Physical Communication 21 (2016) 10–18

Contents lists available at ScienceDirect

Physical Communication
journal homepage: www.elsevier.com/locate/phycom

Full length article

Automatic modulation classification based on high order

cumulants and hierarchical polynomial classifiers
Ameen Abdelmutalab, Khaled Assaleh, Mohamed El-Tarhuni ∗
Electrical Engineering Department, American University of Sharjah, Sharjah, United Arab Emirates

article info abstract

Article history: In this paper, a Hierarchical Polynomial (HP) classifier is proposed to automatically classify
Received 6 November 2015 M-PSK and M-QAM signals in Additive White Gaussian Noise (AWGN) and slow flat
Received in revised form 17 July 2016 fading environments. The system uses higher order cumulants (HOCs) of the received
Accepted 24 August 2016
signal to distinguish between the different modulation types. The proposed system
Available online 5 September 2016
divides the overall modulation classification problem into several hierarchical binary sub-
classifications. In each binary sub-classification, the HOCs are expanded into a higher
Keywords:
Modulation classification
dimensional space in which the two classes are linearly separable. It is shown that there
Hierarchical polynomial classifiers is a significant improvement when using the proposed Hierarchical polynomial structure
High order cumulants compared to the conventional polynomial classifier. Moreover, simulation results are
Adaptive modulation shown for different block lengths (number of received symbols) and at different SNR
values. The proposed system showed an overall improvement in the probability of correct
classification that reaches 100% using only 512 received symbols at 20 dB compared to 98%
and 98.33% when using more complicated systems like Genetic Programming with KNN
classifier (GP-KNN) and Support Vector Machines (SVM) classifiers, respectively.
© 2016 Published by Elsevier B.V.

1. Introduction On the other hand, the classical method of signaling the bit
allocation table (BAT) to inform the receiver about the sub-
Automatic Modulation Classification (AMC) is the pro- carrier modulation level degrades the system throughput
cess of identifying the modulation type of the transmitted significantly [4]. Moreover, AMC has been used in many
signal from the received data samples automatically [1]; it military applications such as: spectrum surveillance, elec-
is an intermediate step between signal detection and de- tronic warfare and threat analysis [5]. For example, AMC
modulation [2]. AMC has received a great deal of research can be used to identify the modulation type of an inter-
and investigation in recent years because of its various cepted enemies’ signal so that the original transmitted
applications in modern communication systems. For ex- message can be extracted [6].
ample, in cognitive radio (CR) systems, AMC is used to AMC algorithms are divided into two main categories:
identify the types of signals in the spectrum. This informa- Likelihood-Based (LB) schemes and Feature-Based (FB)
tion can be used to efficiently utilize the available spectrum schemes. The LB schemes deal with the automatic
and increase the data throughput [3]. In OFDM systems, modulation classification problem as a multi-hypothesis
AMC is used to identify the modulation type in each subcar- problem in which a modulation type with the maximum
rier. Accordingly, the appropriate demodulator is selected. likelihood among all the candidates will be assigned to
the received signal [7]. Many types of LB algorithms were
∗ suggested in the literature, such as the Average Likelihood
Corresponding author.
E-mail addresses: [email protected] (A. Abdelmutalab),
Ratio Test (ALRT) [8], Generalized Likelihood Ratio Test
[email protected] (K. Assaleh), [email protected] (GLRT) [8], Hybrid Likelihood Ratio Test (HLRT) [7] and
(M. El-Tarhuni). Quasi-Hybrid Likelihood Ratio Test (QHLRT) [7]. Normally,
http://dx.doi.org/10.1016/j.phycom.2016.08.001
1874-4907/© 2016 Published by Elsevier B.V.
 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18 11

