1030
9, MAY 1, 2016
Achieving FEC and RLL for VLC: A Concatenated
Convolutional-Miller Coding Mechanism
Xuanxuan Lu and Jing Li (Tiffany), Senior Member, IEEE
Abstract— Run-length limited (RLL) codes are widely used in
visible light communications to avoid long runs of 1s and 0s
that potentially cause the flicker. This letter explores the serial
concatenation of convolutional codes and Miller codes to simul-
taneously achieve forward error correction and RLL control.
Miller codes with high bandwidth efficiency are much ignored in
practice due to their disappointing power efficiency. The novelty
of this letter is that we identity the merit of this previously
unfavorable RLL code (i.e., trellis structure and soft decidability),
exploit some important coding principles (i.e., interleaved serial
concatenation and soft-iterative decoding), and assemble it to a
powerful turbo structure that makes it outperform the existing
favorable choices. A modified Bahl-Cocke-Jelinek-Raviv decoding
algorithm is developed for the proposed concatenation. Analysis
and simulations confirm that the new system is capable of solid
RLL control and a superb performance better than the existing
schemes.
Index Terms— Miller codes, RLL code, iterative decoding,
visible light communication.
I. I NTRODUCTION
A DVANCES in solid-state lighting not only enable light
emitting diodes (LED) to be a promising source for
future lighting, but also make possible the use of visible
light communications (VLC) [1]. VLC is desirable due to an
unregulated bandwidth, no interference with the existing radio
frequency (RF) systems [2]. Consider a practical VLC system
using the visible light spectrum from 380 to 780 nm to provide
the dual-purpose of illumination and communication. Among
the key building blocks are the run-length limited (RLL) code
and the forward error correction code (FEC). The former aims
at eliminating long runs of 1’s and 0’s that could potentially
cause flicker and clock and data recovery detection problems
and the latter is indispensable for providing reliable data
communication.
A number of practical RLL codes have been proposed in
the VLC literature, including the Manchester code, the 4B6B
code and the 8B10B code [3]. On the error control front, there
is the study of advanced FEC coding technology to recover
data from the noise and to achieve dimming control. Notewor-
thy examples include conventional Reed-Muller (RM) codes
based on the bent function [4], rate-compatible convolutional
codes [5], and low-density parity-check (LDPC) codes [6].
While these excellent previous researches have shed good
insight into RLL and FEC separately, little is done about
Manuscript received December 1, 2015; revised January 11, 2016; accepted
January 25, 2016. Date of publication February 3, 2016; date of current
version March 17, 2016. This work was supported by the National Sci-
ence Foundation under Grant 0928092, Grant 1133027, Grant 1343372, and
Grant PITA PIT-15-22.
The authors are with the Department of Electrical and Computer
Engineering, Lehigh University, Bethlehem, PA 18015 USA (e-mail:
xul311@lehigh.edu; jingli@ece.lehigh.edu).
Color versions of one or more of the figures in this letter are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LPT.2016.2524578
1041-1135 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 28, NO.
unified coding solutions. The only exception is a recent
work by Wang and Kim [7], where the authors pioneered a
concatenation of an outer Reed Solomon (RS) code and an
inner 4B6B (or 8B10B) code. By developing a soft decoding
strategy for the 4B6B (and 8B10B) code that produces mul-
tiple decoding candidates for the outer RS code, the authors
showed considerable performance gains over hard-decoding
systems [7].
This letter considers developing joint FEC-RLL coding
mechanisms for VLC systems to simultaneously achieve
good bandwidth efficiency, good error correction and good
run-length control. The specific scheme we propose is a
convolutional-Miller serial concatenation that consists of
an on-off-keying (OOK) modulated Miller code serving as
the inner code, a convolutional code serving as the outer
code. Miller codes has disappointing bit error rate (BER)
performance, but they possess interesting features including
high spectral efficiency (e.g. twice as spectral efficient as
Manchester codes) and efficient soft-decodability. The latter
is particularly worth noting, because unlike other RLL codes
whose soft-decoding involves exhaustive search (with expo-
nential complexity), Miller codes have elegant trellis structures
that lend themselves to linear-time, optimal soft decoding.
