Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2019.2963828, IEEE Wireless
Communications Letters

Modulation Recognition with Graph Convolutional


Network
Yabo Liu, Yi Liu and Cheng Yang

Abstract—In most non-cooperative communication systems, combining above two likelihood algorithms. However, the
modulation recognition is a fundamental and critical technique. complexities of maximum likelihood hypothesis algorithms
Traditional methods of modulation recognition can be categorized are too great. The pattern recognition algorithms contain two
as maximum likelihood hypothesis algorithms and pattern recog-
nition algorithms. However, these methods have high complexities procedures: extracting signal features and classifying. In [4],
or need additional data preprocessing. Recently, neural network the authors extract signal instantaneous statistics and use a
algorithms have shown great potential in modulation recognition. decision tree to recognize modulation modes. [5] performs
In this letter, we propose a method of modulation recognition a wavelet transform on phase shift keying and frequency
by exploiting the graph convolutional network (GCN). However, shift keying modulated signals to extract signal features, then
GCNs cannot be directly used to perform modulation recognition
since modulated signals are not graphs. To convert signals to processes the signal features by pattern recognition algorithms.
graphs, the modulation dataset is divided into multiple subsets. In [6], the authors extract fourth-order statistics of modulated
We design a feature extraction convolutional neural network signals, and analyze the accuracies of pattern recognition
(CNN) and a graph mapping CNN to extract signal features and algorithms various signal-to-noise ratios(SNRs). The types of
map subsets into graphs, respectively. Then we input the graphs extracted signal features have a significant influence on the
into the GCN to predict modulation modes of unlabeled signals.
The experimental results show that the proposed GCN algorithm modulation accuracy.
achieves higher recognition accuracy than CNN algorithm and Over the past few years, deep learning has been widely
K-nearest neighbor (KNN) algorithm, especially when SNR is investigated due to its excellent performances in the field of
low. pattern recognition. Compared with traditional methods such
as pattern recognition, deep learning methods can automatic
Index Terms—modulation recognition, graph convolutional
network, convolutional neural network, graph mapping. extract data features and make predictions. Many deep learning
recognition algorithms based on various networks such as
I. I NTRODUCTION convolutional neural networks (CNNs) and recurrent neural
networks (RNNs) have been investigated. In [7], the authors
Signal modulation techniques are extensively utilized to propose a CNN consists of 2 convolutional layers and 2 dense-
efficient data transmission. In non-cooperative communication ly connected layers to recognize signal modulation modes.
systems, modulation recognition (MR) is a very important They also open source a modulation dataset built by a universal
approach to recognize modulation modes of unknown signals. software radio peripheral. Furthermore, in [8], the authors
This technique plays an important role in many civilian and make a comparison between long short time memory networks
military applications such as cognitive radio and electronic (LSTMs) and CNNs in modulation recognition. Simulation
countermeasures. results show that the depth of networks has little effect on the
Traditional methods of modulation recognition can be cat- recognition accuracies. A method that simultaneously inputs
egorized as maximum likelihood hypothesis algorithms and raw signals and their high-order statistics into a RNN to
pattern recognition algorithms. The maximum likelihood hy- improve recognition accuracy is proposed in [9]. The higher-
pothesis algorithms are based on rigorous mathematical deriva- order statistics of signals are input into a two layer auto-
tions. They can make predictions by maximizing its likelihood encoder in [10], where a layer-by-layer strategy is adopted
functions. Reference [1] proposes an average likelihood ratio to train the auto-encoder.
recognition algorithm for various communication systems. It Recently, some researches are devoting to processing graph
can be used to recognize phase shift keying modulated signals. data with neural networks. One remarkable achievement is
In [2], the authors extend the average likelihood ratio algorithm a connectionist model, graph convolutional network (GCN),
to the generalized likelihood ratio recognition scheme, in which can merge messages transmitting between adjacent
which signal high-order statistics are widely used. The hybrid nodes in graphs [11]. GCNs have been successfully used in
likelihood ratio recognition algorithm is proposed in [3] by text classification, relation extraction and image classification
tasks [12]. This motivates us to utilize GCNs in modulation
Yabo Liu and Yi Liu are with the State Key Laboratory of Integrat-
ed Service Network, Xidian University, Xi’an, 710071, China. (e-mail: recognition. However, GCNs cannot be directly used to pro-
ybliu 1994@163.com, yliu@xidian.edu.cn). Their research was supported in cess signals since they are not graph data. In this letter, we
part by the Natural Science Foundation of China (61671341), the project design two CNNs named as feature extraction CNN (FECNN)
of Shaanxi importance Industry chain (2018ZDCXL-GY-04-06) and the 111
Project (B08038) of China. Cheng Yang is with the China Coal Xi’an Design and graph mapping CNN (GMCNN), respectively to extract
Engineering Co.,Ltd, Xi’an 710054, China. (e-mail: 4658886@qq.com) signal features and map signals into graphs. Then, we present