LB approaches are computationally complex and their • Provides higher probability of correct classification
performance degrades considerably in the existence of compared to most of the proposed systems in the lit-
phase or frequency offset or impulsive noise [9]. erature.
In contrast, FB schemes are simpler to implement and • Has low computational complexity and can be easily
can achieve a very close performance to LB schemes if the implemented.
used features are chosen properly [10]. Different types of • Has an acceptable probability of correct classification
received signal features are suggested in the literature such for the case of slow Rician or Rayleigh fading.
as instantaneous amplitude, phase, and frequency [11],
The rest of the paper is organized as follows: Section 2
wavelet transform [12], Fourier transform [13], high order
shows the signal model and introduces high order
moments (HOMs), high order cumulants (HOCs) [5,14,15],
cumulants. Section 3 explains the system model, Section 4
higher-order cyclic cumulants [16], very high order statis-
shows the simulation results, whereas the conclusion is
tics (VHOS) [17] and constellation diagram [18]. In gen-
presented in Section 5.
eral, the selection of the proper features depends mainly
on the modulation types of interest. The selected fea-
tures are used by a machine learning classifier to deter- 2. Signal model and HOCs
mine the modulation type of a received signal. Various
types of classifiers are used in the field like Artificial In this work, the signals are assumed to be transmitted
Neural Networks (ANN) [19], Support Vector Machines over a slow flat fading channel, which resembles realistic
(SVM) [20], Clustering Algorithms, K-Nearest Neighbors channel conditions especially when using OFDM systems
(KNN) [3], Polynomial Classifier (PC) [21], Threshold-Based to mitigate the effect of frequency selective fading chan-
Classifiers [14,22] and Naïve Bayes Classifier [23]. Fur- nels. The baseband discrete-time received signal contami-
thermore, some work used optimization techniques such nated by Additive White Gaussian Noise (AWGN) in a flat
as Genetic Programming [3,24,25] and Particle Swarm fading environment can be expressed as:
Optimization (PSO) [20] in order to improve the classifi- yn = hn xn + wn , n = 1, . . . , N (1)
cation features. However, classifiers differ in their com-
plexity, accuracy and processing time. Another factor that where xn is the discrete-time transmitted signal, wn is
limits a classifier’s accuracy is the channel conditions the AWGN process with zero mean and two-sided power
N
through which the signal is transmitted. Many authors spectral density 20 , hn is the complex-valued channel gain
investigated AMC schemes in AWGN channels as in assumed to follow a Gaussian distribution, and N is the
[23,24], whereas, other authors have considered more re- number of transmitted symbols per frame. The transmit-
alistic channel models that took into consideration multi- ted signal xn is selected from L possible modulation types.
path fading [26]. In this work we consider BPSK, QPSK, 8-PSK, 16-QAM,
Moreover, AMC can be implemented using one trans- 64-QAM or 256-QAM modulation types (i.e. L = 6). More-
mitter and one receiver, or one transmitter and multiple over, we assume that the noise variance is known or can be
receivers. In the later method, the decision can be made by estimated at the receiver.
each receiver and then passed to a centralized system to The multipath fading effect on the amplitude of the
vote, or all the received signals are directly processed by a received signal is modeled using Rayleigh or Rician fading
centralized system [27]. models. Furthermore, the relative motion between the
It is noted that some modulation types are easier transmitter and the receiver causes a shift in the frequency
to classify such as BPSK and QPSK signals compared to of the received signal fd ‘‘Doppler Shift’’. The effect of
modulation types with dense constellations like 64-QAM different values of Doppler shift on the proposed automatic
and 256-QAM. In both scenarios, the performance of the modulation classification scheme is investigated in this
classifier is affected by the number of samples from work.
which the features are extracted. Intuitively, the more the The Higher order cumulants (HOCs) are estimated from
samples used in extracting the features the more accurate the received signal and used as input features to the pro-
the system performs (for the case of AWGN channels). posed classifier. Generally, HOCs are expressed as func-
However, increasing the number of samples may reduce tions of the signals’ High Order Moments (HOMs). For a
the system performance due to the unexpected phase complex-valued stationary random process y, the pth or-
drift, residual frequency offset, and timing error. This work der moment is defined as [22]
is a continuation of our previous work proposed in [21]
Mpq = E yp−q (y∗ )q
 
where only one multidimensional polynomial classifier (2)
was used to classify the different modulation types. It ∗
where y is the complex conjugate of y, and q is the power
has been noticed that a severe drop in the probability of the conjugate signal y∗ . Table 1 shows the relationship
of correct classification of the single polynomial classifier between the HOCs and HOMs [24].
is observed when the number of considered modulation Feature normalization is required in order for machine
types increases. The probability of correct classification learning algorithms to work properly. For normalization,
(PCC) is defined as the number of correctly classified HOCs are rescaled as described in [28] with each cumulant
signals divided by the total number of signals. raised to the power 2p , where p is the order of the cumulant,
In this work, this problem is solved by using a tree 1 1
structure of binary polynomial classifiers, where each 2
i.e. Ĉ42 = C42 3
and Ĉ63 = C63 .
classifier is trained to identify two classes at time. The For more simplicity, the magnitudes of the cumulants
proposed system has the following advantages: are used instead of their complex values. This step has
12 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18

Table 1
High order cumulants and high order moments proposed in [3].
HOCs HOMs expression

C20 M20
Second order cumulants
C21 M21
2
C40 M40 − 3M20
Fourth order cumulants C41 M40 − 3M20 M21
C42 M42 − |M20 |2 − 2M21
2

3
C60 M60 − 15M20 M40 + 30M20
2
C61 M61 − 5M21 M40 − 10M20 M41 + 30M20 M21
Sixth order cumulants
2 2
C62 M62 − 6M20 M42 − 8M21 M41 − M22 M40 + 6M20 M22 + 24M21 M20
3
C63 M63 − 9M21 M42 + 12M21 − 3M20 M43 − 3M22 M41 + 18M20 M21 M22

