ARTICLE IN PRESS

Mechanical Systems and Signal Processing 23 (2009) 724739

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/jnlabr/ymssp

Review

Rotating machinery prognostics: State of the art, challenges and

opportunities
Aiwina Heng , Sheng Zhang, Andy C.C. Tan, Joseph Mathew
CRC for Integrated Engineering Asset Management, Faculty of Built Environment & Engineering, Queensland University of Technology, Brisbane,
QLD 4001, Australia

a r t i c l e i n f o abstract

Article history: Machinery prognosis is the forecast of the remaining operational life, future con-
Received 20 February 2008 dition, or probability of reliable operation of an equipment based on the acquired
Received in revised form condition monitoring data. This approach to modern maintenance practice promises
18 June 2008
to reduce downtime, spares inventory, maintenance costs, and safety hazards. Given
Accepted 19 June 2008
Available online 1 July 2008
the signicance of prognostics capabilities and the maturity of condition moni-
toring technology, there have been an increasing number of publications on rotating
Keywords: machinery prognostics in the past few years. These publications covered a wide
Condition-based maintenance spectrum of prognostics techniques. This review article rst synthesises and places
Condition monitoring
these individual pieces of information in context, while identifying their merits
Prognostics
and weaknesses. It then discusses the identied challenges, and in doing so,
Reliability
alerts researchers to opportunities for conducting advanced research in the eld.
Current methods for predicting rotating machinery failures are summarised and
classied as conventional reliability models, condition-based prognostics models and
models integrating reliability and prognostics. Areas in need of development or
improvement include the integration of condition monitoring and reliability, utili-
sation of incomplete trending data, consideration of effects from maintenance
actions and variable operating conditions, derivation of the non-linear relationship
between measured data and actual asset health, consideration of failure interactions,
practicability of requirements and assumptions, as well as development of performance
evaluation frameworks.
Crown Copyright & 2008 Published by Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
2. Methods for predicting rotating machinery failures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
2.1. Prognostics approachescondition-based prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
2.1.1. Physics-based prognostics models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
2.1.2. Data-driven prognostics models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
2.2. Integrated approachesprediction based on both reliability and condition data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 730
3. Challenges and opportunities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
3.1. Integration of reliability and prognostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
3.2. Use of censored event data and censored CM data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 732

 Corresponding author. Tel.: +6173 138 5076; fax: +6173 138 1277.
E-mail address: a.heng@qut.edu.au (A. Heng).

0888-3270/$ - see front matter Crown Copyright & 2008 Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.ymssp.2008.06.009
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 725

3.3. Effects of maintenance actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733

3.4. Effects of operating conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
3.5. Non-linear relationship between CM indices and actual asset health. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 734
3.6. Effects of failure interactions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
3.7. Accuracy of assumptions and practicability of requirements. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
3.8. Performance evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
4. Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 736
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 736
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 736

1. Introduction

Operational safety, maintenance cost effectiveness and asset availability have a direct impact on the compe-
titiveness of organisations and nations. Todays complex and advanced machines demand highly sophisticated and
costly maintenance strategies. Domestic plants in the United States spent more than $600 billion to maintain their
critical plant systems in 1981 and this gure doubled within 20 years [1]. An even more alarming fact is that one-third to
one-half of this expenditure is wasted through ineffective maintenance. The trend is similar in many other countries
including Australia [2]. Therefore, there is a pressing need to continuously develop and improve current maintenance
programs.
Current maintenance strategies have progressed from breakdown maintenance, to preventive maintenance, then
to condition-based maintenance (CBM) managed by experts, and lately towards a futuristic view of intelligent
predictive maintenance systems. Breakdown maintenance is the earliest form of maintenance, where no actions are
taken to maintain the equipment until it breaks and consequently needs a repair or replacement. To prevent catastrophic
failures and emergency shutdowns, preventive maintenance was introduced in the 1950s. A typical preventive
maintenance scheme includes setting periodic intervals for machine inspections and maintenance regardless of the
machines health condition. The determination of optimal maintenance interval is critical for this scheme to work
effectively. Bazovsky [3] pioneered the use of mathematical optimization methods in preventive maintenance policies.
Jardine [4] introduced decision models for determining optimal replacement or overhaul interval by analysing reliability
data (e.g. historical breakdown events) and cost data. However, xed time maintenance policies were not well-received by
most practitioners [5]. While these policies do sometimes reduce equipment failures, they are more labour intensive, does
not eliminate catastrophic failures and cause unnecessary maintenance. This is where CBM steps in. It was reported that
99% of mechanical failures are preceded by noticeable indicators [6]. CBM attempts to monitor machinery health based on
condition measurements that do not interrupt normal machine operation. Over the past few decades, technologies in
machine condition monitoring (CM) and fault diagnostics have become more developed. Data such as vibration signatures,
acoustic emissions signatures and oil particle counts can be acquired, processed and analysed through state-of-the-art
sensors, database software and parallel computation technologies. Nevertheless, new technologies often introduce new
types of information that may not have been fully exploited. This development has presented a paramount challenge for
the research community to synthesise and integrate the new information into conventional reliability calculations and
eventually into maintenance scheduling.
Three key elements of effective CBM are data acquisition (i.e. the collection and storage of machine health information),
data processing (i.e. the conditioning and feature extraction/selection of acquired data) and decision making (i.e. the
recommendation of maintenance actions through diagnosis and/or prognosis). Increased automation and mechanisation
have made computerised diagnostics and prognostics systems a valuable tool for maintenance personnel in making
maintenance decisions, or possibly even replace maintenance experts in due time. Todays concept of machine diagnosis
comprises the automated detection and classication of faults, whereas machine prognosis is the automated estimation of
how soon and likely a failure will occur. Prognostics promises to signicantly reduce expensive downtime, spares
inventory, maintenance labour costs and hazardous conditions. However, prognostics is a relatively new research area and
has yet to receive its prominence compared to the other areas of CBM.
Related reviews on prognostics have been reported in the literature. Pusey and Roemer [7] provided a broad overview of
the development in diagnostics and prognostics technologies applicable to high-performance turbo-machines up until year
1999. Jardine et al. [8] provided an overview and a catalogue of publications on data acquisition, data processing,
diagnostics and prognostics of various machines up to year 2005. Vachtsevanos et al. [9] dened and described intelligent
fault diagnostics and prognostics approaches for engineering systems through examples. Ma [10] discussed the need for a
new paradigm shift in CM research for engineering asset management.
This paper focuses on rotating machinery prognostics. Rotating machinery is one of the most common classes of
machines. The number of publications on rotating machinery prognosis has been increasing steadily over the past few
years. Section 2 aims to place these articles in context, showing how they add to the accumulation of knowledge in the area.
Section 3 then discusses the identied challenges and in doing so, alerts researchers to opportunities for advanced research
in the eld of machinery prognostics.
 ARTICLE IN PRESS
726 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

2. Methods for predicting rotating machinery failures

The existing methods for predicting rotating machinery failures can be grouped into the following three main
categories:

1. Traditional reliability approachesevent data based prediction.

2. Prognostics approachescondition data based prediction.
3. Integrated approachesprediction based on both event and condition data.

Table 1 lists the existing models under these three approaches and describes their merits and limitations.
Traditional approaches to reliability estimation are based on the distribution of event records of a population of
identical units. Many parametric failure models, such as Poisson, Exponential, Weibull, and Log-Normal distributions have
been used to model machine reliability. The most popular among them is the Weibull distribution due to its ability to
accommodate various types of behaviour including infant mortality in the bath-tub curve. Reliability analyses have been
extensively studied and developed over the past few decades and numerous books and articles have been published, e.g.
[3,1117]. These classical reliability approaches basically use historical time-to-failure data to estimate the population
characteristics (such as mean-time-to-failure and probability of reliable operation). However, these approaches only
provide general overall estimates for the entire population of identical units. This type of estimations is useful to
manufacturers that produce units in high volumes but are of little value to end users. For example, a maintenance engineer
would be more interested in the ongoing reliability information of a particular piece of component currently running in the
machine, rather than in the mean-time-to-failure of the whole population of such a component.
To estimate the current condition of an operating unit, a more engineering approach to reliability based on the actual
change in unit health is necessary. Recent developments in CM technologies have enabled the collection of non-intrusive
degradation measurements of a unit in operation. These CM data are a rich source of information in reliability evaluation of
individual units. Consequently, the research community has started to develop prognostics models that estimate the future
health of a monitored unit based on acquired CM data. This second category of failure prediction models are reviewed in
Section 2.1. While CM data are corroborative data that reect the current health of an operating item, they do not replace
reliability data that reect population characteristics. The last category comprises models that integrate reliability data into
prognostics and is reviewed in Section 2.2.

2.1. Prognostics approachescondition-based prediction

Most of the existing prognostics models can be divided into two main categories: physics-based models and data-based
models.

