mssp182004187 197

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/239376684

Non-linear noise analysis from a capacitive pressure transmitter

Article  in  Mechanical Systems and Signal Processing · January 2004


DOI: 10.1016/S0888-3270(03)00015-3

CITATIONS READS
13 162

4 authors, including:

José Manuel Chicharro Juan Blázquez


University of Castilla-La Mancha Centro Investigaciones Energéticas, Medioambientales y Tecnológicas
50 PUBLICATIONS   489 CITATIONS    44 PUBLICATIONS   169 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Negative Stiffness System View project

Suspension for high speed rail vehicles View project

All content following this page was uploaded by José Manuel Chicharro on 16 December 2019.

The user has requested enhancement of the downloaded file.


ARTICLE IN PRESS

Mechanical Systems
and
Signal Processing
Mechanical Systems and Signal Processing 18 (2004) 187–197
www.elsevier.com/locate/jnlabr/ymssp

Non-linear noise analysis from


a capacitive pressure transmitter
A. Garc!ıa-Berrocala,*, J.M. Chicharroa, J. Bla! zquezb, M. Balba! sa
a
Applied Physics to Natural Resources Department, ETSIM, Polytechnic University of Madrid, 28003 Spain
b
Nuclear Fission Department, CIEMAT, Madrid, Spain

Abstract

When the capacitive pressure transmitters—Rosemount type—experience a silicone oil-loss from its inner
structure, they do not behave as a linear dynamical system. Such an anomaly is reflected on the noise signal.
In this work, a non-linear model for the sensor and the sensing line is defined. It explains the main features
of the noise signal as measured in service and in the laboratory.
r 2003 Elsevier Science Ltd. All rights reserved.

1. Introduction

In the Nuclear Power Plants the pressure plays a crucial role, so it must be accurately measured;
besides, the Regulatory Commissions oblige to assurance that pressure sensors are measuring
properly [1]. Among the diverse transmitters, the Rosemount type is the most common nowadays
in the Plants. They are used for measuring flow, pressure and level. For the purpose of sensor
surveillance, noise analysis is widely used to estimate the response time.
Those anomalies occasioning large shifts in the signals are detected easily. Some others, hidden
in the noise are more difficult to be detected; that is the case of sensing lines blockages, bubbles in
the sensing line, diaphragm stiffness, etc. The ordinary procedure consists of analysing the noise
signal in order to calculate the response time. This quantity is used as a reference for the predictive
maintenance [2–4].
One of the anomalies most difficult to be detected is the uncommon silicone oil-loss from the
sensor inner structure. The response time is not sensitive to such syndrome. Even more, the sensor
dynamics leaves its linear behaviour; in the laboratory, it is observed that the response time
depends on the sign of the pressure ramp used to determine it. The diaphragm elastic properties

*Corresponding author. Tel.: +34-91-336-70-63; fax: +34-91-336-69-52.


E-mail address: agus@dfarn.upm.es (A. Garc!ıa-Berrocal).

0888-3270/04/$ - see front matter r 2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0888-3270(03)00015-3
ARTICLE IN PRESS

188 A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197

are affected; diaphragm oscillations are asymmetric and, as a consequence, the noise amplitudes
are skewness [5].
Facing that problem [6–9], Rosemount has removed all the sensors in the Plants affected by the
syndrome. The Regulatory Commissions have encouraged the researchers to develope procedures
for oil-loss early detection. Nowadays empirical procedures are found [10]; so, it is interesting to
create theoretical models reproducing the most significant aspects of the noise signal under the
syndrome.
What way the sensor behaves as non-linear? The present model, based on the lack of symmetry
of the sensing diaphragm oscillations, tries to answer that question. It explains why two response
times are detected in the laboratory. It is necessary simulating a non-linear noise for the sensor
response to a driven white noise; as an application, several new procedures for detecting the oil-
loss syndrome might be designed.

2. Experience from the laboratory

The oil-loss problem was firstly detected in 1987 at the Milestone—Unit 3—Nuclear Power
Plant. It was about the differential pressure transmitter Rosemount 1153 HD5PC. The operators
observed:
* A deviation from the redundant sensors indication.
* A decrement of the sensor noise level.
* A very slow response of the sensor after a reactor shut-down.
* Improper sensor calibration.

