Remaining Useful Lifetime Prediction Based on the Damage-marker Bivariate
Degradation Model: A case study on lithium-ion batteries used in electric
vehicles

Jing Feng, Paul Kvam, Yanzhen Tang

PII: S1350-6307(16)30172-8
DOI: doi: 10.1016/j.engfailanal.2016.04.014
Reference: EFA 2868

To appear in:

Received date: 25 August 2015

Revised date: 8 April 2016
Accepted date: 12 April 2016

Please cite this article as: Feng Jing, Kvam Paul, Tang Yanzhen, Remaining Use-
ful Lifetime Prediction Based on the Damage-marker Bivariate Degradation Model:
A case study on lithium-ion batteries used in electric vehicles, (2016), doi:
10.1016/j.engfailanal.2016.04.014

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Remaining Useful Lifetime Prediction Based on the Damage-marker Bivariate

Degradation Model: A case study on lithium-ion batteries used in electric vehicles

Jing Feng1, Paul Kvam2, Yanzhen Tang1

Abstract: Remaining useful lifetime (RUL) refers to the available service time left before the

T
performance of a system degrades to an unacceptable level. Recent innovations to lithium-ion

IP
battery packs have raised expectations with regard to energy storage capability in electric
vehicles (EVs). This has catalyzed new research on RUL prediction, since accurate RUL

R
prediction for lithium-ion batteries used in EV is highly desired for safe and

SC
lifetime-optimized operation. A battery’s maximum releasable capacity (MRC) usually decays
over time, thus it is a primary factor which determines the remaining cycle life of the battery.
However, MRC usually needs to be measured under strict laboratory conditions and cannot be

NU
easily assessed during field use in EVs. This naturally inhibits potential applications of many
online RUL prediction methods that rely on MRC measurements. We found two markers of
MRC decay, named as time-to-voltage-saturation (TVS) and time-to-current-saturation (TCS),
MA
from constant-current constant-voltage charging (CC/CV) curves, which can be used in place
of MRC measurements during field use. We propose a RUL prediction method based on a
damage-marker bivariate degradation model in which one term represents damage (MRC
D

decay), the other represents a composite marker constructed from TVS and TCS. We model
this degradation process using a two-dimensional Wiener process to obtain the RUL
TE

distribution, using method of maximum likelihood for population parameters’ estimation.

Bayesian methods are used to update the estimators of parameters with online data. The
P

effectiveness of the model is verified with public data of four 18650 batteries from NASA.
CE
AC

Keywords: remaining useful lifetime; marker processes; bivariate Wiener process;

lithium-ion battery; electric vehicle

1
National University of Defense Technology, Changsha, HN, China
2
Georgia Institute of Technology, Atlanta, GA, USA
1 / 31
 ACCEPTED MANUSCRIPT

1. Introduction

Nowadays, lithium-ion batteries are used in a variety of portable electronic devices. They
have become indispensable part of our daily life. Due to some of their advantages like high
energy density, low self-discharge rate and long cycle life, and so on [1-3], they are becoming

T
an optimal energy supply for electric vehicles (EVs). And their use is expected to increase

IP
rapidly in the next few decades as the market of EVs grows. Generally, lithium-ion batteries
work along with internal side reactions in their whole lifetime. These reactions consume

R
lithium and generate gaseous and non-conductive by-products that partially prevent batteries
from charging. Under different usage and environment conditions, the rates of these reactions

SC
usually vary. When the maximum available capacity of batteries eventually degrades below
their required level of performance, they can no longer perform their intended functions [1,

NU
4]. To meet the requirements on safety management, research on capacity fading and
remaining useful life (RUL) estimation of lithium-ion batteries has become a hotspot in
prognostics and health management (PHM), reliability engineering and other related fields [3].
MA
The idea is to use the real-time and historical data of batteries to forecast their future states
under certain operating conditions. In accordance with such useful information, users then can
make adequate replacement policies for lithium-ion batteries. In this case, we focus on the
RUL estimation for lithium-ion batteries used in EVs.
D

In recent years, extensive research has been conducted on degradation modeling and RUL
TE

estimation of lithium-ion batteries. Maximum releasable capacity (MRC) and internal

resistance are often used as the health indicators (HIs) of a lithium-ion battery for degradation
modeling and RUL estimation. MRC is the total Amp-hours (Ah) available when the battery
P

is discharged at a certain discharge current (specified as a C-rate) from 100 percent

CE

state-of-charge to the cut-off voltage at 25oC. As the battery ages, its MRC tends to be lower
than the nominal capacity due to the electrochemical degradation. Once the MRC decreases
below a specified threshold (e.g. 70% of the nominal capacity), it should be replaced. In the
AC

light of the above concepts, the RUL of a battery is defined as the number of fully
charge/discharge cycles from the present to the moment when the MRC drops below 70% of
its nominal capacity. This failure threshold is usually pre-specified on the basis of an
industrial standard or customers’ requirement. The models based on MRC or internal
resistance are named as Direct-HIs-Models in this paper, since MRC and internal resistance
are both direct performance indexes of lithium-ion batteries. According to their modeling
features, these models can be further divided into four groups, namely physics-based models,
physical-empirical models, empirical models and data-driven models. Pinson et al. [5]
proposed a physics-based model to describe the degradation of MRC over cycles. Safari et al.
[6] proposed a physical-empirical model based on the fatigue approach for RUL prediction of
lithium-ion batteries. Smith et al. [7] conducted reliability assessment and lifetime prediction
of batteries with empirical models. Meanwhile, data-driven methods, such as autoregressive
and moving average (ARMA) [8], neutral network (NN) [9], support vector machine (SVM),
particle filter (PF), have been widely adopted for the RUL estimation of lithium-ion batteries.
Park et al. [10] established an SVM model for monitoring product quality based on

2 / 31
 ACCEPTED MANUSCRIPT

degradation profiles. Similar to SVM, relevance vector machine (RVM), which is based on
sparse Bayesian learning and kernel learning, has been used to quantify the uncertainty of
RUL estimation [3, 11, 12]. Moreover, PF as well as its extended methods has been applied to
prognostic for lithium-ion batteries [3, 13-16]. Alternatively, in order to improve the
estimation accuracy, Liu et al. [17] presented a Gaussian process regression function to track
and predict the global capacity degradation and local capacity regeneration. Zhang et al. [18]

T
conducted a review on prognostics and health monitoring of lithium-ion batteries. Besides

IP
these above methods, some fusion prognostic models were also developed to improve the
accuracy of battery RUL estimation [3, 19-21]. Kozlowski [19] proposed a RUL prediction

R
approach by combining ARMA model, neutral networks, and fuzzy logic. Xing et al. [20]

SC
proposed an ensemble model for the RUL prediction of lithium-ion batteries by combining a
fused empirical exponential model, polynomial regression model, and PF algorithm.
It is noted that almost all of the aforementioned Direct-HIs-models utilize capacity or

NU
internal resistance as the health indicator of batteries. Liu et al. [3, 22] and Williard et al. [1]
pointed out that it is very difficult to perform battery capacity or internal resistance
measurement in online applications, especially for EV batteries, since the battery capacity is
MA
usually measured in some special ways. For instance, batteries should be deeply discharged
from 100% SOC to 0% SOC or charged from 0% SOC to 100% SOC [23]. Moreover, the
internal resistance cannot be accessed via regular sensors, either. Direct and precise methods
D

of measuring these internal parameters, e.g., electrochemical impedance spectroscopy (EIS)

test, are very costly [23]. That is to say, the precise online measurement of capacity or internal
TE

resistance for EV batteries requires extra charge/discharge tests under strict conditions. This
may cause permanent damage to the health of batteries.
P

Because of the significance of RUL estimation for lithium-ion batteries in online

CE

applications, some Indirect-HIs-models for RUL estimation of batteries have been proposed
recently. The idea is to extract some indirect performance parameters reflecting the health
state of batteries from available monitoring information, such as current, voltage, time and
AC

temperature of charge/discharge process. Based on the sample entropy (SamEn) feature of

discharge voltage, Widodo et al. [24] proposed an intelligent prognostic method for battery
health. The SamEn is treated as a vital feature of learning systems like SVM and RVM. The
output of the learning systems is state of health (SOH) given with capacity. Thus, this model
can automatically predict SOH based on SamEn. However, this technique is time-consuming
and requires the capacity parameter in evaluating the sample entropy indicator. Liu et al. [3]
presented a novel indirect HI, the discharge voltage difference of equal time interval
(DVD_ETI), to measure the capacity degradation in each charge/discharge cycle. The authors
gave the clue of nonlinear transformation and optimization of the extracted HI to enhance the
correlation between the HI and battery capacity. Then, an optimized RVM algorithm is
presented to improve the accuracy and stability of RUL estimation. The aforementioned
indirect HIs are all extracted based on the discharge information. However, negative
influences of some factors, like temperature, vibration, and un-foreseen usage profiles, that
give rise to uncertainties of discharge processes are not addressed in these methods.
As we all know, the usage/discharge profiles of each cycle of lithium-ion batteries used in
EVs are totally different. So the above methods can’t be applicable for EV batteries. Unlike
3 / 31
 ACCEPTED MANUSCRIPT

the discharge process, charge conditions are usually controllable and predictable. During the
charge process, only one protocol, namely constant current constant voltage (CC/CV), is
applied throughout the entire life of batteries [1, 24]. In the CC/CV protocol, a varistor, first,
is used to keep the current constant for batteries until the maximum voltage threshold is
reached. Then, a constant voltage is applied to batteries until the current drops to the
minimum threshold. Thus, how to extract indirect HIs based on charge current, voltage and

