Measurement Science and Technology

PAPER

Optimal measurement angles of the three-probe spindle error motion

separation technique
To cite this article: Yang Jiao et al 2019 Meas. Sci. Technol. 30 095001

View the article online for updates and enhancements.

This content was downloaded from IP address 18.218.56.169 on 24/10/2019 at 11:04

 Measurement Science and Technology

Meas. Sci. Technol. 30 (2019) 095001 (7pp) https://doi.org/10.1088/1361-6501/ab20af

Optimal measurement angles of the

three-probe spindle error motion separation
technique
Yang Jiao , Ming Huang and Pinkuan Liu1
School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240,
People’s Republic of China

E-mail: [email protected] (P Liu)

Received 15 January 2019, revised 3 May 2019

Accepted for publication 9 May 2019
Published 19 July 2019

Abstract
The three-probe error separation technique uses three displacement sensors to separate the
spindle radial error motion from the artefact roundness error. The choice of appropriate sensor
angles is crucial for an accurate and complete separation. This work discusses the optimal
arrangement of the sensors’ angular positions in detail, including the influence of angular
position errors of the sensors on the separation results. The optimization goal is to find a
suitable sensor arrangement to simultaneously maximize the minimum value and the average
value of the transfer function, in order to improve the anti-interference ability and ensure
measurement precision of the three-probe method without harmonic suppression. The result
shows that there is a global optimum for the sensor angles for a specific range of harmonics.
The obtained optimal measurement angles under different harmonic ranges should be useful
for applications of the three-probe method.
Keywords: spindle metrology, radial error motion, multi-probe error separation, optimization

(Some figures may appear in colour only in the online journal)

1. Introduction harmonic components of the roundness error and the error

motion cannot be effectively separated in some cases [5, 6].
The axis of rotation error motion is one of the important tech- Appropriate choice of measurement angles is useful for
nical specifications for assessing the performance of a spindle. avoiding the suppression of low-order harmonics [7]. In view
Usually, the spindle radial error motion is measured directly of this, several authors recently performed optim­izations
by a displacement sensor sensing against a reference artefact for determining the suitable sensor angles. In [5], the so-
[1]. However, the measurement data is also influenced by the called p-mean of the complex harmonic transfer function
artefact’s roundness error, and for high-precision spindles, was chosen as the optimization goal, and one set of optimal
this influence cannot be neglected. Three classes of error angles was obtained with the harmonic range of 2–128. In
separation techniques have been commonly used for sepa- [8], one of the optimal measurement angles was selected
rating the spindle error motion from the artefact roundness by maximizing the minimal amplitude of the transfer func-
error, i.e. reversal, multi-step, and multi-probe methods [2, 3]. tion at the harmonic range of 2–150. Marsh et  al [6] also
Compared to reversal and multi-step methods, the multi-probe
determined an effective set of sensor angles, but they did
method is more suitable for in situ measurement because this
not describe the selection procedure. In [9], optimum sensor
method does not need to reorient the artefact during testing
angles at a series of harmonic ranges were computed by
[4].
minimizing the effects of uncorrelated sensor noise on the
Multi-probe error separation, however, may suffer from
reconstructed roundness error. However, as the optimization
the harmonic suppression problem, which means that some
results vary with types of noise, they can hardly be put into
1
Author to whom any correspondence should be addressed. practice.