Transmitters were declared out of service and returned to Rosemount. After a destructive
inspection it was concluded that the root cause was the silicone oil-loss from the inner sensor
structure.
Because of this event, the US Nuclear Regulatory Commission, several Nuclear Power Plants,
EPRI Institute, related private companies, researchers and mainly Rosemount opened a research
for understanding the syndrome. Methods for an early detection of the syndrome were designed in
order to deal with such a problem.
The silicone oil fills a tiny channel to reduce the pressure between the outer diaphragm—the
isolation one—and the inner diaphragm—the sensing one. Being the transmitter capacitive, the oil
acts also as a dielectric, changing the dielectrical capacity of the sensing diaphragm. Hashemian
and colleagues found that the isolation diaphragm delayed its return to the equilibrium position
after an overpressure event [10]. The oil-loss caused an asymmetric behaviour of the diaphragms
vibration (Fig. 1).
Response times to pressure ramp in laboratory for an ill 1153 model appear on Table 1.
When the set point corresponds to the lower pressure it is observed that the diaphragms return
slowly to its equilibrium position, but only for negative ramps; however, for positive ramps, the
response time does not depend on the set point. It means that diaphragms keep its elasticity on.
The asymmetry of the vibrations is reflected on the noise signal. There is a relationship between
the skewness of the noise amplitude distribution and the oil-loss. Then, a monitoring of the noise
ARTICLE IN PRESS

A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197 189

Fig. 1. Rosemount sensor diagram and its electric analogy for the case of the oil-loss syndrome.

Table 1
Laboratory response times (s) for an 1153 model
Set point Increasing ramp Decreasing ramp
Low 0.23 171
Medium 0.25 19
High 0.25 1.1

based on the skewness is used for an early detection of the syndrome. Still, skewness can be due to
other causes. To make sure that the sensor has the syndrome more conditions are required. For
instance, the observed loss of noise amplitude is confirmed by the reduction of the signal variance.
The response time estimated with the noise signal is not sensitive to the syndrome. It means that
the frequency spectrum is not affected much. This fact helps modelling the syndrome. If the
spectrum does not reveal any change in the coupling between the sensor and the sensing line, the
trouble must be in the sensing diaphragm capacity.

3. Modelling

Modelling is focused to explain the frequency spectrum of the noise signal. It includes the
sensing line and the sensor. In the frequency region of interest, below 20 Hz, the length of the line
is about 25 m with a rather irregular layout. On the other hand, the wavelength for the sound
ARTICLE IN PRESS

190 A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197

Table 2
Typical values for sensor and sensing line parameters
Mechanical system Electrical system Typical values (literature)
Pressure Voltage Pa
Displaced volume Intensity m3
Sensing line inertia Inductance ðLÞ 5.55  107 Pa s2 m3
Sensing line friction Resistance ðRÞ 2.21  108 Pa s m3
Isolating diaphragm Capacity ðMÞ 1.03  1012 m3 Pa1
Oil line friction Resistance ðrÞ 2.5  1012 Pa s m3
Sensing diaphragm Capacity ðCÞ 1.61  1012 m3 Pa1

Fig. 2. Typical power spectrum density (PSD) for noise signal. The fitting line is calculated using an autoregressive
model.

propagation in liquid water is much shorter; it means that the sensing line may be modelled as if it
were lumped. The model can be built writing down the corresponding differential equations or
following an electric analogy. The analogy is more suitable because the natural user of this
technology is a technician in the plant, which is by far more familiar with the language of transfer
function and electrical circuits.
The electrical circuit appearing in Fig. 1 is proposed to model the syndrome. The magnitudes of
the relevant quantities are shown in Table 2.
The data are taken from the literature [11]. They are specific for each sensor, but can be
obtained by fitting the measured power spectrum density (PSD) to the squared modulus of a
transfer function within a real pole and a complex conjugate pair [4]. The line is modelled [2] by an
inductance L and a resistance R; the sensor, by the outer diaphragm M; the inner diaphragm C
and the very large inner resistance r; needed for pressure reduction. In the Fig. 2 it is shown a
typical PSD corresponding to the above magnitudes.
The main features of the syndrome can be explained splitting the sensing line diaphragm into
two capacities. In the case of positive pressure, the normal capacity C1 is assigned; when negative,
it is supposed that the oil-loss changes the capacity up to C2 : To separate both capacities a pair of
diodes is included, making thus the model to be non-linear.
ARTICLE IN PRESS