Motivated by this and exploiting the renowned “serial concate-
nation and soft-iterative paradigm” (which states that, when
done right, short, weak codes can be concatenated to form
long, strong codes), we pull together convolutional codes and
Miller codes and assemble them into a powerful serial concate-
nation. Extrinsic information transfer (EXIT) analysis confirms
the matching harmony of these codes. An overall soft-iterative
decoder is developed to achieve a superior BER performance
that wins over the existing schemes. We thus advocate the
new coding structure in VLC, in terms of solid RLL control,
high spectral efficiency and strong error correction capability.
II. M ILLER C ODE AND S YSTEM M ODEL
Manchester codes with OOK prevail the literature of VLC.
Being the classic choice, Manchester codes are easy to imple-
ment and very effective in controlling run lengths and main-
taining clock synchronization, but suffer from poor bandwidth
efficiency [8]. Since Miller codes have at least one transition
every two bit intervals and there is never more than two tran-
sitions every two bit intervals, they therefore provide effective
run-length control with better bandwidth efficiency. Further,
when the five adjacent channels for the data rates specified
by the 802.15.7 standard coexist, the five carriers may overlap
and become difficult to separate, making Manchester codes
highly vulnerable to error [9].
From the coding perspective, both Manchester codes and
Miller codes may be viewed as rate-1/2 codes. However,
Manchester codes have no memory. In comparison, Miller
codes have memory and take a trellis structure similar to
that of an unterminated convolutional code (see Fig. 1)).
LU AND LI: ACHIEVING FEC AND RLL FOR VLC
Fig. 1. Trellis diagram of Miller codes. At each stage, the dashed branch is
associated with input 0 and the solid line with input 1. The encoded output
bit takes the same value as the value of the next state.
Fig. 2. Super-symbol brightness in Miller codes with different super-symbol
size M = 1, 4, 8, 30.
This makes their effective code length longer, and allows for
optimal linear-time decoding such as the Viterbi algorithm
(for hard decisions) and the Bahl-Cocke-Jelinek-Raviv (BCJR)
algorithm (for soft decision).
A. Flicker Mitigation and Dimming Control
A VLC system has dual functions of illumination and
communication, and the former takes priority. For illumination
to be minimally affected by data communication requires
the proper handling of flicker mitigation (which controls the
fluctuation of the light) and brightness control.
Define the maximum flickering time period (MFTP) as
the maximum time period over which the light intensity
can change without human perception. Then the inverse
of MFTP sets the lower frequency bound for flicker
mitigation. In general, a flicker frequency greater than 200 Hz
(MFTP < 5 ms) is imperceptible by human eyes. Since
Miller codes guarantee transitions every other bit, setting the
data stream with time period <2.5 ms would safely clear
the flicker issue. This is easily satisfied, as today’s VLC can
provide a data rate up to 1 Gbps [2] (i.e. transmission time
per byte <10−6 ms).
It is apparent that Manchester codes, with codewords being
“01” and “10”, always provide a desirable brightness of 50%,
whereas Miller codes, also equipped with codewords
(i.e. symbols) “00” and “11”, are less consistent in per-symbol
brightness intensity. However, if one considers M multiple
consecutive symbols as an “M-super-symbol”, and ensures
that the M-super-symbol meets the MFTP limit, then the
brightness level actually becomes fairly consistent per super-
symbol. As illustrated by the histogram in Fig. 2 (random bit
generation), as the size of the super-symbol increases from
1031
Fig. 3. The proposed system model.
M = 1 to 4, 8 and 30, the brightness intensity of the super-
symbols becomes increasingly uniform and approaches the
desirable level of 50% to keep DC balance.
The specific scheme we propose is a convolutional-Miller
serial concatenation that consists of an on-off-keying (OOK)
modulated Miller code serving as the inner code, a convolu-
tional code serving as the outer code. In VLC systems that
employ OOK, it is generally understood that a good dimming
control is achieved when the system generates about 50%
of the brightness level (i.e. when the “space”(bit 0) and the
“mark”(bit 1) appear with equal probability). With that “50%”
achieved, to further increase or decrease the light intensity
becomes rather straight-forward: one can either insert the
so-called “compensation” symbols (i.e. additional marks or
spaces) in a regular and patterned manner, or, leverage the
popular form of PWM.
B. A Concatenated FEC-RLL Coding Framework
Having analyzed Miller codes, we now detail the proposed
joint FEC-RLL coding solution. As shown in Fig. 3, the system
consists of an outer convolutional code, a random interleaver,
an inner Miller code. For clarity of presentation, we assume
the LEDs use OOK modulation. It is noted that, in general,
the scheme can be employed in the orthogonal frequency
division multiplexing (OFDM) systems with multi-level phase
shift keying (MPSK) and multi-level quadrature amplitude
modulation (MQAM) format to further support the high data
transmission rate.
