Progress In Electromagnetics Research Symposium 2005, Hangzhou, China, August 22-26

Signal Modulation Recognizer Based on Method of

Artificial Neural Networks
M. Richterova
University of Defence, Czech Republic

Abstract
Communication signals travelling in space with different modulation types and different frequencies fall
in a very wide band. Usually, it is required to identify and monitor these signals for many applications.
Some of these applications are in civilian purposes such as signal confirmation, interference identification and
spectrum management. this paper described the new original configuration of subsystems for the automatic
modulation recognition of digital and analog signals. The signal recognizer being developed consists of five
subsystems: (1) adaptive antenna arrays, (2) pre-processing of EM signals, (3) key features extraction, (4)
modulation recognizer and (5) output stage. The choice of maximum value of spectral power density of the
normalized-centred amplitude, standard deviation of the absolute value of the centred non-linear component
of the instantaneous phase, standard deviation of the absolute value of the normalized-centred instantaneous
amplitude, standard deviation of the absolute value of the normalized-centred instantaneous frequency, spectrum
symmetry measure as key features for the digital and analogue modulation recognizer based on the artificial
neural networks (ANNs). The new original structure of the recognizer of digital and analogue signals is described.
The modulation recognizer uses two ANNs with two hidden layers. The results are summarized for real EM
signals.

Introduction
Automatic modulation recognition of communication signals is a rapidly evolving area of signal analysis. In
recent years, the interest of the academic research institutes has focused on the research and development of
modulation recognition algorithms.
this paper described the classifier of analogue and digital modulated signals on the basis of the artificial
neural networks (ANNs). Section 1 deals with key feature extraction. Section 2 describes described the new
original configuration of subsystems for the automatic modulation recognition of digital signals. Section 3
summarizes results of experiments for real signals. The analogue and digital modulation types that can be
classified by the recognizer based on artificial neural networks are: AM, DSB, LSB, USB, FM, ASK2, PSK2,
PSK4, FSK2 and FSK4.

Key Feature Extraction

In the proposed modulation classifiers (MC), tthe key extracted features in the authors method are derived
from the instantaneous phase, amplitude, and frequency.
The first key feature, max , is defined by [1], [2]
2

max = max |DF T (acn (i))| /NS

(1)

where NS , is the number of samples per segment and acn (i) is the value of the normalized-centred instantaneous
amplitude at time instants t = fis , (i = 1, 2, . . . , NS ).
max represents the maximum value of the spectral power density of the normalized-centred instantaneous
amplitude of the intercepted signal.
The second key feature, ap , is defined by [1], [2]
v

2
u
u
X
X
u1
1

ap = t
2N L (i)
|2N L (i)|
(2)
C
C
an (i) i at

an (i) i at

where N L (i) is the value of the centred non-linear component of the instantaneous phase at time instants
t = fis , C is the number of samples in {N L (i)} for which an (i) iat and at is a threshold. N L (i) are the
values of phase characteristic without the contributions of the carrier frequency, i.e.

Progress In Electromagnetics Research Symposium 2005, Hangzhou, China, August 22-26

N L (i) = uw (i)

2 fc
fs

where uw (i) is the unwrapped phase sequence, fc is the carrier frequency and fs is the sampl frequency.
ap is the standard deviation of the absolute value of the centred non-linear component of the instantaneous
phase.
The third key feature, dp , is defined by [1], [2]
v
u

2
u
X
X
u1
1

dp = u
2N L (i)
2N L (i)
(3)
tC
C
an (i) i at

an (i) i at

dp is the standard deviation of the centred non-linear component of the direct (not absolute) instantaneous
phase.
The fourth key feature aa is defined by [1] , [2]
v
!
!2
u
NS
NS
u 1
X
X
1
t
aa =
a2 (i)
|a2 (i)|
(4)
NS i=1 cn
NS i=1 cn
aa is the standard deviation of the absolute value of the normalized- centred instantaneous amplitude of a
signal segment.
The fifth key feature, af ,is defined by [1] , [2]
v

2
u
u
X
X
u1
2 (i) 1
2 (i)|
af = t
fN
|fN
(5)
C
C
an (i)iat

an (i)iat

where fN is the normalized-centred instantaneous frequency. af is standard deviation of the absolute value
of the normalized-centred instantaneous frequency. The four key features, used for the AMRAs (Analogue
Modulation Recognition Algorithms), are also used for the ADMRAs.
Two of the key features used for the DMRAs (Digitally Modulation Recognition Algorithms) are also used
here in the ADMRAs, they are:
aa the standard deviation of the absolute value for normalized-centred instantaneous amplitude of a
signal,
af the standard deviation if the absolute value of the normalized-centred instantaneous frequency evaluated over non-weak interval of a signal segment.
Furthermore, the choice of a , a42 and f42 as key features for the proposed ADMRAs is based on the
following facts:
a - is used to discriminate between the DSB and PSK2 signals as well as to discriminate between the
combined (AM-FM) and PSK4 signals. The PSK signals have no amplitude variation except at the transition
between the successive symbols,
a42 is used to discriminate between the AM signals as one subset and the MASK signal as the second
subset. This key feature is used to measure the compactness of the instantaneous amplitude distribution,
f42 is used to discriminate between the FM signals as one subset and the MFSK signals as the second
subset. This key feature is used to measure the compactness of the instantaneous frequency distribution. The
key features are used as an input layer for the MC based on the ANN. The MC will be described in next section.

