International Journal of Engineering & Technology, 7 (4) (2018) 2638-2645

International Journal of Engineering & Technology

Website: www.sciencepubco.com/index.php/IJET
doi: 10.14419/ijet.v7i4.15994
Research paper

Expectation-maximization-based channel estimation algorithm

for OFDM visible light communication systems
Yaseein Soubhi Hussein 1 *, Mohamad Yusoff Alias 2, Ayad A. Abdulkafi 3, Nazaruddin Omar 4,
Mohd Kamarulzamin Bin Salleh 4
1
Faculty of Computing, Engineering & Technology (FCET), Asia Pacific University of Technology & Innovation, 57000 Kuala Lumpur,
Kuala Lumpur, Malaysia
2
Faculty of Engineering, Multimedia University, 63100 Cyberjaya, Selangor, Malaysia
3
College of Applied Engineering, Tikrit University, Iraq
4
TM Research & Development Sdn Bhd, Innovation Centre, Lingkaran Teknokrat Timur, 63000 Cyberjaya Selangor, Malaysia
*Corresponding author E-mail: [email protected]

Abstract

The tremendous growth of indoor communication requires increased capacity and appropriate quality of services. Visible light communi-
ca-tion (VLC) is a green technology that shows great promise in terms of its ability to meet the demand for communication services.
Orthogo-nal frequency division multiplexing (OFDM) enables VLC to provide a higher data rate and to combat inter-symbol interference.
However, an accurate and efficient channel estimation method is needed for coherent demodulation at the receiver end of an OFDM sys-
tem. In this paper, a new algorithm for OFDM-based VLC systems is proposed. The algorithm is based on expectation maximization and
is called the expectation maximization for visible light communication (EM-VLC) algorithm. The algorithm is implemented to find the
maximum-likelihood (ML) estimation of the channel impulse response and to find unknown parameters. In addition, a low-rank mini-
mum mean square error (lr-MMSE) estimator algorithm is developed and its performance is compared with least squares (LS) and mini-
mum mean square error (MMSE) estimators. The proposed EM-VLC algorithm improves the performance of OFDM VLC systems by
significantly reducing the bit error rate (BER) and consequently increasing system throughput. The simulation results demonstrate that
the EM-VLC algorithm outper-forms the three channel estimation algorithms, LS, MMSE and lr-MMSE.

Keywords: Bit Error Rate; Channel Estimation; Expectation Maximization; Optical OFDM Visible Light Communication.

as the frequency correlation and cyclic prefix (CP) (i.e., they do

1. Introduction not use a training sequence); and 3) semi-blind channel methods,
which are based on a combination of blind channel and training
Visible light communication (VLC) is a viable technology that sequence methods and which use pilot arrangements and the struc-
possesses many attractive qualities, such as an unlicensed spec- tural characteristics of the signals [5].
trum, worldwide availability of unused bandwidth, non- In this paper, a new channel estimation method is developed in an
interference with radio frequency (RF) bands, very high data rates attempt to reduce the bit error rate (BER) and the analytical of
and secure communication [1], [2]. Visible light communication complexity in direct current optical OFDM (DCO-OFDM). The
uses low-cost light-emitting diodes (LEDs) as transmitters, or a remainder of the paper is organized as follows: in section 2, relat-
converter, to translate the modulated electrical signal into an opti- ed works on channel estimation in OFDM-based VLC are intro-
cal signal. Then, a reverse process allows the photodiodes (PDs) to duced. In section 3, the OFDM VLC system and channel estima-
detect the signal and demodulate it via a digital signal processing tion are described. In section 4, the proposed expectation maximi-
technique [3]. Orthogonal frequency division multiplexing zation for visible light communication (EM-VLC) algorithm for
(OFDM) in VLC holds promise as a means of providing a higher OFDM VLC systems is presented and analysed. In section 5, the
data rate, combating inter-symbol interference (ISI) and reducing simulation results are presented and discussed. In section 6, a con-
multipath effects. However, the nonlinearities of the OFDM clusion is provided.
transmitted signal and the wireless optical propagation effects
increase the complexity involved in retrieving the original signal
at the receiver. Hence an accurate and efficient channel estimation
2. Related work
technique is needed. Channel estimation refers to the estimation of
the channel effect on the transmitted signal and the compensation A few approaches, such as those in [6 - 8], have been proposed for
process that is used to recover the transmitted signal at the receiv- channel estimation in OFDM-based VLC systems. The primary
er. Channel estimation techniques can be categorized into 1) train- objective of these approaches, as well as that of a more recent
ing sequence methods (also known as block-type pilot arrange- work [9], is to reduce the BER and thus obtain better system per-
ments), which introduce pilot tones into the subcarriers of the formance. Several other channel estimation methods have also
OFDM symbols [4], 2) blind channel methods, which exploit the been proposed, such as those in [10 - 14]. However, these are for
statistical or structural properties of communication signals, such optical wireless communication (OWC) systems rather than
Copyright © 2018 Yaseein Soubhi Hussein et. al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
International Journal of Engineering & Technology 2639

