Zhou et al.

EURASIP Journal on Advances in Signal

Processing (2019) 2019:21
EURASIP Journal on Advances
[Link] in Signal Processing

R ES EA R CH Open Access

A robust modulation classification

method using convolutional neural
networks
Siyang Zhou1 , Zhendong Yin1 , Zhilu Wu1* , Yunfei Chen3 , Nan Zhao2 and Zhutian Yang1

Abstract
Automatic modulation classification (AMC) is a core technique in noncooperative communication systems. In
particular, feature-based (FB) AMC algorithms have been widely studied. Current FB AMC methods are commonly
designed for a limited set of modulation and lack of generalization ability; to tackle this challenge, a robust AMC
method using convolutional neural networks (CNN) is proposed in this paper. In total, 15 different modulation types
are considered. The proposed method can classify the received signal directly without feature extracion, and it can
automatically learn features from the received signals. The features learned by the CNN are presented and analyzed.
The robust features of the received signals in a specific SNR range are studied. The accuracy of classification using CNN
is shown to be remarkable, particularly for low SNRs. The generalization ability of robust features is also proven to be
excellent using the support vector machine (SVM). Finally, to help us better understand the process of feature
learning, some outputs of intermediate layers of the CNN are visualized.
Keywords: Robust automatic modulation classification, Convolutional neural networks, Deep learning, Feature
learning

1 Introduction likelihood function of the received signal, and FB methods

Automatic modulation classification (AMC) that identi- depend on feature extraction and classifier design.
fies the modulation type of the received signal is an essen- Although LB methods can theoretically achieve the
tial part of noncooperative communication systems. The optimal solution, they suffer from high computational
AMC plays an important role in many civil and military complexity and require prior information from transmit-
applications such as cognitive radio, adaptive communi- ters. In contrast, FB methods can obtain suboptimal solu-
cation, and electronic reconnaissance. tions with much smaller computational complexity and do
In these systems, transmitters can freely choose the not depend on prior information.
modulation type of signals; however, the knowledge of Since the prior information required by LB methods is
modulation type is necessary to the receivers to demodu- often unavailable in practice, researchers have paid more
late the signals so that the transmission can be successful. attention to FB methods over the past two decades. The
AMC is a sufficient way to solve this problem with no two most important parts of FB methods are feature
effects on spectrum efficiency extraction and classifier. Various types of features have
AMC algorithms have been widely studied in the past been studied and used in AMC algorithms. For exam-
20 years. In general, conventional AMC algorithms can ple, instantaneous features [3, 4] were extracted from
be divided into two categories: likelihood-based (LB) [1] the instantaneous amplitude, frequency, and phase in the
and feature-based (FB) [2]. LB methods are based on the time domain. Transformation-based features were calcu-
lated from Fourier and wavelet transforms [5, 6]. The
high-order cumulant (HOC) features [7, 8] are statisti-
*Correspondence: wuzhilu@[Link]
1
School of Electronics and Information Engineering, Harbin Institute of cal features obtained from different orders of cumulants
Technology, No. 92 West Dazhi Street, Nangang District, Harbin, China from the received signals. Additive white Gaussian noise
Full list of author information is available at the end of the article (AWGN) can be completely mathematically eliminated in

© The Author(s). 2019 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License ([Link] which permits unrestricted use, distribution, and
reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons license, and indicate if changes were made.
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 2 of 15

