See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/235350141

System Identification and Estimation Framework for Pivotal Automotive

Battery Management System Characteristics

Article in IEEE Transactions on Systems Man and Cybernetics Part C (Applications and Reviews) · July 2011

CITATIONS
READS
150
738

3 authors:

Bharath Pattipati
Chaitanya Sankavaram
University of Connecticut
University of Connecticut
26 PUBLICATIONS 803
CITATIONS 24 PUBLICATIONS 729 CITATIONS

Krishna Pattipati
University of Connecticut
168 PUBLICATIONS 3,301 CITATIONS

Some of the authors of this publication are also working on these related projects:

Naval Meteorology and Oceanography Command Project View project

CyberSEES: Type 2: Fault Detection, Diagnosis and Prognosis of HVAC systems View project

All content following this page was uploaded by Bharath Pattipati on 20 May 2014.

The user has requested enhancement of the downloaded file.

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS
System Identification and Estimation Framework
for Pivotal Automotive Battery Management
System Characteristics
Bharath Pattipati, Chaitanya Sankavaram, and Krishna R. Pattipati, Fellow, IEEE
Abstract—The battery management system (BMS) is an
integral part of an automobile. It protects the battery from
damage, pre- dicts battery life, and maintains the battery in an
operational con- dition. The BMS performs these tasks by
integrating one or more of the functions, such as protecting the
cell, thermal management, controlling the charge–discharge,
determining the state of charge (SOC), state of health (SOH), and
remaining useful life (RUL) of the battery, cell balancing, data
acquisition, communication with on-board and off-board
modules, as well as monitoring and stor- ing historical data. In
this paper, we propose a BMS that estimates the critical
characteristics of the battery (such as SOC, SOH, and RUL)
using a data-driven approach. Our estimation procedure is based
on a modified Randles circuit model consisting of resistors, a
capacitor, the Warburg impedance for electrochemical
impedance spectroscopy test data, and a lumped parameter
model for hybrid pulse power characterization test data. The
resistors in a Randles circuit model usually characterize the self-
discharge and internal resistance of the battery, the capacitor
generally represents the charge stored in the battery, and the
Warburg impedance repre- sents the diffusion phenomenon. The
Randles circuit parameters are estimated using a frequency-
selective nonlinear least squares estimation technique, while the
lumped parameter model param- eters are estimated by the
prediction error minimization method. We investigate the use of
support vector machines (SVMs) to pre- dict the capacity fade
and power fade, which characterize the SOH of a battery, as well
as estimate the SOC of the battery. An al- ternate procedure for
estimating the power fade and energy fade from low-current
Hybrid Pulse Power characterization (L-HPPC) test data using
the lumped parameter battery model has been pro- posed.
Predictions of RUL of the battery are obtained by support vector
regression of the power fade and capacity fade estimates.
Survival function estimates for reliability analysis of the battery
are obtained using a hidden Markov model (HMM) trained us-
ing time-dependent estimates of capacity fade and power fade as
observations. The proposed framework provides a systematic
way for estimating relevant battery characteristics with a high-
degree of accuracy.
Manuscript received September 27, 2010; accepted October 10, 2010. This
work was supported in part by Toyota Technical Center, Ann Arbor, MI
48105 and in part by the National Science Foundation (Cyber-Physical
systems) under Grant ECCS-0931956. A preliminary version of the paper was
published in the Proceedings of the IEEE AUTOTESTCON as “Automotive
Battery Management Systems” by B. Pattipati, K. Pattipati, J. P.
Christopherson, S. M. Namburu, D.
V. Prokhorov, and L. Qiao, pp. 581–586, Sep. 2008. This paper was recom-
mended by Associate Editor R. Subbu.
The authors are with Electrical and Computer Engineering Department,
Uni- versity of Connecticut, Storrs, CT 06269-2157 USA (e-mail:
bharath@engr. uconn.edu; chaitanya@engr.uconn.edu;
krishna@engr.uconn.edu).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSMCC.2010.2089979
1
Index Terms—Battery management system (BMS), hidden
Markov model (HMM), remaining useful life (RUL), state of
charge (SOC), state of health (SOH), support vector machines
(SVMs).
NOMENCLATURE
A Transition matrix.
B Emission matrix.
Cdl Double layer capacitance.
N Number of states of hidden Markov Model (HMM).
Q Generator matrix.
S Reliability (survival) function.
RHF High-frequency resistance.
Rtc Charge-transfer resistance.
s Hidden states.
X Observations.
ZW Warburg impedance.
π State probability vector.
I. INTRODUCTION
B
ATTERY technology has come a long way since the in-
vention of the first voltaic cell in the 1800s. Because of
the increased interest in hybrid vehicles, a battery
management system (BMS) has become one of the chief
components in an automobile. The goals of BMS are to
maximize the run-time per discharge cycle, as well as the
number of life cycles attainable for the life of the battery [1].
Automotive battery management is very demanding, because
it has to work in real time in rapidly varying charge–discharge
conditions as the vehicle accelerates and brakes, as well as
work in a harsh and uncontrolled environ- ment. In addition, it
must interface with other on-board systems, such as the engine
management, climate controls, communica- tions, and safety
systems.
The function of a BMS in hybrid electric vehicles (HEVs)
and plug-in hybrids (PHEVs) is multifaceted. They include
moni- toring the conditions of individual cells that make up
the battery, maintaining all the cells within their operating
limits, protect- ing the cells from out-of-tolerance conditions,
compensating for any imbalances in cell parameters within the
battery chain, pro- viding information about the State of
Charge (SOC), State of Health (SOH), and Remaining Useful
Life (RUL) of the bat- tery, providing the optimum charging
algorithm, responding to changes in the vehicle operating
mode, and so on.
The main motivation for this paper is to develop a system-
atic procedure for estimating three critical characteristics of a
battery, namely, SOC, SOH, and RUL. Important questions
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

such
1094-6977/$26.00 © 2010 IEEE
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

2 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS

as how long a battery lasts before it fails, how much charge is TABLE I
SUMMARY OF DIFFERENT TECHNIQUES FOR THE DETERMINATION OF SOC
left in the battery, performance of a battery to complete a cer- ALONG WITH THEIR APPLICATIONS, ADVANTAGES, AND DRAWBACKS [3],
tain task, how much time a battery is charged before use, the [4]
different degradation processes affecting the battery, and so
on, are different paradigms encapsulated in these critical
battery characteristics. Knowing the amount of charge left in a
battery compared to when it was fully charged gives the user
an indi- cation of how much longer a battery will continue to
perform before it needs recharging. The SOC, a measure of
remaining capacity in the battery, is used to ensure optimum
control of the charging/discharging process. During the
lifetime of a battery, its “health” deteriorates gradually due to
irreversible physical and chemical changes that take place
with usage and age until eventually the battery is no longer
usable or dead. The SOH is an indication of the point that has
been reached in the life cycle of the battery and a measure of
its condition relative to a fresh battery. The SOH is a measure
of a battery’s capabil- ity to deliver its specified output. This is
vital for assessing the readiness of emergency power
equipment, and is an indicator of whether maintenance actions
are needed. We will employ capacity fade and power fade as
measures of SOH of a battery. The ability to accurately predict
the RUL is the key to proactive, condition-based maintenance
of batteries.