Original Configuration of Subsystems for the Automatic Modulation

Recognition of Analogue and Digital Signals
The modulation classifier being developed consists of five subsystems: (1) adaptive antenna arrays, (2) preprocessing of signals, (3) key features extraction, (4) modulation recognizer and (5) output stage. Fig. 1 shows
the original configuration of subsystems for the automatic modulation recognition.
2.1 Adaptive Antenna Arrays and Pre-processing of the Real Signals

Progress In Electromagnetics Research Symposium 2005, Hangzhou, China, August 22-26

Figure 1: The subsystem for the automatic modulation recognition of ADMRAs

The adaptive antenna arrays are used for the capture of real signals. Generally, the planar, circular or
cylindrical antenna arrays are applied. The real signal inputs into the subsystems for the pre-processing, then
it is filtered and segmented. Every segment has 4096 samples. The segments of real signals are stored in a
database.
2.2 Subsystem for Key Features Extraction
This subsystem is designed by the author in the Matlab. The original algorithm for the key features
extraction of the real signal can utilize either the off-line key features extraction or the on-line key features
extraction. The algorithm processes the segments of the real signals from the database.
2.3 Modulation Classifier Based on ANN
This original software system in Matlab V6.5, Release 13 and NN-Toolbox, ver. 4 for Windows 98, 2000 and
XP is presently under construction. This system serves for the automatic recognition of analogue and digital
modulated signal.
The modulation classifier is composed of two ANNs based on the algorithm of the backward propagation.
The first ANN has 8-nodes in the input layer, 12-nodes in the first hidden layer, 10-nodes in the second hidden
layer and 8-nodes in the output layer. The second ANN has 3-nodes in the input layer, 6-nodes in the first and
second hidden layers and 3-nodes in the output layer. The architecture of the ANNs is shown in Fig. 2.

Figure 2: Two ANNs are used in the MC structure

Progress In Electromagnetics Research Symposium 2005, Hangzhou, China, August 22-26

Experimental Results
The performance evaluations of the proposed MC (see Fig. 1) are introduced for five digital modulated
and five analogue signals. We have classified 40 000 segments of real modulated signals. The results of the
performance are summarized in Table 1.
Table 1: Performance for the two hidden layers of ANN
Real
signal
ASK2
PSK2
PSK4
FSK2
FSK4
AM
DSB
USB
LSB
FM

ASK2
80.2
-

PSK2
11.2
86.1
-

PSK4
81.9
-

FSK2
83.8
7.3
-

FSK4
7.4
6.2
87.3
-

UNK
1.2
13.9
18.1
10.0
5.4
2.0
12.0
11.4
3.1
5.0

AM
78.0
3.8
-

DSB
20.0
84.2
-

USB
76.2
21.8
-

LSB
12.4
75.1
-

FM
95.0

In Table 1, the ASK2 and FSK2 signals are classified as FSK4. The results in this paper show that the
classifier has 75% probability of correctly recognizing real signals.

Conclusion
This paper describes the original classifier based on the ANN for the recognition of analogue and digital
modulated signals. Experimental results show that modulation classifier is capable of recognizing real signals
correctly with more than 75 % probability.
Better results and effectivity of classification can be reached by using combination methods for the modulation recognition but also by detection of other key features.
More details about other modulation classifiers are presented in [1], [2], [3], [4].
REFERENCES

1. Azzouz, E. E., A. K. Nandi, Recognition of Analogue Modulations, Signal Processing, Vol. 46, No. 2,
211-222, 1995.
2. Azzouz, E. E., A. K. Nandi, Automatic Modulation Recognition of Communication Signals, Kluwer Academic
Publishing, 1996.
3. Richterova, M., Contribution to Modulation Recognition of Digital Signals. Ph.D. Thesis, Military Academy
in Brno, Brno, 2001 (in Czech).
4. Richterova, M., The Modulation Recognition of Radiocommunication Systems, In Proceedings. A XXVIII.
Sesiune de Comunicari Stiintifice cu Participare Internationala. ATM, Bucharest, 1999.
5. Demuth, H., M. Beale, Neural Network Toolbox. For Use with MATLAB, Users Guide, ver. 4, The MathWorks, Inc., MA 01760-2098.