2162-2337 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2019.2963828, IEEE Wireless
Communications Letters

subset feature embedding matrixe

a signal modulation recognition algorithm based on the GCN.


P
To the best of our knowledge, this work is the first attempt labeled
to apply GCN to modulation recognition. We evaluate the signals
feature
performance of the proposed GCN algorithm, and compare
its recognition accuracy with those of the K-nearest neighbor
extraction
CNN
+ ?
q
(KNN) algorithm and CNN algorithm proposed in [7]. Experi- unlabeled
mental results show that the proposed GCN algorithm achieves signal

higher recognition accuracy, especially when SNR is low. ?


one-hot
feature embeddings
encodings
II. M ODEL OF GCN BASED M ODULATION R ECOGNITION
In applications of classifying with a typical GCN, every Fig. 1. The process of feature extraction by the FECNN.
input graph consists both labeled and unlabeled nodes. The
GCN can classify unlabeled nodes by merging labeled nodes’ convolution layers are 1 × 3 and 2 × 3, respectively. Both
messages. In this work, we need to choose some labeled of the kernel strides are (2, 1). The numbers of convolutional
and unlabeled signals to generate each graph. Hence before kernels are 32 and 64, respectively. The densely connected
converting signals to graphs, we divide a modulation dataset layer has 64 neurons. Fig. 1 shows the process of extracting
into many subsets. Each subset can be described as signal features by the FECNN. Firstly, a subset divided from
the dataset is input into the FECNN. The output of the
X = {{(x1 , y1 ), ..., (xp , yp )} , {x̄1 , ..., x̄q }} , (1)
network are feature embeddings of input signals. Later, in
where p is the number of labeled signals x and q is the number order to involve signals’ label information during the training
of unlabeled signals x̄. Both xi and x̄i are signals in modu- process, the embedding features are concatenated with one-hot
lation dataset. yi is the label corresponding to xi . {ȳ1 , ..., ȳq } encodings of their labels. For p labeled signals in a subset,
are the labels of unlabeled signals {x̄1 , ..., x̄q }. Every label their one-hot encodings are derived from their real labels. For
belongs to {1, ..., C}. C is the number of modulation modes. q unlabeled signals in a subset, we set all elements of their one-
The GCN algorithm is trained in a supervised manner. The hot encoding to 1/C to blur its label. By the feature extraction
categories of labeled signals are input into the GCN algorithm process, each subset is transformed into a feature embedding
as signal features during forward propagation and the real matrix.
categories of unlabeled signals are used to supervise the B. Graph mapping
training of the GCN algorithm during backward propagation.
The target of each step in training process can be expressed After extracting signal features, we need to map every
as subset into a graph G(V , E), where vi corresponds to the i-th
1 signal and E represents the collection of edges. We use the
min L(ϕ(X), {ȳ1 , ..., ȳq }), (2)
q GMCNN to measure differences between every two signals
in a subset and output a graph. The input of the GMCNN
where L(·) is a loss function and ϕ(·) is the function
is a (p + q) × (p + q) × (64 + C) three dimensional matrix
parametrized with the GCN algorithm. The output of the GCN
M . mi,j represents the absolute distance between the feature
algorithm ϕ(X) contains the predict labels of unlabel signals
embeddings of i-th and j-th signals in a subset. The GMCNN
in subset X. This algorithm can be divided into three steps:
is composed of 4 convolutional layers. The first, second
feature extraction, graph mapping and graph convolution. We
and third convolutional layers are followed by a leaky Relu
propose an FECNN, a GMCNN and a GCN to realize these
activation function and batch-normalization. The numbers of
operations, respectively. All the neural networks are trained
convolutional kernels are 192, 192, 96 and 1, respectively. All
jointly in a supervised manner while reducing the classification
of their kernel sizes are 1 × 1, and their strides are set to
bias of unlabeled training signals. The only loss function
1. The 1 × 1 convolution can merge information of multiple
in the networks is a cross entropy loss function which is
channels and maintain the size of feature map. By utilizing 1
used between unlabeled signals’ predicted and real labels.
× 1 convolution, the GMCNN will output a (p + q) × (p + q)
During the jointly supervised training process, the FECNN
adjacency matrix which illustrates the structure of a graph.
and GMCNN are forced to extract as much as possible useful
The calculation of edges can be summarized as
signal information and produce reasonable graph structures to
help the GCN make correct predictions. ei,j = φ(vi , vj ) = GM CN N (abs(vi − vj ))), (3)
A. Feature extraction where ei,j is the edge between i-th and j-th nodes. φ(·) is
The FECNN used to extract signal features consists of 2 the function parametrized with the GMCNN. This approach
convolutional layers and one densely connected layer. The of computing adjacency matrix has the symmetry property
input of the FECNN is a two-dimensional matrix whose two φ(vi , vj ) = φ(vj , vi ) (4)
rows are real and imaginary parts of the signal, respectively.
and identity property
Each convolutional layer is followed by a leaky Relu activation
function to fit nonlinear distribution data. The kernel sizes of φ(vi , vi ) = 0. (5)

