Content uploaded by Xun Zhang
Author content
All content in this area was uploaded by Xun Zhang on Jul 02, 2019
Content may be subject to copyright.
978-1-7281-2150-5/19/$31.00 ©2019 IEEE
PAPR reduction based on deep autoencoder for
VLC DCO-OFDM system
Lina Shi
1
, Xun Zhang
1
*, Wenxiao Wang
1
, Yue Zhang
3
, Zhan Wang
1
, Andrei Vladimirescu
1
, Yue Zhang
2
, Jintao
Wang
3
1
Institut supérieur d'électronique de Paris, Paris, France
2
Department of Engineering, University of Leicester, UK
3
Department of Electronic Engineering, Tsinghua University, China
*Contact: lina.shi@isep.fr
Abstract—DC-biased optical orthogonal frequency division
multiplexing (DCO-OFDM) is widely used in visible light
communication (VLC) systems to reduce the effects of inter-
symbol interference (ISI) and to achieve high speed data
transmission. However, DCO-OFDM VLC systems suffer from
high peak-to-average power ratio (PAPR) issue which causes
serious degradation of system performance. In this paper, a
novel deep autoencoder-based PAPR reduction scheme for VLC
DCO-OFDM is proposed. The autoencoder could adaptively
adjust the constellation symbols in order to minimize the PAPR
while maintaining an acceptable bit error rate (BER) for VLC
transmission. The simulation results show that PAPR is reduced
to 6.5dB by comparing with conventional methods.
Keywords—VLC, DCO-OFDM, autoencoder, deep learning,
PAPR
I. I
NTRODUCTION
The explosive growth of internet connected devices and
various intelligent services lead to an increasing demand for
wireless capacity. As the next generation communication
technology, 5G is designed to meet this demand, providing
ultra-broadband, ultra-low latency, ultra-high reliability, and
massive connectivity for diverse scenarios, especially in the
indoor environment [1]. Hence a new broadband indoor
access solution in buildings is in urgent need. Among the
short-range wireless communication technologies, VLC is
considered a promising technology for future 5G indoor
communications to enable gigabit-per-second data
transmission and to meet the ever-increasing capacity
requirements [2] [3]. VLC uses a largely so-far neglected
spectrum which is between 400 and 800 THz (780–375 nm)
for data transmission as well as indoor illumination utilizing
light-emitting diodes (LEDs). Moreover, VLC is considered
an economical and practical indoor wireless access
technology. It is also more security for both privacy and health
[4-6] comparing to other radio frequency-based
communication technologies.
The bottleneck for VLC high data rate transmission is the
limited modulation bandwidth of commercial LED, typically
a few megahertz [7]. In the literature [8,9], most of them focus
on the use of higher spectral efficiency modulation method to
address this issue, such as OFDM. However, to adapt the
intensity modulation and direct detection (IM/DD)
requirement in VLC system, many variants of OFDM
modulation need to be adopted to generate the real and non-
negative OFDM signals. Among these variants, DCO-OFDM
has been widely used thanks to the highest spectral efficiency
with simple implementation [10]. But, DCO-OFDM signals
also suffers from high PAPR, which is an intrinsic drawback
of OFDM technology. High PAPR of DCO-OFDM signal
shows a considerable sensitivity to the LED nonlinear device.
It will thence inevitably lead to a double-sided clipping effect
that imposes BER performance degradation to the system.
Current PAPR reduction approaches have been classified
into three categories: clipping [11, 12], signal scrambling [13,
14], and coding schemes [15]. For example, amplitude
clipping is one of the simplest schemes to reduce PAPR thanks
to its low algorithm complexity, however the distorted original
signal leads to a higher BER. In addition, partial transmission
sequence (PTS) and selective mapping (SLM) are more
favorable PAPR reduction scheme of signal scrambling
techniques, but additional information (side information) need
to provide in order to reconstruct the tranmitted signal
properly. The transmission performance is greatly influenceed
by the quality of side information.
