Received July 24, 2019, accepted August 8, 2019, date of publication August 19, 2019, date of current version

September 11, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2936213

A Health Monitoring Method Based on Multiple

Indicators to Eliminate Influences of Estimation
Dispersion for Lithium-Ion Batteries
JIA TANG, QISHENG LIU , SHIQI LIU, XIONG XIE, JICHENG ZHOU, AND ZHONGXIANG LI
School of Electrical Engineering and Automation, Wuhan University, Wuhan 430072, China
Corresponding author: Qisheng Liu (qishengliu@whu.edu.cn)
This work was supported by the National Key Research and Development Plan under Grant 2017YFB1201003.

ABSTRACT The state-of-health (SOH) is a critical parameter to determine the degradation degree of lithium-
ion batteries, thus plays an important part in energy management problems of electric vehicles (EVs). Due
to the estimation dispersion of health indicators, it is difficult to monitor conveniently and robustly the SOH
of on-board batteries. In this paper, a health monitoring method for the application of battery management
system (BMS) has been presented to solve this issue. First of all, the necessity of multiple indicators (MIs) is
demonstrated through the analysis of estimation dispersion of single indicator (SI), and three health indicators
are extracted from actual operating conditions of EVs. After that, the mixed membership function (MMF)
according to fuzzy logic theory is applied to integrate the estimation effects of different health indicators.
Then, the analytic hierarchy process (AHP) using dispersion information of health indicators is introduced
to calculate the weight coefficient of each indicator. Finally, the performance and robustness of the proposed
method have been verified by numerous experiments. The results indicate the average estimation errors of
five tested batteries are less than 3%, and the influences of estimation dispersion are largely eliminated when
compared with the SI-based and other MIs-based methods.

INDEX TERMS Li-ion batteries, capacity fading, health monitoring, estimation dispersion, multiple
indicators.

I. INTRODUCTION be a sophisticated system with a variety of specific inter-

Over the last decade, EVs have become the inevitable trend nal electrochemical reactions. The aging process mainly
of automobile industry because of the unique performance depends on the property changes of electrodes and elec-
advantage and environment friendly attribute [1]–[3]. Li-ion trolyte materials during operating conditions [5], [9]. Many
batteries are gradually considered to be the most promising experimental means such as electrochemical impedance spec-
candidate for power supply in EVs due to the superiority of troscopy (EIS), X-ray diffraction (XRD) and scanning elec-
high energy density, high output voltage, low self-discharge tron microscopy (SEM) have been introduced to find out the
rate and long cycling life [4], [5]. Health monitoring, as an degradation mechanisms of batteries essentially [10]–[12].
indispensable function of BMS, plays an important part in However, these electrochemical methods are difficult to meet
ensuring reliable operation and providing evidence for cas- the demand in practical on-board BMS application due to
cade utilization [6], [7]. However, the convenient and robust expensive analytical equipment, complex operating proce-
estimation of SOH is still a huge challenge. One of the main dure and difficult configuration of key parameters [13]. For
reasons lies in that the health indicators are difficult to obtain this reason, more attentions have been focused on exploring
under actual operating conditions of EVs and present notable the health indicators to reflect batteries degradation degree
estimation dispersion [8]. through measurable signals like voltage, current and so on.
SOH is influenced remarkably by the degradation degree Prior researches have shown batteries lose their abilities
of batteries. Nevertheless, Li-ion batteries are considered to to store energy and deliver/absorb power during the aging
process [14], [15]. But the Reference [16] points out power
The associate editor coordinating the review of this article and approving fading has no significant impact on driving performance of
it for publication was Bin Zhou. EVs, even when substantial levels of power drop to 30%

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
122302 VOLUME 7, 2019
J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

