没有合适的资源?快使用搜索试试~ 我知道了~
首页一种经典的时域滤波器 。
一种经典的时域滤波器 。
需积分: 0 16 下载量 181 浏览量
更新于2023-05-21
评论
收藏 1.17MB PDF 举报
1维数据平滑滤波相关的论文,引用率很高。非常实用。 应用场合,针对mcu采样ADC数据进行平滑滤波(时域)
资源详情
资源评论
资源推荐
Interferences.
Methylene chloride
or
chloroform-soluble carboxylic acids
(acetylsalicylic acid, salicylic acid,
etc.) which inhibited 4-,UP response
were readily extracted from these
phases using
0.ln’
sodium hydroxide.
Span-type excipients, which gave
a
slight response
to
the 4-AAP reagent,
were eliminated by virtue of their insol-
ubility in
10%
aqueous sodium chloride;
while interfering Tween-type excipients
were precipitated
(3)
using the Tween
reagent. Propylene glycol, when formu-
lated at
a
level of
5.25%,
contributed ca.
2 to
370
to the observed absorbance.
This was minimized by
a
more favorable
distribution of the interference within
the aqueous phases used in the pro-
cedure.
Placebo analyses indicated that no
interference was obtained with such
common excipients
as
stearic acid,
stearyl alcohol, cetyl alcohol, petrola-
tum, methyl or propyl-p-hydroxy-
benzoates, sesame oil, or thimerosal.
Lipotropic agents (betaine or choline),
all the common vitamins, and neomycin
sulfate failed to interfere.
Scope
of
Reaction.
In addition to
the other types of steroids reported
to react with the 4-AAP reagent, the
following steroids reacted at elevated
temperature (boiling point of methanol)
and at increased concentration (2.0 mg.
per
10.0
ml. of reagent): 2a-hydroxy-
methyl
-
17p
-
hydroxy
-
17a
-
methyl-
5a-androst-%one (306 mp)
;
and 2a-
hydroxymethyl
-
17p
-
hydroxy
-
501-
androst-3-one (306 mp). Steroids with-
out
a
keto group failed to give any
response. The quantitative aspects of
the above responses to the 4-AAP
reagent were not investigated.
Correlations
of
Chromophore Wave-
length with Structure.
The additive
effect of various ring
A
and
B
sub-
stituents on the chromophore of the
parent saturated-3-keto steroid is
presented in Table
11.
LITERATURE CITED
(1)
Cohen,
H.,
Bates,
R.
W.,
J.
Am.
Pharm.
Assoc.,
Sci.
Ed.
40, 35 (1951).
(2) Hiittenrauch,
R.,
Z.
Physiol. Chem.
326,166 (1961).
(3) Pitter, P.,
Chem.
&
Ind.
42, 1832
(1962).
RECEIVED
for
review September 4, 1963.’
Accepted March 13,
1964.
Smoothing and Differentiation
of
Data
by
Simplified Least Squares Pro,cedures
ABRAHAM SAVITZKY and MARCEL
J.
E.
GOLAY
The Perkin-Elmer Corp., Norwalk, Conn.
b
In attempting to analyze, on
dig i ta
I
computers, data
f
rom basica
II
y
continuous physical experiments,
numerical methods of performing fa-
miliar operations must be developed.
The operations of differentiation and
filtering are especially important both
as an end in themselves, and as a pre-
lude to further treatment of the data.
Numerical counterparts
of
analog de-
vices that perform these operations,
such as RC filters, are often considered.
However, the method of least squares
may be used without additional com-
putational complexity and with con-
siderable improvement in the informo-
tion obtained. The least squares cal-
culations may be carried out in the
computer by convolution of the data
points with properly chosen sets
of
integers. These sets of integers and
their normalizing factors are described
and their use
is
illustrated in spectro-
scopic applications. The computer
programs required are relatively sim-
ple. Two examples are presented as
subroutines in the FORTRAN language.
