Hindawi

Mathematical Problems in Engineering

Volume 2020, Article ID 2678310, 13 pages
https://doi.org/10.1155/2020/2678310

Research Article
Automatic Modulation Classification Based on Deep Learning for
Software-Defined Radio

Peng Wu,1 Bei Sun ,1 Shaojing Su,1 Junyu Wei,1,2 Jinhui Zhao,3 and Xudong Wen4
1
College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China
2
College of Electric and Information Engineering, Hunan University of Technology, Zhuzhou 412007, China
3
Teaching and Research Support Center, National University of Defense Technology, Changsha 410073, China
4
93920 Unit of the Chinese People’s Liberation Army, Hanzhong, China

Correspondence should be addressed to Bei Sun; [email protected]

Received 13 May 2020; Revised 7 September 2020; Accepted 24 October 2020; Published 27 November 2020

Academic Editor: Rafal Zdunek

Copyright © 2020 Peng Wu et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

With the development of artificial intelligence technology, deep learning has been applied to automatic modulation classification
(AMC) and achieved very good results. In this paper, we introduced an improved deep neural architecture for implementing radio
signal identification tasks, which is an important facet of constructing the spectrum-sensing capability required by software-
defined radio. The architecture of the proposed network is based on the Inception-ResNet network by changing the several kernel
sizes and the repeated times of modules to adapt to modulation classification. The modules in the proposed architecture are
repeated more times to increase the depth of neural network and the model’s ability to learn features. The modules in the proposed
network combine the advantages of Inception network and ResNet, which have faster convergence rate and larger receptive field.
The proposed network is proved to have excellent performance for modulation classification through the experiment in this paper.
The experiment shows that the classification accuracy of the proposed method is highest with the varying SNR among the six
methods and it peaks at 93.76% when the SNR is 14 dB, which is 6 percent higher than that of LSTM and 13 percent higher than
that of MentorNet, Inception, and ResNet purely. Besides, the average accuracy from 0 to 18 dB of the proposed method is 3
percent higher than that of GAN network. It will provide a new idea for modulation classification aiming at distraction time signal.

1. Introduction It is also essential for identifying the source of received

wireless signals [6, 7].
With the rapid development of communication technology, At present, AMC can be divided into two categories:
the wireless communication environment is becoming more likelihood-based (LB) and feature-based (FB) [8]. The LB
and more complex and communication signals with various modulation classifier recognizes the modulation of signal by
types of modulation are becoming more diverse and com- comparing the likelihood function value of received signal
plex [1]. Automatic modulation classification (AMC) plays within the known modulation pool [6]. It has been used for
an important role in modern wireless communication [2]. modulation classification in multiple channel environment
The signal analysis and processing can be carried out only with high accuracy [9]. While it needs some parameters to be
when the modulation of the signal is recognized [3]. It finds known in advance, such as carrier frequency, code rate, and
applications in various commercial and military areas. For channel parameters, it becomes very complex when un-
example, recognition of the modulation type is used in known parameters are introduced. Thus, it is difficult to
software-defined radio (SDR) in order to adapt to various design a system for the acquisition of signal. Some re-
communication systems without requiring control overhead searchers have studied the way to simplify the likelihood
in the shortest time [4, 5]. Under such conditions, advanced function, which will lead to information absence and in-
automatic modulation classification techniques are required. accurate results [10]. As LB method is sensitive to parameter
2 Mathematical Problems in Engineering

estimation deviations or model mismatches, it is not ap- two convolutional neural network- (CNN-) based DL
plicable in many practical communication scenarios. models, AlexNet and GoogLeNet [25]. Cheong et al. ex-
For FB method, features of the received signal are tended the previous work on automatic modulation clas-
extracted and the modulation of signal can be identified by sification (AMC) by using deep neural networks (DNNs)
either comparing features with threshold values or feeding and evaluated the performance of these architectures on
feature to pattern recognizer [11, 12]. In many traditional signals [26]. Xie et al. proposed two algorithms, including
pattern recognition methods, it is necessary to extract fea- M2M4-aided algorithm and multi-label DL based algorithm,
tures of signal manually, such as instantaneous statistics, to combat the varying SNR [27]. Peng et al. converted the
high-order statistics, time-frequency characteristics, asyn- raw modulated signals into images that had a grid-like to-
chronous delay sampling characteristics, etc. [13]. Then, pology and fed them to CNN for network training [28].
these features are used as input of the classifier, such as The dataset used in this paper was released on Pro-
decision tree and support vector machine [14]. Although it is ceedings of the GNU Radio Conference in 2016 for appli-
simple with less computation, it shows poor performance for cations of machine learning (ML) to the Radio Signal
non-linear problems. Apart from this, features selected Processing domain [29]. Li et al. used GAN network with
manually may not reflect the characteristics of signals with data argument for modulation classification testing on this
different modulation and improper feature selection will dataset and the highest accuracy was no more than 90% [30].
reduce the classification recognition accuracy of the LSTM, Inception, and residual network were also used for
classifier. modulation classification on this dataset, while the highest
In recent years, great progress has been made in artificial accuracy among them was around 88%. Miao Du proposed a
intelligence and a single computer chip’s computing power new network structure called fully dense neural network
has been greatly improved, which promotes deep learning (FDNN) for the automatic modulation classification and the
algorithms to be widely used in modulation classification average accuracy with FDNN is 89.6% from 0 to 18 dB [31].
[15, 16]. It solves the core problem of how to automatically Yuan Zeng proposed CNN architecture with spectrogram
select and extract the features of samples. Besides, it realizes images as signal representation and achieved good recog-
the combination of simple features into more efficient and nition accuracy in 2019, while the average accuracy from 0 to
complex features to achieve classification recognition [17]. 18 dB of it was less than 90% [32]. Wu et al. proposed
In addition, deep neural networks have a multi-layer convolutional neural network and multi-feature fusion for
structure, which can better extract features of signal avoiding automatic modulation classification. The overall classifica-
the tedious manual selection of data features [18]. At present, tion accuracy of it is less than 93% [33]. In addition, Zhang
CNN, Google’s Inception, and residual network (ResNet) et al. proposed an automatic digital modulation classifica-
have been used in modulation classification with good tion network based on curriculum learning and the overall
results. classification accuracy is less than 90% [34]. The highest
CNN-based modulation classification method was accuracy of the proposed method in this paper is 93.76% at
proposed and it can be used for the sampling sequence of the 14 dB and the average accuracy from 0 to 18 dB with the
intermediate frequency signal directly [19]. Rajendran et al. proposed method is 93.04%, which has excellent
proposed that recurrent neural network (RNN) can also be performance.
used for modulation classification of the sampling sequence This paper is organized as follows: Section 2 introduces
of the intermediate frequency signal [20]. Hu et al. studied the related work. Section 3 formulates the basic principle of
the effect of different noise on the modulation classification the ResNet. Section 4 gives the basic principle of Inception
by using RNN [21]. network. Section 5 details the proposed method in this
Generated adversarial network (GAN) was proposed for paper. Section 6 presents the results of simulations and
data argument with RGB three-channel constellation in experiments to support the theoretical analysis.
modulation classification filed [22]. Wang et al. established a
2-level convolutional neural network (CNN) architecture to 2. Related Work
distinguish 16QAM from 64QAM. In the second CNN
architecture, 16QAM and 64QAM constellation diagrams 2.1. Basic Principle of Signal Modulation. At present, digital
were used as input to obtain a higher recognition rate [23]. modulation technology has been used widely in wireless
However, this method relies on the constellation diagram communication. Although signal processing techniques
and can only identify the baseband signal. The accuracy is such as modulation and demodulation of digital modulation
very low when modulated signals are sampled on the radio are more difficult than analog modulation and the signal
frequency directly. A convolutional long short-term deep processing system is more complex for digital modulation,
neural network (CLDNN) was proposed by combining the digital modulation technology has strong anti-interference
architectures of CNN and long short-term memory (LSTM) and it is easy to use modern digital signal processing
into a deep neural network, which takes advantage of CNNs, technology to process and analyze the signal. Therefore,
LSTMs, and conventional deep neural network architectures digital modulation signals are widely used in practice. The
and enables the learning of long-term dependencies [24]. A modulation classification of digital modulation signals is
ResNet architecture was proposed to distinguish between studied in this paper.
different modulation types by adding the bypass connections A general expression for the received baseband complex
and showed good performance [17]. Shengliang P proposed envelope is
Mathematical Problems in Engineering 3

