Nonlinear Subband Spline Adaptive Filter
Chang Liu, Xueliang Liu
, Zhi Zhang and Xiao Tang
School of Electronic Engineering, Dongguan University of Technology, No.1 DaXueRoad, Dongguan, China
chaneaaa@163.com, liuxueliang83@163.com ,{zhangz, tangx}@dgut.edu.cn
Keywords: Subband adaptive filter, Nonlinear spline adaptive filter, adaptive filter algorithm, system identification.
Abstract: It has been reported that a novel class of nonlinear spline adaptive filter (SAF) obtains some advantages in
modeling the nonlinear systems. In this paper, a nonlinear subband structure based on the spline adaptive
filters, called subband spline adaptive filter (SSAF) is presented. The proposed structure is composed of a
series of subband spline filters, each one comprises a linear time invariant (LTI) filter followed by an
adaptive look-up table (ALUT). In addition, the computational complexity is also analyzed. Some exper-
imental results in the context of the nonlinear system identification demonstrate the robustness of the
proposed structure.
1 INTRODUCTION
In many practical engineering applications, the
nonlinear system identification is an important and
difficult task. Much well-established theory for
linear system identification is unavailable when it
comes to nonlinear case, so techniques to model the
nonlinear behavior have been received more attenti-
on in recent decades (Mathews, 2000). In order to
model the nonlinearity, several adaptive nonlinear
structures have been introduced. Truncated Volterra
adaptive filter (VAF) (Schetzen, 1980) is one of the
most popular nonlinear model. However, its
computational complexity be-comes huge with the
increase of the nonlinear order. Neural Networks
(NNs) (Haykin, 2009) can make a good des-cription
of the nonlinear relation between the input signal
and the current output adequately, but it suffers from
a large computational cost and diffi-culties in on-line
adaptation. Block-oriented archi-tecture (Giri, 2010),
including the Wiener model, Hammers-tein model
and cascade model, originates from the different
combination of the linear time invariant (LTI) filters
and memoryless nonlinear functions. Recently,
Scarpiniti et al. has proposed a novel class of
nonlinear spline adaptive filter (SAF) structure,
which also contains the Wiener spline filter
(Scarpiniti, 2013), the Hammerstein spline filter
(Scarpiniti, 2014) and the cascade spline filter
(Scarpiniti, 2015). In this kind of structure, the
nonlinearity is modelled by a spline function which
can be repress-ented by the adaptive look-up table
(ALUT), and the linear time invariant (LTI) filter is
used for determining the memory effect. Both the
control points belonging to ALUT and the
coefficients of the LTI are adapted by using the
sophisticated adaptive algorithms such as the least
mean square (LMS) algorithm, normalized least
mean square (NLMS) algorithm and affine
projection algorithm (APA).
In this paper, extending the subband idea into the
spline adaptive filter (SAF), a nonlinear subband
spline structure, called subband spline adaptive filter
(SSAF) is proposed. Each subband spline filter is
composed of a LTI filter followed by an ALUT.
Then main advantage of the proposed subband
model is its improved convergence performance
because of the decorrelating properties with no sign-
ificant computational increasement.
2 SPLINE ADAPTIVE FILTER
The block diagram of a SAF is shown in Fig.1,
which consists of an adaptive finite impulse respo-
nse (FIR) filter followed by a nonlinear network. In
the nonlinear network, the spline interpolater,
connected behind the adaptive LUT, determines the
number and the spacing of control points (knots)
contained in the LUT.
The input of the SAF at time n is
()
x
n , ()
s
n
represents the output of the linear networks which is
given by
Liu, C., Liu, X., Zhang, Z. and Tang, X.
Nonlinear Subband Spline Adaptive Filter.
In 3rd International Conference on Electromechanical Control Technology and Transportation (ICECTT 2018), pages 163-167
ISBN: 978-989-758-312-4
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
163
() ()(),
T
s
nnn= wx (1)
()
x
n
()
s
n
()
y
n
Figure 1: Block diagram of SAF.
where
01 1
() [ , , , ]
T
B
nww w
=w L represents the weight
vector of the FIR filter with length
B
, and
[(),( 1), ,))]((1
T
sn nns snB−−+=x L is the input vector
of the linear network.