OFDM. The channel estimation methods in OWC are training- In this section, the OFDM-based VLC system is presented togeth-
sequence-based and can be performed by utilizing least squares er with its channel model. The OFDM VLC system is illustrated
(LS) or minimum mean square error (MMSE) algorithms [15] as in Fig. 1. The typical OFDM signal is bipolar and complex, so it is
well as other modified MMSE algorithms, such as low-rank min- impossible to transmit such a signal via an optical wireless chan-
imum mean square error (lr-MMSE) algorithms [16]. Furthermore, nel using intensity modulated/direct detection (IM/DD). In VLC,
some other methods that have been proposed for purposes other the signal must be unipolar and real. Thus, Hermitian symmetry in
than channel estimation could also be used for channel estimation. the inverse fast Fourier transform (IFFT) and a DC-bias current
These methods, which were proposed in [17] and [18], aimed to are needed to make the signal real and positive, respectively. This
overcome the nonlinearities in OFDM VLC systems, including the is referred to as the DC-biased optical OFDM (DCO-OFDM)
peak to average power ratio (PAPR). scheme [26].
In [6], a method that combines discrete Fourier transformation In the VLC system, a LED serves as a transmitter and a PD acts as
(DFT) with post-processing and LS, called DFT-LS, was proposed a receiver. The number of multipath components is large and the
for DCO-OFDM with the aim of improving system performance propagation channel varies temporally due to the movement of
by eliminating the noise beyond the maximum delay of the chan- users and other natural propagation processes of light, such as
nel. The method uses a LS approach for channel estimation with a reflection and diffraction. The noise due to background light can
comb-type pilot arrangement and linear interpolation. While DFT- be exhibited as additive white Gaussian noise (AWGN) within the
LS performs better than the ordinary LS estimator in terms of model. There is no interference with other RF communication
BER at different orders of constellation, it is less accurate than an systems [27].
MMSE estimator is. [19].
In [7], an adaptive channel estimation method based on a LS dis-
crete Fourier transform (LS-DFT) and orthogonal matching pur-
suit (OMP) was proposed as a means to enhance the reliability of
asymmetrically clipped DC-biased optical (ACO-OFDM) VLC
systems. The first part of the algorithm is similar to that proposed
in [6] in which a post-process DFT-LS/LS-DFT is used to obtain
the channel response in the frequency domain after eliminating the
noise effect [16]. The second part of the algorithm is an iterative
greedy OMP algorithm that is used to obtain the channel response
in the time domain. The results show that LS-DFT achieves a
better BER than the DFT-LS presented in [6]. However, an addi-
tional computational cost is incurred. The results of an investiga-
tion into the differences between DCO-OFDM and ACO-OFDM
in VLC systems can be found in [20].
In [8], an adaptive and specific grid-type pilot arrangement was Fig. 1: Block Diagram of DCO-OFDM VLC System.
proposed as a way to evaluate the channel response in DCO-
OFDM VLC systems. This method is based on sending known At the transmitter end of the system, the mapping to an M-QAM
subcarriers from the transmitter in one or more OFDM symbols to constellation has to be done before the Hermitian symmetry, X H,
the receiver. At the receiver, the channel response can be found by is applied. The IFFT operation results in 2N points in the frequen-
estimating the subcarrier changes. The use of more pilots can im- cy domain. Then, the modulated OFDM symbol vector, x, in the
prove channel estimation, but such an approach sacrifices trans- time domain, xm, is converted from parallel to serial (P/S), adding
mission capacity. However, in this method, fewer pilot insertions the CP and digital-to-analogue (D/A) conversion, as shown in
are needed, so it has better transmission capacity than methods Equation (1):
that use training symbols (block-type pilots). Moreover, this
2 N 1 2hm
method can adapt the pilot tones according to the BER perfor-
X
1 j
xm  H ,h e N for 0  m  2 N  1
mance, at the cost, the signal-to-noise ratio (SNR) cannot maintain N h 0 (1)
the transmission. In [9], a more recent work, an adaptive statistical
channel estimation of Bayesian MMSE was proposed in which the Where h is the hth-subcarrier symbol of XH.
variable statistic window is exploited, but this approach has high The bipolar and real value of OFDM, xt, is obtained in the time
computational complexity. domain. Then, the DC bias is applied to obtain a new unipolar or
In this paper, a new algorithm for OFDM-based VLC systems is positive signal as an input to the LED. Finally, the signal derives
proposed. It is based on EM and is called the expectation maximi- the LED that can convert the electrical signal into an optical inten-
zation for visible light communication (EM-VLC) algorithm. In sity. With the PD acting as a receiver, the signal can be directly
addition, a lr-MMSE algorithm is developed for OFDM VLC detected. The reverse process is carried out in addition to channel
systems. The proposed EM-VLC algorithm is an iterative method estimation to retrieve the original transmitted signal. The output
for computing the maximum-likelihood (ML) channel estimation received signal of the channel can be expressed as:
and finding unknown parameters. It consists of two steps: the first
step involves expectation (E) for unknown parameters and the y(n)  x(n)  h(n)  w(n)
second step involves maximization (M) to maximize the expecta- (2)
tion of the log-likelihood function. The EM algorithm has been
widely used in RF for channel estimation, for instance in [21 - 25]. Where h (n) is the channel impulse response. It is assumed that the
However, EM for indoor VLC is different from that for conven- channel is linear and 2N is the received time-domain signal length.
tional RF due to several limitations, and it is these limitations that An AWGN channel is represented by w (n).
motivated the present study. The simulation results of the pro- Then, the output in the frequency domain is obtained, in which the
posed algorithm demonstrate its ability to significantly reduce the FFT of h (n) and w (n) in Equation (2) becomes H (k) and W (k)
BER. Moreover, it is a powerful channel estimation solution for in Equation (3), respectively, as follows:
OFDM VLC systems.
Y (k )  X (k ) H (k )  W (k ) k  0,....,N  1 (3)
3. The system model
3.2. Channel estimation
3.1. The OFDM VLC system
2640 International Journal of Engineering & Technology