Convolutional codes, listed in the IEEE 802.15.7 standard,
are preferable for their low complexity, soft decodability,
flexible code rates and frame lengths. The Miller code provides
the much-needed run-length control for VLC and, at the same
time, serves as a (recursive) convolutional inner code for the
serial concatenation. The random interleaver connecting the
inner code and the outer code helps to break up the error
bursts (if any) and (together with the soft-iterative decoder) to
provide the renowned interleaving gain.
Let x be the Convolutional-Miller-coded and OOK-
modulated signal that goes through the VLC channel.
The receiver gets y = x + n, where n is additive white
Gaussian noise with variance σ 2 . As illustrated in Fig. 3,
the soft-iterative decoder runs the inner-subdecoder and
the outer-subdecoder multiple times to refine the decoding
results. The outer-subdecoder (for the convolutional code) can
make direct use of the existing soft-in soft-out (SISO) trellis
decoding algorithms, such as the BCJR algorithm [10]. The
inner-subdecoder (for the Miller code) can leverage a similar
decoding idea, but since it is based on OOK modulation
(rather than the prevailing bipolar modulation in conventional
wireless/wireline communication systems), modification must
be made for the existing algorithms to work.
1032 IEEE PHOTONICS TECHNOLOGY LETTERS, VOL. 28, NO. 9, MAY 1, 2016

III. S OFT D ECODING FOR M ILLER RLL C ODES Combining (3) and (4), we get
A. BCJR Algorithm in VLC γk (u k , s  , s) = Ak Bk exp(u k L a (u k ))
   
With the soft-iterative decoding framework in place, the pk2 + qk2 p k y p k + q k yq k
key now boils down to the soft inner subdecoder for the × exp − exp .
2σ 2 σ2
Miller code. Below we discuss how to modify the conventional  

BCJR algorithm [10] to provide maximal a posterior (MAP) γke (u k ,s  ,s)

decoding for the inner Miller code in VLC.
Let u = (u 1 , u 2 , · · · , u K ) be the the input sequence to the The LLR of the information bit u k at the output of the
Miller code, and let x = (( p1 , q1 ), ( p2 , q2 ), · · · , ( p K , q K )) modified BCJR Miller subdecoder, L(u k ), is thus given by
be the Miller coded sequence, where ( pk , qk ) are the  P(u k = 1|y)
branch output corresponding to the input u k at time k. Let L miller (u k ) = ln
P(u = 0|y)
y = (y1 , y2 , · · · , y K ) = ((y p1 , yq1 ), (y p2 , yq2 ), · · · , (y p K ,q K ))  k
be the noise-corrupted sequence at the input to the decoder. αk−1 (s  )γk (u k , s  , s)βk (s)
u =1
= 
k
The BCJR decoding algorithm includes the calculation of
the forward metric αk , the backward metric βk , and the transi- αk−1 (s  )γk (u k , s  , s)βk (s)
u k =0
tion metric γk , ∀k = 1, 2, · · · , K [10]. The forward and back- 
ward metrics, αk and βk , are not directly related to the channel αk−1 (s  )γke (u k , s  , s)βk (s)
u =1
= L a (u k )+ 
k
and are calculated in the same recursive  way as in a conven- . (5)
tional BCJR decoder [10]: αk (s) = αk−1 (s )γk (u k , s  , s), αk−1 (s  )γke (u k , s  , s)βk (s)
u k =0
 s 
βk−1 (s  ) = βk (s)γk (u k , s  , s), where s and s  stand for the
s
trellis states at time k and (k − 1). B. Iterative Decoding of RLL and FEC Codes
The key modification is in the calculation of the transition The soft-iterative decoder of Miller codes and convolutional
metric γk , which must account for the underlying channel and codes employs two BCJR subdecoders: the modified version
the modulation at hand. Let γk (u k , s  , s) denote the transition for the inner Miller code we just described, and the one for
probability of taking in information bit u k and going from the outer convolutional code. Recall that we have a serial
state s  at time (k − 1) to state s at time k. We have concatenation, so the outer convolutional code views the inner
Miller code (together with the actual VLC channel) as the
γk (u k , s  , s) = P(u k )P(yk (s  , s)|u k ), (1) “virtual channel”. The overall decoder proceeds as follows (we
use the superscript (l) to denote the number of iteration):
where yk (s  , s) denotes the noisy signal associated with the • Inner Decoding: In the lth iteration (l = 1, 2, · · · ),
s  → s transition at time k (see the trellis structure in Fig. 1), (l)
perform Miller BCJR decoding to get L miller (u k ) using (5),
and P(u k ) denotes the a prior probability of sending u k . Let (l)
L a (u k ) be the a prior log-likelihood ratio (LLR) of u k , which where the a prior information L miller,a (u k ) is set to 0 for
is set to be zero for the first iteration (for equal-probable l = 1, and is otherwise set to the extrinsic information,
(l−1)
sources), and is updated using the extrinsic information from L conv,e (u k ).
