This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 1

An Adaptive Prognostic Approach Via Nonlinear

Degradation Modelling: Application to Battery Data
Xiao-Sheng Si

Abstract—Remaining useful life (RUL) estimation via degrada- The prognostic part of PHM is typically characterized by
tion modelling is considered as one of the most central components estimating the remaining useful life (RUL) of a system, con-
in prognostics and health management. Current RUL estimation ditional on the available information at hand. In fact, the
studies focus mainly on linear stochastic models and the results
under nonlinear models are relatively limited in literature. Even estimation of the RUL has been considered as one of the most
in nonlinear degradation modelling, the estimated RUL is aimed central components in PHM [12]–[19]. Effective and accurate
at a population of systems of the same type or depend only on the RUL estimation can avoid catastrophic events, extend life-
current degradation observation. In this paper, an adaptive and cycle, and schedule timely health care actions [20]–[22]. Many
nonlinear prognostic model is presented to estimate RUL using a technological systems are subject, during their operating life,
system’s history of the observed data to date. Specifically, a general
nonlinear stochastic process with a time-dependent drift coefficient to a gradual deterioration process that progressively degrades
is first adopted to characterize the dynamics and nonlinearity of their performance until a failure occurs [13], [23], [24]. For
the degradation process. In order to render the RUL estimation example, lithium-ion batteries have been widely used in many
depend on the degradation history to date, a state space model fields, including consumer electronics, electric vehicles, marine
is constructed and Kalman filtering is applied to update one key systems, aircrafts, satellites, etc., due to their high power
parameter in the drifting function through treating this parameter
as an unobserved state variable. To update the hidden state density, low weight, long lifetime, no memory effect and other
and other parameters in the state space model simultaneously advantages. Nevertheless, an inevitable problem is that lithium-
and recursively, the expectation maximization algorithm is used ion battery performance degrades with cycling and aging. Such
in conjunction with Kalman smoother to achieve this aim. The functionality degradation of lithium-ion batteries may cause re-
probability density function of the estimated RUL is derived with duced performance and even catastrophic failure. For example,
an explicit form and some commonly used results under linear
models turn out to be its special cases. Finally, the implementation in 2013, all Boeing 787 Dreamliners were indefinitely grounded
of the presented approach is illustrated by numerical simulations due to battery failures that occurred on two planes. However,
and an application for estimating the RUL of lithium-ion batteries the progression of degradation of systems such as lithium-ion
is used to demonstrate the superiority of the method. batteries and electromechanical systems is typically stochastic
Index Terms—Degradation, lifetime estimation, nonlinear, pre- in nature, resulting in the difficulty to estimate the RUL with
diction method, prognostics and health management, battery. certainty. To characterize the uncertainty of the estimated RUL,
the form of the probability density function (PDF) of the RUL
is often required. Thus, this paper pays a particular attention to
I. I NTRODUCTION
the estimation of the PDF of the required RUL.
ITH the ever-increased high requirement of reliability
W and safety for critical systems, accurately assessing the
pending failure of a system has become an active research area
The degradation processes of technological systems are usu-
ally affected by two kind of variability, namely individual
variability and temporal variability [25], [26]. The individual
over the past decades. This also leads to an emerging concept variability determines heterogeneity among the degradation
called prognostics and health management (PHM) [1]–[3]. PHM paths of different systems. The temporal uncertainty is referred
is an enabling discipline consisting of technologies and methods to as the inherent uncertainty associated with the progression
to assess the reliability of a system in its actual life cycle of the degradation over time. For the individual variability, it
conditions, to determine the advent of failure, and to mitigate can be usually modelled by introducing system-specific random
system risk. The past decade has witnessed an increasing effects by some model parameters [27], [28]. Some efforts in
research interest on various aspects of PHM due primarily to this respect have been made by proposing Bayesian procedures
the fact that PHM have been extensively applied in a variety for real-time estimation of the RUL of degrading units on the
of fields including electronics, smart grid, batteries, bearings, basis of sensor-based degradation models [29]. However, they
motor drives, electromechanical structures, analog electronic assumed that the variance parameter of the model is fixed for
circuits, power industry, aerospace and military application, different units and thus is estimated by off-line data. In practice,
public health management [4]–[11]. different units have different degradation magnitudes and hence
different variance parameters. In addition, the fixed variance
Manuscript received June 18, 2014; revised September 25, 2014 and Novem-
ber 16, 2014; accepted December 20, 2014. parameter dominates the variance of the estimated RUL for a
Copyright (c) 2014 IEEE. Personal use of this material is permitted. However, particular system since random effects on model parameters of
permission to use this material for any other purposes must be obtained from this kind of models are very small for a particular system [25].
the IEEE by sending a request to pubs-permissions@ieee.org.
Xiao-Sheng Si is with the Department of Automation, Xi’an Institute of For each single system, the temporal variability determines
High-Tech., Xi’an, Shaanxi 710025, China (e-mail: sixiaosheng@gmail.com). the random fluctuation over time of the degradation process.

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 2

As advocated by [30] and [31], this can be described by an Cox Proportional Hazard (PH) model where the degradation
appropriate stochastic process which provides flexibility on signals are used as covariate information. Then, by updating
describing the characteristics of the temporal effects. There- the parameters in the degradation model, the RUL prediction is
fore, this paper mainly considers the stochastic process based achieved by formulating the relationship between the survival
modelling approaches for prognostics. The central idea in this function and the hazard rate function. Therefore, the proposed
kind of approaches is that the lifetime can be defined as the prediction method in [40] focuses mainly on the case of hard
first hitting time (FHT) of the degradation process crossing a failure. In other words, both the degradation signals and time-to-
failure threshold [32], i.e. soft failure. event data are required to be available. However, time-to-event
If the degradation process is monotonic, the relationship data might be scarce or even non-existent for systems which
between the degradation and the FHT can be easily established are costly or time-consuming to collect time-to-event data. In
[33]. For non-monotonic degradation process, a Wiener process this case, the hazard rate and its associated parameters are
is a commonly used stochastic process-based model with a difficult to determine. In addition, the mixed-effects regression
constant degradation drift and with the dynamic part driven by models might not be effectively model the temporal variability
the standard Brownian motion. However, most of the works on in stochastic degradation processes [30], [31].
the RUL estimation based on the Wiener process assume that There are also prognostic studies with nonlinear models,
the degradation process has a linear path or can be linearized by but under simulation based framework [41]–[43]. For example,
logarithmic or time-scale transformations [16], [24], [28], [34]– Zhou et al. in [42] proposed a nonlinear state space model to
[36]. Such assumption restricts the applications of these prog- estimate the RUL, where the RUL was estimated by Monte
nostic methods in several complex circumstances. For example, Carlo method and the unknown parameters were estimated
during the aircraft flight process, the engine load varies as the through expectation-maximization (EM) algorithm. Orchard et
aircraft is exposed to several distinct flight conditions including al. in [43] proposed the concept of artificial evolution to
takeoff, maximum climb, maximum cruise, ground idle, etc. estimate unknown model parameters within a particle-filtering-
Moreover, if inclement weather is encountered during flight, the based framework and an adaptation mechanism to modify the
aircraft may change altitude, accelerate, or decelerate to avoid variance of the uncertainty source that defines the performance
turbulence. Another example is the lithium-ion batteries which of the artificial evolution algorithm. To do so, the health state
contains three different operational profiles (charge, discharge and model parameters can be jointly estimated by particle
and impedance). These environmental/operational conditions filter. Similar ideas can also be found in [44]–[46]. However,
affect the degradation processes of systems and lead to the the particle filter is a Monte-Carlo-based approach, and thus
varying degradation rate of systems, i.e. nonlinear stochastic only the numerical result of the estimated RUL distribution
degradation. can be obtained and an explicit PDF estimate of the RUL is
As for nonlinear stochastic deteriorating systems, a nonlinear difficult to achieve. In the context of PHM, it is necessary and
stochastic degradation model for RUL estimation was presented valuable to derive the PDF of the RUL with an explicit form
in [26] by introducing a time-varying degradation rate function. so as to provide real-time RUL estimation for the subsequent
In [26], the RUL distribution was derived under the assumption maintenance scheduling.
that if the degradation process hits the threshold at a certain The purpose of this paper is to develop a general nonlinear
time t, then the probability that such a process crossed the degradation process model that can generate an explicit RUL
threshold level before time t can be negligible. However, distribution and adapt to the history of observed degradation
only the current degradation observation is considered in the data.In particular, the goal is to shed light on three fundamen-
estimated RUL and the parameters in the model cannot be tal issues frequently encountered in soft failure prognosis:(i)
adaptively updated via newly monitored degradation data. There nonlinear degradation modeling without requiring data trans-
are reported models utilizing the degradation data to date for formation and time-to-event data, (ii) obtaining and adaptively
evaluating and updating the RUL [29], [34], [37]. Using the updating the explicit expression of the RUL distribution derived
history of observed degradation, degradation models based on from the degradation process with newly observed data, and (iii)
linear Wiener processes were proposed by [37] and [34] with an updating estimated parameters of the degradation model.
adapted drift to estimate the RUL. The work [29] presented an To address the above issues, an adaptive and nonlinear
exponential-like degradation model with a Brownian error for prognostic model is presented for RUL estimation, in which
RUL estimation, where the historical data to date of individual the dynamics and nonlinearity of the degradation process are
equipment are incorporated by Bayesian mechanism. Following modeled by a time-dependent drift coefficient. This degradation
[29], many variants have been reported (see a review in [38]). model can cover conventionally linear models. In order to make
However, such models are limited to those which can transform the RUL estimation depending on the history of the observa-
the exponential path into a linear path. The work in [26] was tions, a state space model is constructed and Bayesian filtering
extended by [39], but the result for RUL estimation was limited is applied to update one parameter in the drifting function
to the nonlinear model with the power function. Recently, through treating this parameter as a hidden state variable. The
Zhou et al. in [40] proposed a promising prognostic framework PDF of the RUL is derived with an explicit form and some
for individual units subject to hard failure, based on joint commonly used linear model-based methods can be shown to
modeling of degradation signals and time-to-event data. In this be special cases of the proposed model. In addition, the EM
work, the degradation signals are modeled using a mixed-effects algorithm in conjunction with the Kalman filter is utilized to
regression model and time-to-event data are modeled using the update the drifting parameter and other parameters in the state

