Proceedings of ICMPM 2005, 12-14 December 200

Chatter mode analysis in end-milling

Moon-Chul Yoon' and Do-Hun Chin

School of Mechanical Engineering
Pukyong National University, Busan, Korea
mcyoon@pknu.ac.kr

Graduate school of Mechanical Engineering

Pukyong National University, Busan, Korea
chindh@maill.pknu.ac.kr

Abstract: A chatter in end-milling process is a common process where the quality ofthesurface after end-millingis
often critical issue. For the investigation of the chatter modes, the power spectrum of the parametric time series
method was originally studied and analyzed by considering cuting dynamics of end-milling that has several mixed
conditions of different revolution. In this study, the paper describes a methodology for developing several time
series such as AR forward-backhward, burg, least square, Yule Walker, geometric lattice, instrumental variable),
ARX (least square, instrumenial variable), ARMAX, ARMA, Box Jenkins, Output Error, to estimate the chatter
mode using their spectral analysis, and the results were compared with one another. As a result, it was proven that
all time series methods can be used for chatter mode estimation and among them, the ARX, ARMAX and
instrumental variable methods (iv4) are a little bit more desirable and reliable one than the other algorithms for the
calculation of the chatter mode in the end-milling operation. Also, among the three cutting forces, the cuting force
F: has more powerful characteristics of chatter than the cutting forces Fx and Fy in the sense that there is no
shifted or pseudo mode in the estimated power spectra of the cuttingforce.

Keywords: ARMA, FFT, Power spectrum, Time series, Tool dynamometer

1. Introduction

Increasing interest and its results on chatter mechanics have evolved over the past decades. Dilley et al. showed that
the effect of chisel edge on chatter frequency using FFT in drilling [1]. Insperger et al. described the analytical and
experimental identification of the chatter frequencies in milling processes and the frequency diagram are atftached to
the stability charts of the mechanical models [2]. Kim et al also showed the chatter prediction method in this study
13]. With the advancement of a chatter monitoring, the conventional FFT (Fast Fourier Transform) method was
generally used. But it needs huge number of data for reliable monitoring of end-milling force. Also it is difficult to
acquire a system identifying equation for later approach. To investigate the chatter mechanism by considering a
cutting dynamic between the end-mill and workpiece, it is reliable and essential to identify the chatter mechanism at
first for latter approach such as for calculating the chatter frequency and its damping ratio Among several
identification methods, the time series parametric modeling is a desirable method in the sense that it satisfies this
need. In this study, several time series algorithms will be analyzed and discussed with one another and they were
extended to the spectral analysis by comparing with the FFT spectral analysis. By using these time series the weak
chatter frequency in cutting mechanics is detected easily for the signal. It shows more precise and close chatter
mode that was calculated directly from the auto regressive parameter. The chatter modes calculated by the several
time series algorithms in an end-milling operation are well consistent with each other. But some algorithms have
some drawbacks. In this study, the chatter mode was estimated mathematically using with several algorithms Such
as AR (forward backward, burg, least square, Yule Walker, geometric lattice, instrumental variable algorithm),
ARMA, ARX (least square, instrumental variable), ARAMAX, Box Jenkins and Output Error [4-14]. The advantages
and limitations of each algorithm needs to be compared with one and another. So the behaviors of chatter modes are
also discussed. In general, the dominant chatter frequency(1X) appears less than 1500 Hz and its harmonics
(2x, 3X,) may be damped in higher frequency range and it is shown in this study. The exactmess of each
algorithm for the chatter mode is also discussed by comparing with conventional FFT. For the reliability of several
the forces in three directions are applied in this study. To acquire the cutting forces Fx, Fy and
algornthms, cutting
Fz experimental test is performed using tool dynamometer. In the model that has the moving average input, a new

560
 Proceedings of ICMPM 2005, 12-14 December 2005, India

noise input was used for the modeling input. Even if a noise input is used it generally causes no adverse effect inna
total FRF spectrum because it has a flat spectrum in the frequency domain.

2. Time Series Algorithms

2.1. ARMA and AR Model

The general formulation of an ARMA(n,m) model which has the relationship between input and output is considered
for the end-milling process. The process may be defined as the ARMA(n, m) (Auto Regressive Moving Average) as
follows.

(1 a, B +
a, B +... + a,B ")y(1)
= (1 + b, B + b, B +... + b, B " )e(t)
(1)
Where y(t) is defined as the output data and
a; is the autoregressive parameter, b; is moving average parameter and
the integersandm are the order of the
n
autoregressive and moving average respectively. Also B is the
backshift operator such as By(t)= y{t-1) and e(t) is assumed the white noise parameters
input. If the integer m equals zero, the
ARMA(n,m) model may be reduced to the AR(n) model. The AR(»)
model can be identified by using only
autoregressive parameters and can be written in the following way:

(1+a,B+a,B +.+a,B")yt)= et)

(2)
2.2. ARMAX and ARX Model

The ARMAX(7,m,l,nk) model is defined as follows: A(B) is called an autoregressive parameter and the B(B) and
CB) are moving average parameters.