2162-2337 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2019.2963828, IEEE Wireless
Communications Letters

(feature embedding matrixe)

Graph mapping CNN


Graph mapping CNN
Subset

GCN
Softmax

GCN
Recognition result

Layer 1 Layer 2

Fig. 2. The process of graph mapping by the GMCNN

C. Graph convolution into graphs, the dataset is divided into many subsets. We
Since Scarselli et al. proposed the prototype of graph neural set p and q in Eq. (1) be 5×8 and 1, respectively. That is,
network (GNN) in 2009 [13], researchers have developed each subset is composed of 41 signals, including 5 labeled
many variations of GNN such as GCN, graph attention net- signals per modulation mode and only one unlabeled signal. In
work, graph auto-encoder, graph generative network and graph experiment settings, the loss function and activation function
spatial-temporal network. In this work, we consider a simple in the GCN are cross-entropy loss and leaky Relu. Batch size,
implementation of GCN which uses the adjacency operator learning rate and SGD momentum are set to 10, 0.001 and
0.5, respectively. Since the three networks are trained jointly,
they share the same training setup. We first investigate the
!
X
xnk+1 = ρ xkm wk , (6) recognition accuracy of the GCN algorithm versus training
m∈Nn iteration at 10dB SNR. Then, we calculate the recognition
accuracies of all modulation modes with the GCN algorithm.
where xk+1
is the feature vector of n-th node in (k + 1)-th
n
Finally, we compare the recognition accuracy of the GCN
layer, Nn is the set of n-th nodes adjacent nodes, xkm is the
algorithm, KNN algorithm and CNN algorithm proposed in
feature vector of the m-th node in the k-th layer, wk is the
[7].
weighting parameter which connects the n-th node with its
adjacent nodes and ρ(·) is an activation function. During the B. Results
graph convolution, each node sends message to its adjacent
nodes, and each node will merge the messages received from Fig. 3 illustrates the recognition accuracy of the GCN algo-
its adjacent nodes. Then, an activation function is followed rithm versus training iteration at 10dB SNR. From the figure,
to enhance the ability of the GCN to fit data distribution. To the recognition accuracy increases with training progress when
normalize the input data, we use a softmax function on each the iteration is below 9000. Then the recognition accuracy
row of adjacency matrixes in advance. As the GCN deepens, tends to be constant.
the receptive field becomes larger and more messages will 0.8
be included in the prediction. Due to the graphs adjacency
matrixes calculated by the GMCNN are densely connected, 0.7

