A Nonlinear Memory Power Amplifier Behavior Modeling and Identification Based on

Memory Polynomial Model in Soft-defined Shortwave Transmitter

Hanxin Zhou, Guojin Wan and Limin Chen
Department of Electronic Information Engineering, Nanchang University Nanchang, 330031, China
Email: [email protected]

Abstract—General Volterra series based models have been knowledge of the nonlinear behavior of shortwave power
succ-essfully applied for radio frequency (RF) power amplifier amplifier systems.
behave-oral modeling, but their high complexity tends to limit Several techniques can be applied to model the dynamic
their applications to “weakly” nonlinear systems. To model a
soft-defined shortwave PA with strong nonlinearities and long nonlinear behavior of PA. As an example, authors in [1]
memory effects, we propose a memory polynomial model (MP utilized two-box model and identified the behavioral model
model) which is a simplified Volterra series based model for of the PA with a static nonlinear block followed by a
RF power amplifiers by employing a diagonal pruning dynamic linear filter which distinguishes Wiener model.
algorithm, therefore dramatically reducing the complexity of Authors in [2] have used Volterra series to model PA with
the behavioral model. The comparison between the simulation memory effect which is limited to weakly nonlinear devices
results by MP Model and the measured result by the test
platform based on DSP and FPGA shows that MP model and
due to the highly expensive computational task and
adaptive RLS algorithm can describe the nonlinear memory complicated algorithm involved. In [3] Wiener-Hammer-
power amplifier behavior characterizes for the soft-defined stein model and its sub-models have been introduced and a
short-wave amplifier system accurately and effectively. link between the generalized Volterra series model and
Furthermore, the identification method for soft-defined these models is discussed. In each of the main basic
shortwave transmitter is simple, few parameters, steady structure of the Wiener-Hammerstein type models there
algorithm, and practical.
exists one static nonlinear block. Authors in [4] propose an
Keywords: MP model; short-wave power amplifier; behavioral efficient Wiener model for a power amplifier (PA) .The
model; RLS algorithm; nonlinear identification; Wiener model is formed by a linear filter and a memoryless
nonlinearity in which AM/AM and AM/PM characteristics
I. INTRODUCTION are approximated as piecewise linear and piecewise constant
Software radio always determines the radio technology functions, respectively. Authors in [5] model the PA with
trends and it has many advantages not only in military some kinds of the truncated volterra series and validate that
communications field but also civilian mobile the Parallel Hammerstein model is much closer to the
communications domain. However, all the radio transmitters referred PA.
have inherent nonlinear character and memory effect. In
soft-defined shortwave communication systems, when using II. AMPLIFIER MEMORY EFFECT
non-constant envelope signals with linear modulation, the The PA circuit has memory effects, resulting from the
actual shortwave power amplifier has inherent nonlinearity node voltage and current depend on not only the current
which causes the distortion within the signal bandwidth, input but also the historical signals due to the existence of
lowers the data transfer rate, reduces channel capacity, parameters with dynamic distribution[5].Power amplifier
deteriorate the bit error rates of the receivers and brings characteristic implies the evaluation of parameters, like
spectrum regeneration effect as well as has serious memory effects, which are not directly measurable from
interference for adjacent channel. In order to reduce these their input and/or output signal. The existence of memory
undesired effects and meet power and spectral efficiency as effects is often identified by imbalances between the
required, linearization techniques are necessary to corresponding upper and lower distortion products [7]. In
compensate for these nonlinear effects. In most existing wideband systems, the memory effects of PA are significant
predistortion techniques, shortwave PA is usually and the electrical memory effects are the main source [10].
considered as memoryless devices. Nevertheless, in Memory effects due to the existence of components which
wideband communication systems, the memory effects store energy, such as inductors and capacitors, Impedance of
introduced by wideband signals are significant and the inductors and capacitors is relevant to frequency, PA
performance of the traditional memoryless predistortion is memory effect is reflected as a non-linear distortion
seriously degraded. In order to use more efficiently of associated with the signal bandwidth and power.
shortwave high power amplifier, it is necessary to construct
precise behavior model of shortwave amplifier to exemplify
the nonlinear and the memory effects of it. For that reason,
special efforts have been invested in order to attain a deeper

978-1-4244-3709-2/10/$25.00 ©2010 IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on June 14,2022 at 08:09:25 UTC from IEEE Xplore. Restrictions apply.
 wave power amplifier will raise with the bandwidth’s
increase of input signal.

