Digital signal modulation recognition method based on

high-order cumulants and wavelet transform

Anyi Wang Peiru Liu∗
College of Communication and Information Engineering, College of Communication and Information Engineering,
Xi’an University of Science and Technology, Xi’an, Xi’an University of Science and Technology, Xi’an,
Shaanxi, China Shaanxi, China
[email protected] [email protected]

ABSTRACT 1 INTRODUCTION
In view of the current situation that the recognition rate of digi- Information communication is the main tool of human social con-
tal signal modulation recognition method is unsatisfactory at low tact, and its main purpose is to realize information transmission.
Signal-to-Noise Ratio(SNR), a recognition method based on high- Whether in the traditional wired communication system or in the
order cumulants and wavelet transform is proposed to realize the wireless one, due to the limitation of channel transmission char-
automatic modulation recognition of 8 kinds of digital signals such acteristics, the baseband signal containing many low frequency
as 2ASK, 4ASK, 8ASK, 2PSK, 4PSK, 8PSK, 16QAM and 32QAM. components is not suitable for direct transmission through the
Based on the high-order cumulants principle and wavelet trans- channel. The baseband signal needs to be modulated, so it can be
form theory, the characteristic parameters f 1 ∼f 5 are constructed effectively transmitted [1]. In recent years, users’ requirements
by the elaborate analysis of the characteristic extraction of these for information transmission rate and stability have continued to
signals. Through simulation experiments, the characteristic param- increase. Therefore, signal modulation forms have gradually diver-
eter changes of different types of modulation signals at different sified and become more complicated. This has led to the need for
SNR are obtained, and design the classifier of Back Propagation corresponding demodulation schemes at the receiving part for dif-
(BP) neural network to classify the signals. The simulation results ferent modulation methods. However, the premise of demodulation
show that this method can improve the average correct recognition is to correctly identify the signal modulation method [2]. Therefore,
rates of 8 digital modulation signals reaching up to above 97% when modulation recognition of different types of signals has become an
the SNR is higher than 0dB, which greatly improves the signal urgent problem to be solved. Currently, among the existing signal
recognition performance at low SNR. modulation recognition methods, the maximum likelihood method
based on hypothesis testing requires more prior information, and
CCS CONCEPTS the statistical expressions are complicated [3][4]. The identifica-
tion method based on the high-order statistical characteristics of
• Information systems; • Information systems applications;
the signal has simple basic principles, strong operability, and can
• Mobile information processing systems;
suppress additive white Gaussian noise [5-7]. However, among the
commonly used digital signal types, there are some types that are
KEYWORDS difficult to classify and recognize by high-order statistical features.
Digital Signal Modulation Recognition, High-order Cumulants, In the traditional method, the differential operation is performed
Wavelet Transform, Back Propagation Neural Network before the cumulant calculation, but this method has poor signal
recognition performance under the condition of low SNR [8]. In the
ACM Reference Format:
recognition method based on high-order cumulants, improving the
Anyi Wang and Peiru Liu. 2020. Digital signal modulation recognition
recognition rate under the condition of low SNR is an urgent prob-
method based on high-order cumulants and wavelet transform. In 2020 3rd
International Conference on Algorithms, Computing and Artificial Intelligence lem to be solved. At present, deep learning and neural networks
(ACAI 2020), December 24–26, 2020, Sanya, China. ACM, New York, NY, USA, have been studied by many scholars, and have been gradually ap-
6 pages. https://doi.org/10.1145/3446132.3446423 plied to the field of signal modulation recognition, but the research
about this still needs to be carried out [9-12].
This paper analyzes the signal feature extraction method and
∗ Corresponding author: [email protected] classifier design issues, proposes a hybrid modulation recognition
method combining high-order cumulants and wavelet transform,
and designs a Back Propagation (BP) neural network classifier to
Permission to make digital or hard copies of all or part of this work for personal or achieve the high-efficiency identification of three types modula-
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation tion signals, such as MASK, MPSK and MQAM. Compared with
on the first page. Copyrights for components of this work owned by others than ACM traditional feature extraction methods, the method proposed in this
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
paper improves the recognition rate of digital modulation signals
fee. Request permissions from [email protected]. under low SNR, and broadens the application scenarios of digital
ACAI 2020, December 24–26, 2020, Sanya, China signal modulation recognition.
© 2020 Association for Computing Machinery.
ACM ISBN 978-1-4503-8811-5/20/12. . . $15.00
https://doi.org/10.1145/3446132.3446423
ACAI 2020, December 24–26, 2020, Sanya, China Anyi Wang and Peiru Liu

