See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4325242

Automatic Modulation Recognition of Digital Signals using Wavelet Features

and SVM

Conference Paper · March 2008

DOI: 10.1109/ICACT.2008.4493784 · Source: IEEE Xplore

CITATIONS READS

28 548

5 authors, including:

Cheol-Sun Park Jun-Ho Choi

27 PUBLICATIONS   142 CITATIONS   
Agency for Defense Development
40 PUBLICATIONS   179 CITATIONS   
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Jun-Ho Choi on 18 November 2015.

The user has requested enhancement of the downloaded file.

 Automatic Modulation Recognition of Digital
Signals using Wavelet Features and SVM
Cheol-Sun Park, Jun-Ho Choi, Sun-Phil Nah, Won Jang
EW Lab., Agency for Defense Development, Korea
{csun, junhochoi, spnah, jangwon }@add.re.kr

Dae Young Kim

Dept. of Info. & Comm. Eng., Chungnam Nat’l Univ., Korea
[email protected]

Abstract ⎯ This paper presents modulation classification numerical simulations and compare with that of decision tree
method capable of classifying incident digital signals without a classifier, and in Section 5, the paper is concluded.
priori information using WT key features and SVM. These key
features for modulation classification should have good
properties of sensitive with modulation types and insensitive with
2. Wavelet Key Features Extraction
SNR variation. In this paper, the 4 key features using WT
coefficients, which have the property of insensitive to the The key features for modulation classification in pattern
changing of noise, are selected. The numerical simulations using recognition approach must be selected. These features should
these features are performed. We investigate the performance of have robust properties of sensitive with modulation types and
the SVM-DDAG classifier for classifying 8 digitally modulated insensitive with SNR variation [2].
signals using only 4 WT key features (i.e., 4 level scale), and Wavelet key feature extraction is proposed here. The main
compare with that of decision tree classifier to adapt the characteristic of wavelet is that it can provide localized
modulation classification module in software radio. Results
frequency information of a signal, which is very useful for
indicated an overall success rate of 95% at the SNR of 10dB in
SVM-DDAG classifier on an AWGN channel. classification. Due to some desirable properties, the wavelet
basis constructed by Daubechies became the foundation for
Keywords ⎯ Modulation Classification (MC), Wavelet the most popular techniques for signal analysis and
Transformation (WT), Support Vector Machine (SVM), Decision representation in a wide range of applications. Digital
Directed Acyclic Graph (DDAG), Decision Tree (DT). modulated waveform is a cyclostationary signal that contains
transients in amplitude, frequency or phase and the WT is
1. Introduction quite suitable at extracting transient information. Another
attractive feature of WT is that it can be computed using fast
algorithm (e.g., Fast WT) and hence allowing identification of
An automatic radio signal classifier finds its use in military modulation types in real time [3].
and civilian communications applications including signal
confirmation, interference identification, spectrum monitoring, The continuous wavelet transform (CWT) of a signal x(t )
signal surveillance, electronic warfare, military threat analysis, is defined as
electronic counter measure, and software radio system.
Different modulation schemes have the characteristic of CWT (τ , s ) = ∫ x (t ) ψ s* (t ) dt
different transients in amplitude, frequency or phase. The
wavelet transform (WT) is a powerful tool for analyzing 1 ⎛ t −τ ⎞ (1)
= ∫ x (t ) ψ
*
non-stationary signals, which include digital communication ⎜ ⎟ dt
signals, and the WT magnitude of communication signals vary s ⎝ s ⎠
with modulation types [1]. The WT has capability to extract where the function ψ (t ) is called mother wavelet, ψ is its
*

transient information which can be exploited for modulation

classification. complex conjugate, and s is the scaling constant. The baby
In this paper, we investigate the performances of support wavelet ψ s (t ) comes from time-scaling and translation of the
vector machine (SVM) classifier with WT key features for 8 mother wavelet.
types of digital modulated signals, and compare with that of Different than STFT (Short Time Fourier Transform) in
decision tree classifier to adapt the modulation claasification which the window length is fixed, the window size of WT
module in software radio. increases as the analyzing frequency decreases [3]. The choice
The paper is organized as follows. In Section 2, the wavelet of a mother wavelet depends on its application. Due to its
features for classification are presented. In Section 3, the simple form and ease of computation, we selected the Haar
modulation classification using SVM is presented. In Section wavelet (db1) for modulation classification module of
4, we investigate the performance of the SVM classifier using software radio application.

