IWMF2014, 9th INTERNATIONAL WORKSHOP ON MICROFACTORIES

OCTOBER 5-8, 2014, HONOLULU, U.S.A. / 1

Thermal error compensation for a high

precision lathe
Byung-Sub Kim# and Jong-Kweon Park
Department of Ultra Precision Machines and Systems, Korea Institute of Machinery & Materials, Daejeon, South Korea
# Corresponding Author / E-mail: bkim@kimm.re.kr, TEL: +82-42-868-7109, FAX: +82-42-868-7180
KEYWORDS : Thermal error model, Compensation, Thermoelastic process, Desktop 5-axis milling machine

High precision machines require very stable operational environment: temperature control and
vibration isolation. Tight temperature control for machines usually demand high cost to operate air
conditioners. Some of high precision machines require the ambient temperature changes to maintain
within 0.1 degrees. In this paper, we present a thermal error compensation scheme and
experimental results for improving machining accuracy of a high precision lathe. The testbed lathe
has X- and Z-axes and they are driven by linear motors and hydrostatic oil bearing. Due to the
temperature changes of the ambient air and supplied oil to the hydrostatic bearing, thermal
deformation is generated and measured to be around 0.4 m. To identify the dynamic relations
between the temperature changes and the thermal drift, a state-space model is used in which state
variables are constructed from the input measured temperatures and the output thermal drift data.
The identified model is implemented in a servo control loop and the predicted thermal error is
compensated by subtracting the predicted thermal drift from the servo command. In our simulation, a
thermal error of 97 nanometers RMS over 3 hours is reduced to 55 nanometers RMS. Experimental
results show an average of 24% reduction in thermal drift and support the validity of our approach.

1. Introduction

the thermal deformation in the machine tool. Li7 adopted an

auto-regressive (AR) model to predict spindle thermal errors
from spindle speed. Hong8 showed that the pseudo-hysteresis
effect is the main factor causing poor robustness of the static
error model and proposed to use a linear output error (OE)
model to predict the machine tool thermal deformation.
We present a thermal error compensation scheme in which
a state-space model is used for thermoelastic process modeling.
In our simulation, the state-space modeling approach showed
better robustness to capture the dynamic nature of thermal
deformation than other linear parametric models including AR
and OE models. Our machining process usually takes one or
two hours and during the machining a thermal drift appears
with roughly a period of 13-15 minutes and a magnitude of
300-400 nanometers.

The machine tool accuracy directly affects on the

dimensional accuracy of the machined products. In response to
the increasing requirement of product quality, the nature of
thermally induced errors and geometric errors has been
investigated by many researchers over the past few decades1,2.
The thermal error accounts for 40-70% of the total machining
errors and it is the most significant factor influencing the
machine tool accuracy1. Thermally induced errors are quasistatic errors that change slowly in time and are closely related
to the machine tool structure. The commonly used empirical
thermal error model assumes that the thermal deformation of a
machine tool at a particular instant depends on the temperature
distribution in the machine tool at that particular time and
describes the relation between the thermal deformation and the
temperature distribution3-5. These static error model
approaches often tend to be unreliable when working
conditions are different from the tested conditions.
To provide more robust models, researchers have taken into
consideration the dynamics of the thermoelastic process in a
machine tool. Moriwaki6 suggested using an empirically
determined transfer function model describing the process
parameters between spindle speed, ambient temperature, and

2. Testbed Machine: a High Precision Lathe

The testbed high precision lathe has X- and Z-axes and
these axes are driven by linear motors and hydrostatic oil
bearing. The spindle in the testbed machine has a built-in
motor with air bearing. Air is supplied to the spindle for
cooling. A picture of the testbed is shown in Fig. 1. The stroke

239

IWMF2014, 9th INTERNATIONAL WORKSHOP ON MICROFACTORIES

OCTOBER 5-8, 2014, HONOLULU, U.S.A. / 2

than 50 nanometers and the yaw error less than 0.15 arcsec as
shown in Fig. 3. The reason why the Z-axis was not used to
compensate for the horizontal straightness error of the X-axis,
was that the encoder resolution of the Z-axis was not small
enough and we could get finer dynamic response from ACC in
reduction of the straightness error. The next target for high
precision machining goes to the thermally induced error.

of the X-axis is 200 mm and that of the Z-axis 50 mm. The

linear encoders for position feedback have a 10 nanometer
resolution for the X- and Z-axes. The temperature of the
machining room is maintained by an air conditioner with 0.6
C variations. Covering the lathe with metal panels reduces
the ambient temperature change around the machine to be
0.3 C. The temperature of the oil from the hydrostatic oil
bearing is controlled by an oil-cooler within 0.13 degrees,
when the temperature is measured from the oil drain for the
accessibility reason. When the oil bearing and the spindle are
turned off, the distance between the spindle and the tool post
has a thermally induced oscillation of 70 nanometers reflecting
the ambient temperature changes. When the oil bearing and a
servo control are on and the spindle rotates at a speed of 1,000
rpm, thus the lathe is ready to operate, the magnitude of the
oscillation increases to be around 300-400 nanometers.