a great advantage in reducing the processing time in variety of fields such as speech recognition [30], cognitive
the training stage because the classifier weights are real- radio systems [31] and biomedical application [32].
valued in this case instead of being complex. Another Fig. 1 shows a binary classification example in the
advantage of using the cumulants magnitude is reduced original features space d1 and d2 . It can be noticed that
vulnerability to shifts in the constellation. As the phase the two classes are nonlinearly separable and can only be
shift does not affect the magnitude of the cumulants while separated by a quadratic function. While, the new set of
it may affect their imaginary part [28]. Finally, we remark features x1 , x2 and x3 derived from an expansion of d1
that some HOCs are more useful in separating a particular and d2 can linearly separate the two classes in a higher
group of modulation types than the others, hence, it is dimensional space (3 √ dimensions in this case), where x1 =
desirable to decide on which HOCs to use automatically by d21 , x2 = d22 and x3 = 2d1 d2 .
the proposed scheme. Similar to any supervised learning algorithm, a polyno-
mial classifier has two main stages, as shown in Fig. 2. First,
3. Proposed system a training stage in which features from labeled training sig-
nals are used in calculating the classifier weights, and then
The objective of this work is to develop an AMC a testing stage where unlabeled signals are applied to the
scheme that would allow the receiver to decide (classify) classifier to identify the classes of these signals.
on which of the L possible modulation types is used by The proposed HP classifier is shown in Fig. 3, where the
the transmitter to send a given block of data. In this classification of the six modulation types is done in hierar-
work, AMC is treated as a pattern recognition problem, chical binary-classifications stages. In each stage, modula-
where features are extracted from the received signal tions having features of similar values are clustered in one
and used in a classifier to decide upon the modulation class and the rest are placed in a second class. Accordingly,
level. The features used in this work are the HOCs of the the received signal is firstly classified to be PSK or QAM. In
received signal and the classifier used is a hierarchical the next stage, if the signal is PSK, two new classes are in-
polynomial (HP) classifier. It is noted that the performance troduced, BPSK as a class and QPSK and 8-PSK as another
of any pattern recognition scheme depends on the selected class. If the signal is classified as BPSK the classification
features applied to the classifier and the capability of procedure is completed, otherwise, another binary classi-
the classifier itself. The performance of the classifier is fication between QPSK and 8-PSK is introduced. The same
measured by its probability of correct classification and principle applies for the QAM types as shown in Fig. 3. This
computational complexity. In this section, we introduce process helps in optimizing the weights in each binary clas-
the proposed HP classifier and discuss its training and sification stage leading to an overall improvement in the
testing stages. classification accuracy. In contrast, the conventional poly-
nomial classifier (non-hierarchical) used in [21] has less
3.1. Hierarchical polynomial classifier classification accuracy, especially for high values of L. To il-
lustrate this, let us consider a case of six modulation types
The Weierstrass approximation theorem states that: BPSK, QPSK, 8-PSK, 16-QAM, 64-QAM and 256-QAM, a con-
‘‘every real-valued continuous function on a finite closed ventional polynomial classifier will try to find the optimum
interval [a, b] can be uniformly approximated by polyno- set of weights for each modulation type, six weight vectors
mials with real coefficients’’ and ‘‘every complex-valued in this case. For example, the weight vector corresponding
continuous function on a finite closed interval [a, b] can to BPSK modulation is calculated based on the assumption
be uniformly approximated by polynomials with complex that BPSK is one class while QPSK, 8-PSK, 16-QAM, 64-QAM
coefficients’’ [29]. According to this theorem, a polynomial and 256-QAM are another class. This is a valid assump-
classifier can be used to approximate the nonlinear bound- tion knowing that BPSK has a very different constellation
aries between different classes. A polynomial classifier is a compared to the others, accordingly, conventional poly-
machine learning algorithm that expands the original set nomial classifier always gives the right decision when the
of features in a given space to a higher dimensional space transmitted signal is BPSK. In contrary, the weight vector
in which the different classes become linearly separated. corresponding to 64-QAM is calculated based on the as-
Due to its simplicity and high accuracy, it has been used in sumption that 64-QAM is one class and BPSK, QPSK,
 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18 13

Fig. 1. Classification features in the original space and in the high dimensional space.

(a) Offline training. (b) Online testing.

Fig. 2. Training and testing stages.

order polynomial classifiers could be used, but for imple-

mentation simplicity, a second order polynomial classifier
is normally implemented, where the new set of features is
the original features plus their products and squared val-
ues. For instance, for an input feature vector d composed
of normalized HOCs defined as d = [d1 , d2 , . . . , dM ], the
expanded feature vector p is expressed as [21]:

1, d1 , . . . , dM , d1 × d2 , . . . , d1 × dM ,
 

p = d2 × d3 , . . . , d2 × dM , . . . , dM −1 × dM , (3)
d21 , d22 , . . . , d2M
1×R

where M is the number of HOCs used as input features, and

R is the dimension of the expanded feature vector.
Fig. 3. Hierarchical polynomial classifier.
The proposed scheme is capable of classifying any
number of modulation schemes using the appropriate
8-PSK, 16-QAM and 256-QAM are another class. The prob- hierarchical approach with two classes at time. To explain
lem arises in this case, because the classifier assumes that how a polynomial classifier works, let us consider a case
very different modulation types like BPSK and 256-QAM of two modulation types, for example BPSK and QPSK. In
belong to the same class while 64-QAM which is very sim- order to train the polynomial classifier, K training signals
ilar to 256-QAM is a totally different class. This is why are transmitted using BPSK and another K training signals
conventional polynomial classifier tends to miss-classify using QPSK modulations. Second, fourth and sixth order
64-QAM as 256-QAM. The hierarchical polynomial classi- cumulants are estimated from each training signal and the
fier solved this problem by using a single binary classifica- corresponding p vector is calculated using (3). After that, a
tion at a time, two set of weights are calculated to classify new matrix vl that has all the expanded feature vectors for
64-QAM and 256-QAM, and there is no confusion intro- modulation type l is defined as:
duced by using the other modulation types in the training.
(l)
 
p1
 (l) 
3.2. Training stage p2 
vl = 
 .. 
 l = 1, 2. (4)
 . 
Second, fourth and sixth order cumulants are used as (l)
pK
features from each received signal with known modula- K ×R

tion type. The obtained features are expanded into a higher By rearranging both of v1 for BPSK modulation and v2
dimensional space using a polynomial classifier in order for QPSK in a new matrix V, we get:
to produce more features and allow for easier separation  
of the classes. The order of the polynomial classifier de- v1
V= . (5)
termines the dimensionality of the space. Although higher v2
2K ×R
14 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18

