Why is the Remaining Useful Life Prediction Uncertain?

Shankar Sankararaman1 and Kai Goebel2

1
SGT Inc., NASA Ames Research Center, Moffett Field, CA 94035, USA
[email protected]

2
NASA Ames Research Center, Moffett Field, CA 94035, USA
[email protected]

A BSTRACT Since prognostics deals with predicting the future behavior of

engineering systems, there are several sources of uncertainty
This paper discusses the significance and interpretation of
which influence such future prediction, and therefore, it is
uncertainty in the remaining useful life (RUL) prediction of
rarely feasible to obtain an estimate of the RUL with com-
components used in several types of engineering applications,
plete precision. In fact, it is not even meaningful to make
and answers certain fundamental questions such as Why is
such predictions without computing the uncertainty associ-
the RUL prediction uncertain?, How to interpret the uncer-
ated with RUL. As a result, researchers have been developing
tainty in the RUL prediction?, and How to compute the un-
different types of approaches for quantifying the uncertainty
certainty in the RUL prediction?. Prognostics and RUL pre-
associated with the RUL prediction and prognostics in gen-
diction are affected by various sources of uncertainty. In or-
eral.
der to make meaningful prognostics-based decision-making,
it is important to analyze how these sources of uncertainty Existing methods for quantifying uncertainty in prognostics
affect the remaining useful life prediction, and thereby, com- and remaining useful life prediction can be broadly classified
pute the overall uncertainty in the remaining useful life pre- as being applicable to two different types of situations: of-
diction. The classical (frequentist) and Bayesian (subjective) fline prognostics and online prognostics. Methods for offline
interpretations of uncertainty and their implications on prog- prognostics are based on rigorous testing before and/or after
nostics are explained, and it is argued that the Bayesian inter- operating an engineering system, whereas methods for online
pretation of uncertainty is more suitable for remaining useful prognostics are based on monitoring the performance of the
life prediction in the context of condition-based monitoring. engineering system during operation. For example, there are
Finally, it is demonstrated that the calculation of uncertainty several research papers which discuss uncertainty quantifica-
in remaining useful life can be posed as an uncertainty propa- tion in crack growth analysis (Sankararaman, Ling, Shantz,
gation problem, and the practical challenges involved in com- & Mahadevan, 2011; Sankararaman, Ling, & Mahadevan,
puting the uncertainty in the remaining useful life prediction 2011), structural damage prognosis (Farrar & Lieven, 2007;
are discussed. Coppe, Haftka, Kim, & Yuan, 2010), electronics (Gu, Barker,
& Pecht, 2007), and mechanical bearings (Liao, Zhao, &
1. I NTRODUCTION Guo, 2006), primarily in the context of offline testing. En-
gel et. al (Engel, Gilmartin, Bongort, & Hess, 2000) discuss
The prediction of remaining useful life (RUL) is an important
several issues involved in the estimation of remaining useful
functional aspect of an efficient prognostics and health man-
life in online prognostics and health monitoring. While some
agement (PHM) system. The RUL prediction is not only nec-
of the initial studies on remaining useful life prediction lacked
essary to verify if the mission goal(s) can be accomplished
uncertainty measures (Celaya, Saxena, Kulkarni, Saha, &
but also important to aid in online decision-making activi-
Goebel, 2012), researchers have recently started investigating
ties such as fault mitigation, mission replanning, etc. Since
the impact of uncertainty on estimating the remaining useful
the prediction of RUL is critical to operations and decision-
life. For example, there have been several efforts to quantify
making, it is imperative that the RUL be estimated accurately.
the uncertainty in remaining useful life of batteries (Saha &
Goebel, 2008) and pneumatic valves (Daigle & Goebel, 2010)
Shankar Sankararaman et al. This is an open-access article distributed un-
der the terms of the Creative Commons Attribution 3.0 United States Li- in the context of online health monitoring. Different types
cense, which permits unrestricted use, distribution, and reproduction in any of sampling techniques (Daigle, Saxena, & Goebel, 2012)
medium, provided the original author and source are credited.

1
 Annual Conference of the Prognostics and Health Management Society 2013

and analytical methods (Sankararaman, Daigle, Saxena, & nent/system. There are several sources of uncertainty which
Goebel, 2013) have been proposed to predict the uncertainty affect the prediction of future behavior, and in turn, the re-
in the remaining useful life. maining useful life prediction. As a result of these sources
of uncertainty, it is practically impossible to precisely esti-
While the importance of uncertainty quantification in prog-
mate the remaining useful life prediction. In order to make
nostics and RUL estimation have been widely understood,
meaningful prognostics-based decision-making, it is impor-
there have been few efforts to understand and appropriately
tant to analyze how the various sources of uncertainty affect
interpret such uncertainty. Celaya et al. (Celaya, Saxena, &
the remaining useful life prediction and compute the overall
Goebel, 2012) discussed the interpretation of RUL in the con-
uncertainty in the remaining useful life prediction.
text of Kalman filtering-based prognostics techniques, and
explained that it is not approriate to arbitrarily force the vari-
2.1. Activities Related to Uncertainty in PHM
ance of RUL to be small. It is necessary to further delve into
this topic in order to completely analyze the importance and In the context of prognostics and health management, un-
impact of uncertainty in prognostics. certainties have been discussed from representation, quan-
tification, and management points of view (deNeufville, R.,
This paper poses three fundamental questions in order to un-
2004; Hastings, D. and McManus, H., 2004; Ng & Abram-
derstand uncertainty in prognostics, particular in the context
son, 1990; Orchard, Kacprzynski, Goebel, Saha, & Vachtse-
of remaining useful life (RUL) prediction:
vanos, 2008; Tang, Kacprzynski, Goebel, & Vachtsevanos,
1. Why is the RUL prediction uncertain? 2009). While these three are different processes, they are of-
ten confused with each other and interchangeably used. In
2. How do we interpret the uncertainty in RUL?
this paper, the various tasks related to uncertainty quantifi-
3. How do we calculate the uncertainty in RUL? cation and management are classified into four, as explained
The answers to the above questions are sought from multiple below. These four tasks need to performed in order to ac-
points of view. First, the topic of uncertainty in prognostics curately estimate the uncertainty in the RUL prediction and
is discussed from a qualitative point of view in Section 2; the inform the decision-maker regarding such uncertainty.
various sources of uncertainty in prognostics are discussed
and the different activities related to uncertainty quantifica- 1. Uncertainty Representation and Interpretation: The
tion and management are outlined. Second, the interpretation first step is uncertainty representation and interpretation,
of uncertainty is discussed from a statistical point of view in which in many practical applications, is guided by the
Section 3. While statistics and probability methods have been choice of modeling and simulation frameworks. There
in existence for over 200 years, there has always been a dis- are several methods for uncertainty representation that
agreement (amongst mathematicians and statisticians alike) vary in the level of granularity and detail. Some common
regarding the interpretation of probability. It is important to theories include classical set theory, probability theory,
understand this disagreement before attempting to interpret fuzzy set theory, fuzzy measure (plausibility and belief)
uncertainty in prognostics. Third, the interpretation of un- theory, rough set (upper and lower approximations) the-
certainty in prognostics and RUL prediction is analyzed in ory, etc. Amongst these theories, probability theory has
detail in Section 4, and it is explained all different interpre- been widely used in the PHM domain (Celaya, Saxena,
tations of probability may not be suitable for prognostics and & Goebel, 2012); even within the context of probabilistic
health monitoring purposes. Fourth, it is demonstrated that methods, uncertainty can be interpreted and perceived in
calculating the uncertainty in RUL is, fundamentally, an un- two different ways: frequentist (classical) versus subjec-
certainty propagation problem and the challenges in this re- tivist (Bayesian). Sections 3 and 4 outline the differences
gard are outlined in Section 5. In this context, it is examined between these two schools of thought and argues that the
whether it is possible to analytically construct the probability Bayesian approach provides a more suitable interpreta-
distribution of remaining useful life prediction in certain sim- tion for uncertainty in PHM.
ple example problems (consisting of linear models and Gaus-
2. Uncertainty Quantification: The second step is uncer-
sian variables) and it is demonstrated that it is impossible to
tainty quantification, that deals with identifying and char-
estimate closed-form analytical solutions without rigourous
acterizing the various sources of uncertainty that may af-
mathematical considerations even for such simple example
fect prognostics and RUL estimation. It is important that
problems.
these sources of uncertainty are incorporated into models
and simulations as accurately as possible. The common
2. U NCERTAINTY IN P ROGNOSTICS
sources of uncertainty in a typical PHM application in-
Prognostics is the art of predicting future component/system clude modeling errors, model parameters, sensor noise
behavior, identifying possible failure models, and thereby and measurement errors, state estimates (at the time at
computing the remaining useful life of the compo- which prediction needs to be performed), future loading,