D eviation arcsec

D eviation m

0 .1

S tra ig h tn e s s e rro r
0 .0

-0 .1
0 .5

Y a w in g e rro r
0 .0

-0 .5

50

100

150

200

P o s itio n m m

Fig. 3 Straightness and yaw error profiles corrected by ACC

3. Linear Thermoelastic System Identification

To obtain a linear thermal error model describing the
relation from the temperature inputs to the thermally induced
error output, four temperature sensors were attached to the
testbed lathe and the sensor values were measured at every 1
second. The temperature sensor locations are spindle housing,
table, oil drain, and ambient air around the machine. The red
dots in Fig. 4 show specific locations for temperature
measurement except the ambient temperature. A temperature
sensor for the ambient temperature is hung approximately 200
mm over the tool holder.

Fig. 1 Picture of a high precision lathe

Spindle
Tool holder

Fig. 2 Measured error on a workpiece surface due to thermal

deformation and straightness error

After a diamond turning test, it was found that there was a

wavy form error on the machined surface induced by thermal
deformation and straightness error as shown in Fig. 2. Before
tackling the thermal error, we used two actively controlled
capillaries (ACC)9, which acted on one side of the X-axis, to
reduce the horizontal straightness and yaw errors of the X-axis.
The ACC has a piezo actuator to change the supplied oil
pressure to the hydrostatic bearing. The oil pressure change
along the X-axis offers two-degrees-of-freedom motion to the
X-axis stage. The input voltage values to the piezo actuators
were empirically determined so that the straightness and the
yaw errors of the X-axis could be reduced under a certain level.
The straightness error of the X-axis was corrected to be less

240

Fig. 4 Temperature sensor locations shown as red dots

Thermal drift and accompanying temperature sensor

values for 4 hours are shown in Fig. 5. The thermal drift is the
relative displacement change between the workpiece and the

IWMF2014, 9th INTERNATIONAL WORKSHOP ON MICROFACTORIES

0.1
0
-0.1
120
Time (minutes)

Spindle Housing

0.2

180

240

Table

0.2

60

Temp. change ( C)

-0.2
0

0
-0.1
240

0
-0.1
-0.2
0
0.2

60 120 180
Time (minutes)
Oil Drain

240

0.4

60 120 180
Time (minutes)
Ambient

Temp. change ( C)

-0.2
0

0.1

Thermal drift (micron)

0.1

0.2
0
-0.2
-0.4
0

60 120 180
Time (minutes)

240

0.1
0

0.4

-0.2
0

60 120 180
Time (minutes)

240

Fig. 5 Thermal drift (top) and accompanying temperature sensor

values (below four plots)

When a linear regression model was used, the predicted

thermal drift did not match well with the real data. Any
compensation based on the linear regression model seemed to
make the error bigger in our particular case. To consider the
dynamics of the thermoelastic process in the testbed, different

Predictions with normal temperature data

0.2
0
-0.2
-0.4
0

-0.1
Thermal drift (micron)

Temp. change ( C)

system identification methods were tested and evaluated. They

were ARX(auto-regressive exogenous input), ARMAX(autoregressive moving-average exogenous input), BJ(BoxJenkins), OE, and state-space models.
A state-space model did not always show best accuracy in
prediction of thermal drift among tested dynamic models, but
it showed more robust results than other models. For example,
an OE model and a state-space model, which were built based
on the data set depicted in Fig. 5, were applied to another data
set and their predictions were compared in Fig. 6. The solid
line shows the measured thermal drift of 97 nanometers RMS
over three hours. The top plot shows the predictions by the OE
model (dashed line) and the state-space model (dash-dot line)
with unperturbed temperature data. If the thermal drift is
compensated as much as predicted by the OE model, the
thermal error will be reduced to 87 nanometers RMS, and if
the state-model is used for compensation, the thermal error
will be 55 nanometers RMS in this particular simulation. We
can see that the both models can track the general shape of the
thermal drift. When a small constant bias is injected in the
temperature data, the simulation results are compared in the
bottom plot. We added 0.1 degrees to the ambient temperature
and subtracted 0.1 degrees from the oil drain temperature. The
state-space model seems to predict the thermal drift with
perturbed-mean temperature but the OE model goes in the
wrong direction. Similar phenomena could be seen with the
ARX and BJ models. Based on our observation, the statespace model was selected as an appropriate thermal error
model for our experiments. When a thermoelastic model is
built, we remove the mean values from the input temperature
data before computation. Since it is an off-line computation
after collecting the temperature data, the accurate mean values
can be removed. But there is no guarantee that the mean
values in off-line computation will reappear as those in the
real-time on-line data. That is why the thermal error model
should be insensitive to the constant offset of the temperature
mean values.