HE
PRIMARY
OUTPUT
of any experi-
Tment in which quantitative
information is to be extracted is infor-
mation which measures the phenomenon
under observation. Superimposed upon
and indistinguishable from this informa-
tion are random errors which, regardless
of their source, are characteristically
described as noise. Of fundamental
importance
to
the esperimenter is the
removal of as much of this noise
as
possible without,
at
the same time,
unduly degrading the underlying in-
formation.
In much experimental work, the infor-
mation may be obtained in the form of
a
two-column table of numbers,
A
us.
B.
Such a table is typically the result
of
digitizing
a
spectrum
or
digitizing other
kinds of results obtained during the
course of an experiment.
If
plotted,
this table of numbers would give the
familiar graphs of
TOT
us.
wavelength,
pH
us.
volume of titrant, polarographic
current
us.
applied voltage,
SMR
or
ESR
spectrum,
or
chromatographic
elution curve, etc. This paper is con-
cerned with computational methods for
the removal of the random noise from
such information, and with the simple
evaluation of the first few derivatives
of the information with respect to the
graph abscissa.
The bases for the methods
to
be dis-
cussed have been reported previously,
mostly in the mathematical literature
(4,
6,
8,
9).
The objective here is to
present specific methods for handling
current problems in the processing of
such tables of analytical data. The
methods apply as well to the desk
calculator,
or
to simple paper and pencil
operations for small amounts of data, as
they do to the digital computer for
large amounts of data, since their major
utility is to simplify and speed up the
processing of data.
There are two important restrictions
on the way in which the points in the
table may be obtained. First, the points
must be at
a
fixed, uniform interval in
the chosen abscissa.
If
the independent
variable is time, as in chromatography
or
NMR
spectra with linear time sweep,
each data point must be obtained at the
same time interval from each preceding
point.
If
it
is
a
spectrum, the intervals
may be every drum division
or
every
0.1
wavenumber, etc. Second, the
curves formed by graphing the points
must be continuous and more
or
less
smooth-as in the various examples
listed above.
ALTERNATIVE METHODS
One of the simplest ways to smooth
fluctuating data is by a moving average.
In this procedure one takes
a
fised
number of points, adds their ordinates
together, and divides by the number
of
points to obtain the average ordinate at
the center abscissa of the group. Sest,
the point at one end of the group is
dropped, the next point at the other end
added, and the process is repeated.
Figure
1
illustrates how the moving
average might be obtained. While
there is a much simpler way to compute
the moving average than the particular
one described, the following description
is correct and can be extended to more
sophisticated methods as will be seen
shortly. This description is based on
the concept of a convolute and of a
convolution function. The set of
numbers at the right are the data or
ordinate values, those at the left, the
abscissa information. The outlined
VOL.
36,
NO.
8,
JULY
1964
1627
1800.0 705
1799.8
712
1799.6
7 17
1799.4
718
1799.2
721
1799.0
------/-
-<I
722
1798.8
1
x~-2
‘12
’
725
1798.6
I
xo-l
c-l
I
730
1798.4
1
xo
co
I
735
1798.2
xo+l
C1
1
736
1798.0 x0+2
c2
I
741
,
A
,
XO
11
1797.8
,
r
746
-*--*--‘e
17637.6 750
...4
..
xo
Figure
1.
Convolution operation
Abscissa points at left, tabular data at right.
In box area the convolution integers,
cz.
Opera-
tion
is
the multiplication
of
the data points by the
corresponding
C,,
summation of the resulting
products, and division by
a
normalizer, resulting
in
a
single convolute at the point
Xo.
The
box
is then moved down one line, and the process
repeated
block in the center may be considered
to be
a
separate piece of paper on which
are written a new set of abscissa
numbers, ranging from
-2
thru zero to
+2.
The
C’s
at the right represent the
convoluting integers.
