1

Health Prognosis for Electric Vehicle Battery

Packs: A Data-Driven Approach
Xiaosong Hu, Senior Member, IEEE, Yunhong Che, Xianke Lin, Member, IEEE, and Zhongwei Deng

Abstract—Accurate, reliable, and robust prognosis of the state and it is therefore used in many applications, especially in EVs
of health (SOH) and remaining useful life (RUL) plays a [7, 8]. In the multi-stage constant current charging process, it is
significant role in battery pack management for electric vehicles. of great significance to develop effective extraction methods to
However, there still exist challenges in computational cost, storage
extract the health indicator (HI) from partial charging curves,
requirement, health indicators extraction, and algorithm design.
This paper proposes a novel dual Gaussian process regression which will help reduce data storage requirements and
model for the SOH and RUL prognosis of battery packs. The computational cost.
multi-stage constant current charging method is used for aging Battery SOH can be defined as the ratio of actual capacity to
tests. Health indicators are extracted from partial charging curves, initial capacity or the increase of internal resistance [9]. RUL
in which capacity loss, resistance increase, and inconsistency represents the remaining cycles or time before the end of life
variation are examined. A dual Gaussian process regression
(EOL) is reached [10]. The EOL of a battery is generally
model is designed to predict SOH over the entire cycle life and
RUL near the end of life. Experimental results show that the defined as the time when the battery SOH reaches 70% or 80%,
predictions of SOH and RUL are accurate, reliable, and robust. according to the Center for Advanced Life Cycle Engineering
The maximum absolute errors and root mean square errors of at the University of Maryland [11] and the NASA’s Prognostics
SOH predictions are less than 1.3% and 0.5%, respectively, and Center of Excellence [12]. Generally, the methods for battery
the maximum absolute errors and root mean square errors of health prognosis can be divided into three categories:
RUL predictions are 2 cycles and 1 cycle, respectively. The
electrochemical model-based methods, empirical model-based
computation time for the entire training and testing process is less
than 5 seconds. This study shows the prospect of health prognosis methods, and data-driven methods [10, 13, 14]. The
using multiple health indicators in automotive applications. electrochemical model-based methods describe the internal
electrochemical reactions based on aging-related partial
Index Terms—Battery pack, Multiple health indicators, State of differential equations [15]. The pseudo-two-dimensional model
health, Gaussian process regression, Remaining useful life. and the single-particle model are the two commonly used
electrochemical models [16, 17]. However, they do not
I. INTRODUCTION consider the degradation mechanisms in the battery. Many

L ithium-ion batteries have been widely used in many

applications, such as electric vehicles (EVs), plug-in
hybrid electric vehicles (PHEVs), mobile phones, micro-grids,
degradation models have been developed [18-20], which
incorporate complex side reactions into
pseudo-two-dimensional model and the single-particle model
the

and satellites, due to their high energy and power density, low [21]. The degradation model can capture the internal
self-discharge, and long lifespan [1, 2]. In actual applications, electrochemical changes during aging. However, they are very
batteries are usually connected in series and parallel to meet the complicated, the parameters are extremely difficult to calibrate,
energy and power requirements[3, 4]. However, the battery will and the computational cost is high. All these issues hinder the
degrade over time, resulting in a reduced capacity, increased application of this method in BMS.
internal resistance, and cell inconsistencies [5, 6]. Therefore, it Empirical model-based methods fit empirical/semi-empirical
is critical for the battery management system (BMS) to predict degradation models to the calendar or cycling aging data [22,
the state of health (SOH) and remaining useful life (RUL) 23]. Based on the fitted model, SOH can be obtained, and RUL
accurately, reliably, and robustly, which helps prevent overuse, can be predicted by extrapolating the fitted curve to EOL. Many
ensure safety, and support second-life use. The multi-stage optimization algorithms, such as extended Kalman filter [24],
constant current charging method has the potential to speed up unscented Kalman filter [25], least-square [26], particle filter
the charging process and improve energy transfer efficiency, [27, 28], and genetic algorithm [29], can be used to track the
capacity fade or resistance increase. Zhang et al. [27] fitted an
This work was in part supported by National Natural Science Foundation of exponential capacity model by the particle filter to predict the
China (Grant No. 51875054 and No. U1864212) and Chongqing Natural RUL from different starting points. Yang et al. [30] developed
Science Foundation for Distinguished Young Scholars (Grant No.
cstc2019jcyjjq0010), Chongqing Science and Technology Bureau, China.
a semi-empirical model based on the coulombic efficiency and
(Corresponding author: X. Hu) particle filter, and verified its performance under dynamic
X. Hu, Y. Che, and Z. Deng are with the Department of Automotive discharging profiles. Due to the simplicity of data fitting, it is
Engineering, Chongqing University, Chongqing 400044, China. (e-mail:
easy to implement these methods in BMS. However, data
xiaosonghu@ieee.org; cyh@cqu.edu.cn; dengzhongw@cqu.edu.cn)
X. Lin is with the Department of Automotive, Mechanical and fitting requires a large amount of aging data, and these data
Manufacturing Engineering, University of Ontario Institute of Technology, require a lot of testing and time. Another drawback is that the
Oshawa, ON L1G 0C5, Canada. (e-mail: xianke.lin@uoit.ca)
 2