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 3

space model simultaneously and recursively. Finally, several Under model expressed by (1), and based on the concept of
numerical examples and an application to lithium-ion batteries the FHT, the FHT T of {X(t), t ≥ 0} crossing threshold w can
data are provided to validate the proposed approach. be defined as
The remainder parts are organized as follows. Section II
T = inf {t : X(t) ≥ w| X(0) < w} . (2)
develops an adaptive and nonlinear prognostic model for RUL
estimation. In Section III, the parameter estimation/updating As a result, the key for estimating the RUL is to derive the
approach is presented. An illustrative example to the power PDF of lifetime T , denoted as f T |ϕ ( t| ϕ).
model is given in Section IV. In Section V, numerical examples
and an application are provided to demonstrate the developed
B. Derivation of the RUL distribution
method. Section VI concludes the paper.
From [26], the approximated distribution of T can be ob-
tained by the following lemma.
II. N ONLINEAR MODEL DESCRIPTION AND RUL
Lemma 1: For the degradation process {X(t), t ≥ 0} given
ESTIMATION
by (1), if µ(t; ϑ) is a continuous function of time t in [0, ∞),
A. Modeling description then the PDF of the FHT T of X(t) defined by (2) can be
Let {X(t), t ≥ 0} denote the stochastic process describing approximated with an explicit form as follows:
the progression of degradation over operating time t when ( ) [ 2
]
∼ 1 SB (t) λ SB (t)
starting from a known initial state X(0) = x0 . In order to model f T |ϕ ( t| ϕ) = √ + µ(t; ϑ) exp − ,
2πt t σB 2t
the nonlinear stochastic process of systems such as lithium-ion (3)
∫t
batteries, the degradation at time t is denoted as follows where SB (t) = (w − λ · 0 µ(τ ; ϑ)dτ )/σB .
∫ t Lemma 1 is obtained in the context of using a Wiener
X(t) = x0 + λ · µ(τ ; ϑ)dτ + σB B(t), (1) process for degradation modeling under an assumption that, if
0
the degradation process is hitting the threshold w at a certain
where the degradation process X(t) is driven by a standard time t exactly, then the probability that such a process crossed
Brownian motion (BM) B(t)with a nonlinear drift λ · µ(t; ϑ). the threshold level before time t is negligible. This assumption
In (1), λ · µ(t; ϑ) and σB are the drift and diffusion coeffi- has been well demonstrated by recent prognostics studies in
cients, respectively; µ(t; ϑ) is a nonlinear function over t with [26], [47]–[49]. From Lemma 1, it can be easily found that
unknown parameter vector ϑ, which can be used to characterize most of the current models based on the Wiener process with a
time-variable and nonlinear performance of systems such as linear drift can be covered by the proposed model. That is to say,
lithium-ion batteries and electromechanical systems; λ is a f T |ϕ ( t| ϕ) can be reduced to the inverse Gaussian distribution
proportional parameter controlling the speed of the nonlinear provided that λ · µ(t; ϑ) is a constant.
degradation while ϑ is used to determine the shape of the degra- However, in (3), the estimated PDF of the FHT does not
dation progression. Without loss of generality, it is assumed consider the real-time observations of the degradation process.
X(0) = x0 = 0. Further, let ϕ = [λ, ϑ, σB ] denote the model Considering the potential for updating the knowledge of the
parameters. process when new degradation observation X(ti ) = xi become
The motivation of using degradation model (1) is twofold. available, for t ≥ ti , the degradation process over time since ti
First, the engineering systems possibly experiences different can be revised as
operating conditions such as charge, discharge and impedance
conditions for lithium-ion batteries, and thus the degradation (∫ t ∫ ti )
paths of systems exhibit varying degradation rates. Second, X(t) = xi + λ µ(τ ; ϑ)dτ − λ · µ(τ ; ϑ)dτ +
0 0
the model (1) uses a relatively general Wiener process with
a nonlinear drift part. It is general since it can cover linear σB (B(t) − B(ti ))
∫ t
models used in the literature and include a variety of nonlinear
=xi + λ · µ(τ ; ϑ)dτ + σB B(t − ti ) (4)
paths through selecting the forms of µ(t; ϑ). For example, if ti
µ(t; ϑ) is a constant, (1) is reduced to the Wiener process with a
for t ≥ ti .
constant drift, which has been widely used to model degradation
In such case, the residual t − ti corresponds to the realization
processes with linear paths [16], [24], [34], [37].
of the RUL at time ti if t is the FHT of {X(t), t ≥ ti }. Given
Now, we illustrate the main modeling principle of estimating
X(ti ) = xi , the RUL at ti can be defined as
the RUL based on (1). From [31] and [32], the concept of the
FHT is used to define the lifetime and then infer the RUL. In Li = inf { li : X(ti + li ) ≥ w| X(ti ) = xi , ϕ} , (5)
other words, when the degradation X(t) modeled as (1) reaches
with the PDF fLi |xi ,ϕ (li |xi , ϕ).
a pre-set critical level w, the system can be declared to be
Taking the transformation li = t − ti with li ≥ 0 for (4), the
failed and thus there has no useful lifetime left. Therefore, it is
process {X(t), t ≥ ti } can be transformed with time scale over
natural to view the event of lifetime termination as the point that
the residual time li , i.e. RUL, as
X(t) exceeds the threshold level wfor the first time. For critical
∫ li +ti
equipments, it is usually mandatory for putting this into practice
and once the observed degradation is equal or above the set X(li +ti ) = xi +λ· µ(τ ; ϑ)dτ +σB (B(li + ti ) − B(ti )) .
ti
threshold level that the system must be stopped for inspecting. (6)

