vertical lathe has
large inertia, spindle motor current is insensitive to the
change of cutting state. However, the feed motor is smart
and sensitive to the change of cutting force and cutting state.
Accordingly, we can choose the feed motor current signal to
analyze and detect chatter.
This paper proposes a novel method for chatter identifi-
cation in the turning process. Feed motor current signal is
adopted for on-line detection. The feature vector of current
signal used for chatter identification is constructed on the
basis of empirical mode decomposition (EMD). In addition,
acceleration signal of tool is taken as reference and used to
compare with the results from current signal. The pattern
classification results in terms of current signal are compared
with acceleration signal. Subsequently, a support vector
machine is designed for pattern classification. After super-
vised learning of support vector machine (SVM), an excel-
lent performance is achieved for the chatter identification.
Figure 1 shows the structure diagram of chatter detection
system based on current signal.
Table 1 Summary of chatter detection research
Process Physical quantity Sensor Signal processing Chatter classification criteria Ref.
Turning tool vibrations eddy current sensors
spectral analysis of tool
micro-movements
number of excited frequencies
increases
[10]
Turning cutting forces
dynamometer
mounted on the slides
CER of the cutting force signals
define a threshold of CER,
CER below it
[12]
Grinding grinding force AE
force sensor under the tail stock,
AE sensor on the spindle
entropy and NCIR of
force signal and AEs RMS
define thresholds of NCIR and
entropy, high NCIR and low entropy
[13]
Turning tool vibrations
Accelerometer mounted on the
tailstock
ST of acceleration signal
damping index in a narrow
band drastically reduced
[14]
Milling spindle current current sensors
R-value of the spindle
current signal
define a threshold of R-value,
R-value above it
[15]
Milling
spindle vibrations,
cutting forces
accelerometers mounted on the
machine head, axial force sensor
wavelet decomposition
and calculation of statistical
parameters
chatter classification by using a
neural network and fuzzy logic
[16]
Turning cutting forces dynamometer under the tool turret PSD of cutting force
define thresholds,
larger ratio of area of PSD
[17]
Milling tool vibrations accelerometer near tool holder DWT, WTMM
chatter index based on WTMM
exceeds a threshold
[18]
Boring tool vibrations
accelerometer at free end of
boring bar
WT of acceleration signal chatter classification based on SVM [19]
Turning sound microphone placed on the compound PE of the audio signals a sharp drop in PE value [20]
CER, coarse-grained entropy rate; NCIR, normalized coarse-grained information rate; ST, s-transform; PSD, power spectral density; DWT, discrete
wavelet transform; WTMM, wavelet transform modulus maxima; WT, wavelet transform; PE, permutation entropy.
Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12 3121

Figure 1 Structure diagram of chatter detection system.
2 Feature extraction based on empirical mode
decomposition
2.1 Empirical mode decomposition (EMD)
The EMD method is proposed by Huang et al. [21] as a new
signal decomposition method for nonlinear and nonstation-
ary signal. It provides an alternative to traditional time-
frequency or time-scale analytical methods, such as the
short-time Fourier transform and wavelet decomposition.
The EMD decomposes a signal into a collection of oscilla-
tory modes, called intrinsic mode functions (IMF), which
represent fast to slow oscillations in the signal. Each IMF
can be viewed as a subband of a signal. Thus, the EMD can
be viewed as a subband signal decomposition. Traditional
signal analytical methods, such as Fourier and wavelet-
based methods, require some predefined basis functions to
represent a signal. The EMD relies on a priori known basis,
and it depends only on the data itself and is completely un-
supervised. Compared with traditional analysis methods, the
EMD is intuitive, direct, posterior and adaptive. Due to
these special properties, the EMD has been used to address
many science and engineering problems [22, 23].
The algorithm to create IMFs is established with the
definition of local maxima and minima of the time series of
the signal x(t). The local maxima x
max
(t) and minima x
min
(t)
are connected by a cubic spline line to produce respectively
upper envelope u(t) and lower envelope l(t). The upper and
lower envelopes should meet the following condition:
( ) ( ) ( ). l t x t u t < <
Here, mean is denoted as m
1
(t) and is given by

1
( ) ( )
( ) .
2
m
u t l t
t
+
= (1)
The difference between the signals x(t) and m
1
(t) is the first
component, h
1
(t), i.e.