Channel estimation is the key feature of the OFDM VLC system. Where E [•] is the expectation operator. First, it is expected that to
The received signal is influenced by the channel characteristics determine the estimator value L, to obtain the MMSE channel
and an accurate channel estimation is needed to recover the origi- estimation. Therefore, the value of L must be satisfied by Equation
nal transmitted bits [28]. Thus, channel estimation is an essential (9) below:
part of OFDM systems. A good channel estimation algorithm

should have both low complexity and high accuracy. Generally, L  arg min ( L)
two common techniques can be employed for channel estimation. L (9)
The first involves using training symbols when pilots (known In a similar manner to LS, MMSE uses differentiation and sets to
information) are sent when no data is being sent. The second in- zero with respect to each channel coefficient L, as below:
volves using pilot aided where the pilot inserted together with the
data. In the frequency domain, the pilot spacing depends on the 
L

 ( H  LY )( H  LY ) H  0 
frequency coherence, which is related to the delay spread, so it (10)
must be determined carefully [29].
In the VLC system, the shadowing effect can be alleviated by By solving Equation (10), the value of operator L can be obtained
using numerous LEDs. However, increasing the number of LEDs as
aggravates the multipath effect, which leads to ISI. Inter-symbol
interference is a significant issue in OFDM VLC systems because  RH X H
L
it can reduce the system’s data rate as well as its BER [30]. Hence XRH X H  RW
(11)
channel estimation is a critical issue and achieving coherent de-
modulation is a challenging task. The two common channel esti- Where RW = E [WWH] is the covariance matrix of noise, RH = E
mation methods for OFDM VLC systems – LS and MMSE – are [HHH] is the covariance matrix of the channel coefficient in the
described below. Then, the lr-MMSE estimator developed for this frequency domain and σ2 is the noise variance. Thus, the MMSE
work is presented. channel estimation can be expressed as
3.2.1. Least squares    RH X H 
H MMSE  LY   Y
 XRH X  RW
H
The main reason for using the LS algorithm is to find an optimal 
estimator for the unknown parameters by minimizing the LS er-
 
rors [31]. Generally, LS is used to get an initial channel estimation  
 RH  H
1
at the pilot subcarriers [32]. The channel statistical model in the   
LS
2 1
frequency domain for the OFDM VLC system can be modelled as   RH      
 XX   
H
  (12)
Y  XH  W (4)
From Equation (12), it is can be seen that the MMSE channel
Where the Gaussian noise is denoted by W with variance σ2. estimation method mitigates the channel noise effects relatively
better than the LS method. However, in the MMSE method more
Then, the input matrix is X = diag({Xk}). To find the LS channel prior knowledge is needed about the channel coefficient matrix
estimation, the sum of the square errors of channel estimation is (auto-covariance) and noise covariance matrix, RH and RW, re-
computed in the frequency domain. By using differentiation and spectively. In practice, these unknown parameters of channel fre-
setting Equation (4) to zero with each channel coefficient H, the quency response make MMSE channel estimation almost impos-
LS channel estimation in the frequency domain can be calculated sible in real time [29]. Therefore, the MMSE method has greater
as follows: complexity than the LS method.
 Y
H LS  3.2.3. Low-rank minimum mean square error
X (5)
The low-rank approximation for MMSE channel estimation de-
The channel estimation objective and estimation error can be veloped for this paper simplifies the computational process of
found by substituting Equation (4) into Equation (5): MMSE to reduce complexity in three ways. The first simplifica-
H 1
tion involves replacing ( XX ) by its expectation E  {( XX ) } .
H 1

 W
H LS  H  It is assumed that the constellations of the signal on all tones are
X (6)
the same and that there is an equal probability on all constellation
points, thus:
Where W/X is the estimated error and H is the objective of chan-
nel estimation. The LS is simple and easy to implement. However, 2
it has low accuracy. E  {( XX H ) 1}  E  
 1X  I
 k (13)
3.2.2. Minimum mean square error
Where I represents the identity matrix.
The main aim of using the MMSE algorithm is to develop an op- This defines the mean SNR as
timal estimator for finding the unknown parameters by minimizing
the mean square error (MSE). Defining the linear estimator opera- SNR  E{ X k } /  2
2

tor L [16] finds the MMSE as (14)

 2

H MMSE  LY   E{ X k }E  1 X 
2

(7) Then, using  k , the MMSE channel estimation can be

rewritten as
Based on Equation (7), the MSE of the channel estimation for the
 
OFDM VLC system can be modelled as   
H MMSE  RH  H
1
    
LS

   RH    I  
( L)   ( H  H MMSE ) 
2  SNR   
  (15)
  (8)
The singular value decomposition (SVD) is performed over Her-
mitian RH :
International Journal of Engineering & Technology 2641

stellation. According to the received signal in Equation (4), the

RH  UU H (16) conditional probability density function (PDF) of output Y given
X and H is as follows:
1  1 
Where U denotes a unitary matrix that contains singular vectors f (Y | X , H )  exp  | Y  HX |2 
2 2  2
2

and  is referred to as a diagonal matrix that contains singular (19)
values that are λ1 ≥ λ2 ≥ · · · ≥ λM on its diagonal.
The use of SVD significantly decreases the complexity of the Next, Gaussian random variables with variance σ2 and zero mean
calculation process of the matrix. First, the estimator finds the are considered. Then, assuming that the probability and averaging
required number of ranks, denoted by p in Equation (17), which of all M symbols in the constellation are equal for the conditional
should be bigger than (J + 1) [33]. Then, the rest of the parameters PDF in Equation (19), the received signal can be expressed Y
that appear in Equation (17) are as given; the singular values, λk, given H over X as follows:
the channel auto-covariance matrix, RH, the noise variance, SNR,
 

