Conference PaperPDF Available

Combining Process and Condition Monitoring Data

Authors:

Abstract and Figures

Early detection of fluctuations in operating conditions and fault detection can be done with similar methods. Signal processing is needed for the condition monitoring measurements, and interpolation is some process measurements and especially laboratory analysis. Effective time delays are very important in process data. Feature extraction uses statistical analysis, and the methods can be based on generalised norms and moments. Intelligent condition and stress indices are calculated from these features by nonlinear scaling. The new scaling approach, which also uses the norms and moments, improves sensitivity to small fluctuations. In the condition monitoring cases, the condition indices are consistent with the vibration severity criteria, which originate from VDI 2056. Only one norm was needed in the cavitation analysis, and the resulting index can be used in power control. The same methodology provides good results in detecting fluctuations of flavour ingredients in brewing, in predicting web break sensitivity in paper machines and in intelligent analysers of the process conditions in wastewater treatment. Linguistic equation (LE) models of the normal case suit for detecting fluctuations, and case-based reasoning (CBR) is used if specific case models can be developed. The overall procedure includes following steps: (1) select informative features, (2) scale the features, (3) calculate intelligent indices, and (4) combine indices in models.
Content may be subject to copyright.
Combining Process and Condition Monitoring Data
M.Sc. (Eng) Esko Juuso and Prof. Kauko Leiviskä
University of Oulu, Control Engineering Laboratory, Finland
Abstract
Early detection of fluctuations in operating conditions and fault detection can be done with
similar methods. Signal processing is needed for the condition monitoring measurements, and
interpolation is some process measurements and especially laboratory analysis. Effective time
delays are very important in process data. Feature extraction uses statistical analysis, and the
methods can be based on generalised norms and moments. Intelligent condition and stress
indices are calculated from these features by nonlinear scaling. The new scaling approach,
which also uses the norms and moments, improves sensitivity to small fluctuations. In the
condition monitoring cases, the condition indices are consistent with the vibration severity
criteria, which originate from VDI 2056. Only one norm was needed in the cavitation
analysis, and the resulting index can be used in power control. The same methodology
provides good results in detecting fluctuations of flavour ingredients in brewing, in predicting
web break sensitivity in paper machines and in intelligent analysers of the process conditions
in wastewater treatment. Linguistic equation (LE) models of the normal case suit for detecting
fluctuations, and case-based reasoning (CBR) is used if specific case models can be
developed. The overall procedure includes following steps: (1) select informative features, (2)
scale the features, (3) calculate intelligent indices, and (4) combine indices in models.
Keywords: Intelligent methods, statistical analysis, fault diagnosis, condition monitoring,
process industry
1Introduction
Production of high quality products, more efficient use of energy and raw materials, and
stable operation in different operating conditions require advanced automatic systems for
detecting faults and fluctuations in process operation. Interactions between process variables
Revised manuscript sent to congress materials
and various time delays depend strongly on operating conditions and can dramatically limit
the performance and even destabilise the closed loop system. Paper machines can be run at
desired speed with the least possible amount of breaks and good uniform quality when
runnability is good [1]. Efficient use of fermentation vessels is crucial in brewing economy
since fermentation is the most time consuming step in the production of beer, especially
continuous fermentation requires fast and precise response to any changes in process
conditions seen as fluctuations of flavour ingredients [2].
In general, recognizing the most important machines for production and theirs appropriate
health monitoring bring in considerable benefits through more operating time and fewer
shutdowns. Early detection of fluctuations is important as overhaul before breakdown is in
many cases more effective than run to failure method. Automatic fault detection provides
more time to identify root causes when a system takes care of some routine tasks [3].
Measurable changes can happen e.g. in machines vibration, temperature, power consumption
and number of particles in lubricant. Generally, vibration measurements are the most widely
used methods. Generalised moments were calculated for a specific sample time in [4]. The
root mean square (rms) values and peak values are most widely used features, see [5] for more
application experiences. Generalised norms, which were introduced in [6], have been used in
scaled form in condition indices [7]. Nonlinear scaling is an essential part of many
applications [8, 9]. Recently, the scaling methodology has been improved [10], and the new
methodology has been extended to normal process variables [11].
Theoretical development and industrial applications on fault diagnosis are continuously active
in many areas. Faults in sensors, actuators, controllers, control loops, process equipment,
process parameters, as well as operator-induced faults have been described various ways, and
also a wide variety of methods have been used in fault detection [12]. Functions and features
can be realised with various methodologies, and there are also numerous ways to combine the
overall system [13]. The fault detection problem is, in principle, a classification problem that
classifies the state of the process to normal or faulty. This is based on thresholds, feature
selection and classification and trend analysis. Fault localisation starts from the symptoms
generated in the previous stage and it either reports fault causes to the operator or starts
automatically the corrective actions.
Detecting operating conditions can be done with similar method as fault detection although
the classes do not necessarily correspond to any fault [13]. Connection between vibration and
machines health can be used for fault diagnosis by using advanced signal processing and
feature extraction methods. Increasing number of measurement points and more demanding
problems require automatic fault detection. Model-based approach is used in complex
systems. Final goal is to use these indicators in control. Indirect measurements have been used
directly in control, e.g. fuel quality in lime kiln control [8] and water quality in water
treatment [14]. Simulation tests on using cavitation indices in power control of a water turbine
are promising for combining condition monitoring and control [15].