Next step is to calculate the classifier weights, wo1 for

BPSK and wo2 for QPSK. In order to do so, the weights are
selected to minimize the mean square error as:

wol = arg min ∥Vwl − tl ∥2 (6)

wl

where wl is an R × 1 vector and tl is the target matrix

defined as:
 
1
t1 = , when calculating wo1 (7)
0
2K ×1
 
0
t2 = , when calculating wo2 (8)
1 Fig. 4. Features in the original dimensional space.
2K ×1

and
BPSK, QPSK and 8-PSK are used to calculated matrix v1 , and
0 = [0, 0, . . . , 0]T1×K (9) K training signals form 16-QAM, 64-QAM and 256-QAM
are used to calculated matrix v2 . Following the same pro-
1 = [1, 1, . . . , 1]T1×K . (10)
cedure, wo1 and wo2 are calculated in the training stage. In
Eq. (6) can be simply written as: the testing stage if s1 > s2 the received signal is PSK and if
s2 > s1 the received signal is QAM. If the signal is decided
Vwol = tl . (11)
to be PSK, another binary modulation classification prob-
Then: lem is considered. This time, another polynomial classifier
wol = VĎ tl (12) is used where matrix v1 has the expanded feature vector
of BPSK modulation, and matrix v2 has the expanded fea-
Ď
where V is the pseudo inverse of matrix V [33]. Finally, ture vector of both QPSK and 8-PSK. Accordingly, in the
after finding wo1 and wo2 , the classifier is ready for the testing stage, if s1 > s2 the received signal is BPSK and if
testing stage. s2 > s1 the received signal is either QPSK or 8-PSK, and an-
other polynomial classifier is used to determine whether
3.3. Testing stage the modulation type is QPSK or 8-PSK using the same illus-
trated concept. The same idea is used if the signal is decided
In the testing stage, the objective is to find the identity to be QAM, as shown in Fig. 3.
of an unlabeled modulated signal (either BPSK or QPSK for
the example under consideration). To classify the modula-
tion type, the HOCs feature vector d is first extracted from 4. Simulation results
the received signal and the expanded vector p is calculated
using the second order expansion in (3). Then, vector p is In this section, the performance of the proposed AMC
multiplied by classifier weights wo1 and wo2 obtained during scheme is examined under different channel conditions
the training stage to give the scores s1 and s2 , respectively, and is compared to other methods from the literature.
to decide on the modulation type. Ideally, the weights are First, the advantage of the proposed classifier over the
optimized during the training stage to give s1 = 1 and traditional threshold-based method is investigated. A
s2 = 0 if the modulation is type one (BPSK), and to give simulation involving the generation of 1000 different
s1 = 0 and s2 = 1 if the modulation is type two (QPSK). realizations of 16-QAM and 64-QAM signals is conducted
Since the received symbols are noisy, the decision is made at 20 dB signal-to-noise (SNR), with each signal represents
based on the maximum values of s1 and s2 , meaning that a block of N = 2000 symbols. The conventional method
if s1 > s2 then the modulation is BPSK, and if s2 > s1 the for classifying 16-QAM and 64-QAM signals uses the value
modulation is QPSK. Hence, of fourth order cumulant C42 or the value of the sixth order
cumulant C63 as defined in [34]. Using a threshold to decide
Class identity l = arg max {sl } . (13)
l on the modulation type is not the optimal solution. Fig. 4
The previous example explained the modulation classi- shows that, there is no optimum threshold value for C42 or
fication process for only two modulation types, BPSK and C63 that can accurately separate the two modulation types.
QPSK. In this work, we consider a more general classifi- On the other hand, Fig. 5 shows the two output scores
cation problem among all of BPSK, QPSK, 8-PSK, 16-QAM, when using the proposed polynomial classifier with the
64-QAM and 256-QAM. Using the same concept introduced same input features C42 and C63 . In this scenario, when
before, we divide the modulation classification process 16-QAM is transmitted (the first 1000 signals), then the
into binary sub-classification stages. In the first stage PSK first score s1 is always greater than the second score s2 . The
signals are treated as one class and QAM signals as another opposite happens when the transmitted signal is 64-QAM
class. In other words, we define a problem of two modu- (signals numbered 1001–2000), where the second score s2
lation groups, PSK modulation group that has BPSK, QPSK is always greater than the first score s1 . Using the values
and 8-PSK signals, and QAM modulation group that has of s1 and s2 , a probability of correct classification close to
16-QAM, 64-QAM and 256-QAM. K training signals form 100% is obtained.
 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18 15

Fig. 6. Probability of correct classification in AWGN only for CP and HP

classifiers.