2.1.1. Physics-based prognostics models

Physics-based models typically involve building technically comprehensive mathematical models to describe the
physics of the system and failure modes, such as crack propagation and spall growth. These models attempt to combine
system-specic mechanistic knowledge, defect growth formulas and CM data to provide knowledge-rich prognosis
output.
A common physics-based approach is crack growth modelling. Li et al. [19,20] related rolling element bearing defect
growth rate to the instantaneous defect area size and material constants based on Paris formula. Li and Choi [21], and Li
and Lee [22] used Paris law to model spur gear crack growth. A 2D Finite Element Analysis (FEA) model was integrated to
calculate stress and strain elds based on gear tooth load, geometry and material properties. Oppenheimer and Loparo [23]
modelled rotor shaft crack growth using the Forman law of linear elastic fracture mechanics. However, since knowledge of
the instantaneous defect area size is usually unavailable without interrupting the machine operation, these crack growth
models assumed that defect area size could be directly estimated from vibration data.
Orsagh et al. [24,25] used a stochastic version of the Yu-Harris bearing life equation to predict spall initiation and the
Kotzalas-Harris progression model to estimate the time to failure. Kacprzynski et al. [26] enhanced the above system by
proposing a framework for physics-based prognostics integrating material-level models, system-level data fusion
algorithms and parameter tuning techniques. A spiral bevel pinion gear of a helicopter gearbox was used as a case study. It
was stated that uncertainty in factors such as gear geometry, contact, load and material properties limit the reliability of
prognostics systems. Information such as statistical variation of fatigue and fracture properties, as well as compressive
stresses, has a dominant impact on the prediction accuracy of physics-based models.
Sentient Corporation developed a programme called Contact Analysis for Bearing Prognostics (CABPro) [27], which uses
FEA to calculate material stress eld and then determines the cycles to failure based on damage mechanics principles. Qiu
et al. [28] considered the rolling element bearing system as a single-degree-of-freedom vibratory system. The failure
natural frequency and the acceleration amplitude were related to the running time and failure time established from
damage mechanics. Like other physics-based models, these techniques require the estimation of various physics
parameters.
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 727

Table 1
List of rotating machinery health prediction methods and their merits and limitations

Approach Merits Limitations

1. Reliabilityuse event data, e.g. replacement/failure times of historical units

Traditional reliability models (e.g. Weibull,  Population characteristics information  Only provide general, overall estimates for the entire
Poisson, Exponential, Log-Normal enable longer-range forecast population of identical unitsnot necessarily
distributions, see e.g. [3,1117])  Do not require CM accurate for individual operating units

2. Prognosisuse CM data, e.g. vibration measurements of operating units

(i) Physics-based prognostics models  Can be highly accurate if physics of  Real-life system physics is often too stochastic and
models remain consistent across systems complex to model
 Require less data than data-driven  Defect-specic
techniques

Paris law crack growth modelling [19,20]  Least-square scheme enables adaptation  Defect area size is assumed to be linearly correlated
of model parameters to changes in to vibration RMS level
condition  Least-square scheme similar to single-step
adaptation in time series prediction
 Material constants to be determined empirically

Paris law crack growth modelling with  FEA enables material stress calculation  Performance relies on the accuracy of crack size
FEA [21,22] based on bearing geometry, defect size, estimation based on vibration data
load and speed.  Computationally expensive

Forman law crack growth modelling [23]  Relates CM data and crack growth physics  Simplifying assumptions need to be examined
to life models  Model parameters yet to be determined for complex
conditions (e.g. in shaft loading zone and plastic
zones)

Fatigue spall initiation and progression  Calculates the time to spall initiation and  Various physics parameters need to be determined
model [2426] the time from spall initiation to failure
 Cumulative damage since installation is
estimated with consideration of
operating conditions

Contact analysis for bearing prognostics  FEA enables material stress calculation  Various physics parameters need to be determined
[27] based on bearing geometry, defect size,  Computationally expensive
load and speed

Stiffness-based damage rule model [28]  Relates bearing component natural  Least-square scheme similar to single-step
frequency and acceleration amplitude to adaptation in time series prediction
the running time and failure time  Various material constants need to be determined

(ii) Data-driven prognostic models  Do not require assumption or empirical  Generally required a large amount of data to be
estimation of physics parameters accurate

Simple trend projection models [2931]  Ease of calculation  Rely on past degradation pattern and can lead to
inaccurate forecasts in times of change

Time series prediction using ANNs  Fast in handling multivariate analysis  Assume that condition indices deterministically
[33,3842]  Provide non-linear projection represent actual asset health
 Do not require a priori knowledge  Assume that failure occurs once the condition index
exceeds a presumed threshold
 Short prediction horizon

Exponential projection using ANN [43]  Estimates actual failure time instead of  Assumes that all bearing degradation follow an
condition index at future time steps exponential pattern
 Longer prediction horizon  Requires training one ANN for each historical dataset

Data interpolation using ANN [44]  Longer prediction horizon  Requires training one ANN for each historical dataset

Particle ltering [45]  Provides non-linear projection  Poor performance with high dimensional data

Regression analysis and fuzzy logic [46]  Emphasizes the most recent condition  Does not provide indication of time to failure or
information probability of failure
 Fuzzy logic enables condition
classication based on histories
 ARTICLE IN PRESS
728 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

Table 1 (continued )

Approach Merits Limitations

Recursive Bayesian technique [47]  Estimates reliability using CM data of  Accuracy relies strongly on the correct
individual assets, rather than event data determination of thresholds for various trending
features

Hidden Markov Model and Hidden Semi-  Can be trained to recognise different  Lack of relation of the dened health-state change
Markov Model [48,49] bearing fault types and states point to the actual defect progression since it is often
impractical to physically observe a defect in an
operating unit.
 Prognosis projection relies on a failure threshold

Bearing dynamics model using system  Tracks defect severity based on features  Reasonably accurate only when the signal-to-noise
identication [50] that are not affected by operating ratio is high, e.g. damage is severe and running
condition and nearby equipments speed is relatively high

3. Integration of reliability and prognosisuse both event and CM data

Models combining reliability and  Utilise available information more fully  Require both event and condition data to be accurate
prognostics for increased accuracy
 Longer-range prediction

IP and PF interval representation using  Combines reliability and CM data to  Assumes an underlying distribution
Weibull distribution [51] narrow down the time-to-failure window

EXAKT (combining Proportional Hazards  Combines economic considerations with  Accurate transition probability calculation required
Models, transition probability and a cost failure prediction in aiding maintenance a relatively large amount of CM histories
model) [5255,60,61] decision  PHM assumes that hazard changes proportionately
 Identies the order of signicance of with covariates and the proportionality constant is
trending data features the same at all time

Proportional covariates model [62]  Can be used in cases of sparse or no  Assumes that hazard changes proportionately with
historical failure data covariates and the proportionality constant is the
same at all time

Conditional residual time distribution  Considers the whole CM history rather  Requires the determination of a threshold level that
model and proportional residual model than only the current information indicates defect initiation, which is hard to identify
[63,64] and seldom recorded in practice

Symptom reliability model [66]  Denes reliability in symptom domain  Yet to be validated since sufciently large database
 Can include system and operating of symptoms and covariates is scarce in practice
parameters in reliability modelling to
permit more precise prediction

Intelligent product-limit estimator [67]  Uses ANN recognise the nonlinear  Requires sufcient CM histories for ANN training
relationship between actual asset health
and measured condition data
 Non-parametricavoids potentially large
errors due to incorrect assumptions
about the underlying failure distribution

For most industry applications, physics-based models might not be the most practical solution since the fault type in
question is often unique from component to component and is hard to be identied without interrupting operation.
However, physics-based models may be the most suitable approach for cost-justied applications in which accuracy
outweighs most other factors and physics models remain consistent across systems, such as in air vehicles [18]. They also
generally require less data than data-driven models.