empirical model of one type of batteries cannot be directly 1) A novel dual GPR model is designed for battery pack
generalized to other types of batteries. health prognosis. A small training data set is used for the
Data-driven methods have been developed rapidly in recent training of the SOH GPR model, while good prediction results
years, due to their strong adaptability to nonlinear systems and are achieved. A SOH threshold is used to determine whether the
model-free feature. The correlation between battery battery pack is close to EOL. When the SOH threshold is
degradation and extracted HIs largely determines the exceeded, the training of the RUL GPR model is stopped, and
effectiveness of this approach. Therefore, effective HI the RUL prediction is activated for the battery pack.
extraction methods are very important. Battery degradation will 2) The multi-stage charging method is used in the battery
lead to changes in the charging time [31, 32], pack aging tests, and HIs are extracted from the partial charging
charging/discharging voltage curves, such as the slope of curves. The correlation is evaluated by Pearson correlation
voltage curve [31], inflection point position [33, 34], and analysis. Low sampling frequency and simple calculations
voltage curve drop [35], as well as changes of the temperature reduce the requirements for storage and computing power.
curves [33, 36]. Recently, increment capacity analysis (ICA) [5, 3) Capacity loss, internal resistance increase, and
10, 37] and differential voltage (DV) analysis [36, 38] are inconsistency between battery cells during the aging process
widely used to extracted HIs due to the reflection of the phase are examined, and the extracted HIs can clearly reflect the
transitions and phase equilibria. During the degradation process, changes in the above features of the battery pack, which is very
the magnitude and position of the peaks, and the peak area of important for the health prognosis of the battery pack.
ICA/DV curves will change, which reflects the health condition 4) A battery pack consists of four cells in series is used for
of the battery[39]. In [40], the ICA is directly used as the index aging tests, and the test data is used to evaluate the accuracy,
for real-world electric buses to assess the aging conditions. reliability, robustness, and computational cost of the proposed
However, ICA/DV curves are generally noisy, and therefore method.
they are smoothed by filters, such as moving averages, The remainder of this paper is organized as follows: The
numerical derivative methods, and Gaussian filters [37, 41]. experimental setup and method for battery pack aging tests are
After HI extraction, the machine learning algorithms are used introduced in Section Ⅱ. The HI extraction method based on
to build the regression model, such as support vector machine partial charging curves and the correlation analysis are
(SVM) [5], relevance vector machine (RVM) [33], artificial presented. Next, the dual GPR prediction is introduced in
neural network (ANN) [34, 42], and Gaussian progress Section Ⅲ. After that, Section Ⅳ discusses the results of the
regression (GPR) [10, 31]. Among these, the GPR has the battery pack health prognosis based on the dual GPR model.
excellent flexibility and probabilistic predictions that attract Finally, the main conclusions are summarized in Section Ⅴ.
much attention. Wang et al. [43] extracted HIs from ICA peak
and its position, Richardson et al. [44] extracted HIs from II. EXPERIMENTS
charging voltage curve over short periods, then the GPR was The experimental platform designed for the battery pack
adopted to predict the health status. The results are accurate and aging test is shown in Fig.1. It includes a Digatron battery tester
robust, which clearly reflects the advantages and adaptation of together with a data logger, a thermal chamber, a battery pack,
the GPR. The data-driven methods can provide accurate and and a computer. The Digatron battery tester is used to charge or
reliable predictions. However, most of those methods are discharge the battery pack with predefined profiles, and the
time-consuming and computationally costly. Moreover, HIs current range and voltage range are 0-300 A and 0-600V,
extracted from one type of batteries are not suitable for other respectively. Current and voltage measurements have ±20 mA
types. and ±40 mV errors, respectively. The data logger collects the
From the above introduction, it can be found that although testing data, such as current, voltage, temperature, and charging
many methods have been proposed for battery health prognosis, capacity. The voltage and temperature measurement errors are
there are still many research gaps. Specifically, besides CC-CV less than ±5 mV and ±1 ℃, respectively. The thermal chamber
charging method, the multi-stage constant current charging
provides the required constant ambient temperature. In this
method is also used in many applications, especially in EVs.
paper, the sampling frequency and the testing temperature are
The HI extraction method for this charging method is not
set to 1/60 Hz and 25 ℃, respectively. The battery pack consists
developed yet. HI extraction and data-driven prediction
methods usually require a large amount of computation and of four battery cells in series, and the detailed specifications are
storage, which hinders their applications in the BMS. In listed in Table 1. The computer is connected to the tester and
addition, existing studies are mainly focused on the health stores the experimental data. The computer is configured with
prognosis of battery cells, but few studies provide solutions for Windows10, 16 GB RAM, and Intel (R) Core (TM) i7-9700K
battery packs. Besides capacity loss and internal resistance CPU.
increase, the change of internal inconsistencies will also have a Studies have shown that the multi-stage constant current
significant impact on the aging process of the battery pack, charging method can extend the cycle life, shorten the charging
which has been ignored in published papers. time, and improve the energy transfer efficiency [7]. In this
In order to fill the above research gaps, this paper proposes a paper, the multi-stage constant current charging method is used
novel dual GPR model for battery pack SOH and RUL for battery pack aging tests. Fig. 2 shows the current and
prognosis. The main contributions are summarized as follows: voltage curves for a complete test cycle during aging tests. A
 3

five-stage constant current charging method (1.25 C charging voltage, in Section Ⅲ A. After that, the detailed description of
until any cell reaches 3.78 V, next, 0.85 C charging until any the GPR algorithm is introduced in Section Ⅲ B. Finally, a dual
cell reaches 4.08 V, next, 0.5 C charging until any cell reaches GPR model is designed for health prognosis of the battery pack
4.125 V, next, 0.2 C charging until any cell reaches 4.135 V, in Section Ⅲ C.
finally, 0.1 C charging until any cell reaches 4.2 V) is used to
A. Health Indicator Extraction
charge the battery pack until any battery cell reaches its upper
cut-off voltage. This charging policy is designed according to In this subsection, the extracted HIs will be introduced in
the real application in EVs. 1 C rate is used to discharge the detail. The correlation between HIs and SOH/RUL of the
battery pack until any battery cell reaches its lower cut-off battery pack is analyzed based on the Pearson correlation.
voltage. There is a 30-minute rest between charging and
discharging. 1) Health Indicator Extraction Based on the Partial Charging
The aging cycle is repeated until the battery pack reaches its Curve
EOL. In this study, the capacity of the battery pack is defined as During the battery pack aging process, the internal resistance
and capacity will gradually change [10]. In addition, parameter
the discharge capacity at 1 C. 20% capacity loss of the battery
inconsistencies, such as internal resistance inconsistencies,
pack indicates EOL.
capacity inconsistencies, and SOC inconsistencies, will also
change gradually, which have a significant impact on the
battery pack aging process [45]. Therefore, it is an important
task to extract HIs that can reflect the inconsistencies and their
changes over cycling. This paper, for the first time, proposes a
novel HI extraction method based on the partial multi-stage
charging curves, which can capture not only the capacity loss
and resistance increase, but also the inconsistency change.
Fig.3(a) and Fig.3(b) shows the charging voltage curves at
different cycles. The damage of active materials and the growth
of SEI film will cause the capacity decrement and resistance
Fig. 1. Configuration of the experimental platform. increment. Changes in the internal chemical/physical properties
of the battery pack will result in the change of charging curves
during the aging process, especially the first three charging
stages. Therefore, these three charging stages are used to
extract the HIs, which can reduce the memory storage by using
the partial charging curves. Specifically, the twelve different
HIs are shown in Fig.3(c) and are introduced in detail below.
1) Charging capacities of the first three stages (C1~C3): The
battery pack capacity will gradually decrease during the aging
process due to the decrement of active materials, which will
correspondingly cause a change in the charging capacity in the
first three stages. As shown in Fig.4(a) (C1~C3 represent the
charging capacities of the first, second, and third stages,
respectively), the charging capacities of the first three stages
show satisfactory monotonicity over cycling, which reflects the
Fig.2. The current and voltage curves over a complete test cycle. aging state. It shows that the first two charging capacities
decrease as the cycle number increases, while the capacity of
Table 1. The specifications of the battery pack. the third charging stage increases over cycling. This
Parameters Value phenomenon may be the result of a joint effect of reduced
Configuration 1P4S capacity and increased internal resistance. Because the
Battery cell chemistry Li(Ni0.8Co0.1Mn0.1)O2 cathode charging current in the first two stages is large, increased
and graphite anode resistance will cause the voltage to rise faster, and therefore, the
Battery cell rated voltage (V) 3.61 charging capacity is reduced in the first two stages. In the third
Battery cell voltage range (V) 2.8 ~ 4.2 stage, the charging current is further reduced, and the remaining
Working temperature range (℃) -5 ~ 60 capacity will be charged. These three HIs can also be reflected
Battery cell rated capacity (Ah) 177 by the charging time (see Fig.3). As the battery pack ages, the
Battery cell energy (kWh) 0.638 location of the peak voltage moves to the left, which means that
less charging time is needed.
III. METHODOLOGY 2) Slopes of the charging voltage curves of the first three
stages (K1~K3): The internal resistance increases during the
In this section, we first discuss the HI extraction method
aging process due to the effects of SIE films grow and increase,
based on the partial charge curve, then perform correlation
which makes the voltage increase faster. Therefore, the voltage
analysis to evaluate multiple variables at different stages, such slope would increase as the results of resistance increment, as
as charge capacity, voltage curve slope, voltage drop, and peak
 4