The frequency range of phenomena studied in fault detection varies tremendously; from
milliseconds in vibration analysis to hours and days in analysing the slowly developing
process faults. Especially for slow processes met in chemical, pulp and paper and biotechnical
industries, temporal reasoning and trend analysis are very valuable tools to diagnose and
control the process.
This paper presents a nonlinear scaling approach which can be used as indirect measurements
and extended to control and case-based reasoning. Results of applications in condition
monitoring and process analysis are summarised.
2Signal processing and feature extraction
Features can be defined by
,)
1
()( /1
1
)(/1 p
N
i
p
pp
p
p
i
x
N
MM å
=
==
a
a
t
a
t
(1)
where
Â
Î
a
is the order of derivation, the order of the moment
Â
Î
p is non-zero, and
t
is
the sample time. The sample has Nsignal values. This norm, which has the same dimensions
as the signal )(
a
x, is defined in such a way that
.
¥
£
¥
p
The norm (1) is a Hölder mean,
also known as the power mean. The norm values increase with increasing order, i.e.
,)()( /1/1 qqpp MM
a
t
a
t
£(2)
if
q
p
<
. The increase is monotonous if all the signals are not equal. The absolute mean, the
rms value and the absolute harmonic mean are special cases where the order
p
is 1, 2 and -1,
respectively. When the order 0
®
p, we obtain from (1) the absolute geometric mean. The
norm (1) represents all the norms from the minimum to the maximum, which correspond the
orders
=
p
and
¥
=
p
, respectively. [6]
The sample time
t
is an essential parameter in the calculation of norms. The sample time and
the number of samples should be chosen on the basis of process and faults. Short sample
times and relatively small requirements for frequency ranges make this approach feasible for
on-line analysis and control. The computation of the norms can be divided into the
computation of equal sized sub-blocks, i.e. the norm for several samples can be obtained as
the norm for the norms of individual samples. The same result is obtained using the norms:
[ ]
,)(
1
)(
1/1
1
/1
1
/1
p
K
ii
p
S
p
K
i
p
p
i
p
S
p
p
KSS
SM
K
M
K
Mú
û
ù
ê
ë
é
=
þ
ý
ü
î
í
ì
=åå ==
a
t
a
t
a
t
(3)
where S
K is the number of samples
{
}
N
i
i
x1
)(
=
a
.
In condition monitoring, the signals are velocity )1(
x, acceleration )2(
x and higher derivatives,
)3(
xand )4(
x. The other signals are obtained from acceleration through analogue or numerical
integration and derivation, see [6]. Variables can be selected by comparing health indices,
denoted by SOL. These indices are calculated for each feature by dividing the reference value
by the feature value obtained for the case under consideration. The reference value is the
corresponding feature value in the good conditions. This is done separately for each rotation
speed, and the decisions are based on the average.
The norm (1) was introduced for signals in [6], but it can also be used for normal process
measurements [11]. The derivation is not used for the process data if the measurements are
not very frequent. In addition, data acquisition systems normally use the norm (1) with 1
=
p,
and the averages for longer intervals are calculated by (3). If all the signal values are positive,
also the norms generalised to any real-valued order produce real-valued features. To obtain
dimensionless features, these norms must be normalised.
3Detecting operating conditions
Process and condition monitoring data is combined in detecting operating conditions (Fig. 1):
normal process measurements are directly used in feature extraction, signal processing is
needed for the condition monitoring data, and some infrequent measurements need to be
interpolated.
Figure 1. Detecting operating conditions and faults.
3.1 Nonlinear scaling
The basic idea of the linguistic equation (LE) methodology is the nonlinear scaling developed
to extract meanings of variables from measurement signals. The scaling function scale the real
values of variables to the range of [-2, +2] which combines normal operation [-1, +1] with
handling of warnings and alarms. The scaling function contains two monotonously increasing
functions: one for the values between -2 and 0, and one for the values between 0 and 2. Both
expertise and data can be used in developing the mapping functions (membership definitions).
Functions consist of two second-order polynomials )( j
Xf j
- and )( jj Xf +, which are defined
by points
{
}
.)2),(max(),1,)((),0,(),1,)((),2),(min( jjhjjlj xcccx -- (4)
The analysis of the corner points is based the norms (1) and the skewness of the data. The
value range of j
x is divided into two parts by the central tendency value j
c and the core
area, [ jhjl cc )(,)( ], is limited by the central tendency values of the lower and upper part.
There are problems when the value range is very wide or the distribution is very concentrated.
The new approach introduced in [10] is based on the generalised moment
(
)
,
)(
k
X
k
p
p
k
MXE
s
ga
ta
ú
û
ù
ê
ë
é-
=(5)
where X
s
is calculated about the origin, and kis a positive integer. Since this analysis is
done for the features, derivation is not used, i.e. 0
=
a
. The central tendency value is chosen
by the point where the skewness 3
g
changes, i.e. 0
3=
g
. Then the data set is divided into two
parts: a lower part and an upper part. The same analysis is done for these two data sets. The
estimates of the corner points,
(
)
j
l
cand
(
)
j
h
c, are the points where the direction of the
skewness changes. The iteration is performed with generalised norms.