2.1.2. Data-driven prognostics models

Data-driven approaches attempt to derive models directly from routinely collected CM data instead of building models
based on comprehensive system physics and human expertise. They are built based on historical records and produce
prediction outputs directly in terms of CM data.
The conventional data-driven methods include simple projection models, such as exponential smoothing [29] and
autoregressive model [30]. One major advantage of these techniques is the simplicity of their calculations, which can be
carried out on a programmable calculator. However, most of these trend forecasting techniques assume that there is some
underlying stability in the monitored system. They also rely on past patterns of degradation to project future degradation.
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 729

This reliance could lead to inaccurate forecasts in times of change. Most of these models follow the changing pattern with a
time lag of at least one observation. Cempel [31] introduced the Tribo-vibroacoustical (TVA) model, which can estimate the
time to failure of a machine as well as forecasting the vibration amplitude or condition. The model was compared with a
constant trend parabolic model, an exponential trend model and an adaptive trending model in predicting a rolling
bearings peak vibration acceleration. It was reported that none of the forecasting techniques was able to predict the
sudden change in the life curve.
Articial neural network (ANN) is currently the most commonly found data-driven technique in the prognostics
literature. An ANN consists of a layer of input nodes, one or more layers of hidden nodes, one layer of output nodes and
connecting weights. The network learns the unknown function by adjusting its weights with repetitive observations of
inputs and outputs. Numerous studies across various disciplines have demonstrated the merits of ANNs, including the
abilities to (a) perform faster than system identication techniques in multivariate prognosis [32]; (b) perform at least as
good as the best traditional statistical methods, without requiring untenable distributional assumptions [33,34]; and (c)
capture complex phenomenon without a priori knowledge [35]. A widely known limitation of ANNs is the lack of
transparency, or rather the lack of documentation on how decisions are reached in a trained network. However, it was
argued in Ref. [35] that increase in model complexity reduces the transparency of both traditional statistical models and
ANN models. It is just that ANNs are more capable in modelling complex phenomenon and consequently need a more
complex structure to represent the phenomenon. Rules can actually be extracted from trained ANNs to explain how
decisions are reached, see e.g. [36,37].
The most simple ANN-based machinery prognostics approach was time series prediction models. Tse and Atherton [33]
and Yam et al. [38] used recurrent neural networks (RNNs) to trend CM indices and forecast successive index value at the
next time step. Wang and Vachtsevanos [39] developed a recurrent wavelet neural network (RWNN) to predict rolling
element bearing crack propagation. The network performed satisfactorily in trending an articially seeded and manually
enlarged crack, provided that sufcient data points were used and network retraining was carried out after each time step.
Wang et al. [40] used a Neuro-Fuzzy (NF) network to predict spur gear condition value one step ahead. The fuzzy
interference structure is determined by experts, whereas the fuzzy membership functions are trained by the neural
network. The NF system performed much better than RNN when there was sufcient training data. However, it could not
predict well when the train-set was small or when there were fast dynamic uctuations, such as during the chipping of
gear tooth surface material or just prior to gear failure. An adaptive training technique was later proposed by Wang [41] to
improve the NF model. The addition of feedback links to the NF model was able to increase the forecasting accuracy.
However, further work is needed to extend the prediction horizon from single step to multiple steps ahead. Feed forward
neural network (FFNN) has also been used to perform single-step-ahead prediction of rolling element bearing condition by
Shao and Nezu [42]. Multiple-step-ahead predictions were also performed simply by feeding the predicted value back into
the network input until the desired prediction horizon was reached. The authors also proposed some rules to vary the data
sampling period according to the change ratio of consecutive condition index values.
So far all the ANN models mentioned above only produced estimates of the future asset condition indices. Gebraeel et al.
[43] attempted to predict the actual bearing failure time instead of the future condition index. Their thrust bearing
prognosis was based on the assumption that all bearing degradation signals possess an inherent exponential growth, and
FFNNs were used to project the degradation by computing exponential parameters that give the best exponential t.
Remaining useful life was then found by solving for the projected failure time at the predetermined failure threshold.
Ninety-two percent of the failure time predictions were within 20% of the actual bearing life. Huang et al. [44] built on the
above method and used an FFNN to predict the failure time of single-row deep-groove ball bearings based on 100
interpolated points. It was stated that 85% of the failure time predictions were within 20% of the actual bearing lifetime. It
was pointed out that the features effective for trending the degradation of thrust bearings in Ref. [43] either have low
sensitivity to incipient faults or are not effective for trending the degradation of single row, deep groove ball bearings under
highly accelerated tests.
Particle ltering has also been employed to provide non-linear projection in forecasting the growth of a crack on a
turbine engine blade [45]. The current fault dimension was estimated based on the knowledge of the previous state of the
process model. The a priori state estimate was then updated using new CM data. To extend this state estimation to multi-
step-ahead prediction, a recursive integration process based on both Importance Sampling and Kernel probability density
function approximation was applied to generate state predictions to the desired prediction horizon.
Jantunen [46] used high-order regression functions to mimic bearing fault development and also to save trending data
in a compact form. Fuzzy classication limits were set to dene eight classes of predicted conditions or health states of a
rolling element bearing. Simplied fuzzy logic was then used to classify the bearing health state. Indication of time to
failure or probability of failure was not provided.
Zhang et al. [47] proposed a recursive Bayesian technique to calculate failure probability based on the joint density
function of different CM data features. This method enabled reliability analysis and prediction based on the degradation
process of historical units, rather than on failure event data. The prediction accuracy of this model relied strongly on the
correct determination of thresholds for the various trending features.
The use of Hidden Markov Models (HMMs) in bearing fault prognosis was investigated by Zhang et al. [48]. In an HMM,
a system is modelled to be a stochastic process in which the subsequent states have no causal connection with previous
states. In the work of Zhang et al., one HMM was trained to recognise one type of cone-and-cup bearing faults based on the
 ARTICLE IN PRESS
730 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

corresponding vibration data. The HMMs were also trained to estimate the fault states, such as normal, nick,
scratches, more nicks and failure. The similarity between current state and failure state was used as the bearing
degradation index, which was then extrapolated to estimate the time of exceeding a predetermined failure threshold.
HMMs do not represent temporal structure adequately since their state durations follow an exponential distribution. Dong
and He [49] have thus proposed a segmental Hidden Semi-Markov Model (HSMM). Unlike HMMs, which generate a single
observation for each state, HSMMs generate a segment of observations and estimate the durations from training data.
However, the other possible limitations of HMMs remain. One example limitation is the difculty in relating the dened
health-state change point to the actual defect progression since it is often impractical to physically observe a defect in an
operating unit.
Li and Shin [50] argued that it is often impractical to estimate bearing fault severity directly from vibration data because
such data can be affected by many factors other than defect severity, such as machine operating conditions and vibrations
from nearby machine components. They proposed a system identication method to identify a bearing dynamics model
using measured bearing vibration. This dynamics model was then inverted to recover the impact impulse response of the
measured vibration signal. The estimated relative severity index was reasonably accurate only when the signal-to-noise
ratio was high with a relatively high running speed.
Data-driven models may often be the more available solution in many practical cases in which it is easier to gather data
than to build accurate system physics models.

2.2. Integrated approachesprediction based on both reliability and condition data

The previous section presented conventional reliability models as well as condition-based prognostics models made
possible by the new CM technologies. While CM data are a rich source of information for fault/failure prediction, it should
be noted that they do not render reliability data unnecessary. Several valuable models have considered integrating
reliability data into prognostics.
Goode et al. [51] proposed a statistical method to predict the remaining useful life of pumps in a hot strip steel mill.
Alarm limits were rst determined using the Statistical Process Control (SPC) theory, with the assumption that healthy
state data follow a normal distribution. The Installation to Potential failure (IP) and the Potential failure to Functional
failure (PF) intervals were then represented using Weibull distribution. Time to failure was calculated in the PF interval
taking into account vibration data. This work presented a relatively simple approach to forecasting failure with utilisation
of both reliability and CM data.
Jardine et al. [5255] applied the Proportional Hazards Model (PHM) to forecasting the reliability of rolling element
bearings and engines. PHMs assume that hazard changes proportionately with covariates (asset condition in this case) and
that the proportionality constant is the same at all time. The use of PHM for incorporating CM information in reliability
calculation started in the 1980s [5659]. Jardine et al. used Weibull distribution to model the baseline hazard function in
PHM. The covariates were assumed to follow a non-homogeneous Markov stochastic process [60,61]. A software called
EXAKT [61] was developed to calculate the optimal maintenance or replacement time intervals based on historical event
and condition data, as well as cost data.
Sun et al. [62] pointed out that the change in CM covariates is a response to the change in system hazard. Therefore,
system hazard should be modelled as the explanatory variable in the proportional hazards theory, where as CM
information should be modelled as the response variables. A Proportional Covariates Model (PCM) was thus proposed.
PCM can be used to estimate hazard functions of mechanical components or systems in cases of sparse or no historical
failure data, provided that the covariates are proportional to the hazard. A suggested research direction was to use PCM for
reliability prediction based on both failure/covariates data and online CM data.
Wang [63] and [64] stated that models like PHM and PCM only use the current asset condition information (or functions
of current information), rather than the whole monitoring history, to predict future asset health. Therefore, he modelled a
conditional residual time distribution to estimate the residual life distribution of rolling element bearings based on the
stochastic ltering theory [65]. All the test bearings were initially assumed to follow the Weibull delay time distribution
modelled, and as more CM information became available, the distribution was updated. However, this model required the
determination of a threshold level to indicate defect initiation point, which is hard to identify and seldom recorded in
practice.
Cempel et al. [66] dened reliability in symptom (S) (condition data) domain, assuming the symptom of a set of
continuously operating units was measured periodically and had a uniform limit value and breakdown value. Hazard
function was dened as the relative number of units reaching breakdown value per unit increment of symptom. A logistic
vector containing covariates such as system foundation stiffness and running load was also introduced into the reliability
model to permit more precise determination of the system condition and symptom limit value. The model, though not yet
been applied to rotating machinery, provided a unique concept in estimating reliability based on condition data. The model
can only be validated when a sufciently large database of real-life symptoms and covariates are available. Methods for
assessing the covariate function needed for logistic vector implementation also remain an open question.
Heng et al. [67] introduced an intelligent reliability model called the Intelligent Product Limit Estimator, which was able
to include suspended CM data in machinery fault prognosis. The accurate modelling of suspended data was found to be of
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 731