2
 Annual Conference of the Prognostics and Health Management Society 2013

operating and environmental conditions, etc. The goal Most of the research in the PHM community pertains to the
in this step is to address each of these uncertainties sep- topics of uncertainty quantification and propagation; few ar-
arately and quantify them using probabilistic/statistical ticles have directly addressed the topic of uncertainty man-
methods. The Kalman filter is essentially a Bayesian tool agement. Even within the realm of uncertainty quantification
for uncertainty quantification, where the uncertainty in and propagation, the estimates of uncertainty have sometimes
the states is estimated continuously as a function of time, been misinterpreted. For example, when statistical principles
based on data which is also typically available continu- are used to estimate a parameter, there is an emphasis on cal-
ously as a function of time. culating the estimate with the minimum variance. When this
principle is applied to RUL estimation, it is important not
3. Uncertainty Propagation: The third step is uncertainty to arbitrarily reduce the variance of RUL itself. Celaya et
propagation and is most relevant to prognostics, since al. (Celaya, Saxena, & Goebel, 2012) explored this idea and
it accounts for all the previously quantified uncertain- explained that the variance of RUL needs to be carefully cal-
ties and uses this information to predict (1) future states culated by accounting for the different sources of uncertainty.
and the associated uncertainty; and (2) remaining useful
life and the associated uncertainty. The former is com- 2.2. Sources of Uncertainty
puted by propagating the various sources of uncertainty In many practical applications, it may even be challenging
through the prediction model. The latter is computed us- to identify and quantify the different sources of uncertainty
ing the estimated uncertainty in the future states along that affect prognostics. Some researchers have classified the
with a Boolean threshold function which is used to in- different sources of uncertainty into different categories in
dicate end-of-life. In this step, it is important to under- order facilitate uncertainty quantification and management.
stand that the future states and remaining useful life pre- While it has been customary to classify the different sources
dictions are simply dependent upon the various uncer- of uncertainty into aleatory (physical variability) and epis-
tainties characterized in the previous step, and therefore, temic (lack of knowledge), such a classification may not be
the distribution type and distribution parameters of future suitable for condition-based monitoring purposes; this point
states and remaining useful life should not be arbitrar- will be explained in detail in the next section. A completely
ily chosen. Sometimes, a normal (Gaussian) distribution different approach for classification, particularly applicable
has been assigned to the remaining useful life prediction; to condition-based monitoring, is outlined below:
such an assignment is erroneous and the true probability
distribution of RUL needs to be estimated though rig-
1. Present uncertainty: Prior to prognosis, it is important
orous uncertainty propagation of the various sources of
to be able to precisely estimate the condition/state of the
uncertainty through the state space model and the EOL
component/system at the time at which RUL needs to be
threshold function, both of which may be non-linear in
computed. This is related to state estimation commonly
practice.
achieved using filtering. Output data (usually collected
through sensors) is used to estimate the state and many
4. Uncertainty Management: The fourth and final step is filtering approaches are able to provide an estimate of the
uncertainty management, and it is unfortunate that, in uncertainty in the state. Practically, it is possible to im-
several articles, the term Uncertainty Management has prove the estimate of the states and thereby reduce the
been used instead of uncertainty quantification and/or uncertainty, by using better sensors and improved filter-
propagation. Uncertainty management is a general term ing approaches.
used to refer to different activities which aid in managing
uncertainty in condition-based maintenance during real- 2. Future uncertainty: The most important source of un-
time operation. There are several aspects of uncertainty certainty in the context of prognostics is due to the fact
management. One aspect of uncertainty management at- that the future is unknown, i.e. both the loading and
tempts to answer the question: Is it possible to improve operating conditions are not known precisely, and it is
the uncertainty estimates? The answer to this question important to assess the uncertainty in loading and envi-
lies in identifying which sources of uncertainty are sig- ronmental conditions before performing prognostics. If
nificant contributors to the uncertainty in the RUL pre- these quantities were known precisely (without any un-
diction. For example, if the quality of the sensors can be certainty), then there would be no uncertainty regarding
improved, then it may be possible to obtain a better state the true remaining useful life of the component/system.
estimate (with lesser uncertainty) during Kalman filter- However, this true RUL needs to be estimated using a
ing, which may in turn lead to a less uncertain RUL pre- model; the usage of a model imparts additional uncer-
diction. Another aspect of uncertainty management deals tainty as explained below.
with how uncertainty-related information can be used in 3. Modeling uncertainty: It is necessary to use a func-
the decision-making process. tional model in order to predict future state behavior.