of initial value. Thus, capacity fading is seen as a more caused the estimation dispersion of SI. Hence, the existing
appropriate indicator to describe SOH of batteries. Unfortu- SOH estimators, merely established based on SI, not only
nately, the static capacity can only be measured through a full probably has limited reference value, but also has poor gen-
discharge of the fully charged battery. It seems unlikely to be eralization performance.
realized in the operation scenario of EVs. As a result, many In recent years, many researchers have realized the above
researchers have suggested to determine the capacity fading problems and shifted their focus to the health monitoring
by employing the correlation between the remaining capacity method with multiple indicators (MIs). The methods based
and other health indicators which are easy to obtain. on MIs, which are extracted from IC curves, are proposed
Establishing the equivalent circuit model (ECM) of bat- in [31]–[34], and the most commonly used indicators are the
teries and identifying the specific impedance parameters are peak position, peak intensity and area of curve. In [35], four
the most frequently used way for obtaining health indicators. specific health indicators are obtained from CC curves to
The increase in ohmic resistance of batteries has been widely reflect the phenomenon of capacity fading. In [36], the health
used to represent capacity fading [17], [18]. In addition, indicators are extracted from voltage response under pulse
the methods to monitor SOH based on diffusion resistance, conditions, and all the possible combination of these indi-
diffusion capacitor and energy storage capacitor are proposed cators are investigated to find out the most effective one to
in Ref. [19]–[21], respectively. In Reference [8], the authors estimate SOH. Furthermore, to integrate the effects of dif-
have found the solid electrolyte interphase (SEI) resistance ferent health indicators, the methods based on empirical fit-
seems to be a better choice when compared with ohmic ting, entropy weight coefficient, grey relational analysis and
resistance, because it shows a better monotonic correlation multidimensional look-up table are introduced in [37]–[40],
with the remaining capacity. respectively. These methods can enhance the estimation accu-
Another workable method for SOH estimation of batteries racy and robustness, but pay attention insufficiently to the dis-
is the analysis of terminal voltage response under certain persion information of health indicators. In addition, the most
excitation. This approach is actually quite similar to the iden- existing methods use merely the indicators with high corre-
tification of the battery impedance parameters [2]. Previous lation, which are extracted from the same measurement pro-
studies have introduced the health monitoring methods based file and may be subject to the same interference. Therefore,
on terminal voltage response after impulse load [22], [23]. applicability problem is still exist with these MIs-based SOH
In Reference [24], a parameter extracted from open-circuit estimators.
voltage (OCV) curve is correlated to batteries aging process, Aware of the issues discussed above, this paper devotes
thus it can be viewed as the health indicator. Moreover, attention to seek several health indicators, and each of them
the concept of entropy in the field of information theory has can easy to obtain from the CC charging curve or load profile.
been introduced for extracting statistical information from A health monitoring method based on MIs is proposed to
batteries charge/discharge test data. The methods based upon eliminate estimation dispersion of SI. This method adopts
sample entropy and multiscale entropy are proposed to deter- the mixed membership function (MMI) to integrate differ-
mine the SOH and predict the remaining useful life (RUL) ent indicators and AHP to calculate the weight coefficients.
in [25], [26], respectively. Through verification of experimental results and comparison
The relatively complete constant-current (CC) charging of other methods, the effectiveness of the proposed method
curve can be measured in actual operating conditions of to eliminate the influences of estimation dispersion has been
EVs. Therefore, some researchers pay attention to searching proved.
for characteristic information to represent SOH form these The remaining part of this paper is arranged as follows:
curves. The incremental capacity (IC) and differential volt- Section 2 analyzes the estimation dispersion of SI and intro-
age (DV) analysis can act as this role and have been exten- duces three potential health indicators. In Section 3, a health
sively used as a laboratory-based approach to investigate the monitoring method based on MIs is proposed. The proposed
aging mechanisms of batteries [27]–[29]. Take the IC curve method is verified in Section 4, which is followed by the
for instance, the region of each peak indicates the capacity limitations and conclusion in Section 5 and 6, respectively.
participating in the relevant process of phase transforma-
tion. Consequently, health monitoring can be implemented
through observing the evolution of IC curve over the whole II. HEALTH INDICATORS
life of batteries. Beyond that, a novel SOH estimation method A. ESTIMATION DISPERSION OF SI
based on group method of data handling is proposed in Ref- The relationship of five tested batteries between the ohmic
erence [30], where the voltage of CC charging sub-process is resistance and remaining capacity is drawn in Fig. 1. The
applied as the input of estimator. black solid line represents the fitting curve with minimum
The methods mentioned above have showed that there are deviation and red dashed lines represent the fitting curves
certainly some correlations between the health indicators and with specified confidence interval. It is apparent from the
remaining capacity, but the capacity fading cannot be deter- figure that a specific value of remaining capacity corre-
mined exactly by considering only a single indicator (SI). sponds actually to a wide range of ohmic resistance and
This result from the fact that battery inconsistencies have vice versa. Therefore, it is impossible to represent the

VOLUME 7, 2019 122303

 J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

FIGURE 1. The relationship between the ohmic resistance and remaining

capacity.

relationship between the two completely through a certain

fitting curve. Such consideration leads to the conclusion
that there exists obvious estimation dispersion when ohmic
resistance is selected as a single health indicator. In addi-
tion, the Fig. 1 reflects the dispersion information is con-
tained in the whole area surrounded by two fitting curves
with specified confidence interval, which is of significance
to determine the degree of importance of different health
indicators. Although the above analyses are limited to ohmic
resistance, the phenomenon of estimation dispersion has also
been found for other health indicators, which can be observed
from three selected indicators in following section 2.2. Thus,
it is essential to select several health indicators to eliminate
the negative influences of estimation dispersion.

B. SELECTION PRINCIPLES
In this paper, there are three key principles to consider before
the selection of health indicators. Firstly, each indicator can
be obtained conveniently in actual operating conditions of FIGURE 2. The relationship between selected indicators and remaining
capacity.
EVs. Secondly, each indicator should exhibit the monotonic
correlation with remaining capacity as a whole. Thirdly,
indicators have demonstrated the agreeable practicality for
the relevance between different indicators is less signifi-
the application scenario of EVs, which will be introduced in
cant. According to these principles, average transient resis-
detail after the each calculation process of them.
tance (ATR), corrected standard deviation (CSD) and partial
peak area (PPA) have been selected as health indicators. All
C. CALCULATION OF ATR
three health indicators can be obtained from the CC charging
curve or load profile. Fig. 2a∼c present the relationship Li-ion batteries can be described by the Thevenin ECM as
between three selected indicators and remaining capacity, shown in Fig. 3, which includes three parts: the voltage source
respectively. The monotonic correlation between each indi- E, the ohmic resistance R0 and the parallel RC network [41],
cator and remaining capacity can be observed clearly from [42]. According to the circuit theory, the terminal voltage VT
each subgraph. Furthermore, the fitting curves and equations could be expressed as follows:
indicate that there are significantly different change rules for VT = E − R0 IT − VP (1)
different health indicators during the period of capacity fad-
ing. The reason behind this is that ATR, CSD and PPA reflect where IT is terminal current and VP is polarization voltage
respectively the changes of ohmic resistance, fluctuation across parallel RC network.
degree of voltage response and evolution trend of IC curve, When shifts in IT occur abruptly, E and Vp remain con-
which suggests there is less relevance between three selected stant, thus the changes of VT depend completely on voltage
indicators. This conclusion can also be proved by comparing variation across R0 . Considering the period of additional
the distribution of scatter points of the same battery in the voltage change, mainly as a result of Rp and Cp , is far longer
different subgraph. To sum up, all three principles mentioned than the 10 Hz sampling interval, the sampling value can
above have been satisfied. In addition, the three selected be used instead of the actual value of instantaneous voltage

122304 VOLUME 7, 2019

J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

FIGURE 3. Thevenin ECM of batteries.