r(t) � s t; ui 􏼁 + n(t), (1) where s(t; ui ) denotes baseband complex envelope of the
received signal with no noise, and n(t) is noise. s(t; ui ) can
be expressed as

K
s t; ui 􏼁 � ai ej2πΔft ejθ 􏽘 ejϕk s(i)
k g(t − (k − 1)T − εT), 0 ≤ t ≤ KT, (2)
k�1

where ai represents the amplitude of signal, Δf denotes the classification accuracy is computed over the classification
offset of carrier frequency, θ is the time-invariant carrier output for a testing dataset.
phase, ϕk is the phase jitter, T is the symbol period, ε is the Figure 1 illustrates a diagram of deep learning for
normalized epoch for time offset between the transmitter modulation classification. Generally speaking, the classifi-
and signal receiver, g(t) � Ppulse (t) ⊗ h(t) is the composite cation method mainly includes three parts: signal pre-
effect of the residual channel with h(t) denoting the channel processing, feature extraction, and evaluation. The deep
impulse response and ⊗ denoting mathematical convolu- neural network can automatically combine simple basic
tion, and Ppulse (t) is the transmit pulse shape. In equation features into more complex features gradually, achieve ef-
(2), ui is the multidimensional vector that includes the fective feature expression of data samples, and maintain high
deterministic unknown signal or channel parameters for the recognition accuracy. In this paper, a supervised recognition
ith modulation type. The goal of modulation classification is method is used. Firstly, a large number of labeled samples
to recognize the modulation type i from the received signal are used to train the deep neural network, and then the
r(t) [30]. The Amplitude Shift Keying (ASK), Frequency trained model is used to recognize the unknown samples.
Shift Keying (FSK), Phase Shift Keying (PSK), and Quad-
rature Phase Shift Keying (QPSK) are commonly used 3. Residual Network
modulation types.
Although neural networks with more layers have better
learning ability, the degeneration occurs sometimes, which
2.2. Deep Neural Network for Classification Modulation. leads to low accuracy. An effective approach so far, which
Convolutional layers are a common element in all state-of- won ImageNet 2015, is residual networks. A residual net-
the-art deep neural networks. A convolutional layer usually work adds one layer’s output to the output of the layer two
consists of convolutional filters. The size of convolution layers deeper, as shown in Figure 2. Vanishing gradients are
kernels is typically very small, such as 1 × 1 through 5 × 5 resolved by normalization techniques that have been widely
sizes. The transfer function for a standard convolutional adopted and that network depth is instead limited by
layer [6] is given below: training complexity of deep networks which can be sim-
plified with residual functions [36].
⎝b + 􏽘 k ∗ x ⎞
yi � f⎛ ⎠
j ij i , (3) ResNet improves the original network structure by
i adding connection identity mapping. The learning function
is defined as F(x) + x and the input is added to the learned
where yi is the output feature map for the ith filter, b and k
function. By adopting this method, the parameters and
represent learned bias and filter weight parameters, xi
calculation amount of the network will not increase, and the
represents the input activations, ∗ denotes the convolution
training speed and recognition accuracy of the model will be
operation, and f(··) denotes a (typically non-linear) acti-
significantly improved.
vation function [35].
A visible trend in neural networks for classification task
is building deeper networks to learn more complex functions 4. Inception Network
and hierarchical feature relationships. Deep networks enable
The inception architecture is one successful approach to
more complex functions to be learned more readily from raw
increasing the depth of network and learning ability. This
data.
network consists of repeated inception modules. As shown
Typically, applying deep neural networks to solve
in Figure 3, each inception module contains four parallel
modulation classification is a matter of
paths with the output being the concatenation of the four
(i) Designing a network architecture parallel outputs. The first path is a bank of 1 × 1 convolutions
(ii) Training the network to select weights which that forward along selected information. The 1 × 1 convo-
minimizes loss lutions are a form of selective highway networks that simply
pass information forward with no transformation. The
(iii) Validating and testing in practice to solve problem
second and third paths are 1 × 1 convolutions followed by a
To this end, we use various machine learning classifiers bank of 3 × 3 and 5 × 5 convolutions to make the network
based on deep neural network architectures, where a have larger receptive field. The last parallel path is a 3 × 3
training dataset is used to train the network, and then the pooling layer followed by 1 × 1 convolutions. Intermediate
4 Mathematical Problems in Engineering

