11 голос за00 голосов против

14 просмотров8 стр.MATLAB - A Fundamental Tool for Scientific Computing and Engineering Applications - Volume 1

May 06, 2015

© © All Rights Reserved

PDF, TXT или читайте онлайн в Scribd

MATLAB - A Fundamental Tool for Scientific Computing and Engineering Applications - Volume 1

© All Rights Reserved

14 просмотров

11 голос за00 голосов против

MATLAB - A Fundamental Tool for Scientific Computing and Engineering Applications - Volume 1

© All Rights Reserved

Вы находитесь на странице: 1из 8

in surface metrology

Y.-B. Yuana,c,*, X.-F. Qianga, J.-F. Songb, T.V. Vorburgerb

a

National Institute of Standards and Technology (NIST), Gaithersburg, MD 20899 USA

c

Guest Researcher, NIST

Received 30 November 1998; received in revised form 11 August 1999; accepted 16 August 1999

Abstract

A fast recursive algorithm for determining the Gaussian filtered mean line was deduced using the central limit theorem and an

approximation method. This recursive algorithm uses a small number of multiplications per loop and otherwise such simple computer

operations as addition and subtraction, and therefore, can achieve a very high computational speed. Special cases are also presented in which

the relatively inefficient multiplication operation in the computer can be replaced by the efficient digit shifting operation, and the filtering

computational efficiency is enhanced further. High-order algorithms are proposed for practical use to improve filtering accuracy. The

forward filtering and backward filtering implementation of the recursive algorithm results in zero phase distortion of the filtered mean

line. A new relationship between the Gaussian filtering method and the classical 2RC filtering method is also established using this

algorithm. Published by Elsevier Science Inc.

Keywords: Surface roughness; Mean line; Digital filter; Recursive algorithm; Gaussian filter

1. Introduction

The mean line is the reference line used for surface

assessment. In surface metrology, the determination of the

mean line is a fundamental procedure for describing and

assessing surface characteristics. Almost all surface parameters are defined and calculated based on the mean-linesystem or M-system [1,2].

The least-squares method and the 2RC filter are the

traditional methods for determining mean lines [3]. The

former method produces mean lines that are sensitive to the

location of the sampling lengths on the surface profile and

are generally discontinuous from one cut-off length to the

next because of their being calculated piecewise for each

cut-off length within the entire evaluation length [3]. The

latter method distorts the true profile because of its nonlinear phase characteristic [3]. In the 1960s, Whitehouse [4,5]

worked on this problem and developed the concept and

Contribution of the National Institute of Standards and Technology,

not subject to copyright.

* Corresponding author. Tel.: 301-975-5356.

E-mail address: yibao.yuan@nist.gov (Y.-B. Yuan).

0141-6359/00/$ see front matter Published by Elsevier Science Inc.

PII: S 0 1 4 1 - 6 3 5 9 ( 9 9 ) 0 0 0 3 1 - 8

for zero-phase mean line filtering theory in surface metrology. In 1979, Raja and Radhakrishnan solved practical

problems in the implementation of the phase-corrected filter

using digital filtering techniques and designed several different types of digital filters used for processing surface

profiles [6].

The ISO 11562 standard proposed the Gaussian filtered

mean line as the reference line for profiles measured with

contact (stylus) instruments [7]. The Gaussian filter was

recognized as an optimal filter because of its zero-phase

characteristic and its minimum product of time width and

frequency width [8]. However, the low efficiency of the

algorithms for calculating the Gaussian filtered mean line

has been a problem for researchers wishing to use Gaussian

filters in their calculations.

