This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY 1

Observability Analysis and State Estimation of Lithium-Ion

Batteries in the Presence of Sensor Biases
Shi Zhao, Stephen R. Duncan, Member, IEEE, and David A. Howey, Member, IEEE

Abstract— This brief investigates the observability of one of constraints (PDAEs) that describe the thermodynamics, reac-
the most commonly used equivalent circuit models (ECMs) for tion kinetics, and transport within a battery cell.
lithium-ion batteries and presents a method to estimate the Due to the high computational complexity of applying
state of charge in the presence of sensor biases, highlighting
the importance of observability analysis for choosing appropri- state and parameter estimation algorithms to PDAEs, electro-
ate state estimation algorithms. Using a differential geometric chemical models are more difficult to implement in a BMS,
approach, necessary and sufficient conditions for the nonlinear despite the fact that they are able to accurately describe the
ECM to be observable are derived and are shown to be different battery dynamics in a wider operating range. There have been
from the conditions for the observability of the linearized model. some attempts to estimate the battery states using simpli-
It is then demonstrated that biases in the measurements, due to
sensor aging or calibration errors, can be estimated by applying a fied electrochemical models. For example, the single particle
nonlinear Kalman filter to an augmented model where the biases model (SPM), a simplification of the P2D model, is used
are incorporated into the state vector. Experiments are carried in many studies for state estimation [4]–[6]. The problem
out on a lithium-ion pouch cell and three types of nonlinear with the SPM is that it is valid only under low C-rates and
filters, the first-order extended Kalman filter (EKF), the second- the accuracy deteriorates when the applied current becomes
order EKF, and the unscented Kalman filter, are applied using
experimental data. The different performances of the filters are highly dynamic. State estimation of the full P2D model using
explained from the point of view of observability. a modified extended Kalman filter (EKF) has recently been
reported in [7], where the electrochemical model is solved
Index Terms— Battery, equivalent circuit model (ECM),
Kalman filtering, observability, sensor bias, state estimation. by the Chebyshev orthogonal collocation method to reduce
the computational cost. However, parameter estimation of the
model remains a significant challenge.
I. I NTRODUCTION For these reasons, currently a typical BMS uses an ECM,
although the parameters of the model are less easy to interpret

A BATTERY management system (BMS) consists of hard-

ware and software that ensure the safe charging and
discharging of battery cells [1]. A key function of a BMS
in relation to the physics. SOC estimation based on ECMs has
been studied extensively and one major theme for research in
this area is to improve the accuracy of an ECM by either
is accurate estimation of the state of charge (SOC). As battery employing more sophisticated parametrization algorithms or
SOC is not directly measureable, it has to be inferred from the increasing the complexity of the model itself [2], [8]. As for
available measurements, such as the applied current, terminal state estimation algorithms, nonlinear Kalman filters such as
voltage, and surface temperature. A mathematical model that the first-order EKF and sigma-point Kalman filters are among
characterizes the dynamics of a battery is essential for state the most popular choices [8], [9].
estimation. Although there are some articles that examine the observ-
Generally speaking, there are two types of battery mod- ability of battery models [4], [10], generally, it is an over-
els: equivalent circuit models (ECMs) and electrochemical looked topic. In many papers on battery state estimation, the
models. ECMs are low-order models parameterized from time- observability of the models is not considered [2]. Observability
domain or frequency-domain experimental data using system analysis is important because it is not possible to estimate the
identification techniques. The accuracy of an ECM is usu- states of a battery if the model is not observable, regardless of
ally quantified by how closely the model output matches how well the model matches the input–output behavior of the
the experimental measurements when the battery is under a battery. For this reason, when choosing an ECM for a BMS,
dynamic current load [2]. On the other hand, electrochem- the emphasis should not be placed solely on finding the model
ical models such as the pseudo-2-D (P2D) model [3] are that gives the smallest output error compared with the exper-
derived from electrochemical principles and usually charac- imental measurements; the observability of the model should
terized by partial differential equations coupled with algebraic also be considered. Moreover, even if the model is observable,
the state estimation algorithm still needs to be chosen with
Manuscript received October 5, 2015; revised February 1, 2016; accepted
March 5, 2016. Manuscript received in final form March 11, 2016. This work care to account for the reason for the model observability.
was supported by the U.K. Research Councils through the RCUK Energy Pro- For example, if the nonlinear model is locally observable
gramme’s STABLE-NET Project under Grant EP/L014343/1. Recommended everywhere but its linearized version is not observable, then a
by Associate Editor H. R. Pota. (Corresponding author: Shi Zhao.)
The authors are with the Department of Engineering Science, Univer- first-order EKF may fail to track the SOC accurately [10].
sity of Oxford, Oxford OX1 3PJ, U.K. (e-mail: [Link]@[Link]; In this brief, we use a differential geometric approach
[Link]@[Link]; [Link]@[Link]). to analyze the nonlinear observability of a second-order
Color versions of one or more of the figures in this paper are available
online at [Link] RC model for batteries and derive the necessary and sufficient
Digital Object Identifier 10.1109/TCST.2016.2542115 condition for the model to be observable. It is shown that
1063-6536 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