Effects of failure
interactions

Complexities of
real-life systems

Integration of Determination of
reliability and failure threshold
prognostics
Accuracy of
assumptions
and practicability
of requirements Effects of operating
conditions
Accurate and full
utilization of data
Use of censored Variable Operating
events and CM Patterns
data

Effects of maintainance
actions

Fig. 1. Challenges in rotating machinery prognostics.

great importance, since in practice machines are rarely allowed to run to failure and data are commonly suspended. The
model consists of an FFNN whose training targets are asset survival probabilities estimated using a variation of the
KaplanMeier estimator and the true survival status of historical units. This work presented a concept of utilising available
information more fully and accurately, and of providing longer-range prediction in a probabilistic sense with minimal
assumptions.
All of the models reviewed in this subsection combined reliability and condition information in failure prediction. They
generally produce longer-range failure forecasts than models that only use individual asset condition information.
Fig. 1.
Other papers related to machinery prognostics are listed in Refs. [32,6877].

3. Challenges and opportunities

This section discusses how the previous studies, though they have aided the advancement of the discipline, have made
only a limited contribution to developing an effective prognostics model. Several aspects need to be further investigated
before prognostics systems can be reliably applied in real-life situations. Firstly, methods for utilising available data more
fully and accurately need to be explored. Secondly, real-life machines are often subject to variable operating patterns,
which include repairs and change of operating parameters. The effects of these operational complexities, if not properly
considered, may greatly reduce the accuracy of prognosis output. Thirdly, the inherent structural complexity of real-life
machines also hinders practical applications of many prognostics models, which are only designed to predict a specic
failure mode of a component without considering the interaction of the component with other components or with the
operating environment. Several challenges stem from one or more of the three issues mentioned above. To suggest the
many opportunities for prognostics research, we list eight areas in need of new ideas and improvement: the incorporation
of CM data into reliability analyses; the utilisation of incomplete trending data; the consideration of effects from
maintenance actions and variable operating conditions; the deduction of the non-linear relationship between measured
condition and actual degradation; the considerations of failure interactions; the accuracy of assumptions and practicability
of requirements, as well as the development of performance measurement frameworks. These challenges are depicted in
Fig. 2 and discussed in the following sections.

3.1. Integration of reliability and prognostics

As mentioned in Section 2 of this review, reliability models based solely on event data have been reasonably well
developed for machinery life estimation. However, these approaches only provide general or average estimates for the
entire population of identical units to facilitate time-based maintenance. Maintenance or repair at pre-established intervals
tends to incur even higher scheduled downtime. Besides, failure behaviour of each unit is a function of changes in work
schedule, operating environment and other duty parameters, and of failure interaction between components. Therefore,
current condition of an operating unit needs to be monitored online. Nevertheless, CM data, though reect the state of
 ARTICLE IN PRESS
732 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

Preventive replacement
(often mistakenly taken Actual failure point
Start of operation as failure point) if left undisturbed
tr tf
t0

tf -tr

It is the ability of estimating the units

future survival beyond the time of
fault detection or preventive
replacement that is critical to optimal
maintenance scheduling

Fig. 2. Timeline of the operational life of an asset.

individual operating units, do not replace reliability data that reect population characteristics. CM data mainly provide
information for short-term condition prediction only. Several data-driven prognostics models [33,3842] enabled machine
prognosis using time series prediction. These models mainly performed single-step-ahead predictions to estimate the
vibration signal feature value at the next immediate time step. These techniques require further research because prognosis
with such short a prediction horizon is not of much help for optimal maintenance scheduling. Sufciently long lead time is
often needed for effective and economical preparation of spares and human resources. A longer prediction horizon is also
necessary for deciding whether a unit will last until the next maintenance opportunity (such as the end of a batch
production or the next scheduled inspection).
Therefore, other than the unit-specic, online condition information measured, the prediction of remaining useful life
should also be dependent on the operational age or time that the unit has survived, as well as on the general characteristics
of the population to which the unit belongs. For example, given the same condition information, a bearing that is in its
early operating period and is from a family with a long nominal life should have a longer remaining useful life than a
bearing that is in the wear-out period from a family with a much shorter nominal life. Although several models have
attempted to use CM information for reliability estimation as discussed in Section 2, the integration of CM information with
reliability analysis has not been well explored. Some valuable literature [7882] outside the research area of rotating
machine prognostics has proposed some very interesting approaches. For example, Knezevic [78] estimated the probability
density of the condition parameter values of all historical units at a certain operational age. The probability of reliable
operation at that instant of operational age was represented by the instantaneous probability that the condition parameter
values fell below the failure threshold. Most of these approaches [7881] are theoretical formulations and the
opportunities here lie in adapting and validating these techniques in the area of rotating machinery prognostics.

3.2. Use of censored event data and censored CM data

Many prognostics models, especially data-driven models, require abundant historical event data such as the time of
failures. In practice, however, industrial and military communities would rarely allow their assets to run to failure. Most of
the time, once a defect is detected in a unit, the unit is replaced or overhauled before it fails. Therefore the cut-off point at
which the unit will cease to function is not always known or recorded. It is only known that the unit has survived up to the
time of replacement or repair but there is no information as to when it would have failed if left undisturbed. Data of this
sort are called censored or suspended data. Publications related to the modelling of suspended CM data in existing
prognostics models are very limited.
When all the historical suspension times are treated as historical failure times, the prognostics model will produce
biased estimates (underestimation) of the time to failure. Since in many instances a unit is replaced/overhauled once a fault
is detected, treating the replacement/overhaul times as failure times defeats the purpose of prognosis because it is the
duration the failing unit can survive beyond this point that is of interest (Fig. 2). It is not uncommon that a machine
components remaining useful life (from the point where a defect is detected) is substantially more than its nominal life.
A prognostics specialists goal is to recommend a maintenance schedule that does not interrupt production or wastefully
replace units that still have useful remaining life. It is the ability to estimate this remaining lifetime that is critical to
optimal maintenance scheduling.
On the other hand, prediction models that carefully omit suspended data from the training examples will worsen the
problem of data unavailability. Degradation data are already scarce due to irregular measurement recording and/or the
huge amount of time they take to accumulate. For example, a bearing may last several years even under harsh operating
conditions. Therefore, a good prognostics model must be able to maximise the use of available data. The survival
probability of a historical unit after suspension may be estimated based on the population characteristics. Recently Heng
et al. [67] demonstrated an approach of including both suspended CM data and suspended event data in training an
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 733

intelligent prognostics model. This concept aims to stimulate thoughts on utilising available prognostics information more
fully and accurately and opens up new opportunities for the development of more robust variations of the proposed
approach.

3.3. Effects of maintenance actions

One of the main objectives of implementing CBM is to optimise maintenance schedules and actions. This optimisation
can be achieved through the estimation of asset life and the effectiveness of maintenance actions. Therefore, it is necessary
to extend prognostics research effort to monitor the change in unit condition after a maintenance action and estimating the
consequent change in unit reliability. Maintenance actions, such as repairs and re-lubrication, do not always restore the
health of a unit to as good as new conditions. Besides, the unit may sometimes deteriorate at a different rate after such
maintenance actions. These effects need to be considered when developing a prognostics model.
In the eld of Reliability Centered Maintenance (RCM), the reliability of repairable systems has been extensively
modelled. Monga and Zuo [83] introduced a time variable parameter to dene the different start points of a system hazard
function after different repairs. They also proposed a failure rate deterioration factor and modelled system reliability to be
inversely proportional to the number of repairs [84]. Several Markovian models and Poissonian process models have been
proposed to model the reliability of repairable systems. Markov models [8590] were developed to model a repairable unit
as a system with a nite state space. They described events such as repairs, failures and standby as different states and
assumed that the transition between these states to be a stochastic process with Markov property. Poissonian process
models [9196] generally assume that failures of repairable systems can be considered as a series of random discrete
events that follow a Poissons distribution. Sun et al. [97,98] also developed a split system approach to model imperfect
repairs. The basic concept of the split system approach is to separate repaired and unrepaired components within a system
when modelling the system reliability after maintenance activities, since not all the components of a system are repaired
during each maintenance action. A comprehensive review of existing reliability models and maintenance policies can be
found in [99].
Considerable research effort needs to be contributed to developing prognostics models that can account for the effects
of maintenance actions. One way of responding to this opportunity is by enhancing the above-mentioned reliability models
that account for maintenance effects with CM information such as the past examples of post-maintenance changes in
condition indices.