A. Previous Research
1) Critical Battery Characteristics and Their Estimation:
Key factors affecting the SOC are the charge–discharge rates,
hysteresis, temperature, cell age, and self-discharge due to
inter- nal resistance of the battery [2]. SOC is not directly
measurable and needs to be inferred from other measurements
in the battery system. Several methods for estimating the SOC
of a battery have been used. Some are specific to particular
cell chemistries. Most depend on measuring some convenient
parameter, which varies with the SOC. Some of the prominent
techniques for estimating the SOC are summarized in Table I.
Pop et al. [5] proposed a real-time SOC evaluation system
for Li-ion batteries combining direct measurement of
electromotive force (EMF) during the equilibrium state and
Coulomb counting during the charge–discharge states.
Kim et al. [6] implemented a sliding-mode observer for ro-
bust tracking under nonlinear conditions based on a simple
RC battery model. The performance was validated by the
urban dy- namometer driving schedule test and the SOC error tance, voltage, and self-discharge. Hence, knowing changes in
was less than 3% for most cases. In a more recent work [7], he the SOH with time may be viewed as enabling one to assess
proposed a dual sliding-mode observer consisting of a fast the increase in irreversible losses that is inherent in the aging
time-varying observer to estimate the SOC, terminal voltage, of batteries [8].
and polarization effects and a slow time-varying observer to In practice, SOH is estimated from a single measurement of
estimate the SOH in terms of capacity fade and resistance either the cell impedance or the cell conductance. In pursuit of
deterioration. accuracy, others advocate measuring several cell parameters,
Determining the SOH of a battery adaptively is important all of which vary with the age of the battery, and estimate
for optimal energy management and on-board diagnostics for SOH from a combination of these factors. Examples are
effi- cient operation of HEVs and PHEVs. Generally, SOH is capacity, in- ternal resistance, self-discharge, charge
used to deduce how well the battery system is functioning acceptance, discharge capabilities, the mobility of electrolyte,
relative to its nominal (rated) and end (failed) states [8]. The and cycle counting, if possible. The absolute readings will
SOH reflects the general condition of a battery and its ability depend on the cell chem- istry involved. Weighting is added to
to deliver the specified performance in comparison with a new individual factors based on experience, the cell chemistry,
battery. It takes into account factors such as charge and the importance of the
acceptance, internal resis-
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS
particular parameter in the application for which the battery is
used. If any of these variables provide marginal readings, the
end result will be affected. A battery may have a good
capacity, but the internal resistance may be high. In this case,
the SOH estimation will be lowered accordingly. Similar
demerit points are added, if the battery has high self-discharge
or exhibits other chemical deficiencies. The points scored for
the cell are com- pared with the points assigned to a new cell
to give a percentage result or figure of merit.
An alternative method of specifying the SOH is to base the
estimation on the usage history of the battery rather than on
some measured parameter. The number of charge-discharge
cy- cles completed by the battery is an obvious measure, but
this does not necessarily take into account any extreme
operating conditions experienced by the battery, which may
have affected its functionality. It is, however, possible to
record the duration of any periods during which the battery
has been subject to abuse from out-of-tolerance voltages,
currents or temperatures, as well as the magnitude of the
deviations. From this, the SOH can be determined by using a
weighted average of the measured parameters.
Bhangu et al. [9] employed extended Kalman filters to esti-
mate the variation in bulk capacitance, which characterizes the
ability of the battery to store charge over time, indicating the
loss of cell capacity and, consequently, a means for
monitoring SOH.
Do et. al. [10] proposed a real-time identification method
based on an extended Kalman filter for observation of the pa-
rameters of the modified Randles circuit battery model. They
suggested an increase in the number of parallel RC-circuits to
model the Warburg impedance and improve the low-frequency
performance. However, experimentation on Gen 2 Li-ion cell
data via the modified Randles scheme with more than one
paral- lel RC-circuit, leads to ill-conditioning (over-
parameterization) of the optimization problem. In addition,
extended Kalman fil- ters have divergence issues, problems of
initialization, and the covariance matrix tends to be too
optimistic.
Vebrugge et al. [11] concluded that a regressed high-
frequency resistance (from weighted recursive least-squares),
which must be normalized to nominal values and is a function
of both temperature and SOC, is not sensitive to erroneous
values of SOC and that an adaptive acquisition of the high-
frequency resistance of the battery can be used to assess the
SOH. This in- crease in the high-frequency resistance can be
anticipated for all commercially viable PHEV and HEV
batteries upon extensive cycling.
Goser et al. [12] derived a neuro-fuzzy scheme in which
a Kohenen Self-Organizing Map (SOM) is used to train the
data in a fuzzy system. The SOM is an unsupervised
clustering technique that seeks to organize the data. By using
an SOM to feed the data into a fuzzy system, the system is
made both reactive and adaptive (i.e., it unifies both neural
networks and fuzzy systems). The architecture of the system
implemented by Goser et al. [12] is as follows. The
fuzzification is split into two layers, leading to a rule base and
then subsequently to defuzzification. This two-layer network
is then trained in part by using neural network algorithms.
Buchmann et al. [13] implemented a neuro-fuzzy based BMS
that can track the battery
3
SOH in 3 minutes. The neural network was trained on
fuzzified data, and the outputs are then defuzzified.
Saha et al. [14] proposed a Bayesian framework (based on
Relevance Vector Machines and Particle Filters) for predicting
the SOC, SOH, and RUL and for providing uncertainty
bounds based on correlations between battery performance
(capacity) and model parameters (resistance). However, the
battery model used for estimating the battery characteristics
was simplistic be- cause the low-frequency Warburg
impedance was merely repre- sented by a resistance.
Moreover, this resistance is in series with the charge-transfer
resistance making it unidentifiable i.e., they cannot be
uniquely estimated. Furthermore, the Particle filter framework
incorporates exponential growth models for predict- ing the
internal battery parameters (resistance) and is Central
Processing Unit (CPU) intensive.
Conventional BMS approaches have mostly focused on ad-
dressing the estimation of SOC or SOH with limited attention
to RUL. Consequently, we propose a general framework for
esti- mating these critical characteristics (SOC, SOH, and
RUL) and validate it using Li-ion battery data.
2) Key Battery Modeling Approaches: Several battery
mod- els have been proposed in the literature, which are
briefly sum- marized in Table II. An appropriate trade-off
between the model complexity and performance is a key
aspect in battery modeling. In this paper, we focus on
equivalent circuit battery models con- sisting of voltage and/or
current sources, resistors, capacitors (or constant-phase
element, CPE), inductors, and/or Warburg impedance.
Electrochemical models [16], [17] are typically time-
consuming and are best suited for understanding the physical
design aspects of electrodes and electrolyte.
Peukert’s equation [18] is only applicable to batteries dis-
charged at constant temperature and constant discharge cur-
rent. When applied to a battery with a variable discharge rate
and changing operating temperature, it generally results in an
underestimation of the remaining capacity. Specifically, Li-ion
batteries, whose capacity is strongly dependent on the battery
temperature, which in turn is a function of duty cycle, battery
design, and environmental conditions, the Peukert equation is
not applicable.
Similarly, the Shepherd equation [19] has three major short-
comings. The model parameters could only be calculated from
the experimental discharge curves with constant current. Sub-
sequently, the applicability of this model appeared to be
limited to temperature range from 10 ◦C to 30 ◦C, and finally,
the aging effect on the battery model parameters had not been
consid- ered. Modifications to the equation have been
proposed in the literature to improve certain aspects of battery
performance, but limitations still remain.
Other battery models such as hydrodynamic models and
finite-element type battery models are complex and CPU-
intensive requiring high memory and significant computation
time.
In this paper, equivalent-circuit battery models have been
employed as they are more intuitive in circuit simulations,
robust to operating conditions, provide real-time
implementation, and enable the battery nonlinearities to be
modeled using circuit parameters such as resistors, capacitors,
inductors, etc.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
4 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS
TABLE II
BATTERY MODELS AND KEY FEATURES [15]
B. Data for Battery Modeling and Characterization
The validation data used in this paper had been collected as
part of the advanced technology development (ATD) program,
where performance testing was conducted on second-
generation Li-ion cells (i.e., Gen 2 cells) [20]. The 18650-size
Gen 2 cells, with baseline and variant chemistry, were cycle-
life tested at Idaho National Laboratory (INL); only baseline
cells data were used in this paper. Reference performance tests
(RPTs) are a set of tests performed at periodic intervals during
life testing to establish the condition and rate of performance
degradation of
TABLE III
GEN 2 BASELINE CELLS RATINGS AND LIMITATIONS [20]
batteries under test. These tests should be performed prior to
the beginning of life (BOL) testing, at defined periodic
intervals, and at the end of life (EOL) testing of all batteries
undergoing either cycle life testing or calendar life testing.
Life testing was interrupted every 4 weeks (i.e., 33 600
cycle- life profiles) for RPTs consisting of C1 /1 1 static
capacity test, a low-current hybrid pulse power
characterization (L-HPPC) test, and EIS test, at 60% SOC and
25 ◦C [20]. The static ca- pacity test measures the battery
capacity in ampere-hours at a constant-current discharge rate
corresponding to the manu- facturer’s rated C1 capacity in
ampere-hours. Discharge is ter- minated on a manufacturer-
specified discharge voltage limit. After the static capacity
tests, all cells were subjected to the EIS test. EIS was used to
determine impedance changes in the electrode–electrolyte
interface. EIS measurements were con- ducted by discharging
the cells from a fully charged state to the specified open-
circuit voltage (OCV) corresponding to 60% SOC, followed
by an 8- to 12-hour rest at OCV (to reach electro- chemical
equilibrium), and then the impedance is measured over a
frequency range of 10 kHz to 0.01 Hz [21], [22]. The EIS
mea- surements consisted of 61 data points in the frequency
range; computed using a constant AC voltage (5 mV) and
measuring the current response. The HPPC test is intended to
determine the dynamic power capability over the battery’s
usable charge and voltage range using a test profile that
incorporates both discharge and regen pulses [21].
Characterization testing was conducted at 25 ◦C with C1 /1
static capacity tests to establish performance parameters, such
as capacity, resistance, power, and energy. These tests consist
of a constant-current discharge from a fully charged state
using a fraction of the rated capacity defined at the 1-hour
rate. The Baseline cell degradation gener- ally increased with
increasing test temperature, SOC, and test time. The chemistry
of Gen 2 baseline cell used in the data anal- ysis of this paper
may be found in [20]. The battery ratings and limitations used
for testing are shown in Table III.
1
C 1 /1 Rate: The current corresponding to the rated capacity in ampere-
hours for a 1-hour discharge. For example, if the battery’s rated 1-hour
capacity is 5 Ah, then the C1 /1 is 5 A [2].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS
Fig. 1. SOC, SOH, and RUL estimation framework.
C. Scope and Organization of the Paper
Our approach to estimating the SOC, SOH, and RUL based
on EIS data employs a modified Randles circuit model of a
battery. This model consists of a resistor at high frequency, a
paral- lel RC circuit for modeling the charge-transfer
phenomenon at medium frequencies, and a Warburg
impedance to model the dif- fusion phenomenon at low
frequencies. The circuit parameters are estimated from the EIS
data using nonlinear least squares (NLLS) estimation
techniques. In order to overcome the ill- conditioning
associated with the NLLS problem, we divide the estimation
process into high-frequency region, charge-transfer region,
and diffusion region. The long-term temporal variations of
battery resistance are modeled by an auto-regressive sup- port
vector machine (SVM). Exploiting the linear correlation
between the battery resistance and C1 /1 capacity, the
nonlinear SVM models are used to forecast the battery’s SOC,
as well as the capacity fade and the power fade. An
alternative estima- tion procedure based on the U.S. Advanced
Battery Consortium (USABC) lumped parameter model
(LPM) [21] for predicting the power fade and energy fade
from L-HPPC test data is also proposed to validate the
estimates from EIS data. Unlike the method in [21], our
parameter estimation technique for the LPM model, namely,
prediction error minimization (PEM) method, ensures that the
relevant circuit parameters are nonnegative. The estimates of
capacity fade and power fade, in turn, are used to estimate the
RUL of the battery via moving average SVMR. An important
advantage of using SVMR is that the determination of the
model parameters corresponds to a convex optimization
problem, and hence, the solutions are globally optimal (given
the assumptions) and unique along with being sparse and
having simple geometric interpretations. Unlike Artificial
Neural Net-
5
works, the computational complexity of SVMR does not
depend on the dimensionality of the input space and are less
prone to over-fitting. Survival functions and RUL estimates of
the battery are also predicted via a hidden Markov Model
(HMM) trained using time-dependent power fade and capacity
fade estimates. The HMM provides a further validation of our
RUL estimates. The estimation framework proposed in this
paper is summarized in Fig. 1.
The primary contributions of this paper are four fold:
1) frequency-selective NLLS estimation to overcome ill-
conditioning associated with the modified Randles circuit pa-
rameters from the EIS data, 2) constrained optimization to en-
sure that the identified LPM parameters are nonnegative, 3)
novel auto-regressive SVMR-based capacity and power fade
estimation, and 4) SVMR and HMM-based RUL estimation.
The paper is organized as follows. Section II discusses
our approaches for identifying the battery model parameters.
Section III describes procedures for estimating the SOC, SOH,
and RUL characteristics of batteries based on the models
identi- fied in Section II. Section IV validates our estimation
approach on second-generation Li-ion cell (i.e., Gen 2 cell)
data collected at the INL. Section V concludes with a
summary and future research directions.
II. BATTERY MODELING
A. Identification of Impedance-Based Battery Model
In a previous work, Kuhn et al. proposed a modified
Randles scheme [23] shown in Fig. 2 as an equivalent circuit
model of a battery.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