HOC features. Cyclostationary features are based on the Stacked sparse autoencoders are deployed as its classi-
spectral correlation function (SCF) derived from Fourier fier. The modulation set contains 7 modulation types, and
transform of the cyclic autocorrelation function [9, 10]. the generalization ability is also studied. The classifica-
The highest values of SCF for different cyclic frequencies tion accuracy reaches 90.4% when the SNR is between
are taken by the cyclic domain profile and used to train − 10 and 0 dB. O’Shea et al. [18] trained the CNN with
the classifiers. the received based band signals directly, and the classifi-
The classifier is another important part of FB methods. cation accuracy was higher than those trained by HOC
The decision tree [3] is the most widely applied linear clas- features. Some features extracted by CNN were also dis-
sifier in early years. Linear classifiers are notably easy to played. A heterogeneous model based on real-measured
implement but not feasible for linearly inseparable prob- data is proposed in [21], and the performance is enhanced
lems. Many nonlinear classifiers are applied in AMC, e.g., by combining CNN with recurrent neural networks.
K nearest neighbor [11], neural networks [12], and sup- The existing methods are all based on the assumption
port vector machine (SVM) with kernels [13]. SVM is that the SNRs of training and testing are equal. How-
considered to have advantages when the number of sam- ever, the result of SNR estimation is often inaccurate in
ples is limited and can provide better generalization ability practice, the actual channel SNR may also be unstable or
at the same time. Thus, SVM has become the most useful rapid varying under certain conditions. In this case, cur-
classifier for AMC problems in recent years. rent schemes are often lack of generalization ability. To
The performance of FB methods primarily depends solve this problem, a CNN-SVM model for AMC is pro-
on the extracted feature set. Features must be manu- posed in this paper. Considering the advantages of the
ally designed to accommodate the corresponding set of powerful capability of feature learning for deep learn-
modulation and channel environment and may not be ing networks, CNN is deployed to explore new features
feasible in all conditions. Moreover, looking for effective that are suitable for classification under various SNRs. In
features requires great consideration. Considering these this paper, CNN directly handles the received signals at
factors, deep learning (DL) methods, which can auto- mid-frequency from − 10 to 20 dB, and is able to create
matically extract features, have been adopted. DL is a new features robust to SNR variation. The generalization
branch of machine learning and has achieved remarkable ability of AMC under varying SNR conditions can be sig-
success because of its excellent classification ability. DL nificantly improved by these features. The advantages and
has been applied in many fields such as image classifi- contributions of our proposed method in this paper are
cation [14] and natural language processing [15]. Several stated as follows:
typical DL networks such as a deep belief network [16],
• Most current methods identify a limited set of
stacked auto encoder [17], and convolutional neural net-
modulation types, whereas the set of modulations
work (CNN) [18] have been applied in AMC. DL networks
considered in this paper is more complicated and
are commonly deployed as classifiers in most current DL
contains 15 different types in total.
methods. They address different aforementioned features.
• Received signals are directly handled by the DL
The classification accuracy of DL methods has proven
network at intermediate frequency (IF); however,
to be higher than other classifiers, particularly when the
most existing methods still require extra processing
signal-to-noise ratio (SNR) is low.
or transformation before classifying signals.
Currently, most DL-based AMC methods are still imple-
• The method can provide an outstanding classification
mented in two steps: preprocessing and classification.
accuracy under a large SNR range; however, most
The preprocessing can be either transforms or feature
existing method is only feasible under a certain SNR
extraction. DL networks are applied as classifiers to han-
level.
dle preprocessed signals. An AMC method based on
• The CNN built in this paper plays the role of the
DBN was proposed in [19]. The modulation set consists
feature extractor, whereas most DL methods only
of 11 modulation types. Spectra in different orders are
regard DL networks as powerful classifiers. The
calculated for the classification. The classification accu-
features learned by the CNN are displayed and
racy is higher than that of conventional neural networks.
analyzed. The contribution of different convolutional
Zhu and Fujii proposed a high-accuracy classification
kernels is also visualized to better understand the
scenario [20], where 10 different HOC features were
feature learning process.
extracted from 5 modulation types, and SDAE was used
to classify these features. Mendis et al. [16] proposed a The remainder of the paper is organized as follows:
DBN-based method using the SCF of the received sig- the basic model and details of our proposed method is
nals. The classification accuracy is 95% when the SNR is explained in Section 2, followed by the simulation results
− 2 dB. Dai et al. [17] proposed an interclass classifica- and discussion in Section 3. The paper is finally concluded
tion method using the ambiguity function of the signal. in Section 4.
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 3 of 15

2 System model and proposed method modulation phase, respectively, and g(t) is the gate func-
The AMC is an intermediate process that occurs between tion represented as:
signal detection and demodulation at the receiver. The 
1 if 1  t  Ts
structure of our proposed AMC method in com- g(t) = . (3)
0 other
parison with the conventional ones is illustrated in
Fig. 1. For M-QAM (M = 4, 16, 64) signals, we have
Preprocessing in Fig. 1 refers to sampling and quanti- 
zation for IF signals. The procedures inside the dashed s(t) =Am an g(t − nTs ) cos(2πfc t + ϕ0 )
frame, which include the feature extraction, feature selec- n
 , (4)
tion, and classifier, are replaced by the CNN proposed + Am bn g(t − nTs ) cos(2πfc t + ϕ0 )
here. The CNN is pre-trained offline with proper amount n
of samples before it is deployed. Furthermore, as long as  √  √
where an , bn ∈ 2m − 1 − M , m = 1, 2, ..., M, and
the SNR range of the communication channel is known,
two carriers are modulated by an and bn , respectively.
the CNN can learn the features that adapt to the cor-
The OFDM signal, which is the output of a multicarrier
responding condition. This property makes our method
system, can be expressed as
independent from the SNR estimation.
  
s(t) =  (an + jbn ) exp(j2πfn t)
2.1 Signal model n
In this paper, signals are processed in IF and are corrupted  , (5)
= an cos(2πfn t) − bn sin(2πfn t)
by AWGN. Then, the received signal can be denoted as
n

r(t) = s(t) + n(t), (1) where an and bn are the in-phase component and orthog-
onal component of the symbol sequence on the n-th
where s(t) is the transmitted signal of different modula- subcarrier, respectively, and fn is the frequency of the n-th
tion types, n(t) is AWGN, and SNR is defined as Ps /Pn subcarrier.
(Ps is the power of signal and Pn is the power of noise). The LFM signals in a period are denoted as
The modulation set studied in this paper includes M-ASK, 
Ts Ts
M-FSK, M-PSK (M = 2, 4, 8), M-QAM (M = 4, 16, 64), s(t) = Am cos[ 2π(f0 + kt)t] t∈ − , , (6)
2 2
OFDM, MSK, and LFM. For M-ASK, M-FSK, and M-PSK
(M = 2, 4, 8) signals, s(t) is expressed as where k and f0 are defined as the chirp rate and initial
frequency, respectively.
 Finally, for MSK signals, we have
s(t) = Am an g(t −nTs ) cos[ 2π(fc +fm )t +ϕ0 +ϕm ] ,
n πan (k)
(2) s(t) = cos 2πfc t + + ϕk
2Ts , (7)
(k − 1)Ts  t  kTs
where Am , an , Ts , fc , fm , ϕ0 , and ϕm are the modulation
amplitude, symbol sequence, symbol period, carrier fre- where an (k) denotes the k-th symbol in the symbol
quency at IF, modulation frequency, initial phase, and sequence, and ϕk is the phase constant of the k-th symbol.