Table 1: Theoretical values of higher order cumulants of digital signals

Signal Type C20 C21 C40 C41 C42 C60 C61 C63
2ASK E E -2E2 -2E2 -2E2 16E3 16E3 13E3
4ASK E E -1.36E2 -1.36E2 -1.36E2 8.32E3 8.32E3 9.16E3
8ASK E E -1.24E2 -1.24E2 -1.24E2 7.19E3 7.19E3 8.76E3
2PSK E E -2E2 -2E2 -2E2 16E3 16E3 13E3
4PSK 0 E E2 0 -E2 0 4E3 4E3
8PSK 0 E 0 0 -E2 0 0 4E3
16QAM 0 E -0.68E2 0 -0.68E2 0 2.08E3 2.08E3
32QAM 0 E -0.19E2 0 -0.69E2 0 2.11E3 2.11E3

2 FEATURE PARAMETER SELECTION second-, fourth-, and sixth-order cumulants of x(k) are:
In the statistical pattern recognition method, the extraction of char- C 20 = cum (x, x) = M 20
acteristic parameters has an important influence on the accuracy
C 21 = cum x, x ∗ = M 21


and efficiency of signal modulation recognition, and different char-
2
acteristic parameter extraction methods correspond to different C 40 = cum (x, x, x, x) = M 40 − 3M 20
modulation recognition methods. Common modulation recognition C 41 = cum x, x, x, x ∗ = M 41 − 3M 20 M 21


methods include: methods based on high-order cumulants, wavelet
C 42 = cum x, x, x ∗ , x ∗ = M 42 − M 20
2 2


− 2M 21
transform, and constellation diagrams. This paper proposes a hy-
3
brid modulation recognition method based on the combination of C 60 = cum (x, x, x, x, x, x) = M 60 − 15M 40 M 20 + 30M 20
high-order cumulants and wavelet transform, which can extract C 61 = cum x, x, x, x, x, x ∗


the characteristic parameters of the 8 modulation signals of MASK, 3 2
= M 61 − 5M 40 M 21 − 10M 20 M 41 + 30M 20 M 20
MPSK and MQAM under the condition of low SNR by BP neural
C 63 = cum x, x, x, x ∗ , x ∗ , x ∗


network classifier.
2 3
= M 63 − 6M 41 M 20 − 9M 42 M 21 + 18M 20 M 21 + 12M 21

2.1 Feature parameter selection based on The p-order mixing moment of the random sequence x(k) is:
high-order cumulants Mpq = E x(k)p−q x ∗ (k)q
 