3.4. Effects of operating conditions

In most real-life situations, machines are subject to varying operating conditions. This form of variation is a major
contribution to the change in the energy of measured CM signals. In machining processes for example, non-uniformity of
raw material causes the load on the tooling system to vary [100]. This load variation is transmitted to the output shaft of
the gearing system and is manifested as energy variation in vibration sensor signals. Therefore, an effective prognostics
system should only be sensitive to the changes in condition measurements caused by the deterioration of a monitored unit,
and insensitive to the inuence of any non-deterioration source of variation. Prognostics models found in the literature
have not directly considered the effects of varying operating conditions. For example, Qiu et al. [28,101] related remaining
bearing life to bearing stiffness, which decreases as the spall grows. However rotor stiffness is a function of many factors,
such as speed, load, lubrication and temperature, which are rarely constant in practice. Thus the model might only function
well under controlled experimental conditions and requires further research for practical applications. A more general
example is when an increase in operating speed causes the machine vibration amplitudes to rise. A prognostics model
might perceive this rise in vibration as increased degradation of the monitored unit. Conversely, if the operating speed is
decreased when a monitored unit is deteriorating, a prognostics model might recognise this effect as a decrease in machine
hazard even though a failure is imminent. It is not uncommon that the speed of a machine changes continuously during
normal operation, whether it is an electric machine being fed by a variable speed drive or an induction machine operating
under variable load conditions. This change in speed, if not taken into account, might result in errors in life predictions.
In a comparison of vibration amplitudes of healthy and damaged bearings, Stack et al. [102] observed that vibration
amplitude uctuations due to speed variations under normal operating conditions are less pronounced and usually go
unnoticed when a rolling element bearing is healthy. But as bearing health degrades, these variations in vibration
amplitudes become quite signicant. The authors suggested obtaining measurements at the same machine speed and
comparing them to baseline values when performing fault detection. In practice, multiple variables often need to be
accounted for. Marble and Morton [27] ran multiple bearings to failure under various speeds and loads. It was observed
that the effect of speed on spall propagation was much smaller than that of load. It was also shown that the accrued run
time prior to damage (spall) initiation had a signicant effect on the spall propagation rate.
The effects of variations in operating conditions have not been properly explored and modelled for fault prognosis. This
issue limits the application of prognostics models to real-life machines and thus needs to be addressed. One possible
solution may be to use condition data features that are less dependent on operating conditions. In the literature on fault
detection, load-independent techniques have been proposed. McFadden [103] performed a Hilbert transform on band-pass
 ARTICLE IN PRESS
734 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

ltered, time synchronously averaged signals for gear fault detection. This technique cancelled out vibration amplitude
through phase modulation and therefore was independent of load condition. Parker Jr. et al. [104] studied the use of
bispectrum-based statistical change detection algorithms for gear and bearing fault monitoring. It was found that the
bicoherence function reduced the variation in the bispectrum across different torque levels and loads. Stander et al. [105]
proposed order tracking and time synchronously averaging gearbox vibration data to compensate for the variation in
rotational speed induced by uctuating loads. A pseudo-Wigner-Ville distribution was then applied to the data to identify
the inuence of the uctuating load. These load-independent fault detection techniques need to be explored and extended
for application in prognostics. For example, the load demodulation normalisation technique proposed by Stander and
Heyns [105] was validated using limited data recorded at several inspections. To determine the suitability of this technique
for fault prognostics, the performance or stability of this method over a full cycle of fault progression needs to be
investigated. It is also helpful to express remaining useful life in number of cycles/revolutions, instead of calendar time
such as minutes or days. The variation in speed or operation schedule changes the time taken by a rotating machine to
complete the same number of revolutions. Other useful references include papers on accelerated life tests [106108], which
involve collecting data under several conditions and then extrapolating the results for other possible conditions. Simple
stress-life models such as Arrhenius or the Power Law may be combined with a life distribution model [106]. One difculty
of in accelerated life tests is the large number of unknown parameters in the equations. A simple 2-stress, 4-level test
may lead to a series of simultaneous non-linear equations with 10 or more unknowns. Another consideration is the
associated costs.

3.5. Non-linear relationship between CM indices and actual asset health

CM data are commonly taken to indicate the health of a monitored unit. However, the measured condition indices do
not always deterministically represent the actual health of the monitored unit. The challenges and opportunities here lie in
developing prognostics models that recognise the non-linear relationship between a units actual survival condition and
the measured CM indices.
Most of the existing prognostics techniques use CM indices to represent the health of the monitored unit and then use
methods such as regression or time series prediction to estimate the units future health, or rather the future CM indices. In
these techniques, a threshold for the CM data is predened to represent a failure (as depicted in Fig. 3). The prediction
accuracy then relies strongly on the assumption that failure takes place at the instant of time when the relevant CM index
exceeds the predetermined threshold. However, it is not uncommon that a system fails when its condition measurement is
still below a predened failure threshold or even temporarily decreasing. Conversely a system may still be performing its
required function when its condition measurements already fall outside the tolerance range. Missed alarms and false
alarms are signicant issues in practical applications of prognostics systems. Several methods have been proposed for
determining thresholds for fault detection based on mathematical models instead of solely on the asset suppliers
suggestions or maintenance personnels past experiences. Decoste [109] proposed a dynamic threshold approach to set
limits that can adapt to different normal baseline conditions. The variable threshold was modelled by parametric functions
which can be learnt and adapted from historical condition trending data.
In the eld of prognostics, more thought should be devoted to developing prognostics models that can deduce the non-
linear relationship between a units actual survival condition and the measured CM indices. Articial intelligence models
can be trained to learn from past examples. Hence, there are research opportunities to use the past measured condition
data as model training input and the actual unit health as target output. By repetitively presenting various pairs of training
input and target to the intelligent models, the models may learn to recognise how unit degradation is veiled in the non-
deterministic changes in CM measurements and disregard uctuations caused by non-deterioration factors.

Failure
threshold
CM parameter index

Operating time
Failure time

Fig. 3. Assumption that failure takes place at the instant of time when the condition index exceeds a predened threshold.
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 735

3.6. Effects of failure interactions

Given that a machine often consists of multiple components or subsystems, the ability to monitor and predict the
degradation of one component might not be sufcient to predict overall machinery failures. For example, the degradation
of a component may initiate or accelerate the failure of another component and vice versa. In critical cases, this pheno-
menon can result in catastrophic consequences. Literature on prognostics models for multi-component systems appear to
be absent. In the eld of reliability however, there have been several attempts to model interactive failures [97,110113]. For
example, Greig [110] proposed a method for modelling multiple sequential failures in a system. According to this model,
the failure of a component changes the system topology and consequently increases the failure probabilities of the other
components in the system. Sun et al. [113] proposed an analytical model where the interactive failure rate of a component
was estimated by its independent failure rate plus a portion of the failure rates of its inuencing components. It was stated
that the failure rate of a component accelerates when the failure rates of its inuencing components increase. If this type of
failure interaction effects is not considered, the probability of failure will be underestimated.
As CBM systems mature and provide data that actually reect the health state of individual operating units, it is
important to combine this new information with the existing wealth of knowledge on interactive failures.

3.7. Accuracy of assumptions and practicability of requirements

Since prognosis involves projecting into the future and that the future cannot be determined with absolute certainty,
assumptions and simplications are often inevitable in prognostics modelling. Nevertheless, care must be taken to
minimise these assumptions and simplications.
Physics-based approaches can be very accurate when a correct and accurate model is available. However, the
characteristics and relationships of all related components in a system and its environment are often too complicated to be
modelled [32]. It has been reported that the wear of rotating machinery components is still not fully understood at present
[114,115]. One of the main issues with empirical or physics-based models is thus their inherent uncertainty due to the large
amount of accompanying assumptions. It was shown in numerical simulation tests and bearing life tests that physics-based
models without parameter ne-tuning is unable to capture and adapt to the stochastic characteristics of defect growth
[19,20]. It was also shown that a small amount of parameter difference results in large prediction errors with the increase of
bearing running cycles. Currently most physics-based prognostics methods focus on the prediction of crack propagation.
However, in many cases other failure modes tend to dominate and the maintenance engineer needs to correctly identify the
fault type in question. Even if that has been accomplished, defect growth is not a deterministic process. Virkler et al. [116]
showed that even under well-controlled experimental conditions, crack growths of a set of identical components were
vastly different. Crack growth models are also difcult to apply in practice because they require knowledge of the exact
geometry or/and orientation of the crack, which are usually very irregular and cannot be identied without disassembling
the machine component. Some models have attempted to account for these uncertainties by adopting a stochastic or
adaptive component such as a recursive least-squares scheme. Such adaptation techniques are similar to the adaptation in
data-driven time series prediction.
Gebraeel et al. [43] based their rolling element bearing prognosis on the assumption that all bearing degradation signals
possess an inherent exponential growth. It is unclear that if these complex ANN models offer analysis more accurate than
that provided by exponential curve tting and extrapolation methods. Exponential extrapolations often have a large region
of uncertainty. Due to the probabilistic nature of bearing integrity and operating condition, defect propagation rates tend to
be stochastic. If the growth of a bearing defect differs from the assumed exponential pattern, the bearing degradation will
be poorly extrapolated. Even for proper assumptions about the exponential defect growth, the prediction condence
interval of extrapolations often diverges to unrealistic values. Extrapolating beyond that range can lead to misleading
results. Besides, successive prediction outputs may vary vastly and seem confusing. Without probability indication, it is
rather difcult for maintenance personnel to make maintenance decisions based on these prediction outputs. It is however
noted that, although mathematical prognostics models have yet to prove entirely satisfactory in practice, those used for
processing and managing reliability information are critical for reliability prediction work.
Considerations should also be made in order to nd the optimum trade-off between accuracy and costs. For example,
FEA methods might be able to calculate the stress on the material surrounding a fault based on component geometry and
operating conditions, but are very expensive in terms of modelling and computational time.
It is important to keep assumptions and simplications realistic because real-life machines are complex and fault
propagation is probabilistic in nature. Affordability in terms of computation time, accuracy, memory and data storage
should also be considered at all stages including model design stage.