Fig. 1 The proposed AMC method in comparison with the conventional ones
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 4 of 15

2.2 Convolutional neural network the nearby outputs [23]. Max pooling is used in this paper,
CNNs are simply NNs that use convolution in place of which is an operation that reports the maximum output
general matrix multiplication in at least one of their layers within a pooling window [26]. Assume that the output of
[22]. Typical CNN architectures consist of three differ- a convolutional layer hl is max-pooled. The output hl+1 is
ent types of layers: convolutional layer, pooling layer, and shown as
fully connected layer. There is an extra softmax regression     
layer deployed as the classifier at the last layer of the CNN hl+1
k (i) =max h l
k m l+1
(i−1)+1 , h l
k m l+1
(i − 1) + 2 ,
in supervised learning. In this paper, we replace the fully  
..., hlk ml+1 (i − 1) + Lp ,
connected layer with a global average pooling layer, so that
there is no fully connected layer. (10)

2.2.1 Convolutional layer where i  Llo − Ll+1

p /ml+1 + 1, Ll+1
p is the length of the
In convolutional layers, there are several convolution ker- pooling window; ml+1 is the margin between two adjacent
nels (also known as filters) to process the received signal. pooling windows, which is also known as the stride.
Since the received signal is a 1-dimensional vector in Global average pooling [27] is applied after the last con-
AMC, the kernel is also a 1-dimensional vector. Suppose volutional layer. It takes the average of each feature map,
that the l-th layer of an NN is a convolutional layer, Ns , Lls , and the output vector is directly fed into the softmax
Nkl , and Llk represent the number of inputs, length of the layer. Similarly, we assume that the output of the former
input, number of kernels, and length of kernels of the l-th convolutional layer is hl , which contains the output of
layer, respectively. The convolution operation [23] in the Nkl kernels. The output of global average pooling hLk is
l-th layer is described as follows: represented as
hlk = f xl × Wkl + blk Lo l
1 
∞
hLk = l hl (i), k = 1, 2, ..., Nkl . (11)
 (8) Lo i=1 k
l
x × Wkl (i) = x(a)Wkl (i − a),
a=−∞
2.2.3 Batch normalization
is the set of inputs, W ∈ RNk ×Lk is the
l l
where x ∈ R Ns ×Lls The batch normalization (BN) layer can accelerate deep
set of kernels, and b ∈ RNs is the bias for each output. The network training by reducing the internal co-variate shift
output of the k-th k = 1, 2, ..., Nkl kernel is denoted by [28]. The internal covariate shift is defined as the change
in distribution of output of each layer during train-
(8), and xl × Wkl is the convolution between xl and Wkl . ing. The changes are commonly caused by unbalanced
Assume that the length of the output is Llo . The output nonlinear mapping (e.g., ELU activation). In stochas-
hl ∈ RNk ×Lo is the set of output, which is also known as the
l l
tic gradient descent, a single mini-batch is represented
feature map. f (·) is the activation function to achieve the as B = {x1 , x2 , ..., xm }, and the output yi is normal-
nonlinear mapping of outputs, which is often the sigmoid ized by the BN layer. Suppose that the mean and vari-
or tanh function. In this paper, the exponential linear unit ance of B are denoted as μB and σB , respectively.
(ELU) [24] is selected as the activation function, which is The procedures of BN are shown in Table 1.
denoted as In the BN process, parameters γ and β must be learned
 with the training process of CNN.  is a small quan-
x if x  0
f (x) = . (9) tity added to variance to avoid dividing by zero. BN is
α(ex − 1) if x < 0
The ELU is a simple piecewise function derived from
Table 1 Procedures of batch normalization
the rectified linear unit (ReLU). It is designed to overcome
Input: B, γ , β
gradient vanishing [25] while accelerating the convergence
speed. Output: y = {y1 , y2 , ..., ym }
1. Calculate the mean and variance value of B:
2.2.2 Max-pooling and global average pooling 
μB = m1 m i=1 xi ,
The pooling layer is another important type of layer in 
σB2 = m1 m i=1 (xi − μB )
2
the CNN. As mentioned, the convolutional layer performs
several convolutions to produce a set of outputs, each of 2. Normalize
xi −μB
which runs through a nonlinear activation function (ELU). x̂i = 
σB2 +
Then, a pooling function is used to further modify the out-
3. Scale and shift
put of the layer. A pooling function replaces the output of
the net at a certain location with a summary statistic of yi = γ x̂i + β
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 5 of 15