1 1
( ) ( ) ( ). h t x t m t = (2)
Ideally, if h
1
(t) is an IMF, the following definitions must be
fulfilled [21, 24]:
(1) In the whole data set, the number of extrema and the
number of zero-crossing must either equal or differ at most
by one.
(2) At any point, the mean value of the envelope defined
by local maxima and the envelope defined by the local
minima is zero.
If h
1
(t) is not an IMF, then h
1
(t) is treated as the original
signal and repeats the sifting process up to k times, as it is
required to reduce the extracted signal to an IMF. Then, h
k
(t)
becomes an IMF, that is

1
( ) .
k k k
h t h m

= (3)
Then, it is designated as
1
( ),
k
c h t = the first IMF compo-
nent from the original signal x(t). Some stopping criteria are
used to terminate the sifting process. A commonly used
criterion is the standard deviation S:

2
1
2
1
1
( ) ( )
,
( )
T
k k
t
k
S
h t h t
h t

=

=
(

(
(

_
(4)
where T is the total length of the signal. A typical value for
S can be smaller than 0.3.
Separating c
1
from x(t), we could get

1 1
( ) , r x t c = (5)
where r
1
is viewed as the original data, and repeating the
above processes, the second IMF component c
2
of x(t) could
be got. Repeating the process as described above for n times,
n-IMFs of signal x(t) could be got. Then,

2 1 2
1
.
n n n
r r c
r r c

=
)
(6)
The whole procedure terminates when the residue r
n
is
either a constant, and monotonic slope, or a function with
only one extremum. Combining eqs. (5) and (6) yields the
EMD of the original signal,

1
( ) .
n
j n
j
x t c r
=
= +
_
(7)
The main characteristic of the EMD method is that it can
process nonlinear and nonstationary signal smoothly, and
retain data characteristics during decomposition. Conse-
quently, the EMD method is used to decompose motor cur-
rent signal into a series of different feature scale signals,
and get the essential message about cutting state.
2.2 Energy indexenergy of IMF
Chatter occurrence during machining operations can lead to
the change of cutting forces, and motor current of the CNC
can reflect this change brought by chatter. In the proposed
method for chatter detection in turning, feed motor current
3122 Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12
signal is decomposed into a series of different features scale
signals, the variations of IMFs energy represent the cutting
state in machining processes. So we can choose the energy
of these IMFs which are sensitive to chatter as the feature
values. Feature extraction based on EMD consists of the
following main steps:
(1) Sampling current signals by a frequency of 10240 Hz,
collecting a segment of data of 1 s for processing, and slid-
ing data length by 0.1 s.
(2) Calculation of the root mean square (RMS) of the
sampled feed motor current signal is made for converting
the AC current to the equivalent DC current and the RMS of
feed motor current signal can be calculated by

2 2 2
RMS U V W
1
[ ]
3
. I I I I = + + (8)
(3) Decomposing the sampled signal into a set of mono-
component signals, IMFs, as described in Section 2.1.
(4) Choosing the IMFs which are sensitive to chatter
(signal amplitude which is obviously different in chatter is
the most crucial).
(5) Calculating the energy of the selected IMFs by

2
( ) d .
i i
E c t t
+

=
|
(9)
(6) Summing the energy calculated in eq. (9) as the fea-
ture value by
.
i
E E =
_
(10)
2.3 Kurtosis indexkurtosis of IMF
Kurtosis is a dimensionless numerical statistic, it can be
used to reflect the distribution characteristics of the signal.
When there are instantaneous mutations in signals, it will
result in higher kurtosis value. Utilizing this characteristic,
kurtosis index is usually used for detecting the mutations or
transient information in signals. Chatter occurrence will
lead to the mutations of drive motor current, and kurtosis
index can be used to extract these instantaneous mutations.
Kurtosis coefficient K is a normalized fourth order center
distance, and can be represented as

| |
4
4
( ) ( )d
,
x t x p x x
K
o
+

=
|
(11)
where x(t) is the instantaneous amplitude of the signal, x
is the average value, p(x) is the probability density, and o is
the standard deviation.
For the discrete signal data, kurtosis coefficient K is:

