Blind Modulation Classification: A Concept Whose

Time Has Come

Octavia A. Dobre1, Ali Abdi1, Yeheskel Bar-Ness1, and Wei Su2
1
CCSPR, Dept. of ECE, New Jersey Institute of Technology, Newark, NJ 07102, USA
2
RDECOM, Fort Monmouth, NJ 07703, USA
[email protected], [email protected], [email protected], [email protected]

Abstract – In this paper we address the problem of identifying the focus on classification algorithms for phase shift keying
modulation format of an incoming signal. We review many existing (PSK), quadrature amplitude modulations (QAM), amplitude
techniques for digital modulation recognition in a systematic way,
which helps the reader to see the main features of each technique. shift keying (ASK), and frequency shift keying (FSK).
The goal is to provide useful guidelines for choosing appropriate The rest of the paper is organized as follows. A general
classification algorithms for different modulations, from the large signal model is introduced in Section II. Algorithms derived
pool of available techniques. Furthermore, the performance of a within the LB and FB frameworks are presented in Section III
benchmark classifier is presented, as well as its sensitivity to several and Section IV, respectively. Numerical results are provided in
model mismatches. At the end, open problems and possible
directions for further research are briefly discussed. Section V, and conclusions are finally drawn in Section VI.

I. INTRODUCTION II. SIGNAL MODEL

Blind modulation classification (MC) is an intermediate The MC algorithms proposed in the literature employ
step between signal detection and demodulation, with information extracted from the baseband [1]-[9], [13]-[20] or
application in both commercial and military communication intermediate frequency band [11]-[12]. A general expression
systems. In civilian systems, MC is specially applied in for the baseband received waveform is given by
software defined radio (SDR), to cope with the variety of
r (t ) = s (t ; u i ) + n(t ) , 0 ≤ t ≤ KT , (1)
communication systems. In military communication systems,
advanced techniques are required for real-time signal where
interception and processing, which are vital for decisions
s (t ; u i ) = ai e jθ e j 2 π∆ft ∑ k =1 sk( i ) e jφk g (t − (k − 1)T − εT )
K
involving electronic warfare operations and other tactical (2)
actions. This has emerged the need for smart receivers, which
is the noise-free baseband complex envelope of the received
employ blind signal processing techniques. A major task of
signal, resulting from the i th modulation format,
such systems is the blind recognition of the modulation of an
i = 1,..., N mod , ai = Es σ2s E p , σ2s = M i−1 ∑ m =1| sm( i ) |2 is the
M i

incoming signal. MC is a challenging problem, particularly in (i ) (i )

variance for the ideal noise-free constellation corresponding to

a non-cooperative environment, since in addition to complex
the i th modulation format, M i is the number of equi-probable
channels, there are many unknown parameters, such as carrier ∞
points in the i th signal constellation, E p = ∫−∞ | pTX (t ) |2 dt
frequency, symbol timing, etc., that have to be extracted from
represents the pulse energy, with pTX (t ) the transmitter pulse
the received signal. The design of a classifier essentially
shape, Es is the baseband signal energy, ∆f is the frequency
involves two steps: preprocessing of the incoming signal and
offset, θ is the time-invariant carrier phase, {φk }kK=1 represents
proper selection of the classification algorithm. Preprocessing
the phase jitter, g (t ) = pTX (t ) ⊗ h(t ) 1, with h(t ) as the channel
tasks can include noise removal, estimation of carrier
impulse response and ⊗ denoting convolution, T is the
frequency, symbol period, signal and noise powers,
symbol period, [0, KT ] is the K symbol observation interval,
equalization, etc. The required preprocessing and its accuracy
ε denotes the timing offset w.r.t. the receiver reference clock
depend on the classification algorithm chosen in the second
( 0 ≤ ε < 1 ), sk(i ) = sk( i,)I + jsk( i,)Q 2 is the symbol transmitted within
step. For the classification part there are two general
the kth period and n(t ) is aggregate noise: receiver noise, as
approaches: likelihood-based (LB) [1]-[8] and feature-based
well as cochannel interference and jamming. We adopt the
(FB) methods [9]-[20]. A good classifier has to provide a high
notation s (t ; ui ) to stress the signal dependence on the
correct classification rate for a large range of signal-to-noise
unknown quantities, ui = [ai ∆f θ T ε g (t ) {φk }kK=1 {sk(i ) }kK=1 ]† ,
ratio (SNR), in a short observation interval. In addition, it
where † is the transpose operator. Note that the term
needs to fulfill these requirements: robustness with respect to ____________________________________________________________
(w.r.t.) preprocessing inaccuracy, capability to recognize a 1
large number of modulations in environments with different The pulse shape g (t ) contains information of both transmitter pulse shape
and channel. For example, the additive white Gaussian noise (AWGN) channel
propagation characteristics, real-time functionality, and low is characterized by h(t ) = δ (t ) , with δ (.) as the Dirac delta function, whereas
computational complexity. A classifier should either decide for a flat block fading h(t ) = α e jϕ δ (t ) , with α and ϕ the channel amplitude
what modulation has been received, out of N mod equally likely and phase, respectively, constant over the observation interval.
2
candidates, or that the modulation cannot be recognized as one For FSK the data symbols depend on t . For others, such as ASK, PSK and
QAM, sk( i ) ’s do not depend upon t . To simplify the notation, we most often
of the catalog. The former case is considered here, as being the drop the t -dependence, unless otherwise specified.
only treated in the literature. Due to the lack of space, we
“unknown quantity” refers to both unknown signal parameters, respectively, as well as flat block Rayleigh fading channel,
such as carrier frequency and timing offsets, as well as unknown with v i = [α ϕ {sk(i ) }kK=1 ]† [8]. Note that the pulse shape was
data symbols. The data symbols {sk(i ) }kK=1 are taken from a finite- assumed rectangular7, i.e., pTX (t ) = 1 for 0 ≤ t < T and zero
alphabet specific to the i th modulation format, i = 1,..., N mod . otherwise, and the data symbols {sk( i ) }kK=1 were treated as
For alphabets associated to different digital modulations see, for independent and identical distributed r.v.’s. The LFs derived
example, [21] Ch. 4. Without loss of generality, we consider for v i = [{sk( i ) }kK=1 ]† [1] and v i = [θ {sk(i ) }kK=1 ]† [8], are given
here unit variance constellations, i.e, {sk( i ) }kK=1 / σ s . To simplify (i ) respectively by
the notation, only {sk(i ) }kK=1 will be used in the sequel.
Λ (Ai ) [r (t )] = ∏ k =1 E s( i ) exp  2 S N 0−1 Re[ Rk(i ) ] − STN 0−1 | sk(i ) |2  , (6)
K