Modulated Preprocessing Automatic feature extraction Trained model Evaluation

signal received by deep neural network
Figure 1: Flow diagram of deep learning for modulation classification.

for modulation classification through the experiment in this

x paper.
For the residual versions of the Inception networks,
cheaper Inception blocks than the original Inception are
Weight layer used. Each Inception block is followed by filter-expansion
layer (1 × 1 convolution without activation) which is used for
F (x) ReLU
x scaling up the dimensionality of the filter bank before the
Identity addition to match the depth of the input. This is needed to
Weight layer compensate for the dimensionality reduction induced by the
Inception block.
Figure 4(a) shows the overall schema of the proposed
+ network; the input dimension of input data is 1 × 2 × 128
which is set in the dataset for modulated signal. There are
F (x) + x ReLU three Inception-ResNet modules in Figure 4(a); the structure
of each Inception-ResNet module is shown in Figures 5(a)–
Figure 2: Residual network diagram. 5(c), respectively. The module A is repeated 10 times. 20
times and 10 times are repeated for module B and module
C. In order to make residual operation feasible, 1 × 1 con-
Filter volution is added to match the depth of the network in each
concatenation Inception-ResNet module. The schema for the stem of the
proposed method is shown in Figure 4(b), where there are 6
convolutional layers and 1 max pooling layer.
3 × 3 conv 5 × 5 conv 1 × 1 conv
1 × 1 conv
3×3 5.2. Loss Function. For a multi-class classification task such
1 × 1 conv 1 × 1 conv max pooling
as modulation recognition, the objective function is often
categorical cross-entropy. Categorical cross-entropy
Previous layer
(equation (4)) is a measure of difference between two
probability distributions. For deep learning classification
Figure 3: Inception unit diagram. tasks, the probability distribution is usually a Softmax
(equation (5)) of the output of the classifier network which is
then converted to one-hot encoding for classification pur-
inception modules in the network are connected to Softmax poses. The error is calculated in what is known as the for-
classifiers that contribute to the network’s global loss for ward-pass and weights are adjusted using the chain rule to
training. These classifiers are believed to help in increasing find error contribution for each parameter in what is known
the model’s ability to learn features [37]. as the backward-pass. This kind of network output layer,
optimization, and loss function have been used very suc-
5. Proposed Method cessfully for multi-class vision tasks such as object recog-
nition on the ImageNet dataset [39].
5.1. The Structure of the Proposed Network. The architecture
of the proposed network is based on the Inception-ResNet H(p, q) � − 􏽘 p(x)log q(x). (4)
network by changing the several kernel sizes and the re- x
peated times of modules to adapt to modulation classifi-
In equation (4), p(x) and q(x) are two probabilities and
cation for software-defined radio [38]. The architecture of
H(p, q) is a value to measure the difference of two
the proposed network consists of stem network module,
probabilities.
three Inception-ResNet modules, and reduction modules.
This network combines the advantages of the Inception ezj
module and ResNet module, which has faster convergence σ(z)j � zk , for j � 1, . . . , Kzj . (5)
􏽐K
k�1 e
rate. The network is proved to have excellent performance
Mathematical Problems in Engineering 5

Softmax Output: 10 1 × 3 conv

(stride 2) Output: 192 × 2 × 13

Dropout Output: 1536 1 × 3 conv

(192 V) Output: 192 × 2 × 28

Average pooling Output: 1536

1 × 1 conv
(80 V) Output: 80 × 2 × 30

10×
Output: 1536 × 2 × 2
Inception-ResNet-C
1 × 3 max pooling
(sride 2) Output: 64 × 2 × 30

Reduction Output: 2080 × 2 × 2

1 × 3 conv
(64) Output: 64 × 2 × 61
20×
Output: 1088 × 2 × 6
Inception-ResNet-B

1 × 3 conv
(32 V) Output: 32 × 2 × 61
Reduction Output: 1088 × 2 × 6

1 × 3 conv
10× Output: 32 × 2 × 63
Output: 320 × 2 × 13 (32 stride 2 V)
Inception-ResNet-A

Stem Output: 192 × 2 × 13 Input

(1 × 2 × 128) 1 × 2 × 128

Input
1 × 2 × 128 1 × 2 × 128

(a) (b)

Figure 4: The overall schema and stem of the proposed network. (a) The overall schema of proposed network. (b) The schema for stem of the
proposed network.

In equation (5), the standard exponential function is The time segments were sampled randomly from the
applied to each element of the input vector z and these output stream of each simulation and stored in an output
values are normalized by dividing by the sum of all these vector commonly used for Keras, Theano, and TensorFlow.
exponentials; this normalization ensures that the sum of the The dataset uses a 4D real float 32 vector, taking the form
components of the output vector σ(z) is 1. Nexamples × Nchannels × Dim1 × Dim2, where each exam-
ple consists of 128 complex floating point time samples.
6. Experiment and Discussion Nchannels � 1, a representation which commonly is used for
RGBA values in imagery, Dim1 � 2 holding I and Q chan-
6.1. Dataset. We use the RadioML2016.10b dataset as a basis nels, and Dim2 � 128.
for evaluating the modulation recognition task. The dataset
adopted in this paper was first released at the 6th Annual
GNU Radio Conference in 2016. The dataset allows machine 6.2. Training Process. The dataset is divided into two parts:
learning researchers with new ideas to dive directly into an 960000 samples are used for training the deep neural network
important technical area without the need for collecting or and 240000 samples for validation. All models and training are
generating new datasets and allows for direct comparison to done with the Keras with TensorFlow as a deep learning library
efficacy of prior work. The dataset is generated with GNU using a TITAN RTX 24G GPU. The Adam optimizer was used
Radio, consisting of digital and analog modulations at for all architectures, and the loss function was the categorical
varying signal-to-noise ratios. Details about the generation cross-entropy function. We also used ReLU activation functions
of this dataset can be found in [30]. Figure 6 shows a high- for all layers, except the last dense layer where we used Softmax
level framework of the data generation. activation functions. We used a minimum batch size of 128 and
6 Mathematical Problems in Engineering

