2014 IEEE 15th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

Neural Networks-Based Turbo Equalization Of a

Satellite Communication Channel

Hasan Abdulkader∗ , Bouchra Benammar∗ , Charly Poulliat∗ , Marie-Laure Boucheret∗ and Nathalie Thomas∗
∗
University of Toulouse, INPT-ENSEEIHT/IRIT
Email: hasan.abdulkader, bouchra.benammar, charly.poulliat, marie-laure.boucheret, nathalie.thomas}@enseeiht.fr

Abstract—This paper proposes neural networks-based turbo

equalization (TEQ) applied to a non linear channel. Based on a
Volterra model of the satellite non linear communication channel, Channel Digital

HPA
Coding mod
Receiver
we derive a soft input soft output (SISO) radial basis function
(RBF) equalizer that can be used in an iterative equalization Channel
in order to improve the system performance. In particular, it is
shown that the RBF-based TEQ is able to achieve its matched Fig. 1. Model of the communication system.
filter bound (MFB) within few iterations. The paper also proposes
a blind implementation of the TEQ using a multilayer perceptron
(MLP) as an adaptive model of the nonlinear channel. Asymptotic Besides, it can be easily implemented using detected symbols
analysis as well as reduced complexity implementations are also at the equalizer output. These symbols are fed back to the
presented and discussed. equalizer input allowing to reduce the size of the RBF hidden
layer. We also propose a blind equalization scheme. In [10] the
I. I NTRODUCTION transmission channel has been estimated using linear filter. In
our case, non linear tools are needed to model the non linear
Essential satellite communication requirements are spectral channel. We propose to estimate the HPA response using a
and power resources. High efficient modulations are often used MLP neural network.
to enhance the capacity of communication systems with limited The paper is organized as follows: first we describe the satellite
bandwidth and power resources. On board a satellite, a HPA is non linear channel, then we introduce the turbo equalizer in
used to amplify the incoming signal to be sent back to earth. section II. In section III, we derive the RBF-Based TEQ in the
It is desirable to operate near the HPA saturation in order to case of a Volterra non linear channel and its low complexity
provide sufficient power to the transmitted signal. However, implementation. The blind equalization structure is briefly
the HPA performs non linearly near its saturation point and described in section IV. In section V, simulation results are
thus becomes a bottleneck to the system capacity. Constant presented and analysed using EXIT chart. Finally conclusions
modulus modulations are frequently used in such a channel to are drawn in section VI.
reduce the impact of non linear amplification.
Turbo equalization is a scheme attempting to iteratively miti- II. S YSTEM MODEL
gate and overcome the ISI incurred in the channel [1]–[3]. It
is based on the turbo-principle widely used in modern digital In this paper, we consider a non linear communication
communication receivers. In the literature, many schemes channel given by a Volterra filter [11], [12]. Figure 1 illustrates
apply this principle to jointly and iteratively decode, equalize the proposed communication system model. Independent and
or demodulate received signals. Optimal MAP and suboptimal identically distributed bits bi are first encoded by a channel
MMSE algorithms have been widely studied for linear ISI encoder, interleaved and as coded bits ci fed to a digital
channels equalization [1]–[3]. modulator that maps blocks of log2 (M ) coded bits into a
Neural networks (NN) are efficiently used to carry out complex complex symbol x. Let M be the cardinality of the digital
and non linear problems [4]. NN equalization of communica- modulation constellation. After pulse shaping, the symbol
tion channels has been intensively investigated [5]–[7] using sequence is sent to a satellite transponder consisting of an HPA
different architectures such as MLP, RBF, recurrent neural and input/output filters. The amplified signal is transmitted
network (RNN), etc. In [8] RBF neural networks has been downlink to the receiver with an additive Gaussian noise and is
applied as SISO device to equalize linear channels. Based matched-filtered before being sampled at the baud rate. Many
on the optimal decision theory, it showed good performance models of the non linear satellite channel have been proposed
allowing a reduced complexity implementation. Later, it has in literature; some as a complex gain [4], [5], [13] and others as
been used for turbo equalization of linear channels and has a non linear Volterra series expansion [14]. This study assumes
shown good performance for different modulation techniques a Volterra series-based channel model. As such, the received
[9], [10]. symbols can be written as follows:
In this paper, we show how RBF-Based equalization can L−1
X
be successfully applied to non linear channels equalization y(k) = h0 x(k) + hi x(k − i)+
when Volterra-based non linear channel models are considered. i=1
The RBF-Based TEQ can be used to implement the optimal L−1
(1)
X
∗
decision rule given in [8]–[10]. We use the reduced complex- hi,j,l x(k − i)x(k − j)x (k − l) + n(k)
ity algorithm for its interesting performance and simplicity. i,j,l=0

