5

Condition-based Maintenance Modelling

Wenbin Wang

5.1 Introduction
The use of condition monitoring techniques within industry to direct maintenance
actions has increased rapidly over recent years to the extent that it has marked the
beginning of what is likely to prove a new generation in production and
maintenance management practice. There are both economic and technological
reasons for this development driven by tight profit margins, high outage costs and
an increase in plant complexity and automation. Technical advances in condition
monitoring techniques have provided a means to achieve high availability and to
reduce scheduled and unscheduled production shutdowns. In all cases, the
measured condition information does, in addition to potentially improving decision
making, have a value added role for a manager in that there is now a more
objective means of explaining actions if challenged.
In November 1979, the consultants, Michael Neal & Associate Ltd published
‘A Guide to Condition Monitoring of Machinery’ for the UK Department of Trade
and Industry, Neal et al (1997). This groundbreaking report illustrated the
difference in maintenance strategies (e.g., breakdown, planned, etc.) and suggested
that condition based maintenance, using a range of techniques, would offer
significant benefits to industry. By the late 1990’s condition based maintenance
had become widely accepted as one of the drivers to reduce maintenance costs and
increase plant availability. With the advent of e-procurement, Business to Business
(B2B), Customer to Business (C2B), Business to Customer (B2C) etc., industry is
fast moving towards enterprise wide information systems associated with the
internet. Today, plant asset management is the integration of computerised
maintenance management systems and condition monitoring in order to fulfill the
business objectives. This enables significant production benefits through objective
maintenance prediction and scheduling. This positions the manufacturer to remain
competitive in a dynamic market.
Today there exists a large and growing variety of condition monitoring
techniques for machine condition monitoring and fault diagnosis. A particularly
2 Book Title

popular one for rotating and reciprocal machinery is the vibration analysis.
However, irrespective of the particular condition monitoring technique used, the
working principle of condition monitoring is the same namely, condition data
becomes available which needs to be interpreted and appropriate actions taken
accordingly. There are generally two stages in condition based maintenance. The
first stage is related to condition monitoring data acquisition and its technical
interpretations. There have been numerous papers contributing to this stage, as
evidenced by the proceedings of COMADEM over recent years. This stage is
characterised by engineering skill, knowledge and experience. Much effort of the
study at this stage has gone into determining the appropriate variables to monitor,
Chen et al (1994), the design of systems for condition monitoring data acquisition,
Drake et al (1995), signal processing, Wong et al (2006), Samanta et al (2006),
Harrison (1995), Li and Li (1995), and how to implement computerised condition
monitoring, Meher-Homji et al (1994). These are just a few examples and no
modelling is explicitly entered into the maintenance decision process based upon
the results of condition monitoring. For detailed technical aspects of condition
monitoring and fault diagnosis, see Collacott (1997). The second stage is
maintenance decision making, namely what to do now given that condition
information data and its interpretations are available. The decision at this stage can
be complicated and entails consideration of cost, downtime, production demand,
preventive maintenance shutdown windows, and most importantly, the likely
survival time of the item monitored. Compared with the extensive literature on
condition monitoring techniques and their applications, relatively little attention
has been paid to the important problem of modelling appropriate decision making
in condition based maintenance.
This chapter focuses on the second stage of condition monitoring, namely
condition based maintenance modeling as an aid to effective decision making. In
particular, we will highlight a modelling technique used recently in condition based
maintenance, e.g. residual life modelling via stochastic filtering, Wang and
Christer (2000). This is a key element in modeling the decision making aspect of
condition based maintenance. The chapter is organised as follows. Section 5.2
gives a brief introduction to condition monitoring techniques. Section 5.3 focuses
on condition based maintenance modeling and discuss various modeling
techniques used. Section 5.4 presents the modelling of the residual life conditional
on observed monitoring information using stochastic filtering. Section 5.5
concludes the chapter with a discussion of topics for future research.

5.2 Condition monitoring techniques

For many years condition monitoring has been defined as “The assessment on a
continuous or periodic basis of the mechanical and electrical condition of
machinery, equipment and systems from the observation and/or recordings of
selected measurement parameters”, Collacott (1997). One of the obvious analogies
is the temperature measurement of a human body where the observation is the
temperature and the system is the human body. Just as doctors strongly recommend
periodic checks of key health parameters such as blood pressure, pulse, weight
 Chapter Title 3

and/or temperature for an early indication of potential health problems, for

industrial equipment some measurements can be taken and the likely condition of
the plant assessed.
Today, there exists a large and growing variety of forms of condition
monitoring techniques for machine condition monitoring and fault diagnosis.
Understanding the nature of each monitoring technique and the type information
measured will certainly help us when establishing a decision model. Here we
briefly introduce five main techniques and among them, vibration and oil analysis
techniques are the two most popular ones.

5.2.1 Vibration based monitoring

Vibration based monitoring is the main stream of current applications of condition

monitoring in industry. Vibration based monitoring is an on (off) line technique
used to detect system malfunction based on measured vibration signals.
Generally speaking, vibration is the variation with time of the magnitude of a
quantity that is descriptive of the motion or position of a mechanical system, when
the magnitude is alternatively greater than and smaller than some average value or
reference.
Vibration monitoring consists of essentially in identifying two quantities:
- the magnitude (overall level) of the vibration
- the frequency content (and/or time waveform)
The magnitude is basically used for establishing the severity of the vibration
and the frequency content for the cause or origin. Vibration velocity has been seen
as the most meaningful magnitude criterion for assessing machine condition,
though displacement or acceleration is also used. The magnitude of vibration is
usually measured in Root Mean Square (rms). If T denotes the period of vibration
and V (t ) is the vibration (say, velocity) measured at time t, then

T
Vrms = 1
T ∫0
(V (t )) 2 dt ,

which is proportional to the energy of vibration, Charles and Reeves (1998).