the local observability of the model at a particular state is some input functions u. By contrast, for a linear system,
different from the observability of the linearized model around y(x 1, u, t) = y(x 2 , u, t) should hold for any u if the two
this state. We then show that the battery states can still be states are distinguishable. This means that in general it is
estimated accurately even when there exist unknown biases in more difficult to distinguish two states of a nonlinear system
the sensors. This is because the augmented model that incorpo- because it may not be trivial to find an input u that gives
rates the unknown biases into the states is locally observable. different output functions. Moreover, for a linear system, local
Experiments are carried out on a lithium nickel manganese observability implies global observability. However, this is no
cobalt oxide (NMC) pouch cell (Kokam 740 mAh) and three longer the case for a nonlinear system.
types of nonlinear Kalman filters: the first-order EKF, the
second-order EKF, and the unscented Kalman filter (UKF) are B. Observability Rank Test
applied to estimate the SOC of the cell using experimental
The observability of a nonlinear system can be determined
data. The result shows that while the first-order EKF may
by a rank test that involves Lie derivatives. Suppose that
not be able to estimate the SOC when there is a sensor bias,
h(x) = [h 1 (x) . . . h p (x)]T is a p-dimensional vector function
the estimation given by the UKF has a much better accuracy.
on X and its j th component h j (x) is a real-valued smooth
The difference is explained by an observability analysis of the
function. Denote the gradient of h j as dh j , that is
augmented model.  
The brief is organized as follows. Section II introduces the ∂h j ∂h j ∂h j
dh j = ...
definition and criteria for nonlinear observability. The ECM ∂ x1 ∂ x2 ∂ xn
under consideration is briefly discussed and its observability then the Lie derivative of h j with respect to f is a real-valued
is analyzed in Section III. Section IV presents the observability function defined by
analysis of augmented battery models with sensor biases
incorporated into the states. Experimental validation is carried 
n
∂h j
L f h j = dh j · f = fi
out in Section V, and Section VI concludes this brief. ∂ xi
i=1

II. N ONLINEAR O BSERVABILITY where f (x) = [ f 1 (x) . . . f n (x)]T . The zeroth-order

The observability of a linear time-invariant system described Lie derivative L 0f h j is h j itself and the second-order
by a state-space model can be readily determined by checking Lie derivative L 2f h j is defined by L 2f h j = L f L f h j .
whether the observability matrix has full rank. By contrast, The following theorem [11] gives the rank test for local
the observability of a nonlinear system is less straightforward observability of the nonlinear system.
to determine. Although it is tempting to study the observ- Theorem 1: System (1) is locally observable at x 0 ∈ X if
ability by linearizing the system model, there are impor- there are n linearly independent rows in the set
tant differences between linear observability and nonlinear
(dL zs L zs−1 . . . L z1 h j )(x 0 )
observability [11].
where s ≥ 0, z k ∈ { f, g1 , . . . , gm } for k = 1, . . . , s and
A. Distinguishability and Observability j = 1, . . . , p and with s = 0, the expression is defined as
equivalent to dh j (x 0 ).
Denote X as an open subset of Rn . The nonlinear system
under consideration is described by the following state-space The observability rank test for a linear system can be derived
from this theorem. It should be noted that in the theorem,
model:
there is no upper bound for s, which means that the number of
m
ẋ = f (x) + u i gi (x) (1a) Lie derivatives in the calculation is not bounded from above for
i=1
a general nonlinear system (1). In addition, the rank condition
y = h(x) (1b) given by the theorem is only a sufficient but not a necessary
condition for system (1) to be locally observable.
where x ∈ X is the state, u i ∈ R is the input, y ∈ R p is
the output, f : X → Rn , gi : X → Rn , and h : X → R p III. O BSERVABILITY OF AN ECM
are all smooth functions. We now introduce the following
We now study the observability of battery models using the
definition [11].
rank test introduced in the previous section.
Definition 1: Consider system (1) and two states x 1 and x 2 .
Denote the system output at time t with initial state x i and
input u as y(x i , u, t), where i = 1, 2. x 1 and x 2 are said to A. Battery ECM
be distinguishable if there exists an input function u such that The model considered here is the second-order RC model
y(x 1 , u, t) = y(x 2 , u, t) for a finite t. System (1) is locally shown in Fig. 1, which is one of the most widely used ECMs
observable at x 1 if there exists a neighborhood N of x 1 such in the literature [2]. The open-circuit voltage (OCV) VOC is a
that the only state in N that is not distinguishable from x 1 is x 1 nonlinear function of the SOC and Weng et al. [12] summarize
itself. The system is said to be locally observable if it is some of the popular OCV models, which are all smooth
observable at every x ∈ X. nonlinear functions. The VOC is fitted to approximate the
One subtlety is that two states of a nonlinear system OCV–SOC data collected using the galvanostatic intermittent
may be distinguishable even if y(x 1 , u, t) = y(x 2 , u, t) for titration technique (GITT) or to charge and discharge 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 pagination.