978-1-4799-4903-8/14/$31.00 ©2014 IEEE 494

 2014 IEEE 15th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

T
of symbols x = [x(k), . . . , x(k − L − m + 2)] . The above
Decod. equation can be rewritten in the form p (x|y) ∝ p (y|x) p(x)
EQ.
since y corresponds to the vector x. Henceforth, we use
equality instead of ∝ since the computation of LLR eliminates
other constant multiplicative factors. We are interested in
estimating x(k − d) given by the rule:
X
p (x(k − d)|y) = p (y|x) p(x̃) (3)
Fig. 2. Structure of the turbo equalizer.
x̃

with x̃ = x\x(k −1) is the vector x without the entry x(k −1).
where n is an additive white Gaussian noise (AWGN) with Probability p (y(k)|y(k)) follows normal distribution with zero
variance σn2 and hi , hijl kernel coefficients of first and third mean and variance σn2 .
order taking into account all chain filters. xk is the transmitted 1 2 2
symbol at time k, L−1 is the channel memory and (.)∗ stands p (y(k)|y(k)) = √ e−|y(k)−y(k)| /σn (4)
for the conjugate of complex signal. The second and third RHS 2πσ 2
terms of equation 1 represent respectively the linear ISI and In the following we create the link to the RBF structure. The
the non linear ISI generated by the nonlinear amplification of expression of the RBF output is:
the HPA combined with chain filters memory. SISO equaliz- nh
ers/decoders proved efficiency by accepting a priori probability
X
yrbf = φ(y, ci )wi (5)
of received bits and generating new probabilities, a posteriori, i=1
depending on knowledge of the channel response/encoder
structure. A turbo equalizer iteratively exchanges a priori where wi a weight of the output layer connected to hidden
information with a decoder (see figure 2 ). Both equalizer and neuron i and φ(.) a radial function similar to the Gaussian
decoder enhance the bit probability as the number of iterations function, 2
increases until a convergence is reached. The probability of an φ (y, ci ) = e−|y−ci | /σ (6)
encoded bit ci is measured by the log likelihood ratio (LLR)
with σ a parameter characterizing the width of φ. Identifying
defined by equation (2), [1]–[3] as:
equations (3)-(4) with equations (5)-(6), it becomes apparent
p(ci = 0) that choosing c = y and wi = p(x̃i ) allows the RBF to model
L(ci ) = ln (2) the optimal decision rule.
p(ci = 1)
Henceforth the RBF has a priori computed centres given the
Hence probabilities of ci can be computed by: satellite model and the output weights are computed iteratively
eL(ci ) 1 given the a priori LLRs (extrinsic of the decoder). In the first
p(ci = 0) = , p(ci = 1) = iteration, the a priori LLRs are set to zero or equivalently the a
1 + eL(ci ) 1 + eL(ci ) priori probabilities set to p(ci = ∓1) = 0.5. The computation
The equalizer receives a priori LLRs La1 , uses them together of weights follows the rule given by:
with the received symbols to generate extrinsic LLRs Le1 .
On the other hand, the decoder uses La2 together with the wi = p (y i (k), y i (k − 1), . . . , y i (k − m + 1)) (7)
channel code to generate extrinsic coded bits LLRs Le2 and m−1
Y L+m−2
Y
information bits estimates. In this paper, we will consider wi = p (y i (k − j)) = p (xi (k − j)) (8)
trellis based codes using BCJR algorithm [15]. j=0 j=0