and current following a current pulse. Therefore, R0 can FIGURE 4. The first 120s load curves with different life cycles.
be represented by the transient resistance RT , which can be
calculated as follows: differences between the upper-limit (dashed line) and lower-
VT,k − VT,k−1 VT+ − VT− limit (dot dash line) voltage for each curve are different. The
RT = ≈ R0 = (2) curve corresponding to a higher aging degree has a greater
IT,k − IT,k−1 IT+ − IT−
difference, which illustrates there is a more violent fluctua-
where k and k-1 represent current and previous sample time, tion. Therefore, the quantification of fluctuation degree can
respectively. The positive and negative sign represent after be used to determine the capacity fading of batteries to some
and before the change instant of current, respectively. extent.
Actually, the modelling uncertainties and measurement In this paper, the concept of standard deviation is intro-
errors are inevitably introduced to this open-loop calculation. duced to describe the fluctuation degree of voltage response
Therefore, this paper adopts the method that repeats the calcu- sequences. The calculation formula of standard deviation σ
lation process for the first 120s load profile, removes the gross is regulated as follows:
errors and averages the results to eliminate measurement v
u N
error. The formula of ATR is shown as the following: u1 X
σ =t (ui − u)2 (4)
M
1 X N
i=1
ATR = RT,i (3)
M where ui are the voltage response sequences, N is the length
i=1
of sequences, and ū is the average value of ui .
where M is the total number of effective times when
However, standard deviations of voltage response
change of IT exceeds the setting threshold in the first 120s
sequences are not exactly comparable when different exci-
load profile.
tations are applied to batteries. The CSD is proposed to solve
Compared with the identification algorithm of specific
this problem, which introduces simultaneously fluctuation
impedance, the use of ATR can avoid the shortcoming of
information of current sequences. CSD can be obtained as
computing expensive due to its calculation process only
follows:
includes a small number of addition and multiplication oper-
ation. Furthermore, the convergence and stability of ATR are σV
CSD = (5)
independent of some specific adjustable parameters. There- σI
fore, for the application scenario of EVs which has a stronger where the σV and σI represent the standard deviations of
emphasis on the estimation stability, ATR seems a much voltage and current sequences in the same first 120s load
better option. profile, respectively.
Compared with other health indicators based on voltage
D. CALCULATION OF CSD response, which mostly require the specific excitation at
Under the same excitation, the first 120s load curves with certain state-of charge (SOC) point, the CSD can be calcu-
different life cycles are plotted in Fig. 4. The load curves lated directly through the actual load profile. This is a major
move to low voltage direction with the increase of life cycles. advantage for the application scenario of EVs because of
Note that the starting voltages corresponding to different life avoiding the standardized procedures of pulse test. More-
cycles are different, which is mainly caused by the slight over, the CSD differs from the previous methods based on
voltage drop after the constant voltage (CV) step in charging information entropy, requires no matrix inversion and is more
procedure of batteries. This kind of differences are signif- applicable to the embedded system.
icantly affected by the ohmic resistance that is gradually In the calculation process of ATR and CSD, the criteria
increasing as the aging of battery, and even have been fur- of first 120s load profile need to be further explained. The
ther enhanced in voltage relaxation process. In addition, the two indicators are significantly influenced by the impedance

VOLUME 7, 2019 122305

 J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

FIGURE 6. IC curves with different life cycles.

FIGURE 5. The terminal voltage curves of CC charging process with
different life cycles.

collection interval. Actually, the performance of PPA to esti-

characteristics of batteries, and thus are affected by the range mate the remaining capacity increase with a larger range
of SOC in the same way. The influences of SOC should be of data collection interval. Therefore, there is a trade-off
eliminated as far as possible when the ATR and CSD are used between the performance of PPA and range of data collection
as health indicator to describe the capacity fading. Therefore, voltage. In light of this, the range of data collection interval is
the first 120s load profile is selected as data acquisition inter- not mandatory, and can instead be selected according to the
val to ensure the SOC remain approximately constant and actual situation. Note that the selected interval corresponds
every calculation process are completed at the specific value to the voltage platform that is the region with the most
of 100% SOC after charging process. In addition, the criteria significant changes in charging curves, as shown in Fig. 5.
of first 120s load profile can effectively ensure the enough Therefore, the PPA can be calculated as follows:
number of load variation to obtain the changes of voltage
response. Actually, the criteria of first 120s load profile can Z 3.8 m
X
also be adjusted according to the actual application scenarios. PPA = wdV = wi (7)
3.65 i=1
E. CALCULATION OF PPA
where w is the value of dQ/dV , wi is the discrete values of
Fig. 5 shows the terminal voltage curves of CC charging w corresponding to different voltage interval, and m is the
process with different life cycles. The line with higher life number of voltage intervals in the peak region of 3.65∼3.8V.
cycles always shows a larger slope, which means the time for Compared with other indicators based on IC curves, such
the same battery charged from the same voltage to the cut- as peak intensity and peak position, the PPA seems a better
off voltage (4.2V) gradually diminishes while battery health choice to ensure the monotonicity with remaining capacity.
deteriorates. Thus, capacity fading can also be described The reason behind this is that there is difficulty in accurately
by extracting effective information from these charging identifying the peak position and peak intensity, which can
curves. be possibly submerged by measurement noise. The PPA is
The IC analysis can be used to extract the relevant informa- calculated by integrating the IC curves at certain data collec-
tion of capacity fading. In this study, the IC curves are derived tion interval, and thus can effectively avoid the problem of
by means of 1/2 C charging curves by the point-counting the single-valued measurement.
method [13]. The calculation formula of dQ/dV can be
written as:
III. HEALTH MONITORING METHOD
dQ nI A. ESTABLISHMENT OF THE ESTIMATION SYSTEM
= (6)
dV 3600 f 1V Establishment of the estimation system to integrate the above
where I is the charging current, f is the sampling frequency, three indicators is necessary because of the effects of each
1V is the width of voltage interval, and n is the number of indicator to estimate SOH are different. For this purpose,
voltage sampling points in the corresponding interval. a health monitoring method based on MIs is proposed to
The evolution of IC curves with different life cycles are establish a comprehensive SOH estimator. The flowchart of
illustrated in Fig. 6. There is significant drop in the peak the proposed method is plotted in Fig. 7.
region of 3.5∼3.8V, which can be used to indicate the First, evaluation sets are defined as S = {V1 , V2 , V3 },
capacity fading of batteries. Considering that it is rarely which represent different degrees of capacity fading.
to fully discharge the batteries in actual application and Then, three selected health indicators can be calculated
start charging from such low voltage like 3.5V, the peak and normalized after every electrical measurement of CC
region between 3.65V to 3.8V has been selected as the data charging and first 120s dynamic load. The purpose of

122306 VOLUME 7, 2019

J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

FIGURE 7. The flowchart of the proposed health monitoring method.