4
1
1
,
N
i
i
t
K
x x
N o
=
=
| |
|
\ .
_
(12)
where x
i
is the instantaneous amplitude of the signal, x is
the average value, N is the total length of the signal, and o
t

is the standard deviation.
The kurtosis index K of the motor current signal can be
calculated through the same steps described in Section 2.2.
3 Support vector machine
SVM is an effective machine learning method for classifi-
cation problems, especially for small sample of training
vectors, and eventually results in better generalization per-
formance than most traditional methods [25]. SVM maps
input vectors into a higher dimensional feature space and
one can construct an optimal separating hyperplane in this
space. This basically involves solving a quadratic program-
ming problem, while gradient-based training methods for
neural network architectures on the other hand suffer from
the existence of many local minima. A brief review of some
basic work on SVM for binary classification problem is
given herein. For all the further details we can refer to Vap-
niks book [26].
First, we review the simplest case, i.e. the linearly sepa-
rable case. Given a training set of N data points
1
{ , } ,
N
i i i
y x
=

where
n
i
x R e is the ith input vector and { 1, 1}
i
y e + is
the ith output pattern. There exists a separating hyperplane,
the function of which is
0, x b + = (13)
where
n
R e is a normal vector, the bias b is a scale. Two
parallel hyperplanes can be represented as

1, if 1;
1, if 1,
i
i
x b y
x b y

+ > = +
+ < =
(14)
which is equivalent to
( ) 1, 1, 2, , .
i i
y x b i N + > = (15)
SVM tries to maximize the margin between two classes,
where the margin width between the two parallel hyper-
planes equals

2 . Therefore, in a linearly separable case,
one can find the optimal hyperplane by solving the follow-
ing quadratic optimization problem:

2 1
Min , s.t. ( ) 1.
2
i i
y x b + > (16)
By introducing Lagrange multipliers ( 1, 2, , )
i
i N o =
for the constraint, the primal problem becomes a task of
finding the saddle point of Lagrange. Thus, the dual
problem becomes

1 ,
1
1
Max ( ) ( ),
2
s.t. 0, 0.
n n
i i j i j i j
i i j
n
i i i
i
L y y x x
y
o o o o
o o
=
=
=
= >
_ _
_
(17)
Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12 3123
By applying Karush-Kukn-Tucker (KKT) conditions, the
following relationship holds:
[ ( ) 1] 0.
i i i
y x b o + = (18)
If o
i
>0, the corresponding data points are called support
vectors (SVs). Hence, the optimal solution for the normal
vector is given by

*
1
M
i i i
i
y x o
=
=
_
(19)
where M is the number of SVs. By choosing any SVs
( , ),
k k
x y one can obtain
* *
.
k k
b y x = After
* *
( , ) b is
determined, the discrimination function can be given by

*
1
( ) sgn ( )
M
i i i
i
f x y x x b o
=
| |
= +
|
\ .
_
(20)
where sgn(
.
) is the sign function.
In the linearly non-separable case, the support vector
method approach aims at constructing a classifier of the
form:

*
1
( ) sgn ( ) ,
N
i i i
i
f x y x x b o
=
| |
= +
|
\ .
_
(21)
where ( , ) is the kernel function. SVM can map input
vector
n
i
x R e into a higher dimensional feature space,
and can thus solve the linearly non-separable case based on
the kernel function. For ( , ) , typically one can have the
following choices:
Linear kernel ( , ) ;
i j i j
x y x x =

(22a)
Polynomial kernel of degree g

( , ) ( ) , 0;
g
i j i j
x y x x r = + > (22b)
Radial basis function

2
( , ) exp{ }, 0;
i j i j
x y x x = > (22c)
Sigmoid kernel
( , ) tanh( ), 0;
i j i j
x y x x r = + > (22d)
where r, and g are kernel parameters.
Unlike most of the traditional methods which implement
the empirical risk minimization principle, SVM implements
the structural risk minimization principle, which can even-
tually result in better generalization performance. Therefore,
in this work, a model of SVM classifier to recognize chatter
based on feature vector extracted from motor current signal
is proposed.
4 Experimental setup for chatter detection
The schematic diagram of the experimental setup used in
our study is depicted in Figure 2. The experiments are per-
formed on a NC Vertical Lathe (VL-850HR+P, Yu Shine
Precision Machine Co Ltd, Taiwan, China) with AC per-
manent magnet synchronous motors. Based on a tool bar
with a diameter of 32 mm and a length diameter ratio of 6,
chatter easily emerges due to the low stiffness of the tool
bar. In order to detect chatter, three current sensors
(CSNF161, Honey well, USA) are used to measure the
three-phases current of feed motor: I
U
, I
V
, I
W
. A piezoelec-
tric acceleration sensor (356A15, PCB Piezotronics, USA)
is placed at the free end of the tool bar to measure the vibra-
tion of tool. The material of workpiece is aluminum alloy,
and the shape of workpiece is round plate with major di-
ameter (540 mm), inner diameter (150 mm), and thickness
(25 mm). The clamping of workpiece and setup of accelera-
tion sensor are shown in Figure 3. A high speed signal ac-
quisition system (LMS Test. Lab, Signature Testing func-
tion module) is used to collect acceleration signal and cur-
rent signal synchronously. So it is favorable for comparative
analysis of acceleration signal and current signal.
A good chatter detection system must be adapted to the
variations of the cutting parameters. To evaluate the per-
formance of the proposed method, a series of experiments

Figure 2 Schematic diagram of the experimental setup.

