This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
1

Lithium-ion battery health prognosis

based on a real battery management system used
in electric vehicles
Rui Xiong, Senior Member, IEEE, Yongzhi Zhang, Student Member, IEEE, Ju Wang, Hongwen He,
Senior Member, IEEE, Simin Peng, Member, IEEE, and Michael Pecht, Fellow, IEEE

 filters (KF) and particle filter (PF). These methods can estimate
Abstract—This paper developed an effective health indicator to the battery SOH online based on the measured current, voltage,
indicate lithium-ion battery state of health and moving-window- and temperature. The relationship between the open circuit
based method to predict battery remaining useful life. The health voltage (OCV) and the state of charge (SOC) at different battery
indicator was extracted based on the partial charge voltage curve SOHs and temperatures is usually required and thus collected
of cells. Battery remaining useful life was predicted using a linear offline. To simulate the complicated electrochemical
aging model constructed based on the capacity data within a
moving window, combined with Monte Carlo simulation to
characteristics of lithium-ion batteries, calculation-intensive
generate prediction uncertainties. Both the developed capacity algorithms are required, which restricts the application of these
estimation and remaining useful life prediction methods were methods in a battery management system (BMS).
implemented based on a real battery management system used in In the second category, feature engineering [11, 12] is used
electric vehicles. Experimental data for cells tested at different based on battery signals to extract health indicators (HIs). In
current rates, including 1C and 2C, and different temperatures, [11], the authors proposed a novel strategy for neutral vector
including 25 ℃ and 40 ℃, was collected and used. The variable decorrelation, which is an important topic in non-
implementation results show that the capacity estimation errors Gaussian data dimension reduction and feature selection. The
were within 1.5%. During the last 20% of battery lifetime, the root proposed methods are based on the neutrality of the data. The
mean square errors of remaining useful life predictions were
within 20 cycles, and the 95% confidence intervals mainly cover
advantages of the proposed decorrelation strategies are
about 20 cycles. impressive, and can be further investigated to implement to
battery HI extractions. Ref. [13] used the discharge terminal
Index Terms—Lithium-ion batteries; electric vehicles; voltage to extract a feature as the battery HI. Because the
remaining useful life; health indicator; moving window; battery terminal voltage response is highly dependent on the loaded
management system. current, the extracted HI from discharge terminal voltage only
operates when the battery is discharged at a constant current rate.
I. INTRODUCTION For lithium-ion battery application in EVs, where the discharge
current changes dramatically as the working conditions change,
L ITHIUM-ION batteries are widely applied in consumer
electronics, including cell phones, laptops, electric vehicles
(EVs), and military and aerospace electronics [1-4]. The
it is difficult to extract a reliable battery HI based on the
discharge voltage curve. In this case, an effective battery HI can
capacity of a lithium-ion cell, which is often used to indicate its be extracted from the charge voltage curve since the charge
state of health (SOH), degrades as the lithium-ion cell is cycled process for EVs is usually standardized. Li et al. [14] and Wang
with charging and discharging. Most companies set 80% of the et al. [15] both extracted the differential voltage (DV) curve
initial capacity as the end of life (EOL) criteria, since the cell’s during charging to indicate battery SOH. Hu et al. [16] extracted
capacity decreases at a more rapid rate after the EOL [5]. The five charge-related features, and Wu et al. [17] used the
on-board lithium-ion battery health prognosis technique importance sampling technique to extract 11 samples from the
enables an early report of the EOL for replacement purposes. voltage curve during charging as the battery HI. Refs. [14-17]
Battery SOH estimation methods fall into two categories. In extracted battery HIs from the complete charge voltage curve
the first category, the battery SOH is estimated based on the and thus required that the battery was charged from the lower
equivalent circuit models [6-8] or electrochemical models [9,10] cut-off voltage in practice. However, the lithium-ion batteries
in combination with advanced filter techniques such as Kalman for EVs are generally charged from a higher starting voltage

Copyright (c) 2015 IEEE. Personal use of this material is permitted. R. Xiong, Y. Zhang, J. Wang and H. He are with the National Engineering
However, permission to use this material for any other purposes must be Laboratory for Electric Vehicles, Department of Vehicle Engineering, School
obtained from the IEEE by sending a request to pubs-permissions@[Link]. of Mechanical Engineering, Beijing Institute of Technology, Beijing, 100081,
This work was supported in part by the National Natural Science Foundation China (rxiong@[Link], yzzhangbit@[Link], hwhebit@[Link]).
of China (Grant No. 51507012, 51507150) and Beijing Municipal Natural S. Peng is with School of Electrical Engineering, Yancheng Institute of
Science Foundation (Grant No. 3182035). The systemic experiments of the Technology, Yancheng, 224051, China (psmsteven@[Link]).
lithium-ion batteries were performed at the Advanced Energy Storage and M. Pecht is with Center for Advanced Life Cycle Engineering (CALCE),
Application (AESA) Group, Beijing Institute of Technology. (Corresponding University of Maryland, College Park, MD 20742, USA (pecht@[Link]).
author: Y. Zhang).