shown in Fig.3(a). It shows that the curve tends to lean to the charge current changes at the time when the voltage of any
left as the cycle number increases. This phenomenon can be battery cell reaches the set points. Therefore, the pack voltage
reflected by the slope of the charging curve. Thus, the slopes of calculated as the sum of the battery cells will change according
the first three charging stages are extracted, and the results are to the inconsistency's variation. When the battery pack is fresh,
shown in Fig.4(b) (K1~K3 represent slopes of the voltage the inconsistency is low, and the pack voltage is near the
curves in the first, second, and third stages, respectively). K1 multiple of the battery cell which reached the setpoints first.
changes monotonically with the cycle number, while K2 and However, as the inconsistency enlarges, the voltages of other
K3 fluctuate over cycling. In this paper, the slope is defined as battery cells would deviate smaller than the cell which reached
the slope of the five points before the peak voltage, as shown the setpoints first during the aging process. Therefore, the pack
below, voltage would also reduce according to the analysis above. As
U − U i −6 shown in Fig.3, as the cycle number increases, the three peak
K = i −1 (1) voltages decrease. The HIs are extracted as shown in Fig.4(d)
5
(U1~U3 represent voltage values of the first, second, and third
where K is the slope of the voltage curve, U i −1 and U i − 6 are the peaks, respectively). The voltage shows a gradual downward
voltage values at point 1 and point 6 before the peak voltage, trend in the first 500 cycles, and an accelerated trend in the last
respectively. 200 cycles, which will accelerate the battery pack aging
3) Voltage drop at the current switching point (dU1~dU3): process.
The resistance can be calculated according to the ohm law. It is worth noting that HIs are extracted from the partial
Generally, a hybrid pulse power characterization test is used to charging voltage curve, and the sampling interval is 1 minute.
test the resistance, in which a pulse current is suddenly applied Therefore, very little memory storage is required. In addition,
and the voltage change is used to calculate the resistance. the switching point can be easily monitored with no need for
Similarly, in our charging method, there is some sudden current complex curve transformations and HI extractions, such as ICA
switching as well. At the switching point of the multi-stage and DV analysis. Therefore, very little computational cost is
constant current charging, the charging current will change required. This method has great potential for real EVs and
greatly, resulting in a large change in the voltage (see Fig.3). PHEVs applications. Moreover, the capacity loss, resistance
Therefore, the internal resistance can be extracted as HIs based increase, and inconsistency change can be reflected by the
on the current switching dynamics, as shown in Fig.4(c) extracted HIs. The inconsistency in the battery pack is
(dU1~dU3 represent voltage drops at the first, second, and third considered for the first time in this paper for HIs extraction to
current switching points, respectively). The voltage drop can be better track the aging process.
calculated as follows,
dU = U i − U i +1 (2) 2) Correlation Analysis
The correlation between the input and output of a GPR
where, dU is the voltage drop, U i and U i +1 are the peak
model has a significant impact on the prediction performance.
voltage and the following voltage, respectively. It can be seen The Pearson correlation coefficient has many advantages, such
that the extracted HIs increase over cycling, which reflects the as being unaffected by local mutations and easy to calculate[46].
increase in the internal resistance of the battery pack. Thus, Pearson correlation coefficients are used to assess the
4) Peak voltage at the current switching point (U1~U3): The correlation between HIs and SOH / RUL. The formula is
inconsistencies exist among battery cells and will change defined as follows [47],
greatly during the aging process, which complicates the battery
 i =1 ( xi − x )(yi − y )
n

pack aging mechanism. The battery cells have a good rxy = (3)
 i =1 i  i =1 i
n n
consistency when the battery pack is fresh. However, SOC ( x − x ) 2
(y − y ) 2

inconsistency, internal resistance inconsistency, and capacity

1 n 1 n
inconsistency between battery cells will increase during aging,
which will cause the voltage inconsistency between battery
xi =  xi , yi = n  i =1 yi
n i =1
(4)
cells. According to the charge policy described in Section Ⅱ, the

(a) (b) (c)

Fig.3. Multi-stage charging voltage curve of the battery pack: (a) Enlarged view of the first voltage peak; (b) Enlarged view of the second and third voltage peaks,
(c) health indicator extraction.
 5

(a). C1~C3 over cycling.

(b). K1~K3 over cycling.

(c). dU1~dU3 over cycling.

(d). U1~U3 over cycling.