6 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS

Fig. 2. Modified randles circuit model.

Fig. 4. Lumped parameter battery model.

region method and is based on the interior-reflective Newton

method [25], [26]. The noninteger function presented in (2)
was used to model the Warburg impedance.
The optimization algorithm does not converge to a solu-
tion (optimal parameter estimate) sometimes due to severe ill-
conditioning. Hence, we decoupled the EIS data according to
the three different frequency regions of the Nyquist plot
(namely high-frequency region, charge-transfer region, and
diffusion re- gion) shown in Fig. 3 and then, implemented
Fig. 3. Nyquist plot of battery impedance. NLLS on each independent data-set to obtain the parameter
estimates (RHF characterizes the high-frequency region, Rtc
and Cdl charac- terize the charge-transfer phenomenon, and γ
Here, the Warburg impedance ZW represents the diffusion represents the diffusion phenomenon). The three frequency
phenomenon; RHF (high-frequency resistance) denotes the
regions are chosen based on the frequency samples from the
elec- trolyte and connection resistances, and ǁthe Rtc Cdl EIS data as follows:
parallel circuit models the charge-transfer phenomenon.
Hence, the frequency-dependent impedance expression for the
High-frequency region → RHF → 10 000 − 125.89 Hz
above cir- cuit becomes
Charge-transfer region → Rtc ǁCdl → 100 − 0.12589 Hz
Rtc Diffusion region → ZW → 0.1 − 0.01 Hz.
Z ω R HF + ZW (ω) . (1)
( )= + dl
1 + jωR tc Hence, the inputs of the parameter estimation procedure are
Typically, Warburg impedance is theoretically expressed as optimization uses a trust
a noninteger function [24]
ZW (ω) = γω 1/2 (1 − j) (2)
where γ is a parameter, which depends on the electrochemical
phenomenon.
Nyquist plots of resistance versus (negative) reactance as a
function of frequency are used to depict impedance changes
in the electrode–electrolyte interface using equivalent circuit
models. As shown in Fig. 3, the Nyquist curve generally
shows a semi-circle of radius r governed by the charge-
transfer phe- nomenon, and a 450 slope (theoretically), where
the diffusion phenomenon is predominant. These profiles tend
to grow as a function of cell age and are particularly sensitive
to the chosen battery chemistry, the nature of the solid
electrolyte interface, and the temperature. The main difficulty
in modeling a battery is the diffusion phenomenon at low
frequencies.
In this paper, we implemented an NLLS estimation
technique using the lsqnonlin function in the Optimization
Toolbox of MATLAB to identify the parameters of the
equivalent circuit of the battery cell shown in Fig. 2. The
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

1) frequency; 2) real impedance; and 3) imaginary

impedance. The outputs are the parameters of modified
Randles circuit, namely, RHF , Rtc , Cdl , and ZW .

B. Identification of Lumped Parameter Battery Model

A model that encompasses the complex nonlinear
behavior of a battery, and describes all facets of its
performance over the en- tire life and over any energy
storage cycle, contains parameters that are difficult (or
impossible) to estimate from the available test data. Hence,
USABC proposed a battery model that is linear at a given
point in life based on a repeatable test cycle, such as the
HPPC test [22]. The USABC LPM [21] is shown in Fig. 4.
The parameters of the model are OCV = ideal voltage rep-
resenting open-circuit battery voltage, R0 = battery
internal ohmic resistance, RP = battery internal polarization
resistance, C = shunt capacitance, τ = RP C = polarization
time constant, IL = load current (positive for discharge
and negative for charge), IP = polarization current =
solution of the differential equation dIP /dt = (IL IP )/τ , −
initial condition, IP (0− ) = 0, and 1/OCVJ =
capacitance that accounts for variation in
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS 7

Fig. 5. Load current and terminal voltage for one-pulse of L-HPPC data.
Fig. 6. Estimated and actual terminal voltage for L-HPPC data.