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
2

than the lower cut-off voltage. Therefore, an effective battery system (BMS) to validate the performance of the developed
HI should be extracted from the commonly used voltage ranges capacity estimation and RUL prediction methods.
during charging for EVs. A. Acceleration Aging Test
The current battery remaining useful life (RUL) prediction Figure 1 shows the equipment used to conduct the
techniques are mainly implemented as follows: First, a acceleration aging tests of lithium-ion batteries. The setup
nonlinear aging model is initialized based on the offline training includes an Arbin BT-5HC test system to charge/discharge
data, then the model is applied online combined with advanced batteries, three thermal chambers to control battery
filters such as PF to predict the battery RUL. He et al. [18] temperatures, and a computer for data monitoring and storage.
proposed a method for battery RUL prediction using the
Dempster-Shafer theory and the Bayesian Monte Carlo (BMC) Arbin BT-5HC Thermal Chamber
method. Model parameters were initialized offline by
combining sets of training data based on Dempster-Shafer
theory. The BMC method was then used to update the model
parameters and predict the RUL based on the available data Computer
through online monitoring of the battery capacity. Based on He
et al.’s methodology [18], a series of papers including Xing et
al. [19], Miao et al. [20], Liu et al. [21], and Su et al. [22] were
published to improve either the aging model accuracy [19] or
the filtering performance [20-22] for more accurate RUL
prediction. Lyu et al. [23] contributed by adding some
electrochemical characteristics to the aging model. Wang et al.
[24] used nonlinear-drifted Brownian motion to construct a
battery aging model, and then employed the unscented extended Fig. 1 Battery test equipment
Kalman filter (UEKF) to predict the battery RUL. This method
presented higher prediction accuracy than the standard PF- The aging test used high-energy 18650 lithium-ion batteries
based prognostic method. In refs. [18-24], offline training data manufactured by Panasonic, labeled NCR18650PF. The cells
is required to initialize an aging model. The lithium-ion had a rated capacity of 2.7 Ah. The nominal voltage was 3.6 V,
batteries for EVs work under complicated conditions with with the upper and lower cut-off voltages being 4.2 V and 2.5
changeable loads. Therefore, it is difficult to design V, respectively. The materials consisted of graphite on the
acceleration aging tests to collect effective offline training data anode and Li(NiCoAl)O2 on the cathode. The cell test
for lithium-ion batteries under similar working conditions as conditions are summarized in Table 1. In the cycling test, the
those in practice. charging profile of all cells was a constant current (CC)-
This paper presents three improvements to the current constant voltage (CV) process with a constant current rate of
techniques for battery health prognosis. First, the battery HI was 0.5C, followed by a constant voltage of 4.2 V. The cells were
extracted from a partial charge voltage curve within commonly discharged at 1C and 2C current rates, respectively, until the
used ranges. A model was used to describe the relationship low cut-off voltage was reached, then a 0.5C current rate was
between the extracted HI and capacity, and based on the loaded on all cells until the low cut-off voltage was reached
constructed model, the capacity can be estimated online with again. The rest time between charging and discharging was 30
low calculation burden. Second, a moving-window-based min. The characterization tests, including the capacity
method was developed to predict the battery RUL independent calibration test, the dynamic stress test (DST), and the open
of offline training data, and this method was combined with circuit voltage (OCV) test, were conducted on all cells every 50
Monte Carlo (MC) simulation to generate RUL prediction or 100 cycles. The charge profile of the capacity calibration test
uncertainties. Finally, unlike the conventional simulation study was a CC-CV process with a constant current and voltage of
of battery states estimation, experimental study was conducted 0.5C and 4.2 V, respectively, and the cutoff current was 0.05C.
in this paper to predict battery SOH and RUL based on a real After a rest of 1 h, the cell was discharged at a 0.5C rate until
BMS. Experimental results showed that the BMS predicted the lower cutoff voltage 2.5 V was reached. The
battery states accurately and efficiently. charge/discharge process was repeated three times. The
The rest of this paper is organized as follows: Section 2 discharge capacity in the last cycle was considered the battery
introduces the acceleration aging test of lithium-ion batteries capacity and used to indicate the cell’s health state. The DST
and the BMS. Section 3 describes the theory for battery HI test and OCV test results have not been used in this paper.
extraction and RUL prediction. Section 4 describes the Table 1 Test conditions for the NCR18650PF battery
developed methods for battery HI extraction and RUL Test Category Test Condition 1 Test Condition 2
prediction, and shows the BMS-based prediction results, Cycle test Charge: CCCV, charged Charge: CCCV, charged
followed by the conclusions in Section 5. at 0.5C rate up to 4.2 V at 0.5C rate up to 4.2 V
0.05C rate-cutoff 0.05C rate-cutoff
II. EXPERIMENT Charge rest: 30 min Charge rest: 30 min
This section introduces the experiments conducted, including Discharge: 1C rate and Discharge: 1C rate and
the acceleration aging tests to investigate the lithium-ion battery 2C rate, 2.5 V cutoff 2C rate, 2.5 V cutoff
aging characteristics and the setup of a real battery management Discharge rest: 30 min Discharge rest: 30 min

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
3

Temperature: 25 ℃ Temperature: 40 ℃ A. HI extraction

Characterization Capacity calibration- Capacity calibration- Figure 3 shows the terminal voltage curves during charging
tests OCV test-DST test OCV test-DST test of cells at 1C rate and 40 ℃. Fig. 3(a) shows the terminal
voltage responses of a single cell at different cycles. When the
B. Battery Management System number of cycles increased, the terminal voltage was higher at
The BMS used to validate the developed methods consisted the same charge moment and the cell reached the higher cut-off
of a microcontroller unit (MCU), a debugger, a controller area voltage earlier than the cell with less number of cycles. This
network (CAN) bus, and a host computer (Figure 2). The MCU early cut-off of CC charge results in less battery capacity.
model is MPC5644A, which is commonly used in vehicles Therefore, the charge capacity resulting from the CC process
owing to its high computation capability. The maximum clock can indicate the cell’s health state. Fig. 3(b) compares the
frequency (MCF) of the MCU is 150 MHz, indicating high voltage curves of two cells during charging and shows that the
calculation speed. The MCU’s flash memory is 4 MB, and its curves are similar at the same number of cycles. This similarity
random-access memory (RAM) is 192 KB. The MCU illustrates that the charge capacity in the CC process indicates a
implements single-precision floating-point calculations with 7 robust health state against different cells.
significant digits of precision. Any stored single value in the Generally, to improve the usage safety and extend the
MCU requires 32 bits of memory. The maximum lifetime, the battery will be recharged when its SOC decreases
communication speed (MCS) of the CAN bus is 1 M/s, to 10%-20%. The test data showed that the battery SOC is
indicating high communication capability. The coding for the around 20% with a terminal voltage of 3.6V. Therefore, this
developed methods was written using the host computer and study selected the charge capacity within a commonly used
then were debugged and downloaded to the MCU. The MCU voltage range from 3.6 V to 4.2 V to indicate the battery health
communicated with the host computer via the CAN bus. The state. Figure 4 shows the voltage curves during charging after
debugged codes run in the MCU. Except for code writing, the 3.6 V at different cycles. As the cell ages, the time it takes for
host computer can also be used for data storage and results the terminal voltage to reach 4.2 V decreases and thus the
display. charge capacity obtained from 3.6 V to 4.2 V also decreases.
The charge capacity from 3.6 V to 4.2 V, which indicates the
cell’s health state, is thus used as a battery HI.
t t1
HI   i  t  dt
CAN Bus
(1)
t  t0

where i(t) represents the charge current and is constant in this

study; 𝑈 represents the terminal voltage; and 𝑇 represents the
time taken to charge the battery from 3.6 V to 4.2 V. Note that
dynamic current can also be input into Eq. (1) to extract the
battery HI when the charging current is not constant.
Host Computer
MCU
Debugger
Fig. 2 Battery management system

III. CAPACITY ESTIMATION AND RUL PREDICTION

This section introduces the theory behind the developed
capacity estimation and RUL prediction techniques. An
effective HI was extracted based on the partial voltage curve
during charging. A moving-window-based technique was
developed for battery RUL prediction by identifying Fig. 4 Charge voltage curves from 3.6 V to 4.2 V at different cycles at 1C
distinguishable aging stages of lithium-ion batteries. rate and 40 ℃.