Only monotonous, increasing functions can produce realisable systems to be used in a
continuous form. In order to keep the functions monotonous and increasing, the derivatives of
the functions should always be positive. The monotonic increase of the scaling functions is
satisfied with
))(()()max(
),)(()min()(
jjhjjhj
jljjjjl cccx
ccxc
-=-
-=-
+
-
a
a
(6)
if the coefficients -
j
a
and +
j
a
are both in the range ú
û
ù
ê
ë
é3,
3
1. Corrections are done by changing
the borders of the core area, the borders of the support area or the centre point. The local
linearity requirement is taken into account, if possible. [10]
3.2 Intelligent indices
A condition index can be based on several features, which are all scaled to the range
[
]
2,2-:
()
[
]
),()max( )(1
1
1
1
a
a
ta
kk
mn
nk kk
p
k
n
kkC FfwMfwI -
+
+=
-
=åå += (7)
where k
w is the coefficient and 1-
k
f the scaling function of the feature k. Features include
maximum norms
(
)
k
p
M
a
t
max and other features )(
a
k
F, e.g. bins of the histograms.
Features can have specific frequency ranges. Index (7) is obtained from the features of the
signal )(
a
x, but an index can also combine the features of different physical signals. The
condition index )(
a
C
I is a number between -2 and 2: high values correspond to good
conditions, and -2 means not allowable conditions.
The index (7) can also represent a stress index )(
a
S
I, where high values correspond to high
stress conditions, and -2 means very low stress. High-stress operating conditions should be
avoided in order to keep the process in good condition.
3.3 Model-based case detection
Linguistic equations (LE) models provide a flexible environment for fault diagnosis
applications, software sensors, risk analysis and detection of sensor failures [16, 17]. Only
linear equations are needed in LE models, since the nonlinear effects are handled with the
scaling functions. Complex systems also use case-based reasoning (CBR), which is a problem
solving paradigm for finding out the solution to a new problem by remembering a previous
similar situation and by reusing information and knowledge of that situation. In structural
point of view, CBR is a cyclical method that stresses reuse of solutions to similar problems,
where solutions are maintained in carefully indexed memory [1].
4Applications
The first level in detection of operating conditions is to find out if the process is in normal
operation. Deviation from the good operation can be detected by condition and stress indices.
In complex systems with several faults, specialised case models are needed to identify the
fault cases. Degrees of membership for the activated cases can be used for estimating some
process or product quality features.
4.1 Water turbine
The new scaling function improves the sensitivity of the cavitation index for short periods of
cavitation (Fig. 2). The previous methods provided three levels: (1) cavitation-free, (2) short
periods of cavitation and (3) cavitation. The new scaling approach divides level 2 into two
levels (Tab. 1). These results are consistent with the vibration severity criteria, which
originate from VDI 2056 [18, 19]. The cavitation-free situations and strong cavitation are
detected with the relative )max( 75.2
4
3M. The new index provides a clear indication when
the short term cavitation becomes stronger. The levels presented in Table 1 can also be
defined with fuzzy membership functions. [10]
Figure 2. The cavitation index for a Kaplan water turbine [10].
Table 1. Cavitation index and vibration severity criteria [10].
Cavitation index Cavitation level Severity
1
)4( -<
C
ICavitation-free Good
01 )4( <£- C
IShort periods of weak cavitation Usable
10 )4( <£ C
IShort periods of cavitation Still acceptable
1
)4( ³
C
ICavitation Not acceptable
In the simulation test [15], power control minimises the cavitation risk by dividing the load
between three turbines, whose conditions are normal, bad and very good. Each turbine has
three operating modes: low, normal and high power. In the normal area, a cavitation free
power level is taken as an operating point. The low and high operating areas are defined by
local minima of the cavitation indices. The control system has a feedforward controller, which
allocates the load to the turbines by means of cavitation indices, and a feedback controller,
which is based on the LE approach. Each turbine has a P type LE controller which is adapted
to the operating conditions by the scaling functions. The cumulative time in the strong
cavitation provides an indication of possible damage to be used in selecting the turbine for
low power operation. The characteristic curves are adapted to the recent indices in order to
handle the changes in the condition of the turbines. For power stations with many turbines,
alternatives to reduce cavitation risks are evaluated by simulation to optimise maintenance
actions.
4.2 Lime kiln
The scaled norms ))(max( 1
1
4
151 Mf - and ))(max( 25.4
25.4
4
151 Mf - are highly sensitive to
faulty situations in the supporting rolls (Fig. 3). Surface damage and alignment problems are
clearly detected and in the present system also identified with two norms of different order
obtained from the signal )4(
x. An early indication of friction increase is also achieved. The
data set covers the following cases: (1) surface problems, (2) good conditions after grinding,
(3) misalignment, (4) stronger misalignment, (5) very good conditions after repair work, and
(6) very good conditions one year later. Maintenance was done for one of the supporting rolls.
All the rolls can be analysed using the same features throughout the data set.
The overall condition index )4(
C
I can be defined as an average of the scaled features. All the
very good cases are close to the lower corner
[
]
2,2
-
-
:.1
)4( -<
C
I The index is below zero
for the good cases and also for cases with small fluctuations. These cases are clearly usable.
For still acceptable cases, the index is below one. All the faults and clear friction cases are in
the area where ,1
)4( >
C
I i.e. the severity level is ‘not acceptable’. The upper corner
[
]
2,2 is
related to the case where temperature was reduced by cooling. The very high signal levels are
caused by noise, but the hidden faults can be detected by stopping the cooling for a moment.