Fig. 7. Probability of correct classification in AWGN only for different

Fig. 5. The two scores (super features) of the polynomial classifier. number of symbols.

4.1. Hierarchical polynomial classifier number of received symbols to extract the classification
features improves the probability of correct classification.
In this section and the next section, the expanded For example, using N = 5000 symbols leads to a 4 dB im-
feature vector p is calculated from the nine HOCs in provement in SNR compared to N = 1000 symbols. From
Table 1 and used to distinguish between the different Figs. 6 and 7, it is clear that the probability of correct clas-
modulation types. In Fig. 6, a classification problem among sification is a function of SNR of the channel and signal’s
BPSK, QPSK, 8-PSK, 16-QAM, 64-QAM and 256-QAM is block length, this is due to the variation in HOCs values
considered under AWGN channel conditions. First, we have based on these parameters. While HOCs have average con-
compared the performance of the proposed HP classifier stant values (usually considered as a reference to identify
with the conventional polynomial (CP), i.e. nonhierarchical, the modulation type), the variation from these average val-
classifier, under the assumption that no SNR information ues increases with low SNR and small block lengths result-
is available at the receiver side. During the training ing in signals misclassification [35].
stage, signals with different known SNR values are used Table 2 shows the confusion matrix for the six mod-
to calculate an average classifier weights. The average ulation types (BPSK, QPSK, 8-PSK, 16-QAM, 64-QAM and
weights are then used during the testing stage to classify 245-QAM) when N = 10 000 and at SNR of 10 dB. Each
the modulation type of unknown signals regardless of the row represents the actual modulation type (2000 signals
received signal SNR. The results show that the HP classifier are transmitted per each modulation type), and the col-
provides better correct classification of the modulation umn represents the predicted modulation identity by the
type than the CP scheme, especially at higher values of classifier. For example, when sending 2000 signals of mod-
SNR. However, due to the lack of SNR information, the ulation type 16-QAM, the system was able to recognize
performance of both schemes reaches an upper limit it as 16-QAM all the time. Whereas, when sending 2000
indicating the non-optimality of selecting the classifier signals of modulation type 64-QAM, The system classified
weights. Then, we investigated the case when the SNR is 1887 of them as 64-QAM and misclassified 113 signals as
known at the receiver. Accordingly, classifier weights are 256-QAM. Table 3 shows the confusion matrix for N =
calculated for each SNR value in the training stage. Then, 5000 and at SNR of 5 dB. It is clear that the number of mis-
based on the estimated SNR of the received signal, the classified signals in this case is higher compared to Table 2.
corresponding classifier weights are used in the testing This is expected, since the number of transmitted symbols
stage. The results presented in Fig. 6 show that when per signal is less than the case in Table 2 and the SNR is
perfect SNR information is available at the receiver then lower. Moreover, Most of the modulation misclassification
significant improvement in the classification accuracy can is between 64-QAM and 256-QAM, since they have very
be achieved; reaching to about 100% accuracy for SNR similar constellations especially at low SNR. However, for
above 12 dB. The advantage of using the proposed HP over the sake of simplicity and to be able to compare our work
the CP in the two scenarios is clearly demonstrated. to the published work in the literature, we calculate the av-
The effect of using different number of symbols per data erage probability of correct classification for the six modu-
block on the probability of correct classification of HP clas- lation types, which equals to 98.04% in Table 2 and 86.52%
sifier is shown in Fig. 7. It is shown that using a larger in Table 3.
16 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18

Fig. 8. Probability of correct classification in slow Rician fading channels. Fig. 10. Probability of correct classification versus the number of symbols
for different values of Doppler shift.

In particular, we used four modulation types (BPSK, QPSK,

16-QAM and 64-QAM) with 512, 1024 and 2048 symbols
per signal. The proposed HP classifier achieved a probabil-
ity of correct classification of 100% at 20 dB for the three
given scenarios N = 512, 1024 and 2048 symbols. For
N = 512 at 10 dB HP classifier managed to achieve a
96.49% accuracy, whereas, GP-KNN, SVM and Naïve Bayes
classifiers reported accuracy of 94%, 91.23% and 90.17%, re-
spectively. It is clear that the proposed HP classifier outper-
formed the three other classifiers in terms of probability of
Fig. 9. Probability of correct classification in slow Rayleigh fading correct classification. However, the probability of correct
channels.
classification is not the only factor that favors one classi-
fier over another. A crucial factor is the system complexity
The performance of the proposed scheme is investi-
and the order of calculations involved in making the classi-
gated in slow flat fading channel conditions with magni-
fier decision. Since, modulation classification is a real-time
tude follows Rayleigh (no line of sight) or Rician fading
(with line of sight component) whereas the phase is uni- identification problem, systems with simple structure and
formly distributed and is assumed to be constant during low complexity are preferred.
the symbol period. Without loss of generality, the Rician The complexity of a classifier can be analyzed based
factor, which represents the ratio between the power of the on the complexity of the training stage (offline) and the
direct path and the power of the reflected paths, is assumed complexity of the testing stage (online). In the training
to be 6 dB. The parameters used are the same as in [28] stage, GP-KNN and SVM use iterative approaches and
with a symbol rate of 1 × 106 symbols per consume significant processing time to provide the final
 second, average
path gains [0, −10] dB, path delays 0, 1 × 10−7 s, training model, especially for the case of GP-KNN where