Cell 1
Cell 2

Fig. 3 Voltage curves during charging at 1C rate and 40 ℃ of: (a) the same cell at different cycles; (b) cells 1 and 2 at different cycles.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
4

B. RUL prediction this case, when the SOH is lower than 90%, a linear model
The lithium-ion battery SOH is represented by the present describing the degradation trend of stage 2 can be constructed.
capacity divided by the initial capacity, which is expressed as This linear aging model can then be extrapolated to predict the
Cp cell’s EOL. Fig. 5(b) shows the capacity degradation rate per
SOH  100% (2)
Ci cycle at different stages at 25 ℃, in which the degradation rate
decreases as the battery aging stage evolves. Fig. 5(c) shows the
where 𝐶i represents the initial capacity and 𝐶p represents the battery SOH at different cycles at 40 ℃, in which, unlike cells
present capacity. In this paper, the battery reaches the EOL tested at 25 ℃, the cell EOL falls within stage 3. In Fig. 5(c),
when the its capacity degrades to 80% of the initial value. stage 3 starts at 400 cycles and about 85% SOH, and ends at
Battery capacity degrades at different stages and at each stage, 800 cycles and about 80% SOH. In this case, a linear aging
a linear capacity degradation trend is observed [25]. Baumhöfer model can be constructed for the battery RUL prediction when
et al. [26] tested cells with a carbon anode and a NMC cathode. the SOH is lower than 85%. Fig. 5(d) shows the capacity
The aging test results showed a slow and linear capacity degradation rate per cycle at different stages at 40 ℃, which
degradation in the first stage, and when the cell SOH was lower also decreases as the aging stage evolves. Fig. 5 shows that the
than about 85%, the cell degradation reached the second stage battery capacity degrades faster under larger current C-rates.
with a much higher degradation rate. Dubarry et al. [27] tested The aging mechanisms behind this phenomenon are out of the
cells with a composite anode comprising research scope of this paper and should be investigated in the
{LiMn1/3Ni1/3Co1/3O2+LiMn2O4} and a graphite cathode. The future.
cells were tested at different C rates, and two degradation stages
similar to those in ref. [24] were present. The second stage was IV. APPLICATION IN A REAL BATTERY MANAGEMENT SYSTEM
reached at more than 500 cycles. Han et al. [28] found two
linear capacity degradation stages in cells with a LTO anode This section introduces the algorithms of the developed
and a LiNixCoyMn1-x-yO2 cathode. Gao et al. [29] found three techniques and the application results based on a real battery
capacity degradation stages in cells with a graphite anode and a management system. The performance of the lithium-ion
LiCoO2 cathode, and stage 2 and stage 3, respectively, started battery HI extraction and RUL prediction techniques are both
from about 95% SOH and 87% SOH. The capacity degradation validated through experimental data of 8 cells at different
rate decreased as the battery SOH degraded from stage 1 to current rates, including 1C and 2C, and different temperatures,
stage 3. including 25 ℃ and 40 ℃.
This paper tested cells consisting of a graphite anode and a A. HI extraction
Li(NiCoAl)O2 cathode, whose capacity degradation Figure 6 shows a relationship of the cell capacity vs. HI,
characteristics have not been analyzed in the literature. where the HI was extracted at different cycles based on Eq. (1).
Experimental results showed three nearly linear capacity The cell was tested at 1C and 25 ℃, and the capacity and HI
degradation trends, which are similar to the results in ref. [29]. both degraded as the cell aged.
Figure 5 shows the capacity degradation results of the The fitting curve in Fig. 6 was obtained by using a third-order
graphite/Li(NiCoAl)O2 battery at 1C and 2C rates and 25 ℃ polynomial as follows,
and 40 ℃. Fig. 5(a) shows the battery capacity aging at 25 ℃, C  p0  p1 HI  p2 HI 2  p3 HI 3 (3)
where the battery EOL occurs at the end of the second stage of where 𝐶 represents the cell capacity and 𝑝𝑖 (𝑖 = 0, 1,2,3) are
around 500 cycles. Stage 2 starts at 150 cycles, where the SOH the fitting coefficients, which are identified using Matlab’s
is about 90% and covers about 70% of the battery lifetime. In curve fitting tool. The root mean square error (RMSE) of the

Fig. 5 Battery capacity degradation at different C rates and temperatures: (a) Battery SOH at 25 ℃; (b) Battery capacity degradation per cycle at 25 ℃; (c)
Battery SOH at 40 ℃; (d) Battery capacity degradation per cycle at 40 ℃.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
5

fitting is 8.06 ×10-3Ah, indicating a high modeling accuracy of 0.71%. To evaluate the robustness of the extracted HI against
the third-order polynomial. These fitting coefficients are saved different C rates, experimental data at 2C and 25 ℃ was also
in the MCU to estimate the battery capacities based on extracted used, and the capacity estimation results are shown in Figs. 7(c)
HI values. and (d). The capacity estimates at 2C are close to the real data
at different cycles, and the estimation errors are almost all
within 1.5 %. The RMSE of capacity estimation is 1.16%. The
capacity estimation accuracy at 2C indicates that at each
temperature, only one cell is required to estimate the capacities
of cells at the same temperature without considering the C rate.
To further evaluate the estimation performance of the
extracted HI, the experimental data at 40 ℃ was used. Figure 8
shows the relationship of HI vs. capacity of one cell at 1C and
40 ℃, and Eq. (3) was used to fit this relationship. The RMSE
of the fitting is 8.53 ×10-3Ah, again indicating a high modeling
accuracy.
Fig. 6 Relationship of capacity and HI of one cell at 1C and 25 ℃. The fitted polynomial based on aging data in Fig. 8 was
To realize the on-board capacity estimation, the HI values stored in the MCU and used to estimate the capacities of other
were first extracted from the charge data during the CC process. cells at 40 ℃, and the capacity estimation results are shown in
At each sampling moment, the signal, which included one set Figure 9. Figs. 9(a-b) show the capacity estimates of cells tested
of current and voltage required for the HI extraction, was sent at 1C and 40 ℃, and the capacity estimation errors of these cells
to the MCU. The sampling interval was 1 s, which was the same are shown in Fig. 9(d). The capacity estimates of cells at 1C and
as that of the current and voltage collected in the acceleration 40 ℃ are close to the real data, and Fig. 9(d) shows that the
aging test of lithium-ion batteries. The extracted HI was capacity estimation errors at 1C and 40 ℃ are almost all less
calculated by accumulating the product of current and sampling than 1.5%. The RMSEs of these capacity estimates at 1C and
40 ℃ are, respectively, 0.75% and 0.74%. Figs. 9(c-d) show the