Figure 3. Scaled norms, p = 1 and 4.25, to different condition of the points 1-16: surface
damage (t), misalignment (), friction (n), very good (Ï), good (), small fluctuations
(È), and special cases (Ø) [10].
4.3 Experimental systems
Machine condition monitoring presented in [20] was based on models developed for normal
operation and nine fault cases including rotor unbalance, bent shaft, misalignment, and
bearing faults. Five different rotation speeds were used, and measurements were taken with
seven separate accelerometers simultaneously in a test rig. Two sensors measured axial
vibration and five accelerometers radial vibration in vertical direction. The classification
results of the experimental cases were very good and logical in [20]. Also a special artificial
immune system (AIS) algorithm, the AbNET, provided comparable results [21].
A large number of equations and features are not necessarily needed. The large number of
features was reduced by genetic algorithms in [22]. The number of variables and sensors
required to diagnose each fault can be reduced considerably by selecting the most sensitive
features [23]. Misalignment and bearing faults were detected by the peak value )3(
p
x obtained
from two sensors (Fig. 4). More features are needed to identify the faults. In this case, three
sensors are needed from seven, and totally seven features are used to calculate eight health
indices. Good conditions and all six fault types can be identified by seven rules, and the
strength of misalignment (Fig. 4(a)) and unbalance [23] can be calculated. The results are
further improved by using the rotation speed is used in the fault models.
(a) Misalignment (b) Bearing faults
Figure 4. Health indices (SOL) calculated for misalignment and bearing faults in a test rig: the
minimum-maximum variation of the index values is caused by the rotation speed [23].
4.4 Continuous brewing
A model-based system has been developed detecting for monitoring and diagnostics of
fermentation and flavour formation in immobilized yeast fermentation process has been
presented in [2]. Some fluctuations of flavour ingredients were detected for only short time
periods as the process was stable (Fig. 5). Time span of these fluctuations was usually too
short for development of specialised models. A more detailed model with seven equations for
all the prereactor measurements indicates clear differences in the end of the test period. By
creating grounds for prediction of quality factors the models increase options to control the
product quality in different cases.
Figure 5. Degree of membership for the normal model of continuous brewing [2].
4.5 Web break sensitivity
The web break sensitivity stands for possibility of web to break, predicting the amount of
breaks during one day from the measurements collected from the paper making process before
the actual paper machine. The paper making process is typically nonlinear with many and
long delays, numerous process feedbacks at several levels, several closed control loops,
interactions between physical and chemical factors, and various unmeasurable factors. Delays
and interactions change in time and with process conditions. [24]
The case-based reasoning (CBR) is needed in the identification of different operating
situations in paper machines. Because similar break sensitivities can result from a multitude
of dissimilar cases, the case base uses a division into categories which correspond to different
levels of break sensitivity. The system is based on LE models and fuzzy logic. Each equation
provides a new fact with a degree of membership, and the resulting set of facts is used in the
fuzzy reasoning of process cases and break categories. The break category is defined by the
case with the highest degree of membership. Although, the case base is fairly small the results
from the on-line tests were relatively good compared to real break sensitivity. The predicted
break sensitivity is an indirect measurement, which provides an early indication of process
changes. The list of variables in the active cases can be used to avoid harmful operating
conditions. [1]
4.6 Wastewater treatment
Wastewater treatment within Finnish pulp and paper industry is most commonly done in an
activated sludge plant, which is a complex biological process, where several physical,
chemical, and microbiological mechanisms simultaneously affect purification results. Limits
of the emissions are defined by authorities. A lot of process measurements are available, but
measurements do not include sufficient information on special features of the influent nor on
microbial composition of the sludge. Populations of microorganisms are highly important, e.g.
sludge bulking cause especially poor treatment efficiency results when biosludge escapes
from secondary clarification.
Figure 6. Effect of the operating conditions on the treatment result of an active sludge
treatment plant: reduction of chemical oxygen demand (COD) and diluted sludge volume
index (DSVI) [11].
Biological water treatment depends strongly on changes in inlet water quality. Load and
nutrient should be balanced since both an exceptionally high load and excess nutrients cause
problems. The operating conditions are modified by oxygen, temperature and flow. Much
slower changes in biological state influence drastically on the purification result and
subsequent process phases. The nonlinear scaling approach presented above were used in [11]
for the variables obtained from process measurements and laboratory analysis. Interpolation
was needed for some variables. The balance of the load and the nutrients was evaluated by the
difference of the corresponding scaled values. The worst case with low reduction and settling
problems arise, when there multiple warnings and alarms (Fig. 6). Correspondingly, good
reduction and very good settling was achieved when there were very few warnings. As it
takes some time to lose good conditions and also recover from problematic conditions, the
intelligent indices are useful for process control.
5Conclusions
Early detection of fluctuations in operating conditions and fault detection can both be based
on generalised norms and moments. Signal processing is needed for the condition monitoring
measurements, and interpolation for some process measurements. The new scaling approach,
which also uses the norms and moments, improves sensitivity to small fluctuations. LE
models of the normal case suit for detecting fluctuations, and CBR is used if specific case
models can be developed. The overall procedure includes following steps: (1) select
informative features, (2) scale the features, (3) calculate intelligent indices, and (4) combine
indices in models. As the analysis is based the same methodology in all these applications,
monitoring of the machines can be combined with process data. Smooth operation and high
quality of products is the main goal of all these applications, and this can be achieved by
combining these indicators with process control in the same way as it has been done for
smaller indicators used in lime kiln control and water treatment.