and maximum Doppler shift of 3 Hz, 30 Hz, or 83 Hz. different function pools are used to calculate the final
Figs. 8 and 9 show the probability of correct classification training model. However, for our proposed system and
for N = 5000 with different values of maximum Doppler Naïve Bayes classifier, the training process is straight
shifts in Rician and Rayleigh fading, respectively. The maxi- forward without any iterations involved. In general, the
mum Doppler shift and the SNR of the channel are assumed complexity of the training stage does not usually matter
to be known at the receiver side. Accordingly these val- if the calculations in the testing stage are simple, because
ues are used to train the classifier. It is clear that the prob- the training is performed offline.
ability of correct classification is higher for Rician fading In the testing stage, nine features are applied to the
channels compared to Rayleigh fading channels. This is be- proposed HP classifier, and the classifier expands them
cause of the existence of the strong signal due to the line of forming a new feature vector with 55 features. It is noted
sight component. It is noted that a higher Doppler shift re- that the only complexity in this method is calculating
sults in lower probability of correct classification. Finally, the expanded feature vector and then multiplying it by
Fig. 10 shows The classification performance as a function the classifier weights. The complexity of the proposed
of the number of received symbols and Doppler shift for HP classifier can be estimated as O(2f ) where f is the
flat Rayleigh fading channels. dimension of the expanded feature vector, and the factor
2 is for the two binary sub-classification required for each
4.2. Comparison to related work decision. This complexity is constant even if the number
of the training samples is high. On the other hand, for
Table 4 compares the probability of correct classifica- the case of the GP-KNN system, the complexity is much
tion of the Naïve Bayes classifier, Support Vector Machines higher, since it contains the complexity of finding the final
(SVM) classifier, Genetic programming with KNN classifier super feature (the calculations to find this super feature
and the proposed Hierarchical Polynomial (HP) classifier. depends on the pool function) used like (plus, minus,
To ensure fair comparison, the proposed classifier is simu- times, reciprocal, negator, abs, sqrt, sin, cos, tan, asin, acos,
lated using the same parameters used by the other authors. tanh) [24]. Furthermore, in GP-KNN classifier, the distance
 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18 17

Table 2
Confusion matrix for N = 10 000 at SNR = 10 dB (98.04%).
BPSK QPSK 8-PSK 16-QAM 64-QAM 256-QAM

BPSK 2000 0 0 0 0 0
QPSK 0 2000 0 0 0 0
8-PSK 0 0 2000 0 0 0
16-QAM 0 0 0 2000 0 0
64-QAM 0 0 0 0 1887 113
64-QAM 0 0 0 0 122 1878

Table 3
Confusion matrix for N = 5000 at SNR = 5 dB (86.52%).
BPSK QPSK 8-PSK 16-QAM 64-QAM 256-QAM

BPSK 2000 0 0 0 0 0
QPSK 0 2000 0 0 0 0
8-PSK 0 0 2000 0 0 0
16-QAM 0 0 0 1811 189 0
64-QAM 0 0 0 29 1276 695
64-QAM 0 0 0 3 701 1296

Table 4
Comparison to other systems in the literature.
N SNR (dB) Probability of correct classification (%)
Naïve [23] SVM [23] GP-KNN [3] HP (proposed method)

0 63.91 64.53 65 65.78

512 10 90.17 91.23 94 96.49
20 94.66 98.33 98 100
0 69.68 70.3 70 71.99
1024 10 94.4 94.81 98 99.10
20 98.28 98.89 100 100
0 76.75 75.73 75 77.60
2048 10 97.89 97.92 100 99.96
20 99.68 99.78 100 100

is calculated between the calculated super feature and number of received symbols to extract the classification
all the training samples; for example the number of the features results in a higher probability of correct classifica-
reference points in [36] is 400 points, accordingly, after tion. For channels with slow flat fading, the proposed sys-
finding the super feature the distances between this super tem showed no degradation due to the phase shift in the
feature and the 400 reference points are calculated. Hence, constellation. Then, a relationship between the number of
the overall complexity of GP-KNN is O(N ) where N = 400 the received symbols used to extract the classification fea-
plus the complexity of generating the super feature. tures and the probability of correct classification is inves-
For the SVM and Naïve Bayes classifiers in the testing tigated for different scenarios of Doppler shifts. Finally, the
stage the complexity is less than GP-KNN classifier, but advantage of the proposed system is investigated in terms
their overall classification accuracy is relatively lower. of accuracy and calculations complexity compared to other
Finally, it is rational that the more complex computations work in the literature.
the classifier performs the higher probability of correct
classification it provides. Yet in our proposed system, References
we managed to achieve high classification accuracy and
maintain simple classifier structure. [1] V. Orlic, M. Dukic, Algorithm for automatic modulation classification
in multipath channel based on sixth-order cumulants, in: 9th In-
ternational Conference on Telecommunication in Modern Satellite,
5. Conclusion Cable, and Broadcasting Services, TELSIKS’09, 2009, pp. 423–426.
[2] C. Yin, B. Li, Y. Li, B. Lan, Modulation classification of MQAM signals
based on density spectrum of the constellations, in: 2nd Interna-
In this paper, a hierarchical polynomial classifier is
tional Conference on Future Computer and Communication, ICFCC,
used for automatic modulation classification of M-PSK Vol. 3, 2010, pp. V3–57–V3–61.
and M-QAM signals. Higher order cumulants features are [3] M. Aslam, Z. Zhu, A. Nandi, Automatic modulation classification
extracted from the received signal and applied to the using combination of Genetic Programming and KNN, IEEE Trans.
Wireless Commun. 11 (8) (2012) 2742–2750.
proposed classifier. The proposed scheme expands the [4] L. Häring, Y. Chen, A. Czylwik, Automatic modulation classification
original feature vector into a higher dimensional space in methods for wireless OFDM systems in TDD mode, IEEE Trans.
which the classes are easily classified. It has been shown Commun. 58 (9) (2010) 2480–2485.
[5] L. Liu, J. Xu, A novel modulation classification method based on
that using the channel information at the receiver side high order cumulants, in: International Conference on Wireless
to train the classifier improves the probability of cor- Communications, Networking and Mobile Computing, WiCOM,
rect classification significantly. Moreover, using higher 2006, pp. 1–5.
18 A. Abdelmutalab et al. / Physical Communication 21 (2016) 10–18