The convolution integral method [9,10] is the most

widely used algorithm for calculating the Gaussian filtered

mean line, but it has low efficiency. In 1993, Luo et al. [11]

developed two- (2-D) and three-dimensional (3-D) Gaussian filtering algorithms using a Fast Fourier Transform

(FFT) algorithm under a project supported by the European

Community. In 1994, Vanherck [12] reported on an inter-

In that report, a triangle impulse response and a series of

rectangular impulse responses were used to simulate the

impulse response of the Gaussian filter to improve the

computational efficiency. These approximate methods resulted in high computational efficiency with a trade off

against the accuracy of the Gaussian filter. It was reported

[12] that the determination of the Gaussian-filtered mean

line, when using a PC 486/25 MHz computer over five

cut-off lengths with a total of only 2000 data points, took 22

seconds. In 1996, Krystek presented a fast Gaussian filtering

algorithm in which the symmetry of the Gaussian function

and a recursive algorithm for Gaussian filtering coefficients

were studied to reduce the amount of multiplications in

Gaussian filtering. This approach greatly speeds up the

convolution algorithm [13]. Using Krysteks method, 15

seconds were taken for the Gaussian filtering for a surface

profile of seven cut-off lengths with a total of 9333 data

points on a PC 486/33 MHz computer. The computational

efficiency of Krysteks method is comparable to that of the

ordinary FFT filter algorithm [13]. In 1996, Hara et al. [14]

used the second-order Butterworth high-pass filter to simulate the transmission characteristic of the Gaussian filtering

method. The amplitude characteristic of that filter is very

close to the Gaussian filter. However, the phase characteristic is a nonlinear function, resulting in phase distortion, as

a cost of enhancing the computational efficiency.

Based on the central limit theorem, approximation methods, and digital signal processing principles, we present a

new recursive filter algorithm for determining the Gaussian

filtered mean line. This approach can ensure both amplitude

accuracy and a zero-phase characteristic and can also realize

high computational efficiency.

63

or [Eq. (4)]

aa yi k ab xi r

N

yi

k1

r0

where ak, br, M and N are the constants of the digital filter.

From Eq. (3), it can be seen that the mathematical model

of the digital filter only deals with multiplication and addition operations. It is well known that the efficiency of

multiplication is much lower than that of addition. Therefore, the computational speed mostly depends upon the

number of multiplications performed once the mathematical

model of a digital filter is developed. As a result, reducing

or eliminating the multiplication operations is the key issue

for increasing the computational efficiency. This idea has

been successfully used for developing the FFT algorithm

[10] and is the approach by which we improve the efficiency

of the Gaussian filter.

2.2. Approximation of Gaussian function

The Gaussian function [Eq. (5)]

Gu e u

(5)

filter (FIR DF) [9,10]. However, to reduce the number of

multiplications, an infinite impulse response digital filter

(IIR DF) would be very useful for realizing Gaussian filtering. Therefore, Eq. (5) is rewritten as follows: [Eq. (6)]

Gu 1 u 2

1 4 1 6

u u

2

6

(6)

1

1 u2

(7)

Gu 1 u 2

2. Basic theory

(4)

function.

Given that the impulse response of a digital filter is h(i),

that the input and output digital signals are x(i) and y(i),

respectively, and that the frequency spectra obtained

through z-transform are H(ej), X(ej), Y(ej) correspondingly, then the input output relation of the digital filter can

be represented as [9,10] either [Eq. (1)]

yi xi*hi

(1)

Ye j Xe j He j

(2)

equation [9,10]; that is, [Eq. (3)]

a yi k b xi r

N

k0

r0

(3)

undergo convolution operations together, the results generally become smoother as the number of functions increases,

and the final result approaches a Gaussian function. The

strict description of this trend resulting from such a repeated

convolution is the central limit theorem [15].

Both the impulse response in the time domain and the

Fourier transform in the frequency domain of the Gaussian

filter are Gaussian functions. The transform of a self-convolution in the time domain is equal to a self-multiplication

in the frequency domain. Analogously, the inverse transform of a self-convolution in the frequency domain is equal

to a self-multiplication in the time domain. The results of

the self-convolution and the self-multiplication are also

Gaussian functions. That is, for multiplication: [Eq. (8)]

64

2

e u e u e u

e mu

m

2

(8)

If the function 1/(1 u2) is used as the approximation of

H(c/) in Eq. (13), then, we can define the function [Eq.