References
1. Juuso, E. and Ahola, T., 2008, ‘Case-based Detection of Operating Conditions in
Complex Nonlinear Systems’, Proceedings of the 17th World Congress of the
International Federation of Automatic Control (IFAC), Seoul, Korea, July 6-11, IFAC, pp.
11142-11147. http://www.ifac-papersonline.net/
2. Juuso, E., and Kronlöf, J., 2005, ‘Model-based monitoring of immobilized yeast
fermentation using fuzzy logic and linguistic equations’, Proceedings of the 16th Triennial
World Congress, Prague, July 3-8, IFAC, pp. 97-102. http://www.ifac-papersonline.net/
3. Lahdelma, S., Strackeljan, J., and Behr, D., 1999, ‘Combination of Higher Order
Derivatives and a Fuzzy Classifier as a New Approach for Monitoring Rotating
Machinery’, COMADEM 1999, Sunderland, pp. 231-241.
4. Lahdelma, S., and Juuso, E., 2008, ‘Vibration Analysis of Cavitation in Kaplan Water
Turbines’, Proceedings of the 17 World Congress of the International Federation of
Automatic Control (IFAC), Seoul, Korea, July 6-11, pp. 13420-13425. http://www.ifac-
papersonline.net/
5. Lahdelma, S., and Juuso, E., 2008, ‘Advanced Signal Processing in Mechanical Fault
Diagnosis’, Proceedings of Fifth International Conference on Condition Monitoring &
Machinery Failure Prevention Technologies CM 2008 – MFPT 2008, Edinburgh, 15-18
July, pp. 879-889.
6. Lahdelma, S., and Juuso, E., 2008, ’Signal Processing in Vibration Analysis’, Proceedings
of CM 2008 – MFPT 2008, Edinburgh, pp. 867-878.
7. Juuso, E., and Lahdelma, S., 2008, ’Intelligent Condition Indices in Fault Diagnosis’,
Proceedings of CM 2008 – MFPT 2008, Edinburgh, pp 698-708.
8. Juuso, E. K., 1999, ‘Fuzzy Control in Process Industry: The Linguistic Equation
Approach’, In: Verbruggen, H. B., H.-J. Zimmermann and R. Babuska, editors, Fuzzy
Algorithms for Control, International Series in Intelligent Technologies, pp. 243-300.
Kluwer, Boston.
9. Juuso, E. K., 2004, ‘Integration of Intelligent Systems in Development of Smart Adaptive
Systems’, International Journal of Approximate Reasoning. Vol. 35, Issue 3, March 2004,
pp. 307-337.
10. Juuso, E., and Lahdelma, S., 2010, ‘Intelligent Scaling of Features in Fault Diagnosis’,
Proceedings of 7th International Conference on Condition Monitoring & Machinery
Failure Prevention Technologies CM 2010 – MFPT 2010, Stratford-upon-Avon, June 22-
24, 15 p.
11. Juuso, E. K., 2010, ‘Data-based development of hybrid models for biological wastewater
treatment in pulp and paper industry, Proceedings of 7
th EUROSIM Congress on
Modelling and Simulation, Prague, September 6-10, 10 p.
12. Juuso, E., and Leiviskä, K., 2004, ‘Roadmap Contribution IBA A Intelligent Systems in
Production Industries’, EUNITE - European Network on Intelligent Technologies for
Smart Adaptive Systems, 75 p. http://ntsat.oulu.fi/index.php?24
13. Juuso, E., and Leiviskä, K., 2005, ‘Combining Monitoring and Process Data in Detecting
Operation Conditions in Process Industry’, in Maintenance, Condition Monitoring and
Diagnostics, Proceedings of the 2nd International Seminar, Oulu, Finland, pp. 145-156.
14. Ainali, I., Piironen, M., and Juuso, E., 2002, ‘Intelligent Water Quality Indicator for
Chemical Water Treatment Unit’, Proceedings of SIMS 2002 - the 43rd Scandinavian
Conference on Simulation and Modelling, Oulu, Finland, 26-27 September, pp. 247-252.
15. Juuso, E., and Lahdelma, S., 2009, ‘Cavitation Indices in Power Control of Kaplan
Water Turbines’, Proceedings CM 2009 – MFPT 2009, 6th International Conference on
Condition Monitoring & Machinery Failure Prevention Technologies, Dublin, Coxmoor,
Oxford, UK, pp 831-842.
16. Juuso, E. K., 1994, ’Fault Diagnosis based on Linguistic Equation Framework‘, Preprints
of IFAC Symposium on Fault Detection, Supervision and Safety for Technical Processes
SAFEPROCESS'94, Espoo, Finland, 13-16 June, pp. 374-379. Hakapaino, Helsinki.
17. Juuso, E. K., 1997, ’Intelligent Methods in Diagnostical Process Analysis‘, Proceedings of
the XIV IMEKO World Congress, New measurements – Challenges and Visions,
Tampere, June 1-6, Volume VII, pp. 1-6.
18. VDI 2056 Beurteilungsmaβstäbe für mechanische Schwingungen von Maschinen, VDI-
Richtlinien, Oktober 1964.