0.2

Temp. change ( C)

Thermal drift (micron)

tool holder, which is measured by a gap sensor, and all the

values in the plots are adjusted to have zero mean values. The
positive displacement value means that the workpiece and the
tool holder come closer in our experimental setup. The thermal
drift shown in Fig. 5 is 70 nanometers RMS. Notice the
symmetric configuration of the high precision lathe. Due to
symmetry, the relative displacement change between the
workpiece and the tool holder does not show noticeable
difference along the X-axis. In Fig. 5, we can see some pattern
in the thermal drift and the ambient temperature change.
Correlation coefficients indicate that the ambient temperature
(0.52) has the most influence on the thermal drift, and the next
is the table and the oil drain temperature (0.48 and -0.15,
respectively) in our experimental setup. The temperature of the
spindle housing appears to have least influence on the thermal
drift. It is because the spindle was continuously provided
cooling air through dedicated air tubes. The ambient
temperature around the lathe seemed to include the
characteristics of the table temperature. Thus, we chose the oil
drain temperature and the ambient temperature as the inputs to
our linear model and the thermal drift as the output. In our
simulation, adding the table temperature as one of the input
parameters to our thermoelastic model did not help improving
the accuracy of prediction for the thermal drift.

OCTOBER 5-8, 2014, HONOLULU, U.S.A. / 3

0.4

Measured

Predic. by OE

30

60

30

60

Predic. by SS

90
120
150
Time (minutes)
Predictions with perturbed-mean temperature data

180

0.2
0
-0.2
-0.4
0

90
120
Time (minutes)

150

180

Fig. 6 Comparisons of measured thermal drift and predictions by

a 2nd order OE and state-space models: with normal
temperature data (top) and with perturbed-mean temperature
data (bottom)

241

IWMF2014, 9th INTERNATIONAL WORKSHOP ON MICROFACTORIES

OCTOBER 5-8, 2014, HONOLULU, U.S.A. / 4

4. Experimental Results

State-space models are models that use state variables to

describe a system by a set of first-order difference equations,
rather than by one or more nth-order difference equations.
State variables x(k) can be reconstructed from the measured
input-output data, but are not themselves measured during an
experiment (10). State-space models are not derived from
physical equations, so the states in a state-space model have
no direct physical meaning. More information on the statespace system identification method can be found in (11).
Using Matlab software, a second order state-space model
was obtained as follows,

In our experimental setup, the NC controller for the

testbed high precision lathe is a UMAC system from Delta
Tau Data Systems Inc. The servo control update rate is
approximately 2.26 kHz and the sampling rate for thermal
drift prediction is 1 Hz. Since there was a big difference
between these two rates, another DSP (Digital Signal
Processing) system was used to run the thermoelastic model
in real-time. The amount of thermal drift, which was
predicted from on-line temperature sensor values, was
returned to the NC system by the DSP system through an AD
(Analog-Digital) board. The NC system subtracted the readin thermal drift from the position command and made a
compensated actual position command for the Z-axis.
Experimental results are shown in Fig. 8. The thermal
error compensation had been alternatively on and off at every
60 minutes and the consecutive thermal drift was recorded for
four hours. The colored boxes in the figure enclose the
maximum and the minimum of the thermal drift in each period
for an easy comparison of peak-to-peak values. When the
compensation was off, the thermal drifts were 61.4 and 65.9
nanometers RMS from the second and the fourth period,
respectively. During the compensation on, they reduced to
45.2 (first period) and 51.5 (third period) nanometers RMS.
On average, the thermal error compensation obtained a 24%
reduction in thermal drift. The dashed line in the figure shows
the real-time prediction of the thermal drift without sending a
compensating signal to the NC system. We can see how
closely the identified state-space model predicts the thermal
drift in real-time and how big an error can be made by a small
phase mismatch.

x(k+1) = A x(k) + B u(k) + K e(k)

y(k) = C x(k) + D u(k) + e(k)
where x(k) is a state vector, u(k) is a temperature input
vector, u(k) = [Ambient temp.(k), Oil drain temp.(k)]T, e(k) is a
noise vector with Gaussian distribution, y(k) is a thermal error
drift output, k is a discrete-time step such that time t =
ksampling period. Our sampling period was 1 second. The
matrixes A, B, C, D, K and an initial condition were
A = [9.9490e-1, -1.7756e-2; 4.1983e-4, 9.9686e-1],
B = [6.4115e-4, -1.4311e-3; -6.5977e-6, -2.7177e-4],
C = [5.9741, -9.3924e-3],
D = [0, 0],
K = [1.7009e-1, -7.8078e-2]T,
x(0) = [1.3373e-2, -7.9594e-3]T.
Simulation result shown in Fig. 6 has carried out with no
noise and a zero initial condition. It may be interesting to see
the frequency response of the identified state-space model.
The Bode plots are shown in Fig. 7. It cant be claimed that
the identified model exactly describes the real thermoelastic
process in our experimental setup, but there are common
dynamic characteristics captured by other dynamic models as
well as the state-space model. From the frequency responses,
we can say that the effect of the ambient temperature change
on the thermal drift is faster (small phase lag) and stronger
(large DC gain) than that of the oil drain temperature change.
Reflecting on the area they affect, it is not against our
common sense. From the step response test, a step change in
temperature takes about 21 (by ARX) to 33 minutes (by
state-space) until its effect fully appears in the thermal drift.