1361-6501/19/095001+7$33.00 1 © 2019 IOP Publishing Ltd  Printed in the UK

Meas. Sci. Technol. 30 (2019) 095001 Y Jiao et al

The above studies are useful for determining the effec-

tive probe angles, although the recommended angles for each
researcher are different. Moreover, a systematic analysis of the
optimum angles at different harmonic ranges for ensuring meas-
urement precision without harmonic suppression has not been
developed. Since the main aim is to find the sensor angles that
make the transfer functions at important harmonics (harmonics
with larger magnitudes, usually the low-order harmonics) as
large as possible [7], the optimization criterion proposed in [8] is
feasible. However, the transfer function in its most general form
[7, 8] is not suitable to be used in the optimization problem.
Based on the above-mentioned valuable works, this paper
attempts to thoroughly investigate the optimal measurement
angles of the three-probe method. The influence of the dis- Figure 1.  Schematic diagram of the multi-probe error separation.
placement sensor’s angular position error is also discussed
since it is important for the establishment of a suitable optim­ where S(n) and R(n) are the Fourier coefficients of s(θ) and
ization problem. r(θ), respectively, and B(n) is the transfer function of the
three-probe method. Given S(n) and B(n), the artefact round-
ness error can, in principle, be derived by using the inverse
2.  Multi-probe error separation and angular posi-
DFT (IDFT):
tion errors of probes
r (θ) = IDFT [R (n)] = IDFT [S (n)/B (n)] .
(6)
2.1.  Multi-probe error separation technique
Finally, by substituting r(θ) into equation  (1), the X and Y
The multi-probe error separation method is briefly reviewed components of the radial error motion can be obtained:
in this section. Figure 1 shows the configuration of a three-dis-
x (θ) = s1 (θ) − r (θ)
placement-sensor system. The measured data for each sensor (7)
y (θ) = [s2 (θ) − r (θ + β2 ) − x (θ) cos β2 ]/sin β2 .
sk(θ) (k  =  1, 2, 3) is a combination of the spindle radial error
motion in X and Y directions, x(θ) and y (θ), and the artefact It should be emphasized that multi-probe error separation
roundness error r(θ), with a phase shift due to the sensor’s highly depends on the angular arrangement of the displace-
angular position, as shown in equation (1), assuming a clock- ment sensors. For a specific harmonic, according to equa-
wise rotation. tion (6), when its transfer function B(n) is equal to zero, the
corresponding Fourier coefficient R(n) cannot be obtained.
sk (θ) = r (θ + βk ) + x (θ) cos βk + y (θ) sin βk
(1)
This is the well-known phenomenon of harmonic suppression.
Here, βk is the angular position of the sensor Sk. For conve- Furthermore, when B(n) is close to zero, a small error on S(n)
nience, β1 is usually set to 0, and therefore s1(θ)  =  r(θ)  +  x(θ). will cause a relatively large error on R(n) (i.e. the noise ampli-
A weighted sum of the three measurements can be used to fication effect). In these situations, the reconstructed results of
remove the influence of the X and Y radial error motions: the radial error motions will not be accurate. Therefore, it is
3

important to find suitable sensor angles that make the values
s (θ) =
(2) bk sk (θ), of B(n), at the specific harmonic range, as far away from zero
k=1 as possible.
where the coefficients b1, b2, and b3 should satisfy In addition, the coefficients b1, b2, and b3 used here are
slightly different from their general forms in which b1  =  1,
3 3
  b2  =  −sin β3/sin (β3  −  β2), and b3  =  sin β2/sin (β3  −  β2)
bk cos βk = 0,
(3) bk sin βk = 0. [4, 6–8]. Hence, the transfer function B(n) used in this paper is
k=1 k=1
also different from its general form BG(n), and we have
Solving equation (3) gives
BG (n) = B (n)/sin (β3 − β2 ).
(8)
b1 : b2 : b3 = sin (β3 − β2 ) : sin (β1 − β3 ) : sin (β2 − β1 ) .
(4) This slight modification is important for the determination of
For convenience, b1  =  sin (β3  −  β2), b2  =  −sin β3, and a suitable global optimization problem.
b3  =  sin β2 are used in this paper (β1  =  0).
After eliminating the influence of the spindle’s error 2.2.  Influence of angular position errors
motion, only the contribution of the roundness error is left Measurement results of the error separation techniques are
in equation  (2). By applying the discrete Fourier transform inevitably influenced by various errors. Several researchers
(DFT), equation (2) can then be written as have made useful contributions to the error budget of the
3
 spindle error motion measurement [4, 8, 10]. For the three-
S (n) =
(5) bk einβk R (n) = B (n) R (n) , probe method, the angular position error (also known as
k=1 positioning error), the sensitivity error of the sensor, and the