Based on the previous literature review, in this paper, we
consider a deep autoencoder-based PAPR reduction scheme
for DCO-OFDM system to address high PAPR issue while
maintaining a acceptable BER. The deep learning as an
emerging technique to solve the problem of high PAPR in
OFDM system becomes more and more attractive in recent
year because of its strong auto-adaptive ability. A successful
case by using an artificial neural network (ANN) to reduce the
complexity for PAPR reduction has been realized by Sohn
[16] and Sohn and Kim [17]. In [18], a PAPR reduction
solution based on deep learning is proposed in RF-based
OFDM system. But to the best of our knowledge, few work
considers applying deep learning method to DCO-OFDM
visible light system. Thus, in this paper, we propose a high
PAPR reducing scheme based on the autoencoder architecture
of deep learning for VLC DCO-OFDM system.
The main contributions of this paper are summarized in
two aspects as follows:
• We proposed a novel PAPR reduction scheme in VLC
DCO-OFDM system by combining a deep
autoencoder architecture. The constellation symbols
are fed into the deep autoencoder and adaptively
trained to minimize PAPR while retaining the BER.
• We comparatively evaluate the performance of the
proposed scheme with the conventional PAPR
reduction schemes in terms of Complementary
cumulative distribution function (CCDF). The
simulation results confirm that the proposed scheme
can reduce PAPR by 6.59 dB.
The remainder of this paper is organized as follows.
Section II gives an overview of autoencoder and presents the
detailed information of the proposed scheme. Section III
contains the simulation results and a discussion of the
proposed scheme compared to the conventional methods.
Finally, conclusions are reported in Section IV.
II. A
UTOENCODER AND SYSTEM MODEL
A. Autoencoder
Autoencoder which is frequently used for denoising
corrupted data is suitable to dealing with the non-linear
distortions such as high PAPR. Therefore, in this paper a deep
autoencoder network is trained to mitigate the PAPR issue in
VLC DCO-OFDM system.
Autoencoder has multiple hidden layers of representation
[20]. The most attractive advantage of autoencoder is that the
features of each hidden layer are not designed manually, they
are learned from the input data automatically. Therefore,
autoencoder gets lots of attention recent years and it is applied
in diverse wireless communication fields such as channel
coding, channel compensation and modulation recognition
[21].
The general autoencoder architecture consists two
components: encoder and decoder, where x, f(x), g(x) and is
the input, encoder, decoder and reconstruction of the original
input x respectively. Both the encoder and the decoder consists
of the identical hidden layers, each hidden layer includes
Dense layer, Activation function and Dropout. One example
is shown in Fig.1 marked in red box.
Fig. 1. Proposed deep autoencoder architecture (general autoencoder:
marked in the red box; noise layer: optical wireless channel)
In our scheme, we initially set L
f
= L
g
=3, the output of i th
dense layer of the encoder can be defined as ℎ
=
+
,
where
is a weight and
is a bias of i th layer of the
encoder. Then the rectifier linear unit (ReLU) is used as our
activation function of encoder. Finally, dropout is used for
addressing the overfitting problem [23] for our proposed
scheme.
Compared with the traditional autoencoder, a
normalization layer is used to integrate output power
constraint at the transmitter in a communication system [24].
The mathematical expression of normalization layer:
ℎ
=
[
]
+ is given in [18], where and
are the scaling and shift factor respectively and the values
of and will be decided by training. Additionally, in order
to prevent the division by zero, υ=0.001.
Based on the previous theoretical analysis of each hidden
layer, the output of encoder can be mathematically expressed
as follows:
()=
(
…
+
…+
)
(1)
where (∙) is an activation function (ReLU) and
is a
weight and
is a bias of
th layer of the encoder,
respectively.
Similarly, the output of decoder can be written as equation
(2):
=()=
(
…
+
…
+
)
(2)
where (∙) is an activation function (sigmoid) and
is a
weight and
is a bias of
th layer of the decoder,
respectively.
