Professional Documents
Culture Documents
www.elsevier.com/locate/ijmactool
Abstract
This paper proposes a new o line error compensation model by taking into accounting of geometric and
cutting force induced errors in a 3-axis CNC milling machine. Geometric error of a 3-axis milling machine
composes of 21 components, which can be measured by laser interferometer within the working volume.
Geometric error estimation determined by back-propagation neural network is proposed and used separately
in the geometric error compensation model. Likewise, cutting force induced error esti-mation by backpropagation neural network determined based on a flat end mill behavior observation is proposed and used
sep-arately in the cutting force induced error compensation model. Various experiments over a wide range of
cutting conditions are carried out to investigate cutting force and machine error relation. Finally, the
combination of geometric and cutting force induced errors is modeled by the combined back-propagation
neural network. This unique model is used to compensate both geo-metric and cutting force induced errors
simultaneously by a single model. Experimental tests have been carried out in order to validate the
performance of geometric and cutting force induced errors compensation model.
# 2004 Elsevier Ltd. All rights reserved.
Keywords: CNC machine tool accuracy; Geometric error; Cutting force induced error; Machining error; Error compensation
doi:10.1016/j.ijmachtools.2004.0
4.016
1. Introduction
High accuracy CNC milling machines are required in
many manufactures because the demand of precision
components and consistency of quality are growing. The
most important factor of the precision components is the
accuracy of machine tools. Mainly, position errors are
originated from geometric, cutting force, dynamic loading,
etc. [1,2]. Ramesh et al. [1] analyzed various sources of
geometric errors that were usually encountered on
machine tools and the methods of error compensation
employed in machines.
Ehmann et al.
[3]
presented a method to
build a gen-eral error
model of a multi-axis
machine
of
arbitrary
configuration. It is based
on the assumption of rigid
body motion and utilizes
homogenous
transformation matrices.
Okafor et al.
[4]
presented derivation of a
general volumetric error
model, which synthesized
both
geometric
and
thermal errors of a
vertical milling machine
using
homogenous
transformation matrices
of slide axis. They
measured 21 geometric
error compo-nents and
compensated the error to
validate their error model.
displacement
Ahn et al. [6] included backlash error to a volu-metric diagonal
measurement
by
laser
error model.
inter-ferometer,
when
Chen et al. [7] proposed a method to obtain all 21 error
multiple
machine
axis
components
by
performing
simple
displacement
moved
simul-taneously,
measurement along 15 lines using laser interferometer.
with
automatic
optic
Chen et al. [8] developed a system which was capable of a
1284
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
carried out in order to
of machining error
validate the performance
measurements of some determineddue to cut-ting force
of cutting force induced
is
measured
by
a
diagonal lines in the working volume.
errors compensation.
camera.
Various
The other major cause of inaccuracyexperiments over a
range
of
in CNC mill-ing machine is error due towide
conditions
cutting force. The error in workpiece is cutting
caused either by excessive deformatio-are carried out to
n at the tool and workpiece interfaceverify cutting force
due to cutting action or by deformation and machining error
of machine tool structure. Ikua et al. on a 3 axis milling
[9,10] proposed a theoretical model tomachine. The third
discusses
pre-dict cutting forces and machiningsection
error of convex and concave surfaces the combination of
and
in ball-end milling. The machin-inggeometric
force
errors resulted from force induced toolcutting
errors
deflections were calculated at variousinduced
by
the
parts of the machined sur-face. Kim etmodeled
combined
backal.
[15] presented a mathematicalpropagation neural
model to estimate cutting forces andnetwork. It is used
the resultant surface form errors into compensate both
ball-end milling in various cutting con-geometric
and
ditions. Cheng et al.
[16] studiedcutting
force
contour errors of a complete CNC induced errors simmachine system which covered allultaneously by a
groups
of
functions,
includingsingle
model.
trajectory planning, trajectory tracking,Experimental tests
cutting process and machine structure are
dynamics. Yun et al.