However, since vibration signals from machines are, in general, periodic in
nature, a great deal of information is contained in its frequency spectrum form. The
frequency spectrum is usually obtained digitally using a digital analyser or
computer via a mathematical algorithm known as “Fast Fourier Transform” (FFT).
The spectrum analysis of vibration signals is commonly used in the fault diagnosis
of rotating machines. Potentially, all machines can benefit from vibration
monitoring except, perhaps, those running at very low speed (below about 20
rev/min), and those where isolation (or damping) occurs between the source and
the sensor.
From observed vibration signals we often see a typical two-stage process where
the signals may stay flat over the normal operation period and then display some
increasing trend when a defect has initiated, Wang (2002). Another factor coming
4 Book Title

into play when establishing a vibration based maintenance model is the casual
relationship between the measured signals and the state of the plant. It is the defect
which causes the abnormal signals, but not vice versus, Wang (2002). This factor
plays an important role when selecting an appropriate model for describing such a
relationship.

5.2.2 Oil based monitoring

A detailed analysis of a sample of engine, transmission and hydraulic oils is a

valuable preventive maintenance tool for machines. In many cases it enables the
identification of potential problems before a major repair is necessary, has the
potential to reduce the frequency of oil changes, and increase the resale value of
used equipment.
Oil based monitoring involves sampling and analyzing oil for various
properties and materials to monitor wear and contamination in an engine,
transmission or hydraulic system etc. Sampling and analyzing on a regular basis
establishes a baseline of normal wear and can help indicate when abnormal wear or
contamination is occurring. Oil analysis works as follows. Oil that has been inside
any moving mechanical apparatus for a period of time reflects the possible
condition of that assembly. Oil is in contact with engine or mechanical components
as wear metallic trace particles enter the oil. These particles are so small they
remain in suspension. Many products of the combustion process also will become
trapped in the circulating oil. The oil becomes a working history of the machine.
Particles caused by normal wear and operation will mix with the oil. Any
externally caused contamination also enters the oil. By identifying and measuring
these impurities, one can get an indication of the rate of wear and of any excessive
contamination. An oil analysis also will suggest methods to reduce accelerated
wear and contamination.
The typical oil analysis tests for the presence of a number of different materials
to determine sources of wear, find dirt and other contamination, and even check for
the use of appropriate lubricants. Today there exist a variety of forms of oil based
condition monitoring methods and techniques to check the volume and nature of
foreign particles in oil for equipment health monitoring. There are spectrometric oil
analysis, scan electron microscopy/energy dispersive x-ray analysis, energy
dispersive x-ray fluorescent, low powered optical microscopy, and ferrous debris
quantification. One purpose of the oil analysis is to provide a means of predicting
possible impending failure without dismantling the equipment. One can "look
inside" an engine, transmission or hydraulic systems without taking it apart.
For oil based monitoring, there is no such a clear cut distinction between
normal and abnormal operating based on observed particle information in the oil
samples. The foreign particles that accumulate in the lubricant oil increase
monotonically so that we may not able to see a two-stage failure process as seen in
the vibration based monitoring. The casual relationship between the measured
amount of particles in the oil and the state of the plant may also be bilateral in that,
for example, the wear may cause the increase of observed metals in the oil, but the
metals and other contaminants in the oil may also accelerate the wear. This marks a
 Chapter Title 5

difference when modeling the state of the plant in oil based monitoring compared
to vibration based.

5.2.3 Other monitoring techniques

The other popular condition monitoring techniques are infrared thermography,

accustics and the motor current analysis.
The basis of infrared thermography is quite simple. All objects emit heat or
infrared electro-magnetic energy, but only a very small proportion of this energy is
visible to a naked eye. At low temperatures in order to ‘see’ the heat being emitted
an infrared camera must be used. The camera detects the invisible thermal energy
and converts it to a visible image on a screen. The image can then be analyzed to
identify any abnormality.
The acoustic emission (AE) based method is used widely for monitoring the
condition of rotating machinery. Compared to traditional vibration based methods,
the high frequency approach of AE has the advantage of a significant improvement
in signal to noise ratio. It can also be used for non-rotating machinery where defect
activities do not generate distinct repetition frequencies and hence FTT analysis
cannot be used. An item to note is that AE transducers need to have a relatively
narrow band to be able to detect high frequency faults.
The Motor current noise signature analysis methods and apparatus for
monitoring the operating characteristics of an electric motor-operated device, such
as a motor-operated valve, have been frequently used for early detection of rotor
related faults in AC induction Motors. Frequency domain signal analysis
techniques are applied to a conditioned motor current signal to distinctly identify
various operating parameters of the motor driven device from the motor current
signature. The signature may be recorded and compared with subsequent
signatures to detect operating abnormalities and degradation of the device. This
diagnostic method does not require special equipment to be installed on the motor-
operated device, and the current sensing may be performed at remote control
locations, e.g., where the motor-operated devices are used in unaccessible or
hostile environments.
All the techniques briefly introduced above can offer some help for indicating
the current state or condition of the plant monitored. Based on the technical
analysis of the observed condition monitoring data, a maintenance decision has to
be made to maintain the plant in a cost effective way. We discuss in the next
section, how modeling can be used to support such a decision making utilizing
available monitoring information.