M 1 1 1
the receiver pre-calculates, β, and the unitary matrix, U. f (Y | H ) 
i 1 M 2 2
exp 
 2
2
| Y  H . X i |2 
 (20)
In this paper, the lr-MMSE channel estimation algorithm was
developed so that it would be compatible with the OFDM VLC
For indoor VLC, assuming the channel is not time dependent over
system. Therefore, β is a constant and must be divided by 2 due to
a D of OFDM symbols, the transmitted and received symbols in
its unipolar signal and it depends only on the signal constellation.
Therefore, the M × M diagonal matrix,  , is obtained with the
p the vector can be denoted as Y = [Y1,Y2, …YD]T and X =
following parameters: [X1,X2,…XD]T, respectively.
The objective in this work is to find the estimated value of the
  transmitted bits, X from the received signal, Y as expressed in
 k 

 2 , k  0,1,....,p  1
Equation (4), and to reduce the BER caused by the wireless optical
 k  
k  channel. The EM-VLC algorithm is used to estimate X and max-
 SNR 
imize f Y | X  by calculating the mean of the log-likelihood func-

0 k  p,....,M  1 (17)
tion f Y , h | X  over the unknown parameters h . Therefore, it is es-
To calculate the LS according to Equation (4), the pilots X and sential to determine the channel frequency response H in order to
received signals Y are used during transmission, where the lr- estimate the transmitted signal X. Here, (Y , h ) represents complete
MMSE channel estimation with the rank p is given by data and (Y ) denotes incomplete data. It is not easy to expect the
  channel frequency response based on the ‘incomplete’ data. How-
Hlr  MMSE  U pU H H LS ever, the PDF of ‘incomplete’ data can be converted into the PDF
(18)
of ‘complete’ data, and expressed as
From Equations (17) and (18), it can be seen that the adapted lr- D

MMSE estimator proposed above is significantly less complex f (Y | H , X )   f (Y d

| H, X d )
d 1 (21)
than the MMSE estimator and also performs better at low SNR
values. The reduction in the complexity of channel estimation
Therefore, the incomplete data log-likelihood function is given by
provided by the proposed lr-MMSE is due to the optimal value of
the subspace dimension within J+1, and maintain defining the D
channel as well [34], compared to the conventional Ir-MMSE log f (Y | H , X )   log f (Y d
| H, X d )
estimator where there is a problem of recovering the Ir-MMSE d 1 (22)
matrix and uncertainty quantification [35].
The PDF of ‘complete’ data can be written in the same manner as
in Equation (22) by using the log-likelihood function as follows:
4. Proposed EM-VLC algorithm
D
 
  log M
1
The new channel estimation method for OFDM VLC systems log f (Y , X | H )  f (Y d | H , X d ) 
d 1 
proposed in this paper is a semi-blind technique that uses an itera- (23)
tive method [28] to determine the ML estimation of the channel.
This iterative method of channel estimation can be employed to In each iterative process of p, where p can be any integer number
reduce the channel estimation errors. The greater the number of (0, 1, 2,), H can be estimated by maximizing f (Y | H ) and by itera-
iterations, the better the performance achieved. However, this tively increasing the likelihood function at each step.
results in a greater computational cost. To mitigate this drawback, The EM-VLC algorithm consists of the following two steps to
the proposed EM-VLC algorithm uses the same error probabilities estimate H from Y :
that have already been used by the decoders. For instance, in turbo Expectation (E):
decoding, the maximum a posteriori decoder [15,36] is utilized to
find the expectations of the symbols to be transmitted. This is Q( H | H p )  E{log f (Y , X | H ) | Y , H p }
(24)
exactly the approach that has been adopted in the proposed EM-
VLC algorithm for channel estimation. Thus, the computational Maximization (M):
complexity of the EM-VLC algorithm is reduced significantly.
Moreover, no prior knowledge of the channel condition is required. H p 1  arg max Q( H | H p )
The proposed algorithm is an iterative method for computing the H (25)
ML channel estimation and finding unknown parameters. It is
executed in two steps. The first step involves expectation (E) for In the E step, Hp represents the latest estimate of H. The estimated
unknown parameters and the second step involves maximization value of the log-likelihood function of H is calculated by taking
(M) to maximize the expectation of the log-likelihood function. the estimation over X , considering the value of Y and using the
These two steps are repeated until the estimated values converge most recent estimated value of H. In the subsequent M step, Hp+1
[37]. is determined as in Equation (25). The final estimation of (p+1)st
Thus, it is assumed that X stands for one of the M-ary symbols in of H becomes H(p+1), which is the additional manipulation of
the frequency domain of the constellation size M, in which X  H(p+1). These two steps should result in a sufficiently small dif-
{X1, X2,.….XM}, where Xi refers to the ith symbol of the con- ference in value between H(p+1) and Hp. As the H is found, fre-
2642 International Journal of Engineering & Technology