ReLU activation

1 × 1 conv
(320 linear)

3 × 3 conv
(64)
3 × 3 conv
(32) 3 × 3 conv
1 × 1 conv (48)
(32) 1 × 1 conv
(32)
1 × 1 conv
(32)

ReLU activation

(a)

ReLU activation ReLU activation

+ +

1 × 1 conv 1 × 1 conv
(1088 linear) (1536 linear)

7 × 1 conv 3 × 1 conv
(192) (256)

1 × 1 conv 1 × 7 conv 1 × 1 conv 1 × 3 conv

(192) (160) (192) (224)

1 × 1 conv 1 × 1 conv
(128) (192)

ReLU activation ReLU activation

(b) (c)

Figure 5: The schema for the three Inception-ResNet modules of the proposed network. (a) The schema for Inception-ResNet module (A).
(b) The schema for Inception-ResNet module (B). (c) The schema for Inception-ResNet module C.

Source Transmitter Channel Dataset

alphabet model model storage

Figure 6: A frame of data generation.

a learning rate of 0.01 and 0.001. Figure 7 shows the flow di- (a) Batch Normalization
agram of our experiment. Normalize the input data between − 1 and 1. This
trick can be related to
6.3. Trick during the Process. In order to get better perfor-
mance and avoid overfitting, we introduce several tricks in x − mean(x)
x′ � , (6)
training. max(x) − min(x)
Mathematical Problems in Engineering 7

Training Training the proposed

dataset network

RML2016b
Preprocessing
dataset

Validation
Model trained Evaluation
dataset

Figure 7: The flow diagram of our experiment.

where x is the input data for the network and x′ is

normalized data. This trick can accelerate training Training performance
and prevent overfitting. 1.2
(b) Early stopping
A compromise is to train on the training dataset but 1.1
to stop training at the point when performance on a
validation dataset starts to degrade. This simple, 1.0
effective, and widely used approach to training

Loss
neural networks is called early stopping. Stopping 0.9
the training of a neural network early before it has
overfit the training dataset can reduce overfitting and 0.8
improve the generalization of deep neural networks.
0.7
The challenge of training a neural network long
enough to learn the mapping, but not so long that it 0.6
overfits the training data.
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5
Model performance on a holdout validation dataset
can be monitored during training and training can Epoch
be stopped when generalization error starts to
Train loss + error
increase.
The use of early stopping requires the selection of a Val_error
performance measure to monitor, a trigger to stop Figure 8: The results of training and validation loss in our
training, and a selection of the model weights to experiment.
use.
In this paper, we stop training process when the parameters are compared in order to get higher classification
validation loss does not decrease within 10 epochs. accuracy. Here, the influence of different learning rates and
As shown in Figure 8, the validation loss is 0.8820 at batch sizes is considered. The initial learning rate (lr) is
the 8th epoch and the validation loss is more than initialized to 0.001 and 0.01, and the batch size is initialized
0.8820 in the following 10 epochs. In this connection, to 128, 512, and 1024. It is detected that too high or too low
we adopt the parameters at the 8th epoch to evaluate learning rates and batch sizes could yield homogeneous
the performance of the model. prediction and low accuracy. Figure 9 illustrates the overall
(c) Dropout classification accuracy of the six methods from − 20 dB to
18 dB with different parameters.
Assuming that the probability of discarding neurons
As shown in Figure 9, overall classification accuracy of
in each layer is p, then the remaining probability is
the proposed method and five other methods from − 20 dB to
1 − p. This trick can prevent the network from
18 dB with different parameters is compared. In this ex-
overfitting to a certain extent. In the dropout module
periment, the highest classification accuracy in Figure 9 for
in this paper, the remaining probability is 0.8.
each method is adopted as baseline. The baseline result is
gotten from CNN, pure ResNet, pure Inception, LSTM, and
6.4. Result and Discussion MentorNet network on the RML2016b dataset. The accuracy
of the baseline and the proposed method on the validation is
6.4.1. Baseline. In this paper, the performance of the pro- shown in Table 1.
posed method is compared with five methods including
CNN, ResNet, Inception, LSTM, and one of the state-of-the-
art methods MentorNet in [34]. The optimizer adopted in 6.4.2. Overall Classification Accuracy. Figure 10 shows the
this paper is Adam. The default value of Adam optimizer overall classification accuracy of CNN, LSTM, Inception,
(beta1 � 0.9, beta2 � 0.999, epsilon � 1e − 8, decay � 0) is set. ResNet, MentorNet, and the proposed method on the val-
For each method, the performances under different idation dataset from − 20 dB to 18 dB. It can be seen that the
8 Mathematical Problems in Engineering

Classification accuracy of CNN Classification accuracy of ResNet Classification accuracy of Inception

1.0 1.0 1.0
Classification accuracy

Classification accuracy

Classification accuracy
0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

–20 –15 –10 –5 0 5 10 15 –20 –15 –10 –5 0 5 10 15 –20 –15 –10 –5 0 5 10 15
Signal-to-noise ratio Signal-to-noise ratio Signal-to-noise ratio

Ir = 0.001, batch size = 128 Ir = 0.001, batch size = 128 Ir = 0.001, batch size = 128
Ir = 0.001, batch size = 512 Ir = 0.001, batch size = 512 Ir = 0.001, batch size = 512
Ir = 0.001, batch size = 1024 Ir = 0.001, batch size = 1024 Ir = 0.001, batch size = 1024
Ir = 0.01, batch size = 128 Ir = 0.01, batch size = 128 Ir = 0.01, batch size = 128
Ir = 0.01, batch size = 512 Ir = 0.01, batch size = 512 Ir = 0.01, batch size = 512
Ir = 0.01, batch size = 1024 Ir = 0.01, batch size = 1024 Ir = 0.01, batch size = 1024

(a) (b) (c)

