An Observer-Based Fault Diagnosis in Battery Systems of Hybrid Vehicles

Günyaz Ablay

Electrical & Electronics Engineering, Abdullah Gül University, Kayseri, Turkey

[email protected]

Abstract Ni-MH battery systems of the HEVs for fault detection and
isolation under noisy measurements. Fault diagnosis of the
Hybrid electric vehicles (HEVs) currently use Nickel-Metal possible sensor and parameter faults are studied with statistical
Hydride (Ni-MH) batteries which have advantages of design residual evaluations to minimize false alarms.
flexibility, superior power, environmental acceptability and The diagnostic algorithm monitors the battery system for any
recyclability, long life, wide-range operating temperature possible faults and malfunctions. It helps in detecting deviations
and low cost. No matter how good a battery is, a failure can of the battery performance from its normal behavior and also
always occur in a battery leading to serious inconvenience, isolates the possible causes for the faults. In practice, output
performance deterioration and costly replacement. Thus, it signals of the system under consideration are often directly
is desirable to be able to detect the underlying degradation evaluated and compared with a given threshold [6]. However,
and to predict level of unsatisfactory performance. By using such an approach is insufficient in critical processes due to
current, voltage and temperature measurements of Ni-MH possible masking of faults by control systems, and capability to
batteries, they can be modeled so that the internal dynamics react only after a relatively large change in the measured
of the batteries can be estimated and state of health of the variable [7]. If the system is modeled, and the difference
batteries can be predicted for secure and long-life between the measurement and its estimation (residual) is
operations. An observer-based fault diagnosis approach is obtained, then it is theoretically possible to detect every fault
designed to analyze the state of health of the Ni-MH battery with residual generation [8]. Fig. 1 illustrates the general
system of HEVs in this study. Real-world input data is used structure of a fault diagnosis system using model-based
to assess the efficiency of the approach in the existence of methods. The residuals are usually generated from the difference
uncertainties. The possible sensor faults and unexpected between the predictions of the mathematical models of the
parameter deviations are diagnosed efficiently with process under consideration and the actual process input and
statistical evaluation of the generated residuals. outputs in the observer-based methods. Then a residual
evaluation subsystem is used to provide fault information by
1. Introduction comparing the residuals with their nominal values.
Disturbance Faults
The increasing needs for high energy, power, life-cycle, fuel
economy, wide-range operating temperature and
Inputs Outputs
environmentally acceptable batteries in electric and hybrid Actuators Process Sensors
electric vehicles (EVs and HEVs) are the driving forces for
rapid growth of Ni-MH batteries. Ni-MH batteries are dominant
battery technology for HEV applications by having the best Residaul
overall performance in the wide-range requirements set by +
Generator
automobile companies [1]. The functionality and reliability of Noise
Ni-MH batteries as one of the major energy storage device can Residuals
be enhanced with fault diagnosis and health monitoring. Normal Residual
Ni-MH batteries are fabricated in sealed maintenance free Behavior Evaluation
cells so that it is not possible to have close measurement probes.
Hence, the only possible measurements are current and voltage Fault Info
at the poles of the battery. Using these measurements, it is
necessary to acquire state of health of a Ni-MH battery system Fig. 1. A general diagram of the model-based fault diagnosis.
in a HEV in order to prevent serious damage and failures in the
battery. In the literature, most of the studies are provided for the This paper is organized as follows: Section 2 provides a
fault diagnosis and health monitoring of Li-ion batteries with background on Ni-MH battery system of HEV. Fault diagnosis
several methods, including extended Kalman filter, strategy is given in Section 3, and application results are given
autoregressive moving average model and fuzzy logic, which in Section 4. Conclusion of the study is provided in Section 5.
are summarized in [2]. There are a few studies on state
estimation and fault diagnosis of Ni-MH battery systems. A 2. Battery System and Modeling
model-based algorithm is developed to estimate battery
hysteresis for state of charge estimation in [3]. A model-based The HEV which combine an internal combustion engine and
fault diagnosis algorithm for Ni-MH battery is used in [4]. A an electric motor must be supplied by a battery system. The Ni-
combination of artificial network and fuzzy interface is used to MH battery plays a significant role in HEV by supplying power
estimate state of charge of Ni-MH batteries in [5]. In this study, to its electric motor and its electrical accessories. This
an observer-based fault diagnosis methodology is considered for rechargeable battery with mode of constantly charge and