OCV with the time integral of load current IL . Thus

VL = −OCV Σ t Σ III. ESTIMATION OF BATTERY CHARACTERISTICS

∫ IL dt + OCV − RP IP − R0 IL
0 A. SVMR Prediction
J
= battery terminal voltage. (3) The circuit parameters vary with time. We model the tem-
The objective is to estimate the lumped battery model poral variations of total resistance R(k) = (RHF (k)+ Rtc (k))
param- eters OCV, OCVJ , R0 , RP , and τ using the L-HPPC in week k via an auto-regressive SVMR of length L (typically,
test data. However, since the L-HPPC test repeats each pulse L = 2). Specifically
profile at several fixed depth of discharge values, the R(k) = f ({R(i)}k−1 ). (5)
parameters are es- timated at each depth-of-discharge (DOD). i=

Formulation of the linear LPM in state-space representation is We also use moving average SVMRs to predict the capacity
as follows: fade from C1 /1 capacities and RUL from capacity and power
⎡ ⎡ ⎤⎡ ⎡ ⎤
x˙ 1 fade estimates.
⎤0 0 0 x1 1
⎤
⎣ x˙ 2 ⎦ = ⎣ 0 −1/τ 0 ⎦ ⎣ x2 ⎦ + ⎣ 1/τ ⎦ IL
x 0 0 0 x 0 B. State of
⎡ ⎤ State of Charge (SOC) is defined as the available capacity
x1 in a battery expressed as a percentage of the actual (or
⎣ ⎦V = [ c c x3 1] estimated) rated capacity. This is normally referenced to a
x +
∫ dt IL 1 2 constant-current discharge at a C1 /1 rate. That is
x1 (t) =
2 L
IL (σ)dσ; x2 (t) = IP (t), x3 (t) = estimated capacity − capacity removed
OCV0 SOC (%) = estimated capacity ×100
0
c1 = OCVJ , c2 = −RP , d = −R0 , τ =RP C. (4) estimated and measured terminal volt- age and the R2 -statistics
for 1 week (i.e., eight pulses) of the L-HPPC data are shown in
The load current IL and terminal voltage VL for one pulse
Fig. 6.
(one DOD) of L-HPPC test (shown in Fig. 5) are inputs to the
system identification procedure.
The model parameters are estimated using the iterative
prediction-error minimization (PEM) method (see Appendix
B). The algorithm implementation is similar to the
autoregressive moving-average algorithm, with modifications
to the compu- tation of prediction errors and gradients, and the
cost function (sum of squares of measurement prediction
errors) minimization is implemented in such a way that the
variables are constrained to be nonnegative. The R2 -statistic
obtained was approximately 99.5% for all pulses. The
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
(6)
t ∫
0 where capacity removed = I (τ )dτ . For
discharge, capacity removed = IL .t, where IL = discharge
current in Amperes, and t = time in hours. constant-currentL
The sum of high-frequency (or ohmic) resistance and
charge- transfer resistance is correlated with the C1 /1
capacities from the C1 /1 static capacity test data. SVMR
[27] is then used to make future predictions of the
resistance using the model in (5), and consequently, the
estimated capacity is obtained from these correlation plots.
The capacity removed depends on the battery usage.

C. State of Health (SOH)

State of health (SOH) is the ability of a cell to store
energy, source, and sink high currents and retain charge
over extended
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

8 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS

periods, relative to its initial or nominal capabilities [9]. It is

a figure of merit that describes the degree of degradation of a
battery, and gives a quantitative measure that replaces fuzzy
statements, such as “fresh,” “aged,” “old,” and “worn out”
[28]. In this paper, we characterize the SOH of a battery by its
power fade and capacity fade.
1) Power Fade: The loss of cell power due to an increase in
cell impedance during aging is known as power fade [2]. In
this paper, we determine the power fade of the battery from
the EIS test. The actual power from the EIS test is calculated
as
V2
P = (7)
R
where V is the voltage (5 mV for EIS test), and R is the total
resistance (RHF + Rtc ) obtained from EIS data using NLLS or
predicted via SVMR using (5). Hence, power fade is
computed as follows: . Σ
Power(k) R(0)
Power fade = 1 − =1− (8)
Power( R(
where Power(0) is the power at the BOL, and Power(k) is the
power at the desired time (week). Consequently, power fade at
the chosen week is obtained via SVMR predictions of the total
resistance using (8).
An alternate procedure for estimating the power fade and
energy fade from HPPC test data is shown in Fig. 7.
A detailed procedure for determining the available energy
and available power, along with the resistance characteristics,
OCV and the pulse power capability, from the L-HPPC data
has been discussed in [20] and [21]. The Power fade and
Energy fade from the BOL to some later point in time is
calculated using the equations Fig. 7. Block diagram for estimating power fade and energy fade from L-
. Available power (k) HPPC test data using LPM model.
Power fade (%) = 1 100 (9)
Available power (0)
The capacity fade of a battery is estimated by predicting the
C1 /1 capacities using an autoregressive SVMR
Available energy (k)
Energy fade (%) = .1 100 (10) C(k) = g({C(i)}k−1 ). (12)
Σ Available energy (0)
i=
where time ‘0’ refers to−the BOL RPT, and time ‘k’ × refers to Once the capacity is estimated, the capacity fade is
an RPT performed at a later time, where the power fade and computed from (11a). Alternately, one can compute the
energy fade are to be determined. capacity fade from (11b) by appealing to the linear correlation
2) Capacity Fade: The gradual loss of capacity of a sec- between the capacity and battery resistance.
ondary battery with cycling is known as capacity fade [2]. Ca-
pacity fade is also the percent loss in C1 /1 discharge capac- D. RUL via SVMR
ity [22]
In this paper, we predict the RUL of the battery using a
. Capacity(k) moving average SVMR for different thresholds on capacity
Capacity fade (%) = 1 100 (11a)
Σ Capacity(0) fade and power fade. Formally, at week k, we estimate
(− ( ) (0)) × RUL(k) = h({P (i),C(i)}k ) (13)
αRk − R i=1
Capacity fade (%) =
×100 (11b)
where C(k) and P (k) are the capacity fade and power fade,
β
where Capacity(0) is the BOL Capacity, and Capacity(k) is total resistance, i.e., C(k) = β − α(R(k) − R(0)), it can also be
the Capacity at the desired time (week) k. Hence, capacity written as (11b).
fade at the chosen week can be computed by SVMR
prediction of the C1 /1 capacities from the static capacity test.
Given that the capacity is approximately linearly related to the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

respectively, and RUL(k) is the RUL at time k.

E. Estimation of Survival Function and RUL Using HMMs

Survival function is defined as the probability that the
system operates without fail beyond a specified time
instant ‘t’, given that it is operational at time ‘0’. It is also
known as the reliability
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS 9

function and is given by

∫ ∞
S(t) = P (T > t) = f (τ )dτ (14) algorithm or the Viterbi algorithm [29]. Here, we determined
t the jth component of (π(t)) from the Viterbi algorithm using

(
(j δ
ΣN
where T is a continuous random variable denoting ‘time-to- n =1 δn (t)
failure’ (TTF) having pdf f (t). Here, we estimate the where δj (t) is the probability that the observations up to
survival functions of batteries using HMMs, which, in turn, time ‘t’ end in state ‘j’ and is given by
are used to predict the RUL of batteries.
An HMM is characterized by the number of states (N ), δj (t) = max P (s(1) s(2) ... s(t) = j,
s(1),s(2),...,s(t−1)
observations (x), state transition probability matrix (A), emis-
sion probability matrix (B), and initial state probability vector x(1) x(2) ... x(t)|λ)
(π). HMM parameters are generally denoted as λ = (A, B, π), = max [δ −
(t 1)a ]b (x(t)). (18)
where λ represents the parameter set of HMM. A brief 1≤i≤N i ij

descrip- tion of HMMs is provided in Appendix C. Capacity We can also estimate π(t)’s from the normalized forward
fade and power fade are the observations of the HMM. Since variables (via forward algorithm) using
the ob- servations x(t) at time t (e.g., week t) are continuous,
they are
modeled as Gaussian mixture with m components having αj (t) = max [αi(t − 1)aij ] bj (x(t)) (19)
1≤
means i≤N
{μ }, covariances {Σ }, and mixture priors {w }. Formally
i i i
and normalizing the forward variables as in (17). Once the
m
state probabilities are obtained, the estimation of survival (or
bi(x) = ΣN (x; μi , Σ i ) . wi . (15)
relia- bility) function is computed using
i=1