The expression of the received signal is:
When various types of digital modulation signals occur with
equal probability of symbols, the theoretical values of the high-order
√
ak Eд (t − kTs ) e [j(2π ωc t +θ )] + n (t) cumulants of MASK, MPSK, and MQAM digital modulation signals
Õ
r (t) = s (t) + n (t) =
k
can be calculated by formulas (4)∼(11). The detailed theoretical
values are shown in Table 1. From Table 1, except for 2ASK and
2PSK, the remaining signals to be identified can be distinguished
ak is the transmitted symbol sequence, ω c is the carrier frequency, using high-order cumulants.
E is the energy of signal, and θ is the initial phase. In this article, the In the process of feature extraction, in order to reduce the influ-
initial phase of the signal is 0. s(t) and n(t) are independent of each ence of phase jitter and amplitude change on signal recognition,
other, s(t) is the transmitted signal, and n(t) is the white Gaussian the absolute value and ratio methods are used to improve the recog-
noise. nition accuracy. There are 8 types of modulation signals to be iden-
For a zero-mean random sequence x (k) , its n(n=p+q) moment tified in this paper. The feature extraction parameters constructed
and cumulant are defined as: based on the high-order cumulant theory are f1 , f2 , f3 , f4 , and their
theoretical values are shown in Table 2
|C |
f 1 = |C41 | , divide the signal into {MASK, 2PSK} and {4PSK, 8PSK,
k 42
Mn,q = mom x 1k1 , x 2k2 , . . . , xp p , (xp+1 )kp+1 , . . . , (xp+q )kp+q
h i
∗ ∗ MQAM} by the characteristic parameter f 1 ;
|C 63 | 2
f2 = , using the feature parameterf 2 , {MASK, 2PSK} is
|C 40 | 3
divided into {2ASK, 2PSK}, {4ASK} and {8ASK}, pre-processing the
MASK signal before classification, and by reducing the influence of
k Gaussian noise on the signal, it can achieve better results;
Cn,q = cum x 1k1 , x 2k2 , . . . , xp p , (xp+1 )kp+1 , . . . , (xp+q )kp+q
h i
∗ ∗
|C |
f 3 = 61 3 , divide {4PSK, 8PSK, MQAM} into {4PSK}, {16QAM,
|C 21 |
32QAM} and {8PSK} through the characteristic parameterf 3 ;
|C |
* means function conjugate. For a zero-mean stationary random f 4 = |C40 | , according to the characteristic parameter f 4 , {16QAM,
42
sequence x(k), due to its independent and identical distribution, the 32QAM} is divided into {16QAM} and {32QAM}.
Digital signal modulation recognition method based on high-order cumulants and wavelet transform ACAI 2020, December 24–26, 2020, Sanya, China

Table 2: Theoretical value of characteristic parameters based on high-order accumulations

Signal Type f1 f2 f3 f4
2ASK 1 21.125 16 1
4ASK 1 33.356 8.32 1
8ASK 1 40.4433 7.19 1
2PSK 1 21.125 16 1
4PSK 0 16 4 1
8PSK 0 / 0 0
16QAM 0 13.7594 2.08 1
32QAM 0 649.089 2.11 0.275362

2.2 Feature parameter selection based on Table 3: Theoretical value f5 digital signal characteristic pa-
wavelet transform rameters

The feature parameters based on high-order cumulants cannot

classify {2ASK, 2PSK}. In traditional algorithms, the distinction Signal Type 2ASK 2PSK
between 2ASK and 2PSK signals is generally done by introducing f5 2 1
the signal’s instantaneous characteristic parameters. According to
the research of J. J. Guo et al. [13], this method has an unsatisfactory
recognition rate of digital signal modulation in the case of low SNR.
For the modulation recognition problem of 2ASK and 2PSK signals, wavelet transform scale. According to the known ratio of fc and
when the SNR is 1dB, the recognition rate is only about 40%, and fs , the value of acan be calculated when GSNR takes the maximum
when the SNR is between 1 and 4dB, this method’s recognition rate value.
for these two signals is lower than 65%. Through the feature parameters f1 ∼f4 based on high-order cu-
This paper uses characteristic parameters based on wavelet trans- mulants, the signals have been divided into {2ASK, 2PSK}, {4ASK},
form to classify and recognize {2ASK, 2PSK} signals. According to {8ASK}, {4PSK}, {16QAM}, {32QAM}, {8PSK}, and then using fea-
the definition of wavelet transform and the expression of Haar ture parameters based on wavelet transform to solve {2ASK, 2PSK}
wavelet transform, the wavelet transform amplitude of MASK and classification and recognition problems.
MPSK modulated signals can be obtained, as shown below: Through the analysis of equations, it can be known that the
√ modulation information of the digital signal can be extracted by the
4 ai 2π fc a
 
|CW TM AS K (a, τ )| = √ sin2 amplitude information of the signal after wavelet transformation.
a2π fc 4 This property can be used to construct characteristic parameters of
√ different signal modulation. The histogram of the wavelet transform
4 E 2π fc a
 