3.8. Performance evaluation

The literature on prognostics is continually growing. The standardisation of prognostics performance measures,
however, has not matched this growth. Vachtsevanos [117] has dened a possible class of performance metrics that stems
primarily from the main objectives of prognosis and the uncertainty of prognostics algorithms. The metrics include
 ARTICLE IN PRESS
736 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

prediction-to-failure times, prediction accuracy, precision and condence, as well as sensitivity to input changes. More
effort is required in the development of measurement frameworks for supporting prognostics performance evaluation.
Opportunities here include developing frameworks that consider various real-life requirements and include both
quantitative and qualitative metrics. Quantitative metrics may include the average difference between the predicted failure
time and the actual failure time, as well as the average computational time. Example qualitative metrics include the
practicability of model requirements and the ability to provide condence level indication. Studies on the issues of
performance measurement errors and uncertainties, as well as the pros and cons of different measurement strategies, also
make contributive investigations.

4. Concluding remarks

The ability to forecast machinery failure is vital to reducing maintenance costs, operation downtime and safety hazards.
This paper synthesised the progress in the research and development of rotating machinery prognostics. The review
indicates that most of the existing prognostics studies were conducted in research laboratories. It is often easy to neglect
certain practical considerations when developing models in laboratory environments, where access to insights into real-life
situations is often lacking. Nevertheless, the ultimate goal remainsto establish reliable prognostics systems that can be
applied in real-life situations and benet industry. To realise the greatest economic and social benets, it is important to
design every aspect of a prognostics system by considering the asset managers perspective.
The challenges remaining in the eld of rotating machinery prognostics were discussed in this paper. The occurrences of
machinery failures are difcult to predict due to the inherent structural and operational complexities of real life systems.
Failure interaction between components, the probabilistic nature of fault symptoms and growth, as well as varying
operating conditions and duty parameters also add to the complexity of data. To suggest the many opportunities for
prognostics research, we discussed eight areas in need of solution and improvement: (1) the integration of CM and
reliability data, (2) the proper utilisation of incomplete trending data, (3) the consideration of the effects from maintenance
actions, (4) the consideration of varying operating conditions, (5) the deduction of the non-linear relationship between
condition data and asset health, (6) the consideration of failure interactions, (7) the practicability of requirements and
assumptions, and (8) the development of a uniform performance measurement framework.
It should also be noted that one of the most overlooked measures in the development of maintenance applications has
been the degree of user-friendliness. These applications can be complex and overwhelming to the users, who often do not
have in-depth knowledge about the underlying system physics or model algorithms. The review presented in this paper
also suggests that, as is true in so many aspects of life, a single tool is usually incapable of solving all problems. One also has
to content with a wide variety of applications, with different deterioration mechanisms, operating conditions, data
dimensions, computation costs, as well as memory and storage capacities. Collaboration between research groups from
different areas of expertise should be encouraged. The lack of effective capture and management of eld data also places a
huge constraint on the research and development of intelligent maintenance systems. Common datasets consisting of
relevant CM data, maintenance event records and cost data may be made available to all researchers in the eld. In this
way, the technical and economic feasibility of each prognostics model can be evaluated and compared consistently.
Researchers should agree on adopting a uniform set of benchmark criteria for convenient model comparison. This practice
will signicantly accelerate the advancement of prognostics research.

Acknowledgements

The authors gratefully acknowledge the nancial support provided by the QUT Faculty of Built Environment &
Engineering and the Cooperative Research Centre for Integrated Engineering Asset Management (CIEAM).

References

[1] Ferret, The basics o predictive/preventive maintenance, http://www.ferret.com.au/articles/ad/0c0259ad.asp, accessed March 1, 2006.
[2] Westwick-Farrow, Predictive/preventive maintenance, http://www.inmotiononline.com.au/inmotion/latest_news/article.asp?id=7611, accessed
March 2, 2006.
[3] I. Bazovsky, Reliability Theory and Practice, Prentice-Hall, Englewood Cliffs, NJ, 1961.
[4] A.K.S. Jardine, Maintenance, replacement and reliability, Wiley, New York, 1973.
[5] A. Kelly, Maintenance and its management, Proceedings of the Communication Conference, Surrey, England, 1989.
[6] H.P. Bloch, F.K. Geitner, Machinery failure analysis and trouble shooting, Gulf Professional Publishing, 1997.
[7] H.C. Pusey, M.J. Roemer, An assessment of turbomachinery condition monitoring and failure prognosis technology, The Shock and Vibration Digest
31 (1999) 365371.
[8] A.K.S. Jardine, D. Lin, D. Banjevic, A review on machinery diagnostics and prognostics implementing condition-based maintenance, Mechanical
Systems and Signal Processing 20 (2006) 14831510.
[9] G. Vachtsevanos, F. Lewis, M. Roemer, A. Hess, B. Wu, Intelligent fault diagnosis and prognosis for engineering systems, Wile, Hoboken, NJ, 2006.
[10] L. Ma, Condition monitoring in engineering asset management, Proceedings of Asia-Pacic Vibration Conference, Sapporo, Hokkaido, Japan, 2007.
[11] J.F. Lawless, Statistical Models and Methods for Lifetime Data, Wiley-Interscience, New York, 2002.
[12] K.C. Kapur, L.R. Lamberson, Reliability in Engineering Design, Wiley, New York, 1977.
[13] E.A. Elsayed, Reliability engineering, Prentice-Hall, Englewood Cliffs, NJ, 1996.
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 737