The HMM parameters λ are learned from the training data S(t + n) = π(t) An ef (20)
using the Baum–Welch algorithm [29]–[31]. The number of where ef is a vector of size ‘N ’ with 0’s in failing states and
states of HMM is chosen based on how well the HMM 1’s in the operational states. The mathematical derivation of
predicts the power fade and capacity fade as determined by (20) is given in Appendix D.
the coefficient of determination (R2 -statistic) which in turn
depends on the squared error loss of the capacity fade and IV. RESULTS AND DISCUSSION
power fade estimates. Formally, the predicted observations at
time step ‘t + 1’ can be computed recursively via Second-generation Li-ion cells with baseline chemistry
. Σ were chosen as the validation data in this paper. Life testing
p x(t + 1)|xt was inter- rupted every four weeks (data were collected for a
Σ total of 140
= . Σ weeks) for RPTs consisting of static capacity test, an L-HPPC
p x(t + 1), s(t + 1)|xt
s(t+1) test, and EIS test. Five baseline cells (8, 9, 11, 14, and 15) out
Σ of a test matrix of 30 cells have been chosen for our
= Σ . t Σ experiments because of time-period of collection of data (140
p (x(t + 1)|s(t + 1)) p (s(t + 1)|s(t)) p s(t)|x
s(t+1) weeks). The other cells in the test matrix were removed earlier
s(t)
(at different
Σ time-periods) as they were considered “dead” (i.e., approxi-
1 mately, power fade = 50% and capacity fade = 30%).
= p(xt p (x(t + 1)|s(t +
s(
) 1))
Σ . Σ A. Parameter Estimation Results
t+ 1)

× p (s(t + 1)|s(t)) p xt , s(t)

s(t)
Estimates of the modified Randles circuit parameters RHF ,
Σ Rtc , Cdl , and Warburg parameter γ, for different battery cells are
1 obtained as a function of time (weeks) using frequency-
= p(xt p (x(t + 1)|s(t +
s(
1)) selective NLLS from the EIS test data. The time-dependent
)
behavior of (RHF + Rtc ) will be used to estimate the
degradation of battery
t+ 1)
Σ
× Σp (s(t + 1)|s(t)) α(s(t)) 1Σ (x(t + 1)|s(t + 1) = j) a
s(t)
N N = p
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

cells. The battery cells as a function of time (weeks) is shown in Fig. 8

degradatio (Cdl and Warburg parameter γ are not shown here, since they
n of some show negligible change over the aging process and are not
α (t) (16) of the used in our estimation framework).
ij i
p(xt j =1 i=1
The battery’s internal impedance parameter identified from
) the EIS test chosen for correlation analysis is the total internal
where xt = [x(1), x(2),..., x(t)] is the set of observations un- resistance of the battery (i.e., R = RHF + Rtc ). Fig. 8 shows
til time t(1 ≤t K), αi(t) = p(xt, s(t) = i) is the forward the variation of resistance as a function of weeks. The SVMR
variable, and ‘s’ is the hidden state. model (with C = 1000 and σ = 0.001) with L = 2 for the total
An estimate of state probabilities (π(t)) at the current time resistance parameter R showed excellent R2 -fit values of 97%
in- stant ‘t’ can be computed from the HMM model via the for as many as 15–20 week-ahead predictions. Hence, SVMR
forward
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
10 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS
Fig. 8. Total resistance versus battery age in number of weeks.
Fig. 9. Average C1 /1 discharge capacity for baseline cycle-life cell groups.
proved to be an ideal algorithm for future predictions of the
circuit parameters. This implies that EIS tests can be done as
part of routine vehicle maintenance (e.g., oil changes).
The C1 /1 discharge capacities were extracted from the C1 /1
static capacity test data. Fig. 9 shows the average C1 /1
discharge capacity for different baseline cells that were aged at
60% SOC. Fig. 10 shows the high degree of linear correlation
between the C1 /1 discharge capacity and the internal
resistance pa- rameter R for different cells. For the linear
model
− C(k) = β α(R(k)
−
R(0)), typical
∈ values for α and β for the various
∈ cells were
α [9.52, 11.03] and β = C (0) [0.93, 1]. Hence,
linear model parameters of the correlation plots for different
temperatures and SOCs can be stored as part of a look-up table
for instantaneous computation and prediction of internal resis-
tance and battery capacity.
Fig. 10. Correlation between resistance and battery capacity.
Fig. 11. Percentage SOC as a function of time for baseline cell 8.
B. SOC Results
The resistance predictions are used to obtain the
correspond- ing C1 /1 discharge capacity from the linear
correlation plots in Fig. 10. Subsequently, the SOC is obtained
using (6). Fig. 11 shows the variation of the SOC for a
constant-current discharge of baseline cell 8 from beginning-
of-life (BOL) to EOL. Conse- quently, SVMR could also be
used to predict SOC even under nonlinear driving conditions
with a good accuracy.
C. SOH Results
The rate of capacity loss (capacity fade) generally increases
with increasing temperature and SOC. The SVMR model
(with C = 100 and σ = 0.0015) with L = 2 for the C1 /1
capacities showed excellent R2 -fit values of 95% for as many
as 10 week- ahead predictions. Fig. 12 shows the capacity
fade of different baseline-chemistry cells obtained using (11).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS 11
thresholds, and

Fig. 12. Capacity fade prediction for different cells.

Fig. 13. Power fade estimation for different cells.

The power fade of different cells predicted using the ap-

proaches described in Section III is shown in Fig. 13. The
power fade computed from EIS data [using (8)], and L-HPPC
data [us- ing (9) and (10)] showed similar results, with the EIS
predictions providing an underestimate by approximately 2%
for all cells. Hence, battery health monitoring in terms of its
power fade and capacity fade can be obtained using an
autoregressive SVMR, which enables condition-based
maintenance of the batteries.

D. RUL Results
Fig. 15 shows the RUL for an EOL criterion of
approximately 50% power fade and 30% capacity fade, at
which time, the battery is assumed to have failed
operationally.
As seen in Fig. 15, the RUL of the battery decreases with
time and the EOL criterion is based on the application-
dependent capacity fade and power fade. The SVMR can be
trained for various thresholds of capacity and power fade
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

Fig. 14. Survival function estimation via HMM.

consequently, the RUL can be obtained for a specific

context- dependent threshold. The SVMR regularization
parameter was found to be 100 (i.e., C = 100) and σ of the
RBF kernel function was found to be 1 for RUL
predictions.

E. Survival Function and Reliability Analysis Results

An eight-state HMM model is trained using capacity
fade and power fade of a battery as observations and the
survival function is estimated for the cells. A two-
component Gaussian mixture model was used as the
emission model. Fig. 14 shows an estimated survival
function from HMMs. It is a monotonically decreasing
function that depicts the reliability of a battery at any time
‘t’. The R2 fit was about 99.91%.
The expected remaining life at any time ‘t’ can be
computed from the survival function by defining a
threshold on reliability. Mathematically, it is written as
follows:
.. Σ Σ
RUL(t) = arg min π(t)An ef ε0 (21) ≤
n

where ε0 denotes the threshold of functional failure. Given

that a 50% power fade and a 30% capacity fade constitutes
a func- tional failure of the battery (i.e., EOL), ε0 is
approximately 0.2578. Fig. 15 shows the expected
remaining life of the bat- tery with ε0 as 0.2578.
Consequently, SVMR- and HMM-based RUL predictions
are quite similar with SVMR providing better estimates
(based on R2 and MSE measures) and can be used for
prognostic health management of the battery module. Nev-
ertheless, HMMs provide consistently comparable results
and outperform SVMR towards the EOL of the battery
(i.e., it gives estimates with smaller variance for the 136th-
and 140th-week predictions than SVMR).

V. CONCLUSION
Next-generation BMSs will feature online tracking and
mon- itoring of pivotal battery characteristics, such as
SOC, SOH, RUL, survival function, area-specific
impedance (ASI), and
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

12 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS

Further investigation is required on the validation of the

esti- mation framework for real-time driving schedule
conditions and dependence of circuit parameters on
temperature and cycling conditions, as well as different
battery chemistries. In addition, a systematic evaluation of
different algorithms for the estima- tion of essential battery
characteristics on benchmark battery data is suggested as a
future area of research.