 
LIKELIHOOD BASED APPROACH TO MC
k
III.
and
Within the LB framework, MC is formulated as a multiple
composite hypothesis-testing problem. Under the hypothesis Λ (Ai ) [r (t )] = E{s( i ) }K  e − STN0 ηK I 0 (2 S N 0−1 | ξ(Ki ) |)  ,
−1 (i)
(7)
H i , the i th modulation is assigned to the incoming signal, k k =1

i = 1,..., N mod . This approach is based on the likelihood-ratio where S = Es / T is the signal power, I 0 (.) is the zero order
test (LRT), which uses the likelihood function (LF) of r (t ) modified Bessel function of the first kind, E s is nothing but a (i )

over the interval [0, KT ] . Depending on the model chosen for

k
finite summation over all the M i possible constellation points
the unknown quantities, three LB-MC techniques were of the i th modulation, divided by M i , for the k th interval
proposed in the literature: average LRT (ALRT) [1], [8], and E{s } [.] is the averaging performed over K data
(i ) K

symbols. Also, ξ(Ki ) = ∑ k =1 Rk(i ) and η(Ki ) = ∑ k =1| sk( i ) |2 , where

K K
generalized LRT (GLRT) [6] and hybrid LRT (HLRT) [6]-[7].
k k =1

With ALRT, the unknown quantities are treated as random kT

variables (r.v.’s) and the LF is computed by averaging over them, Rk(i ) = ∫ r (t ) sk( i )* (t )dt , k = 1,..., K . (8)
( k −1)T

Λ (Ai ) [r (t )] = ∫ Λ[r (t ) | v i , H i ] p ( v i | H i )dv i , (3) For linear modulations sk( i ) (t ) is constant over the period
kT
(k − 1)T to kT , and thus Rk(i ) = sk( i )*rk , where rk = ∫( k −1)T r (t )dt is
where Λ[ r (t ) | v i , H i ] is the conditional LF of r (t ) , the output of the matched filter at kT . The decision was made
conditioned on the unknown vector v i , and p( v i | H i ) is the a based on (4), with the threshold set to one. By comparing (6)
priori probability density function (PDF) of v i under H i . If and (7), one can easily notice that the complexity in computing
p ( v i | H i ) coincides with the true PDF, ALRT results in an the LF increases with the unknown carrier phase θ .
optimal classifier in the Bayesian sense, viz., it maximizes the For many cases of interest, the computational complexity
average probability of correct classification. and even mathematical intractability of the ALRT-based
In a two-hypothesis classification problem, the decision is classifier, as well as the need for prior knowledge, can render
made according to the ALRT impractical. Hence, approximations of the LF were
H1 investigated, leading to the so-called quasi-ALRT classifiers [3]-
Λ l(1) [r (t )] / Λ l(2) [r (t )] <> ηl , (4) [5]. Such algorithms were derived for linear modulation
H2
classification in AWGN channel, with v i = [θ {sk( i ) }kK=1 ]† 5 [2]-[3]
where ηl is a threshold, l = A( ALRT ), G (GLRT), H (HLRT) . and v i = [θ ε {sk(i ) }kK=1 ]† 8 [2], and for FSK signal identification,
Extension of (4) to multiple hypotheses is straightforward (see, with v i = [{φk }kK=1 {sk( i ) }kK=1 ]† 6 [4] and v i = [ε {φk }kK=1 {sk(i ) }kK=1 ]† 8
for example, [23] Ch. 2). [5]. For example, the approximation of the LF used in [3]9 for
For AWGN, using the complex Gaussian distribution of PSK and QAM signal, with v i = [θ {sk( i ) }kK=1 ]† and AWGN, is
n(t ) and for the i th hypothesis H i , one can show that the
conditional LF is given by (see, for example, [21] Ch.6)3 Λ(Ai ) [r(t )] ≈

