Вы находитесь на странице: 1из 8

See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/281289645

Uncertainty estimation in form error


evaluation of freeform surfaces for precision
metrology
ARTICLE in PROCEEDINGS OF SPIE - THE INTERNATIONAL SOCIETY FOR OPTICAL ENGINEERING AUGUST 2015
Impact Factor: 0.2

READS

39

5 AUTHORS, INCLUDING:
Xiangchao Zhang

Hao Zhang

Fudan University

Ordnance Engineering College

36 PUBLICATIONS 78 CITATIONS

448 PUBLICATIONS 5,626 CITATIONS

SEE PROFILE

SEE PROFILE

Xiaoying He
Fudan University
49 PUBLICATIONS 294 CITATIONS
SEE PROFILE

Available from: Xiangchao Zhang


Retrieved on: 29 December 2015

Uncertainty estimation in form error evaluation of freeform surfaces


for precision metrology
Xiangchao Zhang1*, Hong Xiao2, Hao Zhang1, Xiaoying He1 and Min Xu1
1. Shanghai Engineering Research Centre of Ultra-Precision Optical Manufacturing,
Fudan University, Shanghai, 200438, P. R. China
2. Laboratory of Precision Manufacturing Technology, China Academy of Engineering Physics,
Mianyang, 621900, P.R. China
ABSTRACT
Freeform surfaces are widely applied in precision components to realize novel functionalities. In order to evaluate the
form qualities of the manufactured freeform parts, matching/fitting is required. The uncertainty of the obtained form
error measures needs to be estimated so as to assess the reliability of the evaluating process. The conventional GUM
approach is not suited for complex nonlinear models. In this paper we develop a Monte-Carlo method to estimate the
uncertainty of the fitted position, shape and form error parameters. Based on the correlation analysis, the effects of some
major factors, including objective functions, noise amplitudes, surface shapes etc can be determined, and subsequently
the significant factors influencing the evaluated form errors can be specified. By appropriate planning of the measuring
and matching procedures, the uncertainty of the evaluation results can be effectively reduced, and thereby improving the
reliability of freeform surface characterization.
Keywords: Precision metrology, freeform surface, uncertainty, Monte-Carlo method

1. INTRODUCTION
With the rapid development of advanced design and manufacturing technologies, freeform components are increasingly
applied in modern opto-mechanical systems, because of their compact sizes, small weights, flexibility in
design/utilization and some attractive capabilities of system integration, realizing various novel functions and remedying
the drawbacks of traditional components. These freeform components realize the intended functionalities by their forms;
thereby the evaluation of the form errors with respect to their nominal shapes has become a central task of freeform
surface metrology.
In the ISO standards and commercial precision instruments, various fitting objectives and algorithms have been
developed. In ISO 5459, the objective functions of least squares, minimum zone, one-sided Chebyschev, maximum
inscribed, minimum circumscribed and the Lp-norm are defined [1, 2]. As for the freeform surfaces, the L1 norm, the
least squares and the minimum zone fitting are commonly adopted in accordance with the specific properties and
applications of the measured data,
N

L1 norm

min d i
i =1
N

least squares

min d i2

(1)

i =1

minimum zone

min max d i
i

In the equation, di denotes the signed deviation of data point pi with respect to the nominal shape and N is the number of
data points. Among these three criteria, the least squares approach is most widely used due to its ease of computation and
unbiasedness for the normally distributed noise, but bias will be caused in the obtained form error metrics. The solutions
of the L1 norm and the minimum zone fitting are unbiased for long-tailed and uniformly distributed errors, respectively.
In addition, the minimum zone approach can obtain the smallest parameter of form deviation, complying with the
*

Corresponding author. Tel.: +86-21-51630347. Fax.: +86-21-65641344. Email: zxchao@fudan.edu.cn.

definitions in the ISO 1101 [3]. However, these two objectives are non-differentiable, and thus very difficult to be solved,
especially for the freeform surfaces with complex mathematical representations. Recently some algorithms have been
proposed to solve the L1 norm and minimum zone fitting/matching problem [4-7]. Researchers mainly focused on the
form error parameters obtained, but the reliability of the solutions, i.e. the rotational angles, translation vectors, axis
orientations, shape parameters etc were not examined rigorously.
In order to verify the reliability and stability of the fitting algorithms, the uncertainty of the solutions have to be
estimated. The approaches to the uncertainty estimation have been investigated by many groups and scholars. Some of
the typical principles or methods are provided by Evaluation of measurement data-Guide to the expression of
uncertainty in measurement (GUM) [8] and Monte Carlo (MC) method [9]. The GUM technique can be employed to
assess simple processes, but it is impractical or impossible for the nonlinear minimum zone fitting of freeform surfaces.
Some researchers estimated the uncertainty of straightness and flatness calculation using MC method [10, 11]. The MC
method is a sampling technique that provides rich information by propagating the distributions for the input quantities.