Fig. 7 Capacity estimation results at 25 ℃: (a) Estimates of cell 1 at 1 C; (b) Estimates of cell 2 at 1 C; (c) Estimates of cell 3 at 2 C; (d) Estimation errors
of cells 1-3.
interval between 3.6 V and 4.2 V. The extracted HI values were capacity estimates at 2C and 40 ℃, and the capacity estimation
then input into Eq. (3) for estimating the cell’s capacity at the
same temperature.
To evaluate the effectiveness of the extracted HI, Eq. (3) is
parameterized based on Fig. 6 to estimate the capacities of three
other cells at the same temperature, and the capacity estimation
results based on the BMS are shown in Figure 7. The extracted
HI of each cell at each specified cycle was input into the
parameterized polynomial for estimating the capacity. Figs.
7(a-b) show the capacity estimates of the cells at 1C and 25 ℃
and the capacity estimation errors are shown in Fig. 7(d). The
capacity estimates of the cells at 1C are all close to the real data,
and the capacity errors are almost all within 1%. The RMSEs
of these two capacity estimates are, respectively, 0.87% and Fig. 8 Relationship of capacity and HI of one cell at 1C and 40 ℃.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
6

Fig. 9 Capacity estimation results at 40 ℃: (a) Estimates of cell 1 at 1 C; (b) Estimates of cell 2 at 1 C; (c) Estimates of cell 3 at 2 C; (d) Estimation
errors of cells 1-3.

errors are almost all within 1.5%, indicating accurate capacity RUL prediction accuracy can decrease if there is a slow
estimation. The RMSE of capacity estimation at 2C and 40 ℃ response to the capacity degradation change.
is 1.26%, indicating the developed capacity estimation method
is robust against different discharge C rates.
B. RUL prediction
This section first introduces practical problems of predicting
battery RUL and develops algorithms for addressing these
problems. Then, the developed algorithms were applied to
predict RUL of cells at different current rates and temperatures
based on the BMS.
1) Related algorithms
Based on the identification results of lithium-ion battery
capacity degradation in Section 3.2, a linear aging model can be
constructed based on capacity data at the corresponding aging
stage. Two problems should be considered: first, the battery Fig. 10 Moving-window-based RUL prediction method.
EOL and the collected capacities for linear aging model In Fig. 10, the capacities within each moving window are
construction can be at different aging stages, and second, the used to build a linear aging model as follows:
battery capacities at the same aging stage generally fluctuate
around strictly linear degradation. These problems can result in C  b + kb  ε 0 1
(4)
a linear aging model with large fitting errors and in reduced where 𝐶 represents the battery capacities within each window;
battery RUL prediction accuracy. Considering that the 𝑘 represents the corresponding cycles; 𝜀 represents the
constructed linear aging model predicts a more accurate RUL independent random errors that are normally distributed with a
because the collected capacities are closer to the battery EOL, zero mean and variance of 𝜎 2 ; and 𝑏0 and 𝑏1 are the modeling
a moving-window-based RUL prediction method (Fig. 10) was coefficients, which can be identified using the least squares
developed to improve RUL prediction accuracy. When RUL algorithm. The following equations are the result:
prediction started, a linear aging model was constructed based
on the capacities within a window, which moved forward by bˆ1   i 1 i
 n

k  k Ci  C  (5)


 i 1 i  
n 2
neglecting old capacities and collecting new capacities. The  k  k
window size, which indicates the capacity samples collected 
bˆ0  C  bˆ1k (6)
within the window, can be determined according to the specific
application. When a small window is used, this model can where n represents the window size, 𝑘̅ = ∑𝑖=1 𝑘𝑖 ⁄𝑛, and 𝐶̅ =
𝑛
capture the capacity degradation characteristics rapidly, ∑𝑛𝑖=1 𝐶𝑖 ⁄𝑛.
however, the resulting RUL predictions are easily affected by The parameter uncertainties, which are represented as the
the newly collected capacity data and thus can fluctuate variances of the coefficients 𝑏0 and 𝑏1 , are estimated as
strongly. When a large window is used, the model predicts
robust RULs against newly collected capacities, however, the

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
7

 s2 distribution N( 𝑏̂1 , std(𝑏1 ) ). N( · ) represents the Gaussian

Var  b1   n (7)
i 1  i
 k k distribution. Once the uncertainty sampling of b0 and b1 is
2

 completed, each generated data set (b0, b1) is input into the
   aging model to make multiple-step-ahead SOH predictions in
 2 1 k2 
Var  b0   s   n 2  (8) step 3. If the predicted SOH is less than the threshold value, the

 
n
 i 1  i  
k  k EOL is obtained in step 4. For each prediction, there are a total
of 103 predicted RULs, which can be used to generate the RUL
where 𝑠 2 , which represents the estimate of 𝜎 2 of the error term
PDF.
𝜀 in Eq. (4), is estimated as 𝑠 2 = 𝑆𝑆𝑅 ⁄(𝑛 − 2). SSR represents
the sum of squared residuals and can be expressed as:
2) Application to experimental data
    C  b  b k 
n n
SSR  
2
Ci  Cˆ i
2
(9) Because the cell capacity was calibrated every 50 or 100
i 0 1 i
i 1 i 1 cycles, the experimental data points were not enough to
RUL prediction uncertainties are also required in practice to calculate the values of b0 and b1. To generate enough capacity
evaluate the prediction precision. This paper used MC data points, the cell capacities at other cycles were obtained
simulation to generate the probability distribution function through linear interpolation according to the experimental data.
(PDF) of the RUL prediction. Figure 11 shows the developed This data pre-processing is practical, because many methods
battery RUL prediction method. The on-board capacity data is have been developed for battery capacity estimation in real-
collected based on the MCU. In step 1, the data within the time based on measured currents and voltages [31, 32]. Figures
moving window is used to identify the model parameters (𝑏0 12 and 13 show the BMS-based battery RUL prediction results