For
the moving
average each
C
is numerically equal to
one. To perform
a
convolution of the
ordinate numbers in the table of data
with a set of convoluting integers,
C,,
each number in the block is multiplied
by the corresponding number in the
table of data, the resulting products are
added and this sum is divided by five.
The set of ones is the convoluting
function, and the number by which we
divide, in this case,
5,
is the normalizing
factor. To get the next point in the
moving average, the center block is slid
down one line and the process repeated.
The concept of convolution can
be
generalized beyond the simple moving
average. In the general case the
C’s
represent any set
of
convoluting
integers. There is an associated
normalizing
or
scaling factor. The pro-
cedure is to multiply
C--?
times the
number opposite it, then
C-1
by its
number, etc., sum the results, divide by
the normalizing factor, if appropriate,
and the result is the desired function
evaluated at the point indicated by
Co.
For the next point, we move the set
of
convoluting integers down and repeat,
etc. The mathematical description
of
this process is:
The index
j
represents the running
index of the ordinate data in the original
data table.
For
the moving average, each
C,
is
equal to one and
N
is the number of
convoluting integers. However, for
many types of data the set of all
l’s,
which yields the average, is not
particularly useful. For example, on
going through
a
sharp peak, the average
would tend to degrade the end of the
peak. There are other types of smooth-
ing functions which might be used, and
a
few of these are indicated in Figure
2.
Figure
2A
illustrates the set where all
values have the same weight over the
interval-essentially the moving
average.
The function in Figure
2B
is an
exponential set which simulates the
familiar
RC
analog time constant-
Le., the most recent point is given the
greatest weight, and each preceding
point gets
a
lesser weight determined by
the law of exponential decay.
Future
points have no influence. Such
a
function treats
future
and
past
points
differently and
so
will obviously intro-
duce
a
unidirectional distortion into the
numerical results,
as
does the
RC
filter in an actual instrument.
When dealing with sets of numbers in
hand, and not an actual run on an
instrument where the data is emerging
in serial order, it is possible to look
ahead
as
well
as
behind. Then we can
convolute with
a
function that treats
past
and
future
on an equal basis, such
as the function in Figure
20.
Here the
most weight is given to the central
point, and points on either side of the
center are symmetrically weighed
exponentially. This function acts like
an idealized lead-lag network, which is
not practical to make with resistors,
capacitors, and
so
on.
The usual spectrum from
a
spectro-
photometer is the resultant of two con-
volutions of the actual spectrum
of
the
material, first with
a
function represent-
ing the slit function of the instrument,
which is much like the triangular con-
volute shown in Figure
2C,
and then this
first convolute spectrum is further
convoluted with
a
function representing
the time constant of the instrument.
The triangular convoluting function
could in many cases yield results not
significantly different from the sym-
metrical exponential function.
Figure
3
illustrates the way in which
each of these functions would act on
a
typical set
of
spectroscopic data. Curve
3A
is replotted directly from the instru-
mental data. It is
a
single sharp band
’
recorded under conditions which yield
a
reasonable noise level. The isolated
point just to the right
of
the band has
the value of
666
on the scale of zero to
1000
corresponding to approximately
0
to
100%
transmittance. This point is
introduced to illustrate the effect on
these operations of
a
single point which
has
a
gross error. The numbers along
the bottom are the digital value at the
1
El
I
..4-
I
x.
B
I
Figure
2.
Various convolute functions
A.
Moving average.
B.
Exponential func-
tion.
C.
Symmetrical triangular function, rep-
resenting idealized spectrometer slit function.
3.
Symmetrical exponential function
lowest point of the plot, and one may
consider that the peak goes down to
34.2%
transmittance. The base line at
the top is at about
797,.
Curve
3B
is a nine-point moving
average of the data.
As
expected, the
peak is considerably shortened by this
process. Especially interesting is the
step introduced by the isolated error.
In effect, it has the shape
of
the bodike
convolute in Figure
2d,
which is exactly
what one would expect from the con-
voluting process
(3).