B. RLS Algorithm

x(n) y ( n) e( n )
Memory PA
+
Fig. 1 Power spectrum of PA 380 W output for 125kHz two-tone spacing
-
MP model
z ( n)

Adaptive
Estimation

Fig. 4 Block diagram of the PA idenlification

Define soft-defined short-wave amplifier system’s input
Fig. 2 Power spectrum of PA 1780 W output for 125kHz two-tone spacing
vector, nuclear vector respectively as:
III. SYSTEM MODEL AND ALGORITHM
X ( n ) = [ x ( n ), x ( n − 1)," , x ( n − M + 1),
A. System model x 3 ( n ), x 3 ( n − 1)," , x 3 ( n − M + 1)," , (2)

In order to describe all the unknown characteristics of x N ( n ), x N ( n − 1)," , x N ( n − M + 1)]T

short-wave power amplifier, we show the amplifier with a W ( n ) = [ω1,0 , ω1,1 ," , ω1, M −1 ,
relatively simple and memorable nonlinear behavior model, (3)
namely MP model which can be written as formula (1), and ω3,0 , ω3,1 ," , ω3, M −1 ," ,
we identify MP model parameters with the RLS algorithm. ω N ,0 , ω N ,1 ," , ω N , M −1 ]T
The AB short-wave power amplifier’s nonlinear
characteristics mainly depend on the odd terms (1, 3, 5, 7), Where, N = 2 K − 1 .So we can transfer (1) into vector
because the even terms are much smaller than the odd terms. product form:
So we only consider the low-order odd terms in this paper to z (n) = W T (n) X (n) (4)
simplify the model and make it easy to implement, we We let W ( n) has the optimal estimation to well
didn’t calculate the even terms, DC terms, and higher-order describe the amplifier’s characteristics, and we do iteration
odd terms. with normalized mean square error (NMSE). NMSE shows
x (n − 1) x (n − 2) x (n − 3) x (n − M + 1) the precision that the model describe the amplifier’s
x (n)
z −1 z −1 z −1 characteristics. The formula (4) is nonlinear to x (n) , while
is linear to the unknown parameter. We can update the
P0 (.) P1 (.) P2 (.) P3 (.) PM −1 (.) model parameter W ( n) with recursive least squares (RLS)
y0 ( n) y1 ( n) y2 ( n ) y3 (n) yM −1 ( n)
as (5) to (8).
z (n)
The RLS algorithm for parameter identification in MP
Fig.3 The structure of MP model for short-wave power amplifier model is
K (n) =
λ − 1 P ( n − 1) X ( n ) (5)
Where, Pi (x) = x + x3 +"+ x2k−1,(k =1,2,", K;i = 0,1,2,", M −1) 1 + λ − 1 X T ( n ) P ( n − 1) X ( n )

The MP model is written e ( n ) = y ( n ) − W T ( n − 1) X ( n ) (6)

M −1 K W ( n ) = W ( n − 1) + K ( n ) e ( n ) (7)
∑ ∑ω
2 ( k − 1)
z (n) = x(n − m ) x(n − m ) (1)
m = 0 k =1
2 k − 1, m
P ( n ) = λ − 1 P ( n − 1) − λ − 1 K ( n ) X T ( n ) P ( n − 1) (8)
Where, x (n) is the input envelope’s amplitude of short-
wave power amplifier. z ( n) is the output’s estimate value. C. The selection of order N and memory depth M
M represents the depth of memory, 2 K − 1 is the It’s very important to select the optimal order and the
nonlinear order. ω2 k −1, m is a real number, the actual short- memory depth, otherwise the model would not be allowed.
wave amplifier has memory effect and has the characteristic We can select N and M in actual modeling process
of relying on frequency, memory effect means the output of according to the nonlinear strength degree that short-wave
the amplifier is related with both current input and amplifier shows. But we select them according to Akaike
preceding input and output. The memory effect of short- Information Criterion here. Record AIC as Akaike
Information:

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on June 14,2022 at 08:09:25 UTC from IEEE Xplore. Restrictions apply.
 AIC = ln[ J ⋅ (1 + 2n1 / L)] (9) When The highest order N = 5 ，memory depth M
Where, n1 is the number of model parameter, L is the are different values， AIC curves are shown in Fig.6.
number of data used in identification. From Fig.6. when M ≥ 7 ， AIC are small. Obviously,
L −1
memory depth M = 7 is better. So M = 7 , when the
J = ∑ e 2 (n )
highest order N range from 1 to 11. The AIC variable curve
n=0 (10)
J is the residual sum of squares for MP model. is shown in fig.7.
The steps for determining the highest order N and memory
depth M are:
1) Fixed N as a singular value, M starts from 1 and the
following plus 1 in turn, the M corresponding to the
minimum AIC is the memory depth of MP model.
2) Fixed M as the selected value in step 1), N starts
from 1 and the following plus 2 in turn, the N
corresponding to the minimum AIC is the highest order
N of MP model. Fig.7 M = 7 , the curve of AIC when N is variable

The modeling effect is best when M = 7 , N = 7 .

IV. SIMULATION RESULTS Set the center frequency of 2MHz. use the single FM tone
In order to verify the effectiveness of MP (memory signal of maximum offset 5kHz as input recording 50
polynomials) models, we test it with the significant memory groups I/O data sample. Then randomly selected one sample
effect high-power Class AB shortwave power amplifier. to do the off-line training. the rest of the samples to do the
Using 50 groups I/O data which are the testing of single testing. Using NMSE to measure the accuracy of the model.
tone noise signal. each sample a total of 1,000 I / O data. The modeling result are shown in fig.8.
randomly selected as training samples, the other groups used
as the error testing, to analysis model’s generalization
ability and stability.
Vector coefficient W ( n) of initial values are
W (0) = [1, 0," , 0]T ， the forgetting factor of RLS
algorithm λ = 0.9998 ， δ = 0.005 。

Fig.8 2Mhz single tone FM noise signal, Maximum frequency offset

5kHz , the comparison between actual PA output waveform and
MP model output waveform.

Fig.5 Sample of time domain waveform

2MHz single-tone signal excitated the referred PA. we
obtained the actual output waveform as shown in Fig .5 .
The waveform is not similar to the smoothed cosine curve,
but with an inflexion, and the degree of distortion is Fig. 9 NMSE convergence curve
asymmetric between the left and right sides of peaks, After a number of iterations, we get the final W (n) .
Simulating waveforms as shown in fig.8, we can find that Fig.10.shows output power spectrum y ( n) .
the MP model can be achieved on the modeling memory PA
basically.

Fig.10 Comparison of power spectrum density at PA output

and seventh-order polynomial PA model
Fig. 6 The curve of AIC when the highest order, N = 5,
memory depth M has different values.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on June 14,2022 at 08:09:25 UTC from IEEE Xplore. Restrictions apply.
 practical. MP model’s structure is simple and its
computational complexity is small, and the RLS algorithm
has a high identification precision. Both the evaluative
performance in time domain and frequency domain and the
test results by the test platform of DSP and FPGA validate
that the MP model is more simple and effective in actual
short-wave amplifier modeling system.
ACKNOWLEDGMENT
This work was supported by the National Natural
Science Foundation of China (No.60762005), the Natural
Science Foundation of Jiangxi Province for Youth (No.2009
Fig.11. Shows the convergence process of the parameters W (n) in MP GQS0070), the Scientific and technological project of Jiang-
model which is based on the RLS algorithm which has attenuation xi Province (No.2009BGB01600 and 2009BGB02200) and
factor. the Science and Technology Foundation of Department of
Education in Jiangxi Province (No.GJJ08013, GJJ10362,
GJJ10374 and GJJ10070).
REFERENCES

[1] T. Liu, S. Boumaiza ,M. Helaoui, H.B Nasr, F. M. Ghannouch .