m −1
L+m−2 nY
!
III. RBF-BASED T URBO E QUALIZER Y
wi = p (ci (k − nm j − u)) (9)
Radial basis function neural networks have gained much j=0 u=0
interest in engineering applications. They are used in system
modelling and identification since they have a universal ap- with nm = log2 (M ) the number of bits per constellation sym-
proximation property [16]. They are also used in classification bol. In [10], the implementation of the RBF-Based equalizer
and equalization where they show optimum decision property is realized by M independent RBFs. The RBF i computes
[10]. The RBF is structured in three layers: an input layer, p (x(k − d) = si |y), i = 1, . . . , M where:
a hidden layer and an output layer. The hidden layer of nh p (x(k − d) = si |y) = p (si )
X
p (y|x) p (x̃)
neurons has radial activation functions centred around centres
x̃
c.
Let m be the equalizer memory. The expression of the equal- The number of hidden neurons in each RBF is given by
izer input vector is y = [y(k), y(k − 1), . . . , y(k − m − 1)]
T nh = M L+m−2 . Figure 3 illustrates the structure of one branch
where y(k) are the noisy received symbols. From equation (1), of the RBF-Based equalizer. A decision of the maximum
we can write the Bayes rule as [8]: p (x(k − d) = si |y) defines the best estimation of transmitted
symbol x(k − d) when a Hard-Output equalizer is used.
p (y|y) ∝ p (y|y) p (y) Finally the extrinsic LLRs, noted by Le1 , will be computed by
the following rule:
where p (y|y) is the probability of the vector y given y =
T
[y(k), y(k − 1), . . . , y(k − m + 1)] which is is the vector of p (ci = 0|y)
Le1 (ci ) = ln − La1 (ci ) (10)
noise-free channel output corresponding to a transmitted vector p (ci = 1|y)

495
 2014 IEEE 15th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

Fig. 3.
.
.
.
} .
.
.

The structure of a branch of the RBF equalizer.

(a) Blind receiver

The computation of ln (p (ci = 0|y)) − ln (p (ci = 1|y)), will

be detailed as follows:
nm
X
p (ci = 0|y) = p (x(n − d) = sj |y) δij
j=1

where δij = 1, iff ci of sj is equal to 0. It can be also rewritten

as:
nh
nm
!
X X
p (ci = 0|y) = p(x(n − d) = sj ) p (y|y) p(x̃l ) δij
j=1 l=1
(11) (b) Blind MLP
The computational complexity of the extrinsic LLR calculation
Fig. 4. Estimation of the satellite channel in the receiver.
can be reduced by using the generalized Jacobian algorithm
[10]. Computation of ln p (ci = 0|y), for instance, becomes:
 !  IV. B LIND T URBO E QUALIZER
nm nh
X X ky−yl k2
ln(p (ci = 0|y) = ln  e− 2σ2 eln(p(xl )) δij  In blind equalization the centres of the RBF are iteratively
j=1 l=1
updated based on an estimation of the channel model. Since
(12) the satellite channel is non linear, it should be modelled by a
where ln (p(xl )) = ln (p (x̃l )) + ln (p (x(n − d) = sj)). non linear tool like a NN. We choose to model the channel
Similarly, we can write: by a multilayer perceptron MLP with one input, one hidden
 and one output layer. A tapped-delay line is used at the MLP
nm nh
!  input to add a memory to the MLP (cf. figure 4(b)). Centres of
X X ky−yl k2 0
ln(p (ci = 1|y) = ln  e− 2σ2 eln(p(xl )) δij  the RBF-Based TEQ are computed using the MLP model. On
j=1 l=1 line, MLP is implemented into the TEQ and trained using soft
(13) symbols x̂ estimated by the decoder output, after interleaving,
0
where δij = 1, iff bit ci of symbol sj is equal to 1. Then the as input of the MLP (Figure 4(a)). The output of the MLP
Jacobian algorithm J(. . . , J(θ3 , J(θ2 , θ1 ))) will be applied to ŷ is compared to the channel output y to produce the error
compute equations (12)-(13) with: signal e = y − ŷ, which is used by the backpropagation (BP)
algorithm to adjust the MLP weights. Learning rules use a
J(θ2 , θ1 ) = ln eθ2 + eθ1