2
Meas. Sci. Technol. 30 (2019) 095001 Y Jiao et al

sensor noise are often the main error influences. In general,

sensitivity error can be reduced by carefully calibrating the
sensor, and using one sensor instead of three (rotating the
sensor at three different orientations) [8]. Additionally, the
influence of sensor noise on the filtered synchronous radial
error motion can also be reduced by increasing the number of
sampling points per revolution and the number of revolutions
[2, 11, 12]. Angular position errors of the probes, however, are
difficult to eliminate in most practical cases. Although their
influences have been preliminarily discussed in [8, 10], a thor-
ough investigation of the principle of the influences is neces-
sary since it is also important for the optimization problem.
Assuming that the actual angular position of displacement
sensor Sk has an error γk relative to its nominal position, meas-
urement data for each sensor can then be expressed as
sk (θ) = r (θ + βk + γk ) + x (θ) cos (βk + γk )
+ y (θ) sin (βk + γk ) . (9) Figure 2.  Variation of the coefficient |exp(inγ)  −  1| with sensor’s
 angular position error γ at five different frequency components.
The weighted sum of the three measurements can be rewritten ΔR(n) contains contributions of both roundness error and
as radial error motion. However, their contributions are dis-
3 tinctly different, as can be seen from the coefficients of
s (θ) = bk sk (θ). exp(inβ) R(n) and –sin(β) X(n)  +  cos(β) Y(n). For a given
(10) k=1 γ, the coefficient exp(inγ)  −  1 changes with Fourier order n,
Performing the DFT on equation (10) gives while the coefficient associated with the error motion term
remains constant (γ).
3
 î Figure 2 presents the values of the coefficients |exp(inγ)  −  1|
S (n) = bk ein(βk +γk ) R (n) + cos (βk + γk )
as a function of the angular position error γ, with the harmonics
k=1
(11) n  =  10, 50, 100, 200, and 300, respectively. The coefficient
X (n) + sin (βk + γk ) Y (n)].
 increases linearly as γ increases. Furthermore, by comparing
The measurement errors on the Fourier coefficients of the |exp(inγ)  −  1| at different harmonics, it can be clearly seen
reconstructed roundness error are therefore obtained: that high-order frequency components are more sensitive to
the angular position errors. For the 300th harmonic, an angular
1
∆R (n) = R (n) − R (n) =
(12) [S (n) − S (n)] . position error γ  =  0.1° (6′) can lead to |exp(inγ)  −  1|  >  0.5,
B (n)
which means that the 300th harmonic of the measurement
Assuming cos γ  ≈  1 and sin γ  ≈  γ for a small γ, inserting error, ΔR(300), contains about 50% information of the har-
equations (11) and (5) into (12), and using equation (3), yields monic R(300) of actual roundness error; therefore, the corre­
3 sponding harmonic components of the reconstructed r′(θ),
1   inγk  x′(θ), and y ′(θ) are no longer believable. Since the angular
∆R (n) = bk e − 1 einβk R (n)
B (n) position errors are unavoidable, in practical applications, the
k=1
+γk [− sin βk X (n) + cos βk Y (n)]}. (13) harmonic range used in error separation must be designed in
 accordance with the sensors’ positioning accuracies.
Equation (13) is important since it reveals the principle of the In addition, since |γ|  ≈  |exp(iγ)  −  1|, the contribution of
angular position errors’ influences. By utilizing the IDFT, the radial error motion to the total measurement error is negli-
measurement error on the reconstructed roundness error is gible when compared to the contribution of roundness error.
obtained: For a clear demonstration of the angular position error’s influ-
ence, a numerical example of roundness error and radial error
∆r (θ) = r (θ) − r (θ) = IDFT (∆R (n)) .
(14) motion in X and Y directions is used, as illustrated in figure 3.
The measurement errors on the reconstructed X and Y radial The measurement errors Δr(θ), Δx(θ), and Δy (θ) when the
error motions can finally be computed: angular position errors are γ1  =  0, γ2  =  0.1°, and γ3  =  −0.1°
are presented in figure 4(red lines). By setting exp(inγk)  −  1
∆x (θ) = s1 (θ) − s1 (θ) − ∆r (θ) to 0 (k  =  1, 2, 3) in equation (13), measurement errors only
1
∆y (θ) = sin β2 {s2 (θ) − s2 (θ) − ∆r (θ + β2 ) − ∆x (θ) cos β2 } . contributed by radial error motion can be calculated. These
(15) are indicated by the blue lines in figure  4, which are very
From equation (13), it can be found that for the three-probe small compared to the total measurement errors and can be
method, the total error caused by angular position errors is neglected. In brief, the roundness error (especially for its
a linear superposition of the measurement errors due to γ1, high-order frequency components) is a major contributor to
γ2, and γ3, respectively. Moreover, the measurement error the measurement error caused by angular position errors.