A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197 191

The transfer function HðsÞ ¼ V0 ðsÞ=Vi ðsÞ of the linear circuit is [5]
1=Cs
HðsÞ ¼   ð1Þ
r þ 1=Cs
ðLs þ RÞ 1 þ þ r þ 1=Cs
1=Ms
where C ¼ C2 if V0 p0 and C ¼ C1 if V0 > 0: The way of being non-linear is due then to a pair of
transfer function. It is necessary to explain why the response times to negative and positive ramps
are different. It is derived from Eq. (1) that when C grows HðsÞ decreases. In order to explain the
observed noise variance reduction the model postulates that C2 > C1 necessarily.
More explicitly, the response to the ramp of the linear circuit are given by L1 ½HðsÞ=s2 ; being
1
L the inverse Laplace operator. The response time is calculated from the ramp response. The
response time ðtÞ for Rosemount sensors is mainly due to the sensor and not to the sensing line [3].
Its value is approximately t ¼ rC; so, denoting by tþ and t the response times to the positive and
negative pressure ramps, their ratio is proportional to C1 =C2 ; yielding a measure of the oil-loss
importance. As noise signals are measured in Plant but not the ramps, the next step consists of
translating the above ideas into the noise analysis procedures.

4. The noise analysis

Noise signals are analysed in the time domain [12]. Following the standard procedure, an
autoregressive (AR) model is fitted to the noise time series. If xj stands for the recorded noise
value at instant tj ; then
X n
xj ¼ ak xjk þ ej ; j ¼ 0 ; y; N ð2Þ
k¼1

where N is the length of the statistical noise time series, ak the autoregressive coefficients, n; the
fitting order, and ej the driven gaussian white noise. The AR coefficients are calculated supposing
that the driven noise eðtÞ has minimum variance and solving the corresponding linear system of
equations. The order was chosen following the Akaike information criterion [13].
Once the AR parameters are fitted, the driven noise ej is replaced in Eq. (2) by the unit step
function, obtaining the step response function. From the step response, the response time is
calculated straightforward.
Eq. (2) is only valid for linear systems. Being the transmitter regarded as bilinear there will
be two response times: the tþ ; corresponding to the positive ramp ðC2 ¼ 0Þ and the t
corresponding to the negative ramp ðC1 ¼ 0Þ: Both times can be obtained from the transfer
function straightforward. The response times can also be obtained simulating noise in both cases;
when C2 ; the set aþ is obtained; the set a corresponds to the case C1 ¼ 0: Both sets are listed in
Table 3. The process used to get the AR coefficients comes from the transfer equation (1) as
follows:
* Calculating the impulse response from HðsÞ:
* Calculating the autocorrelation function of the impulse response.
* Solving the Yule–Walker equation [12] to obtain the AR coefficients.
ARTICLE IN PRESS

192 A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197

Table 3
AR coefficients for both diaphragms stiffness: aþ correspond to the normal case (C2 ¼ 0; and ignoring the diodes),
a are for the oil-loss case ðC2 =C1 ¼ 2Þ: Sampling time 0.01 s
AR coefficients 1 2 3 4 5 6
aþ 2.035 2.489 2.156 1.010 0.351 0.075
a 2.048 2.497 2.168 1.010 0.348 0.073

Fig. 3. Simulated pressure noise amplitude for the normal case.

The noise signals have been simulated by solving numerically the non-linear stochastic
differential equations corresponding to the sensor dynamics. In Fig. 3 it is shown the noise
corresponding to the normal case; it can be observed that noise is gaussian. For the case of the oil-
loss syndrome, the noise, which is generated with the same driving noise that normal case, is
plotted in Fig. 4. The lack of symmetry between positive and negative amplitudes means that the
noise has skewness, and consequently, it is not gaussian. This particular case is simulated for
C2 =C1 ¼ 2: The variance reduction is also observed.

