0% found this document useful (0 votes)
10 views9 pages

Tinga

The document discusses how physical failure models can improve maintenance modeling and predictions by accounting for usage parameters and reducing uncertainty. It explains different maintenance strategies from condition-based to time-based and how physical models allow linking usage and loads to failure distributions. More accurate predictions result when strategies incorporate more information on usage severity and conditions from sensors and models.

Uploaded by

fredy
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)
10 views9 pages

Tinga

The document discusses how physical failure models can improve maintenance modeling and predictions by accounting for usage parameters and reducing uncertainty. It explains different maintenance strategies from condition-based to time-based and how physical models allow linking usage and loads to failure distributions. More accurate predictions result when strategies incorporate more information on usage severity and conditions from sensors and models.

Uploaded by

fredy
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/ 9

Advanced Maintenance Modelling

Advanced maintenance modelling based on physical failure


mechanisms

T. Tinga
Netherlands Defence Academy
PO Box 10000
1780 CA Den Helder, The Netherlands

Abstract

The efficiency of a preventive maintenance process largely depends on an accurate


prediction of the component replacement intervals. Component prognostics are
therefore applied both in fixed interval maintenance concepts and condition based
maintenance concepts. The traditionally used methods in maintenance modelling are
based on statistical descriptions of service failure data or stochastic formulations for
degradation processes. Since these methods lack a physical background, they are
generally not able to account for changes in usage of the system. The present paper
illustrates how physical failure models can assist in improving existing reliability
engineering methods. Quantitative knowledge of the failure mechanism combined
with the monitoring of usage severity is shown to enable the definition of usage
parameters that considerably increase the accuracy of the maintenance interval
predictions. Moreover, the use of understanding the physical behaviour for condition
based maintenance is discussed. The paper illustrates how the prediction of the
remaining life benefits from this knowledge.

Keywords: physics of failure, failure mechanism, prognostics, condition based


maintenance.

1. Introduction
In maintenance modelling, the planning of maintenance activities (replacement,
repair, inspection) is traditionally based on failure distributions of components. These
failure distributions are normally obtained by collecting service failure data of the
specific system. To obtain a sufficiently large data set, failures of a range of similar
systems are combined into one set. However, the usage and loading of the individual
systems often varies considerably, which results in a relatively wide failure
distribution. Consequently, application of this distribution in maintenance modelling
yields predictions with a large uncertainty.
In previous work [1, 2] it was demonstrated that application of physical failure
models in combination with monitoring of usage or loads leads to a considerable
reduction of this uncertainty. As a result, more accurate predictions of maintenance
intervals can be obtained. In the present paper, this work will be extended. It will be
demonstrated how existing reliability engineering methods can be improved by
choosing the appropriate usage parameters or by developing more sophisticated

1
Proceedings of the 38th ESReDA Seminar, Pecs, May 4-5, 2010

physical model-based usage parameters. Moreover, it is shown that care must be


taken when failure distributions are transformed to another usage parameter. Finally,
it is discussed how physical failure models can assist in the prognostic part of
condition based maintenance.

2. Reliability and physical models


The key issue of preventive maintenance concepts is the determination of the length
of maintenance intervals. To guarantee the effectiveness of a maintenance concept
(quantified by a certain allowable probability of failure), it has been shown [1] that
more conservatism is required when the uncertainty about the degradation process
increases. This was demonstrated with a gas turbine blade case study, in which five
preventive maintenance strategies have been compared. These strategies are described
below, where the uncertainty about the actual level of degradation increases form the
first to the last:
 Condition based maintenance (CBM): the condition of the system is assessed in
a (almost) direct way, by using appropriate sensors (e.g. crack length sensors).
The amount of uncertainty is limited and maintenance can be performed just-in-
time.
 Load based maintenance (LBM): the usage severity during operation of a
system is taken into account by monitoring the actual relevant internal loads in the
component (e.g. temperature, strain, etc.). A suitable physical failure model then
yields the life consumption associated to the measured loads. Since the actual
condition is calculated in stead of assessed directly, the amount of uncertainty is
somewhat larger than for CBM. The uncertainty depends on the accuracy of the
damage model used.
 Usage severity based maintenance (USBM): the usage severity during operation
of a system is taken into account by monitoring the variation in usage (e.g. power
setting, rotational speed). This requires knowledge of the relation between usage
severity and life consumption, which is provided by suitable physical models. The
uncertainty is again somewhat larger than for LBM, since not only the damage
accumulation, but also the internal loads are calculated. Therefore, the numerical
models used determine the amount of uncertainty.
 Usage based maintenance (UBM): the usage is incorporated in the determination
of the maintenance intervals, e.g. by using the number of operating hours or
cycles. However, the severity of the usage is not accounted for, so in addition to
the uncertainties in the load calculation and life assessment, also uncertainty in
the usage severity affects the final result.
 Calendar time based maintenance (CTBM): fixed periods of calendar time are
used, e.g. every month or every year. This means that no information about the
usage or loads is considered. It is even unknown whether or not the system is
actually operated in a certain period. As a consequence, the uncertainty in the
failure distribution is very large.
The large gain in maintenance process efficiency of the first three strategies was
shown to be caused by the reduction of the uncertainty in the component failure
distribution [1]. The distributions for the five strategies are represented by the normal
distributions shown in Figure 1, demonstrating that the distributions get narrower
when more information about the usage or loads is available. The very sharp
distribution associated to CBM indicates that the moment of failure can be predicted
quite accurately, whereas that is much more difficult for UBM or CTBM. Note that