Λ[r (t ) | u i , N 0 , H i ] = exp {∑ ∞
n=1
( SN0−1 )n K ∑q =0  υn−2q ( q!(n − q)!) | ms(i ) ,n,q || m
n / 2

−1
}
ˆ r ,n,q (0n−1 ) | ,


{
exp 2 N 0−1 Re  ∫ r (t ) s* (t ; ui )dt  − N 0−1 ∫ s (t ; ui ) dt ,
 0
KT

 0
KT 2
} (5) (9)
where ms , n , q = E[( s (i ) ) n − q ( s ( i )* ) q ] is the n th-order/ q conjugate
(i )

moment of the i th constellation, mˆ r , n , q (0 n −1 ) = K −1 ∑ kK=1 rkn − q (rk* ) q

where Re[.] is the real part and N 0 (W/Hz) is the two-sided
is the sample estimate of the n th order/ q -conjugate moment
power spectral density (PSD) of the complex AWGN n(t ) .
at the zero-delay vector 0 n −1 10, . denotes rounding to the
The autocorrelation function of n(t ) is given by
nearest integer, υn − 2 q is 1 if q = n / 2 and 2 for q < n / 2 , and !
E[n(t ) n* (t + τ)] = N 0δ(τ) , where E[.] is the mathematical
denotes factorial. Eq. (9) can be further simplified for symmetric
expectation, and * denotes complex conjugation.
constellations, as the n th order moments are equal to zero for n
In the available literature, the mathematical expression of
odd. Furthermore, one can easily show that the lowest order
____________________________________________________________
the LF, Λ (Ai ) [r (t )] , was derived for certain simplified scenarios,
i.e., AWGN channel, with v i = [{sk( i ) }kK=1 ]† 4 [1], 7
A rectangular pulse shape was assumed in the scanned literature, unless
v i = [θ {sk }k =1 ]
(i ) K †5
[8], and v i = [{φk }kK=1 {sk(i ) }kK=1 ]† 6 [4],
____________________________________________________________ otherwise mentioned.
8
The timing offset ε is modeled as a r.v. U [0,1) .
3
Note that v i = [u†i N 0 ]† , as N 0 is also unknown. 9
Eq. (9) was obtained by writing (5) given in [3] in terms of signal moments.
4
In AWGN, with all parameters perfectly known, we set 10
For the definition of the n th-order/ q conjugate moment, mr ,n ,q (τn−1 ) , and
θ = ∆f = ε = {φk }kK=1 = 0 . The unknowns are specified in v i , and here v i = ui . the cumulant, cr ,n ,q (τn−1 ) , as well as the relations between the moments and
5
The carrier phase θ was modeled as a r.v. U [−π, π) . cumulants, see [22] Ch.2. A τn−1 = 0 n−1 delay vector is an (n − 1) × 1 vector,
6
With φk shown in (2) as a phase jitter and modeled as a r.v. U [ −π, π) . with all the elements equal to zero.
statistic to distinguish between M -PSK and M ' -PSK Eq. (4) was used for decision, with the thresholds set based on
( M ' > M ) is the M th-order/ zero-conjugate moment, ms , M ,0 . (i ) the empirical histogram method12.
This property does not hold for QAM, for which ms , n , q = 0 , (i) An advantage of GLRT and HLRT over ALRT is that as
where n is a multiple of four and q odd, or n is not a multiple by-products, the estimates of the unknown quantities are of
of four and q even. By resorting to only the lowest order interest for data demodulation. In addition, GLRT and HLRT
statistic, suboptimal, but implementable and manageable can be applicable to different environments, e.g., Rician and
classifiers were proposed to discriminate PSK and QAM signals Rayleigh fading [7]. Also, as one can easily notice that GLRT
in [2] and [3], respectively. For example, the metric displays some implementation advantages over ALRT and
ν M =| ∑ k =1 rkM | was used to distinguish between M -PSK and
K
HLRT, as it avoids the calculation of exponential functions
M ' -PSK ( M ' > M ). The decision was made by comparing the and does not require the knowledge of noise power to compute
chosen statistic against a threshold η , which is set either the LF. However, with GLRT maximization over data symbols
theoretically11 or empirically12. By comparing the metric ν M with can lead to equal LFs for nested signal constellations, e.g.,
(7), one can notice that the complexity of the quasi-ALRT 16-QAM and 64-QAM, which in turn leads to incorrect
classifier is much lower than that of the ALRT classifier. classification [6]. Averaging over data symbols in HLRT
The following approximation of the LF was used with the removes the nested constellations problem of GLRT [6],
classifiers designed for unknown timing offset ε 8 [5] though, with several unknown parameters, HLRT does not
seem to be a good solution either, as finding the ML estimates
Λ (Ai ) [r (t )] ≈ D −1 ∑ d = 0 Λ[r (t ) | ε d , H i ],
D −1
(10) of several parameters can be very time consuming. Low-
complexity estimators can be used instead, leading to the so-
where D is the number of levels to which the timing offset called quasi-HLRT classifiers. For example, methods of
uncertainty interval is quantized and εd = d / D , moments were employed in [8] to estimate the channel
d = 0,..., D − 1 . This approximation improves as D → ∞ , since
parameters, when discriminating QAM signals in flat block
then the summation converges to an integral. The value of D fading channels. Then, (4) was used for decision, with the
directly determines the classifier complexity, as it introduces threshold set to one. A quasi-HLRT multi-antenna classifier,
more terms in (10). Interestingly, the quasi-ALRT classifiers which exploits receive diversity for performance enhancement,
are actually FB classifiers derived within the LB framework. was also reported in [8].
When a PDF cannot be assigned to the unknown We summarize the afore-mentioned LB classifiers in Table
parameters, a logical procedure is to estimate the unknown I, emphasizing the type of modulations, unknown parameters
parameters assuming H i is true and use these estimates in the and channel used.
LRT as if they were correct. If maximum likelihood (ML)
estimates are used, the result is called GLRT. Then, the LF is IV. FEATURE BASED APPROACH TO MC
given by
The design of a FB-MC algorithm essentially needs feature
Λ G(i ) [r (t )] = max Λ[r (t ) | v i , H i ]. (11) selection and decision-making. Features common to different
vi
modulations are selected in an ad-hoc way and the decision is
HLRT is a mixture of ALRT and GLRT, with the LF given by made based on their differences. Some examples of features
are statistics of the instantaneous amplitude, phase and
Λ (Hi ) [r (t )] = ∫ max Λ[r (t ) | v i , H i ] p ( v i2 | H i )dv i2 . (12) frequency [9], [14], statistics of the signal itself [13], [16]-
vi1
[20], the number of extrema in the PDF of the magnitude and
where v i = [ v †i v†i ]† .
1 2
peak magnitude of the signal wavelet transform [11], etc.
GLRT and HLRT were examined for linear modulation Different methods can be used for decision-making, such as
classification in AWGN ( v i = [θ {sk( i ) }kK=1 ]† 5) [6], with the LFs PDF-based [13]-[14], Euclidian distance [17]-[20], binary
given respectively by decision tree [9]-[10], neural networks [9], etc. The FB