small learning rate µ. After each iteration the centres of the
  RBF-TEQ are updated using the new MLP coefficients. The
= max (θ2 , θ1 ) + ln 1 + e−|θ2 −θ1 | (14) output of the MLP is computed by: The following equations
present simplified learning rules of the hidden layer parameter
In [10], authors proposed to reduce the complexity of the wh and output layer parameters wo respectively:
RBF equalizer using feedback from the output of the RBF
into the input layer. The feedback is used to select a subset wo = wo + µe.yh
of hidden neurons involved into the output computation. The wh = wh + µδ.x̂
reduced complexity algorithm considers the first two symbols
with δ = e.wo .f 0 the error of BP algorithm computed at the
of the vector x known and detected on the RBF output. Finally,
hidden layer, and f 0 is the derivative of the activation function.
the number of hidden neurons used to compute each branch
equals M L+m−F B−2 instead of M L+m−2 , with F B being the
number of symbols fed back. For example, given simulation
V. S IMULATION R ESULTS
parameters of section V; L = 4, m = 3, F B = 2 and M = 8,
the feedback allows a reduction of the hidden neurons to 83 We consider the Volterra filter as a model of the satellite
neurons for each RBF instead of 85 . nonlinear channel, with kernel coefficients given in table 1

496
 2014 IEEE 15th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

(a) Performance of blind equalizers

(a) EXIT charts for RBF equalizers

(b) EXIT chart details (b) Performance of non blind equalizer

Fig. 5. EXIT Chart curves of the RBF-Based TEQ. Fig. 6. Performance of RBF-Based TEQ.

TABLE I. C OEFFICIENTS OF VOLTERRA KERNELS

1st order 3rd order determines the value on the vertical axis. The computation of
h0 = 0.8529 + 0.4502j h001 = 0.0979 − 0.0979j the mutual information takes into account the extrinsic of the
h1 = 0.0881 − 0.0014j h002 = 0.1091 − 0.0615j
h2 = −0.0336 − 0.0196j h003 = −0.1119 − 0.0252j
equalizer and the decoder only. EXIT chart allows to analyse
h3 = 0.0503 + 0.0433j h110 = −0.0280 − 0.0475j the performance of the iterative system. Figure 5 shows the
Eb
h330 = 0.0503 − 0.0503j EXIT chart within different conditions of noise ( N 0
= 3, 6 and
9dB). Curves show the equalizer mitigation of both the linear
and the nonlinear ISI since the equalizer output IE increases
[11], [12]. The memory depth of the channel is L − 1 = 3, and monotonically when the a priori IA increases.
we consider the equalizer of memory m = 3 and a delay d = 2. Figure 6(a) shows the performance of the blind implementation
Thus, the number of hidden neurons in each RBF is equal to of the reduced RBF-Based turbo equalizer. Centres of the
M L+m−2 neurons, with M = 8 the number of symbols in the RBF are deduced using the MLP. At the beginning, the MLP
modulation 8-PSK. The system uses a convolutional encoder is sufficiently trained to approximate the channel model. It
with polynomials given in octal by (5, 7) with rate = 0.5 and though continues to be trained inside the receiver blindly
a BCJR decoder with trellis termination [15]. using soft estimation of transmitted symbols together with the
Extrinsic information transfer (EXIT) chart [1] and [2] are channel output. A small learning rate, of order 1e−4 , is used
used to analyse the performance of the equalizer. An EXIT in learning rules to prevent the divergence of MLP. At the end
Chart [17] plots the mutual information of the equalizer versus of each iteration, the RBF centres are updated according to
the mutual information of the decoder which is considered as the new weights of MLP. The BER, in fig. 6(a), is close to the
a priori. The decoder output determines the value of the hori- matched filter bound after few iterations. Although 2 iterations
zontal axis of the EXIT chart and the output of the equalizer are sufficient to get close to the MFB, a higher iteration number

497
 2014 IEEE 15th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC)