19. Collacott, R. A., 1977, ‘Mechanical Fault Diagnosis and condition monitoring’, Chapman
and Hall, London.
20. Juuso, E. K., Kivistö, M., and Lahdelma, S., 2004, ’Intelligent Condition Monitoring
Using Vibration Signals‘, Proceedings of EUNITE 2004, Aachen, Germany, pp. 381-390.
Verlag Mainz, Aachen.
21. Strackeljan, J., and Leiviskä, K., 2008, ‘Artificial immune system approach for the fault
detection in rotating machinery’, Proceedings CM 2008 – MFPT 2008, pp. 1365-1375.
22. Rami-Yahyaoui, O., Gebus, S., Juuso, E., and Ruusunen, M., 2008, ’Tuning of linguistic
equations for failure mode identification using genetic algorithms’, Proceedings CM 2008
– MFPT 2008, pp. 1543-1553.
23. Juuso, E., Ruusunen, M., and Perigot, G., 2010, ‘Linguistic Equation Models for Failure
Mode Identification from Multisensor Vibration Analysis, Proceedings of CM 2010
MFPT 2010, Stratford-upon-Avon, 13 p.
24. Ahola, T., Kumpula, H., and Juuso, E., 2004, ’Prediction of paper machine runnability by
identification of operating situations. Proceedings of AFNC'04 - The 2nd IFAC
Workshop on Advanced Fuzzy/Neural Control, 16-17 September, Oulu, Finland, pp. 205-
210.
... Degrees of membership for the activated cases can be used for estimating some process or product quality features. (25) ...
... Detecting operating conditions and faults(25) ...
Conference Paper
Full-text available
Automatic fault detection with condition indices enables reliable condition monitoring to be combined with process control. Useful information on different faults can be obtained by selecting suitable features from generalised norms, which are defined by the order of derivation, the order of the norm and sample time. The nonlinear scaling based on generalised norms and skewness extends the idea of dimensionless indices to nonlinear systems and provides good results for the automatic generation of condition indices. Condition indices, which are used in the same way as the process measurements in process control, detect differences between normal and faulty conditions and provide an indication of the severity of the faults. Feature specific health indices, which are calculated as ratios of feature values in the reference condition and the faulty case, are used in selecting efficient features. In the multisensor vibration analysis, the number of sensors and features were drastically reduced. The number of features is further reduced by optimal orders for the derivatives and norms. The complexity of the models is simultaneously reduced. For the supporting rolls of a lime kiln, an efficient indication of faulty situations is achieved with two features. All the rolls can be analysed with the same approach throughout the data set. The results of both the applications are consistent with the vibration severity criteria: good, usable, still acceptable, and not acceptable. Three standard deviations obtained for the signal x(4) on three frequency ranges were needed to detect unbalance and bearing faults. The norms based on the signal x(4) provide the best results in all the frequency ranges.
... Intelligent methods extend the idea of dimensionless indices to nonlinear systems: the basic idea is nonlinear scaling, which was developed to extract the meanings of variables from measurement signals . In the present systems, the scaling functions are developed by using generalised moments and norms (Juuso and Lahdelma, 2010;Juuso, 2013) and tuned with genetic algorithms (Juuso, 2009 The condition monitoring applications are similar with detecting operating conditions in the process industry (Juuso and Leiviskä, 2010). Detection of operating conditions can be extended by means of a Case-Based Reasoning (CBR) type application with linguistic equation (LE) models and fuzzy logic (Juuso, 1994(Juuso, , 1999. ...
Book
Full-text available
The papers published in these proceedings are presented in the International Conference on Maintenance, Condition Monitoring and Diagnostics, and Maintenance Performance Measurement and Management, MCMD 2015 and MPMM 2015, to be arranged in Oulu, Finland, in 30th September – 1st October, 2015. Arranged by the University of Oulu and POHTO – The Institute for Management and Technological Training, the present conference is supported by a variety of Finnish industrial enterprises.
... In the present systems, the scaling functions are developed by using generalised moments and norms (Juuso and Lahdelma, 2010;Juuso, 2013) and tuned with genetic algorithms (Juuso, 2009). The condition monitoring applications are similar with detecting operating conditions in the process industry (Juuso and Leiviskä, 2010). Process and condition monitoring data is combined in detecting operating conditions: measurements which require signal processing are denoted as signals, normal process measurements are directly used in feature extraction, and in addition some infrequent measurements need to be interpolated. ...
Article
The early detection of fluctuations in operating conditions and fault detection is done with similar methods. The feature extraction uses statistical analysis based on generalised norms and moments. Intelligent stress indices are calculated from these features by nonlinear scaling. The scaling approach uses the norms and moments to produce indices, which are consistent with the vibration severity criteria. Nonlinear scaling can be used for finding suitable control limits for the features and indices. Harmful high levels of stress are efficiently detected with control limits adjusted to the process requirements. The limits can be explained by fuzzy set systems and categorical information is included through knowledge-based analysis. The statistical process control (SPC) can be extended to nonlinear and non-Gaussian data: the new generalised SPC is suitable for a large set of statistical distributions. It operates without interruptions in short run cases and adapts to the changing process requirements. The approach is tested in two application cases: a rolling mill and an underground load haul dump (LHD) machine.