Unlike the traditional autoencoder, we add a noise layer to
simulate an additive white Gaussian noise (AWGN) channel
in a communication system [25].
We train the autoencoder network for minimizing the
reconstruction error, this factor is defined as loss function
ℒ(,). ℒ evaluates the differences between our original
input x and the consequent reconstruction , where is the
output of the decoder g(f(x)). The detailed information of loss
function will be described in subsection B.
B. System model
Fig.2 shows an overview of the proposed system model.
Compared with the tranditional DCO-OFDM system, an
autoencoder is applied into the original system architecture to
mitigate the PAPR issue.
Fig. 2. Block diagram of proposed PAPR reduction scheme based on deep
autuencoder
We assume that in our DCO-OFDM system 2N
subcarriers are used to generate OFDM symbols. It can be
seen from Fig.2 that after serial to parallel operation and
constellation mapping operation, the parallel mapping data
=[
,
,…
]
is used as the input of encoder.
Then the encoded symbols (
(∈[,])
) are constrained
to have Hermitian symmetric in order to obtain a real signal
after the 2N-IFFT operation. The definition of Hermitian
symmetric is shown in the expression below:
=[(
),(
),…(
),(
)
∗
,…(
)
∗
,(
)
∗
]
(3)
where S represents the new vector after Hermitian symmetric,
(*) denotes the complex conjugate of a vector. After DC bias
operations, the positive real DCO-OFDM symbols in time
domain =[
,
,…
]
are generated for VLC
transmission.
The PAPR definition is described in equation (4).
{}=max
(|
|) (4)
where the denominator is the average power of DCO-OFDM
symbols.
Then these symbols are transmitted through a wireless
channel before arriving at the receiver. Finally, the received
signal passes through reverse processes and is decoded using
the decoder of the autoencoder ((
)). The reconstructed
symbol at the receiver((
)) can be written as follows:
=(ℎ∘∘∘∘ℎ
∘(
)) (5)
Where ∈[1,], ℎ is Hermitian symmetric operation,
ℎ
is inverse Hermitian symmetric operation, is a VLC
wireless channel model.
The training process is simply described for minimizing
the joint loss function ℒ defined in equation (6). ℒ consists
two loss functions ℒ
and ℒ
, where is weight parameter to
decide which loss function is dominant. We initially set to
0.001.
ℒ(
,
)=ℒ
(
,
)+ℒ
(
) (6)
ℒ
(
,
)=(
−(ℎ∘∘∘
∘ℎ
∘(
)+);
)
(7)
ℒ
(
)={(
);
} (8)
The function of ℒ
is to enable the BER do not deteriorate
and reconstruct the transmitted signals form the received data.
Here, is the noise. ={,}, W is a weight matrix and b
is a bias vector of autoencoder.
The stochastic gradient descent (SGD) method is used to
update W and b. The function of ℒ
is to minimize PAPR.
Through the training of
and
, the joint loss function is
minimized.
III. S
IMULATION RESULTS AND DISCUSSION
In this section, the preliminary simulation results are
presented. The parameters of the autoencoder network are
summarized as follows.
Fig. 3. Parameters of the proposed network
We consider a DCO-OFDM system with 128 subcarriers
and QPSK modulation. A VLC wireless channel we used is an
AWGN channel. We used 640000 independant random bits
for tranining, 128000 bits for validation and 128000 bits for
testing, respectively. In perticualr, 100000 DCO-OFDM
symbols are used to generate all simulation results. The CCDF
of the PAPR are used to represent system performance.
We evaluated the proposed scheme by contrast with two
well known conventional PAPR reduction schemes listed as
follows:
1) Classical Selected Mapping (SLM) method. The
detailed theory principle can be found at [27].
2) Classical amplitude clipping with a clipping radio γ.
According to the clipping equation given in [28], the DCO-
OFDM signals will be direcly upper or lower clipped when
the generated DCO-OFDM signals exceed the given upper or
lower level.