Classification accuracy of LSTM Classification accuracy of MentorNet Classification accuracy of proposed method
1.0 1.0 1.0
Classification accuracy

Classification accuracy

Classification accuracy
0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

–20 –15 –10 –5 0 5 10 15 –20 –15 –10 –5 0 5 10 15 –20 –15 –10 –5 0 5 10 15
Signal-to-noise ratio Signal-to-noise ratio Signal-to-noise ratio

Ir = 0.001, batch size = 128 Ir = 0.001, batch size = 1024 Ir = 0.001, batch size = 128
Ir = 0.001, batch size = 512 Ir = 0.01, batch size = 1024 Ir = 0.001, batch size = 512
Ir = 0.001, batch size = 1024 Ir = 0.001, batch size = 512 Ir = 0.001, batch size = 1024
Ir = 0.01, batch size = 128 Ir = 0.001, batch size = 128 Ir = 0.01, batch size = 128
Ir = 0.01, batch size = 512 Ir = 0.01, batch size = 512 Ir = 0.01, batch size = 512
Ir = 0.01, batch size = 1024 Ir = 0.01, batch size = 128 Ir = 0.01, batch size = 1024

(d) (e) (f )

Figure 9: The overall classification accuracy of the six methods from − 20 dB to 18 dB with different parameters. (a) The overall classification
accuracy of CNN with different parameters. (b) The overall classification accuracy of ResNet with different parameters. (c) The overall
classification accuracy of Inception with different parameters. (d) The overall classification accuracy of LSTM with different parameters. (e)
The overall classification accuracy of MentorNet with different parameters. (f ) The overall classification accuracy of the proposed method
with different parameters.

Table 1: Test accuracy of baseline and the proposed method.

SNR (dB) CNN (%) ResNet (%) Inception (%) LSTM (%) MentorNet (%) Proposed method (%)
− 10 25.71 26.92 23.61 29.56 24.86 30.40
−4 54.32 48.53 50.86 58.46 62.45 74.30
0 77.64 75.41 73.65 83.19 76.36 90.81
4 80.53 80.25 78.87 85.96 78.59 92.84
10 81.61 81.69 79.91 87.09 80.16 93.29

classification accuracy of the six methods increases gradually varying SNR among the six methods. The classification
and remains stable with the increase of SNR. The classifi- accuracy of the proposed method reaches 90% at 0 dB, while
cation accuracy of the proposed method is highest with the the accuracy of the other methods is less than 90%. The
Mathematical Problems in Engineering 9

Classification accuracy of different networks

1.0

0.9

0.8

Classification accuracy
0.7

0.6

0.5

0.4

0.3

0.2

0.1
–20 –15 –10 –5 0 5 10 15 20
Signal-to-noise ratio

Proposed network Inception

LSTM ResNet
CNN MentorNet
Figure 10: The overall classification accuracy of the six methods.

classification accuracy of the proposed method peaks at networks. It can be seen that the classification accuracy of
93.76% when the SNR is 14 dB and it is 6 percent higher than each modulation type using the six methods shows a general
that of LSTM and 13 percent higher than that of MentorNet, upward trend gradually and fluctuates within a narrow range
Inception, and ResNet purely. The performance of the with the increase of SNR. The classification accuracy of 8PSK,
proposed method is verified by this experiment. BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, and QPSK is
more than 95% using the proposed method, which has the
highest number of modulation types whose classification
6.4.3. Classification Accuracy of Each Kind of Modulated accuracy is more than 95% among the six methods. We can
Signal. Figure 11 shows the normalized classification of each also see that the classification accuracy of GFSK and QAM64
kind of modulated signal including 8PSK, AM-DSB, BPSK, is close to 100% using the proposed method, while that of the
CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK, and WBFM other four methods is less than 80%. In addition, the clas-
by the six methods when the SNR is 18 dB. In Figure 11, the sification accuracy of WBFM is close to 100% using the
vertical column represents the true label of the modulated proposed method, while that of the other four methods is less
signal and the horizontal row represents the predicted label than 40%. We can also see that the classification accuracy of
gotten from the deep neural network. All data is normalized. AM-DSB increases and then declines from − 20 dB to 18 dB
In this connection, the data on the diagonal is the classifi- using the proposed method, which shows the proposed
cation accuracy. For example, in Figure 11(a), there are two method is not well used to recognize AM-DSB modulation
non-zero values in the top row, that is, 0.95 and 0.05, which type. Overall, the proposed method has the best performance
indicates that the classification accuracy is 95% for 8PSK in classifying the modulated signals for the RML2016b dataset
modulated signal by the proposed method and there are 5% among the six methods adopted in this paper.
8PSK signals that are recognized as QPSK signals. In this
connection, it can be found that the classification accuracy of
8PSK, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, 6.4.4. Computation Complexity. Table 2 shows the com-
QPSK, and WBFM gotten from the proposed method is putation complexity of different methods. Total parameters
highest among the six methods when the SNR is 18 dB. indicate the number of parameters in different models.
From Figure 11, we can get the classification accuracy of Training and inference time refer to the cost of time in
each modulated signal gotten from the six methods when the training and validation process. FLOPs (floating point of
SNR is 18 dB. In this context, we can get the classification operations) can also represent the complexity of deep neural
accuracy of each modulated signal using the proposed network. From the table, we can see that the number of total
method, LSTM, CNN, Inception, and ResNet from − 20 dB to parameters in the proposed network is highest among the six
18 dB, which is shown in Figure 12. methods. The training time, inference time, and FLOPs are
Figure 12 indicates the classification accuracy of each also higher than other methods. In this connection, it will
modulated signal from − 20 dB to 18 dB using the proposed take more time to use the proposed method to get the trained
method, LSTM, CNN, Inception, ResNet, and MentorNet model.
10 Mathematical Problems in Engineering

1.0
8PSK 0.88 0.00 0.01 0.02 0.00 0.00 0.00 0.01 0.08 0.00
8PSK 0.97 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.00
AM-DSB 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
AM-DSB 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.8
0.8
BPSK 0.01 0.00 0.97 0.00 0.00 0.00 0.00 0.00 0.02 0.00
BPSK 0.01 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00
CPFSK 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
CPFSK 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.6
0.6