5.3 Condition based maintenance modelling

There is a basic, but not always clearly answered question in condition monitoring,
that is: what is the purpose of condition monitoring? Have we lost sight of the
ultimate need? Condition monitoring is not an end itself, it involves an expenditure
entered into by the managers in the belief that it will save them money. How is this
saving achieved? It can be obtained by using monitored condition information to
6 Book Title

optimise the maintenance to achieve minimum breakdown of the plant with

maximum availability for production, and to ensure that maintenance is only
carried out when necessary. This is what one calls condition based maintenance
which contrasts with the traditional break down or time based maintenance policies
where maintenance is only carried out when it becomes necessary utilizing
available condition information. But in reality, all too often we see effort and
money spent on monitoring equipment for faults which rarely occur, and we also
see planned maintenance being carried out when the equipment is perfect healthy
though the monitored information indicates something is ''wrong''. A study of oil
based condition monitoring of gear boxes of locomotives used by Canadian Pacific
Railway indicated, Aghjagan (1989), that since condition monitoring was
commissioned (entailed 3-4 samples per locomotive per week, 52 weeks per year),
the incidence failure of gear boxes while in use fell by 90%. This is a significant
achievement. However, when subsequently stripped down for
reconditioning/overhaul, there was nothing evidently wrong in 50% of cases.
Clearly, condition monitoring can be highly effective, but may also be very
inefficient at the same time. Modelling is necessary to improve the cost
effectiveness and efficiency of condition monitoring.

5.3.1 The decision model

This is an extension to the age-based replacement model in that the replacement

decision will be made not only dependent upon the age, but also upon the
monitored information, plus other cost or downtime parameters. If we take the cost
model as an example, then the decision model amounts to minimising the long run
expected cost per unit time. We use the following notation:
c f : the mean cost per failure;
c p : the mean cost per preventive replacement;
c m : the mean cost per condition monitoring;
t i : the ith and the current monitoring point;
Yi : monitored information at t i with y i of its observed value;
ℑi : history of observed condition variables to t i , ℑi = { y1 ,..., yi } ;
X i : the residual life at time t i ;
pi ( xi | ℑi ) : pdf of X i conditional on ℑi ;

The long term expected cost per unit time, C (t ) , given that a preventive
replacement is scheduled at time t> t i is given by, Wang (2003),

(c f − c p ) P(t − ti | ℑi ) + c p + icm
C (t ) = t −ti
(5.1)
ti + (t − ti )(1 − P (t − ti | ℑi )) + ∫ xi pi ( xi | ℑi )dxi
0
 Chapter Title 7

t −ti
where P (t − ti | ℑi ) = P ( X i < t − ti | ℑi ) = ∫
0
pi ( xi | ℑi )dxi , which is the
probability of a failure before t conditiional on ℑi . The right hand side of (5.1) is
the expected cost per unit time formulated as a renewal reward function, though the
life times are independent but not identical.
The time point t is usually bounded within the time period from the current to
the next monitoring since a new decision shall be made once a new monitoring
reading becomes available at time t i +1 .
In general, if a minimum of C (t ) is found within the interval to the next
monitoring in terms of t , then this t should be the optimal replacement time. If no
minimum is found, then the recommendation would be to continue to use the plant
and evaluate (5.1) at the next monitoring point when new information becomes
available. For a graphical illustration of the above principle see Fig. 5.1.

C(t)

No replacement is recommended

Optimal replacement time

t i Current time t* Next monitoring time t i +1 t

Fig. 5.1 A graph to show the optimal replacement time

Obviously the key element in (5.1) is the determination of pi ( xi | ℑi ) , which is

the topic of the next two sections,

5.3.2 Modelling pi ( xi | ℑi )

Before we proceed to the discussion of the modelling of pi ( xi | ℑi ) , there are few

issues that need clarification.
The first relates to the concept of direct and indirect monitoring, Christer and
Wang (1995). In direct monitoring, the actual condition of the item, say the depth
of a brake pad, can be observed, and a critical level, say C , can be set up. While in
the indirect monitoring case we can only collect measures related to the actual
8 Book Title

condition of the item monitored in a stochastic manner. For example in the

vibration monitoring case, if a high vibration signal is observed we may suspect the
item’s condition might be bad, but we may neither know the exact condition of it,
nor its quantification. For direct monitored systems, Markov models are popular,
see Black et al (2005), Chen and Trivedi (2005), and Love (2000). Counting
processes have also been used for modeling the deterioration of directly monitored
plant, see Aven (1996), Jenson (1992). Christer and Wang (1992) used a random
coefficient model for a direct monitored case. It is noted however that the majority
of condition monitoring applications are indirect monitoring such as the five
popular monitoring techniques discussed earlier. It is therefore, in this chapter that
our attention is paid to the indirect monitoring cases.
The second issue is the appropriate definition of the plant state. This also
relates to the first issue whether the monitoring is direct or indirect. In direct
monitoring, the actual observed condition of the item is clearly the plant state.
While in the indirect monitoring case we can only observe measures indirectly
related to the actual condition of the item monitored as discussed earlier. The most
simple and intuitive definition is a set of categorical states ranging, say from 0
(new) to N (failed) as seen from Markov based models, Baruah and Chinnam
(2005). Wang (2006a) also used a generic term of wear to represent the state of the
monitored plant, which is particularly useful in modeling wear related problems in
condition monitoring. Wang and Christer (2000) first used the residual life at the
time of checking as a measure of the state of the monitored unit of interest. This
definition provides an immediate modeling means to directly establish a link
between the measured information and the residual life of interest. It is noted
however, that this residual life is usually not observable which increases modeling
complexity. A model of pi ( xi | ℑi ) introduced later will be based on this
definition.
Various different methods or models have been proposed in literature to
formulate and calculate pi ( xi | ℑi ) . Proportional Hazard Modeling (PHM, one
particular and natural form for modelling the hazard) is a popular one, Kumar and
Westberg (1997), Love and Guo 1991, Makis and Jardine (1991), Jardine et al
(1998), Banjevic et al (2001). Accelerated life models, Kalbfleisch and Prentice
(1980), Wang and Zhang (2005) could also be used here, and may be more
appropriate since the analogy between accelerated life testing, where these models
originate, and condition monitoring is a close one. It should be noted that
accelerated life models and proportional hazard models are identical when the time
to failure distribution is Weibull, that is when the hazard function is given by

h ( t ) = αβ t β − 1 .

