SPECIAL SECTION ON ARTIFICIAL INTELLIGENCE FOR

PHYSICAL-LAYER WIRELESS COMMUNICATIONS

Received February 20, 2019, accepted February 28, 2019, date of publication March 5, 2019, date of current version March 25, 2019.
Digital Object Identifier 10.1109/ACCESS.2019.2903130

Deep Learning for Signal Demodulation in

Physical Layer Wireless Communications:
Prototype Platform, Open Dataset, and Analytics
HONGMEI WANG1 , ZHENZHEN WU1 , SHUAI MA 1,2 , SONGTAO LU 3, (Member, IEEE),
HAN ZHANG 4 , GUORU DING 5 , AND SHIYIN LI1
1 School of Information and Control Engineering, China University of Mining Technology, Xuzhou 221116, China
2 StateKey Laboratory of Integrated Services Networks, Xidian University, Xi’an 710071, China
3 Department of Electrical and Computer Engineering, University of Minnesota Twin Cities, Minneapolis, MN 55455, USA
4 Department of Electrical and Computer Engineering, University of California at Davis, Davis, CA 95616, USA
5 College of Communications Engineering, Army Engineering University, Nanjing 210007, China

Corresponding author: Shuai Ma ([email protected])

The work of H. Wang was supported by the National Natural Science Foundation of China under Grant 61601464. The work of S. Ma and
S. Li was supported in part by the Fundamental Research Funds for the Central Universities under Grant 2017QNA32, in part by the
National Natural Science Foundation of China under Grant 61701501 and Grant 61771474, in part by the Natural Science Foundation of
Jiangsu Province under Grant BK20170287, in part by the China Postdoctoral Science Foundation under Grant 2016M600452, in part by
the State Key Laboratory of Integrated Services Networks, Xidian University, under Grant ISN19-07, in part by the Key Laboratory of
Cognitive Radio and Information Processing, Ministry of Education, Guilin University of Electronic Technology, under Grant
CRKL180204, and in part by the Key Laboratory of Ocean Observation-Imaging Testbed of Zhejiang Province.

ABSTRACT In this paper, we investigate deep learning (DL)-enabled signal demodulation methods and
establish the first open dataset of real modulated signals for wireless communication systems. Specifically,
we propose a flexible communication prototype platform for measuring real modulation dataset. Then,
based on the measured dataset, two DL-based demodulators, called deep belief network (DBN)-support
vector machine (SVM) demodulator and adaptive boosting (AdaBoost)-based demodulator, are proposed.
The proposed DBN-SVM based demodulator exploits the advantages of both DBN and SVM, i.e., the
advantage of DBN as a feature extractor and SVM as a feature classifier. In DBN-SVM based demodulator,
the received signals are normalized before being fed to the DBN network. Furthermore, an AdaBoost-
based demodulator is developed, which employs the k-nearest neighbor as a weak classifier to form a
strong combined classifier. Finally, the experimental results indicate that the proposed DBN-SVM based
demodulator and AdaBoost-based demodulator are superior to the single classification method using DBN,
SVM, and maximum likelihood-based demodulator.

INDEX TERMS Machine learning, DBN-SVM based demodulator, AdaBoost based demodulator.

I. INTRODUCTION complex problems, has attracted increasing attention from

Conventional wireless communication systems are generally both academia and industry. DL has been applied in image
designed in accordance with the rigorous mathematical the- recognition [4], [5], computer vision [6], natural language
ories and accurate system models [1]. However, because of processing [7] and spectrum prediction [8], etc. In addition,
increasing wireless service requirements, such as the use some literatures have focused on using DL to optimize perfor-
of smartphones, virtual reality, and internet of things (IoT), mance of wireless communication systems [9]–[11]. In [9],
it is challenging to characterize future complex wire- an unsupervised learning-based fast beamforming method
less communication networks accurately by using tractable is proposed to maximize the weighted sum rate under the
mathematical models or system models [2]. Recently, total power constraint. In [10], a deep recurrent neural net-
deep learning (DL) [3], as an effective method to handle work based algorithm is proposed to tackle energy efficient
resource allocation problem for heterogeneous IoT. In [11],
The associate editor coordinating the review of this manuscript and
a three dimensional message-passing algorithm based on
approving it for publication was Guan Gui. deep learning scheme is proposed to minimize the weighted

2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
30792 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