it is not necessary to set too much layers in the GCN. In


this work, we choose the number of layers in the GCN to be 0.6
Accuracy

2. The graph convolution process is shown in Fig. 2. The


circles represent signals in a subset. The adjacency matrix 0.5

is recalculate before every graph convolution layer. Finally,


0.4 Test accuracy
the feature vector of the unlabeled signal in a subset can be
obtained, and it will be mapped into the final classification by
0.3
the softmax function. 0.5 1 1.5 2
Iteration ×10 4
III. E XPERIMENTAL R ESULTS
Fig. 3. The recognition accuracy of the GCN algorithm versus training
A. Experimental setup iteration at 10dB SNR.
To evaluate the proposed GCN based modulation recog- We also investigate the recognition performance of modula-
nition algorithm, we consider several experiments in this tion modes with the GCN algorithm. The recognition accura-
section. The dataset we generated for this work contains 8 cies of 8 modulation modes with the GCN algorithm are shown
types of digital modulated signals with SNRs varying from in Fig. 4. The recognition accuracies of all modulation modes
-14dB to +10dB. The modulation modes are 2ASK, 4ASK, increase with SNR. Because low-order modulated signals are
2FSK, 4FSK, BPSK, QPSK, 16QAM and 64QAM which are less complex compared with high-order modulated signals, it
widely used in modern communication systems. Every signal is easier for the GCN to extract and learn signal features
is composed of 256 samples. Before mapping the signals from low-order modulated signals. The accuracies of low-

2162-2337 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LWC.2019.2963828, IEEE Wireless
Communications Letters

0.9 TABLE I
T HE COMPARISON OF PARAMETERS AND AVERAGE INFERENCE TIMES
0.8
BETWEEN THE GCN AND CCN ALGORITHMS .

Algorithm Number of parameters Inference time (ms)


0.7
GCN 437071 0.78
Accuracy

CNN 1455768 0.16


0.6
2ASK
4ASK
0.5 2FSK
4FSK
BPSK
IV. C ONCLUSIONS
0.4 QPSK
16QAM In this letter, we propose an initial effort to apply GCN to
64QAM
modulation recognition. In order to convert modulated signals
0.3
-10 -5 0 5 10 to graphs, the dataset is divided into multiple subsets. Each
SNR(dB)
subset consists of 40 labeled signals and one unlabeled signal.
Fig. 4. The recognition accuracies of 8 modulation modes with the GCN We design an FECNN and a GMCNN based on CNNs to
algorithm. extract signal features and map subsets to graphs, respec-
tively. Then we input the graphs into a GCN to recognize
the modulation mode of unlabeled signal. The experimental
order modulated signals are higher than those of high-order results show that the proposed GCN algorithm has a better
modulated signals. performance compared with the CNN algorithm and KNN
algorithm, especially when SNR is low.
0.8

GCN R EFERENCES
0.7 CNN
KNN [1] K. Kim and A. Polydoros, “Digital modulation classification: the
0.6 bpsk versus qpsk case,” in MILCOM 88, 21st Century Military
Communications-What’s Possible?’. Conference record. Military Com-
Accuracy

0.5 munications Conference. IEEE, 1988, pp. 431–436.