[11,12]
developed
a
new
method
that
calculated
cutting-condition-independent coecient and its forces over
a wide range of cutting conditions and
predicted
the
three-dimen-sional
machined surface errors generated
during the peripheral end milling
process.
This paper is organized into four
sections. The first section discusses
the geometric error model of a 3 axis
milling machine. The model presented
by Ahn et al. [6] is applied. A total of
21 geometric error components are
measured by laser interferometer. This
model is implemented in an ARDTB400 CNC milling machine. Geometric
error estimation determ-ined by backpropagation
neural
network
is
proposed and used in the geometric
error
compensation
model.
The
geometric error is compensated o-line
by mod-ifying the CNC command.
Laser interferometer mea-sures the
compensated results of two straight
lines moving on the x-y plane
simultaneously in both for-ward and
backward directions. The second
section discusses the cutting force
induced error model. It is estimated by
back-propagation
neural
network
determined based on a flat end mill
behavior obser-vation. The model is
used in the cutting force induced error
compensation model. A cutting force
sensor is developed and applied to
measure the cut-ting forces. The result
2. Geometric error
model
For
the
3-axis
milling
machine,
there are 21 error
components.
The
geometric
error
model
is
constructed
by
using a rigid body
model, small angle
approxi-mation
of
the
error,
and
homogeneous
transforma-tions. In
this
paper,
the
geometric
error
model of Ahn et al.
[6] is applied as
follows:
Px dxx dxy
dxz
ezxy
eyxz
eyyz
Sxyy
S z
xz
d e
yy zx
d e
yz zx
d e
yz zy
d e
zy yx
dzzeyx dzzeyy
exyezxz ezxSyzz
ezySyzz 1
d e
d e
d e
y yx yy
zz xx
d e
zy xx
zz xy
xy zx
e S y
zx xy
d e
xz zx
e S z
zx xz
d e
xz zy
ezySxzz dyz
Syzz exxz exyz
eyyezxz 2
P d d d e yd e d e
z
zx
d e
zy
d e
zz
xx
d e
yy xx
d e
xy yx
e e z
1 xz yx xz yy yz xx yz xy xx xy
2 eyxeyyz eyxSxyy exxSyzz eyySxzz
exysyzz eyxSxzz
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
sinhtk
; B1
cosar
1285
Cutting force
components acting
on one flute at an
arbitrary cutter rotation angle can be calculated as:
2
4
Fxi;j;k
Fy i;j;k
B1
and the
2
4
Fxi;j;k
Fy i;j;k
Fz i;j;k
3
5
2
B
21
4
A11
A31
A
A
A13
12
22
A33
32
i;j;k Kn i;j;k C3
i;j;k Kn i;j;k C4
Kf
3
5
A12 cos
where A11 C1cos/ artci;j;k;
i;j;k; A13 sin/ artci;j;k;
A21 C1sin
tci;j;k; A22 sin/tci;j;k; A23 cos/
artci;j;k;
A31 C2tci;j;k A32 cothhtci;j;k; A33 0;
coshh
sinhh
sinhtk
; C2 sinhtk
coshc
A33
32
Kf
i;j;k Kn
Kf
i;j;k Kn
B1
XX
23
i;j;k C3
A21
A
A
A
A31
12
22
A13
A
23
A3
32
Kni;j;k
Kf
i;j;k C4
A11
4
3
5
i;j;k Kn i;j;k C3
Kf
i;j;k Kn i;j;k C4
23
Kni;j;k
K
22
A
A31
Kni;j;k
Fz i;j;k
3. Mathematical model of
cutting force and machining
error
Fxi;j;k
Fy i;j;k
A13
12
Cutting
force
components on x, y
and z direction at an
arbitrary cutter rotation
angle can be obtained
by summing up the
forces acting on each
flute at that
particular cutter rotation angle:
2
shows a schematic of six degrees of
freedom error motion of a machine system.
