IEEE SIGNAL PROCESSING LETTERS, VOL. 23, NO. 7, JULY 2016 1003
Fold-based Kolmogorov–Smirnov
Modulation Classifier
Fanggang Wang, Octavia A. Dobre, Chung Chan, and Jingwen Zhang
Abstract—Modulation classification is crucial in applications
such as electronic warfare and interference cancellation. In this
letter, a novel feature-based Kolmogorov–Smirnov classifier is pro-
posed for the identification of the modulation formats. The received
signal is first preprocessed with a folding operation that helps iden-
tify the modulation formats based on their different axes of symme-
try. Simulation results show that the performance of the proposed
classifier is close to that of the optimal likelihood-based classifier,
while its robustness to noise uncertainty is improved and its com-
putational complexity is reduced compared to that of the optimal
likelihood-based classifier.
I. INTRODUCTION
M
ODULATION classification represents a step performed
before demodulation when the modulation format is un-
known to the receiver, and plays a key role in various military
and civilian applications, such as electronic warfare, spectrum
surveillance, software-defined and cognitive radios, and inter-
ference cancellation [1]–[5]. Likelihood and feature-based clas-
sifiers represent two main classes of modulation classification
techniques, being surveyed in [1]–[5]. The maximum likeli-
hood classifier [6], [7] and the cumulant-based classifier [8]
are seminal works for these two classes, respectively. Recently,
research on modulation classification has focused on: 1) the
investigation of the classification problems in non-ideal sce-
narios, e.g., in [9]–[11]; and 2) finding powerful features and
tools to improvethe performance of the feature-based classifiers,
e.g., in [12] and [13]. Specifically, in [9], a likelihood-based
classifier is proposed for classification in time-correlated non-
Gaussian channels, which involves a whitening filter. In [10],
the cumulant-based classifier is extended to multipath channels.
In [11], a method relying on cyclic cumulant is proposed, which
is robust to carrier phase and frequency offsets. In [12], the
concept of using the empirical cumulative distribution function
of the received signal for modulation classification is first pro-
posed, which is further exploited in [13]. This technique has the
Manuscript received November 04, 2015; revised May 17, 2016; accepted
May 22, 2016. Date of publication May 24, 2016; date of current version June
22, 2016. This work was supported in part by the National Natural Science
Foundation under Grant 61571034 and under Grant U1334202, the Fundamen-
tal Research Funds for the Central Universities under Grant 2015JBM112,
the State Key Laboratory of Rail Traffic Control and Safety under Grant
RCS2016ZT013, and the Natural Sciences and Engineering Research Coun-
cil of Canada (NSERC) under Grant 327285. The associate editor coordinating
the review of this manuscript and approving it for publication was Prof. Peter
K. Willett.
F. Wang and J. Zhang are with the Beijing Jiaotong University, Beijing
100044, China (e-mail: wangfg@bjtu.edu.cn; 12111009@bjtu.edu.cn).
O. A. Dobre is with the Memorial University, NL A1B 3X5, Canada (e-mail:
odobre@mun.ca).
C. Chan is with the Chinese University of Hong Kong, Hong Kong (e-mail:
cchan@inc.cuhk.edu.hk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LSP.2016.2572666
advantage of low computational complexity and robustness to
nonideal scenarios when compared with the optimal likelihood-
based approach; however, there exists an obvious gap in the
classification performance.
In this letter, we propose a feature-based Kolmogorov–
Smirnov (K-S) classifier that further improves the classification
performance of those in [12] and [13], reducing the perfor-
mance gap from the optimal approach. To illustrate the idea
of the proposed classifier, suppose that the signal of interest is
real-valued and modulated by either binary phase-shift keying
(BPSK) or amplitude-shift keying (ASK) with four points in
the signal constellation, i.e., 4-ASK. The signal constellation
of BPSK consists of two points, while that of 4-ASK consists
of four points, and the samples obtained from the noisy signal
are scattered around the points in the signal constellation. Let
us consider the case of low noise power. Given that the number
of samples is large enough, they would concentrate around the
signal points and it would then be easy to identify the mod-
ulation formats by counting the number of “hot spots” where
the samples concentrate. However, if there are not enough sam-
ples, it becomes difficult to identify and count these hot spots.
Furthermore, the problem becomes more difficult for higher
order modulation formats. Hence, the goal is to find a suit-
able feature that identifies different modulation formats with
a small number of samples. We propose such a feature based
on the following folding operation. For BPSK, by folding the
negative axis towards the positive axis, the two signal constel-
lation points merge into a single point. All the samples now
concentrate around the positive signal point, and the number
of samples around this point basically doubles. For 4-ASK,
the folding operation has the same effect, merging four signal
points into two. Then, through a second folding operation, the
two signal points merge into one.
1
As a result, the samples
concentrate around a single point whose location depends on
the modulation format. Then, classification can be performed
by observing how these samples concentrate using the K-S
test [12].
The rest of this letter is organized as follows: the system
model and the K-S test are presented in Section II; the pro-
posed classifiers for quadrature amplitude modulation (QAM)
and PSK are introduced in Sections III and IV, respectively,
while discussions are provided in Section V; simulation re-
sults are shown in Section VI and conclusions are drawn in
Section VII.
II. S
YSTEM DESCRIPTION
Consider the following discrete-time signal model
y
n
= x
n
+ w
n
,n=1,...,N (1)
1
A better folding scheme is described in Section III; here we provide folding
of the constellation points as an example for ease of understanding.
1070-9908 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications
standards/publications/rights/index.html for more information.