FIGURE 8. The shapes of three typical membership fuctions.

normalization is to eliminate the dimensional differences of results [43]. Because the MF can be considered to be a
each indicator. The formulas of normalization are as follow: generalized fitting relation between remaining capacity and
the health indicators. There are three typical shapes of the
xi − xi,min MF: linear type, conical type, bell-shaped type, which are
xi0 = (i = 1, 2) (8) displayed in Fig. 8. It is obvious that the slope of the rising
xi,max − xi,min
xi − xi,max edge of membership function gets larger, the membership
0
xi = (i = 3) (9) increases faster. Similarly, the slope of the falling edge gets
xi,min − xi,max
smaller, the membership decreases faster. Therefore, the three
where xi is the value of the ith indicator, xi,max and xi,min MF are applicable to different situations. The linear type
are respectively the maximum and minimum value of the ith MF performs better in the situation that capacity fading has
indicator, and x 0 i is the normalization value of xi . Note that a linear relation with the indicator. The conical type and
the formula of PPA is different from the other two indicators bell-shaped MF adapted to the situation that capacity fading
because of its decreasing trend with the capacity fading, rate gradually increases/decreases with indicator increases,
as shown in Fig. 2c. respectively.
x 0 i (i = 1, 2, 3) are used as input for respective membership It is manifested from the Fig. 2 that the fitting relation
function to generate fuzzy set Z3×3 that can be considered to between different health indicators and remaining capacity
be an overall estimation of battery SOH. are not exactly the same. Thus, the mathematical forms of
Furthermore, the quantitative values αi (i = 1, 2, 3) of the MF should be determined according to the shapes of the
dispersion of three indicators are used to calculate the weight fitting curves. In this paper, the concept of mixed member-
coefficients βi (i = 1, 2, 3) of each indicator by AHP, and ship function (MMF) is proposed to adapt to different health
details of calculation procedure are introduced in Section 3.3. indicators. The conical type MF is applied for indicator ATR,
Finally, the comprehensive SOH can be outputted by the the linear type for CSD, and the bell-shaped type for PPA.
weighted average defuzzification, which contains the capac-
ity fading information of three health indicators to eliminate C. CALCULATION OF INDICATOR WEIGHTS
the estimation dispersion of SI. The estimation effects of each health indicator are different,
and even have time-varying characteristics. Therefore, apply-
B. SELECTION OF MEMBERSHIP FUNCTION ing different weight coefficients to different indicators and
Actually, the mathematical forms of the membership func- adjusting weight coefficients according to actual estimation
tions (MF) have a great impact on the accuracy of estimation dispersion of each indicator is necessary [44].

VOLUME 7, 2019 122307

 J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

with larger αi has an inevitably smaller important degree in

the discriminant matrix, which can be observed from Eq. (11).

D. COMPREHENSIVE SOH
The process of defuzzification is implemented by weighted
average algorithm for the above-mentioned fuzzy set. The
calculation formula is as follow.
3
βi zij
P
3
i=1
X
yj = = βi zij (j = 1, 2, 3) (12)
3
βi
P i=1
FIGURE 9. The relationship between remaining capacity and the
i=1
normalization value of ATR.
where yj is the weighting estimation value of the jth column
In this paper, AHP is introduced to calculate the weight elements of fuzzy set, which corresponds to evaluation set
coefficients, and the weight coefficients are recalculated after Vj , zij is the element of the ith row and the jth column of
each electrical measurement of CC charging and first 120s fuzzy set.
dynamic load. Following are details about the calculation The comprehensive SOH value can be calculated by
procedure. Eq. (13), which synthesizes the capacity fading information
Firstly, quantification of estimation dispersion. Take indi- of all three evaluation set.
cator ATR for instance, the relationship between remain- 3
X
ing capacity and the normalization value of ATR is plotted SOH = yj Vj (13)
in Fig. 9. More exactly, assuming the normalization value of j=1
ATR is taken as x 0 , then the value of remaining capacity in the
fitting curves with minimum deviation and with confidence E. GENERALIZATION OF PROPOSED METHOD
interval can be defined as Cfit , Cmax and Cmin . The difference Although the three selected indicators mainly aim at batteries
value between Cmax and Cmin is obviously affected by the applied in EVs, there is no such restriction on the proposed
estimation dispersion. Therefore, the quantization value αi health monitoring method. The reason behind this is that the
of estimation dispersion of the ith health indicator can be proposed method focuses primarily on using the informa-
defined by Eq. (10). A higher value of αi indicates a higher tion of dispersion to assign weights level to different health
estimation dispersion. indicators, and reduce the influence of such indicators with
high dispersion at the current estimation moment. Therefore,
Ci,max − Ci,min
αi = (10) the key issue to apply the proposed method to other batteries
Ci,fit operation conditions lies in that the selected indicators have
Then, establishment of discriminant matrix D3×3 . The ele- certain applicability for corresponding application scenarios.
ment of the ith row and the jth column of D3×3 is expressed In addition, the criteria mentioned in calculation process of
as dij , which represents the important degree of the ith health indicators, such as the criteria of the first 120s load profile and
indicator compared with the jth indicator. According to the the data collection interval of 3.65∼3.8V, is not mandatory,
theory of AHP, the higher important degree of the indicator and should instead be developed according to the actual
corresponds to the higher weight coefficient. In addition, demand of EVs. Moreover, the number of health indicators
the higher estimation dispersion of the indicator indicates the can be selected freely according to the specific application
lower accuracy of SOH estimation. Therefore, αi can be acted condition. Generally, the greater the number of indepen-
as evidence of important scale of D3×3 . The formula of dij is: dent health indicators and the better its monotonicity with
remaining capacity, the higher the estimation performance,
αj
dij = (11) especially the robustness. But the problems of computational
αi complexity and storage resource will become more serious
where αi and αj are the quantization value of estimation as the number of indicators increase. Thus the number of
dispersion of the ith and jth health indicator, respectively. selected indicators should ultimately be determined by the
Finally, calculation of eigenvector. According to AHP, the balance between estimation performance and database size.
eigenvector corresponding to maximum eigenvalue of the
discriminant matrix can reveal the importance order of each IV. EXPERIMENTS AND VERIFICATION
indicator. The indicator weights βi can be obtained by nor- A. EXPERIMENTAL PLATFORM AND PROCEDURES
malization of eigenvector. Actually, it is necessary to carry out The battery test platform is built to complete the related
the consistency check for the discriminant matrix. However, experiments as shown in Fig. 10. The tested batteries are fixed
the important scale is based on αi , which means the consis- on the clamping apparatus and placed together in the thermal
tency check is satisfied automatically. Because the indicator chamber. A Neware BTS-4000 tester realizes the functions

122308 VOLUME 7, 2019

J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

FIGURE 10. The battery test platform.