ZHAO et al.: OBSERVABILITY ANALYSIS AND STATE ESTIMATION OF LITHIUM-ION BATTERIES 3

And it can be shown by mathematical induction that

 
1 1
dL f h = −
k
− 0
(−τ1 )k (−τ2 )k
 
1 d k+1 VOC
dL kg h = 0 0
(−Q)k d Z k+1

for all k ∈ Z+ and Lie derivatives involving both f and g

such as L f L g h are constants, and thus can be ignored in the
Fig. 1. Schematic of the ECM.
rank test.
Model (2) is locally observable at a point x 0 if the set
cell at a low constant current (e.g., 1/25C) under constant tem- ⎡ ⎤
perature. The parameters R1 , R2 , C1 , C2 , and Rs are identified dh
⎢dL h ⎥
such that the model output matches the experimental measure- ⎢ f ⎥
⎢ ⎥
ments as closely as possible. In this model, these parameters ⎢ dL g h ⎥
⎢ 2 ⎥
are set to be constants, but we note that the model accuracy O = ⎢dL h ⎥ (4)
⎢ f ⎥
can be improved if the parameters are updated iteratively and ⎢ 2 ⎥
⎢ dL g h ⎥
allowed to vary with the SOC and temperature, at the cost of ⎣ ⎦
..
higher computational complexity. If the parameters are state .
dependent, the observability of the model can still be analyzed
using Theorem 1, although the computation becomes more evaluated at x 0 has n = 3 linearly independent row vectors.
laborious. One thing to observe is that the rank condition does not hold
The state-space representation of the model in the continu- if τ1 = τ2 . Physically, this is because when τ1 = τ2 , the
ous form is given by two RC pairs can be combined into a single one, so that the
⎡ 1 ⎤ voltage across either RC pairs cannot be uniquely determined.
⎡ 1 ⎤
⎡ ⎤ − 0 0 ⎡ ⎤ ⎢ C1 ⎥ In practice, the two time constants are of different magnitudes
V̇1 ⎢ R1 C 1 ⎥ V1 ⎢ ⎥ and in the following analysis, we assume that τ1 = τ2 .
⎣V̇2 ⎦ = ⎢ ⎢ 0 1 ⎥⎣ ⎦ ⎢ 1 ⎥
⎥ V2 + ⎢ ⎥I (2a) Under this assumption, O(x 0 ) has full column rank if and
⎣ − 0⎦ ⎢ C2 ⎥
Ż R2 C 2 Z ⎣ 1⎦ only if there exists a k ∈ Z+ such that
0 0 0 −
Q  
d k VOC
V = VOC (Z ) − V1 − V2 − I Rs (2b) (x 0 ) = 0. (5)
d Zk
where V1 and V2 are the voltages across the first and second
RC pairs, respectively, Z ∈ [0%, 100%] is the normalized This means that battery model (2) is observable if all the
SOC, Q is the battery capacity, I is the current applied to the derivatives of VOC are not zero simultaneously.
battery, and V is the terminal voltage. Note that the sign of the As an aside, note that since the state equation (2a)
input I is taken to be positive when the battery is discharging is linear and the system is controllable, according to
and negative during charging. For ease of reference, define [13, Proposition 3.38], the rank condition is also a necessary
τ1 = R1 C1 and τ2 = R2 C2 . condition for model (2) to be locally observable at x 0 . Thus,
The state equation (2a) is linear and the only nonlinearity if (5) is not satisfied, the system is not locally observable.
in the model comes from VOC in the output equation (2b). In the literature, it is common to simply equate the
To apply the nonlinear observability theorem to the system, local observability of a nonlinear system with the observ-
we first rewrite the model in the following form: ability of its linearized system. Linearizing model (2)
at x 0 gives
ẋ = f (x) + gu (3a)
y = h(x) − Rs u (3b) ẋ = Ax + Bu
where x = [V1 , V2 , Z ]T , u = I, y = V, f (x) = y = C x + Du
[−(1/τ1 )V1 − (1/τ2 )V2 0]T , g = [(1/C1 ) (1/C2 ) − (1/Q)]T ,
and h(x) = VOC (Z ) − V1 − V2 . The feed-through term Rs u where
in the output equation does not affect the applicability of ⎡ 1 ⎤
1 ⎡ ⎤
Theorem 1 to the model.
− 0 0 ⎢ C1 ⎥
⎢ τ1 ⎥ ⎢ ⎥
⎢ ⎥ ⎢ 1 ⎥
A=⎢ , B =⎢ ⎢ ⎥
0⎥
1
B. Observability Analysis ⎣ 0 − ⎦ ⎢ C2 ⎥
⎥
τ2 ⎣ 1 ⎦
Using the notation introduced in Section II-B, the gradient 0 0 0
of h(x) with respect to x is −Q
     