deployed before the activation function when it is pro- is set to Nc , and the subcarriers are modulated by 4PSK.
posed, but experiments prove that BN should occur after Additionally, we denote the SNR of the training samples
the activation function [29]. As a result, BN is applied after and the SNR of the testing samples as SNRtr and SNRte ,
each activation function in this paper. respectively.
The CNN that we built for AMC consists of 16 convolu-
2.2.4 Softmax regression tional layers, whose structure is similar to that of VGG-19
The last layer of the CNN in supervised learning is the [31] (as shown in Table 3). The input size of the CNN
softmax regression layer. Softmax regression is a multi- can be calculated by fs · Nc /fd = 4000. The parameters
class classifier generalized from logistic regression, whose of the convolutional layers and pooling layers in the l-th
output is a set of probability distributions of different
layer take the form of Nkl , Llk , Llp , ml , respectively. Sig-
classes. Considering an n-class classification problem, the
input of the softmax regression is hL , which is the out- nals are normalized to [ − 1, 1] with zero mean and are
put of the global average pooling layer, and the output of then processed by CNN. The CNN in this paper is imple-
softmax regression yo can be denoted as: mented by the DL library, Keras [32], with Theano [33] as
its backend.
exp (WLc hL + bLc )
P(yo = c|hL , WL , bL ) = n , (12)
i=1 exp(WLi hL + bLi )
3.2 Classification with CNN
where c = 1, 2, ..., n, WL , bL is the weight and bias between In this section, signals are directly classified by CNN, and
the former output and the softmax. The neuron with the the results are from the final softmax layer. The process
maximum output is selected as the classification result, is displayed in Fig. 2. The classification accuracy under
which is also the output of the entire CNN. The loss func- fixed SNR level is firstly displayed in Table 4 for SNRte =
tion of the CNN is defined as J(W , b). Then, the training SNRtr . We generate 20000 training samples and 1000 test-
process is described as ing samples for each modulation type at every SNR level.
arg min J(W , b). (13) As observed from Table 4, the classification accuracy is
W ,b 90% when SNRtr = − 10 dB. This finding demonstrates
The problem in (13) can be solved by a gradient excellent performance for AMC methods. The classifica-
descent. Partial derivatives are calculated using the back- tion accuracy of all individual classes reaches almost 100%
propagation method [30] and used to update W and b. when SNRtr  5 dB. Because our channel is AWGN, the
The process is as follows: signals with amplitude modulation suffer most from the
decreasing SNRtr . The accuracy of 4ASK and 8ASK dra-
∂J(W , b)
W :=W −α matically deteriorates when SNRtr  0 dB. Only 48.8% of
∂W (14) 4ASK are correctly classified under − 10 dB. The accu-
∂J(W , b) racy of 16QAM and 64QAM also rapidly decreases when
b:=b−α ,
∂b SNRtr  − 4 dB.
where α is known as the learning rate, which controls the The detailed classification result when SNRtr = − 10 dB
update step of the parameters. is shown in Table 5. Signals with identical classes but
different orders (e.g., 2ASK, 4ASK, and 8ASK) may be
3 Numerical results and discussion mixed, but there is very little interclass misclassification.
3.1 Simulation parameters For example, nearly half of 4ASK signals are classified
All signals are generated based on the description in as 2ASK and 8ASK, but all of them are M-ASK signals.
Section2, and the parameters of modulation are shown in The intraclass classification result for M-QAM and M-
Table 2. The number of subcarriers in the OFDM signal ASK signals may be unsatisfactory, whereas the interclass
accuracy remains nearly 100%.
Table 2 Modulation parameters
The classification accuracy of CNNs with different
numbers of layers versus SNR is also provided. The result
Parameter Symbol Value
is illustrated in Fig. 3. When the SNR is low, increas-
Code rate fd 2 MHz ing the number of layers can significantly improve the
Carrier frequency fc 70 MHz classification performance of the CNN. However, for
Sampling frequency fs 400 MHz SNRte  0 dB, CNN with five convolutional layers can
Frequency interval (M-FSK) f 10 MHz correctly classify over 99% of all signals. The results for
SNRtr  2 dB are not plotted because they are above
Initial frequency (LFM) f0 (1, 10) MHz
99.5% for all three CNNs. Deeper CNNs can signifi-
Chirp rate k (12, 60) MHz
cantly improve the classification accuracy under low SNR
Number of codes Nc 20 conditions.
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 6 of 15

Table 3 Structure of CNN

Layer number Layer type Parameters Layer number Layer type Parameters
1 Conv (35, 4) 12 Conv (50, 4)
2 Convl (35, 4) 13 Conv (50, 4)
3 Pooling (2, 2) 14 Conv (50, 4)
4 Conv (40, 4) 15 Conv (50, 4)
5 Conv (40, 4) 16 Pooling (2, 2)
6 Pooling (2, 2) 17 Conv (60, 4)
7 Conv (45, 4) 18 Conv (60, 4)
8 Conv (45, 4) 19 Conv (60, 4)
9 Conv (45, 4) 20 Conv (60, 4)
10 Conv (45, 4) 21 Global pooling None
11 Pooling (2, 2) 22 Softmax 15