ISBN 978-89-5519-136-3 -387- Feb. 17-20, 2008 ICACT 2008

 Fig. 1 is the entire WT modulation classification procedure
for 8 digital modulated signals. The identifier first finds the
0.3
magnitude of the WT of an incoming signal without any
normalization. After median filtering, we compute the 0.25
standard deviation of the results. Classification method using

Standard Deviation
these standard deviations of WT coefficients will determine 0.2
the type of incoming signals.
The WT decomposition process can be iterated with 0.15 2ASK
successive approximations being decomposed in turn, so that 4ASK

one signal is broken down into many lower-resolution 0.1

2FSK
4FSK
components as shown in Fig. 2. This is called the WDT 2PSK

(wavelet decomposition tree). In this paper, WDT with scale 0.05 4PSK
MSK
factor 4 is used for modulation classification. 16QAM
0
0 5 10 15 20 25 30
SNR(dB)
Fig. 3b. Graph of the SNR vs sd_cD2

0.8

0.7

Standard Deviation
0.6
Fig. 1. The procedure of automatic modulation recognition using WT
0.5

0.4
2ASK
4ASK
0.3
2FSK
4FSK
0.2
2PSK
4PSK
0.1
MSK
16QAM
0
0 5 10 15 20 25 30
SNR(dB)
Fig. 3c. Graph of the SNR vs sd_cD3

Fig. 2. Wavelet Decomposition Tree

1.2

We selected only 4 key features for modulation 1

classification. These key features (i.e. standard deviations of
Standard Deviation

detail coefficients at each level) using Haar WT are shown in 0.8

Fig. 3a – 3d.
0.6
2ASK
4ASK
2FSK
0.4
4FSK
2PSK
0.5
0.2 4PSK
MSK
16QAM
0.4
Standard Deviation

0
0 5 10 15 20 25 30
SNR(dB)
0.3 Fig. 3d. Graph of the SNR vs sd_cD4
2ASK
4ASK
0.2 2FSK 3. Modulation Classification using SVM
4FSK
2PSK
0.1 4PSK SVM is an empirical modeling algorithm, and is the
MSK
16QAM
state-of-the-art for the existing classification methods. The
0 SVM is basically a two-class classifier based on the ideas of
0 5 10 15 20 25 30
SNR(dB) “large margin” and “mapping data into a higher dimensional
Fig. 3a. Graph of the SNR vs sd_cD1 space,” and the kernel functions in the SVM.

ISBN 978-89-5519-136-3 -388- Feb. 17-20, 2008 ICACT 2008

 The first objective of the SVM classification is the The decision tree classifier (DTC) [5], which is known as
maximization of the margin between the two nearest data usually easy to implement and have low complexity, is built
points belonging to two separate classes. The second objective for performance comparison of SVM-DDAG classifier. In
is to constraint that all data points belong to the right class. It is principle the DTC learning algorithms analyze one feature
a two-class solution which can use features in multi variable at a time for all the data classes. In Fig. 6, the
dimensions. SVM classifies the points from two linearly thresholds (i.e., t1-t7) of decision tree classifier are selected
separable sets in two classes by solving a quadratic using the Mahalanobis distance [6].
optimization problem in order to find the optimal separating
hyperplane between these two classes. This hyperplane
maximizes the distance from the convex hulls of each class.
These techniques can be extended to the nonlinear cases by
embedding the data in a nonlinear space using kernel functions.
The robustness of SVM classification originates from the
strong fundamentals of statistical learning theory.
In modulation classification using SVM, we used linear,
polynomial-of-power-2 and exponential radial basis function
(RBF) kernels. One of examples using exponential RBF kernel
in SVM classification shows as shown in Fig. 4.
Since SVM is basically a binary classifier, it is not straight
forward to apply it to multi-class classification problems. The
most typical method for multi-class problem is to classify one
class from the other classes (refer 1-v-r), another typical Fig. 5. SVM-DDAG for 4-classes
method is to combine all possible two-class (pair wise)
classifiers (refer 1-v-1). It’s known as 1-v-1 type SVM is In numerical simulation, the carrier frequency and sampling
superior to 1-v-r with respect to its learning time, but rate were chosen to 150 kHz and 1200 kHz, respectively. A
execution time for classification of 1-v-1 is much worse than random symbol sequence at symbol rate equal to 25k symb/s is
1-v-r. The SVM-DDAG (Decision Directed Acyclic Graph) used as a modulating signal for digital modulations. So, the
method yields comparable accuracy and memory usage to the number of samples per symbol duration was 48. No a priori
other two methods, but yields substantial improvement in both knowledge was assumed for the classifier. To distinguish 8
training and evaluation time [4]. We applied SVM-DDAG digital modulation types, simulation runs were carried out with
method using WT for our 8 multi-class modulation 4,096 samples (equivalent to 3.4 ms) at SNR ranging from 0
classification problem (see Fig. 5). dB to 30 dB. The probabilities of correct classification (Pcc)
obtained from 400 independent ensembles at each SNR in
SVM-DDAG classifier are plotted for each modulation types
0.6 2FSK class as shown in Fig. 7.
4FSK class
Support Vector
0.55 desion boundary
2FSK class margin
4FSK class margin
0.5
σcd4

0.45

0.4

0.35

0.3
0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7
σcd3
Fig. 4. Classification 2FSK vs 4FSK in SVM

4. Numerical Simulations
Fig. 6. Flowchart of Decision Tree Classifier
In this Section, the performance of the proposed scheme is
investigated in the Matlab environment. The 8 digital Results indicated an overall success rate of over 90% at the
modulation types (i.e. 2ASK, 4ASK, 2FSK, 4FSK, 2PSK, SNR of 10dB in 2 classification schemes as shown in Fig. 8.
4PSK, MSK, and 16QAM) are classified. Digital signals are Especially, it was shown that SVM-DDAG classifier can
generated randomly and then modulated into band-limited achieve the good results with high Pcc (i.e., >= 95%) over
signal with band-limited AWGN. region of 10dB SNR. Fig. 9 is Pe (Probability Classification