d VOC d VOC
dh = −1 − 1 . C = −1 −1 (x 0 ) , D = −Rs .
dZ dZ
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

The observability of this linearized model can be determined Such a model is called an augmented model, while (2) is called
by checking the rank of the observability matrix the original model.
⎡   ⎤ The augmented model can be rewritten into the form in (3)
d VOC
−1 −1 (x ) where x = [V1 , V2 , Z , Ve ]T , u = I, y = Vm , f (x) =
⎡ ⎤ ⎢ dZ
0 ⎥
C ⎢ ⎥ [−(1/τ1)V1 − (1/τ2 )V2 0 0]T , g = [(1/C1 ) (1/C2 ) −
⎢ 1 1 ⎥
⎣CA ⎦ = ⎢ 0 ⎥. (1/Q) 0]T , and h(x) = VOC (Z ) − V1 − V2 + Ve . Note that the
⎢ τ τ ⎥
CA 2 ⎢ 1 2 ⎥ augmented model is not controllable.
⎣ 1 1 ⎦
− 2 − 2 0 For the case that there is an unknown bias in the current
τ1 τ2
sensor but no bias in the voltage sensor, the model can be
This matrix has full rank if and only if (d VOC /d Z )(x 0 ) = 0. reformulated in the same manner. Denote the bias by Ie and
However, this is not a necessary condition for nonlinear the inaccurate current measurement by Im , then I = Im − Ie is
system (2) to be locally observable at x 0 , so that system (2) the actual applied current. As a result, battery model (2) now
may be observable even if its linearization is not. becomes
IV. E STIMATION W ITH S ENSOR B IASES ⎡ ⎤ ⎡ 1 ⎤
1 1
− 0 0 −
The accuracy of state estimation is known to be dependent ⎡ ⎤ ⎢ τ1 C1 ⎥ ⎡ ⎤ ⎢ C1 ⎥
V̇1 ⎢ ⎥ ⎢ 1 ⎥
⎢ 1 1 ⎥ V1
on the quality (precision and accuracy) of the sensors that give ⎢V̇2 ⎥ ⎢ 0 − 0 − ⎥⎢V2 ⎥ ⎢ ⎢
⎥
⎥
the input and output measurements. It is commonly assumed ⎢ ⎥ =⎢ τ C ⎥⎢ ⎥ + ⎢ 2 ⎥ Im (7a)
⎣ Ż ⎦ ⎢ 2 2 ⎥⎣ ⎦
Z ⎢
C
⎥
that the current and voltage sensors in model (2) are subject ⎢ 1 ⎥ ⎢ 1⎥
I˙e ⎢ 0 0 0 ⎥ Ie ⎣ − ⎦
to zero-mean Gaussian noise, where the sensor precisions ⎣ Q ⎦ Q
are described by the variances of the Gaussian distributions. 0 0 0 0 0
However, the sensors may also have some unknown biases V = VOC (Z ) − V1 − V2 − Im Rs + Ie Rs . (7b)
due to aging or calibration errors, which are referred to as a
sensor bias fault, and there are mainly two ways to detect the When there are biases in both the current and voltage
fault: hardware redundancy and analytical redundancy [14]. sensors, an augmented model including both Ie and Ve in
In the hardware redundancy approach, redundant sensors are the state vector can be easily derived. For brevity, the model
used to measure one variable and a fault is detected if there is is not given here. In the following analysis, the focus is
a notable difference between the measurements from different mainly on the first case, which is described by model (6).
sensors. In the analytical redundancy approach, the system is The analysis can be readily extended to other cases and is not
analyzed using multiple-model-based techniques. repeated.
When there is a sensor bias fault, it is expected that the
accuracy of the state estimation deteriorates when the effect
of a sensor bias builds up over time. One approach that is B. Observability Analysis
able to estimate the sensor biases and the battery SOC at the
same time, without incorporating any additional sensors into It can be shown that for model (6)
the system, is to augment model [15]. As in [14] and [15], the  
sensor biases are modeled as constants here. d VOC
dh = −1 − 1 1
dZ
 