T
time information is indeed critical in online RUL estimation for EV batteries, which would

IP
reduce uncertainties in the usage profile. To the best of our knowledge, little work has been
conducted on this field. Williard et al. [1] developed two HIs based on the CC/CV protocol,

R
that is, length of time of constant current charge phase (CCCT) and length of time of constant

SC
voltage charge time (CVCT). During each cycle, capacity, resistance, CCCT and CVCT were
measured. And then they were all taken into consideration to give a fusion result for SOH
estimation of batteries in online applications.

NU
In conclusion, the online RUL estimation of EV lithium-ion batteries is usually performed
on the basis of the following three facts.
(1) Direct HIs like MRC can accurately reflect the damage of batteries, but they can only
MA
be measured offline. In other words, they cannot be monitored online.
(2) The discharge profiles are totally different for each cycle due to variable discharge
current, temperature, etc. Only one charge protocol (e.g. CC/CV) is applied throughout the
D

entire life of batteries. In the rest of the paper, we will study the situation that the batteries
are fully discharged with time-varying profile before charged to full capacity, which is
TE

common in practical applications.

(3) The charge information, such as voltage, current and time, can be measured accurately
P

in each cycle, both for offline batteries and online batteries.

CE

To overcome the problems and challenges which exist in the above facts, we propose
indirect HIs extracted methods based on charge profile information, and established a
damage-marker bivariate Wiener model for the online degradation analysis and RUL
AC

estimation of EV lithium-ion batteries. We aim to make full use of the monitoring charge
information and reduce the difficulty in measuring capacity and internal resistance in online
applications. Here, “damage” means the MRC of each cycle, which is a direct HI. It is
assumed that this index can only be measured under offline conditions during each cycle with
some special constant current discharge protocols, but cannot be measured in online
applications due to the complexity of discharge profile. Furthermore, “marker” refers to some
indirect health indexes, which is extracted from monitoring charge information and has close
correlation with damage. This marker can objectively be monitored both under offline and
online conditions. Actually, the idea of marker was first introduced in survival analysis by
Jewell and Singpurwalla et al [25, 26], which incorporates auxiliary information to improve
lifetime estimation. Based on the idea of marker, Whitmore et al. [27] proposed a bivariate
Wiener process for the life prediction problem in which one performance characteristic
represents a marker and another one is a latent (unobserved) variable. Later, Lee et al. [28]
introduced a useful Bayesian extension of the bivariate Wiener process model. However, the
statistical parameter estimation methods in above literature were based on the assumption that
the damage level, which cannot be observed either online or offline, determines the RUL. In
4 / 31
 ACCEPTED MANUSCRIPT

our case, the latent damage level will be monitored under some offline conditions, which
cannot be obtained online.
In fact, numerous examples in reliability engineering reveal that the damage process of
products often cannot be directly observed in field applications. Nevertheless, some other
closely related markers can indicate the health degradation of products, since they can directly
be observed. In such cases, it will greatly improve RUL prediction under online conditions by

T
unearthing the correlation between the marker and damage from offline data, and making full

IP
use of observed online marker information.
In our work, we first develop the framework of RUL prediction with the damage-marker

R
bivariate degradation model. Then we extract two HIs, time-to-voltage-saturation (TVS) and

SC
time-to-current-saturation (TCS), from the charge current, voltage and time, and construct a
composite marker. The efficiency of the marker is demonstrated by the Pearson correlation
analysis with offline data. Further, the parameters of the damage-marker model are estimated

NU
by the Bayesian method based on laboratory offline prior data. Finally, the RUL estimation is
conducted with the conditional distribution of damage on marker.
The main contributions of our work can be summarized as (1) two types of Health indices
MA
(HIs) called TVS and TCS and one composite marker are proposed based on charging current,
voltage and time information instead of maximum releasable capacity in each cycle. And the
efficiency of the extracted marker is confirmed via correlation analysis. (2) two-dimensional
D

Wiener process is used to realize RUL estimation with uncertainty representation. (3)We give
the model specification and verification method in the case study of NASA lithium-ion
TE

batteries, and this provides a way to determine the applicability of our model.
P

The layout of this paper is organized as follows. The framework of RUL estimation with
CE

the damage-marker process is introduced in Section 2. The HI extraction and composite

method from charge curve are given in Section 3. The development of the damage-marker
bivariate degradation model and RUL distribution model are described in detail in Section 4.
AC

The RUL model parameters are estimated in Section 5. Using NASA lithium-ion batteries
data, Section 6 illustrates the efficiency of the proposed method for the RUL estimation of EV
lithium-ion batteries. Finally, the conclusion and future research are presented in Section 7.

2. Framework of RUL prediction

To intuitively know about the whole procedure of the RUL prediction method based on the
damage-marker degradation process, a flow chart is illustrated as follows. Fig. 1 shows the
framework of our RUL prediction method for EV lithium-ion batteries.

5 / 31
 ACCEPTED MANUSCRIPT

Datebase
Historical laboratory database Field monitoring Database

· Charge curves data · MRC degradation data · Charge curves data

T
Step one

IP
Extract markers offline Extract markers online
Marker I: Time-to-voltage-saturation(TVS) Marker I: Time-to-voltage-saturation(TVS)
Marker II: Time-to-current-saturation(TCS) Marker II: Time-to-current-saturation(TCS)

R
One composite marker One composite marker
(historical marker data) (field marker data)

SC
Step two
Degradation Modeling
· Damage-marker bivariate model
· Distribution of MRC conditional on integrate marker
·

NU
Remaining useful life(RUL) distribution model

Step three
Bayesian estimation of model parameters
MA
Output
RUL prediction
· point estimation
· Confidence interval estimation
· Model verification
D

Figure 1. Framework of RUL prediction for EV lithium-ion batteries based on

TE

damage-marker degradation process

P

For the sake of convenience of modeling, we make the following assumptions for our
study based on characteristics of EV lithium-ion battery.
CE

(1) The damage and corresponding marker can be monitored in laboratory testing, from
which the data obtained are called “offline data”.
AC

(2) During the use of batteries, only marker data can be monitored. The parameters of the
damage-marker degradation model could be estimated based on the damage (MRC) and
composite marker (TVS and TCS) extracted from the charge information obtained in
laboratory. One of the model parameters describes the relationship between “damage” and
“marker”. Then, the model can be used to predict online damage through online monitoring
marker information. At last, the online RUL estimation of lithium-ion batteries could be
realized.
In the last section, we study a case of EV lithium-ion batteries to demonstrate the RUL
prediction method based on the damage-marker bivariate degradation model. Throughout the
paper, the “lifetime” of a testing unit is referred to as the number of charge/discharge cycles of
the unit observed before failure. For simplicity, we occasionally use “time” instead of
“lifetime”.

3. Marker extraction method

In this section, we show how to extract markers from charge curves that could reflect the
6 / 31
 ACCEPTED MANUSCRIPT

degradation of the MRC of lithium-ion batteries under repeated charge/discharge cycles.