[6] B. Ramkumar, Automatic modulation classification and blind equal- [30] W. Campbell, K. Assaleh, C. Broun, Speaker recognition with
ization for cognitive radios (Ph.D. thesis), Virginia Polytechnic polynomial classifiers, IEEE Trans. Speech Audio Process. 10 (4)
Institute and State University, Virginia. (2002) 205–212.
[7] F. Hameed, O. Dobre, D. Popescu, On the likelihood-based approach [31] Y. Hassan, M. El-Tarhuni, K. Assaleh, Comparison of linear and
to modulation classification, IEEE Trans. Wireless Commun. 8 (12) polynomial classifiers for co-operative cognitive radio networks,
(2009) 5884–5892. in: 21st IEEE International Symposium on Personal Indoor and
[8] O. Dobre, A. Abdi, Y. Bar-Ness, W. Su, Survey of automatic Mobile Radio Communications, PIMRC, 2010, pp. 797–802.
modulation classification techniques: classical approaches and new [32] K. Assaleh, H. Al-Nashash, A novel technique for the extraction of
trends, IET Commun. 1 (2) (2007) 137–156. fetal ECG using polynomial networks, IEEE Trans. Biomed. Eng. 52
[9] F. Wang, X. Wang, Fast and robust modulation classification via (6) (2005) 1148–1152.
[33] T. Shanableh, K. Assaleh, Feature modeling using polynomial
Kolmogorov-Smirnov test, IEEE Trans. Commun. 58 (8) (2010)
classifiers and stepwise regression, Neurocomputing 73 (10–12)
2324–2332.
[10] A. Hazza, M. Shoaib, S. Alshebeili, A. Fahad, An overview of (2010) 1752–1759.
feature-based methods for digital modulation classification, in: 1st [34] V. Orlic, M. Dukic, Automatic modulation classification: Sixth-order
International Conference on Communications, Signal Processing, cumulant features as a solution for real-world challenges, in: 20th
and their Applications, ICCSPA, 2013, pp. 1–6. Telecommunications Forum, TELFOR, 2012, pp. 392–399.
[35] A. Abdelmutalab, Learning-based automatic modulation classifica-
[11] K. Assaleh, K. Farrell, R. Mammone, A new method of modulation
tion (Master’s thesis), American University of Sharjah, 2015.
classification for digitally modulated signals, in: IEEE Military Com- [36] M. Aslam, Pattern recognition using genetic programming for
munications Conference, MILCOM’92, Vol. 2, 1992, pp. 712–716. classification of diabetes and modulation data (Ph.D. thesis), the
[12] W. Dan, G. Xuemai, G. Qing, A new scheme of automatic modulation University of Liverpool, Liverpool, 2013.
classification using wavelet and wsvm, in: 2nd International Con-
ference on Mobile Technology, Applications and Systems, 2005, p. 5.
[13] F. Xie, C. Li, G. Wan, An efficient and simple method of MPSK mod-
ulation classification, in: 4th International Conference on Wireless Ameen Abdelmutalab received the B.Sc. de-
Communications, Networking and Mobile Computing WiCOM’08, gree from the University of Khartoum, Sudan in
2008, pp. 1–3. 2011 and the M.Sc. degree from the American
[14] A. Swami, B. Sadler, Hierarchical digital modulation classification University of Sharjah, UAE in 2015. He is cur-
using cumulants, IEEE Trans. Commun. 48 (3) (2000) 416–429. rently a Ph.D. candidate at Concordia University,
[15] Y. Han, G. Wei, C. Song, L. Lai, Hierarchical digital modulation Montreal, Canada. After completing his B.Sc., he
recognition based on higher-order cumulants, in: 2012 Second worked as a teaching assistant at the Univer-
International Conference on Instrumentation, Measurement, Com- sity of Khartoum and the University of Science
puter, Communication and Control, IMCCC, 2012, pp. 1645–1648. and Technology in Sudan. He also worked as a
http://dx.doi.org/10.1109/IMCCC.2012.398. graduate teaching assistant during his studies at
[16] O. Dobre, Y. Bar-Ness, W. Su, Higher-order cyclic cumulants for the American University of Sharjah. His research
high order modulation classification, in: Military Communications interests include wireless communications, cognitive radio, machine-
Conference, MILCOM, Vol. 1, 2003, pp. 112–117. learning applications and image and video processing.
[17] W. Su, Feature space analysis of modulation classification using very
high-order statistics, IEEE Commun. Lett. 17 (9) (2013) 1688–1691. Khaled Assaleh is currently a Professor of elec-
[18] C. Zhendong, J. Weining, X. Changbo, L. Min, Modulation recognition trical engineering at the American University of
based on constellation diagram for M-QAM signals, in: 11th IEEE Sharjah, UAE where he is also the interim vice
