Thanks to visit codestin.com
Credit goes to www.scribd.com

0% found this document useful (0 votes)
18 views7 pages

Gorbunova 2020 J. Phys. Conf. Ser. 1679 022037

The paper discusses an autoregressive parametric method for modeling autowave processes, particularly in the context of oscillations in liquid-propellant engines. It highlights the method's ability to provide high-resolution spectral estimates without requiring complex signal processing techniques. The authors demonstrate the effectiveness of their approach through the design of a digital model that captures the dynamic system's essential properties.

Uploaded by

ri.bazhenov
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
18 views7 pages

Gorbunova 2020 J. Phys. Conf. Ser. 1679 022037

The paper discusses an autoregressive parametric method for modeling autowave processes, particularly in the context of oscillations in liquid-propellant engines. It highlights the method's ability to provide high-resolution spectral estimates without requiring complex signal processing techniques. The authors demonstrate the effectiveness of their approach through the design of a digital model that captures the dynamic system's essential properties.

Uploaded by

ri.bazhenov
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

Journal of Physics: Conference Series

PAPER • OPEN ACCESS

An autoregressive parametric method applied to the autowave process


modelling
To cite this article: T Gorbunova et al 2020 J. Phys.: Conf. Ser. 1679 022037

View the article online for updates and enhancements.

This content was downloaded from IP address 95.70.60.13 on 26/11/2020 at 13:11


APITECH II IOP Publishing
Journal of Physics: Conference Series 1679 (2020) 022037 doi:10.1088/1742-6596/1679/2/022037

An autoregressive parametric method applied to the autowave


process modelling

T Gorbunova1, R Bazhenov2, M Tumanova3, L Alekseeva4 and I Korosteleva4


1
Moscow State University of Civil Engineering, 26 Yaroslavskoye shosse, Moscow,
129337, Russia
2
Sholom-Aleichem Priamursky State University, 70a Shirokaya str., Birobidzhan,
679015, Russia
3
K.G. Razumovsky Moscow State University of Technologies and Management, 73
Zemlyanoy Val str., Moscow, 109004, Russia
4
Far Eastern State Transport University, 47 Serysheva str., Khabarovsk, 680000,
Russia

E-mail: [email protected]

Abstract. The authors of the paper research the topic of the autoregressive parametric spectral
estimation method which is applied to oscillation (vibration). It is illustrated by the study of
dynamically stable the combustion modes in a liquid-propellant engine. The scholars suppose a
hypothesis that if a model parameter is selected correctly, the method offers a higher resolution.
It does not require a dialogue-based weighted signal, and compensated parser/analyzer filter
related to the accurate estimates of the oscillating system attributes. The survey proves that
parametric spectral estimation amounts to solving an optimization problem with and seeking the
autoregressive model value parameters of waveform shaping, in which the model would be as
close as possible to the observed signal in reality. Based on the study, the authors design a digital
model of the dynamic system that meets all the required properties.

1. Introduction
The investigation of the oscillating process represented by autoregressive models (AR models) has been
actively observed recently by parametric methods of spectral estimation [1-3]. These methods assume
an available procedure for generating an intelligence signal, namely the signal must be produced as a
resonant response of a linear narrow-band system to a broad-band noise impact. For instance, according
to the diagnostic model, it occurs on dynamically stable combustion modes in a liquid-propellant engine.
The simplest diagnostic model of a combustion chamber as a potentially self-oscillating system in
the ‘noise’ modes of its operation [4] can represent a nonconservative dynamic second-order system,
stimulated by random broad-band (white) noise:
d2 Yν 𝑑𝑌𝜈 2 2 ( ) 𝜔𝑜𝜈
dt2
+ 2𝛿𝜈 𝑀 𝑑𝑡
+ 𝜔𝑜𝜈 𝑌𝜈 = 𝜔𝑜𝜈 𝜉 𝑡 , 𝛿𝜈
≫ 1, δν (𝑀) = (δνd − δνg ) > 0 , (1)

where
𝑌𝜈 is the time realization of a narrow-band random process,
ωoν is a radian frequency of natural/free vibrations (without damping),

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
APITECH II IOP Publishing
Journal of Physics: Conference Series 1679 (2020) 022037 doi:10.1088/1742-6596/1679/2/022037