Curve
3C
is
for
a triangular function
which obviously forces both the peak it-
self and the isolated error into a
triangular mold.
Curve
30
is the result of convoluting
with the numerical equivalent of a
conventional
RC
esponential time con-
stant filter using only five points. The
peak is not only shortened, but is also
shifted to the right by one data point,
or
0.002
micron and the isolated data
point is asymmetric in the same manner.
The convolution with a symmetrical
lead-lag exponential, as in Figure
3E,
does not distort the peak but does still
reduce its intensity.
Sote that while all of these functions
have had the desired effect of reducing
the noise level, they are clearly unde-
sirable because of the accompanying
degradation of the peak intensity.
METHOD
OF
LEAST SQUARES
The convoluting functions discussed
so
far are rather simple and do not
extract as much information as is pos-
sible. The experimenter, if presented
with
a
plot
of
the data points, would
tend to draw through these points a line
which best fits them. Xumerically,
this can also be done, provided one can
adequately define what is meant by
best
fit.
The most, common criterion is
that of least squares which may be
simply stated as follows:
A
set of points is to be fitted to
some
curve-for example, the curve
a3x3
+
a2x2
+
alx
+
a.
=
y.
The
a’s
are to be
selected such that when each abscissa
point, is substituted into this equation,
1628
ANALYTICAL CHEMISTRY
363
354
342
368
3
56
342
Figure
3.
Spectral band convoluted
by the various
9
point functions
The number at the bottom of each peak refers
to the lowest recorded point, and is a measure
of the ability to retain the shape of the peak.
A.
Raw data with single isolated error point.
E.
Moving average. C. Triangular func-
tion.
D.
Normal exponential function.
E.
Symmetrical exponential function.
F.
Least
squares smoothing function
the square of thr, differences between
the computed nuniberr,
y,
and the
obyerved numbers is
a
minimum for the
total
of
the observations used in deter-
mining the coefficients. A11 of the
error is assumed
to
be in the ordinate
and none in thr atiscissa.
Consider thr block of seven data
points enclosed by the left bracket in
Figure
4.
If
these fall along a curve
that can be described approsiniately by
the equation
;.hewn,
thrn there are
specific procedures-which are described
in most books on numerical analysis-
to find the
a's.
One then substitutes
back into the resulting equation the
abscisha at the central point indicated
by the circle. The value which is
obtained
by
this procedure is the best
value at that point based on the least
squares criterion, on the function which
was
chohrn. and on the group of points
esamined.
This procedure can be repeated for
I
ap
x)
+
a1
x2
+
a,
x
+
ao=y
I
Figure
4.
moving polynomial smooth
Representation
of
a 7-point
each group of seven points, dropping one
at the left and picking up one at the
right each time.
A
somewhat later
block is indicated at the right. In the
usual case, there is found a different set
of coefficients for each group of seven
points. Even with a high-speed com-
puter this is
a
tedious proposition
at
best.
Sote, however, that finding the
a
coefficients is required only
&s
a
means
for
determining the final best value
at
just one point, the central point of the
set.
A
careful study of the least squares
procedure using these constraints, leads
to the derivation of
a
set of integers
which provide a weighting function.
Kith this set of integers the central
point can be evaluated
by
the con-
voluting procedure discussed above.
This procedure is esactly equivalent to
the least squares. It is not approximate.
The derivation is presented in
Appendis
I.
For
either
a
cubic
or
a
quadratic function, the set of integers
is the same, and the set for up to
25
points is shown in Table
I
of Appendix
I1
with the appropriate normalizing
factors.
X
most instructive exercise is
to tabulate a simple function such as
y
=
z3
over any interval, apply these
smoot,hing convolutes and compare
these new values with the original.
The answers will be found to be exact.
In Figure
3F
this least squares con-
voluting procedure has been applied to
the data of Figure
3A,
using
a
9-point
cubic convolute. The value
at
the peak
and the shape of the peak are es-
sentially undistorted.