Step 1 Step 2 Step 3 Step 4

On-board Parameter identification of Uncertainty Sampling Aging Model EOL Threshold
Capacity aging model (Eq. (4)) Eqs. (7) and (8) (Eq. (4)) Comparison
Data

bˆ0 ,std  b0  b0 
N bˆ0 ,std  b0   Multiple-Step-
RUL
bˆ ,std  b 
1 1 b1 
N bˆ ,std  b 
1 1  Ahead Prediction

Fig. 11 RUL prediction method.

and 𝑏1 ) of Eq. (4). 𝑏̂0 and 𝑏̂1 are obtained based on Eqs. (5) and using the moving-window-based method at 25 ℃ and 40 ℃,
(6). The standard deviations of 𝑏0 and 𝑏1 , which are respectively. At each temperature, the RUL predictions were
represented as std(𝑏0 ) and std(𝑏1 ) in Fig. 11, are obtained by obtained based on window sizes of 50, 100, 150, and 200 cycles.
extracting square root of the corresponding variances (Eqs. (7) The battery RUL predictions at two temperatures were both
and (8)). Step 2 generates 103 sample sets of (b0, b1) based on updated every 10 cycles. At 25 ℃, the battery RUL was
the probability distributions of these two parameters using MC predicted from the 150th cycle, whereas at 40 ℃, because the
simulation [30]. That is, a total of 103 samples of b0 is generated window-based method can predict RUL by forgetting old
based on the probability distribution N(𝑏̂0 , std(𝑏0 )), whereas a information, the battery RUL was predicted from the same
total of 103 samples of b1 is generated based on the probability cycle. To ensure the prediction stability, the first RUL

Fig. 12 Battery RUL predictions at 25 ℃ and: (a-c) 1C; (e) 2C.

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
8

prediction was obtained when 50 cycles’ data points were Figs. 12(a-c) show the RUL prediction results at 1C rate,
available. Therefore, the first battery RUL was predicted at the whereas Fig. 12(d) shows the RUL prediction results at 2C rate.
200th cycle. For a window size larger than 50 cycles, such as As expected, the RUL prediction using a small window size
150 cycles, no data was neglected when the available data size fluctuates more strongly than that using a large window size
was smaller than 150 cycles, and the method started to forget owing to the data variations, which can be observed in Fig. 12.
old data from the 300th cycle. The error thresholds indicate The started cycles with RUL prediction errors falling within the
predictions 50 cycles away from the real data. Tables 2 and 3 error thresholds based on different window sizes are the same.
show the started cycle with a RUL prediction error falling For cells 1 and 2, this started cycle is 200, whereas for cells 3
within the error thresholds. Because the battery replacements of and 4, this started cycle is 210. Table 2 shows that, for each cell,
EVs are highly dependent on the RUL predictions near the the RMSE of RUL predictions during the last 20% cycles is the
battery EOL, the two tables also show the RMSEs of the battery smallest at a window size of 200, which phenomenon indicates
RUL predictions during the last 20% of cycles of the battery that, extending the window size is an effective way to reduce
lifetime. The RMSE of RUL predictions is defined as the RUL prediction error against data variations. Therefore, a
window size of 200 can be selected at 25 ℃ for accurate RUL
  RUL  RUL r ,i 
iL 2
RMSE  i  ks p ,i N (10) predictions with RMSEs vary around 10 cycles.
where ks is the started cycle for battery RUL predictions, L is Fig. 13 shows the BMS-based RUL prediction results at
the cycle indicating battery EOL, RULp,i is the predicted value 40 ℃, where Figs. 13(a-c) show the RUL prediction results at
at cycle i, RULr,i is the real value at cycle i, and N is the total 1C rate and Fig. 13(d) show the RUL prediction results at 2C
number of predicted RULs from cycle ks to cycle L. Since the rate. Fig. 13 shows that the method with a small window size
RUL is predicted every 10 cycles in the study, N should be predicts RULs with strong variations owing to the data
equal to ⌈(L-ks+1)/10⌉. For RMSE of predictions during the last fluctuations. However, with a small window size, the method
20% of cycles, ⌊ ks/L ⌋ should be equal to 80%. ⌈·⌉ and ⌊·⌋ are can catch the new data information efficiently, and construct a
mathematical symbols to, respectively, round a decimal up and model that predicts the battery RUL accurately. Table 3 shows
down to the nearest integer. that the predicted RUL with a smaller window size generally
fell within the error thresholds earlier for each cell, and the
Table 2 Battery RUL prediction results at 25 ℃
Started Cycle Falling Within Error Threshold RMSE During Last 20% Cycles
Window size Cell 1 Cell 2 Cell 3 Cell 4 Cell 1 Cell 2 Cell 3 Cell 4
50 200 200 210 210 18 14 21 54
100 200 200 210 210 13 14 15 30
150 200 200 210 210 13 15 10 16
200 200 200 210 210 12 14 8 9

Fig. 13 Battery RUL predictions at 40 ℃ and: (a-c) 1C; (e) 2C.

Table 3 Battery RUL prediction results at 40 ℃

Started Cycle Falling Within Error Threshold RMSE During Last 20% Cycles
Window size Cell 1 Cell 2 Cell 3 Cell 4 Cell 1 Cell 2 Cell 3 Cell 4
50 430 570 480 500 19 24 19 9
100 470 510 490 520 24 31 18 12
150 510 530 520 540 29 33 20 20
200 540 550 550 550 38 25 25 31

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
9

Real RUL 50 200

(a) (c)
300 300
Cell 1: 1C, 25℃ Cell 3: 1C, 40 ℃
250 250

200 200

RUL
150

RU L
150
100 100
RUL PDF
50 50
0 0
-50 -50
200 250 300 350 400 400 450 500 550 600 650 700
Equivalent Full Cycle Equivalent Full Cycle
(b) (d)
300 300
Cell 2: 2C, 25℃ Cell 4: 2C, 40 ℃
250 250
200 200
RUL

