You are on page 1of 8

On the Utility of Sixth-Order Cyclic Cumulants for RI?

Signal Classification

Chad M. Spooner
Mission Research Corporation
spooner@mrcmry.com

Abstract out here the distinction between two kinds of signal


classification problems. In the first problem, most or
Several researchers have recently proposed the use all of the signal parameters are assumed known, and
of higher-order cumulants as features for automatic the problem reduces to determining the modulation
communication-signal classijication. Cumulants can type. The choice is typically constrained to a small
be associated with a stationary signal model or with number of candidate signal classes. This problem is
a cyclostationary signal model, in which case they amenable to optimization of various sorts as well as to
are periodically time-varying with Fourier compo- feature-based classification schemes (see [2] and ref-
nents called cyclic cumulants. These researchers limit erences therein). In the second problem, very little
the maximum order to foul; and it is commonly be- is assumed known regarding the data to be processed.
lieved that sixth- and higher-order cumulants pro- That is, the number of active spectral bands, number
vide no additional benejt due to their relatively large and types of signals within the bands, and the sig-
measurement-error variance. However, we have ob- nals' parameters are all unknown prior to processing.
served experimentally that classification performance The set of candidate signal classes is very large. We
can be improved by the use of sixth-order cyclic cu- are primarily interested in this latter problem, dubbed
mulants. In this paper, we present the results of a the general search problem (GSP),but the results con-
simulation study of the cumulant variancefor signals tained herein apply to the former problem as well.
corrupted by cochunnel interference and AWGN. Situ-
Of particular interest for the GSP are features that
ations in which the use of sixth-order cumulants pro-
can be used for signal detection, parameter estimation,
vides a classification beneJit are provided and the ben-
and classification, and that can be used in a cochannel-
eft is quantified.
signal situation. Of the many proposed features, we
have focused on nth-order cyclic cumulants [3,4] be-
1 Introduction cause they provide all of the desired properties for the
GSP. Their drawbacks are computational, mathemat-
For noncooperative communication problems, includ- ical, and conceptual complexity. Some authors have
ing signal detection, classification, interception, and also suggested that an addition drawback is the in-
emitter location, the problem of automatic signal de- creasing variance of the cumulant as n increases. This
tection and classification is of great importance both point brings us to the paper topic. Contrary to what ap-
because the successful solution of this problem is used pears to be a commonly held belief, sixth-order cumu-
as an input to the remaining problems and because sig- lants are indeed useful for signal classification, even
nal classification information is itself a valuable prod- for relatively short observation intervals. Evidence for
uct. Moreover, for cooperative communication, the this conclusion is presented in the remainder of this
ability to quickly and automatically classify a signal is paper.
increasingly important for multifunction receivers and This paper is organized in the following manner.
software radio. In Section 2, we present the time-domain probabil-
As we have done in previous papers [1,2], we point ity structure of communication signals [3], expressed

0-7803-7 147-X/01/$10.0002001 EEE 890


in terms of time-varying cumulant functions. Rele- where m of the factors are conjugated, and the corre-
vant prior classification work involving cumulants is sponding nth-order cumulant function is given by
reviewed in Section 3, and the variance of cumulant
P
estimates is empirically investigated in Section 4. Fi-
nally, the results of several classification experiments
7 ;n, m)
~ z ( t7 = WP) R",~( t 7 7 v j ; nj7 m j ) ,
P,, j=1
are provided in Section 5. Conclusions are drawn in
Section 6. where the sum is over all distinct partitions { ~ j } ; =of~
the index set { 1 , 2 , . . . ,n } and k ( p ) = (-l)P-l(p -
2 Characterizing the Signals of Interest l)! [3]. Since the nth-order moment functions are
polyperiodic functions of time, so too are the nth-order
We are interested in all communications signals, in- cumulant functions, and we can represent the latter as
cluding wireline, terrestrial wireless, mobile radio, and a Fourier series,
satellite signals. Our generic data model is given by
M-1
z(t)= aksk(t - t k ) ei2nfkt+'i$k + w ( t ) , (1) C ~ <n,
Um); = c,B([T~ ...Tn-101; n, m)
k=O
where each Component signal Sk is the complex en-
(a) where C!(U;n, m) is called a cyclic cmulmr (CC)
velope of a corresponding RF signal, and w ( t )is white and ,f3 is an nth-order cyclefrequency (CF).
Gaussian noise (WGN). Note that this data model does A key propeq of CCs that makes them desirable
not restrict the signals to linear modulation types (e.g., for our present purpose is signal selectivity:
digital QAM). However, digital QAM is of practical
importance, and will be the focus of this paper. For z ( t ) = zo(t) + ... + rcM-l(t) + w ( t )
M-1
digital QAM and PSK signals, the complex-envelope
signal is given by CA*, 7 ;n, m) = CZk(t,7 ;n, m)
k=O
00
+Cw(t, 7 ;n7 m)
Sk(t) = b j d t -j T k ) , (2)
j=-w provided /3 is
and C ! ( ~ ; n , m )= CZk(7;n,m),
B
where { b j } is the complex-valued symbol sequence, unique to signal z k (t).
drawn from an alphabet of symbols (the constellation)
that defines the modulation type, p ( . ) is a complex- Cumulants of Digital QAM/PSK.
valued Fourier-transformablepulse function, and 1 / T k For the signal model (2), the CCs are nonzero only for
is the symbol rate. ,f3 = q / T k , and are given by