True label
GFSK 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.02
True label

GFSK 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.02
PAM4 0.00 0.00 0.01 0.00 0.00 0.97 0.01 0.01 0.00 0.00
PAM4 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.4
0.4
QAM16 0.08 0.00 0.00 0.00 0.00 0.00 0.22 0.69 0.01 0.00
QAM16 0.02 0.00 0.00 0.00 0.00 0.00 0.28 0.70 0.00 0.00
QAM64 0.05 0.00 0.00 0.00 0.00 0.00 0.13 0.81 0.01 0.00
QAM64 0.01 0.00 0.00 0.00 0.00 0.00 0.17 0.82 0.00 0.00 0.2
0.2
QPSK 0.16 0.00 0.03 0.01 0.00 0.00 0.00 0.00 0.80 0.00
QPSK 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.98 0.00
WBFM 0.00 0.63 0.00 0.00 0.08 0.00 0.00 0.00 0.00 0.30
WBFM 0.00 0.61 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.37 0.0
0.0

8PSK

AM-DSB

BPSK

CPFSK

GFSK

PAM4

QAM16

QAM64

QPSK

WBFM
8PSK

AM-DSB

BPSK

CPFSK

GFSK

PAM4

QAM16

QAM64

QPSK

WBFM
Predicted label
Predicted label

(a) (b)
1.0 1.0
8PSK 0.85 0.00 0.00 0.01 0.00 0.00 0.02 0.00 0.12 0.00 8PSK 0.95 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.00

AM-DSB 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 AM-DSB 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02
0.8 0.8
BPSK 0.01 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 BPSK 0.01 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00

CPFSK 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 CPFSK 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
0.6 0.6
True label

True label

GFSK 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.01 GFSK 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00

PAM4 0.00 0.00 0.01 0.00 0.00 0.99 0.00 0.00 0.00 0.00 PAM4 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00
0.4 0.4
QAM16 0.04 0.00 0.00 0.00 0.00 0.00 0.56 0.40 0.00 0.00 QAM16 0.05 0.00 0.00 0.00 0.00 0.00 0.73 0.22 0.00 0.00

QAM64 0.02 0.00 0.00 0.00 0.00 0.00 0.44 0.53 0.00 0.00 QAM64 0.03 0.00 0.00 0.00 0.00 0.00 0.34 0.62 0.01 0.00
0.2 0.2
QPSK 0.22 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.76 0.00 QPSK 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00

WBFM 0.00 0.62 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.35 WBFM 0.00 0.57 0.00 0.00 0.08 0.00 0.00 0.00 0.00 0.43
0.0 0.0
8PSK

AM-DSB

BPSK

CPFSK

GFSK

PAM4

QAM16

QAM64

QPSK

WBFM

8PSK

AM-DSB

BPSK

CPFSK

GFSK

PAM4

QAM16

QAM64

QPSK

WBFM
Predicted label Predicted label

(c) (d)
1.0 1.0
8PSK 0.97 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.01 0.00 8PSK 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

AM-DSB 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 AM-DSB 0.00 0.64 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.36
0.8 0.8
BPSK 0.01 0.00 0.98 0.00 0.00 0.00 0.00 0.00 0.00 0.00 BPSK 0.00 0.00 0.99 0.00 0.00 0.00 0.00 0.00 0.00 0.00

CPFSK 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 CPFSK 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00
0.6 0.6
True label

True label

GFSK 0.00 0.00 0.00 0.00 0.98 0.00 0.00 0.00 0.00 0.03 GFSK 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00

PAM4 0.00 0.00 0.01 0.00 0.00 0.99 0.00 0.00 0.00 0.00 PAM4 0.00 0.00 0.00 0.00 0.00 0.99 0.00 0.00 0.00 0.00
0.4 0.4
QAM16 0.02 0.00 0.00 0.00 0.00 0.00 0.01 0.97 0.00 0.00 QAM16 0.00 0.00 0.00 0.00 0.00 0.00 0.98 0.01 0.00 0.00

QAM64 0.01 0.00 0.00 0.00 0.00 0.00 0.01 0.98 0.00 0.00 QAM64 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.98 0.00 0.00
0.2 0.2
QPSK 0.25 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.74 0.00 QPSK 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.99 0.00

WBFM 0.00 0.63 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.34 WBFM 0.00 0.19 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.81
0.0 0.0
8PSK

AM-DSB

BPSK

CPFSK

GFSK

PAM4

QAM16

QAM64

QPSK

WBFM

8PSK

AM-DSB

BPSK

CPFSK

GFSK

PAM4

QAM16

QAM64

QPSK

WBFM

Predicted label Predicted label

(e) (f )

Figure 11: The confusion matrix of the six methods when the SNR is 18 dB. (a) Normalized confusion matrix of CNN (SNR � 18 dB). (b)
Normalized confusion matrix of Inception (SNR � 18 dB). (c) Normalized confusion matrix of LSTM (SNR � 18 dB). (d) Normalized
confusion matrix of ResNet (SNR � 18 dB). (e) Normalized confusion matrix of MentorNet (SNR � 18 dB). (f ) Normalized confusion matrix
of the proposed method (SNR � 18 dB).
Mathematical Problems in Engineering 11

1 1
0.9 0.9
0.8 0.8

Classification accuracy
Classification accuracy

0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4

0.3 0.3
0.2 0.2
0.1 0.1
0 0
–20 –18 –16 –14 –12 –10 –8 –6 –4 –2 0 2 4 6 8 10 12 14 16 18 –20 –18 –16 –14 –12 –10 –8 –6 –4 –2 0 2 4 6 8 10 12 14 16 18

SNR (dB) SNR (dB)

8PSK GFSK QAM64 8PSK GFSK QAM64

AM-DSB PAM4 QPSK AM-DSB PAM4 QPSK

BPSK QAM16 WBFM BPSK QAM16 WBFM

CPFSK CPFSK

(a) (b)
1 1
0.9 0.9
0.8 0.8
Classification accuracy

Classification accuracy
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
–20 –18 –16 –14 –12 –10 –8 –6 –4 –2 0 2 4 6 8 10 12 14 16 18 –20 –18 –16 –14 –12 –10 –8 –6 –4 –2 0 2 4 6 8 10 12 14 16 18
SNR (dB) SNR (dB)