APPENDIX A
SUPPORT VECTOR MACHINE REGRESSION
The main objective in SVM regression (SVMR) is to es-
timate a function f (x) that is as close as possible to
the
target values {y(k)}K for the corresponding {x(k)}K ,
k k
where x(k) R∈ . In the current context, y(k) = R(k), and
L

T
x(k) = [R(k −1),R(k − 2),..., R(k − L)] . A similar anal-
ogy applies to capacity and RUL. The training data are
arranged as follows:
Fig. 15. RUL prediction via SVMR and HMM.
F = {(x(1), y(1)) , (x(2), y(2)) , . . . , (x(K), y(K))} . (22)
By a nonlinear mapping Φ, x(k) is mapped into a feature
power density to facilitate efficient diagnostic and prognos- space and a linear estimate is constructed in this space as
tic maintenance of batteries. Here, we proposed a framework follows:
for estimating and predicting these salient battery performance y(k) = f (x(k), w) = wT Φ (x(k)) + b. (23)
measures. Three commonly used RPTs were used for our
anal- ysis, namely, the C1 /1 static capacity test, the EIS test, To determine the weight vector w and bias b, the following
and the L-HPPC test. Our approach to estimating the SOC, regularized risk function is minimized:
SOH, and RUL is based on a modified Randles equivalent 1 Σ
2
circuit model Φ (w) = ǁwǁ + C K |y(k) − f (x(k), w)| ; C > 0,ε
> 0.
of a battery. This model consists of a high-frequency 2 ε
k =1
resistance, a parallel RC circuit for modeling charge-transfer
(24)
phenomenon at medium frequencies, and Warburg impedance
to model the
diffusion phenomenon at low frequencies. The circuit param- The first term in the objective function represents the model
eters are estimated from the EIS data using NLLS estimation complexity and the second term represents the model
techniques. The ill-conditioned parameter estimation problem accuracy. Here, C is a regularization parameter to control the
has been addressed by decoupling the high-frequency, charge- trade-off between these two terms. We should note that the
transfer, and diffusion regions of the Nyquist plot. We larger the value of C, the more the error is penalized. Thus, C
proposed an autoregressive SVMR using an ε-sensitive loss should be chosen with care to keep away from over fitting (C,
function for predicting the resistance. The C1 /1 capacity was an important tuning parameter in SVMR, was found to
obtained from the C1 /1 static capacity test. Hence, the SOC generally lie between 100–1000 in our experiments). We use
can be obtained using the resistance-capacity mapping or from the ε-sensitive loss function for the second term in (24) given
the C1 /1 predic- tion via an autoregressive SVMR. by [27]
Consequently, a framework was proposed to predict the
|y(k) − f (x(k), w)|ε
capacity fade and power fade, which
characterizes the SOH of a battery. An alternate procedure for .
0, if |y(k) − f (x(k), w)|ε < ε (25)
estimating the power fade and energy fade from L-HPPC test = |y(k) − f (x(k), w)|ε − ε,
data using the LPM is also proposed. The RUL predictions otherwise
were
made by setting different application-dependent thresholds on which means that we tolerate errors up to ε (ε was set to 10−3
capacity fade and power fade. Survival function and RUL es- in our experiments) and that errors beyond ε have a linear
timation of the battery was performed using HMM model pre- effect. This error function is therefore more tolerant to noise
dictions as well. The proposed framework for the Gen 2 Li- and is thus more robust [27]. By using Lagrange multiplier
ion battery data showed a good accuracy for estimating the techniques, the minimization of Φ (w) leads to the following
impor- tant battery characteristics. dual optimization problem:
Consistent estimates of residual useful life (RUL) via two ΣK K
Σ
different techniques (SVMR, HMM) provide partial cross- max − ε (ψk∗ + ψk ) + (ψk∗ − ψk ) y(k)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
α,α∗
validation, as well as demonstrate the accuracy of RUL esti- k =1 k =1
mates. In addition, our use of steady-state EIS and transient K K
HPPC data to separately estimate power fade and capacity fade 1Σ Σ ) K (x(k),x(m)) (26)
provides another cross-validation of our − 2 (ψ∗k − ψ k ) (ψ∗m − m
ψ k =1 m
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS 13

subject to where A denotes a suitable norm, such as A (ξ) =ǁξ 2 (two-

K Norm). The maximum likelihood estimate is obtained by a
Σ(ψk
∗
ψk ) = 0, 0 < ψk , ψk∗ < C, k = 1, . . . , K. (27) par- ticular choice of the norm. Let the observed data be
k =1 − produced by
This problem is solved by quadratic programming
techniques [27]. The regression takes the form VL (t) = f (Zt−1 , θ)+ e(t) (33)
Σ
f (x) = K (ψ∗ − ψ ) K (x(k),x)+ b. (28) where {e(t)} is a sequence of independent random variables
−
k k
with probability density function p(x). If A(x) = log p(x),
k =1
then θˆK is equal to the maximum likelihood estimate, as the
The nonlinear mapping (or kernel function) K is used to criterion in (32) is the negative logarithm of the likelihood
trans- form the original input x to a higher dimensional feature func- tion (apart from the θ-independent terms). The
space Ω as K (x(k), x(m))( = Φ (x(k)) , Φ (x(m))) , where
( ., . numerical cal- culation of the minimizing argument involves
de- notes dot product. In practice, various kernel functions are substantial com- putations, and probably a complex search
used, such as polynomial, radial basis functions (RBF) [27], over a function with several local minima. Damped Gauss–
and sig- moid functions. Here, we use the RBF kernel function Newton method is gen- erally used to perform the numerical
(σ was usually found to be in the range − of 0.001 1 in our search. For a scalar output and A(ξ) = 1 ξ2 , this takes the form
. numerical difficulties Σ
experiments) because it has less 2
− ǁx(k) − x(m)ǁ
2 θˆ(i+1) = θˆ(i) + δi Hi−1 dˆi
K (x(k), x(m)) = exp (29)
2σ2
dˆi = −WKJ (θˆ(i) )
APPENDIX B WKJ (θ(i) ) = −∇WK (θ)
K
PREDICTION ERROR MINIMIZATION METHOD 1 Σ (t) −
= (V (t|θ))ζ(t, θ)
Prediction-error methods [33] are a broad family of param- Lˆ
Kt=1 V L
eter estimation methods that canand
be have
applied to reasonably ar- ∂
bitrary model parameterizations a close relationship ˆ
with the maximum likelihood method. Some of the
advantages of PEM are that it has excellent asymptotic
properties of un- 1 Σ
K
H i= W JJ (θˆ(i) ) ≈ ζ(t, θˆ(i) )ζ T (t, θˆ(i) ) (34)
biased estimates and minimal uncertainty and that it can K
K
t
han- dle systems that operate in a closed loop (the input is
determined as output feedback). The method also has disad-
vantages such as it requires an explicit parameterization of where δi is a scalar chosen so that WK (θ(i+1) ) < WK (θ(i) ). As
the model, the minimization may be laborious, and involve the data size increases (K →), the estimate converges to the
search surfaces that may have local minima, and hence, value θ∗ that would minimize the expected value of the norm
requires good initial parameter values. Let Z K = IL (1), VL (1), of the prediction errors
{
IL (2), VL (2),... ., IL (K), VL (K) denote the measured dataset
} θ∗ = arg min E{A(ξ(t, θ))}. (35)
up to time K. The general predictor model is given by
θ
VˆL (t|θ) = f . Z t−1 , θΣ ; 1 ≤ t ≤ K (30)
If {ξ(t , θ∗)} is approximately white noise, then the random
√
where θ = {c1 = OCV , c2 = −Rp , d = −R0 , τ = Rp C} is
J
vector K θK (θˆ∗ converges
− to the normal distribution with zero-
a vector of the parameters of the lumped parameter battery mean and covariance matrix of θˆK is approximately given by
model (LPM).
The goal of PEM is to determine an estimate of θ (which is P = ρ[E{ζ(t, θ∗ )ζT (t, θ∗)}]−1 (36)
de- noted θˆK ) from the model parameterization, i.e., f (Z t ,
θ), and the observed dataset Z K to minimize the distance where
between the predictor model output {VˆL (t|θ)}K and the
t
measured output
{VL (k)}K in a suitable norm 2 ∗
k
θˆK ρ = E{ξ (t, θ )}
= arg min (θ) (31)
WK θ
∗ dˆ
Σ
WK (θ) = K A(VL (t) − VˆL (t|
θ)) √
t Hence, θˆk converges to θ∗ at a rate K, and if
A(ξ(t, θ)); ξ(t, θ) = VL (t) VˆL (t θ) (32)
K
Σ =
−
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
t=1 disturbances are Gaussian, the covariance matrix (36) equals
the Cramer–Rao lower bound. Therefore, PEM gives the
optimal asymptotic properties.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

14 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS

sequence is as follows:
p (s(1),..., s(n), x(1),... , x(n)) Σ
Σn−1 Σ n
= πs(1) as(t)s(t+1) Σ bs(t)x(t) (42)
t=1 t=1

and this can be considered its defining property.