“Behavior Modeling Procedure of Wideband RF Transmitters
Exhibiting Memory Effects” Proc. IEEE MTT-S Int. Microw.
Symp.Dig. Long Beach, CA ,Jun.12-17, 2005
[2] M. Schetzen, The Volterra and Wiener Theories Nonlinear Systems
New Yorn: Wiley, 1980.
Fig.12 the statistical histogram of testing result [3] D.R. Morgan,. Z.Ma, J.Kim, M.G.Zierdt,J.Pastalan. "A Generalized
Memory Polynomial Model for Digital Predistortion of RF Power
The actual power amplifier’s output waveform is Amplifiers", IEEE Trans. Sign.Proc,vol.54, no. 10, Oct. 2006.
apparently asymmetric, which is caused by the nonlinear [4] Sungho Choi, Eui-Rim Jeong, and Yong H. Lee, “Adaptive
effects of odd terms 1, 3, 5, 7. We realize that the MP predistortion with direct learning based on piecewise linear
approximation of amplifier nonlinearity”, IEEE Trans. Signal Process.,
model’s order can be 7 and the memory depth should be at vol.3, no.3, pp.397-403, June 2009.
least 7 with repeated experiments. The frequency domain [5] Zhixing Le, Guojin Wan, Limin Chen, and Xianghua Lu. Research on
the truncated volterra series model for shortwave power amplifiers
effect in Fig.10 shows the PA output and tells us the power linearization [C]. The Ninth International Conference on Electronic
difference of 1, 3, 5, 7 odd harmonic terms in MP model are Measurement & Instruments, Beijing, China, 2009, 282～285.
respectively 0.0023dB, 0.0031dB, -0.0395dB, -0.2036dB, [6] J. Kim and K. Konstantinou, “Digital predistortion of wideband signals
based on power amplifier model with memory,” Electron. Lett., vol.
which all are smaller than 1dB. As the evaluation index 37, no. 23, pp. 1417–1418, 2001.
NMSE is -36dB, we can conclude that the MP model has a [7] João Paulo Martins, Nuno Borges Carvalho, José Carlos Pedro,
rather high precision. “Intermodulation distortion of third-order nonlinear systems with
memory under multisine excitations,” IEEE Transaction on
V. CONCLUSION Microwave Theory and Techniques, vol. 55, no. 6, pp. 1264-
1271,June 2007
In order to complete the soft-defined short-wave [8] A. Zhu and T. J. Brazil, “Behavioral modeling of RF power amplifiers
based on pruned Volterra series,” IEEE Microw.Wireless Compon.
amplifier’s pre-distortion correction, we first need to Lett., vol. 14, no. 12, pp. 563–565, Dec. 2004.
construct an accurate nonlinear behavior short-wave [9] Sungho Choi, Eui-Rim Jeong, and Yong H. Lee, “Adaptive
amplifier model to describe the short-wave amplifier's predistortion with direct learning based on piecewise linear
approximation of amplifier nonlinearity”, IEEE Trans. Signal Process.,
nonlinear characteristics and strong memory effects. We use vol.3, no.3, pp.397-403, June 2009.
MP model and the adaptive RLS algorithm to systematically [10] Limin Chen, Zhixing Le, Guojin Wan, and Songyun Liu. DDS-based
injected tone canceling method for short-wave PA memory effects
model and identify the software radio shortwave transmitter. measurement [C]. The 2009 International Conference on Wireless
The analysis of simulation by Matlab based on MP model Communications and Signal Processing, Nanjing, China, 2009.
and the result of experiments with the test platform of DSP [11] M. E. Gadringer, D. Silveira, and G. Magerl, “Efficient power
amplifier identification using modified parallel cascade Hammerstein
and FPGA show that the MP model, along with the adaptive models,” in IEEE Radio Wireless Symp. Dig., Jan. 2007, pp. 305–308.
RLS algorithm, can achieve an accurate model for the short- [12] D. R. Morgan, Z. Ma, J. Kim, M. G. Zierdty, and J. Pastalan, “A
wave amplifier system, the model can well describe generalized memory polynomial model for digital predistortion of RF
power amplifiers,” IEEE Trans. Signal Process, vol. 54, no. 10, pp.
characteristics of strong nonlinear and memory effect. This 3852–3860, Oct. 2006.
method has obvious advantages with simple identification,
few parameters, steady algorithm, and it is real very

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on June 14,2022 at 08:09:25 UTC from IEEE Xplore. Restrictions apply.