3.1. CC/CV charge pattern of lithium-ion batteries

The standard lithium-ion battery is composed of four basic elements: the cathode (or positive
electrode, generally constructed from lithium compounds), anode (or negative electrode,
generally constructed from special carbon), electrolyte and separator [29]. While the battery is

T
being charged, lithium-ions are forced to move into the anode, which leads to a positive

IP
charge. During the discharge process, the ion flow moves in the opposite direction, namely
from the anode to the cathode. Because this reaction is reversible in secondary batteries,

R
secondary batteries are capable of being charged and reused up to hundreds of

SC
charge/discharge cycles.
In regular field use, the EV lithium-ion battery is discharged at varying current, due to
differences in drivers, road conditions, and so on. In comparison with the discharge process,

NU
the charge process of batteries is usually performed under a constant-current constant-voltage
(CC/CV) protocol controlled by the charge system. CC/CV systems are significantly
necessary for batteries, since batteries are vulnerable to damage when the charge voltage
MA
exceeds the nominal maximum voltage. The constant-current charge rate specified by
manufacturer is the maximum charge rate at which the battery can normally work. Special
precautions are needed in order to maximize the charge rate and to ensure that the battery is
D

fully charged rather than overcharged. Therefore, it is suggested that the charge regime adopt
TE

a constant voltage before the battery voltage reaches its nominal maximum voltage. Note that
this implies that chargers for lithium-ion batteries must be capable of controlling both the
charge current and battery voltage. A typical charge curve [29], which describes the
P

relationship between the change process of current and voltage and charge time, is shown in
CE

Figure 2. As we see, the battery is first charged at a constant current (Stage 1). Meanwhile, the
battery voltage increases over time until it reaches the nominal maximum voltage. Then, the
charge voltage is kept constant while the current decreases over time (Stage 2). The charge
AC

process is terminated when the current is less than 3% of the rated current. Such curves can be
recorded during each charge cycle of EV batteries.

7 / 31
 ACCEPTED MANUSCRIPT

T
R IP
SC
NU
Figure 2. A typical charge curve of a lithium-ion battery
MA
3.2. Extract markers from the charge curve

Our work is based on the battery dataset provided by the Prognostics Center of Excellence
D

(CoE) at NASA Ames[30]. Four 18650 lithium-ion batteries (denoted by # 5, #6, #7 and #18,
TE

respectively) were tested using three different operational profiles (including charge,
discharge and impedance) at room temperature. Charge was carried out at a constant current
(CC) of 1.5A until the battery voltage reached 4.2V, and then continued in a constant voltage
P

(CV) mode until the charge current dropped to 20mA. Discharge was conducted at a constant
CE

current (CC) of 2A until the battery voltage fell to 2.7V, 2.5V, 2.2V and 2.5V for batteries #5,
#6, #7 and #18, respectively. The testing was stopped when the batteries reached the
end-of-life (EOL) criteria, which was a 30% fading in rated capacity, for instance, from 2Ahr
AC

to 1.4Ahr. It is clear that the charge mode adopted in the experiments is CC/CV, which is the
same as the one applied to EV lithium-ion batteries.
For battery #5, three charge/discharge cycles, namely 2nd, 91st and 151st, are selected from
its 170 charge/discharge cycles. And the corresponding charge voltage curves are depicted in
Figure 3 We define the time interval from the beginning to the time at which voltage reached
4.2V as time-to-voltage-saturation (TVS). Let denote the TVS, where represents
the number of charge/discharge cycles. From Figure 3, we could see that decreases as
increases. This means that the charge voltage quickly reached the saturation point as the
battery’s capacity decreased. This phenomenon is due to the increase of the battery’s internal
resistance over time. Based on this fact, we use TVS as the first marker of MRC.
As shown in Figure 4, we analyze the charge current curves using the data from NASA in
a similar manner. We first define the time interval from the time at which the charge current
starts to drop from the rated current of 1.5A to the terminal charge time at which the charge
current is 20mA as time-to-current-saturation (TCS). Let denote the TCS, where
represents the number of charge/discharge cycles. From Figure 4, we can see that

8 / 31
 ACCEPTED MANUSCRIPT

increases with increasing. This means that the area behind the charge current curve, which
is related to the capacity that the battery can hold, decreases as the battery ages. This
phenomenon is due to the decreasing number of active lithium-ion. Based on the above
information, we adopt TCS as the second marker of MRC.
Overall, it is clearly seen that the decay of MRC over time can be indirectly reflected by

T
the decreasing TVS and the increasing TCS. And it is relatively easy to obtain measurements
of TVS and TCS from the CC/CV charge curve during the use of EV lithium-ion batteries.

R IP
SC
NU
MA

Figure 3. Extraction of time-to-voltage-saturation (TVS) from CC/CV charge voltage curve

D
P TE
CE
AC

Figure 4. Extraction of time-to-current-saturation (TCS) from CC/CV charge current curve

Just as Liu et al said in [3], the main challenge for the RUL estimation of lithium-ion
batteries working under time-varying conditions is to find robust HIs. Our markers are
extracted under the assumption that the fully discharge and fully charge profile appear
alternatively. Because the discharge profile of each cycle is time-varying with different
discharge rate, temperature etc., it may have different damage to the health of the battery. The
damage can be reflected in the TVS and TCS extracted from the next charge profile. That is to
say, different discharge modes can cause different damage, and induce different change of
TVS and TCS. This phenomenon can be explained by aforementioned physical failure
mechanisms of lithium-ion batteries. From this point of view, TVS and TCS are robust
markers in the application of fully CC/CV charge and fully discharge condition with
time-varying profiles. This condition is somewhat common in practical applications, since we
9 / 31
 ACCEPTED MANUSCRIPT

usually charge the battery to full after the capacity of the battery is thoroughly empty. Under
this condition, we need to know the RUL of the EV battery.
In the following part, we will propose a composite marker based on TVS and TCS, and
verify the efficiency of the composite marker via the Pearson correlation analysis between
composite marker and direct damage (MRC).

T
IP
4. Damage-marker degradation model and RUL prediction model

R
Wiener processes are widely used in mathematical models to describe the degradation process
(e.g., wear, corrosion) of products in the field of engineering [31]. In the chosen model, the

SC
entire degradation process is generally divided into many degradation increments, each of
which has a small effect on the entire degradation process and serve normal assumptions.

NU
Owing to the major impact of cumulative damage failures on degradation failures, we
establish the damage-marker degradation model based on a two-dimensional Wiener process.
In this section, we first introduce the damage-marker degradation model. And then we derive
MA
the RUL conditional distribution function.

4.1. Modeling the damage-marker degradation based on the Wiener process

The damage-marker degradation model introduced by Whitmore et al. [27] assumes that every
D

testing unit path is represented by a bivariate Wiener process W(t) = {X(t), Y(t)}, t > 0, with
TE

initial values W(0)={X(0),Y(0))}={0,0}, drift μ and covariance Σ, where

 X2  XY
  ( X , Y ) ,   , > 0.
 XY Y2
P

The term X(t) represents a sequence of a unit’s unobservable damage states, namely the latent
CE

degradation process. And the term Y(t) denotes a marker process that is correlated with the
degradation process X(t). The strength of the association between the two terms of the
AC

bivariate Wiener process is described by the correlation coefficient , where

   XY /( X   Y ) . The bivariate Wiener model with five parameters is denoted as W (t |  ) ,
where   ( X , Y , X , Y ,  ) , and the matrix  is assumed to be positive infinite. The test
unit fails when the damage process X(t) reaches the failure threshold L>0 at the first time. We
denote this number-of-cycles threshold (equivalent to a “first-hitting time”) by a random
variable T. It is well known that T follows an inverse Gaussian (IG) distribution [32] , and the
probability density function (pdf) of T can be written as
fT (t )  L exp( ( L   X t ) 2 /( 2 X2 t )) / 2 X2 t 3 . (1)
This stochastic model depends on a single marker. If there are two or more markers, the
univariate model cannot be applicable. To solve the problem, a multivariate model can be
established. However, it is not easy to get the multivariate extension. Therefore, under the
assumption that each marker has a degree of correlation with the underlying damage process,
a composite marker based on the selection of candidate markers should be taken into
consideration. The construction of the composite marker denoted by Y(t), will be detailed in
Section 5.
10 / 31
 ACCEPTED MANUSCRIPT

Followed by this argument, we propose a two-dimensional Wiener process for

characterizing this degradation problem. When this model is applied to a battery’s RUL
estimation, t denotes the number of cumulative charge-recharge cycles, X(t) refers to the
degradation process of MRC, and Y(t) refers to the degradation process of composite marker
constructed from TVS and TCS.

T
4.2. RUL distribution based on the damage-marker model

IP
The RUL of the online unit is a function of its latent damage process. This process cannot

R
be directly observed during normal online operation, so we have to use online data from the
marker to conduct analysis. Before we outline the method for estimating the RUL distribution

SC
under these conditions, some mathematical notations are listed as follows:
 Let n be the number of units tested in the lab.
 Let {X i (tij ), Yi (tij )} be the observed laboratory offline data for the i th unit at the t ij th
cycle, j  1,2,, mi , i  1,2,, n .
NU
 The measured online data is denote as Yt0  (Y (t1 ),, Y (tk )) , where Y (ti ) is the
MA
observation of Y(t) at the t i th cycles, and 0  t1  t 2    t k  t0 . We assume that the unit
does not fail at time t0.
 Let L be the failure threshold.
D

 Let x(t0) be the pseudo damage level of the unit at the t0th cycle, which is not observable
TE

in the field. According to the definition, x(t0)<L.