With reference to the spline interpolation scheme
(Guarnieri, 1999), the output of the whole system
()yn and ()
n can be related by a local polynomial
function
()
in
u
ϕ
, which depends on the span index i
and the local parameter
u . The two parameters are
defined as follows
() () ,
n
usnxsnx Δ
⎢⎥
⎣⎦
(2)
() ( 1)2,
n
isnxQ+
⎢⎥
⎣⎦
(3)
where
x
Δ is the uniform space between two control
points for the function
()
in
u
ϕ
, Q is the total number
of control point and
⎢⎥
⎣⎦
denotes the floor operator.
The output of the whole system can be expressed as
,
() ( ) ,
T
in n in
yn u==φ uCq
(4)
where
C is the 44× spline basis matrix if the three-
order spline function is used. Two suitable types of
spline basis are B-spline and Catmull-Rom (CR)
spline (
Scarpiniti, 2013) which are given by
13 31 13 31
3 630 2 54-1
11
,
30 30 10 1 0
62
1410 0200
BCR
CC
−− −−
⎡⎤
⎢⎥
−−
⎢⎥
= =
⎢⎥
−−
⎢⎥
⎣⎦
(5)
where
32
[, ,,1]
nnn
T
n
uuu=u ,
,123
[, , , ]
T
in i i i i
qq q q
++ +
=q is the
control point vector and superscript
()
T
denotes
transposition.
3 SUBBAND SPLINE ADAPTIVE
FILTER (SSAF)
Fig.2 shows the block diagram of the proposed
SSAF.
0
()Φ denotes the unknown nonlinear system
which generates the desired signal
[
]
0
)(()d
x
nn = Φ
()vn , where ()
x
n is the system input. ()vn is the
background noise, assumed to be zero mean ,and
independent of
()
x
n , its variance is
2
v
σ
. The input
signal
()
x
n
and desired signal
()dn
are partitioned
into
M
subband signals
()
m
x
n
and
()
m
dn
via the
analysis filters
()
m
Hz, 0,1, , 1.mM=−L The subband
signals,
()
m
dn and ()
m
x
n are critically decimated to
a lower sampling rate commensurate with their
bandwidth. We use the variable
n to index the
original sequences, and
k to index the decimated
sequence for all subband signals. The decimated
output for the
thm subband filter can be computed as
0
()dn
()dn
()vn
0
()Hz
1
()Hz
1
()
M
Hz
0
()dk
1
()dn
1
()dk
1
()
M
dk
+
0
()
x
n
0
()Hz
1
()Hz
1
()
x
n
1
()
M
x
n
1
()
M
dn
()
x
n
0
()xk
1
()
x
k
1
()
M
x
k
0
()kw
Σ
Σ
-
0
()yk
0
()ek
Σ
Σ
-
-
1
()yk
1
()
M
yk
1
()ek
1
()
M
ek
0
()Fz
1
()
F
z
1
()
K
Fz
Σ
0
()en
1
()en
1
()
M
en
()en
0
0, 0,
(, )
kk
x
q
i
φ
()Sn
o
w
1
1, 1,
(, )
kk
x
q
i
φ
1
()
M
k
w
1
1, 1,
(, )
M
Mk Mk
xq
−−i
φ
0
()
s
k
1
()
s
k
1
()
M
s
k
0
()Φ
0
(, )xqφ
()dn
1
()
M
Hz
M
M
M
M
M
M
1
()kw
M
M
M
Figure 2: Block diagram of SSAF.
,,,
() ( )= ,
mm
T
mimkmkik
yk u= φ uCq (6)
where
3
,,
2
,,
[,,,1]
T
mk mk mkmk
uuu=u
,
,(1)(2)
[,=, ,
mmm m
ik i i i
qq q
++
q
(3)
]
m
i
T
q
+
represents the
thm
subband control
point vector at the decimated time
k
. The
corresponding subband local parameter and subband
span index
m
i are defined as
,
() () ,
mkmmmm
uskxskx Δ
⎢⎥
⎣⎦
(7)
() ( 1)2,
mm m m
iskx Q+
⎢⎥
⎣⎦
(8)
where
m
Q
is the total number of the control points
for the
thm
subband LUT and
m
x
Δ is the uniform
space, which can be selected to different values for
0,1, , 1.mM=−L
()
m
s
k is the output of the
thm
subband linear combiner which is given by
() () (),
T
mmm
s
kkk= wx (9)
where
,0 ,1 , 1
[(),(),,((]))
T
mmmmB
wkwk w kk
=w L
denotes
the weight vector of the subband FIR filter with
length
.