3
 Annual Conference of the Prognostics and Health Management Society 2013

Further, as mentioned before, the end-of-life is also de- of interpretations: von Mises frequentist (Von Mises, 1981)
fined using a Boolean threshold function which indicates and Poppers propensity (Popper, 1959); the former is more
end-of-life by checking whether failure has occurred or easily understood and widely used.
not. These two models are combinedly used to predict
In the context of physical probabilities, randomness arises
the RUL, and may either be physics-based or data-driven.
only due to the presence of physical probabilities. If the true
It may be practically impossible develop models which
value of any particular quantity is deterministic, then it is not
accurately predict reality. Modeling uncertainty repre-
possible to associate physical probabilities to that quantity.
sents the difference between the predicted response and
In other words, when a quantity is not random but unknown,
the true response (which can neither be known nor mea-
then tools of probability cannot be used to represent this type
sured accurately), and comprises of several parts: model
of uncertainty. For example, the mean of a random variable,
parameters, model form, and process noise. While it may
sometimes referred to as the population mean, is determinis-
be possible to quantify these terms until the time of pre-
tic. It is meaningless to talk about its probability distribution.
diction, it is practically challenging to know their values
In fact, for any type of parameter estimation, the underlying
at future time instants.
parameter is assumed to be deterministic and only an esti-
4. Prediction method uncertainty: Even if all the above mate of this parameter is obtained. The uncertainty in the
sources of uncertainty can be quantified accurately, it is parameter estimate is addressed through confidence intervals.
necessary to quantify their combined effect on the RUL The interpretation of confidence intervals, as explained in the
prediction, and thereby, quantify the overall uncertainty forthcoming subsection, is sometimes confusing and mislead-
in the RUL prediction. It may not be possible to do this ing. Further, the uncertainty in the parameter estimate cannot
accurately in practice and leads to additional uncertainty. be used for further uncertainty quantification. For example,
This topic will be revisited again, later in Section 5. if the model parameters of a battery model are estimated un-
der a particular loading condition, then this uncertainty can-
While the different sources of uncertainty and the various not be used for quantifying the battery-response for a similar
uncertainty-related activities have been explained in detail, it loading condition. This is a serious limitation, since it is not
is important to understand how to interpret this uncertainty. possible to propagate uncertainty after parameter estimation,
This topic is detailed in the next section. which is often necessary in system-level uncertainty quantifi-
cation (Sankararaman, 2012).
3. I NTERPRETING U NCERTAINTY
Clearly, there are two limitations of the frequentist interpre-
A probabilistic approach to uncertainty representation and tation of probability. First, a truly deterministic but unknown
quantification has been most commonly used in the prognos- quantity cannot be assigned a probability distribution. Sec-
tics and health management domain. Though probabilistic ond, uncertainty represented using confidence intervals can-
methods, mathematical axioms and theorems of probability not be used for further uncertainty propagation. The second
have been well-established in the literature, there is consid- interpretation of probability, i.e. the subjective interpretation,
erable disagreement among researchers on the interpretation overcomes these limitations.
of probability. There are two major interpretations based on
physical and subjective probabilities, respectively. It is essen- 3.2. Subjective Probability
tial to understand the difference between these two interpre-
tations before attempting to interpret the uncertainty in RUL Subjective probabilities (de Finetti, 1977) can be assigned
prediction. to any statement. It is not necessary that the concerned
statement is in regard to an event which is a possible out-
3.1. Physical Probability come of a random experiment. In fact, subjective probabil-
ities can be assigned even in the absence of random exper-
Physical probabilities (Szabo, 2007), also referred to objec- iments. The Bayesian methodology is based on subjective
tive or frequentist probabilities, are related to random phys- probabilities, which are simply considered to be degrees of
ical experiments such as rolling dice, tossing coins, roulette belief and quantify the extent to which the statement is
wheels, etc. Each trial of the experiment leads to an event supported by existing knowledge and available evidence. In
(which is a subset of the sample space), and in the long run of recent times, the terms subjectivist and Bayesian have
repeated trials, each event tends to occur at a persistent rate, become synonymous with one another. Calvetti and Som-
and this rate is referred to as the relative frequency. These ersalo (Calvetti & Somersalo, 2007) explain that random-
relative frequencies are expressed and explained in terms of ness in the context of physical probabilities is equivalent to
physical probabilities. Thus, physical probabilities are de- lack of information in the context of subjective probabil-
fined only in the context of random experiments. The the- ities. In this approach, even deterministic quantities can be
ory of classical statistics is based on physical probabilities. represented using probability distributions which reflect the
Within the realm of physical probabilities, there are two types

4
 Annual Conference of the Prognostics and Health Management Society 2013

subjective degree of the analysts belief regarding such quan- ples (ri ; i = 1 to n) are measured. It is important to under-
tities. As a result, probability distributions can be assigned to stand that different RUL values are obtained due to inherent
parameters that need to be estimated, and therefore, this in- variability across the n different specimens, thereby confirm-
terpretation facilitates uncertainty propagation after parame- ing the presence of physical probabilities. Assume that these
ter estimation. Interestingly, subjective probabilities can also random samples belong to an underlying probability density
be applied in situations where physical probabilities are in- function (PDF) fR (r), with expected value E(R) = and
volved (Sankararaman, 2012). variance V ar(R) = 2 . The goal of uncertainty quantifica-
tion is to characterize this probability density function based
The concept of likelihood and its use in Bayes theorem are
on the available n data. Theoretically, an infinite amount of
key to the theory of subjectvive probability. The numerical
data is necessary to accurately estimate this PDF; however,
implementation of Bayes theorem may be complicated in
due to the presence limited data, the estimated PDF is not ac-
some practical cases, and several sampling techniques have
curate. As a result, both frequentists and subjectivists express
been developed by researchers to address this issue. Today,
uncertainty regarding the estimate itself. However, frequen-
Bayesian methods are used to solve a variety of problems in
tists and subjectivists quantify and express this uncertainty in
engineering. Filtering techinques such as particle filtering,
completely different ways.
Kalman filtering, etc. are also primarily based on the use of
Bayes theorem, and sequential sampling. For the sake of illustration, assume that the entire PDF can be
equivalently represented using its mean and variance; in other
3.3. Summary words, assume that the random variable R follows a two-
parameter distribution. Therefore, estimating the parameters
Both the frequentist and subjectivist approaches have been
and is equivalent to estimating the PDF. In the context of
well-established in the literature, in order to aid uncertainty
physical probabilities (frequentist approach), the true un-
quantification. In fact, both the approaches may yield sim-
derlying parameters and are referred to as population
ilar results (but different interpretations) for a few standard
mean and population standard deviation respectively. Let
problems involving Gaussian variables. Sometimes, both ap-
and s denote the mean and the standard deviation of the
proaches may be suitable for a given problem at hand; for
available n data. As stated earlier, due to the presence of lim-
example, Kalman filtering has a purely frequentist interpreta-
ited data, the sample parameters ( and s) will not be equal
tion based on least squares minimization as well as a purely
to the corresponding population parameters ( and ). The
Bayesian interpretation which relies on continuously updat-
fundamental assumption in this approach is that, since there
ing the uncertainty in the state estimates using Bayes theo-
are true but unknown population parameters, it is meaning-
rem. It is acceptable to interpret uncertainty using the fre-
less to talk about the probability distribution of any popula-
quentist approach or the Bayesian approach, provided the in-
tion parameter. Instead, the sample parameters are treated as
terpretation is suitable for the problem at hand. The following
random variables, i.e., if another set of n data were available,
section further explores this idea in the context of PHM and
then another realization of and s would have been obtained.
RUL estimation.
Using the sample parameters ( and ) and the number of
data available (n), frequentists construct confidence intervals
4. U NDERSTANDING U NCERTAINTY IN RUL
on the population parameters.
Consider the problem of estimating the remaining useful life
Confidence intervals can be constructed for both and
prediction, in the context of prognostics and health man-
(Haldar & Mahadevan, 2000). It is important that these
agement. Researchers have pursued two different classes of
intervals be interpreted correctly. As stated earlier, the inter-
methods for this purpose; while the first method is based on
pretation of confidence intervals may be confusing and mis-
reliability-testing, the second method is based on condition-
leading. A 95% confidence interval on does not imply that
monitoring and future behavior prediction. There is a sig-
the probability that lies in the interval is equal to 95%;
nificant difference in the interpretation of uncertainty, when
such a statement is wrong because is purely deterministic
RUL is estimated using these two different approaches. Un-
and physical probabilities cannot be associated with it. The
derstanding this difference is important for prognostics and
random variable here is in fact , and the interval calculated
decision-making, and this is focus of the present section.
using . Therefore, the correct implication is that the proba-
bility that the estimated confidence interval contains the true
4.1. Testing-Based Prognostics
population mean is equal to 95%.
Consider a simple numerical example where the remaining
Alternatively, it is also possible to address the problem of
useful life needs to be calculated at a given time instant. As-
computing fR (r) purely from a subjective (Bayesian) point of
sume that a set of run to failure experiments have been per-
view. One important difference now is that the Bayesian ap-
formed with high level of control, ensuring same usage and
proach does not clearly differentiate between sample param-
operating conditions. The time to failure for all the n sam-