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 4

As a result, the RUL at time ti is equal to the FHT of the as a time-dependent variable, conditional on the observed data
process {Y (li ), li ≥ 0} crossing threshold wi = w − xi , where up to ti . The degradation equation can be reconstructed with a
Y (li ) = X(li + ti ) − xi and Y (0) = 0. That is to say, given ti , state-space model as
∫ li +ti {
λi = λi−1 + η
Y (li ) = λ · µ(τ ; ϑ)dτ + σB W (li ), (7) xi = xi−1 + λi−1 Ωi (ϑ) + σB εi
, (10)
ti
∫t
where W (li ) = B(li + ti ) − B(ti ). where Ωi (ϑ) = h(ti ; ϑ) − h(ti−1 ; ϑ), h(ti ; ϑ) = 0 i µ(τ ; ϑ)dτ ,
In order to derive fLi |xi ,ϕ (li |xi , ϕ), it is necessary to prove the error term in the state equation is distributed as η ∼
that the stochastic process, {W (li ), li ≥ 0}, with W (li ) = N (0, Q), and εi = [B(tk ) − B(tk−1 )] ∼ N (0, ti − ti−1 ). The
B(li + ti ) − B(ti ) is still a BM. This is guaranteed by the use of ti − ti−1 as the variance of εi is required by the property
following lemma. of BM. Here it is assumed that the initial drift λ0 follows a
Lemma 2: Given ti , the stochastic process, {W (t), t ≥ 0}, normal distribution with mean µ0 and variance P0 . As such,
with W (t) = B(t + ti ) − B(ti ) for any t ≥ 0 is still a standard the drift parameter is considered as a hidden “state” and can
BM, where {B(t), t ≥ 0} is a standard BM. only be estimated from the historical information to date, X0:i .
The proof of Lemma 2 can be easily achieved by checking In the first equation of (10), λi follows a Gaussian distribution
the properties of the standard BM. As such, the estimated which can be estimated by a recursive filter based on X0:i . In
PDF of the RUL, fLi |xi ,ϕ (li |xi , ϕ), conditional on the current the following, denote its mean as λ̂i = E( λi | X0:i ) and its
observation and ϕ can be obtained by the following theorem. variance as Pi|i = var( λi | X0:i ). In the framework of Bayesian
Theorem 1: Under the same conditions as Lemma 1, the PDF filtering, λ̂i and Pi|i can be easily obtained by Kalman filtering
of the RUL can be formulated at time ti with the available in a recursive manner, as summarized below.
current degradation measurement xi as Algorithm 1 (Kalman filtering algorithm for estimating λi ):
wi − λ (υ(li ) − li µ(li + ti ; ϑ))
fLi |xi ,ϕ (li |xi , ϕ) ∼
= √ × Step 1: Initialize µ0 , P0 .
σB 2πli3
[ ] Step 2: State estimation at time ti
2
(wi − λυ(li )) λ̂i|i−1 = λ̂i−1
exp − 2 , (8)
2σB li Pi|i−1 = Pi−1|i−1 + Q
∫ l +t Ki = Pi|i−1 Ωi (ϑ)(Ω 2 2
− ti−1 ))−1 .
( i (ϑ) Pi|i−1 + σB (ti )
with υ(li ) = tii i µ(τ ; ϑ)dτ and wi = w − xi . λ̂i = λ̂i−1 + Ki xi − xi−1 − λ̂i−1 Ωi (ϑ)
Proof: Based on Lemma 1, (6) and (7), it is known that
Step 3: Updating variance Pi|i = Pi|i−1 − Ωi (ϑ)Ki Pi|i−1 .
{Y (li ), li ≥ 0} is still a nonlinear degradation process satisfying
Based on (10) and the Gaussian nature of the Kalman filter,
(1) with a drift part λ · µ(li + ti ; ϑ) and initial value Y (0) = 0.
the PDF of λi conditional on X0:i is still Gaussian with
Therefore, the RUL estimation at ti can be calculated as the [ ]
FHT of {Y (li ), li ≥ 0} crossing wi . In addition, it is easy to 1 (λi − λ̂i )2
verify that {Y (li ), li ≥ 0} satisfies[ all the conditions of Lemma p( λi | X0:i ) = √ exp − , (11)
∫ l +t ] 2πPi|i 2Pi|i
1; so it is obtained SB (li ) = wi − tii i µ(τ ; ϑ)dτ /σB .
From (3), the proof of Theorem 1 is completed. where the dependence between λi and X0:i is contained in λ̂i
and Pi|i .
Remark 1: In above updating mechanism, we only make
C. Adaptive RUL estimation
λ adaptive with the observed data to date. For a more re-
It is noted that (11) only uses the current monitoring value alistic implementation, making both λ and ϑ adapted with
xi when making extrapolations about the future trend of the the history may be more appropriate, but this consideration
degradation process. Now, consider making the estimated RUL will require more complicated estimating procedure such as
depend on X0:i = {x0 , x1 , ..., xi }, which is the history of extended Kalman filter or particle filter. This consideration
degradation observations for the system up to ti . Toward this deserves future research.
end, from the FHT concept, the definition of the RUL Li at ti From the state equation in (10), it is known that λi is random
in (7) can be revised as and characterized by the PDF p( λi | X0:i ), as denoted by (11).
Li = inf { li : X(ti + li ) ≥ w| X0:i } , (9) Consequently, to derive f Li |X0:i ( li | X0:i ), it is obtained
∫
with the PDF f Li |X0:i ( li | X0:i ). f Li |X0:i ( li | X0:i ) = fLi |X0:i ,λi (li |X0:i , λi )p( λi | X0:i )dλi ,
To incorporate the history of the observations, the concept of
(12)
combined parameter and state estimation, suggested in [50], has
where fLi |X0:i ,λi (li |X0:i , λi ) denotes the estimated RUL condi-
been frequently adopted in prognostic studies such as [43]–[46].
tional on λi and X0:i . It is worth noting that other parameters
This paper also utilizes this concept to achieve joint state and
are omitted in the conditional part of fLi |X0:i ,λi (li |X0:i , λi )
parameter estimation and to make the history of the observations
for notation simplicity since they are not random. In order to
be incorporated into the estimation. Specifically, an updating
calculate fLi |X0:i ,λi (li |X0:i , λi ), the following result is given.
procedure for the parameter λ in the drifting part λ · µ(t; ϑ)
Theorem 2: Once X0:i is available at ti , the following holds,
is introduced by a random walk model λi = λi−1 + η over
time, where η ∼ N (0, Q). In this case, the parameter λ evolves fLi |X0:i ,λi (li |X0:i , λi ) = fLi |X(ti )=xi ,λi (li |λi , xi ). (13)

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 5

Proof: Due to the Markov property of the standard BM, it From (16), once a new degradation observation is available
is obtained and parameter λ is updated, the RUL of this monitored system
d can be adaptively updated. Comparing (16) with (8), it is found
fLi |X0:i ,λi (li |X0:i , λi ) = Pr(Li ≤ li |X0:i , λi ) that the uncertainty in estimating λi , i.e. Pi|i , propagates into
dli
d the PDF of the RUL f Li |X0:i ( li | X0:i ) and Pi|i can be updated
= Pr(inf { li : X(li + ti ) ≥ w| X0:i } ≤ li |X0:i , λi ) by Kalman filter as well.
dli
d For the model (16) to be used in real time estimation, several
= Pr(inf { li : X(li + ti ) ≥ w| X0:i } ≤ li |X(ti ) = xi , λi ) unknown parameters, including σB 2
, ϑ, µ0 , P0 and Q, need to
dli
be estimated. The selection of these parameters will surely
= fLi |X(ti )=xi ,λi (li |λi , xi ). (14)
affect the performance of the RUL estimation. For instance, the
This completes the proof. selected variance Q of the random walk noise determines both
In order to calculate f Li |X0:i ( li | X0:i ), it is required to the rate of the convergence of λ and the estimation performance
compute the integration of (12). To avoid the lengthy derivation, once convergence is achieved. A large Q will yield quick
the following lemma is directly given [16]. convergence but tracking with too wide a variance, whereas too
Lemma 3: If Z ∼ N (µ, σ 2 ) and w, A, B ∈ R, C ∈ R+ , then small a Q may yield a very slow convergence. Therefore, it is
the following holds: desirable to estimate these parameters to obtain the best possible
[ ( )] performance. The mechanism to estimate these parameters will
(B − Z)2 be discussed in the next section.
EZ (A − Z) · exp −
2C
√ ( ) ( )
C 2
σ B + µC (B − µ)2 III. A DAPTIVE PARAMETER ESTIMATION
= A− × exp − .
σ2 + C σ2 + C 2 (σ 2 + C) [ 2 θ is used to] denote the unknown parameter
In this section,
(15) vector, θ = σB , ϑ, µ0 , P0 , Q , and Υi = {λ0 , λ1 , . . . , λi } is
From (12) and Theorem 2, the following conclusion is the history of the proportional parameter up to ti . In order to
obtained to derive the PDF of the RUL conditional on X0:i , achieve simultaneous estimation of the proportional parameter λ
i.e. f Li |X0:i ( li | X0:i ). and θ, this paper considers to calculate the maximum likelihood
Theorem 3: As for the degradation process defined in (1), estimation (MLE) of θ once the new observation xi is available.
given X0:i , the PDF of the RUL conditional on X0:i can be In this case, the log-likelihood function for the degradation data
formulated as X0:i can be written as
∑i
wi Λ(li ) − αi (li ; ϑ)∆(li ) ι(θ) = log[p(X0:i |θ)] = log[p(xj |X0:j−1 , θ)]. (19)
f Li |X0:i ( li | X0:i ) ∼
= √ × j=1
2πli2 Λ3 (li ) [ 2 ]
 ( )2  In the following, θ̂i = σB,i , ϑi , µ0,i , P0,i , Qi is used to
 w i − λ̂ i υ(l i )  represent the MLE of θ conditional on X0:i at ti . Under model
exp −  (16) (10), p( X0:i | θ) is induced by the missing data, since the data
2Λ(li )
Υi about parameter λ is used as the state in state-space model
(10). In general, it is difficult to write the function p( X1:i | θ)
∫where αi (li ; ϑ) = υ(li ) − li µ(li + ti ; ϑ), υ(li ) =
li +ti with an explicit form when there exists a hidden variable.
µ(τ ; ϑ)dτ , Λ(li ) = Pi|i υ(li )2 + σB 2
li , and ∆(li ) =
ti As such, it is natural to consider the effect of missing data
2
Pi|i υ(li )wi + λ̂i σB li . Υi to search for the optimum of the log-likelihood function
Proof: In order to complete the proof, of the parameters. The EM algorithm [51] provides a natural
fLi |X0:i ,λi (li |X0:i , λi ) should be first derived from (13). framework for approaching a parameter estimation problem
As a result of Theorem 1, fLi |X0:i ,λi (li |X0:i , λi ) can be involved hidden variables.
derived from (8) through replacing λ with λi . From (11)-(13), In the EM algorithm, by exploiting the relationship between
it is obtained p(X0:i |θ )and p( X0:i , Υi | θ), it is possible to generate a se-
∫
f Li |X0:i ( li |[X0:i ) = fLi |X(ti )=xi ,λ]i (li |λi , xi )p( λi | X0:i )dλi quence of parameter estimates that converge to the MLE of
= E λi |X0:i fLi |X(ti[)=xi ,λi (li |λi , xi ) [ ]]
the parameters. The complete-data log-likelihood is defined
= √ 3 E λi |X0:i (wi − λi αi (li ; ϑ)) exp −
1 (wi −λi υ(li ))2 as ℓ(θ) = log p( X0:i , Υi | θ). Then the estimation procedure
2 l
2σB
σB 2πli [ [ ( )2
i
]] consists of two steps: E-step and M-step,
( ) wi
−λi • E-step: Calculate
= √ αi (li ;ϑ) wi
− λ exp −
υ(li )
E
λi |X0:i αi (li ;ϑ) i 2 l /(υ 2 (l )
2σB
σB 2πli3 i i
(k)
(17) ℓ( θ| θ̂i ) = E Υi |X0:i ,θ̂(k) {log p( X0:i , Υi | θ)} , (20)
i

where αi (li ; ϑ) = υ(li ) − li µ(li + ti ; ϑ). [ ]