{∑ }}
classifiers are summarized in Table II, including the features
Λ G(i ) [r (t )] = max
θ
K
k =1 s ( i ) {
max Re[ sk( i )* rk e − jθ ] − 2−1 ST | sk(i ) |2 , (13) used, modulation types, unknown parameters and channel.
k
Subsequently, the FB algorithms are presented from the
and perspective of a hierarchical approach, i.e., the modulation
class of the incoming signal is first identified (e.g., QAM,
Λ (Hi ) [r (t )] = PSK, ASK, FSK), and then the modulation order of each
max
θ
{∏ K
k =1
E s( i )  exp  2 S N 0−1 Re[ sk(i )* rk e − jθ ] − STN 0−1 | sk( i ) |2   .
k   } ( M ).

(14) A. FB Algorithms for Modulation Class Identification

____________________________________________________________ • Algorithms based on the instantaneous amplitude, phase and
11
frequency
In order to maximize the probability of correct classification, the threshold The most intuitive way to identify the modulation class of
η is chosen to satisfy p (i ) (η) = p (i ') (η) , where p (i ) (η) and p ( i ') (η) are the
PDFs of the chosen metric under the hypotheses H i and H i ' , respectively. an incoming signal is by using information extracted from the
12
This threshold is set to maximize the average probability of correct instantaneous amplitude, phase and frequency. In [9], the
classification over a large number of data and noise realizations. It is assumed maximum of the discrete Fourier transform (DFT) of the
that such simulations should be run off-line and the threshold can be stored as
a function of the noise and signal parameters. In a practical implementation,
the threshold is therefore obtained from a look-up table.
normalized centered13 amplitude was used to discriminate FSK • Algorithms based on the wavelet transform
and (ASK and PSK) classes, the variance of the absolute Different PSK signals give rise to different sets of peak
normalized centered phase was used to distinguish between values in the magnitude of the Haar wavelet transform. The
M -PSK ( M > 2 ) and real-valued constellation signals (BPSK modulation order ( M ) of the PSK signal was selected by
and ASK), and the variance of direct (not absolute) normalized matching the histogram of the peaks with theoretical PDFs
centered phase to distinguish between BPSK and ASK classes. corresponding to different orders [11]. The number of modes
A binary decision tree structure was employed to discriminate in the PDF of the HWT magnitude was investigated for FSK
between classes. Discrimination within each class will be briefly signal recognition in [11]; the input was identified as M -FSK
discussed in Subsection B. At each tree node, the decision was if there were M / 2 + 1 to M modes in the histogram.
made by comparing the selected statistic against a threshold12. A • Algorithms based on cumulants
decision-tree classifier was also proposed in [10], and a practical To identify the modulation order of ASK, PSK and QAM
system was developed based on this algorithm, at the signals, features based on the fourth-order cumulants of the
Communication Research Center, Canada [24]. received signal were investigated in [13]. For example, the
• Algorithms based on the wavelet transform normalized cumulant of fourth-order/ zero-conjugate,
The utility of the wavelet transform (WT) to localize cr ,4,0 (03 ) / cr2,2,1 (0) 10, was employed to identify the order of
transients in the instantaneous frequency, amplitude and phase QAM signals. The PDF of the sample estimate was used for
of the received signal was also exploited for MC. The distinct decision-making.
behavior of the Haar WT (HWT) magnitude for PSK, QAM • Algorithms based on cyclostationarity
and FSK signals was explored for class identification in Signal cyclostationarity was exploited for classification,
[11]-[12]. For a PSK signal the HWT magnitude is a constant, with two manifestations, i.e., spectral line generation when
with peaks occurring at phase changes. On the other hand, passing the signal through different nonlinearities [15] and
because of the frequency and amplitude variation in an FSK periodical fluctuation of cumulants up to the n th-order
and QAM signal, respectively, the HWT magnitude is a ( n = 8,6,4,2 ) with time [16]-[20]. A generic classifier based
staircase function with peaks at phase changes, although these on a feature vector whose components are the magnitudes of
peaks do not provide useful information for non-continuous the cyclic cumulants (CCs)14 up to the n th-order ( q -
phase FSK signals. If only the phase is retained for a QAM conjugate, q = 0,..., n ), raised to the power of 2 / n 15, as n