There are two problems with proportional hazards modeling or accelerated life
models in condition based maintenance. The first is that the current hazard is
determined partially by the current monitoring measurements and the full
monitoring history is not used. The second is the assumption that the hazard or the
life is a function of the observed monitoring data which acts directly on the hazard
 Chapter Title 9

via a covariate function. Both problems relate to the modeling assumption rather
than the technique. The first one can be overcome if some sort of transformation of
the observed data is used. The second problem remains unless the nature of
monitoring indicates so. It is noted however that for most condition monitoring
techniques, the observed monitoring measurements are concomitant types of
information which are a function of the underlying plant state. A typical example is
in vibration monitoring where a high level of vibration is usually caused by a
hidden defect but not vice versus as we have discussed earlier. In this case the
observed vibration signals may be regarded as concomitant variables which are
caused by the plant state. Note that in oil based monitoring things are different as
the metal particles and other contaminants observed in the oil can be regarded both
as concomitant variables and covariates as we discussed earlier. In this case a
model considers both variables might be appropriate.
The last decade has seen an increased use of stochastic filtering and Hidden
Markov Models (HMM) for modelling pi ( xi | ℑi ) in condition based
maintenance, Hontelez et al (1996), Christer et al (1997), Wang and Christer
(2000), Bunka et al (2000), Dong and He (2004), Lin and Markis (2003, 2004),
Baruah and Chinnam (2005), Wang (2006a). These techniques overcome both
problems of PHM and provide a flexible way to model the relationship between the
observed signals and unobserved plant state. HMM can be seen as a specific type
of stochastic filtering models that are usually used for discrete state and
observation variables. If the noise factors in the model are not Gaussian, then a
closed form for pi ( xi | ℑi ) is generally not available and one has to resort to
numerical approximations. A comparison study using both filtering, Wang (2002),
and PHM, Markic and Jardine (1991), based on vibration data revealed that the
filtering based model produced a better result in terms of prediction accuracy,
Matthew and Wang (2006).

It should be noted also that if the monitored variables also influence the state to
some extent, then both HMM and PHM should be used to tackle the problem.
Alternatively an interactive HMM can also be formulated where a bilateral
relationship is assumed between the observed and unobserved. In the next section,
we shall discuss in details a specific filtering model used for the derivation of
pi ( xi | ℑi ) . This model is simple to use and is analytically tractable.

5.4 Conditional residual life prediction

First we define the true state of plant is the residual life conditional upon measured
condition related information to date, such as, vibration, temperature, etc..
Next we assume these conditional information are functions of the residual life,
that is, it is the residual life which controls the behavior of the measured
conditional information, but not vice-versa (this assumption can be relaxed).
Generally we expect a short residual life (depending on the severity of the defect)
will generate a high signal level in some of the measures of condition variables,
10 Book Title

though in a typical stochastic fashion. In theory, we may have the following

relationship;

Defect short residual life higher than normal signal may be observed.

If the severity of the defect is represented by the length of the residual life, the
relationship between the residual life and observed condition related variables
follows.

5.4.1 Conditional residual life prediction

The model is built based on the following assumptions.

1. Plant items are monitored regularly at discrete time points.

2 There are two periods in the plant life where the first period is the time
length from new to the point when the item was first identified to be
faulty, and the second period is the time interval from this point to failure
if no maintenance intervention is carried out. The second period is often
called the failure delay time. It is also assumed that these two periods are
statistically independent with each other.
3. A threshold level is established to classify the item monitored to be in a
potential faulty state if the condition information signal is above the level.
Such a threshold level is usually determined by engineering experience or
by a statistical analysis of measured condition related variables.
4. The conditional information obtained at time t i , y i , during the failure
delay time is a random variable which depends on xi .
Assumptions 1 and 2 can often be observed in condition monitoring practice.
Assumption 3 can be relaxed and a model which can both identify the starting
point of the second stage and residual life prediction can be established, Wang
(2006b). For now to keep the model simple we still use assumption 3. Assumption
4 was first proposed in Wang and Christer (2000), which states that the rapid
increase in the observed condition information is partly due to the shortened
residual life because of the hidden defect. However this relationship is
contaminated with random noise. Assumption 4 is the fundamental principle
underpinning our model. For a detailed discussion on assumption 4 see Wang and
Christer (2000).
Because the interest in residual life prediction is over the failure delay time
(assuming it exists) and the information collected over the normal working period
may not be beneficial for the residual life prediction, we revise our notation on t i
as the ith and the current monitoring time since the item was suspected to be faulty
but still operating (noted that the order starts from the moment when the item was
first identified to be possibly faulty). This implies that t1 is the first monitoring
point which may indicate that the second stage has started. However, some
monitoring may not be able to display a two-stage process such as oil based
 Chapter Title 11

monitoring. If this is the case, we can simply set the threshold level to be zero. Fig.
5.2 shows a typical condition monitoring practice.