Fig. 2 shows squareness errors between
each pair of axis.
21
k
Fz i;j;k
A11
1286
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
6EI L Lf 3L Lf L zF &
F
6EIf
Square) algorithm.
The algorithm is
provided
with
training set:
fp1;t1g; fp2;t2g;:::;
3
2
3
hzF zi Lf z 3Lf z Lf fpQ;tQg 9
2
L Lf 2L Lf L zF &Lfwhere pq and tq are the
input
and
target
output of the training
2EI 8
set.
The
algorithm
modifies the network
where ds is deflection of the shank, df ispara-meters in order
deflection of the flute, /s is deflectionto minimize the mean
angle of the shank, Lf is length of thesquare error:
network
training,
weights and bia-ses
are modified until
the
approximated
mean square error
is minimum.
5. Experiments and
results
5.1. Geometric error
5.1.1. Geometric error
determination
Three
positional
errors
(dxx;dyy;dzz),
6
angular
errors
(eyx;ezx;exy;ezy;exz;eyz),
6
straightness
errors
(dyx;dzx;dxy;
dzy;dxz;dyz)
and 3 squareness errors
(Sxy;Sxz;Syz)
were
measured
by
laser
interferometer.
3
angular
errors
(exx;eyy;ezz) were derived
from
vertical
straightness
Fig. 4. 12-1
Network.
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
both
forward
and
backward
errors as referred in thedirections.
The
paper of Ahn et al. [6].geometric
All error components inerror of x, y
z axis
both
forward
andand
was
backward directions werecalculated
Eqs.
measured.
The
laserby
system consisted of the(1)(3) with
to
following
components:respect
the
ML10 measurement laser,measuring
EC10
environmentalposition x,
and z.
compensation
unit,yFunction
interferometer, reflectorapproximat
mirror set, and computer.ion
by
The error models asbackfunctions of positions in x,propagatio
n
neural
y and z axis were investi-network
gated.
Figs. 5 and 6was used
to
show some of the errorapproximat
compo-nents.
Fig. 5e
show the position andgeometric
error
straightness errors alongmodel. The
x
axis
in
backwardbackdirection. Fig. 6 show thepropagation
angular error along y axisneural
in backward direction.network
of
Esti-mation of the errorstructure
geometric
model at an intermediateerror model
position
between
thewas shown
measured
nodes
wasin Fig. 7.
determined
by
back-The
x,
propagation
neuralpositions
y and z of
network.
the
measuring
point were
5.1.2. Geometric error model the input of
the
A total of 21 errornetwork
components
wereand
measured
directly
by
laser interferometer in
Fig. 6.
Angular
error
along y
axis in
backwar
both
forward
and
backward
directions.
the
After
geometri training the
net-work,
c error inthe
x, y and zgeometric
axis wereerror could
be
the
determined
this
output ofusing
model.
the
network.
5.1.3.
The
transfer Geometric
function error
of
the
hidden compensation
layer was To
Hyperevaluate the
bolic
Tangent algorithm,
Sigmoid the
and
transfer geometric
function error of two
of
thestraight
output
layer waslines moving
Linear.
on x-y plane
The
both
number in
of nodesforward and
in
thebackward
hidden directions,
layer was
as
varied
until theillustrated in
approxim
Fig. 8, was
ated
mean
compensate
square d.
These
error was
lines
were
satisfied
at eightyrepresented
nodes for
d direction.
Fig. 8.
1287
by
parametric
equations.
The
compensati
on algorithm
modi-fied
the CNC Gcode
command
by
subtracting
the
expected
error
from
the nominal
position to
reduce the
error.
Laser
interferomet
er was used
to measure
position
errors of the
machine
following
a
straight line
from
point
Poxo;yo to
point
P1x1;y1.
Due
to
position
error,
the
dierence
between
nominal and
real
distances
was
identified by
laser
interferomet
er.
The
modified
Line 1
and line 2
in both
forward
and
backward
directions
.