sum of the secondary interference power for cognitive without any prior knowledge of the channel model. More-
radio networks. Recent works [12], [13] have interpreted over, the performance of the two data-driven demodulators
an end-to-end wireless communication system as an auto- are evaluated on different modulation schemes through real
encoder. This is promising for applications of DL to wireless measured data. The main contributions of this paper are as
communications. follows:
Demodulation is one of the fundamental modules for wire- • A flexible end-to-end wireless communication proto-
less communications systems for high-speed transmission type platform is developed for application in real phys-
with a low bit error rate. Theoretically, optimum demod- ical environments, which can generate real signals. The
ulators of conventional wireless communication systems prototype is used to establish measured modulation
are designed for additive white Gaussian noise (AWGN) datasets from real communication systems in actual
channels [1]. Moreover, both channel state information (CSI) physical environments in eight modulation schemes,
and channel noise distribution are usually required. Most i.e., binary phase shift keying (BPSK) and multiple
previous studies [14]–[19] have assumed that each receiver quadrature amplitude modulation (M -QAM) modula-
can accurately estimate the fading coefficients. However, tion, where M = 2φ and φ = {2, 3, 4, 5, 6, 7, 8}.
practical wireless communication channels may suffer from The received SNR of the eight modulated signals
multi-path fading, impulse noise, spurious or continuous are measured from 3 dB to 25 dB. An open online
jamming, and numerous other complex impairments, which real modulated dataset is established, available at
deteriorate demodulation performance significantly. Because https://pan.baidu.com/s/1biDooH6E81Toxa2u4D3p2g
of the limited length of the training sequence, the estimate or https://drive.google.com/open?id=1jXO9OMZOyV
CSI will have limited accuracy [20]. Especially, for fast- MOYvQSn3WVmlfQoQbonKuo, where the transmis-
fading scenarios, it is difficult to accurately estimate CSI sion distance of the eight modulated signals is measured
because the fading coefficients change rapidly during the data in an indoor environment. To the best of our knowledge,
transmission period. Designing optimum demodulators for this is the first open dataset of real modulated signals for
different channel models is challenging because the channel wireless communication systems.
model may not be known at the receiver end. • Then, based on the measured data, two DL-based
Given the above issues, DL-based model-free demodu- demodulators are proposed, namely, DBN-SVM based
lators have attracted a considerable amount of attention, demodulator and AdaBoost based demodulator. The
where the requirements for a priori knowledge can be widely proposed DBN-SVM based demodulator, which has a
relaxed or even removed [21]. Because the information of novel demodulation architecture, exploits the advan-
the modulated signals is represented by the amplitude and tages of both DBN and SVM, i.e., the advantage of DBN
phase, feature extraction is of critical importance for signal as a feature extractor and SVM as a feature classifier. To
demodulation. accelerate the convergence rate, the received signals are
DL-based demodulators have been investigated in first normalized before being fed to the DBN network so
conventional radio frequency (RF) systems. In [22], a deep that the features of the received signals can be extracted,
convolutional network demodulator (DCND) is proposed the SVM is utilized to classify these features.
to demodulate mixed modulated signals, which can fur- • An AdaBoost based demodulator, which utilizes multi-
ther reduce the bit error rate compared with the coherent ple KNNs as a weak classifier to form a strong combined
demodulation method. Fan and Wu [23] show that the pro- classifier, is developed. The proposed AdaBoost based
posed demodulator based on deep belief network (DBN) demodulator increases the weights for the error demod-
is feasible for an AWGN channel with a certain channel ulated symbols and decreases the corresponding weights
impulse response and a Rayleigh non-frequency-selective flat for correctly demodulated symbols during the iterations.
fading channel. In [24], a DL-based detector is proposed for • Finally, the demodulation performance of the two
signal demodulation in short-range multi-channels without proposed data-driven demodulators are investigated.
a signal equalizer. Mohammad et al. [25] show that deep Specifically, the demodulation accuracies of the two
convolutional neural networks (DCNN) for frequency-shift DL-based demodulators decrease over the respective
keying (FSK) demodulation can substantially reduce error transmission and modulation orders for a fixed trans-
bit probabilities over an AWGN Rayleigh-fading channel. mission distance. The experimental results also show
To the best of our knowledge, most of existing DL-based that the demodulation accuracy of the DBN-SVM based
demodulation schemes are based on simulated data rather demodulator is higher than those of DBN-based and
than real measured data. SVM-based demodulators. Moreover, the demodula-
This paper presents a data-driven framework for DL-based tion accuracy of the AdaBoost based demodulator is
demodulators. Specifically, two data-driven demodulation higher than that of the DBN-SVM based demodulator
methods based on DBN-support vector machine (SVM) and at the lower SNR regions, and the accuracies of the
adaptive boosting (AdaBoost) [26] are developed for end-to- two demodulators are similar at high SNRs. For the
end wireless communication systems. These methods learn high SNR scenario, a high-order modulation is generally
and extract features from the received modulation signals preferred.

VOLUME 7, 2019 30793

 H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

T
ŷ1 , ŷ2 , . . . , ŷNL is given by


yi − ymin
ŷi = , 1 ≤ i ≤ NL, (3)
ymax − ymin
where ymin = min yi , and ymax = max yi .
1≤i≤NL 1≤i≤NL
FIGURE 1. End-to-end wireless system model. Because the information of the BPSK and M -QAM are
represented by amplitudes and phases, DL is used to extract
information features from the received signals. Specifically,
The remainder of this paper is organized as follows. with the sampled signal vector y, two DL-based demodulators
Section II describes the system model. Section III explores are proposed: DBN-SVM based demodulator and AdaBoost
the structures of the DBN-SVM and AdaBoost, including based demodulator. The DL-based demodulators consist of
detailed descriptions of the data stream and how to make two phases: training phase and testing phase. During the
classification decisions. In section IV, the data analysis results training phase, the parameters of the DL-based demodulators
are provided and analyzed. Finally, the conclusions from the are optimized with the training dataset. Then, in the testing
study are drawn in Section V. phase, the demodulators demodulate the received signal and
Notations: Boldfaced lowercase and uppercase letters rep- recover the transmitted information.
resent vectors and matrices, respectively. The transpose of Let zl denote the label signal of the lth period, where
1 1 l ∈ L and 8 is the label set, i.e., 8 = {z1 , z2 , . . . , zL },
a matrix is denoted as (·)T . L = {1, 2, . . . , L}, L1 =
1 1 which is
determined by the modulation scheme. Let T1 =
{1, . . . , L1 }, Mk = {1, 2, . . . , Mk }, Nk = {1, 2, . . . , Nk },
ŷ1 , z1 , ŷ2 , z2 , . . . , ŷL1 , zL1 denote the labeled train-



1 1 1
D = {1, . . . , D}, Q = {0, 1, . . . , M̄ }, and M = T
ing signal set, where ŷl = ŷ1+(l−1)N , ŷ2+(l−1)N , . . . , ŷlN