Our aim is to estimate the RUL of the unit at the t0th cycle. Let Tt0 denote the unit’s RUL.
P

The damage process {X(t),t }of the unit is characterized by a Wiener process. From the
definition of RUL, we note that Tt0 satisfies the following equation
CE

Tt0  inf{t | X (t  t0 )  L, X (t0 )  x(t0 ), t  0} . (2)

From the independent increment and homogeneous Markov properties of the Wiener process,
AC

we have
Tt0  inf{t | X (t  t 0 )  X (t 0 )  L  x(t 0 ), t  0}
. (3)
 inf(t | X (t )  L  x(t 0 ), t  0)
Thus, we could infer that Tt0 follows the inverse Gaussian distribution given in Equation (1).
Here L is replaced by L-x(t0). The pdf of RUL at time t0 is
fT (t )  ( L  x(t0 )) exp(( L  x(t0 )   X t ) 2 /( 2 X2 t )) / 2 X2 t 3 (4)
Because X(t) is a latent degradation process and x(t0) is not observable online, Equation (4) is
not sufficiently self-contained. However, the observed value of Yt0 is closely correlated with
x(t0), which can be described by the two-dimensional Wiener process. Next, we draw
attention to the statistical inference of x(t0) based on information obtained from Yt0 .
In order to model the uncertainty in the process, tt is natural to treat x(t0) as a random
variable. In the following sections, lowercase notations like yi denote observed values of the

11 / 31
 ACCEPTED MANUSCRIPT

corresponding random variable Yi . For consistency, we restate x(t0) as X(t0). It follows,
therefore, that the focus is how to get the distribution of X(t0) from Yt0 .
It is simple to show that the conditional distribution of X (t0 ) | Yt0 is a normal distribution
from the regression property of two-dimensional normal variable[30], that is,
X (t0 ) | Yt0 ~ N ( X |Y (t0 ), X2 |Y (t0 ))

T
(5)

IP
with
 X |Y (t0 )  E[ X (t0 ) | Yt ]  E[ X (t0 ) | Y (t0 )  y(t0 )]   X t0   X ( y(t0 )  Y (t0 )) /  Y
0
(5a)

R
 X2 |Y (t0 )  Var[ X (t0 ) | Yt ]  Var[ X (t0 ) | Y (t0 )  y(t0 )]  (1   2 ) X2 t0
0
(5b)

SC
If an unit has not failed by t0 cycles, then from Equations (4) and (5), we can express the
pdf of Tt0 conditional on marker data Yt0 , denoted as f Tt 0 (t | Yt0 ) ,

NU
L
f Tt 0 (t | Yt0 )   [( L  s) exp( ( L  s   X t ) 2 /( 2 X2 t )) / 2 X2 t 3 ] 


[exp( ( s   X |Y (t 0 )) 2 /( 2 X2 |Y (t 0 )) / 2 X2 |Y (t 0 ) ]ds

MA
L
  ( L  s) exp( ( L  s   X t ) 2 /( 2 X2 t )  ( s   X |Y (t 0 )) 2 /( 2 X2 |Y (t 0 )))ds /( 2  X2  X2 |Y (t 0 )t 3 )


(6)
where  X |Y (t0 ) and  X2 |Y (t0 ) are listed in Equation (5a) and (5b), respectively.
D

The unknown parameters from (6) include the five parameters of two-dimensional Wiener
TE

process W (t |  ) , where   ( X , Y ,  X ,  Y ,  ) . These parameters can be estimated with

offline and online data. Here, offline means the batteries were not installed in EV and were
P

tested in the laboratory. The offline data refers to the MUC, TCS and TVS information
CE

extracted from laboratory test data. Online means the batteries has been installed in EV. For
these batteries, the discharging cycle is actually the use profile of EV and the discharge
current and power are time-varying. When the batteries are out of energy, the EV will be
AC

stopped and the batteries be charged to full. The charging voltage, current and time
information could be recorded with the charging facility. The online information refers to the
TCS and TVS extracted from CC/CV charge cycle. Especially noted that we could not get the
MRC data from charge cycle and time-varying discharge cycle when the batteries are installed
in the EV. The estimate method of the parameters was detailed in Section 5. The integral in
Equation (6) could be computationally solved using numerical integration techniques, for
example the trapezoid method[33].

4.3. RUL prediction

Based on the conditional pdf f Tt 0 (t | Yt0 ) , we can obtain the point estimate and confidence
interval (CI) of the predicted RUL at t0 cycles.
For the point estimation of the predicted RUL at t0, we use the median of the RUL
distribution computed at t0, denoted by sˆt0 , which can be obtained by solving
sˆt0
0
fTt 0 (t | Yt0 )dt  0.5 (7)

12 / 31
 ACCEPTED MANUSCRIPT

Here the median is chosen over the mean as an approximate value for the point estimate of the
predicted RUL since the distribution of the RUL is often highly skewed. The median is more
robust than mean. We can construct effective approximate confidence intervals (CIs) for the
predicted based on Equation (7); for any constant , let sˆ |t0 be the time satisfying
sˆ |t0
 f Tt 0 (t | Yt0 )dt  1   (8)

T
0

Therefore, the 100% CI of the predicted RUL at t0 is represented as [sˆ(1 / 2)|t0 , sˆ( / 2)|t0 ] ,

IP
where sˆ(1 / 2)|t0 and sˆ( / 2)|t0 can be calculate numerically from Equation (8).

R
SC
5. Parameter estimation for the RUL distribution

NU
5.1. Estimation of with offline data

Parameter  describe the correlation relationship between X(t) and Y(t)， which supposed
MA
to be not changed during the whole life. X(t) refers to the degradation process of MRC offline.
Thus, parameters μ could be obtained directly via Maximum Likelihood Estimation
(MLE) based on the offline data. The estimators of μ need not update with online data.
th
unit (1≤ i ≤ n), we observe {X(t),Y(t)}at various cycles 0  ti1  ti 2    timi ,
D

For the i
denoted by {x(tij ), y(tij )} , i=1,2,…,n; j=1,2,…,mi. We tacitly assume that not all of the n
TE

units have failed during the test. The data of the i th uint are summarized as:
 Measurement interval data: ti1  ti1 , tij  tij  ti ( j 1) , j  2,3,, mi ;
P

 Degradation increment of X(t): xi1  x(ti1 ), xij  x(tij )  x(ti ( j 1) ), j  2,3,, mi ;
CE

 Degradation increment of Y(t): yi1  y(ti1 ), yij  y(tij )  y(ti ( j 1) ), j  2,3,, mi .
First, we compute the MLE of  X , X from the degradation increment data from the Wiener
2
AC

process X(t). Since the observed set xij represent independent normal samples, the likelihood
function of  X , X is expressed
2

n mi
L(  X , X2 ; DataX )   exp( (xij   X tij ) 2 /( 2 X2 tij )) / 2 X2 tij (9)
i 1 j 1

and the MLEs of  X , X are solved as

2

n n n mi n n
̂ X   xim /  tim , ˆ X2  ( ((xij ) 2 / tij )  ( xim ) 2 /  tim ) / N ,
i i i i
(10)
i 1 i 1 i 1 j 1 i 1 i 1

n
where N   mi .
i 1

Next, we compute the MLEs of Y , Y from the degradation increments Y(t):

2

n n n mi n n
̂Y   yim /  tim , ˆ Y2  ( ((yij ) 2 / tij )  ( yim ) 2 /  tim ) / N .
i i i i
(11)
i 1 i 1 i 1 j 1 i 1 i 1

13 / 31
 ACCEPTED MANUSCRIPT

From the independent increment property of the two-dimensional Wiener process, x p1q1
and y p2q2 are independent if p1  p2 or q1  q2 . In other words, the MLE of the correlation
coefficient  can only be estimated with the degradation increment data of X(t) and Y(t) with
the same subscript as
n mi
ˆ   ((xij  ˆ X tij )(yij  ˆ Y tij ) / tij ) /( Nˆ X ˆ Y )

T
(12)
i 1 j 1

IP
5.2. Estimation of Y and  Y with offline and online data

R
SC
As we know, Y(t) refers to the degradation process of composite marker. The prior estimator
of Y , Y can be obtained with offline data. Then the estimators need update with online data.
Let Yti  { yt1 , yt2 ,, yti } represent the online marker data for the unit in field use. Denote  Y
2

NU
as D, then parameters Y and D should be updated with these observed online data Yti .
Because of its convenient way for updating estimates, we pragmatically adopt the Bayesian
MA
framework when combining the online and offline data. The Bayesian estimates of Y and D
reflect the individual degradation characteristics, and the updating steps are outlined below.
D

Step1. Determination of prior distribution  (Y , D) .

