Professional Documents
Culture Documents
samples to be processed.
Abstract—Many physical phenomena are described by In [1] and [2] it is stated that the accuracy of the
equations whose solution is an exponential decay. Determining frequencies estimated by Prony’s method depends on the
the amplitudes, damping factors and oscillating frequencies sampling rate of the analyzed signal. Guidelines for choosing
from experimental data is a common task in various areas of the optimal sampling rate are presented, but they are limited
science and technology. Prony’s method is a well-known tool for
to a particular application.
estimating these parameters. However in some cases, especially
when the signal-to-noise ratio is small, Prony’s method may In this paper an analysis of the influence of the parameter
inaccurately estimate damping factors and oscillating values will be discussed, and a simple but powerful algorithm
frequencies. Hence, this paper investigates possible ways for for choosing the optimal sampling rate is presented. Our
improving the performance of Prony’s method when analyzing algorithm is based on minimizing condition numbers of
the output signal of a pressure transducer. It is shown that by matrices arising in the Prony’s method.
choosing optimal parameter values to reduce the condition The main contribution of this research is to improve the
number it is possible to increase the accuracy of Prony’s performance of Prony’s method for the spectral analysis of
method. The algorithm for choosing the optimal sampling rate is pressure transducer signals. It is achieved by pre-processing
presented. the signal data, roughly estimating the frequencies using a
Keywords—Condition number, Prony’s method, sampling Fourier transform, choosing the optimal sampling rate and
rate, frequency estimation, model order. then analyzing the signal by Prony’s method.
The layout of the paper is as follows. Section 2 provides a
I. INTRODUCTION brief description of Prony’s method and potential ways to
improve it. In section 3 the algorithm for choosing the
pressure transducer is an elastic body consisting of optimal sampling rate is presented. Results from simulations
A several mechanical components. The frequencies of the
natural vibrations of these components depend primarily
are discussed in section 4. Finally, conclusions and future
plans are stated.
on the transducer structure rather than on ambient conditions.
If these frequencies are estimated and isolated from those
frequencies caused by the process, it will be possible to detect II. PRONY’S METHOD
the condition of the transducer itself. The key technical
A. Conventional Algorithm of Prony’s Method
challenge is to estimate the parameters of a pressure
transducer output signal (such as frequencies or damping Prony’s method is based on fitting a sum of exponential
factors) with very high accuracy in order to be able to detect terms to the measured data.
small shifts. The conventional method of spectral analysis is It can identify the amplitudes a j , damping factors V j
the Fourier transform which suffers from various
shortcomings, including the resolution limit depending on the ( V j t 0 ), frequencies f j and phases M j of the signal
response duration. n V t
Prony’s method is an alternative to the Fourier transform s (t )= ¦ a j e j cos(2Sf j t M j ) , (1)
j =1
widely applied to spectral analysis. It provides high-accuracy
estimation for amplitudes, damping factors, frequencies and which is a sum of n non-increasing cosine curves. Without
phases of signals. loss of generality, we may assume that 0 < f1 < < f n .
Unfortunately, in practice there are some disadvantages The signal (1) can also be written in the following form:
with Prony’s method because it is very sensitive to noise. p
¦Ae
J jt
Therefore special techniques are needed to analyze real s (t ) , (2)
j
signals with Prony’s method. Moreover, the accuracy of the j 1
method depends on parameters choices, such as the model
Here there are p 2n complex exponentials (two for
order, the sampling rate of the signal and the number of
each of cosine curve).
In general, the Prony’s method involves three steps
Manuscript received February 11, 2012. (see [3]):
O.Yu. Bushuev is with the Computer-measurement department, South
Ural State University, Chelyabinsk, Russia (corresponding author, phone: +7-
1) Construction of a linear prediction (LP) model from the
904-306-45-81; e-mail: bushuev@init.susu.ac.ru). observed data and solving it.
O.L. Ibryaeva is with the Department of differential equations and 2) Computation of the roots of the characteristic
dynamical systems, South Ural State University, Chelyabinsk, Russia (e-mail: polynomial equation generated by the LP model.
oli@6v6power.ru).
¦s
equations (7) and (8). It is well-known that for the relative
D k = 0, m = 0,1, , p 1. (6)
k =0
k m error Gx of the solution of the linear system Ax = b the
following inequality is true:
Given that D p = 1 (6) can be rewritten as:
GA Gb
p 1 Gx d condA , (9)
1 condA GA
¦s
k =0
k m D k = s p m , m = 0,1, , p 1. (7)
where GA A data, and
is the relative error of matrix
1
After solving equations (7), coefficients Dk of (5) are condA || A || || A || is the condition number of A
obtained. (recall that condA t 1 for any A and for any matrix norm
The second step involves the computation of the roots || A || ).
z j of the characteristic polynomial (5). Then frequencies f j The smaller the condition number of the matrixes in (7)
and damping factors Vj are estimated. Finally, using z j the and (8), the better the results obtained.