B
( ) [ ( ), ( 1), , ( 1)]
T
mm mm
ksksk skB−−+=s L
is
the subband input vector.
The subband output error can be expressed as
ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation
164
,
() () ()
=() ( ),
m
mmm
mimk
ek dk yk
dk u
=−
φ
(10)
It is realized that the adaptation in each subband
is carried out independently as shown in Fig.2.
Therefore, the updating equations of the linear and
nonlinear networks for each subband can be derived
by minimizing the cost function
()
,,
,=
m
mmkik
θ
wq
{
}
2
=()
m
Ee k , where
{
}
E
is the expectation
operator. For analytical simplicity, using the
instantaneous error instead of the expectation, we
get
()
2
,,
,= 0,1,,1,
m
mmkik m
ekm M
θ
() = wq L (11)
and taking the partial derivative
()
,,
,
m
mmkik
θ
wq with
respect to
,mk
w
, we have
()
,,
,
,
,,
,
,
()
()
=2
()
()
2
=()(),
m
m
m
mmkik
imk
mk
m
m
T
mk mk m
mimkm
m
u
u
s
k
ek
usk
k
ek u k
x
θ
− ()
∂∂
− ()
Δ
wq
φ
w
w
φ x
(12)
where
,
()
m
imk
u
φ
is the partial derivative of the local
activation function for the
thm subband,
,
()
m
imk
u
φ
,,
=,
m
T
mk i k
uCq
&
1
,,,
2
[3 ,2 ,1,0]
mkmk
T
mk
uu=u
&
, so the updating
equation of the linear networks for each subband can
be given as
w
,,
(1)=() (),
m
T
mm mmkikm
m
kkek k
x
μ
++ ()
Δ
ww uCqx
&
(13)
where
w
μ
is the step-size for the linear network
adaptation.
For the subband nonlinear networks, the
derivative calculation of
()
,,
,
m
mmkik
θ
wq to
,
m
ik
q
can
be defined by
()
,,
,
,
,,
,
()
=2 2() ,
m
m
mm
mmkik
imk
T
mmmk
ik ik
u
ek ek
θ
− () =
∂∂
wq
φ
Cu
qq
(14)
The updating equation of the nonlinear networks
for each subband can be written as
,1 , q ,
=(),
mm
T
ik ik m mk
ek
μ
+
+qq Cu (15)
4 COMPUTATIONAL
COMPLEXITY
Note that (13) and (15) are the updating equations of
the LMS algorithm for the proposed SSAF. For each
iteration, only four control points for each subband
are changed because of the local behavior of the
spline function. This leads to a large computational
savings. The computational complexity of the pro-
posed SSAF solution is mainly evaluated in terms of
the number of multiplications per sample.
Note that
for each subband , the control points of the LUT and
the weights of the FIR filter are updated every M
samples due to the critical sampling. Considering
that there are M subband signals par-ticipating in the
adaptation, it requires 2B+1 multi-plications for the
linear updating equation (13) and B multiplications
for the output estimation of the linear network. For
the spline output calculation and adap-tation, like the
conventional SAF (Scarpiniti, 2013), we take into
account of the repetitive appearance of the terms
,
m
ik
Cq ,
,
T
mk
Cu
in (6), (13) and (15), it only needs 4K
q
multiplications by the date reuse of the past
computations, where K
q
(less than 16) is the
constant which can be defined with reference to the
implementation spline structure in (Guarnieri, 1999).
In addition, the subband input signal and desired
signal partition needs
2
M
P
multiplications, where
P
is the length of the analysis and synthesis filters.
For error signal synthesis, it needs
M
P
multiplications. Therefore, compared with the SAF
scheme, the proposed one only requires extra
3
M
P
multiplications for the subband signal analysis and
synthesis.