2
Advanced Maintenance Modelling

all five distributions have the same mean value here, which is generally not the case
in practice, as will be discussed later.
The failure distributions f(t) in Figure 1a can also be plotted as reliability
curves as is shown in Figure 1b. The reliability R(t) is defined as

t
R (t ) = 1 − F (t ) = 1 − ∫ f ( x)dx (1)
0

where f(x) and F(t) are the probability density (PDF) and cumulative density (CDF)
functions, respectively. The reliability curves can be used to determine the moment in
time where the reliability drops below a critical level, e.g. 95% as indicated by the
dashed line. At this point action should be taken to prevent regular failures. The
figure shows that an increasing uncertainty requires action at an earlier point in time
and thus yields a less efficient maintenance process.

3.00E-04
CBM 1.0
LBM
2.50E-04
USBM
UBM 0.8
2.00E-04 CTBM

0.6
PDF

R(t)

1.50E-04

0.4 CBM
1.00E-04
LBM
USBM
5.00E-05 0.2 UBM
CTBM

0.00E+00 0.0
0 10000 20000 30000 40000 50000 60000 70000 0 10000 20000 30000 40000 50000
Calendar time (hrs) Calendar time (hrs)

Figure 1. a) Failure distributions associated to different maintenance concepts; b) Reliability


curves for the five strategies.

2.1 Usage parameters


The total uncertainty in a failure distribution is the combined effect of a range of
uncertainties in the underlying processes, like usage, loads and material properties.
However, if a number of these uncertainties can be quantified, the total uncertainty
can be reduced considerably. The challenge is therefore to find the usage parameter
that provides the best correlation with the failure behaviour.
That is actually the way in which the distribution width is reduced in Figure
1a when going from CTBM to CBM. For example, for a fatigue failure dominated
component, the number of operating hours (UBM) generally does not correlate very
well with the service life. However, the correlation with the number of stress cycles
(LBM) is much better, which means that number of cycles is a better usage parameter
than operating hours in this case. As a result, plotting the failure distribution in terms
of number of cycles provides a much narrower failure distribution. Depending on the
actual usage and loads, the time to failure can be predicted quite accurately, as
represented by a narrow distribution located somewhere along the original wide UBM
distribution. This is illustrated in Figure 2, where three LBM distributions, for a low,
middle and high usage severity (i.e. number of cycles), are plotted together with the
UBM distribution.

3
Proceedings of the 38th ESReDA Seminar, Pecs, May 4-5, 2010

The selection of the appropriate usage parameter is only possible when the
physical failure mechanisms are understood and the internal loads that govern this
mechanism are known. In previous work [3] an overview of several common failure
mechanisms and their governing loads is given. In summary, these are:
• Fatigue  number and magnitude of stress or strain cycles
• Creep  temperature, stress level and time
• Wear  normal load and sliding distance
• Corrosion  exposure time and environmental conditions
This knowledge enables the first improvement of the maintenance process efficiency,
since replacement intervals of components can be calculated in terms of the variations
in the dominating load: for fatigue in terms of cycles, for creep and corrosion in terms
of operating hours. However, in this way only the usage is taken into account,
whereas the usage severity (how damaging is each cycle or operating hour ?) is not
considered. To quantify the differences in usage severity, a physical failure model can
be used to define a more sophisticated usage parameter. For the case study of the
creep dominated gas turbine blade that was used in [1], the development of a
sophisticated usage parameter will be discussed in the next section.