quency channel response. Then, the ML of the transmitted signal is presented. Figure 2 shows the BER produced by the proposed
can be expressed as algorithm when using a FFT size of 64 and a pilot spacing of 8.
The proposed algorithm performs better than the other three chan-
X (n)  arg min | Y (n)  H (n) X (n) |2 nel estimation methods, LS, MMSE and Ir-MMSE. This is due to
X M , 0  n  2N  1 (26)
the fast convergence of the EM-VLC algorithm. A comparatively
low BER is achieved by the EM-VLC algorithm at different SNR
The final estimation of the transmitted signal, X, can be found by
values. As explained in section 4, and based on the analytical re-
 Y ( n) 
sults, the EM-VLC algorithm has less complexity than the MMSE
X (n)  Quantization  channel estimation method because the former can estimate un-
 H ( n)  , 0  n  2N  1 (27) known parameters (noise variance and channel covariance) and
retrieve transmitted bits.
Two factors that have a high impact on the performance of the The MMSE estimator requires prior knowledge of the channel
EM-VLC algorithm are the number of iterations and the initializa- covariance and noise variance. Furthermore, the MMSE estimator
tion step. Greater accuracy can be achieved with a greater number has higher complexity than the LS estimator. However, the LS
of iterations. However, as the number of iterations increases the estimator is less accurate than the MMSE estimator. Thus, the
complexity increases as well. Initialization can be performed ei- higher the SNR, the better the performance of the MMSE estima-
ther with pilot insertion or without pilots by using the statistical tor in comparison to the LS estimator.
properties of the signal, such as in direct decision channel estima- On the other hand, the enhanced lr-MMSE estimator is a modified
tion. In this paper, we use LS estimator pilot insertion for initiali- MMSE estimator and it has less complexity due to some simplifi-
zation, using a grid-type pilot as in [8] for a more appropriate cations, as explained in section 3.2.3. Thus this estimator is, in
initial value for X 0 . essence, a compromise between the MMSE and LS estimators in
It is possible to obtain results that are more accurate with pilot respect of performance and complexity. Therefore, the BER of the
insertion [38], and less time is required for convergence. On the lr-MMSE estimator lies between that of the MMSE and LS esti-
other hand, a direct decision is undesirable in real-time estimation mators. The worst BER occurs when no channel estimation meth-
because the data detection of the current estimation relies on the od is applied.
previous OFDM symbol, then the new detected data is used for the In this paper, the highest BER at all SNR values occurs when no
estimation of the current channel [39]. Practically, the EM-VLC channel estimation is used, while the lowest BER at all SNR val-
algorithm is a highly robust technique for channel estimation ues occurs when the EM-VLC algorithm is used. For example, at a
when the available data are incomplete. This is because it has high SNR of 13 dB, the proposed EM-VLC algorithm has a BER of 10-
accuracy and less complexity compared to other channel estima- 1. This value shows a gain of [1], [2]. [5] And [3] dB of BER over
tion methods. the other channel estimation methods, MMSE, lr-MMSE and LS,
respectively.
5. Simulation results and discussion
5.1. Simulation setup

The optical channel with IM/DD in the VLC was modelled as a

baseband linear system [40] with the channel response that was
defined above in Equation (2). In the simulation setup, linear scal-
ing and DC bias were used so that the OFDM signal was within
the dynamic range and non-negativity was constrained, respective-
ly. This setup is called DCO-OFDM. The transmitted bits were
randomly generated and M-ary quadrature amplitude modulation
(M-QAM) symbols were formed by serial to parallel conversion.
The QAM symbols were assigned to subcarriers. Thus, the light Fig. 2: BER Performance of EM-VLC Algorithm for 64 FFT and Pilot
source was a modulated signal that could be transmitted over the Spacing of [8].
OWC channel. At the receiver, the reverse process was executed
after the DC bias as removed. The VLC channel is not subject to Figures 3 and 4 provide the results of a comparison of the BER of
fast fading effects because it uses a very small wavelength com- the proposed EM-VLC algorithm with that of the LS, MMSE and
pared to the detector area [41]. However, it is subject to slow fad- lr-MMSE algorithms, in which six and four pilot insertions were
ing, which is the result of shadowing from furnishings or other used, respectively. The results show that increasing the pilot den-
indoor objects. The effect of shadowing can be overcome by in- sity (i.e., less pilot spacing) can improve estimation accuracy.
creasing the number of LEDs. The received optical signal power, However, this introduces a lower transmitted bit rate and results in
distributed on the receiver plane was estimated based on the line- poorer spectral efficiency. Moreover, the pilot arrangement or
of-sight channel model and by ignoring the reflection off the walls. pilot pattern, which is a grid type in this paper, can also affect
The simulation parameters are provided in Table 1. channel estimation. Hence, an optimum pilot arrangement is a
trade-off between accuracy and the energy wasted [29].
Table 1: Simulation Parameters It is clear from Fig. 3 and Fig. 4 that, in the proposed method, the
Parameter Specification BER is maintained at almost the same value when the pilot spac-
System DCO-OFDM-based VLC ing is decreased. This is because pilot insertion is only used to find
IFFT size (N) 16, 32, 64, 128, 512 the initial value. Therefore, unnecessary pilot insertions can be
Channel OWC with WGN avoided when using the EM-VLC algorithm, which improves the
Modulation schemes 16, 64 QAM
transmitted bit rate and results in higher spectral efficiency. The
Pilot insertion type Grid type
Pilot spacing 4, 6, 8 other estimators need a high number of pilots to track the channel
Initial channel estimation LS estimator changes for better channel estimation, which is especially the case
for the MMSE estimator. In the proposed method, computational
5.2. Results and discussion complexity is significantly reduced because complexity is ex-
tremely dependent on the number of subcarriers.
In this section, a comprehensive analysis of the performance of the For instance, in Fig. 3, the EM-VLC algorithm produces a lower
proposed EM-VLC algorithm for OFDM VLC channel estimation BER compared to the other three algorithms. For instance, it has a
BER of 10-1 at a SNR of 12 dB, whereas MMSE, lr-MMSE and
International Journal of Engineering & Technology 2643

LS need have a higher SNR of 14.5, 15 and 16.5 dB, respectively,

in order to produce a similar BER. In Fig. 4, the number of pilots
is higher (i.e., there is less pilot spacing), so the BER is higher.
For example, the EM-VLC algorithm achieves a BER of 10-1 at 12
dB, whereas MMSE, lr-MMSE and LS need to have a higher SNR
of 15, 15.5 and 17 dB, respectively, in order to produce a similar
BER. This is because the fewer the pilot insertions, the higher the
bit rate that can be transmitted.

