0% found this document useful (0 votes)
282 views12 pagesComputationally Efficient Algorithms For Cyclic Spectral Analysis
Talk about algorithm for Cyclic Spectral Analysis
Uploaded by
Hoàng Mạnh LinhCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
0% found this document useful (0 votes)
282 views12 pagesComputationally Efficient Algorithms For Cyclic Spectral Analysis
Talk about algorithm for Cyclic Spectral Analysis
Uploaded by
Hoàng Mạnh LinhCopyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
/ 12
ere!
eae
Seer
eran
orithms are
Computationally Efficient
Algorithms for
Cyclic Spectral Analysis
RANDY S. ROBERTS, Wi
and HERSCHEL H.
‘spectral analysis algorithms becomes increas.
ingly evident as eyelie spectral analysis grows in
{importance as. signal analysis tool I]. For many signal
analysis problems the computational complexity of
‘yelie spectral analysis far exceeds that of conventional
spectral analysis. The reason for the computational
complexity of cyelie spectral analysis lies in the nature
of the estimation problem. Essentially, cyclic spectral
analysis algorithms estimate the correlation between
spectral components of signals. In the simplest case the
Spectral components of a real-valued signal are corre
lated, whereas in the most general case spectral com
ponents of two complex-valued signals are correlated
Is the potentially large number of correlation com
putations, rather than computing the spectral com-
ponents, that makes cyclic spectral analysis,
‘computationally complex. Over the last stx years several
computationally efficient cyelle spectral analysis algo
rithms have evolved from the
briginal methods introduced
In [2] and {3}. The objective of
this paper is to present these
a need for computationally efficient eyelte
” veo
IAM A, BROWN,
LOOMIS, JR
algorithms and describe their signal processing, com:
pllational, and structural properties.
‘Cyclle spectral analysis algorithms generally fll into
‘wo classes: those that average in frequency (frequency
smoothing) and those that average in time (time
smoothing). Although both classes of algorithms
‘Produce similar approximations to the cycle spectrum,
lume smoothing algorithms are considered to be more
‘computationally eMcient for general eyclic spectral
Analysis. Frequency smoothing algorithms can be com:
pultationally superior to time smoothing algorithms in
certain reslrieted cases, e.g. for estimating the cyclic
spectrum fora few values of cycle frequency or estimat
{ng the eyelle spectrum for small time frequency resolu
tion product {8}. With computational efficiency for
general elie spectral analysis as our primary motiva
tion, we focus our attention in this. paper on time
‘smoothing algorithms.
‘Our discussion of time smoothing algorithms begins
by describing an algorithm
based on the time smoothed
cyelie cross perfodogram. This
algorithm is considered (0 be
mint the most fundamental ime smoothing algorithm and
serves to ilustrate the general characteristics of me
Smoothing algorthms. A mathematical deseripion of
the baste algorthm that lends ise to the study of
algorithm attributes such as frequency and cycle fre
‘quency resolution, and also computational complex
is developed. The mathematical description is quite
{general and we apply itt other algorithms aswell From
there we develop successively more sophisteated (and
Substantially less computationally complex) algorithms
by refining the base algorithm. After several intermedi
atealgorthms we arrive at two computationally ecient
lgortis: the FFT Accumulation Method (PAM) and
the Sirp Spectral Correlation Algorithm (SSCA).
In order to fully examine the algorithms we consider
the problem of estimating the yeic eoss spectrum of #0
compexvalued sequences, This problem isthe most
general and computationally demanding problem of dig
taleyee spectral analyst. Simplieation of the resulting
expressions to special eases of the cross cyclic spectrum
of two complex-valued sequences, such as the eylle
Spectrum of a single real-valued sequence, are easly
found by replacing references to 4") with (2) where
needed. Computational and structural simplifications
fs fom the specialization are desribed.
BASIC TIME SMOOTHING
ALGORITHMS,
An implementation of the time
‘smoothed cyclic cross periodogram
All time smoothing algorithms are based on the time
smoothed cyclic cross periodogram (2;
SymdeMiansoaawynrar), —
* Complete cycle special analysis of wo complex smquences
ol and yl requires the estimation of both Sg(0 and 58
soni
The time smoothed cyclic cross periodogram has the
physical interpretation of correlating spectral compo-
nents of xin) with spectral components of ylrd over a
lume span of at seconds, The spectral components
XilnJea/2) and Yiinyea/2). also called complex
‘demodulates, are the complex envelopes of nartow-
band, bandpass components of a signal. Figure 1 1h
lustrates the processing for a special case of Equation
(1), namely, the ime smoothed eyclic periodogram of
real sighal. A data tapering window of length T seconds
slides over the data for a time span of At seconds. At
each instant the complex demodulates of the data
‘within the window are computed. (Details of computing
‘the complex demodulates are considered later. For now.
note that the demodulates are lowpass sequences and
hhave bandwidths on the order of the reciprocal of the
data tapering window. Le, af = 1/T Hz} After the
‘complex demodulates have been computed. they are
FT Raatng ie iia ate ea
‘This work was supported in part by a grant from BSL Tne.