RU L
150 150
100 100
50 50
0 0
-50 -50
200 250 300 350 400 450 400 450 500 550 600 650
Equivalent Full Cycle Equivalent Full Cycle
Fig. 14 RUL prediction results based on the moving-window method of cells tested at: (a) 1C and 25 ℃; (b) 2C and 25 ℃; (c) 1C and 40 ℃; (d) 2C and
40 ℃.
corresponding RMSE during the last 20% cycles was also charge/discharge for EV batteries takes several hours or several
smaller. This phenomenon is different than that at 25 ℃, when days, and the developed method can thus predict the RUL at
the predicted RULs with different window sizes generally fell each cycle in real applications.
within the error thresholds at the beginning of the prediction. 3) Advancements to the current techniques
Also, with a smaller window size, the method predicted less The developed RUL prediction method advances the current
accurate RULs near the battery EOL at 25 ℃. This difference techniques from the following three aspects. First, the
of prediction is because that, at 25 ℃, cells under different developed method operates independent of any offline training
current rates degraded linearly after 150 cycles with few turning data, whereas PF-based and machine learning methods
points, whereas, at 40 ℃, one obvious turning point occurred at generally require a large amount of offline data, which data is
the 400th cycle for each cell. Therefore, less accurate RUL obtained under similar working conditions as that of online data.
predictions can be caused at 40 ℃ with a larger moving window, Second, model parameters of the developed method can be
which covered the turning point for a longer time. In this case, identified accurately and easily based on Eqs. (5) and (6).
a small window size such as 50 can be selected for RUL However, it is generally difficult to locate optimal model
predictions at 40 ℃. parameter sets for both PF-based and machine learning methods.
In practice, a suitable window size should be determined Finally, compared with the current methods, the computational
based on the variation characteristics of capacity data. If the burden of the developed method is low owing to its simple
capacities show strong variations, a large window size such as calculation processes.
200 should be selected to improve the RUL prediction
robustness. If the capacities show obvious different slopes of V. CONCLUSIONS
two adjacent linear aging stages, a small window size such as State of health (SOH) prognostics of lithium-ion batteries
50 should be selected to capture the data variations rapidly. requires accurate online capacity estimation and remaining
Fig. 14 shows the predicted RUL PDF at different cycles for useful life (RUL) prediction. This study developed methods to
four cells selected from Figs. 12 and 13. Figs. 14(a)-(b) show estimate battery capacity with low computation burden and to
the RUL prediction results with a window size of 200 for cells predict battery RUL independent of offline training data.
at 25 ℃, whereas Figs. 14(c)-(d) show the RUL prediction An effective health indicator (HI) based on the partial charge
results with a window size of 50 for cells at 40 ℃. Fig. 14 shows voltage curve and a model that describes the relationship
that the developed method predict RUL with both high accuracy between the HI and the capacity at different aging states were
and precision, where the 95% confidence intervals generally developed. Only one set of offline training data at each
cover number of cycles within 20. In Fig. 14(a), the 95% temperature was required to construct the model. Based on
confidence interval of RUL prediction at 250 cycles is [256, which model, the capacity of other cells at the same temperature
272], and at 350 cycles, the 95% interval decreases to [41, 55]. was estimated online with low calculation burden. The RUL
The run time to predict each RUL and generate each RUL PDF prediction method was developed by identifying the specific
using MC simulation based on the BMS is generally within 5 s. degradation characteristics of cells. Experimental results show
This time includes the time spent to download the algorithms that a battery typically goes through three distinguishable linear
from the host computer, the calculation time of the MCU, and degradation stages, and its end of life (EOL) at 25 ℃ and 40 ℃
the delay of the CAN bus for data communication. Therefore, occurs in the second and third stages, respectively. Therefore, a
the calculation time cost by the MCU for each RUL PDF linear aging model within the moving-window was developed
generation can be shorter. In practice, one cycle of