(14a)]

m1

2 u2/m

e u e u e u

e

m

m

2

(9)

2

emu can be approximated by (1 u2)m [Eq. (10)]

2

1 u 2 m 1 mu 2

1 mu 2

mm 1 4

u

12

m 2u 4

remainders

12

e mu remainders

2

(10)

very interesting that the Gaussian term in Eq. (10) becomes

narrower, and from the central limit theorem, the remainders become less significant.

2.4. Gaussian filter

The weighting function of the Gaussian filter can be

given as [2] [Eq. (11)]

ht

1 t/c2

e

c

(11)

is the cut-off wavelength (in the units of t), and is a

constant. If 0.4697, then when c, the transmission

ratio of the Gaussian filter is 50%.

The Fourier transform of the Gaussian function is also a

Gaussian function, and the amplitude characteristic of a

Gaussian filter can be obtained by means of the Fourier

transform of h(t), [Eq. (12)]

1

H

2

H 1 c/

1

1

1 / c 2

3

e f/fc

(14c)

1

1 / c 2

3

2CR filter and its corresponding 2RC filter, because their

amplitude transmission characteristics can be implemented

without any phase distortion in the form of phase-corrected

or zero-phase-shift digital filters [8,16,17]. If we let 1 3,

then Eq. (14a) is the standard amplitude transmission characteristic of a 2RC filter for calculating the mean line in

surface roughness measurements.

To ensure that the transmission ratio H1(c/), when

c, is equal to 50%, the value of 1 in Eq. (14a) is

constrained to be equal to 1. Now a series of Gaussian filter

approximations can be constructed by self-multiplication of

the function Eq. (14a) in the frequency domain with the

same constraint [Eqs. (15), (16), (17), (18), (19)]

1

H 2 c /

1 2 c / 2

H 4 c /

e /c

(14b)

so the amplitude transmission characteristic of its corresponding mean line filter with the form of a 2RC low-pass

filter can be calculated by this way [Eq. (14c)]

1

1 3 c / 2

1

1 4 c / 2

1

H 5 c /

1 5 c / 2

e c/

(12)

H c/ e c/

(13)

From Eqs. (12) and (13), it can be seen that this form of the

Gaussian filter is actually a low-pass filter.

(14a)

filter whose amplitude transmission characteristic is [2] [Eq.

(14b)]

H 3 c /

ht e jtdt

1

1 1 c/ 2

1

H 6 c /

1 6 c / 2

(15)

4

(16)

(17)

16

(18)

32

(19)

0.0905, 5 0.0443, 6 0.0219. The transmission characteristic deviations of these approximation filters from the

Gaussian filter are shown in Fig. 1.

65

Fig. 1. The transmission characteristic deviation of the theoretical approximation filters H2 H6 from the Gaussian filter.

the time domain and self-multiplication in the frequency

domain, the smaller the approximation error and the higher

the approximation accuracy. For example, it can be seen

that the approximation error is about 4% for H3(c/), but

is no more than 0.5% for H6(c/ ).

A 4 2

1

1 4/ c 2

H a4 j H a4j

H a4 j 2

Considering both the approximation accuracy and the

processing efficiency, the functions H4(c/) and H5(c/

) are selected as bases for the design of the Gaussian

digital filters. The accuracy of H5(c/) is higher than that

of H4(c/ ), but its efficiency is lower than the latter by

about a factor of 2, because it contains about twice as many

multiplications.

a conjugate pair of complex functions Ha4(j). A4(2) or

Ha4(j)2 is an amplitude-squared function.

According to the relationship between the Laplace transform and the Fourier transform, we let s j. Then [Eq.