The above results are obtained under the assump- 3.3 Feature learning with CNN
tion of SNRte = SNRtr . However, SNRtr (commonly For most existing DL-based AMC methods, DL networks
obtained from SNR estimation) is often inaccurate in are treated as classifiers. However, DL networks also have
practice. Moreover, SNRte should be the actual channel the powerful capability of feature learning. Only the last
SNR, which may also be unstable or rapidly varying layer of a CNN (softmax layer) is a classifier; thus, the
under certain conditions (e.g., satellite communication). input of softmax layer hL is equal to the features learned by
To study this problem, the CNN is trained under a cer- the CNN. Thus, we can analyze these features by observ-
tain SNR range to make the trained CNN robust to SNR ing hL (the output of the global average pooling layer). The
variations. In this case, the training SNR is denoted as multi-dimensional scaling (MDS) method [34] is applied
SNRtr ∈[ SNRmin max
tr , SNRtr ] and separately takes values of to map hL , which is a 60-dimensional vector, into a 2-D
[ − 10, 0] dB, [ − 5, 15] dB, and [ 0, 20] dB. The line of axis for convenient observation. Features under SNRtr =
classification accuracy when SNRtr = SNRte is plotted for − 10 dB and SNRtr = 5 dB are normalized and visualized
comparison. The results are shown in Fig. 4. in Fig. 5.
The CNNs trained in a certain SNR range are robust The features of 4PSK and 8PSK signals are completely
to SNR variations when SNRte is in the range of SNRtr . mixed in Fig. 5a, which also shows why these two cat-
The classification accuracy is also notably close to that egories are poorly classified when SNRtr = − 10 dB.
under SNRte = SNRtr . The generalization ability can The situation is similar to 4ASK and 8ASK signals. In
stretch to the higher SNR range when SNRte is not in the contrast to Fig. 5a, the distribution of CNN-learned fea-
range of SNRtr . For CNN trained under [ − 10, 0] dB, the tures in Fig. 5b is much better because of the increase in
classification accuracy can still reach 96% under 20 dB. SNR. Most signals in the same categories are distributed
The CNNs can be robust to SNR variation; thus, they can in the same cluster, and the margins among different
be deployed in a certain SNR range. clusters are evident, which implies that the extracted fea-

Fig. 2 Classification with CNN

Table 4 Classification accuracy of each modulation type (SNRte = SNRtr )
Type M-ASK M-FSK M-PSK M-QAM
OFDM LFM MSK Total
SNRte 2 4 8 2 4 8 2 4 8 4 16 64
− 10 0.950 0.553 0.650 1.0 1.0 0.997 1.0 0.956 0.907 0.999 0.726 0.802 0.985 1.0 1.0 0.9017
Zhou et al. EURASIP Journal on Advances in Signal Processing

−8 0.979 0.579 0.670 1.0 1.0 0.997 1.0 0.997 0.953 1.0 0.845 0.848 1.0 1.0 1.0 0.9245
−6 0.997 0.632 0.712 1.0 1.0 0.999 1.0 0.998 0.991 1.0 0.939 0.886 1.0 1.0 1.0 0.9436
−4 1.0 0.762 0.791 1.0 1.0 1.0 1.0 1.0 0.993 1.0 0.98 0.957 1.0 1.0 1.0 0.9653
−2 1.0 0.881 0.917 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.990 1.0 1.0 1.0 0.9859
0 1.0 0.972 0.972 1.0 1.0 1.0 1.0 0.999 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.9962
(2019) 2019:21

2 1.0 0.992 0.985 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.9984
5 1.0 1.0 0.997 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.9998
10 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
15 1.0 1.0 0.998 1.0 1.0 0.999 1.0 1.0 0.999 1.0 1.0 1.0 1.0 0.996 1.0 0.9995
20 1.0 1.0 0.999 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.996 1.0 0.9999
Page 7 of 15
Table 5 Confusion matrix of the testing samples (SNRtr =− 10 dB)
Classification result
Actual type M-ASK M-FSK M-PSK M-QAM
OFDM LFM MSK
2 4 8 2 4 8 2 4 8 4 16 64
2ASK 950 33 16 0 0 0 0 0 0 0 0 1 0 0 0
4ASK 65 553 382 0 0 0 0 0 0 0 0 0 0 0 0
Zhou et al. EURASIP Journal on Advances in Signal Processing

8ASK 21 329 650 0 0 0 0 0 0 0 0 0 0 0 0

2FSK 0 0 0 1000 0 0 0 0 0 0 0 0 0 0 0
4FSK 0 0 0 0 1000 0 0 0 0 0 0 0 0 0 0
8FSK 0 0 0 0 1 997 0 0 0 0 0 0 0 0 2
2PSK 0 0 0 0 0 0 1000 0 0 0 0 0 0 0 0
(2019) 2019:21

4PSK 0 0 0 0 0 0 0 956 44 0 0 0 0 0 0
8PSK 0 0 0 0 0 0 0 93 907 0 0 0 0 0 0
4QAM 0 0 0 0 0 0 0 0 0 999 0 1 0 0 0
16QAM 0 0 0 0 0 0 0 0 1 8 726 265 0 0 0
64QAM 0 0 0 0 0 0 0 0 0 6 192 802 0 0 0
OFDM 0 0 0 0 0 0 0 0 0 0 0 0 985 15 0
LFM 0 0 0 0 0 0 0 0 0 0 0 0 0 1000 0
MSK 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1000
Page 8 of 15
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 9 of 15