Fig.4. Health indicators over cycling.

where n denotes the sample size. x is the input HIs, and y is are presented in Table 2. It shows that the extracted HIs are
th
the predicted SOH/RUL values at the i sample time. xi and closely related to SOH and RUL. Most correlation coefficients
between HIs and battery pack SOH are larger than 0.8, except
yi are the mean values. for the slopes of the charging voltage curves of the second and
The correlation coefficients between the extracted HIs and third charging stages. More than half of them are greater than
the battery pack SOH/RUL are shown in Fig.5, and the values 0.9, which makes it possible to predict SOH with a small set of
training data. The extracted HIs have a good correlation with
 6

RUL, which can be seen from the results that more than half of and is easy for implementation. The GPR method can combine
them are greater than 0.8, and some of them are even greater prior distribution and noise, obtain prior knowledge through
than 0.9. The maximum correlation coefficients for both SOH Bayesian theory, and output predicted mean, variance, and
and RUL are from charging capacities, which indicates the confidence interval [49, 50]. Thus, it is a powerful tool in many
capacity loss is the most important factor affecting battery prediction applications and is used in this paper to predict the
aging. The results between K1 and SOH/RUL are both larger SOH and RUL of the battery pack.
than 0.9, which illustrates the internal resistance plays an Generally, the noise is assumed to be additive, independent
important role in the aging process. The correlation coefficient and Gaussian, therefore the relationship between input x and
between U1/U2 and SOH is larger than 0.93, and the out y is given as follows,
correlation coefficient between U1/U2 and RUL is larger than
y = f ( x) +  ,  ~  (0,  n2 ) (5)
0.82. This result indicates that the inconsistency and its change
do affect the battery pack aging process. These satisfactory where  is the white noise with a variance of  . f ( x) is a 2
n
results of the correlation analysis provide a good basis for the latent function, which has a probability distribution as follows:
dual GPR design and support for achieving satisfactory
f ( x) ~ GP (m( x), k ( x, x ')) (6)
prediction results.
where m( x) is the mean function, and k ( x, x ') is the
covariance function.
m( x) = E[ f ( x)] (7)
k ( x, x ') = E[( f ( x) − m( x))( f ( x ') − m( x '))T ] (8)
For any finite set of input points, x = x1 , x2 , ... , xND , the
probability distribution p ( f ( x1 ), f ( x2 ), ... , f ( xND )) is joint
Gaussian distribution [51]. Generally, the mean function is set
to zero. The squared exponential covariance function is usually
used as the kernel function for battery pack health prediction.
The function is expressed as follows,
1
k f =  2f exp(− ( x − x ')l −1 ( x − x ')) (9)
2
where the covariance  2f represents the output amplitude the
Fig.5. Correlation coefficients between health indicators and SOH / RUL. diagonal matrix, l is the characteristic length scale. Thus, the
prior distribution of the observations of Eq (6) is
Table 2. Correlation coefficients. y ~ N (0, K f ( x, x ) +  n2 I n ) (10)
Health Indicators SOH RUL
C1 0.99 0.91 where y is the observation series [ y1 , y2 ... yn ] , which
C2 0.92 0.98
C3 0.96 0.85
denotes the SOH and RUL in this paper. I n is an
K1 0.98 0.91 n-dimensional unit matrix. Here the hyper-parameters
K2 0.68 0.64  = [ f ,  n , l ] in the covariance matrix can be optimized by
K3 0.68 0.67
dU1 0.83 0.77 maximizing the log-likelihood function expressed by [49, 52]:
dU2 0.80 0.63 1
dU3 0.91 0.86 L = log p ( y x , ) = − log(det( K f ( x , x ) +  n2 I n ))
U1 0.94 0.83 2 (11)
1 T 1
U2
U3
0.93
0.85
0.82
0.73
− y [ K f ( x , x ) +  n2 I ]−1 y − log 2
2 2
B. GPR Algorithm One of the most widely used methods for finding the parametric
optimal solution is the conjugate gradient method. The basic
Many machine learning algorithms have been used in the
idea is to find the maximum value of the objective function by
SOH estimation, and the popular methods mainly include ANN,
obtaining the derivative of the log-likelihood function [10, 31].
SVM, RVM, and GPR. Each of them has its own advantages
 
  log p( y x, Θ) =
and disadvantages, which has been summarized in Ref. [48].
One common shortage for the first three is that it is easy to  i
cause over-fitting problems. The uncertainty management of 1   T  ( K f ( x, x ) +  n2 I n )  

 tr   − ( K f ( x, x ) +  n I n )
−1
  (12)
2
ANN and SVM is poor, while the RVM and GPR provide a
 2 
   i 

probabilistic prediction. Compared to large data that is needed  = [ K f ( x, x ) +  n2 I n ]−1 y
for RVM model training, the GPR has more potentials of the 
accurate estimation with fewer data for model training. 
Moreover, it is easy for RVM to fall into local optimization Then the prediction task can be carried out by the posterior
while GPR is much better. In addition, the computational distribution based on the GPR model. Thus, the joint prior
efficiency of GPR is generally high, and has good flexibility
 7

distribution of the observations y along with the predicted predict the remaining cycle numbers. In this case, the training
* set for RUL GPR model training includes the data until the
value y * at a test point x is expressed by SOH threshold is triggered. With this large amount of training
 y  ~ N  0,  K f ( x, x ) +  n I n K f ( x, x )   data, the RUL GPR model can provide accurate and reliable
2 *


  K f ( x, x* )T 
K f ( x* , x* )  
(13)
 y 
* predictions near EOL. Based on the dual GPR strategy, the
  battery pack health can be predicted more efficiently and
Considering the joint Gaussian prior distribution of y , reliably. At last, the error calculation module checks four error
the posterior distribution of p( y* x, y, x * ) is indexes to assess the accuracy and reliability of the prediction
results.
analytically derived as follows, The error analysis is conducted in this paper to assess the
p( y* x , y, x * ) = N ( y* y * , cov( y* )) (14) prediction performance of SOH and RUL. The 95% confidence
*
where the predicted mean value y and the predicted interval (CI) is used to represent the uncertainty of the GPR
* model, and the equation is as follows,
covariance value cov( y ) are expressed as follows
y * = K f ( x , x * )T [ K f ( x , x ) +  n2 I n )]−1 y (15) 95%CI = yˆ  1.96* 2 ( yˆ ) (17)
cov( y ) = K f ( x , x ) −
* * *
where, 95%CI is the confidence interval, ŷ and  ( yˆ ) are 2
(16)
K f ( x, x * )T [ K f ( x, x ) +  n2 I n )]−1 K f ( x, x * ) the predicted value and variance, respectively. Three variables
Until now, the GPR prediction process is completely called maximum absolute error (MaAE), mean absolute error
introduced. The prediction value and the uncertainty are (MeAE), root mean square error (RMSE) are used to evaluate
represented by the mean y * and covariance cov( y * ) , the prediction performance. Specifically, the three variables are
defined as follows,
respectively. The input matrix is a 12-dimensional HIs, and the
MaAE = max( yˆ i − yi ) (18)
output is the predicted SOH or RUL.
1