Figure 3 Installation of the workpiece and sensor.
3124 Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12
have been conducted under different cutting conditions.
Table 2 lists the cutting conditions of this work.
5 Experimental results and discussion
5.1 Analysis of feed motor current signal
Figure 4 displays the feed motor current signals and tool vi-
bration signals collected at a linear velocity of 150 m min
1
.
The feed speed is 0.1 mm r
1
and the depth of cut is 0.5 mm.
It can be seen that the amplitude of acceleration signals in-
creases sharply during the period of unstable cutting.
Meanwhile, the feed motor current signals are seriously
distorted. The changes in the feed motor current signature
may occur due to chatter. On the purpose of effective rec-
ognition of chatter based on current signals, these changes
must be extracted. Figure 5 shows the chatter marks and
their distribution in workpiece. It can be seen that chatter
marks mainly appear in three areas of the workpiece surface
Table 2 Experimental condition
Tool Kentanium; K
r
(negative rake)= 95
Tool bar L (length)=200 mm; d
m
(diameter)=32 mm
Cutting condition
linear velocity 100, 120, 150, 200 m min
1
,
feed rate 0.1, 0.2, 0.3 mm r
1
,
depth of cut 0.1, 0.2, 0.3, 0.4, 0.5 mm
Turning method
constant linear velocity and variable spindle
speedwithout coolant

where stiffness is smaller than other local. And among them,
the regional one is badly damaged. From the distribution of
chatter marks, we can know that it is coincident with tool
acceleration signals during the period of unstable cutting.
The root mean square (RMS) of feed motor signals be-
fore and after chatter is shown in Figure 6. The figure shows
that during unstable cutting, RMS of the feed motor current
signals fluctuates as that of tool vibration shown in Figure 4.
And RMS of the feed motor current signals during the
steady-state cutting is stable as tool vibration signals during
the same period.
5.2 Chatter detection based on current signal
To test the performance of the new detection method, a lot
of real machining signals are analyzed. Figure 7 shows 14
components of the IMFs obtained by using the EMD
method for feed motor current signals during steady-state
cutting (Figure 7(a)) and in the case of chatter (Figure 7(b)).
The components are listed from high to low frequency. The
last one (res.) is the residue signal and it is a monotonic
function and represents the central tendency of the signal.
It can be seen that the amplitudes of IMFs vary from one
to another. They demonstrate that lower order IMFs capture
fast oscillation modes while higher order IMFs typically
represent slow oscillation modes. Lower order IMFs (e.g.
c1c4) contain most of the noise, and they can be easily
separated from the data. The higher order IMFs (e.g.
c13c14) have no physical sense and can be removed from


Figure 4 Feed-motor current and tool acceleration signal. (a) Steady-state cut; (b) unsteady-state cut.
Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12 3125

Figure 5 Chatter marks and their distribution in workpiece.
the analysis, because their amplitudes are much smaller than
other IMFs. Obviously, the first 7 modes are almost similar
in the two cutting states and thus they may represent the
information about the background signal. From c8 to c12,
there are significant differences. These modes fluctuate
violently and have higher amplitudes during the period of

Figure 6 RMS of feed-motor current signal for steady-state cut (a) and
chatter (b).


Figure 7 Decomposition by EMD of the feed-motor current signal (510 s) shown in Figure 6 for steady-state cut (a) and chatter (b).
3126 Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12
chatter. So these IMFs contain the main effective informa-
tion which is sensitive to the detection of chatter. In order to
see the differences between the two cases clearly, Figure 8
shows the instantaneous amplitudes of c9 components.
However, the order of the most informative IMFs depends
on the length of the data for analysis and the cutting condi-
tions. To select these IMFs automatically, different length
feed motor current signals under different cutting conditions
are analyzed thoroughly. It can be found that if the length of
the data for analysis is a constant value, the order of the
informative IMFs is in a certain range. The most informa-
tive IMFs that indicate cutting state can be selected auto-
matically by removing the residual and considering only the
components in this range. So the signal to disturbance ratio
can be greatly improved.
In order to calculate the energy and kurtosis of the in-
formative IMFs, we collect a segment of 1s data for proc-
essing with EMD method, and to slide data length by 0.1s.
Based on analysis of large quantities data of 1s, the most
informative IMFs in the range of c6c10 are found. Energy
index E