|CW TM P S K (a, τ )| = √ sin2 amplitude of 2ASK and 2PSK is shown in Figure 1. Because the
a2π fc 4
2ASK signal has two different amplitude values, there are two peaks
Where E is the energy of the transmitted symbol waveform, fc in Figure 1. For the 2PSK signal, its wavelet transform amplitude
is the carrier frequency, and ai is the amplitude of the i symbol of is a constant value, so there is only one peak in the amplitude
the ASK signal. histogram. This method can be used to identify {2ASK, 2PSK}.
From equations, it can be seen that within the same symbol, the The feature parameter f5 is the number of peaks in the histogram
wavelet transform amplitude values of the two modulation types of the amplitude value of the signal after wavelet transform to
of signals are all fixed values. The amplitude of wavelet transfor- distinguish {2ASK} and {2PSK}, as shown in Table 3
mation of ASK signal is related to ai . In different symbols, the PSK The signal classification and recognition process based on char-
signal’s wavelet transform amplitude values are the same constant. acteristic parameters f1 ∼f5 is shown in Figure 2
According to the above analysis, the amplitude of ASK signal’s
wavelet transform is stepped. There is a direct relationship between
3 CLASSIFIER DESIGN BASED ON NEURAL
the scale of wavelet transform and the gain of SNR. For signals
under high SNR, the amplitude transformation is obvious, and the NETWORK
peaks in the histogram of wavelet transform are obvious too, which Compared with decision tree classifiers, neural network classifiers
is beneficial to improve the signal recognition rate. According to have stronger pattern recognition capabilities, can quickly adapt to
Y. Xu’s research [14], the SNR gain of wavelet transform has the environmental changes, have good robustness and fault tolerance
following relationship with the scale of wavelet transform: when dealing with nonlinear problems, and have faster recognition
speed and high recognition rate for digital modulation signals [15].
4fs2 4 πa fc
 
GS N R = sin This paper selects BP neural network as the classifier, and uses a
aπ 2 fc2 2fs typical three-layer BP network structure model, including: input
In the above formula, GSNR is the SNR gain of the wavelet trans- layer, hidden layer, and output layer. The network model is shown
form, fc is the carrier frequency, fs is the sampling rate, and a is the in Figure 3
ACAI 2020, December 24–26, 2020, Sanya, China Anyi Wang and Peiru Liu

Figure 1: 2ASK, 2PSK wavelet transform amplitude histogram

too large, it will lead to "overfitting". On the contrary, if the number

of hidden layer neurons is too small, it will lead to "Underfitting".
The empirical formula is shown as follow:
√
m = n +l +a

m is the number of nodes in the hidden layer, n is the number of

nodes in the input layer, l is the number of nodes in the output
layer, and the value of a is an integer between 1 and 10.

4 SIMULATION ANALYSIS
Based on the above analysis, this article uses the MATLAB platform
for modulation signal simulation and signal feature extraction, and
uses the Tensor Flow machine learning software library to com-
plete the construction and training of the BP neural network. The
following simulation experiments and analysis are carried out for
the digital signal modulation recognition problem studied in this
Figure 2: Identification process f1 ∼f5 signal classification article:
based on characteristic parameters Experiment 1: Research on the variation of characteristic pa-
rameters with different SNR.
Using the random sequence generated by MATLAB as the signal
source, the signal types include: 2ASK, 4ASK, 8ASK, 2PSK, 4PSK,
8PSK, 16QAM, 32QAM, the carrier frequency is 10 kHz, the sam-
pling frequency is 60 kHz, and the symbol rate is 1000 b/s, the SNR
ranges from -5 to 20dB, and the added noise is Gaussian white
noise.
The variation of characteristic parameters f1 ∼f4 with the differ-
ent SNR is shown in Figure 4. From the Figure 4(a), it can be seen
that f1 is relatively stable when the SNR is higher than -2.5dB. The
signal can be divided into two types, and the distinction between
the types is more obvious; as shown in Figure 4 (b), f2 is relatively
stable when the SNR is higher than -2dB, and the signals can be
Figure 3: Neural network model divided into three categories, and the various types can be clearly
distinguished; as shown in Figure 4(c), f3 tends to be stable when
the SNR is higher than 0dB. The signals can be divided into three
This paper has 5 characteristic parameters from f1 to f5 , and 8 categories, and the differences between the categories are obvious;
modulation types of signals to be recognized. Therefore, the neural as shown in Figure 4(d), the characteristic parameter f4 tends to be
network structure adopted is: the input layer contains 5 neurons, stable when the SNR is higher than 0dB, and 16QAM and 32QAM
corresponding to the 5 characteristic parameters [f1 f2 f3 f4 f5 ]T . can be distinguished.
The output layer contains 8 neurons, in the form of One-Hot, for The variation of the characteristic parameter f5 with the different
example: the output is [1 0 0 0 0 0 0 0]T . The more hidden layer neu- SNR is shown in Figure 5(a). It shows that 2PSK and 2ASK can be
rons, the higher the network training accuracy, but if the number is effectively identified under low SNR conditions.
Digital signal modulation recognition method based on high-order cumulants and wavelet transform ACAI 2020, December 24–26, 2020, Sanya, China