FIGURE 11. The procedures of life cycle test.

FIGURE 13. The voltage and current waveforms of three SDS in first 120s.

Fig. 11 reveals the procedures of life cycle test. At the

beginning of life (BOL), static capacity test (SCT) and ref-
erence performance test (RPT) are carried out to investigate
the initial performance of batteries. The RPT consists of a
resistance test, a hybrid pulse test and three representative
simulated driving schedule (SDS) tests. The three SDS con-
FIGURE 12. The current waveform of discharge steps.
sists of federal urban driving schedule (FUDS), inspection
& maintenance driving schedule (IMDS) and dynamic stress
of charging/discharging, protection and signal acquisition. test (DST), which refer to PNGV battery test manual and
A computer is connected with the tester through data cable. EPA dynamometer drive schedules [45], [46]. Then it begins
The operation of data transfer, load mode editing and graphics to enter the aging cycles. Single life cycle consists of two
output can be completed by the human-computer interface. patterns: the main charging and the main discharging pat-
The sampling frequency of the tester can be set according to tern. The charging pattern adopts a simple CC mode, and
actual demand, up to 10 Hz. All experiments are conducted the discharging contains several discharge steps with the
at 25 ◦ C. The commercial cylindrical batteries with layered same current excitation, as shown in Fig. 12. The design
LiNix Coy Mn1-x-y O2 (NCM) materials are selected to test due of discharge step refers to the life cycle test method of EV
to relatively balanced performance between energy density power batteries in [47]. Each step releases 2% of the nominal
and safety. The detailed specifications are listed in Table 1. capacity of battery. After several cycles, SCT is executed to

VOLUME 7, 2019 122309

 J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

TABLE 1. The specification of tested batteries.

FIGURE 14. Estimation results of all tested batteries.

record the capacity attenuation. Once the capacity attenuation of EVs. Three representative SDS are implemented to verify
accumulates more than 3% when compared with the last RPT, the consistency of ATR and CSD under the different load
another RPT will be completed to explore the changes of profile. The voltage and current waveforms of three SDS
battery performance under the current aging degree. In this in first 120s are plotted in Fig. 13. It is apparent that the
paper, total capacity attenuation by more than 25% of initial waveform changes are different from one another.
capacity is defined as the end of life (EOL). Table 2 and 3 give the verification results of ATR and CSD
of battery # 3, respectively. The maximum deviation (MD) is
defined in Eq. (14), and the smaller the MD value, the better
B. VERIFICATION OF HEALTH INDICATORS the consistency of the health indicator.
For three selected health indicators mentioned above, ATR
xi − xj
and CSD are calculated through the first 120s load profile. MD = max × 100% (i 6= j) (14)
3
However, the load profile is commonly fraught with uncer- 1 P
n xn
tainty, which mainly depends on actual operating conditions n=1

122310 VOLUME 7, 2019

J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

TABLE 2. ATR verification results of battery # 3 under three SDS.

TABLE 3. CSD verification results of battery # 3 under three SDS.

FIGURE 15. Estimation error results compared with SI-based methods.

TABLE 4. Estimation error results compared with SI-based methods.

where x1 , x2 and x3 are the same indicator value calculated

under different SDS, respectively.
The results indicate the both selected indicators are basi-
cally at the same level under different schedules when the
life cycles are the same. The MD values of ATR are less than
5%, and CSD less than 3%. It suggests that both indicators
have sufficient reliability to represent attenuation degree of
batteries even under different load profile.

C. VERIFICATION OF ESTIMATION RESULTS

By using the relationship between the remaining capacity
and three selected indicators, the proposed health monitoring over the whole predefined life-span. In addition, it is apparent
method can estimate SOH of batteries after every CC charg- from the numeric results in the figures that the estimation
ing and first 120s load profile process. It is quite important errors have an obvious characteristic of randomness. There-
to investigate the validity and robustness of the proposed fore, all these analyses clearly prove the fact that the proposed
method against all tested batteries. The estimation results for method is effective to characterize the capacity fading and
five tested batteries are shown in Fig. 14. The actual value exhibits good robustness and generalization performance.
of remaining capacity has been calculated by Ah counting To more clearly illustrate the performance of the proposed
method. The first point to note is the same changing trend health monitoring method to eliminate the estimation dis-
between the estimation value and actual value. Although the persion of SI, the comparison results of estimation error
estimation value appears to increase instead of decrease in a between the proposed method and three SI-based methods
very few cases, the estimation results still sustain a high accu- are plotted in Fig. 15, and the detailed numerical results are
racy and the relative error for all tested batteries are within 5% listed in Table 4. The estimation values of SI-based methods

VOLUME 7, 2019 122311

 J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

FIGURE 16. Estimation error results compared with other MIs-based methods.