(22)]

A 4s 2 H a4s H a4s

filter

The filter H4(c/) [see Eq. (17)] consists of a cascade

of eight A4(c/) filters, which we will call an 8-cascade

filter, where each one, A4(c/), is given by [Eq. (20)]

A 4 c /

1

1 4 c / 2

1

1 4/ c2 s 2

(22)

the pole in the left half s-plane 1/ 4 c, correspond

to Ha4(s). Then [Eq. (23)]

1

H a4s

(20)

(21)

4

s

(23)

c

66

digital filter equivalent to H4. By using a bilinear transformation [9,10], the analog filter Ha4(s) can be transformed

into the digital filter H4(z) [Eq. (24)]

H 4 z

ri ri 1 b xi xi 1

2ri 1

y mN t i y mN t i 1 b rN t i

rN t i 1 2y mN t i 1

10441 z 1

(24)

100N c 1044 100N c 1044 z 1

length c (or sampling length), H4(z) is an elementary cascade for the desired Gaussian filter, which is obtained by an

8-cascade of {H4(z) H4(z1)} assembled as a pair according to the zero-phase-shift filtering principle [8,16,17]. The

zero-phase-shift filtering principle is shown by the equation

[Eq. (25)]

H 4 z H 4 z 1 zej H 4e j 2 e j0

(25)

within a spatial wavelength , n (/c) Nc. If the

Gaussian filter consists of an 8-cascade of such zero-phaseshift filters, each of them is expressed in the spatial domain

by difference equations [Eqs. (26), (27)]

ri a ri 1 b xi xi 1

(30)

It can be seen that the number of multiplications in Eqs. (29)

and (30) are reduced to one-half of those used in Eqs. (26)

and (27). That means the computational efficiency is increased by nearly 2:1.

For the special cases when b 1044/(100Nc 1044)

2k the multiplication by 2k can be implemented as a digit

shift based on binary storage. Then, all multiplication operations left in the filter equations are replaced by bit shift

operations, so that the Gaussian filter algorithm achieves its

highest efficiency. The values of Nc that suit this condition

are [Eq. (31a)]

N c 1044 2 k 1044/100

(31a)

Eq. (31a), when [Eq. (31b)]

k 5, N c 324

(26)

k 6, N c 658

y mN t i a y mN t i 1 b rN t i

rN t i 1

(29)

k 7, N c 1326

(27)

(31b)

(32), (33)]

a 100N c 1044/100N c 1044

(28a)

b 1044/100N c 1044

(28b)

{x(i)} is the input of each filter;

{r(i)} is the medial result of each filter;

and {ym(i)} is the output of each filter.

For simplicity, the filtering initial state conditions may be

given as x(0) 0, r(0) 0, and ym(Nt) 0. Of course, other

values for these initial state quantities can also be selected,

because the end effects in a filtering process are unavoidable, no matter what the initial state quantities are. The end

effects converge to zero when the filtering algorithm is

stable. One strategy for dealing with them is to determine

first the spatial extent of the end effects, then throw the end

effects away. So the selection of the initial state quantities is

not important. After the eight-cascade zero-phase-shift filter, the center part of {ym(i)} becomes the Gaussian filtered

mean line for the measured surface profile. Eq. (26) is called

the forward filter, or the direct filter; whereas, Eq. (27)

is called the backward or correction filter.

Because a 1 2b, Eqs. (26) and (27) can be rewritten

as [Eqs. (29), (30)]

ri ri 1 2 k xi xi 1

2ri 1

(32)

y mN t i y mN t i 1 2 k rN t i

rN t i 1 2y mN t i 1

(33)

Eqs. (32) and (33) show that the Gaussian filtered mean line

can be calculated merely by using such simple operations as

addition, subtraction, and digital shifting.

3.2. Sixteen-cascade, zero-phase-shift, Gaussian mean line

filter

Higher-order approximations of Gaussian mean line filters [H5(c/), H6(c/), etc.] can also be designed using

the same methods described above. For example, the Gaussian filter from H5(c /) consists of a 16-fold cascade of

zero-phase-shift filters or a 32-fold cascade of elementary

filters. The elementary filter is [Eq. (34)]

H 5 z

14931 z 1

100N c 1493 100N c 1493 z 1

(34)

Each zero phase filter has the same general form as that

described in Eqs. (29) and (30). The only difference is [Eq.