[2] D. Boiteau and C. Le Martret, “A general maximum likelihood frame-
0.4
work for modulation classification,” in Proceedings of the 1998 IEEE
International Conference on Acoustics, Speech and Signal Processing,
0.3
ICASSP’98 (Cat. No. 98CH36181), vol. 4. IEEE, 1998, pp. 2165–2168.
0.2
[3] L. Hong and K. Ho, “Bpsk and qpsk modulation classification with
unknown signal level,” in MILCOM 2000 Proceedings. 21st Century
0.1 Military Communications. Architectures and Technologies for Informa-
-20 -10 0 10 20 tion Superiority (Cat. No. 00CH37155), vol. 2. IEEE, 2000, pp. 976–
SNR(dB) 980.
[4] A. K. Nandi and E. E. Azzouz, “Algorithms for automatic modulation
Fig. 5. The comparison of recognition performance among GCN algorithm, recognition of communication signals,” IEEE Transactions on commu-
CNN algorithm and KNN algorithm. nications, vol. 46, no. 4, pp. 431–436, 1998.
[5] K. Ho, W. Prokopiw, and Y. Chan, “Modulation identification of digital
The comparison of recognition accuracies among GCN signals by the wavelet transform,” IEE Proceedings-Radar, Sonar and
Navigation, vol. 147, no. 4, pp. 169–176, 2000.
algorithm, CNN algorithm and KNN algorithm is illustrated [6] A. Swami and B. M. Sadler, “Hierarchical digital modulation classifica-
in Fig. 5. All of these three algorithms use the same dataset tion using cumulants,” IEEE Transactions on communications, vol. 48,
without any preprocessing. We can see that KNN performs no. 3, pp. 416–429, 2000.
[7] T. J. OShea, J. Corgan, and T. C. Clancy, “Convolutional radio modula-
worst without manual extracting signal features, while neural tion recognition networks,” in International conference on engineering
networks can automatic extract signal features. Therefore the applications of neural networks. Springer, 2016, pp. 213–226.
recognition accuracies of two neural network algorithms are [8] N. E. West and T. O’Shea, “Deep architectures for modulation recog-
nition,” in 2017 IEEE International Symposium on Dynamic Spectrum
much higher than that of KNN algorithm. The recognition Access Networks (DySPAN). IEEE, 2017, pp. 1–6.
accuracy of GCN algorithm is almost equal to that of CNN [9] M. Zhang, Y. Zeng, Z. Han, and Y. Gong, “Automatic modulation
algorithm when SNR is above 2dB, while higher than that recognition using deep learning architectures,” in 2018 IEEE 19th
International Workshop on Signal Processing Advances in Wireless
of CNN algorithm when SNR is below 2dB. Specifically, Communications (SPAWC). IEEE, 2018, pp. 1–5.
the recognition accuracy of GCN algorithm is around 10% [10] M. Bouchou, H. Wang, and M. E. H. Lakhdari, “Automatic digital
higher than that of CNN algorithm when SNR is -20dB. This modulation recognition based on stacked sparse autoencoder,” in 2017
IEEE 17th International Conference on Communication Technology
is because the GCN can merge features of all signals in a (ICCT). IEEE, 2017, pp. 28–32.
graph, and compare the test signal with all modes of training [11] J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun, “Spectral networks and
signals to make predictions. The number of parameters and locally connected networks on graphs,” arXiv preprint arXiv:1312.6203,
2013.
average inference times based on Tesla V100 GPU between [12] V. Garcia and J. Bruna, “Few-shot learning with graph neural networks,”
the GCN and CCN algorithms is compared in Table I. From arXiv preprint arXiv:1711.04043, 2017.
the table, the GCN algorithm has fewer parameters than the [13] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini,
“The graph neural network model.” IEEE Transactions on Neural
CCN algorithm. However, the GCN algorithm spends more av- Networks, vol. 20, no. 1, p. 61, 2009.
erage inference time. This is because the graph convolutional
operator implemented by Pytorch is not efficient enough.

2162-2337 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

You might also like