Figure 4: Characteristic parameters f1 , f2 , f3 , f4 with SNR changing

Figure 5: (a) Characteristic parameters f5 with SNR changing, (b) Average recognition rate of BP neural network model.

In general, the simulation values of characteristic parameters of Experiment 3: The recognition rate of the BP neural network
digital modulation signals have obvious differences, and they are model under different SNR.
all similar to their theoretical values, which proves the correctness This experiment studies the performance of the algorithm un-
of the characteristic parameters selected based on the high-order der different SNR by adding ideal Gaussian white noise. For the
cumulant and wavelet transform theory. composition of the training set, the SNR range of each modulation
In the neural network constructed in this paper, the number of signal is between -5 and 20dB, the SNR step is 1dB, and the signal
input layer nodes n is 5, and the number of output layer nodes l is characteristics of each modulation type corresponding to each type
8. According to formula (16), it can be concluded that the value of of SNR are 1000 groups, that is, each modulation signal charac-
m ranges from 5 to 14, depending on the number of hidden layer teristic is 26000 groups. The test set contains 2600 groups of each
nodes. Compared with the recognition rate of the neural network, modulation signal.
it can be concluded from Table 4 that the number of hidden layer In the Gaussian white noise channel, the average recognition
nodes set in this paper should be 10. rate of the BP neural network model is shown in Figure 5(b), and
ACAI 2020, December 24–26, 2020, Sanya, China Anyi Wang and Peiru Liu

Table 4: Neural network recognition rate corresponding to speed, training time and recognition accuracy has been improved.
the number of nodes in different hidden layers (%) The simulation experiment results show that the digital signal
modulation recognition method proposed in this paper has better
Signal Type The Number of Nodes in hidden layers performance in Gaussian white noise channel. When the SNR is
5 8 10 12 greater than 0dB, the recognition rate of the system can reach over
97%. This method provides new methods and ideas for solving
2ASK 97.3 99.3 100 100
the problem of automatic modulation and recognition of digital
4ASK 94.6 95.3 100 100
signals in Gaussian white noise channels, and has high engineering
8ASK 91.2 93.2 97.1 98.7
practical value placeholder later in the process with the citation
2PSK 91.2 98.4 100 99.0
label as generated by the Word macros in the "master template.
4PSK 93.3 94.3 100 100
8PSK 90.4 94.6 99.1 99.2 ACKNOWLEDGMENTS
16QAM 86.3 92.7 99.3 99.6 This work is supported by National Key Industry Innovation Chain
32QAM 95.0 98.4 100 100 Project of China under Grant no: 2019ZDLSF07-06.