(35a)]

b 1493/100N c 1493

(35a)

(35b)

same manner as that shown in Eqs. (32) and (33), but the Nc

values will be different. When k 5, Nc 463; when k

6, Nc 940.

3.3. Transmission characteristics of the practical

Gaussian filters

2

n

16

1044 2 1 cos

2

n

2

n

(36)

where n is the number of sampled points within a spatial

wavelength , n (/c) Nc.

The sixteen-cascade zero-phase-shift Gaussian filters

transmission characteristic is [Eq. (37)]

H e

j

2

n

32

1493 2 1 cos

2

n

2

n

16

(37)

If Nc is chosen to be large enough, for example, Nc 200,

it can be shown that both digital approximation filters above

approach the Gaussian filtering characteristic well. For any

value of Nc, not just the special values that eliminate multiplication operations, both IIR type digital filters are fast

enough for mean line filtering. For example, if we assume

that Nc 1600, a typical set-up condition for one of our

instruments, then [Eq. (38)]

H 4 e j Nc

16

49.98%

50.00%

(39)

deviations of the digital approximation filters [Eq. (40)]

H e

j

2

n

16

(40)

H e

j

2

n

32

(41)

with those shown in Fig. 1.

3.4. Example

transmission characteristic is [Eq. (36)]

H e

32

H 5 e j Nc

1493) 2k; that is [Eq. (35b)]

N c 1493 2 k 1493/100

67

(38)

consists of ten harmonic components with known frequencies ranging from 0.1/ c to 25/c. The detailed filter parameters are the cut-off length, c 0.8 mm, sampling

interval, 0.5 m, traversing length of the filtered profile,

7c 5.6 mm, and the total number of sampled data points,

Nt 11,200. The simulation results show that the mean

lines of H416, H532 and theoretical Gaussian filter calculated directly from the ten sinusoidal harmonic components

are almost identical, except for the two ends of the profile.

The small differences among these mean lines and their

amplified versions are shown in Fig. 3(b) and (c), respectively. They result from the transmission characteristic deviations (Fig. 2) of the two approximation filters from the

Gaussian filter. The end effect or transition for filtering is

unavoidable for all types of filters. The shapes of the end

effects are determined by their impulse responses and initial

states. However, the width of the end effect is mainly

determined by the impulse response of a filter. For Gaussian

filtering, the span of the end effect is about half of the

cut-off length; that is, c/2. Conventionally, we measure

unfiltered profile data with a traversing length of 7c and

use the 5c profile data in the center of the 7c profile for

surface assessment. Conversely, we omit the 2c profile

data distorted at two ends because of the end effect of the

filtering process. This approach can ensure the surface assessment accuracy. In this example, because Nc 1600, the

bit shifting approach is not used.

4. Summary

The computational efficiency for determining the Gaussian filtered mean line is mainly determined by four factors:

the complexity of the algorithm itself, the speed of the

computer, the quantity of data to be processed, and the

program language and programming technique. Of all these

factors, the first is a key issue for improving the computational efficiency.

68

Fig. 2. The transmission characteristic deviations of the digital approximation filters H416 and H532 from the Gaussian filter.

Based on a simple approximation function and the central limit theorem, the complicated method for calculating

the Gaussian filtered mean line is considerably simplified.

The new recursive algorithms are accurate and are also fast,

a small amount of addition, subtraction, and multiplication

or digital shifting operations without phase distortion. The

approximation method demonstrates a relationship between

Fig. 3. (a) Surface profile and its filtered mean lines; (b) the mean line deviations of the practical approximation filters from the Gaussian filter; (c) the

amplified mean line deviations of (b).

the 2RC filtered mean line and Gaussian filtered mean line.

The amplitude transmission characteristic of a Gaussian

filtered mean line is well approximated by those of an 8RC

zero-phase-shift filtered mean line or a 16RC zero-phaseshift filtered mean line.