A. ECM With Sensor Biases 1 1
dL kf h = − − 0 0
We first consider the case that there is an unknown bias (−τ1 )k (−τ2 )k
 
in the voltage sensor, while there is no bias in the current 1 d k+1 VOC
sensor. Denote the voltage sensor bias by Ve and the inaccurate dL kg h = 0 0 0
(−Q)k d Z k+1
voltage measurement by Vm , then Vm −Ve represents the actual
battery terminal voltage denoted by V . The difficulty is the for all k ∈ Z+ . Once again, Lie derivatives involving both
bias Ve is unknown and if we simply ignore it, the estimation f and g can be ignored in the rank test since they are all
based on model (2) and Vm will be inaccurate. The problem constants.
can be solved by including the bias Ve into the state vector Model (6) is locally observable at a point x 0 if the set O
and reformulating the model. Specifically, since V = Vm − Ve , [see (4) in Section III-B] evaluated at x 0 has n = 4 linearly
model (2) can now be reformulated as independent row vectors. This rank condition is satisfied if and
⎡ 1 ⎤
only if there exists an integer k ≥ 2 such that
⎡ 1 ⎤
⎡ ⎤ ⎢C ⎥
− 0 0 0 ⎡ ⎤ ⎢ 1 ⎥ 
d k VOC

V̇1 ⎢ τ1 ⎥ V1 ⎢ ⎥
⎢V̇2 ⎥ ⎢ ⎥⎢ ⎥ ⎢ 1 ⎥ (x 0 ) = 0. (8)
⎢ ⎥=⎢ ⎥ V2 ⎥ ⎢ C ⎥
1 d Zk
− ⎢
0 0⎥ ⎣ ⎦ + ⎢ 2 ⎥ I
⎣ Ż ⎦ ⎢ ⎢
0
τ2 ⎥ Z ⎢ ⎥
(6a)
⎣ 0 ⎦ ⎢ 1 ⎥
V̇e 0 0 0 Ve ⎢− ⎥ If condition (8) does not hold, then dL ig h is a
⎣ Q⎦
0 0 0 0 zero row vector for any i ∈ Z+ , which means that
0 O(x 0 ) can at most have rank 3. Therefore, the neces-
Vm = VOC (Z ) − V1 − V2 − I Rs + Ve . (6b) sity is proved. Now suppose the condition is satisfied,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHAO et al.: OBSERVABILITY ANALYSIS AND STATE ESTIMATION OF LITHIUM-ION BATTERIES 5

the following matrix:

⎡ ⎤
d VOC
−1 −1 (x 0 ) 1
⎡ ⎤ ⎢ ⎢ dZ ⎥
⎥
dh ⎢ 1 1 ⎥
⎢ dL f h ⎥ ⎢ 0 0⎥
⎢ ⎥ ⎢ τ1 τ2 ⎥
⎢ ⎥=⎢ ⎢
⎥
⎥
⎣ dL f h ⎦ ⎢− 1
2
− 2
1
0 0⎥
⎢ τ2 τ2 ⎥
dL k−1
g h ⎢ 1 ⎥
⎣ 1 d k VOC ⎦
0 0 (x 0 ) 0
(−Q) k−1 d Zk
has full rank. This can be verified by calculating the
determinant of the matrix. Thus, the condition is also
a sufficient one.
We note that since system (6) is not controllable, the rank Fig. 2. OCV as a function of SOC.
condition is no longer a necessary condition for the system
to be locally observable at x 0 . Therefore, condition (8) is a TABLE I

necessary and sufficient condition for O(x 0 ) to have full rank, VALUES OF THE OCV P OLYNOMIAL C OEFFICIENTS
which is only a sufficient condition for system (6) to be locally
observable at x 0 .
It is interesting to note that the observability of system (6)
crucially depends on the nonlinearity of VOC . If the OCV
is a linear function of SOC such as in a capacitor, the
rank condition is not satisfied at any x 0 and system (6) is
unobservable.
A similar analysis can be applied to system (7). Although
the calculation is more laborious, it turns out that the condition
for system (7) to be locally observable at a point x 0 is the same
as that for system (6).

V. E XPERIMENTAL VALIDATION
TABLE II
In this section, using experimental data, several nonlinear VALUES OF M ODEL PARAMETERS
Kalman filters based on model (6) are implemented to estimate
the states of a lithium-ion battery when there is an unknown
bias in the voltage sensor.