Probability Structure.
Virtually all communications signals are usefully and
accurately modeled as cycZosrutionary time-series (or
random processes) [3,4,5]. This common ground pro- The cumulant Cb,n,m is the cumulant for the constella-
vides a starting point for the general decision-theoretic tion random variable; it is tabulated in Table 1.
analysis sketched in [2]. A theory of cyclostation-
ary signals is provided for second-order probabilistic 3 Relevant Prior Work
parameters in [5] and for higher-order parameters in
[3, 41. The most relevant parameters for our present Several researchers have proposed using cumulants
purpose are reviewed here. or cyclic cumulants to perform signal classification.
The nth-order moment function is defined by However, none have used sixth- or higher-order cumu-
lants. Relevant signal classification work is briefly re-
&(t, 7 ;n, m) = E nn

j=1
z(*)j(t + Ti), viewed in this section. To set the stage for the presen-
tation of experimental results employing sixth-order

89 1
n=4 I n=6 To specify f, the authors must specify A, which is
Const. m = 0 . 4 I m = 2 II m = l . 5 I m = 3 done by numerical optimization. The authors mention
QPSK I 1.o -1.0 -4.0 I 4.0 sixth-order cumulants in their final discussion: “Theo-
I
MPSK I 0.0 -1.0 0.0 I 4.0 I retically, it could be worth to imagine a discriminating
feature involving sixth-order statistics, but it may not
be reasonable from a estimation point of view [sic].”

Reichert et al.
In the work of Reichert et al [10, 111, the authors pro-
pose passing the received data through various nonlin-
Table 1. The cumulants for several complex- earities, detecting the resulting sine-wave components,
valued constellations. All constellationshave and classifying the signal based on the pattern of ob-
a mean of zero and variance of one. served sine-wave frequencies. It is assumed that the
signal is free from cochannel interference and a small
number of signal classes is considered. First-, second-,
CCs, the classification system employed by the author fourth-order, and the absolute-value nonlinearities are
is described in some detail. used to generate the required sine waves. Sixth-order
nonliiearities are not mentioned.
Swami et al.
In the work of Swami, Sadler, et al [6, 71, the classi- Spooner et al.
fication feature consists of the magnitude of a single In the author’s work [1, 21, the focus is on exploit-
fourth-order noncyclic cumulant, such as IC: (0;4,2) I ing CCs to solve the GSP. Therefore, multiple inter-
or lC:(O; 4,O) I. The idea is to use the lowest- fering signals are of great interest, as is the single-
order cumulant that provides discrimination for dig- signal problem, and the number of signal classes is
ital QAM/PSK signals. The focus is on small-size large, including the important QAM/PSK classes, as
problems involving two, four, or eight classes. Per- well as SQPSK, ?r/4-DQPSK, AM, CPM, CPFSK,
haps most importantly, true cochannel interference is and specific modem modulation formats such as GSM,
not considered, and the symbol rate and symbol-clock CDMA, IS-54, etc. In 123, the author derived an
phase are assumed to be known, so that symbol-rate M L classifier that uses uZ2 of the involved signals’
sampling can be performed. Since the focus is nar- CCs. This classifier is briefly reviewed in the follow-
row with respect to the GSP, probability models for ing paragraphs.
the cumulant estimates can be used to develop optimal Suppose our signal classes are denoted by integers
threshold tests. Sixth-order cumulants are mentioned in the set B, and the data model is (1). Then the single-
in passing with regard to potentially better discrimina- signal classification decision (A4= 1) is given by
tion for 256QAM versus 64QAM.
; =
Marchand et al.
In the work of Marchand et al [8,9], the focus is again
on small-size single-signal problems involving digital
QAM signals. The authors believe that the best per-
formance will be had by combining CCs of different where
orders, restricted to two and four, and they propose a sumed that we ca$ estimate the power U: and the delay
statistic that is a linear combination of the symbol-rate t o for each element of B .
ccs: The weighting function W ( ~ , n , r n , / I i , is
~ ,in-
~)
versely proportional to the variance of the estimate of
n, m)2/n.
e * n v m (U;