Fig. 5: Comparative BER Performance for 512 FFT and Pilot Spacing of
[4].

Figure 6 shows that the proposed EM-VLC algorithm produces

almost the same BER at FFT sizes ranging from 16 to 256. This
indicates that the proposed method has another significant ad-
vantage, i.e.,the accuracy of the EM-VLC estimator is independent
of the increase in the number of samples for each signal. In addi-
Fig. 3: Comparative BER Performance for 64 FFT and Pilot Spacing of tion, the EM-VLC algorithm maintains the lowest BER compared
[6].
to all the other estimators (MMSE, lr-MMSE and LS) at different
FFT sizes because it does not rely on a large number of pilots
To evaluate the effectiveness of the proposed EM-VLC algorithm
because it only uses pilots at initialization.
for OFDM VLC systems, several simulations were carried out for
different FFT sizes and modulation schemes. Figure 5 shows that
the proposed EM-VLC algorithm produces a lower BER than the
MMSE, lr-MMSE and LS estimators when the FFT size increases.

Fig. 6: BER of EM-VLC Algorithm with 64 QAM at Different FFT.

Figure 7 depicts the BER produced by the compared estimators at

different QAM. Obviously, the BER decreases gradually as the
Fig. 4: Comparative BER Performance for 64 FFT and Pilot Spacing of SNR increases for all types of QAM. As can be seen from Fig. 7,
[4]. the proposed EM-VLC algorithm outperforms all of the other
channel estimation methods for all types of QAM. This stems
In Figure 5, the EM-VLC algorithm has a BER of 10-1 at a SNR of from the EM-VLC algorithm’s ability to estimate unknown pa-
13 dB with a FFT size of 512, which is the same as the BER and rameters in a very low number of iterations before convergence.
SNR it produced at the FFT size of 64 (see Fig. 2). The MMSE, lr-
MMSE and LS estimators at a BER of 10 -1 achieve about a 0.9,
1.7 and 0.2 dB reduction in SNR, respectively, with a FFT size of
512, compared with a FFT size of 64 in Fig. 2. In other words, the
proposed EM-VLC algorithm can produce the best BER at all FFT
sizes.
However, the other estimators (MMSE, lr-MMSE and LS) do
perform better when the FFT size is larger (for example, 512 FFT
compared to 64 FFT). This is because when the FFT size is in-
creased, more samples are added for each signal. Consequently,
the signal is smoother and more accurate.

Fig. 7: Comparative BER Performance with 16 QAM and 64 QAM.

It is also clear from Fig. 7 that the BER increases when the QAM
order increases for each channel estimation algorithm. For in-
stance, with 16-QAM the performance of the MMSE estimator is
better than that of the LS estimator; the MMSE estimator can
achieve the same BER of 10-1 but at a lower SNR of 3 dB. In other
words, the higher the SNR, the better the performance of the
MMSE estimator compared to the LS estimator. Moreover, the LS
estimator is susceptible to channel noise and variance in the esti-
2644 International Journal of Engineering & Technology