The matrix in equation (8) is the well-known
complex amplitudes A j are estimated by solving the set of Vandermonde matrix (see [5]-[6]):
p equations (4): § 1 1 1 ·
¨ ¸
p
¨ z1 z2 zp ¸
sm = ¦A j z mj , m = 0,1, , p 1. V =¨ ,
¸
(8)
j =1 ¨ p 1 ¸
B. Ways to Improve the Performance of Prony’s Method
¨z
© 1 z 2p 1 z pp 1 ¸¹
Prony’s method in its original form is very sensitive to
noise. In practice, techniques can be used to improve its while the matrix in equation (7) can be written as
performance. Firstly, more than 2n experimental noisy points
can be used to determine 2n parameters and to search for § A1Ap Az p1 p1 ·
coefficients D k in (7) by the least squares method. ¨ 1 1 Apzp ¸
¨ ¸
Another approach (described in [4]) is to use Prony’s ¨ Az A z Az p p ¸
method first with model order q larger than the expected M=¨ 11 pp 1 1 Apzp ¸
number of exponential components p. The result is a set of q ¨ ¸
exponentials that are candidates for the signal components. ¨ ¸
¨ p1 ¸
Apzpp1
The next step is to determine a smaller subset of exponentials 2p 2 2p 2
which gives the best fit to the observed data using the least ¨ Az
1 Az
1 1 Apzp ¸¹
© 1
squares criterion. To determine the number of exponential
components in TLS-Prony’s method Singular Value
635
It can be verified that this matrix has a factorization
M = VA V T , where VT is the transposed Vandermonde
§ A1 0 0 ·
¨ ¸
¨0 A2 0 ¸
matrix and A = ¨
¸
.
¨ ¸
¨0 0 A p ¸¹
©
should be located on the complex plane as further as possible circle, then cond V = 1.
from each other. It can be shown that if a1 = = an and roots z1 , , z p
An example of the dependence between the sampling rate
are uniformly placed on a unit circle, then cond M = 1.
Fs (Fs = 1/T) and the condition numbers is presented in
Fig.1. In the example a signal that is a sum of three damped Real cosine curves lead to complex conjugate nodes
sinusoids is considered. The frequencies are 400 Hz, 750 Hz z n 1 = z n* , , z p = z1* . It means that z1 , , z p , ( p 2n )
and 1500 Hz and the damping factors are 100, 50, 150 are uniformly placed on the unit circle, if and only if they are
respectively. the vertexes of a regular polygon which is symmetrical with
It can be seen that the optimal sampling rate for the respect to X-axis, i.e.
example is in the range around 3500 Hz.
S 3S (2n 1)S
In the next section we will describe an algorithm for arg z1 = , arg z 2 = , , arg z n = .
choosing the optimal sampling rate for the given signal. 2n 2n 2n
Note that arg z k 2Sf k T , k 1, n, where T is the
sampling period.
636
§ 1 ·
¨ 4n ¸
§ f1 · ¨ ¸
¨ ¸ ¨ 3 ¸
¨ f2 ¸
F B ¨ 4n ¸
¨ ¸ ¨ ¸
¨ ¸
fn ¹ ¨ ¸
© ¨ 2n 1 ¸
¨ ¸
© 4n ¹
© fn ¹ ¨ 2n 1 ¸
The values of frequencies and damping factors, influencing
the optimal sampling rate, are presented in Table I. We add
¨ ¸
© 4n ¹ white noise to the signal to obtain a noisy signal with signal-
Numerical experiments suggest the use of the Euclidean to-noise ratio (SNR) of 30 dB.
norm for the expression (12). In Fig. 2 logarithms of the Parameters of the signal (amplitudes, frequencies, damping
condition numbers of M and V and the Euclidean norm of factors and phases) are estimated by means of the original
(12) for different sampling rates are presented. Prony’s method. To check the proposed algorithm we take
Fig. 2 shows that the minimum of the Euclidean norm the signal with different sampling rates:
1) Fs = 2900 Hz, that is about the Nyquist rate for the
|| F T B ||2 corresponds to the minimum value of the
signal;
condition numbers of M and V. Note that this norm was 2) Fs = 3800 Hz, that is calculated according to the
selected in [9] in the similar problem of minimizing the algorithm and is optimal for the signal;
condition number of the Vandermonde matrix. 3) Fs = 4700 Hz.