As {Y (t ) : t  0} is a Wiener process, the increment Y (t )  Y (t  t )  Y (t ) is a Normal
TE

random variable with mean Y  t and variance  Y2  t . As we all know, the use of conjugate
prior allows all the results in Bayesian analysis to be derived in closed form and it is preferred
P

in engineering application especially in the condition of plentiful history data, we assume the
CE

prior distribution of   (Y , D) , denoted by  (Y , D) , be Normal-Inverse Gamma (N-IG)

distribution, which is a natural conjugate prior for the normal distribution and is very
flexible[32].  (Y , D) is
AC

 (Y , D)  D 1/ 2 e ( Y  0 ) /( 20 D )

 D ( 0 1) e 0 / D .
2
(13)
We need to estimate the parameters (0 , 0 ,  0 ,  0 ) in  (Y , D) through offline data of Y(t).

In fact, we could get the MLE of Y and D  Y2 for each unit tested in the lab with its offline
data of Y(t) alone by equation (11), denoted by ̂ i and D̂i , i  1, 2, , n . As D is an Inverse
Gamma distributed random variable, we could get MLE of ( 0 ,  0 ) by solving the following
likelihood equations
n n n
ln  0  (  0 )  ln(( (1/ Di ) / n) /( (1/ Dˆ i ))1/ n ) ,  0   0 n /  (1/ Di ) (14)
i 1 i 1 i 1

where  ( )  ' ( ) / ( ) and (  ) is the gamma function. At the same time, considering
that Y ~ N ( ,D) , we write the log-likelihood function of (v0 ,0 ) as

14 / 31
 ACCEPTED MANUSCRIPT

n
ln L( 0 , 0 )  n ln( 2 ) / 2   Dˆ i / 2  n ln  0 / 2
i 1
n n n
(15)
 (1 /( 2 0 ))( ( ˆ / Dˆ i )  2 0  ( ˆ i / Dˆ i )   02  (1 / Dˆ i ))
i
2

i 1 i 1 i 1

Then we can get MLE of (v0 ,0 ) as

T
IP
n n n n n
v0   ( ˆ i / Dˆ i ) /  (1/ Dˆ i ) ,  0  ( ( ˆ i2 / Dˆ i )  ( ˆ i / Dˆ i ) 2 /  (1/ Dˆ i )) / n (16)
i 1 i 1 i 1 i 1 i 1

R
Step 2. Calculation of posterior distribution  (Y , D | DataY ) .

SC
We could get the posterior distribution  (Y , D | DataY ) with the following theorem 1.
Theorem 1 Assume that the online monitoring data is updated from Yti  { yt1 , yt2 ,, yti } to

NU

Yti m  Yti  Ynew  Yti  { yti 1 , , yti  m } and that the prior distribution  (Y , D) is given in
equation (16). Then the posterior distribution denoted by  (Y , D | Yti  m ) is still an N-IG
MA
distribution with updated parameters (1 , 1 ,1 , 1 ) , denoted by
 (Y , D | Yt )  D 1/ 2 e ( 
im
Y v1 ) 2 /( 21D )
 D ( 1 1) e 1 / D (17)
D

where,
1  0 /(1  0  m  t ) , v1  (v0 0  m  yi ) /(1 0  m  t ) , (17a)
TE

1   0  mu /( 2  t )  m(yi  v0  t ) 2 /( 2  t  (1  0  m  t )) , 1   0  m / 2 , (17b)
and
P

m is the number of observations in Ynew .

CE

t is the constant measurement interval of Ynew : t  ti  j  ti  j 1 , j  1,2,, m .

The proof of Theorem 1 is detailed as follows.
AC

1
Let b  Y  t , Db  D  t and b  . The prior distribution of (Y , D) is
t

 (Y , D)  D 1/ 2  exp((Y  v0 ) 2 / 20 D)  D (  1)  exp( 0 / D)0

According to the property of random vector function, we could get the distribution of
(b , Db ) as
 (b , Db ) | J | (b Db ) 1/ 2  exp( (b b  v0 ) 2 /( 20 b Db ))  (b Db ) (  1)  exp(  0 /(b Db )) 0

b 0
where, J is the Jacobi matrix and | J |  b2 is a constant. Thus, we have
0 b
 (b , Db )  Db 1/ 2  exp((b  t  v0 ) 2 /( 20  t  Db ))  Db (  1)  exp( 0  t / Db ) (18) 0

Let y j  yti  j  yti  j 1 , j  1,2,, m , then y  {y1 , y2 ,, ym } are independent and identical
distributed samples of normal distribution N (b , Db ) , with sample mean y and sample
variance u as
15 / 31
 ACCEPTED MANUSCRIPT

m m
y   y j / m , u   (y j  y ) 2 / m
j 1 j 1

Considering that random vector (y, u) is a sufficient statistic of ( b , Db ) , We could get the
posterior distribution of ( b , Db ) , denoted by  (b , Db | y) , by Bayesian law
 (b , Db | y)   (b , Db | y, u)  f (y, u | b , Db )   (b , Db )

T
(19)
where, f (y, u | b , Db ) is the likelihood function of (y, u) given parameters ( b , Db ) .

IP
As we know that y ~ N (b , Db / m) , that is,

R
f (y | b , Db )  Db1/ 2  exp( m  (y  b ) 2 / 2Db ) .

SC
Also the pdf of u given parameters ( b , Db ) is
f (u | b , Db )  u ( m2) / 2  exp(m  u /( 2Db ))  Db( m1) / 2 , u  0 .

NU
The variables y and u are independent, thus we have
f (y , u | b , Db )  f (y | b , Db )  f (u | b , Db )
 Db1/ 2  exp( m  (y  b ) 2 / 2 Db )  u ( m2) / 2  exp( m  u /( 2 Db ))  Db( m1) / 2
MA
(20)
By substituting Equation (18) and (20) into (19), we arrived at the distribution
 (b , Db | y, u)  Db1/ 2  exp((b  1* ) 2 /( 21*  Db ))  D (  1)  exp(1* / Db ) ,
*
D

where,
TE

1*  (0  t ) /(1  m 0  t ) , 1*  (0  0  m  y ) /(1  m 0  t ) ,

1*   0  t  m  u / 2  m(y  0 t ) 2 / 2(1  m 0  t ) , 1*   0  m / 2 .
P

We again took the random vector (Y , D) as the function of the vector (b , Db ) , we could get
CE

the posterior distribution of (Y , D) from  (b , Db | y, u) with the Jacobi matrix. This proves
Theorem 1.
AC

Note that without loss of generality, we assume the observed time intervals of Ynew to be
identical. If the time intervals are not all the same, we could partition Ynew so that that each
group has the same time interval. Theorem 1 could be used recursively as long as we know
the original prior value of parameters (0 , 0 ,  0 ,  0 ) based on offline data.
Note that yi and u denote the sample mean and sample variance of newly observed
degradation increment of Y(t), that is,
im im
yi   ( yt j  yt j 1 ) / m , u 
j i 1
[( y
j i 1
tj  yt j 1 )  yi ]2 / m .

Step 3 Posterior estimation of Y and D.

For estimating Y and D, we need the marginal posteriors from joint distribution
 (Y , D | Yt ) ,
i m


 (Y | Yt )    (Y , D | Yt )dD ,
im im
0

16 / 31
 ACCEPTED MANUSCRIPT


 ( D | Yt )    (Y , D | Yt )dY .
im im
0

If the quadratic loss function is applied, the Bayesian estimation of Y and  Y2  D

conditional on observation history Ytim are the mean of  (Y | Ytim ) and  ( D | Ytim ) ,
respectively:

T
  
ˆY | Yt   Y   (Y | Yti  m )d Y    Y  (Y , D | Yt )dDd Y (21)

IP
im 0 0 0 im

  
ˆ Y2 | Yt   D   ( D | Yti  m )dD    D   (Y , D | Yti  m )dY dD (22)

R
im 0 0 0

The integrals in Equations (21) and (22) can be solved with numerical integral techniques

SC
as multiple trapezoidal integration method[34] using Matlab as soon as we get the specific
mathematical form of  (Y , D | Ytim ) with Theorem 1.