{0, 1, . . . , M − 1}.
denotes the normalized signal of the lth period, and L1
denotes the total number of training signal periods (L1 < L).
II. SYSTEM MODEL
An end-to-end wireless communication system1 is consid-
III. DBN-SVM BASED DEMODULATOR
ered, which includes a single antenna transmitter and a single
As an unsupervised features extraction method, the DBN can
antenna receiver, as illustrated in Fig. 1. By adopting the
efficiently extract high-level and hierarchical features from
BPSK or M -QAM digital modulation schemes, the transmit-
the measured signal, while the SVM minimizes the structure
ted signal x (t) is given as
risk and shows good learning and generalization performance
with a small amount of samples. Inspired by those advantages
x (t) = Vm cos (2π fc t +θm ), m = 1, . . . , M , 1 ≤ t ≤ T ,
of the two approaches, a combination of DBN and SVM for
(1) demodulation is proposed. The DBN-SVM demodulator is
shown in Fig. 2, the DBN is used as a feature generator and
where Vm , θm and T denote the amplitude, phase, and period the SVM is used as a classifier.
of the signal x (t), respectively; fc is the carrier frequency.
Let g (t) denote the multipath channel between the trans- A. DBN
mitter and the receiver, which may suffer nonlinear distortion,
The proposed DBN includes three stacked restricted Boltz-
interference, and frequency selective fading. At the receiver,
mann machines (RBM) [31], i.e., RBM1 , RBM2 , and RBM3 ,
the received signal y (t) is given by
as shown in Fig. 2. Specifically, RBMk is an undirected,
bipartite graphical model, and it composes a visible layer
y (t) = g (t) x (t) + nr (t), (2)
vk = [vk,1 , vk,2 , . . . , vk,Mk ]T and a hidden layer hk =
[hk,1 , hk,2 , . . . , hk,Nk ]T , where vk,α and hk,β are the αth neu-
where nr (t) denotes the received noise.
ron of vk and the βth neuron of hk , respectively, α ∈ Mk ,
Then, the received analog signal y (t) is converted to
1 β ∈ Nk , k ∈ {1, 2, 3}. The visible layer vk and hidden layer
the digital signal via the vector signal analyzer. Let y = hk are fully connected via a symmetric undirected weighted
[y1 , y2 , . . . , yNL ] denotethe total sampled digital signal vec-
T
T
matrix Wk = wk,1 , wk,2 , . . . , wk,Nk , where wk,β =

tor, where yn = y n−1 N T is the nth sample, N is the number (1) (2) (M )
[wk,β , wk,β , . . . , wk,βk ]T is a weight vector between vk and
of samples of one period, and L denotes the number of signal hk,β . For the three RBM, there is no intralayer connections
periods. between either the visible layer or the hidden layer.
Before the demodulation process, the received signal y is For RBMk , the energy E (vk , hk ) is defined by combining
normalized to [0, 1], which can accelerate the DL network the configuration of both vk and hk as follows
1
processing speed [30]. Senerally, the normalized data ŷ =
E (vk , hk ) = −hTk Wk vk − aTk vk − bTk hk , (4)
1 The term end-to-end wireless system model implies that signal features T
where ak = ak,1 , ak,2 , . . . , ak,MK is an offset vector of vk ,

are learned from a single deep neural network, without the complex multi- T
and bk = bk,1 , bk,2 , . . . , bk,NK is an offset vector of hk .

stage expert machine learning processing [12], [13], [27]–[29].

30794 VOLUME 7, 2019

H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

by
 
p hk,β = 1|vk = sigmoid bk,β + vk T wk,β ,


(8a)
 
Nk
(α)
X
hk,β wk,β ,


p vk,α = 1 |hk = sigmoid ak,α + (8b)
β=1

1
where sigmoid (x) = 1+e1 −x , α ∈ Mk , β ∈ Nk , hk,β , and
vk,α ∈ [0, 1].
Then, the variables Wk , ak , and bk are updated by the
following equations [33]
(α) (α) ∂ log p (vk )
wk+1,β ← wk,β +η (α)
, α ∈ Mk , β ∈ Nk , (9a)
∂wk,β
∂ log p (vk )
ak+1,α ← ak,α +η , α ∈ Mk , (9b)
∂ak,α
∂ log p (vk )
bk+1,β ← bk,β +η , β ∈ Nk , (9c)
∂bk,β
where η > 0 is the learning rate.
By employing the gradient descent method, RBM1 is
FIGURE 2. Structure of DBN-SVM based demodulator. trained first, where v1 = ŷl and l ∈ L1 . Then, let
v2 = h1 , and RBM2 is trained. Similarly, after train-
ing RBM2 , let v3 = h2 , and RBM3 is trained. More-
Based on E (vk , hk ), the probability of vk is given by over, when RBM3 is trained, the parameters of DBN can
1 X E(vk ,hk ) be obtained, i.e., {Wk , ak , bk }k∈{1,2,3} . Then, the parameters
p (vk ) = e , (5) {Wk , ak , bk }k∈{1,2,3} are further fine-tuned by the supervised
Zk
hk back propagation (BP) algorithm [34].
After DBN is trained, it outputs the extracted feature ȳl1 =
eE(vk ,hk ) is the normalization factor.
P
where Zk = T
h3 , where l1 ∈ L1 . Let Ȳ = ȳ1 , ȳ2 , . . . , ȳL1 denote the

vk ,hk
During the training phrase, the goal of the RBMk is to output feature set.
maximize the log-likelihood function as follows
X B. OVO-SVM
max log p (vk ). (6) With the extracted feature set Ȳ , the one-versus-one
Wk ,ak ,bk
vk (OVO)-SVM is adopted for further classification, which
To solve equation (6), the gradient descent method is used achieves multiclassification by solving the two-classification
to iteratively calculate the variables Wk , ak , and bk , where the subproblems [35], [36]. As shown in Fig. 2, OVO-SVM
corresponding partial derivative with respect to Wk , ak , and exploits M̄ nonlinear two-class SVMs, i.e., SVM0 ,. . . ,SVMM̄ ,
bk can be written as to classify M categories for M -QAM modulation, where
1
M̄ = M (M2 −1) − 1.
To map pedestrian features to a high dimensional space,
∂ log p (vk )