Article
Purpose – The purpose of this paper is to develop a comprehensive approach to efficiently integrate maintenance and operation by combining process and condition monitoring data with performance measures. Design/methodology/approach – Intelligent stress, condition and health indicators have been developed for control and condition monitoring by combining generalised moments and norms with efficient nonlinear scaling. The data analysis resulting nonlinear scaling functions can also be used to handle performance measures used for management. The generalised norms provide limits for an advanced statistical process control. Findings – The data‐driven analysis methodology demonstrates that management‐oriented indicators can be presented in the same scale as intelligent condition and stress indices. Control, condition monitoring, maintenance and performance monitoring are represented as interactive feedback loops. Practical implications – Performance analysis can be based on real‐time information by using various stress, condition and health indices as inputs. Similar approaches can be used for outputs: quality indices, harmonised indices, key performance indicators, process capability indices and overall equipment effectiveness. Since consistent linguistic explanations based on nonlinear scaling are available for all these indices, the analysis can be further deepened with LE modelling. Efficient monitoring with intelligent indices provides a good basis for control and condition‐based maintenance and performance monitoring. Originality/value – The paper extends the nonlinear scaling methodology and linguistic equations to intelligent performance measures. The methodology provides a consistent way to also represent all information with linguistic terms.
Conference Paper
Full-text available
Machine condition monitoring enables reliable and economical way of action for maintenance operations in modern industrial plants. Increasing number of measurement points and more demanding problems require automatic fault detection. Advanced signal processing methods exposed failures earlier and then it's possible to plan more operating time and less shutdowns. Intelligent methods have been increasingly used in model based fault diagnosis and intelligent analysers. Intelligent methods provide various techniques for combining a large number of features. A test rig was used to simulate different fault types and changes in operating conditions. Linguistic equation (LE) models were developed for the normal operation and nine fault cases including rotor unbalance, bent shaft, misalignment and bearing faults. Classification is based on the degrees of membership developed for each case from the fuzziness of the LE models. The classification results of the experimental cases are very good and logical. As even very small faults are detected by a slight increase of membership, the results are very promising for early detection of faults. Together with the compact implementation and the operability of the normal model, this makes the extension to real world problems feasible.
Conference Paper
Full-text available
Mechanical vibration is dynamic phenomena, i.e. their intensity varies with time. Most failures of rotating machinery are rooting in the damage of rolling element bearings. The widespread applications of rolling element bearings in both industry and commercial life require advanced technologies to efficiently and effectively monitor their health status. In practice, it is very difficult to avoid vibration during machine running conditions. Therefore, to know the level of damage on bearing due to machine running conditions, vibration measurement should be carried out by suitable sensors/transducers. Many problems arising in motor operations are linked to bearing faults. In this paper, bearing vibration frequency features are discussed for motor bearing fault diagnosis. This paper then presents an approach for motor rolling bearing fault diagnosis using neural networks and time/frequency-domain bearing vibration analysis. Vibration simulation is used to assist in the design of various motor rolling bearing fault diagnosis strategies.Record the vibration spectrum, specify the peaks corresponds to the bearing components, Record each component peak and frequency. By using the software and the standard limits, determine the trend of each peak. Determine the bearing state (good –need service –need change).
Conference Paper
Full-text available
Cavitation is harmful to water turbines and may cause shutdowns for several weeks. The real-time detection of cavitation risk is increasingly important, and even narrow cavitation-free power ranges can be utilised in load optimisation. Higher derivative signals x(α), α = 3 or 4, calculated from acceleration signals x(2) are very suitable for detecting impacts. Generalised moments and their lp norms, ‖τMαp‖, detect the normal operating conditions, which are free of cavitation, and also provide an early indication of cavitation risk. On-line cavitation monitoring is based on cavitation indices calculated from a moving maximum of the lp norms obtained from samples. Data compression is very efficient, as the detailed analysis only requires feature values with a short sample time, τ = 3s. The absolute mean, i.e. the order of moment p = 1, is very good in normal operating conditions and in the high power range. The cavitation indicator also provides warnings on short periods of cavitation if the power is not too low. This is readily suitable for intelligent sensors where a good solution is to use analogue signals x(4). The optimal norm (p = 2.75) is only needed in the low power range. Power control minimises the cavitation risk by dividing the load between three turbines, whose conditions are normal, bad and very good. Each turbine has three operating modes: low, normal and high power. In the normal area, a cavitation free power level is taken as an operating point. The low and high operating areas are defined by local minima of the cavitation indices. The control system has a feedforward controller, which allocates the load to the turbines by means of knowledge-based cavitation indices, and a feedback controller, which is based on the linguistic equation (LE) approach. Each turbine has a P type LE controller which is adapted to the operating conditions by the scaling functions. The cumulative time in the strong cavitation provides an indication of possible damage to be used in selecting the turbine for low power operation. The characteristic curves are adapted to the recent indices in order to handle the changes in the condition of the turbines. For power stations with many turbines, alternatives to reduce cavitation risks are evaluated by simulation to optimise maintenance actions.
Technical Report
Full-text available
Early fault mode identification presents a great interest for defect prevention or for planning maintenance operations. In case of non-linear systems, the amount and accuracy of measured parameters can be an issue. It is essential to identify the fault type, but it is also relevant to reduce the number of variables and thus the number of sensors. Redundant variables can result in a reduced generalization and a subsequent loss of predictive ability of the model. This report presents a case study in which a test rig equipped with multiple sensors is used to simulated different fault types. Fault identification is based on linguistic equations to create linear sub-regions. This approach is then enhanced by the use of genetic algorithms for variables selection. This approach enabled proper identification of faults with a minimum amount of sensors.