5
 Annual Conference of the Prognostics and Health Management Society 2013

eters and population parameters. The probability distribu- that it is important to interpret the uncertainty in EOL and
tion of is directly computed using the available data (recall RUL subjectively.
that this was impossible in the frequentist approach since
is the true parameter and precise but unknown), and this un- 4.3. Why is the RUL Prediction Uncertain?
certainty is referred to as the analysts degree of belief for the
In light of the above discussion, it is necessary to revisit the
underlying true parameter . Similarly, the probability distri-
question Why is the RUL uncertain? from a new perspec-
bution of can also be computed. Recall that one realization
tive. While Section 2 explained that RUL is uncertain be-
of the parameters ( and ) uniquely define the PDF fR (r).
cause there are several sources of uncertainty which influ-
However, since the parameters are themselves uncertain, R is
ence RUL estimation, now it is clear that the uncertainty in
now represented by a family of distributions (Sankararaman
RUL could arise due to variability across multiple specimens
& Mahadevan, 2011), reflective of the fact that there is lim-
(testing-based prognostics scenario) or simply due to subjec-
ited data. This family of distributions will shrink to the true
tive uncertainty regarding a single specimen (condition-based
underlying PDF as the number of available data increases.
prognostics). The following section discusses the computa-
tion of RUL in detail by presenting a detailed framework for
4.2. Condition-Based Prognostics
uncertainty quantification in prognostics, and explains how to
Most of the discussion pertaining to testing-based prognostics calculate the uncertainty in remaining useful life prediction.
is not applicable to condition-based monitoring and prognos-
tics. The distinctive feature of condition-based monitoring 5. U NCERTAINTY Q UANTIFICATION IN RUL
is that each component/subsystem/system is considered by
First, a general computational framework for uncertainty
itself, and therefore, variability across specimens is non-
quantification in prognostics and remaining useful life pre-
existent. Any such variability is spurious and must not be
diction in presented. Second, it is illustrated as to how the
considered. At any generic time instant tP at which prognos-
problem of computing uncertainty in the remaining useful life
tics needs to be performed, the component/subsystem/system
prediction can be viewed as an uncertainty propagation prob-
is at a specific state. The actual state of the system is purely
lem. Third, the need of rigorous mathematical algorithms for
deterministic, i.e., the true value is completely precise, how-
uncertainty quantification in RUL is demonstrated using cer-
ever unknown. Therefore, if a probability distribution is as-
tain numerical examples. Finally, the challenges involved in
signed for this state, then this distribution is simply reflective
computing RUL uncertainty are discussed in detail.
of the analysts knowledge regarding this state and cannot be
interpreted from a frequentist point of view. Thus, by virtue
5.1. Computational Framework for Prognostics
of definition of condition-based monitoring, physical proba-
bilities are not present here, and a subjective (Bayesian) ap- Suppose that it is desired to perform prognostics and pre-
proach is only suitable for uncertainty quantification. dict the RUL at a generic time-instant tP . Daigle and
Goebel (Daigle & Goebel, 2011) explain that it is important to
The goal in condition-based prognostics is, at any generic
develop an architecture for model-based prognostics for prac-
time instant tP , to predict the remaining useful life of the
tical engineering purposes. This paper considers the architec-
component/subsystem/system as condition-based estimate of
ture in Fig. 1, where the whole problem of prognostics can be
the usage time left until failure. First, measurements until
considered to consist of the following three sub-problems:
time tP are used to estimate the state at time tP . Then,
using a forecasting method (which may be model-based or 1. Present state estimation
data-driven), future state values (corresponding to time in- 2. Future state prediction
stants greater than tP ) are computed. In order to forecast fu- 3. RUL computation
ture state values, it is also necessary to assume future loading
conditions (and operating conditions) which is a major chal- 5.1.1. State Estimation
lenge in condition-based prognostics. Typically, the analyst
The first step of estimating the state at tP serves as the pre-
subjectively assumes statistics for future loading conditions
cursor to prognosis and RUL computation. Consider the state
based on past experience and existing knowledge; thus, the
space model which is used to continuously predict the state
subjective interpretation of uncertainty is clearly consistent
of the system, as:
across the entire condition-based monitoring procedure, and
therefore, inferences made out of condition-based monitor- x(t) = f (t, x(t), (t), u(t), v(t)) (1)
ing also need to be interpreted subjectively. This forecast-
ing is stopped when failure is reached, as indicated by the where x(t) Rnx is the state vector, (t) Rn is the
aforementioned boolean threshold function. This indicates parameter vector, u(t) Rnu is the input vector, v(t) Rnv
the end-of-life (EOL) and the EOL can be directly used to is the process noise vector, and f is the state equation.
compute the remaining useful life (RUL) prediction. Note The state vector at time tP , i.e., x(t) (and the parameters (t),

6
 Annual Conference of the Prognostics and Health Management Society 2013

Continue future state prediction until failure

x(t)
u(t) y(t) x(tP )
System 1. Estimation 2. Prediction Check for failure
t > tP

3. RUL Computation

Figure 1. Model-Based Prognostics Architecture

if they are unknown) is (are) estimated using output data col- TEOL is equal to 1 when any of the constraints are violated.
lected until tP . Let y(t) Rny , n(t) Rnn , and h de- Then, the End of Life (EOL, denoted by E) at any time instant
note the output vector, measurement noise vector, and output tP is then defined as the earliest time point at which the value
equation respectively. Then, of TEOL becomes equal to one. Therefore,

y(t) = h(t, x(t), (t), u(t), n(t)) (2) E(tP ) , inf{t R : t tP TEOL (x(t), (t), u(t)) = 1}. (4)

Typically, filtering approaches such as Kalman filtering, par- The Remaining Useful Life (RUL, denoted by R) at time in-
ticle filtering, etc. may be used for such state estimation. It stant tP is expressed as:
must be recalled that these filtering methods are collectively
known as Bayesian tracking methods, not only because they R(tP ) , E(tP ) tP . (5)
use Bayes theorem for state estimation but also are based on
the subjective interpretation of uncertainty. In other words, Note that the output equation (Eq. 2) or output data (y(t)) is
any time instant, there is nothing uncertain regarding the true not used in the prediction stage, and EOL and RUL are depen-
states. However, the true states are not known precisely, and dent only on the state estimates at time tP ; though these state
therefore, the probability distributions of these state variables estimates are obtained using the output data, the output data
are estimated through filtering. The estimated probability dis- is not used for EOL/RUL calculation after state estimation.
tributions are simply reflective of the subjective knowledge For the purpose of implentation, f in Eq. 1 is transformed into
regarding those state variables. the corresponding discrete-time version.