{Gun partial matching suppore rom the Calornia State
SICKO Program ft: W. A Gardiner.
‘eeu 9 One-sided
Bandpass filters
ABBE),
wat
yea)
i orrLs
"Fig implementation ofthe tne soothed yale oss
periadogram
correlated by time averaging their conjugate product
‘over an interval of At seconds to produce estimates of
the eyelie spectrum, To estimate the cyclie spectrum at
the point (o.oo). demodulates separated in Irequeney
by an amount ap and centered about a midpoint of fo
are correlated, The quantities At and Afare referred to
as the time and frequency resolutions of the point
estimate,
system based on Equation (1) illustrated in Fig. 2
(for now let the time decimation parameter L = 1). in
this approach the complex demodulates are obtained.
by filtering the input sequences with one-sided
bandpass filters and frequency shifting the Milter out-
Pius to baseband. To compute the point estimate at (fo.
ta) the center frequencies ofthe filters are sett f1 =f
+09 / 2 and fo= fo 00/2. Following the notation in (4
the input fliers in Pig. 2 have memory lengths of T
1N’ Tz seconds and therefore bandwidths on the order
of 1/Ps f/N ez. (Tes the sampling interval and fe is
‘he sampling frequency.) Mathematically. computation
of the complex demodulates is expressed as,
Ann Larnain-nje 2%
where alr) is a data tapering window of length
T-=N’T, seconds. For convenience, the quantity Sa is
defined to be the bandwidth ofthe input filters: hence,
{or this system the bandwidth ofthe input filters fs also
the frequency resolution of the estimate:
f= aa= fa/N’
Aer the complex demodulates have been computed
they are correlated over a time span of Af seconds, In
Fig. 2 the correlation operation Is performed by a com-
plex multiplier followed by a lowpass filter. The lowpass
filter has a memory length of At = NT; seconds and a
bandwidth on the order of 1/t= fe/N Hz. The correla
{ton operation is expressed as
thn fos
Teor ewan @)
here gin 1s a data tapering window of width A¢= NTs
seconds. The resulting output sequence S$} (n. se!
0 vero
the spectral cross correlation estimate at (facts) and is
composed of a bias lif spectral correlation is present)
‘whose magnitude indicates the degree of spectral cor:
relation, and a random component that ean be inter-
preted as measurement nolse. For a reliable estimate it
Is necessary to have At))T, or equivalently. to have the
fme-frequency resolution product be much greater
than unity: Atf)) 1. It is shown in [2] that the time
smoothed eyelle cross periodogram converges to the
cyclic cross spectrum in the limit, as At-> = followed by
f 0, ifthe time windows ain) and gin) are properly
normalized. Therefore. 1, amy = Zan
1 we have
dour “
Bt Sornaostuo “
where the cyte ross spectrum S$} (6) fs defined by
L wee o
sti
and the eyelie cross correlation Rijtigis defined by
im (stn skryytenTge®R*)
‘See [2] and [5] for further details on the above
relationships.
‘An interesting representation of the basic system is
found by combining expressions for the complex
demodulates with Equation (3) to get [4]
SHensnzLT marwrayrerre ee" — 7)
‘Equation 718 a quadratic transformation of the input
sequences with the kernel mig). By manipulating
Equations (2), (3), and (7). the Kernel for the basic
system is found to be
sda
mar eorternarperirarymirn: ey
Im general, most cyclic spectral analysts algorithms
can be mathematically deseribed tn the form of Equa:
tuon (6), that is, as a quadratic transformation of the
input sequences with a kemel that depends on the
system parameters fo, do, J. and AF [2], (5). The inter
esting feature of this representation is that the system
parameters are contained solely within mig.. Hence,
Equation (8) describes how system parameters in
uence estimation of the eyclic cross spectrum. Al
though the system Kernel provides a compact
‘mathematical representation of eyelie spectral analysis,
algorithms, a transformed version ofthis representation
provides greater insight into the algorithms.
‘mn terms of the rotated Fourier Series Transform (FST)
of the kemel mgr )
map-S,E mane eon gmewert,
the output is expressed as [4]
raion Shasouen [ Me Bora" eaL0)
oor
Where the Ex-Indicated summation Is over the region
{6/2 < B - aos Jx/2, and Rin isa random component
that can be interpreted as measurement noise. Note
that Equation (10) also describes. how system
parameters (represented by Mf )) influence the
estimation of Saf), the underlying cyclic cross
spectrum. Cyele features of S#Af) that are within the
domain of Maf) are summed in eyeie frequency and
Integrated in frequency to produce the output value of
the estimate. Cycle features outside the domain of
‘Mla,f) are suppressed. Thus, by understanding the
Dropecties of Mla,f) we ean understand how system
parameters affeet the estimate. As a general rule, if
Af) 1 then the kernel transform Affa,/) can often be
approximated by the separable form (2)
MlasOysforagtty AF) on
‘The components of the kernel, Gis) and Hayr(f)
are frequency windows that are simply related to the
Sts of data tapering windows used in the algorithm
under consideration, and have approximate
‘bandwidths of 1/A¢ and 1/7 Hz respectively. Typically,
-Maf} is a two dimensional pulse-tike funetion sur
rounded by sidelobes. The sidclobes of Mia/) are
fenerally small compared to the maintobe but can be
{roublesome n some measurements, The region of su
port of the main lobe of Mla, f) is called a Cyclic
Spectrum Analyzer (CSA) cell. Equation (11) indicates
‘Ghat a CSA cell has a width on the order of 1/ tin eycle
frequency ancl a length on the order of 1/Tin frequency.