3
Meas. Sci. Technol. 30 (2019) 095001 Y Jiao et al

function as large as possible in order to improve the anti-

interference ability and ensure measurement precision of the
three-probe error separation method without harmonic sup-
pression. The general transfer function form BG(n) is not suit-
able for use in the optimization task due to the effect of the
denominator sin (β3  −  β2) in its expression. The reasons are
briefly described below. From section  2.1, the amplitude of
the transfer function B(n) can be represented as
Ä 2
|B (n)| = [sin (β3 − β2 ) − sin β3 cos nβ2 + sin β2 cos nβ3 ]
ä
2 1/2
+[sin β3 sin nβ2 − sin β2 sin nβ3 ] . (16)

Then, |BG(n)| can be written as
(17)
|BG (n)| = |B (n)|/|sin (β3 − β2 )|.

Figure 3. The X and Y components of radial error motion and the Firstly, due to the denominator sin (β3  −  β2), the values of
artefact roundness used in the error analyses. |BG(n)| increase rapidly when β3  −  β2 is close to 0° or 180°.
In terms of the optimization goal’s definition alone, it seems
that, under the transfer function form BG(n), the optimal
angles β2 and β3 are mainly located near the regions where
β3  −  β2  =  0° or 180°. This is contradictory to the practical
situation. Additionally, for a given arrangement of sensors, by
choosing each sensor as the reference sensor (β1  =  0), three
combinations of the measurement angles β2 and β3 (corre­
sponding to three values of sin (β3  −  β2)) can be obtained. The
values of |B(n)| for these combinations are the same. However,
the values of |BG(n)| would be different due to the influence of
sin (β3  −  β2). This is another problem when BG(n) is used in
the optimization, since the separated results should be iden-
tical for a specific arrangement of the three sensors, no matter
which one is chosen as the reference sensor. In conclusion,
the transfer function B(n) is better suited for the optimization
problem.
In order to determine the most suitable angles β2 and β3
that make |B(n)| at the required harmonic range as large as
possible, this paper attempts to simultaneously maximize
the minimum value of |B(n)| and the average value of |B(n)|.
Hence, this is a 2D, two-objective, global optimization
problem. The goal is to
T
Find β = (β2 , β3 )
Figure 4.  Measurement errors (Δr, Δx, and Δy ) on the to maximize f1 (β) = min |B (n)| n = 2, ..., NC
reconstructed roundness and error motions when γ1  =  0, γ2  =  0.1°, f2 (β) = mean |B (n)| n = 2, ..., NC
and γ3  =  −0.1°. Δr1, Δx1, and Δy 1 denote the negligible parts of
the measurement errors contributed only by radial error motion. where NC is the cutoff frequency, which can be empirically
determined according to practical situations. For solving this
Therefore, for high-precision spindle error motion meas- problem, the multi-objective problem needs to be transformed
urement, even when the error separation technique is used, into one with a single objective. First, the functions f j (β)
the accuracy of the selected artefact should also be better than (j   =  1, 2) are normalized:
that of the error motion. In addition, the restriction of angular
position errors and the appropriate choice of harmonic range fj (β) − fj min (β)
fj (β) =
(18) ,
are also very important for the multi-probe method. fj max (β) − fj min (β)
where f j max(β) and f j min(β) are the global maximum and min-
3.  Optimization of measurement angles imum of the function f j (β), respectively. Then, the general
objective function can be defined as
The optimization goal is to find a suitable arrangement of dis-
f  (β) = f1 (β) f2 (β) .
(19)
placement sensors that makes the amplitudes of the transfer

4
Meas. Sci. Technol. 30 (2019) 095001 Y Jiao et al

Figure 5.  Distribution of the objective functions f ′(β) for three different cutoff frequencies. The measurement angles located in red regions
are preferred.