yi
x1

y3
y2 x3
y1 x2
Threshold level

0 t1 t2 t3 failure

Fig. 5.2 Condition monitoring practice

It is noted from Fig. 5.2 that the conditional information obtained before t1 is not
used since they are irrelevant to the decision making process. It is noted however,
that the time to t1 is one of important information sources to be used in
determining the condition monitoring interval, Wang (2003).
Since the residual life at t i is the residual life at t i −1 minus the interval
between t i and t i −1 provided the item has survived to t i and no maintenance
action has been taken, it follows that

⎧ X − (t i − t i −1 ) if X i −1 > t i − t i −1
X i = ⎨ i −1 . (5.2)
⎩ not defined else

The relationship between Yi and X i is yet to be identified. From assumption 4 we

know that it can be described by a distribution, say, p( y i | xi ) . We will discuss
this later when fitting the model to data.
We wish to establish the expression of pi ( xi | ℑi ) , and therefore a
consequential decision model can be constructed on the basis of such a conditional
probability, see (5.1). Since ℑi = { y1 , y2 ,..., yi } = { yi , ℑi −1} , then pi ( xi | ℑi )
can be expressed as pi ( xi | ℑi ) = p ( xi | yi , ℑi −1 ) . It follows that
12 Book Title

p( xi , yi | ℑi −1 )
pi ( xi | ℑi ) = p( xi | yi , ℑi −1 ) = (5.3)
p( yi | ℑi −1 )

By using the multiplicative rule, the joint distribution, p ( xi , yi | ℑi −1 ) is given as

p( xi , yi | ℑi −1 ) = p( yi | xi , ℑi −1 ) p ( xi | ℑi −1 ) (5.4)

Since given both xi and ℑi −1 , y i depends on xi only from assumption 4 so (5.4)

reduces to

p( xi , yi | ℑi −1 ) = p( yi | xi , ℑi −1 ) p ( xi | ℑi −1 ) = p ( yi | xi ) p( xi | ℑi −1 ) (5.5)

Integrating out the xi term in (5.5) we have

∞ ∞
p( yi | ℑi −1 ) = ∫ p( xi , yi | ℑi −1 )dxi = ∫ p( yi | xi ) p( xi | ℑi −1 )dxi (5.6)
0 0

We focus our attention to p ( xi | ℑi −1 ) which appears both in (5.4) and (5.6).

From (5.2) we havexi −1 = g ( xi ) = xi + (ti − ti −1 ) conditional on
X i −1 > ti − ti −1 . Then the distribution of X i | ℑi −1 can be expressed by a
transformation of variables from X i to X i −1 , Freund (2004), as

dg ( xi )
p( xi | ℑi −1 ) = pi −1 ( g ( xi ) | ℑi −1 , X i −1 > ti − ti −1 ) (5.7)
dxi

dg ( xi )
Since = 1 and
dx i

pi −1 ( g ( xi ) | ℑi −1 )
pi −1 ( g ( xi ) | ℑi −1 , X i −1 > ti − ti −1 ) = ∞
(5.8)
∫
t i − t i −1
pi −1 ( xi −1 | ℑi −1 )dxi −1

we finally have
 Chapter Title 13

pi −1 ( xi + ti − ti −1 | ℑi −1 )
p ( xi | ℑi −1 ) = ∞
(5.9)
∫ t i − t i −1
pi −1 ( xi −1 | ℑi −1 )dxi −1

Using (5.5), (5.6) and (5.9), (5.3) becomes

p ( yi | xi ) pi −1 ( xi + ti − ti −1 | ℑi −1 )
pi ( xi | ℑi ) = ∞
(5.10)
∫ 0
p ( yi | xi ) pi −1 ( xi + ti − ti −1 | ℑi )dxi −1

(5.10) is a recursive equation which starts at time t1 . At time t1 , using (5.10) we

have

p( y1 | x1 ) p0 ( x1 + t1 − t0 | ℑ0 )
p1 ( x1 | ℑ1 ) = ∞
(5.11)
∫ 0
p ( y1 | x1 ) p0 ( x1 + t1 − t0 | ℑ0 )dx1

Since ℑ0 is usually 0 or not available, so

p0 ( x1 + t1 − t0 | ℑ0 ) = p0 ( x1 + t1 − t0 ) , then if p0 ( x0 ) and p( y1 | x1 ) can
be specified, (5.11) can be determined. Similarly we can proceed to determining
pi ( xi | ℑi ) if pi −1 ( xi −1 | ℑi −1 ) and p( y i | xi ) are available from the previous
step calculation at time t i −1 .
Now the task is how to specify p 0 ( x0 ) and p ( y i | xi ) .

5.4.2 Specification of p 0 ( x0 ) and p ( y i | xi ) .

p0 ( x0 ) is just the delay time distribution over the second stage of the plant life.
Here we use the Weibull dsitribution as an example in this context. In practice or
theory, the distribution density function p 0 ( x0 ) should be chosen from the one
which best fits to the data or from some known theory.
The set-up of the p ( y i | xi ) term requires more attention. Here we follow the
one used in Wang (2002), where y i | xi is assumed to follow a Weibull
− cx
distribution with the scale parameter being equal to the inverse of A + Be i . In
this way we establish a negative correlation between y i and xi as expected., that
is E (Yi | X i = xi ) ∝ A + Be − cxi . The pdf is given below
14 Book Title

η
y
yi −( i
)η
η −1 A+ Be − cxi
p ( y i | xi ) = ( ) e . (5.12)
A + Be −cxi A + Be −cxi
This is a concept called floating scale parameter, which is particularly useful in this
case, Wang (2002). There are other choices to model the relationship between y i
and xi , but will not be discussed here, and can be found in Wang (2006a).