1288
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
Test no.
Spindle speed
(rpm)
Depth of
cut (mm)
Test 1
Test 2
Test 3
Test 4
Test 5
Test 6
Test 7
Test 8
Test 9
Test 10
Test 11
Test 12
835
835
835
835
835
835
835
835
835
835
835
835
1
1
1
1
2
2
2
2
3
3
3
3
Feed rate
(mm/min)
50
100
150
200
50
100
150
200
50
100
150
200
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
1289
Transfer function of
the hidden layer was
Hyperbolic
Tangent
Sigmoid and transfer
function of the output
layer was Linear. The
number of nodes in the
hidden
layer
was
varied
until
the
approximated
mean
square
error
was
satisfied
at
twenty
nodes for both forward
and
backward
directions.
After
training the network,
the geometric error
could be determined
using this model.
5.2.3. Cutting force
induced error
compensation
To
evaluate
the
algorithm, the machining
error of four straight line
slots cutting on x-y
plane moving from point
A to B to C to D and to
E, as illustrated in Fig.
15, was compensated.
The
compensation
algor-ithm
modified
CNC G-code command
by
subtracting
the
expected error from the
nominal
position
to
reduce error. A set of 4
cutting
tests
were
carried out. Table 2 lists
the cutting conditions.
All the tests were
conduc-ted with the
20 lm for test 4.
5.3. Error model
combination
approximation by backpropagation
neural
network. The backpropa-gation
neural
network structure of
the combined model is
shown in Fig. 17.
The x, y, z positions,
tool tip feedrate and
tool axial depth of cut
were the input of the
network. Geometric
1290
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
geometric
6. Conclusions
and
cutting
A
new
line
modeled
model
combined
and
taking
into
geometric
cutting
compensate
machine
was
geometric
this
cutting
in
Geometric
cutting
force
1
2
1.5
1
65
35
70
130
both
and
force
by
model.
a
The
experimental results
show
that
applying
the
geometric
error
compensation,
machine
by
the
accuracy
could be improved.
Estimation
835
835
835
835
neural
nonlinear.
Test 1
Test 2
Test 3
Test 4
propagation
Feedrate
(mm/min)
back-
and
Depth of
cut (mm)
the
induced errors in a
paper.
Spindle
speed (rpm)
by
force
Test No.
error compensation
account
Table 2
Cutting conditions
force
of
the
geometric error at
Cutting
force (kg)
From
the
experiment,
the
average
working
volume was
10.0
improvement of line
13.0
carried
15.0
out
of
was
separately by back-
approximately
propagation
and
network.
neural
Likewise,
estimation
of
cutting
force
induced
also
the
error
was
carried
out
28
lm
forward
backward directions,
respect-ively.
was
propagation
and
combination
The
of
The
improvement of line
approximately
network.
in
and
separately by backneural
34
40
forward
lm
44
in
and
backward directions,
show
machine
accuracy
was
improved
signifi-cantly.
that
the
C. Raksiri, M. Parnichkun / International Journal of Machine Tools & Manufacture 44 (2004) 12831291
1291
Acknowledgements
This research project is financially
supported by Mitutoyo Association for
Science and Technology.
References
[1]
[2]
M.
Rahman,
J.
Heikkala,
K.
Lappalainene, Modeling measure-ment and
error compensation of multi-axis machine
tools. Part I: theory, International Journal of
Machine Tools & Manufacture 40 (2000)
15351546.
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
H.F.F. Castro, M.
Burdekin,
Dynamic
calibration
of
the
posi-tioning accuracy
of machine tools and
coordinate
measuring machines
using
a
laser
interferometer,
International Journal
of Machine Tools and
Manufacture
43
(2003) 947954.
[11]
[12]
[13]
Cho,
[14]
Estimation of cutter
deflec-tion and form
error
in
ball-end
milling
process,
International Journal
of Machine Tools and
Manufacture
43
(2003) 917924.
[17]