The minimum of the Euclidean norm can be easily found The noisy signal (SNR = 30 dB) and its reconstruction
† with the estimated parameters for these sampling rates are
using a pseudoinverse matrix F :
presented in Fig. 3-5. Solid lines correspond to the signal, and
dotted lines correspond to the reconstruction.
Topt = F † B. (13) The figures show that the value Fs = 3800 Hz appears to be
the optimal sampling rate for the signal. Estimated parameters
Now we can present the algorithm for choosing the optimal obtained from the noisy signal with this sampling rate are
sampling rate. The algorithm is broken down into three steps. very close to their actual values.
1. Pre-process the signal, estimating the number n and TABLE I
rough values of the frequencies f1 < < f n . SIGNAL PARAMETERS
637
in [10]. The author concludes that the value of cond V can
be small if:
1) the number p of nodes z1, …, zp is small;
2) nodes z1, …, zp are uniformly placed on a circle.
It means that the proposed algorithm of choosing the
optimal sampling rate will result in minimum condition
number not only for the original Prony’s method but also for
TLS-Prony’s method since conditions 1) and 2) are desirable
for both methods.
V. CONCLUSION
In this paper ways to improve the performance of Prony’s
method were described. It was shown that one of the reasons
Fig. 3. Noisy signal (SNR = 30 dB) and its reconstruction, Fs=2900 Hz
of its inaccuracy is the improper choosing of a signal
sampling rate since it can cause the large condition numbers
of matrixes arising in Prony’s method. Some ideas of how to
decrease the condition number were proposed.
An algorithm for choosing the optimal sampling rate of a
signal being analyzed was presented. It was shown how this
algorithm can be applied for analyzing a signal with several
examples. The algorithm may be applied when the pre-
processing of the signal is possible, for example, while
analyzing an output signal of a pressure transducer.
REFERENCES
[1] J.C.-H. Peng and N.-K.C. Nair, “Effects of Sampling in Monitoring
Power System Oscillations using On-line Prony Analysis”, in Proc.
Fig. 4. Noisy signal (SNR = 30 dB) and its reconstruction, Fs=3800 Hz 2008 Australasian Universities Power Engineering Conf. (AUPEC'08),
Sydney, 2008, pp. 1-5.
[2] R.W. Kulp, “An Optimum Sampling Procedure for Use with the Prony
Method”, IEEE Transactions on Electromagnetic Compatibility, Vol.
EMC-23, No. 2, May 1981
[3] S.L. Marple, Jr., “Digital spectral analysis with applications”,
Englewood Cliffs, NJ, Prentice-Hall, Inc., 1987, 512 p.
[4] A.A. Istratov and O.F. Vyvenko, “Exponential analysis in physical
phenomena”, in Review of Scientific Instruments, vol. 70, num. 2,
February 1999.
[5] P. Ferreira, “Super-resolution, the recovery of missing samples and
Vandermonde matrices on the unit circle”, in Proc. Workshop on
Sampling Theory and Applications, Loen, Norway, 1999, pp. 216- 220.
[6] L. Berman and A. Feuer, “On perfect conditioning of Vandermonde
matrices on the unit circle”, in Electronic Journal of Linear Algebra,
vol. 16, pp. 157-161, 2007.
[7] O.L. Ibryaeva, A.S. Semenov and A.L. Shestakov, “Estimation of the
condition number of the matrix in Prony’s method”, in Chelyabinsk
Fig. 5. Noisy signal (SNR = 30 dB) and its reconstruction, Fs=4700 Hz research center, Trans. on Math., vol. 2 (48), pp.1-5, 2010 (in
Russian).
Note that we use the original Prony’s method. It estimates [8] W. Gautschi, “On inverses of Vandermonde and confluent
Vandermonde matrices”,in Numer. Math. , vol. 4, pp. 117–123, 1962.
12 parameters of the signal using its 2p=12 samples. [9] L. Berman and A. Feuer, “Robust patterns in recurrent sampling of
TLS-Prony’s method (that uses N>2p samples) will result in a multi-band signals”, in IEEE Transactions on Signal Processing, 2008,
higher accuracy. In this case matrixes and V are vol. 56, issue 6, pp. 2326-2333.
rectangular. For the condition number of the rectangular [10] F. Bazan, “Conditioning of rectangular vandermonde matrices with
nodes in the unit disk”, SIAM Journal on Matrix Analysis and
Vandermonde matrix upper and lower estimates are obtained Applications, vol. 21, pp. 679–693, 2000.
638