5.4.3 Estimating the model parameters within pi ( xi | ℑi )

| ℑi ) we need to know the values for the model

To calculate the actual pi ( xi
parameters. They are the parameters of p 0 ( x0 ) and p ( y i | xi ) . The most
popular way to estimate them is using the method of maximum likelihood.
At each monitoring point, t i , two pieces information are available, namely, y i
and X i −1 > t i − t i −1 , both conditional on ℑi −1 . The pdf. for yi | ℑi −1 is given by
(5.7) and the probability function of X i −1 > ti − ti −1 | ℑi −1 is given by

∞
P ( X i −1 > ti − ti −1 | ℑi −1 ) = ∫ pi −1 ( xi −1 | ℑi −1 ) dxi −1 (5.13)
t i − t i −1

If the item monitored failed at time t f after the last monitoring at time t n , the
complete likelihood function is then given by

L(Θ) = ⎛⎜ ∏i =1 p( yi | ℑi −1 ) ∫ pi −1 ( xi −1 | ℑi −1 )dxi −1 ) ⎞⎟ pn (t f − tn | ℑn )
n ∞

⎝ t i − t i −1 ⎠

(5.14)

where Θ is the set of parameters to be estimated. Taking log on both sides of

(5.14) and maximising it in terms of unknown parameters should give the
estimated values of those parameters. However, computationally it has to be solved
numerically since (5.14) involves many integrals which may not have analytical
solutions.

5.4.4 A case study

Fig. 5.3 shows the data of overall vibration level in rms of six bearings, which is
from a fatigue experiment, Wang (2002). It can be seen from Fig. 5.3 that the
bearing lives vary from around 100 hours to over 1000 hours, which shows a
 Chapter Title 15

typical stochastic nature of the life distribution. The monitored vibration signals
also indicate an increasing trend with bearing ages in all cases, but with different
paths. An important observation is the pattern of vibration signals which stays
relatively flat in the early stage of the bearing life and then increases rapidly (a
defect may have been initiated). This indicates the existence of the two stage
failure process as defined earlier.

Fig. 5.3 Vibration data of six bearings

The initial point of the second stage in these bearings is identified using a control
chart called the Shewhart average level chart and the threshold levels of the
bearings are shown in table 5.1, Zhang (2004).

Table 5.1 Threshold level for each bearing

Bearing Threshold level
1 5.06
2 5.62
3 4.15
4 5.14
5 3.92
6 4.9

Assuming both distributions for p 0 ( x0 ) and p ( y i | xi ) are Weibull where

β
p( x0 ) = αβ (αx0 ) β −1 e − (αx0 )

and
16 Book Title

η
y
yi −( i
)η
η −1 A + Be − cxi
p ( y i | xi ) = ( ) e
A + Be −cxi A + Be −cxi

then starting from t1 and after recursive filtering we have

∏
β
( xi + ti ) β −1 e − (α ( xi + t i )) ψ k ( xi , ti )
i

pi ( xi | ℑi ) = ∞
k =1
(5.15)
∏
β

∫ ( z + ti ) β −1 e − (α ( z + t i )) ψ ( z , ti )dz
i

0 k =1 k

where

− C ( z + t i − t k ) −1 η
e − ( yk ( A+ Be ) )
ψ k ( z, ti ) = .
A + Be −C ( z + ti −tk )

To estimate the parameters in p 0 ( x0 ) and p ( y i | xi ) we need write down the

likelihood function as (5.14). The actual process to estimate these unknown
parameters is complicated and involves heavy numerical manipulation which we
omit and interested readers can get the details in Zhang (2004). The estimated
result is listed in table 5.2.

Table 5.2 Estimated parameter values in p 0 ( x0 ) and p ( y i | xi )

α̂ βˆ Â B̂ Ĉ ηˆ
0.011 1.873 7.069 27.089 0.053 4.559

Based on the estimated parameter values in table 5.2 and (5.15) the predicted
residual life at some monitoring points given the history information of bearing 6
in Fig. 5.3 is plotted in Fig. 5.4.
 Chapter Title 17

Fig. 5.4 Predicted condition residual life of bearing 6

In Fig.5.4 the actual residual lives at those checking points are also plotted with
symbol *. It can be seen that actual residual lives are well within the predicted
residual life distribution as expected.
Given the estimated values for parameters and associated costs such as
c f = 6000 , c p = 2000 and cm = 30 , Wang and Jia (2001), we have the
expected cost per unit time for one of the bearings at various checking time t,
shown in Fig. 5.5.
Expectd cost per unit time

27
t=80.5 hrs
t=92.5 hrs
23
t=104 hrs
t=116.5 hrs
19 t=129 hrs

15
0 10 20 30
Planned replacement time

Fig. 5.5. Expected cost per unit time v planned replacement time in hours from the current
time t
18 Book Title

In can be seen from Fig. 5.5. that at t=116.5 and 129 hours both planned
replacements are recommended within the next 30 hours.
To illustrate an alternative decision chart in terms of the actual condition
monitoring reading, we transformed the cost related decision into actual reading in
Fig. 5.6 where the dark grey area indicates that if the reading falls within this area a
preventive replacement is required within the planning period of consideration.
The advantage of Fig. 5.6 is that it can not only tell us whether a preventive
replacement is needed but also show us how far the reading is from the area of
preventive replacement so that appropriate preparation can be done before the
actual replacement.

14
Preventive replacement area
12
Observed CM reading

10

4
No preventive replacement area
2

0
80.5 92.5 104 116.5 129
Tim e (age in hour) of CM reading taken

Fig. 5.6 Decision chart using observed CM reading.