𝑡 is the time.
Accordingly, δν𝑑 , δνg are the coefficients of dissipation and generation of acoustic energy are the
modes of normal vibrations, which are functions of M combustion chamber operating mode parameters,
ξ(t) is a time-independent normal random broad-band impact (turbulent combustion noise).
Each of the equations (1) of the linear diagnostic model is similar to the equation of a simple (R, L,
C) resonant oscillatory circuit (resonator filter) stimulated by time-independent broad-band noise [5, 6]
and known from Radio engineering. The amplitude-frequency response (AFR) module of such a filter
is defined by the corresponding expression (2)
1/2 1/2
ω4o (2πf )4
|W| = ( 2
(ω2−ωo )2+4δ2ω2
) = (16π4 (f2 −f2)20+16π2 f2δ2 ) . (2)
0

2. Materials and methods


The discrete digital simulator of the resonator filter (2) is a second-order autoregressive model (Yule
model [6]) declared by the equation
Yi,n = 𝛼1 Yi,n−1 + α2 Yi,n−2 + xn , (3)
where Yi,n…. is a sample at the output of i-filter, i = 1,2,3 … … … … …,
4r∗cos2πfi ∗h
α1 = , (4)
1+r

α2 = −r , (5)
r = exp(−2δd ∗ h) , (6)
fi is a resonant frequency of the filter,
h is a sample spacing, sec.,
δd is a filtered damping coefficient, sec.−1 .
Z is a rearrangement of the equation (3) can be represented as
Yn = α1 Yn ∗ Z−1 + α2 Yn ∗ Z−2 + xn . (7)
Therefore there is an AFR filter
1
W(Z) = 1−α −1 −α −2 . (8)
1 ∗Z 2 ∗Z

Putting Z = ej2π∗f∗Δ , where Δ is the time step, one gets


1 1
W(f) = [1−α −j2π∗f∗Δ −α e−2jπ∗f∗Δ ] = [1−α −jωΔ −α e−2jωΔ ] . (9)
1e 2 1e 2

AFR filter module


1 1/2
|W(f)| = ( ) (10)
1+α21+α22−2α1 (1−α2 )∗cos(2πfΔt)−2α2∗cos (4πfΔt)

Comparison of the module results of the amplitude-versus-frequency response characteristics


constructed in analogy with the expressions (2) and (10) stands for their identity property.
The digital simulator (3) is used as a waveform generator when testing the designed autoregressive
method for spectral estimation of diagnostic value parameters.
It is shown in [7] that equation (3) gives time-independent solutions for 𝑌𝑛 , provided that 𝑎1 and 𝑎2
are within a triangular region.
𝑎1 + 𝑎2 < 1, 𝑎1 − 𝑎2 > −1, −1 < 𝑎2 < 1
In this case, parametric spectral estimation amounts to solving an optimization problem, i.e.,
searching for the parameters of the autoregressive model of signal generation, whereby it would be as
close as possible to the actually observed signal.

2
APITECH II IOP Publishing
Journal of Physics: Conference Series 1679 (2020) 022037 doi:10.1088/1742-6596/1679/2/022037

In the AR model of the assumed timing series, the discrete values 𝑥(𝑛) are represented by a linear
combination of the prior values and the size of error ε(n) [8, 9].
x(n) = 𝑎1 𝑥(n−1) + 𝑎2 𝑥(𝑛−2) … + 𝑎𝐾 𝑥(𝑛−𝐾) + 𝑎𝑃 𝑥(𝑛−𝑃) + 𝜀(𝑛) = ∑PK=1 𝑎K x(n−K) + ε(n) , (11)
where P is a model order.
The value ε(n) is the difference between the true/ideal value of 𝑥𝐾 и value of x̂K , forecasted by the
AR model (prediction error).
It is assumed that 𝜀(𝑛) is white noise with a Gaussian probability density function of transient values
and a uniform power spectrum Nε =const. x(n) can be interpreted as the output of a recursive AR filter
of P order, where white noise enters. AFR of such a filter is defined by the expression
−1
H(f) = (1 − ∑PK=1 𝑎K ∗ e−iKωT ) , (12)
where T is the sampling period/time.
In accordance with it, the power spectral density of the AR model signal
σ2ε (n)
P(ω) = P 2 , (13)
|1−∑K=1 𝑎K ∗e−iKωT |

where σ2ε (n) is the white noise dispersion.