Fig. 3 The classification accuracy of CNNs with different number of layers versus SNR

tures are well suited for classification. The cluster of LFM dimension of the feature under different SNR levels is
signals consists of several small clusters because parts of separately plotted (SNRtr ∈[ − 5, 15] dB) in Fig. 7, where
the modulation parameters of LFM signals are randomly feature 1, feature 2, feature 3, and feature 4 correspond to
generated. the output of the four neurons. We can find that for each
We obtain several CNNs that are robust to SNR vari- modulation type, at least one feature rarely changes with
ations by training them in a certain SNR range as in the the SNR (e.g., feature 1 and feature 4 of 2ASK and feature
previous subsection. We can learn noise-robust features 2 of OFDM and 4PSK). These features are robust to SNR
in a notably similar manner. The dimension reduction to variation; thus, they are expected to provide an excellent
hL is accomplished using the neural networks by adding a generalization ability under the varying SNRte .
hidden layer containing four neurons between the global A linear support vector machine (SVM) is deployed
average pooling layer and softmax layer (see Fig. 6). Thus, to test the generalization ability of the learned features.
the learned features will be 4-dimensional vectors. Each Unlike the previous subsection, the SVM is trained for a

Fig. 4 Classification accuracy of the CNN for different SNR ranges

Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 10 of 15

Fig. 5 Distribution of learned features under SNRtr = − 10 dB and SNRtr = 5 dB (visualized by MDS).a Distribution of CNN-learned features
(SNRtr = − 10 dB). b Distribution of CNN-learned features (SNRtr = 5 dB)

fixed SNR level. In this case, SNRtr takes the values of outstanding. The SVM trained for − 5 dB can correctly
− 5, 0, 5, 10, 15 dB in turn, and the classification accuracy classify 99.1% of signals under 20 dB, and the classification
is tested for [ −10, 20] dB. The results are illustrated in accuracy of signals for − 5 dB is 90% for the SVM trained
Fig. 8. for 15 dB. In this way, the classifiers trained by CNN-
The classification accuracies of all SVMs are notably learned features can reduce their dependency on the SNR
similar. The SVM trained for − 5 dB performs slightly estimation.
worse than the others because the values of the learned The method in [35], which focuses on selecting proper
features fluctuate the most for [ − 5, 0] dB, and most features from manually extracted feature set under vary-
of them are stable when SNRtr  0 dB (see Fig. 7). ing SNR conditions, is chosen for comparison. The signals
The generalization ability of CNN-learned features is are re-generated according to the modulation set in [35],
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 11 of 15

Fig. 6 Feature learning with CNN

and CNN is deployed to extract features under [0, 20] dB. 3.4 Visualization of feature learning process
Then, a linear SVM is deployed to evaluate the perfor- We have proven that the CNN can learn efficient features
mance of our method. SNRtr takes the values of 0, 10, for classification. In this section, the process of feature
20 dB in turn and classification accuracy is tested for [0, learning by CNN is analyzed by visualizing the outputs
20] dB. From the result in Fig. 9, we can observe that the of the intermediate layers. The signals modulated by the
performance is significantly improved, especially under same symbols are in the dashed box with identical colors.
low SNRs, indicating the superiority of CNN-learned We select some typical outputs of intermediate layers and
features. plot them in Fig. 10. For 8FSK signals in Fig. 10a, identical

a c

b d

Fig. 7 Value of learned features versus SNR (SNRtr ∈[-5,20] dB). a Value of feature 1 versus SNR. b Value of feature 2 versus SNR. c Value of feature 3
versus SNR. d Value of feature 4 versus SNR
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 12 of 15

Fig. 8 Classification accuracy of SVM trained using the CNN-learned features

symbols correspond to identical modulation frequencies. ily distinguished in layer 6 when SNRtr = 15 dB, but the
Each convolution kernel retains only a portion of the fre- differences are not obvious when SNRtr = 2 dB. Hence,
quency components. The frequency component of sym- we need more layers when the SNR is low.
bol 6 (in the magenta box) is maintained in the 38th kernel Similar to the frequency information in Fig. 10a, the
of layer 6 but filtered out in the 29th kernel of layer 11, e.g., kernel can also learn the phase information in Fig. 10c.
each kernel concerns only a part of the information from For the 16QAM signal in Fig. 10d, the amplitude infor-
the received signals. By comparing Fig. 10a with Fig. 10b, mation and phase information are recorded by the 20th
we also find that feature learning becomes harder with kernel of layer 6 and the 33rd kernel in layer 11,
the decrease in SNRtr . Different frequencies can be eas- respectively. Thus, symbols 1 and 14, which are mod-

Fig. 9 Comparison with the method in [35]

Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 13 of 15

a c

b d

Fig. 10 Visualization of the outputs of intermediate layers. a Signal type = 8FSK, SNRtr = 15 dB. b Signal type = 8FSK, SNRtr = 2 dB. c Signal type =
8PSK, SNRtr = 15 dB. d Signal type = 16QAM, SNRtr = 15 dB

ulated with identical amplitudes but different phases, generalization ability for certain SNR ranges have been
can be distinguished through a combination of different considered. The numerical results show that the clas-
kernels. sification accuracy can reach 90% under − 10 dB and
is notably close to 100% when the training SNR is
4 Conclusion higher than 5 dB. We have also improved the general-
In this paper, an AMC method based on a CNN has been ization ability by training the CNN under a certain SNR
proposed. First, we have used the CNN as a powerful range. The CNN trained under [ − 10, 0] dB can cor-
classifier. In total, 15 different modulation types have rectly classify 96% of all signals when the testing SNR
been studied, and the classification for fixed SNR and is 20 dB.
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 14 of 15