The transformation is carried out in this way. At each monitoring point of t i , by

gradually changing the value of y i in pi ( xi | ℑi ) used in (5.1) until a preventive
replacement is recommended by the model within the planning period, and then
mark this value of y i as the threshold value at time t i . Connecting these threshold
values at those monitoring points forms the boundary between the light and dark
grey areas. Finally mark the actual reading of y i on the graph to see which area it
falls in.

5.5 Future research directions

5.5.1 Multi-component systems

Previous condition based prognosis models developed in the literature mainly

focused on a single failure mode system subject to routine monitoring and
 Chapter Title 19

replacement such as bearings, pumps and motors, and various probability

distributions are used to describe the life time of the component. In the case of a
high value and high risk system with many components such as aircraft engines
and gas turbines, how to assess the health condition and make prognosis based on
condition information obtained from all components is still an open question. It is
typical with a multi-component system that many observed signal parameters are
available and the times between failures are neither independent nor identical.

5.5.2 Idebtification of the initial point of a random defect

With the delay time concept, see chapter 14, system life is assumed to be classified
into two stages. The first is the normal working stage where no abnormal condition
parameters are to be expected. The second starts when a hidden defect is first
initiated with possible abnormal signals. The identification of the initial point in
the evolution of such a defect is important and has a direct impact on the
subsequent prediction model. Most research on fault diagnosis focuses on the
location of the fault, the possible cause of the fault, and of course, the type of fault.
This serves for the engineering purpose of deciding what to repair, but does not aid
the decision of when to do the task. This initial point defect identification has
received very little attention in prognosis literature. Wang (2006b) addressed this
problem to some extent using a combination of the delay time concept and the
HMM. Much work still remains. It is possible that a multi-stage (>2) failure
process could be used, which might be more appropriate to some cases.

5.5.3 The definition of plant state

The definition of the underlying state and the relationship between the observed
monitoring parameters and the state of the system are issues which still need
attention. In the model presented in this chapter, the state of the system is defined
as the residual life, which is assumed to influence the observed signal parameters.
Whilst the modelling output appears to make sense, there are a few potential
problems with the approach. The first is the issue that the life of the plant is fixed
at birth (installation) but unknown. This is termed as playing the God. Secondly,
the residual life is not the direct cause of the observed abnormal signals. These are
more likely caused by some hidden defects which are linked to the residual life in
this chapter. To correct the first problem we can introduce another equation
describing the relationship between X i and X i −1 deterministically or randomly.
This will allow X i to change during use, which is more appropriate. If the
relationship is deterministic, then a closed form of (5.3) is still available, but if it is
random, HMM must be used and no closed form of (5.3) exists unless the noises
associated are normally distributed. The second problem can be overcome if we
adopt a discrete or continuous state hidden Markov chain to describe the system
deterioration process where the state space of the chain represents the system state
under question.
20 Book Title

5.5.4 Information fusion

There is now a considerable amount of condition monitoring and process control

information available in industry, thanks to the recent development in condition
monitoring technology. It is noted that not all information are useful, or because of
correlation they may provide similar information. There are two ways to deal with
this. One is to use some statistical methods to reduce the dimension of the original
data such as principal component analysis, and the other is to use multi-variate
distributions. The principal component analysis method has been used in Wang and
Zhang (2005), but unless the first principle component accounts for most of the
variation in the original data we still need to deal with a data set with more than
two dimensions. The use of multi-variate distributions in prognosis has not been
reported apart from the normal distribution which has the drawback of producing
negative values.
A final point worth mentioning is that in practice observed condition
monitoring variables could be concomitant variables or covariates with respect to
the system state. A model which can handle both type of information is ideal, but
very few attempts have been made, Hussin and Wang (2006).

5.6 Summary and Conclusions

This chapter introduces the concept of condition monitoring, key condition
monitoring techniques, condition based maintenance and associated modelling
support in aid of condition based maintenance. Particular attention is paid to the
residual time prediction based on available condition information to date. An
important development made here is the establishment of the relationship between
the observed information and underlying condition which is the residual life in this
case. This is achieved by letting the mean of the observed information at ti be a
function of the residual life at that point conditional on X i = xi . The
mathematical development is based on a recursive algorithm called filtering where
all past information is included. The example illustrated is based on real data which
came from a fatigue experiment. However, data from industry has showed the
robustness of the approach and the residual life predictions conducted so far are
satisfactory.

5.6 References
Aghjagan, H.N., 1989, Lubeoil analysis expert system, Canadian Maintenance Engineering
Conference, Toronto.
Aven T, 1996, Condition based replacement policies - a counting process approach, Rel.
Eng. & Sys. Safety, 51(3), 275-281.
Banjevic D., Jardine A.K.S., Makis V., Ennis M., 2001, A control-limit policy and software
for condition based maintenance optimization, INFOR 39(1), 32-50.
 Chapter Title 21