[14] A.Z. Keller, U.D. Perera, A.R.R. Kamath, Reliability analysis of CNC machine tools, Reliability Engineering 3 (1982) 449473.
[15] M.J. Crowder, A. Kimber, T. Sweeting, R. Smith, Statistical Analysis of Reliability Data, Chapman & Hall/CRC, London/, 1994.
[16] A. Schomig, O. Rose, On the suitability of the Weibull distribution for the approximation of machine failures, Proceedings of the Industrial
Engineering Research Conference, Portland, OR, 2003.
[17] P.G. Groer, Analysis of time-to-failure with a Weibull model, Proceedings of the Maintenance and Reliability Conference, Knoxville, TN, USA, 2000,
pp. 59.0159.04.
[18] M.J. Roemer, G.J. Kacprzynski, E.O. Nwadiogbu, G. Bloor, Development of diagnostic and prognostic technologies for aerospace health management
applications, Proceedings of the IEEE Aerospace Conference, Big Sky, Montana, USA, 2001, pp. 31393147.
[19] Y. Li, S. Billington, C. Zhang, T. Kurfess, S. Danyluk, S. Liang, Adaptive prognostics for rolling element bearing condition, Mechanical Systems and
Signal Processing 13 (1999) 103113.
[20] Y. Li, T.R. Kurfess, S.Y. Liang, Stochastic prognostics for rolling element bearings, Mechanical Systems and Signal Processing 14 (2000) 747762.
[21] C.J. Li, S. Choi, Spur gear root fatigue crack prognosis via crack diagnosis and fracture mechanics, Proceedings of Proceedings of the 56th Meeting of
the Society of Mechanical Failures Prevention Technology, Virginia Beach, VA, 2002, pp. 311320.
[22] C.J. Li, H. Lee, Gear fatigue crack prognosis using embedded model, gear dynamic model and fracture mechanics, Mechanical Systems and Signal
Processing 19 (2005) 836846.
[23] C.H. Oppenheimer, K.A. Loparo, Physically based diagnosis and prognosis of cracked rotor shafts, Proceedings of SPIE Component and Systems
Diagnostics, Prognostics, and Health Management II, Bellingham, 2002, pp. 122132.
[24] R.F. Orsagh, J. Sheldon, C.J. Klenke, Prognostics/diagnostics for gas turbine engine bearings, Proceedings of the IEEE Aerospace Conference 7 (2003)
30953103.
[25] R. Orsagh, M. Roemer, J. Sheldon, A comprehensive prognostics approach for predicting gas turbine engine bearing life, Proceedings of the 10th
National Turbine Engine High Cycle Fatigue Conference, New Orleans, LO, 2004, pp.
[26] G.J. Kacprzynski, A. Sarlashkar, M.J. Roemer, A. Hess, B. Hardman, Predicting remaining life by fusing the physics of failure modeling with
diagnostics, JOM Journal of the Minerals, Metals and Materials Society 56 (2004) 2935.
[27] S. Marble, B. P. Morton, Predicting the remaining life of propulsion system bearings, Proceedings of the IEEE Aerospace Conference, 2006, pp. 18.
[28] J. Qiu, B.B. Set, S.Y. Liang, C. Zhang, Damage mechanics approach for bearing lifetime prognostics, Mechanical Systems and Signal Processing 16
(2002) 817829.
[29] W. Batko, Prediction method in technical diagnostics, Doctoral Thesis, Cracov Mining Academy, 1984.
[30] K. Kazmierczak, Application of autoregressive prognostic techniques in diagnostics, Proceedings of the Vehicle Diagnostics Conference, Tuczno,
Poland, 1983.
[31] C. Cempel, Simple condition forecasting techniques in vibroacoustical diagnostics, Mechanical Systems and Signal Processing 1 (1987) 7582.
[32] J. Lee, A systematic approach for developing and deploying advanced prognostics technologies and tools: methodology and applications,
Proceedings of the Second World Congress on Engineering Asset Management, Harrogate, UK, 2007, pp. 11951206.
[33] P. Tse, D. Atherton, Prediction of machine deterioration using vibration based fault trends and recurrent neural networks, Transactions of the ASME:
Journal of Vibration and Acoustics 121 (1999) 355362.
[34] R. Joshi, C. Reeves, Beyond the Cox model: articial neural networks for survival analysis part II, Proceedings of the Eighteenth International
Conference on Systems Engineering (2006) pp.
[35] D.G. Bostwichk, H.B. Burke, Prediction of individual patient outcome in cancer, Cancer Supplement 91 (2001) 16431646.
[36] R. Setiono, W.K. Leow, and J.Y.L. Thong, Opening the neural network black box: An algorithm for extracting rules from function approximating
articial neural networks, Proceedings of International Conference on Information Systems, Brisbane, Australia, 2000, pp. 176186.
[37] A.B. Tickle, R. Andrews, M. Golea, J. Diederich, The truth will come to light: directions and challenges in extracting the knowledge embedded within
trained articial neural networks, IEE Transactions on Neural Networks 9 (1998) 10571068.
[38] R.C.M. Yam, P.W. Tse, L. Li, P. Tu, Intelligent predictive decision support system for CBM, The International Journal of Advanced Manufacturing
Technology 17 (2001) 383391.
[39] P. Wang, G. Vachtsevanos, Fault prognostics using dynamic wavelet neural networks, Articial Intelligence for Engineering Design, Analysis and
Manufacturing 15 (2001) 349365.
[40] W.Q. Wang, M.F. Golnaraghi, F. Ismail, Prognosis of machine health condition using neuro-fuzzy systems, Mechanical Systems and Signal Processing
18 (2004) 813831.
[41] W. Wang, An adaptive predictor for dynamic system forecasting, Mechanical Systems and Signal Processing 21 (2007) 809823.
[42] Y. Shao, K. Nezu, Prognosis of remaining bearing life using neural networks, Proceedings of the Institution of Mechanical Engineers, Part I: Journal of
Systems and Control Engineering, 2000, pp. 217230.
[43] N. Gebraeel, M. Lawley, R. Liu, V. Parmeshwaran, Residual life predictions from vibration-based degradation signals: a neural network approach,
IEEE Transactions on Industrial Electronics 51 (2004) 694700.
[44] R. Huang, L. Xi, X. Li, C. Richard Liu, H. Qiu, J. Lee, Residual life predictions for ball bearings based on self-organizing map and back propagation
neural network methods, Mechanical Systems and Signal Processing 21 (2007) 193207.
[45] M. Orchard, B. Wu, and G. Vachtsevanos, A particle lter framework for failure prognosis, Proceedings of Proceedings of the World Tribology
Congress, Washington, DC, 2005.
[46] E. Jantunen, Prognosis of rolling bearing failure based on regression analysis and fuzzy logic, Proceedings of the VETOMAC-3 and ACSIM-2004, New
Delhi, India, 2004, pp. 836846.
[47] S. Zhang, L. Ma, Y. Sun, J. Mathew, Asset health reliability estimation based on condition data, Proceedings of the 2nd WCEAM and the 4th ICCM,
Harrogate, UK, 2007, pp. 21952204.
[48] X. Zhang, R. Xu, C. Kwan, S. Y. Liang, Q. Xie, L. Haynes, An integrated approach to bearing fault diagnostics and prognostics, Proceedings of American
Control Conference, Portland, OR, USA, 2005, pp. 27502755.
[49] M. Dong, D. He, A segmental hidden semi-Markov model (HSMM)-based diagnostics and prognostics framework and methodology, Mechanical
Systems and Signal Processing 21 (2007) 22482266.
[50] C.J. Li, H. Shin, Tracking bearing spall severity through inverse modeling, Proceedings of the ASME International Mechanical Engineering Congress,
Anaheim, CA, USA, 2004, pp. 16.
[51] K.B. Goode, J. Moore, B.J. Roylance, Plant machinery working life prediction method utilizing reliability and condition-monitoring data, Proceedings
of Institution of Mechanical Engineers 214 (2000) 109122.
[52] A.K.S. Jardine, P.M. Anderson, D.S. Mann, Application of the Weibull proportional hazards model to aircraft and marine engine failure data, Quality
and Reliability Engineering International 3 (1987) 7782.
[53] A.K.S. Jardine, P. Ralston, N. Reid, J. Stafford, Proportional hazards analysis of diesel engine failure data, Quality and Reliability Engineering
International 5 (1989) 207216.
[54] N. Montgomery, T. Lindquist, M.-A. Garnero, R. Chevalier, A.K.S. Jardine, Reliability functions and optimal decisions using condition data for EDF
primary pumps, Proceedings of the International Conference on Probabilistic Methods Applied to Power Systems, Stockholm, Sweden, 2006.
[55] P.O. Sundin, N. Montgomery, A.K.S. Jardine, Pulp mill on-site implementation of CBM decision support software, Proceedings of International
Conference of Maintenance Societies, Melbourne, Australia, 2007.
[56] A. Bendell, Proportional hazards modelling in reliability assessment, Reliability Engineering 11 (1985) 175183.
[57] A. Bendell, D.W. Wightman, E.V. Walker, Applying proportional hazards modelling in reliability, Reliability Engineering and System Safety 34 (1991)
3553.
 ARTICLE IN PRESS
738 A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739