(α)
= vk,α p hk,β = 1|vk a Gaussian kernel is introduced, which can be expressed as
∂wk,β
ȳl − ȳl 2
!
X
p (vk ) p hk,β = 1|vk vk,α , Gq ȳl1 , ȳl2 = exp − 1 2
,



− (10)
vk
2σq2
α ∈ Mk , β ∈ Nk , (7a) where σq > 0 is the bandwidth of the Gaussian kernel and
∂ log p (vk ) X
q ∈ Q.
= vk,α − p (vk )vk,α , α ∈ Mk , (7b)
∂ak,α v
According to the nonlinear SVM theory [37], the nonlinear
k
∂ log p (vk )
two-class SVMq problem can be formulated as
= p hk,β = 1|vk
∂bk,β 1X
L1 X
L1

XL1
X min cq,l1 cq,l2 zl1 zl2 Gq ȳl1 , ȳl2 − cq,l1 (11a)
p (vk ) p hk,β = 1|vk , β ∈ Nk . (7c)


− cq 2
l1 =1l2 =1 l1 =1
vk
L1
X
According
to [32], the conditional

probability s.t. cq,l1 zl1 = 0, (11b)
p hk,β = 1|vk and p vk,α = 1 |hk are respectively given l1 =1

VOLUME 7, 2019 30795

 H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

Algorithm 1 The DBN-SVM Based Demodulator

1. Given the labeled training signal ŷl , l ∈ L1 ;
2. Initialize v1 , h1 , W1 , a1 and b1 ;
3. For k = 1, . . . , 3 do
4. Train kth RBM according to formula (4) − (6);
5. Update Wk , ak and bk are according to formula
(7) − (9);
6. End for;
7. Get the extracted feature vector set Ȳ , and classified by
OVO-SVM; FIGURE 3. Structure of AdaBoost based demodulator.
8. Update Adderm , m ∈ M;
9. Output label ẑ:
ẑ = arg max {um }. weakly classifiers into a stronger classifier. In this section,
m∈M
we exploit the k-Nearest Neighbor (KNN) classifiers as the
weak classifier for constructing the AdaBoost.
As shown in Fig. 3, the proposed AdaBoost consists D
0 ≤ cq,l1 ≤ K , l1 ∈ L1 , (11c) KNN classifiers.
The labeled training signal set is denoted
by T = ŷ1 , z1 , ŷ2 , z2 , . . . , ŷL1 , zL1 .
T 


where cq = cq,1 , cq,2 , . . . , cq,L1 and q ∈ Q.


By hsolving linear programming (11), the optimal solution Let wd = [wd (1), wd (2), . . . , wd (L1 )]T denote the
i T weight vector of dth KNN, where 0 ≤ wd (l) ≤ 1, l ∈ L1 ,
c∗q = c∗q,1 , c∗q,2 , . . . , c∗q,L1 is obtained. Then, the nonlinear L1
wd (l) = 1, d ∈ D. For the 1st KNN, w1 (l) = L11 ,
P


two-class SVMq decision function fq ȳl1 , with l1 ∈ L1 , q ∈ and
l=1
Q, is given as l ∈ L1 . Based on the weight vector wd , the dth KNN re-
samples
 the training set T
and generates a
new training set
L1 ȳi − ȳl 2
! !
Td = ỹd1 , zd1 , ỹd2 , zd2 , . . . , ỹdL1 , zdL1 , dl ∈ L1 .
X

fq ȳl1 = γ ∗ 1
+ bq , (12)
∗


cq,i zi exp −
2σq2 Then, a vector in Td is searched with the minimum distance
i=1
from ŷl , i.e.,
L1
 
1 ȳi −ȳl 2
b∗q = c∗q,l1 zi exp
P 1
where zl1 − − 2σ 2 is a biased
l ∗ = arg min ỹdi − ŷl 2 , l ∈ L1 , d ∈ D.

i=1 q (15)
variable [38], and i∈L1
(

1 1, if x ≥ 0 Because the label of ỹdl∗ is zdl∗ , fd ŷl = zdl∗ , which implies
γ (x) = (13)
0, if x < 0. that the classification result of dth KNN for ŷl is zdl∗ .
Let χd denote the weight sum of misclassified samples of
Let τq,m̄ and τq,m denote the output of the SVMq , and dth KNN as follows
τq,m̄ and τq,m are the inputs of the Adderm̄ and the Adderm ,
L1
respectively, where τq,m̄ , τq,m ∈ {0, 1}, τq,m̄ + τq,m = 1, m̄, X
χd = wd (l) I fd ŷl , zl ,



m ∈ M, and m̄ 6 = m. Then, for Adderm , the number of votes d ∈ D, (16)
is updated by um = um + τq,m , where the initial value of um l=1

is 0, and m ∈ M. where I (x, y) is the indicator function, i.e.,

−1
Then, with the number of votes {um }M m=0 , the output label (
ẑ is obtained as follows 1, if x 6= y
I (x, y) = (17)
ẑ = arg max {um }. (14) 0, if x = y.
m∈M
Then, for (d + 1)th KNN, weight w(d+1) =
Finally, ẑ is mapped to the demodulation result ŝ. T
w(d+1) (1), . . . , w(d+1) (L1 ) is updated as