Baruah P., and Chinnam R.B., 2005, HMM for diagnostics and prognostics in maching
processes, I. J. Prod. Res., 43(6), 1275-1293.
Black M., Brint, A.T., and Brailsford J.R., 2005, A semi-Markov approach for modelling
asset deterioration, J. Opl. Res. Soc. 56(11), 1241-1249.
Bunks C., McCarthy D., and Al-Ani T., Condition based maintenance of machine using
hidden Markov models, 2000, Mech. Sys. & Sig. Pro., 14(4), 597-612.
Charles W. Reeves, 1998, The vibration monitoring handbook, Coxmoor Publishing
Company, Oxford, 1998.
Chen, W., Meher-Homji, C.B. and Mistree, F., 1994, COMPROMISE: an effective
approach for condition-based maintenance management of gas turbines. Engineering
Optimization, 22, 185-201.
Chen D. and Trivedi K.S., 2005, Optimization for condition based maintenance with semi-
Markov decision process, Rel. Eng. & Sys. Safety, 90(1), 25-29.
Christer A.H and Wang W., 1995, A simple condition monitoring model for a direct
monitoring process, E. J. Opl. Res., 82, 258-269.
Christer A.H. and Wang W., 1992, A model of condition monitoring inspection of
production plant, I. J. Prod. Res., 30, 2199-2211.
Collacott, R.A., 1977, Mechanical fault diagnosis and condition monitoring, Chapman and
Hall Ltd., London.
Dong M., and He D., 2004, Hidden semi-Markov models for machinery health diagnosis
and prognosis, Trans. North Amer. Manu. Res. Ins. of SME, 32, 199-206.
Drake, P.R., Jennings, A.D., Grosvenor, R.I. and Whittleton, D., 1995, acquisition system
for machine tool condition monitoring. Quality and Reliability Engineering
International 11, 15-26.
Freud J.E., 2004, Mathematical statistics with applications, Pearson Prentice and Hall,
London.
Harrison, N., 1995, Oil condition monitoring for the railway business. Insight 37, 278-283.
Hontelez J.A.M., Burger H.H. and Wijnmalen D.J.D., 1996, Optimum condition based
maintenance policies for deteriorating systems with partial information, Rel. Eng. & Sys.
Safety, 51(3), 267-274.
Hussin B, and Wang, W., 2006, Conditional residual time modelling using oil analysis: a
mixed condition information using accumulated metal concentration and lubricant
measurements, to appear in Proc. 1st Main. Eng. Conf, Chendu, China.
Jardine A.K.S., Makis V., Banjevic D., Braticevic D., and Ennis M., 1998, A decision
optimization model for condition based maintenance, J. Qua. Main. Eng., 4(2), 115-121.
Jensen U., 1992, Optimal replacement rules based on different information level, Naval Res.
Log. 39, 937-955.
Kalbfleisch, J.D. & Prentice, R.L., 1980, The Statistical Analysis of Failure Time Data.
Wiley, New York.
Kumar, D., and Westberg U., 1997, Maintenance scheduling under age replacement policy
using proportional hazard modelling and total-time-on-test plotting, Euro. J. Opl. Res.,
99, 507-515.
Li, C.J. & Li, S.Y., 1995, Acoustic emission analysis for bearing condition monitoring.
Wear 185, 67-74.
Lin D., and Makis V., 2003, Recursive filters for a partially observable system subject to
random failures, Adv. Appl. Prob., 35(1), 207-227.
Lin D., and Makis V., 2003, Filters and parameter estimation for a partially observable
system subject to random failures with continuous-range observations, , Adv. Appl.
Prob., 36(4), 1212-1230.
Love, C.E. & Guo, R., 1991, Using proportional hazard modelling in plant maintenance.
Quality and Reliability Engineering International, 7, 7-17.
22 Book Title

Makis V. and Jardine A.K.S., 1991, Computation of optimal policies in replacement models,
IMA J. Maths. Appl. Business & Industry, 3, 169-176.
Matthew C., and Wang W., 2006, A comparison study of proportional hazard and stochastic
filtering when applied to vibration based condition monitoring, submitted to Int. Tran
OR.
Meher-Homji, C.B., Mistree, F. and Karandikar, S., 1994, An approach for the integration of
condition monitoring and multi-objective optimization for gas turbine maintenance
management. International Journal of Turbo and Jet Engines, 11, 43-51.
Neal M., and Associates, 1979, Guide to the condition monitoring of machinery, DTI,
London.
Samanta, B., Al-Balushi, K.R., Al-Araimi, S.A. 2006, Artificial neural networks and
genetic algorithm for bearing fault detection Soft Computing, 10 (3), 264-271.
Wang W., 2002, A model to predict the residual life of rolling element bearings given
monitored condition monitoring information to date, IMA. J. Management Mathematics,
13, 3-16.
Wang W., 2003, Modelling condition monitoring intervals: A hybrid of simulation and
analytical approaches, J. Opl. Res Soc, 54, 273-282.
Wang W., 2006a, A prognosis model for wear prediction based on oil based monitoring, to
appear in J. Opl. Res Soc,
Wang W., 2006b, Modelling the probability assessment of the system state using available
condition information, to appear in IMA. J. Management Mathematics.
Wang W. and A.H. Christer , 2000, Towards a general condition based maintenance model
for a stochastic dynamic system, J. Opl. Res. Soc. 51, 145-155.
Wang W., and Jia, Y., 2001, A multiple condition information sources based maintenance
model and associated prototype software development, proceedings of COMADEM
2001, Eds. A. Starr and Raj B.K.N. Rao, Elsevier, 889-898.
Wang W., and Zhang W., 2005, A model to predict the residual life of aircraft engines based
on oil analysis data, Naval Logistics Research, 52, 276-284.
Wong, M.L.D., Jack, L.B., Nandi, A.K., 2006, Modified self-organising map for
automated novelty detection applied to vibration signal monitoring Mech. Sys.
& Sig. Proc., 20(3), 593-610.
Zhan Y. Makis V., and Jardine A.K.S., 2006, Adaptive state detection of gearboxes under
varying load conditions based on parametric modeling, Mech. Sys. & Sig. Prod. 20(1),
188-221.
Zhang W., 2004, Stochastic modeling and applications in condition based maintenance,
PhD, thesis, University of Salford, UK.
Zhang Z.G. and Love C.E., 2000, A simple recursive Markov chain model to determine the
optimal replacement policies under general repairs, Com. and OR, 27(4), 321-333.