Thus, the AR method of spectral estimation amounts to estimating the coefficients 𝑎𝐾 AR-модели
of the AR model of P order, estimating the power 𝜎𝜀2 of white noise, and calculating the power spectral
density according to the formula (13).
The introduced parametric method for spectral estimation of resonant frequencies and vibration
damping coefficients consists of three stages. At the first stage, the coefficients 𝑎1 … 𝑎𝑃 f the
autoregressive model (of P order) are estimated. The authors chose a modified covariance method out
of the known methods for estimating these coefficients. At the second stage, the estimated values of the
parameters (𝑎1 , 𝑎2 … 𝑎𝑃 ) are inserted into the abstract/theoretical expression for the AFR squared
absolute value of the AR model (12).
At the third stage, in accordance with the diagnostic model, narrow-band rises of the squared absolute
value of AFR module of the AR model are approximated by a theoretical expression for the squared
absolute value of AFR module of a second-order linear resonant circuit, followed by an estimation of
δ
the oscillation damping coefficient d̂ (d̂ = ).
fP
1
|AFR|2 ~ (ω−ω 2 +δ2 , (14)
P)

whereωP is the resonant circular frequency, and δ is the dissipation loss coefficient.
Estimation of the required order of the autoregressive model is performed by comparing the obtained
estimates of oscillation damping coefficients with the corresponding estimates obtained by regular
methods (spectral, correlation).
In practice, the authors do not know the real correlation function of the signal explored, so they use
the estimates of the autocorrelation function (ACF) to minimize the prediction error. There is a number
of autoregressive estimation methods known, which differ mainly in the approach to processing edge/tip
effects, i.e., in the method of involving into the calculation of those edge/tip samples of the signal for
which there is no shifted twain when calculating the ACF.
All methods give virtually the same research results; differences begin to appear in case of short
tones when reviewing long sequences of samples.
It is worth mentioning that the expression for the power spectral density of AR model signal is the
same as the corresponding expression of the maximum entropy method, which is commonly used, for
example, in time series mining/analysis in astrophysics [7]. Thus, if there is a relevant adoption of P
order of the autoregressive process, an estimate of the spectral density that is supposed to be the best
from the maximum entropy method can be obtained automatically.

3
APITECH II IOP Publishing
Journal of Physics: Conference Series 1679 (2020) 022037 doi:10.1088/1742-6596/1679/2/022037

The authors choose the last method among the widely accepted ones for estimating the coefficients
of aK in the AR model (Yule-Walker, Berg, covariance, modified covariance). In this method, linear
prediction parameters are estimated by minimizing the arithmetic mean-variance of fore and aft linear
prediction errors (figure 1)
1
σ2f b = 2 (σ2f + σ2b ) , (15)
Where
N−1 P 2 N−1 P 2
1 1
σ2f = ∑ |x(n) + ∑ afk x(n+K) | , 𝜎𝑏2 = ∑ |x(n) + ∑ afk x(n+K) | .
N−P N−P
n=P K=1 n=P K=1

Figure 1. Flow chart of the linear


prediction error filter fore and aft.

Estimation of AR parameters using the maximum-likelihood method results in solving simultaneous


equations
Cx (1,1) Cx (1,2) … . Cx (1, P) a (1) Cx (1,0)
C (2,1) Cx (2,2) … . Cx (2, P) a (2) C (2,0)
( x )( ⋮ ) = ( x ), (16)
⋮ ⋮ ⋮ ⋮
Cx (P, 1) Cx (P, 2) … . Cx (P, P) a (P) Cx (P, 0)
where the items of the covariance matrix are written as
1
Cx(j,K) = 2(N−P) (∑N−1 N−1−P
n=P x(n−j) ∗ x(n−K) + ∑n=1 x(n+j) ∗ x(n+K) ) . (17)

White noise variance estimation


N−1 P N−1−P P
1
σ2ε = σ2f b,min = [∑ (x(n) + ∑ afK x(n−K) ) ∗ x(n) + ∑ (x(n) + ∑ afK x(n+K) ) x(n) ] ,
2(N − P)
n=P K=1 n=P K=1

or
σ2ε = σ2f b,min = 𝐶𝑥(0,0) + ∑𝑃𝐾=1 𝑎𝐾 ∗ 𝐶𝑥(𝑜,𝐾) . (18)
Several linear/serial algorithms for calculating the coefficients of the AR model are known (Yule-
Walker, Levinson–Durbin, Berg, etc.) [8]. Combustion chamber noise modelling is designed in this
study through a high-order AR process (P=50 ... 300). The modified covariance method (MCM) is
chosen among the characteristics of algorithms for estimating the AR process parameters. It has a high-
resolution capability. In particular, when it analyzes relatively short signal representations, it has no
decoupling/division of spectral/wavelength peaks and some other advantages [3, 8, 9].
The authors use a sample of fast computation of the AR model coefficients proposed by Marple [3]
as a pre-image of MCM algorithm, It shows a greater improvement of computer-based accuracy (for
example, than solving simultaneous equations with decomposition according to Cholesky).