Eb
still remains useful to enhance the BER for N 0
values smaller [13] A. A. M. Saleh, “Frequency-independent and frequency-dependent non-
then 5dB. By comparing the two parts of figure 6, It appears linear models of twt amplifiers,” Communications, IEEE Transactions
on, vol. 29, no. 11, pp. 1715–1720, November 1981.
the blind equalizer has almost as good performance as the non
blind equalizer. [14] S. Benedetto, E. Biglieri, and R. Daffara, “Modeling and performance
evaluation of nonlinear satellite links-a volterra series approach,” IEEE
Transactions on Aerospace and Electronic Systems, vol. AES-15, no. 4,
VI. C ONCLUSION pp. 494 –507, july 1979.
[15] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linear
This paper presents an RBF-Based turbo equalizer for codes for minimizing symbol error rate (corresp.),” Information Theory,
satellite non linear communication channel applications. We IEEE Transactions on, vol. 20, no. 2, pp. 284–287, Mar 1974.
described the turbo equalisation and the blind turbo equaliza- [16] J. Park and I. W. Sandberg, “Universal approximation using radial-basis-
tion for the non linear channel using neural networks. The non function networks,” Neural Computation, vol. 3, no. 2, pp. 246–257,
linear channel is modelled by a multilayer perceptron trained 1991.
on line using soft symbols and the channel output. The deriva- [17] J. Hagenauer, “The EXIT chart-Introduction to extrinsic information
transfer in iterative processing,” in Proceeding 12th EUSIPCO, 2004,
tion of soft input soft output representing extrinsic versus a pp. 1541–1548.
priori has been done. The reduced complexity implementation
has been presented and results outlined in the simulation.
We plotted EXIT chart curves of the RBF-TEQ. They showed
a monotonic increase of the extrinsic when a priori increases.
This proves the ability of the RBF-equalizer to reduce ISI and
to achieve the matched filter bound performance if iterative
detection-decoding is used. The blind scheme of the reduced-
complexity of RBF-Based TEQ, applied to the Volterra non
linear channel, has good performance and BER is close to the
MFB of the full RBF TEQ. BER converges to the MFB in
reached after almost 2 iterations.

R EFERENCES
[1] M. Tuchler and A. Singer, “Turbo equalization: An overview,” Informa-
tion Theory, IEEE Transactions on, vol. 57, no. 2, pp. 920–952, 2011.
[2] M. Tuchler, R. Koetter, and A. Singer, “Turbo equalization: principles
and new results,” Communications, IEEE Transactions on, vol. 50,
no. 5, pp. 754–767, 2002.
[3] R. Koetter, A. Singer, and M. Tuchler, “Turbo equalization,” Signal
Processing Magazine, IEEE, vol. 21, no. 1, pp. 67–80, 2004.
[4] H. Abdulkader, F. Langlet, D. Roviras, and F. Castanie, “Natural
gradient algorithm for neural networks applied to non-linear high power
amplifiers,” International Journal on Adaptive Control and Signal
Processing, vol. 16, pp. 557–576, 2002.
[5] S. Bouchired, M. Ibnkahla, D. Roviras, and F. Castanie, “Equalization
of satellite umts channels using rbf networks,” in Personal, Indoor and
Mobile Radio Communications, 1998. The Ninth IEEE International
Symposium on, vol. 3, 1998, pp. 1250–1254 vol.3.
[6] B. Lu and B. Evans, “Channel equalization by feedforward neural
networks,” in Circuits and Systems, 1999. ISCAS ’99. Proceedings of
the 1999 IEEE International Symposium on, vol. 5, 1999, pp. 587–590
vol.5.
[7] G. Kechriotis, E. Zervas, and E. Manolakos, “Using recurrent neural
networks for adaptive communication channel equalization,” Neural
Networks, IEEE Transactions on, vol. 5, no. 2, pp. 267–278, 1994.
[8] S. Chen, B. Mulgrew, and S. McLaughlin, “Adaptive bayesian equalizer
with decision feedback,” Signal Processing, IEEE Transactions on,
vol. 41, no. 9, pp. 2918–2927, 1993.
[9] M. Yee, T. Liew, and L. Hanzo, “Burst-by-burst adaptive turbo-coded
radial basis function-assisted decision feedback equalization,” Commu-
nications, IEEE Transactions on, vol. 49, no. 11, pp. 1935–1945, 2001.
[10] M.-S. Yee, B. Yeap, and L. Hanzo, “Radial basis function-assisted turbo
equalization,” Communications, IEEE Transactions on, vol. 51, no. 4,
pp. 664–675, 2003.
[11] B. Benammar, N. Thomas, C. Poulliat, M.-L. Boucheret, and M. Dervin,
“On linear mmse based turbo-equalization of nonlinear volterra chan-
nels,” in Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE
International Conference on, 2013, pp. 4703–4707.
[12] D. Ampeliotis, A. Rontogiannis, K. Berberidis, M. Papaleo, and
G. Corazza, “Turbo equalization of non-linear satellite channels using
soft interference cancellation,” in Advanced Satellite Mobile Systems,
2008. ASMS 2008. 4th, 2008, pp. 289–292.

498