238
discharge has several excellent advantages: flexible cell sizes
from 30 mAh - 250 Ah, safe operation at high voltage (+320 V),  1 1
high volumetric energy and power, maintenance free, excellent =−  + 
     (2)
thermal properties, environmentally acceptable and recyclable  =  −  − 
materials, and simple and inexpensive charging and electronic
control circuits [1]. The electrochemistry of a Ni-MH battery is
where  (in A) is the current at the input, and  (in V) is the
given with the following overall reaction of discharge/charge,
output voltage. All the parameters of (2) are functions of battery
temperature and SOC, and are usually different in charge and
 +  ⇌  + () (1) discharge phases.
State of Charge Model: SOC provides information about
where  is a hydrogen absorbing alloy in the form of
 . remaining useful energy and the remaining usable time of the
Many batteries are connected to each other in the form of battery battery, but its measurement is very difficult. While several SOC
packs to supply high power. As shown in Fig. 2, the battery models have been appeared in the literature [2], [15], one of the
system consists of Ni-MH battery packs and temperature most SOC calculation methods is the current integration [2] as
controller. There are also other electronic systems including given below:
rectifier to charge battery and inverter to adjust battery output,
but these electronic controllers can be considered separately. 1
 =  −  () (3)

Current, I Voltage, V
Ni-MH
Fan Output Temperature, T where  (in A) is the battery current,  is the initial state of
Controller Battery
charge, and  (in Ah) is the nominal capacity. This approach
requires dynamic measurement of the battery current.
Practically, a Ni-MH battery for HEV is never fully charged and
Sensor discharged, and its operating range with respect to SOC is
around 40%-80%.
Fig. 2. Schematic diagram of the battery system. Thermal Model: In electric and hybrid vehicles, it is critical
to achieve performance and extended life of batteries via
2.1. Battery Model thermal management. Suitable modeling for predicting thermal
behavior of battery systems in vehicles can also help to improve
The electric vehicle batteries have been commonly modeled battery design and development process. Therefore, thermal
with two ways: equivalent circuit battery models [9], [10], and models for the batteries have been developed based on thermal
electrochemical battery models [11], [12], [13]. Here, we energy balance of the batteries, and they are coupled with
consider an equivalent circuit battery model that is based on electrochemical or electric models [12], [13]. Based on the
current/voltage relations of the Ni-MH battery. Considering electric model and thermal energy balance of the battery, a
battery (load) current as an input, and battery voltage and simple thermal model can be built as
temperature as outputs, the battery model consists of three sub-
models: electrical model, thermal model, and state of charge   +   ℎ
=  − ( −  ) (4)
(SOC) model.   
Electrical Model: The most commonly used battery models
are ideal model, linear model and Thevenin equivalent model where  (in J/oC) is the effective heat capacity per cell, ℎ (in
[9]. The Thevenin equivalent model with varying parameters J/oC) is the effective heat transfer per cell, and  (in oC) is the
represents a battery with higher accuracy, and thus it is bulk temperature. The effective heat transfer capacity per cell is
considered in this study. This model is made of electrical values given in terms of temperature controller (fan) settings as
of the open-circuit voltage (  ), internal resistance (  ),
capacitance ( ), and the overvoltage resistance ( ) [10]. The ℎ = ℎ (1 + 0.5 ) (5)
Thevenin equivalent model of the battery is depicted in Fig. 3.
where ℎ = 0.07 (/℃) for natural convection, and the fan
Ro setting  depending on the temperature is
R
I
 = 0 off mode ;  = 2 if  ≥ 35 ℃
" (6)
+  = 1 if  ≥ 30 ℃ ;  = 3 if  ≥ 40 ℃
+
Eo Co V
2.2. Dynamical Analysis of the Battery Model
-
The state space form of the battery model can be written by

Fig. 3. Battery Thevenin equivalent model. ̇ −' 0  6 0

$ & = $ * & , - + $ * &  + $ &  (7)
̇ 0 −'  6  '
Using current-voltage relations from Fig. 3, the dynamic
electric equations of the circuit are
where '* = 1/  , ' = ℎ / , 6* = 1/ and 6 = ( +
 )/. It is obvious from system matrix that both states have

239
distinct and stable eigenvalues with 9:<> 9 ≫ |: @ | since all A residual evaluation chart is given in Table 1. As seen from
parameters have positive values and the current and ambient the Table 1, all sensor and internal faults can be detected and
temperature are bounded, i.e. sup|()| < B for B < ∞ and isolated with the given logic.
sup| ()| < B for B < ∞ . These characteristics of the
battery model can be used to design an effective observer-based Table 1. Fault diagnosis chart
fault diagnosis scheme.
Residuals Residual 1 Residual 2
Faults M< =  − E M@ =  − E
3. Fault Diagnosis in the Battery Systems
Temperature Sensor 0 1
Voltage Sensor 1 0
The Ni-MH battery system is a complicated system due to its
feedback loop and parameter-varying model. The equivalent Resistance Change 1 1
electrical model parameters of the battery are functions of
temperature and SOC, which makes fault detection problem It is well-known that the generated residuals can be disturbed
hard. Moreover, the temperature feedback controller can mask by the measurement noise or disturbance. Hence, a systematical
the temperature sensor faults so that the detection problem can threshold selection should be based on the statistical evaluation
pose some other challenges. of the residuals under the absence of the fault and the presence
of the fault. In this study, the thresholds for each residual of
measurements are calculated by comparing probability density
3.1. Residual Generation
functions of the residuals for faulty and non-faulty cases. The
probability density estimations of voltage residual are illustrated
An observer based approach is considered to generate residuals
in Fig. 4. The probability density estimation is obtained in the
in this study. Based on the system model given in (2)-(6), an
absence of fault and for drift-like intermediate fault (see Fig. 6).
observer can be designed by
A similar case is considered for temperature residual M@ .
E 1 1
=− E +  + H* ( − E)
 E G G 1.5