A. Battery Test
The experiments were carried out on a 740-mAh Kokam
NMC lithium-ion pouch cell (SLPB533459). The data were
The parameters are identified using MATLAB’s system iden-
collected using a BioLogic SP-150 potentiostat and the tem-
tification toolbox and are given in Table II. The volt-
perature is fixed at 20 ◦C using a thermal chamber. A GITT
age measurements and the predicted output by the model
procedure with current 0.1C was used to determine the rela-
under the FUDS load are plotted in Fig. 3(c). To further
tionship between the OCV and SOC, and 50 data points were
validate the model, an ARTEMIS European urban driving
recorded in both the charge and discharge processes [16].
cycle (UDC) [18] shown in Fig. 3(b) is applied to the same
As shown in Fig. 2, there is a discrepancy between the charge
fully charged cell and simulation results under the UDC
and discharge curves due to hysteresis. The average of these
load using the identified parameters are compared with the
two curves is taken as the OCV–SOC
function and is approx-
experimental voltage measurements, as shown in Fig. 3(d). The
imated by a polynomial VOC = 12 k
k=0 ak Z , where the coeffi- maximal error is 21.1 mV and the RMSE is 4.78 mV, which
cients are given in Table I and the fitted polynomial is plotted
corresponds to 0.13% of the mean battery terminal voltage.
in Fig. 2. The maximal approximation error of this polynomial
in the SOC range [10%, 100%] compared with the aver-
aged OCV–SOC curve is 17.97 mV and the root-mean-square B. State Estimation Using Various Kalman Filters
error (RMSE) is 3.85 mV. To verify that the SOC can be estimated in the presence of
Before applying state estimation algorithms to either unknown sensor biases, we implement several nonlinear filters
model (2) or (6), the model parameters need to be identified. to model (6) using the experimental measurements of the cell
We apply a federal urban driving schedule (FUDS) current under the UDC load. The biased voltage measurements are
profile [17] shown in Fig. 3(a) to the cell with 100% initial obtained by adding a constant to the recorded voltage mea-
SOC and record the applied current and terminal voltage. surements. The constant is the voltage sensor bias, which is
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 3. Current profiles and model validation results. (a) FUDS current Fig. 4. State estimation based on model (2) using the first-order EKF when
profile, training dataset. (b) UDC current profile, validation dataset. there is a bias in the voltage sensor. (a) Evolution of the true battery terminal
(c) Experimental and model outputs under the FUDS load. (d) Experimental voltage (blue curve), the measured voltage (green curve), and the prediction
and model outputs under the UDC load. by the EKF (red curve). (b) Evolution of the true SOC (blue curve) and the
estimated SOC (red curve).

considered unknown at the beginning of the state estimation

process. The SOC calculated by Coulomb counting using the everywhere for the polynomial function VOC considered here.
accurate lab equipment is regarded as the true SOC [12]. Thus, it is possible to estimate the SOC and the bias at the
The filtering algorithms implemented here are the first- same time based on inaccurate measurements. However, the
order EKF, the second-order EKF, and the UKF. Since these first-order EKF still performs badly even though the estimation
algorithms are well known and widely used for battery state is now based on model (6), as shown in Fig. 5. The initial error
estimation, the details are not discussed here and we refer covariance P0 is taken as a diagonal matrix with diagonal
the reader to [19] for a comprehensive introduction. The entries [0.01 0.0016 0.01 0.0625], the process noise Q 0 is
tuning parameters of the filters include the initial state error taken as a diagonal matrix with vector [10−8 10−8 10−8 10−8 ]
covariance P0 , the process noise covariance Q 0 , and the in the diagonal, the measurement noise standard deviation is
measurement noise covariance R. For the UKF, there is an taken as σV = 6 mV, which means that R = 3.6 × 10−5 , and
extra parameter κ, which is used to tune the weights of the the time step is set to 1 s. The initial value of the estimated
sigma points [20]. bias is set to zero, which means that it is estimated that there
Consider the case that there is a bias Ve = 100 mV in the is no bias at the beginning. One might attribute the failure of
measured voltage. After one complete driving cycle (1000 s), the EKF to a poor implementation such as inappropriate initial
the true SOC of the cell is about 85% and the corresponding values for the covariance matrices. However, tuning the EKF
OCV is 4.025 V. Suppose the state estimation process starts does not seem to improve the estimation accuracy much.
at this point and due to the presence of the sensor bias, the The observability analysis of model (6) shows that it is
initial value of the estimation on the SOC is set to 95%, which locally observable because of the presence of nonlinearity.
means that there is an error of about 10% at the beginning. The It is easy to verify that the observability is lost once the model
guessed initial SOC corresponds to an OCV of 4.123 V, which is linearized. In the first-order EKF, the estimation at every
is 98 mV higher than the true value. If the estimation is based time step is based on a linearization of the model. Thus, it
on model (2), which means that the inaccurate measurement is is not surprising that the battery states cannot be estimated
trusted, it is not surprising that the estimated states are far from accurately.
their actual values. As shown in Fig. 4, the SOC estimated by The estimation accuracy may be improved by using the
the first-order EKF has considerable errors, due to the biased second-order EKF, which retains the quadratic terms in non-
voltage measurement on which the estimation is based. Note linear functions. Fig. 5 shows that the second-order EKF with
that the estimated voltage in Fig. 4(a) is calculated from the the same filter parameters does have some advantages over the
output equation (2b) using the a posteriori state estimate. The first-order EKF for this problem.
estimated SOC has an RMSE of 15.6% between t = 1000 s Among the three types of nonlinear Kalman filters, the UKF,
and t = 2000 s. Using the second-order EKF or the UKF does which makes use of the unscented transform to account for the
not improve the estimation accuracy as the sensor bias is still nonlinearity, provides the most accurate estimation. With the
present. same filter parameters and κ = 4, the estimated bias soon
The analysis in Section IV-B gives a sufficient condi- converges to the true value and its RMSE between t = 1000 s
tion for model (6) to be locally observable at a point. and t = 2000 s is 3.41 mV, which is of the order of the voltage
It can be easily verified that this condition is satisfied measurement noise standard deviation. The SOC estimation
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHAO et al.: OBSERVABILITY ANALYSIS AND STATE ESTIMATION OF LITHIUM-ION BATTERIES 7