AAs described by Equation (10), features of $4) within
CSA cell are trarismitted to the output while features
‘usi¢e of the region are suppressed. Thus. the width
(ofa CSA cell determines the eyele Frequency resolution
‘a and the length ofthe cell determines the Irequency
resolution af Figure 3 depicts an idealized (te. m0
sidelobes or skirts) CSA cell located at. o, For @
ropes measurement the eycle frequency resolution
‘must be small enough to resolve the evele features of
SHI) and the frequency resolution roust be small
‘enough to resolve Si) in frequency. Note that eyele
Region of support) OL
of (0
"Fig. 3A Ogee Spectrum Analer (CSAT Sal he
bafrequeney plane
features exerfor oa CSA cll but within the stelobes
of Ma, f) can cntsbute to the estimate of S$). Tis
desirable eft is called eyte leakage nd forint
This by designing May) to have suffcenty bw
Sdctoves
The kernel transform is a useful means for studying
the equeny reschion and jee frequency resolution
tan algoriin As an example, the Kemel transfor
forthe Base system is
mercado) ((4- 282) 02
where A{f} and Gla) are the FST of a(n ) and gin}
(Note that the bandwidth of A(/)is on the order of Aa
41/Tand that of Gia) on the order of 1/0) Since fi = Jo
+ a/2 and fe =, a9 /2, he transform kernel can also
be expressed as
May)
roaai(enes (ene $) an
Examination of Equations (12) or (19) reveals that
most of the energy in Ma,f) is indeed confined to the
TABLE I
BASIC TIME SMOOTHING
Number of Complex Multipieations
Computation Cyclic Cross Spectrum of Two ‘cycle Spectrum of
Section Complex Signals |___a Sine Reat Signal
Filtering 2NN NN’
Frequency Shift Nt Nt |
Correlate NN’ NENA,
‘ompitatonal complexity of ie basic te smoothing algorhm in rms of Uwe numberof complex liplcaions Nave Wad
complete coverage ofthe biFequency plan reves NN” point estates
acs
esac « Gia-ay-fyL)
eSACell
(ea)
f
Fig Th region of suppor of Mla Jor te Basi tiie
smoothing agri
region Lf- ols Aa/2 and ter ao! $ 1/281 (ef. Fig. 3).
Hence, the frequency resolution of this algorithm ts
again Seen to be Af= Aa f/N’. and eycle frequency
resolution is Aa= 1/At = /'N. Finally, the time-frequen:
«resolution product is Athf= N/N
Im addition to determining the resolutions of a cycle
spectral analysis algorithm the CSA cell concept 1s
‘useful or determining the computational complexity oF
an algorithm. Consider the problem of estimating the
eyclic cross spectrum of two complex signals
‘everywhere in the bifrequency plane at a given time
instant. To perform the estimation approximately NN’
point estimates must be calculated (this is the area of
the region of support of Sif) divided by the area of a
CCSA cell). All ofthe input data can be filtered with two
banks of N one-sided bandpass filters for a total of
N?N’ complex multiplications and the complex
demodulates calculated with an additional 2? com-
plex muluiplications. Additionally, NN” complex mul-
plications are needed to compute the spectral
correlations for all pint estimates (assuming that g(a)
is a rectangular window with unity height). Table 1
summarizes the computational complexity of the algo:
‘thm in terms of the number of complex rmulipications
required to compute estimates ofthe cyclic cross spectrum
oftwo complex signals and the eyelie spectrum of single
real signal
“To puit these numbers into perspective a numerical
examples considered. Say that we want to estimate the
cyclic cross spectrum of two complex sequences for Af
21/8 {Te 1) and At f= 16384. For these values. N'=8
and N=131072. Ifthe computations are performed on a
‘uniprocessor computer (ie., one ALU) capable of
producing one complex product every 150 ns, then the
‘computation would take approximately 18.6 hrs. For
the case of estimating the eyelic spectrum of a real
signal the symmetry relations Sgi-))= Sf) and
"(= SUN" tmply that S%()) need only be estimated
@ esr
fm one quadrant of the bifrequeney plane, Thus, the
previous estimation problem with a single real signal
‘would require approximately 7-9 hrs of computation.
This example serves to illustrate the “brute force” ap
proach to digital cyclic spectral analysis. Algorithms
that are considerably more efficient are presented in the
remainder ofthis paper.
‘Time smoothing with decimation
‘The computational efficiency of the previous algo-
rithm ean be improved by decimating the outputs of the
bandpass filters by a suitable factor. Equivalently, data
{s shifed into the filters in blocks of L samples where
L