have been obtained by the simple fitting relationship between while the peak to peak error has declined more than 3.0%.
remaining capacity and three selected health indicators. Furthermore, the distribution of error boxes indicates the pro-
It can be seen from the results of both average and max- posed method displays better randomness of estimation error,
imum estimation error that the health monitoring method and all five boxes which correspond specific battery are not
based on MIs exhibits better performance compared with the located in the same side of zero demarcation line (red dashed
three SI-based methods. The average and maximum estima- lines). This result also reveals that the proposed method has
tion errors of proposed method are respectively within 3% better performance advantage for eliminating the estimation
and 5% for all five tested batteries, and both of them reflect dispersion. The proposed method gives an effective solution
a noticeably stable trend, which is the goal wanted to achieve of the problem because MIs and dispersion information of
from a state estimation point of view. However, there exist different health indicators are used in the estimation process,
the great differences of estimation error between the differ- thus has a better reliability and stability.
ent tested batteries when the SI-based methods are applied.
Some estimation results present intolerably serious error, for V. LIMITATIONS
example, the maximum estimation error of battery # 2 even Despite having obvious advantages in terms of good estima-
exceeds 10% when the ATR-based method is applied. But tion accuracy and robustness, the proposed method still has
some estimation results are about the same as the MIs-based several limitations. First, prior information between remain-
method, such as both average and maximum error of battery ing capacity and health indicators is necessary to make sure
# 4. Actually this is due to the estimation dispersion of SI. fuzzy procedure and weights calculation match the charac-
When coupled with the information from both Fig. 2 and teristics of batteries. This requires plenty of data accumu-
Fig. 14, leads to the conclusion that the closer the fitting curve lated from batteries attenuation experiments. If the proposed
and the distribution of scatter points of health indicator are, method is to be applied in BMS, it would be reasonable for
the higher accuracy of the SI-based method can be acquired. the battery manufactures to establish the associated database
Therefore, the SI-based methods are not always effective due and provide the statistical information. The second possible
to the inconsistency of batteries. limitation is the demand for specific data collection inter-
In addition, Fig. 16 describes the estimation error results val to calculate the selected health indicators. The complete
of the proposed and other two commonly used MIs-based acquisition of selected indicators is the prerequisite for the
methods in the form of boxplot. The red lines inside the boxes effective implementation of the proposed health monitoring
represent the average errors of the certain battery, and two of method. The batteries applied in EVs usually have a rel-
dark red dashed lines are used to indicate the peak to peak atively dynamic load profile and wide range of depth of
errors based on different methods. It is apparent that the pro- discharge (DOD), which makes the criteria of indicators cal-
posed method has smaller estimation error than others when culation satisfy more easily. However, there still exists failure
the same health indicators are selected to estimate the remain- risk in SOH estimation process when batteries are not cycled
ing capacity, especially the peak to peak error. Compared with in designated data collection interval or the data sampling
the empirical fitting and multidimensional look-up table, the frequency of the BMS can not meet the condition of 10 Hz.
total error of the proposed method maintains under 2.5%, Therefore, the other health indicators may need to be taken