an error in the estimated model output compared with that in

the measured output is caused by the current sensor error or
the voltage sensor error.

VI. C ONCLUSION
The importance of observability analysis of nonlinear ECMs
for battery state estimation has not received much attention.
This brief analyzes the nonlinear observability of a widely
used second-order RC model for lithium-ion batteries. Using
the ECM as an example, it is pointed out that the local
observability of a nonlinear dynamic system at a certain
point is not the same as the observability of the system
linearized around the point. We then present a method to
estimate the battery SOC when there are unknown biases
in the sensors. The estimation is based on an augmented
model that incorporates the sensor biases into the state vector.
Fig. 5. State estimation based on model (6). (a) Voltage sensor bias The observability of the augmented model is shown to be
(blue curve) and the estimated bias by the first-order EKF (green curve), the dependent on the model nonlinearity. Experimental validation
second-order EKF (black curve), and the UKF (red curve). (b) Evolution of shows that it is vital to take the nonlinearity of the model into
the true SOC (blue curve) and the estimated SOC by the first-order EKF
(green curve), the second-order EKF (black curve), and the UKF (red curve). account in the estimation algorithm, highlighting the impor-
tance of observability analysis for state estimation. We also
show that accurate voltage measurement is critical for SOC
error quickly decreases to less than 1% and the RMSE between
estimation in a BMS using the original model, whereas current
t = 1000 s and t = 2000 s is 0.33%. We note that the results
measurement accuracy has a smaller impact as the Kalman
are very similar when the voltage sensor bias Ve is negative
filters are relatively robust against current sensor bias (due
or zero or when the initial guess of the SOC is smaller than
to their knowledge of the OCV–SOC ground truth and ECM
the true value.
parameters). If the model parameters and OCV curve are prone
A bias in the current sensor has smaller impact on the
to change over time, this makes the problem significantly more
performances of the nonlinear filters. When there is only a
challenging, but is a topic for future investigation.
bias Ie = −100 mA in the current measurement, the SOC
estimation given by the first-order EKF based on model (2)
has an RMSE of 2.80% between t = 1000 s and t = 2000 s. ACKNOWLEDGMENT
This means that the state estimation algorithm is relatively The authors would like to thank C. Birkl for providing his
robust against current sensor biases [21]. Nevertheless, the help in the experiments.
estimation accuracy can be further improved if the UKF is
applied to model (7), with the RMSE of the estimated SOC in R EFERENCES
the same time range being reduced to 0.43%. Furthermore, the
current sensor bias Ie can be tracked accurately by the UKF. [1] D. A. Howey and S. M. M. Alavi, “Rechargeable battery energy
storage system design,” in Handbook of Clean Energy Systems, vol. 5.
As a comparison, the second-order EKF based on (7) has a New York, NY, USA: Wiley, 2015.
similar accuracy and the first-order EKF is the least accurate [2] X. Hu, S. Li, and H. Peng, “A comparative study of equivalent circuit
in estimating the SOC and the bias. models for Li-ion batteries,” J. Power Sour., vol. 198, pp. 359–367,
Jan. 2012.
The purpose of this section is not to claim that the UKF is [3] M. Doyle, T. F. Fuller, and J. Newman, “Modeling of galvano-
always superior to the first-order EKF, for if there are no sensor static charge and discharge of the lithium/polymer/insertion cell,”
biases, then the first-order EKF based on model (2) gives as J. Electrochem. Soc., vol. 140, no. 6, pp. 1526–1533, 1993.
[4] S. J. Moura, N. A. Chaturvedi, and M. Krstić, “Adaptive partial
accurate an estimation as the UKF does. The objective is rather differential equation observer for battery state-of-charge/state-of-health
to show that it is important to analyze the observability of a estimation via an electrochemical model,” J. Dyn. Syst., Meas., Control,
model before choosing an estimation algorithm, with particular vol. 136, no. 1, p. 011015, Oct. 2013.
[5] D. Di Domenico, A. Stefanopoulou, and G. Fiengo, “Lithium-ion
relevance to battery SOC estimation. In addition, it presents a battery state of charge and critical surface charge estimation using an
method to estimate the sensor bias without incorporating any electrochemical model-based extended Kalman filter,” J. Dyn. Syst.,
additional sensors into the system. Meas., Control, vol. 132, no. 6, p. 061302, 2010.
[6] A. Bartlett, J. Marcicki, S. Onori, G. Rizzoni, X. G. Yang, and T. Miller,
Note that in the case that there are biases in both the voltage “Electrochemical model-based state of charge and capacity estimation
and current sensors, it is probably not possible to accurately for a composite electrode lithium-ion battery,” IEEE Trans. Control Syst.
estimate the SOC and the biases at the same time using Technol., vol. 24, no. 2, pp. 384–399, Mar. 2016.
[7] A. M. Bizeray, S. Zhao, S. R. Duncan, and D. A. Howey, “Lithium-
an augmented model, even when the model appears locally ion battery thermal-electrochemical model-based state estimation using
observable. This might be because usually the observability orthogonal collocation and a modified extended Kalman filter,” J. Power
matrix is ill conditioned despite its full rank. Intuitively, this Sour., vol. 296, pp. 400–412, Nov. 2015.
[8] G. L. Plett, “Extended Kalman filtering for battery management systems
makes sense: if we have both a current sensor bias and of LiPB-based HEV battery packs: Part 1. Background,” J. Power Sour.,
a voltage sensor bias, there is ambiguity regarding whether vol. 134, no. 12, pp. 252–261, 2004.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