Figure 7.  Transfer functions |B(n)| for three optimal arrangements

Figure 6.  Transfer functions |B(n)| for three optimal solutions of β2 and β3 obtained in step sizes of 1°, 0.1°, and 0.01°,
obtained by using step sizes 1°, 0.1°, and 0.01°, respectively, when respectively, when NC  =  50.
the cutoff frequency is NC  =  20.
Table 1.  Optimal measurement angles at several different cutoff
Thus, the two-objective problem is reduced to a single-objec- frequencies.
tive one, which can be formulated as
Step size 1° Step size 0.1°
maximize f  (β) NC β2 (°) β3 (°) β2 (°) β3 (°)
subject to 0 < β2  β3 − β2  360◦ − β3 .
◦
10 30 120 30 120
The boundary constraints are different from the constraints in 15 43 103 42.6 104.7
[8]. The reason for this is that the values of |B(n)| are the same 20 49 113 48.6 112.9
for a given set of probe angles regardless of which sensor is 30 52 112 51.3 111.3
chosen as the reference sensor. Therefore, the constraints used 50 61 127 52.4 111.4
here can help to avoid repeated solutions and to reduce com- 100 55 113 59.4 125.5
putational time. 150 52 118 63.2 128.6
In addition, the best measurement angles vary according to 200 113 234 111.8 232.4
the harmonic range (i.e. the selection of NC). Since amplitudes 300 59 126 59.8 123.1
of high-order frequency components of radial error motion are
usually small for most engineering applications, a relatively condition is already difficult to satisfy. Therefore, it seems
small cutoff frequency, such as NC  =  20, 50, or 100, may be unnecessary to calculate the optimal angles for NC  >  300. In
enough to reconstruct the error motion. Moreover, as dis- this study, optimal values for NC ranging from 10 to 300 are
cussed in section 2.2, for multi-probe error separation, higher calculated.
harmonics of reconstructed error motion are more sensitive to For this 2D optimization problem, a feasible method for
the unavoidable angular position errors. As an example, for the finding the optimum solution is to list the values of f ′(β) for
harmonic n  =  300, in order to make the contribution of R(n) to all combinations of β2 and β3 in a given step size. Figure 5
the measurement error ΔR(n) less than 20%, the angular posi- shows the distribution of f ′(β) for the angles β2 and β3  −  β2
tion errors need to be controlled within the range of about  ±2′ in a step size of 0.1°, with the cutoff frequency NC  =  20, 50,
(see equation  (13) and figure  2); however, in practice, this and 100, respectively. Red regions represent the favorable

5
Meas. Sci. Technol. 30 (2019) 095001 Y Jiao et al

Figure 8.  Transfer functions |B(n)| for three optimal arrangements obtained in steps of 1°, 0.1°, and 0.01°, respectively, when NC  =  100.

Figure 9.  Variation of optimal angles with cutoff frequency, in steps of (a) 1° and (b) 0.1°.

Figure 10.  Angular regions where the global optimal arrangements

for sensors S2 and S3, varying with the cutoff frequency, most
commonly appear, according to figure 9.

measurement angles, while blue regions indicate the bad solu-

tions under which harmonic suppression and/or noise amplifi-
cation may occur. From this figure it is clear that the objective
function is highly nonlinear, non-convex, and has many local
maxima. By comparing figures  5(a)–(c), it can also be seen
Figure 11.  Variation of min |B(n)| and mean |B(n)|, at optimal
that as NC increases, the nonlinearity of the objective func-
arrangements in figure 9, with the cutoff frequency NC, in step sizes
tion and therefore the number of local maxima also increase. 1° and 0.1°, respectively.
However, it should be noted that the objective function does
have a global optimum for β2 and β3, which can be directly using step sizes 1°, 0.1°, and 0.01°, respectively. Since integer
found from the distribution of f ′(β) by using the MATLAB angles are often used in practical applications, the global
program. optimal angles computed in step sizes of 1° should be useful.
For different cutoff frequencies NC, the global optima of In addition, as seen from these figures, the results based on
objective functions are different. As examples, figures  6–8 step sizes of 0.1° and 0.01° in each case are very close. Hence,
give the transfer functions |B(n)| under nine different global the optimal arrangements obtained based on step sizes of 1°
optimal arrangements when NC  =  20, 50, and 100, and by and 0.1° should be enough to satisfy the actual requirements.