signal, it behaves like a PSK signal and thus, the HWT goes to infinity, and computed at all possible cycle frequencies
magnitude is constant. On the other hand, as PSK and FSK (CFs) and delay vectors, τ n -1 , was proposed in [17]. The
signals are of constant amplitude, amplitude normalization has difficulty of implementing such a classifier led to simplified
no effect on their HWT magnitude. After peaks removal, the algorithms [17]-[18]. For example, the features investigated in
variance of the HWT magnitude with amplitude normalization [18] were based on the magnitudes of the CCs up to the sixth-
was used to distinguish between FSK and (QAM and PSK) order ( n = 6,4,2 , q = 0,..., n ), at the CF (n − 2q )∆f + 1/ T and
classes, while the variance of HWT magnitude without a delay vector τ n -1 = 0n -1 , which ensures the maximum of the
amplitude normalization to distinguishing between QAM and features. Such features are robust to the carrier phase and
PSK classes. The decision was made by comparing the timing offset [17]-[19]. The decision was made based on the
selected feature against a threshold [11]-[12]. distance between the vector of sample estimates of the features
• Algorithm based on cumulants and the vector of prescribed features, with the Euclidian norm
To discriminate amongst BPSK, ASK, M -PSK ( M > 2 ) employed as metric. The CC-based features were estimated
and QAM classes, the normalized fourth-order/two-conjugate from K ρ 16 samples, taken over the observed K symbol
cumulant of the received signal, cr ,4,2 (03 ) / cr2,2,1 (0) 10, was interval [25]. Classification of single or multiple incoming
investigated in [13]. The PDF of the sample estimate was used signals present at the receiver was investigated in [18].
for decision-making. Multiple signals which overlap in time and frequency but have
distinct CFs, can be distinguished using CCs (selectivity
B. FB Algorithms for Modulation Order Identification property of CCs) [17]. Eight-order CC-based features were
• Algorithms based on the instantaneous amplitude, phase and investigated in [19] for classifying real- and complex-valued
frequency constellations, respectively. Features based on the n th-order
Information extracted from the instantaneous amplitude, CCs ( n = 4,6,8 , q = n / 2 ), robust to carrier phase and timing
phase and frequency was also exploited to identify the offset, as well as frequency offset and phase jitter, were
modulation order ( M ) [9]. The variance of the absolute value proposed for QAM recognition in [20]. The minimum
of the normalized centered instantaneous amplitude Euclidian distance was also employed for decision in [19]-
(frequency) was used to distinguish between 2-ASK and [20]. A raised-cosine pulse shape was considered in [17]-[20].
____________________________________________________________
4-ASK (2-FSK and 4-FSK) [9]. The feature was compared 14
For the definition of the n th-order CCs of a cyclostationary process, as well
against a threshold12 for decision making, at a node, as part of as the theoretical expression for the n th-order CCs of linear modulations see,
the binary decision tree classifier mentioned in Section IV.A. for example, [17], [19] and [25].
15
Statistical moments of the phase were investigated to Raising the n th-order CC magnitudes to the power of 2 / n forces the
identify PSK signals in [14], with the PDF of the sample feature to take values within the same order of magnitude. Therefore, the
estimates used for decision-making. classic Euclidian distance can be use for decision.
16
The received signal is oversampled (multiple samples over a symbol period)
____________________________________________________________ in order to exploit signal cyclostationarity. The sampling frequency is ρ/T,
13
The term “centered” specifies that the average is removed from the data set. with ρ a positive integer, called the oversampling factor.
 V. NUMERICAL RESULTS noticed that classification performance degradation does not
Subsequently, we present numerical results for QAM signal depend on the modulation order of the interference.
identification, with 16-QAM, 32-QAM and 64-QAM as In summary, results of the sensitivity analysis, i.e., the need
candidate modulations ( N mod = 3 )17. In classifying N mod for high SNR (14dB and 13.5dB compared to 11dB) show that
modulations, we use the average probability of correct the ALRT-based classifier is not robust to model mismatches.
classification as a performance measure. This is defined as VI. CONCLUSION
∑ i =1 Pc(i|i ) , where Pc(i|i ) is the probability to declare
N
Pcc = N mod
−1 mod