122312 VOLUME 7, 2019

J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

as a temporarily alternative choice for higher reliability when [11] H. J. Chang, N. M. Trease, A. J. Ilott, D. Zeng, L.-S. Du, A. Jerschow,
the proposed method is adopted in BMS. Additionally, some and C. P. Grey, ‘‘Investigating li microstructure formation on li anodes for
lithium batteries by in Situ 6 Li/7 Li NMR and SEM,’’ J. Phys. Chem. C,
environmental factors and load conditions have significant vol. 119, no. 29, pp. 16443–16451, 2015.
impacts on the battery characteristics, which might change [12] P. K. A. Babu, A. S. Waghmare, S. M. Mulla, U. S. Karle, and M. R. Saraf,
applicability of the selected health indicators. This also will ‘‘Material characterization of Lithium-ion battery cells by scanning elec-
tron microscopy & X-ray diffraction techniques,’’ in Proc. IEEE Transp.
be the main content of the future research. Electrific. Conf. (ITEC-India), Dec. 2017, pp. 1–5.
[13] X. Han, M. Ouyang, L. Lu, J. Li, Y. Zheng, and Z. Li, ‘‘A comparative
study of commercial lithium ion battery cycle life in electrical vehicle:
VI. CONCLUSION
Aging mechanism identification,’’ J. Power Sources, vol. 251, pp. 38–54,
The estimation dispersion of health indicators always makes Apr. 2014.
it difficult for convenient and robust estimation of SOH. [14] M. Dubarry, V. Svoboda, R. Hwu, and B. Y. Liaw, ‘‘Capacity and power
fading mechanism identification from a commercial cell evaluation,’’
In this paper, a health monitoring method based on MIs J. Power Sources, vol. 165, no. 2, pp. 566–572, 2007.
is proposed to address this problem. Three health indica- [15] X. Li, J. Kang, Y. Yang, F. Yan, C. Du, and M. Luo, ‘‘A study on
tors which are extracted from the CC charging curve or capacity and power fading characteristics of Li(Ni1/3 Co1/3 Mn1/3 )O2 -
based lithium-ion batteries,’’ Ionics, vol. 22, no. 11, pp. 2027–2036, 2016.
the first 120s load profile are used to reduce the estimation
[16] S. Saxena, C. Le Floch, J. MacDonald, and S. Moura, ‘‘Quantifying EV
dispersion of SI. Then, the concepts of MMF and AHP battery end-of-life through analysis of travel needs with vehicle powertrain
have been adopted to integrate the estimation effects and to models,’’ J. Power Sources, vol. 282, pp. 265–276, May 2015.
[17] J. Remmlinger, M. Buchholz, M. Meiler, P. Bernreuter, and K. Dietmayer,
determine the weight coefficients of different health indica-
‘‘State-of-health monitoring of lithium-ion batteries in electric vehicles
tors, respectively. After verification, the proposed method is by on-board internal resistance estimation,’’ J. Power Sources, vol. 196,
able to predict the capacity fading accurately, and average no. 12, pp. 5357–5363, 2011.
estimation errors are within 3% over the whole life-span [18] H. Dai, Z. Sun, and X. Wei, ‘‘Estimation of internal states of power lithium-
ion batteries used on electric vehicles by dual extended Kalman filter,’’
for five tested batteries. Moreover, error results of the J. Mech. Eng., vol. 45, no. 6, pp. 95–101, 2009.
proposed method demonstrate the effectiveness to reduce [19] J. Kim and B.-H. Cho, ‘‘State-of-charge estimation and state-of-health
the influences of estimation dispersion through compared prediction of a li-ion degraded battery based on an EKF combined with a
per-unit system,’’ IEEE Trans. Veh. Technol., vol. 60, no. 9, pp. 4249–4260,
with SI-based and other MIs-based methods. In the future, Nov. 2011.
the work will be focused on investigating the sensitivity of [20] Z. Chen, C. C. Mi, Y. Fu, J. Xu, and X. Gong, ‘‘Online battery state of health
above-mentioned health indicators to the load conditions and estimation based on Genetic Algorithm for electric and hybrid vehicle
applications,’’ J. Power Sources, vol. 240, pp. 184–192, Oct. 2013.
environmental factors such as temperature and mechanical [21] C. Zhang, Y. Zhang, and Y. Li, ‘‘A novel battery state-of-health estimation
vibration. method for hybrid electric vehicles,’’ IEEE/ASME Trans. Mechatronics,
vol. 20, no. 5, pp. 2604–2612, Oct. 2015.
[22] M. Coleman, W. G. Hurley, and C. K. Lee, ‘‘An improved battery character-
REFERENCES ization method using a two-pulse load test,’’ IEEE Trans. Energy Convers.,
[1] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, ‘‘A review of vol. 23, no. 2, pp. 708–713, Jun. 2008.
lithium-ion battery state of charge estimation and management system in [23] A. Banaei and B. Fahimi, ‘‘Real time condition monitoring in Li-Ion
electric vehicle applications: Challenges and recommendations,’’ Renew. batteries via battery impulse response,’’ in Proc. IEEE Vehicle Power
Sustain. Energy Rev., vol. 78, pp. 834–854, Oct. 2017. Propuls. Conf., Sep. 2010, pp. 1–6.
[2] W. Waag, C. Fleischer, and D. U. Sauer, ‘‘Critical review of the methods [24] L. Lavigne, J. Sabatier, J. M. Francisco, F. Guillemard, and A. Noury,
for monitoring of lithium-ion batteries in electric and hybrid vehicles,’’ ‘‘Lithium-ion open circuit voltage (OCV) curve modelling and its ageing
J. Power Sour., vol. 258, no. 14, pp. 321–339, Jul. 2014. adjustment,’’ J. Power Sources, vol. 324, pp. 694–703, Aug. 2016.
[25] A. Widodo, M.-C. Shim, W. Caesarendra, and B.-S. Yang, ‘‘Intelligent
[3] E. Samadani, M. Mastali, S. Farhad, R. A. Fraser, and M. Fowler, ‘‘Li-ion
prognostics for battery health monitoring based on sample entropy,’’ Expert
battery performance and degradation in electric vehicles under different
Syst. Appl., vol. 38, no. 9, pp. 11763–11769, Sep. 2011.
usage scenarios,’’ Int. J. Energy Res., vol. 40, no. 3, pp. 379–392, 2016.
[26] Y. Chen, J. Bao, Z. Xiang, and J. Wei, ‘‘Predictability analysis of lithium-
[4] L. Datong, Z. Jianbao, G. Limeng, and P. Yu, ‘‘Survey on lithium-ion
ion battery remaining useful life with multiscale entropy,’’ in Proc. Int.
battery health assessment and cycle life estimation,’’ Chin. J. Sci. Instrum.,
Conf. Comput. Inf. Sci., Jun. 2013, pp. 1052–1055.
vol. 36, no. 1, pp. 1–16, Jan. 2015.
[27] C. Weng, Y. Cui, J. Sun, and H. Peng, ‘‘On-board state of health monitoring
[5] A. Barré, B. Deguilhem, S. Grolleau, M. Gérard, F. Suard, and D. Riu, of lithium-ion batteries using incremental capacity analysis with support
‘‘A review on lithium-ion battery ageing mechanisms and estimations for vector regression,’’ J. Power Sources, vol. 235, no. 4, pp. 36–44, Aug. 2013.
automotive applications,’’ J. Power Sour., vol. 241, no. 11, pp. 680–689, [28] M. Kassem, J. Bernard, R. Revel, S. Pélissier, F. Duclaud, and C. Delacourt,
Nov. 2013. ‘‘Calendar aging of a graphite/LiFePO4 cell,’’ J. Power Sources, vol. 208,
[6] J. Shen, Y. He, and Z. Ma, ‘‘Progress of model based SOC and SOH pp. 296–305, Jun. 2012.
estimation methods for lithium-ion battery,’’ J. Chem. Ind. Eng., vol. 69, [29] L. Wang, C. Pan, L. Liu, Y. Cheng, and X. Zhao, ‘‘On-board state of health
no. 1, pp. 309–316, Jan. 2018. estimation of LiFePO4 battery pack through differential voltage analysis,’’
[7] L. Lu, X. Han, J. Li, J. Hua, and M. Ouyang, ‘‘A review on the key issues Appl. Energy, vol. 168, pp. 465–472, Apr. 2016.
for lithium-ion battery management in electric vehicles,’’ J. Power Sour., [30] J. Wu, Y. J. Wang, X. Zhang, and Z. H. Chen, ‘‘A novel state of health
vol. 226, no. 3, pp. 272–288, Mar. 2013. estimation method of Li-ion battery using group method of data handling,’’
[8] R. Xiong, J. Tian, H. Mu, and C. Wang, ‘‘A systematic model-based degra- J. Power Sources, vol. 327, pp. 457–464, Sep. 2016.
dation behavior recognition and health monitoring method for lithium-ion [31] Y. C. Zhang, O. Briat, J.-Y. Delétage, C. Martin, and N. Chadourne,
batteries,’’ Appl. Energy, vol. 207, pp. 372–383, Dec. 2017. ‘‘Efficient state of health estimation of Li-ion battery under several age-
[9] M. M. Kabir and D. E. Demirocak, ‘‘Degradation mechanisms in Li-ion ing types for aeronautic applications,’’ Microelectron. Rel., vols. 88–90,
batteries: A state-of-the-art review,’’ Int. J. Energy Res., vol. 41, no. 14, pp. 1231–1235, Sep. 2018.
pp. 1963–1986, 2017. [32] Y. Li, M. A. Monem, R. Gopalakrishnan, M. Berecibar, E. N. Maury,
[10] W. Waag, S. Käbitz, and D. U. Sauer, ‘‘Experimental investigation of N. Omar, P. Bossche, and J. V. Mierlo, ‘‘A quick on-line state of health
the lithium-ion battery impedance characteristic at various conditions and estimation method for Li-ion battery with incremental capacity curves
aging states and its influence on the application,’’ Appl. Energy, vol. 102, processed by Gaussian filter,’’ J. Power Sources, vol. 373, pp. 40–53,
no. 2, pp. 885–897, 2013. Jan. 2018.