Probability density estimation

Non-faulty residual
E E + E  ℎ J Faulty residual
=  − KE −  L + H K − E L (8)
   1
E E E
 =  −  −  E
J = ℎ K1 + 0.5G L
ℎ
0.5
J ) denotes the estimated values of (, , ℎ )
where (E , E , ℎ
respectively. The observer gains are denoted with H* and H . 0
These observer gains can always be selected appropriately to -1 0 1 2 3
bring the initial errors to zero because the battery system is MR<V
dynamically stable as analyzed in Section 2.2. Since the battery
system has two outputs, voltage and temperature, it is possible Fig. 4. Probability density estimation for voltage
to detect and isolate single faults in temperature and voltage residual.
sensors. More importantly, the possible internal resistance
variations can be detected and isolated. It is seen from Fig. 4 that especially the standard deviations
Residuals are acquired with the following features: for faulty and non-faulty cases are completely different in the
probability density estimation figures, and probability density
estimations for non-faulty cases are similar to Gaussian
M< =  − E distribution, so thresholds can be selected based on the means
(9)
M@ =  − E and standard deviations in the fault-free cases. While there is
always a trade-off in the selection of thresholds, to avoid false
alarms, thresholds for residuals M< and M@ can be selected as
where M< and M@ are residuals for voltage and temperature
< = P< + 4.4Q< , and @ = P @ + 4. 4Q@ where P is the mean
measurements, respectively.
value and Q is the standard deviation of related residuals under
fault-free cases. On the other hand, the 4.4Q value can be
3.2. Residual Evaluation decreased to have more sensitive fault detections.
The generated residuals can be evaluated based on some
threshold selections. There will always be a trade-off between 4. Numerical Results
the avoidance of false alarms and the detection of small faults in
the existence of noise and modeling errors. Practically, a fault The numerical simulation results are acquired with
can be detected only if it causes the residual evaluation function MATLAB/Simulink programs by using experimental current
to surpass the threshold, namely [6], measurements. The data used in this work were acquired as part
of the activities of the Center for Automotive Research (CAR) at
the Ohio State University. It is assumed that all measurements
|MN | ≥ N (10) are exposed to normally distributed measurement noise with a
range of ±10% of the measured states. The time responses of
where MN is the function of the ’th residual and N is the selected the current, voltage and temperature for non-faulty cases are
threshold for ’th residual function. depicted in Fig. 5. The Ni-MH battery has a continuous

240
discharge and charge characteristics, so all other state variables 6

Sensor Fault (oC)

Temperature
are also having a continuous stochastic fluctuations. This makes
4
fault diagnosis problem difficult since during these fluctuations
system parameters are also varying. 2

150 0
0 500
00 (a) 100
1000 1500 2000 2500 3000 3500
Current (A)

100
50 6

Residual (oC)
Temperature
0
4
-50
0 500 1000 1500 2000 2500 3000 3500
2
400
Voltage (V)

0
300 0 500
00 (b) 100
1000 1500 2000 2500 3000 3500
Time (sec)

200
Fig. 7. Injected temperature sensor fault and generated
0 500 1000 1500 2000 2500 3000 3500 temperature residual.
Temperature (oC)

40
-3
x 10

Resistance (Ohm)
30
3.5
o
Tamb= 20.0 ( C) normal

Change in
20 3 faulty
0 500 1000 1500 2000 2500 3000 3500
Time (sec) 2.5
Fig. 5. Time responses of the measured states under 2
normal conditions. 0 500
00 (a) 100
1000 1500 2000 2500 3000 3500
4

Residual (V)
Three different faults are considered as seen in Table 1. Fig.
6 illustrates a voltage sensor fault and the generated residual M< . Voltage 2
It is seen that once the voltage sensor fault occurs, it is
immediately detected and isolated via residual surpassing the 0
threshold < . 0 500
00 (b) 100
1000 1500 2000 2500 3000 3500
3
Residual (oC)
Temperature