and kurtosis index K

of these informative IMFs can
be calculated through the method proposed in Sections 2.2
and 2.3. Figure 9 shows the energy index E and kurtosis
index K of feed motor current signal comparing with the
standard deviation D of acceleration signal. The current data
and acceleration data come from the cutting conditions:
linear velocity is 100 m min
1
, feed rate is 0.2 mm r
1
, depth
of cut is 0.5 mm. It can be seen clearly that chatter occurred
at about the 3rd second from the acceleration signal and its
standard deviation D. When chatter occurs, the energy index
E and kurtosis index K of current signal change abruptly.
During the period of steady-state cut, the energy index E

and kurtosis index K keep smooth and steady, and at a rela-
tively low level. But when chatter occurs, they increase
sharply, and become unsteady and variable. So the energy
and kurtosis of these informative IMFs can distinguish


Figure 8 Instantaneous amplitudes of c9 components.
cutting state clearly, and the energy index E

and kurtosis
index K

can be used to classify chatter. In order to prove the
reliability of these indexes, a lot of real machining signals
from different cutting conditions are analyzed. Figure 10
shows the results from other cutting conditions: linear ve-
locity is 150 m min
1
, feed rate is 0.1 mm r
1
, depth of cut is
0.5 mm. In this analysis the energy index E

and kurtosis
index K

are found to reflect the change of cutting state in
time and effectively.
Based on the EMD method, the energy index E

and kur-
tosis index K of feed motor current can construct a feature


Figure 9 (a) Acceleration signal; (b) energy index of current signal and
standard deviation of acceleration signal; (c) kurtosis index of current signal
and standard deviation of acceleration signal. (Linear velocity 100 m min
1
,
feed rate 0.2 mm r
1
, depth of cut 0.5 mm).

Figure 10 (a) Acceleration signal; (b) energy index of current signal and
standard deviation of acceleration signal; (c) kurtosis index of current
signal and standard deviation of acceleration signal. (Linear velocity 150 m
min
1
, feed rate 0.1 mm r
1
, depth of cut 0.5 mm).
Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12 3127
vector [E K] for chatter prediction. Although this method is
not based on dynamical model, this vector has a compre-
hensive response of the development in time domain and
energy spectrum. So it is very applicable to the chatter de-
tection and recognition.
6 Chatter recognition based on SVM
The main target of pattern classification is to identify the
category of current signal based on the feature vector [E K]
by SVM. In this way, we can detect the cutting state effec-
tively. A brief review of some basic work on SVM for bi-
nary classification problem is given in Section 3. In order to
solve the linearly non-separable problem, kernel function
was introduced. The idea of the kernel function is to enable
operations to be performed in the potentially high dimen-
sional feature space rather than in the input space. Hence
the inner product does not need to be evaluated in the fea-
ture space. This provides a way of addressing the curse of
dimensionality. In order to obtain a good classification re-
sult, choice of kernel function is also a very important seg-
ment. In general, radial basis function (RBF) is a first rea-
sonable choice. The RBF kernel nonlinearly maps samples
into a higher dimensional space. Unlike the linear kernel, it
can handle the case when the relation between class labels
and attributes is nonlinear, and the RBF kernel has less nu-
merical difficulties [27]. So a RBF kernel was selected for
classification. To classify the two patterns, a SVM was con-
structed using a MATLAB toolboxLIBSVM developed
by Hsu and Lin [28], a simple and effective support vector
machines tool for classification, regression and distribution
estimation. For parameter selection, LIBSVM provides a
simple tool to check a grid of parameters. Moreover, for
multi- classifica- tion, under a given parameter vector (C, ),
LIBSVM uses the one-against-one method to obtain the
cross validation (CV) accuracy. The parameter selection
suggests the same (C, ) for all k(k1)/2 decision functions.
LIBSVM is simple, effective, and its faster computational
time is a key requirement for on-line chatter identification.
According to tool vibration signals, the states in machin-
ing process can be classified into two categories: stable state
(Pattern 1) and chatter state (Pattern +1). The feed motor
current data are chosen to refer to the tool vibration signals
according to the following rules: when the vibration signals
have a long-term and low-amplitude smooth cutting, it is
considered to be in Pattern 1; when the vibration signals
have prodigious amplitude or the rhythm-vibration phe-
nomenon occurs, it is considered to be in Pattern +1.
Referring to the tool vibration signals and the rules men-
tioned above, 120 section feed motor current signals of two
categories were acquired from machining experiments un-
der different linear velocities, cutting depths and feed rates.
Each pattern had 60 signals. The energy index E