892
The derived classifier operates by summing the 4 Variance of Cumulant Estimates
weighted differences between measured and theoret-
ical CCs for all orders. This structure is too costly to If a single signal is embedded in WGN, then the
implement because of the requirement for infinite or- asymptotic variance for the CC estimates can be deter-
ders n and because even for finite orders n, the in- mined [4], and estimates of the variance for relatively
volved CCs are multidimensional functions. There- short observation intervals can be had [4]. These for-
fore, we present two realizable approximations in the mulas depend on the SNR, observation time, and the
following paragraphs. CCs for the involved signal. That is, the variance of
The Order-ReducedClassifier (ORCL the CC estimates can be quite different for distinct sig-
To obtain the ORC, we restrict the order n to n = 2 nal types, such as rectangular-pulse BPSK versus low-
excess-bandwidth 64QAM, even when the S N R and
and we consider only the CC magnitudes in the final
result. This leads to the classifier with the following observation interval are identical. Thus, prior to clas-
structure, sification, it does not appear possible to know the vari-
ance of the involved CCs. Therefore, to create a prac-
tical version of the DRC, we can either approximate
or neglect the weighting functions. We are then faced
with the legitimate question of how to choose Nmaz.
In the classifier (3), we combine different orders of
warped CCs, that is, CCs raised to the power 2/n. This
where we have assumed that the factor warping involves no loss of information and achieves
W ( q 2 ,m, / 3 ~ , is~ approximately
) constant and two important goals. The first is that the power level
St(f)is the spectral correlationfindon [5]. of the signal appears linearly in the various warped CC
expressions, which facilitates its estimation. The sec-
The Delay-Reduced Classifier (DRC). ond is that the higher-order CCs, which are generally
To obtain the DRC, we limit the maximum order n to larger in magnitude than the lower-order CCs, are not
some number Nma, and severely limit the range of ec unduly emphasized in the formation of the statistic.
to one delay vector per order, As in [9], we realize that the inclusion of an ad-
Nmm n ditional cumulant order has two opposing effects on
performance. The first is that the theoretical feature-
vector separation is increased, which should improve
performance. The second is that the variance of the
feature-vector increases, which should decrease per-
formance. The question for any particular problem is:
Which effect dominates for each value of n?
To validate the arguments of this section, we present
The multi-signal versions of the ML classifier, the some CC variance measurements. The variances of
ORC, and the DRC are straightforward 121. The detec- many CCs were measured for BPSK, QPSK, and MSK
tion and classification system makes use of the ORC signals in WGN, as well as for QPSK with a cochan-
and the DRC together with CF detection and grouping ne1 MSK interferer. Space is too scarce to provide all
algorithms. results here; a few for QPSK are shown in Figure 1.
The figures show normalized variances (variance di-
Discussion. vided by the mean) for the CCs and their warped ver-
The classifier (3) incorporates all orders of CCs, but sions. We draw the following conclusions. The vari-
weights them according to their measurement vari- ances depend weakly on the signal type and strongly
ance. If this variance is known, it can be used in the on the observation interval, CF, SNR, SIR, and cu-
classifier, and CCs with large variance will be auto- mulant order. The results for QPSK in WGN (SNR
matically deemphasized in the sums. of 6 dB) and QPSK plus MSK (SIR of 6 dB) are a p