the outer convolutional subdecoder during each round of • Outer Decoding: The outer subdecoder performs the BCJR
iterative decoding. We have: decoding by always setting its a prior information to 0 and
taking in the LLR computed by the Miller code L (l) miller (u k ) as
P(u k = 1) P(u k = 1) the reception from the virtual channel. The BCJR algorithm
L a (u k ) = ln = ln , (2) (l)
P(u k = 0) 1 − P(u k = 1) produces an LLR information, L conv (u k ), from which a binary
decision of the overall system can be made. If additional
and hence decoding iteration is desired, the extrinsic can be computed
P(u k ) = Ak exp(u k L a (u k )), (3) via L (l) (l) (l)
conv,e (u k ) = L conv (u k ) − L miller (u k ) and be passed back
to the inner subdecoder for further refinement.
where Ak = 1+exp(L 1
a (u k ))
is a constant.
Recall that ( pk , qk ) is the coded output corresponding to C. EXIT Analysis
the transition (s  → s) at time k. The probably P(yk (s  , s)|u k )
We employ the useful tool of EXIT analysis, and show
in (1) is thus calculated by
an EXIT chart in Fig. 4 to show the convolutional-Miller
P(yk (s  , s)|u k ) concatenation is capable of providing iterative gain. In the plot,
    the x-axis represents the extrinsic mutual information Iconv,e
1 (y p − pk )2 1 (yq − qk )2 from the outer convolutional code, or, equivalently, the a priori
= √ exp − k 2 √ exp − k 2
2πσ 2 2σ 2πσ 2 2σ mutual information to the inner Miller code I Miller,a ; the
  
p +q
2 2
p k y p k + q k yq k y-axis represents the extrinsic mutual information I Miller,e
= Bk exp − k 2 k exp , (4) from the inner Miller code, or, equivalently, the a priori mutual
2σ σ2 information to the outer convolutional code Iconv,a . As we see,
 2 2  the two EXIT curves (depicting the transfer characteristics
y p +yq of the two component codes) match nicely in shape and in
where Bk = 1
2πσ 2
exp − k2σ 2 k is also a constant.
position, leaving a consistent, open tunnel between them, thus
(Note (3) and (4) is the major difference between the modified allowing the iterative decoder to continually refine its decision
BCJR algorithm and the conventional BCJR algorithm.) along the iterations until it converges to the desired point.
LU AND LI: ACHIEVING FEC AND RLL FOR VLC
Fig. 4. EXIT chart for the iterative decoding of Miller and the convolutional
code, E b /N0 = 2 dB.
Fig. 5. BER performance of convolutional-Miller concatenation with soft-
iterative decoding. A rate-1/2 convolutional code (5, 7) is used, the block
sizes 150 and 3000 bits.
IV. S IMULATION R ESULTS
We now evaluate the proposed joint convolutional-Miller
scheme with soft-iterative decoding. Two types of convolu-
tional codes are tested: rate-1/2 (5, 7) code and punctured
rate-1/3 (5, 7) and (15, 17) codes. Two reference schemes are
pulled in for comparison: Ref(i) – the same convolutional-
Miller concatenation using hard decoding, and Ref(ii) –
RS-4B6B concatenation proposed in [7], which consists of an
RS (15, 7) code, an 4B6B RLL codes, and a multi-candidate
decoder.
Fig. 5 demonstrates effectiveness of the soft-iterative par-
adigm. We observe the renowned interleaving gain phenom-
enon, and in both cases, 4 iterations of soft decoding brings
up more than 3 dB gains over sequential hard decoding
at the bit error rate (BER) of 10−4 . In Fig. 6, we com-
pare the proposed schemes with the two references schemes.