(k) 2(k) (k) (k) (k) (k)
Therefore, setting the followings and applying Lemma 3 will where θ̂i = σB,i , ai , µ0,i , P0,i , Qi denotes the
complete the proof after necessary simplification. parameter estimation in the kth iteration conditional on
X0:i .
wi wi σ 2 li • M-step: Calculate
A= ,B = , C = 2B . (18)
αi (li ; ϑ) υ(li ) υ (li ) (k+1) (k)
θ̂i = arg max ℓ( θ| θ̂i ). (21)
θ

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 6

To be more precise, in the E-step, to obtain the expected E Υi |X0:i ,θ̂(k) (λj λj−1 ) can be calculated as follows
(k) i
value of ℓ(θ) with respect to the kth iteration of θ̂i and the
observed data X0:i , there is E Υi |X0:i ,θ̂(k) (λj ) = λ̂j|i ,
i

(k)
E Υi |X0:i ,θ̂(k) (λ2j ) = λ̂2j|i + Pj|i ,
ℓ( θ| θ̂i ) = E Υi |X0:i ,θ̂(k) {log p( X0:i , Υi | θ)} i .
i E Υi |X0:i ,θ̂(k) (λj λj−1 ) = P j|j Sj−1 +
= E Υi |X0:i ,θ̂(k) {log p( X0:i | Υi , θ)} + E Υi |X0:i ,θ̂(k) {log p( Υi | θ)} .
i

i i
Sj (Mj+1|i − P j|j )Sj−1 + λ̂j|i λ̂j−1|i
(22) (26)
(k)
Together with (25) and (26), ℓ( θ| θ̂i ) can be written as
Following (13), the full likelihood function can be formulated (k)
as 2ℓ( θ| θ̂i ) = E Υi |X0:i ,θ̂(k) [2ℓi (θ)]
[ i
∑i ( (λj −λj−1 )2 )
= E Υi |X0:i ,θ̂(k) − log P0 − (λ0 −µ
2
0)
− j=1
( i (
P 0
∫ tj )2 )]
Q

log ℓ(θ) = log p( X0:i | Υi , θ) + log p( Υi | θ) = log p( λ0 | θ)+ ∑ i x j −x j−1 −λ j−1 t µ(τ ;ϑ)dτ
− j=1 log σB 2
+ j−1
− i log Q
∏i ∏i 2 (t −t
σB j j−1 )

log p( λj | λj−1 , θ) + log p( xj | λj−1 , θ) (23) C0|i −2λ̂0|i µ0 +µ20

j=1 j=1
= − log P − −
∑i ( )
0 P0
Cj|i −2Cj,j−1|i +Cj−1|i
From state space model (10), it is ob- j=1 log Q + Q −
tained λj | λj−1 ∼ N (λj−1 , Q), x)j | λj−1 ∼ ( ∫ tj )
( ∫ tj ∑ i 2
(x j −x j−1 ) 2
+( tj−1 µ(τ ;ϑ) dτ )2 Cj−1|i
N xj−1 + λj−1 tj−1 µ(τ ; ϑ)dτ , σB (tj − tj−1 ) 2
and j=1 log σ B + σB (tj −tj−1 )
2 +
( ∫ tj )
λ0 ∼ N (µ0 , P0 ). Using (23) and ignoring the constant ∑i 2λ̂j−1|i (xj −xj−1 )· t
j−1
µ(τ ;ϑ)dτ
terms, the complete likelihood function can be written as j=1 σB2 (t −t
j j−1 )

∑ ( ) (27)
(λj −λj−1 ) 2
2 log ℓ(θ) = − log P0 − (λ0 −µ
2
0) i
− log Q + After obtaining ℓ( θ| θ̂
(k)
), the unknown parameters θ̂
(k+1)
( P0
(
j=1
∫t )2 )
Q i
(k)
i
∑i xj −xj−1 −λj−1 t j µ(τ ;ϑ)dτ can be obtained by maximizing ℓ( θ| θ̂i ) with respect to θ.
− j=1 2 log σB + 2 (t −t
σB j
j−1
j−1 ) However, it is difficult to obtain the explicit form of θ̂i
(k+1)

(24) through maximizing (27) due to the effect of the nonlinear part
The conditional expectation, ℓ( θ| θ̂i ), given in (20) can be involved
(k) by the parameter ϑ. In order to tackle this problem,
(k+1)
written as θ̂i is calculated through the technique of a profile log-
(k)
likelihood function as follows.
ℓ( θ| θ̂i ) = E (k) [2 log ℓ(θ)]
(k+1)
Υi |X0:i ,θ̂i Firstly, given ϑ, the results of the estimated parameter θ̂i
[ ∑i ( )
(λ0 − µ0 )2 (λj − λj−1 )2in the (k+1)th step can be summarized in the following.
=E (k) − log P0 − − log Q +
Υi |X0:i ,θ̂i P0 j=1 Q (k+1)
 ( ∫ tj )2  µ0,i = λ̂0|i ,
∑i xj − xj−1 − λj−1 tj−1 µ(τ ; ϑ)dτ (k+1)
−

2 log σB +

 .
P0,i = C0|i − λ̂20|i = P0|i ,
j=1 σB2 (t − t
j−1 ) (k+1) ∑i
j
Qi = 1i j=1 (Cj|i − 2Cj,j−1|i + Cj−1|i ),
( ∫t )
(25)
2(k+1) ∑i (xj −xj−1 )2 +( t j µ(τ ;ϑ)dτ )2 Cj−1|i
σB,i (ϑ) = 1i j=1 j−1
tj −tj−1 −
Clearly, to derive the expectation implies to ( ∫ tj )
work out E Υi |X0:i ,θ̂(k) (λj ), E Υi |X0:i ,θ̂(k) (λ2j ) and 1
∑i 2λ̂j−1|i (xj −xj−1 ) t
j−1
µ(τ ;ϑ)dτ
i i i j=1 tj −tj−1 .
E Υi |X0:i ,θ̂(k) (λj λj−1 ), which are the conditional expectation
i (28)
given X0:i . In this paper, Rauch-Tung-Striebel (RTS) smoother
with Cj|i = E Υi |X0:i ,θ̂(k) (λ2j ), Cj,j−1|i =
is used to provide an optimal estimate of the above conditional i
2(k+1)
expectations [52]. The smoothing algorithm is summarized as E Υi |X0:i ,θ̂(k) (λj λj−1 ). It is noted here that σB,i (ϑ)
i
follows, where M j|i = cov( λj , λj−1 | X0:i ). (k+1)
is a function of ϑ, while µ0i , P0i
(k+1) (k+1)
, Qi are free of
Algorithm 2 (RTS smoothing algorithm): (k)
the effect of ϑ provided that θ̂i is given.
Step 1: Forwards iteration by Kalman filter algorithm Secondly, substituting (27) into (28) gives the profile log-
Step 2: Backwards iteration likelihood function as
−1
Sj = P j|j Pj+1|j (k)
2ℓ( ϑ| θ̂(i ) = − log P0|i − 1− )
λ̂j|i = λ̂j + Sj (λ̂j+1|i − λ̂j+1|j ) = λ̂j + Sj (λ̂j+1|i − λ̂j ) ∑i (k+1) Cj|i −2Cj,j−1|i +Cj−1|i
Pj|i = Pj|j + Sj2 (Pj+1|i − Pj+1|j ) j=1 log Qi + (k+1)
Qi
Step 3: Initialize
( ) ( ∫t )
∫ tj ∑i 2(k+1) (xj −xj−1 )2 +( t j µ(τ ;ϑ)dτ )2 Cj−1|i .
Mi|i = 1 − Ki tj−1 µ(τ ; ϑ)dτ P i−1|i−1 − j=1 log σB,i (ϑ) + 2(k+1)
j−1
σB,i (ϑ)(tj −tj−1 )
Step 4: Backwards iteration for smoothing covariance ( ∫t )
∑i 2λ̂j−1|i (xj −xj−1 ) t j µ(τ ;ϑ)dτ
Mj|i = P j|j Sj−1 + Sj (Mj+1|i − P j|j )Sj−1
j−1
+ j=1 2(k+1)
σB,i (ϑ)(tj −tj−1 )
(29)
From the RTS smoothing algorithm, the condition- (k)
As such, the estimate of ϑ, i.e. ϑi , can be obtained by
al expectations, E Υi |X0:i ,θ̂(k) (λj ), E Υi |X0:i ,θ̂(k) (λ2j ), and maximizing the profile log-likelihood function in (29) through
i i