As
always, the
isolated point assumes the shape of the
convoluting function. The FORTRAY
language comput'er program for per-
forming this operation is presented in
Program
I
of Appendix
111.
Going beyond simple curve fitting,
one can find in the literature on numeri-
cal analysis a variety of least squares
procedures for det,ermining the first de-
rivative. These procedures are usually
based on interpolation formulas and are
for data
at
any arbitrary interval.
Again,
if
we restrict ourselves to evaluat-
ing the function only at the center point
of a set of equally spaced observations,
then there esist sets of convoluting
integers for the first derivative as well.
(These actually evaluate the derivative
of the least squares best function.)
h
complete set of tables for
derivatives
up
to the fifth order for
polynomials up to the fifth degree, using
from
5
to
25
points, is presented in
.ippendis
11.
These are more than
adequate for most work, since,
if
the
points are taken sufficiently close to-
gether, then practically any smooth
curve will look more or less like a
quadratic in the vicinity of a peak,
or
like
a
cubic in the vicinity of
a
shoulder.
More complete tables can be found in
the statistical literature
(b,
4,
6,
9).
SMOOTHING-C~~~.
1
y'-Ouodrotic
I
'
y'-Cubic
*+
..
?,-
Figure
5.
9-Point convoluting functions
(orthogonal polynomial) for smoothing
and first, second, and third derivative
Program
I1
of Appendis
111
shows the
use
of
these tables to obtain the coef-
ficients of
a
polynomial for finding the
precise center of an infrared band.
The shapes of the 9-point con-
volutes for a few of the functions are
illustrated in Figure
5.
Of special
interest is the linear relation of the first
derivative convolute for a quadratic.
This
is
quite unique operationally be-
cause in processing a table of data, only
one multiplication is necessary for each
convolution. The remainder of the
points are found from the set calculated
for
the previous point by simple subtrac-
tion. In Figure
6B
is shown the first
derivative of the spectrum in Figure
6A,
obtained using
a
9-point convolute.
The derivatives are useful in cases
such as our methods of band finding on
a computer
(Y),
in studies of derivative
spectra, in derivative therniogravimet-
ric analysis, derivative polarography,
etc.
CONCLUSIONS
With the increase in the application
of
computers to the analysis
of
digit,ized
data, the convolution methods described
are certain to gain wider usage. With
these methods, the sole function of the
computer is to act as a filter to smooth
the noise fluctuations and hopefully to
introduce no distortions into the re-
corded data
(3).
This problem of distortion is difficult,
to assess. In any of the curves of Figure
3,
there remain small fluctuations in the
Figure
6.
volute
17-Point first derivative con-
A.
Original spectrum.
B.
First derivative
spectrum.
VOL.
36,
NO.
8,
JULY
1964
1629
剩余12页未读,继续阅读
kerwinliu
- 粉丝: 32
- 资源: 5
上传资源 快速赚钱
- 我的内容管理 收起
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
会员权益专享
最新资源
- RTL8188FU-Linux-v5.7.4.2-36687.20200602.tar(20765).gz
- c++校园超市商品信息管理系统课程设计说明书(含源代码) (2).pdf
- 建筑供配电系统相关课件.pptx
- 企业管理规章制度及管理模式.doc
- vb打开摄像头.doc
- 云计算-可信计算中认证协议改进方案.pdf
- [详细完整版]单片机编程4.ppt
- c语言常用算法.pdf
- c++经典程序代码大全.pdf
- 单片机数字时钟资料.doc
- 11项目管理前沿1.0.pptx
- 基于ssm的“魅力”繁峙宣传网站的设计与实现论文.doc
- 智慧交通综合解决方案.pptx
- 建筑防潮设计-PowerPointPresentati.pptx
- SPC统计过程控制程序.pptx
- SPC统计方法基础知识.pptx
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功
评论0