6. RUL estimation of EV lithium-ion batteries

NU
MA
6.1. Battery Database

The case study in this section was based on the charge-discharge test data of four 18650
D

lithium-ion batteries introduced in Section 3.2. To demonstrate the effectiveness of our

method, we divided the four batteries into two parts as (1) field monitoring condition (#18)
TE

and (2)lab test condition (#5, #6, #7). The data collection plan of the case study is listed in
Table 1. The effectiveness of our method is demonstrated with test data of aforementioned
P

four batteries. Actually, we use three of them as training data and another one as test data.
From the following two reasons, we could say that the data of four batteries can verify the
CE

applicability of our method. One reason is that our method is based on bivariate wiener
process, the degradation increment data is the basis for model parameter estimation. The ith
AC

degradation increment is calculated from the difference between the marker data of the ith
cycle and the marker of the (i+1)th cycle. Every battery has undergone more than 100 cycles.
Thus, the number of increment data is large enough to estimate the parameters of bivariate
wiener process. On the other hand, the results of parameter estimation from three batteries are
used as prior information in our method, it will timely update with online data of the new
model. As we use Bayesian analysis to realize the update of model parameter, the influence of
prior information will become weaker and weaker if the performance of the offline batteries
and online batteries are differ from each other drastically.
Table 1 Data collection for NASA case study
Offline (lab, #5, #6, #7) pseudo online
(field, #18)
Data used for marker TVS,TCS TVS,TCS
Modeling and damage MRC /
estimation
Data used for Model MRC
verification
17 / 31
 ACCEPTED MANUSCRIPT

Note that the accurate of RUL estimation is close related with afore working time. To
verify the efficiency of our method, we arbitrarily chose one battery from the four batteries
and arbitrarily chose three different times to get RUL. Hereafter in this section, our aim is to
obtain the RUL distribution f t0 (t | Yt0 ) of battery #18 at three different times (cycles)t0=t1=40,

T
t0=t2=85, and t0=t3=120. Choose three different times

IP
That is to say, the modeling and estimation process was based on the online data
(TVS,TCS of #18) up to ti (i=1,2,3) along with all the offline data(TVS,TCS and MRC of

R
#5,#6,#7). The efficiency of the model is verified based on MRC data of #18. The RUL
estimation process of the case study included five steps: (1)data preprocessing, (2)composite

SC
marker construction, (3)model specification, (4)RUL estimation, and (5)model verification.

6.2. Data preprocessing

NU
Data preprocessing is critical for guaranteeing reasonable and accurate prediction results.
The first step of preprocessing consists of scaling the data such that marker I(TVS), marker
MA
II(TCS) and damage(MRC) are scaled in the range of 0 to 1. By proper scaling, we prevent
the statistical model from being overly sensitive to changes in inputs (e.g., having some
elements dominate over the others and gaining too much influence regarding the accuracy of
the prediction model.) which assures robustness.
D

Let C(t) be the maximum releasable capacity (MRC) at cycle t, and C(0) be the MRC of the
TE

first cycle. The MRC process C(t) decreases with cycle t and can be scaled to X(t) as
X (t )  (C (0)  C (t )) / C (0) . (23)
P

The lifetime could be conveniently defined as the first time that X(t) hits 0.3 according to the
failure definition of lithium ion battery.
CE

Let M I (t ) be the time-to-current-saturation (TCS) of cycle t (the subscript I here refers to

current), and M I (0) be the TCS of the first cycle. The marker process M I (t ) increases in t
AC

and could be scaled as

YI (t )  (M I (t )  M I (0)) / M I (0) . (24)
Let M U (t ) designate the time-to-voltage-saturation (TVS) of cycle t (the subscript U here
refers to voltage), and M U (0) be the TVS of the first cycle. The marker process M U (t )
decreases with cycle t and could be scaled to YU (t ) as,
YU (t )  (M U (0)  M U (t )) / M U (0) . (25)
The original data and scaled data of MRC, TVS and TCS of the four batteries are described
in Figure 5, Figure 6, and Figure 7.

18 / 31
 ACCEPTED MANUSCRIPT

T
R IP
SC
NU
Figure 5. (a) Maximum releasable capacity (MRC, damage) change with cycle according to
MA
original data; (b) Maximum releasable capacity (MRC, damage) change with cycle according
to scaled data.
D
P TE
CE
AC

Figure 6. (a) Time-to-Current-Saturation(TCS, marker I) change with cycle according to

original data; (b) Time-to-Current-Saturation(TCS, marker I) change with cycle according to
scaled data.

19 / 31
 ACCEPTED MANUSCRIPT

T
R IP
SC
NU
Figure 7. (a) Time-to-Voltage-Saturation (TVS, marker II) change with cycle according to
original data; (b) Time-to-Voltage-Saturation (TVS, marker II) change with cycle according to
MA
scaled data.

6.3. Composite Marker

D

From the analysis in Section 3.2, we know that the decrease of TVS with cycles is due to
TE

the increase of the battery’s internal resistance with cycles. The increase of TCS with cycle is
due to the decrease of number of active Li+ in the battery. That is to say, the markers YI (t ) and
P

YU (t ) are both closely related to the demage level X(t) from the failure physical mechanism of
CE

lithium-ion battery. Decisions are based on a summary measure of performance which

corresponds to a function on the underlying process, YI (t ) and YU (t ) . C.T.Barker et al[35].
use the l2 norm to combine multiple wiener degradation variables into one composite
AC

degradation variable. In our study, the function used to summarize the state of the underlying
markers’ degradation process YI (t ) and YU (t ) , is the l2 norm Y(t).

Y (t )  (YI2 (t )  YU2 (t )) / 2 . (26)

Then, two markers TVS and TCS are integrated into one composite marker, denoted byY(t).
It can be easily seen that Y(t) still lies between 0 and 1. The composite marker for each
battery is depicted in Figure 8. It can be seen that Y(t) increase with cycle t. We need to
further verify that the composit marker has close correlation with damage process, which will
be given in the next section.

20 / 31
 ACCEPTED MANUSCRIPT

T
R IP
SC
Figure 8. Composite marker Y(t) (constructed by YI (t) and YU (t))change with cycle t

NU
6.4. Model specification

In order to judge whether the vector data {X (t ), Y (t ) : t  0} can be modeled with

MA
two-dimensional Wiener process, we analyzed the data in three ways: (1)test the correlation
between X(t) and Y(t); (2)test the degradation increment data of X(t) to be normal distributed;
(3)test the degradation increment data of Y(t) to be normal distributed.
The pearson correlation coefficients between X(t) and Y(t) for each battery were calculated
D

to be 0.9392, 0.9561, 0.8402 and 0.8426 from the scaled data. The correlation relationships
TE

were depicted in Figure 9. We could see that the linear correlation relationship between X(t)
and Y(t) is strong for each battery. This also verified the effectiveness of composite marker. In
P

other words, there are strong underlying correlation between the composite marker and
damage degradation process.
CE
AC

Figure 9. (a)The correlation relationship between composite marker Y(t) and damage X(t) for
battery #5; (b)The correlation relationship between composite marker Y(t) and damage X(t)

21 / 31
 ACCEPTED MANUSCRIPT

for battery #6; (c)The correlation relationship between composite marker Y(t) and damage X(t)
for battery #7; (d)The correlation relationship between composite marker Y(t) and damage
X(t) for battery #18

The normal probability paper test of degradation increment data X (t ) and Y (t ) for

T
three assumed offline batteries (#5, #6, #7) are ploted in Figure 10, respectively. The purpose
of a normal probability plot is to graphically assess whether the data in X could come from a

IP
normal distribution. If the data are normal the plot will be linear. From figure 10, we could

R
see that the increment data are both fit for normal distribution(here, the time intervals for the
degradation increments are all the same and be t  1).

SC
NU
MA
D
P TE

Figure 10. (a) Normal Paper test of degradation increment data X (t ) (data of batteries #5,
CE

#6, #7); (b) Histogram of degradation increment data Y (t ) (data of batteries #5, #6, #7)
AC

From above discussion, we could see that the process X(t) has close relationship with Y(t),
and both of them could be modeled by Wiener process.

6.5. RUL estimation

The RUL of battery #18 at times t1=40, t2=85 and t3=120 were estimated with the following
three steps.

Step1 Estimation of population parameters with MLE based on offline data.

We obtained the MLE of parameters  X , X2 ,  XY based on the offline data
{X (t ), Y (t ) : t  0} of batteries #5, #6, #7 by Equation (10),(11)and(12). Then we got MLE of
Y ,  Y2 based on each battery’s data {Y (t ) : t  0} by Equation (11), that is, ˆ Y ,ˆ Y2 , i  1,2,3 .
i i

These estimations of Y ,  Y will be used in the second step to specify the hyper-parameters
2

v0 ,0 , 0 ,  0 in the N-IG prior distribution. All the MLE results were listed in Table 2.