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 7

a search algorithm. In this paper, “fminsearch” function in where ˜li is the actual RUL obtained at ti and f Li |X0:i ( li | X0:i )
MATLAB is utilized to optimize (29), which is relatively easy is the according conditional PDF of the RUL estimated by (16).
(k)
to perform. Now, substituting ϑi into (28), the estimates of The total mean-squared error (TMSE) is the sum of the MSE
2(k+1) (k) (k+1) (k+1) (k+1) at each CM point over the whole life cycle. If there are N
σB,i (ϑi ), µ0,i , P0,i , Qi are obtained, respective-
ly. In contrast with directly maximizing (27) to obtain θ̂i
(k+1)
, observations, the TMSE of the RUL estimation based on these
the number of the unknown parameters in the search algorithm observations can be calculated as
is reduced by adopting the above profile log-likelihood function ∑
N

technique. Then, the algorithm iterates the E-step and M-step TMSE = MSEi . (34)
i=1
until a criterion of convergence is satisfied.
Due to the above definition and the effect of the integral
operator, MSEi is actually a metric to characterize the ac-
IV. A N ILLUSTRATIVE EXAMPLE
cumulated estimation error between the actual RUL and the
For simplicity and an illustration purpose, this section consid- estimated RUL described by a probabilistic distribution. This
ers the following power model to illustrate the implementation metric will be used in the following numerical example for
of the presented approach, model comparisons.
X(t) = λta + σB B(t). (30)
A. Numerical example
It is noted that (30) is a nonlinear model fallen into the The purpose of this numerical example is twofold. The first
type defined in (1) with µ(t; ϑ) = ata−1 and ϕ = [ϑ, σB ] = is to illustrate the implementation of the proposed method by
2
[λ, a, σB ]. This kind of power model has been found in many comparing the estimation results with existing models. The
applications such as modeling crack length and battery data second is to verify the usefulness of the proposed parameters
[53]–[55]. Of course, different forms of µ(t; ϑ) could be used, estimation algorithm. Specifically, a nonlinear Wiener process
if it is suggested by the specific application of interest. as X(t) = λta + σB B(t) is directly considered to generate the
By Lemma 1, the PDF of the FHT corresponding to (30) can simulation data by the so-called Euler approximation. Hence
be formulated directly as the simulated data are √generated by X(i+1)∆t = Xi∆t +
w − λt a
(1 − a) [ / ] λa (i∆t)
a−1
∆t + σB Y ∆t where Y ∼ N (0, 1) and ∆t is
f T |ϕ ( t| ϕ) ∼
2
= √ exp −(w − λta ) 2σB 2
t . the discretization 2
step. Here, set λ = 0.05, a = 2, σB = 0.16
σB 2πt 3
(31) and ∆t = 0.1. When the threshold is w = 10.36, one simulated
Once the history of degradation observations X0:i up to ti path is obtained as shown in Figure 1, where the FHT can be
is available, the PDF of the estimated RUL can be formulated approximated as 14.8 after 148 sampling.
by Theorem 3 as
 ( )2 
12

The actual sampling path

w Λ(l ) − ∆(l )  wi − λ̂i υ(li )  The predicted path
f Li |X0:i ( li | X0:i ) ∼
10
i i i
= √ 2 3 exp −
The degradation path

,
2πli Λ (li ) 2Λ(l i ) 8

(32) 6

with αi (li ; ϑ) = υ(lti ) − ali (li + ti )a−1 , υ(li ) = (li + ti )a − tai ,

Λ(li ) = Pi|i υ(li )2 + σB 2
li , ∆(li ) = Pi|i υ(li )wi + λ̂i σB 2
li and 4

wi = w − xi . 2

It is noted that Pi|i and λ̂i can be obtained via Algorithm 1

by setting h(ti ; ϑ) = tai . Accordingly, the parameter estimation 0
0 5 10 15
The sampling point
algorithm in Section III can be applied to the model (30) so as
to implement adaptive RUL estimation. Fig. 1. Comparison of the prediction with the actual sampling path

V. N UMERICAL EXAMPLE AND CASE STUDY It can be found that there is no simple method which
can transform the above nonlinear simulation model into a
In this section, a numerical simulation and a case study
linear process under the same time scale. In order to show
for battery data are provided to compare the performance of
the superiority of the proposed approach, the following four
the presented method with some existing models. Here a loss
competing models, which have been widely used in degradation
function is employed to enable a direct comparison with any
modeling fields, are considered for a comparative purpose.
alternative modeling approach that produces a conditional PDF
(i) Model 1. It is a Wiener process with a linear drift such
for the underlying RUL estimation. The loss function is the
as X(t) = λt + σB B(t). To achieve a fair comparison, all the
mean-squared error (MSE) about the actual RUL obtained at
sampling data are used to obtain the MLE of λ and σB , denoted
each observation point [16], defined as
by λ̂ and σ̂B . Once these parameters are obtained, they are fixed
∫∞ in the process of estimating the RUL distribution.
MSEi = (li − ˜li )2 f Li |X0:i ( li | X0:i ) dli , (33) (ii) Model 2. This model also uses the same process as Model
0 1. However, at each sampling point all the sampling data to date

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 8

are used to compute the MLE λ̂ and σ̂B , and then the estimated the estimated λ̂i based on all sampling points is 0.049 in
RUL distribution at this sampling point is obtained. Namely, λ̂ the simulation, which is also close to the true value 0.05.
and σ̂B are updated in a sequential fashion. These results verify the usefulness of the developed parameters
(iii) Model 3. Based on the same process as Model 1 and estimation method. For the results for RUL estimation, the
Model 2, the method developed in [37] is adopted to estimate comparisons with Models 1-4 are shown in Figure 3 from the
λ and σB , in which a Kalman filter and the EM algorithm are 144th sampling point to the 147th point.
used together.
(iv) Model 4. Here, the presented model and parameters The proposed method
2
The proposed method
2
Model 1
estimation method developed in this paper are used, but the Model 2

PDFs of the RUL

PDFs of the RUL
1.5
1.5

RUL distribution is calculated by (8), where the unknown 1

1

0.5 0.5

parameters, λ and σB , are replaced by λ̂i and the estimate of 0 0

σB in Section III. In this model, the PDF of the RUL is not 14.6

14.5 2
3
4
5 14.6

14.5 2
3
4
5

1 1

fully dependent on the observation history to date since this The sampling point
14.4 0
The RUL The sampling point
14.4 0
The RUL

model only replaces λ with its mean λ̂i and does not consider The proposed method 2
The proposed method
2
Model 3 Model 4

PDFs of the RUL

its distribution. Thus, this model is different from the proposed

PDFs of the RUL

1.5
1.5

model which calculates the RUL distribution from (16). 1

1

0.5 0.5

In addition, the prediction equation for degradation in Models 0 0

1-3 is x̂i = xi−1 +λ̂·∆t, while the prediction equation for the p- 14.6

14.5 2
3
4
5 14.6

14.5 2
3
4
5

a−1 1 1

resented approach and Model 4 is x̂i = xi−1 + λ̂i a (i∆t) ∆t The sampling point
14.4 0
The RUL The sampling point
14.4 0
The RUL

where ∆t is the sampling interval and a is estimated by the

Fig. 3. Comparative results of the proposed method with Models 1-4 in RUL
method developed in Section III. estimation (the asterisk denotes the end of life)
The one step prediction of the degradation using the proposed
method is illustrated by Figures 1. It can be seen that the It can be seen from Figure 3 that the proposed method has
proposed method can recover the degradation path very well and obvious advantage over linear Wiener process based Models 1-
the MSE between the predictions and the actual data is 0.0113. 3 and Model 4 that does not consider the observation history.
In contrast, the MSEs using Models 1-3 are 0.0179, 0.0175, In particular, the estimated PDF of the RUL from the presented
0.0158, respectively. This shows that the presented method method is much tighter than other competing models. This is the
produces better prediction accuracy for degradation data. The desired result since all the data have been used for estimating
evolving path of the estimated parameter vector θ, consisting the RUL in the proposed approach and all model parameters
2
of σB , a, µ0 , P0 , Q, is illustrated in Figure 2. can be updated once new observation is available. Thus, the
uncertainty in the estimated RUL can be reduced naturally.
0.45 0.01

0.4 0.009 To further compare the performance of the proposed method

0.35 0.008

against Models 1-4, the MSE and TMSE about the RUL
Estimated µ0

Estimated P0

0.007
0.3
0.006
0.25

0.2
0.005 associated with these models are calculated. The evolving paths
0.004
0.15

0.1
0.003

0.002
of the MSE at each observation point for all methods are shown
0.05

0
0.001

0
in Figure 4. Note that only the results of the last 88 observation
points (i.e. from the 61th sampling point to the 148th point)
0 5 10 15 0 5 10 15
The sampling point The sampling point
0.5 0.25

0.45

0.2
are plotted since Figure 2 indicates that the changes of the
0.4

estimation will be smaller after the 60th observation point.