5. Results

The first symptom of the syndrome is the loss of noise amplitude. In Fig. 5 it is shown how
the variance of noise signal decreases when the capacity C2 increases. The reference capacity C1 is
left unchanged. Redundant sensors which do not experience oil-loss keep the noise variance
constant. That is a key point for the syndrome early diagnosis. It is necessary although not
sufficient.
ARTICLE IN PRESS

A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197 193

Fig. 4. Simulated pressure noise amplitude for the case of the oil-loss syndrome.

Fig. 5. Variance reduction versus the capacity change DC=C1 ¼ ðC2 2C1 Þ=C1 :

The skewness of the noise amplitude distribution versus the change of the capacity is plotted in
Fig. 6. The increasing skewness is due to the lack of symmetry in the noise signal and to the
variance reduction. Noise from redundant sensors is gaussian.
Noise analysis in the frequency domain is less useful for the syndrome early detection. The
resonant peak on the power spectrum density is formed due to the sensing line and the isolation
diaphragm. Both remain unaffected by the syndrome. In spite of that, a tiny shift of the real pole
to the left can be observed. Fig. 7 is plotted for the case C2 =C1 ¼ 2; it should be compared with
Fig. 2 corresponding to C2 =C1 ¼ 1: The resonance peak is unaffected by the syndrome.
The response time estimated by the noise analysis in the time domain does not change very
much. Because the syndrome is modelled decreasing the sensing diaphragm stiffness, the response
time should increase as an overall effect. In fact, the dynamical system is non-linear so the
ARTICLE IN PRESS

194 A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197

Fig. 6. Skewness in the pressure noise amplitude distribution versus the capacity change.

Fig. 7. Noise power spectrum density (PSD) for C2 =C1 ¼ 2: Tiny difference with the case for C2 =C1 ¼ 1:

response time is not univocal defined; but, being the response time the required quantity for the
sensor surveillance, it has been calculated as if the system were linear. Results appear in Fig. 8. It
is concluded that response time is not as sensitive as skewness for the syndrome early detection.
Although, when the aþ and a AR coefficients were used to estimate the response time to a
positive and negative ramp, respectively, two different values were obtained, according to the
laboratory experience. Responses to the ramps are plotted in Fig. 9 for the case C2 =C1 ¼ 2:
The response times obtained were: tþ ¼ 0:24 s; t ¼ 0:47 s: This result suggests that the ratio
ðt  tþ Þ=t (being t the response time regarding the noise as linear) is an indicator of the
syndrome intensity.
Consequently, it is proposed decomposing xðtÞ; the noise signal, into two signals: the positive
one, xþ ðtÞ; and the negative one, x ðtÞ: Explicitly, xþ ðtÞ ¼ xðtÞ for xðtÞ > 0 and xþ ðtÞ ¼ 0 for
xðtÞp0: A similar definition for x ðtÞ:
It should be borne in mind that xþ ðtÞ and x ðtÞ are fictitious noise signals. They are analysed
with the only purpose of the oil-loss syndrome surveillance. The response times obtained, as a
consequence of the analysis, do not correspond to the real response times, mainly because it does
ARTICLE IN PRESS

A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197 195

Fig. 8. Response times regarding the noise as linear versus the capacity change.

Fig. 9. Response to pressure ramps for the case C2 =C1 ¼ 2: Increasing ramp: (a) input, (b) output. Decreasing ramp: (c)
input, (d) output.

not exists a real response time within a non-linear system. Under those circumstances, response
times may be regarded as indexes, instead of physical quantities, indicating how severe the
syndrome is.
What can one expect from an AR linear analysis of xþ ðtÞ and x ðtÞ which are not linear by
definition? Obviously not a signal reproducibility suitable for forecasting: the AR coefficients
cannot reproduce the constant part of the signals. Nevertheless the AR coefficients are able to
reproduce many special features. For instance, it can be tested that when xðtÞ comes from a linear
system, the AR analysis of xþ ðtÞ and x ðtÞ yields the same response time, in spite xþ ðtÞ and x ðtÞ
are clearly non-linear.
According, in spite of the noise comes from a non-linear transmitter, both resultant signals are
analysed as if the noise were linear, obtaining the response times tðxþ Þ and tðx Þ: The ratio
jtðx Þ  tðxþ Þj=t > 0:1 may detect the oil-loss syndrome. Results appear in Fig. 10.
Summarising, the syndrome is recognised using three indicators:

* The variance reduction,


* The skewness,
* The ratio ½tðx Þ  tðxþ Þ=t:
ARTICLE IN PRESS

196 A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197

Fig. 10. Ratio Dt=t ¼ ½tðx Þ  tðxþ Þ=t versus the capacity change. tðx Þ and tðxþ Þ are the response times of noise
signal split in its positive and negative parts. t is the response time regarding the noise as linear.

All indicators are measurable with standard noise analysis techniques. Plant procedures might
be added to the actual sensor surveillance procedures in order to know if a given sensor suffers the
syndrome and when it has to be removed.

6. Conclusions

A model is proposed to explain quantitatively the laboratory results for a Rosemount pressure
transmitter under the oil-loss syndrome. The electric equivalent circuit for this mechanical system,
the sensor and the sensing line, has two diodes to account for the observed non-linear responses.
Noise analysis match the observed results in the laboratory and the Plant. It is able to recognise
the syndrome quantifying its intensity. Results can be translated to the Plant for a safer sensor
surveillance.

Acknowledgements

The authors would like to thank Dr. P. Vilarroig for careful review of the manuscript.

References

1. Nuclear plant aging research program plan, US Nuclear Regulatory Commission, NUREG-1144, September 1987.
2. J.A. Mullens, J.A. Thie, Pressure noise in pressurized water reactors, Nuclear Regulatory Commission, NUREG/
CR 4389, 1985, pp. 76–78.
.
3. O. Glocker, D.F. Cooke, G.J. Czuppon, K.K. Kpoor, Estimating the response time of pressure/flow transmitters
and RTDs in CANDU reactors, Proceeding of the International Meeting on Reactor Noise, IMORN-28, NCSR
Demokritos, Athens, Greece, 2000, pp. 271–286.
ARTICLE IN PRESS

A. Garc!ıa-Berrocal et al. / Mechanical Systems and Signal Processing 18 (2004) 187–197 197

4. J. Bl!azquez, J. Ballestr!ın, Pressure transmitter surveillance: the dominant real pole case, Progress in Nuclear Energy
29 (3) (1995) 139–145.
5. J. Bla! zquez, O. Vela, A useful model for the capacitive sensor oil-loss problem, Proceedings of the International
Meeting on Reactor Noise, IMORN-27, CIEMAT, Valencia, Spain, 1997, pp. 3–10, ISBN 84-7834-337-7.
6. Loss of fill oil in transmitter manufactured by Rosemount, NRC Bulletin No 90-01, December 1992, US Nuclear
Regulatory Commission, Washington, DC.
7. Technical guidance for detection of oil-loss failure of Rosemount’s pressure transmitters, Electric Power Research
Institute, EPRI NP-7121, December 1990, Palo Alto, CA.
8. Nuclear Operating Group, Rosemount Eng. Company, Status report on oil-loss Rosemount model 1153 series B/D
and 1154 transmitters, Rosemount Report No. D9200129, Minneapolis, MN, 1992.
9. H.M. Hashemian, et al., Effect of aging on response time of Nuclear Plant pressure sensors, Nuclear Regulatory
Commission, NUREG/CR-5383, Washington, DC, 1989.
10. H.M. Hashemian, D.W. Mitchell, R.E. Fain, K.M. Petersen, Long term performance and aging characteristics of
Nuclear Plant pressure transmitters, Nuclear Regulatory Commission, NUREG/CR-5851, 1993, pp. 190–217.
11. J. Barbero, J. Bl!azquez, O. Vela, Bubbles in the sensing line of Nuclear Power Plants pressure transmitters: the shift
of spectrum resonances, Nuclear Engineering and Design 19 (2000) 327–334.
12. M.B. Priestly, Spectral Analysis and Time Series, Academic Press Ltd., London, 1981, pp. 291–384.
13. H. Akaike, Anew look at the statistical model identification, IEEE Transactions on Automatic Control AC-19
(1974) 716–728.

View publication stats

You might also like