N
C. Dual GPR Design MeAE = yˆ − yi
i =1 i
(19)
N
Predicting battery pack SOH and RUL is critical to the BMS.
1
In real applications, these two variables play different roles.
 ( yˆ − yi )
N 2
RMSE = i =1 i
(20)
SOH provides the health condition while RUL provides the N
remaining cycle life. Therefore, SOH is important over the where N is the size of the testing set, max reparents the
entire cycle life, and RUL becomes important near the EOL. It maximum value in the matrix.
is difficult to predict these two variables, and a well-designed
system needs to be built to predict the battery pack health over
the entire cycle life.
In this paper, a novel dual GPR system is designed to predict
the battery pack health conditions. The flowchart of the system
is shown in Fig.6. The SOH GPR model is designed to predict
the battery pack SOH over the entire cycle life, and the RUL
GPR model is activated to predict the number of remaining
battery cycles near the EOL. Specifically, the measured signals,
such as current, battery cell/pack voltage, and charging
capacity (or charging time), are fed into the health indicator
extraction module, and the HIs are extracted from the partial
stage charging voltage curve, as introduced in Section Ⅲ A.
These HIs are a 12-dimensional matrix with charging capacities
of three stages (C1~C3), three voltage curve slopes (K1~K3),
three voltage drops at switching points (dU1~dU3), and three
peak voltages (U1~U3). Then, HIs are fed into the battery pack
SOH and RUL prediction modules to train the SOH GPR and
RUL GPR models to make predictions. The data is divided into
two sets, one for training and the other for prediction. The
detailed GPR algorithm is introduced in Section Ⅲ B. Before
prediction, the training data set is used to train the regression
models and then the HIs are used to predict the health
conditions. The training set for SOH GPR model training in this
paper is set to 10%, 30%, 50% and 70% of the experimental Fig.6. Flowchart of the dual GPR based battery pack health prognosis.
data. The predicted SOH is compared with a threshold to
determine if the EOL is approaching. If the SOH comparison IV. RESULTS AND DISCUSSIONS
determines that the EOL is approaching, the RUL model In this section, we first evaluate the battery pack SOH
training is stopped, and the RUL prediction is activated to prediction results, where 10% of the aging data set is used for
 8

GPR model training. The model robustness using different application prospects for battery pack health prediction of EVs
training data sets is evaluated in Section Ⅳ A. Then, the battery and PHEVs.
pack RUL prediction results and computational cost are In order to evaluate the robustness of the proposed SOH
assessed in Section Ⅳ B. prediction method for battery pack based on GPR with HIs
extracted from the partial charging curve, different sizes of
A. Battery pack SOH prediction results
training datasets are used to demonstrate the prediction
The curves of SOH prediction results and prediction errors performance. The predicted results and the absolute error
based on GPR is given in Fig.7(a), where 10% of the curves are shown in Fig.7(b), Fig.7(c), Fig.7(d), respectively,
experiment data is used for model training. The solid lines and the training set accounts for 30%, 50%, and 70% of the
represent the estimated values based on the GPR model, while experimental data, respectively. The MaAE, MeAE, RMSE of
the dotted lines represent the truth values. The shadow zones predicted results are shown in Table 3. It can be concluded that
represent the 95% confidence intervals (calculated by Eq (17)). with the increase in training data, the predicted results become
In general, the narrower this interval is, the more convergent more and more accurate and reliable. MaAE, MeAE, and
and reliable the predicted results are. The MaAE, MeAE, and RMSE tend to decrease and the confidence interval tends to
RMSE of the predictions are given in Table 3. It can be seen converge gradually. When 70% of the data was used for
from the figure that the predicted SOH value can follow the true training, the MaAE and MeAE RMSE of predicted results are
value smoothly and satisfactorily. The results show that the less than 0.8%, 0.1%, and 0.12%, respectively, and the range of
MaAE, MeAE, and RMSE are less than 1.3%, 0.4%, and 0.5%, confidence interval is less than 0.3%. This shows the GPR
respectively. The confidence interval is narrow, with a range based model can have a more accurate and reliable prediction if
less than 2.5%. Therefore, it can be concluded that the GPR enough training set is used. Thus, it can be concluded that the
based battery pack SOH prediction is accurate and reliable. proposed prediction method is able to provide an accurate,
Moreover, even the training set is not very large, satisfactory reliable and robust prediction of battery pack SOH.
prediction results can be achieved. This has important

(a) (b) (c) (d)

Fig.7. Battery pack SOH predictions and errors using different training sets: (a) 10% of experimental data;
(b) 30% of experimental data; (c) 50% of experimental data; (d) 70% of experimental data.

well converge to the real value. The MaAE is less than 2 cycles,
B. Battery pack RUL prediction results
the MeAE is less than 1 cycle, and the RMSE is also less than 1
The dual GPR model for battery pack health prediction cycle. Moreover, the confidence interval of the predicted
proposed in this paper can provide predictions of the remaining results is small, which indicates that the predicted results of the
cycle numbers when the predicted SOH value reaches a battery pack RUL are accurate and reliable. The data size of the
threshold. In order to verify the prediction performance of the training set for the GPR-based RUL model is determined by the
proposed method, the RUL prediction of the battery pack is estimated SOH under different training sets, and the
activated when the SOH prediction value reaches 83% (the corresponding results of MaAE, MeAE, and RMSE are listed in
remaining cycle number is about 10%). The curves of Table 3. As you can see, larger training sets lead to more
prediction results and absolute error are shown in Fig.8, and the accurate SOH predictions. With different RUL training set
results of MaAE, MeAE, and RMSE are listed in Table3. In the sizes, the accuracy of RUL predictions also varies. However,
figure, solid line, dotted line, and shadow zones respectively the MaAE, MeAE, and RMSE are still less than 2 cycles,1
represent predicted value, real value and confidence intervals. cycle, and 1 cycle, respectively. Thus, it can be concluded that
It can be found that RUL predicted results based on GPR can
 9