8PSK GFSK QAM64 8PSK GFSK QAM64

AM-DSB PAM4 QPSK AM-DSB PAM4 QPSK
BPSK QAM16 WBFM BPSK QAM16 WBFM
CPFSK CPFSK

(c) (d)
1 1
0.9 0.9
0.8 0.8
Classification accuracy

Classification accuracy

0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
–20 –18 –16 –14 –12 –10 –8 –6 –4 –2 0 2 4 6 8 10 12 14 16 18 –20 –18 –16 –14 –12 –10 –8 –6 –4 –2 0 2 4 6 8 10 12 14 16 18
SNR (dB) SNR (dB)

8PSK GFSK QAM64 8PSK GFSK QAM64

AM-DSB PAM4 QPSK AM-DSB PAM4 QPSK
BPSK QAM16 WBFM BPSK QAM16 WBFM
CPFSK CPFSK

(e) (f )

Figure 12: The classification accuracy of each modulated signal using the six methods from − 20 dB to 18 dB. (a) The classification accuracy
of each modulated signal using CNN. (b) The classification accuracy of each modulated signal using Inception. (c) The classification
accuracy of each modulated signal using LSTM. (d) The classification accuracy of each modulated signal using ResNet. (e) The classification
accuracy of each modulated signal using MentorNet. (f ) The classification accuracy of each modulated signal using the proposed method.
12 Mathematical Problems in Engineering

Table 2: Computation complexity comparison between the six methods.

CNN ResNet Inception LSTM MentorNet Proposed method
Total parameters 2830170 3424976 10142726 846376 3327400 50278762
Training time(s)/epoch 44.40 48.90 38.16 91.75 45.54 590.18
Inference time(s) 2.57 2.44 2.19 5.73 3.07 46.94
FLOPs 8490523 10274951 30428197 2539146 1664319 150897325

7. Conclusion [3] Y. Wang, H. Zhang, Z. Sang, L. Xu, C. Cao, and T. A. Gulliver,

“Modulation classification of underwater communication
In this paper, we introduced an improved deep neural ar- with deep learning network,” Computational Intelligence and
chitecture for implementing radio signal identification tasks, Neuroscience, vol. 2019, pp. 1–12, 2019.
which is an important facet of constructing the spectrum- [4] B. Kim, J. Kim, H. Chae, D. Yoon, and J. W. Choi, “Deep
sensing capability required by software-defined radio. The neural network-based automatic modulation classification
goal was to achieve feature extraction by learning from technique,” in Proceedings of the International Conference on
original sampled signals on training dataset and evaluate the Information and Communication Technology Convergence
(ICTC), Jeju Island, South Korea, October 2016.
performance on validation dataset. We also compared the
[5] E. Axell, G. Leus, E. Larsson, and H. Poor, “Spectrum sensing
classification accuracy of the proposed method with that of
for cognitive radio: state-of-the-art and recent advances,”
LSTM, CNN, Inception, ResNet, and one of the state-of-the- IEEE Signal Processing Magazine, vol. 29, no. 3, pp. 101–116,
art methods MentorNet. The experiment shows that the 2012.
classification accuracy of the proposed method is highest [6] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, “Survey of
with the varying SNR among the six methods and it peaks at automatic modulation classification techniques: classical
93.76% when the SNR is 14 dB and is 6 percent higher than approaches and new trends,” IET Communications, vol. 1,
that of LSTM and 13 percent higher than that of MentorNet, no. 2, pp. 137–156, 2007.
Inception, and ResNet purely. The proposed method has [7] D. Zhang, W. Ding, B. Zhang et al., “Automatic modulation
excellent performance for modulation classification. We classification based on deep learning for unmanned aerial
believe that the significance of this paper far exceeds the vehicles,” Sensors, vol. 18, no. 3, p. 924, 2018.
modulation recognition task itself and it will offer a new idea [8] Y. Kumar, M. Sheoran, G. Jajoo, and S. K. Yadav, “Automatic
for modulation classification. modulation classification based on constellation density using
deep learning,” IEEE Communications Letters, vol. 24, no. 6,
pp. 1275–1278, 2020.
Data Availability [9] J. Zhang, D. Cabric, F. Wang, and Z. Zhong, “Cooperative
modulation classification for multipath fading channels via
The data used to support the findings of this study are in- expectation-maximization,” IEEE Transactions on Wireless
cluded within the article. Communications, vol. 16, no. 10, pp. 6698–6711, 2017.
[10] E. Nachmani, Y. Bachar, E. Marciano, D. Burshtein, and
Conflicts of Interest Y. Be’ery, “Near maximum likelihood decoding with deep
learning,” 2018, https://arxiv.org/abs/1801.02726.
The authors declare no conflicts of interest. [11] A. Afan and Y. Y. Fan, “Unsupervised feature learning and
automatic modulation classification using deep learning
model,” Physical Communication, vol. 25, pp. 75–84, 2017.
Authors’ Contributions [12] T. J. O’Shea, N. West, and M. Vondal, “Semi-supervised radio
S. S. contributed to data curation; P.W. and X.W. con- signal identification,” in Proceedings of the IEEE International
tributed to methodology; J.Z. contributed to resources; B.S. Conference on Advanced Communication Technology, Bong-
pyeong, South Korea, February 2017.
contributed to software; J.W. contributed to supervision.
[13] M. Abu-Romoh, A. Aboutaleb, and Z. Rezki, “Automatic
modulation classification using moments and likelihood
Acknowledgments maximization,” IEEE Communications Letters, vol. 22, no. 5,
pp. 938–941, 2018.
This work was supported by Youth Science Foundation of [14] X. Sun, S. Su, Z. Zuo, X. Guo, and X. Tan, “Modulation
Hunan Natural Science Foundation under Grant No. classification using compressed sensing and decision tree-
2019JJ50121. support vector machine in cognitive radio system,” Sensors,
vol. 20, no. 5, p. 1438, 2020.
References [15] Z. Zheng, T. Huang, H. Zhang, S. Sun, J. Wen, and P. Wang,
“Towards a resource migration method in cloud computing
[1] G. Q. Yang, “Modulation classification based on extensible based on node failure rule,” Journal of Intelligent & Fuzzy
neural networks,” Mathematical Problems in Engineering, Systems, vol. 31, no. 5, pp. 2611–2618, 2016.
vol. 9, p. 6416019, 2017. [16] X. Shi, Z. Zheng, Y. Zhou et al., “Graph processing on GPUs: a
[2] F. Wang, S. Huang, H. Wang, and C. Yang, “Automatic survey,” ACM Computing Surveys, vol. 50, 2017.
modulation classification exploiting hybrid machine learning [17] T. O’Shea, T. James, T. Roy, and T. Clancy, “Over-the-air deep
network,” Mathematical Problems in Engineering, vol. 2018, learning based radio signal classification,” IEEE Journal of
pp. 1–14, 2018. Selected Topics in Signal Processing, vol. 12, pp. 168–179, 2018.
Mathematical Problems in Engineering 13