Then, the features that the CNN learns from the 2. A. Hazza, M. Shoaib, S. A. Alshebeili, A. Fahad, in Communications, Signal
received signals have been analyzed. Features are mapped Processing, and Their Applications (ICCSPA), 2013 1st International
Conference On. An overview of feature-based methods for digital modulation
to a 2-D axis using MDS, where we observe that most sig- classification (IEEE, 2013), pp. 1–6. [Link]
nals in the same categories are distributed in the same document/6487244
3. E. E. Azzouz, A. K. Nandi, Automatic identification of digital modulation
cluster. The margins among different clusters are also evi-
types. Signal Process. 47(1), 55–69 (1995)
dent; thus, they are well suited for classification. Robust 4. J. J. Popoola, R. van Olst, A novel modulation-sensing method. IEEE Veh.
features learned under [ − 5, 15] dB are also studied. Technol. Mag. 6(3), 60–69 (2011)
5. J. Liu, Q. Luo, in Communication Technology (ICCT), 2012 IEEE 14th
Robust features are insensitive to the SNR variation, so International Conference On. A novel modulation classification algorithm
they have strong generalization ability. The SVM trained based on daubechies5 wavelet and fractional fourier transform in
by these robust features under − 5 dB can correctly clas- cognitive radio (IEEE, 2012), pp. 115–120. [Link]
abstract/document/6511199
sify 99.1% of signals when the testing SNR is 20 dB. As 6. Y. Lv, Y. Liu, F. Liu, J. Gao, K. Liu, G. Xie, in Computer and Information
a result, CNNs trained in this way can be robust to SNR Technology (CIT), 2014 IEEE International Conference On. Automatic
variation. modulation recognition of digital signals using CWT based on optimal
scales (IEEE, 2014), pp. 430–434. [Link]
Additionally, we visualize some typical outputs of the document/6984692
intermediate layers. We find that each kernel in the con- 7. D. Das, A. Anand, P. K. Bora, R. Bhattacharjee, in Signal Processing and
Communications (SPCOM), 2016 International Conference On. Cumulant
volutional layer can learn different information from the
based automatic modulation classification of QPSK, OQPSK, π/4-QPSK
received signal. The information includes the phase, fre- and 8-PSK in MIMO environment (IEEE, 2016), pp. 1–5. [Link]
quency, amplitude, and other information that is difficult [Link]/abstract/document/7439996
8. A. Hazza, M. Shoaib, A. Saleh, A. Fahd, Robustness of digitally modulated
for us to understand. signal features against variation in HF noise model. EURASIP J. Wirel.
Commun. Netw. 2011(1), 24 (2011)
9. U. Satija, M. Manikandan, B. Ramkumar, in Industrial and Information
Abbreviations
Systems (ICIIS), 2014 9th International Conference On. Performance study of
AMC: Automatic modulation classification; AWGN: Additive white Gaussian
cyclostationary based digital modulation classification schemes (IEEE,
noise; CNN: Convolutional neural network; DL: Deep learning; ELU: Exponential
2014), pp. 1–5. [Link]
linear unit; FB: Feature based; HOC: High-order cumulant; IF: Intermediate
10. P. M. Rodriguez, Z. Fernandez, R. Torrego, A. Lizeaga, M. Mendicute, I. Val,
frequency; LB: Likelihood based; ReLU: Rectified linear unit; SCF: Spectral
Low-complexity cyclostationary-based modulation classifying algorithm.
correlation function; SNR: Signal to noise ratio; SVM: Support vector machine
AEU Int. J. Electron. Commun. 74, 176–182 (2017)
11. M. W. Aslam, Z. Zhu, A. K. Nandi, Automatic modulation classification
Funding using combination of genetic programming and KNN. IEEE Trans. Wirel.
The research in this article is supported by “the National Natural Science Commun. 11(8), 2742–2750 (2012)
Foundation of China” (Grant nos. 61571167, 61471142, 61102084 and 12. K. Hassan, I. Dayoub, W. Hamouda, M. Berbineau, in Intelligent Transport
61601145) Systems Telecommunications,(ITST), 2009 9th International Conference On.
Automatic modulation recognition using wavelet transform and neural
Availability of data and materials network (IEEE, 2009), pp. 234–238. [Link]
The datasets used during the current study are available from the document/5399351
corresponding author on reasonable request. 13. V. Orlic, M. Dukic, Multipath channel estimation algorithm for automatic
modulation classification using sixth-order cumulants. Electron. Lett.
Authors’ contributions 46(19), 1348–1349 (2010)
SZ proposed the framework of the whole algorithm. ZYin handled all the 14. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V.
simulations. ZW made full contribution in the acquisition of funding. YC was a Vanhoucke, A. Rabinovich, in Proceedings of the IEEE conference on computer
major contributor in writing the manuscript. NZ made all the figures and vision and pattern recognition. Going deeper with convolutions, (2015), pp. 1–9.
tables in the manuscript. ZYang made the visualization in the last section. All [Link]
authors read and approved the final manuscript. Szegedy_Going_Deeper_With_2015_CVPR_paper.html
15. Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y.
Cao, Q. Gao, K. Macherey, et al., Google’s neural machine translation
Competing interests
system: bridging the gap between human and machine translation. arXiv
The authors declare that they have no competing interests.
preprint arXiv:1609.08144 (2016)
16. G. J. Mendis, J. Wei, A. Madanayake, in Communication Systems (ICCS), 2016
Publisher’s Note IEEE International Conference On. Deep learning-based automated
Springer Nature remains neutral with regard to jurisdictional claims in modulation classification for cognitive radio (IEEE, 2016), pp. 1–6. https://
published maps and institutional affiliations. [Link]/abstract/document/7833571/
17. A. Dai, H. Zhang, H. Sun, in Signal Processing (ICSP), 2016 IEEE 13th
Author details International Conference On. Automatic modulation classification using
1 School of Electronics and Information Engineering, Harbin Institute of stacked sparse auto-encoders (IEEE, 2016), pp. 248–252. https://
Technology, No. 92 West Dazhi Street, Nangang District, Harbin, China. [Link]/abstract/document/7877834
2 School of Engineering, University of Warwick, Coventry CV4 7AL, UK. 3 School 18. T. J. O’Shea, J. Corgan, T. C. Clancy, in International Conference on
of Information and Communication Engineering, Dalian University of Engineering Applications of Neural Networks. Convolutional radio
Technology, No.2 Linggong Road, Ganjingzi District, Dalian, China. modulation recognition networks (Springer, 2016), pp. 213–226. https://
[Link]/chapter/10.1007/978-3-319-44188-7_16
19. J. Fu, C. Zhao, B. Li, X. Peng, in The Proceedings of the Third International
Received: 17 October 2018 Accepted: 4 March 2019
Conference on Communications, Signal Processing, and Systems. Deep
learning based digital signal modulation recognition (Springer, 2015),
pp. 955–964. [Link]
References 08991-1_100
1. J. L. Xu, W. Su, M. Zhou, IEEE Trans. Syst. Man Cybernet. Part C Appl. Rev. 20. X. Zhu, T. Fujii, Modulation classification for cognitive radios using stacked
41(4), 455–469 (2011) denoising autoencoders. Int. J. Satell. Commun. Netw. 35(5), 517–531 (2017)
Zhou et al. EURASIP Journal on Advances in Signal Processing (2019) 2019:21 Page 15 of 15