After the entire network is trained, the parameters Wk , ak ,
and bk of the DBN, and cq , b∗q , σq of the OVO-SVM are
wd (l) exp −αd I fd ŷl , zl




optimized, where k ∈ {1,
2, 3} and q ∈
Q. Then, the
test wd+1 (l) = , l ∈ L1 , d ∈ D,
signal T2 = ŷL1 +1 , zL1 +1 , ŷL1 +2 , zL1 +2 , . . . , ŷL , zL is
 Qd
converted to the feature vector ȳ, where L2 is the number (18)
of test signal periods. The details of the DBN-SVM based   L1
1−χd
where αd 1
wd (l)
P
demodulator are listed in Algorithm 1. = 2 ln χd , and Qd =
l=1
exp −αd I fd ŷl , zl is the normalization factor. If ŷl is




IV. AdaBoost BASED DEMODULATOR
classified correctly, i.e., I fd ŷl , zl = 0, wd+1 (l) = wQd (l)



AdaBoost is a general method used to improve machine learn- d
.
wd (l) exp(−αd )
ing algorithms [39], which integrates multiple independent Otherwise, wd+1 (l) = Qd .

30796 VOLUME 7, 2019

H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

Algorithm 2 The KNN Based AdaBoost Demodulator

1. Given the labeled training signal set T ;
2. Initialize signal weight w1 (l) = L11 ;
3. For d = 1, . . . , D do
4. Train dth KNN according
to weights wd ;
5. Get weak classifier fd ŷl ∈ M with error rate
L1
χd = wd (l) I fd ŷl , zl
P


d ∈ D;
l=1
6. Update:
wd (l) exp(−αd I (fd (ŷl ),zl ))
wd+1 (l) = , l ∈ L1
 Qd
1−χd
where αd = 2 ln χd , and
1

L1
wd (l) exp −αd I fd ŷl , zl ;
P




Qd =
l=1
7. End for;
8. Output the final decision classifier: FIGURE 4. End-to-end wireless communication system prototype.
D
αd 1 − I fd ŷl , zl .

P



F ŷl = zˆl = arg max
z∈8 d=1
TABLE 1. Experimental equipment and parameters.

After training D KNNs, AdaBoost classifies ŷl as follows

D
X
αd 1 − I fd ŷl , zl ,





F ŷl = zˆl = arg max (19)
zl ∈8 d=1

where α
d
is the coefficient of 1 − I fd ŷl , zl

and
I fd ŷl
, zl
can be regarded as the voting value, i.e., if
I fd ŷl , zl = 0, fd ŷ classifies signal ŷl into class zl , oth-

The carrier frequency fc and the sampling rate fs are
erwise, ŷl does not belong to class zl . The class with the
max- 2.4 GHz and 100 MHz/s, respectively. For each M -QAM
imum sum of weighted voting value, αd 1 − I fd ŷl , zl ,


modulation scheme, the number of sample points N has four
for all classifiers, is identified as the classification result ẑl cases, i.e., N = 10, 20, 40, and 80.
of the Adaboost classifier, and then ẑl is mapped to demod- To reduce the generalization error, the collected data set
ulation result ŝl . The details of the KNN-based AdaBoost contains 10000 transmit signal periods, in which 8000 periods
demodulator are listed in Algorithm 2. are used for training and 2000 periods are used for testing.

V. EXPERIMENTAL RESULTS AND DISCUSSIONS B. EXPERIMENTAL RESULTS

In this section, the performance of the proposed DBN-SVM based and AdaBoost based demodulators are
DBN-SVM based demodulator and AdaBoost based demod- trained on these training sets. The DBN-SVM based demod-
ulator is investigated. Also the performance of the the DBN ulator training ends after 110 epochs, after which the training
based, SVM based, and maximum likelihood (MLD) based loss almost does not decline, and the AdaBoost based demod-
demodulation methods are presented for comparison. ulator training ends when the iteration error is less than 10−3 .
In the experiment, signal sets with different SNRs, ranging
A. THE END-TO-END WIRELESS COMMUNICATION from 3 to 25 dB, are chosen as the validation sets; the DBN
SYSTEM PROTOTYPE based, SVM based, and MLD based demodulation methods
As shown in Fig. 4, an end-to-end wireless communication are used for comparison.
system prototype is first established to collect the dataset, In Fig. 5 and Fig. 6, the demodulation performance versus
which consists of a source, a RF vector signal generator, SNR of the proposed demodulator and the three baseline
a transmitter antenna, a receiver antenna, and a vector signal schemes are compared by the demodulation of 4-QAM and
analyzer. The parameters of the devices of the proposed end- 16-QAM, respectively. The demodulation accuracy of the
to-end wireless communication system prototype are listed models increases as SNR increases. In particular, Fig. 5
in Table 1. indicates that the demodulation accuracy of all methods
The volume environment is a 15×5×3 m3 office, where


are close to 100% when SNR ≥ 15dB, and the proposed
15, 5, and 3 denote the length, width, and height, respectively. AdaBoost based demodulator is significantly superior to the
Note that the distance between the transmitter and the receiver other models when SNR ≤ 13dB. Besides, the proposed
is approximately 10 meters. The power of the background DBN-SVM based demodulator has better performance than
noise is 78 dBm. the DBN-based and SVM-based demodulation methods.

VOLUME 7, 2019 30797

 H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

FIGURE 5. Demodulation accuracy comparison of different demodulators

with N = 40 and 4-QAM. FIGURE 7. Demodulation accuracy of AdaBoost versus SNR with 16-QAM.