International Conference on Electronic Measurement Instruments, provost for research and graduate studies. He re-
ICEMI, Vol. 1, 2013, pp. 70–74. ceived his Ph.D. in electrical engineering from
[19] J. Popoola, R. Van Olst, Automatic classification of combined analog Rutgers, The State University of New Jersey in
and digital modulation schemes using feedforward neural network, 1993. Upon completing his Ph.D. he worked as a
in: AFRICON, 2011, pp. 1–6. Research Professor at the CAIP Center of Rutgers
[20] M. Valipour, M. Homayounpour, M. Mehralian, Automatic digital University for one year. He then had an 8-year
modulation recognition in presence of noise using SVM and PSO, career in industry with Motorola, Inc. in Phoenix,
in: Sixth International Symposium on Telecommunications, IST, Arizona and with Rockwell Semiconductor Sys-
2012, pp. 378–382. tems (Conexant/Skyworks) in Newport Beach, California. Assaleh holds
[21] A. Abdelmutalab, K. Assaleh, M. El-Tarhuni, Automatic modu- 12 US patents and has published over 100 articles in signal processing and
lation classification using polynomial classifiers, in: 25th IEEE pattern recognition and their applications. His research interests include
International Symposium on Personal Indoor and Mobile Radio bio-signal processing, biometrics, speech processing, and pattern recog-
Communications, PIMRC, 2014, pp. 785–789. nition. Dr. Assaleh has served in the organization committees of several
[22] M. Mirarab, M. Sobhani, Robust modulation classification for PSK international conferences in different capacities. He has also served as a
/QAM/ASK using higher-order cumulants, in: 6th International guest editor for several journal special issues.
Conference on Information, Communications Signal Processing,
2007, pp. 1–4.
[23] M. Wong, S.K. Ting, A. Nandi, Naïve Bayes classification of adaptive Mohamed El-Tarhuni received the B.Sc. and
broadband wireless modulation schemes with higher order cumu- M.Sc. degrees in electrical engineering from the
lants, in: 2nd International Conference on Signal Processing and University of Benghazi, Libya, in 1986 and 1990,
Communication Systems, ICSPCS, 2008, pp. 1–5. respectively, and the Ph.D. degree in electrical
[24] M. Aslam, Z. Zhu, A. Nandi, Automatic digital modulation classifica- engineering from Carleton University, Ottawa,
tion using Genetic Programming with K-Nearest Neighbor, in: Mil- Canada, in 1997. From 1987 to 1993, he was with
itary Communications Conference, MILCOM, 2010, pp. 1731–1736. the General Electric Company of Libya, Beng-
[25] N. Ahmadi, R. Berangi, Modulation classification of QAM and PSK hazi, Libya, as a Telecommunications Engineer
from their constellation using genetic algorithm and hierarchical and Manager of the Communications Depart-
clustering, in: 3rd International Conference on Information and ment. From 1997 to 2000, he was with Nortel
Communication Technologies: From Theory to Applications, ICTTA, Networks, Ottawa, ON, Canada, as a member of
2008, pp. 1–5. Scientific Staff, working on third-generation wireless communication sys-
[26] H.-C. Wu, M. Saquib, Z. Yun, Novel automatic modulation classifi- tems using CDMA technology. He has been with the American University
cation using cumulant features for communications via multipath of Sharjah (AUS), Sharjah, United Arab Emirates (UAE), since 2000, where
channels, IEEE Trans. Wireless Commun. 7 (8) (2008) 3098–3105. he is currently a Professor and Associate Dean for the College of Engi-
[27] W. Su, Modulation classification of single-input multiple-output neering. From 2001 to 2006, he was the Director for the Cisco Regional
signals using asynchronous sensors, IEEE Sens. J. 15 (1) (2015) Networking Academy, AUS. He was the Head of Electrical Engineering
346–357. Department from 2006 to 2014. His current research interests include
[28] N.P. Geisinger, Classification of digital modulation schemes using wireless and mobile radio systems, cognitive radio, and physical layer is-
linear and nonlinear classifiers (Master’s thesis), Naval Postgraduate sues. Dr. El-Tarhuni has served on the organizing and technical commit-
School, California, 2010. tees of many international conferences. He served as the IEEE Industry
[29] P. Borwein, T. Erdelyi, Polynomials and Polynomial Inequalities, Relations Officer for the IEEE UAE Section from 2002 to 2006. He is a se-
Springer, 1995. nior member of the IEEE.