(1+ a,B + a,B +...+ a,B")y(t)

=
(b, + b2B +
b,B +...+b,B)u (t -
nk
+(1 + c,B +c,B +..+ C,B ')e (t)
(3)
If the ARMAX(n,mlnk) model is reduced to a model in which the moving average part C(B) of ARMAX becomes a
unity by assuming that the past noise has no correlation with current one then the B(B) term disappears and the
following ARX(n,m, nk) model is obtained:

(1+ a ,B + a, B +...+ a,B")y()

=
(b, +b, B + b, B +..+
b,B"Jult- nk) + et) (4)
All other prediction method of parameters can be used in the same procedure for AR or ARMA methods.

2.3. BJ Model

If the end-milling process is modeled by assuming that the output has a different correlation with the input u(t) and
noise input e(t). Such a model can be defined as the Box Jenkins, BI (n,m,l,knk), model as follows:

y t ) - t b ,B+...+b,B") -
B") elt
,+B *.te,B")
nk) + G B t . t c ,
LB"-nk)
14 SB+..+ SB)
(+SB++ (l+d,B +.+ d, B*)
(5)

561
 Proceedings of ICMPM 2005, 12-14 December 2005, India

functions that can be represented using the back shift operator B.

Where B(B), CB), D(B) and F(B) are polynomial

2.4. OE Model

Jenkins model by assuming that there is no correlation between

If the part of C(B)/D(B) part equals unity in the Box
the past noise and current one, the output error, OE(n,1, nk),
model can be written as follows.

(b, +b,B++ b5(t nk) +e(0)

y)= (1+ B
-

++ fB')
6)
2.5. Nanural Mode and Spectrum Estimation

can be calculated in the

From the all üme series models defined above, the chatter frequency and its damping ratio
denominator part of the transfer function of these models. In each model, by breaking up the denominator into
calculated as the
factors, the factor terms such as (1-1,B)(1-4;B) may be obtained and the chatter mode can be
If the sampling period equals 7,, the chatter frequency and its damping ratio can be summarized as
following way.
follows. [4,12]
abs(In(4 ) (7)

S-cos(angle(In(4,)) (8)

The transfer function of each time series model can be summarized in direct form as follows.

HN(B) (+hB+bB +b,B°+..+b,B")

DB) (1+a,B+a,B* +..+a,B°) (9)

H(B) N(B) .0+6,B+b,8' +.+bB)

H(B) =
=

1+a,B+a,B*+... +a,B')
DB) (10)
And the power spectrum can be calculated using these equations. By substituting 8 i n t o Eq. (9) or(10),then
the transfer function may be composed of the real and imaginary part. By calculating the root of square sum of real
and imaginary part, the power spectrum of the model can be calculated.

3. Experiments and Diseussion

3.1. Experimetal Methods

To acquire the cutting force during the end-milling operation, the experiments were performed in a vertical
machining center (DOOMAC 40V). The cutting force signals Fr, Fy and Fz were acquired from the tool
dynamometer at the revolution conditions of 500 pm 1300 pm, axial depth of 4 mm and feedrate of 0.02
mm/rev. The end-mill used has a shape of pl5mm x 48.7mm (diameter x overhang) with four flutes. The acquired
cutting force was used for each modelling for spectrum analysis. In the analysis, the visual ct++ was used for data
arrangement for resampling. The tool dynamometer was used for acquiring cuting forces in three directions (i.e. Fx,
Fy and Fz). The type of tool dynamometer used for acquiring the cutting forces in three directions is the piezo ype
dynamometer (kistler : 9272B). The cutting conditions selected in this experiment are summarized in Table 1.

562
 Proceedings of ICMPM 2005, 12-14 December 2005, India

y
O )
500 1300
700 900 1100
rpm rpm rpm rpm rpm

******

(a) (b)
Fig.1 Configuration of workpiece for experiments
The workpiece was prepared by considering five steps as in Fig. 1 for causing chatter at the right fifth stage in the
workpiece. In each stage, the rotational specd changes from 500 rpm to 1300 rpm as shown in Table 1. The feedrate
of 0.02 mm/rev was selected in all stages from first stage(D) to fifth stage(5). In experiments, a chatter occur at
the fifth stage(5) around 20.5 sec. All data of the total force including all stages were used for time series
modelling.
Table 1 Cutting condition for experiments
Workpiece SM45C
Endmill 15p,4 fluted HSS
Feedrate 0.02mm/rev
Axial depth of cut 4mm
Radial depth of cut 7.5mm
Rotational speed 500rpm, 700rpm,900rpm,1100rpm,1300rpm
Overhang 48.7mm_
Cutting method Up-milling
3.2. Mode estimation