5 EXPERIMENTAL RESULTS
To confirm the performance of the proposed scheme
in this paper, we present the experimental results of
the proposed scheme for the nonlinear system
identification. All the following results are obtained
by averaging over 50 Monte Carlo trials. The
performance is measured by use of mean square
error (MSE) defined as
2
10
10log [ ( )]en
. The input
signal is generated by the process
2
() ( 1) 1 ()
x
nxn n
ωωβ
=−+ (16)
where
()n
β
is the White Gaussian noise signal with
zero mean and unitary variance, the parameter
ω
is
selected in the range
[0, 0.95] , which interprets the
degree of correlation for the adjacent samples. The
FIR filter coefficients for the SAF and SSAF model
Nonlinear Subband Spline Adaptive Filter
165
are initialized as
-1
=[ ,0,...,0]
α
w with length
B
and
01
α
<≤, while the spline model is initially set to a
straight line with a unitary slope. For convenience,
only B-spline basis is applied in the simulations,
however, the similar results can also be achieved
using the CR-spline basis.
5.1 Experimental 1
The unknown system model
0
()Φ
is the Wiener
spline model which comprises a FIR filter
[0.6, 0.4,0.25, 0.15,0=.1]
o
T
−−w and a nonlinear spline
function represented by a 23 control points length
LUT
0
q ,
x
Δ and
m
x
Δ
are set to 0.2 and
0
q is
defined by
0
[ 2.2, 2.0, 1.8, 1.6, 1.4, 1.2, 1.0, 0.8, 0.91,
0.4, 0.2, 0.05,0, 0.4,0.58,1.0,1.0,1.2,1.4,1.6,1,8, 2.0, 2.2],
=−−−−−−−−−
q
(17)
For signal partitioning in this experiment, The
cosine-modulated filter banks with subband number
2, 4,K = and 8 are used and the prototype filters’
length of analysis filter increases with the number of
subband, the prototype filters’ length ,
P , are 32 , 64
and 128 for
2, 4,M = and 8 correspondently. The
default values
0.5
ω
= , 0.1
α
= , and 5B = are emp-
loyed and 10000 samples are used. An independent
White Gaussian noise signal,
()vn , is added to the
output of the unknown system, with 23-dB, 26-dB,
30dB signal to noise ratio (SNR) for
2, 4,8M =
respectively. The step sizes are selected to ensure
that the conventional SAF and the SSAF obtain the
similar steady-state MSE. The performances of the
SAF and proposed SSAF are compared for the diff-
erent numbers of subband in Fig.3. It can be seen
that the proposed SSAF supplies the faster conver-
gence rate than the SAF. This is due to the
decorrelating properties of the subband scheme for
colored input signals.
Figure 3: MSE curves of the SAF and SSAF.
5.2 Experimental 2
In the second experiment, we compare the MSE
performance of the polynomial model (Stenger,
2000), 2-rd order Volterra model (Kuech, 2002), 2-
rd order subband Volterra model (Burton, 2009),
SAF model (Scarpiniti, 2013) and proposed SSAF
model. The unknown system to be identified
0
()Φ
consists of two blocks, the first one is the 3-order
IIR filter
123
12 3
0.0154 0.046 0.0462 0.0154
() ,
1 1.99 1.572 0.4583
zzz
Hz
zz z
−−
−−
++ +
=
−+
(18)
and the second block nonlinearity
( ) sin( [ ]).yn xn= (19)
The number of subband is set to 4 for the SSAF
and 2-rd order subband Volterra model. The cosine-
modulated filter banks with subband number
4
M
=
are used and the prototype filters’ length is selected
to 64. The step sizes are set to
qw
= =0.01
μμ
for both
the HSAF and the SSAF, step size =0.001
μ
is used
for the polynomial model and the step size is set to
0.01 for the 2-rd order Volterra and 2-rd order
subband Volterra model. The signal to noise ratio is
30SNR dB= . The default values 0.6
ω
= , 0.1
α
= ,
==0.2
m
xxΔΔ and 15B = are employed and 50000
samples are used. A comparison of the MSE learn-
ing curves is reported in Fig. 4, it can be clearly
noted that the robustness of the proposed SSAF.
Figure 4: Comparison MSE of the different models in
experiment 2.
ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation
166
5.3 Experimental 3
The third experiment aims to test the effectiveness
of the proposed structure in case of a high degree of
nonlinearity. The unknown system in system ident-
ification
0
()Φ comprises three blocks. The first and
the last blocks are IIR filters whose transfer function
can be written as
12
1
12
12
12
0.2851 0.5704 0.2851
()
1 0.1024 0.4475
0.2025 0.2880 0.2025
,
1 0.6591 0.1498
zz
Hz
zz
zz
zz
−−
−−
−−
−−
⎛⎞
++
=
⎜⎟
−+
⎝⎠
⎛⎞
++
×
⎜⎟
−+
⎝⎠
(20)
12
3
12
12
12
0.2025 0.2880 0.2025
()
1 1.01 0.5861
0.2025 0.2880 0.2025
,
1 0.6591 0.1498
zz
Hz
zz
zz
zz
−−
−−
−−
−−
⎛⎞
++
=
⎜⎟
−+
⎝⎠
⎛⎞
++
×
⎜⎟
−+
⎝⎠
(21)
and the nonlinear portion of this model is expressed
by
2
2()
() .
1()
xn
yn
xn
=
+
(22)
It is noted that this model is capable of
describing the behavior model of the radio frequency
amplifier for satellite communications (Scarpiniti,
2013). For both the SAF and the SSAF, linear filter
length
B is set to 15. The parameter
ω
is set to 0.2.
All the other parameters are set to similar values as
in Experimental 2. Fig. 5 shows that the comparison
of the MSE learning curves for several models, it is
clear in this case that the deteriorating performance
of the Polynomial model and the 2-rd order Volterra
model, verifying that these two models only cope
with the case of mile nonlinearity. Furthermore, the
proposed SSAF obtains the best convergence
performance with respect to other models.
Figure 5: Comparison MSE of the different models in
experiment 3.
6 CONCLUSIONS
In this paper, a novel SSAF structure for nonlinear
system identification is presented. It consists of a
series of subband nonlinear filters, the adaptation of
these subband filters is carried out independently.
The computational complexity is analyzed based on
the LMS algorithm. Some experimental results in
the context of the nonlinear system identification
show the effectiveness of the proposed structure.
ACKNOWLEDGEMENTS
This research was supported by the National Natural
Science Foundation of China (61501119).
REFERENCES
Mathews V.J., Sicuranza G.L., 2000. Polynomial Signal
Processing.
John Wiley & Sons, Hoboken, New
Jersey.
Schetzen M., 1980.
The Volterra and Wiener Theories of
Nonlinear Systems
, John Wiley & Sons, New York,.
Haykin S., 2009.
Neural Networks and Learning Machines,
Pearson Publishing, 2
nd
ed.
Giri F., Bai E.W., (Eds.), 2010.
Block-Oriented Nonlinear
System Identification,
Springer-Verlag, Berlin.
Scarpiniti M., Conniniello D., Parisi R., Uncini A., 2013.
Nonlinear spline adaptive filtering, Signal Proce-ssing
93(4) 772–783.
Scarpiniti M., Conniniello D., Parisi R., Uncini A., 2014.
Hammerstein uniform cubic spline adaptive filtering:
Learning and convergence properties
, Si-gnal
Processing 100(4) 112–123.
Scarpiniti M., Conniniello D., Parisi R., Uncini A., 2015.
Novel cascade spline architectures for the id-
entification of nonlinear systems
, IEEE Tran-sanctions
on Circuits and Systems-I: regular papares 62(7) 1825-
1835.
Guarnieri S., Piazza F., Uncini A., 1999.
Multilayer
feedforward networks with adaptive spline activation
function
, IEEE Transanctions on Neural Networks,
10(3) 672-683.
Burton T. G., Goubran R.A., 2009. Beaucoup F.,
Non-
linear system identification using a sub-band adaptive
Volterra filter
, IEEE Transanctions on Instrumentation
and Measurement, 58(5) 1389-1397.
Stenger S., Kellerman W., 2000.
Adaptation of a
memoryless preprocessor for nonlinear acoustic echo
cancelling
, Signal Processing, 80(9) 1747–1760.
Kuech F., Kellerman W., 2002.
Nonlinear line echo
cancellation using a simplified second order Volterra
filter
, in Proceedings of the ICASSP, vol. 2, Orlando,
Flo-rida, 1117–1120.
Nonlinear Subband Spline Adaptive Filter
167