[9] G. L. Plett, “Sigma-point Kalman filtering for battery management [16] C. R. Birkl, E. McTurk, M. R. Roberts, P. G. Bruce, and D. A. Howey,
systems of LiPB-based HEV battery packs: Part 1: Introduction and “A parametric open circuit voltage model for lithium ion batteries,”
state estimation,” J. Power Sour., vol. 161, no. 2, pp. 1356–1368, 2006. J. Electrochem. Soc., vol. 162, no. 12, pp. A2271–A2280,
[10] X. Lin, A. G. Stefanopoulou, Y. Li, and R. D. Anderson, “State of 2015.
charge imbalance estimation for battery strings under reduced volt- [17] Electric Vehicle Battery Test Procedures Manual, Idaho Nat. Eng.
age sensing,” IEEE Trans. Control Syst. Technol., vol. 23, no. 3, Environ. Lab., Idaho Falls, ID, USA, 1996.
pp. 1052–1062, May 2015. [18] M. André, “The ARTEMIS European driving cycles for measuring car
[11] M. Vidyasagar, Nonlinear Systems Analysis, vol. 42. Philadelphia, PA, pollutant emissions,” Sci. Total Environ., vols. 334–335, pp. 73–84,
USA: SIAM, 2002. Dec. 2004.
[12] C. Weng, J. Sun, and H. Peng, “A unified open-circuit-voltage model [19] D. Simon, Optimal State Estimation: Kalman, H∞ , and Nonlinear
of lithium-ion batteries for state-of-charge estimation and state-of-health Approaches. New York, NY, USA: Wiley, 2006.
monitoring,” J. Power Sour., vol. 258, pp. 228–237, Jul. 2014. [20] S. Julier, J. Uhlmann, and H. F. Durrant-Whyte, “A new method for
[13] H. Nijmeijer and A. van der Schaft, Nonlinear Dynamical Control the nonlinear transformation of means and covariances in filters and
Systems. New York, NY, USA: Springer-Verlag, 1990. estimators,” IEEE Trans. Autom. Control, vol. 45, no. 3, pp. 477–482,
[14] A. T. Vemuri, “Sensor bias fault diagnosis in a class of nonlinear Mar. 2000.
systems,” IEEE Trans. Autom. Control, vol. 46, no. 6, pp. 949–954, [21] J. K. Barillas, J. Li, C. Günther, and M. A. Danzer, “A comparative
Jun. 2001. study and validation of state estimation algorithms for Li-ion batteries
[15] F. Gustafsson, Adaptive Filtering and Change Detection. New York, NY, in battery management systems,” Appl. Energy, vol. 155, pp. 455–462,
USA: Wiley, 2000. Oct. 2015.