and kurto-
sis index K for each signal were calculated with EMD
method. The distribution of feature vector [E K] of each
pattern is shown in Figure 11. 40 signals of each pattern
were taken out from the 120 signals on the purpose of
training, and the remaining 20 signals of each pattern were
used for test. The training of the SVM was performed in
Matlab. The efficiency of the chatter recognition model
after training was tested by the input of the 80 training sig-
nals and the 40 test signals respectively.
In order to prove the reliability and applicability of the
chatter recognition model, experiments were carried out by
adopting the method of changing the cutting conditions,
such as fixture, workpiece and cutting parameters. 132 sec-
tion feed motor current signals were acquired from these
experiments to test and verify the chatter recognition model.
A total of 6 groups cutting parameters, each group cutting
parameter has 22 current signals. Figure 12 shows the dis-
tribution of all these feature vectors [E K] of each pattern.
Parts of the results are listed in Table 3. It demonstrates that


Figure 11 Distribution of feature vector [E K] of each pattern for training
and testing the chatter recognition model.

Figure 12 Distribution of feature vector [E K] of each pattern from new
experiments for verifying the chatter recognition model.
3128 Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12
Table 3 Pattern classification results of the chatter recognition model
Input
No.
E K
Output Target Result
Linear velocity
(m min
1
)
Feed rate
(mm r
1
)
Cutting depth
(mm)
1 32.7135 2.8227 1 1 correct 130 0.3 0.8
2 28.9705 2.8985 1 1 correct 130 0.3 0.8
3 26.5276 3.2352 1 1 correct 130 0.3 0.8
4 26.2586 2.8605 1 1 correct 130 0.3 0.8
5 290.629 3.1195 +1 +1 correct 130 0.3 0.8
6 284.968 4.1060 +1 +1 correct 130 0.3 0.8
7 234.683 2.9498 +1 +1 correct 130 0.3 0.8
8 380.846 3.1630 +1 +1 correct 130 0.3 0.8
9 433.817 4.0686 +1 +1 correct 130 0.3 0.8
10 20.8446 3.1821 1 1 correct 160 0.1 0.5
11 26.5663 2.9206 1 1 correct 160 0.1 0.5
12 29.0302 3.1026 1 1 correct 160 0.1 0.5
13 18.5127 2.9643 1 1 correct 160 0.1 0.5
14 38.8059 4.0185 +1 +1 correct 160 0.1 0.5
15 53.0238 3.7320 +1 +1 correct 160 0.1 0.5
16 79.1480 4.7464 +1 +1 correct 160 0.1 0.5
17 114.451 5.0388 +1 +1 correct 160 0.1 0.5
18 51.3077 4.2334 +1 +1 correct 160 0.1 0.5
19 22.6092 2.7947 1 1 correct 160 0.2 0.4
20 23.3091 2.8803 1 1 correct 160 0.2 0.4
21 27.4309 2.6593 1 1 correct 160 0.2 0.4
22 24.5725 2.9635 1 1 correct 160 0.2 0.4
23 18.2346 2.5681 1 1 correct 160 0.2 0.4
24 92.7392 3.6702 +1 +1 correct 160 0.2 0.4
25 197.955 3.3718 +1 +1 correct 160 0.2 0.4
26 236.297 3.7305 +1 +1 correct 160 0.2 0.4
27 255.160 4.2908 +1 +1 correct 160 0.2 0.4
28 218.906 4.7624 +1 +1 correct 160 0.2 0.4
29 20.7560 3.0038 1 1 correct 160 0.3 0.2
30 19.1263 2.9321 1 1 correct 160 0.3 0.2
31 16.5012 2.9882 1 1 correct 160 0.3 0.2
32 17.8797 2.9115 1 1 correct 160 0.3 0.2
33 15.5732 3.2661 1 1 correct 160 0.3 0.2
34 46.5545 3.1003 +1 +1 correct 160 0.3 0.2
35 63.0903 3.3704 +1 +1 correct 160 0.3 0.2
36 43.2717 4.2037 +1 +1 correct 160 0.3 0.2
37 49.5488 3.0752 +1 +1 correct 160 0.3 0.2
38 42.4244 3.0221 +1 +1 correct 160 0.3 0.2
39 24.4285 3.0627 1 1 correct 190 0.3 0.2
40 33.7392 2.9987 +1 1 in correct 190 0.3 0.2
41 24.6545 2.9719 1 1 correct 190 0.3 0.2
42 31.4310 3.4046 1 1 correct 190 0.3 0.2
43 23.9407 3.3470 1 1 correct 190 0.3 0.2
44 54.9162 3.4730 +1 +1 correct 190 0.3 0.2
45 51.4100 3.6469 +1 +1 correct 190 0.3 0.2
46 42.2602 3.5756 +1 +1 correct 190 0.3 0.2
47 46.6221 3.3755 +1 +1 correct 190 0.3 0.2
48 45.9620 4.0435 +1 +1 correct 190 0.3 0.2
49 25.9378 2.5628 1 1 correct 220 0.3 0.2
50 29.9697 2.9401 1 1 correct 220 0.3 0.2
51 22.2719 2.9121 1 1 correct 220 0.3 0.2
52 23.1481 3.1567 1 1 correct 220 0.3 0.2
53 34.0250 3.1234 +1 1 incorrect 220 0.3 0.2
54 246.890 3.7628 +1 +1 correct 220 0.3 0.2
55 210.485 5.0534 +1 +1 correct 220 0.3 0.2
56 256.991 7.0680 +1 +1 correct 220 0.3 0.2
57 293.460 5.7599 +1 +1 correct 220 0.3 0.2
58 229.780 9.0551 +1 +1 correct 220 0.3 0.2
Liu H Q, et al. Sci China Tech Sci December (2011) Vol.54 No.12 3129
most samples have been recognized correctly with an accu-
racy rate of above 95%. It can be seen that chatter recogni-
tion model based on the EMD method and SVM has an ex-
cellent performance for chatter premonition identification,
which is robust for different cutting conditions.
7 Conclusion and future work
In this paper, the application of energy index and kurtosis
index for the detection of onset of chatter in turning using
feed motor current signal was verified. Moreover, an intel-
ligent chatter recognition system based on EMD and SVM
was investigated. Experiments were conducted under many
different cutting conditions. A feature vector [E K] was
constituted by