that the i th modulation is received when indeed it has been A useful summary of many existing LB- and FB-MC
originally transmitted. The Pc(i|i ) is estimated based on 1000 algorithms is provided in this paper, which gives the reader an
Monte Carlo simulations. The number of processed symbols is overview of the approaches used so far. The LB approach
K = 100 , the pulse shape is rectangular, and the SNR per provides an optimal solution to the MC problem (ALRT), in
symbols is defined as γ s = ST N 0 . We set T = 1 and S = 1 , the sense that it maximizes the average probability of correct
and change the SNR by varying N 0 . classification. However, the complexity of the optimal
The ALRT-based classifier, as defined by (4) and (6), with solution, in many cases of interest naturally, gives rise to
η A = 1 , serves as a benchmark, against which performances of proposing suboptimal algorithms, e.g., the quasi-ALRT
other classifiers are compared. We present in Table III the classifiers. Using ML estimates of the unknown quantities,
performance of this classifier, as well as a sensitivity effect to GLRT and HLRT were investigated as two alternatives.
several model mismatches. Such analysis provides bounds of Although GLRT has some advantages, it fails in identifying
performance, imposed by different preprocessing operations. nested constellations. On the other hand, HLRT can be
In an ideal scenario (AWGN and all parameters assumed implemented with less complexity than ALRT, and still
perfectly known), a Pcc of one is attained at 14dB SNR (III-1), achieving a reasonable performance. The complexity is further
and a Pcc of 0.9 at 11dB (III-2). Henceforth, the SNR is set to reduced in quasi-HLRT classifiers, which rely on low-
14dB, unless otherwise mentioned. An acceptable performance complexity yet accurate parameter estimators. Obviously,
( Pcc above 0.9) is reached for a normalized carrier frequency there is a trade-off between the complexity and performance,
offset ∆fT lower than 3.3x10-4 (III-3). When investigated for which depends on the estimation method. Although a FB
model mismatch, the carrier phase is fixed over a realization, method may not be optimal, it is usually simple to implement,
but varies randomly, U [−θlim , θlim ) from realization to with near-optimal performance, when designed properly.
realization. An acceptable performance is achieved for Accurate preprocessing is required for the effective
θlim ≤ 9o (III-4). A similar result ( φlim ≤ 12o ) is obtained when implementation of most of the known MC algorithms.
studying performance degradation due to a phase jitter, with Devising low-complexity blind algorithms for joint parameter
the phase φk modeled as a r.v. uniformly distributed over estimation is a topic of interest in MC. In addition,
[−φlim , φlim ) , which varies from symbol to symbol (III-5). For development of classification methods which rely less on
a rectangular pulse shape, one can easily show that a preprocessing is another topic for further investigation. New
synchronization error of ε translates, after matched filtering, classification problems have raised as a result of emerging
to an equivalent two-path channel [1 − ε ε] . An acceptable Pcc wireless technologies, such as single carrier versus multicarrier
is still achieved for ε = 0.074 (III-6). With an error ∆S in modulation recognition, classification of signals transmitted
estimating the signal power S , an acceptable performance is using single and multiple antennas, identification of space-time
achieved for | ∆S / S |≤ 22% (III-7). The effect of the impulsive modulation formats, etc. These issues mean that MC in real-
noise was investigated using a contaminated Gaussian noise world environments continues to be a dynamic research field.
PDF, (1 − p)N (0, σ 2n ) + pN (0,100σ 2n ) [26]. Here the addition
REFERENCES
refers to “mixture”, i.e., the process is realized from N (0, σ 2n )
with probability 1 − p and from N (0,100σ2n ) with probability [1] W. Wei and J. M. Mendel, “Maximum-likelihood classification for
digital amplitude-phase modulations,” IEEE Trans. Commun., vol. 48,
p . The impulsive noise causes occasional impulsive events of pp. 189-193, 2000.
a magnitude that is considerably greater than the background [2] C. Y. Huang and A. Polydoros, “Likelihood methods for MPSK
noise. We set the SNR based on the total noise variance, i.e., modulation classification,” IEEE Trans. Commun., vol. 43, pp. 1493-