the RUL prediction of the battery pack based on the dual GPR Table 3. The error results and computational cost of the dual GPR model
models proposed in this paper has good robustness. Parameters Value Value Value Value
SOH Training set 10% 30% 50% 70%
In addition, the running time of the battery pack health SOH MeAE (%) 0.363 0.318 0.127 0.087
prediction system based on dual GPR is evaluated. The results SOH MaAE (%) 1.244 1.149 0.805 0.751
are listed in Table 3 (MATLAB 2018b, entire prognosis SOH RMSE (%) 0.466 0.429 0.153 0.115
process time), and it should be noted that the processing time is RUL Training cycles 626 627 649 645
RUL MeAE (cycles) 0.794 0.793 0.732 0.713
the average value after the program is run 10 times. It can be RUL MaAE (cycles) 1.947 1.947 1.717 1.757
found that the computational cost based on the method RUL RMSE (cycles) 0.962 0.961 0.871 0.867
proposed in this paper is low, because the sampling frequency Time (s) 3.97 4.07 4.29 4.79
of the data set is very low, and the HI extraction method is
efficient. It can be seen from the table that as the amount of data V. CONCLUSION
in the training set increases, the accuracy increases and the time It is very important for BMS to predict battery pack SOH and
also increases, but it is still less than 5 s. This shows that the RUL accurately, reliably, and robustly. Many factors affect the
battery pack health prediction based on this method can aging process of the battery pack, which poses a challenge to
accurately predict the SOH and RUL with lower calculation
health prognosis. A novel dual GPR-based battery pack health
and storage requirements. According to the research [53],
prognosis method is proposed in this paper. The main
which provided the hardware-in-loop verification of the SOH
estimation, the MUC used in that paper is MPC5644A, and the conclusions are summarized as: (1) The multi-stage constant
flash memory is 4 MB, and its random-access memory (RAM) charging method is used for aging experiments, and a low
is 192 KB. Then, we compare that the code (transformed into C sampling frequency is used, which is of great significance for
code) and HIs data are 3 KB and 22 KB, respectively. Therefore, fast charging applications and reducing storage requirements.
the method proposed in this paper has huge potential to be (2) HIs are extracted from the partial charging curve, and only a
applied in real BMS. The results show that even only 10% data small amount of computation is needed, which has a great
is used for model training, the RMSE is less than 0.5% and 1 potential to be applied in BMS. (3) changes in capacity, internal
cycle for SOH and RUL prognosis, respectively. And 10% data resistance and inconsistencies are considered in model
only requires 3 KB memory. Therefore, we suggest that 10% development, which provides a new direction for battery pack
data for model training in real applications, which is a very aging research. (4) A novel dual GPR model is developed to
small memory requirement for BMSs and also reduces the predict the SOH over the entire life cycle and RUL near the
computational cost. EOL of the battery pack, respectively. (5) The accurate, reliable,
and robust prediction results are achieved under different
training sets. The MaAE, MeAE, and RMSE are less than 1.3%,
0.4%, and 0.5% for SOH prediction, respectively, where only
10% of the experimental data is used for model training. The
MaAE, MeAE, and RMSE are 2 cycles, 1 cycle, and 1 cycle for
RUL prediction, respectively. Moreover, the required
calculation time is small (less than 5 seconds). Although this
paper provides a novel and effective data-driven method for
health prognosis, some areas for improvements are recognized
for the future work. For example, temperature changes and
inconsistencies need to be considered, only four battery cells
are tested, and different types of battery packs should be
verified, which will be illustrated in our future work.

(a) ACKNOWLEDGMENTS
This work was in part supported by the National Natural
Science Foundation of China (Grant No. 51875054 and No.
U1864212) and Chongqing Natural Science Foundation for
Distinguished Young Scholars (Grant No. cstc2019jcyjjq0010),
Chongqing Science and Technology Bureau, China.

REFERENCES
[1] L. Zhang, X. Hu, Z. Wang, F. Sun, and D. G. Dorrell, "A review of
supercapacitor modeling, estimation, and applications: A
control/management perspective," Renewable and Sustainable Energy
Reviews, vol. 81, pp. 1868-1878, January 2018.
[2] B. Nykvist and M. Nilsson, "Rapidly falling costs of battery packs for
electric vehicles," Nature climate change, vol. 5, no. 4, pp. 329-332,
March 2015.
(b)
Fig.8. Battery pack RUL (a) predictions and (b) errors.
 10