mation error. Indeed, the performance of the LS estimator is high- using adaptive channel estimation algorithm. IET Communications,
ly dependent on the noise level. On the other hand, the MMSE 9(4), 501-507. https://doi.org/10.1049/iet-com.2014.0649.
estimator is more powerful in terms of its ability to eliminate noise [8] Lin, W.-F., Chow, C.-W., & Yeh, C.-H. (2015). Using specific and
adaptive arrangement of grid-type pilot in channel estimation for
and because it relies on channel frequency correlation it obtains a
white-lightLED-based OFDM visible light communication system.
better estimation. However, it does have higher complexity than Optics Communications, 338, 7-10.
the LS estimator. As mentioned above, the lr-MMSE estimator https://doi.org/10.1016/j.optcom.2014.09.080.
developed in this paper is a compromise between the MMSE and [9] Chen, X., & Jiang, M. (2017). Adaptive Statistical Bayesian MMSE
LS estimators in terms of performance and complexity. Hence it Channel Estimation for Visible Light Communication. IEEE Trans-
produces a better BER than the MMSE, except at higher SNR, and actions on Signal Processing, 65(5), 1287-1299.
better BER than the LS estimator. In light of the above, the per- https://doi.org/10.1109/TSP.2016.2630036.
formance of the proposed EM-VLC algorithm is superior to that of [10] Wang, J.-b., Jiao, Y., Dang, X.-y., Chen, M., Xie, X.-x., & Cao, L.-l.
(2011). Training sequence based channel estimation for indoor vis-
all the other methods compared in this paper.
ible light communication system. Optoelectronics Letters, 7, 213-
216. https://doi.org/10.1007/s11801-011-0143-7.
6. Conclusion [11] Wang, J.-B., Jiao, Y., Song, X., & Chen, M. (2011). Optimal train-
ing sequences for indoor wireless optical communications. Journal
of Optics, 14(1), 015401. https://doi.org/10.1088/2040-
The receiver of an OFDM-based VLC system needs to be 8978/14/1/015401.
equipped with high-precision channel estimation to improve data [12] Wang, J.-B., Jiao, Y., Song, X., Zhao, X., Gu, M., Chen, M., et al.
detection performance. Therefore, in this paper, a new EM-VLC (2011). Complementary sequences–based channel estimation for
channel estimation algorithm for OFDM-based VLC systems was diffuse wireless optical communications. Optical Engineering,
proposed and evaluated. In addition, a low-rank minimum mean 50(7), 075003-075003-075006.
square error channel estimation algorithm (lr-MMSE) was devel- [13] Pham, Q. N., Rachim, V. P., an, J., & Chung, W.-Y. (2017). Ambi-
ent Light Rejection Using a Novel Average Voltage Tracking in
oped and assessed. The proposed EM-VLC algorithm is an itera-
Visible Light Communication System. Applied Sciences, 7(7), 670.
tive method for computing the ML channel estimation and finding https://doi.org/10.3390/app7070670.
unknown parameters. It was evaluated and compared with three [14] Wang, J.-B., Xie, X.-X., Jiao, Y., & Chen, M. (2012). Training se-
different channel estimation methods, MMSE, lr-MMSE and LS, quence based frequency-domain channel estimation for indoor dif-
as well as with no channel estimation. Extensive simulations were fuse wireless optical communications. EURASIP Journal on Wire-
conducted to assess the robustness of the proposed algorithm. less Communications and Networking, 2012(1), 1-10.
The results demonstrated that the proposed EM-VLC algorithm https://doi.org/10.1186/1687-1499-2012-326.
produced a much lower BER and had less complexity than the [15] Zhuang, X., & Vook, F. W. Iterative channel estimation and decod-
ing for a turbo-coded OFDM system via the EM algorithm. In
MMSE, lr-MMSE and LS channel estimation methods. This is
Acoustics, Speech, and Signal Processing (ICASSP), 2002 IEEE
because EM-VLC is not dependent on pilot insertion, except at International Conference on, 2002 (Vol. 3, pp. III-2337-III-2340):
initialization, and yet this independence does not affect its ability IEEE
to find unknown parameters. While the LS estimator is simple it [16] Krishna, E. H., Sivani, K., & Reddy, K. A. OFDM channel estima-
produces a higher BER than MMSE. On the other hand, although tion and equalization using multi scale independent component
the MMSE estimator produces a lower BER than LS, it has higher analysis. In Signal Processing, Informatics, Communication and
complexity. The lr-MMSE estimator was developed in this study Energy Systems (SPICES), 2015 IEEE International Conference on,
as a compromise between the MMSE and LS estimators in order 2015 (pp. 1-5): IEEE.
[17] Popoola, W. O., Ghassemlooy, Z., & Stewart, B. G. (2014). Pilot-
to achieve a better balance between performance and complexity.
assisted PAPR reduction technique for optical OFDM communica-
Thus, in practice, the proposed EM-VLC algorithm provides a tion systems. Journal of Lightwave Technology, 32(7), 1374-1382.
great trade-off between accuracy and complexity. In addition, it is https://doi.org/10.1109/JLT.2014.2304493.
more suitable for real-time estimation. In future work, the EM- [18] Popoola, W. O., Ogunkoya, F. B., & Sinanović, S. Mitigating non-
VLC algorithm could be extended to make it suitable for a more linearities under average power constraint in visible light communi-
sophisticated OFDM-based VLC system, such as multiple-input cation. In Wireless Communications and Mobile Computing Con-
multiple-output (MIMO). ference (IWCMC), 2014 International, 2014 (pp. 815-820): IEEE.
[19] Hussein, Y. S., Alias, M. Y., & Abdulkafi, A. A. On performance
analysis of LS and MMSE for channel estimation in VLC systems.
References In 2016 IEEE 12th International Colloquium on Signal Processing
& Its Applications (CSPA), 2016 (pp. 204-209): IEEE.
[1] Singh, S., & Bharti, R. (2014).163m/10Gbps 4QAM-OFDM visible [20] Dissanayake, S. D., & Armstrong, J. (2013). Comparison of aco-
light communication. IJETR, 2, 225-228. ofdm, dco-ofdm and ado-ofdm in im/dd systems. Journal of Light-
[2] Burchardt, H., Serafimovski, N., Tsonev, D., Videv, S., & Haas, H. wave Technology, 31(7), 1063-1072.
(2014). VLC: Beyond point-to-point communication. Communica- https://doi.org/10.1109/JLT.2013.2241731.
tions Magazine, IEEE, 52(7), 98-105. [21] Ma, X., Kobayashi, H., & Schwartz, S. C. (2004). EM-based chan-
https://doi.org/10.1109/MCOM.2014.6852089. nel estimation algorithms for OFDM. EURASIP Journal on Ap-
[3] Li, Y., Tsonev, D., & Haas, H. Non-DC-biased OFDM with optical plied Signal Processing, 2004, 1460-1477.
spatial modulation. In Personal Indoor and Mobile Radio Commu- [22] Karkhanechi, H. M., & Levy, B. C. EM-based channel estimation
nications (PIMRC), 2013 IEEE 24th International Symposium on, for space-time block coded MIMO OFDM systems. In Signal Pro-
2013 (pp. 486-490): IEEE. cessing Systems Design and Implementation, 2006. SIPS'06. IEEE
[4] Xia, C., Hainfeng, W., & Honglin, H. A Low Complexity Pilot- Workshop on, 2006 (pp. 177-181): IEEE.
Aided Channel Estimation Resisting Residual Timing Errors for [23] Kabaoğlu, N. (2013). EM Based Data Detection Algorithm for
OFDM Systems. In Personal, Indoor and Mobile Radio Communi- Downlink of a Single Hop Relaying Network. Wireless Personal
cations, 2007. PIMRC 2007. IEEE 18th International Symposium Communications, 1-14. https://doi.org/10.1007/s11277-012-0562-1.
on, 3-7 Sept. 2007 2007 (pp. 1-5). [24] Al-Naffouri, T. Y. (2007). An EM-based forward-backward Kal-
https://doi.org/10.1109/PIMRC.2007.4394644. man filter for the estimation of time-variant channels in OFDM.
[5] Guvenc, I., Gezici, S., Sahinoglu, Z., & Kozat, U. C. (2011). Relia- IEEE Transactions on Signal Processing, 55(7), 3924-3930.
ble communications for short-range wireless systems: Cambridge https://doi.org/10.1109/TSP.2007.894271.
University Press. https://doi.org/10.1017/CBO9780511974366. [25] Vizziello, A., Savazzi, P., & Favalli, L. Mitigation of intercarrier
[6] Yang, X., Min, Z., Xiongyan, T., Jian, W., & Dahai, H. A post- interference in OFDM systems over slow fading channels based on
processing channel estimation method for DCO-OFDM visible the EM algorithm. In MELECON 2010-2010 15th IEEE Mediterra-
light communication. In Communication Systems, Networks & nean Electrotechnical Conference, 2010 (pp. 1320-1325): IEEE.
Digital Signal Processing (CSNDSP), 2012 8th International Sym- [26] Armstrong, J., & Schmidt, B. J. (2008). Comparison of asymmetri-
posium on, 2012 (pp. 1-4): IEEE. cally clipped optical OFDM and DC-biased optical OFDM in
[7] Zhang, T., Guo, S., Chen, H., Zhong, F., & Ma, C. (2014). Enhanc- AWGN. Communications Letters, IEEE, 12(5), 343-345.
ing the bit error rate of indoor visible light communication systems https://doi.org/10.1109/LCOMM.2008.080193.
International Journal of Engineering & Technology 2645