the energy index E

and kurtosis index K

which were extracted from feed motor current signal based
on the EMD method. This vector can effectively reflect the
change of cutting state in time, and the classification results
by LIBSVM indicate that it is suitable for chatter premoni-
tion recognition. The system with combination of the
EMDs feature extraction capability and the SVMs pattern
classification capability has an accuracy rate of above 95%
for cutting states recognition. Additionally, energy and kur-
tosis analysis of feed motor current signal is a fairly low-
cost, non-contact, and non-destructive technique which en-
hances its suitability for on-line detection of chatter without
disturbing the machining process. For these advantages, this
system for chatter detection can be applied to real-time
turning process with the help of suitable chatter suppression
methods and control mechanism.
For future work, although the feature value based on
EMD of feed motor current signal mentioned above has a
good performance, type and parameters of servo system
may affect the detection results. It is still a challenge of how
to ascertain these effects and eliminate them. Therefore, it is
suggested that the following research is to investigate ap-
plication of this method for different types of servo system
and find out methods to eliminate effects. In addition, the
optimum feature vector will be another interesting work in
future.
This work was supported by the Major State Basic Research Develop-
ment of China (Grant No. 2011CB706803), National Natural Science
Foundation of China (Grant No. 50875098), and Important National
Science & Technology Specific Projects of China (Grant No.
2009ZX04014-024).
1 Tang A J, Liu Z Q. Effect of helix angle and normal rake angle on
stability in end milling. Adv Mater Res, 2009, 69-70: 394398
2 Tang A J, Liu Z Q. Three-dimensional stability lobe and maximum
material removal rate in end milling of thin-walled plate. Intl J Adv
Manuf Technol, 2009, 43 (1-2): 3339
3 Budak E. An analytical design method for milling cutters with non-
constant pitch to increase stability, Part 2: Application. Transactions
of the ASME. J Manuf Sci Eng, 2003, 125(1): 3538
4 Semercigil S E, Chen L A. Preliminary computations for chatter con-
trol in end milling. J Sound Vib, 2002, 249 (3): 622633
5 Han D, Zhu L M. Global optimization of tool path for five-axis flank
milling with a cylindrical cutter. Sci China Ser E-Tech Sci, 2009, 53:
24492459
6 Bi Q Z, Wang Y H, Zhu L M, et al. Wholly smoothing cutter orienta-
tions for five-axis NC machining based on cutter contact point mesh.
Sci China Tech Sci, 2010, 53: 12941303
7 Zhu L M, Ding H, Xiong Y L. Third-order point contact approach for
five-axis sculptured surface machining using non-ball-end tools (I):
Third-order approximation of tool envelope surface. Sci China Tech
Sci, 2010, 53: 19041912
8 Zhu L M, Ding H, Xiong Y L. Third-order point contact approach for
five-axis sculptured surface machining using non-ball-end tools (II):
Tool positioning strategy. Sci China Tech Sci, 2010, 53: 21902197