N 0 = (1 − p)σ 2n + p (100σ2n ) , and p = 10−2 . Another type of non- 1504, 1995.
Gaussian noise used in the sensitivity analysis was generated [3] C. Long, K. Chugg, and A. Polydoros, “Further results in likelihood
classification of QAM signals,” in Proc. MILCOM, 1994, pp. 57-61.
by a gun fired repeatedly (Machine gun noise) [27]. With non- [4] B. F. Beidas and C. L. Weber, “Higher-order correlation-based approach
Gaussian noise, 13.5dB SNR is required to achieve a Pcc of to modulation classification of digitally frequency-modulated signals,”
0.9 (III-8). With a PSK signal as interference, i.e., transmitted IEEE Journal on Sel. Areas in Commun, vol. 13, pp. 89-101, 1995.
on the same carrier frequency and at the same symbol rate as [5] B. F. Beidas and C. L. Weber, “Asynchronous classification of MFSK
the QAM signal, a signal to interference ratio (SIR) greater signals using the higher order correlation domain,” IEEE Journal on Sel.
Areas in Commun, vol. 46, pp. 480-493, 1998.
than 18.4 dB is required to identify the QAM signals with a [6] P. Panagiotou, A. Anastasoupoulos, and A. Polydoros, ”Likelihood ratio
Pcc higher than 0.9 (III-9). The SIR is defined as the signal tests for modulation classification,” in Proc. MILCOM, 2000, pp. 670-
power over the interference power. Interestingly, we have
____________________________________________________________ 674.
[7] O. A. Dobre, J. Zarzoso, Y. Bar-Ness, and W. Su, “On the classification
17
Of course, when higher order modulations are included in the modulation of linearly modulated signals in fading channel”, in Proc. CISS, 2004.
pool, higher SNRs and/or a larger number of symbols are needed to achieve [8] A. Abdi, O. A. Dobre, R. Chauchy, Y. Bar-Ness, and W. Su,
the same performance. We have simulated these modulations to draw some “Modulation classification in fading channels using antenna arrays,” in
basic, yet insightful conclusions. Proc. MILCOM, 2004.
[9] E. E. Azzouz and A. K. Nandi, Automatic Modulation Recognition of
 Communication Signals. Kluwer Academic, 1996. [19] O. A. Dobre, Y. Bar-Ness, and W. Su, “Higher-order cyclic cumulants
[10] D. Boudreau et al., “A fast automatic modulation recognition algorithm for high order modulation classification,” in Proc. MILCOM, 2003, pp.
and its implementation in a spectrum monitoring application,” in Proc. 112-117.
MILCOM, 2000, pp. 732-736. [20] O. A. Dobre, Y. Bar-Ness, and W. Su, “Robust QAM modulation
[11] K. C. Ho, W. Prokopiw, and Y. T. Chan, “Modulation identification of classification algorithm based on cyclic cumulants,” in Proc. WCNC,
digital signals by the wavelet transform,” IEE Proc. Radar, Sonar and 2004, pp. 745-748.
Navig., vol. 47, pp. 169-176, 2000. [21] J. G. Proakis, Digital Communications, 4th ed., New York: McGraw-Hill,
[12] L. Hong and K. C. Ho, “Identification of digital modulation types using 2001.
the wavelet transform,” in Proc. MILCOM, 1999, pp. 427-431. [22] C. L Nikias and A. P. Petropulu, Higher-Order Spectra Analysis: A
[13] A. Swami and B. M. Sadler, “Hierarchical digital modulation Nonlinear Signal Processing Framework. Englewood Cliffs, NJ:
classification using cumulants,” IEEE Trans. Commun., vol. 48, pp. 416- Prentice-Hall, 1993.
429, 2000. [23] H. L. Van Trees, Detection, Estimation and Modulation Theory- Part I.
[14] S. S. Soliman and S. Z. Hsue, “Signal classification using statistical New York: Wiley, 1968.
moments,” IEEE Trans. Commun., vol. 40, pp. 908-916, 1992. [24] Communication signal analyzer, Communication Research Center,
[15] J. Reichert, “Automatic classification of communication signals using Canada, May 2000, www-ext.crc.ca.
higher order statistics,” in Proc. ICASSP, 1992, pp. 221-224. [25] V. Dandawade and G. B. Giannakis, ”Asymptotic theory of mixed time
[16] P. Marchand, J. L. Lacoume, and C. Le Martret, “Multiple hypothesis averages and kth- order cyclic-moment and cumulant statistics”, IEEE
classification based on cyclic cumulants of different orders,” in Proc. Trans. Inform. Theory, vol. 41, no.1, pp. 216-232, 1995.
ICASSP, 1998, pp. 2157-2160. [26] M. E. Johnson, Multivariate Statistical Simulation. New York: Wiley,
[17] C. M. Spooner, “Classification of co-channel communication signals 1987.
using cyclic cumulants,” in Proc. ASILOMAR, 1996, pp.531-536. [27] Signal Processing Information Base, Rice University, Houston, TX, URL
[18] C. M. Spooner, “On the utility of sixth-order cyclic cumulants for RF http://spie.ece.rice.edu/spib/data/signals/.
signal classification,” in Proc. ASILOMAR, 2001, pp. 890-897.