Cutting Force wignal F Culting Force Sipnal Fy CngForca Sgral Fz

1000 1000

19 68 20 25 65 Tms ec 19 68
Time sec)

(a) (c) (e)

00 O008
Tooth passing Frequency Tooth passing Frequency
Tooth passing
3H Frequency o He Chater tequenoy
o T Cucs tregue
hader roquency 47 Hz,60 H2
4H 60 H

H
57 Ke

0 C0
400 000 800 1000 20 10
Frequancy ( ) Frequency (H FquenY IR

(b) (d) (
Fig. 2 Measured cutting force in three direction Ex(a), Fy(c) and Fz(e) in all stages ranging O to and their
power spectra (b),(d) and () using FFT

563
 Proceedings of ICMPM 2005, 12-14 December 2005, India

Fig. 2 shows thecutting force in five stages ( with each cutting

the forces Fx, Fy and Fz. In these experiments,
sampling frequency was selected as 5 kHz and its Nyquist frequency for
possible to detect the chatter frequency using power spectrum of FFT within viewing
and analyzing is 2.5
kHz. It is
this Nyquist
the power spectrum for times series frequency. Also, to obtain
modeling, the order of an autoregressive and moving average part was selected
as an order of (10,5) (however n=5,
m=3, -10 and k-10 for BJ model) respectively. For the case of a small
the order in moving average, it revealed some value of
pseudo mode. So the order of a moving average parameteT is selected
as 5 for no pseudo or
shifting. The modes of a basic tooth passing frequency in end-milling (33Hz for 500 rpm,
47Hz for 700 rpm, 60 Hz for 900
rpm, 73 Hz for 1100 rpm, 87Hz for 1300 rpm) were included in the
spectrum of cutting forces Fx, Fy and Fz in each directions(Fig. 3(6), (a) and power
of the cutting force Fx for (f). Fig. 2 shows the power spectrum
Fig. 2(a). It is considered that the harmonic tooth passing
teeth of an end-mill in spectrum of FFT. The components appeared due to the
cutting force Fx shows chatter in the fifth
stage of the cutting force
signal. Its chatter frequency appears at 1X (about 620-740 Hz) and its
harmonic appears at 2X
algorithms respectively. So the cutting force Fx can be acquired(about
1800 Hz) according to several 1293 Hz
chatter analysis with a time series. In this and used for the
study, to analyze vibrating property of the cutting force, the
frequency was set to 5 kHz in this cutting force, Fx. In FFT analysis, the sampling
components decreases in the higher frequency range compared to that of the first
amplitude of following hamonic

power spectrum of the time series the global trend is similar, but some weak synchronous frequency. In the
mode is magnmified and can be seen
more clearly than FFT
(for ARMAX, Box Jenkins, Output error, Iv4). In experiments the dominant first chatter
mode is about 1X. Among the time series
exact chatter mode (i.e. 1X :
algorithms it was proven that the ARMAX is most suitable to represent the
a; =
611.13 Hz and i= 0.0557) in
and its harmonic one (1X or 2x) in end-milling process. Moreover the chatter mode
spectrum may be calculated mathematically using the Eqs. (7) and (8). The
other weak modes were also calculated and shown in Table which
2, cannot be seen in spectrum of Fig. 3 in all
modes. These calculated results are summarized in Table 2.
chatter mode appears in this
Fig. 3 (a) shows the spectrum by FFT method. The
spectrum. Fig. 3 (b)-(1) shows the spectra of each time series
(g)), ARXh), ARMAX), Box Jekins(j), Output Error(k), iv4(0) respectively. In our view,algorithms
i.e. AR (b)-
the AR, ARX, ARMAX
and Output Error model were
pro' en reasonable method for chatter mode 1X. But in Box Jenkins method, some
pseudo and shifted modes appear at the spectrum in Fig. 3(1). Also, among AR models,
chatter mode 1X more than the harmonic second mode 2X. In our AR(iv) intensifies the first
investigation, the most exact chatter detection
algorithm was proven ARMAX model in this experiments. For the models (ARX, ARMAX, BJ and OE
need new input, the independent white noise was also used as an model) that
input u(). The mode shown in spectrum conforms
well to the calculated modes and these were summarized in Table 2.