ISBN 978-89-5519-136-3 -389- Feb. 17-20, 2008 ICACT 2008

Error) version of log scale for Fig. 8 to show the details of the 5. Conclusion
performance for specific region (i.e., 0-10dB SNR).
The detailed classification results at the SNR of 10dB in 2 In this paper, we proposed a robust SVM-DDAG classifier
classifiers are provided in the confusion matrix as listed in using 4 WT coefficients capable of recognizing 8 digitally
Table 1 – 2. The global Pccs of decision tree classifier and modulated signals without a priori information.
SVM-DDAG classifier are obtained 90.19% and 95.47% at We investigate the performance of the proposed classifier
the SNR of 10dB, respectively. using numerical simulation and compare with that of decision
tree classifier. In numerical simulation, 2 classifiers used the
1 only 4 key features using Haar WT magnitude.
Results indicated an overall success rate of over 90% at the
0.9
SNR of 10dB in 2 classification schemes as shown in Fig. 8.
Prob. of Correct Classification

0.8
These good results came from the large effects of properties of
0.7 the WT coefficients, which have the property of insensitive to
0.6 the changing of noise. Especially, it was shown that
0.5
SVM-DDAG classifier due to the statistical strong learning
2ASK theory of SVM and the robust property of WT magnitude for a
0.4 4ASK
2FSK
wide range of SNR can achieve over the 95% of Pcc at the
0.3
4FSK SNR of 10dB.
2PSK
0.2
4PSK
MSK
Table 1. Confusion Matrix of Decision Tree (%)
0.1
16QAM Actual Classified Modulation Type @ SNR = 10dB
0
0 5 10 15 20 25 30
Modulation 2ASK 4ASK 2FSK 4FSK 2PSK 4PSK MSK 16QAM
SNR(dB) 2ASK 100
Fig. 7. Pcc of 8 digital modulation types at SNR from 0dB – 30dB in 4ASK 98.25 1.75
SVM-DDAG 2FSK 97.5 2.25
4FSK 16.5 0.5 81.75 1.25
2PSK 99.5 0.5
1 4PSK 3.5 1.0 79.5 16.0
MSK 32.25 67.75
16QAM 2.25 0.75 97.0
Prob. of Correct Classification

0.9

Table 2. Confusion Matrix of SVM-DDAG (%)

0.8 Decision Tree Actual Classified Modulation Type @ SNR = 10dB
SVM-DDAG Modulation 2ASK 4ASK 2FSK 4FSK 2PSK 4PSK MSK 16QAM
2ASK 100
0.7 4ASK 95.5 1.75 2.75
2FSK 3.5 95.75 0.75
4FSK 0.5 99.5
0.6 2PSK 100
4PSK 0.75 9.5 84.25 5.5
MSK 0.25 7.0 92.75
0.5 16QAM 4.0 96.0
0 5 10 15 20 25 30
SNR(dB)
Fig. 8. Pcc of 2 classifiers at SNR from 0dB – 30dB REFERENCES

0
[1] X. Z. Feng, J Yang, F L Luo, J Y Chen, and X P Zhong, “Automatic
10
Modulation Recognition by Support Vector Machines using Wavelet
Decision Tree
SVM-DDAG
kernel,” Journal of Physics: Int. Symp. on Instrumentation Science and
Technology, 2006, pp. 1264-1267
[2] Cheol-Sun Park, Won Jang, Sun-Phil Nah, and Dae Young Kim,
Prob. of Classification Error

“Automatic Modulation Recognition using Support Vector Machine in

Software Radio Applications,” in Proc. 9th IEEE ICACT, Feb. 2007, pp.
9-12
[3] K. C. Ho, W. Prokopiw, and Y.T. Chan, “Modulation Identification by
-1
the Wavelet Transform,” in Proc. IEEE MILCOM, Nov. 1995, pp.
10
886-890
[4] John C. Platt, Nello Cristianini, and John Shawe-Taylor, “Large Margin
DAGs for Multiclass Classification,” in Advances in Neural Information
Processing Systems, MIT Press, 2000, pp. 547-553
[5] Jiang Yuan, Zhang Zhao-Yang, and Qiu Pei-Liang, “Modulation
Classification of Communication Signals,” in Proc. IEEE MILCOM, Oct.
0 1 2 3 4 5 6 7 8 9 10
2004, pp. 1470-1476
SNR(dB) [6] Ilan Druckmann, Eugen I. Plotkin, and M.N.S. Swamy, “Automatic
modulation type recognition,” in Proc. IEEE Canadian Conf. on Electrical
Fig. 9. Pe of 2 classifiers at SNR from 0dB – 10dB and Computer Eng., May 1998, pp. 65-68

ISBN 978-89-5519-136-3 -390- Feb. 17-20, 2008 ICACT 2008

View publication stats