The parameter set of the HMM model, namely, A, B, and
Fig. 16. (a) Three-state HMM. (b) N -states of an HMM are coupled to an π, represent the probability of moving from the current state to
observation process x of cardinality M . another, the probability of a new observation given the current
state, and the likelihood of initial state, respectively. The
forward variable will be used to evaluate the likelihood of a
APPENDIX C
sequence of observations. The forward variable of an HMM is
HIDDEN MARKOV MODELS defined as
As the HMM is a doubly embedded stochastic signal α i (t)= p(x(1), x(2),... , x(t), s(t)= |i λ)= p(xt, s(t)= i
model, the premise behind an HMM is that the true underlying λ).
pro- cess, which is represented as a Markov chain (43)
depicting the It is easily checked that the following recursion holds for
evolution
di- of true states (events) as a function of time, is not
the forward variable
Σ Σ
rectly observable (hidden states), but it can be N
probabilistically inferred through another set of stochastic αj (t + 1) Σ αi(t)aij bj x(t+1) (44)
processes (observa- tions). An excellent tutorial on HMMs = i=1
can be found in [29] and [30]. The Baum–Welch re-estimation
algorithm [31], which is in fact an application of the EM
with the initial condition
algorithm [32], makes it a
convenient tool for modeling dependent observations. HMMs αj (1) = πj bj x(1) (45)
have been successfully applied in speech recognition, DNA Additional details on HMMs including the backward algo-
se- quence analysis, robot control, and signal detection, to rithm, the Viterbi algorithm for finding the most likely
name a few. Three fundamental problems solved by HMM sequence of hidden states that could have generated an
models are observed sequence, and the Baum–Welch algorithm for
1) evaluation of the probability of observation sequences learning HMM parameters may be found in [29]–[31].
given a specific model, 2) decoding the most-likely evolution
of model states, and 3) estimation of HMM parameters to best APPENDIX D
elucidate a sequence of observations.
A discrete HMM is parameterized by DERIVATION OF RELIABILITY FUNCTION
FROM HIDDEN MARKOV MODEL PARAMETERS
Λ = (A, B, π) (38)
It is well known that if the continuous time Markov chain
where
A = [aij ] = [p (s(t + 1) = j|s(t) = i)] (39) is uniformized with an intensity parameter Λ = supi (−qii),
then the infinitesimal generator matrix Q = [qij ] = Λ(A − I),
(i, j = 1 , . . . , N ) is the transition probability matrix of the where A is the transition matrix, and I is a N N matrix. The
un- derlying Markov chain elements of A = [aij ] and Q = [qij ] are related×via
B = [bij ] = [p (x(t) = j|s(t) = i)] (40) q
aij = ij
(i = 1 , . . . , N ; j = 1 , . . . , M ) is the emission matrix (which Λ
is qii
also known as confusion and observation matrix), and aii = 1 + . (46)
Λ
π = [πi = p(s(1) = i)] (41) Now, let us define πi(t) to be the probability that the sys-
(i = 1,...,N ) is the initial probability distribution of the un- tem is in state ‘i’ at time t. Then, the 1 ×N row vector
derlying Markov states at time t = 1. A graphical model of an π(t) = [π1 (t), π2 (t),..., πN (t)] of state probabilities satisfies
HMM is shown in Fig. 16(a) for N = 3 and its dependence on the Chapman–Kolmogorov (C–K) equation
the observation process x(t) is shown in Fig. 16(b). dπ(t)
= π(t)Q (47)
Implicit to the above notation is the finite number of states dt
(N ) and finite alphabet of observations (M ). The HMMs can
be generalized to allow for continuous emissions, implying with the initial condition π(0) = π0 . Equivalently
that bij ∞ n
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
Σ (Λt)
in (40) is a probability density function. A convenient choice π(t) = π eQt = π(n)e−Λt (48)
of the initial probability is the stationary distribution of the 0
n!
n =0
underlying Markov states, so that the resulting sequence can
be regarded as stationary. The joint probability of an HMM where π(n + 1) = π(n)A.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

PATTIPATI et al.: SYSTEM IDENTIFICATION AND ESTIMATION FRAMEWORK FOR PIVOTAL AUTOMOTIVE BMS CHARACTERISTICS 15

Then, the reliability function for the random variable T is [6] I.-S. Kim, “Nonlinear state of charge estimator for hybrid electric
the probability that the system is still in operational state and vehicle battery,” IEEE Trans. Power Electron., vol. 23, no. 4, pp. 2027–
2034, Jul. 2008.
is given by [7] I.-S. Kim, “A technique for estimating the state of health of lithium
batter- ies through a dual-sliding-mode observer,” IEEE Trans. Power
S(t) = Pr(T > t) = Pr {failure time >t} Electron., vol. 25, no. 4, pp. 1013–1022, Apr. 2010.
[8] M. Verbrugge and E. Tate, “Adaptive state of charge algorithm for
= π(t) ef nickel metal hydride batteries including hysteresis phenomena,” J.
Power Sources, vol. 126, pp. 236–249, 2004.
= π0 e Qt e f (49) [9] B. S. Bhangu, P. Bentley, D. A. Stone, and C. M. Bingham, “Nonlinear
observers for predicting state-of-charge and state-of-health of lead-acid
where ef is a vector of size ‘N ’ with 0’s in failing states and batteries for hybrid-electric vehicles,” IEEE Trans. Veh. Technol., vol.
54, no. 3, pp. 783–794, May 2005.
1’s in the operational states. Hence, the reliability function at [10] D. V. Do, C. Forgez, K. K. Benkara, and G. Friedrich, “Impedance
any time ‘t + τ ’ can be written as follows: observer for a Li-ion battery using Kalman filter,” IEEE Trans. Veh.
Technol., vol. 58, no. 8, pp. 3930–3937, Oct. 2009.
S(t + τ ) = π(t)eQτ ef (50) [11] M. Verbrugge, D. Frisch, and B. Koch, “Adaptive energy management
of electric and hybrid electric vehicles,” J. Electrochem. Soc., vol. 152,
Subsequently, the failure probability distribution function is no. 2, pp. A333–A342, 2005.
[12] K. Goser, K. Schuhmacher, M. Hartung, K. Heesche, B. Hesse, and A.
given by Kanstein, “Neuro-fuzzy systems for engineering applications,” in Proc.
IEEE AFRICON, vol. 2, 1996, pp 759–764..
B(t) = 1 − S(t) [13] I. Buchmann, “Artificial intelligence reads battery state-of-health in
three minutes,” presented at the 16th Annu. Battery Conf., Long Beach,
= 1 − π0 e Qtef (51) CA, 2001.
[14] B. Saha, K. Goebel, S. Poll, and J. Christopherson, “Prognostics
whereas the density function is as follows: methods for battery health monitoring using a Bayesian framework,”
IEEE Trans. Instrum. Meas., vol. 58, no. 2, pp. 291–296, Feb. 2009.
dB(t) [15] ThermoAnalytics Inc., ‘‘Battery modeling for HEV simulation,”
b(t) = = −π0 QQt . (52) HEVsim Technical Manual [Online]. Available:
dt e ef http://www.thermoanalytics.com/ docs/batteries.html.