564
 Proceedings of ICMPM 2005, 12-14 December 2005, India

Table 2 Chatter
frequency of end-milling dynamics using cutting force Fx

mode
|AR(10) (burg) (») s622.28
t 2nd 3rd 4th Sth
1293.7 1801.9
(G) 0.0533 0.0848 0.0982
AR(10) (fb) 622.32 1293.8 1801.9
0.0532 0.0847 0.0981
AR(10) (g) 622.28 1293.8 1801.9
0.0533 0.0848 0.0982
AR(10) (iv) 1845.3 1289.9 630.41 93.470
0.0674 0.0379 0.0461 0.5121
AR(10) (ls) 622.31 1293.8 1801.9
0.0532 0.0847 0.0982
|AR(10) (yw) 622.25 1293.8 1802.1
0.0534 0.0849 0.0983
ARMA(10,3) 616.37 1297.4 1812.1 93.470
0.0714 0.0981 0.0996
ARMAX(10,5.3,0) 611.13
0.5121
66.959 1297.2 1808.7
0.0557 0.2603 0.0985 0.1010
ARX(10,3,0) (ls) 622.34 1293.8 1801.9 93.470
0.0533 0.0847 0.0982 0.5121
|ARX(10,3,0) (iv4) 1662.4 1966.8 608.00 386.50
0.0104 0.1754 0.0190 0.2420
BJ(10,3,10,3,0) (c/d) 43.107 601.32 1300.8 1825.1
0.2492 0.0818 0.1209 0.1246
BJ(10,3,10,3,0) (b/f) 2076.2 1717.5 760.43 453.14 30.319
0.0062 0.0078 0.0089
OE(3,10,0) 1777.2 0.0612 0.0047
1466.8 605.39 72.783
10.0009 30.319
0.0555 0.0005 0.0054 0.0047

FFT
AR-bug
AR-6
R-gotia
eeno hgan tre

J0
(a) b)
() (d)

565
 Procecdings of ICMPM 2005, 12-14 December 2005, India

(d) (e) (0

ARX ARMA
ARMAI Oupu E

(h) (i) (k)

B denirs

Fresen

m)
Fig. 3 Powver spectra of each method with the sampling frequency of 5 kHz for cutting force Fx

4. Concusions

1) Several time series algorithms have been developed for a chatter mode and
power spectrum estimation in an end-
milling. The advantages and drawbacks of time series modeling can be verified by analyzing the characteristics of
the power spectrum. These results may be used for the estimation of
chatter frequency that happen frequently
in the
real end-milling.
2) The ARMAX models were proven to be most suitable for the estimation of
malfunction and these can be used for
obtaining the chatter mode with less error in a more reliability. The first chatter mode(1X) locating around 670 Hz
and damping factor were calculated with less bias. The time series
method is also a good one to obtain the chatter
mode by modeling with a lower
sampling frequency and to discriminate the real chatter mode by using Eqs. (7) and
(8), which is more exact than the graphical spectrum analysis which reveals a damped and shifted mode.

5. References

1. D. N. Dilley, P. V. Bayly, A. J. Schaut, Effects of the chisel

Sound and Vibration, 2005, (281), pp. 423-438.
edge on the chatter frequency in drilling, J. of
2. T. Insperger, G. Stépán, P. V.
Bayly, B. P. Mann, Multiple chatter frequencies in milling processes, J. of
Sound and Vibration, 2003, (262),. pp. 333-345.
3. S. K. Kim, S. Y. Lee, Chatter
prediction of end-milling in a vertical machining center, J. of Sound and
Vibration, 2001, (241), pp. 567-586.
4. S. M. Pandit and S. M. Wu, Time series and
5.
system analysis with applications, John Wiley and Sons, 1993.
L. Ljung, System identification
theory
-
for the user, Prentice Hall, 1999.

566
 Proceedings of ICMPM 2005, 12-14 December 2005, India

6. L. Ljung, T. Glad, Modeling of dynamic systems, Prentice Hall, 1999.

7. T. Soderström and P. Stoica, Instrumental variable methods for system identification, Berlin:Springer
Verlag, 1993.
8. S. M. Kay, Modem spectral estimation: theory and application, Prentice Hall, 1990.
9. P. Young, Recursive estimation and time series analysis, Springer Verlag, 1993.
10. M. C. Yoon et al., A study on diagnostics of machining system with ARMA modeling and spectrum
analysis, Journal of KSMTE, 8 (1999) 42-51.
11. M. C. Yoon et al, A study on the modeling and diagnostics on chatter in end-milling operation, Journal of
KSPE, 2001, (18), pp. 101-108.
12. M.C. Yoon et al., Natural mode estimation ofcutting dynamics with time series modeling, J. of MPT, 155-
156 (2004) pp. 1373-1382.
13. M. C. Yoon, D. H. Chin, Detection of chatter in end-milling operation using wavelet transform, Proceeding
of ICAMT2004, May, 2004.
14. S. M. Kay, Modem spectral estimation, Theory & Application, Prentice-Hall, Englewood Cliffs, NJ, 1988.