Conference Paper
Full-text available
Early detection of fluctuations in operating conditions are important in maintaining smooth production in process industry. Detecting can be done with similar method as fault detection although the classes do not necessarily correspond to any fault. Case-based reasoning (CBR) is used for same purpose for finding out the solution to a new problem by remembering a previous similar situation. Model-based approaches, especially intelligent methods, provide useful extensions for these approaches. Linguistic equations (LE) are suitable for modelling multivariable nonlinear systems. Indicators have been built for several applications by combining LE models wit fuzzy logic. The same methodology provides good results in detecting fluctuations of flavour ingredients in brewing, in predicting web break sensitivity in paper machines and in condition monitoring of machines.
Conference Paper
Full-text available
Automatic fault detection enables reliable condition monitoring even when long periods of continuous operation are required. Dimensionless indices provide useful information on different faults, and even more sensitive solutions can be obtained by selecting suitable features. These indices combine two or more features, e.g. root-mean-square values and peak values. Additional features can be introduced by analysing signal distributions, for example. The features are generated directly from the higher order derivatives of the acceleration signals, and the models can be based on data or expertise. Generalised moments and norms introduce efficient new features which even alone can provide good solutions with automation systems, but combining several easily calculated features is an efficient approach for intelligent sensors. The nonlinear scaling used in the linguistic equation approach extends the idea of dimensionless indices to nonlinear systems. Indices are obtained from these scaled values by means of linear equations. Indices detect differences between normal and faulty conditions and provide an indication of the severity of the faults. They can even classify different faults in case-based reasoning (CBR) type applications. Additional model complexity, e.g. response surface methods or neural networks, does not provide any practical improvements in these examples. The indices are calculated with problem-specific sample times, and variation with time is handled as uncertainty by presenting the indices as time-varying fuzzy numbers. The classification limits can also be considered fuzzy. Condition indices can be obtained from the degrees of membership which are produced by the reasoning system. Practical long-term tests have been performed e.g. for diagnosing faults in bearings, in supporting rolls of lime kilns and for the cavitation of water turbines. The indices obtained from short samples are aimed for use in the same way as the process measurements in process control. The new indices are consistent with the measurement index MIT and the health index SOL developed for condition monitoring.
Conference Paper
Full-text available
Automatic fault detection with condition indices enables reliable condition monitoring to be combined with process control. Useful information on different faults can be obtained by selecting suitable features. Generalised norms can be defined by the order of derivation, the order of the moment and sample time. These norms have the same dimensions as the corresponding signals. The nonlinear scaling used in the linguistic equation approach extends the idea of dimensionless indices to nonlinear systems. Condition indices are obtained from short samples by means of the scaled values and linear equations. Indices, which are used in the same way as the process measurements in process control, detect differences between normal and faulty conditions and provide an indication of the severity of the faults. The generalised norms represent the norms from the minimum to the maximum in a smoothly increasing way. The new nonlinear scaling methodology based on generalised norms and skewness provides good results for the automatic generation of condition indices. Additional sensitivity is achieved for the values which differ only slightly from the centre values. For cavitation, the new approach provides four levels from cavitation-free to clear cavitation. For the supporting rolls of a lime kiln, it provides an efficient indication of faulty situations. Sensitivity is also improved for small fluctuations. All the supporting rolls can be analysed using the same approach throughout the data set. The results of both the applications are consistent with the vibration severity criteria: good, usable, still acceptable, and not acceptable. Warning and alarm limits can be defined and fault types can also be identified with fuzzy set systems and specialised condition indices.
Chapter
The ability of operators to carry out preventive and corrective maintenance involves a balance between reliability and maintainability associated with the identification of hazard systems together with the provisioning of adequate spares. The assessment of such conditions is a design function involving the study of fault trees with the use of a failure modes, effects and criticality analysis (FMECA), itself an extension of the techniques of quality assurance [1].
Chapter
Vibratory effects are sensed quite easily by most people, they are tiring and unpleasant, at the extreme they can be frightening. It is not surprising that vibration analysis occupies a prominent place in machine diagnostics.
Conference Paper
A multisensor approach has been used for fault diagnosis in a test rig that consists of an electric motor and a transmission between two axes with SKF single row ball bearings 6002. Two sensors measured axial vibration and five accelerometers radial vibration in the vertical direction. The rig was to simulate nine independent fault modes: rotor unbalance at two levels, three coupling misalignment cases between the motor and input shaft, bent shaft, and three bearings faults. Excellent classification results have been achieved with case-based reasoning (CBR) approach using case models of normal operation and nine fault situations. Each case model consists of seven linguistic equations. Each sensor specific equation is based on the rotation speed, five features obtained from vibration measurements and a bias term. A large number of equations and features are not necessarily needed. The number of variables and sensors required to diagnose each fault can be reduced considerably by selecting the most sensitive features. Finally, three sensors are needed from seven, and totally seven features are used to calculate eight health indices Good conditions and all six fault types can be identified by seven rules, and the strength of misalignment and unbalance can be calculated. All the labels can be defined with fuzzy numbers. The LE system reduces to seven equations for diagnosing the faults and two equations for calculating the strengths of misalignment and unbalance. The classification results are in line with the previous studies with the same data set