5.1.2. State Prediction 5.2. RUL Estimation through Uncertainty Propagation

Having estimated the state at time tP , Eq. 1 is used to predict Thus, it is clear that RUL predicted at time tP , i.e., R(tP )
the future states of the component/system. This differential depends on
equation can be discretized and used to predict x(t + 1) as a
function of x(t). Therefore, using this recursive relation, the 1. Present state estimate (x(tP )); using the present state es-
state at any future time instant t > tP can be calculated. timate and the state space equations in Eq. 1, the future
states (x(tP ), x(tP + 1), x(tP + 2), ..., x(tP + R(tP )))
5.1.3. RUL Computation can be calculated.
RUL computation is concerned with the performance of the 2. Future loading (u(tP ), u(tP + 1), u(tP + 2), ..., u(tP +
component that lies outside a given region of acceptable be- R(tP ))); these values are needed to calculate the future
havior. The desired performance is expressed through a set of state values using the state space equations.
nc constraints, CEOL = {ci }ni=1 c
, where ci : Rnx Rn 3. Parameter values from time tP until time tP + R(tP )
nu
R B maps a given point in the joint state-parameter (denoted by (tP ), (tP + 1), ..., (tP + R(tP ))).
space given the current inputs, (x(t), (t), u(t)), to the
Boolean domain B , [0, 1], where ci (x(t), (t), u(t)) = 1 if 4. Process noise (v(tP ), v(tP + 1), v(tP + 2), ..., v(tP +
the constraint is satisfied, and 0 otherwise (Daigle & Goebel, R(tP ))).
2013).
For the purpose of RUL prediction, all of the above quan-
These individual constraints may be combined into a single tities are independent quantities and hence, RUL becomes a
threshold function TEOL : Rnx Rn Rnu B, defined dependent quantity. Let X = {X1 , X2 , ...Xi , ...Xn } denote
as: the vector of all the above dependent quantities, where n is
the length of the vector X, and therefore the number of un-
(
1, 0 {ci (x(t), (t), u(t))}n c
i=1
TEOL (x(t), (t), u(t)) = (3) certain quantities that influence the RUL prediction. Then the
0, otherwise.
calculation of RUL (denoted by R) can be expressed in terms

7
 Annual Conference of the Prognostics and Health Management Society 2013

of a function, as: While Gaussian distributions and linear state space models
R = G(X) (6) (linear f in Eq. 1) may be commonly used in the prognostics
The above functional relation in Eq. 6 can be graphically ex- and health management domain, it is important to understand
plained, as shown in Fig. 2. that using linear state space models is not equivalent to G
being linear. In other words, the use of the threshold function
For example, consider the case where the component/system along with the linear state models automatically renders G
is subjected to uniform loading (characterized by one vari- non-linear.
able, the amplitude which remains constant with time), mod-
eled using one parameter (which is time-invariant), and char- In order to illustrate this important point, and to empha-
acterized using two states (the state estimates at time tP and size the importance of using rigorous computational meth-
Eq. 1 can be used to predict the state values at any future time ods, consider a simple example where the state state equation
instant). Then, excluding the effect of process noise, there is given by:
are n = 4 quantities that affect the RUL prediction. Note x(t + 1) = a.x(t) + b. (7)
that there are R(tP ) + 1 process noise terms for each state; Assume that a suitable time-discretization has been chosen
therefore, the inclusion of process noise increases the value for the purpose of implementation. It is desired to predict fu-
of n, and therefore the dimensionality of the problem. This ture behavior and compute RUL at tP = 0, and state value at
raises a practical concern and has been addressed in an earlier this time is denoted by x(0) which is a Gaussian random vari-
publication by replacing the time-variant process noise using able. Further a and b are constants (i.e., not random) which
an equivalent time-invariant process noise (Sankararaman & are used to predict future states. It can be easily demonstrated
Goebel, 2013). In the rest of the paper, a generalized frame- that the state value at any future time instant can be expressed
work is presented without using this equivalent time-invariant as a function of x(0).
process noise concept.
X
j=n1
Knowing the values of X, it is possible to compute the value x(n) = an .x(0) + aj b (8)
of R, using Fig. 2 that is equivalently represented by Eq. 6. j=0
The quantities contained in X are uncertain, and the focus in
It is clear from Eq. 8 that the state value at any future time
prognostics to compute their combined effect on the RUL pre-
instant is a linear function of x(0), and therefore is also Gaus-
diction, and thereby compute the probability distribution of
sian. In order to compute the remaining useful life, it is neces-
R. The problem of estimating the uncertainty in R is equiva-
sary to chose a threshold function. Depending on the choices
lent to propagating the uncertainty in X through G, and it is
of a and b, x(n) may either be an increasing function or a
necessary to use computational methods for this purpose.
decreasing function. If x(n) is a decreasing function, then
the threshold function will indicate that failure occurs when
5.3. Need for Computational Approaches
the state value x becomes smaller than a critical lower bound
The problem of estimating the uncertainty in R using uncer- (l). Alternatively, if x(n) is an increasing function, then the
tainty propagation techniques is a non-trivial problem, and threshold function will indicate that failure occurs when the
needs rigorous computational approaches. This involves es- state value x becomes greater than a critical upper bound u.
timating the probability density function of R (PDF, denoted Without loss of generality, any of the two cases may be cho-
by fR (r)) or equivalently the cumulative distribution func- sen for illustrative purposes. For example, consider that x(n)
tion of R (CDF, denoted by FR (r)). In some rare cases, it is is decreasing and failure happens when x < l. Therefore, the
possible to analytically obtain the distribution of R. Some of remaining useful life (r, an instance of the random variable
such special cases are listed below: R) is equal to the smallest n such that x(n) < l. Therefore
RUL can be calculated as
1. Each and every quantity contained in X follows a nor- X
j=n1
mal (Gaussian) distribution, and the function G can be r = inf{n : an .x(0) + aj b < l}, (9)
expressed as a weighted linear combination of the quan- j=0
tities in X. In this case, R also follows a normal distri-
bution, and its statistics can be calculated analytically. Assuming that the chosen time-discretization level is in-
finitesimally small, it is possible to directly estimate the RUL
2. Each and every quantity contained in X follows a log- by solving the equation:
normal distribution, and if the logarithm of the function
G can be expressed as a weighted combination of the X
j=r1
ar .x(0) + aj .b = l. (10)
quantities in X, then log(R) follows a normal distribu-
j=0
tion whose statistics can be estimated analytically. In
other words, R also follows a lognormal distribution.