Estimated σB2

Estimated Q

0.35
0.15
0.3

0.25
0.1
45 90
0.2
The proposed method The proposed method
MSEi at each sampling point

MSEi at each sampling point

40 80
0.15
0.05 Model 1 Model 2
35 70
0.1

30 60
0 2 4 6 8 10 12 14 2 4 6 8 10 12 14
The sampling point The sampling point 25 50

20 40

15 30

2 10 20

10
Estimated a

1.5
0
6 7 8 9 10 11 12 13 14 15 6 7 8 9 10 11 12 13 14 15
The sampling point The sampling point
1 140
30

The proposed method The proposed method

MSEi at each sampling point

MSEi at each sampling point

0.5 25 Model 3 120 Model 4

100
20
0
2 4 6 8 10 12 14
80
The sampling point 15
60

10
40

Fig. 2. Estimated parameters 2 , Q, a

µ0 , P0 , σB with the sampling data 5
20

accumulated 0
6 7 8 9 10 11 12 13 14 15
0
6 7 8 9 10 11 12 13 14 15
The sampling point The sampling point

From Figure 2, it is observed that the parameter estimation

Fig. 4. Comparative results of the proposed method with Models 1-4 in MSE
method developed in Section III works well. In particular, the of the RUL
2
estimates of σB and a can respectively approach the true values
0.16 and 2 closely as the sampling data are accumulated. In From Figure 4, it is found that the proposed method maintains
addition, λ is made be adapted to the observed data and be the MSE of the RUL estimation in a relatively low level
estimated from the Kalman filter, but the sample mean of compared with linear Wiener process based Models 1-3 and

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 9

approaches a small MSE more quickly than Models 1-3. In air. In addition, it is observed that the degradation path of
addition, the proposed method avoids the abrupt change of battery exhibits a trend with variable degradation rate. Thus, a
MSE occurring in Model 4 since the presented method fully degradation model with a time-varying degradation rate could
uses the observation history. Furthermore, the TMSEs of the be suitable. All these observations encourage that the type of
presented method and Models 1-4 are 249.2, 1789.4, 2348.3, stochastic processes in (1), driven by BM with a nonlinear drift,
910.8, 513.4, respectively. Obviously, the presented method has is appropriate to characterize the path of the battery degradation
the least TMSE and thus has a better RUL estimation using the process. Therefore, the presented model (1) is used to analyze
measure MSE than the competing models. the batteries data in this paper.
For comparative purpose, this paper also considers the case
B. Lithium-ion battery life prognosis that the power law model is directly fitted to the battery data,
In this experiment, the presented method is applied for termed as Model 5. In Model 5, the parameters are estimated
RUL estimation of lithium-ion batteries to demonstrate the by a Bayesian method as [40] and the lifetime is estimated
effectiveness. The used degradation datasets are provided by by formulating the distribution of the degradation quantity as
the NASA Ames Prognostics Center of Excellence [56]. In the [29]. In order to verify the presented method, this paper first
used dataset, a set of four lithium-ion batteries (#5, #6, #7 and fits the four batteries data to the presented model and compare
#18) were run through 3 different operational profiles (charge, the fitting accuracy with Models 1-5 in Section V-A. Note
discharge and impedance) at room temperature. Charging was that the proposed method is derived based on the case that
carried out in a constant current mode at 1.5A until the battery the degradation process has an increasing trend. However, the
voltage reached 4.2V and then continued in a constant voltage capacity of the battery has a decreasing trend over time. Thus,
mode until the charge current dropped to 20mA. Discharge to apply the proposed method, the original data are transformed
was carried out at a constant current level of 2A until the by make the initial capacity minus all the capacity data of
battery voltage fell to 2.7V, 2.5V, 2.2V and 2.5V for batteries battery. Accordingly, the critical failure threshold should be
#5, #6, #7, and #18 respectively. Impedance measurement was changed as the initial capacity minus 1.4 Ahr. Based on the
carried out through an electrochemical impedance spectroscopy transformed batteries data, the predicted paths of the batteries
frequency sweep from 0.1Hz to 5kHz. Repeated charge and capacity by the proposed method can be obtained after the
discharge cycles result in accelerated aging of the batteries inverse transformation, which are illustrated in Figure 5. For
while impedance measurements provide insight into the internal battery #5, the MSEs between the predicted degradation and
battery parameters that change as aging progresses. Based on the actual degradation for the presented method and Models
the analysis of the available performance measures of the 1-5 are 1.03E-4, 1.92E-4, 1.65E-4, 1.62E-4, 1.53E-4, 1.58E-4,
Lithium-ion batteries, the capacity can be used to characterize respectively. This shows the better accuracy in the degradation
the long-term degradation process induced by the charge- prediction of the developed method. Accordingly, the estimated
discharge operational cycle [56]–[58]. The experiments were parameters evolving with the cycle for battery #5 are shown
stopped when the batteries reached end-of-life criteria, which in Figure 6. It is observed from Figure 6 that, as the battery
was a 30% fade in rated capacity (from 2 Ahr to 1.4 Ahr). The degradation data are accumulated, the model parameters will
capacity data of four batteries are illustrated in Figure 5. In this converge quickly. In addition, if there are some great changes
case, the lifetime of battery #5 is about 125 cycles when the in the degradation data, the estimated parameters can reflect
2
failure threshold is 1.4 Ahr. such data changes as expected (see σB ,Q,and a for examples).
These observations reflect the adaptive ability of the developed
2.1
B0005
parameters estimation algorithm.
2 predicted B0005
B0006 Based on the updated model parameters, the degradation
1.9
predicted B0006
signals of the battery capacity are further used for estimating
Battery capacity (Ahr)

1.8 B0007

1.7
predicted
B0018
B0007 the RUL of batteries. In this paper, the data of the battery #5
1.6
predicted B0018 are taken for example. For comparisons, the mediums of the
1.5 RUL, the associated 95% confidence intervals (CI) and relative
1.4 errors (RE) of the estimated RUL for #5 are computed at the
1.3
30th , 60th , and 90th percentiles of the lifetime, as summarized
1.2
in Table I. Here, the median is chosen as an approximate value
1.1
0 20 40 60 80 100
Cycle time (cycle)
120 140 160 180 for the point estimate of the RUL since the distribution of the
RUL is often highly skewed and thus it is reasonable to use the
Fig. 5. Degradation paths and predicted paths of four batteries
median as a measure of central tendency rather than the mean.
In addition, let ˜li denote the actual RUL at ti , and let ˆli be the
Figure 5 shows that the degradation signals of batteries
estimated RUL by its median ( at ti . Then,
) the RE for the RUL
experience a lot of fluctuations. In physics, Brownian motion at ti is defined as REi = ˜li − ˆli /˜li × 100. This measure
process aims at modelling the movement of small particles is used for model comparisons.
in fluids and air with tiny fluctuations. The tiny increase or The results in Table I indicate that all estimation results for
decrease in degradation over a small time interval behaves nonlinear models are superior to the results of linear models. In
similarly to the random walk of small particles in fluids and nonlinear models, Model 5 is a mixed-effect regression model

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 10

TABLE I
MRUL, 95%CI, RE OF THE ESTIMATED PARAMETERS

30th percentile 60th percentile 90th percentile

MRUL 95%CI RE MRUL 95%CI RE MRUL 95%CI RE
Model 1 68.8 (63.5,96.3) 21.4% 39.5 (37.2,56.1) 21.0% 14.2 (11.2,17.9) 13.6%
Model 2 67.1 (62.4,95.8) 23.3% 39.1 (35.3,57.4) 21.8% 14.5 (12.1,19.4) 16.0%
Model 3 69.4 (64.2,95.4) 20.7% 40.8 (37.8,55.3) 18.4% 13.9 (10.1,16.8) 11.2%
Model 4 72.5 (69.3,92.8) 17.1% 43.8 (40.2,53.5) 12.4% 11.4 (9.4,14.1) 8.8%
Model 5 70.2 (65.2,93.5) 19.9% 42.6 (38.1,54.2) 14.8% 11.2 (8.1,14.5) 10.4%
Proposed model 74.8 (71.1,91.5) 14.5% 45.8 (41.3,52.1) 8.4% 11.9 (10.2,13.6) 4.8%