893
proximately equivalent. However, the presence of the with exact knowledge of the C F s for the signal to be
MSK interferer cannot be determined before applying classified. Three problems are considered: QPSK vs.
a joint detection and classification algorithm. Judging 16QAM, 16QAM vs. V29, and QPSK vs. V29. The
solely on the observed noise-floor level and the power results are shown in Figure 3. Also included in the
in the signal band (both easily observed), the signal's plots of Figure 3 are results from [9] and [6]. These
S N R would appear to be 20 dB,not 6 dB,and any pre- comparisons are approximate, since the pulse func-
dictions of CC variances based on the apparent S N R tions are quite different and the S N R s differ by a dB
would be grossly in error. or so. Nevertheless, the results reveal that OUT use of
Note also that the normalized variances for the fourth-order cumulants is comparable to these other
warped CCs are generally about the same, even when researchers' uses, indicating that perhaps their meth-
the normalized variances for the unwarped CCs differ ods might be improved by using sixth-order cumu-
greatly. This implies that the inclusion of the warped lants. Note also the rather large performance improve-
sixth-order CCs does not introduce a greater amount of ment obtained by using sixth-order cumulants for the
variability to the classification feature than that intro- shorter collects and lower SNRs.
duced by the lower-order CCs. In other words, inclu-
sion of the sixth-order CCs should generally be bene- Three-Class Problems.
ficial even if the weighting function is neglected. Next we consider two three-class problems:
64QAMI16QAMN29 and 64QAM/16QAM/QPSK.
The results are shown in Figure 4. It is clear from
5 ClassificationExperiments
these plots that the inclusion of sixth-order CCs
provides a substantial benefit whether the signal's CFs
We present here the results of several classification
are known a priori or not.
experiments. To better complement the published
For the 64QAM/16QAM/QPSK problem, we have
work referenced herein, we focus primarily on small
included some results from [8]. These results corre-
problems involving a single signal in noise. Inband
spond to a rectangular pulse function and an S N R of
SNRs of 3,6, and 9 dB are considered. Unless noted 10 dB, so that a direct comparison is not possible. As
otherwise, all signals employ a square-root raised- before, the results indicate that our use of fourth-order
cosine pulse function with 35% excess bandwidth, ten
cumulants is reasonable and that others' methods may
samples per symbol interval, and a carrier offset of benefit from inclusion of sixth-order cumulants.
0.0123456789 Hz.
A Four-Class nlo-signal Problem.
'ho-Class Detection & Classification Problems. Finally, consider a four-class problem in which two
We consider three two-class problems in which the members of the classes appear in a single frequency
system must jointly estimate the signal's CFs, CC band. The signal types, symbol rates, carrier offsets,
magnitudes, and modulation type. The three problems SIRS, and SNRs are randomly chosen within spec-
are l6QAh4 vs. 64QAM, 16QAM vs. V29, and QPSK ified intervals for each trial, and multiple trials are
vs. V29, and the results are shown in Figure 2. Evi- performed. The system is given no prior informa-
dently, the inclusion of the sixth-order CCs never de- tion except the knowledge of the four classes. For the
grades performance and usually improves it. The de- four classes consisting of BPSK, QPSK, n/4-DQPSK,
gree to which it improves performance depends on the and 8PSK, we expect very little difference between
ability of the fourth-order CCs to provide class dis- Nmaz = 4 and Nma = 6 because the CF patterns for
crimination. From Table 1, we expect the biggest im- these four classes are very different for Nm > 2. For
provement for 16QAM vs. 64QAM, the next biggest average observation lengths of 2000, 4OO0, and 6000
for 16QAM vs. V29, and the least for QPSK vs. V29. symbols, we obtain average probabilities of detection
This is verified by Figure 2. and correct classification of 0.91, 0.98, and 0.99 for
both values of Nmaz. On the other hand, for the very
Two-Class Classification Problems. difficult problem of the four classes of QPSK, 8QAM,
For this set of experiments, we provide the system 16QAM, and V29, we obtain 0.38, 0.53, and 0.65 for

894
Beta =Symbol Rate, Magnitude Statistics Beta = 4 X Carrier, MagnitudeStatistics
lo-' lo-'
< ...................... ..--
I......I ........... ..........
......I...........
j ......
.....................i ..........I......I - * - n d . w
;
a d U

"-"
......<...................................................................................
10.5 1'1 11.5 1'2 1i.5
log2(Syrnbols)
1'3 li.5 1'4 ' lot
...... ..........> .....................!....
...... ...........> ................................
10.5
!t

1'1 11.5
I
i

1'2 1i.5
log2(Symbols)
'.....I .........
<...I.
..I.'.......................
...... ..........>. .........,...
1;3
i.

1i.5 1'4

Beta ISymbol Rate, Magnitude Statistics Beta = 4 x Carrier, MagnitudeStatistics


lo-'

8 lo'
f
9
'&U 10'
.-
1

210'
.... .........> ..............
. . . . . . * . . . . . . . . . . . . . . . . . . . . . . * . . . . . . . . . . . > . . . . . . . . . . < . . . . . . I!..
...... ...................... '...................... i ........... .........> ..........,...
i..