All the convolutional-Miller concatenation schemes tested here
(soft decoding or hard decoding) use a (5, 7) or (15, 17) outer
convolutional code with data length 3000 that is punctured
to rate 23 , resulting in an overall code rate of 23 12 = 13 . For
Ref(ii) the RS-4B6B scheme [7], 16 candidates are preserved
6 = 45 ,
7 4 14
in the decoding process, and the overall code rate is 15
1
which is slightly smaller than 3 . It is encouraging to see that
the proposed convolutional-Miller concatenation with soft-
iterative decoding clearly outperform these existing schemes
(in addition to providing a slightly code rate).
The encoding and the decoding complexity of our scheme
is linear to the code length. The encoder can run extremely
fast via linear shift-register (LSR) circuits (which is much
faster than the look-up table encoder used in, for example,
4B6B code). The decoder performs the BCJR algorithm
for the FEC and the RLL code. It is recognized that the
1033
Fig. 6. BER comparison between rate-1/3 Convolutional-Miller codes
(using punctured convolutional code (5, 7) and (15, 17)) and rate-14/45
RS-4B6B code [7].
BCJR algorithm has about 3 times the complexity of a Viterbi
algorithm for the same code, but the BCJR algorithm is
advantageous in that it does not involve a trace-back procedure
as with the Viterbi algorithm and, best of all, it lends itself
to a very natural and efficient parallelization (which signifi-
cantly expedite the decoding process). The overall decoding
complexity increases linearly with the number of iterations,
and we have shown that 4 iterations are sufficient to extract
most of the coding gain.
V. C ONCLUSION
We have presented a convolutional-Miller concatenated
scheme as a strong FEC/RLL candidate for VLC. We showed
that Miller codes not only provide a high bandwidth efficiency
than Manchester codes, but also lend themselves to a serial
concatenation that can provide substantial interleaving gains
using soft-iterative decoding. The proposed scheme presented
in this letter are for the OOK-modulated systems. An inter-
esting future direction would be to consider systems with
advanced modulation to achieve higher data transmission rate.
R EFERENCES
[1] S. Rajagopal, R. D. Roberts, and S.-K. Lim, “IEEE 802.15.7 Visible
light communication: Modulation schemes and dimming support,” IEEE
Commun. Mag., vol. 50, no. 3, pp. 72–82, Mar. 2012.
[2] A. Jovicic, J. Li, and T. Richardson, “Visible light communication:
Opportunities, challenges and the path to market,” IEEE Commun. Mag.,
vol. 51, no. 12, pp. 26–32, Dec. 2013.
[3] IEEE Standard for Local and Metropolitan Area Networks—Part 15.7:
Short-Range Wireless Optical Communication Using Visible Light,
IEEE Standard 802.15.7, 2011, pp. 248–271.
[4] S. Kim and S.-Y. Jung, “Modified Reed–Muller coding scheme made
from the bent function for dimmable visible light communications,”
IEEE Photon. Technol. Lett., vol. 25, no. 1, pp. 11–13, Jan. 1, 2013.
[5] J. Kim and H. Park, “A coding scheme for visible light communication
with wide dimming range,” IEEE Photon. Technol. Lett., vol. 26, no. 5,
pp. 465–468, Mar. 1, 2014.
[6] S. H. Lee and J. K. Kwon, “Turbo code-based error correction scheme
for dimmable visible light communication systems,” IEEE Photon.
Technol. Lett., vol. 24, no. 17, pp. 1463–1465, Sep. 1, 2012.
[7] H. Wang and S. Kim, “New RLL decoding algorithm for multiple
candidates in visible light communication,” IEEE Photon. Technol. Lett.,
vol. 27, no. 1, pp. 15–17, Jan. 1, 2015.
[8] M. Hecht and A. Guida, “Delay modulation,” Proc. IEEE, vol. 57, no. 7,
pp. 1314–1316, Jul. 1969.
[9] M. Hecht and A. Guida, “Evaluation of the noise effects on visible light
communications using Manchester and Miller coding,” in Proc. 12th Int.
Conf. Develop. Appl. Syst., Suceava, Romania, May 2014, pp. 85–89.
[10] S. Lin and D. J. Costello, Error Control Coding, vol. 31. London, U.K.:
Pearson Education, 2013, ch. 12, pp. 1211–1216, no. 8.