2. ESTIMATION PREEDURE OF FITTING UNCERTAINTY


The purpose of uncertainty estimation for surface fitting is to identify the significant factors influencing the fitting
quality, and assess the reliability of the fitted results, so that the performance of the utilized fitting algorithms can be
validated.
As for the fitted results, the obtained rotation angles, translation components and shape parameters are stored. Three
form error parameters are used to evaluate the relative deviations between the fitted surface and the input noisy data: the
arithmetic average (AA), the root-mean-squares (RMS) and the peak-to-valley (PV),

AA
RMS
PV

1 N
di
N i =1
1 N 2
di
N i =1
max d i min d i

(2)

It is known that the objective function of surface fitting should be consistent with the form error parameter to be
calculated. In surface metrology the form deviations are usually assessed along the z direction only. It is known the
solutions of such algebraic fitting are unstable and sensitive to measurement noise [12]. Moreover, the obtained error
parameters are prone to be biased and over-estimated when the surfaces are highly curved [13]. As a result we adopt
orthogonal distance fitting here, i.e. the form deviations are calculated along the normal vectors of the nominal surfaces.
To assess the influence of the different configurations, including the objective functions, noise amplitudes and surface
shapes, the fitting program is run repetitively for M times, with random noise added onto the sampled data. To make the
calculation results faithfully reflect the properties of the fitting algorithms, it is important to ensure that no bias is
introduced by the added noise, i.e. the mean shapes and positions of the noisy data should always be identical with the
reference datums.
N

shape preserving

=0

i =1
N

rest in x direction

=0

i =1
N

rest in y direction

(3)

yi = 0
i =1
N

rest in z direction

=0

i =1

Here xi , yi and zi are the x, y and z components of the form deviation di, respectively.
In practice, outliers and defects are usually removed in advance, hence they are not considered here.

Then the distribution of the parameters of interest is analyzed. The mean, standard deviation, skewness and kurtosis are
calculated respectively to assess the bias, dispersion, symmetry and peakedness of the distribution of each calculated
parameter,

ME =

mean
standard deviation
skewness
kurtosis

1
N

i =1

1 N
( xi ME )2
N 1 i =1
N
1
SK =
( xi ME ) 3

( N 1) SD 3 i =1
N
1
KU =
( xi ME )4
( N 1) SD 4 i =1
SD =

(4)

The implementing procedure of Monte Carlo estimation of the fitting uncertainty is presented in Figure 1.
change
configuration

generate
surface data
M times

add random
noise
surface fitting

analyze
distribution
establish
correlation
Figure 1. Procedure of Monte-Carlo estimation of fitting uncertainty

3. UNCERTAINTY ESTIMATION OF BICONIC SURFACE FITTING


The representations of freeform surfaces are very complex. To make the estimation results typical and representative, a
biconic surface is adopted [4], as shown in Figure 2.
z=

x 2 / Rx + y 2 / R y
1 + 1 (1 + k x ) x 2 / Rx2 (1 + k y ) y 2 / Ry2

+ a2 i x 2 i + b2 j y 2 j
i =2

(5)

j =2

In the equation, Rx=50 mm and Ry=60 mm are the radii of curvature and cx= -0.3 and cy=0.5 are conic factors. The
coefficients of polynomial terms are set as a4=3.5e-6, b4=-2.8e-6, a6=-1.9e-10 and b6=6.3e-9. A set of 200 200 points
are taken as measured data.