Before PAPR comparison, an appropriate signal-noise
ratio (SNR) level needs to be selected for traning the proposed
scheme. As shown in Fig.4, the BER of the proposed scheme
at different SNR levels is presented. We can see that training
the proposed scheme at SNR = 15 dB can provide lower BER
than other cases over the whole range of SNR. Therefore, we
set up SNR = 15 dB to obtain PAPR results for performance
evaluation.
Fig. 4. BER of the proposed scheme at different SNR levels
The PAPR comparison results are inllustrated in Fig.5.
Firstly, our scheme can reach a 6.5 dB of PAPR compared to
traditional DCO-OFDM wihout PAPR reduction method with
13.83 dB. Secondly, it can be shown that the PAPR of the
proposed reduction scheme is also much lower than others.
For example, to reach a CCDF of 10
-4
, our proposed scheme
reduced 3.98 dB of PAPR by comparing with SLM scheme
(U = 4), 6.59 dB of PAPR by comparing with clipping scheme
(γ = 10 %), respectively.
Fig. 5. CCDF of PAPR comparison for proposed and conventional schemes
IV. C
ONCLUSION
In this paper, a novel deep autoencoder-based PAPR
reduction scheme for DCO-OFDM VLC is proposed. The
autoencoder could adjust adaptively the constellation symbols
in order to minimize the PAPR as well as maintain a lower
BER for VLC system. Based on the PAPR comparison results,
the proposed scheme outperforms conventional DCO-OFDM
VLC schemes maximum 6.59 dB.
A
CKNOWLEDGMENT
The authors gratefully acknowledge the financial support
of the National Key R&D Program of China (Grant
No.2017YFE011230) and the EU Horizon 2020 program
towards the Internet of Radio-Light project H2020-ICT
761992.
R
EFERENCES
[1] Alliance, N. G. M. N. (2015). 5G white paper. Next generation mobile
networks, white paper, 1-125.
[2] Cosmas, J., Zhang, Y., & Zhang, X. (2017, May). Internet of Radio-
Light: 5G Broadband in Buildings. In European Wireless 2017; 23th
European Wireless Conference; Proceedings of (pp. 1-6). VDE.
[3] Jawad, N., Salih, M., Ali, K., Meunier, B., Zhang, Y., Zhang, X., ... &
Shi, L. (2019). Smart Television Services Using NFV/SDN Network
Management. IEEE Transactions on Broadcasting.
[4] Meunier, B., & Cosmas, J. (2018, June). 5G Internet of Radio Light
Virtual Reality System. In 2018 IEEE International Symposium on
Broadband Multimedia Systems and Broadcasting (BMSB) (pp. 1-5).
IEEE.
[5] Shi, L., Li, W., Zhang, X., Zhang, Y., Chen, G., & Vladimirescu, A.
(2018, December). Experimental 5G New Radio integration with VLC.
In 2018 25th IEEE International Conference on Electronics, Circuits
and Systems (ICECS) (pp. 61-64). IEEE.
[6] Feng, L., Hu, R. Q., Wang, J., Xu, P., & Qian, Y. (2016). Applying
VLC in 5G networks: Architectures and key technologies. IEEE
Network, 30(6), 77-83.
[7] L. SHI; X. ZHANG; A VLADIMIRESCU. OFDMA-TDM position
scheme for VLC indoor positioning system, GDR SoC-SiP Paris
France, 2018
[8] Huang, X., Shi, J., Li, J., Wang, Y., Wang, Y., & Chi, N. (2015,
March). 750Mbit/s visible light communications employing 64QAM-
OFDM based on amplitude equalization circuit. In Optical Fiber
Communications Conference and Exhibition (OFC), 2015 (pp. 1-3).
IEEE.
[9] Ghosh, A., Thomas, T. A., Cudak, M. C., Ratasuk, R., Moorut, P.,
Vook, F. W., ... & Nie, S. (2014). Millimeter-wave enhanced local area
systems: A high-data-rate approach for future wireless networks. IEEE
Journal on Selected Areas in Communications, 32(6), 1152-1163.