10
2
Voltage Sensor
Fault (V)

1
5
0
0 500
00
(c) 10
1000 1500 2000 2500 3000 3500
0 Time (sec)
0 500
50 1000 1500 2000 2500 3000 3500
(a) Fig. 8. Resistance deviation and its diagnosis with
15 voltage and temperature residuals.
Residual (V)

10
Voltage

5. Conclusion
5
An observer-based fault diagnosis approach is designed for
0
0 500
50 1000 1500 2000 2500 3000 3500 state of health monitoring of batteries of HEVs. All possible
(b) Time (sec) sensor faults and unexpected resistance deviation are efficiently
Fig. 6. Injected voltage sensor fault and generated detected and isolated in the presence of parameter uncertainties.
voltage residual. Since it is not possible to observe what happens inside the
battery pack, detection of unexpected internal resistance
A temperature sensor fault and its detection via temperature variations that gives information inner dynamics like corrosion
residual are depicted in Fig. 7. Once the fault occurs, as seen in is especially an important result.
Fig.7b, the residual follows the sensor fault and exceeds the
threshold @ . 6. Acknowledgement
An unexpected deviation in the internal resistance as a I would like to thank Prof. Giorgio Rizzoni, Director of the
parameter fault is shown in Fig. 8. A small and drift-like Center for Automotive Research (CAR) at the Ohio State
deviation in the resistance results in alarms in voltage (Fig. 8b) University, for his valuable guidance and contributions on this
and temperature (Fig. 8c) residuals. Since internal resistance study.
deviation gives information about the inner dynamics of the
battery pack, such diagnosis of parameter fault is an important 7. References
aspect of the diagnosis method in terms of efficiency and
usability.
[1] M. Fetcenko, S. Ovshinsky, B. Reichman, K. Young, C.
Fierro, J. Koch, A. Zallen, W. Mays and T. Ouchi, "Recent
advances in NiMH battery technology," Journal of Power

241
 Sources, vol. 165, pp. 544-551, 2007.
[2] J. Zhang and J. Lee, "A review on prognostics and health
monitoring of Li-ion battery," Journal of Power Sources,
vol. 196, pp. 6007-6014, 2011.
[3] X. Tang, X. Zhang, B. Koch and D. Frisch, "Modeling and
Estimation of Nickel Metal Hydride Battery Hysteresis for
SOC Estimation," in Int. Conf. on Prognostics and Health
Management, Danver, CO, 2008.
[4] C. Suozzo, S. Onori and G. Rizzoni, "Model-based fault
diagnosis for NiMH," in 1st Annual Dynamic Systems and
Control Conference, Michigan, USA, 2009.
[5] B. Sun, L. Wang and C. Liao, "SOC estimation of NiMH
battery for HEV based on adaptive neuro-fuzzy inference
system," in IEEE Vehicle Power and Propulsion
Conference, Harbin, China, 2008.
[6] P. Frank and X. Ding, "Survey of robust residual generation
and evaluation methods in observer-based fault detection
systems," J. Proc. Cont., vol. 7, no. 6, pp. 403-424, 1997.
[7] R. Isermann, Fault-Diagnosis Systems, Springer-Verlag,
2006.
[8] C. Hajiyev and F. Caliskan, Fault Diagnosis and
Reconfiguration in Flight Control Systems, Kluwer
Academic, 2003.
[9] A. C. Margaret and M. Salameh, "Determination of lead-
acid battery via mathematical model techniques," IEEE
Trans. on Energy Conversion, vol. 7, no. 3, pp. 442-446,
1992.
[10] Y. Kim and H. Ha, "Design of interface circuits with
electrical battery models," IEEE Trans. on Industrial
Electronics, vol. 44, no. 1, pp. 81-86, 1997.
[11] B. Paxton and J. Newmann, "Modeling of nickel/metal
hydride," J. Electrochem. Soc., vol. 144, no. 11, pp. 3818-
3831, 1997.
[12] W. B. Gu and C. Y. Wang, "Thermal-Electrochemical
Modeling of Battery Systems," Journal of The
Electrochemical Society, vol. 147, no. 8, pp. 2910-2922,
2000.
[13] V. Srinivasan and C. Y. Wang, "Analysis of
Electrochemical and Thermal Behavior of Li-Ion Cells,"
Journal of The Electrochemical Society, vol. 150, no. 1, pp.
A98-A106, 2003.
[14] A. Emadi, M. Ehsani and J. M. Miller, Vehicular Electric
Power Systems, Marcel Dekker, 2004.
[15] S. Lee, J. Kim, J. Lee and B. H. Cho, "The state and
parameter estimation of an Li-ion battery using a new OCV-
SOC concept," in IEEE Power Electronics Specialists
Conference, Orlando, FL, 2007.

242