Figure 2. Biconic surface

3.1 Effect of fitting objectives


Three fitting criteria the L1 norm, least squares (LS) and minimum zone (MZ) are compared. A surrogate function
approach based on the majorize-minimize theory [4], the fast Levenberg-Marquardt algorithm [14] and the differential
evolution method [7] are adopted to solve these three optimization problems, respectively. It is worth noting that the
obtained shape parameters and form errors of the orthogonal distance fitting are irrelevant to the surface orientation;
therefore no translation/rotation is implemented on the data set before fitting.
Random Gaussian noise with standard deviation of 0.6 m is superimposed on the sampled data points, so as to make the
data uncertainty consistent with the Carl Zeiss Prismo Ultra CMM. The programs are implemented repetitively for 300
times, and the distributions of the obtained rotation angles, translation components and shape parameters are presented.
It can be seen that the L1 norm and LS fitting can always obtain correct positions. The relative rotation angles are within
10-3 . The lateral and vertical translation errors are less than 0.9m and 80 nm, respectively. This is due to the symmetry
of the data sets and the added random noise, thus no misalignment will be caused in surface fitting. But the MZ fitting
results depend on several feature points only, henceforth the variations of these extreme points can make the fitted
surface shifted or tilt. The SD of the translation components Tx, Ty and Tz and rotation angles x, y and z are 66.9m,
25.0m, 8.0m, 0.033, 0.088 and 0.043, respectively.
As for the fitted shape parameters, the L1 and LS method behave almost the same. In addition, the shape parameters
associated with x behaves similarly with those associated with y, e.g. Rx and Ry, kx and ky, and so forth. As a result only
the x shape parameters are presented in Tables 1 and 2. For the sake of clarity, their units are omitted.
All the fitted shape and form error parameters obey symmetric Gaussian distribution, expect for kx obtained by the MZ
fitting. The uncertainties of the MZ shape parameters are much greater than the LS fitting. However, the obtained form
error parameters have similar uncertainties. The obtained AA and RMS values are equal, but the PV parameters of MZ
fitting is smaller than the LS values by 47%, implying the LS method tends to over-estimate the form deviation if
measured with a zone width.
parameters

ME

SD

SK

KU

Rx

50.0000

0.0008

-0.0632

3.2891

kx

-0.30000

0.0003

-0.0033

2.7053

a4

3.50e-6

5.96e-10

-0.1692

3.0918

a6

-1.90e-10

4.94e-13

-0.2099

3.4022

AA

4.78e-4

2.84e-6

0.0959

3.0826

RMS

6.00e-4

3.23e-6

0.1067

3.4439

PV

0.0047

2.51e-4

0.6251

3.3067

Table 1. Distributions of the fitted parameters of least squares

parameters

ME

SD

SK

KU

Rx

50.0111

0.2752

-0.0191

3.1276

kx

-0.3033

0.0592

0.1645

2.1946

a4

3.51e-6

1.6420e-7

-0.0277

2.8722

a6

-1.95e-10

8.64e-11

0.0786

2.3807

AA

4.79e-4

2.85e-6

0.0826

3.0216

RMS

6.00e-4

3.24e-6

0.1094

3.4112

PV

0.0025

1.80e-4

1.1379

5.4346

Table 2. Distributions of the fitted parameters of minimum zone

3.2 Effect of shapes


The curvedness of the measured surfaces severely influences the reliability of the fitted results [12]. For the sake of
expression clarity, the surface parameters remain unchanged, while the width of the data set in the x direction is
decreased. The ratio between the x and y widths are set to be 0.2, 0.4, 0.6, 0.8 and 1 respectively, and the MZ fitting with
differential evolution is implemented 200 times for each case. Gaussian random noise with standard deviation of 0.6m
is added into the data.
It is found that the bias and uncertainties of the motion parameters are not affected by the width ratio. As the data sets are
symmetric about the rotational axis and the added noise has zero mean, hence the program can always obtain the correct
positions for the fitted surfaces. In addition, the obtained form error parameters AA, RMS, and PV are not apparently
affected by the width ratio either.
As for the fitted shape parameters, we define two factors: the relative bias s / s0 1 100% and relative uncertainty

( s ) / s0 100% . Here s and ( s ) are the mean and standard deviation of the fitted shape parameter s, and s0 is its
ideal value.
2

10

10

Rx
kx
a4
a6

10
0

10

10

Ry
ky
b4
b6

10

-2

10

Rx
kx
a4
a6

-4

10

0.2

0.4

0.6
(a) bias

Ry
ky
b4
b6
0.8

10

-1

10

0.2

0.4

0.6

0.8

(b) uncertainty

Figure 3. Relative bias and uncertainty at different width ratios (%)