[3] L. Hu, X. Hu, Y. Che, F. Feng, X. Lin, and Z. Zhang, "Reliable state of current analysis," Applied Energy, vol. 212, pp. 1589-1600, February
charge estimation of battery packs using fuzzy adaptive federated 2018.
filtering," Applied Energy, vol. 262, March 2020. [24] J. Kim, S. Lee, and B. H. Cho, "Complementary Cooperation Algorithm
[4] W. Shi, X. Hu, C. Jin, J. Jiang, Y. Zhang, and T. Yip, "Effects of Based on DEKF Combined With Pattern Recognition for SOC/Capacity
imbalanced currents on large-format LiFePO 4 /graphite batteries Estimation and SOH Prediction," IEEE Transactions on Power
systems connected in parallel," Journal of Power Sources, vol. 313, pp. Electronics, vol. 27, no. 1, pp. 436-451, January 2012.
198-204, May 2016. [25] Q. Wang, Y. Jiang, and Y. Lu, "State of Health Estimation for
[5] C. Weng, X. Feng, J. Sun, and H. Peng, "State-of-health monitoring of Lithium-ion Battery Based on D-UKF," International Journal of
lithium-ion battery modules and packs via incremental capacity peak Hybrid Information Technology, vol. 8, no. 7, pp. 55-70, July 2015.
tracking," Applied Energy, vol. 180, pp. 360-368, October 2016. [26] G. K. Prasad and C. D. Rahn, "Model based identification of aging
[6] C. Zhang, Y. Jiang, J. Jiang, G. Cheng, W. Diao, and W. Zhang, "Study parameters in lithium ion batteries," Journal of Power Sources, vol. 232,
on battery pack consistency evolutions and equilibrium diagnosis for pp. 79-85, June 2013.
serial- connected lithium-ion batteries," Applied Energy, vol. 207, pp. [27] L. Zhang, Z. Mu, and C. Sun, "Remaining Useful Life Prediction for
510-519, 2017. Lithium-Ion Batteries Based on Exponential Model and Particle Filter,"
[7] T. Ikeyaa, N. Sawadab, J.-i. Murakamic, K. Kobayashid, M. Hattorie, N. IEEE Access, vol. 6, pp. 17729-17740, April 2018.
Murotanif, S. Ujiieg, K. Kajiyamah, H. Nasui, H. Narisokoj, Y. [28] Q. Miao, L. Xie, H. Cui, W. Liang, and M. Pecht, "Remaining useful
Tomakik, K. Adachil, Y. Mitaa, and K. Ishiharaa, "Multi-step constant life prediction of lithium-ion battery with unscented particle filter
current charging method for an electric vehicle nickel metal hydride technique," Microelectronics Reliability, vol. 53, no. 6, pp. 805-810,
battery with high energy efficiency and long cycle life," Journal of June 2013.
Power Sources, vol. 105, pp. 6-12, August 2002. [29] J. Bi, T. Zhang, H. Yu, and Y. Kang, "State-of-health estimation of
[8] A. B. Khan and W. Choi, "Optimal Charge Pattern for the lithium-ion battery packs in electric vehicles based on genetic
High-Performance Multistage Constant Current Charge Method for the resampling particle filter," Applied Energy, vol. 182, pp. 558-568,
Li-Ion Batteries," IEEE Transactions on Energy Conversion, vol. 33, no. November 2016.
3, pp. 1132-1140, February 2018. [30] F. Yang, X. Song, G. Dong, and K.-L. Tsui, "A coulombic
[9] L. Lu, X. Han, J. Li, J. Hua, and M. Ouyang, "A review on the key efficiency-based model for prognostics and health estimation of
issues for lithium-ion battery management in electric vehicles," Journal lithium-ion batteries," Energy, vol. 171, pp. 1173-1182, March 2019.
of Power Sources, vol. 226, pp. 272-288, March 2013. [31] D. Yang, X. Zhang, R. Pan, Y. Wang, and Z. Chen, "A novel Gaussian
[10] X. Li, Z. Wang, and J. Yan, "Prognostic health condition for lithium process regression model for state-of-health estimation of lithium-ion
battery using the partial incremental capacity and Gaussian process battery using charging curve," Journal of Power Sources, vol. 384, pp.
regressio," Journal of Power Sources, vol. 421, pp. 56-67, March 2019. 387-395, April 2018.
[11] Battery Data Set. CALCE [Online]. Available: [32] Z. Guo, X. Qiu, G. Hou, B. Y. Liaw, and C. Zhang, "State of health
https://web.calce.umd.edu/batteries/index. . estimation for lithium ion batteries based on charging curves," Journal
[12] Battery Data Set. NASA Ames Prognostics Data Repository [Online]. of Power Sources, vol. 249, pp. 457-462, March 2014.
Available:https://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-datareposit [33] P. Guo, Z. Cheng, and L. Yang, "A data-driven remaining capacity
ory/#battery. estimation approach for lithium-ion batteries based on charging health
[13] X. Hu, F. Feng, K. Liu, L. Zhang, J. Xie, and B. Liu, "State estimation feature extraction," Journal of Power Sources, vol. 412, pp. 442-450,
for advanced battery management: Key challenges and future trends," February 2019.
Renewable and Sustainable Energy Reviews, vol. 114, 2019. [34] J. Wu, C. Zhang, and Z. Chen, "An online method for lithium-ion
[14] M. S. H. Lipu, M. A. Hannan, A. Hussain, M. M. Hoque, P. J. Ker, M. H. battery remaining useful life estimation using importance sampling and
M. Saad, and A. Ayob, "A review of state of health and remaining neural networks," Applied Energy, vol. 173, pp. 134-140, July 2016.
useful life estimation methods for lithium-ion battery in electric [35] K. A. Severson, P. M. Attia, N. Jin, N. Perkins, B. Jiang, Z. Yang, M. H.
vehicles: Challenges and recommendations," Journal of Cleaner Chen, M. Aykol, P. K. Herring, D. Fraggedakis, M. Z. Bazant, S. J.
Production, vol. 205, pp. 115-133, December 2018. Harris, W. C. Chueh, and R. D. Braatz, "Data-driven prediction of
[15] C. Zou, L. Zhang, X. Hu, Z. Wang, T. Wik, and M. Pecht, "A review of battery cycle life before capacity degradation," Nature Energy, vol. 4,
fractional-order techniques applied to lithium-ion batteries, lead-acid no. 5, pp. 383-391, March 2019.
batteries, and supercapacitors," Journal of Power Sources, vol. 390, pp. [36] Y. Merla, B. Wu, V. Yufit, N. P. Brandon, R. F. Martinez-Botas, and G.
286-296, June 2018. J. Offer, "Novel application of differential thermal voltammetry as an
[16] S. Santhanagopalan, Q. Guo, P. Ramadass, and R. E. White, "Review of in-depth state-of-health diagnosis method for lithium-ion batteries,"
models for predicting the cycling performance of lithium ion batteries," Journal of Power Sources, vol. 307, pp. 308-319, March 2016.
Journal of Power Sources, vol. 156, no. 2, pp. 620-628, June 2006. [37] Y. Li, M. Abdel-Monem, R. Gopalakrishnan, M. Berecibar, E.
[17] A. Jokar, B. Rajabloo, M. Désilets, and M. Lacroix, "Review of Nanini-Maury, N. Omar, P. van den Bossche, and J. Van Mierlo, "A
simplified Pseudo-two-Dimensional models of lithium-ion batteries," quick on-line state of health estimation method for Li-ion battery with
Journal of Power Sources, vol. 327, pp. 44-55, September 2016. incremental capacity curves processed by Gaussian filter," Journal of
[18] X. Lin and W. Lu, "A battery model that enables consideration of Power Sources, vol. 373, pp. 40-53, January 2018.
realistic anisotropic environment surrounding an active material [38] M. Berecibar, M. Garmendia, I. Gandiaga, J. Crego, and I. Villarreal,
particle and its application," Journal of Power Sources, vol. 357, pp. "State of health estimation algorithm of LiFePO4 battery packs based
220-229, July 2017. on differential voltage curves for battery management system
[19] X. Lin, J. Park, L. Liu, Y. Lee, A. M. Sastry, and W. Lu, "A application," Energy, vol. 103, pp. 784-796, May 2016.
Comprehensive Capacity Fade Model and Analysis for Li-Ion [39] B. Jiang, H. Dai, X. Wei, and T. Xu, "Joint estimation of lithium-ion
Batteries," Journal of The Electrochemical Society, vol. 160, no. 10, pp. battery state of charge and capacity within an adaptive variable
A1701-A1710, August 2013. multi-timescale framework considering current measurement offset,"
[20] X. Lin, X. Hao, Z. Liu, and W. Jia, "Health conscious fast charging of Applied Energy, vol. 253, November 2019.
Li-ion batteries via a single particle model with aging mechanisms," [40] C. She, Z. Wang, F. Sun, P. Liu, and L. Zhang, "Battery Aging
Journal of Power Sources, vol. 400, pp. 305-316, October 2018. Assessment for Real-World Electric Buses Based on Incremental
[21] A. Eddahech, O. Briat, E. Woirgard, and J. M. Vinassa, "Remaining Capacity Analysis and Radial Basis Function Neural Network," IEEE
useful life prediction of lithium batteries in calendar ageing for Transactions on Industrial Informatics, vol. 16, no. 5, pp. 3345-3354,
automotive applications," Microelectronics Reliability, vol. 52, no. May 2020.
9-10, pp. 2438-2442, September 2012. [41] L. Wang, C. Pan, L. Liu, Y. Cheng, and X. Zhao, "On-board state of
[22] Y. Xing, E. W. M. Ma, K.-L. Tsui, and M. Pecht, "An ensemble model health estimation of LiFePO4 battery pack through differential voltage
for predicting the remaining useful performance of lithium-ion analysis," Applied Energy, vol. 168, pp. 465-472, April 2016.
batteries," Microelectronics Reliability, vol. 53, no. 6, pp. 811-820, [42] C. Sbarufatti, M. Corbetta, M. Giglio, and F. Cadini, "Adaptive
June 2013. prognosis of lithium-ion batteries based on the combination of particle
[23] J. Yang, B. Xia, W. Huang, Y. Fu, and C. Mi, "Online state-of-health filters and radial basis function neural networks," Journal of Power
estimation for lithium-ion batteries using constant-voltage charging Sources, vol. 344, pp. 128-140, March 2017.
 11