22 / 31
 ACCEPTED MANUSCRIPT

Table 2 Maximum Likelihood Estimation (MLE) results of parameters in the degradation

model
ML
E ˆ X ˆ X2  XY ˆY ˆY2
Data

T
#5, #6, #7 0.0019 7.4508e-5 0.8945
#5 0.0024 0.0020

IP
#6 0.0035 0.0029

R
#7 0.0018 0.0030

SC
Step2 Update of individual characteristics based on online data with Bayesian method
We got the estimation of the hyper-parameters v0 ,0 ,  0 ,  0 in the N-IG prior distribution

NU

 (Y , D) based on the MLE ̂Yi and Dˆ i  ̂Yi ( i  1,2,3 ) that have been obtained in step 1,
according to Equation (14)and (16).
Let vˆi ,ˆi , ˆ i , ˆi be the posterior estimation of the hyper-parameters in the N-IG posterior
MA
distribution  (Y , D | Yti ) , then vˆi ,ˆi , ˆ i , ˆi could be sequentially estimated based on
Theorem 1, i  1,2,3 . The probability density function (pdf) of the N-IG prior distribution
D

and the N-IG posterior distribution function at t = t1,t = t2 and t = t3were plotted in Figure 11.
TE

We could see that the prior and posterior distributions change with the updated
hyper-parameters estimated in different times, but all the pdfs have the same shape. The
Bayesian estimation of Y and  Y at time t i can be obtained based on the N-IG posterior
2
P

distribution  (Y , D | Yti ) according to Equation (18) and (19). The Baysian estimation of
CE

 X |Y and  X2 |Y conditional on Y  Yt could be obtained from Equation (5a) and (5b). All the
i

results obtained in step 2 were listed in Table 3 according to their corresponding time.
AC

23 / 31
 ACCEPTED MANUSCRIPT

T
R IP
SC
NU
MA
Figure 11. (a) Prior Normal-InverseGamma(N-IG) distribution; (b) Posterior N-IG
distribution at time t1; (c) Posterior N-IG distribution at time t2; (d) Posterior N-IG
D

distribution at time t3.

TE

Table 3 Bayesian estimation results

t v̂ ̂ ̂ ˆ ˆY ˆY2 ˆ X |Y ˆ X2 |Y
P

t=0 0.0025 1.8441e-4 13.3185 29.5008

CE

t=t1 0.0026 1.8323e-4 13.3291 47.0008 3.0175e-14 0.5368 0.0768 5.9548e-4

t=t2 0.0026 1.8173e-4 13.3487 69.5008 2.6270e-20 0.4406 0.1629 0.0013
t=t3 0.0026 1.8058e-4 13.3730 87.0008 2.8340e-25 0.3938 0.2294 0.0018
AC

Step3 RUL distribution based on online updated marker data

The probability density function (pdf) of the RUL at time t1, t2 and t3 could be obtained
according to Equation (6) and the cumulative distribution function (cdf) of RUL could be
obtained according to definite integral of the pdf from 0 to t. The pdf and cdf of RUL at
different times are depicted in Figure 12. Let ŝti , sˆ(1 / 2)ti , sˆ( / 2)ti and sti be the point
estimate, the lower bound estimate, the upper bound estimate, and true value of RUL at cycle
t i , i  1,2,3 . The estimations are listed in Table 4.

24 / 31
 ACCEPTED MANUSCRIPT

T
R IP
SC
NU
MA
Figure 12. (a)The probability density function(pdf) of RUL at different times; (b)The
cumulative probability function (cdf) of RUL at different times
D

Table 4 Point, Confidence Interval Estimation and True Value of RUL at ti

sˆti sˆ(1 /2)|ti sˆ( /2)|ti
TE

sti
t1=40 95 23 118 94
t2=85 28 5 55 49
P

t3=120 11 5 24 14
CE

Note: The true lifetime of battery #18 is 134; the unit of all the numbers is cycle.

6.6. Model verification

AC

To verify the accuracy of our model, we first estimated the value of X(t) based on our model
with the observed marker data (TCS and TVS) of battery #18. Then the estimation of X1(t) is
compared with the real observed MRC X2(t) of battery #18. We want to see whether X2(t) be
close to X1(t) or X2(t) lies within the confidence interval of X1(t).
For any ti  0 , X (ti ) is a normal distributed random variable,
X (ti ) ~ N ( X ti     X ( yti  Y ti ) /  Y , (1   2 ) X2 ti ) , i  1,2, . (27)
Here,  X , X and  are estimated with the offline data{X(t),Y(t)} of batteries #5, #6, #7 by
MLE; Y , Y are estimated from the online data {Y (t ),0  t  ti } of battery #18 by Bayesian
methods.
The point estimation of X (ti ) , denoted by Xˆ (ti ) , is
Xˆ (ti )   X ti     X ( yti  Y ti ) /  Y , i  1,2, (28)

The (1   )% confidence interval (CI) of X (ti ) , denoted by [ Xˆ L (ti ), Xˆ U (ti )] , is

25 / 31
 ACCEPTED MANUSCRIPT

Xˆ L (ti )   X ti     X ( yti  Y ti ) /  Y  z / 2 (1   2 ) X2 ti ,
Xˆ U (ti )   X ti     X ( yti  Y ti ) /  Y  z / 2 (1   2 ) X2 ti , i  1,2, (29)
where z / 2 is the upper  / 2 percentile of the standard normal distribution.

T
The point estimation curve and CI estimation curves of X (ti ) for 1-  =0.95 were plotted

IP
in Figure 13. We could see that the actual values of X (ti ) were contained in the estimated
interval at almost all the times.

R
SC
NU
MA
D
P TE

Figure 13. Observed values along with point estimates and 95% confidence intervals for X(t)
CE

7. Conclusion and future research

AC

For lithium-ion batteries, maximum releasable capacity (MRC) can only be derived with
some special offline tests. So it is hard to conduct RUL prediction methods depending on
capacity measurement hard to apply since operation interruptions and additional equipment
are out of the question during ordinary EV usage. This paper presents a RUL prediction
method based on the damage-marker bivariate degradation model, as well as its brief
application background, marker extraction method, model description and development
method, parameter estimation method, and model verification method, based on the test data
provided by the Prognostics CoE at NASA Ames. We obtained the probability density
function of RUL and displayed the RUL prediction results (point estimation, interval
estimation).
The proposed method could be used to predict RUL online for lithium-ion battery packs
using online marker data (TCS and TVS) without extra capacity measurements. In order to
apply our proposed methods to practical applications, an amount of offline data is needed.
The offline data include damage data (MRC) and the corresponding marker data (TCS and
TVS). These data can be easily sampled, since the charge and discharge performance of
26 / 31
 ACCEPTED MANUSCRIPT

batteries are typically tested before they are put into use. Another prerequisite for the method
is that the damage degradation process and the marker degradation process are both Wiener
processes. Thus, our damage-marker model and RUL prediction model are both developed
based on two-dimensional Wiener process.
At present, the online RUL prediction based on the damage-marker degradation model is
increasingly prevalent in degradation tests of highly complex devices, because the real

T
damage process is usually latent and hard to monitor online. We can usually observe the

IP
damage and corresponding marker in laboratory tests, which makes it possible to model the
relationship between the marker and cumulative damage. In summary, there are two

R
promising directions for our future research. One is to extend the RUL prediction method

SC
based on other distributions (e.g. Gamma process etc.) in case the Wiener process does not
adequately model the damage/marker process exhibited by the data. The other one is thatThe
charge activity of lithium-ion batteries is influenced by temperature. However, in this paper,

NU
we assume that the environment temperatures of lab test and field use are almost the same.
Therefore, for another direction, we will consider the impact of temperature on the marker
and damage with accelerated function (e.g. Arrhenius Law, etc.).
MA
Acknowledgements
D

This work is supported by the National Natural Science Foundation of China (Grant No.
61273041).
TE

References
P

[1] N. Williard, W. He, M. Osterman, and M.Pecht, “Comparative Analysis of Features for
CE

Determining State of Health in Lithium-Ion Batteries,” Int.J.Progn.Heal.Manag., Vol.4,

pp.1-7,2013.
[2] V. Klass, M. Behm, G. Lindbergh. Evaluating real-life performance of lithium-ion battery
AC

packs in electric vehicles. Journal of Electrochemical Society, 159(11), 2012: 1856-1860.