[18] T. Wang, C.-K. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, [33] H. Wu, Y. Li, L. Zhou, and J. Meng, “Convolutional neural
“Deep learning for wireless physical layer: opportunities and network and multi-feature fusion for automatic modulation
challenges,” China Communications, vol. 14, no. 11, classification,” Electronics Letters, vol. 55, no. 16, pp. 895–897,
pp. 92–111, 2017. 2019.
[19] T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional radio [34] M. Zhang, Z. Yu, H. Wang, H. Qin, W. Zhao, and Y. Liu,
modulation recognition networks,” in Proceedings of the In- “Automatic digital modulation classification based on cur-
ternational Conference on Engineering Applications of Neural riculum learning,” Applied Sciences, vol. 9, no. 10, p. 2171,
Networks, Aberdeen, UK, September 2016. 2019.
[20] S. Rajendran, W. Meert, D. Giustiniano, V. Lenders, and [35] S. Ramjee, S. Ju, and D. Yang, “Fast deep learning for au-
S. Pollin, “Deep learning models for wireless signal classifi- tomatic modulation classification,” 2019, https://arxiv.org/
cation with distributed low-cost spectrum sensors,” IEEE abs/1901.05850.
Transactions on Cognitive Communications and Networking, [36] N. E. West and T. J. O’Shea, “Deep architectures for mod-
vol. 4, no. 3, pp. 433–445, 2018. ulation recognition,” in Proceedings of the IEEE International
[21] S. Hu, Y. Pei, and P. P. Liang, “Robust modulation classifi- Symposium on Dynamic Spectrum Access Networks, Piscat-
cation under uncertain noise condition using recurrent neural away, NJ, USA, March 2017.
network,” in Proceedings of the IEEE Global Communications [37] I. Laina, C. Rupprecht, and V. Belagiannis, “Deeper depth
prediction with fully convolutional residual networks,” 2016,
Conference, Abu Dhabi, United Arab Emirates, December
https://arxiv.org/abs/1606.00373.
2018.
[38] A. A. Litonjua and D. R. Gold, “Asthma and obesity: common
[22] B. Tang, Y. Tu, Z. Zhang, and Y. Lin, “Digital signal modu-
early-life influences in the inception of disease,” Journal of
lation classification with data augmentation using generative
Allergy and Clinical Immunology, vol. 121, no. 5, pp. 1075–
adversarial nets in cognitive radio networks,” IEEE Access,
1084, 2008.
vol. 6, pp. 15713–15722, 2018.
[39] C. Szegedy, S. Ioffe, and V. Vanhoucke, “Inception-v4, in-
[23] Y. Wang, M. Liu, J. Yang, and G. Gui, “Data-driven deep
ception-resnet and the impact of residual connections on
learning for automatic modulation recognition in cognitive learning,” 2016, https://arxiv.org/abs/1602.07261.
radios,” IEEE Transactions on Vehicular Technology, vol. 68,
no. 4, pp. 4074–4077, 2019.
[24] T. N. Sainath, O. Vinyals, A. W. Senior, and H. Sak, “Con-
volutional, long short-term memory, fully connected deep
neural networks,” in Proceedings of the IEEE International
Conference on Acoustics, Brisbane, Australia, April 2015.
[25] P. Shengliang, J. Hanyu, W. Huaxia, A. Hathal, Z. Yu, and
S. M. Mazrouei, “Modulation classification based on signal
constellation diagrams and deep learning,” IEEE Transactions
on Neural Networks and Learning Systems, vol. 30, pp. 718–
727, 2019.
[26] P. S. Cheong, M. Camelo, and S. Latré, “Evaluating deep
neural networks to classify modulated and coded radio sig-
nals,” in Proceedings of the 13th EAI International Conference,
CROWNCOM, Ghent, Belgium, April 2018.
[27] X. Xie, Y. Nil, S. Peng, and Y. D. Yao, “Deep learning based
automatic modulation classification for varying snr envi-
ronment,” in Proceedings of the 2019 28th Wireless and Optical
Communication Conference (WOCC), Beijing, China, May
2019.
[28] S. Peng, H. Jiang, H. Wang, H. Alwageed, and Y. D. Yao,
“Modulation classification using convolutional neural net-
work based deep learning model,” in Proceedings of the 2017
26th Wireless and Optical Communication Conference
(WOCC), IEEE, Newark, NJ, USA, April 2017.
[29] T. J. O’Shea and N. West, “Radio machine learning dataset
generation with gnu radio,” in Proceedings of the GNU Radio
Conference, Boulder, Colorado, September 2016.
[30] M. Li, O. Li, G. Liu, and C. Zhang, “Generative adversarial
networks-based semi-supervised automatic modulation rec-
ognition for cognitive radio networks,” Sensors, vol. 18, no. 11,
p. 3913, 2018.
[31] M. Du, Q. Yu, and S. Fei, “Fully dense neural network for the
automatic modulation recognition,” 2019, https://arxiv.org/
abs/1912.03449.
[32] Y. Zeng, M. Zhang, F. Han, Y. Gong, and J. Zhang, “Spectrum
analysis and convolutional neural network for automatic
modulation recognition,” IEEE Wireless Communications
Letters, vol. 8, no. 3, pp. 929–932, 2019.