−6
0.03 x 10
2

0.025
1.8

1.6
VI. C ONCLUSION
Estimated µ0

0.02 1.4
Estimated P0

0.015
1.2 In this paper, an adaptive and nonlinear prognostic model
1

0.01
0.8

0.6
is presented to estimate the RUL using the history of the
0.005
0.4

0.2
observed data to date. Three issues in RUL estimation are jointly
0
0 20 40 60 80
Cycle time (cycle)
100 120 140 160
0
0 20 40 60 80 100
Cycle time (cycle)
120 140 160 studied: (i) nonlinear degradation modeling without requir-
1

0.9
x 10
−3

1
x 10
−5

ing data transformation and time-to-event data, (ii) obtaining

0.9

0.8 0.8 and adaptively updating the explicit expression of the RUL
2

0.7 0.7
Estimated Q
Estimated σB

0.6

0.5
0.6

0.5
distribution derived from the degradation process with newly
0.4

0.3
0.4

0.3
observed data, and (iii) updating estimated parameters of the
0.2

0.1
0.2

0.1
degradation model. Specifically, a general stochastic process
0
0 20 40 60 80
Cycle time (cycle)
100 120 140 160
0
0 20 40 60 80 100
Cycle time (cycle)
120 140 160 with a nonlinear and time-dependent drift coefficient is adopted
1.4

1.2
to characterize the dynamics and nonlinearity of the degradation
1
process. This degradation model can cover the conventionally
Estimated a

0.8
linear model and thus answers issue (i). In order to make the
0.6

0.4
RUL estimation depending on the history of the observations,
0.2 a state space model is constructed to update one key parameter
0
0 20 40 60 80
Cycle time (cycle)
100 120 140 160
in the drifting function through appending this parameter as a
system state. The PDF of the RUL is derived with an explicit
Fig. 6. 2 , Q, a with the data of Battery #5
Estimated parameters µ0 , P0 , σB form accordingly and some commonly used linear model-based
methods turn out to be special cases of the proposed model.
This addresses issue (ii). For issue (iii), the EM algorithm in
conjunction with the Kalman filter is utilized to update the
drifting parameter and other parameters in the state space model
simultaneously and recursively once new observation becomes
while Model 4 and the presented model are stochastic process
available. This overcomes the disadvantage of other similar
based models. As expected, the results imply that stochastic
models which only estimated the model parameters once and
process based models are more superior in the estimation accu-
then fixed regardless subsequently newly available observations.
racy. In stochastic process based models, Model 4 calculates the
Finally, a numerical example and a case study for battery
PDF of the RUL through (8) and is not fully dependent on the
data are provided to demonstrate the proposed method. With
observation history to date since this model only replaces λ with
comparisons to four competing models that are mostly reported
its mean λ̂i and does not consider its distribution. Therefore,
in the literature, it is concluded that the presented method can
the accuracy of the RUL estimation by the presented method is
generate better results and incorporating the observation history
much improved as opposed to Models 1-5 and the superiority
can improve the accuracy of the RUL estimation.
of the presented method is validated.

Overall, the numerical example and case study effectively ACKNOWLEDGMENT

provides the supporting evidence of the presented modeling
The authors would like to thank the Associate Editor and four
development—nonlinear degradation modeling and RUL esti-
anonymous reviewers for their valuable suggestions and critical
mation with the observation history dependency. On one hand,
comments which have considerably helped in the revision of
the results demonstrate that the conventional linear Wiener
earlier version of the paper. The research reported here is
process-based model cannot give satisfactory results in the non-
partially supported by NSFC under grant numbers 61174030,
linear case even though the updating mechanism is incorporated
61473094 and 61374126.
in Model 3. But the proposed method works well. On the
other hand, the comparative results verify that incorporating
the observation history to date can improve the accuracy of the R EFERENCES
RUL estimation indeed. Both are evaluated by the measures of [1] M. Pecht, Prognostics and Health Management of Electronics. John Wiley,
MSE and TMSE. New Jersey, 2008.