Similarly, in the discrete case (i.e., when t is discrete), the [16] D. W. Dees, V. S. Battaglia, and A. Belanger, “Electrochemical
modeling of lithium polymer batteries,” J. Power Sources, vol. 110, no.
probability mass function is defined: 2, pp. 310– 320, Aug. 2002.
[17] J. Newman, K. E. Thomas, H. Hafezi, and D. R. Wheeler, “Modeling of
p(n) = Prob {failure time > nΔ} lithium-ion batteries,” J. Power Sources, vol. 119–121, pp. 838–843,
Jun. 2003.
− Prob {failure time > (n + 1) Δ} [18] D. Doerffel and S. A. Sharkh, “A critical review of using the Peukert
equation for determining the remaining capacity of lead-acid and
= π(n) ef − π(n + 1) ef lithium- ion batteries,” J. Power Sources, vol. 155, no. 2, pp. 395–400,
Apr. 2006.
= π(n) [I − A] ef [19] W. X. Shen, C. C. Chan, E. W. C. Lo, and K. T. Chau, “Estimation of
battery available capacity under variable discharge currents,” J. Power
= π(0)An [I − A]ef (53) Sources, vol. 103, no. 2, pp. 180–187, Jan. 2002.
[20] J. P. Christopherson, I. Bloom, E. V. Thomas, K. L. Gering, G. L. Hen-
where Δ is the discretized interval (e.g., week). This is riksen, V. S. Battaglia, and D. Howell, Advanced technology
analogous to development program for lithium-ion batteries: Gen 2 performance
evaluation, Final Report, INL/EXT-05-00913, Jul. 2006.
[21] FreedomCAR Battery Test Manual for Power-Assist Hybrid Electric Vehi-

S(t + n) = π(t)An cles, DOE/ID-11069, Oct. 2003.

(54)
ef
ACKNOWLEDGMENT
The authors would like to thank the Idaho National Labora-
tory for access to the Advanced Technology Development
Gen2 cell data.
REFERENCES
[1] E. Bordeaux, “Portable Power Management—A Holistic Perspective,”
Intersil Corp., 2002.
[2] D. L. Linden and T. B. Reddy, Handbook of Batteries, 3rd ed. New
York: McGraw-Hill, 2002.
[3] S. Piller, M. Perrin, and A. Jossen, “Methods for state-of-charge
determi- nation and their applications,” J. Power Sources, vol. 96, pp.
113–120, 2001.
[4] V. Pop, H. J. Bergveld, P. H. L. Notten, and P. P. L. Regtien, “State-of-
the-art of battery state-of-charge determination,” Meas. Sci. Technol. J.,
vol. 16, pp. R93–R110, 2005.
[5] V. Pop, H. J. Bergveld, P. H. L. Notten, J. H. G. Op het veld, and P. P.
L. Regtien, “Accuracy analysis of the state-of-charge and remaining
run- time determination for lithium-ion batteries,” Meas. J., vol. 42, pp.
1131– 1138, 2009.
[22] PNGV Battery Test Manual, Rev.3, DOE/ID-10597, 2001.
[23] E. Kuhn, C. Forgez, and G. Friedrich, “Modeling diffusive phenomena
using non-integer derivatives: Application NiMH batteries,” Eur. Phys.
J. Appl. Phys., vol. 25, pp. 183–190, 2004.
[24] D. Landolt, “Corrosion et chimie de surfaces des me´taux,” Traite´
des Mate´riaux, Presses Polytechniques et Universitaires Romandes,
1993.
[25] T. F. Coleman and Y. Li, “An interior, trust region approach for
nonlinear minimization subject to bounds,” SIAM J. Optim., vol. 6, pp.
418–445, 1996.
[26] T. F. Coleman and Y. Li, “On the convergence of reflective Newton
meth- ods for large-scale nonlinear minimization subject to bounds,”
Math. Programm., vol. 67, no. 2, pp. 189–224, 1994.
[27] C. M. Bishop, Pattern Recognition and Machine Learning. New York:
Springer-Verlag, 2006.
[28] E. Meissner and G. Richter, “The challenge to the automotive battery
industry: The battery has become an increasingly integrated component
within the vehicle electric power system,” J. Power Sources, vol. 144,
pp. 438–460, 2005.
[29] L. R. Rabiner, “A tutorial on hidden Markov models and selected appli-
cations in speech recognition,” Proc. IEEE, vol. 77, no. 2, pp. 257–286,
Feb. 1989.
[30] L. Rabiner and B. Juang, “An introduction to hidden Markov models,”
IEEE ASSP Mag., pp. 4–16, Jan. 1986.
[31] L. Baum, T. Petrie, G. Soules, and N. Weiss, “A maximization technique
occurring in the statistical analysis of probabilistic functions of Markov
chains,” Ann. Math. Stat., vol. 41, pp. 164–171, 1970.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of
16 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS—PART C: APPLICATIONS AND REVIEWS
[32] T. Moon, “The expectation-maximization algorithm,” IEEE Signal Pro-
cessing Mag., vol. 13, no. 6, pp. 47–60, Nov. 1996.
[33] L. Ljung, “Prediction error estimation methods,” Circuits, Syst., Signal
Process., vol. 21, pp. 11–21, Jan. 2002.
Bharath Pattipati received the B.E. degree in elec-
trical and electronics engineering from M.S.
Ramaiah Institute of Technology, Bangalore, India,
in 2005 and the Master’s degree in electrical and
computer engi- neering from the University of
Connecticut, Storrs, in 2009, where he is currently
working toward the Ph.D. degree in electrical and
computer engineering. His current research interests
include the applica- tion of systems theory and
optimization techniques to complex large-scale
systems, application-driven analysis of neural
networks, pattern recognition, and
fault diagnosis and prognostics.
Chaitanya Sankavaram received the B.Tech. de-
gree in electrical and electronics engineering from
Sri Venkateswara University, Tirupathi, India, in
2005. She is currently working toward the Ph.D.
degree in electrical and computer engineering with
the Univer- sity of Connecticut, Storrs.
She was a Project Engineer with Wipro
Technolo- gies, Bangalore, India, for two years. Her
current re- search interests include fault diagnosis
and prognosis, proportional hazard models,
reliability analysis, data mining, pattern
recognition, and optimization theory.
Krishna R. Pattipati (S’77–M’80–SM’91–F’95)
re- ceived the B.Tech. degree in electrical
engineering with highest honors from the Indian
Institute of Tech- nology, Kharagpur, India, in 1975
and the M.S. and Ph.D. degrees in systems
engineering from the Uni- versity of Connecticut,
Storrs, in 1977 and 1980, re- spectively.
From 1980 to 1986, he was with ALPHATECH,
Inc., Burlington, MA. Since 1986, he has been with
the University of Connecticut, where he is currently
a Professor of electrical and computer engineering.
He was a Consultant to Alphatech, Inc., Aptima, Inc., and IBM Research and
Development. He is a Co-Founder of Qualtech Systems, Inc., which is a small
business specializing in intelligent diagnostic software tools. His research in-
terests include the areas of adaptive organizations for dynamic and uncertain
environments, multiuser detection in wireless communications, signal
process- ing and diagnosis techniques for complex system monitoring, and
multiobject tracking.
Dr. Pattipati was selected by the IEEE Systems, Man, and Cybernetics
(SMC) Society as the Outstanding Young Engineer of 1984 and was the
recipient of the Centennial Key to the Future award. He also contributed to
discrete-optimization algorithms for large-scale systems and team decision
making. He served as the Editor-in-Chief of the IEEE TRANSACTIONS ON SYSTEMS,
MAN, AND CYBER-
NETICS: PART B-CYBERNETICS during 1998–2001, Vice-President for Technical
Activities of the IEEE SMC Society (1998–1999), and as Vice-President for
Conferences and Meetings of the IEEE SMC Society (2000–2001). He was
corecipient of the Andrew P. Sage award for the Best SMC Transactions Pa-
per for 1999, the Barry Carlton award for the Best Aerospace and Electronics
Systems Transactions Paper for 2000, the 2002 and 2008 NASA Space Act
Awards for “A Comprehensive Toolset for Model-based Health Monitoring
and Diagnosis,” the 2003 American Association of University Professors
(AAUP) Research Excellence Award and the 2005 School of Engineering
Teaching Excellence Award from the University of Connecticut. He also won
the best technical paper awards at the 1985, 1990, 1994, 2002, 2004, and 2005
IEEE AUTOTEST Conferences, as well as at the 1997 and 2004 Command
and Con- trol Conferences.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of

View publication stats