Table 5: Average recognition rate of regulated signal with BP REFERENCES

neural network model (%) [1] P. Popovski, J. J. Nielsen, C. Stefanovic, et al. Wireless access for ultra-reliable
low-latency communication: Principles and building blocks. Ieee Network, 2018,
32(2), 16-23.
Signal Type 0dB 1dB 2dB 5dB 10dB 15dB [2] Y. Tu and Y. Lin. Deep neural network compression technique towards efficient
digital signal modulation recognition in edge device. IEEE Access, 2019, 7, 58113-
2ASK 100 100 100 100 100 100 58119.
4ASK 99.8 100 100 100 100 100 [3] O. A. Dobre, Signal identification for emerging intelligent radios: classical prob-
lems and new challenges. IEEE Instrumentation & Measurement Magazine. 2015,
8ASK 99.2 98.7 98.5 97.1 100 100 18(2), 11-18.
2PSK 97 99 100 100 100 100 [4] T. A. Almohamad, M. F. M. Salleh, M. N. Mahmud, et al. Simultaneous determina-
4PSK 98.6 100 100 100 100 100 tion of modulation types and signal-to-noise ratios using feature-based approach.
IEEE access, 2018, 6, 9262-9271.
8PSK 97.1 98 98.2 99.1 100 100 [5] A. Ali, F. and Yangyu, S. Liu. Automatic modulation classification of digital
16QAM 99.3 100 99.8 99.3 100 100 modulation signals with stacked autoencoders. Digital Signal Processing, 2017,
71, 108-116.
32QAM 99.9 100 100 100 100 100 [6] A. Swami and B. M. Sadler. Hierarchical digital modulation classification using
cumulants. IEEE Transactions on communications, 2000, 48(3), 416-429.
[7] F. Zhou, R. Hang, Q. Liu, et al. Hyperspectral image classification using spectral-
the recognition rate of each digital signal is shown in Table 5. It can spatial LSTMs. Neurocomputing, 2019, 328, 39-47.
be seen that in the Gaussian white noise channel, when the SNR is [8] K. Jiang, J. Zhang, H. Wu, et al. A Novel Digital Modulation Recognition Algorithm
Based on Deep Convolutional Neural Network. Applied Sciences, 2020, 10(3),
higher than 0dB, the recognition accuracy of the method proposed 1166.
in this paper can reach higher than 97%. [9] C. Mishra and D. L. Gupta. Deep machine learning and neural networks: An
overview. IAES International Journal of Artificial Intelligence, 2017, 6(2), 66-73.
[10] Y. LeCun, Y. Bengio and G. Hinton. Deep learning. Nature, 2015, 521(7553), 436-
5 CONCLUSIONS 444.
[11] T. O’Shea and J. Hoydis. An introduction to deep learning for the physical layer.
This paper studies the digital signal modulation recognition method IEEE Transactions on Cognitive Communications and Networking, 2017, 3(4),
based on high-order cumulant and wavelet transform. The theoreti- 563-575.
cal values of the high-order cumulants of 8 types of digital modula- [12] M. Liu, T. Song, J. Hu, et al. Deep learning-inspired message passing algorithm
for efficient resource allocation in cognitive radio networks. IEEE Transactions
tion signals of 2ASK, 4ASK, 8ASK, 2PSK, 4PSK, 8PSK, 16QAM, and on Vehicular Technology, 2018, 68(1), 641-653.
32QAM are obtained, and 4 new cumulants based on the 2-, 4- and [13] J. J. Guo, H. D. Yin, L. Jiang, et al. Recognition of digital modulation signals via
6-order cumulants are constructed. The feature parameters f 1 to high-order cumulants. Communications Technology, 2014, 47(11), 1255-1260.
[14] Y. Xu, L. Ge and B. Wang, Digital Modulation Recognition Method Based on Tree-
f 4 can divide the 8 digital signals to be identified into 7 categories, Structured Neural Networks, 2009 International Conference on Communication
and the remaining modulation category signals are classified and Software and Networks, Macau, 2009, 708-712.
[15] N. Daldal, Z. Cömert and K. Polat. Automatic determination of digital modulation
identified by the feature parameter f 5 based on wavelet transform. types with different noises using Convolutional Neural Network based on time–
For the selection of classifiers, this paper adopts the BP neural frequency information. Applied Soft Computing, 2020, 86, 105834.Prokop, Emily.
network model without decision threshold, and the convergence 2018. The Story Behind. Mango Publishing Group. Florida, USA.