2.2 Usage parameter for a gas turbine blade


Physical models have been selected to calculate the internal loads and the resulting
creep strain accumulation in a low pressure turbine blade. The blades rotate at speeds
in the order of 10.000 revolutions per minute (rpm) and operate in a hot gas stream
with temperatures in the range from 500 to 1000 oC. Due to the rotation, a centrifugal
force acts on the turbine blades, causing radially directed normal stresses in the root
of the blade. The magnitude of these stresses depends on the mass (m) of the blade,
the rotational speed (ω), the distance from the blade to the engine centre axis (r) and
the area of the blade cross section (A) in the following way

F mω 2 r
σ= = (2)
A A

The rotational speed and gas temperature depend on the power setting of the gas
turbine. Since the turbine blades are solid and uncooled, the blade temperature will
equal the gas temperature in a steady-state situation.
For the present component, creep is the life limiting failure mechanism. Creep
is a high temperature deformation process that depends on both the stress and
temperature level. The creep strain rate for the present material is described by a
Norton creep law and depends on the temperature (T) and stress (σ) as follows

d ε cr
ε&cr = = AT 4σ 2 (3)
dt

where the value of the constant A equals 10-20 (MPa)-4 (oC)-2 (hour)-1. A creep strain
of 1.0 % is defined as the critical amount of creep deformation (εcr), leading to an
unacceptable elongation of the blade. The time to failure (tf) can then be calculated as

ε cr
tf = (4)
ε&cr

4
Advanced Maintenance Modelling

The accumulated amount of damage (D) can be obtained using Robinson’s damage
rule [4]:
∆ti
D=∑ (5)
i t f ,i

where ∆ti is the time period spent at some condition (stress and temperature) and tf,i
the failure time at those conditions. Failure will occur when the damage parameter D
attains the value 1.
The usage of the gas turbine is defined in terms of operating hours per year and
a usage severity based on the fractions of time spent at low, middle and high power
settings, respectively. The induced blade loads and resulting creep damage
accumulation are shown in Table I.

Table I: Turbine blade loads and creep damage accumulation at different power settings.
Power Rotational speed Stress Blade temperature Creep strain rate Failure time
setting [rad/s] [MPa] [oC] [hour-1] [hours]
low 597 102 500 1.27 ⋅10-7 78587
middle 733 154 750 9.72 ⋅10-7 10293
high 1047 313 900 4.11 ⋅10-6 2432

The creep strain rates are a good indication for the usage severity, so these quantities
are used here to define a usage parameter in terms of equivalent operating hours:

cr
cr
ε&low cr
ε&mid ε&high
∆teq = cr ∆tlow + cr ∆tmid + cr ∆thigh (6)
ε&high ε&high ε&high

where ∆tlow , ∆tmid and ∆thigh are the numbers of hours spent at the three power settings
and ε&icr are the associated creep strain rates. Using the numbers form Table I, the
following expression for the equivalent operating hours is obtained:

∆teq = 0.03 ∆tlow + 0.24 ∆tmid + 1.00 ∆thigh (7)

If all operating hours would be run at high power, ∆teq equals the number of operating
hours. If the machine is operated at lower power settings for a certain part of the time,
∆teq will be lower than the number of actual operating hours. From Table I it appears
that at the high power setting the failure time is 2432 hours. Since the usage
parameter in (6) is normalized to this high power setting, the failure time in terms of
equivalent hours is also 2432 hours.

5
Proceedings of the 38th ESReDA Seminar, Pecs, May 4-5, 2010

2.50E-04 35000
LBM
LBM2 30000
2.00E-04 LBM3
UBM
25000

Time to failure (hrs)


1.50E-04 20000
PDF

15000
1.00E-04

10000

5.00E-05
5000

0
0.00E+00
0 0.2 0.4 0.6 0.8
0 10000 20000 30000 40000 50000 60000 70000
Fraction of operating hours at low power
Operating time (hrs)

Figure 2. Three LBM failure distributions (for low, Figure 3. Effect of usage severity on
middle and high usage severity) compared to the gas turbine blade creep life.
UBM distribution.