[27] Carruthers, J. B., & Kahn, J. M. Modeling of nondirected wireless

infrared channels. In Communications, 1996. ICC'96, Conference
Record, Converging Technologies for Tomorrow's Applications.
1996 IEEE International Conference on, 1996 (Vol. 2, pp. 1227-
1231): IEEE.
[28] Cho, Y. S., Kim, J., Yang, W. Y., & Kang, C. G. (2010). MIMO-
OFDM wireless communications with MATLAB: John Wiley &
Sons. https://doi.org/10.1002/9780470825631.
[29] Ozdemir, M. K., & Arslan, H. (2007). Channel estimation for wire-
less OFDM systems. IEEE Communications Surveys & Tutorials,
9(2), 18-48. https://doi.org/10.1109/COMST.2007.382406.
[30] Bandara, K., Niroopan, P., & Chung, Y.-H. PAPR reduced OFDM
visible light communication using exponential nonlinear com-
panding. In Microwaves, Communications, Antennas and Electron-
ics Systems (COMCAS), 2013 IEEE International Conference on,
2013 (pp. 1-5): IEEE.
[31] Arslan, H., & Bottomley, G. E. (2001). Channel estimation in nar-
rowband wireless communication systems. Wireless Communica-
tions and Mobile Computing, 1(2), 201-219.
https://doi.org/10.1002/wcm.14.
[32] Won Gi, J., Kyung Hyun, P., & Yong Soo, C. An efficient channel
estimation technique for OFDM systems with transmitter diversity.
In Personal, Indoor and Mobile Radio Communications, 2000.
PIMRC 2000. The 11th IEEE International Symposium on, 2000
2000 (Vol. 2, pp. 1246-1250 vol.1242).
https://doi.org/10.1109/PIMRC.2000.881618.
[33] Wang, S., & Abdi, A. (2005). Aperiodic complementary sets of se-
quences-based MIMO frequency selective channel estimation.
IEEE Communications Letters, 9(10), 891-893.
https://doi.org/10.1109/LCOMM.2005.10019.
[34] Shen, Y., & Martinez, E. (2006). Channel estimation in OFDM sys-
tems. Application note, Freescale semiconductor.
[35] Carpentier, A., & Kim, A. K. (2015). An iterative hard thresholding
estimator for low rank matrix recovery with explicit limiting distri-
bution. arXiv preprint arXiv:1502.04654.
[36] Ma, X., Kobayashi, H., & Schwartz, S. C. An EM-based estimation
of OFDM signals. In Wireless Communications and Networking
Conference, 2002. WCNC2002. 2002 IEEE, 2002 (Vol. 1, pp. 228-
232): IEEE.
[37] Moon, T. K. (1996). The expectation-maximization algorithm. Sig-
nal processing magazine, IEEE, 13(6), 47-60.
https://doi.org/10.1109/79.543975.
[38] Xie, Y., & Georghiades, C. N. (2003). Two EM-type channel esti-
mation algorithms for OFDM with transmitter diversity. Communi-
cations, IEEE Transactions on, 51(1), 106-115,
https://doi.org/10.1109/TCOMM.2002.807617.
[39] Bossert, M., Donder, A., & Trushkin, A. (1995). Channel estima-
tion and equalization in orthogonal frequency division multiplexing
systems. ITG FACHBERICHT, 485-485.
[40] González, O., Guerra, M., & Martín, R. (2014). Multi-User Visible
Light Communications. Advances in Optical Communications, 36-
63.
[41] Wang, C.-X., Haider, F., GAO, X., You, X.-H., Yang, Y., Yuan, D.,
et al. (2014). Cellular architecture and key technologies for 5G
wireless communication networks. IEEE Communications maga-
zine, 52(2), 122-130. https://doi.org/10.1109/MCOM.2014.6736752.