9 Ding Y, Zhu L M, Zhang X J, et al. Milling stability analysis using
the spectral method. Sci China Tech Sci, 2011, 54: 31303136
10 Devillez A, Dudzinski D. Tool vibration detection with eddy current
sensors in machining process and computation of stability lobes using
fuzzy classifiers. Mech Syst Signal Proc, 2007, 21: 441456
11 Mao X Y, Liu H Q, Li B. Time-frequency analysis and detecting
method research on milling force token signal in spindle current sig-
nal. Sci China Ser E-Tech Sci, 2009, 52(10): 28102813
12 Gradisek J, Govekar E, Grabec I. Using coarse-grained entropy rate
to detect chatter in cutting. J Sound Vib, 1998, 241: 941952
13 Gradisek J, Baus A, Govekar E, et al. Automatic chatter detection in
grinding. Int J Mach Tools & Manuf, 2003, 43(14): 13971403
14 Tansel I N, Wang X, Chen P, et al. Transformations in machining.
Part 2: Evaluation of machining quality and detection of chatter in
turning by using s-transformation. Int J Mach Tools & Manuf, 2006,
46(1): 4350
15 Soliman E, Ismail F. Chatter detection by monitoring spindle drive
current. Int J Adv Manuf Technol, 1997, 13(1): 2734
16 Kuljanic E, Totis G, Sortino M. Development of an intelligent mul-
tisensor chatter detection system in milling. Mech Syst Signal Proc,
2009, 23(5): 17041718
17 Tangjitsitcharoen S. In-process monitoring and detection of chip
formation and chatter for CNC turning. J Mater Proc Technol, 2009,
209(10): 46824688
18 Wang L, Liang M. Chatter detection based on probability distribution
of wavelet modulus maxima. Rob Comp-Int Manuf, 2009, 25(6):
989998
19 Yao Z H, Mei D Q, Chen Z C. On-line chatter detection and identifi-
cation based on wavelet and support vector machine. J Mater Proc
Technol, 2010, 210(5): 713719
20 Nair U, Krishna B M, Namboothiri V N N, et al. Permutation entropy
based real-time chatter detection using audio signal in turning process.
Int J Adv Manuf Technol, 2010, 46(1-4): 6168
21 Huang N E, Shen Z, Long S R. The empirical mode decomposition
and the Hilbert spectrum for nonlinear and nonstationary time series
analysis. Proc R Soc London, 1998, 454: 903995
22 Nunes J C, Guyot S, Delechelle E. Texture analysis based on local
analysis of the bidimensional empirical mode decomposition. Mach
Vis Appl, 2005, 16(3): 177188
23 Bassiuny A M, Li X L. Flute breakage detection during end milling
using HilbertHuang transform and smoothed nonlinear energy op-
erator. Int J Mach Tools & Manuf, 2007, 47(6): 10111020
24 Huang N E, Shen Z, Long S R. A new view of nonlinear water waves:
the Hilbert spectrum. Ann Rev Fluid Mech, 1999, 31: 417457
25 Cao L J, Chua K S, Chong W K, et al. A comparison of PCA, KPCA
and ICA for dimensionality reduction in support vector machine.
Neurocomputing, 2003, 55(1-2): 321336
26 Vapnik V. The Nature of Statistical Learning Theory. New York:
Springer-Verlag, 1995
27 Vapnik V. The support vector method of function estimation. Pro-
ceedings of Neural Networks and Machine Learning, 1998, 239268
28 Hsu C W, Lin C J. A comparison of methods for multiclass support
vector machines. IEEE Trans on Neural Networks, 2002, 13(2): 415
425