'
j

lo6' ld.5 il li.5 1'2 12.5


log2(Syrnbols)
1'3 13.5 1'4 lo-.' ld.5 1'1 li.5 2; li.5
log2(Syrnbols)
1'3 Ii.5 1'4

Figure 1. Normalized CC variances for QPSK. Upper plots correspond to 6 dB inband SNR, lower
plots correspond to an MSK interferer with QPSK SIR of 6 dB.

each value of Nma. For this last problem, the prob- der four. In this paper, we have addressed the issue of
abilities of detection are 0.73, 0.90, and 0.97, indicat- using features that include second-, fourth-, and sixth-
ing that the observation length is sufficient for good order cyclic cumulants for automatic signal detection
detection performance but not good classification per- and classification. We have found that the inclusion of
formance. This, in turn,is due to the low inband SIR sixth-order cumulants almost never degrades detection
values for the problem combined with small feature and classification performance, and provides a sig-
distances. nificant advantage for surprisingly short observation
interval lengths when the involved signals have suf-
6 Conclusions ficiently similar second- and fourth-order cumulants
(e.g., 16QAh4, V29,MQAM). The results herein only
Severalcumulant-based signal classifiers have been re- begin to touch on the topic, and the author hopes other
cently proposed in the open literature. These classi- researchers will examine the possibility of using sixth-
fiers typically use fourth-order cumulants or a com- and higher-order cumulants to perform signal process-
bination of second- and fourth-order cumulants. Not ing tasks such as detection, classification, and param-
much justification has been given for stopping at or- eter estimation.

895
64QAM vs. 16QAM QPSK vs. 16QAM

V29 vs. 16QAM V29 vs. 16QAM

"0 1000 2000 3000 4000 5000 6000 7000 O.#' '
Collect (Symbols) 1000 2000 3000 4000 5000 6000
Collect (Symbols)
QPSK vs. V29
QPSK vs. V29

"0 1000 2000 3000 4000 5000


Collect (Symbols)
6000 7000 .ti1
06

0.5"
. .....i .......

'
1000 2000
..........

3000 4000
Collect (Symbols)
5000 6000

Figure2. Joint detection and classification re-


sults for several two-class problems. Figure 3. Classification results for several
two-class problems with a priori known CFs.

896
64QAM vs. 16QAM vs. V29 References
’I I 1
C. M. Spooner,“Classification of Cochannel Commu-
nication Signals using Cyclic Cumulants,” Proc of the
29th Asilomar Conference on Signals, System, and
Computers, 1995, pp. 531-536.
C. M. Spooner, W. A. Brown, and G. K. Yeung,
“Automatic Radio-Frequency Environment Analy-
sis,” Pmc of the 34th Asilomar Conference on Signals,
Systems, and Computers, 2000, pp. 1181-1 186.
C. M. Spooner and W. A. Gardner, “The Cumulant
Theory of Cyclostationary Time-Series, Part I: Foun-
dation and Part II: Development and Applications,”
IEEE Trans Sig Proc, Vol. 42, pp. 3387-3429, Dec.
64QAM vs. 16QAM vs. V29 1994.
A.V. Dandawatk and G.B. Giannakis, “Asymptotic
theory of mixed time averages and kth-order cyclic
moment and cumulant statistics,” IEEE Trans Inform
Theory, Vol. 41, pp. 216-232, January 1995.
W. A. Gardner. Statistical Spectral Analysis: A Non-
probabilistic Theory. Englewood Cliffs, New Jersey:
Prentice-Hall, 1987.
A. Swami and B. M. Sadler, “Hierarchical Digital
Modulation Classification Using Cumulants,” IEEE
Transactions on Communications, Vol. 48, No. 3, pp.
416-429, March 2000.
A. Swami, S . Barbarossa, and B. M. Sadler, “Blind
Collect (Symbols) Source Separation and Signal Classification,” Pmc of
64QAM vs. 16QAM vs. QPSK the 34th Asilomur Conference on Signals, Systems,
11 . and Computers, 2000, pp. 1187-1191.
P. Marchand,J-L Lacoume, and C. Le Martret, “Mul-
tiple Hypothesis Modulation Classification Based on
Cyclic Cumulants of Different Orders,” Proc. of
ICASSP, 1998.
P. Marchand and C. Le Martret, “Classification of
Linear Modulations by a Combination of Different
Orders Cyclic Cumulants,” Proc. Workshop on HOS,
1997.
J. Reichert, “Automatic Classification of Communi-
cation Signals using Higher-Order Statistics,” Proc of
ICASSe San Francisco, CA, pp. V-221-V-224, April
Collect (Symbols) 1992.
C. Schreyogg, K. Kittel, and U. Kressel, “Robust
Figure 4. Results for two three-class prob- Classification of Modulation Types using Spectral
lems. Upper plot: 16QAM vs. 64QAM vs. Features Applied to HMM,” Proc of MILCOM,Mon-
V29, joint detection and classification. Middle terey, CA, 1997.
plot: same as upper but with apriori known
CFs. Lower plot: 64QAM vs. 16QAM vs.
QPSK, known CFs.

897

You might also like