[10] Gomes, R., Reis, J., Al-Daher, Z., Hammoudeh, A., & Caldeirinha, R.
F. (2018). 5G: performance and evaluation of FS-FBMC against
OFDM for high data rate applications at 60 GHz. IET Signal
Processing.
[11] He, Y., Jiang, M., & Zhao, C. (2018). A Neural Network Aided
Approach for LDPC Coded DCO-OFDM with Clipping
Distortion. arXiv preprint arXiv:1809.01022.
[12] Deng, S. K., & Lin, M. C. (2005, May). OFDM PAPR reduction using
clipping with distortion control. In Communications, 2005. ICC 2005.
2005 IEEE International Conference on (Vol. 4, pp. 2563-2567).
IEEE.
[13] Wang, Y. C., & Luo, Z. Q. (2011). Optimized iterative clipping and
filtering for PAPR reduction of OFDM signals. IEEE Transactions on
Communications, 59(1), 33-37.
[14] Jiang, T., & Zhu, G. (2004). Nonlinear companding transform for
reducing peak-to-average power ratio of OFDM signals. IEEE
Transactions on Broadcasting, 50(3), 342-346.
[15] Jiang, Y. (2010). New companding transform for PAPR reduction in
OFDM. IEEE Communications Letters, 14(4).
[16] Wu, F., Wang, J., Wang, J., & Song, J. (2018, June). A Precoding
Method for Both Out-of-Band Emission and PAPR Reduction in
OFDM Systems. In 2018 IEEE International Symposium on
Broadband Multimedia Systems and Broadcasting (BMSB) (pp. 1-5).
IEEE.
[17] Sohn, I. (2014). A low complexity PAPR reduction scheme for OFDM
systems via neural networks. IEEE Communications Letters, 18(2),
225-228.
[18] Sohn, I., & Kim, S. C. (2015). Neural network based simplified
clipping and filtering technique for PAPR reduction of OFDM
signals. IEEE Communications Letters, 19(8), 1438-1441.
[19] Kim, M., Lee, W., & Cho, D. H. (2018). A Novel PAPR Reduction
Scheme for OFDM System Based on Deep Learning. IEEE
Communications Letters, 22(3), 510-513.
[20] Y. Bengio et al., “Generalized denoising auto-encoders as generative
models,” in Proc. NIPS, Lake Tahoe, NV, USA, Dec. 2013, pp. 899–
907.
[21] G.E. Hinton, R.R. Salakhutdinov, Reducing the dimensionality of data
with neural networks, Science 313 (5786) (2006) 504–507.
[22] Felix, A.; Cammerer, S.; Dörner, S.; Hoydis, J.; ten Brink, S. OFDM-
Autoencoder for End-to-End Learning of Communications Systems.
arXiv 2018, arXiv:1803.05815.
[23] Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating
deep network training by reducing internal covariate shift. arXiv
preprint arXiv:1502.03167.
[24] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., &
Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural
networks from overfitting. The Journal of Machine Learning Research,
15(1), 1929-1958.
[25] Wang, T., Wen, C. K., Wang, H., Gao, F., Jiang, T., & Jin, S. (2017).
Deep learning for wireless physical layer: Opportunities and
challenges. China Communications, 14(11), 92-111.
[26] O’Shea, T., & Hoydis, J. (2017). An introduction to deep learning for
the physical layer. IEEE Transactions on Cognitive Communications
and Networking, 3(4), 563-575.
[27] Hu, W.-W.; Lee, D.-H. PAPR Reduction for Visible Light
Communication Systems Without Side Information. IEEE Photonics J.
2017, 9, 1–11.
[28] Elgala, H.; Mesleh, R.; Haas, H. An LED model for intensity-
modulated optical communication systems. IEEE Photonics Technol.
Lett. 2010, 22, 835–837.