When the three LBM distributions in Figure 2 represent the turbine blade failure
behaviour, they would collapse onto one distribution if they would be plotted versus
the proposed usage parameter (equivalent operating hours). They all have 2434
equivalent hours to failure, but the left hand side curve represents a usage with many
hours at high power, whereas the right hand side curve represents a usage with
extensive operation at low power. If only the operating hours are monitored (UBM),
the usage severity is unknown and the exact life time of one specific system is much
more uncertain, resulting in the observed wide distribution.
Now the effect of the usage on the life time of the blade is known, a sensitivity
study is performed with different usage profiles. It is assumed that always 20% of the
operating hours are run at middle power. Then the fraction of time operating at low
power is varied from 0 to 80% (which means that the remaining hours are run at high
power). Using the equivalent time parameter as defined in (6), the resulting service
life of the gas turbine is calculated, as is shown in Figure 3. It can be observed that for
low fractions of time at low power, the service life approaches the 2434 hours service
life associated to the high power setting. For an increasing fraction of time at low
power, the service life considerably increases.
In this sensitivity study, it is assumed that the power ratio’s are fixed during
the complete service life. In practice this will vary in the course of time, which means
that monitoring the actual ratio’s provides the information to assess the damage
accumulation in the blades. It should be stressed that the proposed usage parameter
can only be applied when the hours at the three power settings are actually monitored.
Therefore, the accuracy and efficiency can be increased by developing and applying
sophisticated usage parameters, but not without an additional effort in monitoring.
Further, the usage parameter presented here takes into account the severity of
the usage to a certain extent. The power settings are divided into three classes only,
which for some applications may be too course. However, it may be relatively easy to
monitor these three classes in a USBM maintenance strategy. If the internal loads can
be monitored directly (LBM), a much finer classification of the usage variations is
possible. Finally, in this section a usage parameter has been presented for a creep
failure dominated component. Similar parameters can be defined for fatigue
(equivalent cycles) or wear (equivalent sliding distance).

6
Advanced Maintenance Modelling

2.3 Life exchange rates


For a system with several critical components, the reliability of the complete system
depends on the reliabilities of the individual components, as is generally calculated
using reliability block diagrams (RBD). For three components that all are critical to
the functionality of the system, the RBD consists of three reliabilities Ri(t) in series.
The system reliability Rs(t) is then obtained by

3
Rs (t ) = ∏ Ri (t ) (8)
i =1

This equation assumes that all individual component reliabilities are given as function
of the same variable t (generally representing operating time). However, in the
previous section it has been stated that the selection of the appropriate usage
parameter is very important. Therefore, it is quite probable that the reliabilities of the
three components are given as a function of three different variables, e.g. hours,
cycles and kilometres. In that case, equation (8) can not be used to calculate Rs(t) and
a Life Exchange Rate Matrix (LERM) is often used [5]. This LERM is a matrix
containing factors to exchange two parameters, e.g. hours and cycles. The entries ri,j
have the following meaning: 1 life unit of parameter i (ti) equals ri,j times a life unit of
parameter j (tj). Therefore, equation (8) can be rewritten as

3 3
Rs (ti ) = ∏ R j (t j ) = ∏ R j (rj ,iti ) (9)
j =1 j =1

Apart from application of the LERM in RBDs, this theory can also be used to
transform a single reliability (or failure distribution) from one to another usage
parameter. For example, a component reliability in terms of cycles can be
transformed into an hour-based reliability if the number of cycles per hour are known.
However, this can not be done unconditionally, as will be discussed next.
In section 2.2 it was argued that always the most specific usage parameter
should be used to minimize the uncertainty in the failure distribution and reliability
functions. The actual properties of the distribution can then be obtained by either a
calculation (using the physical models) or measuring the failure times. However, in
the latter case, the required narrow distribution is only obtained when the actual usage
(severity) is monitored, thus enabling the assessment of the selected usage
parameter. Equation (9) suggests that a distribution in terms of a less specific usage
parameter (e.g. operating hours in stead of cycles for a fatigue case) can be
transformed into the required reliability function, but in that case the required narrow
distribution is not obtained. Only a distribution equivalent to the wider original
distribution is obtained, since the actual severity of the usage is still not specified.
On the other hand, using the LERM to transform a distribution in terms of a
very specific usage parameter to a more general parameter (e.g. from cycles to hours
for the fatigue case) yields a quite narrow distribution based on the specific value of
the exchange rate used. However, this is quite dangerous. If for some reason the
exchange rate changes (e.g. more cycles per hour), the original distribution in terms
of cycles remains the same, but the transformed distribution in terms of operating
hours would have to shift and is thus not representative anymore. Failure will occur at
a lower number of operating hours, since the critical number of cycles will have

7
Proceedings of the 38th ESReDA Seminar, Pecs, May 4-5, 2010

accumulated earlier. If the assumption of the fixed exchange rate is unknown,


application of the hour-based distribution may lead to unconservative maintenance
planning and associated failures.
For example, if for a gas turbine the average cycle to hour ratio has been
determined by monitoring the fleet of a specific operator, the cycle-based reliability
can be transformed into an hour-based reliability using that average exchange rate.
But if one of the gas turbines is operated such that much more start-stops are made
than on average, the time-based reliability is not representative for this machine
anymore. Also, another operator probably uses the same gas turbines quite
differently, which also means that the hour-based reliability can not be used to
perform an accurate analysis.
The conclusion of this subsection is that the use of exchange rates seems to be a
convenient way to transform distributions, but care should be taken when applying it.
On the one hand, the exchange rates can not be used to transform distributions in
terms of general usage parameters into the more precise distributions in terms of a
specific usage parameter. And on the other hand, transforming back specific
distributions into more general distributions is dangerous, since the transformation is
based on one specific exchange rate, which is not representative for many other
situations.