4
APITECH II IOP Publishing
Journal of Physics: Conference Series 1679 (2020) 022037 doi:10.1088/1742-6596/1679/2/022037

3. Method verification/checking
Model (1) makes it possible to produce a time implementation of linear narrow-band noise Y(t) with
given statistical characteristics which are a prototype of acoustic noise of the combustion chamber at the
frequency of ν- normal mode
To implement it, one can make a sample estimate of oscillation decrement and compare the obtained
value with the given one. Using (1), the authors produce a time representation of linear narrow-band
noise Y(t) with the values of the resonant frequency f = 1000 Hz and the value of the oscillation
decrement d = 0.15. The processing results by Matlab [10] are presented in table 1.
Table 1. The effects of signal processing.
𝑓𝑃 a, 𝜀𝑑 f, 𝜀𝑑𝑇 g, 𝜀𝜎 h,
db 𝜎c 𝑑̂d 𝜎̂e
𝐻𝑧 % % %
1000 0.05 285.1 0.06 271.8 20 24.6 4.6
1000 0.10 219.4 0.09 222.7 10 17.4 1.5
1000 0.15 223.2 0.17 226.6 13 14.2 1.5
2000 0.15 214.4 0.16 213.8 6.6 10.0 0.2
4000 0.15 233.8 0.14 222.7 6.6 7.1 4.7
a
𝑓𝑃 is a given decrement value;
b
𝑑 - is a given decrement value;
c
𝜎 - is the root-mean-square value of the narrow-band signal, estimated according to the
1 1/2
formula: 𝜎 = (N ∑N 1 Yi ) , N=1000;

𝑑 is decrement estimation;
e
𝜎̂ is the estimation of the root-mean-square value of the signal by integrating the spectrum of the
AR model;
f
𝜀𝑑 is the decrement estimation error;
g
𝜀𝑑𝑇 is an abstract/ theoretical statistical error;
h
𝜀𝜎 is the difference ratio between the estimates 𝜎 and 𝜎̂.

As can be seen, the estimation decrement error is not more than the theoretical statistical error defined
by the expression𝜀𝑑𝑇 = (𝛿 ∗ 𝑇𝑃 )−1/2 , where 𝛿 is a given value of the damping coefficient; 𝑇𝑃 is the
implementation length.
The difference between the estimates of the mean-square values of the signal obtained straight
according to 𝑌(𝑛) implementation samples and integrating the spectrum of the AR model is insignificant
(no more than 5%).

4. Conclusion
Special test signals are usually used as a data sequence of finite length (representations) with the given
characteristics to compare the capabilities of different methods of spectral estimation of dynamic system
parameters (in the present case, resonant frequencies and vibration decrements). In particular, a
representative of such signals is a model digital signal generated by a second-order linear dynamic
system, which is supplied by white noise at the input. This model allows one to produce a time
representation of linear narrow-band noise 𝑌(𝑛) with the given dynamic and statistic characteristics
which is considered an equivalent of acoustic noise of the combustion chamber at the frequency of ν-
normal mode. When implementing it, one can make a sample estimation of the frequency and vibration
decrement by one or another method and compare the obtained values with the given ones.
The values of frequencies and vibration decrements are a priori unknown while tested by the actual
signal method (such as signals from pressure pulsation sensors). Therefore, these values, which are
conventionally named ‘basic values’, are previously estimated using an alternative method (in this case,
a correlation method according to the rate of decay of the autocorrelation function of the signal at the
resonant frequency). Evidently, the estimates of oscillation decrements based on the AR model are close
to the given values.

5
APITECH II IOP Publishing
Journal of Physics: Conference Series 1679 (2020) 022037 doi:10.1088/1742-6596/1679/2/022037

References
[1] Hayes M H 1996 Statistial Digital Signal Processing and Modeling (New York: Wiley)
[2] Papoulis A and Pillai S U 2002 Probability Random Variables and Stochastic Processes (New
York: M Graw-Hill)
[3] Marple Jr S L 2019 Digital Spectral Analysis (Mineola ; New York: Dover Publications)
[4] Culick F E 2006 Unsteady Motions in Combustion Chambers for Propulsion Systems (Neuilly-
sur-Seine Cedex: N.A.T.O., Research and Technology Organization)
[5] Levin B R 1989 Theoretical Foundations of the Statistical Radio Engineering (Moscow: Radio i
svyaz)
[6] Bykov V 1973 STIN 75 10294
[7] Monogan J E 2015 Political Analysis Using R 157-86
[8] Jenkins G M and Watts D G 1968 Spectral Analysis and Its Applications (Oakland: Holden-Day)
[9] Deergha Rao K and Swamy M N S 2018 Digital Signal Processing: Theory and Practice
(Singapore: Springer Singapore)
[10] Thyagarajan K S 2019 Introduction to Digital Signal Processing Using MATLAB with
Application to Digital Communications (Cham: Springer)

You might also like