[58] H. Ascher, H. Fetngold, Repairable Systems Reliability, Marcel Dekker, New York, 1984.
[59] K. Dhananjay, B. Klefsjo, Proportional hazards model: a review, Reliability Engineering and System Safety 44 (1994) 177188.
[60] D. Banjevic, A.K.S. Jardine, Calculation of reliability function and remaining useful life for a Markov failure time process, IMA Journal of Management
Mathematics 17 (2006) 115130.
[61] A.K.S. Jardine, D. Banjevic, V. Makis, Optimal replacement policy and the structure of software for condition-based maintenance, Journal of Quality
in Maintenance Engineering 3 (1997) 109.
[62] Y. Sun, L. Ma, J. Mathew, W. Wang, S. Zhang, Mechanical systems hazard estimation using condition monitoring, Mechanical Systems and Signal
Processing 20 (2006) 11891201.
[63] W. Wang, A model to predict the residual life of rolling element bearings given monitored condition information to date, IMA Journal of
Management Mathematics 13 (2002) 316.
[64] W. Wang, W. Zhang, A model to predict the residual life of aircraft engines based upon oil analysis data, Naval Research Logistics 52 (2005) 276284.
[65] W. Wang, A.H. Christer, Towards a general condition based maintenance model for a stochastic dynamic system, Journal of Operational Research
Society 51 (2000) 145155.
[66] C. Cempel, H.G. Natke, J.T.P. Yao, Symptom reliability and hazard for systems condition monitoring, Mechanical Systems and Signal Processing 14
(2000) 495505.
[67] A. Heng, A. Tan, J. Mathew, B.-S. Yang, Machine prognosis with full utilization of truncated lifetime data, Proceedings of the Second World Congress
on Engineering Asset Management, Harrogate, UK, 2007, pp. 775784.
[68] G. Wen, X. Zhang, Prediction method of machinery condition based on recurrent neural networks models, Journal of Applied Sciences 4 (2004)
675679.
[69] T.E. Tallian, Data tted bearing life prediction model for variable operating conditions, Tribology Transactions 42 (1999) 241.
[70] D.C. Swanson, A general prognostic tracking algorithm for predictive maintenance, Proceedings of the IEEE Aerospace Conference, 2001.
[71] L. Zhang, Z. Wang, S. Zhao, Short-term fault prediction of mechanical rotating parts on the basis of fuzzy-grey optimising method, Mechanical
Systems and Signal Processing 21 (2007) 856865.
[72] K. Goebel, P. Bonanni, N. Eklund, Towards an integrated reasoner for bearings prognostics, Proceedings of the IEEE Aerospace Conference, Niskayuna,
NY, USA, 2005, pp. 38643657.
[73] Y.-L. Dong, G. Yu-Jiong, K. Yang, Research on the condition based maintenance decision of equipment in power plant, Proceedings of the Machine
Learning and Cybernetics Conference, 2004.
[74] L. Hitchcock, ISO standards for condition monitoring, Proceedings of the Inaugural World Congress on Engineering Asset Management, Gold Coast,
Australia, 2006.
[75] B. Castanier, C. Berenguer, A. Grall, A sequential condition-based repair/replacement policy with non-periodic inspections for a system subject to
continuous wear, Applied Stochastic Models in Business and Industry 19 (2003) 327347.
[76] L. Dieulle, C. Berenguer, A. Grall, M. Roussignol, Sequential condition-based maintenance scheduling for a deteriorating system, European Journal of
Operational Research 150 (2003) 451461.
[77] A. Grall, C. Berenguer, L. Dieulle, A condition-based maintenance policy for stochastically deteriorating systems, Reliability Engineering and System
Safety 76 (2002) 167180.
[78] J. Knezevic, Condition parameter based approach to calculation of reliability characteristics, Reliability Engineering 19 (1987) 2939.
[79] J.-C. Lu, J. Park, Q. Yang, Statistical inference of a time-to-failure distribution derived from linear degradation data, Technometrics 39 (1997) 391.
[80] H. Saranga, J. Knezevic, Reliability analysis using multiple relevant condition parameters, Journal of Quality in Maintenance Engineering 6 (2000)
165176.
[81] S. Lu, H. Lu, W.J. Kolarik, Multivariate performance reliability prediction in real-time, Reliability Engineering & System Safety 72 (2000) 3945.
[82] J. Yan, M. Koc-, J. Lee, A prognostic algorithm for machine performance assessment and its application, Production Planning & Control 15 (2004)
796801.
[83] A. Monga, M.J. Zuo, Optimal design of series-parallel systems considering maintenance and salvage value, Computer & Industrial Engineering 40
(2001) 323337.
[84] A. Monga, M.J. Zuo, R. Toogood, Reliability based design considering preventive maintenance and minimal repair, International Journal for Quality,
Reliability and Safety Engineering 4 (2002) 5571.
[85] H. Pham, Reliability analysis of k-out-of-N systems with partially repairable multi-state components, Microelectronics R 36 (1996) 14071415.
[86] H. Pham, A. Suprasad, R.B. Misra, Availability and mean life time prediction of multistage degraded system with partial repairs, Reliability
Engineering & System Safety 56 (1997) 1691736.
[87] S. Bloch-Mercier, Optimal restarting distribution after repair for a Markov deteriorating system, Reliability Engineering & System Safety 74 (2001)
181191.
[88] C.H. Wang, S.H. Sheu, Determining the optimal production maintenance policy with inspection errors: using a Markov chain, Computer &
Operational Research Society 30 (2003) 117.
[89] G. Becker, L. Camarinopoulos, G. Zioutas, A semi-Markovian model allowing for inhomoginities with respect to process time, Reliability Engineering
& System Safety 70 (2000) 4148.
[90] L.A. Papazoglou, Semi-Markovian reliability model for systems with testable components and general test/outage times, Reliability Engineering &
System Safety 68 (2000) 121133.
[91] R. Gue, C.E. Love, Statistical analysis of an age model for imperfectly repaired systems, Quality and Reliability Engineering International 8 (1992)
133146.
[92] J.W.T. Roberts, J.L. Mann, Failure predictions in repairable multi-component systems, International Journal of Production Economics 29 (1993)
103110.
[93] M. Guida, M. Giorgio, Reliability analysis of accelerated life-test data from a repairable system, IEEE Transactions on Reliability 44 (1995) 337342.
[94] J.L. Coetzee, The role of NHPP models in the practical analysis of maintenance failure data, Reliability Engineering & System Safety 56 (1997)
161168.
[95] G.R. Weckman, R.L. Shell, J.H. Marvel, Modelling the reliability of repairable systems in the aviation industry, Computer & Industrial Engineering 40
(2001) 5163.
[96] G. Pulcini, Modelling the failure data of repairable equipment with bathtub type failure intensity, Reliability Engineering & System Safety 71 (2001)
209218.
[97] Y. Sun, L. Ma, J. Mathew, S. Zhang, Determination of preventive maintenance lead time using hybrid analysis, International Journal of Plant
Engineering and Management 10 (2005) 1318.
[98] Y. Sun, L. Ma, J. Mathew, Prediction of system reliability for single component repair, Journal of Quality in Maintenance Engineering 13 (2007)
111124.
[99] H. Wang, A survey of maintenance policies of deteriorating systems, European Journal of Operational Research 139 (2002) 469489.
[100] Y. Zhan, V. Makis, A.K.S. Jardine, Adaptive model for vibration monitoring of rotating machinery subject to random deterioration, Journal of Quality
in Maintenance Engineering 9 (2003) 351375.
[101] J. Qiu, C. Zhang, B.B. Seth, S.Y. Liang, Failure lifetime prognostic of bearing based on vibration response and damage mechanics, Proceedings of the
JapanUSA Flexible Automation Conference, July 2326, 2000.
[102] J.R. Stack, T.G. Habetler, R.G. Harley, Effects of machine speed on the development and detection of rolling element bearing faults, IEEE Power
Electronics Letters 1 (2003) 1921.
 ARTICLE IN PRESS
A. Heng et al. / Mechanical Systems and Signal Processing 23 (2009) 724739 739

[103] P.D. McFadden, Detecting fatigue cracks in gears by amplitude and phase demodulation of the meshing vibration, Transactions of the ASME Journal
of Vibration, Acousitcs, Stress and Reliability in Design 108 (1986) 165170.
[104] B.E. Parker Jr., H.A. Ware, D.P. Wipe, W.R. Tompkins, B.R. Clark, E.C. Larson, H.V. Poor, Fault diagnostics using statistical change detection in the
bispectral domain, Mechanical Systems and Signal Processing 14 (2000) 561570.
[105] C.J. Stander, P.S. Heyns, Using vibration monitoring for local fault detection on gears operating under uctuating load conditions, Mechanical
Systems and Signal Processing 16 (2002) 10051024.
[106] J.A. McLinn, New analysis methods of multilevel accelerated life tests, Proceedings of the Annual Reliability and Maintainability Symposium (1999)
3842.
[107] P. Johannessona, T. Svenssona, J.d. Marea, Fatigue life prediction based on variable amplitude testsmethodology, International Journal of Fatigue
27 (2005) 954965.
[108] W.Q. Meeker, L.A. Escobar, C.J. Lu, Accelerated degradation tests: modeling and analysis, Technometrics 40 (1998) 8999.
[109] D. Decoste, Learning envelopes for fault detection and state summarization, Proceedings of the IEEE Aerospace Conference, Big Sky, MT, USA, 2000,
pp. 337344.
[110] G.L. Greig, Second moment reliability analysis of redundant systems with dependent failures, Reliability Engineering and System Safety 41 (1991)
5770.
[111] A. Hoyland, M. Rausand, System reliability theory: models and statistical methods, Wiley, New York, 1994.
[112] W. Long, Y. Sato, M. Horigome, Quantication of sequential failure logic for fault tree analysis, Reliability Engineering & System Safety 67 (2000)
269274.
[113] Y. Sun, L. Ma, J. Mathew, S. Zhang, An analytical model for interactive failures, Reliability Engineering & System Safety 91 (2006) 495504.
[114] E. Jantunen, Intelligent monitoring and prognosis of degradation of rotating machinery, Proceedings of the Intelligent Maintenance Systems
IMS2004, Advances in Maintenance and Modelling, Simulation and Intelligent Monitoring of Degradation, Arles, 2004.
[115] A. Muller, M.-C. Suhner, B. Iung, Formalisation of a new prognosis model for supporting proactive maintenance implementation on industrial
system, Reliability Engineering and System Safety 93 (2008) 234253.
[116] D.A. Virkler, B.M. HiUbcry, P.K. Goel, The statistical nature of fatigue crack propagation, ASME Journal of Engineering Materials and Technology 101
(1979) 148153.
[117] G. Vachtsevanos, Performance metrics for fault prognosis of complex systems, The IEEE Conference on Systems Readiness Technology (2003)
341345.