[43] Z. Wang, J. Ma, and L. Zhang, "State-of-Health Estimation for

Lithium-Ion Batteries Based on the Multi-Island Genetic Algorithm and Xiaosong Hu (SM’16) received the Ph.D. degree in
the Gaussian Process Regression," IEEE Access, vol. 5, pp. automotive engineering from the Beijing Institute of
21286-21295, October 2017. Technology, Beijing, China, in 2012. He did scientific
[44] R. R. Richardson, C. R. Birkl, M. A. Osborne, and D. A. Howey, research and completed the Ph.D. dissertation in
"Gaussian Process Regression for In Situ Capacity Estimation of Automotive Research Center at the University of
Lithium-Ion Batteries," IEEE Transactions on Industrial Informatics, Michigan, Ann Arbor, MI, USA, between 2010 and 2012.
vol. 15, no. 1, pp. 127-138, January 2019. He is currently a Professor with the State Key
[45] K. Rumpf, M. Naumann, and A. Jossen, "Experimental investigation of Laboratory of Mechanical Transmissions and at the
parametric cell-to-cell variation and correlation based on 1100 Department of Automotive Engineering, Chongqing
commercial lithium-ion cells," Journal of Energy Storage, vol. 14, pp. University, Chongqing, China. He was a Postdoctoral
224-243, December 2017. Researcher with the Department of Civil and Environmental Engineering,
[46] J. Adler and I. Parmryd, "Quantifying Colocalization by Correlation University of California, Berkeley, CA, USA, between 2014 and 2015, as well
The Pearson Correlation Coefficient is Superior to the Mander’s as at the Swedish Hybrid Vehicle Center and the Department of Signals and
Overlap Coefficient," Cytometry Part A, vol. 77(8), pp. 733-742, March Systems at Chalmers University of Technology, Gothenburg, Sweden, between
2010. 2012 and 2014. He was also a Visiting Postdoctoral Researcher with the
[47] J. L. Rodgers and W. A. Nicewander, "Thirteen Ways to Look at the Institute for Dynamic Systems and Control at Swiss Federal Institute of
Correlation Coefficient," The American Statistician, vol. 42(1), pp. Technology (ETH), Zurich, Switzerland, in 2014. His research interests include
59-66, February 2012. modeling and control of alternative powertrains and energy storage systems.
[48] Y. Li, K. Liu, A. M. Foley, A. Zülke, M. Berecibar, E. Nanini-Maury, J. Dr. Hu was the recipient of several prestigious awards/honors, including
Van Mierlo, and H. E. Hoster, "Data-driven health estimation and Emerging Sustainability Leaders Award in 2016, EU Marie Currie Fellowship
lifetime prediction of lithium-ion batteries: A review," Renewable and in 2015, ASME DSCD Energy Systems Best Paper Award in 2015, and Beijing
Sustainable Energy Reviews, vol. 113, 2019. Best Ph.D. Dissertation Award in 2013.
[49] C. E. Rasmussen, Gaussian processes in machine learning (Summer
School on Machine Learning). Springer, 2003, pp. 63-71.
Yunhong Che received the B.E. degree in Automotive
[50] J. Quiñonero-Candela and C. E. Rasmussen, "A Unifying View of
Engineering from Chongqing University, Chongqing,
Sparse Approximate Gaussian Process Regression," Journal of
China, in 2019. He is currently pursuing the M.S. degree in
Machine Learning Research, vol. 6, pp. 1939-1959, December 2005.
state estimation and life prediction of Li-ion batteries at the
[51] R. R. Richardson, M. A. Osborne, and D. A. Howey, "Gaussian process
Vehicle Power System Lab, Department of Automotive
regression for forecasting battery state of health," Journal of Power
Engineering, Chongqing University.
Sources, vol. 357, pp. 209-219, July 2017.
[52] K. V. Mardia and R. J. Marshall, "Maximum likelihood estimation of
models for residual covariance in spatial regression," vol. 71, no. 1, pp.
135-146, April 1984.
[53] R. Xiong, Y. Zhang, J. Wang, H. He, S. Peng, and M. Pecht,
Xianke Lin received his Ph.D. in Mechanical
"Lithium-Ion Battery Health Prognosis Based on a Real Battery
Engineering from the University of Michigan, Ann Arbor,
Management System Used in Electric Vehicles," IEEE Transactions on
in 2014. He previously finished his B.S. from Zhejiang
Vehicular Technology, vol. 68, no. 5, pp. 4110-4121, May 2019.
University, China, in 2009.
He has extensive industrial experience at Fiat Chrysler
Automobiles, Mercedes-Benz R&D North America, and
General Motor of Canada. Currently, he is an Assistant
Professor in the Department of Automotive, Mechanical
and Manufacturing Engineering at the University of
Ontario Institute of Technology, Oshawa, Canada. His
research activities have concentrated on hybrid powertrain design and control
strategy optimization, Multiscale/Multiphysics modeling and optimization of
energy storage systems.

Zhongwei Deng received his Ph.D. in Mechanical

Engineering from the Shanghai Jiao Tong University,
Shanghai, China, in 2019. He previously finished his B.S.
from Jilin University, China, in 2014. He is currently a
postdoctoral researcher in the Department of Automotive
Engineering, Chongqing University, Chongqing, China.
His research interests focus on data-driven and
electrochemical mechanism modeling, parameter
identification, state estimation, health diagnosis, and
cascade utilization of lithium-ion batteries.