When using the eight-cascade, zero-phase-shift Gaussian

filter to process a 7-cut-off-length surface profile containing

11,200 sampled data points, only 0.66 seconds are required

using a PC 486/33 MHz computer in C language. For the

16-cascade zero-phase-shift Gaussian filter and 32-cascade

zero-phase-shift Gaussian filter, only 1.32 seconds and 2.64

seconds are taken for filtering under the same constraints,

respectively. For the same conditions, 18 seconds are estimated for Krysteks fast Gaussian filtering algorithm [13]

from the calculation, (11,200/9333) 15 s 18 s, where

9333 is the number of filtered data points, and 15 s is the

filtering time in Krysteks example. Because, the computational efficiency of Krysteks method is comparable to an

existing fast convolution method based on FFT techniques

[13], we conclude that the current approach is considerably

faster than that approach, as well.

The fast recursive algorithm described in this paper can

also be used in 3-D surface topography for determining the

Gaussian filtered mean plane, which can improve the computational efficiency even more. The algorithm for determining the 3-D Gaussian filtered mean plane will be discussed elsewhere.

References

[1] ISO 4287-1997, Geometrical product specifications(GPS)Surface

texture: Profile methodTerms, definitions and surface texture parameters,International Organization for Standardization, Geneva.

69

[2] ASME B46.1. Surface texture: Surface roughness, waviness, and lay.

New York: American Society of Mechanical Engineers, 1995.

[3] Whitehouse DJ, Reason RE. The Equation of the Mean Line of

Surface Texture Found by an Electric Wave Filter. Leicester, UK:

Rank Taylor Hobson, 1965: pp. 137.

[4] Whitehouse DJ. The Mean Line Found by a Phase-Corrected Wave

Filter. Leicester, UK: Rank Taylor Hobson, 1965: pp. ivii.

[5] Whitehouse DJ. Survey of reference lines in the assessment of surface

texture. Ann CIRP 1972;21(2): 26773.

[6] Raja J, Radhakrishnan V. Digital filtering of surface profiles. Wear

1979;57: 14755.

[7] ISO 11562-1996, Geometrical Product Specifications(GPS)Surface

Texture: Profile MethodMetrological Characteristics of Phase Correct Filters. International Organization for Standardization, Geneva.

[8] Yuan YB. Research on some fundamental theories of surface topography measurement. Ph.D. thesis, Harbin Institute of Technology,

Harbin, Peoples Republic of China, 1997.

[9] Hamming RW. Digital Filters, 3rd ed. Upper Saddle River, NJ:

Prentice-Hall, 1988.

[10] Shu Hu An. Digital Signal ProcessingTheory, Algorithm, and Realization. Beijing, Peoples Republic of China: Tsinghua University

Press, 1997.

[11] Luo NL, Sullivan PJ, Stout KJ. Gaussian filtering of three-dimensional engineering surface topography. Proc SPIE 1993;2101:52738.

[12] Vanherck P. STCs cooperative work on phase correct Gaussian

filtering in surface roughness measurement. Ann CIRP 1994;43(2):

599 601.

[13] Krystek M. A fast Gauss filtering algorithm for roughness measurements. Prec Eng 1996;19:198 200.

[14] Hara S, Tsukada T, Ito M. The conditions to apply the phase corrected

filter to traced profiles for roughness profiles (1st rept)Application

of the Gaussian high-pass digital filter. J Japan Soc Prec Eng 1996;

62(4):594 98.

[15] Weaver HJ. Applications of Discrete and Continuous Fourier Analysis. Malabar, FL: Krieger Publishing Company, 1992.

[16] Qiang XF, Yuan YB. An optimal digital filter model used in surface

roughness assessment. J Astronaut Metrol Measure 1988;6:1 6.

[17] Qiang XF, Yuan YB. A fast algorithm for finding the mean line in

surface roughness assessment. Chinese J Sci Instrum 1994;15(4):

35358.

## Гораздо больше, чем просто документы.

Откройте для себя все, что может предложить Scribd, включая книги и аудиокниги от крупных издательств.

Отменить можно в любой момент.