21. D. Zhang, W. Ding, B. Zhang, C. Xie, H. Li, C. Liu, J. Han, Automatic

modulation classification based on deep learning for unmanned aerial
vehicles. Sensors. 18(3), 924 (2018)
22. Y. LeCun, et al., in Connectionism in perspective. Generalization and
network design strategies, vol. 19 (Citeseer, 1989)
23. I. Goodfellow, Y. Bengio, A. Courville, Deep learning. (MIT Press, 2016)
24. D.-A. Clevert, T. Unterthiner, S. Hochreiter, Fast and accurate deep
network learning by exponential linear units (elus). arXiv preprint
arXiv:1511.07289 (2015)
25. S. Hochreiter, The vanishing gradient problem during learning recurrent
neural nets and problem solutions. Int. J. Uncertain. Fuzziness
Knowl.-Based Syst. 6(02), 107–116 (1998)
26. K. Jarrett, K. Kavukcuoglu, Y. LeCun, et al., in Computer Vision, 2009 IEEE
12th International Conference On. What is the best multi-stage architecture
for object recognition? (IEEE, 2009), pp. 2146–2153. [Link]
[Link]/csdl/proceedings/iccv/2009/4420/00/[Link]
27. M. Lin, Q. Chen, S. Yan, Network in network. arXiv preprint arXiv:1312.4400
(2013)
28. S. Ioffe, C. Szegedy, Batch normalization: accelerating deep network
training by reducing internal covariate shift. arXiv preprint
arXiv:1502.03167 (2015)
29. D. Mishkin, N. Sergievskiy, J. Matas, Systematic evaluation of CNN
advances on the ImageNet. ArXiv e-prints 1606.02228 (2016)
30. Y. Bengio, et al., Learning deep architectures for AI. Found. Trends® Mach.
Learn. 2(1), 1–127 (2009)
31. K. Simonyan, A. Zisserman, Very deep convolutional networks for
large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014)
32. F. Chollet, et al., Keras. GitHub (2015). [Link]
faq/#how-should-i-cite-keras
33. R. Al-Rfou, G. Alain, A. Almahairi, C. Angermueller, D. Bahdanau, N. Ballas, F.
Bastien, J. Bayer, A. Belikov, A. Belopolsky, et al., Theano: A python
framework for fast computation of mathematical expressions. arXiv
preprint arXiv:1605.02688 (2016)
34. A. Buja, D. F. Swayne, M. L. Littman, N. Dean, H. Hofmann, L. Chen, Data
visualization with multidimensional scaling. J. Comput. Graph. Stat. 17(2),
444–472 (2008)
35. Z. Wu, S. Zhou, Z. Yin, B. Ma, Z. Yang, Robust automatic modulation
classification under varying noise conditions. IEEE Access. 5, 19733–19741
(2017)