FIGURE 6. Demodulation accuracy comparison of different demodulators FIGURE 8. Demodulation performance versus different number of
with N = 40 and 16-QAM. training signal periods with 16-QAM and 32-QAM.

In Fig. 6, compared with Fig. 5, we focus on the same per-

formance index at 16-QAM. It shows the designed AdaBoost approximately achieve 100% with N = 40 or N = 80 when
based demodulator is close to 100% when SNR ≥ 15dB. SNR ≥ 15dB. However, with an increase in the number of
However, other methods cannot approach 100% as SNR sample points, the computational complexity also increases.
increases. Furthermore, among these demodulation methods, Fig. 8 shows the demodulation accuracy achieved by the
the AdaBoost based demodulator obviously outperform the AdaBoost based demodulator versus the number of training
other four methods. It can be observed that the demodulation signal periods, where the number of sampling points is 40 and
accuracy achieved by DBN-SVM based demodulator exceeds SNR = 12dB. The result shows that the demodulation accu-
ones by the DBN-based, SVM-based demodulation methods. racy initially increases with an increase in the number of train-
Although the overall trend of MLD classification accuracy ing signal periods, and then, it reaches saturation when the
increases as SNR increases, it has a obvious fluctuation. number of training signal periods is 5000. It can be observed
The reason is that the practical wireless channels include that, compared with 32-QAM, 16-QAM can achieve higher
complicaful interferences, but the robustness of MLD is poor. accuracy. Meanwhile, 16-QAM can provide stable perfor-
In Fig. 7, the accuracy performance for different sam- mance with relatively fewer training signal periods. Different
pling points at 16-QAM is simulated. It can be observed modulation models have different requirements with different
that the demodulation accuracy increases with the number of number of training signals periods. In general, higher orders
sample points. Furthermore, the demodulation accuracy can require longer training signals periods.

30798 VOLUME 7, 2019

H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

positive impact on the performance of the transmission capac-

ity. The performance gap between the low order and the high
order modulation is clearer when SNR ≤ 15dB. However,
the demodulation accuracy of high order modulation is low,
so there is a trade-off between the demodulation accuracy and
the effective capacity.
VI. CONCLUSION
In this paper, a flexible end-to-end wireless communications
prototype platform was proposed for real physical environ-
ments. Then, the first open measured modulation data dataset
with eight modulation schemes, i.e., BPSK, 4-QAM, 8-QAM,
16-QAM, 32-QAM, 64-QAM, 128-QAM, and 256-QAM,
were established and accessed online. Furthermore, two
DL-based demodulators, i.e., DBN-SVM based demodulator
and AdaBoost based demodulator, were proposed. Based
on the real dataset, the demodulation performance of the
proposed demodulators were tested. Finally, experimental
FIGURE 9. Demodulation accuracy comparison of different modulation results indicated that the proposed demodulators outperform
mode versus SNR with AdaBoost based demodulator.
the DBN based, SVM based, and MLD based demodulators
at various scenarios.
REFERENCES
[1] J. G. Proakis, Digital Communications. New York, NY, USA:
McGraw-Hill, 2001.
[2] M. Shafi et al., ‘‘5G: A tutorial overview of standards, trials, challenges,
deployment, and practice,’’ IEEE J. Sel. Areas Commun., vol. 35, no. 6,
pp. 1201–1221, Jun. 2017.
[3] Y. LeCun, Y. Bengio, and G. Hinton, ‘‘Deep learning,’’ Nature, vol. 521,
no. 7553, pp. 436–444, May 2015.
[4] L. Chang, C. Yu, L. Yan, G. Chen, V. Vokkarane, and Y. Ma, ‘‘DeepFood:
Deep learning-based food image recognition for computer-aided dietary
assessment,’’ in Proc. Int. Conf. Smart Homes Health Telematics, vol. 9677.
Berlin, Germany: Springer, 2016, pp. 37–48.
[5] T. Zhou, S. Yang, L. Wang, J. Yao, and G. Gui, ‘‘Improved cross-label
suppression dictionary learning for face recognition,’’ IEEE Access, vol. 6,
pp. 48716–48725, 2018.
[6] D. Gerónimo, J. Serrat, A. M. López, and R. Baldrich, ‘‘Traffic sign recog-
nition for computer vision project-based learning,’’ IEEE Trans. Educ.,
vol. 56, no. 3, pp. 364–371, Aug. 2013.
[7] T. Tan, Y. Qian, and K. Yu, ‘‘Cluster adaptive training for deep neural
network based acoustic model,’’ IEEE/ACM Trans. Audio, Speech, Lang.
Process., vol. 24, no. 3, pp. 459–468, Mar. 2016.
[8] L. Yu, J. Chen, G. Ding, Y. Tu, J. Yang, and J. Sun, ‘‘Spectrum prediction
FIGURE 10. Effective capacity comparison of different modulation mode based on taguchi method in deep learning with long short-term memory,’’
versus SNR with AdaBoost based demodulator. IEEE Access, vol. 6, pp. 45923–45933, 2018.
[9] H. Huang, W. Xia, J. Xiong, J. Yang, G. Zheng, and X. Zhu, ‘‘Unsuper-
vised learning-based fast beamforming design for downlink MIMO,’’ IEEE
Fig. 9 presents the demodulation accuracies of BPSK Access, vol. 7, pp. 7599–7605, 2018.
and M -QAM modulation schemes. In this experiment, [10] M. Liu, T. Song, and G. Gui, ‘‘Deep cognitive perspective: Resource
allocation for NOMA based heterogeneous IoT with imperfect SIC,’’ IEEE
the AdaBoost based demodulation algorithm was employed, Internet Things J., to be published. doi: 10.1109/JIOT.2018.2876152.
where the number of sampling points is N = 40. The [11] M. Liu, T. Song, J. Hu, J. Yang, and G. Gui, ‘‘Deep learning-inspired
demodulation accuracy for all modulation schemes increases message passing algorithm for efficient resource allocation in cognitive
radio networks,’’ IEEE Trans. Veh. Technol., vol. 68, no. 1, pp. 641–653,
with SNR. Meanwhile, the accuracy achieved by the BPSK- Jan. 2019.
modulation scheme is better than the other seven schemes [12] T. O’Shea and J. Hoydis, ‘‘An introduction to deep learning for the physical
for the same SNR. Furthermore, Fig. 9 also indicates that layer,’’ IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575,
Dec. 2017.
the demodulation accuracy reduces with an increase of the [13] S. Dörner, S. Cammerer, J. Hoydis, and S. T. Brink, ‘‘Deep learning based
modulation order. communication over the air,’’ IEEE J. Sel. Topics Signal Process., vol. 12,
In Fig. 10, the same modulation schemes, demodulation no. 1, pp. 132–143, Feb. 2018.
[14] T. Fath and H. Haas, ‘‘Performance comparison of MIMO techniques for
algorithm, and sampling points are used as in Fig. 9, where optical wireless communications in indoor environments,’’ IEEE Trans.
the effective capacity of different modulation methods versus Commun., vol. 61, no. 2, pp. 733–742, Feb. 2013.
SNR are reported. The effective capacity by BPSK, 4-QAM, [15] T. Q. Wang, Y. A. Sekercioglu, and J. Armstrong, ‘‘Analysis of an
optical wireless receiver using a hemispherical lens with application in
and 8-QAM almost remain unchanged with an increase in MIMO visible light communications,’’ J. Lightw. Technol., vol. 31, no. 11,
SNR. It is found that the modulation order has a considerable pp. 1744–1754, Apr. 12, 2013.