8
 Annual Conference of the Prognostics and Health Management Society 2013

Present State Future Loading Future Parameters Future Process Noise

u(tP ), u(tP + 1), (tP ), (tP + 1), v(tP ), v(tP + 1),

x(tP )
... u(tP + R(tP ) ... (tP + R(tP )) ...v(tP + R(tP ))

t = tP

Discretize Eq. 1

if TEOL = 1
Compute x(t + 1) TEOL R = t tP

Using
x(t) if TEOL = 0 R = G(X)

Assign t = t + 1

Figure 2. Definition of G

The above equation calculates the RUL (r) as a function of the forward.
initial state (x(0)). Hence, the above equation is similar to G
Practical problems in the prognostics and health management
defined earlier in Fig. 2. The difference now is that the only
domain may consist of:
considered source of uncertainty is the state estimate x(0);
model uncertainty, future loading uncertainty, etc. are not 1. Several non-Gaussian random variables which affect the
considered here. The RUL R follows a Gaussian distribution RUL prediction,
if and only if it is linearly dependent on x(0). In other words, 2. A non-linear multi-dimensional state space model,
R follows a Gaussian distribution if and only if Eq. 10 can be
3. Uncertain future loading conditions,
rewritten as:
.r + .x(0) + = 0 (11) 4. A complicated threshold function which may be defined
in multi-dimensional space.
for some arbitrary values of , , and . If it were possible to
estimate such values for , , and , the distribution of RUL The fact that the distribution of RUL simply depends on the
can be obtained analytically. quantities indicated in Fig. 2 implies that it is technically in-
accurate to artifically assign the probability distribution type
In order to examine if this is possible, rewrite Eq. 10 as: (or any statistic such as the mean or variance) to RUL.

1 X
j=r1
x(0) = (l aj .b) (12) 5.4. Illustrations
a r
j=0 Sometimes, the probability distribution of RUL may be ex-
While x(0) is completely on the left hand side of this equa- tremely skewed; for example, the RUL of a lithium-ion bat-
tion, r appears not only as an exponent in the denominator tery used to power an unmanned aerial vehicle was predicted
but is also indicative of the number of terms in the summa- by Sankararaman et al. (Sankararaman et al., 2013) and it
tion on the right hand side of the above equation. Therefore, was observed that the probability distribution was particu-
it is clear that the relationship between r and x(0) is not lin- larly skewed near failure. The PDFs of the End-of-Discharge
ear. Therefore, even if the initial state (x(0), a realization of (EOD) prediction at various time instants (T = 0 seconds
X(0)) follows a Gaussian distribution, the RUL (r, a real- through T = 4000) are shown in Fig. 3 and then the PDF
ization of R) does not follow a Gaussian distribution. Thus, of End-of-Discharge predicted at T = 5000 seconds (which
it is clear that even for a simple problem consisting of lin- corresponds to near-failure) is indicated in Fig. 4. The RUL
ear state models, an extremely simple threshold function, and can be calculated by simply subtracting the prediction time-
only one uncertain variable that is Gaussian, the calculation of instant from the EOD prediction. Note the significant change
the probability distribution of R is neither trivial nor straight- in the shape of the PDF near failure. It is extremely important
to be able to accurately predict the RUL particularly as failure

9
 Annual Conference of the Prognostics and Health Management Society 2013

-3
is approaching, and it is clear from Fig. 4 that assuming a nor- x 10
mal distribution or an arbitrary standard deviation would not 6
be able to achieve this goal; only a theoretically accurate un-
5

Probability Density Function

certainty quantification method can reproduce this probabil-
ity distribution, whose mode almost coincides with its lower 4
bound (left-hand-side tail).
3
-3
x 10
1.5 2
at T=0 s
Probability Density Function

at T=1000 s 1
1 at T=2000 s
at T=3000 s
0
0 200 400 600 800 1000 1200
at T=4000 s EOD (s)
0.5 Figure 5. A Multi-Modal PDF

5.5. Uncertainty Propagation Methods

0
2000 4000 6000 8000 10000 In order to answer the obvious question: How to calculate
EOD (s)
the uncertainty in R and estimate the PDF of R?, it is neces-
Figure 3. EOD Prediction at Multiple Time Instants sary to resort to rigorous computational methodologies which
have been developed by statisticians and researchers in the
-3
field of uncertainty quantification in order to solve a typical
x 10 uncertainty propagation problem. There are different types of
3.5
sampling methods such as Monte Carlo sampling (Caflisch,
3
Probability Density Function

1998), Latin hypercube sampling (Loh, 1996), adaptive sam-

2.5 pling (Bucher, 1988), importance sampling (Glynn & Igle-
hart, 1989), unscented transform sampling (Van Zandt, 2001),
2 etc. Alternatively, there are analytical methods such as the
first-order second moment method (Dolinski, 1983), first-
1.5
order reliability method (Hohenbichler & Rackwitz, 1983),
1 second-order reliability method (Der Kiureghian, Lin, &
Hwang, 1987), etc. In addition, there are also methods such
0.5
as the efficient global reliability analysis (Bichon, Eldred,
0 Swiler, Mahadevan, & McFarland, 2008) method which in-
5000 6000 7000 8000 9000 10000 volve both sampling and the use of analytical techniques. All
EOD (s)
of these methods empirically calculate the probability distri-
Figure 4. EOD Prediction at T = 5000 seconds (near failure) bution of RUL; while some of these methods calculate the
PDF (fR (r)) of RUL, some other methods calculate the CDF
Sometimes, depending on the chosen statistics of future load- (FR (r)), and some other methods directly generate samples
ing conditions, the distribution of EOD may even be multi- from the target probability density function (fR (r)). Due to
modal. For example, Saha et al. (Saha & Goebel, 2008) calcu- some limitations of each of these methods, it may not be pos-
lated future loading statistics that lead to a multi-modal PDF sible to accurately calculate the actual probability distribution
for the EOD, as shown in Fig. 5. of R. Accurate calculation is possible only by using infinite
samples for Monte Carlo sampling. Any other method (for
It is important to capture such characteristics of the RUL
example, the use of a limited, finite number of samples) will
(which is equivalent to the end-of-discharge in Fig 3-5) prob-
lead to uncertainty in the estimated probability distribution,
ability distribution, and this can be accomplished only by us-
and this additional uncertainty is referred to as prediction-
ing accurate uncertainty quantification methodologies with-
method uncertainty. It is possible to decrease (and maybe
out making critical assumptions regarding the shape of the
eventually eliminate) this type of uncertainty either by us-
PDF of the RUL, its mean, median, mode, standard devia-
ing advanced probability techniques or powerful computing
tion, etc. Therefore, the goal must be to accurately calcuate
power.
the probability distribution of R by propagating the different
sources of uncertainty through G as indicated in Fig. 2. It is necessary to further investigate the aforementioned un-