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
10

to predict RUL. The collected data updated within a moving Transactions on Neural Network and Learning Systems, 2018, DOI:
10.1109/TNNLS.2018.2844399.
window was used to construct a linear aging model, which was [13] D. Liu, J. Zhou, H. Liao, Y. Peng, and X. Peng, “A health indicator
used to predict battery EOL by extrapolation. The developed extraction and optimization framework for lithium-ion battery
capacity estimation and RUL prediction techniques were both degradation modeling and prognostics,” IEEE Transactions on Systems,
implemented in a real battery management system (BMS) to Man, and Cybernetics: Systems, vol. 45, no. 6, pp. 915–928, Jun. 2015.
[14] X. Li, J. Jiang, L. Y. Wang, D. Chen, Y. Zhang, and C. Zhang, “A capacity
verify the performance. model based on charging process for state of health estimation of lithium
The developed capacity estimation and RUL prediction ion batteries,” Applied Energy, vol. 177, pp. 537–543, Sep. 2016.
methods operated efficiently and accurately based on the BMS. [15] L. Wang, C. Pan, L. Liu, Y. Cheng, and X. Zhao, “On-board state of
Experimental data of 8 cells tested at different temperatures health estimation of LiFePO4 battery pack through differential voltage
analysis,” Applied Energy, vol. 168, pp. 465–472, Apr. 2016.
including 25 ℃ and 40 ℃, and different C rates including 1C [16] C. Hu, G. Jain, C. Schmidt, C. Strief, and M. Sullivan, “Online estimation
and 2C were collected and used. The on-board capacity of lithium-ion battery capacity using sparse Bayesian learning,” Journal
estimation errors of cells tested at 1C and 2C were mostly of Power Sources, vol. 289, pp. 105–113, Sep. 2015.
within 1.5%, indicating high accurate capacity estimation. The [17] J. Wu, C. Zhang, and Z. Chen, “An online method for lithium-ion battery
remaining useful life estimation using importance sampling and neural
moving-window-based method predicted accurate RUL, and networks,” Applied Energy, vol. 173, pp. 134–140, Jul. 2016.
during the last 20% of the battery’s lifetime cycles, the root [18] W. He, N. Williard, M. Osterman, and M. Pecht, “Prognostics of lithium-
mean square errors (RMSEs) of RUL predictions with suitable ion batteries based on Dempster–Shafer theory and the Bayesian Monte
window sizes were mostly within 20 cycles. Monte Carlo (MC) Carlo method,” Journal of Power Sources, vol. 196, no. 23, pp. 10314–
10321, Dec. 2011.
simulation was used to generate the RUL probability [19] Y. Xing, E. W. Ma, K. L. Tsui, and M. Pecht, “An ensemble model for
distribution function (PDF). Implementation results showed predicting the remaining useful performance of lithium-ion batteries,”
that the 95% confidence interval of RUL predictions were Microelectronics Reliability, vol. 53, no. 6, pp. 811–820, Jun. 2013.
mainly within 20 cycles. The MC simulation time based on the [20] Q. Miao, L. Xie, H. Cui, W. Liang, and M. Pecht, “Remaining useful life
prediction of lithium-ion battery with unscented particle filter technique,”
BMS for each RUL PDF generation was within 5 s, indicating Microelectronics Reliability, vol. 53, no. 6, pp. 805–810, Jun. 2013.
the practicality of the developed method. [21] Z. Liu, G. Sun, S. Bu, J. Han, X. Tang, and M. Pecht, “Particle learning
framework for estimating the remaining useful life of lithium-ion
REFERENCES batteries,” IEEE Transactions on Instrumentation and Measurement, vol.
[1] Y. Zhang, R. Xiong, H. He, and M. Pecht, “Long short-term memory 66, no. 2, pp. 280–293, Feb. 2017.
recurrent neural network for remaining useful life prediction of lithium- [22] X. Su, S. Wang, M. Pecht, L. Zhao, and Z. Ye, “Interacting multiple
ion batteries,” IEEE Transactions on Vehicular Technology, vol. 67, no. model particle filter for prognostics of lithium-ion batteries,”
7, pp. 5695–5705, Jul. 2018. Microelectronics Reliability, vol. 70, pp. 59–69, Mar. 2017.
[2] Y. Song, D. Liu, C. Yang, and Y. Peng, “Data-driven hybrid remaining [23] C. Lyu, Q. Lai, T. Ge, H. Yu, L. Wang, and Na Ma, ” A lead-acid battery's
useful life estimation approach for spacecraft lithium-ion battery,” remaining useful life prediction by using electrochemical model in the
Microelectronics Reliability, vol. 75, pp. 142–153, Aug. 2017. Particle Filtering framework,” Energy, vol. 120, pp. 975–984, Feb. 2017.
[3] R. Xiong, Y. Zhang, H. He, X. Zhou, and M. Pecht, "A double-scale, [24] D. Wang, Y. Zhao, F. Yang, and K. Tsui, “Nonlinear-drifted Brownian
particle-filtering, energy state prediction algorithm for lithium-ion motion with multiple hidden states for remaining useful life prediction
batteries," IEEE Transactions on Industrial Electronics, vol. 65, no. 2, of rechargeable batteries,” Mechanical Systems and Signal Processing,
pp. 1526–1538, Feb. 2018. vol. 93, pp. 531–544, Sep. 2017.
[4] Y. Zhang, R. Xiong, H. He, and M. Pecht, “Lithium-ion battery [25] M. Dubarry, C. Truchot, and B. Y. Liaw, “Synthesize battery degradation
remaining useful life prediction with Box–Cox transformation and modes via a diagnostic and prognostic model,” Journal of Power Sources,
Monte Carlo simulation,” IEEE Transactions on Industrial Electronics, vol. 219, pp. 204–216, Dec. 2012.
2018, DOI: 10.1109/TIE.2018.2808918. [26] T. Baumhöfer, M. Brühl, S. Rothgang, and D. U. Sauer, “Production
[5] J. L. Zhang and J. Lee, “A review on prognostics and health monitoring caused variation in capacity aging trend and correlation to initial cell
of Li-ion battery,” Journal of Power Sources, vol. 196, no. 15, pp. 6007– performance,” Journal of Power Sources, vol. 247, pp. 332–338, Feb.
6014, Aug. 2011. 2014.
[6] J. Bi, T. Zhang, H. Yu, Y. Kang, “State-of-health estimation of lithium- [27] M. Dubarry, C. Truchot, B. Y. Liaw, K. Gering, S. Sazhin, D. Jamison,
ion battery packs in electric vehicles based on genetic resampling particle and C. Michelbacher, “Evaluation of commercial lithium-ion cells based
filter,” Applied Energy, vol. 182, pp. 558–568, Nov. 2016. on composite positive electrode for plug-in hybrid electric vehicle
[7] Z. Wei, K. J. Tseng, N. Wai, T. M. Lim, and M. Skyllas-Kazacos, applications. Part II. Degradation mechanism under 2 C cycle aging,”
“Adaptive estimation of state of charge and capacity with online Journal of Power Sources, vol. 196, no. 23, pp. 10336–10343, Dec. 2011.
identified battery model for vanadium redox flow battery,” Journal of [28] X. Han, M. Ouyang, L. Lu and J. Li, “Cycle Life of Commercial
Power Sources, vol. 112, pp. 469–480, Oct. 2016. Lithium-Ion Batteries with Lithium Titanium Oxide Anodes in Electric
[8] Z. Deng, L. Yang, Y. Cai, H. Deng, and Liu Sun, “Online available Vehicles,” Energies, vol. 7, no. 8, pp. 4895–4909, 2014.
capacity prediction and state of charge estimation based on advanced [29] Y. Gao, J. Jiang, C. Zhang, W. Zhang, Z. Ma, and Y. Jiang, “Lithium-ion
data-driven algorithms for lithium iron phosphate battery,” Energy, vol. battery aging mechanisms and life model under different charging
201, pp. 257–269, Sep. 2017. stresses,” Journal of Power Sources, vol. 356, pp. 103–114, Jul. 2017.
[9] A. Bartlett, J. Marcicki, S. Onori, G. Rizzoni, X. G. Yang, and T. Miller, [30] B. Zhang, L. Tang, J. DeCastro, and K. Goebel, “A verification
“Electrochemical model-based state of charge and capacity estimation methodology for prognostic algorithms,” in Proc. 2010 IEEE
for a composite electrode lithium-ion battery,” IEEE Transactions on Autotestcon Conf., Orlando, FL, USA, Sep. 2010, pp. 1–8.
Control Systems Technology, vol. 24, no. 2, pp. 384–399, Mar. 2017. [31] A. Farmann, W. Waag, A. Marongiu, and D. U. Sauer, “Critical review
[10] W. Sung, D. S. Hwang, J. Nam, J. Choi, and J. Lee, “Robust and efficient of on-board capacity estimation techniques for lithium-ion batteries in
capacity estimation using data-driven metamodel applicable to battery electric and hybrid electric vehicles,” Journal of Power Sources, vol. 281,
management system of electric vehicles,” Journal of The pp. 114–130, May. 2015.
Electrochemical Society, vol. 163, no. 6, pp. A981–A991, Mar. 2016. [32] M. Gholizadeh and F. R. Salmasi, “Estimation of state of charge,
[11] Z. Ma, J. Xue, A. Leijon, Z. Tan, Z. Yang, and J. Guo, "Decorrelation of unknown nonlinearities, and state of health of a lithium-ion battery based
Neutral Vector Variables: Theory and Applications," IEEE Transactions on a comprehensive unobservable model,” IEEE Transactions on
on Neural Networks and Learning Systems, vol. 29, no. 1, pp. 129–143, Industrial Electronics, vol. 61, no. 3, pp. 1335–1344, Mar. 2014.
Jan. 2018.
[12] Z. Ma, Y. Lai, W. B. Kleijn, L. Wang, and J. Guo, “Variational Bayesian
Learning for Dirichlet Process Mixture of Inverted Dirichlet
Distributions in Non-Gaussian Image Feature Modeling,” IEEE