VOLUME 7, 2019 122313

 J. Tang et al.: Health Monitoring Method Based on Multiple Indicators to Eliminate Influences of Estimation Dispersion

[33] X. Li, Z. Wang, and J. Yan, ‘‘Prognostic health condition for lithium battery QISHENG LIU was born in Wuhan, China,
using the partial incremental capacity and Gaussian process regression,’’ in 1962. He received the B.S. degree in electric-
J. Power Sources, vol. 421, pp. 56–67, May 2019. machinery from Hunan University, Chang-
[34] Z. Wang, J. Ma, and L. Zhang, ‘‘State-of-health estimation for lithium- sha, China, in 1984, and the M.S. degree in
ion batteries based on the multi-Island genetic algorithm and the Gaussian high-voltage and the Ph.D. degree in power system
process regression,’’ IEEE Access, vol. 5, pp. 21286–21295, 2017. and automation from Wuhan University, Wuhan,
[35] D. Yang, X. Zhang, R. Pan, Y. Wang, and Z. Chen, ‘‘A novel Gaussian pro- China, in 1992 and 2006, respectively, where he is
cess regression model for state-of-health estimation of lithium-ion battery
currently an Associate Professor with the School
using charging curve,’’ J. Power Sour., vol. 384, pp. 387–395, Apr. 2018.
of Electrical Engineering. His research interests
[36] J. Meng, L. Cai, G. Luo, D.-I. Stroe, and R. Teodorescu, ‘‘Lithium-ion
battery state of health estimation with short-term current pulse test and include power system operation and control, bat-
support vector machine,’’ Microelectron. Rel., vols. 88–90, pp. 1216–1220, tery charger, and renewable energy applications.
Sep. 2018.
[37] K. Qian, B. Huang, A. Ran, Y.-B. He, B. Li, and F. Kang, ‘‘State-of-
health (SOH) evaluation on lithium-ion battery by simulating the voltage
relaxation curves,’’ Electrochimica Acta, vol. 303, pp. 183–192, Apr. 2019.
[38] X. Li, Z. Wang, L. Zhang, C. Zou, and D. Dorrell, ‘‘State-of-health esti- SHIQI LIU received the B.Sc. degree from Wuhan
mation for Li-ion batteries by combing the incremental capacity analysis University, in 2017, where he is currently pursu-
method with grey relational analysis,’’ J. Power Sources, vols. 410–411, ing the Ph.D. degree. His main research interests
pp. 106–114, Jan. 2019. include Lithium battery characteristics, battery
[39] Y. Li, S. Zhong, Q. Zhong, and K. Shi, ‘‘Lithium-ion battery state of management systems in electric vehicles, and bat-
health monitoring based on ensemble learning,’’ IEEE Access, vol. 7, tery balance management systems.
pp. 8754–8762, 2019.
[40] M. Dubarry, M. Berecibar, A. Devie, D. Anseán, N. Omar, and I. Villarreal,
‘‘State of health battery estimator enabling degradation diagnosis: Model
and algorithm description,’’ J. Power Sources, vol. 360, pp. 59–69,
Aug. 2017.
[41] V.-H. Duong, H. A. Bastawrous, K. Lim, K. W. See, P. Zhang, and
S. X. Dou, ‘‘Online state of charge and model parameters estimation of
the LiFePO4 battery in electric vehicles using multiple adaptive forgetting
factors recursive least-squares,’’ J. Power Sources, vol. 296, pp. 215–224, XIONG XIE received the bachelor’s degree from
Nov. 2015. Wuhan University, in 2017. His research interests
[42] S. Liu, J. Wang, Q. Liu, J. Tang, H. Liu, and Z. Fang, ‘‘Deep-discharging include electricity market and economic operation,
Li-ion battery state of charge estimation using a partial adaptive forgetting especially on demand response of spot market.
factors least square method,’’ IEEE Access, vol. 7, pp. 47339–47352, 2019.
[43] A. Zenati, P. Desprez, H. Razik, and S. Rael, ‘‘A methodology to assess the
state of health of lithium-ion batteries based on the battery’s parameters and
a fuzzy logic system,’’ in Proc. IEEE Int. Electr. Vehicle Conf., Mar. 2012,
pp. 1–6.
[44] L. Han, Q. Mei, Y. Lu, and M. Ji, ‘‘Analysis and study on AHP-fuzzy
comprehensive evaluation,’’ China Saf. Sci. J., vol. 14, no. 7, pp. 86–89,
2004.
[45] PNGV Battery Test Manual, document DOE/ID-10597 Rev.3, 2001.
[46] EPA. (2019). In official Website of the United States Environmental
Protection Agency. Accessed: Apr. 23, 2019. [Online]. Available: JICHENG ZHOU received the bachelor’s degree
https://www.epa.gov/vehicle-and-fuel-emissions-testing/dynamometer- from Wuhan University, in 2017. His research
drive-schedules#Economic interest includes the subsynchronous resonance
[47] Cycle Life Requirements and Test Methods for Traction Battery of Electric of doubly fed induacion machine conneceted to
Vehicle, document GB/T 31484-2015, 2015. power systems.

JIA TANG received the bachelor’s degree from ZHONGXIANG LI is currently pursuing the bach-
Wuhan University, in 2017. His research inter- elor’s degree with Wuhan University. His main
ests include Lithium battery characteristics, bat- research interests include Li-ion battery character-
tery state estimation, and fault detection in electric istics and battery management systems in electric
vehicles. vehicles.

122314 VOLUME 7, 2019