10
 Annual Conference of the Prognostics and Health Management Society 2013

certainty propagation methods, and identify whether they can 6. C ONCLUSION

be applied to prognostics health monitoring. Some earlier
This paper discussed the significance and interpretation of un-
publications have investigated the use of certain methods
certainty the context of prognostics and health management.
such as Monte Carlo sampling, unscented transform sam-
The prediction of remaining useful life in engineering sys-
pling, first-order reliability methods, etc. in this regard.
tems is affected by several sources of uncertainty, and it is
important to correctly interpret this uncertainty in order to
5.6. Challenges
facilitate meaningful decision-making. Uncertainty can be
There are several challenges in using different uncertainty interpreted in two ways, either in terms of physical proba-
quantification methods for prognostics, health management bilities from a frequentist point of view or in terms of sub-
and decision-making. It is not only important to understand jective probabilities from a Bayesian point of view. While
these challenges but also necessary to understand the require- a frequentist interpretation may be suitable for testing-based
ments of PHM systems in order to integrate efficient un- prognostics, there are no physical probabilities in the context
certainty quantification along with prognostics and aid risk- of condition-based prognostics. Therefore, uncertainty in the
informed decision-making. Some of the issues involved in context of condition-based monitoring needs to be interpreted
such integration are outlined below: subjectively, and hence, a Bayesian approach is more suitable
for this purpose. It was also explained that Bayesian tracking
1. An uncertainty quantification methodology for prognos- methods for state estimation are so-called not only because
tics needs to be computationally feasible for implementa- they use Bayes theorem but are also based on the principle of
tion in online health monitoring. This requires quick cal- subjective probability.
culations, while uncertainty quantification methods have
This paper also emphasized the importance of accurately
been traditionally known to be time-consuming and com-
computing the uncertainty in the remaining useful life predic-
putationally intensive.
tion. It was illustrated that it may not be analytically possible
2. Sometimes, the probability distribution of RUL may be to calculate the uncertainty in the remaining useful life pre-
multi-modal and the uncertainty quantification method- diction even for certain simple problems involving Gaussian
ology needs to be able to accurately capture such distri- random variables and linear state-prediction models. There-
butions. fore, it is necessary to resort to computational methodologies
3. Existing verification, validation, and certification proto- for such uncertainty quantification and compute the proba-
cols require algorithms to produce deterministic, i.e., re- bility distribution of remaining useful life prediction. In this
peatable calculations. Several uncertainty quantification process, it is important not to make assumptions regarding the
methods are non-deterministic, i.e. produce different (al- shape of the probability distribution of the remaining useful
beit, only slightly if implemented well) results on repeti- life prediction or any of its statistics such as the mean, me-
tion. dian, standard deviation, etc.
4. The uncertainty quantification method needs to be accu- Finally, it was explained that the problem of estimating the
rate, i.e., the entire probability distribution of X needs probability distribution of remaining useful life can be viewed
to be correctly accounted for, and the functional relation- as an uncertainty propagation problem which can be solved
ship defined by G in Fig. 2. Some methods use only using different types of computational approaches. Sev-
a few statistics (usually, mean and variance) of X and eral sampling-based methods, analytical methods and hybrid
some methods make approximations (say for example, methods have been developed by researchers in the field of
linear) of G. Finally, it is important to correctly prop- uncertainty quantification and it is necessary to investigate
agate the uncertainty to compute the entire probability the applicability of these methods to prognostics and health
distribution of RUL. management. Further, several challenges involved in integrat-
5. While it is important to be able to calculate the entire ing uncertainty quantification techniques into prognostics and
probability distribution of RUL, it is also important to health management were outlined. It is clear that further re-
be able to quickly obtain bounds on RUL which can be search is necessary to address these challenges and develop
useful for online decision-making. a comprehensive framework for uncertainty quantification in
prognostics and health management.
Each uncertainty quantification method may address one or
more of the above issues, and therefore, it may even be ACKNOWLEDGMENT
necessary to resort to different methods to achieve different
goals. Future research needs to continue this investigation, The work reported herein was in part funded by the NASA
analyze different types of uncertainty quantification methods System-wide Safety Assurance Technologies (SSAT) project
and study their applicability to prognostics before these meth- under the Aviation Safety (AvSafe) Program of the Aero-
ods can be applied in practice. nautics Research Mission Directorate (ARMD), and by the