Table I. A SUMMARY OF LIKELIHOOD-BASED CLASSIFIERS.

Proposed by Classifier Modulations Unknown parameters Channel
Wei and Mendel [1] ALRT Linear modulations - AWGN
Huang and Polydoros [2] Quasi-ALRT PSK signals Carrier phase and timing offset AWGN
Long et al. [3] Quasi-ALRT QAM signals Carrier phase AWGN
Beidas and Weber [4]-[5] ALRT, Quasi-ALRT FSK signals Carrier phase and timing offset AWGN
Panagiotu et al. [6] GLRT, HLRT Linear modulations Carrier phase AWGN
Dobre et al. [7] HLRT Linear modulations Channel amplitude and phase Flat fading
Abdi et al. [8] ALRT Linear modulations, FSK signals Channel amplitude and phase Flat fading
Multi-antenna Quasi-HLRT QAM signals

Table II. A SUMMARY OF FEATURE-BASED CLASSIFIERS.

Proposed by Features Modulations Unknown parameters Channel
Azzouz and Maximum DFT of normalized centered 2ASK,4ASK,BPSK, - AWGN
Nandi [9] amplitude, variance of normalized centered QPSK,2FSK,4FSK
(absolute) amplitude, phase and frequency
Ho et al. Variance of the HWT magnitude, FSK and PSK classes, - AWGN
[11] the HWT magnitude and peak magnitude as well as within each
histograms classes
Hong and Variance of the HWT magnitude FSK, PSK and QAM - AWGN
Ho [12] classes
Swami and Normalized fourth-order cumulants of the Linear modulations Carrier phase, frequency and AWGN, impulsive noise,
Sadler [13] received signal timing offsets cochannel interferences
Spooner The 2 / n th powers of the magnitudes of the Linear modulations Carrier phase, frequency and AWGN,
[17]-[18] n th-order CCs of the received signal, and minimum shift timing offsets, symbol period, cochannel interferences
n = 6, 4, 2 and q = 0,..., n keying (MSK) signal amplitude and pulse shape
Dobre et al. The 2 / n th powers of the magnitudes of the QAM signals Carrier phase, frequency and AWGN,
[20] n th-order CCs of the received signal, timing offsets, and phase jitter impulsive noise
n = 8, 6, 4 and q = n / 2

Table III. RESULTS OF THE SENSITIVITY ANALYSIS OF THE ALRT-BASED CLASSIFIER TO MODEL MISMATCHES.
Classifier Model mismatch SNR(dB) Pcc
1 ALRT, (4) and (6), with η A = 0 Ideal case 14 1
2 ALRT (as in first row) Ideal case 11 0.9
3 ALRT (as in first row) Carrier frequency offset ( ∆fT = 3.3 × 10−4 ) 14 0.9
4 ALRT (as in first row) Carrier phase ( θlim = 9o , with θ r.v. U [−θlim , θlim ) ) 14 0.9
5 ALRT (as in first row) Phase jitter ( φlim = 12o , with φk r.v. U [−φlim , φlim ) ) 14 0.9
6 ALRT (as in first row) Timing offset ( ε = 0.074 ) 14 0.9
7 ALRT (as in first row)) Error in estimating the signal power ( ∆S / S = 22% ) 14 0.9
8 ALRT (as in first row)) Impulsive noise (mixture model, as well a machine gun noise) 13.5 0.9
9 ALRT (as in first row) Cochannel interferences (PSK signal as interference, SIR=18.4dB) 14 0.9