VOLUME 7, 2019 30799

 H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

[16] K. Ying, H. Qian, R. J. Baxley, and S. Yao, ‘‘Joint optimization of precoder HONGMEI WANG received the Ph.D. degree
and equalizer in MIMO VLC systems,’’ IEEE J. Sel. Areas Commun., from the School of Mechatronics Engineering,
vol. 33, no. 9, pp. 1949–1958, Sep. 2015. Changwon National University, South Korea,
[17] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, ‘‘Deep learning in 2012. She is currently a Lecturer with the
for super-resolution channel estimation and doa estimation based massive School of Information and Control Engineering,
MIMO system,’’ IEEE Trans. Veh. Technol., vol. 67, no. 9, pp. 8549–8560, China University of Mining and Technology. Her
Sep. 2018. research interests include in the areas of design-
[18] J. Wang, J. Yang, J. Xiong, H. Sari, and G. Gui, ‘‘SHAFA: Sparse hybrid
ing system applied for wireless communication
adaptive filtering algorithm to estimate channels in various SNR environ-
ments,’’ IET Commun., vol. 12, no. 16, pp. 1963–1967, Nov. 2018. modem and various systems required advanced
[19] G. Gui, H. Huang, Y. Song, and H. Sari, ‘‘Deep learning for an effec- digital signal processing.
tive nonorthogonal multiple access scheme,’’ IEEE Trans. Veh. Technol.,
vol. 67, no. 9, pp. 8440–8450, Sep. 2018.
[20] S. Schiessl, H. Al-Zubaidy, M. Skoglund, and J. Gross, ‘‘Delay per-
formance of wireless communications with imperfect CSI and finite- ZHENZHEN WU received the B.S. degree in
length coding,’’ IEEE Trans. Commun., vol. 66, no. 12, pp. 6527–6541, electronic and communication engineering from
Dec. 2018. Yantai University, in 2016. She is currently pur-
[21] T. Mitchell, B. Buchanan, G. Dejong, T. Dietterich, P. Rosenbloom, and suing the M.S. degree with the School of Infor-
A. Waibel, Machine Learning. Beijing, China: China Machine Press, 2003. mation and Communication Engineering, China
[22] X. Lin, R. Liu, W. Hu, Y. Li, X. Zhou, and X. He, ‘‘A deep convolu-
University of Mining and Technology. Her cur-
tional network demodulator for mixed signals with different modulation
types,’’ in Proc. IEEE Int. Conf. Dependable, Autonomic Secure Comput., rent research interests include wireless communi-
Nov. 2017, pp. 893–896. cations and machine learning.
[23] M. Fan and L. Wu, ‘‘Demodulator based on deep belief networks in
communication system,’’ in Proc. Int. Conf. Commun., Control, Comput.
Electron. Eng., Jan. 2017, pp. 1–5.
[24] L. Fang and L. Wu, ‘‘Deep learning detection method for signal demod-
ulation in short range multipath channel,’’ in Proc. IEEE 2nd Int. Conf. SHUAI MA received the B.S. and Ph.D. degrees
Opto-Electron. Inf. Process., Jul. 2017, pp. 16–20.
in communication and information systems from
[25] A. S. Mohammad, N. Reddy, F. James, and C. Beard, ‘‘Demodulation
Xidian University, Xian, China, in 2009 and 2016,
of faded wireless signals using deep convolutional neural networks,’’ in
Proc. IEEE 8th Annu. Comput. Commun. Workshop Conf., Jan. 2018, respectively. From 2014 to 2015, he was a Visit-
pp. 969–975. ing Scholar with the Department of Electrical and
[26] Y. Freund and R. E. Schapire, ‘‘Experiments with a new boosting algo- Computer Engineering, Texas A&M University,
rithm,’’ in Proc. 13th Int. Conf. Int. Conf. Mach. Learn., Jul. 1996, College Station, TX, USA. Since 2019, he has
pp. 148–156. been an Associate Professor with the School of
[27] M. Kulin, T. Kazaz, I. Moerman, and E. De Poorter, ‘‘End-to-end learn- Information and Control Engineering, China Uni-
ing from spectrum data: A deep learning approach for wireless signal versity of Mining and Technology. He is also
identification in spectrum monitoring applications,’’ IEEE Access, vol. 6, working at the State Key Laboratory of Integrated Services Networks,
pp. 18484–18501, 2018. Xidian University. His research interests include visible light communica-
[28] B. Karanov et al., ‘‘End-to-end deep learning of optical fiber communica- tion, machine learning, wireless communications, and network information
tions,’’ J. Lightw. Technol., vol. 36, no. 20, pp. 4843–4855, Oct. 15, 2018. theory.
[29] Y. Lecun, U. Müller, J. Ben, E. Cosatto, and B. Flepp, ‘‘Off-road obstacle
avoidance through end-to-end learning,’’ in Proc. Adv. Neural Inf. Process.