The x shape parameters Rx, kx, a4 and a6 are increasingly biased when the data set narrows down in the x direction. The
uncertainties are getting worse simultaneously. This is straightforward to understand. When the surface is narrowed, the
form is flattened, as a result the data correspond to only a small section on the original surface. Slight variation of data in
the z direction can cause remarkable change to the fitted shape parameters. This is consistent with the conclusions of [12].
Interestingly, the y parameters Ry, ky, b4 and b6 show a different trend. When the x coordinates are relatively small, the

coefficients associated with y dominate in the Hessian and Jacobian matrices. As a result the observation matrix will be
less affected by the x terms, and the numerical conditions of the y parameters can in turn be improved.
This finding implies the numerical condition of the observation matrix is critical to the performance of the fitting
algorithms of complex shapes. When the program is unsatisfactory, there are two feasible approaches to improving the
numerically stability of the fitting algorithms. One is subdividing the unknown parameters into groups, and updating
each group alternatively. It is common to have three groups in practice, the six motion parameters (rotation angles and
translation components), the base radii and conic factors, and the polynomial coefficients. The other technique is scaling.
Since the polynomials xk or yk are normally the greatest in the observation matrix, thus these terms determine the
numerical stability of the whole system. It is suggested to scale down the whole data set so as to make the greatest x and
y coordinates into 1. Then the values of the elements in the Hessian and Jacobian matrices can be balanced, thereafter
improving their numerical stability. The fitted shape parameters should be scaled back accordingly after the program
terminates.

4. CONCLUSIONS
For complex freeform surfaces, the objective functions, noise amplitudes, surface shapes, and configuration parameters
of the fitting algorithms severely affect the reliability of the fitted results. By theoretical analysis and numerical
experiments it is found that the solutions, including the motion and shape parameters of the L1 and least squares are
determined by the mean of all the data points, thus they are very stable. On the contrary, the solutions of minimum zone
fitting are determined by only several key points, hence the fitted surface may be severely shifted or tilt. If the tolerance
width is of concern, the minimum zone fitting can obtain much better deviation values, with similar uncertainties with
the L1 norm and least squares fitting. The curvedness of the surface is critical to the numerical stability of the program. It
is feasible to subdivide the unknown parameters or scale down the data sets in practice to improve the numerical
conditions of the observation matrices, so as to control the bias and uncertainty in the solutions of complex surface fitting.

ACKNOWLEDGEMENT
The authors gratefully acknowledge the financial support of National Natural Science Foundation of China (51205064),
Shanghai Natural Science Foundation (12ZR1441100), Specialized Research Fund for the Doctoral Program of Higher
Education (20120071120018) and Open Fund of Key Laboratory of Precision Manufacturing Technology, CAEP
(KF14004).

REFERENCES
1. Zhang, X., Freeform surface fitting for precision coordinate metrology, PhD thesis, University of Huddersfield,
UK (2009).
2. ISO 5459 Geometrical product specifications (GPS) -- Geometrical tolerancing -- Datums and datum systems,
(2011).
3. ISO 1101 Geometrical product specifications-geometrical tolerancing-Tolerances of form, orientation, location and
run-out, (2012).
4. Zhang, X., Jiang, X. and Scott, P.J., Orthogonal distance fitting of precision freeform surfaces based on L1 norm,
in Pavese, F et al, [Advanced Mathematical and Computational Tools in Metrology and Testing], World Scientific, 385390 (2009).
5. Gosavi, A. and Cudney E. A., Form errors in precision metrology: a survey of measurement techniques, Quality
Engineering 24(3), 369-380(2012).
6. Liu, J., Calculation of profile error for complex surface, Measurement 48, 183-186 (2014).
7. Zhang, X., Jiang, X. and Scott, P.J., Minimum zone evaluation of the form errors of quadric surfaces, Precision
Engineering 35,383-389 (2011).
8. JCGM 100, Evaluation of measurement data-Guide to the expression of uncertainty in measurement, (2008).
9. JCGM 101, Evaluation of measurement data-Supplement 1 to the Guide to the expression of uncertainty in
measurement-Propagation of distributions using a Monte Carlo method, (2008).
10. Wen, X., Xu, Y., Li, H., et al. Monte Carlo method for the uncertainty evaluation of spatial straightness based on
new generation geometrical product specification, Chinese Journal of Mechanical Engineering 25(5), 875-881 (2012).

11. Calvo, R., Gomez, E. and Domingo, R., Vectorial method of minimum zone tolerance for flatness, straightness,
and their uncertainty estimation, International Journal of Precision Engineering and Manufacturing 15(1), 31-44 (2014).
12. Sun, W., McBride, J.W. and Hill, M., A new approach to characterizing aspheric surfaces, Precision Engineering
34, 171-179 (2010).
13. Zhang, X., Zhang, H., He, X. and Xu, M., Bias in parameter estimation of form errors, Surface Topography:
Metrology and Properties2(3),035006 (2014).
14. Boggs, P.T., Byrd, R.H. and Schabel, R.B., A stable and efficient algorithm for nonlinear orthogonal distance
regression, SIAM J Stat Comput. 8(6), 1052-1078 (1987).

Вам также может понравиться