[3] D. Liu, J. Zhou, H. Liao, Y. Peng, and X. Peng. A Health Indicator Extraction and
Optimization Framework for Lithium-Ion Battery Degradation Modeling and Prognostics.
IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2015: 2168-2216 DOI:
10.1109/TSMC.2015.2389757.
[4] M. Broussely, Ph. Biensan, F. Bonhomme, Ph. Blanchard, S. Herreyre, K. Nechev, R.J.
Staniewicz. Main aging mechanisms in Li ion batteries. Journal of Power Sources 146,
2005: 90-96.
[5] M. B. Pinson, M. Z. Bazant. Theory of SEI formation in rechargeable batteries: capacity
fade, accelerated aging and lifetime prediction. Journal of Electrochemical Society,
160(2), 2013: A243-A250. (physical-based models)
[6] M. Safari, M. Morcrette, A. Teyssot, and C. Delacourt. Life-prediction methods for
lithium-ion batteries derived from a fatigue approach. Journal of The Electrochemical
Society, 157(6), 2010: 713-720. (physical- empirical-based models)
[7] K.Smith, J.Neubauer, E.Wood, M.Jun, A.Pesaran. Models for battery reliability and
27 / 31
 ACCEPTED MANUSCRIPT

lifetime, applications in design and health management, Center for Transportation

Technologies and Systems, NERL/PR-5400-58550, Battery Congress, April 15-16, 2013
Ann Arbor, Michigan. (empirical-based models)
[8] B.Saha, K.Goebel, and J. Christophersen, Comparison of prognostics algorithms for
estimating remaining useful life of batteries. IEEE Transaction on Inst. Meas. Control,
vol.31, no.3-4, pp,293-308, Jun.2009. (ARMA)

T
[9] J.Liu, A.Saxena, K.Goebel, B.Saha, and W. Wang. An adaptative recurrent neural network

IP
for remaining useful life prediction of lithium-ion batteries, Proc. Annu. Conf.
Prognostics Health Manage. Soc., Protland, OR, USA, 2010, pp.1-9 (NN)

R
[10] J.I.Park, M.K.Jeong. Recursive support vector censored regression for monitoring

SC
product quality based on degradation profiles. Applied Intelligence 35, 2011:63-74.
[11] J. Zhou, D.Liu, Y. Peng, and X.Peng, Combined sparse Bayesian learning strategy for
remaining useful life forecasting of lithium-ion batteries. IEEE Trans. Ind. Electron.,

NU
vol.60, no.11, pp.5260-5269, Nov.2013.
[12] B.Saha, K.Goebel, S.Poll, and J.Christophersen, An integrated approach to battery health
monitoring using Bayesian regression and state estimation, Proc. IEEE Autotestcon,
MA
Baltimore, MD, USA, 2007, pp. 646-653.
[13] B.E.Olivares, M.A.C.Munoz, M.E.Orchard, and J.F.Silva, Particle-filtering-based
prognosis framework for energy storage devices with a statistical characterization of
D

state-of-health regeneration phenomena, IEEE Trans Instrum. Meas., vol.62, no.2,

pp.364-376, Feb.2013.
TE

[14] M.Dalal, J.Ma, and D.He, Lithium-ion battery life prognostic health management system
using particle filtering framework, Proc. Inst. Mech. Eng. Part O J. Risk Reliab. vol.225,
P

no.1, pp.81-90, Mar.2011.

CE

[15] J.Liu, W. Wang, and F.Ma, A regularized auxiliary particle filtering approach for system
state estimation and battery life prediction, Smart Mater, Struct., vol.20,no.7, pp.1-9,
Jul.2011.
AC

[16] Q.Miao, L.Xie, H.Cui, W.Liang, and M.Pecht, Remaining useful life prediction of
lithium-ion battery with unscented particle filter technique, Microelectron. Reliab., Vol.53,
no.6, pp.805-810, Jun.2013.
[17] D.Liu, J.Pang, J.Zhou, Y.Peng, and M.Pecht, Prognostics for state of health estimation of
lithium-ion batteries based on combination Gaussian process function regression,
Microelectron, Reliab., Vol.53, no.6, pp.832-839, Jun.2013.
[18] J.L.Zhang, J.Lee. A review on prognostics and health monitoring of Li-ion battery.
Journal of Power Sources, 196, 2011: 6007-6014.
[19] J.Kozlowski, Electrochemical cell prognostics using online impedance measurements and
model-based data fusion techniques, in Proc. IEEE Aerosp. Conf. vol.7, Big Sky, MT,
USA, 2003, pp.3257-3270.
[20] Y.Xing, E.W.M.Ma, K.L.Tsui, and M.Pecht, An ensemble model for predicting the
remaining useful performance of lithium-ion batteries, Microelectron. Reliab., vol.53,
no.6, pp. 811-820, Jun.2013.
[21] C.Hu, B.D.Youn, P.Wang, and J.T.Yoon, Ensemble of data-driven prognostic algorithms
for robust prediction of remaining useful life, Reliab. Eng. Syst. Safety, vol.103,
28 / 31
 ACCEPTED MANUSCRIPT

pp.120-135, Jul.2012.
[22] D.Liu, H.Wang, Y.Peng, W.Xie, and H.Liao. ”Satellite lithium-ion battery remaining
cycle life prediction with novel indirect health indicator extraction,’’ Energies, vol.6, no.8,
pp.3654-3668, 2013.
[23] A. Widodo, M. –C. Shim, W.Caesarendra, and B.-S.Yang, “Intelligent prognostics for
battery health monitoring based on sample entropy,” Expert Syst. Appl., Vol.38, no.9, pp.

T
11763-11769, Sep.2011.

IP
[24] Battery test manual for plug-in hybrid electric vehicle, September 2010.
[25] N.P.Jewell, J.D.Kalbfleisch. Marker processes in survival analysis. Lifetime Data

R
Analysis, 2, 1996: 15-29.

SC
[26] N.D.Singpurwalla. On competing risk and degradation processes. 2nd Lehmann
Symposium-Optimality, 49, 2006: 229-240.
[27] G.A.Whitmore, M.J.Crowder, J.F.Jawless. Failure inference from a marker process based

NU
on a bivariate Wiener model. Lifetime Data Analysis, 4, 1998: 229-251.
[28] M.T.Lee, M.Shubina, and A.Zaslavsky. Bayesian analysis for markers and
degradation,http://conferences.telecom-bretagne.eu/asmda2005/IMG/pdf/proceedings/
MA
1262.pdf.
[29] Lithium ion rechargeable batteries. Technical Handbook of Sony, Osaki West Technology
Center.
D

[30] Data "B. Saha and K. Goebel (2007). "Battery Data Set", NASA Ames Prognostics Data
Repository, [http://ti.arc.nasa.gov/project/prognostic-data-repository], NASA Ames,
TE

Moffett Field, CA".

[31] S.Ghahramani. Fundamentals of probability with stochastic process(3rd edition), Pearson
P

Education inc., 2005.

CE

[32] R.S.Chhikara and J.L.Folks. The inverse Gaussian distribution: theory, methodology, and
applications. Marcel Dekker, 1989.
[33] A.Gelman, J.B.Carlin, H.S.Stern and D.B.Rubin. Baysian Data Analysis(second edition),
AC

Chapman & Hall/CRC, 2004.

[34] W.Gautschi. Numerical analysis (2nd edition). Springer New York Dordrecht Heidelberg
London, 2012.
[35] C.T.Barker, M.J.Newby, Optimal non-inspection for a multivariate degradation model,
Reliability Engineering and System Safety, 94, 2009:33-43.

Authors’ biographies
Jing Feng is Associate Professor in the College of Information System & Management at
National University of Defense Technology (NUDT), China. She worked at Georgia Institute
of Technology as a visiting scholar from March 2013 to March 2014. She received her B.S.
and M.S. degrees in Mathematics from NUDT at 1998 and 2001, respectively, and Ph.D.
degree in Systems Engineering from NUDT at 2004. Her research interests are in quality &
reliability engineering.

Paul H. Kvam is Professor in the School of Industrial and Systems Engineering at Georgia
Institute of Technology. He received his B.S. degree in Mathematics from Iowa State
29 / 31
 ACCEPTED MANUSCRIPT

University in 1984, M.S. degree in Statistics from the University of Florida in 1986, and Ph.D.
degree in Statistics from the University of California, Davis in 1991. His research interests
focus on statistical reliability with applications to engineering, nonparametric estimation, and
analysis of complex and dependent systems. He has served as an associate editor for IEEE
Transactions on Reliability (1992–2000), Technometrics (1999–2005), The American
Statistician (2005–2013) and Journal of the American Statistical Association (2002–2012).

T
He is a Fellow of the American Statistical Association, as well as an active member of

IP
Institute of Mathematical Statistics and Institute for Operations Research and Management
Science.

R
SC
Yanzhen Tang is a Ph.D. student in the College of Information System & Management at
National University of Defense Technology (NUDT), China. He studied at Arizona State
University as a visiting Ph.D. student from December 2013 to December 2015. He received

NU
his B.Sc. and M.Sc. degrees in Control Science and Engineering from NUDT at 1998 and
2001, respectively. His research interests are in system safety, risk assessment and reliability
engineering.
MA
D
P TE
CE
AC

30 / 31
 ACCEPTED MANUSCRIPT

Highlights

1、Two types of Health indices (HIs) are proposed based on CC/CV charging curve.

2、Damage-Marker bivariate degradation model is proposed for RUL estimation online.

T
IP
3、The composite marker constructed on l2-norm has close relation with Damage(MRC).

R
4、Bayesian parameter estimation method for the bivariate degradation model is given.

SC
5、Bivariate Wiener process is fit to describe degradation of the marker and MRC.

NU
MA
D
P TE
CE
AC

31 / 31