6
Meas. Sci. Technol. 30 (2019) 095001 Y Jiao et al

Figure 9 plots the global optimal measurement angles for Acknowledgment

β2 and β3  −  β2 at cutoff frequency NC ranging from 10 to 300.
Figure 9(a) presents optimal solutions obtained in steps of 1°, This work is supported by the Science Challenge Project
while (b) presents results obtained in steps of 0.1°. Table 1 lists (Grant No. JCKY2016212A506-0105).
the global optima of β2 and β3 for several commonly used har-
monic ranges by using step sizes of 1° and 0.1°. Interestingly,
ORCID iDs
from figure 9 we find that these global optimal angles for sen-
sors S2 and S3 are mainly located in two regions (S2: 46°–64°, Yang Jiao https://orcid.org/0000-0003-3938-6812
S3: 107°–129°; and S2: 108°–114°, S3: 227°–234°), as shown
in figure 10. Moreover, as can be seen from figure 5, most of
the local optimal solutions are also located in these regions. References
Figure 11 shows the minimum and average values of |B(n)|
for global optimal arrangements in figure  9. It presents the [1] ISO 230-7 2015 Test Code for Machine Tools—Part 7:
performance of the two types of global optima obtained in Geometric Accuracy of Axes of Rotation (International
steps 1° and 0.1°, respectively. In general, the average values Organisation for Standardization)
[2] Grejda R D 2002 Use and calibration of ultraprecision axes
remain relatively stable, at around 1.4. However, the minimum of rotation with nanometer level of metrology PhD Thesis
values decrease with the increase in cutoff frequency, which Pennsylvania State University, USA
means that for the three-probe method the harmonic range to [3] Marsh E R, Couey J and Vallance R 2006 Nanometer-level
be selected must be balanced with the robustness of the meas- comparison of three spindle error motion separation
urement accuracy. techniques ASME J. Manuf. Sci. Eng. 128 180–7
[4] Shi S, Lin J, Wang X and Zhao M 2016 A hybrid three-probe
method for measuring the roundness error and the spindle
error Precis. Eng. 45 403–13
4. Conclusion [5] Hale L, Ummethala U and Hench J 2011 Multi-probe error
separation applied to roundness circular flatness and
This work explores the suitable measurement angles for the angularity Proc. ASPE Annual Conf. pp 476–9
[6] Marsh E R 2010 Precision Spindle Metrology 2nd edn
three-probe spindle error motion separation technique. For this (Lancaster, PA: DEStech Publications)
purpose, the influence of the sensors’ angular position errors [7] Moore D 1989 Design considerations in multiprobe roundness
is also discussed. The following conclusions are obtained: measurement J. Phys. E: Sci. Instrum. 22 339–43
[8] Cappa S, Reynaerts D and Al-Bender F 2014 A sub-nanometre
(1) High-order frequency components of reconstructed error spindle error motion separation technique Precis. Eng.
motion are more sensitive to angular position errors. 38 458–71
Roundness error is the main contributor to the measure- [9] Fujimaki K, Sase H and Mitsui K 2008 Effects of sensor noise
ment errors due to angular position errors. Positioning in digital signal processing of the three-point method Meas.
Sci. Technol. 19 015201
accuracies of sensors should be an important factor to be [10] Shu Q, Zhu M, Liu X B and Cheng H 2017 Radial error
considered in selecting the range of harmonics for error motion measurement of ultraprecision axes of rotation
motion reconstruction. with nanometer level precision ASME J. Manuf. Sci. Eng.
(2) There exist global optimal measurement angles for a 139 071017
given objective function with a specific harmonic range. [11] Estler W T, Evans C J and Shao L Z 1997 Uncertainty estimation
for multiposition form error metrology Precis. Eng. 21 72–82
The optimal solutions at several commonly used har- [12] Neugebauer M 2001 Uncertainty analysis for roundness
monic ranges shown in table  1 are useful for practical measurements by the example of measurements on a glass
applications of the multi-probe technique. hemisphere Meas. Sci. Technol. 12 68–76