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 11

[2] J. W. Sheppard, M. A. Kaufman, and T. J. Wilmering, “IEEE standards [26] X. S. Si, W. Wang, C. H. Hu, D. H. Zhou, and M. G. Pecht, “Remaining
for prognostics and health management,” IEEE Aerospace Electron. Syst. useful life estimation based on a nonlinear diffusion degradation process,”
Mag., vol. 22, no. 9, pp. 34-41, 2009. IEEE Trans. Rel., vol. 61, no. 1, pp. 50-67, 2012.
[3] P. Lall, R. Lowe, and K. Goebel, “Prognostics health management of [27] X. -X. Yuan and M. D. Pandey, “A nonlinear mixed-effects model for
electronic systems under mechanical shock and vibration using Kalman degradation data obtained from in-service inspections,” Rel. Eng. Syst.
filter models and metrics,” IEEE Trans. Ind. Electron., vol. 59, no. 11, pp. Safety, vol. 94, pp. 509-519, 2009.
4301–4314, 2012. [28] F. O. Eker, F. Camci, A. Guclu, H. Yilboga, M. Sevkli, and S. Sain, “A
[4] R. K. Singleton II, E. G. Strangas, and S. Aviyente, “Extended Kalman simple state-based prognostic model for railway turnout systems,” IEEE
Filtering for Remaining Useful Life Estimation of Bearings,” IEEE Trans. Trans. Ind. Electron., vol. 58, no. 5, pp. 1718–1726, 2011.
Ind. Electron., DOI: 10.1109/TIE.2014.2336616, 2014. [29] N. Gebraeel, M. A. Lawley, R. Li, and J. K. Ryan, “Residual-life
[5] E. R. Brown, N. N. McCollom, E. E. Moore, and A. Hess, “Prognostics distributions from component degradation signals: A Bayesian approach,”
and health management-A data-driven approach to supporting the F-35 IIE Trans., vol. 37, no. 5, pp. 543-557, 2005.
lightning II,” IEEE Aerospace Conf., 3-10 March, pp. 1-12, 2007. [30] N. D. Singpurwalla, “Survival in dynamic environments,” Stat. Sci., vol.
[6] B. Saha, K. Goebel, S. Poll, and J. Christophersen, “Prognostics methods 10, no. 1, pp. 86-103, 1995.
for battery health monitoring using a Bayesian framework,” IEEE Trans. [31] O. O. Aalen, O. Borgan, and H. K. Gjessing, Survival and Event History
Instrum. Meas., vol. 58, no. 2, pp. 291-296, 2009. Analysis: A Process Point of View. Springer, New York, 2008.
[7] K. Javed, R. Gouriveau, N. Zerhouni, and P. Nectoux, “Enabling health [32] M. L. T. Lee and G. A. Whitmore, “Threshold regression for survival anal-
monitoring approach based on vibration data for accurate prognostics,” ysis: modeling event times by a stochastic process reaching a boundary,”
IEEE Trans. Ind. Electron., vol. 62, no. 1, pp. 647-656, 2015. Stat. Sci., vol. 21, no. 4, pp. 501-513, 2006.
[8] K. L. Tsui, S. Y. Wong, W. Jiang, and C. J. Lin, “Recent research [33] J. I. Park and S. J. Bae, “Direct prediction methods on lifetime distribution
and developments in temporal and spatiotemporal surveillance for public of organic light-emitting diodes from accelerated degradation tests,” IEEE
health,” IEEE Trans. Rel., vol. 60, no. 1, pp. 49-58, 2011. Trans. Rel., vol. 59, no. 1, pp. 74-90, 2010.
[9] E. Zio and M. Compare, “Evaluating maintenance policies by quantitative [34] X. S. Si, W. Wang, C. H. Hu, M. Y. Chen, and D. H. Zhou, “A
modeling and analysis,”Rel. Eng. Syst. Safety, vol. 109, pp. 53-65, 2013. Wiener-process-based degradation model with a recursive filter algorithm
[10] V. Climente-Alarcon, J. A. Antonino-Daviu, E. Strangas, and M. Riera- for remaining useful life estimation,” Mech. Syst. Signal Process., vol. 35,
Guasp, “Rotor bar breakage mechanism and prognosis in an induction no. 1-2, pp. 219-237, 2013.
motor,” IEEE Trans. Ind. Electron., DOI: 10.1109/TIE.2014.2336604, [35] X. Wang, “Wiener processes with random effects for degradation data,”
2014. J. Multivariate Anal., vol. 101, no. 2, pp. 340-351, 2010.
[11] A. S. S. Vasan, B. Long, and M. Pecht, “Diagnostics and prognostics [36] G. A. Whitmore and F. Schenkelberg, “Modelling accelerated degradation
method for analog electronic circuits,” IEEE Trans. Ind. Electron., vol. 60, data using Wiener diffusion with a time scale transformation,” Lifetime
no. 11, pp. 5277–5291, 2013. Data Anal., vol. 3, no. 1, pp. 27-45, 1997.
[12] P. Baraldi, F. Mangili, and E. Zio, “Investigation of uncertainty treatment [37] W. Wang, M. Carr, W. Xu, and A. K. H. Kobbacy, “A model for
capability of model-based and data-driven prognostic methods using sim- residual life prediction based on Brownian motion with an adaptive drift,”
ulated data,” Rel. Eng. Syst. Safety, vol. 112, no. 1, pp. 94—108, 2013. Microelectron. Rel., vol. 51, no. 2, pp. 285-293, 2011.
[38] X. S. Si, W. Wang, C. H. Hu, and D. H. Zhou, “Remaining useful life
[13] N. Chen and K. L. Tsui, “Condition monitoring and residual life prediction
estimation-A review on the statistical data driven approaches,” Eur. J. Oper.
using degradation signals: revisited,” IIE Trans., vol. 45, no. 9, pp. 939-952,
Res., vol. 213, no. 1, pp. 1-14, 2011.
2013.
[39] X. S. Si, C. H. Hu, W. Wang, and M. Y. Chen, “An adaptive and
[14] F. DiMaio, K. L. Tsui, and E. Zio, “Combining Relevance Vector
nonlinear drift-based Wiener process for remaining useful life estimation,”
Machines and exponential regression for bearing residual life estimation,”
in Proc. Prognostics and Health Manag. Conf., Shenzhen, China, 2011,
Mech. Syst. Signal Process., vol. 31, pp. 405-427, 2012.
IEEE Xplore.
[15] M. I. Mazhar, S. Kara, and H. Kaebernick, “Remaining life estimation of [40] Q. Zhou, S. Zhou, X. Mao, and M. Salman, “Remaining useful life
used components in consumer products Life cycle data analysis by Weibull prediction of individual units subject to hard failure,” IIE Trans., vol. 46,
and artificial neural networks,” J. Oper. Manag., vol. 25, pp. 1184-1193, no. 10, pp. 1017-1030, 2014.
2007. [41] B. Olivares, M. Cerda, and J. Silva, “Particle-filtering-based prognosis
[16] X. S. Si, W. Wang, M. Y. Chen, C. H. Hu, and D. H. Zhou, “A degradation framework for energy storage devices with a statistical characterization of
path-dependent approach for remaining useful life estimation with an exact state-of-health regeneration phenomena,” IEEE Trans. Instrum. Meas., vol.
and closed-form solution,” Eur. J. Oper. Res., vol. 226, no. 1, pp. 53-66, 32, no. 2, pp. 364-376, 2013.
2013. [42] Y. F. Zhou, Y. Sun, J. Mathew, R. Wolff, and L. Ma, “Latent degrada-
[17] M. E. Orchard, P. Hevia-Koch, B. Zhang, and L. Tang, “Risk measures for tioncindicators estimation and prediction: A Monte Carlo approach,” Mech.
particle-filtering-based state-of-charge prognosis in lithium-ion batteries,” Syst. Signal Process., vol. 25, no. 1, pp. 222-236, 2011.
IEEE Trans. Ind. Electron., vol. 60, no. 11, pp. 5260–5269, 2013. [43] M. Orchard, F. Tobar, and G. Vachtsevanos, “Outer feedback correction
[18] A. Soualhi, H. Razik, G. Clerc, and D. Doan, “Prognosis of bearing fail- loops in particle filtering-based prognostic algorithms: statistical perfor-
ures using hidden markov models and the adaptive neuro-fuzzy inference mance comparison,” Stud. Inform. Control, vol. 18, no. 4, pp. 295-304,
system,” IEEE Trans. Ind. Electron., vol. 61, no. 6, pp. 2864-2874, 2014. 2009.
[19] X. S. Si, C. H. Hu, X. Y. Kong, and Zhou, D.H. “A residual storage [44] M. Orchard, A particle filtering based framework for fault diagnosis and
life prediction approach for systems with operation state switches,” IEEE failure prognosis, Ph.D. Thesis, Department of Electrical and Computer
Trans. Ind. Electron., vol. 61, no. 11, pp. 6304-6315, 2014. Engineering, Georgia Institute of Technology, 2007.
[20] E. Zio and G. Peloni, “Particle filtering prognostic estimation of the [45] M. J. Daigle and K. Goebel, “Model-based prognostics with concurrent
remaining useful life of nonlinear components,” Rel. Eng. Syst. Safety, vol. damage progression processes,” IEEE Trans. Syst., Man, Cybern. Syst., vol.
96, no. 3, pp. 403-409, 2011. 43, no. 3, pp. 535-546, 2013.
[21] W. Wang and M. Pecht, “Economic analysis of canary-Based prognostics [46] Q. Miao, L. Xie, H. J. Cui, W. Liang, and M. Pecht, “Remaining useful life
and health management,” IEEE Trans. Ind. Electron., vol. 58, no. 7, pp. prediction of lithium-ion battery with unscented particle filter technique,”
3077–3089, 2011. Microelectron. Rel., vol. 53, no. 6, pp. 805-810, 2013.
[22] E. G. Strangas, S. Aviyente,J. D. Neely, and S. S. H. Zaidi, “The effect of [47] Z. Q. Wang, W. B. Wang, C. H. Hu, and X. S. Si, “An additive Wiener
failure prognosis and mitigation on the reliability of permanent-magnet AC process-based prognostic model for hybrid deteriorating systems,” IEEE
motor drives ,” IEEE Trans. Ind. Electron., vol. 60, no. 8, pp. 3519–3528, Trans. Rel., vol. 63, no. 1, pp. 208-222, 2014.
2013. [48] K. L. Son, Mi. Fouladirad, A. Barros, E. Levrat, and B. Iung, “Remaining
[23] Z. S. Ye, L. C. Tang, and H. Y. Xu, “A distribution-based systems useful life estimation based on stochastic deterioration models: A compar-
reliability model under extreme shocks and natural degradation,” IEEE ative study,” Rel. Eng. Syst. Safety, vol. 112, no. 2, pp. 165-175, 2013.
Trans. Rel., vol. 60, no. 1, pp. 246-256, 2011. [49] X. L. Wang, N. Balakrishnan, and B. Guo, “Residual life estimation based
[24] Z. S. Ye, Y. Wang, K. L. Tsui, and M. Pecht, “Degradation data analysis on a generalized Wiener degradation,” Rel. Eng. Syst. Safety, vol. 124, no.
using Wiener processes with measurement errors,” IEEE Trans. Rel., vol. 1, pp. 13-23, 2014.
62, no. 4, pp. 772-780, 2013. [50] J. Liu and M. West, “Combined parameter and state estimation in
[25] M. D. Pandey, X. -X. Yuan, and J. M. van Noortwijk, “The influence simulation-based filtering,” in Sequential Monte Carlo Methods in Practice,
of temporal uncertainty of deterioration on life-cycle management of A. Doucet, N. de Freitas, and N. Gordon, Eds. New York: Springer-Verlag,
structures,” Structure Infrastructure Eng., vol. 5, no. 2, pp. 145-156, 2009. 2001.

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TIE.2015.2393840, IEEE Transactions on Industrial Electronics
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS 12

[51] A. P. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood

from incomplete data via the EM algorithm,”J. Roy. Statist. Soc., Ser. B,
vol. 39, no. 1, pp. 1-38, 1977.
[52] H. E. Rauch, F. Tung, and C. T. Striebel, “Maximum Likelihood estimates
of linear dynamic systems,” AIAA J., vol. 3, no. 8, pp. 1145-1450, 1965.
[53] J. Svensson, Survival estimation for opportunistic maintenance. Ph.D. the-
sis, Mathematical Statistics, Chalmers University of Technology, Göteborg,
2007.
[54] Y. Xing, E.W.M. Ma, K.-L. Tsui, M. Pecht, “An ensemble model for
predicting the remaining useful performance of lithium-ion batteries,”
Microelectron. Rel., vol. 53, pp. 811–820, 2013.
[55] S. Lorén, J. de Maré, “Maintenance for reliability—a case study,” Ann.
Oper. Res., DOI: 10.1007/s10479-011-0873-8, 2012.
[56] B. Saha and K. Goebel, “Battery data set,” NASA Ames Prognostics Data
Repository, Moffett Field, CA: NASA Ames, 2007. [Online]. Available:
http://ti.arc.nasa.gov/project/prognostic-data-repository/
[57] G. Jin, D. E. Matthews, and Z. Zhou, “A Bayesian framework for on-line
degradation assessment and residual life prediction of secondary batteries
in spacecraft,” Rel. Eng. Syst. Safety, vol. 113, no. 1, pp. 7-20, 2013.
[58] M. Shahriari and M. Farrokhi, “Online state-of-health estimation of VRLA
batteries using state of charge,” IEEE Trans. Ind. Electron., vol. 60, no. 1,
pp. 191–202, 2013.

Xiao-Sheng Si is a Lecture at High-Tech Institute

of Xi’an, Xi’an, China. He has published more than
20 articles in several journals including European
Journal of Operational Research, IEEE Transactions
on Reliability, IEEE Transactions on Industrial Elec-
tronics, IEEE Transactions on Fuzzy Systems, IEEE
Transactions on Systems, Man and CyberneticsłPart
A, IEEE Transaction on Automation Science and En-
gineering, Reliability Engineering and System Safety,
and Mechanical Systems and Signal Processing. He is
an active reviewer for several international journals.
His research interests include prognostics and health management, reliability
estimation, and predictive maintenance.

0278-0046 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.