3. Condition based maintenance


Also in condition based maintenance (CBM) concepts the knowledge of physical
failure mechanisms can improve the process efficiency. Due to the recent boost in
performance and availability of sensors and other hardware, condition monitoring
systems are offered by OEMs in several industries as the way to increase maintenance
efficiency. However, many operators, and even some manufacturers, do not realize
that condition based maintenance can only be performed in an efficient way when
both a condition monitoring system (the sensors) and a prognostic method are
available. The sensor data only provides information about the current state of the
component. Just waiting till the moment that a monitored condition parameter
exceeds a critical value means that immediate action is required (i.e. a very short P-F
interval), which is difficult to plan (e.g. personnel, spare parts) and may have serious
consequences for the system availability. Therefore, a prognostic method is required
to determine when future maintenance activities are necessary.
In traditional condition monitoring systems, like vibration analysis, trending
methods and growth models are used to extrapolate trends in monitored condition
parameters (e.g. vibration levels) to determine component replacement or repair
intervals. But similar to the statistical and stochastic methods used in reliability
engineering, these trending methods lack a physical basis and assume some model
form like linear, exponential, etc. (this is often called the black box model).
Moreover, the trends are based on historical data, which enables only accurate
predictions when the future usage of the system is similar to the past usage. For
systems that are operated in a variable way, reliable predictions are rather difficult to
make in this way.
Also in this case, knowledge of the physical failure mechanisms can assist in
improving the maintenance efficiency. Having access to a quantitative relation
between loads and degradation rate enables accurate component prognostics for
condition based maintenance (which turns the black box model into a grey box
model). Changes in usage or loading of the component can easily be incorporated in

8
Advanced Maintenance Modelling

the prognostic method and the prediction of the maintenance activities no longer
depends on past usage only.
The large difference with the prognostic methods used for fixed maintenance
intervals is the availability of the monitoring data. This data specifies the actual
condition of the component at any moment, which provides two important
advantages. Firstly, the prediction only covers a fraction of the total component
service life, starting from a known condition. This considerably increases the
accuracy of the prognosis, especially when approaching the end of life, since the
uncertainties in, for example, usage and material properties are reduced.
Secondly, the monitoring data may be used to validate and update the
prognostic model. Provided that the relevant loads on the component are also
monitored (or can be calculated from any other monitored parameter), the present
monitored condition can be compared to the present calculated condition. Any
deviation between the measured and calculated condition can be used to update the
prognostic model, which reduces the uncertainty in all predictions to come. In this
way, an accurate model is obtained quite rapidly. The only requirements are a
representative physical model and the registration of the governing loads. If the
physical model does not represent the failure mechanism correctly, the calculations
will always deviate from the measurements and updating the model parameters is not
possible. Monitoring the loads requires some additional effort, but ultimately yields a
significant benefit.
To summarize, physical model based methods improve the accuracy of the
prognostics in a condition based maintenance concept, since variations in usage can
easily be incorporated. At the same time, the monitoring data can be applied to
validate and update the used model.

4. Conclusion
In the present paper, the benefit of applying physical failure models in advanced
maintenance modelling is discussed. In reliability engineering, sophisticated usage
parameters can be developed, that considerably reduce the uncertainty in failure
distributions. In condition based maintenance concepts, the physical models can be
applied to perform the prognostics that are required to determine the remaining life of
the system.

References
[1] Tinga, T. The role of physical failure models in optimizing maintenance. in
Asset Optimization and Maintainability. 2009. Baden, Switzerland: European
Safety, Reliability and Data Association.
[2] Tinga, T., Application of physical failure models to enable usage and load based
maintenance. Reliability Engineering and System Safety, submitted (2009): p.
1-20.
[3] Tinga, T., Physical model based component prognostics, in Advanced
Maintenance Modelling, J. Andrews, Editor. 2010, ??: London. p.??
[4] Robinson, E.L., Effect of Temperature Variation on the Long-Time Rupture
Strength of Steels. Transactions of the ASME, 74 (1952), 5: p. 777-781.
[5] Kumar, U.D., Reliability, maintenance and logistic support; a life cycle
approach. 2000, Norwell: Kluwer Academic Publishers.

You might also like