0.2

Compesation ON
RMS = 45.2 nano

Compesation OFF
RMS = 61.4 nano

Compesation ON
RMS = 51.5 nano

Compesation OFF
RMS = 65.9 nano

Thermal drift (micron)

0.1

Real-time
Prediction w/o
Compensation

-0.1

10
Amplitude

10

-1

-0.2
0

10

120
Time (minutes)

180

240

Fig. 8 Experimental results of the thermal error compensation

-3

10 -6
10

-5

10

-4

10

-3

10

-2

10

-1

10

0
Phase (deg)

60

-2

10

5. Conclusions

-100
-200
-300 -6
10

In this paper, a state-space dynamic thermal error model is

presented for thermal error compensation in a high precision
lathe. Based on the system identification theory, a state-space
model is used for dynamic modeling and online prediction for
thermal drift. The thermally induced errors in the conventional

Ambient -> Thermal Drift

Oil Drain -> Thermal Drift
-5

10

-4

-3

10
10
Frequency (Hz)

-2

10

-1

10

Fig. 7 Frequency responses of the identified state-space model

242

IWMF2014, 9th INTERNATIONAL WORKSHOP ON MICROFACTORIES

machine tools are usually more than scores of microns, but the
thermal error dealt in this paper is few hundred nanometers.
Thus, a linear thermal error model in our research should be
able to susceptibly follow any little temperature changes. The
effectiveness of state-space modeling approach is verified
through computer simulation and experiments. In our
simulation, a thermal error of 97 nanometers RMS over 3
hours is reduced to 55 nanometers RMS. Experimental results
show an average of 24% reduction in thermal drift and support
the validity of our approach.

REFERENCES
1. Ferreira, P. M. and Liu, C. R., A method for estimating
and compensating quasi-static errors of machine tools,
ASME Journal of Engineering for Industry, Vol. 115, No. 1,
pp. 149159, 1993.
2. Donmez, M. A., Liu, C. R. and Barash, M. M., General
Methodology for Machine Tool Accuracy Enhancement by
Error Compensation, Precision Eng., Vol. 8, No. 4, pp.
187196, 1986.
3. Yang, S., Yuan, J. and Ni, J., The Improvement of
Thermal Error Modeling and Compensation on Machine
Tools by Neural Network, Int. J. Machine Tools and
Manufacture, Vol. 36, No. 4, pp. 527537, 1996.
4. Yang, J., Yuan, J. and Ni, J., Thermal Error Mode
Analysis and Robust Modeling for Error Compensation on
a CNC Turning Center, Int. J. Machine Tools and
Manufacture, Vol. 39, pp. 13671381, 1999.
5. Yang, J., Yuan, J. and Ni, J., Thermal Error Mode
Analysis and Robust Modeling for Error Compensation on
a CNC Turning Center, Int. J. Machine Tools and
Manufacture, Vol. 39, pp. 13671381, 1999.
6. Moriwaki, T. and Shamoto, E., Analysis of Thermal
Deformation of an Ultraprecision Air Spindle System,
CIRP Annals, Vol. 47, No. 1, pp. 315319, 1998.
7. Li, S., Zhang, Y. and Zhang, G., A Study of Precompensation for Thermal Errors of NC Machine Tools,
Int. J. Machine Tools and Manufacture, Vol. 37, No. 12, pp.
17151719, 1997.
8. Yang, H. and Ni, J., Dynamic Modeling for Machine Tool
Thermal Error Compensation, J. Manufacturing Science
and Engineering, Vol. 125, pp. 245-253, 2003.
9. Park, C. H., Oh, Y. J., Hwang, J. H., and Lee, D. W.,
Motion error compensation method for hydrostatic tables
using actively controlled capillaries, J. Mechanical
Science and Technology, Vol. 20, No. 1, pp. 51-58, 2006.
10. System Identification Toolbox, Users Guide, MathWorks
Inc., 20013.
11. LJung, L. System identification: Theory for the user, Sec.
10.6, Prentice Hall Inc., 1999.

243

OCTOBER 5-8, 2014, HONOLULU, U.S.A. / 5