Syst., 2006, pp. 739–746.
[30] J. Sola and J. Sevilla, ‘‘Importance of input data normalization for the
application of neural networks to complex industrial problems,’’ IEEE SONGTAO LU (S’10–M’18) received the Ph.D.
Trans. Nucl. Sci., vol. 44, no. 3, pp. 1464–1468, Jun. 1997. degree in electrical and computer engineering
[31] G. E. Hinton, ‘‘Training products of experts by minimizing contrastive from Iowa State University, in 2018. He is cur-
divergence,’’ Neural Comput., vol. 14, no. 8, pp. 1771–1800, 2002. rently a Postdoctoral Associate with the Depart-
[32] Z. Yang and X. Pang, ‘‘Research and implementation of text classification
ment of Electrical and Computer Engineering,
model based on combination of DAE and DBN,’’ in Proc. 10th Int. Symp.
University of Minnesota Twin Cities, Minneapolis.
Comput. Intell. Design, vol. 2, Dec. 2017, pp. 190–193.
[33] R. Sarikaya, G. E. Hinton, and A. Deoras, ‘‘Application of deep belief His primary research interests include wire-
networks for natural language understanding,’’ IEEE/ACM Trans. Audio, less communications, optimization, and machine
Speech, Language Process., vol. 22, no. 4, pp. 778–784, Apr. 2014. learning. He was a recipient of the Graduate and
[34] F. Zhang and Q. Zhou, ‘‘Ensemble detection model for profile injection Professional Student Senate Research Award from
attacks in collaborative recommender systems based on BP neural net- Iowa State University, in 2015, the Research Excellence Award from the
work,’’ IET Inf. Secur., vol. 9, no. 1, pp. 24–31, Jan. 2015. Graduate College of Iowa State University, in 2017, and the Student Travel
[35] X. H. Wang, ‘‘Multi-class SVMs based on probability voting strat- Award from the 20th International Conference on Artificial Intelligence and
egy and its application,’’ Comput. Eng., vol. 35, no. 2, pp. 180–183, Statistics, in 2017.
Feb. 2009.
[36] J. C. Platt, N. Cristianini, and J. Shawe-Taylor, ‘‘Large margin DAGs
for multiclass classification,’’ in Proc. Adv. Neural Inf. Process. Syst., HAN ZHANG received the B.E. degree from the
Jan. 2000, pp. 547–553. Beijing University of Posts & Telecommunica-
[37] M. Yu, A. Rhuma, S. M. Naqvi, and J. Chambers, ‘‘An improved directed
tions, Beijing, China, in 2010. He is currently
acyclic graph support vector machine,’’ J. Meas. Sci. Instrum., vol. 2, no. 4,
pursuing the Ph.D. degree with the University of
pp. 367–370, Dec. 2011.
[38] Y. Wang, L. Wang, J. Qiu, P. Shen, and X. Du, ‘‘Research on pedestrian California at Davis, USA. His current research
detection method based on multilayer RBM network and SVM,’’ J. China interests include wireless networks and machine
Railway Soc., vol. 40, no. 3, pp. 95–100, Mar. 2018. learning.
[39] R. A. Mcdonald, D. J. Hand, and I. A. Eckley, An Empirical Comparison of
Three Boosting Algorithms on Real Data Sets with Artificial Class Noise.
Berlin, Germany: Springer, 2003.

30800 VOLUME 7, 2019

H. Wang et al.: DL for Signal Demodulation in Physical Layer Wireless Communications

GUORU DING (S’10–M’14–SM’16) received SHIYIN LI received the Ph.D. degree in infor-
the B.S. degree (Hons.) in electrical engineering mation and communication engineering from the
from Xidian University, Xian, China, in 2008, China University of Mining and Technology,
and the Ph.D. degree (Hons.) in communica- Xuzhou, China, in 2010, where he has been a Pro-
tions and information systems from the College fessor with the School of Information and Control
of Communications Engineering, Nanjing, China, Engineering, since 2010, and is also the Head of
in 2014. Since 2014, he has been an Assistant the Department of Information Engineering. His
Professor with the College of Communications research interests include wireless communication
and a Research Fellow with the National High and network congestion control.
Frequency Communications Research Center of
China. Since 2015, he has been a Postdoctoral Research Associate with the
National Mobile Communications Research Laboratory, Southeast Univer-
sity, Nanjing. He is currently with the College of Communications Engineer-
ing, Army Engineering University of PLA. His research interests include
cognitive radio networks, massive MIMO, machine learning, and big data
analytics over wireless networks.

VOLUME 7, 2019 30801