11
 Annual Conference of the Prognostics and Health Management Society 2013

NASA Automated Cryogenic Loading Operations (ACLO) neering systems planning and design. In Engineering
project under the Office of the Chief Technologist (OCT) of systems symposium mit. Cambridge, MA..
Advanced Exploration Systems (AES). The authors would Der Kiureghian, A., Lin, H.-Z., & Hwang, S.-J. (1987).
like to acknowledge valuable discussions with Dr. Abhinav Second-order reliability approximations. Journal of
Saxena and Dr. Jose Celaya, since these discussions raised Engineering Mechanics, 113(8), 12081225.
several important questions regarding uncertainty in prognos- Dolinski, K. (1983). First-order second-moment approxima-
tics, and ultimately, led to the ideas explored in this paper. tion in reliability of structural systems: critical review
and alternative approach. Structural Safety, 1(3), 211
231.
R EFERENCES Engel, S. J., Gilmartin, B. J., Bongort, K., & Hess, A. (2000).
Bichon, B., Eldred, M., Swiler, L., Mahadevan, S., & McFar- Prognostics, the real issues involved with predicting
land, J. (2008). Efficient global reliability analysis for life remaining. In Aerospace Conference Proceedings,
nonlinear implicit performance functions. AIAA jour- 2000 IEEE (Vol. 6, pp. 457469).
nal, 46(10), 24592468. Farrar, C. R., & Lieven, N. A. (2007). Damage prognosis:
Bucher, C. G. (1988). Adaptive samplingan iterative fast the future of structural health monitoring. Philosophi-
monte carlo procedure. Structural Safety, 5(2), 119 cal Transactions of the Royal Society A: Mathematical,
126. Physical and Engineering Sciences, 365(1851), 623
Caflisch, R. E. (1998). Monte carlo and quasi-monte carlo 632.
methods. Acta numerica, 1998, 149. Glynn, P. W., & Iglehart, D. L. (1989). Importance sam-
Calvetti, D., & Somersalo, E. (2007). Introduction to pling for stochastic simulations. Management Science,
bayesian scientific computing: ten lectures on subjec- 35(11), 13671392.
tive computing (Vol. 2). Springer New York. Gu, J., Barker, D., & Pecht, M. (2007). Uncertainty assess-
Celaya, J., Saxena, A., & Goebel, K. (2012). Uncertainty ment of prognostics of electronics subject to random
representation and interpretation in model-based prog- vibration. In AAAI fall symposium on artificial intelli-
nostics algorithms based on kalman filter estimation. gence for prognostics (pp. 5057).
In Proceedings of the Annual Conference of the PHM Haldar, A., & Mahadevan, S. (2000). Probability, reliabil-
Society (pp. 2327). ity, and statistical methods in engineering design. John
Celaya, J., Saxena, A., Kulkarni, C., Saha, S., & Goebel, K. Wiley & Sons, Incorporated.
(2012). Prognostics approach for power MOSFET un- Hastings, D. and McManus, H. (2004). A framework for un-
der thermal-stress aging. In Reliability and Maintain- derstanding uncertainty and its mitigation and exploita-
ability Symposium (RAMS), 2012 Proceedings-Annual tion in complex systems. In Engineering Systems Sym-
(pp. 16). posium MIT (p. 19). Cambridge MA..
Coppe, A., Haftka, R. T., Kim, N. H., & Yuan, F.-G. (2010). Hohenbichler, M., & Rackwitz, R. (1983). First-order con-
Uncertainty reduction of damage growth properties us- cepts in system reliability. Structural safety, 1(3), 177
ing structural health monitoring. Journal of Aircraft, 188.
47(6), 20302038. Liao, H., Zhao, W., & Guo, H. (2006). Predicting remaining
Daigle, M., & Goebel, K. (2010). Model-based prognostics useful life of an individual unit using proportional haz-
under limited sensing. In Aerospace Conference, 2010 ards model and logistic regression model. In Reliability
IEEE (pp. 112). and Maintainability Symposium, 2006. RAMS06. An-
Daigle, M., & Goebel, K. (2011). A model-based prognostics nual (pp. 127132).
approach applied to pneumatic valves. International Loh, W.-L. (1996). On latin hypercube sampling. The annals
Journal of Prognostics and Health Management, 2(2). of statistics, 24(5), 20582080.
Daigle, M., & Goebel, K. (2013). Model-based prog- Ng, K.-C., & Abramson, B. (1990). Uncertainty management
nostics with concurrent damage progression pro- in expert systems. IEEE Expert Systems, 20.
cesses. Systems, Man, and Cybernetics: Sys- Orchard, M., Kacprzynski, G., Goebel, K., Saha, B., & Vacht-
tems, IEEE Transactions on, 43(3), 535-546. doi: sevanos, G. (2008, oct.). Advances in uncertainty rep-
10.1109/TSMCA.2012.2207109 resentation and management for particle filtering ap-
Daigle, M., Saxena, A., & Goebel, K. (2012). An efficient plied to prognostics. In Prognostics and Health Man-
deterministic approach to model-based prediction un- agement, 2008. PHM 2008. International Conference
certainty estimation. In Annual conference of the prog- on (p. 1 -6). doi: 10.1109/PHM.2008.4711433
nostics and health management society (pp. 326335). Popper, K. (1959). The propensity interpretation of probabil-
de Finetti, B. (1977). Theory of probability, Volumes I and II. ity. The British journal for the philosophy of science,
Bull. Amer. Math. Soc. 83 (1977), 94-97, 00029904. 10(37), 2542.
deNeufville, R. (2004). Uncertainty management for engi- Saha, B., & Goebel, K. (2008). Uncertainty management for

12
 Annual Conference of the Prognostics and Health Management Society 2013

diagnostics and prognostics of batteries using Bayesian B IOGRAPHIES

techniques. In Aerospace Conference, 2008 IEEE (pp.
Shankar Sankararaman received his B.S.
18).
degree in Civil Engineering from the In-
Sankararaman, S. (2012). Uncertainty quantification and in-
dian Institute of Technology, Madras in In-
tegration in engineering systems (Ph.D. Dissertation).
dia in 2007 and later, obtained his Ph.D. in
Vanderbilt University.
Civil Engineering from Vanderbilt Univer-
Sankararaman, S., Daigle, M., Saxena, A., & Goebel, K.
sity, Nashville, Tennessee, U.S.A. in 2012.
(2013). Analytical algorithms to quantify the uncer-
His research focuses on the various aspects
tainty in remaining useful life prediction. In Aerospace
of uncertainty quantification, integration, and management
Conference, 2013 IEEE (pp. 111).
in different types of aerospace, mechanical, and civil engi-
Sankararaman, S., & Goebel, K. (2013, Apr). Uncertainty
neering systems. His research interests include probabilis-
quantification in remaining useful Life of aerospace
tic methods, risk and reliability analysis, Bayesian networks,
components using state space models and inverse
system health monitoring, diagnosis and prognosis, decision-
FORM. In Proceedings of the 15th Non-Deterministic
making under uncertainty, treatment of epistemic uncertainty,
Approaches Conference.
and multidisciplinary analysis. He is a member of the Non-
Sankararaman, S., Ling, Y., & Mahadevan, S. (2011). Un-
Deterministic Approaches (NDA) technical committee at the
certainty quantification and model validation of fatigue
American Institute of Aeronautics, the Probabilistic Methods
crack growth prediction. Engineering Fracture Me-
Technical Committee (PMC) at the American Society of Civil
chanics, 78(7), 14871504.
Engineers (ASCE), and the Prognostics and Health Manage-
Sankararaman, S., Ling, Y., Shantz, C., & Mahadevan, S.
ment (PHM) Society. Currently, Shankar is a researcher at
(2011). Uncertainty quantification in fatigue crack
NASA Ames Research Center, Moffett Field, CA, where he
growth prognosis. International Journal of Prognos-
develops algorithms for uncertainty assessment and manage-
tics and Health Management, 2(1).
ment in the context of system health monitoring, prognostics,
Sankararaman, S., & Mahadevan, S. (2011). Likelihood- and decision-making.
based representation of epistemic uncertainty due to
sparse point data and/or interval data. Reliability En- Kai Goebel is the Deputy Area Lead for
gineering & System Safety, 96(7), 814824. Discovery and Systems Health at NASA
Szabo, L. (2007). Objective probability-like things with and Ames where he also directs the Prognos-
without objective indeterminism. Studies In History tics Center of Excellence. After receiving
and Philosophy of Science Part B: Studies In History the Ph.D. from the University of California
and Philosophy of Modern Physics, 38(3), 626634. at Berkeley in 1996, Dr. Goebel worked at
Tang, L., Kacprzynski, G., Goebel, K., & Vachtsevanos, G. General Electrics Corporate Research Cen-
(2009, march). Methodologies for uncertainty manage- ter in Niskayuna, NY from 1997 to 2006 as a senior research
ment in prognostics. In Aerospace conference, 2009 scientist before joining NASA. He has carried out applied re-
IEEE (p. 1 -12). doi: 10.1109/AERO.2009.4839668 search in the areas of artificial intelligence, soft computing,
Van Zandt, J. R. (2001). A more robust unscented trans- and information fusion and his interest lies in advancing these
form. In International symposium on optical science techniques for real time monitoring, diagnostics, and prog-
and technology (pp. 371380). nostics. He holds 17 patents and has published more than
Von Mises, R. (1981). Probability, statistics and truth. Dover 250 papers in the area of systems health management.
Publications.

13