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2018.2864688, IEEE
Transactions on Vehicular Technology
11

Rui Xiong (S’12–M’14–SM’16) received the electric vehicles, design, and control theory of the hybrid power trains.
[Link]. degree in vehicle engineering and the Ph.D. Dr. He was recipients of the second prize of Chinese National Science
degree in mechanical engineering from Beijing and Technology Award for the work on the development of new-energy
Institute of Technology, Beijing, China, in 2010 electric bus powertrains in 2015, the first prize of Henan Science and
and 2014, respectively. He conducted scientific Technology Award for the work on the development of hybrid bus
research as a joint Ph.D. student in the DOE powertrain in 2013, the first prize of Henan Science and Technology
GATE Center for Electric Drive Transportation at Award for the work on the development of battery electric bus powertrain
the University of Michigan, Dearborn, MI, USA, in 2014 and the second prize of National Defense Technology
between 2012 and 2014. Innovation Award for the work on the development of hybrid powertrains
Since 2014, he has been an Associate Professor in off road vehicles in 2016. He received Best Paper Awards from the
in the Department of Vehicle Engineering, School of Mechanical journal Energies and International conference. He is serving as the
Engineering, Beijing Institute of Technology, Beijing, China. Since 2017, Editorial Board of the Energies. He was the conference chair of the 2017
he has been an Adjunct Professor in the Faculty of Science, Engineering International conference on Energy, Ecology and Environment held in
and Technology, Swinburne University of Technology, Melbourne, Vic., Stockholm, Sweden.
Australia. He has conducted extensive research and authored more
than 100 peer-reviewed articles. He holds eight patents. His research
interests mainly include electrical/hybrid vehicles, energy storage, and Simin Peng(M’16-SM’18) received the B.S
battery management system. degree in automation from Xiangtan University,
Dr. Xiong received the Excellent Doctoral Dissertation from Beijing Xiangtan, China, in 2003, and the Ph.D degree
Institute of Technology in 2014, the first prize of Chinese Automobile in electrical engineering from Shanghai Jiao
Industry Science and Technology Progress Award in October 2015 and Tong University, Shanghai, China, in 2013.
the second prize of National Defense Technology Invention Award in From 2016 to 2017, he was a Visiting Scholar
December 2016. He received the 2018 Best Vehicular Electronics Paper with the University of Maryland, College Park,
Award recognizing the best paper relating to Vehicular Electronics MD, USA. He is an associate Professor with
published in the IEEE Transactions on Vehicular Technology during the Yancheng Institute of Technology, Yancheng,
past five years. He is serving as the Associate Editors of IEEE Access China. His research interests include modeling
and SAE International Journal of Alternative Powertrains, Editorial and control of wind power, microgrid and battery
Board of the Applied Energy, Energies, Sustainability and Batteries. He energy storage system.
was the conference chair of the 2017 International Symposium on
Electric Vehicles held in Stockholm (ISEV2017), Sweden.
Michael G. Pecht (S’78–M’83–SM’90–F’92)
received the B.S. degree in physics in 1976, the
Yongzhi Zhang (S’15) received the [Link]
M.S. degree in electrical engineering in 1978,
in vehicle engineering from Chongqing University,
and the M.S. and Ph.D. degrees in engineering
Chongqing, China, in 2013. He is currently
mechanics from the University of Wisconsin-
working toward the Ph.D. degree in mechanical
Madison, Madison, WI, USA, in 1979 and 1982
engineering with the National Engineering
respectively.
Laboratory for Electric Vehicles, Beijing institute
He is the Founder and Director of the Center for
of Technology, Beijing, China.
Advanced Life Cycle Engineering, University of
He was a Research Scholar with the Center for
Maryland, College Park, MD, USA, which is
Advanced Life Cycle Engineering, University of
funded by over 150 of the world’s leading
Maryland, College Park, MD, USA, from Sept.
electronics companies at more than US$6M/year. He is also a Chair
2016 to Mar. 2018. He is currently a Visiting
Professor of mechanical engineering and a Professor of applied
Scholar in the Future Energy Center, Malardalen University, Sweden.
mathematics, statistics, and scientific computation with the University of
His research interests include prognostics and health management of
Maryland. He has written more than 20 books, 400 technical articles,
lithium-ion batteries.
and has eight patents.
Mr. Zhang received the Best Paper Awards in the International
Dr. Pecht is a Professional Engineer, a Fellow of the American Society
Symposium on Electric Vehicles, Sweden, 2017.
of Mechanical Engineers, a Fellow of the Society of Automotive
Engineers, and a Fellow of the International Microelectronics Assembly
and Packaging Society. He is the Editor-in-Chief of the IEEE ACCESS,
Ju Wang received the M.E. degree in automotive
and served as the Chief Editor of the IEEE TRANSACTIONS ON
engineering from Beijing Institute of Technology,
RELIABILITY for nine years and the Chief Editor for the Microelectronics
Beijing, China, in 2017. He is currently working
Reliability for 16 years. He has also served on three U.S. National
toward the Ph.D degree in vehicle engineering with
Academy of Science studies, two U.S. Congressional investigations in
the National Engineering Laboratory for Electric
automotive safety, and as an expert to the U.S. Food and Drug
Vehicles, Beijing Institute of Technology and
Administration.
Collaborative Innovation Center of Electric Vehicles
in Beijing, Beijing, China.
His research mainly focuses on, power batteries,
battery management system, battery multi-state
estimation.

Hongwen He (M’03–SM’12) received the Ms.E.

degree from the Jilin University of Technology,
Changchun, China, in 2000 and the Ph.D.
degree from the Beijing Institute of Technology,
Beijing, China, in 2003, both in vehicle
engineering.
He is currently a Professor with the National
Engineering Laboratory for Electric Vehicles,
School of Mechanical Engineering, Beijing
Institute of Technology. He has published more
than 100 papers and holds 20 patents. His
research interests include power battery modeling and simulation on

0018-9545 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.